• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains the X86 implementation of the TargetInstrInfo class.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "X86InstrInfo.h"
15 #include "X86.h"
16 #include "X86InstrBuilder.h"
17 #include "X86MachineFunctionInfo.h"
18 #include "X86Subtarget.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/CodeGen/LivePhysRegs.h"
22 #include "llvm/CodeGen/LiveVariables.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineDominators.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineModuleInfo.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/StackMaps.h"
30 #include "llvm/IR/DerivedTypes.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/MC/MCAsmInfo.h"
34 #include "llvm/MC/MCExpr.h"
35 #include "llvm/MC/MCInst.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Target/TargetOptions.h"
41 
42 using namespace llvm;
43 
44 #define DEBUG_TYPE "x86-instr-info"
45 
46 #define GET_INSTRINFO_CTOR_DTOR
47 #include "X86GenInstrInfo.inc"
48 
49 static cl::opt<bool>
50 NoFusing("disable-spill-fusing",
51          cl::desc("Disable fusing of spill code into instructions"));
52 static cl::opt<bool>
53 PrintFailedFusing("print-failed-fuse-candidates",
54                   cl::desc("Print instructions that the allocator wants to"
55                            " fuse, but the X86 backend currently can't"),
56                   cl::Hidden);
57 static cl::opt<bool>
58 ReMatPICStubLoad("remat-pic-stub-load",
59                  cl::desc("Re-materialize load from stub in PIC mode"),
60                  cl::init(false), cl::Hidden);
61 static cl::opt<unsigned>
62 PartialRegUpdateClearance("partial-reg-update-clearance",
63                           cl::desc("Clearance between two register writes "
64                                    "for inserting XOR to avoid partial "
65                                    "register update"),
66                           cl::init(64), cl::Hidden);
67 static cl::opt<unsigned>
68 UndefRegClearance("undef-reg-clearance",
69                   cl::desc("How many idle instructions we would like before "
70                            "certain undef register reads"),
71                   cl::init(64), cl::Hidden);
72 
73 enum {
74   // Select which memory operand is being unfolded.
75   // (stored in bits 0 - 3)
76   TB_INDEX_0    = 0,
77   TB_INDEX_1    = 1,
78   TB_INDEX_2    = 2,
79   TB_INDEX_3    = 3,
80   TB_INDEX_4    = 4,
81   TB_INDEX_MASK = 0xf,
82 
83   // Do not insert the reverse map (MemOp -> RegOp) into the table.
84   // This may be needed because there is a many -> one mapping.
85   TB_NO_REVERSE   = 1 << 4,
86 
87   // Do not insert the forward map (RegOp -> MemOp) into the table.
88   // This is needed for Native Client, which prohibits branch
89   // instructions from using a memory operand.
90   TB_NO_FORWARD   = 1 << 5,
91 
92   TB_FOLDED_LOAD  = 1 << 6,
93   TB_FOLDED_STORE = 1 << 7,
94 
95   // Minimum alignment required for load/store.
96   // Used for RegOp->MemOp conversion.
97   // (stored in bits 8 - 15)
98   TB_ALIGN_SHIFT = 8,
99   TB_ALIGN_NONE  =    0 << TB_ALIGN_SHIFT,
100   TB_ALIGN_16    =   16 << TB_ALIGN_SHIFT,
101   TB_ALIGN_32    =   32 << TB_ALIGN_SHIFT,
102   TB_ALIGN_64    =   64 << TB_ALIGN_SHIFT,
103   TB_ALIGN_MASK  = 0xff << TB_ALIGN_SHIFT
104 };
105 
106 struct X86MemoryFoldTableEntry {
107   uint16_t RegOp;
108   uint16_t MemOp;
109   uint16_t Flags;
110 };
111 
112 // Pin the vtable to this file.
anchor()113 void X86InstrInfo::anchor() {}
114 
X86InstrInfo(X86Subtarget & STI)115 X86InstrInfo::X86InstrInfo(X86Subtarget &STI)
116     : X86GenInstrInfo((STI.isTarget64BitLP64() ? X86::ADJCALLSTACKDOWN64
117                                                : X86::ADJCALLSTACKDOWN32),
118                       (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKUP64
119                                                : X86::ADJCALLSTACKUP32),
120                       X86::CATCHRET,
121                       (STI.is64Bit() ? X86::RETQ : X86::RETL)),
122       Subtarget(STI), RI(STI.getTargetTriple()) {
123 
124   static const X86MemoryFoldTableEntry MemoryFoldTable2Addr[] = {
125     { X86::ADC32ri,     X86::ADC32mi,    0 },
126     { X86::ADC32ri8,    X86::ADC32mi8,   0 },
127     { X86::ADC32rr,     X86::ADC32mr,    0 },
128     { X86::ADC64ri32,   X86::ADC64mi32,  0 },
129     { X86::ADC64ri8,    X86::ADC64mi8,   0 },
130     { X86::ADC64rr,     X86::ADC64mr,    0 },
131     { X86::ADD16ri,     X86::ADD16mi,    0 },
132     { X86::ADD16ri8,    X86::ADD16mi8,   0 },
133     { X86::ADD16ri_DB,  X86::ADD16mi,    TB_NO_REVERSE },
134     { X86::ADD16ri8_DB, X86::ADD16mi8,   TB_NO_REVERSE },
135     { X86::ADD16rr,     X86::ADD16mr,    0 },
136     { X86::ADD16rr_DB,  X86::ADD16mr,    TB_NO_REVERSE },
137     { X86::ADD32ri,     X86::ADD32mi,    0 },
138     { X86::ADD32ri8,    X86::ADD32mi8,   0 },
139     { X86::ADD32ri_DB,  X86::ADD32mi,    TB_NO_REVERSE },
140     { X86::ADD32ri8_DB, X86::ADD32mi8,   TB_NO_REVERSE },
141     { X86::ADD32rr,     X86::ADD32mr,    0 },
142     { X86::ADD32rr_DB,  X86::ADD32mr,    TB_NO_REVERSE },
143     { X86::ADD64ri32,   X86::ADD64mi32,  0 },
144     { X86::ADD64ri8,    X86::ADD64mi8,   0 },
145     { X86::ADD64ri32_DB,X86::ADD64mi32,  TB_NO_REVERSE },
146     { X86::ADD64ri8_DB, X86::ADD64mi8,   TB_NO_REVERSE },
147     { X86::ADD64rr,     X86::ADD64mr,    0 },
148     { X86::ADD64rr_DB,  X86::ADD64mr,    TB_NO_REVERSE },
149     { X86::ADD8ri,      X86::ADD8mi,     0 },
150     { X86::ADD8rr,      X86::ADD8mr,     0 },
151     { X86::AND16ri,     X86::AND16mi,    0 },
152     { X86::AND16ri8,    X86::AND16mi8,   0 },
153     { X86::AND16rr,     X86::AND16mr,    0 },
154     { X86::AND32ri,     X86::AND32mi,    0 },
155     { X86::AND32ri8,    X86::AND32mi8,   0 },
156     { X86::AND32rr,     X86::AND32mr,    0 },
157     { X86::AND64ri32,   X86::AND64mi32,  0 },
158     { X86::AND64ri8,    X86::AND64mi8,   0 },
159     { X86::AND64rr,     X86::AND64mr,    0 },
160     { X86::AND8ri,      X86::AND8mi,     0 },
161     { X86::AND8rr,      X86::AND8mr,     0 },
162     { X86::DEC16r,      X86::DEC16m,     0 },
163     { X86::DEC32r,      X86::DEC32m,     0 },
164     { X86::DEC64r,      X86::DEC64m,     0 },
165     { X86::DEC8r,       X86::DEC8m,      0 },
166     { X86::INC16r,      X86::INC16m,     0 },
167     { X86::INC32r,      X86::INC32m,     0 },
168     { X86::INC64r,      X86::INC64m,     0 },
169     { X86::INC8r,       X86::INC8m,      0 },
170     { X86::NEG16r,      X86::NEG16m,     0 },
171     { X86::NEG32r,      X86::NEG32m,     0 },
172     { X86::NEG64r,      X86::NEG64m,     0 },
173     { X86::NEG8r,       X86::NEG8m,      0 },
174     { X86::NOT16r,      X86::NOT16m,     0 },
175     { X86::NOT32r,      X86::NOT32m,     0 },
176     { X86::NOT64r,      X86::NOT64m,     0 },
177     { X86::NOT8r,       X86::NOT8m,      0 },
178     { X86::OR16ri,      X86::OR16mi,     0 },
179     { X86::OR16ri8,     X86::OR16mi8,    0 },
180     { X86::OR16rr,      X86::OR16mr,     0 },
181     { X86::OR32ri,      X86::OR32mi,     0 },
182     { X86::OR32ri8,     X86::OR32mi8,    0 },
183     { X86::OR32rr,      X86::OR32mr,     0 },
184     { X86::OR64ri32,    X86::OR64mi32,   0 },
185     { X86::OR64ri8,     X86::OR64mi8,    0 },
186     { X86::OR64rr,      X86::OR64mr,     0 },
187     { X86::OR8ri,       X86::OR8mi,      0 },
188     { X86::OR8rr,       X86::OR8mr,      0 },
189     { X86::ROL16r1,     X86::ROL16m1,    0 },
190     { X86::ROL16rCL,    X86::ROL16mCL,   0 },
191     { X86::ROL16ri,     X86::ROL16mi,    0 },
192     { X86::ROL32r1,     X86::ROL32m1,    0 },
193     { X86::ROL32rCL,    X86::ROL32mCL,   0 },
194     { X86::ROL32ri,     X86::ROL32mi,    0 },
195     { X86::ROL64r1,     X86::ROL64m1,    0 },
196     { X86::ROL64rCL,    X86::ROL64mCL,   0 },
197     { X86::ROL64ri,     X86::ROL64mi,    0 },
198     { X86::ROL8r1,      X86::ROL8m1,     0 },
199     { X86::ROL8rCL,     X86::ROL8mCL,    0 },
200     { X86::ROL8ri,      X86::ROL8mi,     0 },
201     { X86::ROR16r1,     X86::ROR16m1,    0 },
202     { X86::ROR16rCL,    X86::ROR16mCL,   0 },
203     { X86::ROR16ri,     X86::ROR16mi,    0 },
204     { X86::ROR32r1,     X86::ROR32m1,    0 },
205     { X86::ROR32rCL,    X86::ROR32mCL,   0 },
206     { X86::ROR32ri,     X86::ROR32mi,    0 },
207     { X86::ROR64r1,     X86::ROR64m1,    0 },
208     { X86::ROR64rCL,    X86::ROR64mCL,   0 },
209     { X86::ROR64ri,     X86::ROR64mi,    0 },
210     { X86::ROR8r1,      X86::ROR8m1,     0 },
211     { X86::ROR8rCL,     X86::ROR8mCL,    0 },
212     { X86::ROR8ri,      X86::ROR8mi,     0 },
213     { X86::SAR16r1,     X86::SAR16m1,    0 },
214     { X86::SAR16rCL,    X86::SAR16mCL,   0 },
215     { X86::SAR16ri,     X86::SAR16mi,    0 },
216     { X86::SAR32r1,     X86::SAR32m1,    0 },
217     { X86::SAR32rCL,    X86::SAR32mCL,   0 },
218     { X86::SAR32ri,     X86::SAR32mi,    0 },
219     { X86::SAR64r1,     X86::SAR64m1,    0 },
220     { X86::SAR64rCL,    X86::SAR64mCL,   0 },
221     { X86::SAR64ri,     X86::SAR64mi,    0 },
222     { X86::SAR8r1,      X86::SAR8m1,     0 },
223     { X86::SAR8rCL,     X86::SAR8mCL,    0 },
224     { X86::SAR8ri,      X86::SAR8mi,     0 },
225     { X86::SBB32ri,     X86::SBB32mi,    0 },
226     { X86::SBB32ri8,    X86::SBB32mi8,   0 },
227     { X86::SBB32rr,     X86::SBB32mr,    0 },
228     { X86::SBB64ri32,   X86::SBB64mi32,  0 },
229     { X86::SBB64ri8,    X86::SBB64mi8,   0 },
230     { X86::SBB64rr,     X86::SBB64mr,    0 },
231     { X86::SHL16rCL,    X86::SHL16mCL,   0 },
232     { X86::SHL16ri,     X86::SHL16mi,    0 },
233     { X86::SHL32rCL,    X86::SHL32mCL,   0 },
234     { X86::SHL32ri,     X86::SHL32mi,    0 },
235     { X86::SHL64rCL,    X86::SHL64mCL,   0 },
236     { X86::SHL64ri,     X86::SHL64mi,    0 },
237     { X86::SHL8rCL,     X86::SHL8mCL,    0 },
238     { X86::SHL8ri,      X86::SHL8mi,     0 },
239     { X86::SHLD16rrCL,  X86::SHLD16mrCL, 0 },
240     { X86::SHLD16rri8,  X86::SHLD16mri8, 0 },
241     { X86::SHLD32rrCL,  X86::SHLD32mrCL, 0 },
242     { X86::SHLD32rri8,  X86::SHLD32mri8, 0 },
243     { X86::SHLD64rrCL,  X86::SHLD64mrCL, 0 },
244     { X86::SHLD64rri8,  X86::SHLD64mri8, 0 },
245     { X86::SHR16r1,     X86::SHR16m1,    0 },
246     { X86::SHR16rCL,    X86::SHR16mCL,   0 },
247     { X86::SHR16ri,     X86::SHR16mi,    0 },
248     { X86::SHR32r1,     X86::SHR32m1,    0 },
249     { X86::SHR32rCL,    X86::SHR32mCL,   0 },
250     { X86::SHR32ri,     X86::SHR32mi,    0 },
251     { X86::SHR64r1,     X86::SHR64m1,    0 },
252     { X86::SHR64rCL,    X86::SHR64mCL,   0 },
253     { X86::SHR64ri,     X86::SHR64mi,    0 },
254     { X86::SHR8r1,      X86::SHR8m1,     0 },
255     { X86::SHR8rCL,     X86::SHR8mCL,    0 },
256     { X86::SHR8ri,      X86::SHR8mi,     0 },
257     { X86::SHRD16rrCL,  X86::SHRD16mrCL, 0 },
258     { X86::SHRD16rri8,  X86::SHRD16mri8, 0 },
259     { X86::SHRD32rrCL,  X86::SHRD32mrCL, 0 },
260     { X86::SHRD32rri8,  X86::SHRD32mri8, 0 },
261     { X86::SHRD64rrCL,  X86::SHRD64mrCL, 0 },
262     { X86::SHRD64rri8,  X86::SHRD64mri8, 0 },
263     { X86::SUB16ri,     X86::SUB16mi,    0 },
264     { X86::SUB16ri8,    X86::SUB16mi8,   0 },
265     { X86::SUB16rr,     X86::SUB16mr,    0 },
266     { X86::SUB32ri,     X86::SUB32mi,    0 },
267     { X86::SUB32ri8,    X86::SUB32mi8,   0 },
268     { X86::SUB32rr,     X86::SUB32mr,    0 },
269     { X86::SUB64ri32,   X86::SUB64mi32,  0 },
270     { X86::SUB64ri8,    X86::SUB64mi8,   0 },
271     { X86::SUB64rr,     X86::SUB64mr,    0 },
272     { X86::SUB8ri,      X86::SUB8mi,     0 },
273     { X86::SUB8rr,      X86::SUB8mr,     0 },
274     { X86::XOR16ri,     X86::XOR16mi,    0 },
275     { X86::XOR16ri8,    X86::XOR16mi8,   0 },
276     { X86::XOR16rr,     X86::XOR16mr,    0 },
277     { X86::XOR32ri,     X86::XOR32mi,    0 },
278     { X86::XOR32ri8,    X86::XOR32mi8,   0 },
279     { X86::XOR32rr,     X86::XOR32mr,    0 },
280     { X86::XOR64ri32,   X86::XOR64mi32,  0 },
281     { X86::XOR64ri8,    X86::XOR64mi8,   0 },
282     { X86::XOR64rr,     X86::XOR64mr,    0 },
283     { X86::XOR8ri,      X86::XOR8mi,     0 },
284     { X86::XOR8rr,      X86::XOR8mr,     0 }
285   };
286 
287   for (X86MemoryFoldTableEntry Entry : MemoryFoldTable2Addr) {
288     AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
289                   Entry.RegOp, Entry.MemOp,
290                   // Index 0, folded load and store, no alignment requirement.
291                   Entry.Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
292   }
293 
294   static const X86MemoryFoldTableEntry MemoryFoldTable0[] = {
295     { X86::BT16ri8,     X86::BT16mi8,       TB_FOLDED_LOAD },
296     { X86::BT32ri8,     X86::BT32mi8,       TB_FOLDED_LOAD },
297     { X86::BT64ri8,     X86::BT64mi8,       TB_FOLDED_LOAD },
298     { X86::CALL32r,     X86::CALL32m,       TB_FOLDED_LOAD },
299     { X86::CALL64r,     X86::CALL64m,       TB_FOLDED_LOAD },
300     { X86::CMP16ri,     X86::CMP16mi,       TB_FOLDED_LOAD },
301     { X86::CMP16ri8,    X86::CMP16mi8,      TB_FOLDED_LOAD },
302     { X86::CMP16rr,     X86::CMP16mr,       TB_FOLDED_LOAD },
303     { X86::CMP32ri,     X86::CMP32mi,       TB_FOLDED_LOAD },
304     { X86::CMP32ri8,    X86::CMP32mi8,      TB_FOLDED_LOAD },
305     { X86::CMP32rr,     X86::CMP32mr,       TB_FOLDED_LOAD },
306     { X86::CMP64ri32,   X86::CMP64mi32,     TB_FOLDED_LOAD },
307     { X86::CMP64ri8,    X86::CMP64mi8,      TB_FOLDED_LOAD },
308     { X86::CMP64rr,     X86::CMP64mr,       TB_FOLDED_LOAD },
309     { X86::CMP8ri,      X86::CMP8mi,        TB_FOLDED_LOAD },
310     { X86::CMP8rr,      X86::CMP8mr,        TB_FOLDED_LOAD },
311     { X86::DIV16r,      X86::DIV16m,        TB_FOLDED_LOAD },
312     { X86::DIV32r,      X86::DIV32m,        TB_FOLDED_LOAD },
313     { X86::DIV64r,      X86::DIV64m,        TB_FOLDED_LOAD },
314     { X86::DIV8r,       X86::DIV8m,         TB_FOLDED_LOAD },
315     { X86::EXTRACTPSrr, X86::EXTRACTPSmr,   TB_FOLDED_STORE },
316     { X86::IDIV16r,     X86::IDIV16m,       TB_FOLDED_LOAD },
317     { X86::IDIV32r,     X86::IDIV32m,       TB_FOLDED_LOAD },
318     { X86::IDIV64r,     X86::IDIV64m,       TB_FOLDED_LOAD },
319     { X86::IDIV8r,      X86::IDIV8m,        TB_FOLDED_LOAD },
320     { X86::IMUL16r,     X86::IMUL16m,       TB_FOLDED_LOAD },
321     { X86::IMUL32r,     X86::IMUL32m,       TB_FOLDED_LOAD },
322     { X86::IMUL64r,     X86::IMUL64m,       TB_FOLDED_LOAD },
323     { X86::IMUL8r,      X86::IMUL8m,        TB_FOLDED_LOAD },
324     { X86::JMP32r,      X86::JMP32m,        TB_FOLDED_LOAD },
325     { X86::JMP64r,      X86::JMP64m,        TB_FOLDED_LOAD },
326     { X86::MOV16ri,     X86::MOV16mi,       TB_FOLDED_STORE },
327     { X86::MOV16rr,     X86::MOV16mr,       TB_FOLDED_STORE },
328     { X86::MOV32ri,     X86::MOV32mi,       TB_FOLDED_STORE },
329     { X86::MOV32rr,     X86::MOV32mr,       TB_FOLDED_STORE },
330     { X86::MOV64ri32,   X86::MOV64mi32,     TB_FOLDED_STORE },
331     { X86::MOV64rr,     X86::MOV64mr,       TB_FOLDED_STORE },
332     { X86::MOV8ri,      X86::MOV8mi,        TB_FOLDED_STORE },
333     { X86::MOV8rr,      X86::MOV8mr,        TB_FOLDED_STORE },
334     { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
335     { X86::MOVAPDrr,    X86::MOVAPDmr,      TB_FOLDED_STORE | TB_ALIGN_16 },
336     { X86::MOVAPSrr,    X86::MOVAPSmr,      TB_FOLDED_STORE | TB_ALIGN_16 },
337     { X86::MOVDQArr,    X86::MOVDQAmr,      TB_FOLDED_STORE | TB_ALIGN_16 },
338     { X86::MOVPDI2DIrr, X86::MOVPDI2DImr,   TB_FOLDED_STORE },
339     { X86::MOVPQIto64rr,X86::MOVPQI2QImr,   TB_FOLDED_STORE },
340     { X86::MOVSDto64rr, X86::MOVSDto64mr,   TB_FOLDED_STORE },
341     { X86::MOVSS2DIrr,  X86::MOVSS2DImr,    TB_FOLDED_STORE },
342     { X86::MOVUPDrr,    X86::MOVUPDmr,      TB_FOLDED_STORE },
343     { X86::MOVUPSrr,    X86::MOVUPSmr,      TB_FOLDED_STORE },
344     { X86::MUL16r,      X86::MUL16m,        TB_FOLDED_LOAD },
345     { X86::MUL32r,      X86::MUL32m,        TB_FOLDED_LOAD },
346     { X86::MUL64r,      X86::MUL64m,        TB_FOLDED_LOAD },
347     { X86::MUL8r,       X86::MUL8m,         TB_FOLDED_LOAD },
348     { X86::PEXTRDrr,    X86::PEXTRDmr,      TB_FOLDED_STORE },
349     { X86::PEXTRQrr,    X86::PEXTRQmr,      TB_FOLDED_STORE },
350     { X86::PUSH16r,     X86::PUSH16rmm,     TB_FOLDED_LOAD },
351     { X86::PUSH32r,     X86::PUSH32rmm,     TB_FOLDED_LOAD },
352     { X86::PUSH64r,     X86::PUSH64rmm,     TB_FOLDED_LOAD },
353     { X86::SETAEr,      X86::SETAEm,        TB_FOLDED_STORE },
354     { X86::SETAr,       X86::SETAm,         TB_FOLDED_STORE },
355     { X86::SETBEr,      X86::SETBEm,        TB_FOLDED_STORE },
356     { X86::SETBr,       X86::SETBm,         TB_FOLDED_STORE },
357     { X86::SETEr,       X86::SETEm,         TB_FOLDED_STORE },
358     { X86::SETGEr,      X86::SETGEm,        TB_FOLDED_STORE },
359     { X86::SETGr,       X86::SETGm,         TB_FOLDED_STORE },
360     { X86::SETLEr,      X86::SETLEm,        TB_FOLDED_STORE },
361     { X86::SETLr,       X86::SETLm,         TB_FOLDED_STORE },
362     { X86::SETNEr,      X86::SETNEm,        TB_FOLDED_STORE },
363     { X86::SETNOr,      X86::SETNOm,        TB_FOLDED_STORE },
364     { X86::SETNPr,      X86::SETNPm,        TB_FOLDED_STORE },
365     { X86::SETNSr,      X86::SETNSm,        TB_FOLDED_STORE },
366     { X86::SETOr,       X86::SETOm,         TB_FOLDED_STORE },
367     { X86::SETPr,       X86::SETPm,         TB_FOLDED_STORE },
368     { X86::SETSr,       X86::SETSm,         TB_FOLDED_STORE },
369     { X86::TAILJMPr,    X86::TAILJMPm,      TB_FOLDED_LOAD },
370     { X86::TAILJMPr64,  X86::TAILJMPm64,    TB_FOLDED_LOAD },
371     { X86::TAILJMPr64_REX, X86::TAILJMPm64_REX, TB_FOLDED_LOAD },
372     { X86::TEST16ri,    X86::TEST16mi,      TB_FOLDED_LOAD },
373     { X86::TEST32ri,    X86::TEST32mi,      TB_FOLDED_LOAD },
374     { X86::TEST64ri32,  X86::TEST64mi32,    TB_FOLDED_LOAD },
375     { X86::TEST8ri,     X86::TEST8mi,       TB_FOLDED_LOAD },
376 
377     // AVX 128-bit versions of foldable instructions
378     { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr,  TB_FOLDED_STORE  },
379     { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
380     { X86::VMOVAPDrr,   X86::VMOVAPDmr,     TB_FOLDED_STORE | TB_ALIGN_16 },
381     { X86::VMOVAPSrr,   X86::VMOVAPSmr,     TB_FOLDED_STORE | TB_ALIGN_16 },
382     { X86::VMOVDQArr,   X86::VMOVDQAmr,     TB_FOLDED_STORE | TB_ALIGN_16 },
383     { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr,  TB_FOLDED_STORE },
384     { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
385     { X86::VMOVSDto64rr,X86::VMOVSDto64mr,  TB_FOLDED_STORE },
386     { X86::VMOVSS2DIrr, X86::VMOVSS2DImr,   TB_FOLDED_STORE },
387     { X86::VMOVUPDrr,   X86::VMOVUPDmr,     TB_FOLDED_STORE },
388     { X86::VMOVUPSrr,   X86::VMOVUPSmr,     TB_FOLDED_STORE },
389     { X86::VPEXTRDrr,   X86::VPEXTRDmr,     TB_FOLDED_STORE },
390     { X86::VPEXTRQrr,   X86::VPEXTRQmr,     TB_FOLDED_STORE },
391 
392     // AVX 256-bit foldable instructions
393     { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
394     { X86::VMOVAPDYrr,  X86::VMOVAPDYmr,    TB_FOLDED_STORE | TB_ALIGN_32 },
395     { X86::VMOVAPSYrr,  X86::VMOVAPSYmr,    TB_FOLDED_STORE | TB_ALIGN_32 },
396     { X86::VMOVDQAYrr,  X86::VMOVDQAYmr,    TB_FOLDED_STORE | TB_ALIGN_32 },
397     { X86::VMOVUPDYrr,  X86::VMOVUPDYmr,    TB_FOLDED_STORE },
398     { X86::VMOVUPSYrr,  X86::VMOVUPSYmr,    TB_FOLDED_STORE },
399 
400     // AVX-512 foldable instructions
401     { X86::VMOVPDI2DIZrr,   X86::VMOVPDI2DIZmr, TB_FOLDED_STORE },
402     { X86::VMOVAPDZrr,      X86::VMOVAPDZmr,    TB_FOLDED_STORE | TB_ALIGN_64 },
403     { X86::VMOVAPSZrr,      X86::VMOVAPSZmr,    TB_FOLDED_STORE | TB_ALIGN_64 },
404     { X86::VMOVDQA32Zrr,    X86::VMOVDQA32Zmr,  TB_FOLDED_STORE | TB_ALIGN_64 },
405     { X86::VMOVDQA64Zrr,    X86::VMOVDQA64Zmr,  TB_FOLDED_STORE | TB_ALIGN_64 },
406     { X86::VMOVUPDZrr,      X86::VMOVUPDZmr,    TB_FOLDED_STORE },
407     { X86::VMOVUPSZrr,      X86::VMOVUPSZmr,    TB_FOLDED_STORE },
408     { X86::VMOVDQU8Zrr,     X86::VMOVDQU8Zmr,   TB_FOLDED_STORE },
409     { X86::VMOVDQU16Zrr,    X86::VMOVDQU16Zmr,  TB_FOLDED_STORE },
410     { X86::VMOVDQU32Zrr,    X86::VMOVDQU32Zmr,  TB_FOLDED_STORE },
411     { X86::VMOVDQU64Zrr,    X86::VMOVDQU64Zmr,  TB_FOLDED_STORE },
412 
413     // AVX-512 foldable instructions (256-bit versions)
414     { X86::VMOVAPDZ256rr,      X86::VMOVAPDZ256mr,    TB_FOLDED_STORE | TB_ALIGN_32 },
415     { X86::VMOVAPSZ256rr,      X86::VMOVAPSZ256mr,    TB_FOLDED_STORE | TB_ALIGN_32 },
416     { X86::VMOVDQA32Z256rr,    X86::VMOVDQA32Z256mr,  TB_FOLDED_STORE | TB_ALIGN_32 },
417     { X86::VMOVDQA64Z256rr,    X86::VMOVDQA64Z256mr,  TB_FOLDED_STORE | TB_ALIGN_32 },
418     { X86::VMOVUPDZ256rr,      X86::VMOVUPDZ256mr,    TB_FOLDED_STORE },
419     { X86::VMOVUPSZ256rr,      X86::VMOVUPSZ256mr,    TB_FOLDED_STORE },
420     { X86::VMOVDQU8Z256rr,     X86::VMOVDQU8Z256mr,   TB_FOLDED_STORE },
421     { X86::VMOVDQU16Z256rr,    X86::VMOVDQU16Z256mr,  TB_FOLDED_STORE },
422     { X86::VMOVDQU32Z256rr,    X86::VMOVDQU32Z256mr,  TB_FOLDED_STORE },
423     { X86::VMOVDQU64Z256rr,    X86::VMOVDQU64Z256mr,  TB_FOLDED_STORE },
424 
425     // AVX-512 foldable instructions (128-bit versions)
426     { X86::VMOVAPDZ128rr,      X86::VMOVAPDZ128mr,    TB_FOLDED_STORE | TB_ALIGN_16 },
427     { X86::VMOVAPSZ128rr,      X86::VMOVAPSZ128mr,    TB_FOLDED_STORE | TB_ALIGN_16 },
428     { X86::VMOVDQA32Z128rr,    X86::VMOVDQA32Z128mr,  TB_FOLDED_STORE | TB_ALIGN_16 },
429     { X86::VMOVDQA64Z128rr,    X86::VMOVDQA64Z128mr,  TB_FOLDED_STORE | TB_ALIGN_16 },
430     { X86::VMOVUPDZ128rr,      X86::VMOVUPDZ128mr,    TB_FOLDED_STORE },
431     { X86::VMOVUPSZ128rr,      X86::VMOVUPSZ128mr,    TB_FOLDED_STORE },
432     { X86::VMOVDQU8Z128rr,     X86::VMOVDQU8Z128mr,   TB_FOLDED_STORE },
433     { X86::VMOVDQU16Z128rr,    X86::VMOVDQU16Z128mr,  TB_FOLDED_STORE },
434     { X86::VMOVDQU32Z128rr,    X86::VMOVDQU32Z128mr,  TB_FOLDED_STORE },
435     { X86::VMOVDQU64Z128rr,    X86::VMOVDQU64Z128mr,  TB_FOLDED_STORE },
436 
437     // F16C foldable instructions
438     { X86::VCVTPS2PHrr,        X86::VCVTPS2PHmr,      TB_FOLDED_STORE },
439     { X86::VCVTPS2PHYrr,       X86::VCVTPS2PHYmr,     TB_FOLDED_STORE }
440   };
441 
442   for (X86MemoryFoldTableEntry Entry : MemoryFoldTable0) {
443     AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
444                   Entry.RegOp, Entry.MemOp, TB_INDEX_0 | Entry.Flags);
445   }
446 
447   static const X86MemoryFoldTableEntry MemoryFoldTable1[] = {
448     { X86::BSF16rr,         X86::BSF16rm,             0 },
449     { X86::BSF32rr,         X86::BSF32rm,             0 },
450     { X86::BSF64rr,         X86::BSF64rm,             0 },
451     { X86::BSR16rr,         X86::BSR16rm,             0 },
452     { X86::BSR32rr,         X86::BSR32rm,             0 },
453     { X86::BSR64rr,         X86::BSR64rm,             0 },
454     { X86::CMP16rr,         X86::CMP16rm,             0 },
455     { X86::CMP32rr,         X86::CMP32rm,             0 },
456     { X86::CMP64rr,         X86::CMP64rm,             0 },
457     { X86::CMP8rr,          X86::CMP8rm,              0 },
458     { X86::CVTSD2SSrr,      X86::CVTSD2SSrm,          0 },
459     { X86::CVTSI2SD64rr,    X86::CVTSI2SD64rm,        0 },
460     { X86::CVTSI2SDrr,      X86::CVTSI2SDrm,          0 },
461     { X86::CVTSI2SS64rr,    X86::CVTSI2SS64rm,        0 },
462     { X86::CVTSI2SSrr,      X86::CVTSI2SSrm,          0 },
463     { X86::CVTSS2SDrr,      X86::CVTSS2SDrm,          0 },
464     { X86::CVTTSD2SI64rr,   X86::CVTTSD2SI64rm,       0 },
465     { X86::CVTTSD2SIrr,     X86::CVTTSD2SIrm,         0 },
466     { X86::CVTTSS2SI64rr,   X86::CVTTSS2SI64rm,       0 },
467     { X86::CVTTSS2SIrr,     X86::CVTTSS2SIrm,         0 },
468     { X86::IMUL16rri,       X86::IMUL16rmi,           0 },
469     { X86::IMUL16rri8,      X86::IMUL16rmi8,          0 },
470     { X86::IMUL32rri,       X86::IMUL32rmi,           0 },
471     { X86::IMUL32rri8,      X86::IMUL32rmi8,          0 },
472     { X86::IMUL64rri32,     X86::IMUL64rmi32,         0 },
473     { X86::IMUL64rri8,      X86::IMUL64rmi8,          0 },
474     { X86::Int_COMISDrr,    X86::Int_COMISDrm,        0 },
475     { X86::Int_COMISSrr,    X86::Int_COMISSrm,        0 },
476     { X86::CVTSD2SI64rr,    X86::CVTSD2SI64rm,        0 },
477     { X86::CVTSD2SIrr,      X86::CVTSD2SIrm,          0 },
478     { X86::CVTSS2SI64rr,    X86::CVTSS2SI64rm,        0 },
479     { X86::CVTSS2SIrr,      X86::CVTSS2SIrm,          0 },
480     { X86::CVTDQ2PDrr,      X86::CVTDQ2PDrm,          TB_ALIGN_16 },
481     { X86::CVTDQ2PSrr,      X86::CVTDQ2PSrm,          TB_ALIGN_16 },
482     { X86::CVTPD2DQrr,      X86::CVTPD2DQrm,          TB_ALIGN_16 },
483     { X86::CVTPD2PSrr,      X86::CVTPD2PSrm,          TB_ALIGN_16 },
484     { X86::CVTPS2DQrr,      X86::CVTPS2DQrm,          TB_ALIGN_16 },
485     { X86::CVTPS2PDrr,      X86::CVTPS2PDrm,          TB_ALIGN_16 },
486     { X86::CVTTPD2DQrr,     X86::CVTTPD2DQrm,         TB_ALIGN_16 },
487     { X86::CVTTPS2DQrr,     X86::CVTTPS2DQrm,         TB_ALIGN_16 },
488     { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm,  0 },
489     { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm,     0 },
490     { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm,  0 },
491     { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm,     0 },
492     { X86::Int_UCOMISDrr,   X86::Int_UCOMISDrm,       0 },
493     { X86::Int_UCOMISSrr,   X86::Int_UCOMISSrm,       0 },
494     { X86::MOV16rr,         X86::MOV16rm,             0 },
495     { X86::MOV32rr,         X86::MOV32rm,             0 },
496     { X86::MOV64rr,         X86::MOV64rm,             0 },
497     { X86::MOV64toPQIrr,    X86::MOVQI2PQIrm,         0 },
498     { X86::MOV64toSDrr,     X86::MOV64toSDrm,         0 },
499     { X86::MOV8rr,          X86::MOV8rm,              0 },
500     { X86::MOVAPDrr,        X86::MOVAPDrm,            TB_ALIGN_16 },
501     { X86::MOVAPSrr,        X86::MOVAPSrm,            TB_ALIGN_16 },
502     { X86::MOVDDUPrr,       X86::MOVDDUPrm,           0 },
503     { X86::MOVDI2PDIrr,     X86::MOVDI2PDIrm,         0 },
504     { X86::MOVDI2SSrr,      X86::MOVDI2SSrm,          0 },
505     { X86::MOVDQArr,        X86::MOVDQArm,            TB_ALIGN_16 },
506     { X86::MOVSHDUPrr,      X86::MOVSHDUPrm,          TB_ALIGN_16 },
507     { X86::MOVSLDUPrr,      X86::MOVSLDUPrm,          TB_ALIGN_16 },
508     { X86::MOVSX16rr8,      X86::MOVSX16rm8,          0 },
509     { X86::MOVSX32rr16,     X86::MOVSX32rm16,         0 },
510     { X86::MOVSX32rr8,      X86::MOVSX32rm8,          0 },
511     { X86::MOVSX64rr16,     X86::MOVSX64rm16,         0 },
512     { X86::MOVSX64rr32,     X86::MOVSX64rm32,         0 },
513     { X86::MOVSX64rr8,      X86::MOVSX64rm8,          0 },
514     { X86::MOVUPDrr,        X86::MOVUPDrm,            TB_ALIGN_16 },
515     { X86::MOVUPSrr,        X86::MOVUPSrm,            0 },
516     { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm,     TB_ALIGN_16 },
517     { X86::MOVZX16rr8,      X86::MOVZX16rm8,          0 },
518     { X86::MOVZX32rr16,     X86::MOVZX32rm16,         0 },
519     { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8,   0 },
520     { X86::MOVZX32rr8,      X86::MOVZX32rm8,          0 },
521     { X86::PABSBrr128,      X86::PABSBrm128,          TB_ALIGN_16 },
522     { X86::PABSDrr128,      X86::PABSDrm128,          TB_ALIGN_16 },
523     { X86::PABSWrr128,      X86::PABSWrm128,          TB_ALIGN_16 },
524     { X86::PCMPESTRIrr,     X86::PCMPESTRIrm,         TB_ALIGN_16 },
525     { X86::PCMPESTRM128rr,  X86::PCMPESTRM128rm,      TB_ALIGN_16 },
526     { X86::PCMPISTRIrr,     X86::PCMPISTRIrm,         TB_ALIGN_16 },
527     { X86::PCMPISTRM128rr,  X86::PCMPISTRM128rm,      TB_ALIGN_16 },
528     { X86::PHMINPOSUWrr128, X86::PHMINPOSUWrm128,     TB_ALIGN_16 },
529     { X86::PMOVSXBDrr,      X86::PMOVSXBDrm,          TB_ALIGN_16 },
530     { X86::PMOVSXBQrr,      X86::PMOVSXBQrm,          TB_ALIGN_16 },
531     { X86::PMOVSXBWrr,      X86::PMOVSXBWrm,          TB_ALIGN_16 },
532     { X86::PMOVSXDQrr,      X86::PMOVSXDQrm,          TB_ALIGN_16 },
533     { X86::PMOVSXWDrr,      X86::PMOVSXWDrm,          TB_ALIGN_16 },
534     { X86::PMOVSXWQrr,      X86::PMOVSXWQrm,          TB_ALIGN_16 },
535     { X86::PMOVZXBDrr,      X86::PMOVZXBDrm,          TB_ALIGN_16 },
536     { X86::PMOVZXBQrr,      X86::PMOVZXBQrm,          TB_ALIGN_16 },
537     { X86::PMOVZXBWrr,      X86::PMOVZXBWrm,          TB_ALIGN_16 },
538     { X86::PMOVZXDQrr,      X86::PMOVZXDQrm,          TB_ALIGN_16 },
539     { X86::PMOVZXWDrr,      X86::PMOVZXWDrm,          TB_ALIGN_16 },
540     { X86::PMOVZXWQrr,      X86::PMOVZXWQrm,          TB_ALIGN_16 },
541     { X86::PSHUFDri,        X86::PSHUFDmi,            TB_ALIGN_16 },
542     { X86::PSHUFHWri,       X86::PSHUFHWmi,           TB_ALIGN_16 },
543     { X86::PSHUFLWri,       X86::PSHUFLWmi,           TB_ALIGN_16 },
544     { X86::PTESTrr,         X86::PTESTrm,             TB_ALIGN_16 },
545     { X86::RCPPSr,          X86::RCPPSm,              TB_ALIGN_16 },
546     { X86::RCPSSr,          X86::RCPSSm,              0 },
547     { X86::RCPSSr_Int,      X86::RCPSSm_Int,          0 },
548     { X86::ROUNDPDr,        X86::ROUNDPDm,            TB_ALIGN_16 },
549     { X86::ROUNDPSr,        X86::ROUNDPSm,            TB_ALIGN_16 },
550     { X86::RSQRTPSr,        X86::RSQRTPSm,            TB_ALIGN_16 },
551     { X86::RSQRTSSr,        X86::RSQRTSSm,            0 },
552     { X86::RSQRTSSr_Int,    X86::RSQRTSSm_Int,        0 },
553     { X86::SQRTPDr,         X86::SQRTPDm,             TB_ALIGN_16 },
554     { X86::SQRTPSr,         X86::SQRTPSm,             TB_ALIGN_16 },
555     { X86::SQRTSDr,         X86::SQRTSDm,             0 },
556     { X86::SQRTSDr_Int,     X86::SQRTSDm_Int,         0 },
557     { X86::SQRTSSr,         X86::SQRTSSm,             0 },
558     { X86::SQRTSSr_Int,     X86::SQRTSSm_Int,         0 },
559     { X86::TEST16rr,        X86::TEST16rm,            0 },
560     { X86::TEST32rr,        X86::TEST32rm,            0 },
561     { X86::TEST64rr,        X86::TEST64rm,            0 },
562     { X86::TEST8rr,         X86::TEST8rm,             0 },
563     // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
564     { X86::UCOMISDrr,       X86::UCOMISDrm,           0 },
565     { X86::UCOMISSrr,       X86::UCOMISSrm,           0 },
566 
567     // MMX version of foldable instructions
568     { X86::MMX_CVTPD2PIirr,   X86::MMX_CVTPD2PIirm,   0 },
569     { X86::MMX_CVTPI2PDirr,   X86::MMX_CVTPI2PDirm,   0 },
570     { X86::MMX_CVTPS2PIirr,   X86::MMX_CVTPS2PIirm,   0 },
571     { X86::MMX_CVTTPD2PIirr,  X86::MMX_CVTTPD2PIirm,  0 },
572     { X86::MMX_CVTTPS2PIirr,  X86::MMX_CVTTPS2PIirm,  0 },
573     { X86::MMX_MOVD64to64rr,  X86::MMX_MOVQ64rm,      0 },
574     { X86::MMX_PABSBrr64,     X86::MMX_PABSBrm64,     0 },
575     { X86::MMX_PABSDrr64,     X86::MMX_PABSDrm64,     0 },
576     { X86::MMX_PABSWrr64,     X86::MMX_PABSWrm64,     0 },
577     { X86::MMX_PSHUFWri,      X86::MMX_PSHUFWmi,      0 },
578 
579     // 3DNow! version of foldable instructions
580     { X86::PF2IDrr,         X86::PF2IDrm,             0 },
581     { X86::PF2IWrr,         X86::PF2IWrm,             0 },
582     { X86::PFRCPrr,         X86::PFRCPrm,             0 },
583     { X86::PFRSQRTrr,       X86::PFRSQRTrm,           0 },
584     { X86::PI2FDrr,         X86::PI2FDrm,             0 },
585     { X86::PI2FWrr,         X86::PI2FWrm,             0 },
586     { X86::PSWAPDrr,        X86::PSWAPDrm,            0 },
587 
588     // AVX 128-bit versions of foldable instructions
589     { X86::Int_VCOMISDrr,   X86::Int_VCOMISDrm,       0 },
590     { X86::Int_VCOMISSrr,   X86::Int_VCOMISSrm,       0 },
591     { X86::Int_VUCOMISDrr,  X86::Int_VUCOMISDrm,      0 },
592     { X86::Int_VUCOMISSrr,  X86::Int_VUCOMISSrm,      0 },
593     { X86::VCVTTSD2SI64rr,  X86::VCVTTSD2SI64rm,      0 },
594     { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
595     { X86::VCVTTSD2SIrr,    X86::VCVTTSD2SIrm,        0 },
596     { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm,    0 },
597     { X86::VCVTTSS2SI64rr,  X86::VCVTTSS2SI64rm,      0 },
598     { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
599     { X86::VCVTTSS2SIrr,    X86::VCVTTSS2SIrm,        0 },
600     { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm,    0 },
601     { X86::VCVTSD2SI64rr,   X86::VCVTSD2SI64rm,       0 },
602     { X86::VCVTSD2SIrr,     X86::VCVTSD2SIrm,         0 },
603     { X86::VCVTSS2SI64rr,   X86::VCVTSS2SI64rm,       0 },
604     { X86::VCVTSS2SIrr,     X86::VCVTSS2SIrm,         0 },
605     { X86::VCVTDQ2PDrr,     X86::VCVTDQ2PDrm,         0 },
606     { X86::VCVTDQ2PSrr,     X86::VCVTDQ2PSrm,         0 },
607     { X86::VCVTPD2DQrr,     X86::VCVTPD2DQXrm,        0 },
608     { X86::VCVTPD2PSrr,     X86::VCVTPD2PSXrm,        0 },
609     { X86::VCVTPS2DQrr,     X86::VCVTPS2DQrm,         0 },
610     { X86::VCVTPS2PDrr,     X86::VCVTPS2PDrm,         0 },
611     { X86::VCVTTPD2DQrr,    X86::VCVTTPD2DQXrm,       0 },
612     { X86::VCVTTPS2DQrr,    X86::VCVTTPS2DQrm,        0 },
613     { X86::VMOV64toPQIrr,   X86::VMOVQI2PQIrm,        0 },
614     { X86::VMOV64toSDrr,    X86::VMOV64toSDrm,        0 },
615     { X86::VMOVAPDrr,       X86::VMOVAPDrm,           TB_ALIGN_16 },
616     { X86::VMOVAPSrr,       X86::VMOVAPSrm,           TB_ALIGN_16 },
617     { X86::VMOVDDUPrr,      X86::VMOVDDUPrm,          0 },
618     { X86::VMOVDI2PDIrr,    X86::VMOVDI2PDIrm,        0 },
619     { X86::VMOVDI2SSrr,     X86::VMOVDI2SSrm,         0 },
620     { X86::VMOVDQArr,       X86::VMOVDQArm,           TB_ALIGN_16 },
621     { X86::VMOVSLDUPrr,     X86::VMOVSLDUPrm,         0 },
622     { X86::VMOVSHDUPrr,     X86::VMOVSHDUPrm,         0 },
623     { X86::VMOVUPDrr,       X86::VMOVUPDrm,           0 },
624     { X86::VMOVUPSrr,       X86::VMOVUPSrm,           0 },
625     { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm,    TB_ALIGN_16 },
626     { X86::VPABSBrr128,     X86::VPABSBrm128,         0 },
627     { X86::VPABSDrr128,     X86::VPABSDrm128,         0 },
628     { X86::VPABSWrr128,     X86::VPABSWrm128,         0 },
629     { X86::VPCMPESTRIrr,    X86::VPCMPESTRIrm,        0 },
630     { X86::VPCMPESTRM128rr, X86::VPCMPESTRM128rm,     0 },
631     { X86::VPCMPISTRIrr,    X86::VPCMPISTRIrm,        0 },
632     { X86::VPCMPISTRM128rr, X86::VPCMPISTRM128rm,     0 },
633     { X86::VPHMINPOSUWrr128, X86::VPHMINPOSUWrm128,   0 },
634     { X86::VPERMILPDri,     X86::VPERMILPDmi,         0 },
635     { X86::VPERMILPSri,     X86::VPERMILPSmi,         0 },
636     { X86::VPMOVSXBDrr,     X86::VPMOVSXBDrm,         0 },
637     { X86::VPMOVSXBQrr,     X86::VPMOVSXBQrm,         0 },
638     { X86::VPMOVSXBWrr,     X86::VPMOVSXBWrm,         0 },
639     { X86::VPMOVSXDQrr,     X86::VPMOVSXDQrm,         0 },
640     { X86::VPMOVSXWDrr,     X86::VPMOVSXWDrm,         0 },
641     { X86::VPMOVSXWQrr,     X86::VPMOVSXWQrm,         0 },
642     { X86::VPMOVZXBDrr,     X86::VPMOVZXBDrm,         0 },
643     { X86::VPMOVZXBQrr,     X86::VPMOVZXBQrm,         0 },
644     { X86::VPMOVZXBWrr,     X86::VPMOVZXBWrm,         0 },
645     { X86::VPMOVZXDQrr,     X86::VPMOVZXDQrm,         0 },
646     { X86::VPMOVZXWDrr,     X86::VPMOVZXWDrm,         0 },
647     { X86::VPMOVZXWQrr,     X86::VPMOVZXWQrm,         0 },
648     { X86::VPSHUFDri,       X86::VPSHUFDmi,           0 },
649     { X86::VPSHUFHWri,      X86::VPSHUFHWmi,          0 },
650     { X86::VPSHUFLWri,      X86::VPSHUFLWmi,          0 },
651     { X86::VPTESTrr,        X86::VPTESTrm,            0 },
652     { X86::VRCPPSr,         X86::VRCPPSm,             0 },
653     { X86::VROUNDPDr,       X86::VROUNDPDm,           0 },
654     { X86::VROUNDPSr,       X86::VROUNDPSm,           0 },
655     { X86::VRSQRTPSr,       X86::VRSQRTPSm,           0 },
656     { X86::VSQRTPDr,        X86::VSQRTPDm,            0 },
657     { X86::VSQRTPSr,        X86::VSQRTPSm,            0 },
658     { X86::VTESTPDrr,       X86::VTESTPDrm,           0 },
659     { X86::VTESTPSrr,       X86::VTESTPSrm,           0 },
660     { X86::VUCOMISDrr,      X86::VUCOMISDrm,          0 },
661     { X86::VUCOMISSrr,      X86::VUCOMISSrm,          0 },
662 
663     // AVX 256-bit foldable instructions
664     { X86::VCVTDQ2PDYrr,    X86::VCVTDQ2PDYrm,        0 },
665     { X86::VCVTDQ2PSYrr,    X86::VCVTDQ2PSYrm,        0 },
666     { X86::VCVTPD2DQYrr,    X86::VCVTPD2DQYrm,        0 },
667     { X86::VCVTPD2PSYrr,    X86::VCVTPD2PSYrm,        0 },
668     { X86::VCVTPS2DQYrr,    X86::VCVTPS2DQYrm,        0 },
669     { X86::VCVTPS2PDYrr,    X86::VCVTPS2PDYrm,        0 },
670     { X86::VCVTTPD2DQYrr,   X86::VCVTTPD2DQYrm,       0 },
671     { X86::VCVTTPS2DQYrr,   X86::VCVTTPS2DQYrm,       0 },
672     { X86::VMOVAPDYrr,      X86::VMOVAPDYrm,          TB_ALIGN_32 },
673     { X86::VMOVAPSYrr,      X86::VMOVAPSYrm,          TB_ALIGN_32 },
674     { X86::VMOVDDUPYrr,     X86::VMOVDDUPYrm,         0 },
675     { X86::VMOVDQAYrr,      X86::VMOVDQAYrm,          TB_ALIGN_32 },
676     { X86::VMOVSLDUPYrr,    X86::VMOVSLDUPYrm,        0 },
677     { X86::VMOVSHDUPYrr,    X86::VMOVSHDUPYrm,        0 },
678     { X86::VMOVUPDYrr,      X86::VMOVUPDYrm,          0 },
679     { X86::VMOVUPSYrr,      X86::VMOVUPSYrm,          0 },
680     { X86::VPERMILPDYri,    X86::VPERMILPDYmi,        0 },
681     { X86::VPERMILPSYri,    X86::VPERMILPSYmi,        0 },
682     { X86::VPTESTYrr,       X86::VPTESTYrm,           0 },
683     { X86::VRCPPSYr,        X86::VRCPPSYm,            0 },
684     { X86::VROUNDYPDr,      X86::VROUNDYPDm,          0 },
685     { X86::VROUNDYPSr,      X86::VROUNDYPSm,          0 },
686     { X86::VRSQRTPSYr,      X86::VRSQRTPSYm,          0 },
687     { X86::VSQRTPDYr,       X86::VSQRTPDYm,           0 },
688     { X86::VSQRTPSYr,       X86::VSQRTPSYm,           0 },
689     { X86::VTESTPDYrr,      X86::VTESTPDYrm,          0 },
690     { X86::VTESTPSYrr,      X86::VTESTPSYrm,          0 },
691 
692     // AVX2 foldable instructions
693 
694     // VBROADCASTS{SD}rr register instructions were an AVX2 addition while the
695     // VBROADCASTS{SD}rm memory instructions were available from AVX1.
696     // TB_NO_REVERSE prevents unfolding from introducing an illegal instruction
697     // on AVX1 targets. The VPBROADCAST instructions are all AVX2 instructions
698     // so they don't need an equivalent limitation.
699     { X86::VBROADCASTSSrr,  X86::VBROADCASTSSrm,      TB_NO_REVERSE },
700     { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm,     TB_NO_REVERSE },
701     { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm,     TB_NO_REVERSE },
702     { X86::VPABSBrr256,     X86::VPABSBrm256,         0 },
703     { X86::VPABSDrr256,     X86::VPABSDrm256,         0 },
704     { X86::VPABSWrr256,     X86::VPABSWrm256,         0 },
705     { X86::VPBROADCASTBrr,  X86::VPBROADCASTBrm,      0 },
706     { X86::VPBROADCASTBYrr, X86::VPBROADCASTBYrm,     0 },
707     { X86::VPBROADCASTDrr,  X86::VPBROADCASTDrm,      0 },
708     { X86::VPBROADCASTDYrr, X86::VPBROADCASTDYrm,     0 },
709     { X86::VPBROADCASTQrr,  X86::VPBROADCASTQrm,      0 },
710     { X86::VPBROADCASTQYrr, X86::VPBROADCASTQYrm,     0 },
711     { X86::VPBROADCASTWrr,  X86::VPBROADCASTWrm,      0 },
712     { X86::VPBROADCASTWYrr, X86::VPBROADCASTWYrm,     0 },
713     { X86::VPERMPDYri,      X86::VPERMPDYmi,          0 },
714     { X86::VPERMQYri,       X86::VPERMQYmi,           0 },
715     { X86::VPMOVSXBDYrr,    X86::VPMOVSXBDYrm,        0 },
716     { X86::VPMOVSXBQYrr,    X86::VPMOVSXBQYrm,        0 },
717     { X86::VPMOVSXBWYrr,    X86::VPMOVSXBWYrm,        0 },
718     { X86::VPMOVSXDQYrr,    X86::VPMOVSXDQYrm,        0 },
719     { X86::VPMOVSXWDYrr,    X86::VPMOVSXWDYrm,        0 },
720     { X86::VPMOVSXWQYrr,    X86::VPMOVSXWQYrm,        0 },
721     { X86::VPMOVZXBDYrr,    X86::VPMOVZXBDYrm,        0 },
722     { X86::VPMOVZXBQYrr,    X86::VPMOVZXBQYrm,        0 },
723     { X86::VPMOVZXBWYrr,    X86::VPMOVZXBWYrm,        0 },
724     { X86::VPMOVZXDQYrr,    X86::VPMOVZXDQYrm,        0 },
725     { X86::VPMOVZXWDYrr,    X86::VPMOVZXWDYrm,        0 },
726     { X86::VPMOVZXWQYrr,    X86::VPMOVZXWQYrm,        0 },
727     { X86::VPSHUFDYri,      X86::VPSHUFDYmi,          0 },
728     { X86::VPSHUFHWYri,     X86::VPSHUFHWYmi,         0 },
729     { X86::VPSHUFLWYri,     X86::VPSHUFLWYmi,         0 },
730 
731     // XOP foldable instructions
732     { X86::VFRCZPDrr,          X86::VFRCZPDrm,        0 },
733     { X86::VFRCZPDrrY,         X86::VFRCZPDrmY,       0 },
734     { X86::VFRCZPSrr,          X86::VFRCZPSrm,        0 },
735     { X86::VFRCZPSrrY,         X86::VFRCZPSrmY,       0 },
736     { X86::VFRCZSDrr,          X86::VFRCZSDrm,        0 },
737     { X86::VFRCZSSrr,          X86::VFRCZSSrm,        0 },
738     { X86::VPHADDBDrr,         X86::VPHADDBDrm,       0 },
739     { X86::VPHADDBQrr,         X86::VPHADDBQrm,       0 },
740     { X86::VPHADDBWrr,         X86::VPHADDBWrm,       0 },
741     { X86::VPHADDDQrr,         X86::VPHADDDQrm,       0 },
742     { X86::VPHADDWDrr,         X86::VPHADDWDrm,       0 },
743     { X86::VPHADDWQrr,         X86::VPHADDWQrm,       0 },
744     { X86::VPHADDUBDrr,        X86::VPHADDUBDrm,      0 },
745     { X86::VPHADDUBQrr,        X86::VPHADDUBQrm,      0 },
746     { X86::VPHADDUBWrr,        X86::VPHADDUBWrm,      0 },
747     { X86::VPHADDUDQrr,        X86::VPHADDUDQrm,      0 },
748     { X86::VPHADDUWDrr,        X86::VPHADDUWDrm,      0 },
749     { X86::VPHADDUWQrr,        X86::VPHADDUWQrm,      0 },
750     { X86::VPHSUBBWrr,         X86::VPHSUBBWrm,       0 },
751     { X86::VPHSUBDQrr,         X86::VPHSUBDQrm,       0 },
752     { X86::VPHSUBWDrr,         X86::VPHSUBWDrm,       0 },
753     { X86::VPROTBri,           X86::VPROTBmi,         0 },
754     { X86::VPROTBrr,           X86::VPROTBmr,         0 },
755     { X86::VPROTDri,           X86::VPROTDmi,         0 },
756     { X86::VPROTDrr,           X86::VPROTDmr,         0 },
757     { X86::VPROTQri,           X86::VPROTQmi,         0 },
758     { X86::VPROTQrr,           X86::VPROTQmr,         0 },
759     { X86::VPROTWri,           X86::VPROTWmi,         0 },
760     { X86::VPROTWrr,           X86::VPROTWmr,         0 },
761     { X86::VPSHABrr,           X86::VPSHABmr,         0 },
762     { X86::VPSHADrr,           X86::VPSHADmr,         0 },
763     { X86::VPSHAQrr,           X86::VPSHAQmr,         0 },
764     { X86::VPSHAWrr,           X86::VPSHAWmr,         0 },
765     { X86::VPSHLBrr,           X86::VPSHLBmr,         0 },
766     { X86::VPSHLDrr,           X86::VPSHLDmr,         0 },
767     { X86::VPSHLQrr,           X86::VPSHLQmr,         0 },
768     { X86::VPSHLWrr,           X86::VPSHLWmr,         0 },
769 
770     // BMI/BMI2/LZCNT/POPCNT/TBM foldable instructions
771     { X86::BEXTR32rr,       X86::BEXTR32rm,           0 },
772     { X86::BEXTR64rr,       X86::BEXTR64rm,           0 },
773     { X86::BEXTRI32ri,      X86::BEXTRI32mi,          0 },
774     { X86::BEXTRI64ri,      X86::BEXTRI64mi,          0 },
775     { X86::BLCFILL32rr,     X86::BLCFILL32rm,         0 },
776     { X86::BLCFILL64rr,     X86::BLCFILL64rm,         0 },
777     { X86::BLCI32rr,        X86::BLCI32rm,            0 },
778     { X86::BLCI64rr,        X86::BLCI64rm,            0 },
779     { X86::BLCIC32rr,       X86::BLCIC32rm,           0 },
780     { X86::BLCIC64rr,       X86::BLCIC64rm,           0 },
781     { X86::BLCMSK32rr,      X86::BLCMSK32rm,          0 },
782     { X86::BLCMSK64rr,      X86::BLCMSK64rm,          0 },
783     { X86::BLCS32rr,        X86::BLCS32rm,            0 },
784     { X86::BLCS64rr,        X86::BLCS64rm,            0 },
785     { X86::BLSFILL32rr,     X86::BLSFILL32rm,         0 },
786     { X86::BLSFILL64rr,     X86::BLSFILL64rm,         0 },
787     { X86::BLSI32rr,        X86::BLSI32rm,            0 },
788     { X86::BLSI64rr,        X86::BLSI64rm,            0 },
789     { X86::BLSIC32rr,       X86::BLSIC32rm,           0 },
790     { X86::BLSIC64rr,       X86::BLSIC64rm,           0 },
791     { X86::BLSMSK32rr,      X86::BLSMSK32rm,          0 },
792     { X86::BLSMSK64rr,      X86::BLSMSK64rm,          0 },
793     { X86::BLSR32rr,        X86::BLSR32rm,            0 },
794     { X86::BLSR64rr,        X86::BLSR64rm,            0 },
795     { X86::BZHI32rr,        X86::BZHI32rm,            0 },
796     { X86::BZHI64rr,        X86::BZHI64rm,            0 },
797     { X86::LZCNT16rr,       X86::LZCNT16rm,           0 },
798     { X86::LZCNT32rr,       X86::LZCNT32rm,           0 },
799     { X86::LZCNT64rr,       X86::LZCNT64rm,           0 },
800     { X86::POPCNT16rr,      X86::POPCNT16rm,          0 },
801     { X86::POPCNT32rr,      X86::POPCNT32rm,          0 },
802     { X86::POPCNT64rr,      X86::POPCNT64rm,          0 },
803     { X86::RORX32ri,        X86::RORX32mi,            0 },
804     { X86::RORX64ri,        X86::RORX64mi,            0 },
805     { X86::SARX32rr,        X86::SARX32rm,            0 },
806     { X86::SARX64rr,        X86::SARX64rm,            0 },
807     { X86::SHRX32rr,        X86::SHRX32rm,            0 },
808     { X86::SHRX64rr,        X86::SHRX64rm,            0 },
809     { X86::SHLX32rr,        X86::SHLX32rm,            0 },
810     { X86::SHLX64rr,        X86::SHLX64rm,            0 },
811     { X86::T1MSKC32rr,      X86::T1MSKC32rm,          0 },
812     { X86::T1MSKC64rr,      X86::T1MSKC64rm,          0 },
813     { X86::TZCNT16rr,       X86::TZCNT16rm,           0 },
814     { X86::TZCNT32rr,       X86::TZCNT32rm,           0 },
815     { X86::TZCNT64rr,       X86::TZCNT64rm,           0 },
816     { X86::TZMSK32rr,       X86::TZMSK32rm,           0 },
817     { X86::TZMSK64rr,       X86::TZMSK64rm,           0 },
818 
819     // AVX-512 foldable instructions
820     { X86::VMOV64toPQIZrr,   X86::VMOVQI2PQIZrm,      0 },
821     { X86::VMOVDI2SSZrr,     X86::VMOVDI2SSZrm,       0 },
822     { X86::VMOVAPDZrr,       X86::VMOVAPDZrm,         TB_ALIGN_64 },
823     { X86::VMOVAPSZrr,       X86::VMOVAPSZrm,         TB_ALIGN_64 },
824     { X86::VMOVDQA32Zrr,     X86::VMOVDQA32Zrm,       TB_ALIGN_64 },
825     { X86::VMOVDQA64Zrr,     X86::VMOVDQA64Zrm,       TB_ALIGN_64 },
826     { X86::VMOVDQU8Zrr,      X86::VMOVDQU8Zrm,        0 },
827     { X86::VMOVDQU16Zrr,     X86::VMOVDQU16Zrm,       0 },
828     { X86::VMOVDQU32Zrr,     X86::VMOVDQU32Zrm,       0 },
829     { X86::VMOVDQU64Zrr,     X86::VMOVDQU64Zrm,       0 },
830     { X86::VMOVUPDZrr,       X86::VMOVUPDZrm,         0 },
831     { X86::VMOVUPSZrr,       X86::VMOVUPSZrm,         0 },
832     { X86::VPABSDZrr,        X86::VPABSDZrm,          0 },
833     { X86::VPABSQZrr,        X86::VPABSQZrm,          0 },
834     { X86::VBROADCASTSSZr,   X86::VBROADCASTSSZm,     TB_NO_REVERSE },
835     { X86::VBROADCASTSSZr_s, X86::VBROADCASTSSZm,     TB_NO_REVERSE },
836     { X86::VBROADCASTSDZr,   X86::VBROADCASTSDZm,     TB_NO_REVERSE },
837     { X86::VBROADCASTSDZr_s, X86::VBROADCASTSDZm,     TB_NO_REVERSE },
838 
839     // AVX-512 foldable instructions (256-bit versions)
840     { X86::VMOVAPDZ256rr,        X86::VMOVAPDZ256rm,        TB_ALIGN_32 },
841     { X86::VMOVAPSZ256rr,        X86::VMOVAPSZ256rm,        TB_ALIGN_32 },
842     { X86::VMOVDQA32Z256rr,      X86::VMOVDQA32Z256rm,      TB_ALIGN_32 },
843     { X86::VMOVDQA64Z256rr,      X86::VMOVDQA64Z256rm,      TB_ALIGN_32 },
844     { X86::VMOVDQU8Z256rr,       X86::VMOVDQU8Z256rm,       0 },
845     { X86::VMOVDQU16Z256rr,      X86::VMOVDQU16Z256rm,      0 },
846     { X86::VMOVDQU32Z256rr,      X86::VMOVDQU32Z256rm,      0 },
847     { X86::VMOVDQU64Z256rr,      X86::VMOVDQU64Z256rm,      0 },
848     { X86::VMOVUPDZ256rr,        X86::VMOVUPDZ256rm,        0 },
849     { X86::VMOVUPSZ256rr,        X86::VMOVUPSZ256rm,        0 },
850     { X86::VBROADCASTSSZ256r,    X86::VBROADCASTSSZ256m,    TB_NO_REVERSE },
851     { X86::VBROADCASTSSZ256r_s,  X86::VBROADCASTSSZ256m,    TB_NO_REVERSE },
852     { X86::VBROADCASTSDZ256r,    X86::VBROADCASTSDZ256m,    TB_NO_REVERSE },
853     { X86::VBROADCASTSDZ256r_s,  X86::VBROADCASTSDZ256m,    TB_NO_REVERSE },
854 
855     // AVX-512 foldable instructions (128-bit versions)
856     { X86::VMOVAPDZ128rr,        X86::VMOVAPDZ128rm,        TB_ALIGN_16 },
857     { X86::VMOVAPSZ128rr,        X86::VMOVAPSZ128rm,        TB_ALIGN_16 },
858     { X86::VMOVDQA32Z128rr,      X86::VMOVDQA32Z128rm,      TB_ALIGN_16 },
859     { X86::VMOVDQA64Z128rr,      X86::VMOVDQA64Z128rm,      TB_ALIGN_16 },
860     { X86::VMOVDQU8Z128rr,       X86::VMOVDQU8Z128rm,       0 },
861     { X86::VMOVDQU16Z128rr,      X86::VMOVDQU16Z128rm,      0 },
862     { X86::VMOVDQU32Z128rr,      X86::VMOVDQU32Z128rm,      0 },
863     { X86::VMOVDQU64Z128rr,      X86::VMOVDQU64Z128rm,      0 },
864     { X86::VMOVUPDZ128rr,        X86::VMOVUPDZ128rm,        0 },
865     { X86::VMOVUPSZ128rr,        X86::VMOVUPSZ128rm,        0 },
866     { X86::VBROADCASTSSZ128r,    X86::VBROADCASTSSZ128m,    TB_NO_REVERSE },
867     { X86::VBROADCASTSSZ128r_s,  X86::VBROADCASTSSZ128m,    TB_NO_REVERSE },
868     // F16C foldable instructions
869     { X86::VCVTPH2PSrr,        X86::VCVTPH2PSrm,            0 },
870     { X86::VCVTPH2PSYrr,       X86::VCVTPH2PSYrm,           0 },
871 
872     // AES foldable instructions
873     { X86::AESIMCrr,              X86::AESIMCrm,              TB_ALIGN_16 },
874     { X86::AESKEYGENASSIST128rr,  X86::AESKEYGENASSIST128rm,  TB_ALIGN_16 },
875     { X86::VAESIMCrr,             X86::VAESIMCrm,             0 },
876     { X86::VAESKEYGENASSIST128rr, X86::VAESKEYGENASSIST128rm, 0 }
877   };
878 
879   for (X86MemoryFoldTableEntry Entry : MemoryFoldTable1) {
880     AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
881                   Entry.RegOp, Entry.MemOp,
882                   // Index 1, folded load
883                   Entry.Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
884   }
885 
886   static const X86MemoryFoldTableEntry MemoryFoldTable2[] = {
887     { X86::ADC32rr,         X86::ADC32rm,       0 },
888     { X86::ADC64rr,         X86::ADC64rm,       0 },
889     { X86::ADD16rr,         X86::ADD16rm,       0 },
890     { X86::ADD16rr_DB,      X86::ADD16rm,       TB_NO_REVERSE },
891     { X86::ADD32rr,         X86::ADD32rm,       0 },
892     { X86::ADD32rr_DB,      X86::ADD32rm,       TB_NO_REVERSE },
893     { X86::ADD64rr,         X86::ADD64rm,       0 },
894     { X86::ADD64rr_DB,      X86::ADD64rm,       TB_NO_REVERSE },
895     { X86::ADD8rr,          X86::ADD8rm,        0 },
896     { X86::ADDPDrr,         X86::ADDPDrm,       TB_ALIGN_16 },
897     { X86::ADDPSrr,         X86::ADDPSrm,       TB_ALIGN_16 },
898     { X86::ADDSDrr,         X86::ADDSDrm,       0 },
899     { X86::ADDSDrr_Int,     X86::ADDSDrm_Int,   0 },
900     { X86::ADDSSrr,         X86::ADDSSrm,       0 },
901     { X86::ADDSSrr_Int,     X86::ADDSSrm_Int,   0 },
902     { X86::ADDSUBPDrr,      X86::ADDSUBPDrm,    TB_ALIGN_16 },
903     { X86::ADDSUBPSrr,      X86::ADDSUBPSrm,    TB_ALIGN_16 },
904     { X86::AND16rr,         X86::AND16rm,       0 },
905     { X86::AND32rr,         X86::AND32rm,       0 },
906     { X86::AND64rr,         X86::AND64rm,       0 },
907     { X86::AND8rr,          X86::AND8rm,        0 },
908     { X86::ANDNPDrr,        X86::ANDNPDrm,      TB_ALIGN_16 },
909     { X86::ANDNPSrr,        X86::ANDNPSrm,      TB_ALIGN_16 },
910     { X86::ANDPDrr,         X86::ANDPDrm,       TB_ALIGN_16 },
911     { X86::ANDPSrr,         X86::ANDPSrm,       TB_ALIGN_16 },
912     { X86::BLENDPDrri,      X86::BLENDPDrmi,    TB_ALIGN_16 },
913     { X86::BLENDPSrri,      X86::BLENDPSrmi,    TB_ALIGN_16 },
914     { X86::BLENDVPDrr0,     X86::BLENDVPDrm0,   TB_ALIGN_16 },
915     { X86::BLENDVPSrr0,     X86::BLENDVPSrm0,   TB_ALIGN_16 },
916     { X86::CMOVA16rr,       X86::CMOVA16rm,     0 },
917     { X86::CMOVA32rr,       X86::CMOVA32rm,     0 },
918     { X86::CMOVA64rr,       X86::CMOVA64rm,     0 },
919     { X86::CMOVAE16rr,      X86::CMOVAE16rm,    0 },
920     { X86::CMOVAE32rr,      X86::CMOVAE32rm,    0 },
921     { X86::CMOVAE64rr,      X86::CMOVAE64rm,    0 },
922     { X86::CMOVB16rr,       X86::CMOVB16rm,     0 },
923     { X86::CMOVB32rr,       X86::CMOVB32rm,     0 },
924     { X86::CMOVB64rr,       X86::CMOVB64rm,     0 },
925     { X86::CMOVBE16rr,      X86::CMOVBE16rm,    0 },
926     { X86::CMOVBE32rr,      X86::CMOVBE32rm,    0 },
927     { X86::CMOVBE64rr,      X86::CMOVBE64rm,    0 },
928     { X86::CMOVE16rr,       X86::CMOVE16rm,     0 },
929     { X86::CMOVE32rr,       X86::CMOVE32rm,     0 },
930     { X86::CMOVE64rr,       X86::CMOVE64rm,     0 },
931     { X86::CMOVG16rr,       X86::CMOVG16rm,     0 },
932     { X86::CMOVG32rr,       X86::CMOVG32rm,     0 },
933     { X86::CMOVG64rr,       X86::CMOVG64rm,     0 },
934     { X86::CMOVGE16rr,      X86::CMOVGE16rm,    0 },
935     { X86::CMOVGE32rr,      X86::CMOVGE32rm,    0 },
936     { X86::CMOVGE64rr,      X86::CMOVGE64rm,    0 },
937     { X86::CMOVL16rr,       X86::CMOVL16rm,     0 },
938     { X86::CMOVL32rr,       X86::CMOVL32rm,     0 },
939     { X86::CMOVL64rr,       X86::CMOVL64rm,     0 },
940     { X86::CMOVLE16rr,      X86::CMOVLE16rm,    0 },
941     { X86::CMOVLE32rr,      X86::CMOVLE32rm,    0 },
942     { X86::CMOVLE64rr,      X86::CMOVLE64rm,    0 },
943     { X86::CMOVNE16rr,      X86::CMOVNE16rm,    0 },
944     { X86::CMOVNE32rr,      X86::CMOVNE32rm,    0 },
945     { X86::CMOVNE64rr,      X86::CMOVNE64rm,    0 },
946     { X86::CMOVNO16rr,      X86::CMOVNO16rm,    0 },
947     { X86::CMOVNO32rr,      X86::CMOVNO32rm,    0 },
948     { X86::CMOVNO64rr,      X86::CMOVNO64rm,    0 },
949     { X86::CMOVNP16rr,      X86::CMOVNP16rm,    0 },
950     { X86::CMOVNP32rr,      X86::CMOVNP32rm,    0 },
951     { X86::CMOVNP64rr,      X86::CMOVNP64rm,    0 },
952     { X86::CMOVNS16rr,      X86::CMOVNS16rm,    0 },
953     { X86::CMOVNS32rr,      X86::CMOVNS32rm,    0 },
954     { X86::CMOVNS64rr,      X86::CMOVNS64rm,    0 },
955     { X86::CMOVO16rr,       X86::CMOVO16rm,     0 },
956     { X86::CMOVO32rr,       X86::CMOVO32rm,     0 },
957     { X86::CMOVO64rr,       X86::CMOVO64rm,     0 },
958     { X86::CMOVP16rr,       X86::CMOVP16rm,     0 },
959     { X86::CMOVP32rr,       X86::CMOVP32rm,     0 },
960     { X86::CMOVP64rr,       X86::CMOVP64rm,     0 },
961     { X86::CMOVS16rr,       X86::CMOVS16rm,     0 },
962     { X86::CMOVS32rr,       X86::CMOVS32rm,     0 },
963     { X86::CMOVS64rr,       X86::CMOVS64rm,     0 },
964     { X86::CMPPDrri,        X86::CMPPDrmi,      TB_ALIGN_16 },
965     { X86::CMPPSrri,        X86::CMPPSrmi,      TB_ALIGN_16 },
966     { X86::CMPSDrr,         X86::CMPSDrm,       0 },
967     { X86::CMPSSrr,         X86::CMPSSrm,       0 },
968     { X86::CRC32r32r32,     X86::CRC32r32m32,   0 },
969     { X86::CRC32r64r64,     X86::CRC32r64m64,   0 },
970     { X86::DIVPDrr,         X86::DIVPDrm,       TB_ALIGN_16 },
971     { X86::DIVPSrr,         X86::DIVPSrm,       TB_ALIGN_16 },
972     { X86::DIVSDrr,         X86::DIVSDrm,       0 },
973     { X86::DIVSDrr_Int,     X86::DIVSDrm_Int,   0 },
974     { X86::DIVSSrr,         X86::DIVSSrm,       0 },
975     { X86::DIVSSrr_Int,     X86::DIVSSrm_Int,   0 },
976     { X86::DPPDrri,         X86::DPPDrmi,       TB_ALIGN_16 },
977     { X86::DPPSrri,         X86::DPPSrmi,       TB_ALIGN_16 },
978 
979     // Do not fold Fs* scalar logical op loads because there are no scalar
980     // load variants for these instructions. When folded, the load is required
981     // to be 128-bits, so the load size would not match.
982 
983     { X86::FvANDNPDrr,      X86::FvANDNPDrm,    TB_ALIGN_16 },
984     { X86::FvANDNPSrr,      X86::FvANDNPSrm,    TB_ALIGN_16 },
985     { X86::FvANDPDrr,       X86::FvANDPDrm,     TB_ALIGN_16 },
986     { X86::FvANDPSrr,       X86::FvANDPSrm,     TB_ALIGN_16 },
987     { X86::FvORPDrr,        X86::FvORPDrm,      TB_ALIGN_16 },
988     { X86::FvORPSrr,        X86::FvORPSrm,      TB_ALIGN_16 },
989     { X86::FvXORPDrr,       X86::FvXORPDrm,     TB_ALIGN_16 },
990     { X86::FvXORPSrr,       X86::FvXORPSrm,     TB_ALIGN_16 },
991     { X86::HADDPDrr,        X86::HADDPDrm,      TB_ALIGN_16 },
992     { X86::HADDPSrr,        X86::HADDPSrm,      TB_ALIGN_16 },
993     { X86::HSUBPDrr,        X86::HSUBPDrm,      TB_ALIGN_16 },
994     { X86::HSUBPSrr,        X86::HSUBPSrm,      TB_ALIGN_16 },
995     { X86::IMUL16rr,        X86::IMUL16rm,      0 },
996     { X86::IMUL32rr,        X86::IMUL32rm,      0 },
997     { X86::IMUL64rr,        X86::IMUL64rm,      0 },
998     { X86::Int_CMPSDrr,     X86::Int_CMPSDrm,   0 },
999     { X86::Int_CMPSSrr,     X86::Int_CMPSSrm,   0 },
1000     { X86::Int_CVTSD2SSrr,  X86::Int_CVTSD2SSrm,      0 },
1001     { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm,    0 },
1002     { X86::Int_CVTSI2SDrr,  X86::Int_CVTSI2SDrm,      0 },
1003     { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm,    0 },
1004     { X86::Int_CVTSI2SSrr,  X86::Int_CVTSI2SSrm,      0 },
1005     { X86::Int_CVTSS2SDrr,  X86::Int_CVTSS2SDrm,      0 },
1006     { X86::MAXPDrr,         X86::MAXPDrm,       TB_ALIGN_16 },
1007     { X86::MAXPSrr,         X86::MAXPSrm,       TB_ALIGN_16 },
1008     { X86::MAXSDrr,         X86::MAXSDrm,       0 },
1009     { X86::MAXSDrr_Int,     X86::MAXSDrm_Int,   0 },
1010     { X86::MAXSSrr,         X86::MAXSSrm,       0 },
1011     { X86::MAXSSrr_Int,     X86::MAXSSrm_Int,   0 },
1012     { X86::MINPDrr,         X86::MINPDrm,       TB_ALIGN_16 },
1013     { X86::MINPSrr,         X86::MINPSrm,       TB_ALIGN_16 },
1014     { X86::MINSDrr,         X86::MINSDrm,       0 },
1015     { X86::MINSDrr_Int,     X86::MINSDrm_Int,   0 },
1016     { X86::MINSSrr,         X86::MINSSrm,       0 },
1017     { X86::MINSSrr_Int,     X86::MINSSrm_Int,   0 },
1018     { X86::MOVLHPSrr,       X86::MOVHPSrm,      TB_NO_REVERSE },
1019     { X86::MPSADBWrri,      X86::MPSADBWrmi,    TB_ALIGN_16 },
1020     { X86::MULPDrr,         X86::MULPDrm,       TB_ALIGN_16 },
1021     { X86::MULPSrr,         X86::MULPSrm,       TB_ALIGN_16 },
1022     { X86::MULSDrr,         X86::MULSDrm,       0 },
1023     { X86::MULSDrr_Int,     X86::MULSDrm_Int,   0 },
1024     { X86::MULSSrr,         X86::MULSSrm,       0 },
1025     { X86::MULSSrr_Int,     X86::MULSSrm_Int,   0 },
1026     { X86::OR16rr,          X86::OR16rm,        0 },
1027     { X86::OR32rr,          X86::OR32rm,        0 },
1028     { X86::OR64rr,          X86::OR64rm,        0 },
1029     { X86::OR8rr,           X86::OR8rm,         0 },
1030     { X86::ORPDrr,          X86::ORPDrm,        TB_ALIGN_16 },
1031     { X86::ORPSrr,          X86::ORPSrm,        TB_ALIGN_16 },
1032     { X86::PACKSSDWrr,      X86::PACKSSDWrm,    TB_ALIGN_16 },
1033     { X86::PACKSSWBrr,      X86::PACKSSWBrm,    TB_ALIGN_16 },
1034     { X86::PACKUSDWrr,      X86::PACKUSDWrm,    TB_ALIGN_16 },
1035     { X86::PACKUSWBrr,      X86::PACKUSWBrm,    TB_ALIGN_16 },
1036     { X86::PADDBrr,         X86::PADDBrm,       TB_ALIGN_16 },
1037     { X86::PADDDrr,         X86::PADDDrm,       TB_ALIGN_16 },
1038     { X86::PADDQrr,         X86::PADDQrm,       TB_ALIGN_16 },
1039     { X86::PADDSBrr,        X86::PADDSBrm,      TB_ALIGN_16 },
1040     { X86::PADDSWrr,        X86::PADDSWrm,      TB_ALIGN_16 },
1041     { X86::PADDUSBrr,       X86::PADDUSBrm,     TB_ALIGN_16 },
1042     { X86::PADDUSWrr,       X86::PADDUSWrm,     TB_ALIGN_16 },
1043     { X86::PADDWrr,         X86::PADDWrm,       TB_ALIGN_16 },
1044     { X86::PALIGNRrri,      X86::PALIGNRrmi,    TB_ALIGN_16 },
1045     { X86::PANDNrr,         X86::PANDNrm,       TB_ALIGN_16 },
1046     { X86::PANDrr,          X86::PANDrm,        TB_ALIGN_16 },
1047     { X86::PAVGBrr,         X86::PAVGBrm,       TB_ALIGN_16 },
1048     { X86::PAVGWrr,         X86::PAVGWrm,       TB_ALIGN_16 },
1049     { X86::PBLENDVBrr0,     X86::PBLENDVBrm0,   TB_ALIGN_16 },
1050     { X86::PBLENDWrri,      X86::PBLENDWrmi,    TB_ALIGN_16 },
1051     { X86::PCLMULQDQrr,     X86::PCLMULQDQrm,   TB_ALIGN_16 },
1052     { X86::PCMPEQBrr,       X86::PCMPEQBrm,     TB_ALIGN_16 },
1053     { X86::PCMPEQDrr,       X86::PCMPEQDrm,     TB_ALIGN_16 },
1054     { X86::PCMPEQQrr,       X86::PCMPEQQrm,     TB_ALIGN_16 },
1055     { X86::PCMPEQWrr,       X86::PCMPEQWrm,     TB_ALIGN_16 },
1056     { X86::PCMPGTBrr,       X86::PCMPGTBrm,     TB_ALIGN_16 },
1057     { X86::PCMPGTDrr,       X86::PCMPGTDrm,     TB_ALIGN_16 },
1058     { X86::PCMPGTQrr,       X86::PCMPGTQrm,     TB_ALIGN_16 },
1059     { X86::PCMPGTWrr,       X86::PCMPGTWrm,     TB_ALIGN_16 },
1060     { X86::PHADDDrr,        X86::PHADDDrm,      TB_ALIGN_16 },
1061     { X86::PHADDWrr,        X86::PHADDWrm,      TB_ALIGN_16 },
1062     { X86::PHADDSWrr128,    X86::PHADDSWrm128,  TB_ALIGN_16 },
1063     { X86::PHSUBDrr,        X86::PHSUBDrm,      TB_ALIGN_16 },
1064     { X86::PHSUBSWrr128,    X86::PHSUBSWrm128,  TB_ALIGN_16 },
1065     { X86::PHSUBWrr,        X86::PHSUBWrm,      TB_ALIGN_16 },
1066     { X86::PINSRBrr,        X86::PINSRBrm,      0 },
1067     { X86::PINSRDrr,        X86::PINSRDrm,      0 },
1068     { X86::PINSRQrr,        X86::PINSRQrm,      0 },
1069     { X86::PINSRWrri,       X86::PINSRWrmi,     0 },
1070     { X86::PMADDUBSWrr128,  X86::PMADDUBSWrm128, TB_ALIGN_16 },
1071     { X86::PMADDWDrr,       X86::PMADDWDrm,     TB_ALIGN_16 },
1072     { X86::PMAXSWrr,        X86::PMAXSWrm,      TB_ALIGN_16 },
1073     { X86::PMAXUBrr,        X86::PMAXUBrm,      TB_ALIGN_16 },
1074     { X86::PMINSWrr,        X86::PMINSWrm,      TB_ALIGN_16 },
1075     { X86::PMINUBrr,        X86::PMINUBrm,      TB_ALIGN_16 },
1076     { X86::PMINSBrr,        X86::PMINSBrm,      TB_ALIGN_16 },
1077     { X86::PMINSDrr,        X86::PMINSDrm,      TB_ALIGN_16 },
1078     { X86::PMINUDrr,        X86::PMINUDrm,      TB_ALIGN_16 },
1079     { X86::PMINUWrr,        X86::PMINUWrm,      TB_ALIGN_16 },
1080     { X86::PMAXSBrr,        X86::PMAXSBrm,      TB_ALIGN_16 },
1081     { X86::PMAXSDrr,        X86::PMAXSDrm,      TB_ALIGN_16 },
1082     { X86::PMAXUDrr,        X86::PMAXUDrm,      TB_ALIGN_16 },
1083     { X86::PMAXUWrr,        X86::PMAXUWrm,      TB_ALIGN_16 },
1084     { X86::PMULDQrr,        X86::PMULDQrm,      TB_ALIGN_16 },
1085     { X86::PMULHRSWrr128,   X86::PMULHRSWrm128, TB_ALIGN_16 },
1086     { X86::PMULHUWrr,       X86::PMULHUWrm,     TB_ALIGN_16 },
1087     { X86::PMULHWrr,        X86::PMULHWrm,      TB_ALIGN_16 },
1088     { X86::PMULLDrr,        X86::PMULLDrm,      TB_ALIGN_16 },
1089     { X86::PMULLWrr,        X86::PMULLWrm,      TB_ALIGN_16 },
1090     { X86::PMULUDQrr,       X86::PMULUDQrm,     TB_ALIGN_16 },
1091     { X86::PORrr,           X86::PORrm,         TB_ALIGN_16 },
1092     { X86::PSADBWrr,        X86::PSADBWrm,      TB_ALIGN_16 },
1093     { X86::PSHUFBrr,        X86::PSHUFBrm,      TB_ALIGN_16 },
1094     { X86::PSIGNBrr128,     X86::PSIGNBrm128,   TB_ALIGN_16 },
1095     { X86::PSIGNWrr128,     X86::PSIGNWrm128,   TB_ALIGN_16 },
1096     { X86::PSIGNDrr128,     X86::PSIGNDrm128,   TB_ALIGN_16 },
1097     { X86::PSLLDrr,         X86::PSLLDrm,       TB_ALIGN_16 },
1098     { X86::PSLLQrr,         X86::PSLLQrm,       TB_ALIGN_16 },
1099     { X86::PSLLWrr,         X86::PSLLWrm,       TB_ALIGN_16 },
1100     { X86::PSRADrr,         X86::PSRADrm,       TB_ALIGN_16 },
1101     { X86::PSRAWrr,         X86::PSRAWrm,       TB_ALIGN_16 },
1102     { X86::PSRLDrr,         X86::PSRLDrm,       TB_ALIGN_16 },
1103     { X86::PSRLQrr,         X86::PSRLQrm,       TB_ALIGN_16 },
1104     { X86::PSRLWrr,         X86::PSRLWrm,       TB_ALIGN_16 },
1105     { X86::PSUBBrr,         X86::PSUBBrm,       TB_ALIGN_16 },
1106     { X86::PSUBDrr,         X86::PSUBDrm,       TB_ALIGN_16 },
1107     { X86::PSUBQrr,         X86::PSUBQrm,       TB_ALIGN_16 },
1108     { X86::PSUBSBrr,        X86::PSUBSBrm,      TB_ALIGN_16 },
1109     { X86::PSUBSWrr,        X86::PSUBSWrm,      TB_ALIGN_16 },
1110     { X86::PSUBUSBrr,       X86::PSUBUSBrm,     TB_ALIGN_16 },
1111     { X86::PSUBUSWrr,       X86::PSUBUSWrm,     TB_ALIGN_16 },
1112     { X86::PSUBWrr,         X86::PSUBWrm,       TB_ALIGN_16 },
1113     { X86::PUNPCKHBWrr,     X86::PUNPCKHBWrm,   TB_ALIGN_16 },
1114     { X86::PUNPCKHDQrr,     X86::PUNPCKHDQrm,   TB_ALIGN_16 },
1115     { X86::PUNPCKHQDQrr,    X86::PUNPCKHQDQrm,  TB_ALIGN_16 },
1116     { X86::PUNPCKHWDrr,     X86::PUNPCKHWDrm,   TB_ALIGN_16 },
1117     { X86::PUNPCKLBWrr,     X86::PUNPCKLBWrm,   TB_ALIGN_16 },
1118     { X86::PUNPCKLDQrr,     X86::PUNPCKLDQrm,   TB_ALIGN_16 },
1119     { X86::PUNPCKLQDQrr,    X86::PUNPCKLQDQrm,  TB_ALIGN_16 },
1120     { X86::PUNPCKLWDrr,     X86::PUNPCKLWDrm,   TB_ALIGN_16 },
1121     { X86::PXORrr,          X86::PXORrm,        TB_ALIGN_16 },
1122     { X86::ROUNDSDr,        X86::ROUNDSDm,      0 },
1123     { X86::ROUNDSSr,        X86::ROUNDSSm,      0 },
1124     { X86::SBB32rr,         X86::SBB32rm,       0 },
1125     { X86::SBB64rr,         X86::SBB64rm,       0 },
1126     { X86::SHUFPDrri,       X86::SHUFPDrmi,     TB_ALIGN_16 },
1127     { X86::SHUFPSrri,       X86::SHUFPSrmi,     TB_ALIGN_16 },
1128     { X86::SUB16rr,         X86::SUB16rm,       0 },
1129     { X86::SUB32rr,         X86::SUB32rm,       0 },
1130     { X86::SUB64rr,         X86::SUB64rm,       0 },
1131     { X86::SUB8rr,          X86::SUB8rm,        0 },
1132     { X86::SUBPDrr,         X86::SUBPDrm,       TB_ALIGN_16 },
1133     { X86::SUBPSrr,         X86::SUBPSrm,       TB_ALIGN_16 },
1134     { X86::SUBSDrr,         X86::SUBSDrm,       0 },
1135     { X86::SUBSDrr_Int,     X86::SUBSDrm_Int,   0 },
1136     { X86::SUBSSrr,         X86::SUBSSrm,       0 },
1137     { X86::SUBSSrr_Int,     X86::SUBSSrm_Int,   0 },
1138     // FIXME: TEST*rr -> swapped operand of TEST*mr.
1139     { X86::UNPCKHPDrr,      X86::UNPCKHPDrm,    TB_ALIGN_16 },
1140     { X86::UNPCKHPSrr,      X86::UNPCKHPSrm,    TB_ALIGN_16 },
1141     { X86::UNPCKLPDrr,      X86::UNPCKLPDrm,    TB_ALIGN_16 },
1142     { X86::UNPCKLPSrr,      X86::UNPCKLPSrm,    TB_ALIGN_16 },
1143     { X86::XOR16rr,         X86::XOR16rm,       0 },
1144     { X86::XOR32rr,         X86::XOR32rm,       0 },
1145     { X86::XOR64rr,         X86::XOR64rm,       0 },
1146     { X86::XOR8rr,          X86::XOR8rm,        0 },
1147     { X86::XORPDrr,         X86::XORPDrm,       TB_ALIGN_16 },
1148     { X86::XORPSrr,         X86::XORPSrm,       TB_ALIGN_16 },
1149 
1150     // MMX version of foldable instructions
1151     { X86::MMX_CVTPI2PSirr,   X86::MMX_CVTPI2PSirm,   0 },
1152     { X86::MMX_PACKSSDWirr,   X86::MMX_PACKSSDWirm,   0 },
1153     { X86::MMX_PACKSSWBirr,   X86::MMX_PACKSSWBirm,   0 },
1154     { X86::MMX_PACKUSWBirr,   X86::MMX_PACKUSWBirm,   0 },
1155     { X86::MMX_PADDBirr,      X86::MMX_PADDBirm,      0 },
1156     { X86::MMX_PADDDirr,      X86::MMX_PADDDirm,      0 },
1157     { X86::MMX_PADDQirr,      X86::MMX_PADDQirm,      0 },
1158     { X86::MMX_PADDSBirr,     X86::MMX_PADDSBirm,     0 },
1159     { X86::MMX_PADDSWirr,     X86::MMX_PADDSWirm,     0 },
1160     { X86::MMX_PADDUSBirr,    X86::MMX_PADDUSBirm,    0 },
1161     { X86::MMX_PADDUSWirr,    X86::MMX_PADDUSWirm,    0 },
1162     { X86::MMX_PADDWirr,      X86::MMX_PADDWirm,      0 },
1163     { X86::MMX_PALIGNR64irr,  X86::MMX_PALIGNR64irm,  0 },
1164     { X86::MMX_PANDNirr,      X86::MMX_PANDNirm,      0 },
1165     { X86::MMX_PANDirr,       X86::MMX_PANDirm,       0 },
1166     { X86::MMX_PAVGBirr,      X86::MMX_PAVGBirm,      0 },
1167     { X86::MMX_PAVGWirr,      X86::MMX_PAVGWirm,      0 },
1168     { X86::MMX_PCMPEQBirr,    X86::MMX_PCMPEQBirm,    0 },
1169     { X86::MMX_PCMPEQDirr,    X86::MMX_PCMPEQDirm,    0 },
1170     { X86::MMX_PCMPEQWirr,    X86::MMX_PCMPEQWirm,    0 },
1171     { X86::MMX_PCMPGTBirr,    X86::MMX_PCMPGTBirm,    0 },
1172     { X86::MMX_PCMPGTDirr,    X86::MMX_PCMPGTDirm,    0 },
1173     { X86::MMX_PCMPGTWirr,    X86::MMX_PCMPGTWirm,    0 },
1174     { X86::MMX_PHADDSWrr64,   X86::MMX_PHADDSWrm64,   0 },
1175     { X86::MMX_PHADDWrr64,    X86::MMX_PHADDWrm64,    0 },
1176     { X86::MMX_PHADDrr64,     X86::MMX_PHADDrm64,     0 },
1177     { X86::MMX_PHSUBDrr64,    X86::MMX_PHSUBDrm64,    0 },
1178     { X86::MMX_PHSUBSWrr64,   X86::MMX_PHSUBSWrm64,   0 },
1179     { X86::MMX_PHSUBWrr64,    X86::MMX_PHSUBWrm64,    0 },
1180     { X86::MMX_PINSRWirri,    X86::MMX_PINSRWirmi,    0 },
1181     { X86::MMX_PMADDUBSWrr64, X86::MMX_PMADDUBSWrm64, 0 },
1182     { X86::MMX_PMADDWDirr,    X86::MMX_PMADDWDirm,    0 },
1183     { X86::MMX_PMAXSWirr,     X86::MMX_PMAXSWirm,     0 },
1184     { X86::MMX_PMAXUBirr,     X86::MMX_PMAXUBirm,     0 },
1185     { X86::MMX_PMINSWirr,     X86::MMX_PMINSWirm,     0 },
1186     { X86::MMX_PMINUBirr,     X86::MMX_PMINUBirm,     0 },
1187     { X86::MMX_PMULHRSWrr64,  X86::MMX_PMULHRSWrm64,  0 },
1188     { X86::MMX_PMULHUWirr,    X86::MMX_PMULHUWirm,    0 },
1189     { X86::MMX_PMULHWirr,     X86::MMX_PMULHWirm,     0 },
1190     { X86::MMX_PMULLWirr,     X86::MMX_PMULLWirm,     0 },
1191     { X86::MMX_PMULUDQirr,    X86::MMX_PMULUDQirm,    0 },
1192     { X86::MMX_PORirr,        X86::MMX_PORirm,        0 },
1193     { X86::MMX_PSADBWirr,     X86::MMX_PSADBWirm,     0 },
1194     { X86::MMX_PSHUFBrr64,    X86::MMX_PSHUFBrm64,    0 },
1195     { X86::MMX_PSIGNBrr64,    X86::MMX_PSIGNBrm64,    0 },
1196     { X86::MMX_PSIGNDrr64,    X86::MMX_PSIGNDrm64,    0 },
1197     { X86::MMX_PSIGNWrr64,    X86::MMX_PSIGNWrm64,    0 },
1198     { X86::MMX_PSLLDrr,       X86::MMX_PSLLDrm,       0 },
1199     { X86::MMX_PSLLQrr,       X86::MMX_PSLLQrm,       0 },
1200     { X86::MMX_PSLLWrr,       X86::MMX_PSLLWrm,       0 },
1201     { X86::MMX_PSRADrr,       X86::MMX_PSRADrm,       0 },
1202     { X86::MMX_PSRAWrr,       X86::MMX_PSRAWrm,       0 },
1203     { X86::MMX_PSRLDrr,       X86::MMX_PSRLDrm,       0 },
1204     { X86::MMX_PSRLQrr,       X86::MMX_PSRLQrm,       0 },
1205     { X86::MMX_PSRLWrr,       X86::MMX_PSRLWrm,       0 },
1206     { X86::MMX_PSUBBirr,      X86::MMX_PSUBBirm,      0 },
1207     { X86::MMX_PSUBDirr,      X86::MMX_PSUBDirm,      0 },
1208     { X86::MMX_PSUBQirr,      X86::MMX_PSUBQirm,      0 },
1209     { X86::MMX_PSUBSBirr,     X86::MMX_PSUBSBirm,     0 },
1210     { X86::MMX_PSUBSWirr,     X86::MMX_PSUBSWirm,     0 },
1211     { X86::MMX_PSUBUSBirr,    X86::MMX_PSUBUSBirm,    0 },
1212     { X86::MMX_PSUBUSWirr,    X86::MMX_PSUBUSWirm,    0 },
1213     { X86::MMX_PSUBWirr,      X86::MMX_PSUBWirm,      0 },
1214     { X86::MMX_PUNPCKHBWirr,  X86::MMX_PUNPCKHBWirm,  0 },
1215     { X86::MMX_PUNPCKHDQirr,  X86::MMX_PUNPCKHDQirm,  0 },
1216     { X86::MMX_PUNPCKHWDirr,  X86::MMX_PUNPCKHWDirm,  0 },
1217     { X86::MMX_PUNPCKLBWirr,  X86::MMX_PUNPCKLBWirm,  0 },
1218     { X86::MMX_PUNPCKLDQirr,  X86::MMX_PUNPCKLDQirm,  0 },
1219     { X86::MMX_PUNPCKLWDirr,  X86::MMX_PUNPCKLWDirm,  0 },
1220     { X86::MMX_PXORirr,       X86::MMX_PXORirm,       0 },
1221 
1222     // 3DNow! version of foldable instructions
1223     { X86::PAVGUSBrr,         X86::PAVGUSBrm,         0 },
1224     { X86::PFACCrr,           X86::PFACCrm,           0 },
1225     { X86::PFADDrr,           X86::PFADDrm,           0 },
1226     { X86::PFCMPEQrr,         X86::PFCMPEQrm,         0 },
1227     { X86::PFCMPGErr,         X86::PFCMPGErm,         0 },
1228     { X86::PFCMPGTrr,         X86::PFCMPGTrm,         0 },
1229     { X86::PFMAXrr,           X86::PFMAXrm,           0 },
1230     { X86::PFMINrr,           X86::PFMINrm,           0 },
1231     { X86::PFMULrr,           X86::PFMULrm,           0 },
1232     { X86::PFNACCrr,          X86::PFNACCrm,          0 },
1233     { X86::PFPNACCrr,         X86::PFPNACCrm,         0 },
1234     { X86::PFRCPIT1rr,        X86::PFRCPIT1rm,        0 },
1235     { X86::PFRCPIT2rr,        X86::PFRCPIT2rm,        0 },
1236     { X86::PFRSQIT1rr,        X86::PFRSQIT1rm,        0 },
1237     { X86::PFSUBrr,           X86::PFSUBrm,           0 },
1238     { X86::PFSUBRrr,          X86::PFSUBRrm,          0 },
1239     { X86::PMULHRWrr,         X86::PMULHRWrm,         0 },
1240 
1241     // AVX 128-bit versions of foldable instructions
1242     { X86::VCVTSD2SSrr,       X86::VCVTSD2SSrm,        0 },
1243     { X86::Int_VCVTSD2SSrr,   X86::Int_VCVTSD2SSrm,    0 },
1244     { X86::VCVTSI2SD64rr,     X86::VCVTSI2SD64rm,      0 },
1245     { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm,  0 },
1246     { X86::VCVTSI2SDrr,       X86::VCVTSI2SDrm,        0 },
1247     { X86::Int_VCVTSI2SDrr,   X86::Int_VCVTSI2SDrm,    0 },
1248     { X86::VCVTSI2SS64rr,     X86::VCVTSI2SS64rm,      0 },
1249     { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm,  0 },
1250     { X86::VCVTSI2SSrr,       X86::VCVTSI2SSrm,        0 },
1251     { X86::Int_VCVTSI2SSrr,   X86::Int_VCVTSI2SSrm,    0 },
1252     { X86::VCVTSS2SDrr,       X86::VCVTSS2SDrm,        0 },
1253     { X86::Int_VCVTSS2SDrr,   X86::Int_VCVTSS2SDrm,    0 },
1254     { X86::VRCPSSr,           X86::VRCPSSm,            0 },
1255     { X86::VRCPSSr_Int,       X86::VRCPSSm_Int,        0 },
1256     { X86::VRSQRTSSr,         X86::VRSQRTSSm,          0 },
1257     { X86::VRSQRTSSr_Int,     X86::VRSQRTSSm_Int,      0 },
1258     { X86::VSQRTSDr,          X86::VSQRTSDm,           0 },
1259     { X86::VSQRTSDr_Int,      X86::VSQRTSDm_Int,       0 },
1260     { X86::VSQRTSSr,          X86::VSQRTSSm,           0 },
1261     { X86::VSQRTSSr_Int,      X86::VSQRTSSm_Int,       0 },
1262     { X86::VADDPDrr,          X86::VADDPDrm,           0 },
1263     { X86::VADDPSrr,          X86::VADDPSrm,           0 },
1264     { X86::VADDSDrr,          X86::VADDSDrm,           0 },
1265     { X86::VADDSDrr_Int,      X86::VADDSDrm_Int,       0 },
1266     { X86::VADDSSrr,          X86::VADDSSrm,           0 },
1267     { X86::VADDSSrr_Int,      X86::VADDSSrm_Int,       0 },
1268     { X86::VADDSUBPDrr,       X86::VADDSUBPDrm,        0 },
1269     { X86::VADDSUBPSrr,       X86::VADDSUBPSrm,        0 },
1270     { X86::VANDNPDrr,         X86::VANDNPDrm,          0 },
1271     { X86::VANDNPSrr,         X86::VANDNPSrm,          0 },
1272     { X86::VANDPDrr,          X86::VANDPDrm,           0 },
1273     { X86::VANDPSrr,          X86::VANDPSrm,           0 },
1274     { X86::VBLENDPDrri,       X86::VBLENDPDrmi,        0 },
1275     { X86::VBLENDPSrri,       X86::VBLENDPSrmi,        0 },
1276     { X86::VBLENDVPDrr,       X86::VBLENDVPDrm,        0 },
1277     { X86::VBLENDVPSrr,       X86::VBLENDVPSrm,        0 },
1278     { X86::VCMPPDrri,         X86::VCMPPDrmi,          0 },
1279     { X86::VCMPPSrri,         X86::VCMPPSrmi,          0 },
1280     { X86::VCMPSDrr,          X86::VCMPSDrm,           0 },
1281     { X86::VCMPSSrr,          X86::VCMPSSrm,           0 },
1282     { X86::VDIVPDrr,          X86::VDIVPDrm,           0 },
1283     { X86::VDIVPSrr,          X86::VDIVPSrm,           0 },
1284     { X86::VDIVSDrr,          X86::VDIVSDrm,           0 },
1285     { X86::VDIVSDrr_Int,      X86::VDIVSDrm_Int,       0 },
1286     { X86::VDIVSSrr,          X86::VDIVSSrm,           0 },
1287     { X86::VDIVSSrr_Int,      X86::VDIVSSrm_Int,       0 },
1288     { X86::VDPPDrri,          X86::VDPPDrmi,           0 },
1289     { X86::VDPPSrri,          X86::VDPPSrmi,           0 },
1290     // Do not fold VFs* loads because there are no scalar load variants for
1291     // these instructions. When folded, the load is required to be 128-bits, so
1292     // the load size would not match.
1293     { X86::VFvANDNPDrr,       X86::VFvANDNPDrm,        0 },
1294     { X86::VFvANDNPSrr,       X86::VFvANDNPSrm,        0 },
1295     { X86::VFvANDPDrr,        X86::VFvANDPDrm,         0 },
1296     { X86::VFvANDPSrr,        X86::VFvANDPSrm,         0 },
1297     { X86::VFvORPDrr,         X86::VFvORPDrm,          0 },
1298     { X86::VFvORPSrr,         X86::VFvORPSrm,          0 },
1299     { X86::VFvXORPDrr,        X86::VFvXORPDrm,         0 },
1300     { X86::VFvXORPSrr,        X86::VFvXORPSrm,         0 },
1301     { X86::VHADDPDrr,         X86::VHADDPDrm,          0 },
1302     { X86::VHADDPSrr,         X86::VHADDPSrm,          0 },
1303     { X86::VHSUBPDrr,         X86::VHSUBPDrm,          0 },
1304     { X86::VHSUBPSrr,         X86::VHSUBPSrm,          0 },
1305     { X86::Int_VCMPSDrr,      X86::Int_VCMPSDrm,       0 },
1306     { X86::Int_VCMPSSrr,      X86::Int_VCMPSSrm,       0 },
1307     { X86::VMAXPDrr,          X86::VMAXPDrm,           0 },
1308     { X86::VMAXPSrr,          X86::VMAXPSrm,           0 },
1309     { X86::VMAXSDrr,          X86::VMAXSDrm,           0 },
1310     { X86::VMAXSDrr_Int,      X86::VMAXSDrm_Int,       0 },
1311     { X86::VMAXSSrr,          X86::VMAXSSrm,           0 },
1312     { X86::VMAXSSrr_Int,      X86::VMAXSSrm_Int,       0 },
1313     { X86::VMINPDrr,          X86::VMINPDrm,           0 },
1314     { X86::VMINPSrr,          X86::VMINPSrm,           0 },
1315     { X86::VMINSDrr,          X86::VMINSDrm,           0 },
1316     { X86::VMINSDrr_Int,      X86::VMINSDrm_Int,       0 },
1317     { X86::VMINSSrr,          X86::VMINSSrm,           0 },
1318     { X86::VMINSSrr_Int,      X86::VMINSSrm_Int,       0 },
1319     { X86::VMOVLHPSrr,        X86::VMOVHPSrm,          TB_NO_REVERSE },
1320     { X86::VMPSADBWrri,       X86::VMPSADBWrmi,        0 },
1321     { X86::VMULPDrr,          X86::VMULPDrm,           0 },
1322     { X86::VMULPSrr,          X86::VMULPSrm,           0 },
1323     { X86::VMULSDrr,          X86::VMULSDrm,           0 },
1324     { X86::VMULSDrr_Int,      X86::VMULSDrm_Int,       0 },
1325     { X86::VMULSSrr,          X86::VMULSSrm,           0 },
1326     { X86::VMULSSrr_Int,      X86::VMULSSrm_Int,       0 },
1327     { X86::VORPDrr,           X86::VORPDrm,            0 },
1328     { X86::VORPSrr,           X86::VORPSrm,            0 },
1329     { X86::VPACKSSDWrr,       X86::VPACKSSDWrm,        0 },
1330     { X86::VPACKSSWBrr,       X86::VPACKSSWBrm,        0 },
1331     { X86::VPACKUSDWrr,       X86::VPACKUSDWrm,        0 },
1332     { X86::VPACKUSWBrr,       X86::VPACKUSWBrm,        0 },
1333     { X86::VPADDBrr,          X86::VPADDBrm,           0 },
1334     { X86::VPADDDrr,          X86::VPADDDrm,           0 },
1335     { X86::VPADDQrr,          X86::VPADDQrm,           0 },
1336     { X86::VPADDSBrr,         X86::VPADDSBrm,          0 },
1337     { X86::VPADDSWrr,         X86::VPADDSWrm,          0 },
1338     { X86::VPADDUSBrr,        X86::VPADDUSBrm,         0 },
1339     { X86::VPADDUSWrr,        X86::VPADDUSWrm,         0 },
1340     { X86::VPADDWrr,          X86::VPADDWrm,           0 },
1341     { X86::VPALIGNRrri,       X86::VPALIGNRrmi,        0 },
1342     { X86::VPANDNrr,          X86::VPANDNrm,           0 },
1343     { X86::VPANDrr,           X86::VPANDrm,            0 },
1344     { X86::VPAVGBrr,          X86::VPAVGBrm,           0 },
1345     { X86::VPAVGWrr,          X86::VPAVGWrm,           0 },
1346     { X86::VPBLENDVBrr,       X86::VPBLENDVBrm,        0 },
1347     { X86::VPBLENDWrri,       X86::VPBLENDWrmi,        0 },
1348     { X86::VPCLMULQDQrr,      X86::VPCLMULQDQrm,       0 },
1349     { X86::VPCMPEQBrr,        X86::VPCMPEQBrm,         0 },
1350     { X86::VPCMPEQDrr,        X86::VPCMPEQDrm,         0 },
1351     { X86::VPCMPEQQrr,        X86::VPCMPEQQrm,         0 },
1352     { X86::VPCMPEQWrr,        X86::VPCMPEQWrm,         0 },
1353     { X86::VPCMPGTBrr,        X86::VPCMPGTBrm,         0 },
1354     { X86::VPCMPGTDrr,        X86::VPCMPGTDrm,         0 },
1355     { X86::VPCMPGTQrr,        X86::VPCMPGTQrm,         0 },
1356     { X86::VPCMPGTWrr,        X86::VPCMPGTWrm,         0 },
1357     { X86::VPHADDDrr,         X86::VPHADDDrm,          0 },
1358     { X86::VPHADDSWrr128,     X86::VPHADDSWrm128,      0 },
1359     { X86::VPHADDWrr,         X86::VPHADDWrm,          0 },
1360     { X86::VPHSUBDrr,         X86::VPHSUBDrm,          0 },
1361     { X86::VPHSUBSWrr128,     X86::VPHSUBSWrm128,      0 },
1362     { X86::VPHSUBWrr,         X86::VPHSUBWrm,          0 },
1363     { X86::VPERMILPDrr,       X86::VPERMILPDrm,        0 },
1364     { X86::VPERMILPSrr,       X86::VPERMILPSrm,        0 },
1365     { X86::VPINSRBrr,         X86::VPINSRBrm,          0 },
1366     { X86::VPINSRDrr,         X86::VPINSRDrm,          0 },
1367     { X86::VPINSRQrr,         X86::VPINSRQrm,          0 },
1368     { X86::VPINSRWrri,        X86::VPINSRWrmi,         0 },
1369     { X86::VPMADDUBSWrr128,   X86::VPMADDUBSWrm128,    0 },
1370     { X86::VPMADDWDrr,        X86::VPMADDWDrm,         0 },
1371     { X86::VPMAXSWrr,         X86::VPMAXSWrm,          0 },
1372     { X86::VPMAXUBrr,         X86::VPMAXUBrm,          0 },
1373     { X86::VPMINSWrr,         X86::VPMINSWrm,          0 },
1374     { X86::VPMINUBrr,         X86::VPMINUBrm,          0 },
1375     { X86::VPMINSBrr,         X86::VPMINSBrm,          0 },
1376     { X86::VPMINSDrr,         X86::VPMINSDrm,          0 },
1377     { X86::VPMINUDrr,         X86::VPMINUDrm,          0 },
1378     { X86::VPMINUWrr,         X86::VPMINUWrm,          0 },
1379     { X86::VPMAXSBrr,         X86::VPMAXSBrm,          0 },
1380     { X86::VPMAXSDrr,         X86::VPMAXSDrm,          0 },
1381     { X86::VPMAXUDrr,         X86::VPMAXUDrm,          0 },
1382     { X86::VPMAXUWrr,         X86::VPMAXUWrm,          0 },
1383     { X86::VPMULDQrr,         X86::VPMULDQrm,          0 },
1384     { X86::VPMULHRSWrr128,    X86::VPMULHRSWrm128,     0 },
1385     { X86::VPMULHUWrr,        X86::VPMULHUWrm,         0 },
1386     { X86::VPMULHWrr,         X86::VPMULHWrm,          0 },
1387     { X86::VPMULLDrr,         X86::VPMULLDrm,          0 },
1388     { X86::VPMULLWrr,         X86::VPMULLWrm,          0 },
1389     { X86::VPMULUDQrr,        X86::VPMULUDQrm,         0 },
1390     { X86::VPORrr,            X86::VPORrm,             0 },
1391     { X86::VPSADBWrr,         X86::VPSADBWrm,          0 },
1392     { X86::VPSHUFBrr,         X86::VPSHUFBrm,          0 },
1393     { X86::VPSIGNBrr128,      X86::VPSIGNBrm128,       0 },
1394     { X86::VPSIGNWrr128,      X86::VPSIGNWrm128,       0 },
1395     { X86::VPSIGNDrr128,      X86::VPSIGNDrm128,       0 },
1396     { X86::VPSLLDrr,          X86::VPSLLDrm,           0 },
1397     { X86::VPSLLQrr,          X86::VPSLLQrm,           0 },
1398     { X86::VPSLLWrr,          X86::VPSLLWrm,           0 },
1399     { X86::VPSRADrr,          X86::VPSRADrm,           0 },
1400     { X86::VPSRAWrr,          X86::VPSRAWrm,           0 },
1401     { X86::VPSRLDrr,          X86::VPSRLDrm,           0 },
1402     { X86::VPSRLQrr,          X86::VPSRLQrm,           0 },
1403     { X86::VPSRLWrr,          X86::VPSRLWrm,           0 },
1404     { X86::VPSUBBrr,          X86::VPSUBBrm,           0 },
1405     { X86::VPSUBDrr,          X86::VPSUBDrm,           0 },
1406     { X86::VPSUBQrr,          X86::VPSUBQrm,           0 },
1407     { X86::VPSUBSBrr,         X86::VPSUBSBrm,          0 },
1408     { X86::VPSUBSWrr,         X86::VPSUBSWrm,          0 },
1409     { X86::VPSUBUSBrr,        X86::VPSUBUSBrm,         0 },
1410     { X86::VPSUBUSWrr,        X86::VPSUBUSWrm,         0 },
1411     { X86::VPSUBWrr,          X86::VPSUBWrm,           0 },
1412     { X86::VPUNPCKHBWrr,      X86::VPUNPCKHBWrm,       0 },
1413     { X86::VPUNPCKHDQrr,      X86::VPUNPCKHDQrm,       0 },
1414     { X86::VPUNPCKHQDQrr,     X86::VPUNPCKHQDQrm,      0 },
1415     { X86::VPUNPCKHWDrr,      X86::VPUNPCKHWDrm,       0 },
1416     { X86::VPUNPCKLBWrr,      X86::VPUNPCKLBWrm,       0 },
1417     { X86::VPUNPCKLDQrr,      X86::VPUNPCKLDQrm,       0 },
1418     { X86::VPUNPCKLQDQrr,     X86::VPUNPCKLQDQrm,      0 },
1419     { X86::VPUNPCKLWDrr,      X86::VPUNPCKLWDrm,       0 },
1420     { X86::VPXORrr,           X86::VPXORrm,            0 },
1421     { X86::VROUNDSDr,         X86::VROUNDSDm,          0 },
1422     { X86::VROUNDSSr,         X86::VROUNDSSm,          0 },
1423     { X86::VSHUFPDrri,        X86::VSHUFPDrmi,         0 },
1424     { X86::VSHUFPSrri,        X86::VSHUFPSrmi,         0 },
1425     { X86::VSUBPDrr,          X86::VSUBPDrm,           0 },
1426     { X86::VSUBPSrr,          X86::VSUBPSrm,           0 },
1427     { X86::VSUBSDrr,          X86::VSUBSDrm,           0 },
1428     { X86::VSUBSDrr_Int,      X86::VSUBSDrm_Int,       0 },
1429     { X86::VSUBSSrr,          X86::VSUBSSrm,           0 },
1430     { X86::VSUBSSrr_Int,      X86::VSUBSSrm_Int,       0 },
1431     { X86::VUNPCKHPDrr,       X86::VUNPCKHPDrm,        0 },
1432     { X86::VUNPCKHPSrr,       X86::VUNPCKHPSrm,        0 },
1433     { X86::VUNPCKLPDrr,       X86::VUNPCKLPDrm,        0 },
1434     { X86::VUNPCKLPSrr,       X86::VUNPCKLPSrm,        0 },
1435     { X86::VXORPDrr,          X86::VXORPDrm,           0 },
1436     { X86::VXORPSrr,          X86::VXORPSrm,           0 },
1437 
1438     // AVX 256-bit foldable instructions
1439     { X86::VADDPDYrr,         X86::VADDPDYrm,          0 },
1440     { X86::VADDPSYrr,         X86::VADDPSYrm,          0 },
1441     { X86::VADDSUBPDYrr,      X86::VADDSUBPDYrm,       0 },
1442     { X86::VADDSUBPSYrr,      X86::VADDSUBPSYrm,       0 },
1443     { X86::VANDNPDYrr,        X86::VANDNPDYrm,         0 },
1444     { X86::VANDNPSYrr,        X86::VANDNPSYrm,         0 },
1445     { X86::VANDPDYrr,         X86::VANDPDYrm,          0 },
1446     { X86::VANDPSYrr,         X86::VANDPSYrm,          0 },
1447     { X86::VBLENDPDYrri,      X86::VBLENDPDYrmi,       0 },
1448     { X86::VBLENDPSYrri,      X86::VBLENDPSYrmi,       0 },
1449     { X86::VBLENDVPDYrr,      X86::VBLENDVPDYrm,       0 },
1450     { X86::VBLENDVPSYrr,      X86::VBLENDVPSYrm,       0 },
1451     { X86::VCMPPDYrri,        X86::VCMPPDYrmi,         0 },
1452     { X86::VCMPPSYrri,        X86::VCMPPSYrmi,         0 },
1453     { X86::VDIVPDYrr,         X86::VDIVPDYrm,          0 },
1454     { X86::VDIVPSYrr,         X86::VDIVPSYrm,          0 },
1455     { X86::VDPPSYrri,         X86::VDPPSYrmi,          0 },
1456     { X86::VHADDPDYrr,        X86::VHADDPDYrm,         0 },
1457     { X86::VHADDPSYrr,        X86::VHADDPSYrm,         0 },
1458     { X86::VHSUBPDYrr,        X86::VHSUBPDYrm,         0 },
1459     { X86::VHSUBPSYrr,        X86::VHSUBPSYrm,         0 },
1460     { X86::VINSERTF128rr,     X86::VINSERTF128rm,      0 },
1461     { X86::VMAXPDYrr,         X86::VMAXPDYrm,          0 },
1462     { X86::VMAXPSYrr,         X86::VMAXPSYrm,          0 },
1463     { X86::VMINPDYrr,         X86::VMINPDYrm,          0 },
1464     { X86::VMINPSYrr,         X86::VMINPSYrm,          0 },
1465     { X86::VMULPDYrr,         X86::VMULPDYrm,          0 },
1466     { X86::VMULPSYrr,         X86::VMULPSYrm,          0 },
1467     { X86::VORPDYrr,          X86::VORPDYrm,           0 },
1468     { X86::VORPSYrr,          X86::VORPSYrm,           0 },
1469     { X86::VPERM2F128rr,      X86::VPERM2F128rm,       0 },
1470     { X86::VPERMILPDYrr,      X86::VPERMILPDYrm,       0 },
1471     { X86::VPERMILPSYrr,      X86::VPERMILPSYrm,       0 },
1472     { X86::VSHUFPDYrri,       X86::VSHUFPDYrmi,        0 },
1473     { X86::VSHUFPSYrri,       X86::VSHUFPSYrmi,        0 },
1474     { X86::VSUBPDYrr,         X86::VSUBPDYrm,          0 },
1475     { X86::VSUBPSYrr,         X86::VSUBPSYrm,          0 },
1476     { X86::VUNPCKHPDYrr,      X86::VUNPCKHPDYrm,       0 },
1477     { X86::VUNPCKHPSYrr,      X86::VUNPCKHPSYrm,       0 },
1478     { X86::VUNPCKLPDYrr,      X86::VUNPCKLPDYrm,       0 },
1479     { X86::VUNPCKLPSYrr,      X86::VUNPCKLPSYrm,       0 },
1480     { X86::VXORPDYrr,         X86::VXORPDYrm,          0 },
1481     { X86::VXORPSYrr,         X86::VXORPSYrm,          0 },
1482 
1483     // AVX2 foldable instructions
1484     { X86::VINSERTI128rr,     X86::VINSERTI128rm,      0 },
1485     { X86::VPACKSSDWYrr,      X86::VPACKSSDWYrm,       0 },
1486     { X86::VPACKSSWBYrr,      X86::VPACKSSWBYrm,       0 },
1487     { X86::VPACKUSDWYrr,      X86::VPACKUSDWYrm,       0 },
1488     { X86::VPACKUSWBYrr,      X86::VPACKUSWBYrm,       0 },
1489     { X86::VPADDBYrr,         X86::VPADDBYrm,          0 },
1490     { X86::VPADDDYrr,         X86::VPADDDYrm,          0 },
1491     { X86::VPADDQYrr,         X86::VPADDQYrm,          0 },
1492     { X86::VPADDSBYrr,        X86::VPADDSBYrm,         0 },
1493     { X86::VPADDSWYrr,        X86::VPADDSWYrm,         0 },
1494     { X86::VPADDUSBYrr,       X86::VPADDUSBYrm,        0 },
1495     { X86::VPADDUSWYrr,       X86::VPADDUSWYrm,        0 },
1496     { X86::VPADDWYrr,         X86::VPADDWYrm,          0 },
1497     { X86::VPALIGNRYrri,      X86::VPALIGNRYrmi,       0 },
1498     { X86::VPANDNYrr,         X86::VPANDNYrm,          0 },
1499     { X86::VPANDYrr,          X86::VPANDYrm,           0 },
1500     { X86::VPAVGBYrr,         X86::VPAVGBYrm,          0 },
1501     { X86::VPAVGWYrr,         X86::VPAVGWYrm,          0 },
1502     { X86::VPBLENDDrri,       X86::VPBLENDDrmi,        0 },
1503     { X86::VPBLENDDYrri,      X86::VPBLENDDYrmi,       0 },
1504     { X86::VPBLENDVBYrr,      X86::VPBLENDVBYrm,       0 },
1505     { X86::VPBLENDWYrri,      X86::VPBLENDWYrmi,       0 },
1506     { X86::VPCMPEQBYrr,       X86::VPCMPEQBYrm,        0 },
1507     { X86::VPCMPEQDYrr,       X86::VPCMPEQDYrm,        0 },
1508     { X86::VPCMPEQQYrr,       X86::VPCMPEQQYrm,        0 },
1509     { X86::VPCMPEQWYrr,       X86::VPCMPEQWYrm,        0 },
1510     { X86::VPCMPGTBYrr,       X86::VPCMPGTBYrm,        0 },
1511     { X86::VPCMPGTDYrr,       X86::VPCMPGTDYrm,        0 },
1512     { X86::VPCMPGTQYrr,       X86::VPCMPGTQYrm,        0 },
1513     { X86::VPCMPGTWYrr,       X86::VPCMPGTWYrm,        0 },
1514     { X86::VPERM2I128rr,      X86::VPERM2I128rm,       0 },
1515     { X86::VPERMDYrr,         X86::VPERMDYrm,          0 },
1516     { X86::VPERMPSYrr,        X86::VPERMPSYrm,         0 },
1517     { X86::VPHADDDYrr,        X86::VPHADDDYrm,         0 },
1518     { X86::VPHADDSWrr256,     X86::VPHADDSWrm256,      0 },
1519     { X86::VPHADDWYrr,        X86::VPHADDWYrm,         0 },
1520     { X86::VPHSUBDYrr,        X86::VPHSUBDYrm,         0 },
1521     { X86::VPHSUBSWrr256,     X86::VPHSUBSWrm256,      0 },
1522     { X86::VPHSUBWYrr,        X86::VPHSUBWYrm,         0 },
1523     { X86::VPMADDUBSWrr256,   X86::VPMADDUBSWrm256,    0 },
1524     { X86::VPMADDWDYrr,       X86::VPMADDWDYrm,        0 },
1525     { X86::VPMAXSWYrr,        X86::VPMAXSWYrm,         0 },
1526     { X86::VPMAXUBYrr,        X86::VPMAXUBYrm,         0 },
1527     { X86::VPMINSWYrr,        X86::VPMINSWYrm,         0 },
1528     { X86::VPMINUBYrr,        X86::VPMINUBYrm,         0 },
1529     { X86::VPMINSBYrr,        X86::VPMINSBYrm,         0 },
1530     { X86::VPMINSDYrr,        X86::VPMINSDYrm,         0 },
1531     { X86::VPMINUDYrr,        X86::VPMINUDYrm,         0 },
1532     { X86::VPMINUWYrr,        X86::VPMINUWYrm,         0 },
1533     { X86::VPMAXSBYrr,        X86::VPMAXSBYrm,         0 },
1534     { X86::VPMAXSDYrr,        X86::VPMAXSDYrm,         0 },
1535     { X86::VPMAXUDYrr,        X86::VPMAXUDYrm,         0 },
1536     { X86::VPMAXUWYrr,        X86::VPMAXUWYrm,         0 },
1537     { X86::VMPSADBWYrri,      X86::VMPSADBWYrmi,       0 },
1538     { X86::VPMULDQYrr,        X86::VPMULDQYrm,         0 },
1539     { X86::VPMULHRSWrr256,    X86::VPMULHRSWrm256,     0 },
1540     { X86::VPMULHUWYrr,       X86::VPMULHUWYrm,        0 },
1541     { X86::VPMULHWYrr,        X86::VPMULHWYrm,         0 },
1542     { X86::VPMULLDYrr,        X86::VPMULLDYrm,         0 },
1543     { X86::VPMULLWYrr,        X86::VPMULLWYrm,         0 },
1544     { X86::VPMULUDQYrr,       X86::VPMULUDQYrm,        0 },
1545     { X86::VPORYrr,           X86::VPORYrm,            0 },
1546     { X86::VPSADBWYrr,        X86::VPSADBWYrm,         0 },
1547     { X86::VPSHUFBYrr,        X86::VPSHUFBYrm,         0 },
1548     { X86::VPSIGNBYrr256,     X86::VPSIGNBYrm256,      0 },
1549     { X86::VPSIGNWYrr256,     X86::VPSIGNWYrm256,      0 },
1550     { X86::VPSIGNDYrr256,     X86::VPSIGNDYrm256,      0 },
1551     { X86::VPSLLDYrr,         X86::VPSLLDYrm,          0 },
1552     { X86::VPSLLQYrr,         X86::VPSLLQYrm,          0 },
1553     { X86::VPSLLWYrr,         X86::VPSLLWYrm,          0 },
1554     { X86::VPSLLVDrr,         X86::VPSLLVDrm,          0 },
1555     { X86::VPSLLVDYrr,        X86::VPSLLVDYrm,         0 },
1556     { X86::VPSLLVQrr,         X86::VPSLLVQrm,          0 },
1557     { X86::VPSLLVQYrr,        X86::VPSLLVQYrm,         0 },
1558     { X86::VPSRADYrr,         X86::VPSRADYrm,          0 },
1559     { X86::VPSRAWYrr,         X86::VPSRAWYrm,          0 },
1560     { X86::VPSRAVDrr,         X86::VPSRAVDrm,          0 },
1561     { X86::VPSRAVDYrr,        X86::VPSRAVDYrm,         0 },
1562     { X86::VPSRAVD_Intrr,     X86::VPSRAVD_Intrm,      0 },
1563     { X86::VPSRAVD_IntYrr,    X86::VPSRAVD_IntYrm,     0 },
1564     { X86::VPSRLDYrr,         X86::VPSRLDYrm,          0 },
1565     { X86::VPSRLQYrr,         X86::VPSRLQYrm,          0 },
1566     { X86::VPSRLWYrr,         X86::VPSRLWYrm,          0 },
1567     { X86::VPSRLVDrr,         X86::VPSRLVDrm,          0 },
1568     { X86::VPSRLVDYrr,        X86::VPSRLVDYrm,         0 },
1569     { X86::VPSRLVQrr,         X86::VPSRLVQrm,          0 },
1570     { X86::VPSRLVQYrr,        X86::VPSRLVQYrm,         0 },
1571     { X86::VPSUBBYrr,         X86::VPSUBBYrm,          0 },
1572     { X86::VPSUBDYrr,         X86::VPSUBDYrm,          0 },
1573     { X86::VPSUBQYrr,         X86::VPSUBQYrm,          0 },
1574     { X86::VPSUBSBYrr,        X86::VPSUBSBYrm,         0 },
1575     { X86::VPSUBSWYrr,        X86::VPSUBSWYrm,         0 },
1576     { X86::VPSUBUSBYrr,       X86::VPSUBUSBYrm,        0 },
1577     { X86::VPSUBUSWYrr,       X86::VPSUBUSWYrm,        0 },
1578     { X86::VPSUBWYrr,         X86::VPSUBWYrm,          0 },
1579     { X86::VPUNPCKHBWYrr,     X86::VPUNPCKHBWYrm,      0 },
1580     { X86::VPUNPCKHDQYrr,     X86::VPUNPCKHDQYrm,      0 },
1581     { X86::VPUNPCKHQDQYrr,    X86::VPUNPCKHQDQYrm,     0 },
1582     { X86::VPUNPCKHWDYrr,     X86::VPUNPCKHWDYrm,      0 },
1583     { X86::VPUNPCKLBWYrr,     X86::VPUNPCKLBWYrm,      0 },
1584     { X86::VPUNPCKLDQYrr,     X86::VPUNPCKLDQYrm,      0 },
1585     { X86::VPUNPCKLQDQYrr,    X86::VPUNPCKLQDQYrm,     0 },
1586     { X86::VPUNPCKLWDYrr,     X86::VPUNPCKLWDYrm,      0 },
1587     { X86::VPXORYrr,          X86::VPXORYrm,           0 },
1588 
1589     // FMA4 foldable patterns
1590     { X86::VFMADDSS4rr,       X86::VFMADDSS4mr,        TB_ALIGN_NONE },
1591     { X86::VFMADDSD4rr,       X86::VFMADDSD4mr,        TB_ALIGN_NONE },
1592     { X86::VFMADDPS4rr,       X86::VFMADDPS4mr,        TB_ALIGN_NONE },
1593     { X86::VFMADDPD4rr,       X86::VFMADDPD4mr,        TB_ALIGN_NONE },
1594     { X86::VFMADDPS4rrY,      X86::VFMADDPS4mrY,       TB_ALIGN_NONE },
1595     { X86::VFMADDPD4rrY,      X86::VFMADDPD4mrY,       TB_ALIGN_NONE },
1596     { X86::VFNMADDSS4rr,      X86::VFNMADDSS4mr,       TB_ALIGN_NONE },
1597     { X86::VFNMADDSD4rr,      X86::VFNMADDSD4mr,       TB_ALIGN_NONE },
1598     { X86::VFNMADDPS4rr,      X86::VFNMADDPS4mr,       TB_ALIGN_NONE },
1599     { X86::VFNMADDPD4rr,      X86::VFNMADDPD4mr,       TB_ALIGN_NONE },
1600     { X86::VFNMADDPS4rrY,     X86::VFNMADDPS4mrY,      TB_ALIGN_NONE },
1601     { X86::VFNMADDPD4rrY,     X86::VFNMADDPD4mrY,      TB_ALIGN_NONE },
1602     { X86::VFMSUBSS4rr,       X86::VFMSUBSS4mr,        TB_ALIGN_NONE },
1603     { X86::VFMSUBSD4rr,       X86::VFMSUBSD4mr,        TB_ALIGN_NONE },
1604     { X86::VFMSUBPS4rr,       X86::VFMSUBPS4mr,        TB_ALIGN_NONE },
1605     { X86::VFMSUBPD4rr,       X86::VFMSUBPD4mr,        TB_ALIGN_NONE },
1606     { X86::VFMSUBPS4rrY,      X86::VFMSUBPS4mrY,       TB_ALIGN_NONE },
1607     { X86::VFMSUBPD4rrY,      X86::VFMSUBPD4mrY,       TB_ALIGN_NONE },
1608     { X86::VFNMSUBSS4rr,      X86::VFNMSUBSS4mr,       TB_ALIGN_NONE },
1609     { X86::VFNMSUBSD4rr,      X86::VFNMSUBSD4mr,       TB_ALIGN_NONE },
1610     { X86::VFNMSUBPS4rr,      X86::VFNMSUBPS4mr,       TB_ALIGN_NONE },
1611     { X86::VFNMSUBPD4rr,      X86::VFNMSUBPD4mr,       TB_ALIGN_NONE },
1612     { X86::VFNMSUBPS4rrY,     X86::VFNMSUBPS4mrY,      TB_ALIGN_NONE },
1613     { X86::VFNMSUBPD4rrY,     X86::VFNMSUBPD4mrY,      TB_ALIGN_NONE },
1614     { X86::VFMADDSUBPS4rr,    X86::VFMADDSUBPS4mr,     TB_ALIGN_NONE },
1615     { X86::VFMADDSUBPD4rr,    X86::VFMADDSUBPD4mr,     TB_ALIGN_NONE },
1616     { X86::VFMADDSUBPS4rrY,   X86::VFMADDSUBPS4mrY,    TB_ALIGN_NONE },
1617     { X86::VFMADDSUBPD4rrY,   X86::VFMADDSUBPD4mrY,    TB_ALIGN_NONE },
1618     { X86::VFMSUBADDPS4rr,    X86::VFMSUBADDPS4mr,     TB_ALIGN_NONE },
1619     { X86::VFMSUBADDPD4rr,    X86::VFMSUBADDPD4mr,     TB_ALIGN_NONE },
1620     { X86::VFMSUBADDPS4rrY,   X86::VFMSUBADDPS4mrY,    TB_ALIGN_NONE },
1621     { X86::VFMSUBADDPD4rrY,   X86::VFMSUBADDPD4mrY,    TB_ALIGN_NONE },
1622 
1623     // XOP foldable instructions
1624     { X86::VPCMOVrrr,         X86::VPCMOVrmr,           0 },
1625     { X86::VPCMOVrrrY,        X86::VPCMOVrmrY,          0 },
1626     { X86::VPCOMBri,          X86::VPCOMBmi,            0 },
1627     { X86::VPCOMDri,          X86::VPCOMDmi,            0 },
1628     { X86::VPCOMQri,          X86::VPCOMQmi,            0 },
1629     { X86::VPCOMWri,          X86::VPCOMWmi,            0 },
1630     { X86::VPCOMUBri,         X86::VPCOMUBmi,           0 },
1631     { X86::VPCOMUDri,         X86::VPCOMUDmi,           0 },
1632     { X86::VPCOMUQri,         X86::VPCOMUQmi,           0 },
1633     { X86::VPCOMUWri,         X86::VPCOMUWmi,           0 },
1634     { X86::VPERMIL2PDrr,      X86::VPERMIL2PDmr,        0 },
1635     { X86::VPERMIL2PDrrY,     X86::VPERMIL2PDmrY,       0 },
1636     { X86::VPERMIL2PSrr,      X86::VPERMIL2PSmr,        0 },
1637     { X86::VPERMIL2PSrrY,     X86::VPERMIL2PSmrY,       0 },
1638     { X86::VPMACSDDrr,        X86::VPMACSDDrm,          0 },
1639     { X86::VPMACSDQHrr,       X86::VPMACSDQHrm,         0 },
1640     { X86::VPMACSDQLrr,       X86::VPMACSDQLrm,         0 },
1641     { X86::VPMACSSDDrr,       X86::VPMACSSDDrm,         0 },
1642     { X86::VPMACSSDQHrr,      X86::VPMACSSDQHrm,        0 },
1643     { X86::VPMACSSDQLrr,      X86::VPMACSSDQLrm,        0 },
1644     { X86::VPMACSSWDrr,       X86::VPMACSSWDrm,         0 },
1645     { X86::VPMACSSWWrr,       X86::VPMACSSWWrm,         0 },
1646     { X86::VPMACSWDrr,        X86::VPMACSWDrm,          0 },
1647     { X86::VPMACSWWrr,        X86::VPMACSWWrm,          0 },
1648     { X86::VPMADCSSWDrr,      X86::VPMADCSSWDrm,        0 },
1649     { X86::VPMADCSWDrr,       X86::VPMADCSWDrm,         0 },
1650     { X86::VPPERMrrr,         X86::VPPERMrmr,           0 },
1651     { X86::VPROTBrr,          X86::VPROTBrm,            0 },
1652     { X86::VPROTDrr,          X86::VPROTDrm,            0 },
1653     { X86::VPROTQrr,          X86::VPROTQrm,            0 },
1654     { X86::VPROTWrr,          X86::VPROTWrm,            0 },
1655     { X86::VPSHABrr,          X86::VPSHABrm,            0 },
1656     { X86::VPSHADrr,          X86::VPSHADrm,            0 },
1657     { X86::VPSHAQrr,          X86::VPSHAQrm,            0 },
1658     { X86::VPSHAWrr,          X86::VPSHAWrm,            0 },
1659     { X86::VPSHLBrr,          X86::VPSHLBrm,            0 },
1660     { X86::VPSHLDrr,          X86::VPSHLDrm,            0 },
1661     { X86::VPSHLQrr,          X86::VPSHLQrm,            0 },
1662     { X86::VPSHLWrr,          X86::VPSHLWrm,            0 },
1663 
1664     // BMI/BMI2 foldable instructions
1665     { X86::ANDN32rr,          X86::ANDN32rm,            0 },
1666     { X86::ANDN64rr,          X86::ANDN64rm,            0 },
1667     { X86::MULX32rr,          X86::MULX32rm,            0 },
1668     { X86::MULX64rr,          X86::MULX64rm,            0 },
1669     { X86::PDEP32rr,          X86::PDEP32rm,            0 },
1670     { X86::PDEP64rr,          X86::PDEP64rm,            0 },
1671     { X86::PEXT32rr,          X86::PEXT32rm,            0 },
1672     { X86::PEXT64rr,          X86::PEXT64rm,            0 },
1673 
1674     // ADX foldable instructions
1675     { X86::ADCX32rr,          X86::ADCX32rm,            0 },
1676     { X86::ADCX64rr,          X86::ADCX64rm,            0 },
1677     { X86::ADOX32rr,          X86::ADOX32rm,            0 },
1678     { X86::ADOX64rr,          X86::ADOX64rm,            0 },
1679 
1680     // AVX-512 foldable instructions
1681     { X86::VADDPSZrr,         X86::VADDPSZrm,           0 },
1682     { X86::VADDPDZrr,         X86::VADDPDZrm,           0 },
1683     { X86::VSUBPSZrr,         X86::VSUBPSZrm,           0 },
1684     { X86::VSUBPDZrr,         X86::VSUBPDZrm,           0 },
1685     { X86::VMULPSZrr,         X86::VMULPSZrm,           0 },
1686     { X86::VMULPDZrr,         X86::VMULPDZrm,           0 },
1687     { X86::VDIVPSZrr,         X86::VDIVPSZrm,           0 },
1688     { X86::VDIVPDZrr,         X86::VDIVPDZrm,           0 },
1689     { X86::VMINPSZrr,         X86::VMINPSZrm,           0 },
1690     { X86::VMINPDZrr,         X86::VMINPDZrm,           0 },
1691     { X86::VMAXPSZrr,         X86::VMAXPSZrm,           0 },
1692     { X86::VMAXPDZrr,         X86::VMAXPDZrm,           0 },
1693     { X86::VPADDDZrr,         X86::VPADDDZrm,           0 },
1694     { X86::VPADDQZrr,         X86::VPADDQZrm,           0 },
1695     { X86::VPERMPDZri,        X86::VPERMPDZmi,          0 },
1696     { X86::VPERMPSZrr,        X86::VPERMPSZrm,          0 },
1697     { X86::VPMAXSDZrr,        X86::VPMAXSDZrm,          0 },
1698     { X86::VPMAXSQZrr,        X86::VPMAXSQZrm,          0 },
1699     { X86::VPMAXUDZrr,        X86::VPMAXUDZrm,          0 },
1700     { X86::VPMAXUQZrr,        X86::VPMAXUQZrm,          0 },
1701     { X86::VPMINSDZrr,        X86::VPMINSDZrm,          0 },
1702     { X86::VPMINSQZrr,        X86::VPMINSQZrm,          0 },
1703     { X86::VPMINUDZrr,        X86::VPMINUDZrm,          0 },
1704     { X86::VPMINUQZrr,        X86::VPMINUQZrm,          0 },
1705     { X86::VPMULDQZrr,        X86::VPMULDQZrm,          0 },
1706     { X86::VPSLLVDZrr,        X86::VPSLLVDZrm,          0 },
1707     { X86::VPSLLVQZrr,        X86::VPSLLVQZrm,          0 },
1708     { X86::VPSRAVDZrr,        X86::VPSRAVDZrm,          0 },
1709     { X86::VPSRLVDZrr,        X86::VPSRLVDZrm,          0 },
1710     { X86::VPSRLVQZrr,        X86::VPSRLVQZrm,          0 },
1711     { X86::VPSUBDZrr,         X86::VPSUBDZrm,           0 },
1712     { X86::VPSUBQZrr,         X86::VPSUBQZrm,           0 },
1713     { X86::VSHUFPDZrri,       X86::VSHUFPDZrmi,         0 },
1714     { X86::VSHUFPSZrri,       X86::VSHUFPSZrmi,         0 },
1715     { X86::VALIGNQZrri,       X86::VALIGNQZrmi,         0 },
1716     { X86::VALIGNDZrri,       X86::VALIGNDZrmi,         0 },
1717     { X86::VPMULUDQZrr,       X86::VPMULUDQZrm,         0 },
1718     { X86::VBROADCASTSSZrkz,  X86::VBROADCASTSSZmkz,    TB_NO_REVERSE },
1719     { X86::VBROADCASTSDZrkz,  X86::VBROADCASTSDZmkz,    TB_NO_REVERSE },
1720 
1721     // AVX-512{F,VL} foldable instructions
1722     { X86::VBROADCASTSSZ256rkz,  X86::VBROADCASTSSZ256mkz,      TB_NO_REVERSE },
1723     { X86::VBROADCASTSDZ256rkz,  X86::VBROADCASTSDZ256mkz,      TB_NO_REVERSE },
1724     { X86::VBROADCASTSSZ128rkz,  X86::VBROADCASTSSZ128mkz,      TB_NO_REVERSE },
1725 
1726     // AVX-512{F,VL} foldable instructions
1727     { X86::VADDPDZ128rr,      X86::VADDPDZ128rm,        0 },
1728     { X86::VADDPDZ256rr,      X86::VADDPDZ256rm,        0 },
1729     { X86::VADDPSZ128rr,      X86::VADDPSZ128rm,        0 },
1730     { X86::VADDPSZ256rr,      X86::VADDPSZ256rm,        0 },
1731 
1732     // AES foldable instructions
1733     { X86::AESDECLASTrr,      X86::AESDECLASTrm,        TB_ALIGN_16 },
1734     { X86::AESDECrr,          X86::AESDECrm,            TB_ALIGN_16 },
1735     { X86::AESENCLASTrr,      X86::AESENCLASTrm,        TB_ALIGN_16 },
1736     { X86::AESENCrr,          X86::AESENCrm,            TB_ALIGN_16 },
1737     { X86::VAESDECLASTrr,     X86::VAESDECLASTrm,       0 },
1738     { X86::VAESDECrr,         X86::VAESDECrm,           0 },
1739     { X86::VAESENCLASTrr,     X86::VAESENCLASTrm,       0 },
1740     { X86::VAESENCrr,         X86::VAESENCrm,           0 },
1741 
1742     // SHA foldable instructions
1743     { X86::SHA1MSG1rr,        X86::SHA1MSG1rm,          TB_ALIGN_16 },
1744     { X86::SHA1MSG2rr,        X86::SHA1MSG2rm,          TB_ALIGN_16 },
1745     { X86::SHA1NEXTErr,       X86::SHA1NEXTErm,         TB_ALIGN_16 },
1746     { X86::SHA1RNDS4rri,      X86::SHA1RNDS4rmi,        TB_ALIGN_16 },
1747     { X86::SHA256MSG1rr,      X86::SHA256MSG1rm,        TB_ALIGN_16 },
1748     { X86::SHA256MSG2rr,      X86::SHA256MSG2rm,        TB_ALIGN_16 },
1749     { X86::SHA256RNDS2rr,     X86::SHA256RNDS2rm,       TB_ALIGN_16 }
1750   };
1751 
1752   for (X86MemoryFoldTableEntry Entry : MemoryFoldTable2) {
1753     AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1754                   Entry.RegOp, Entry.MemOp,
1755                   // Index 2, folded load
1756                   Entry.Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
1757   }
1758 
1759   static const X86MemoryFoldTableEntry MemoryFoldTable3[] = {
1760     // FMA foldable instructions
1761     { X86::VFMADDSSr231r,         X86::VFMADDSSr231m,         TB_ALIGN_NONE },
1762     { X86::VFMADDSSr231r_Int,     X86::VFMADDSSr231m_Int,     TB_ALIGN_NONE },
1763     { X86::VFMADDSDr231r,         X86::VFMADDSDr231m,         TB_ALIGN_NONE },
1764     { X86::VFMADDSDr231r_Int,     X86::VFMADDSDr231m_Int,     TB_ALIGN_NONE },
1765     { X86::VFMADDSSr132r,         X86::VFMADDSSr132m,         TB_ALIGN_NONE },
1766     { X86::VFMADDSSr132r_Int,     X86::VFMADDSSr132m_Int,     TB_ALIGN_NONE },
1767     { X86::VFMADDSDr132r,         X86::VFMADDSDr132m,         TB_ALIGN_NONE },
1768     { X86::VFMADDSDr132r_Int,     X86::VFMADDSDr132m_Int,     TB_ALIGN_NONE },
1769     { X86::VFMADDSSr213r,         X86::VFMADDSSr213m,         TB_ALIGN_NONE },
1770     { X86::VFMADDSSr213r_Int,     X86::VFMADDSSr213m_Int,     TB_ALIGN_NONE },
1771     { X86::VFMADDSDr213r,         X86::VFMADDSDr213m,         TB_ALIGN_NONE },
1772     { X86::VFMADDSDr213r_Int,     X86::VFMADDSDr213m_Int,     TB_ALIGN_NONE },
1773 
1774     { X86::VFMADDPSr231r,         X86::VFMADDPSr231m,         TB_ALIGN_NONE },
1775     { X86::VFMADDPDr231r,         X86::VFMADDPDr231m,         TB_ALIGN_NONE },
1776     { X86::VFMADDPSr132r,         X86::VFMADDPSr132m,         TB_ALIGN_NONE },
1777     { X86::VFMADDPDr132r,         X86::VFMADDPDr132m,         TB_ALIGN_NONE },
1778     { X86::VFMADDPSr213r,         X86::VFMADDPSr213m,         TB_ALIGN_NONE },
1779     { X86::VFMADDPDr213r,         X86::VFMADDPDr213m,         TB_ALIGN_NONE },
1780     { X86::VFMADDPSr231rY,        X86::VFMADDPSr231mY,        TB_ALIGN_NONE },
1781     { X86::VFMADDPDr231rY,        X86::VFMADDPDr231mY,        TB_ALIGN_NONE },
1782     { X86::VFMADDPSr132rY,        X86::VFMADDPSr132mY,        TB_ALIGN_NONE },
1783     { X86::VFMADDPDr132rY,        X86::VFMADDPDr132mY,        TB_ALIGN_NONE },
1784     { X86::VFMADDPSr213rY,        X86::VFMADDPSr213mY,        TB_ALIGN_NONE },
1785     { X86::VFMADDPDr213rY,        X86::VFMADDPDr213mY,        TB_ALIGN_NONE },
1786 
1787     { X86::VFNMADDSSr231r,        X86::VFNMADDSSr231m,        TB_ALIGN_NONE },
1788     { X86::VFNMADDSSr231r_Int,    X86::VFNMADDSSr231m_Int,    TB_ALIGN_NONE },
1789     { X86::VFNMADDSDr231r,        X86::VFNMADDSDr231m,        TB_ALIGN_NONE },
1790     { X86::VFNMADDSDr231r_Int,    X86::VFNMADDSDr231m_Int,    TB_ALIGN_NONE },
1791     { X86::VFNMADDSSr132r,        X86::VFNMADDSSr132m,        TB_ALIGN_NONE },
1792     { X86::VFNMADDSSr132r_Int,    X86::VFNMADDSSr132m_Int,    TB_ALIGN_NONE },
1793     { X86::VFNMADDSDr132r,        X86::VFNMADDSDr132m,        TB_ALIGN_NONE },
1794     { X86::VFNMADDSDr132r_Int,    X86::VFNMADDSDr132m_Int,    TB_ALIGN_NONE },
1795     { X86::VFNMADDSSr213r,        X86::VFNMADDSSr213m,        TB_ALIGN_NONE },
1796     { X86::VFNMADDSSr213r_Int,    X86::VFNMADDSSr213m_Int,    TB_ALIGN_NONE },
1797     { X86::VFNMADDSDr213r,        X86::VFNMADDSDr213m,        TB_ALIGN_NONE },
1798     { X86::VFNMADDSDr213r_Int,    X86::VFNMADDSDr213m_Int,    TB_ALIGN_NONE },
1799 
1800     { X86::VFNMADDPSr231r,        X86::VFNMADDPSr231m,        TB_ALIGN_NONE },
1801     { X86::VFNMADDPDr231r,        X86::VFNMADDPDr231m,        TB_ALIGN_NONE },
1802     { X86::VFNMADDPSr132r,        X86::VFNMADDPSr132m,        TB_ALIGN_NONE },
1803     { X86::VFNMADDPDr132r,        X86::VFNMADDPDr132m,        TB_ALIGN_NONE },
1804     { X86::VFNMADDPSr213r,        X86::VFNMADDPSr213m,        TB_ALIGN_NONE },
1805     { X86::VFNMADDPDr213r,        X86::VFNMADDPDr213m,        TB_ALIGN_NONE },
1806     { X86::VFNMADDPSr231rY,       X86::VFNMADDPSr231mY,       TB_ALIGN_NONE },
1807     { X86::VFNMADDPDr231rY,       X86::VFNMADDPDr231mY,       TB_ALIGN_NONE },
1808     { X86::VFNMADDPSr132rY,       X86::VFNMADDPSr132mY,       TB_ALIGN_NONE },
1809     { X86::VFNMADDPDr132rY,       X86::VFNMADDPDr132mY,       TB_ALIGN_NONE },
1810     { X86::VFNMADDPSr213rY,       X86::VFNMADDPSr213mY,       TB_ALIGN_NONE },
1811     { X86::VFNMADDPDr213rY,       X86::VFNMADDPDr213mY,       TB_ALIGN_NONE },
1812 
1813     { X86::VFMSUBSSr231r,         X86::VFMSUBSSr231m,         TB_ALIGN_NONE },
1814     { X86::VFMSUBSSr231r_Int,     X86::VFMSUBSSr231m_Int,     TB_ALIGN_NONE },
1815     { X86::VFMSUBSDr231r,         X86::VFMSUBSDr231m,         TB_ALIGN_NONE },
1816     { X86::VFMSUBSDr231r_Int,     X86::VFMSUBSDr231m_Int,     TB_ALIGN_NONE },
1817     { X86::VFMSUBSSr132r,         X86::VFMSUBSSr132m,         TB_ALIGN_NONE },
1818     { X86::VFMSUBSSr132r_Int,     X86::VFMSUBSSr132m_Int,     TB_ALIGN_NONE },
1819     { X86::VFMSUBSDr132r,         X86::VFMSUBSDr132m,         TB_ALIGN_NONE },
1820     { X86::VFMSUBSDr132r_Int,     X86::VFMSUBSDr132m_Int,     TB_ALIGN_NONE },
1821     { X86::VFMSUBSSr213r,         X86::VFMSUBSSr213m,         TB_ALIGN_NONE },
1822     { X86::VFMSUBSSr213r_Int,     X86::VFMSUBSSr213m_Int,     TB_ALIGN_NONE },
1823     { X86::VFMSUBSDr213r,         X86::VFMSUBSDr213m,         TB_ALIGN_NONE },
1824     { X86::VFMSUBSDr213r_Int,     X86::VFMSUBSDr213m_Int,     TB_ALIGN_NONE },
1825 
1826     { X86::VFMSUBPSr231r,         X86::VFMSUBPSr231m,         TB_ALIGN_NONE },
1827     { X86::VFMSUBPDr231r,         X86::VFMSUBPDr231m,         TB_ALIGN_NONE },
1828     { X86::VFMSUBPSr132r,         X86::VFMSUBPSr132m,         TB_ALIGN_NONE },
1829     { X86::VFMSUBPDr132r,         X86::VFMSUBPDr132m,         TB_ALIGN_NONE },
1830     { X86::VFMSUBPSr213r,         X86::VFMSUBPSr213m,         TB_ALIGN_NONE },
1831     { X86::VFMSUBPDr213r,         X86::VFMSUBPDr213m,         TB_ALIGN_NONE },
1832     { X86::VFMSUBPSr231rY,        X86::VFMSUBPSr231mY,        TB_ALIGN_NONE },
1833     { X86::VFMSUBPDr231rY,        X86::VFMSUBPDr231mY,        TB_ALIGN_NONE },
1834     { X86::VFMSUBPSr132rY,        X86::VFMSUBPSr132mY,        TB_ALIGN_NONE },
1835     { X86::VFMSUBPDr132rY,        X86::VFMSUBPDr132mY,        TB_ALIGN_NONE },
1836     { X86::VFMSUBPSr213rY,        X86::VFMSUBPSr213mY,        TB_ALIGN_NONE },
1837     { X86::VFMSUBPDr213rY,        X86::VFMSUBPDr213mY,        TB_ALIGN_NONE },
1838 
1839     { X86::VFNMSUBSSr231r,        X86::VFNMSUBSSr231m,        TB_ALIGN_NONE },
1840     { X86::VFNMSUBSSr231r_Int,    X86::VFNMSUBSSr231m_Int,    TB_ALIGN_NONE },
1841     { X86::VFNMSUBSDr231r,        X86::VFNMSUBSDr231m,        TB_ALIGN_NONE },
1842     { X86::VFNMSUBSDr231r_Int,    X86::VFNMSUBSDr231m_Int,    TB_ALIGN_NONE },
1843     { X86::VFNMSUBSSr132r,        X86::VFNMSUBSSr132m,        TB_ALIGN_NONE },
1844     { X86::VFNMSUBSSr132r_Int,    X86::VFNMSUBSSr132m_Int,    TB_ALIGN_NONE },
1845     { X86::VFNMSUBSDr132r,        X86::VFNMSUBSDr132m,        TB_ALIGN_NONE },
1846     { X86::VFNMSUBSDr132r_Int,    X86::VFNMSUBSDr132m_Int,    TB_ALIGN_NONE },
1847     { X86::VFNMSUBSSr213r,        X86::VFNMSUBSSr213m,        TB_ALIGN_NONE },
1848     { X86::VFNMSUBSSr213r_Int,    X86::VFNMSUBSSr213m_Int,    TB_ALIGN_NONE },
1849     { X86::VFNMSUBSDr213r,        X86::VFNMSUBSDr213m,        TB_ALIGN_NONE },
1850     { X86::VFNMSUBSDr213r_Int,    X86::VFNMSUBSDr213m_Int,    TB_ALIGN_NONE },
1851 
1852     { X86::VFNMSUBPSr231r,        X86::VFNMSUBPSr231m,        TB_ALIGN_NONE },
1853     { X86::VFNMSUBPDr231r,        X86::VFNMSUBPDr231m,        TB_ALIGN_NONE },
1854     { X86::VFNMSUBPSr132r,        X86::VFNMSUBPSr132m,        TB_ALIGN_NONE },
1855     { X86::VFNMSUBPDr132r,        X86::VFNMSUBPDr132m,        TB_ALIGN_NONE },
1856     { X86::VFNMSUBPSr213r,        X86::VFNMSUBPSr213m,        TB_ALIGN_NONE },
1857     { X86::VFNMSUBPDr213r,        X86::VFNMSUBPDr213m,        TB_ALIGN_NONE },
1858     { X86::VFNMSUBPSr231rY,       X86::VFNMSUBPSr231mY,       TB_ALIGN_NONE },
1859     { X86::VFNMSUBPDr231rY,       X86::VFNMSUBPDr231mY,       TB_ALIGN_NONE },
1860     { X86::VFNMSUBPSr132rY,       X86::VFNMSUBPSr132mY,       TB_ALIGN_NONE },
1861     { X86::VFNMSUBPDr132rY,       X86::VFNMSUBPDr132mY,       TB_ALIGN_NONE },
1862     { X86::VFNMSUBPSr213rY,       X86::VFNMSUBPSr213mY,       TB_ALIGN_NONE },
1863     { X86::VFNMSUBPDr213rY,       X86::VFNMSUBPDr213mY,       TB_ALIGN_NONE },
1864 
1865     { X86::VFMADDSUBPSr231r,      X86::VFMADDSUBPSr231m,      TB_ALIGN_NONE },
1866     { X86::VFMADDSUBPDr231r,      X86::VFMADDSUBPDr231m,      TB_ALIGN_NONE },
1867     { X86::VFMADDSUBPSr132r,      X86::VFMADDSUBPSr132m,      TB_ALIGN_NONE },
1868     { X86::VFMADDSUBPDr132r,      X86::VFMADDSUBPDr132m,      TB_ALIGN_NONE },
1869     { X86::VFMADDSUBPSr213r,      X86::VFMADDSUBPSr213m,      TB_ALIGN_NONE },
1870     { X86::VFMADDSUBPDr213r,      X86::VFMADDSUBPDr213m,      TB_ALIGN_NONE },
1871     { X86::VFMADDSUBPSr231rY,     X86::VFMADDSUBPSr231mY,     TB_ALIGN_NONE },
1872     { X86::VFMADDSUBPDr231rY,     X86::VFMADDSUBPDr231mY,     TB_ALIGN_NONE },
1873     { X86::VFMADDSUBPSr132rY,     X86::VFMADDSUBPSr132mY,     TB_ALIGN_NONE },
1874     { X86::VFMADDSUBPDr132rY,     X86::VFMADDSUBPDr132mY,     TB_ALIGN_NONE },
1875     { X86::VFMADDSUBPSr213rY,     X86::VFMADDSUBPSr213mY,     TB_ALIGN_NONE },
1876     { X86::VFMADDSUBPDr213rY,     X86::VFMADDSUBPDr213mY,     TB_ALIGN_NONE },
1877 
1878     { X86::VFMSUBADDPSr231r,      X86::VFMSUBADDPSr231m,      TB_ALIGN_NONE },
1879     { X86::VFMSUBADDPDr231r,      X86::VFMSUBADDPDr231m,      TB_ALIGN_NONE },
1880     { X86::VFMSUBADDPSr132r,      X86::VFMSUBADDPSr132m,      TB_ALIGN_NONE },
1881     { X86::VFMSUBADDPDr132r,      X86::VFMSUBADDPDr132m,      TB_ALIGN_NONE },
1882     { X86::VFMSUBADDPSr213r,      X86::VFMSUBADDPSr213m,      TB_ALIGN_NONE },
1883     { X86::VFMSUBADDPDr213r,      X86::VFMSUBADDPDr213m,      TB_ALIGN_NONE },
1884     { X86::VFMSUBADDPSr231rY,     X86::VFMSUBADDPSr231mY,     TB_ALIGN_NONE },
1885     { X86::VFMSUBADDPDr231rY,     X86::VFMSUBADDPDr231mY,     TB_ALIGN_NONE },
1886     { X86::VFMSUBADDPSr132rY,     X86::VFMSUBADDPSr132mY,     TB_ALIGN_NONE },
1887     { X86::VFMSUBADDPDr132rY,     X86::VFMSUBADDPDr132mY,     TB_ALIGN_NONE },
1888     { X86::VFMSUBADDPSr213rY,     X86::VFMSUBADDPSr213mY,     TB_ALIGN_NONE },
1889     { X86::VFMSUBADDPDr213rY,     X86::VFMSUBADDPDr213mY,     TB_ALIGN_NONE },
1890 
1891     // FMA4 foldable patterns
1892     { X86::VFMADDSS4rr,           X86::VFMADDSS4rm,           TB_ALIGN_NONE },
1893     { X86::VFMADDSD4rr,           X86::VFMADDSD4rm,           TB_ALIGN_NONE },
1894     { X86::VFMADDPS4rr,           X86::VFMADDPS4rm,           TB_ALIGN_NONE },
1895     { X86::VFMADDPD4rr,           X86::VFMADDPD4rm,           TB_ALIGN_NONE },
1896     { X86::VFMADDPS4rrY,          X86::VFMADDPS4rmY,          TB_ALIGN_NONE },
1897     { X86::VFMADDPD4rrY,          X86::VFMADDPD4rmY,          TB_ALIGN_NONE },
1898     { X86::VFNMADDSS4rr,          X86::VFNMADDSS4rm,          TB_ALIGN_NONE },
1899     { X86::VFNMADDSD4rr,          X86::VFNMADDSD4rm,          TB_ALIGN_NONE },
1900     { X86::VFNMADDPS4rr,          X86::VFNMADDPS4rm,          TB_ALIGN_NONE },
1901     { X86::VFNMADDPD4rr,          X86::VFNMADDPD4rm,          TB_ALIGN_NONE },
1902     { X86::VFNMADDPS4rrY,         X86::VFNMADDPS4rmY,         TB_ALIGN_NONE },
1903     { X86::VFNMADDPD4rrY,         X86::VFNMADDPD4rmY,         TB_ALIGN_NONE },
1904     { X86::VFMSUBSS4rr,           X86::VFMSUBSS4rm,           TB_ALIGN_NONE },
1905     { X86::VFMSUBSD4rr,           X86::VFMSUBSD4rm,           TB_ALIGN_NONE },
1906     { X86::VFMSUBPS4rr,           X86::VFMSUBPS4rm,           TB_ALIGN_NONE },
1907     { X86::VFMSUBPD4rr,           X86::VFMSUBPD4rm,           TB_ALIGN_NONE },
1908     { X86::VFMSUBPS4rrY,          X86::VFMSUBPS4rmY,          TB_ALIGN_NONE },
1909     { X86::VFMSUBPD4rrY,          X86::VFMSUBPD4rmY,          TB_ALIGN_NONE },
1910     { X86::VFNMSUBSS4rr,          X86::VFNMSUBSS4rm,          TB_ALIGN_NONE },
1911     { X86::VFNMSUBSD4rr,          X86::VFNMSUBSD4rm,          TB_ALIGN_NONE },
1912     { X86::VFNMSUBPS4rr,          X86::VFNMSUBPS4rm,          TB_ALIGN_NONE },
1913     { X86::VFNMSUBPD4rr,          X86::VFNMSUBPD4rm,          TB_ALIGN_NONE },
1914     { X86::VFNMSUBPS4rrY,         X86::VFNMSUBPS4rmY,         TB_ALIGN_NONE },
1915     { X86::VFNMSUBPD4rrY,         X86::VFNMSUBPD4rmY,         TB_ALIGN_NONE },
1916     { X86::VFMADDSUBPS4rr,        X86::VFMADDSUBPS4rm,        TB_ALIGN_NONE },
1917     { X86::VFMADDSUBPD4rr,        X86::VFMADDSUBPD4rm,        TB_ALIGN_NONE },
1918     { X86::VFMADDSUBPS4rrY,       X86::VFMADDSUBPS4rmY,       TB_ALIGN_NONE },
1919     { X86::VFMADDSUBPD4rrY,       X86::VFMADDSUBPD4rmY,       TB_ALIGN_NONE },
1920     { X86::VFMSUBADDPS4rr,        X86::VFMSUBADDPS4rm,        TB_ALIGN_NONE },
1921     { X86::VFMSUBADDPD4rr,        X86::VFMSUBADDPD4rm,        TB_ALIGN_NONE },
1922     { X86::VFMSUBADDPS4rrY,       X86::VFMSUBADDPS4rmY,       TB_ALIGN_NONE },
1923     { X86::VFMSUBADDPD4rrY,       X86::VFMSUBADDPD4rmY,       TB_ALIGN_NONE },
1924 
1925     // XOP foldable instructions
1926     { X86::VPCMOVrrr,             X86::VPCMOVrrm,             0 },
1927     { X86::VPCMOVrrrY,            X86::VPCMOVrrmY,            0 },
1928     { X86::VPERMIL2PDrr,          X86::VPERMIL2PDrm,          0 },
1929     { X86::VPERMIL2PDrrY,         X86::VPERMIL2PDrmY,         0 },
1930     { X86::VPERMIL2PSrr,          X86::VPERMIL2PSrm,          0 },
1931     { X86::VPERMIL2PSrrY,         X86::VPERMIL2PSrmY,         0 },
1932     { X86::VPPERMrrr,             X86::VPPERMrrm,             0 },
1933 
1934     // AVX-512 VPERMI instructions with 3 source operands.
1935     { X86::VPERMI2Drr,            X86::VPERMI2Drm,            0 },
1936     { X86::VPERMI2Qrr,            X86::VPERMI2Qrm,            0 },
1937     { X86::VPERMI2PSrr,           X86::VPERMI2PSrm,           0 },
1938     { X86::VPERMI2PDrr,           X86::VPERMI2PDrm,           0 },
1939     { X86::VBLENDMPDZrr,          X86::VBLENDMPDZrm,          0 },
1940     { X86::VBLENDMPSZrr,          X86::VBLENDMPSZrm,          0 },
1941     { X86::VPBLENDMDZrr,          X86::VPBLENDMDZrm,          0 },
1942     { X86::VPBLENDMQZrr,          X86::VPBLENDMQZrm,          0 },
1943     { X86::VBROADCASTSSZrk,       X86::VBROADCASTSSZmk,       TB_NO_REVERSE },
1944     { X86::VBROADCASTSDZrk,       X86::VBROADCASTSDZmk,       TB_NO_REVERSE },
1945     { X86::VBROADCASTSSZ256rk,    X86::VBROADCASTSSZ256mk,    TB_NO_REVERSE },
1946     { X86::VBROADCASTSDZ256rk,    X86::VBROADCASTSDZ256mk,    TB_NO_REVERSE },
1947     { X86::VBROADCASTSSZ128rk,    X86::VBROADCASTSSZ128mk,    TB_NO_REVERSE },
1948      // AVX-512 arithmetic instructions
1949     { X86::VADDPSZrrkz,           X86::VADDPSZrmkz,           0 },
1950     { X86::VADDPDZrrkz,           X86::VADDPDZrmkz,           0 },
1951     { X86::VSUBPSZrrkz,           X86::VSUBPSZrmkz,           0 },
1952     { X86::VSUBPDZrrkz,           X86::VSUBPDZrmkz,           0 },
1953     { X86::VMULPSZrrkz,           X86::VMULPSZrmkz,           0 },
1954     { X86::VMULPDZrrkz,           X86::VMULPDZrmkz,           0 },
1955     { X86::VDIVPSZrrkz,           X86::VDIVPSZrmkz,           0 },
1956     { X86::VDIVPDZrrkz,           X86::VDIVPDZrmkz,           0 },
1957     { X86::VMINPSZrrkz,           X86::VMINPSZrmkz,           0 },
1958     { X86::VMINPDZrrkz,           X86::VMINPDZrmkz,           0 },
1959     { X86::VMAXPSZrrkz,           X86::VMAXPSZrmkz,           0 },
1960     { X86::VMAXPDZrrkz,           X86::VMAXPDZrmkz,           0 },
1961     // AVX-512{F,VL} arithmetic instructions 256-bit
1962     { X86::VADDPSZ256rrkz,        X86::VADDPSZ256rmkz,        0 },
1963     { X86::VADDPDZ256rrkz,        X86::VADDPDZ256rmkz,        0 },
1964     { X86::VSUBPSZ256rrkz,        X86::VSUBPSZ256rmkz,        0 },
1965     { X86::VSUBPDZ256rrkz,        X86::VSUBPDZ256rmkz,        0 },
1966     { X86::VMULPSZ256rrkz,        X86::VMULPSZ256rmkz,        0 },
1967     { X86::VMULPDZ256rrkz,        X86::VMULPDZ256rmkz,        0 },
1968     { X86::VDIVPSZ256rrkz,        X86::VDIVPSZ256rmkz,        0 },
1969     { X86::VDIVPDZ256rrkz,        X86::VDIVPDZ256rmkz,        0 },
1970     { X86::VMINPSZ256rrkz,        X86::VMINPSZ256rmkz,        0 },
1971     { X86::VMINPDZ256rrkz,        X86::VMINPDZ256rmkz,        0 },
1972     { X86::VMAXPSZ256rrkz,        X86::VMAXPSZ256rmkz,        0 },
1973     { X86::VMAXPDZ256rrkz,        X86::VMAXPDZ256rmkz,        0 },
1974     // AVX-512{F,VL} arithmetic instructions 128-bit
1975     { X86::VADDPSZ128rrkz,        X86::VADDPSZ128rmkz,        0 },
1976     { X86::VADDPDZ128rrkz,        X86::VADDPDZ128rmkz,        0 },
1977     { X86::VSUBPSZ128rrkz,        X86::VSUBPSZ128rmkz,        0 },
1978     { X86::VSUBPDZ128rrkz,        X86::VSUBPDZ128rmkz,        0 },
1979     { X86::VMULPSZ128rrkz,        X86::VMULPSZ128rmkz,        0 },
1980     { X86::VMULPDZ128rrkz,        X86::VMULPDZ128rmkz,        0 },
1981     { X86::VDIVPSZ128rrkz,        X86::VDIVPSZ128rmkz,        0 },
1982     { X86::VDIVPDZ128rrkz,        X86::VDIVPDZ128rmkz,        0 },
1983     { X86::VMINPSZ128rrkz,        X86::VMINPSZ128rmkz,        0 },
1984     { X86::VMINPDZ128rrkz,        X86::VMINPDZ128rmkz,        0 },
1985     { X86::VMAXPSZ128rrkz,        X86::VMAXPSZ128rmkz,        0 },
1986     { X86::VMAXPDZ128rrkz,        X86::VMAXPDZ128rmkz,        0 }
1987   };
1988 
1989   for (X86MemoryFoldTableEntry Entry : MemoryFoldTable3) {
1990     AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1991                   Entry.RegOp, Entry.MemOp,
1992                   // Index 3, folded load
1993                   Entry.Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1994   }
1995 
1996   static const X86MemoryFoldTableEntry MemoryFoldTable4[] = {
1997      // AVX-512 foldable instructions
1998     { X86::VADDPSZrrk,         X86::VADDPSZrmk,           0 },
1999     { X86::VADDPDZrrk,         X86::VADDPDZrmk,           0 },
2000     { X86::VSUBPSZrrk,         X86::VSUBPSZrmk,           0 },
2001     { X86::VSUBPDZrrk,         X86::VSUBPDZrmk,           0 },
2002     { X86::VMULPSZrrk,         X86::VMULPSZrmk,           0 },
2003     { X86::VMULPDZrrk,         X86::VMULPDZrmk,           0 },
2004     { X86::VDIVPSZrrk,         X86::VDIVPSZrmk,           0 },
2005     { X86::VDIVPDZrrk,         X86::VDIVPDZrmk,           0 },
2006     { X86::VMINPSZrrk,         X86::VMINPSZrmk,           0 },
2007     { X86::VMINPDZrrk,         X86::VMINPDZrmk,           0 },
2008     { X86::VMAXPSZrrk,         X86::VMAXPSZrmk,           0 },
2009     { X86::VMAXPDZrrk,         X86::VMAXPDZrmk,           0 },
2010     // AVX-512{F,VL} foldable instructions 256-bit
2011     { X86::VADDPSZ256rrk,      X86::VADDPSZ256rmk,        0 },
2012     { X86::VADDPDZ256rrk,      X86::VADDPDZ256rmk,        0 },
2013     { X86::VSUBPSZ256rrk,      X86::VSUBPSZ256rmk,        0 },
2014     { X86::VSUBPDZ256rrk,      X86::VSUBPDZ256rmk,        0 },
2015     { X86::VMULPSZ256rrk,      X86::VMULPSZ256rmk,        0 },
2016     { X86::VMULPDZ256rrk,      X86::VMULPDZ256rmk,        0 },
2017     { X86::VDIVPSZ256rrk,      X86::VDIVPSZ256rmk,        0 },
2018     { X86::VDIVPDZ256rrk,      X86::VDIVPDZ256rmk,        0 },
2019     { X86::VMINPSZ256rrk,      X86::VMINPSZ256rmk,        0 },
2020     { X86::VMINPDZ256rrk,      X86::VMINPDZ256rmk,        0 },
2021     { X86::VMAXPSZ256rrk,      X86::VMAXPSZ256rmk,        0 },
2022     { X86::VMAXPDZ256rrk,      X86::VMAXPDZ256rmk,        0 },
2023     // AVX-512{F,VL} foldable instructions 128-bit
2024     { X86::VADDPSZ128rrk,      X86::VADDPSZ128rmk,        0 },
2025     { X86::VADDPDZ128rrk,      X86::VADDPDZ128rmk,        0 },
2026     { X86::VSUBPSZ128rrk,      X86::VSUBPSZ128rmk,        0 },
2027     { X86::VSUBPDZ128rrk,      X86::VSUBPDZ128rmk,        0 },
2028     { X86::VMULPSZ128rrk,      X86::VMULPSZ128rmk,        0 },
2029     { X86::VMULPDZ128rrk,      X86::VMULPDZ128rmk,        0 },
2030     { X86::VDIVPSZ128rrk,      X86::VDIVPSZ128rmk,        0 },
2031     { X86::VDIVPDZ128rrk,      X86::VDIVPDZ128rmk,        0 },
2032     { X86::VMINPSZ128rrk,      X86::VMINPSZ128rmk,        0 },
2033     { X86::VMINPDZ128rrk,      X86::VMINPDZ128rmk,        0 },
2034     { X86::VMAXPSZ128rrk,      X86::VMAXPSZ128rmk,        0 },
2035     { X86::VMAXPDZ128rrk,      X86::VMAXPDZ128rmk,        0 }
2036   };
2037 
2038   for (X86MemoryFoldTableEntry Entry : MemoryFoldTable4) {
2039     AddTableEntry(RegOp2MemOpTable4, MemOp2RegOpTable,
2040                   Entry.RegOp, Entry.MemOp,
2041                   // Index 4, folded load
2042                   Entry.Flags | TB_INDEX_4 | TB_FOLDED_LOAD);
2043   }
2044 }
2045 
2046 void
AddTableEntry(RegOp2MemOpTableType & R2MTable,MemOp2RegOpTableType & M2RTable,uint16_t RegOp,uint16_t MemOp,uint16_t Flags)2047 X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
2048                             MemOp2RegOpTableType &M2RTable,
2049                             uint16_t RegOp, uint16_t MemOp, uint16_t Flags) {
2050     if ((Flags & TB_NO_FORWARD) == 0) {
2051       assert(!R2MTable.count(RegOp) && "Duplicate entry!");
2052       R2MTable[RegOp] = std::make_pair(MemOp, Flags);
2053     }
2054     if ((Flags & TB_NO_REVERSE) == 0) {
2055       assert(!M2RTable.count(MemOp) &&
2056            "Duplicated entries in unfolding maps?");
2057       M2RTable[MemOp] = std::make_pair(RegOp, Flags);
2058     }
2059 }
2060 
2061 bool
isCoalescableExtInstr(const MachineInstr & MI,unsigned & SrcReg,unsigned & DstReg,unsigned & SubIdx) const2062 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
2063                                     unsigned &SrcReg, unsigned &DstReg,
2064                                     unsigned &SubIdx) const {
2065   switch (MI.getOpcode()) {
2066   default: break;
2067   case X86::MOVSX16rr8:
2068   case X86::MOVZX16rr8:
2069   case X86::MOVSX32rr8:
2070   case X86::MOVZX32rr8:
2071   case X86::MOVSX64rr8:
2072     if (!Subtarget.is64Bit())
2073       // It's not always legal to reference the low 8-bit of the larger
2074       // register in 32-bit mode.
2075       return false;
2076   case X86::MOVSX32rr16:
2077   case X86::MOVZX32rr16:
2078   case X86::MOVSX64rr16:
2079   case X86::MOVSX64rr32: {
2080     if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
2081       // Be conservative.
2082       return false;
2083     SrcReg = MI.getOperand(1).getReg();
2084     DstReg = MI.getOperand(0).getReg();
2085     switch (MI.getOpcode()) {
2086     default: llvm_unreachable("Unreachable!");
2087     case X86::MOVSX16rr8:
2088     case X86::MOVZX16rr8:
2089     case X86::MOVSX32rr8:
2090     case X86::MOVZX32rr8:
2091     case X86::MOVSX64rr8:
2092       SubIdx = X86::sub_8bit;
2093       break;
2094     case X86::MOVSX32rr16:
2095     case X86::MOVZX32rr16:
2096     case X86::MOVSX64rr16:
2097       SubIdx = X86::sub_16bit;
2098       break;
2099     case X86::MOVSX64rr32:
2100       SubIdx = X86::sub_32bit;
2101       break;
2102     }
2103     return true;
2104   }
2105   }
2106   return false;
2107 }
2108 
getSPAdjust(const MachineInstr & MI) const2109 int X86InstrInfo::getSPAdjust(const MachineInstr &MI) const {
2110   const MachineFunction *MF = MI.getParent()->getParent();
2111   const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
2112 
2113   if (MI.getOpcode() == getCallFrameSetupOpcode() ||
2114       MI.getOpcode() == getCallFrameDestroyOpcode()) {
2115     unsigned StackAlign = TFI->getStackAlignment();
2116     int SPAdj =
2117         (MI.getOperand(0).getImm() + StackAlign - 1) / StackAlign * StackAlign;
2118 
2119     SPAdj -= MI.getOperand(1).getImm();
2120 
2121     if (MI.getOpcode() == getCallFrameSetupOpcode())
2122       return SPAdj;
2123     else
2124       return -SPAdj;
2125   }
2126 
2127   // To know whether a call adjusts the stack, we need information
2128   // that is bound to the following ADJCALLSTACKUP pseudo.
2129   // Look for the next ADJCALLSTACKUP that follows the call.
2130   if (MI.isCall()) {
2131     const MachineBasicBlock *MBB = MI.getParent();
2132     auto I = ++MachineBasicBlock::const_iterator(MI);
2133     for (auto E = MBB->end(); I != E; ++I) {
2134       if (I->getOpcode() == getCallFrameDestroyOpcode() ||
2135           I->isCall())
2136         break;
2137     }
2138 
2139     // If we could not find a frame destroy opcode, then it has already
2140     // been simplified, so we don't care.
2141     if (I->getOpcode() != getCallFrameDestroyOpcode())
2142       return 0;
2143 
2144     return -(I->getOperand(1).getImm());
2145   }
2146 
2147   // Currently handle only PUSHes we can reasonably expect to see
2148   // in call sequences
2149   switch (MI.getOpcode()) {
2150   default:
2151     return 0;
2152   case X86::PUSH32i8:
2153   case X86::PUSH32r:
2154   case X86::PUSH32rmm:
2155   case X86::PUSH32rmr:
2156   case X86::PUSHi32:
2157     return 4;
2158   case X86::PUSH64i8:
2159   case X86::PUSH64r:
2160   case X86::PUSH64rmm:
2161   case X86::PUSH64rmr:
2162   case X86::PUSH64i32:
2163     return 8;
2164   }
2165 }
2166 
2167 /// Return true and the FrameIndex if the specified
2168 /// operand and follow operands form a reference to the stack frame.
isFrameOperand(const MachineInstr & MI,unsigned int Op,int & FrameIndex) const2169 bool X86InstrInfo::isFrameOperand(const MachineInstr &MI, unsigned int Op,
2170                                   int &FrameIndex) const {
2171   if (MI.getOperand(Op + X86::AddrBaseReg).isFI() &&
2172       MI.getOperand(Op + X86::AddrScaleAmt).isImm() &&
2173       MI.getOperand(Op + X86::AddrIndexReg).isReg() &&
2174       MI.getOperand(Op + X86::AddrDisp).isImm() &&
2175       MI.getOperand(Op + X86::AddrScaleAmt).getImm() == 1 &&
2176       MI.getOperand(Op + X86::AddrIndexReg).getReg() == 0 &&
2177       MI.getOperand(Op + X86::AddrDisp).getImm() == 0) {
2178     FrameIndex = MI.getOperand(Op + X86::AddrBaseReg).getIndex();
2179     return true;
2180   }
2181   return false;
2182 }
2183 
isFrameLoadOpcode(int Opcode)2184 static bool isFrameLoadOpcode(int Opcode) {
2185   switch (Opcode) {
2186   default:
2187     return false;
2188   case X86::MOV8rm:
2189   case X86::MOV16rm:
2190   case X86::MOV32rm:
2191   case X86::MOV64rm:
2192   case X86::LD_Fp64m:
2193   case X86::MOVSSrm:
2194   case X86::MOVSDrm:
2195   case X86::MOVAPSrm:
2196   case X86::MOVAPDrm:
2197   case X86::MOVDQArm:
2198   case X86::VMOVSSrm:
2199   case X86::VMOVSDrm:
2200   case X86::VMOVAPSrm:
2201   case X86::VMOVAPDrm:
2202   case X86::VMOVDQArm:
2203   case X86::VMOVUPSYrm:
2204   case X86::VMOVAPSYrm:
2205   case X86::VMOVUPDYrm:
2206   case X86::VMOVAPDYrm:
2207   case X86::VMOVDQUYrm:
2208   case X86::VMOVDQAYrm:
2209   case X86::MMX_MOVD64rm:
2210   case X86::MMX_MOVQ64rm:
2211   case X86::VMOVAPSZrm:
2212   case X86::VMOVUPSZrm:
2213     return true;
2214   }
2215 }
2216 
isFrameStoreOpcode(int Opcode)2217 static bool isFrameStoreOpcode(int Opcode) {
2218   switch (Opcode) {
2219   default: break;
2220   case X86::MOV8mr:
2221   case X86::MOV16mr:
2222   case X86::MOV32mr:
2223   case X86::MOV64mr:
2224   case X86::ST_FpP64m:
2225   case X86::MOVSSmr:
2226   case X86::MOVSDmr:
2227   case X86::MOVAPSmr:
2228   case X86::MOVAPDmr:
2229   case X86::MOVDQAmr:
2230   case X86::VMOVSSmr:
2231   case X86::VMOVSDmr:
2232   case X86::VMOVAPSmr:
2233   case X86::VMOVAPDmr:
2234   case X86::VMOVDQAmr:
2235   case X86::VMOVUPSYmr:
2236   case X86::VMOVAPSYmr:
2237   case X86::VMOVUPDYmr:
2238   case X86::VMOVAPDYmr:
2239   case X86::VMOVDQUYmr:
2240   case X86::VMOVDQAYmr:
2241   case X86::VMOVUPSZmr:
2242   case X86::VMOVAPSZmr:
2243   case X86::MMX_MOVD64mr:
2244   case X86::MMX_MOVQ64mr:
2245   case X86::MMX_MOVNTQmr:
2246     return true;
2247   }
2248   return false;
2249 }
2250 
isLoadFromStackSlot(const MachineInstr & MI,int & FrameIndex) const2251 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
2252                                            int &FrameIndex) const {
2253   if (isFrameLoadOpcode(MI.getOpcode()))
2254     if (MI.getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
2255       return MI.getOperand(0).getReg();
2256   return 0;
2257 }
2258 
isLoadFromStackSlotPostFE(const MachineInstr & MI,int & FrameIndex) const2259 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr &MI,
2260                                                  int &FrameIndex) const {
2261   if (isFrameLoadOpcode(MI.getOpcode())) {
2262     unsigned Reg;
2263     if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
2264       return Reg;
2265     // Check for post-frame index elimination operations
2266     const MachineMemOperand *Dummy;
2267     return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
2268   }
2269   return 0;
2270 }
2271 
isStoreToStackSlot(const MachineInstr & MI,int & FrameIndex) const2272 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI,
2273                                           int &FrameIndex) const {
2274   if (isFrameStoreOpcode(MI.getOpcode()))
2275     if (MI.getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
2276         isFrameOperand(MI, 0, FrameIndex))
2277       return MI.getOperand(X86::AddrNumOperands).getReg();
2278   return 0;
2279 }
2280 
isStoreToStackSlotPostFE(const MachineInstr & MI,int & FrameIndex) const2281 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr &MI,
2282                                                 int &FrameIndex) const {
2283   if (isFrameStoreOpcode(MI.getOpcode())) {
2284     unsigned Reg;
2285     if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
2286       return Reg;
2287     // Check for post-frame index elimination operations
2288     const MachineMemOperand *Dummy;
2289     return hasStoreToStackSlot(MI, Dummy, FrameIndex);
2290   }
2291   return 0;
2292 }
2293 
2294 /// Return true if register is PIC base; i.e.g defined by X86::MOVPC32r.
regIsPICBase(unsigned BaseReg,const MachineRegisterInfo & MRI)2295 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
2296   // Don't waste compile time scanning use-def chains of physregs.
2297   if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
2298     return false;
2299   bool isPICBase = false;
2300   for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg),
2301          E = MRI.def_instr_end(); I != E; ++I) {
2302     MachineInstr *DefMI = &*I;
2303     if (DefMI->getOpcode() != X86::MOVPC32r)
2304       return false;
2305     assert(!isPICBase && "More than one PIC base?");
2306     isPICBase = true;
2307   }
2308   return isPICBase;
2309 }
2310 
isReallyTriviallyReMaterializable(const MachineInstr & MI,AliasAnalysis * AA) const2311 bool X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
2312                                                      AliasAnalysis *AA) const {
2313   switch (MI.getOpcode()) {
2314   default: break;
2315   case X86::MOV8rm:
2316   case X86::MOV16rm:
2317   case X86::MOV32rm:
2318   case X86::MOV64rm:
2319   case X86::LD_Fp64m:
2320   case X86::MOVSSrm:
2321   case X86::MOVSDrm:
2322   case X86::MOVAPSrm:
2323   case X86::MOVUPSrm:
2324   case X86::MOVAPDrm:
2325   case X86::MOVDQArm:
2326   case X86::MOVDQUrm:
2327   case X86::VMOVSSrm:
2328   case X86::VMOVSDrm:
2329   case X86::VMOVAPSrm:
2330   case X86::VMOVUPSrm:
2331   case X86::VMOVAPDrm:
2332   case X86::VMOVDQArm:
2333   case X86::VMOVDQUrm:
2334   case X86::VMOVAPSYrm:
2335   case X86::VMOVUPSYrm:
2336   case X86::VMOVAPDYrm:
2337   case X86::VMOVDQAYrm:
2338   case X86::VMOVDQUYrm:
2339   case X86::MMX_MOVD64rm:
2340   case X86::MMX_MOVQ64rm:
2341   case X86::FsVMOVAPSrm:
2342   case X86::FsVMOVAPDrm:
2343   case X86::FsMOVAPSrm:
2344   case X86::FsMOVAPDrm:
2345   // AVX-512
2346   case X86::VMOVAPDZ128rm:
2347   case X86::VMOVAPDZ256rm:
2348   case X86::VMOVAPDZrm:
2349   case X86::VMOVAPSZ128rm:
2350   case X86::VMOVAPSZ256rm:
2351   case X86::VMOVAPSZrm:
2352   case X86::VMOVDQA32Z128rm:
2353   case X86::VMOVDQA32Z256rm:
2354   case X86::VMOVDQA32Zrm:
2355   case X86::VMOVDQA64Z128rm:
2356   case X86::VMOVDQA64Z256rm:
2357   case X86::VMOVDQA64Zrm:
2358   case X86::VMOVDQU16Z128rm:
2359   case X86::VMOVDQU16Z256rm:
2360   case X86::VMOVDQU16Zrm:
2361   case X86::VMOVDQU32Z128rm:
2362   case X86::VMOVDQU32Z256rm:
2363   case X86::VMOVDQU32Zrm:
2364   case X86::VMOVDQU64Z128rm:
2365   case X86::VMOVDQU64Z256rm:
2366   case X86::VMOVDQU64Zrm:
2367   case X86::VMOVDQU8Z128rm:
2368   case X86::VMOVDQU8Z256rm:
2369   case X86::VMOVDQU8Zrm:
2370   case X86::VMOVUPSZ128rm:
2371   case X86::VMOVUPSZ256rm:
2372   case X86::VMOVUPSZrm: {
2373     // Loads from constant pools are trivially rematerializable.
2374     if (MI.getOperand(1 + X86::AddrBaseReg).isReg() &&
2375         MI.getOperand(1 + X86::AddrScaleAmt).isImm() &&
2376         MI.getOperand(1 + X86::AddrIndexReg).isReg() &&
2377         MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
2378         MI.isInvariantLoad(AA)) {
2379       unsigned BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg();
2380       if (BaseReg == 0 || BaseReg == X86::RIP)
2381         return true;
2382       // Allow re-materialization of PIC load.
2383       if (!ReMatPICStubLoad && MI.getOperand(1 + X86::AddrDisp).isGlobal())
2384         return false;
2385       const MachineFunction &MF = *MI.getParent()->getParent();
2386       const MachineRegisterInfo &MRI = MF.getRegInfo();
2387       return regIsPICBase(BaseReg, MRI);
2388     }
2389     return false;
2390   }
2391 
2392   case X86::LEA32r:
2393   case X86::LEA64r: {
2394     if (MI.getOperand(1 + X86::AddrScaleAmt).isImm() &&
2395         MI.getOperand(1 + X86::AddrIndexReg).isReg() &&
2396         MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
2397         !MI.getOperand(1 + X86::AddrDisp).isReg()) {
2398       // lea fi#, lea GV, etc. are all rematerializable.
2399       if (!MI.getOperand(1 + X86::AddrBaseReg).isReg())
2400         return true;
2401       unsigned BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg();
2402       if (BaseReg == 0)
2403         return true;
2404       // Allow re-materialization of lea PICBase + x.
2405       const MachineFunction &MF = *MI.getParent()->getParent();
2406       const MachineRegisterInfo &MRI = MF.getRegInfo();
2407       return regIsPICBase(BaseReg, MRI);
2408     }
2409     return false;
2410   }
2411   }
2412 
2413   // All other instructions marked M_REMATERIALIZABLE are always trivially
2414   // rematerializable.
2415   return true;
2416 }
2417 
isSafeToClobberEFLAGS(MachineBasicBlock & MBB,MachineBasicBlock::iterator I) const2418 bool X86InstrInfo::isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
2419                                          MachineBasicBlock::iterator I) const {
2420   MachineBasicBlock::iterator E = MBB.end();
2421 
2422   // For compile time consideration, if we are not able to determine the
2423   // safety after visiting 4 instructions in each direction, we will assume
2424   // it's not safe.
2425   MachineBasicBlock::iterator Iter = I;
2426   for (unsigned i = 0; Iter != E && i < 4; ++i) {
2427     bool SeenDef = false;
2428     for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
2429       MachineOperand &MO = Iter->getOperand(j);
2430       if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
2431         SeenDef = true;
2432       if (!MO.isReg())
2433         continue;
2434       if (MO.getReg() == X86::EFLAGS) {
2435         if (MO.isUse())
2436           return false;
2437         SeenDef = true;
2438       }
2439     }
2440 
2441     if (SeenDef)
2442       // This instruction defines EFLAGS, no need to look any further.
2443       return true;
2444     ++Iter;
2445     // Skip over DBG_VALUE.
2446     while (Iter != E && Iter->isDebugValue())
2447       ++Iter;
2448   }
2449 
2450   // It is safe to clobber EFLAGS at the end of a block of no successor has it
2451   // live in.
2452   if (Iter == E) {
2453     for (MachineBasicBlock *S : MBB.successors())
2454       if (S->isLiveIn(X86::EFLAGS))
2455         return false;
2456     return true;
2457   }
2458 
2459   MachineBasicBlock::iterator B = MBB.begin();
2460   Iter = I;
2461   for (unsigned i = 0; i < 4; ++i) {
2462     // If we make it to the beginning of the block, it's safe to clobber
2463     // EFLAGS iff EFLAGS is not live-in.
2464     if (Iter == B)
2465       return !MBB.isLiveIn(X86::EFLAGS);
2466 
2467     --Iter;
2468     // Skip over DBG_VALUE.
2469     while (Iter != B && Iter->isDebugValue())
2470       --Iter;
2471 
2472     bool SawKill = false;
2473     for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
2474       MachineOperand &MO = Iter->getOperand(j);
2475       // A register mask may clobber EFLAGS, but we should still look for a
2476       // live EFLAGS def.
2477       if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
2478         SawKill = true;
2479       if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
2480         if (MO.isDef()) return MO.isDead();
2481         if (MO.isKill()) SawKill = true;
2482       }
2483     }
2484 
2485     if (SawKill)
2486       // This instruction kills EFLAGS and doesn't redefine it, so
2487       // there's no need to look further.
2488       return true;
2489   }
2490 
2491   // Conservative answer.
2492   return false;
2493 }
2494 
reMaterialize(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,unsigned DestReg,unsigned SubIdx,const MachineInstr & Orig,const TargetRegisterInfo & TRI) const2495 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
2496                                  MachineBasicBlock::iterator I,
2497                                  unsigned DestReg, unsigned SubIdx,
2498                                  const MachineInstr &Orig,
2499                                  const TargetRegisterInfo &TRI) const {
2500   bool ClobbersEFLAGS = false;
2501   for (const MachineOperand &MO : Orig.operands()) {
2502     if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
2503       ClobbersEFLAGS = true;
2504       break;
2505     }
2506   }
2507 
2508   if (ClobbersEFLAGS && !isSafeToClobberEFLAGS(MBB, I)) {
2509     // The instruction clobbers EFLAGS. Re-materialize as MOV32ri to avoid side
2510     // effects.
2511     int Value;
2512     switch (Orig.getOpcode()) {
2513     case X86::MOV32r0:  Value = 0; break;
2514     case X86::MOV32r1:  Value = 1; break;
2515     case X86::MOV32r_1: Value = -1; break;
2516     default:
2517       llvm_unreachable("Unexpected instruction!");
2518     }
2519 
2520     const DebugLoc &DL = Orig.getDebugLoc();
2521     BuildMI(MBB, I, DL, get(X86::MOV32ri))
2522         .addOperand(Orig.getOperand(0))
2523         .addImm(Value);
2524   } else {
2525     MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig);
2526     MBB.insert(I, MI);
2527   }
2528 
2529   MachineInstr &NewMI = *std::prev(I);
2530   NewMI.substituteRegister(Orig.getOperand(0).getReg(), DestReg, SubIdx, TRI);
2531 }
2532 
2533 /// True if MI has a condition code def, e.g. EFLAGS, that is not marked dead.
hasLiveCondCodeDef(MachineInstr & MI) const2534 bool X86InstrInfo::hasLiveCondCodeDef(MachineInstr &MI) const {
2535   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
2536     MachineOperand &MO = MI.getOperand(i);
2537     if (MO.isReg() && MO.isDef() &&
2538         MO.getReg() == X86::EFLAGS && !MO.isDead()) {
2539       return true;
2540     }
2541   }
2542   return false;
2543 }
2544 
2545 /// Check whether the shift count for a machine operand is non-zero.
getTruncatedShiftCount(MachineInstr & MI,unsigned ShiftAmtOperandIdx)2546 inline static unsigned getTruncatedShiftCount(MachineInstr &MI,
2547                                               unsigned ShiftAmtOperandIdx) {
2548   // The shift count is six bits with the REX.W prefix and five bits without.
2549   unsigned ShiftCountMask = (MI.getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
2550   unsigned Imm = MI.getOperand(ShiftAmtOperandIdx).getImm();
2551   return Imm & ShiftCountMask;
2552 }
2553 
2554 /// Check whether the given shift count is appropriate
2555 /// can be represented by a LEA instruction.
isTruncatedShiftCountForLEA(unsigned ShAmt)2556 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
2557   // Left shift instructions can be transformed into load-effective-address
2558   // instructions if we can encode them appropriately.
2559   // A LEA instruction utilizes a SIB byte to encode its scale factor.
2560   // The SIB.scale field is two bits wide which means that we can encode any
2561   // shift amount less than 4.
2562   return ShAmt < 4 && ShAmt > 0;
2563 }
2564 
classifyLEAReg(MachineInstr & MI,const MachineOperand & Src,unsigned Opc,bool AllowSP,unsigned & NewSrc,bool & isKill,bool & isUndef,MachineOperand & ImplicitOp) const2565 bool X86InstrInfo::classifyLEAReg(MachineInstr &MI, const MachineOperand &Src,
2566                                   unsigned Opc, bool AllowSP, unsigned &NewSrc,
2567                                   bool &isKill, bool &isUndef,
2568                                   MachineOperand &ImplicitOp) const {
2569   MachineFunction &MF = *MI.getParent()->getParent();
2570   const TargetRegisterClass *RC;
2571   if (AllowSP) {
2572     RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
2573   } else {
2574     RC = Opc != X86::LEA32r ?
2575       &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
2576   }
2577   unsigned SrcReg = Src.getReg();
2578 
2579   // For both LEA64 and LEA32 the register already has essentially the right
2580   // type (32-bit or 64-bit) we may just need to forbid SP.
2581   if (Opc != X86::LEA64_32r) {
2582     NewSrc = SrcReg;
2583     isKill = Src.isKill();
2584     isUndef = Src.isUndef();
2585 
2586     if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
2587         !MF.getRegInfo().constrainRegClass(NewSrc, RC))
2588       return false;
2589 
2590     return true;
2591   }
2592 
2593   // This is for an LEA64_32r and incoming registers are 32-bit. One way or
2594   // another we need to add 64-bit registers to the final MI.
2595   if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
2596     ImplicitOp = Src;
2597     ImplicitOp.setImplicit();
2598 
2599     NewSrc = getX86SubSuperRegister(Src.getReg(), 64);
2600     MachineBasicBlock::LivenessQueryResult LQR =
2601         MI.getParent()->computeRegisterLiveness(&getRegisterInfo(), NewSrc, MI);
2602 
2603     switch (LQR) {
2604     case MachineBasicBlock::LQR_Unknown:
2605       // We can't give sane liveness flags to the instruction, abandon LEA
2606       // formation.
2607       return false;
2608     case MachineBasicBlock::LQR_Live:
2609       isKill = MI.killsRegister(SrcReg);
2610       isUndef = false;
2611       break;
2612     default:
2613       // The physreg itself is dead, so we have to use it as an <undef>.
2614       isKill = false;
2615       isUndef = true;
2616       break;
2617     }
2618   } else {
2619     // Virtual register of the wrong class, we have to create a temporary 64-bit
2620     // vreg to feed into the LEA.
2621     NewSrc = MF.getRegInfo().createVirtualRegister(RC);
2622     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(TargetOpcode::COPY))
2623         .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
2624         .addOperand(Src);
2625 
2626     // Which is obviously going to be dead after we're done with it.
2627     isKill = true;
2628     isUndef = false;
2629   }
2630 
2631   // We've set all the parameters without issue.
2632   return true;
2633 }
2634 
2635 /// Helper for convertToThreeAddress when 16-bit LEA is disabled, use 32-bit
2636 /// LEA to form 3-address code by promoting to a 32-bit superregister and then
2637 /// truncating back down to a 16-bit subregister.
convertToThreeAddressWithLEA(unsigned MIOpc,MachineFunction::iterator & MFI,MachineInstr & MI,LiveVariables * LV) const2638 MachineInstr *X86InstrInfo::convertToThreeAddressWithLEA(
2639     unsigned MIOpc, MachineFunction::iterator &MFI, MachineInstr &MI,
2640     LiveVariables *LV) const {
2641   MachineBasicBlock::iterator MBBI = MI.getIterator();
2642   unsigned Dest = MI.getOperand(0).getReg();
2643   unsigned Src = MI.getOperand(1).getReg();
2644   bool isDead = MI.getOperand(0).isDead();
2645   bool isKill = MI.getOperand(1).isKill();
2646 
2647   MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
2648   unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
2649   unsigned Opc, leaInReg;
2650   if (Subtarget.is64Bit()) {
2651     Opc = X86::LEA64_32r;
2652     leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
2653   } else {
2654     Opc = X86::LEA32r;
2655     leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
2656   }
2657 
2658   // Build and insert into an implicit UNDEF value. This is OK because
2659   // well be shifting and then extracting the lower 16-bits.
2660   // This has the potential to cause partial register stall. e.g.
2661   //   movw    (%rbp,%rcx,2), %dx
2662   //   leal    -65(%rdx), %esi
2663   // But testing has shown this *does* help performance in 64-bit mode (at
2664   // least on modern x86 machines).
2665   BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
2666   MachineInstr *InsMI =
2667       BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY))
2668           .addReg(leaInReg, RegState::Define, X86::sub_16bit)
2669           .addReg(Src, getKillRegState(isKill));
2670 
2671   MachineInstrBuilder MIB =
2672       BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(Opc), leaOutReg);
2673   switch (MIOpc) {
2674   default: llvm_unreachable("Unreachable!");
2675   case X86::SHL16ri: {
2676     unsigned ShAmt = MI.getOperand(2).getImm();
2677     MIB.addReg(0).addImm(1ULL << ShAmt)
2678        .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
2679     break;
2680   }
2681   case X86::INC16r:
2682     addRegOffset(MIB, leaInReg, true, 1);
2683     break;
2684   case X86::DEC16r:
2685     addRegOffset(MIB, leaInReg, true, -1);
2686     break;
2687   case X86::ADD16ri:
2688   case X86::ADD16ri8:
2689   case X86::ADD16ri_DB:
2690   case X86::ADD16ri8_DB:
2691     addRegOffset(MIB, leaInReg, true, MI.getOperand(2).getImm());
2692     break;
2693   case X86::ADD16rr:
2694   case X86::ADD16rr_DB: {
2695     unsigned Src2 = MI.getOperand(2).getReg();
2696     bool isKill2 = MI.getOperand(2).isKill();
2697     unsigned leaInReg2 = 0;
2698     MachineInstr *InsMI2 = nullptr;
2699     if (Src == Src2) {
2700       // ADD16rr %reg1028<kill>, %reg1028
2701       // just a single insert_subreg.
2702       addRegReg(MIB, leaInReg, true, leaInReg, false);
2703     } else {
2704       if (Subtarget.is64Bit())
2705         leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
2706       else
2707         leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
2708       // Build and insert into an implicit UNDEF value. This is OK because
2709       // well be shifting and then extracting the lower 16-bits.
2710       BuildMI(*MFI, &*MIB, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg2);
2711       InsMI2 = BuildMI(*MFI, &*MIB, MI.getDebugLoc(), get(TargetOpcode::COPY))
2712                    .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
2713                    .addReg(Src2, getKillRegState(isKill2));
2714       addRegReg(MIB, leaInReg, true, leaInReg2, true);
2715     }
2716     if (LV && isKill2 && InsMI2)
2717       LV->replaceKillInstruction(Src2, MI, *InsMI2);
2718     break;
2719   }
2720   }
2721 
2722   MachineInstr *NewMI = MIB;
2723   MachineInstr *ExtMI =
2724       BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY))
2725           .addReg(Dest, RegState::Define | getDeadRegState(isDead))
2726           .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
2727 
2728   if (LV) {
2729     // Update live variables
2730     LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
2731     LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
2732     if (isKill)
2733       LV->replaceKillInstruction(Src, MI, *InsMI);
2734     if (isDead)
2735       LV->replaceKillInstruction(Dest, MI, *ExtMI);
2736   }
2737 
2738   return ExtMI;
2739 }
2740 
2741 /// This method must be implemented by targets that
2742 /// set the M_CONVERTIBLE_TO_3_ADDR flag.  When this flag is set, the target
2743 /// may be able to convert a two-address instruction into a true
2744 /// three-address instruction on demand.  This allows the X86 target (for
2745 /// example) to convert ADD and SHL instructions into LEA instructions if they
2746 /// would require register copies due to two-addressness.
2747 ///
2748 /// This method returns a null pointer if the transformation cannot be
2749 /// performed, otherwise it returns the new instruction.
2750 ///
2751 MachineInstr *
convertToThreeAddress(MachineFunction::iterator & MFI,MachineInstr & MI,LiveVariables * LV) const2752 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
2753                                     MachineInstr &MI, LiveVariables *LV) const {
2754   // The following opcodes also sets the condition code register(s). Only
2755   // convert them to equivalent lea if the condition code register def's
2756   // are dead!
2757   if (hasLiveCondCodeDef(MI))
2758     return nullptr;
2759 
2760   MachineFunction &MF = *MI.getParent()->getParent();
2761   // All instructions input are two-addr instructions.  Get the known operands.
2762   const MachineOperand &Dest = MI.getOperand(0);
2763   const MachineOperand &Src = MI.getOperand(1);
2764 
2765   MachineInstr *NewMI = nullptr;
2766   // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's.  When
2767   // we have better subtarget support, enable the 16-bit LEA generation here.
2768   // 16-bit LEA is also slow on Core2.
2769   bool DisableLEA16 = true;
2770   bool is64Bit = Subtarget.is64Bit();
2771 
2772   unsigned MIOpc = MI.getOpcode();
2773   switch (MIOpc) {
2774   default: return nullptr;
2775   case X86::SHL64ri: {
2776     assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
2777     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2778     if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2779 
2780     // LEA can't handle RSP.
2781     if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2782         !MF.getRegInfo().constrainRegClass(Src.getReg(),
2783                                            &X86::GR64_NOSPRegClass))
2784       return nullptr;
2785 
2786     NewMI = BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r))
2787                 .addOperand(Dest)
2788                 .addReg(0)
2789                 .addImm(1ULL << ShAmt)
2790                 .addOperand(Src)
2791                 .addImm(0)
2792                 .addReg(0);
2793     break;
2794   }
2795   case X86::SHL32ri: {
2796     assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
2797     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2798     if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2799 
2800     unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2801 
2802     // LEA can't handle ESP.
2803     bool isKill, isUndef;
2804     unsigned SrcReg;
2805     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2806     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2807                         SrcReg, isKill, isUndef, ImplicitOp))
2808       return nullptr;
2809 
2810     MachineInstrBuilder MIB =
2811         BuildMI(MF, MI.getDebugLoc(), get(Opc))
2812             .addOperand(Dest)
2813             .addReg(0)
2814             .addImm(1ULL << ShAmt)
2815             .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef))
2816             .addImm(0)
2817             .addReg(0);
2818     if (ImplicitOp.getReg() != 0)
2819       MIB.addOperand(ImplicitOp);
2820     NewMI = MIB;
2821 
2822     break;
2823   }
2824   case X86::SHL16ri: {
2825     assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!");
2826     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2827     if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2828 
2829     if (DisableLEA16)
2830       return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV)
2831                      : nullptr;
2832     NewMI = BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r))
2833                 .addOperand(Dest)
2834                 .addReg(0)
2835                 .addImm(1ULL << ShAmt)
2836                 .addOperand(Src)
2837                 .addImm(0)
2838                 .addReg(0);
2839     break;
2840   }
2841   case X86::INC64r:
2842   case X86::INC32r: {
2843     assert(MI.getNumOperands() >= 2 && "Unknown inc instruction!");
2844     unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
2845       : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2846     bool isKill, isUndef;
2847     unsigned SrcReg;
2848     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2849     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2850                         SrcReg, isKill, isUndef, ImplicitOp))
2851       return nullptr;
2852 
2853     MachineInstrBuilder MIB =
2854         BuildMI(MF, MI.getDebugLoc(), get(Opc))
2855             .addOperand(Dest)
2856             .addReg(SrcReg,
2857                     getKillRegState(isKill) | getUndefRegState(isUndef));
2858     if (ImplicitOp.getReg() != 0)
2859       MIB.addOperand(ImplicitOp);
2860 
2861     NewMI = addOffset(MIB, 1);
2862     break;
2863   }
2864   case X86::INC16r:
2865     if (DisableLEA16)
2866       return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV)
2867                      : nullptr;
2868     assert(MI.getNumOperands() >= 2 && "Unknown inc instruction!");
2869     NewMI = addOffset(BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r))
2870                           .addOperand(Dest)
2871                           .addOperand(Src),
2872                       1);
2873     break;
2874   case X86::DEC64r:
2875   case X86::DEC32r: {
2876     assert(MI.getNumOperands() >= 2 && "Unknown dec instruction!");
2877     unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2878       : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2879 
2880     bool isKill, isUndef;
2881     unsigned SrcReg;
2882     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2883     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2884                         SrcReg, isKill, isUndef, ImplicitOp))
2885       return nullptr;
2886 
2887     MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
2888                                   .addOperand(Dest)
2889                                   .addReg(SrcReg, getUndefRegState(isUndef) |
2890                                                       getKillRegState(isKill));
2891     if (ImplicitOp.getReg() != 0)
2892       MIB.addOperand(ImplicitOp);
2893 
2894     NewMI = addOffset(MIB, -1);
2895 
2896     break;
2897   }
2898   case X86::DEC16r:
2899     if (DisableLEA16)
2900       return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV)
2901                      : nullptr;
2902     assert(MI.getNumOperands() >= 2 && "Unknown dec instruction!");
2903     NewMI = addOffset(BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r))
2904                           .addOperand(Dest)
2905                           .addOperand(Src),
2906                       -1);
2907     break;
2908   case X86::ADD64rr:
2909   case X86::ADD64rr_DB:
2910   case X86::ADD32rr:
2911   case X86::ADD32rr_DB: {
2912     assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
2913     unsigned Opc;
2914     if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
2915       Opc = X86::LEA64r;
2916     else
2917       Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2918 
2919     bool isKill, isUndef;
2920     unsigned SrcReg;
2921     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2922     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2923                         SrcReg, isKill, isUndef, ImplicitOp))
2924       return nullptr;
2925 
2926     const MachineOperand &Src2 = MI.getOperand(2);
2927     bool isKill2, isUndef2;
2928     unsigned SrcReg2;
2929     MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
2930     if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
2931                         SrcReg2, isKill2, isUndef2, ImplicitOp2))
2932       return nullptr;
2933 
2934     MachineInstrBuilder MIB =
2935         BuildMI(MF, MI.getDebugLoc(), get(Opc)).addOperand(Dest);
2936     if (ImplicitOp.getReg() != 0)
2937       MIB.addOperand(ImplicitOp);
2938     if (ImplicitOp2.getReg() != 0)
2939       MIB.addOperand(ImplicitOp2);
2940 
2941     NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
2942 
2943     // Preserve undefness of the operands.
2944     NewMI->getOperand(1).setIsUndef(isUndef);
2945     NewMI->getOperand(3).setIsUndef(isUndef2);
2946 
2947     if (LV && Src2.isKill())
2948       LV->replaceKillInstruction(SrcReg2, MI, *NewMI);
2949     break;
2950   }
2951   case X86::ADD16rr:
2952   case X86::ADD16rr_DB: {
2953     if (DisableLEA16)
2954       return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV)
2955                      : nullptr;
2956     assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
2957     unsigned Src2 = MI.getOperand(2).getReg();
2958     bool isKill2 = MI.getOperand(2).isKill();
2959     NewMI = addRegReg(
2960         BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r)).addOperand(Dest),
2961         Src.getReg(), Src.isKill(), Src2, isKill2);
2962 
2963     // Preserve undefness of the operands.
2964     bool isUndef = MI.getOperand(1).isUndef();
2965     bool isUndef2 = MI.getOperand(2).isUndef();
2966     NewMI->getOperand(1).setIsUndef(isUndef);
2967     NewMI->getOperand(3).setIsUndef(isUndef2);
2968 
2969     if (LV && isKill2)
2970       LV->replaceKillInstruction(Src2, MI, *NewMI);
2971     break;
2972   }
2973   case X86::ADD64ri32:
2974   case X86::ADD64ri8:
2975   case X86::ADD64ri32_DB:
2976   case X86::ADD64ri8_DB:
2977     assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
2978     NewMI = addOffset(BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r))
2979                           .addOperand(Dest)
2980                           .addOperand(Src),
2981                       MI.getOperand(2).getImm());
2982     break;
2983   case X86::ADD32ri:
2984   case X86::ADD32ri8:
2985   case X86::ADD32ri_DB:
2986   case X86::ADD32ri8_DB: {
2987     assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
2988     unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2989 
2990     bool isKill, isUndef;
2991     unsigned SrcReg;
2992     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2993     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2994                         SrcReg, isKill, isUndef, ImplicitOp))
2995       return nullptr;
2996 
2997     MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc))
2998                                   .addOperand(Dest)
2999                                   .addReg(SrcReg, getUndefRegState(isUndef) |
3000                                                       getKillRegState(isKill));
3001     if (ImplicitOp.getReg() != 0)
3002       MIB.addOperand(ImplicitOp);
3003 
3004     NewMI = addOffset(MIB, MI.getOperand(2).getImm());
3005     break;
3006   }
3007   case X86::ADD16ri:
3008   case X86::ADD16ri8:
3009   case X86::ADD16ri_DB:
3010   case X86::ADD16ri8_DB:
3011     if (DisableLEA16)
3012       return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV)
3013                      : nullptr;
3014     assert(MI.getNumOperands() >= 3 && "Unknown add instruction!");
3015     NewMI = addOffset(BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r))
3016                           .addOperand(Dest)
3017                           .addOperand(Src),
3018                       MI.getOperand(2).getImm());
3019     break;
3020   }
3021 
3022   if (!NewMI) return nullptr;
3023 
3024   if (LV) {  // Update live variables
3025     if (Src.isKill())
3026       LV->replaceKillInstruction(Src.getReg(), MI, *NewMI);
3027     if (Dest.isDead())
3028       LV->replaceKillInstruction(Dest.getReg(), MI, *NewMI);
3029   }
3030 
3031   MFI->insert(MI.getIterator(), NewMI); // Insert the new inst
3032   return NewMI;
3033 }
3034 
3035 /// Returns true if the given instruction opcode is FMA3.
3036 /// Otherwise, returns false.
3037 /// The second parameter is optional and is used as the second return from
3038 /// the function. It is set to true if the given instruction has FMA3 opcode
3039 /// that is used for lowering of scalar FMA intrinsics, and it is set to false
3040 /// otherwise.
isFMA3(unsigned Opcode,bool * IsIntrinsic=nullptr)3041 static bool isFMA3(unsigned Opcode, bool *IsIntrinsic = nullptr) {
3042   if (IsIntrinsic)
3043     *IsIntrinsic = false;
3044 
3045   switch (Opcode) {
3046     case X86::VFMADDSDr132r:      case X86::VFMADDSDr132m:
3047     case X86::VFMADDSSr132r:      case X86::VFMADDSSr132m:
3048     case X86::VFMSUBSDr132r:      case X86::VFMSUBSDr132m:
3049     case X86::VFMSUBSSr132r:      case X86::VFMSUBSSr132m:
3050     case X86::VFNMADDSDr132r:     case X86::VFNMADDSDr132m:
3051     case X86::VFNMADDSSr132r:     case X86::VFNMADDSSr132m:
3052     case X86::VFNMSUBSDr132r:     case X86::VFNMSUBSDr132m:
3053     case X86::VFNMSUBSSr132r:     case X86::VFNMSUBSSr132m:
3054 
3055     case X86::VFMADDSDr213r:      case X86::VFMADDSDr213m:
3056     case X86::VFMADDSSr213r:      case X86::VFMADDSSr213m:
3057     case X86::VFMSUBSDr213r:      case X86::VFMSUBSDr213m:
3058     case X86::VFMSUBSSr213r:      case X86::VFMSUBSSr213m:
3059     case X86::VFNMADDSDr213r:     case X86::VFNMADDSDr213m:
3060     case X86::VFNMADDSSr213r:     case X86::VFNMADDSSr213m:
3061     case X86::VFNMSUBSDr213r:     case X86::VFNMSUBSDr213m:
3062     case X86::VFNMSUBSSr213r:     case X86::VFNMSUBSSr213m:
3063 
3064     case X86::VFMADDSDr231r:      case X86::VFMADDSDr231m:
3065     case X86::VFMADDSSr231r:      case X86::VFMADDSSr231m:
3066     case X86::VFMSUBSDr231r:      case X86::VFMSUBSDr231m:
3067     case X86::VFMSUBSSr231r:      case X86::VFMSUBSSr231m:
3068     case X86::VFNMADDSDr231r:     case X86::VFNMADDSDr231m:
3069     case X86::VFNMADDSSr231r:     case X86::VFNMADDSSr231m:
3070     case X86::VFNMSUBSDr231r:     case X86::VFNMSUBSDr231m:
3071     case X86::VFNMSUBSSr231r:     case X86::VFNMSUBSSr231m:
3072 
3073     case X86::VFMADDSUBPDr132r:   case X86::VFMADDSUBPDr132m:
3074     case X86::VFMADDSUBPSr132r:   case X86::VFMADDSUBPSr132m:
3075     case X86::VFMSUBADDPDr132r:   case X86::VFMSUBADDPDr132m:
3076     case X86::VFMSUBADDPSr132r:   case X86::VFMSUBADDPSr132m:
3077     case X86::VFMADDSUBPDr132rY:  case X86::VFMADDSUBPDr132mY:
3078     case X86::VFMADDSUBPSr132rY:  case X86::VFMADDSUBPSr132mY:
3079     case X86::VFMSUBADDPDr132rY:  case X86::VFMSUBADDPDr132mY:
3080     case X86::VFMSUBADDPSr132rY:  case X86::VFMSUBADDPSr132mY:
3081 
3082     case X86::VFMADDPDr132r:      case X86::VFMADDPDr132m:
3083     case X86::VFMADDPSr132r:      case X86::VFMADDPSr132m:
3084     case X86::VFMSUBPDr132r:      case X86::VFMSUBPDr132m:
3085     case X86::VFMSUBPSr132r:      case X86::VFMSUBPSr132m:
3086     case X86::VFNMADDPDr132r:     case X86::VFNMADDPDr132m:
3087     case X86::VFNMADDPSr132r:     case X86::VFNMADDPSr132m:
3088     case X86::VFNMSUBPDr132r:     case X86::VFNMSUBPDr132m:
3089     case X86::VFNMSUBPSr132r:     case X86::VFNMSUBPSr132m:
3090     case X86::VFMADDPDr132rY:     case X86::VFMADDPDr132mY:
3091     case X86::VFMADDPSr132rY:     case X86::VFMADDPSr132mY:
3092     case X86::VFMSUBPDr132rY:     case X86::VFMSUBPDr132mY:
3093     case X86::VFMSUBPSr132rY:     case X86::VFMSUBPSr132mY:
3094     case X86::VFNMADDPDr132rY:    case X86::VFNMADDPDr132mY:
3095     case X86::VFNMADDPSr132rY:    case X86::VFNMADDPSr132mY:
3096     case X86::VFNMSUBPDr132rY:    case X86::VFNMSUBPDr132mY:
3097     case X86::VFNMSUBPSr132rY:    case X86::VFNMSUBPSr132mY:
3098 
3099     case X86::VFMADDSUBPDr213r:   case X86::VFMADDSUBPDr213m:
3100     case X86::VFMADDSUBPSr213r:   case X86::VFMADDSUBPSr213m:
3101     case X86::VFMSUBADDPDr213r:   case X86::VFMSUBADDPDr213m:
3102     case X86::VFMSUBADDPSr213r:   case X86::VFMSUBADDPSr213m:
3103     case X86::VFMADDSUBPDr213rY:  case X86::VFMADDSUBPDr213mY:
3104     case X86::VFMADDSUBPSr213rY:  case X86::VFMADDSUBPSr213mY:
3105     case X86::VFMSUBADDPDr213rY:  case X86::VFMSUBADDPDr213mY:
3106     case X86::VFMSUBADDPSr213rY:  case X86::VFMSUBADDPSr213mY:
3107 
3108     case X86::VFMADDPDr213r:      case X86::VFMADDPDr213m:
3109     case X86::VFMADDPSr213r:      case X86::VFMADDPSr213m:
3110     case X86::VFMSUBPDr213r:      case X86::VFMSUBPDr213m:
3111     case X86::VFMSUBPSr213r:      case X86::VFMSUBPSr213m:
3112     case X86::VFNMADDPDr213r:     case X86::VFNMADDPDr213m:
3113     case X86::VFNMADDPSr213r:     case X86::VFNMADDPSr213m:
3114     case X86::VFNMSUBPDr213r:     case X86::VFNMSUBPDr213m:
3115     case X86::VFNMSUBPSr213r:     case X86::VFNMSUBPSr213m:
3116     case X86::VFMADDPDr213rY:     case X86::VFMADDPDr213mY:
3117     case X86::VFMADDPSr213rY:     case X86::VFMADDPSr213mY:
3118     case X86::VFMSUBPDr213rY:     case X86::VFMSUBPDr213mY:
3119     case X86::VFMSUBPSr213rY:     case X86::VFMSUBPSr213mY:
3120     case X86::VFNMADDPDr213rY:    case X86::VFNMADDPDr213mY:
3121     case X86::VFNMADDPSr213rY:    case X86::VFNMADDPSr213mY:
3122     case X86::VFNMSUBPDr213rY:    case X86::VFNMSUBPDr213mY:
3123     case X86::VFNMSUBPSr213rY:    case X86::VFNMSUBPSr213mY:
3124 
3125     case X86::VFMADDSUBPDr231r:   case X86::VFMADDSUBPDr231m:
3126     case X86::VFMADDSUBPSr231r:   case X86::VFMADDSUBPSr231m:
3127     case X86::VFMSUBADDPDr231r:   case X86::VFMSUBADDPDr231m:
3128     case X86::VFMSUBADDPSr231r:   case X86::VFMSUBADDPSr231m:
3129     case X86::VFMADDSUBPDr231rY:  case X86::VFMADDSUBPDr231mY:
3130     case X86::VFMADDSUBPSr231rY:  case X86::VFMADDSUBPSr231mY:
3131     case X86::VFMSUBADDPDr231rY:  case X86::VFMSUBADDPDr231mY:
3132     case X86::VFMSUBADDPSr231rY:  case X86::VFMSUBADDPSr231mY:
3133 
3134     case X86::VFMADDPDr231r:      case X86::VFMADDPDr231m:
3135     case X86::VFMADDPSr231r:      case X86::VFMADDPSr231m:
3136     case X86::VFMSUBPDr231r:      case X86::VFMSUBPDr231m:
3137     case X86::VFMSUBPSr231r:      case X86::VFMSUBPSr231m:
3138     case X86::VFNMADDPDr231r:     case X86::VFNMADDPDr231m:
3139     case X86::VFNMADDPSr231r:     case X86::VFNMADDPSr231m:
3140     case X86::VFNMSUBPDr231r:     case X86::VFNMSUBPDr231m:
3141     case X86::VFNMSUBPSr231r:     case X86::VFNMSUBPSr231m:
3142     case X86::VFMADDPDr231rY:     case X86::VFMADDPDr231mY:
3143     case X86::VFMADDPSr231rY:     case X86::VFMADDPSr231mY:
3144     case X86::VFMSUBPDr231rY:     case X86::VFMSUBPDr231mY:
3145     case X86::VFMSUBPSr231rY:     case X86::VFMSUBPSr231mY:
3146     case X86::VFNMADDPDr231rY:    case X86::VFNMADDPDr231mY:
3147     case X86::VFNMADDPSr231rY:    case X86::VFNMADDPSr231mY:
3148     case X86::VFNMSUBPDr231rY:    case X86::VFNMSUBPDr231mY:
3149     case X86::VFNMSUBPSr231rY:    case X86::VFNMSUBPSr231mY:
3150       return true;
3151 
3152     case X86::VFMADDSDr132r_Int:  case X86::VFMADDSDr132m_Int:
3153     case X86::VFMADDSSr132r_Int:  case X86::VFMADDSSr132m_Int:
3154     case X86::VFMSUBSDr132r_Int:  case X86::VFMSUBSDr132m_Int:
3155     case X86::VFMSUBSSr132r_Int:  case X86::VFMSUBSSr132m_Int:
3156     case X86::VFNMADDSDr132r_Int: case X86::VFNMADDSDr132m_Int:
3157     case X86::VFNMADDSSr132r_Int: case X86::VFNMADDSSr132m_Int:
3158     case X86::VFNMSUBSDr132r_Int: case X86::VFNMSUBSDr132m_Int:
3159     case X86::VFNMSUBSSr132r_Int: case X86::VFNMSUBSSr132m_Int:
3160 
3161     case X86::VFMADDSDr213r_Int:  case X86::VFMADDSDr213m_Int:
3162     case X86::VFMADDSSr213r_Int:  case X86::VFMADDSSr213m_Int:
3163     case X86::VFMSUBSDr213r_Int:  case X86::VFMSUBSDr213m_Int:
3164     case X86::VFMSUBSSr213r_Int:  case X86::VFMSUBSSr213m_Int:
3165     case X86::VFNMADDSDr213r_Int: case X86::VFNMADDSDr213m_Int:
3166     case X86::VFNMADDSSr213r_Int: case X86::VFNMADDSSr213m_Int:
3167     case X86::VFNMSUBSDr213r_Int: case X86::VFNMSUBSDr213m_Int:
3168     case X86::VFNMSUBSSr213r_Int: case X86::VFNMSUBSSr213m_Int:
3169 
3170     case X86::VFMADDSDr231r_Int:  case X86::VFMADDSDr231m_Int:
3171     case X86::VFMADDSSr231r_Int:  case X86::VFMADDSSr231m_Int:
3172     case X86::VFMSUBSDr231r_Int:  case X86::VFMSUBSDr231m_Int:
3173     case X86::VFMSUBSSr231r_Int:  case X86::VFMSUBSSr231m_Int:
3174     case X86::VFNMADDSDr231r_Int: case X86::VFNMADDSDr231m_Int:
3175     case X86::VFNMADDSSr231r_Int: case X86::VFNMADDSSr231m_Int:
3176     case X86::VFNMSUBSDr231r_Int: case X86::VFNMSUBSDr231m_Int:
3177     case X86::VFNMSUBSSr231r_Int: case X86::VFNMSUBSSr231m_Int:
3178       if (IsIntrinsic)
3179         *IsIntrinsic = true;
3180       return true;
3181     default:
3182       return false;
3183   }
3184   llvm_unreachable("Opcode not handled by the switch");
3185 }
3186 
commuteInstructionImpl(MachineInstr & MI,bool NewMI,unsigned OpIdx1,unsigned OpIdx2) const3187 MachineInstr *X86InstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
3188                                                    unsigned OpIdx1,
3189                                                    unsigned OpIdx2) const {
3190   auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & {
3191     if (NewMI)
3192       return *MI.getParent()->getParent()->CloneMachineInstr(&MI);
3193     return MI;
3194   };
3195 
3196   switch (MI.getOpcode()) {
3197   case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
3198   case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
3199   case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
3200   case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
3201   case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
3202   case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
3203     unsigned Opc;
3204     unsigned Size;
3205     switch (MI.getOpcode()) {
3206     default: llvm_unreachable("Unreachable!");
3207     case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
3208     case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
3209     case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
3210     case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
3211     case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
3212     case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
3213     }
3214     unsigned Amt = MI.getOperand(3).getImm();
3215     auto &WorkingMI = cloneIfNew(MI);
3216     WorkingMI.setDesc(get(Opc));
3217     WorkingMI.getOperand(3).setImm(Size - Amt);
3218     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
3219                                                    OpIdx1, OpIdx2);
3220   }
3221   case X86::BLENDPDrri:
3222   case X86::BLENDPSrri:
3223   case X86::PBLENDWrri:
3224   case X86::VBLENDPDrri:
3225   case X86::VBLENDPSrri:
3226   case X86::VBLENDPDYrri:
3227   case X86::VBLENDPSYrri:
3228   case X86::VPBLENDDrri:
3229   case X86::VPBLENDWrri:
3230   case X86::VPBLENDDYrri:
3231   case X86::VPBLENDWYrri:{
3232     unsigned Mask;
3233     switch (MI.getOpcode()) {
3234     default: llvm_unreachable("Unreachable!");
3235     case X86::BLENDPDrri:    Mask = 0x03; break;
3236     case X86::BLENDPSrri:    Mask = 0x0F; break;
3237     case X86::PBLENDWrri:    Mask = 0xFF; break;
3238     case X86::VBLENDPDrri:   Mask = 0x03; break;
3239     case X86::VBLENDPSrri:   Mask = 0x0F; break;
3240     case X86::VBLENDPDYrri:  Mask = 0x0F; break;
3241     case X86::VBLENDPSYrri:  Mask = 0xFF; break;
3242     case X86::VPBLENDDrri:   Mask = 0x0F; break;
3243     case X86::VPBLENDWrri:   Mask = 0xFF; break;
3244     case X86::VPBLENDDYrri:  Mask = 0xFF; break;
3245     case X86::VPBLENDWYrri:  Mask = 0xFF; break;
3246     }
3247     // Only the least significant bits of Imm are used.
3248     unsigned Imm = MI.getOperand(3).getImm() & Mask;
3249     auto &WorkingMI = cloneIfNew(MI);
3250     WorkingMI.getOperand(3).setImm(Mask ^ Imm);
3251     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
3252                                                    OpIdx1, OpIdx2);
3253   }
3254   case X86::PCLMULQDQrr:
3255   case X86::VPCLMULQDQrr:{
3256     // SRC1 64bits = Imm[0] ? SRC1[127:64] : SRC1[63:0]
3257     // SRC2 64bits = Imm[4] ? SRC2[127:64] : SRC2[63:0]
3258     unsigned Imm = MI.getOperand(3).getImm();
3259     unsigned Src1Hi = Imm & 0x01;
3260     unsigned Src2Hi = Imm & 0x10;
3261     auto &WorkingMI = cloneIfNew(MI);
3262     WorkingMI.getOperand(3).setImm((Src1Hi << 4) | (Src2Hi >> 4));
3263     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
3264                                                    OpIdx1, OpIdx2);
3265   }
3266   case X86::CMPPDrri:
3267   case X86::CMPPSrri:
3268   case X86::VCMPPDrri:
3269   case X86::VCMPPSrri:
3270   case X86::VCMPPDYrri:
3271   case X86::VCMPPSYrri: {
3272     // Float comparison can be safely commuted for
3273     // Ordered/Unordered/Equal/NotEqual tests
3274     unsigned Imm = MI.getOperand(3).getImm() & 0x7;
3275     switch (Imm) {
3276     case 0x00: // EQUAL
3277     case 0x03: // UNORDERED
3278     case 0x04: // NOT EQUAL
3279     case 0x07: // ORDERED
3280       return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
3281     default:
3282       return nullptr;
3283     }
3284   }
3285   case X86::VPCOMBri: case X86::VPCOMUBri:
3286   case X86::VPCOMDri: case X86::VPCOMUDri:
3287   case X86::VPCOMQri: case X86::VPCOMUQri:
3288   case X86::VPCOMWri: case X86::VPCOMUWri: {
3289     // Flip comparison mode immediate (if necessary).
3290     unsigned Imm = MI.getOperand(3).getImm() & 0x7;
3291     switch (Imm) {
3292     case 0x00: Imm = 0x02; break; // LT -> GT
3293     case 0x01: Imm = 0x03; break; // LE -> GE
3294     case 0x02: Imm = 0x00; break; // GT -> LT
3295     case 0x03: Imm = 0x01; break; // GE -> LE
3296     case 0x04: // EQ
3297     case 0x05: // NE
3298     case 0x06: // FALSE
3299     case 0x07: // TRUE
3300     default:
3301       break;
3302     }
3303     auto &WorkingMI = cloneIfNew(MI);
3304     WorkingMI.getOperand(3).setImm(Imm);
3305     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
3306                                                    OpIdx1, OpIdx2);
3307   }
3308   case X86::VPERM2F128rr:
3309   case X86::VPERM2I128rr: {
3310     // Flip permute source immediate.
3311     // Imm & 0x02: lo = if set, select Op1.lo/hi else Op0.lo/hi.
3312     // Imm & 0x20: hi = if set, select Op1.lo/hi else Op0.lo/hi.
3313     unsigned Imm = MI.getOperand(3).getImm() & 0xFF;
3314     auto &WorkingMI = cloneIfNew(MI);
3315     WorkingMI.getOperand(3).setImm(Imm ^ 0x22);
3316     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
3317                                                    OpIdx1, OpIdx2);
3318   }
3319   case X86::CMOVB16rr:  case X86::CMOVB32rr:  case X86::CMOVB64rr:
3320   case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
3321   case X86::CMOVE16rr:  case X86::CMOVE32rr:  case X86::CMOVE64rr:
3322   case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
3323   case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
3324   case X86::CMOVA16rr:  case X86::CMOVA32rr:  case X86::CMOVA64rr:
3325   case X86::CMOVL16rr:  case X86::CMOVL32rr:  case X86::CMOVL64rr:
3326   case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
3327   case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
3328   case X86::CMOVG16rr:  case X86::CMOVG32rr:  case X86::CMOVG64rr:
3329   case X86::CMOVS16rr:  case X86::CMOVS32rr:  case X86::CMOVS64rr:
3330   case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
3331   case X86::CMOVP16rr:  case X86::CMOVP32rr:  case X86::CMOVP64rr:
3332   case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
3333   case X86::CMOVO16rr:  case X86::CMOVO32rr:  case X86::CMOVO64rr:
3334   case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
3335     unsigned Opc;
3336     switch (MI.getOpcode()) {
3337     default: llvm_unreachable("Unreachable!");
3338     case X86::CMOVB16rr:  Opc = X86::CMOVAE16rr; break;
3339     case X86::CMOVB32rr:  Opc = X86::CMOVAE32rr; break;
3340     case X86::CMOVB64rr:  Opc = X86::CMOVAE64rr; break;
3341     case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
3342     case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
3343     case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
3344     case X86::CMOVE16rr:  Opc = X86::CMOVNE16rr; break;
3345     case X86::CMOVE32rr:  Opc = X86::CMOVNE32rr; break;
3346     case X86::CMOVE64rr:  Opc = X86::CMOVNE64rr; break;
3347     case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
3348     case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
3349     case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
3350     case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
3351     case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
3352     case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
3353     case X86::CMOVA16rr:  Opc = X86::CMOVBE16rr; break;
3354     case X86::CMOVA32rr:  Opc = X86::CMOVBE32rr; break;
3355     case X86::CMOVA64rr:  Opc = X86::CMOVBE64rr; break;
3356     case X86::CMOVL16rr:  Opc = X86::CMOVGE16rr; break;
3357     case X86::CMOVL32rr:  Opc = X86::CMOVGE32rr; break;
3358     case X86::CMOVL64rr:  Opc = X86::CMOVGE64rr; break;
3359     case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
3360     case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
3361     case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
3362     case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
3363     case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
3364     case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
3365     case X86::CMOVG16rr:  Opc = X86::CMOVLE16rr; break;
3366     case X86::CMOVG32rr:  Opc = X86::CMOVLE32rr; break;
3367     case X86::CMOVG64rr:  Opc = X86::CMOVLE64rr; break;
3368     case X86::CMOVS16rr:  Opc = X86::CMOVNS16rr; break;
3369     case X86::CMOVS32rr:  Opc = X86::CMOVNS32rr; break;
3370     case X86::CMOVS64rr:  Opc = X86::CMOVNS64rr; break;
3371     case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
3372     case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
3373     case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
3374     case X86::CMOVP16rr:  Opc = X86::CMOVNP16rr; break;
3375     case X86::CMOVP32rr:  Opc = X86::CMOVNP32rr; break;
3376     case X86::CMOVP64rr:  Opc = X86::CMOVNP64rr; break;
3377     case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
3378     case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
3379     case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
3380     case X86::CMOVO16rr:  Opc = X86::CMOVNO16rr; break;
3381     case X86::CMOVO32rr:  Opc = X86::CMOVNO32rr; break;
3382     case X86::CMOVO64rr:  Opc = X86::CMOVNO64rr; break;
3383     case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
3384     case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
3385     case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
3386     }
3387     auto &WorkingMI = cloneIfNew(MI);
3388     WorkingMI.setDesc(get(Opc));
3389     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
3390                                                    OpIdx1, OpIdx2);
3391   }
3392   default:
3393     if (isFMA3(MI.getOpcode())) {
3394       unsigned Opc = getFMA3OpcodeToCommuteOperands(MI, OpIdx1, OpIdx2);
3395       if (Opc == 0)
3396         return nullptr;
3397       auto &WorkingMI = cloneIfNew(MI);
3398       WorkingMI.setDesc(get(Opc));
3399       return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
3400                                                      OpIdx1, OpIdx2);
3401     }
3402 
3403     return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
3404   }
3405 }
3406 
findFMA3CommutedOpIndices(MachineInstr & MI,unsigned & SrcOpIdx1,unsigned & SrcOpIdx2) const3407 bool X86InstrInfo::findFMA3CommutedOpIndices(MachineInstr &MI,
3408                                              unsigned &SrcOpIdx1,
3409                                              unsigned &SrcOpIdx2) const {
3410 
3411   unsigned RegOpsNum = isMem(MI, 3) ? 2 : 3;
3412 
3413   // Only the first RegOpsNum operands are commutable.
3414   // Also, the value 'CommuteAnyOperandIndex' is valid here as it means
3415   // that the operand is not specified/fixed.
3416   if (SrcOpIdx1 != CommuteAnyOperandIndex &&
3417       (SrcOpIdx1 < 1 || SrcOpIdx1 > RegOpsNum))
3418     return false;
3419   if (SrcOpIdx2 != CommuteAnyOperandIndex &&
3420       (SrcOpIdx2 < 1 || SrcOpIdx2 > RegOpsNum))
3421     return false;
3422 
3423   // Look for two different register operands assumed to be commutable
3424   // regardless of the FMA opcode. The FMA opcode is adjusted later.
3425   if (SrcOpIdx1 == CommuteAnyOperandIndex ||
3426       SrcOpIdx2 == CommuteAnyOperandIndex) {
3427     unsigned CommutableOpIdx1 = SrcOpIdx1;
3428     unsigned CommutableOpIdx2 = SrcOpIdx2;
3429 
3430     // At least one of operands to be commuted is not specified and
3431     // this method is free to choose appropriate commutable operands.
3432     if (SrcOpIdx1 == SrcOpIdx2)
3433       // Both of operands are not fixed. By default set one of commutable
3434       // operands to the last register operand of the instruction.
3435       CommutableOpIdx2 = RegOpsNum;
3436     else if (SrcOpIdx2 == CommuteAnyOperandIndex)
3437       // Only one of operands is not fixed.
3438       CommutableOpIdx2 = SrcOpIdx1;
3439 
3440     // CommutableOpIdx2 is well defined now. Let's choose another commutable
3441     // operand and assign its index to CommutableOpIdx1.
3442     unsigned Op2Reg = MI.getOperand(CommutableOpIdx2).getReg();
3443     for (CommutableOpIdx1 = RegOpsNum; CommutableOpIdx1 > 0; CommutableOpIdx1--) {
3444       // The commuted operands must have different registers.
3445       // Otherwise, the commute transformation does not change anything and
3446       // is useless then.
3447       if (Op2Reg != MI.getOperand(CommutableOpIdx1).getReg())
3448         break;
3449     }
3450 
3451     // No appropriate commutable operands were found.
3452     if (CommutableOpIdx1 == 0)
3453       return false;
3454 
3455     // Assign the found pair of commutable indices to SrcOpIdx1 and SrcOpidx2
3456     // to return those values.
3457     if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
3458                               CommutableOpIdx1, CommutableOpIdx2))
3459       return false;
3460   }
3461 
3462   // Check if we can adjust the opcode to preserve the semantics when
3463   // commute the register operands.
3464   return getFMA3OpcodeToCommuteOperands(MI, SrcOpIdx1, SrcOpIdx2) != 0;
3465 }
3466 
getFMA3OpcodeToCommuteOperands(MachineInstr & MI,unsigned SrcOpIdx1,unsigned SrcOpIdx2) const3467 unsigned X86InstrInfo::getFMA3OpcodeToCommuteOperands(
3468     MachineInstr &MI, unsigned SrcOpIdx1, unsigned SrcOpIdx2) const {
3469   unsigned Opc = MI.getOpcode();
3470 
3471   // Define the array that holds FMA opcodes in groups
3472   // of 3 opcodes(132, 213, 231) in each group.
3473   static const uint16_t RegularOpcodeGroups[][3] = {
3474     { X86::VFMADDSSr132r,   X86::VFMADDSSr213r,   X86::VFMADDSSr231r  },
3475     { X86::VFMADDSDr132r,   X86::VFMADDSDr213r,   X86::VFMADDSDr231r  },
3476     { X86::VFMADDPSr132r,   X86::VFMADDPSr213r,   X86::VFMADDPSr231r  },
3477     { X86::VFMADDPDr132r,   X86::VFMADDPDr213r,   X86::VFMADDPDr231r  },
3478     { X86::VFMADDPSr132rY,  X86::VFMADDPSr213rY,  X86::VFMADDPSr231rY },
3479     { X86::VFMADDPDr132rY,  X86::VFMADDPDr213rY,  X86::VFMADDPDr231rY },
3480     { X86::VFMADDSSr132m,   X86::VFMADDSSr213m,   X86::VFMADDSSr231m  },
3481     { X86::VFMADDSDr132m,   X86::VFMADDSDr213m,   X86::VFMADDSDr231m  },
3482     { X86::VFMADDPSr132m,   X86::VFMADDPSr213m,   X86::VFMADDPSr231m  },
3483     { X86::VFMADDPDr132m,   X86::VFMADDPDr213m,   X86::VFMADDPDr231m  },
3484     { X86::VFMADDPSr132mY,  X86::VFMADDPSr213mY,  X86::VFMADDPSr231mY },
3485     { X86::VFMADDPDr132mY,  X86::VFMADDPDr213mY,  X86::VFMADDPDr231mY },
3486 
3487     { X86::VFMSUBSSr132r,   X86::VFMSUBSSr213r,   X86::VFMSUBSSr231r  },
3488     { X86::VFMSUBSDr132r,   X86::VFMSUBSDr213r,   X86::VFMSUBSDr231r  },
3489     { X86::VFMSUBPSr132r,   X86::VFMSUBPSr213r,   X86::VFMSUBPSr231r  },
3490     { X86::VFMSUBPDr132r,   X86::VFMSUBPDr213r,   X86::VFMSUBPDr231r  },
3491     { X86::VFMSUBPSr132rY,  X86::VFMSUBPSr213rY,  X86::VFMSUBPSr231rY },
3492     { X86::VFMSUBPDr132rY,  X86::VFMSUBPDr213rY,  X86::VFMSUBPDr231rY },
3493     { X86::VFMSUBSSr132m,   X86::VFMSUBSSr213m,   X86::VFMSUBSSr231m  },
3494     { X86::VFMSUBSDr132m,   X86::VFMSUBSDr213m,   X86::VFMSUBSDr231m  },
3495     { X86::VFMSUBPSr132m,   X86::VFMSUBPSr213m,   X86::VFMSUBPSr231m  },
3496     { X86::VFMSUBPDr132m,   X86::VFMSUBPDr213m,   X86::VFMSUBPDr231m  },
3497     { X86::VFMSUBPSr132mY,  X86::VFMSUBPSr213mY,  X86::VFMSUBPSr231mY },
3498     { X86::VFMSUBPDr132mY,  X86::VFMSUBPDr213mY,  X86::VFMSUBPDr231mY },
3499 
3500     { X86::VFNMADDSSr132r,  X86::VFNMADDSSr213r,  X86::VFNMADDSSr231r  },
3501     { X86::VFNMADDSDr132r,  X86::VFNMADDSDr213r,  X86::VFNMADDSDr231r  },
3502     { X86::VFNMADDPSr132r,  X86::VFNMADDPSr213r,  X86::VFNMADDPSr231r  },
3503     { X86::VFNMADDPDr132r,  X86::VFNMADDPDr213r,  X86::VFNMADDPDr231r  },
3504     { X86::VFNMADDPSr132rY, X86::VFNMADDPSr213rY, X86::VFNMADDPSr231rY },
3505     { X86::VFNMADDPDr132rY, X86::VFNMADDPDr213rY, X86::VFNMADDPDr231rY },
3506     { X86::VFNMADDSSr132m,  X86::VFNMADDSSr213m,  X86::VFNMADDSSr231m  },
3507     { X86::VFNMADDSDr132m,  X86::VFNMADDSDr213m,  X86::VFNMADDSDr231m  },
3508     { X86::VFNMADDPSr132m,  X86::VFNMADDPSr213m,  X86::VFNMADDPSr231m  },
3509     { X86::VFNMADDPDr132m,  X86::VFNMADDPDr213m,  X86::VFNMADDPDr231m  },
3510     { X86::VFNMADDPSr132mY, X86::VFNMADDPSr213mY, X86::VFNMADDPSr231mY },
3511     { X86::VFNMADDPDr132mY, X86::VFNMADDPDr213mY, X86::VFNMADDPDr231mY },
3512 
3513     { X86::VFNMSUBSSr132r,  X86::VFNMSUBSSr213r,  X86::VFNMSUBSSr231r  },
3514     { X86::VFNMSUBSDr132r,  X86::VFNMSUBSDr213r,  X86::VFNMSUBSDr231r  },
3515     { X86::VFNMSUBPSr132r,  X86::VFNMSUBPSr213r,  X86::VFNMSUBPSr231r  },
3516     { X86::VFNMSUBPDr132r,  X86::VFNMSUBPDr213r,  X86::VFNMSUBPDr231r  },
3517     { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr231rY },
3518     { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr231rY },
3519     { X86::VFNMSUBSSr132m,  X86::VFNMSUBSSr213m,  X86::VFNMSUBSSr231m  },
3520     { X86::VFNMSUBSDr132m,  X86::VFNMSUBSDr213m,  X86::VFNMSUBSDr231m  },
3521     { X86::VFNMSUBPSr132m,  X86::VFNMSUBPSr213m,  X86::VFNMSUBPSr231m  },
3522     { X86::VFNMSUBPDr132m,  X86::VFNMSUBPDr213m,  X86::VFNMSUBPDr231m  },
3523     { X86::VFNMSUBPSr132mY, X86::VFNMSUBPSr213mY, X86::VFNMSUBPSr231mY },
3524     { X86::VFNMSUBPDr132mY, X86::VFNMSUBPDr213mY, X86::VFNMSUBPDr231mY },
3525 
3526     { X86::VFMADDSUBPSr132r,  X86::VFMADDSUBPSr213r,  X86::VFMADDSUBPSr231r  },
3527     { X86::VFMADDSUBPDr132r,  X86::VFMADDSUBPDr213r,  X86::VFMADDSUBPDr231r  },
3528     { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr231rY },
3529     { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr231rY },
3530     { X86::VFMADDSUBPSr132m,  X86::VFMADDSUBPSr213m,  X86::VFMADDSUBPSr231m  },
3531     { X86::VFMADDSUBPDr132m,  X86::VFMADDSUBPDr213m,  X86::VFMADDSUBPDr231m  },
3532     { X86::VFMADDSUBPSr132mY, X86::VFMADDSUBPSr213mY, X86::VFMADDSUBPSr231mY },
3533     { X86::VFMADDSUBPDr132mY, X86::VFMADDSUBPDr213mY, X86::VFMADDSUBPDr231mY },
3534 
3535     { X86::VFMSUBADDPSr132r,  X86::VFMSUBADDPSr213r,  X86::VFMSUBADDPSr231r  },
3536     { X86::VFMSUBADDPDr132r,  X86::VFMSUBADDPDr213r,  X86::VFMSUBADDPDr231r  },
3537     { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr231rY },
3538     { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr231rY },
3539     { X86::VFMSUBADDPSr132m,  X86::VFMSUBADDPSr213m,  X86::VFMSUBADDPSr231m  },
3540     { X86::VFMSUBADDPDr132m,  X86::VFMSUBADDPDr213m,  X86::VFMSUBADDPDr231m  },
3541     { X86::VFMSUBADDPSr132mY, X86::VFMSUBADDPSr213mY, X86::VFMSUBADDPSr231mY },
3542     { X86::VFMSUBADDPDr132mY, X86::VFMSUBADDPDr213mY, X86::VFMSUBADDPDr231mY }
3543   };
3544 
3545   // Define the array that holds FMA*_Int opcodes in groups
3546   // of 3 opcodes(132, 213, 231) in each group.
3547   static const uint16_t IntrinOpcodeGroups[][3] = {
3548     { X86::VFMADDSSr132r_Int,  X86::VFMADDSSr213r_Int,  X86::VFMADDSSr231r_Int },
3549     { X86::VFMADDSDr132r_Int,  X86::VFMADDSDr213r_Int,  X86::VFMADDSDr231r_Int },
3550     { X86::VFMADDSSr132m_Int,  X86::VFMADDSSr213m_Int,  X86::VFMADDSSr231m_Int },
3551     { X86::VFMADDSDr132m_Int,  X86::VFMADDSDr213m_Int,  X86::VFMADDSDr231m_Int },
3552 
3553     { X86::VFMSUBSSr132r_Int,  X86::VFMSUBSSr213r_Int,  X86::VFMSUBSSr231r_Int },
3554     { X86::VFMSUBSDr132r_Int,  X86::VFMSUBSDr213r_Int,  X86::VFMSUBSDr231r_Int },
3555     { X86::VFMSUBSSr132m_Int,  X86::VFMSUBSSr213m_Int,  X86::VFMSUBSSr231m_Int },
3556     { X86::VFMSUBSDr132m_Int,  X86::VFMSUBSDr213m_Int,  X86::VFMSUBSDr231m_Int },
3557 
3558     { X86::VFNMADDSSr132r_Int, X86::VFNMADDSSr213r_Int, X86::VFNMADDSSr231r_Int },
3559     { X86::VFNMADDSDr132r_Int, X86::VFNMADDSDr213r_Int, X86::VFNMADDSDr231r_Int },
3560     { X86::VFNMADDSSr132m_Int, X86::VFNMADDSSr213m_Int, X86::VFNMADDSSr231m_Int },
3561     { X86::VFNMADDSDr132m_Int, X86::VFNMADDSDr213m_Int, X86::VFNMADDSDr231m_Int },
3562 
3563     { X86::VFNMSUBSSr132r_Int, X86::VFNMSUBSSr213r_Int, X86::VFNMSUBSSr231r_Int },
3564     { X86::VFNMSUBSDr132r_Int, X86::VFNMSUBSDr213r_Int, X86::VFNMSUBSDr231r_Int },
3565     { X86::VFNMSUBSSr132m_Int, X86::VFNMSUBSSr213m_Int, X86::VFNMSUBSSr231m_Int },
3566     { X86::VFNMSUBSDr132m_Int, X86::VFNMSUBSDr213m_Int, X86::VFNMSUBSDr231m_Int },
3567   };
3568 
3569   const unsigned Form132Index = 0;
3570   const unsigned Form213Index = 1;
3571   const unsigned Form231Index = 2;
3572   const unsigned FormsNum = 3;
3573 
3574   bool IsIntrinOpcode;
3575   isFMA3(Opc, &IsIntrinOpcode);
3576 
3577   size_t GroupsNum;
3578   const uint16_t (*OpcodeGroups)[3];
3579   if (IsIntrinOpcode) {
3580     GroupsNum = array_lengthof(IntrinOpcodeGroups);
3581     OpcodeGroups = IntrinOpcodeGroups;
3582   } else {
3583     GroupsNum = array_lengthof(RegularOpcodeGroups);
3584     OpcodeGroups = RegularOpcodeGroups;
3585   }
3586 
3587   const uint16_t *FoundOpcodesGroup = nullptr;
3588   size_t FormIndex;
3589 
3590   // Look for the input opcode in the corresponding opcodes table.
3591   for (size_t GroupIndex = 0; GroupIndex < GroupsNum && !FoundOpcodesGroup;
3592          ++GroupIndex) {
3593     for (FormIndex = 0; FormIndex < FormsNum; ++FormIndex) {
3594       if (OpcodeGroups[GroupIndex][FormIndex] == Opc) {
3595         FoundOpcodesGroup = OpcodeGroups[GroupIndex];
3596         break;
3597       }
3598     }
3599   }
3600 
3601   // The input opcode does not match with any of the opcodes from the tables.
3602   // The unsupported FMA opcode must be added to one of the two opcode groups
3603   // defined above.
3604   assert(FoundOpcodesGroup != nullptr && "Unexpected FMA3 opcode");
3605 
3606   // Put the lowest index to SrcOpIdx1 to simplify the checks below.
3607   if (SrcOpIdx1 > SrcOpIdx2)
3608     std::swap(SrcOpIdx1, SrcOpIdx2);
3609 
3610   // TODO: Commuting the 1st operand of FMA*_Int requires some additional
3611   // analysis. The commute optimization is legal only if all users of FMA*_Int
3612   // use only the lowest element of the FMA*_Int instruction. Such analysis are
3613   // not implemented yet. So, just return 0 in that case.
3614   // When such analysis are available this place will be the right place for
3615   // calling it.
3616   if (IsIntrinOpcode && SrcOpIdx1 == 1)
3617     return 0;
3618 
3619   unsigned Case;
3620   if (SrcOpIdx1 == 1 && SrcOpIdx2 == 2)
3621     Case = 0;
3622   else if (SrcOpIdx1 == 1 && SrcOpIdx2 == 3)
3623     Case = 1;
3624   else if (SrcOpIdx1 == 2 && SrcOpIdx2 == 3)
3625     Case = 2;
3626   else
3627     return 0;
3628 
3629   // Define the FMA forms mapping array that helps to map input FMA form
3630   // to output FMA form to preserve the operation semantics after
3631   // commuting the operands.
3632   static const unsigned FormMapping[][3] = {
3633     // 0: SrcOpIdx1 == 1 && SrcOpIdx2 == 2;
3634     // FMA132 A, C, b; ==> FMA231 C, A, b;
3635     // FMA213 B, A, c; ==> FMA213 A, B, c;
3636     // FMA231 C, A, b; ==> FMA132 A, C, b;
3637     { Form231Index, Form213Index, Form132Index },
3638     // 1: SrcOpIdx1 == 1 && SrcOpIdx2 == 3;
3639     // FMA132 A, c, B; ==> FMA132 B, c, A;
3640     // FMA213 B, a, C; ==> FMA231 C, a, B;
3641     // FMA231 C, a, B; ==> FMA213 B, a, C;
3642     { Form132Index, Form231Index, Form213Index },
3643     // 2: SrcOpIdx1 == 2 && SrcOpIdx2 == 3;
3644     // FMA132 a, C, B; ==> FMA213 a, B, C;
3645     // FMA213 b, A, C; ==> FMA132 b, C, A;
3646     // FMA231 c, A, B; ==> FMA231 c, B, A;
3647     { Form213Index, Form132Index, Form231Index }
3648   };
3649 
3650   // Everything is ready, just adjust the FMA opcode and return it.
3651   FormIndex = FormMapping[Case][FormIndex];
3652   return FoundOpcodesGroup[FormIndex];
3653 }
3654 
findCommutedOpIndices(MachineInstr & MI,unsigned & SrcOpIdx1,unsigned & SrcOpIdx2) const3655 bool X86InstrInfo::findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1,
3656                                          unsigned &SrcOpIdx2) const {
3657   switch (MI.getOpcode()) {
3658   case X86::CMPPDrri:
3659   case X86::CMPPSrri:
3660   case X86::VCMPPDrri:
3661   case X86::VCMPPSrri:
3662   case X86::VCMPPDYrri:
3663   case X86::VCMPPSYrri: {
3664     // Float comparison can be safely commuted for
3665     // Ordered/Unordered/Equal/NotEqual tests
3666     unsigned Imm = MI.getOperand(3).getImm() & 0x7;
3667     switch (Imm) {
3668     case 0x00: // EQUAL
3669     case 0x03: // UNORDERED
3670     case 0x04: // NOT EQUAL
3671     case 0x07: // ORDERED
3672       // The indices of the commutable operands are 1 and 2.
3673       // Assign them to the returned operand indices here.
3674       return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1, 2);
3675     }
3676     return false;
3677   }
3678   default:
3679     if (isFMA3(MI.getOpcode()))
3680       return findFMA3CommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
3681     return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
3682   }
3683   return false;
3684 }
3685 
getCondFromBranchOpc(unsigned BrOpc)3686 static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
3687   switch (BrOpc) {
3688   default: return X86::COND_INVALID;
3689   case X86::JE_1:  return X86::COND_E;
3690   case X86::JNE_1: return X86::COND_NE;
3691   case X86::JL_1:  return X86::COND_L;
3692   case X86::JLE_1: return X86::COND_LE;
3693   case X86::JG_1:  return X86::COND_G;
3694   case X86::JGE_1: return X86::COND_GE;
3695   case X86::JB_1:  return X86::COND_B;
3696   case X86::JBE_1: return X86::COND_BE;
3697   case X86::JA_1:  return X86::COND_A;
3698   case X86::JAE_1: return X86::COND_AE;
3699   case X86::JS_1:  return X86::COND_S;
3700   case X86::JNS_1: return X86::COND_NS;
3701   case X86::JP_1:  return X86::COND_P;
3702   case X86::JNP_1: return X86::COND_NP;
3703   case X86::JO_1:  return X86::COND_O;
3704   case X86::JNO_1: return X86::COND_NO;
3705   }
3706 }
3707 
3708 /// Return condition code of a SET opcode.
getCondFromSETOpc(unsigned Opc)3709 static X86::CondCode getCondFromSETOpc(unsigned Opc) {
3710   switch (Opc) {
3711   default: return X86::COND_INVALID;
3712   case X86::SETAr:  case X86::SETAm:  return X86::COND_A;
3713   case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
3714   case X86::SETBr:  case X86::SETBm:  return X86::COND_B;
3715   case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
3716   case X86::SETEr:  case X86::SETEm:  return X86::COND_E;
3717   case X86::SETGr:  case X86::SETGm:  return X86::COND_G;
3718   case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
3719   case X86::SETLr:  case X86::SETLm:  return X86::COND_L;
3720   case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
3721   case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
3722   case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
3723   case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
3724   case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
3725   case X86::SETOr:  case X86::SETOm:  return X86::COND_O;
3726   case X86::SETPr:  case X86::SETPm:  return X86::COND_P;
3727   case X86::SETSr:  case X86::SETSm:  return X86::COND_S;
3728   }
3729 }
3730 
3731 /// Return condition code of a CMov opcode.
getCondFromCMovOpc(unsigned Opc)3732 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
3733   switch (Opc) {
3734   default: return X86::COND_INVALID;
3735   case X86::CMOVA16rm:  case X86::CMOVA16rr:  case X86::CMOVA32rm:
3736   case X86::CMOVA32rr:  case X86::CMOVA64rm:  case X86::CMOVA64rr:
3737     return X86::COND_A;
3738   case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
3739   case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
3740     return X86::COND_AE;
3741   case X86::CMOVB16rm:  case X86::CMOVB16rr:  case X86::CMOVB32rm:
3742   case X86::CMOVB32rr:  case X86::CMOVB64rm:  case X86::CMOVB64rr:
3743     return X86::COND_B;
3744   case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
3745   case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
3746     return X86::COND_BE;
3747   case X86::CMOVE16rm:  case X86::CMOVE16rr:  case X86::CMOVE32rm:
3748   case X86::CMOVE32rr:  case X86::CMOVE64rm:  case X86::CMOVE64rr:
3749     return X86::COND_E;
3750   case X86::CMOVG16rm:  case X86::CMOVG16rr:  case X86::CMOVG32rm:
3751   case X86::CMOVG32rr:  case X86::CMOVG64rm:  case X86::CMOVG64rr:
3752     return X86::COND_G;
3753   case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
3754   case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
3755     return X86::COND_GE;
3756   case X86::CMOVL16rm:  case X86::CMOVL16rr:  case X86::CMOVL32rm:
3757   case X86::CMOVL32rr:  case X86::CMOVL64rm:  case X86::CMOVL64rr:
3758     return X86::COND_L;
3759   case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
3760   case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
3761     return X86::COND_LE;
3762   case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
3763   case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
3764     return X86::COND_NE;
3765   case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
3766   case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
3767     return X86::COND_NO;
3768   case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
3769   case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
3770     return X86::COND_NP;
3771   case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
3772   case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
3773     return X86::COND_NS;
3774   case X86::CMOVO16rm:  case X86::CMOVO16rr:  case X86::CMOVO32rm:
3775   case X86::CMOVO32rr:  case X86::CMOVO64rm:  case X86::CMOVO64rr:
3776     return X86::COND_O;
3777   case X86::CMOVP16rm:  case X86::CMOVP16rr:  case X86::CMOVP32rm:
3778   case X86::CMOVP32rr:  case X86::CMOVP64rm:  case X86::CMOVP64rr:
3779     return X86::COND_P;
3780   case X86::CMOVS16rm:  case X86::CMOVS16rr:  case X86::CMOVS32rm:
3781   case X86::CMOVS32rr:  case X86::CMOVS64rm:  case X86::CMOVS64rr:
3782     return X86::COND_S;
3783   }
3784 }
3785 
GetCondBranchFromCond(X86::CondCode CC)3786 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
3787   switch (CC) {
3788   default: llvm_unreachable("Illegal condition code!");
3789   case X86::COND_E:  return X86::JE_1;
3790   case X86::COND_NE: return X86::JNE_1;
3791   case X86::COND_L:  return X86::JL_1;
3792   case X86::COND_LE: return X86::JLE_1;
3793   case X86::COND_G:  return X86::JG_1;
3794   case X86::COND_GE: return X86::JGE_1;
3795   case X86::COND_B:  return X86::JB_1;
3796   case X86::COND_BE: return X86::JBE_1;
3797   case X86::COND_A:  return X86::JA_1;
3798   case X86::COND_AE: return X86::JAE_1;
3799   case X86::COND_S:  return X86::JS_1;
3800   case X86::COND_NS: return X86::JNS_1;
3801   case X86::COND_P:  return X86::JP_1;
3802   case X86::COND_NP: return X86::JNP_1;
3803   case X86::COND_O:  return X86::JO_1;
3804   case X86::COND_NO: return X86::JNO_1;
3805   }
3806 }
3807 
3808 /// Return the inverse of the specified condition,
3809 /// e.g. turning COND_E to COND_NE.
GetOppositeBranchCondition(X86::CondCode CC)3810 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
3811   switch (CC) {
3812   default: llvm_unreachable("Illegal condition code!");
3813   case X86::COND_E:  return X86::COND_NE;
3814   case X86::COND_NE: return X86::COND_E;
3815   case X86::COND_L:  return X86::COND_GE;
3816   case X86::COND_LE: return X86::COND_G;
3817   case X86::COND_G:  return X86::COND_LE;
3818   case X86::COND_GE: return X86::COND_L;
3819   case X86::COND_B:  return X86::COND_AE;
3820   case X86::COND_BE: return X86::COND_A;
3821   case X86::COND_A:  return X86::COND_BE;
3822   case X86::COND_AE: return X86::COND_B;
3823   case X86::COND_S:  return X86::COND_NS;
3824   case X86::COND_NS: return X86::COND_S;
3825   case X86::COND_P:  return X86::COND_NP;
3826   case X86::COND_NP: return X86::COND_P;
3827   case X86::COND_O:  return X86::COND_NO;
3828   case X86::COND_NO: return X86::COND_O;
3829   case X86::COND_NE_OR_P:  return X86::COND_E_AND_NP;
3830   case X86::COND_E_AND_NP: return X86::COND_NE_OR_P;
3831   }
3832 }
3833 
3834 /// Assuming the flags are set by MI(a,b), return the condition code if we
3835 /// modify the instructions such that flags are set by MI(b,a).
getSwappedCondition(X86::CondCode CC)3836 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
3837   switch (CC) {
3838   default: return X86::COND_INVALID;
3839   case X86::COND_E:  return X86::COND_E;
3840   case X86::COND_NE: return X86::COND_NE;
3841   case X86::COND_L:  return X86::COND_G;
3842   case X86::COND_LE: return X86::COND_GE;
3843   case X86::COND_G:  return X86::COND_L;
3844   case X86::COND_GE: return X86::COND_LE;
3845   case X86::COND_B:  return X86::COND_A;
3846   case X86::COND_BE: return X86::COND_AE;
3847   case X86::COND_A:  return X86::COND_B;
3848   case X86::COND_AE: return X86::COND_BE;
3849   }
3850 }
3851 
3852 /// Return a set opcode for the given condition and
3853 /// whether it has memory operand.
getSETFromCond(CondCode CC,bool HasMemoryOperand)3854 unsigned X86::getSETFromCond(CondCode CC, bool HasMemoryOperand) {
3855   static const uint16_t Opc[16][2] = {
3856     { X86::SETAr,  X86::SETAm  },
3857     { X86::SETAEr, X86::SETAEm },
3858     { X86::SETBr,  X86::SETBm  },
3859     { X86::SETBEr, X86::SETBEm },
3860     { X86::SETEr,  X86::SETEm  },
3861     { X86::SETGr,  X86::SETGm  },
3862     { X86::SETGEr, X86::SETGEm },
3863     { X86::SETLr,  X86::SETLm  },
3864     { X86::SETLEr, X86::SETLEm },
3865     { X86::SETNEr, X86::SETNEm },
3866     { X86::SETNOr, X86::SETNOm },
3867     { X86::SETNPr, X86::SETNPm },
3868     { X86::SETNSr, X86::SETNSm },
3869     { X86::SETOr,  X86::SETOm  },
3870     { X86::SETPr,  X86::SETPm  },
3871     { X86::SETSr,  X86::SETSm  }
3872   };
3873 
3874   assert(CC <= LAST_VALID_COND && "Can only handle standard cond codes");
3875   return Opc[CC][HasMemoryOperand ? 1 : 0];
3876 }
3877 
3878 /// Return a cmov opcode for the given condition,
3879 /// register size in bytes, and operand type.
getCMovFromCond(CondCode CC,unsigned RegBytes,bool HasMemoryOperand)3880 unsigned X86::getCMovFromCond(CondCode CC, unsigned RegBytes,
3881                               bool HasMemoryOperand) {
3882   static const uint16_t Opc[32][3] = {
3883     { X86::CMOVA16rr,  X86::CMOVA32rr,  X86::CMOVA64rr  },
3884     { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
3885     { X86::CMOVB16rr,  X86::CMOVB32rr,  X86::CMOVB64rr  },
3886     { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
3887     { X86::CMOVE16rr,  X86::CMOVE32rr,  X86::CMOVE64rr  },
3888     { X86::CMOVG16rr,  X86::CMOVG32rr,  X86::CMOVG64rr  },
3889     { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
3890     { X86::CMOVL16rr,  X86::CMOVL32rr,  X86::CMOVL64rr  },
3891     { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
3892     { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
3893     { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
3894     { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
3895     { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
3896     { X86::CMOVO16rr,  X86::CMOVO32rr,  X86::CMOVO64rr  },
3897     { X86::CMOVP16rr,  X86::CMOVP32rr,  X86::CMOVP64rr  },
3898     { X86::CMOVS16rr,  X86::CMOVS32rr,  X86::CMOVS64rr  },
3899     { X86::CMOVA16rm,  X86::CMOVA32rm,  X86::CMOVA64rm  },
3900     { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
3901     { X86::CMOVB16rm,  X86::CMOVB32rm,  X86::CMOVB64rm  },
3902     { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
3903     { X86::CMOVE16rm,  X86::CMOVE32rm,  X86::CMOVE64rm  },
3904     { X86::CMOVG16rm,  X86::CMOVG32rm,  X86::CMOVG64rm  },
3905     { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
3906     { X86::CMOVL16rm,  X86::CMOVL32rm,  X86::CMOVL64rm  },
3907     { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
3908     { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
3909     { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
3910     { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
3911     { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
3912     { X86::CMOVO16rm,  X86::CMOVO32rm,  X86::CMOVO64rm  },
3913     { X86::CMOVP16rm,  X86::CMOVP32rm,  X86::CMOVP64rm  },
3914     { X86::CMOVS16rm,  X86::CMOVS32rm,  X86::CMOVS64rm  }
3915   };
3916 
3917   assert(CC < 16 && "Can only handle standard cond codes");
3918   unsigned Idx = HasMemoryOperand ? 16+CC : CC;
3919   switch(RegBytes) {
3920   default: llvm_unreachable("Illegal register size!");
3921   case 2: return Opc[Idx][0];
3922   case 4: return Opc[Idx][1];
3923   case 8: return Opc[Idx][2];
3924   }
3925 }
3926 
isUnpredicatedTerminator(const MachineInstr & MI) const3927 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const {
3928   if (!MI.isTerminator()) return false;
3929 
3930   // Conditional branch is a special case.
3931   if (MI.isBranch() && !MI.isBarrier())
3932     return true;
3933   if (!MI.isPredicable())
3934     return true;
3935   return !isPredicated(MI);
3936 }
3937 
3938 // Given a MBB and its TBB, find the FBB which was a fallthrough MBB (it may
3939 // not be a fallthrough MBB now due to layout changes). Return nullptr if the
3940 // fallthrough MBB cannot be identified.
getFallThroughMBB(MachineBasicBlock * MBB,MachineBasicBlock * TBB)3941 static MachineBasicBlock *getFallThroughMBB(MachineBasicBlock *MBB,
3942                                             MachineBasicBlock *TBB) {
3943   // Look for non-EHPad successors other than TBB. If we find exactly one, it
3944   // is the fallthrough MBB. If we find zero, then TBB is both the target MBB
3945   // and fallthrough MBB. If we find more than one, we cannot identify the
3946   // fallthrough MBB and should return nullptr.
3947   MachineBasicBlock *FallthroughBB = nullptr;
3948   for (auto SI = MBB->succ_begin(), SE = MBB->succ_end(); SI != SE; ++SI) {
3949     if ((*SI)->isEHPad() || (*SI == TBB && FallthroughBB))
3950       continue;
3951     // Return a nullptr if we found more than one fallthrough successor.
3952     if (FallthroughBB && FallthroughBB != TBB)
3953       return nullptr;
3954     FallthroughBB = *SI;
3955   }
3956   return FallthroughBB;
3957 }
3958 
AnalyzeBranchImpl(MachineBasicBlock & MBB,MachineBasicBlock * & TBB,MachineBasicBlock * & FBB,SmallVectorImpl<MachineOperand> & Cond,SmallVectorImpl<MachineInstr * > & CondBranches,bool AllowModify) const3959 bool X86InstrInfo::AnalyzeBranchImpl(
3960     MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB,
3961     SmallVectorImpl<MachineOperand> &Cond,
3962     SmallVectorImpl<MachineInstr *> &CondBranches, bool AllowModify) const {
3963 
3964   // Start from the bottom of the block and work up, examining the
3965   // terminator instructions.
3966   MachineBasicBlock::iterator I = MBB.end();
3967   MachineBasicBlock::iterator UnCondBrIter = MBB.end();
3968   while (I != MBB.begin()) {
3969     --I;
3970     if (I->isDebugValue())
3971       continue;
3972 
3973     // Working from the bottom, when we see a non-terminator instruction, we're
3974     // done.
3975     if (!isUnpredicatedTerminator(*I))
3976       break;
3977 
3978     // A terminator that isn't a branch can't easily be handled by this
3979     // analysis.
3980     if (!I->isBranch())
3981       return true;
3982 
3983     // Handle unconditional branches.
3984     if (I->getOpcode() == X86::JMP_1) {
3985       UnCondBrIter = I;
3986 
3987       if (!AllowModify) {
3988         TBB = I->getOperand(0).getMBB();
3989         continue;
3990       }
3991 
3992       // If the block has any instructions after a JMP, delete them.
3993       while (std::next(I) != MBB.end())
3994         std::next(I)->eraseFromParent();
3995 
3996       Cond.clear();
3997       FBB = nullptr;
3998 
3999       // Delete the JMP if it's equivalent to a fall-through.
4000       if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
4001         TBB = nullptr;
4002         I->eraseFromParent();
4003         I = MBB.end();
4004         UnCondBrIter = MBB.end();
4005         continue;
4006       }
4007 
4008       // TBB is used to indicate the unconditional destination.
4009       TBB = I->getOperand(0).getMBB();
4010       continue;
4011     }
4012 
4013     // Handle conditional branches.
4014     X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
4015     if (BranchCode == X86::COND_INVALID)
4016       return true;  // Can't handle indirect branch.
4017 
4018     // Working from the bottom, handle the first conditional branch.
4019     if (Cond.empty()) {
4020       MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
4021       if (AllowModify && UnCondBrIter != MBB.end() &&
4022           MBB.isLayoutSuccessor(TargetBB)) {
4023         // If we can modify the code and it ends in something like:
4024         //
4025         //     jCC L1
4026         //     jmp L2
4027         //   L1:
4028         //     ...
4029         //   L2:
4030         //
4031         // Then we can change this to:
4032         //
4033         //     jnCC L2
4034         //   L1:
4035         //     ...
4036         //   L2:
4037         //
4038         // Which is a bit more efficient.
4039         // We conditionally jump to the fall-through block.
4040         BranchCode = GetOppositeBranchCondition(BranchCode);
4041         unsigned JNCC = GetCondBranchFromCond(BranchCode);
4042         MachineBasicBlock::iterator OldInst = I;
4043 
4044         BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
4045           .addMBB(UnCondBrIter->getOperand(0).getMBB());
4046         BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_1))
4047           .addMBB(TargetBB);
4048 
4049         OldInst->eraseFromParent();
4050         UnCondBrIter->eraseFromParent();
4051 
4052         // Restart the analysis.
4053         UnCondBrIter = MBB.end();
4054         I = MBB.end();
4055         continue;
4056       }
4057 
4058       FBB = TBB;
4059       TBB = I->getOperand(0).getMBB();
4060       Cond.push_back(MachineOperand::CreateImm(BranchCode));
4061       CondBranches.push_back(&*I);
4062       continue;
4063     }
4064 
4065     // Handle subsequent conditional branches. Only handle the case where all
4066     // conditional branches branch to the same destination and their condition
4067     // opcodes fit one of the special multi-branch idioms.
4068     assert(Cond.size() == 1);
4069     assert(TBB);
4070 
4071     // If the conditions are the same, we can leave them alone.
4072     X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
4073     auto NewTBB = I->getOperand(0).getMBB();
4074     if (OldBranchCode == BranchCode && TBB == NewTBB)
4075       continue;
4076 
4077     // If they differ, see if they fit one of the known patterns. Theoretically,
4078     // we could handle more patterns here, but we shouldn't expect to see them
4079     // if instruction selection has done a reasonable job.
4080     if (TBB == NewTBB &&
4081                ((OldBranchCode == X86::COND_P && BranchCode == X86::COND_NE) ||
4082                 (OldBranchCode == X86::COND_NE && BranchCode == X86::COND_P))) {
4083       BranchCode = X86::COND_NE_OR_P;
4084     } else if ((OldBranchCode == X86::COND_NP && BranchCode == X86::COND_NE) ||
4085                (OldBranchCode == X86::COND_E && BranchCode == X86::COND_P)) {
4086       if (NewTBB != (FBB ? FBB : getFallThroughMBB(&MBB, TBB)))
4087         return true;
4088 
4089       // X86::COND_E_AND_NP usually has two different branch destinations.
4090       //
4091       // JP B1
4092       // JE B2
4093       // JMP B1
4094       // B1:
4095       // B2:
4096       //
4097       // Here this condition branches to B2 only if NP && E. It has another
4098       // equivalent form:
4099       //
4100       // JNE B1
4101       // JNP B2
4102       // JMP B1
4103       // B1:
4104       // B2:
4105       //
4106       // Similarly it branches to B2 only if E && NP. That is why this condition
4107       // is named with COND_E_AND_NP.
4108       BranchCode = X86::COND_E_AND_NP;
4109     } else
4110       return true;
4111 
4112     // Update the MachineOperand.
4113     Cond[0].setImm(BranchCode);
4114     CondBranches.push_back(&*I);
4115   }
4116 
4117   return false;
4118 }
4119 
analyzeBranch(MachineBasicBlock & MBB,MachineBasicBlock * & TBB,MachineBasicBlock * & FBB,SmallVectorImpl<MachineOperand> & Cond,bool AllowModify) const4120 bool X86InstrInfo::analyzeBranch(MachineBasicBlock &MBB,
4121                                  MachineBasicBlock *&TBB,
4122                                  MachineBasicBlock *&FBB,
4123                                  SmallVectorImpl<MachineOperand> &Cond,
4124                                  bool AllowModify) const {
4125   SmallVector<MachineInstr *, 4> CondBranches;
4126   return AnalyzeBranchImpl(MBB, TBB, FBB, Cond, CondBranches, AllowModify);
4127 }
4128 
analyzeBranchPredicate(MachineBasicBlock & MBB,MachineBranchPredicate & MBP,bool AllowModify) const4129 bool X86InstrInfo::analyzeBranchPredicate(MachineBasicBlock &MBB,
4130                                           MachineBranchPredicate &MBP,
4131                                           bool AllowModify) const {
4132   using namespace std::placeholders;
4133 
4134   SmallVector<MachineOperand, 4> Cond;
4135   SmallVector<MachineInstr *, 4> CondBranches;
4136   if (AnalyzeBranchImpl(MBB, MBP.TrueDest, MBP.FalseDest, Cond, CondBranches,
4137                         AllowModify))
4138     return true;
4139 
4140   if (Cond.size() != 1)
4141     return true;
4142 
4143   assert(MBP.TrueDest && "expected!");
4144 
4145   if (!MBP.FalseDest)
4146     MBP.FalseDest = MBB.getNextNode();
4147 
4148   const TargetRegisterInfo *TRI = &getRegisterInfo();
4149 
4150   MachineInstr *ConditionDef = nullptr;
4151   bool SingleUseCondition = true;
4152 
4153   for (auto I = std::next(MBB.rbegin()), E = MBB.rend(); I != E; ++I) {
4154     if (I->modifiesRegister(X86::EFLAGS, TRI)) {
4155       ConditionDef = &*I;
4156       break;
4157     }
4158 
4159     if (I->readsRegister(X86::EFLAGS, TRI))
4160       SingleUseCondition = false;
4161   }
4162 
4163   if (!ConditionDef)
4164     return true;
4165 
4166   if (SingleUseCondition) {
4167     for (auto *Succ : MBB.successors())
4168       if (Succ->isLiveIn(X86::EFLAGS))
4169         SingleUseCondition = false;
4170   }
4171 
4172   MBP.ConditionDef = ConditionDef;
4173   MBP.SingleUseCondition = SingleUseCondition;
4174 
4175   // Currently we only recognize the simple pattern:
4176   //
4177   //   test %reg, %reg
4178   //   je %label
4179   //
4180   const unsigned TestOpcode =
4181       Subtarget.is64Bit() ? X86::TEST64rr : X86::TEST32rr;
4182 
4183   if (ConditionDef->getOpcode() == TestOpcode &&
4184       ConditionDef->getNumOperands() == 3 &&
4185       ConditionDef->getOperand(0).isIdenticalTo(ConditionDef->getOperand(1)) &&
4186       (Cond[0].getImm() == X86::COND_NE || Cond[0].getImm() == X86::COND_E)) {
4187     MBP.LHS = ConditionDef->getOperand(0);
4188     MBP.RHS = MachineOperand::CreateImm(0);
4189     MBP.Predicate = Cond[0].getImm() == X86::COND_NE
4190                         ? MachineBranchPredicate::PRED_NE
4191                         : MachineBranchPredicate::PRED_EQ;
4192     return false;
4193   }
4194 
4195   return true;
4196 }
4197 
RemoveBranch(MachineBasicBlock & MBB) const4198 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
4199   MachineBasicBlock::iterator I = MBB.end();
4200   unsigned Count = 0;
4201 
4202   while (I != MBB.begin()) {
4203     --I;
4204     if (I->isDebugValue())
4205       continue;
4206     if (I->getOpcode() != X86::JMP_1 &&
4207         getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
4208       break;
4209     // Remove the branch.
4210     I->eraseFromParent();
4211     I = MBB.end();
4212     ++Count;
4213   }
4214 
4215   return Count;
4216 }
4217 
InsertBranch(MachineBasicBlock & MBB,MachineBasicBlock * TBB,MachineBasicBlock * FBB,ArrayRef<MachineOperand> Cond,const DebugLoc & DL) const4218 unsigned X86InstrInfo::InsertBranch(MachineBasicBlock &MBB,
4219                                     MachineBasicBlock *TBB,
4220                                     MachineBasicBlock *FBB,
4221                                     ArrayRef<MachineOperand> Cond,
4222                                     const DebugLoc &DL) const {
4223   // Shouldn't be a fall through.
4224   assert(TBB && "InsertBranch must not be told to insert a fallthrough");
4225   assert((Cond.size() == 1 || Cond.size() == 0) &&
4226          "X86 branch conditions have one component!");
4227 
4228   if (Cond.empty()) {
4229     // Unconditional branch?
4230     assert(!FBB && "Unconditional branch with multiple successors!");
4231     BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(TBB);
4232     return 1;
4233   }
4234 
4235   // If FBB is null, it is implied to be a fall-through block.
4236   bool FallThru = FBB == nullptr;
4237 
4238   // Conditional branch.
4239   unsigned Count = 0;
4240   X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
4241   switch (CC) {
4242   case X86::COND_NE_OR_P:
4243     // Synthesize NE_OR_P with two branches.
4244     BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(TBB);
4245     ++Count;
4246     BuildMI(&MBB, DL, get(X86::JP_1)).addMBB(TBB);
4247     ++Count;
4248     break;
4249   case X86::COND_E_AND_NP:
4250     // Use the next block of MBB as FBB if it is null.
4251     if (FBB == nullptr) {
4252       FBB = getFallThroughMBB(&MBB, TBB);
4253       assert(FBB && "MBB cannot be the last block in function when the false "
4254                     "body is a fall-through.");
4255     }
4256     // Synthesize COND_E_AND_NP with two branches.
4257     BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(FBB);
4258     ++Count;
4259     BuildMI(&MBB, DL, get(X86::JNP_1)).addMBB(TBB);
4260     ++Count;
4261     break;
4262   default: {
4263     unsigned Opc = GetCondBranchFromCond(CC);
4264     BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
4265     ++Count;
4266   }
4267   }
4268   if (!FallThru) {
4269     // Two-way Conditional branch. Insert the second branch.
4270     BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(FBB);
4271     ++Count;
4272   }
4273   return Count;
4274 }
4275 
4276 bool X86InstrInfo::
canInsertSelect(const MachineBasicBlock & MBB,ArrayRef<MachineOperand> Cond,unsigned TrueReg,unsigned FalseReg,int & CondCycles,int & TrueCycles,int & FalseCycles) const4277 canInsertSelect(const MachineBasicBlock &MBB,
4278                 ArrayRef<MachineOperand> Cond,
4279                 unsigned TrueReg, unsigned FalseReg,
4280                 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
4281   // Not all subtargets have cmov instructions.
4282   if (!Subtarget.hasCMov())
4283     return false;
4284   if (Cond.size() != 1)
4285     return false;
4286   // We cannot do the composite conditions, at least not in SSA form.
4287   if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
4288     return false;
4289 
4290   // Check register classes.
4291   const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
4292   const TargetRegisterClass *RC =
4293     RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
4294   if (!RC)
4295     return false;
4296 
4297   // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
4298   if (X86::GR16RegClass.hasSubClassEq(RC) ||
4299       X86::GR32RegClass.hasSubClassEq(RC) ||
4300       X86::GR64RegClass.hasSubClassEq(RC)) {
4301     // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
4302     // Bridge. Probably Ivy Bridge as well.
4303     CondCycles = 2;
4304     TrueCycles = 2;
4305     FalseCycles = 2;
4306     return true;
4307   }
4308 
4309   // Can't do vectors.
4310   return false;
4311 }
4312 
insertSelect(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,const DebugLoc & DL,unsigned DstReg,ArrayRef<MachineOperand> Cond,unsigned TrueReg,unsigned FalseReg) const4313 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
4314                                 MachineBasicBlock::iterator I,
4315                                 const DebugLoc &DL, unsigned DstReg,
4316                                 ArrayRef<MachineOperand> Cond, unsigned TrueReg,
4317                                 unsigned FalseReg) const {
4318   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
4319   assert(Cond.size() == 1 && "Invalid Cond array");
4320   unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
4321                                  MRI.getRegClass(DstReg)->getSize(),
4322                                  false /*HasMemoryOperand*/);
4323   BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
4324 }
4325 
4326 /// Test if the given register is a physical h register.
isHReg(unsigned Reg)4327 static bool isHReg(unsigned Reg) {
4328   return X86::GR8_ABCD_HRegClass.contains(Reg);
4329 }
4330 
4331 // Try and copy between VR128/VR64 and GR64 registers.
CopyToFromAsymmetricReg(unsigned DestReg,unsigned SrcReg,const X86Subtarget & Subtarget)4332 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
4333                                         const X86Subtarget &Subtarget) {
4334 
4335   // SrcReg(VR128) -> DestReg(GR64)
4336   // SrcReg(VR64)  -> DestReg(GR64)
4337   // SrcReg(GR64)  -> DestReg(VR128)
4338   // SrcReg(GR64)  -> DestReg(VR64)
4339 
4340   bool HasAVX = Subtarget.hasAVX();
4341   bool HasAVX512 = Subtarget.hasAVX512();
4342   if (X86::GR64RegClass.contains(DestReg)) {
4343     if (X86::VR128XRegClass.contains(SrcReg))
4344       // Copy from a VR128 register to a GR64 register.
4345       return HasAVX512 ? X86::VMOVPQIto64Zrr: (HasAVX ? X86::VMOVPQIto64rr :
4346                                                X86::MOVPQIto64rr);
4347     if (X86::VR64RegClass.contains(SrcReg))
4348       // Copy from a VR64 register to a GR64 register.
4349       return X86::MMX_MOVD64from64rr;
4350   } else if (X86::GR64RegClass.contains(SrcReg)) {
4351     // Copy from a GR64 register to a VR128 register.
4352     if (X86::VR128XRegClass.contains(DestReg))
4353       return HasAVX512 ? X86::VMOV64toPQIZrr: (HasAVX ? X86::VMOV64toPQIrr :
4354                                                X86::MOV64toPQIrr);
4355     // Copy from a GR64 register to a VR64 register.
4356     if (X86::VR64RegClass.contains(DestReg))
4357       return X86::MMX_MOVD64to64rr;
4358   }
4359 
4360   // SrcReg(FR32) -> DestReg(GR32)
4361   // SrcReg(GR32) -> DestReg(FR32)
4362 
4363   if (X86::GR32RegClass.contains(DestReg) && X86::FR32XRegClass.contains(SrcReg))
4364     // Copy from a FR32 register to a GR32 register.
4365     return HasAVX512 ? X86::VMOVSS2DIZrr : (HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr);
4366 
4367   if (X86::FR32XRegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
4368     // Copy from a GR32 register to a FR32 register.
4369     return HasAVX512 ? X86::VMOVDI2SSZrr : (HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr);
4370   return 0;
4371 }
4372 
isMaskRegClass(const TargetRegisterClass * RC)4373 static bool isMaskRegClass(const TargetRegisterClass *RC) {
4374   // All KMASK RegClasses hold the same k registers, can be tested against anyone.
4375   return X86::VK16RegClass.hasSubClassEq(RC);
4376 }
4377 
MaskRegClassContains(unsigned Reg)4378 static bool MaskRegClassContains(unsigned Reg) {
4379   // All KMASK RegClasses hold the same k registers, can be tested against anyone.
4380   return X86::VK16RegClass.contains(Reg);
4381 }
4382 
GRRegClassContains(unsigned Reg)4383 static bool GRRegClassContains(unsigned Reg) {
4384   return X86::GR64RegClass.contains(Reg) ||
4385          X86::GR32RegClass.contains(Reg) ||
4386          X86::GR16RegClass.contains(Reg) ||
4387          X86::GR8RegClass.contains(Reg);
4388 }
4389 static
copyPhysRegOpcode_AVX512_DQ(unsigned & DestReg,unsigned & SrcReg)4390 unsigned copyPhysRegOpcode_AVX512_DQ(unsigned& DestReg, unsigned& SrcReg) {
4391   if (MaskRegClassContains(SrcReg) && X86::GR8RegClass.contains(DestReg)) {
4392     DestReg = getX86SubSuperRegister(DestReg, 32);
4393     return X86::KMOVBrk;
4394   }
4395   if (MaskRegClassContains(DestReg) && X86::GR8RegClass.contains(SrcReg)) {
4396     SrcReg = getX86SubSuperRegister(SrcReg, 32);
4397     return X86::KMOVBkr;
4398   }
4399   return 0;
4400 }
4401 
4402 static
copyPhysRegOpcode_AVX512_BW(unsigned & DestReg,unsigned & SrcReg)4403 unsigned copyPhysRegOpcode_AVX512_BW(unsigned& DestReg, unsigned& SrcReg) {
4404   if (MaskRegClassContains(SrcReg) && MaskRegClassContains(DestReg))
4405     return X86::KMOVQkk;
4406   if (MaskRegClassContains(SrcReg) && X86::GR32RegClass.contains(DestReg))
4407     return X86::KMOVDrk;
4408   if (MaskRegClassContains(SrcReg) && X86::GR64RegClass.contains(DestReg))
4409     return X86::KMOVQrk;
4410   if (MaskRegClassContains(DestReg) && X86::GR32RegClass.contains(SrcReg))
4411     return X86::KMOVDkr;
4412   if (MaskRegClassContains(DestReg) && X86::GR64RegClass.contains(SrcReg))
4413     return X86::KMOVQkr;
4414   return 0;
4415 }
4416 
4417 static
copyPhysRegOpcode_AVX512(unsigned & DestReg,unsigned & SrcReg,const X86Subtarget & Subtarget)4418 unsigned copyPhysRegOpcode_AVX512(unsigned& DestReg, unsigned& SrcReg,
4419                                   const X86Subtarget &Subtarget)
4420 {
4421   if (Subtarget.hasDQI())
4422     if (auto Opc = copyPhysRegOpcode_AVX512_DQ(DestReg, SrcReg))
4423       return Opc;
4424   if (Subtarget.hasBWI())
4425     if (auto Opc = copyPhysRegOpcode_AVX512_BW(DestReg, SrcReg))
4426       return Opc;
4427   if (X86::VR128XRegClass.contains(DestReg, SrcReg) ||
4428       X86::VR256XRegClass.contains(DestReg, SrcReg) ||
4429       X86::VR512RegClass.contains(DestReg, SrcReg)) {
4430      DestReg = get512BitSuperRegister(DestReg);
4431      SrcReg = get512BitSuperRegister(SrcReg);
4432      return X86::VMOVAPSZrr;
4433   }
4434   if (MaskRegClassContains(DestReg) && MaskRegClassContains(SrcReg))
4435     return X86::KMOVWkk;
4436   if (MaskRegClassContains(DestReg) && GRRegClassContains(SrcReg)) {
4437     SrcReg = getX86SubSuperRegister(SrcReg, 32);
4438     return X86::KMOVWkr;
4439   }
4440   if (GRRegClassContains(DestReg) && MaskRegClassContains(SrcReg)) {
4441     DestReg = getX86SubSuperRegister(DestReg, 32);
4442     return X86::KMOVWrk;
4443   }
4444   return 0;
4445 }
4446 
copyPhysReg(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,const DebugLoc & DL,unsigned DestReg,unsigned SrcReg,bool KillSrc) const4447 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
4448                                MachineBasicBlock::iterator MI,
4449                                const DebugLoc &DL, unsigned DestReg,
4450                                unsigned SrcReg, bool KillSrc) const {
4451   // First deal with the normal symmetric copies.
4452   bool HasAVX = Subtarget.hasAVX();
4453   bool HasAVX512 = Subtarget.hasAVX512();
4454   unsigned Opc = 0;
4455   if (X86::GR64RegClass.contains(DestReg, SrcReg))
4456     Opc = X86::MOV64rr;
4457   else if (X86::GR32RegClass.contains(DestReg, SrcReg))
4458     Opc = X86::MOV32rr;
4459   else if (X86::GR16RegClass.contains(DestReg, SrcReg))
4460     Opc = X86::MOV16rr;
4461   else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
4462     // Copying to or from a physical H register on x86-64 requires a NOREX
4463     // move.  Otherwise use a normal move.
4464     if ((isHReg(DestReg) || isHReg(SrcReg)) &&
4465         Subtarget.is64Bit()) {
4466       Opc = X86::MOV8rr_NOREX;
4467       // Both operands must be encodable without an REX prefix.
4468       assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
4469              "8-bit H register can not be copied outside GR8_NOREX");
4470     } else
4471       Opc = X86::MOV8rr;
4472   }
4473   else if (X86::VR64RegClass.contains(DestReg, SrcReg))
4474     Opc = X86::MMX_MOVQ64rr;
4475   else if (HasAVX512)
4476     Opc = copyPhysRegOpcode_AVX512(DestReg, SrcReg, Subtarget);
4477   else if (X86::VR128RegClass.contains(DestReg, SrcReg))
4478     Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
4479   else if (X86::VR256RegClass.contains(DestReg, SrcReg))
4480     Opc = X86::VMOVAPSYrr;
4481   if (!Opc)
4482     Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget);
4483 
4484   if (Opc) {
4485     BuildMI(MBB, MI, DL, get(Opc), DestReg)
4486       .addReg(SrcReg, getKillRegState(KillSrc));
4487     return;
4488   }
4489 
4490   bool FromEFLAGS = SrcReg == X86::EFLAGS;
4491   bool ToEFLAGS = DestReg == X86::EFLAGS;
4492   int Reg = FromEFLAGS ? DestReg : SrcReg;
4493   bool is32 = X86::GR32RegClass.contains(Reg);
4494   bool is64 = X86::GR64RegClass.contains(Reg);
4495 
4496   if ((FromEFLAGS || ToEFLAGS) && (is32 || is64)) {
4497     int Mov = is64 ? X86::MOV64rr : X86::MOV32rr;
4498     int Push = is64 ? X86::PUSH64r : X86::PUSH32r;
4499     int PushF = is64 ? X86::PUSHF64 : X86::PUSHF32;
4500     int Pop = is64 ? X86::POP64r : X86::POP32r;
4501     int PopF = is64 ? X86::POPF64 : X86::POPF32;
4502     int AX = is64 ? X86::RAX : X86::EAX;
4503 
4504     if (!Subtarget.hasLAHFSAHF()) {
4505       assert(Subtarget.is64Bit() &&
4506              "Not having LAHF/SAHF only happens on 64-bit.");
4507       // Moving EFLAGS to / from another register requires a push and a pop.
4508       // Notice that we have to adjust the stack if we don't want to clobber the
4509       // first frame index. See X86FrameLowering.cpp - usesTheStack.
4510       if (FromEFLAGS) {
4511         BuildMI(MBB, MI, DL, get(PushF));
4512         BuildMI(MBB, MI, DL, get(Pop), DestReg);
4513       }
4514       if (ToEFLAGS) {
4515         BuildMI(MBB, MI, DL, get(Push))
4516             .addReg(SrcReg, getKillRegState(KillSrc));
4517         BuildMI(MBB, MI, DL, get(PopF));
4518       }
4519       return;
4520     }
4521 
4522     // The flags need to be saved, but saving EFLAGS with PUSHF/POPF is
4523     // inefficient. Instead:
4524     //   - Save the overflow flag OF into AL using SETO, and restore it using a
4525     //     signed 8-bit addition of AL and INT8_MAX.
4526     //   - Save/restore the bottom 8 EFLAGS bits (CF, PF, AF, ZF, SF) to/from AH
4527     //     using LAHF/SAHF.
4528     //   - When RAX/EAX is live and isn't the destination register, make sure it
4529     //     isn't clobbered by PUSH/POP'ing it before and after saving/restoring
4530     //     the flags.
4531     // This approach is ~2.25x faster than using PUSHF/POPF.
4532     //
4533     // This is still somewhat inefficient because we don't know which flags are
4534     // actually live inside EFLAGS. Were we able to do a single SETcc instead of
4535     // SETO+LAHF / ADDB+SAHF the code could be 1.02x faster.
4536     //
4537     // PUSHF/POPF is also potentially incorrect because it affects other flags
4538     // such as TF/IF/DF, which LLVM doesn't model.
4539     //
4540     // Notice that we have to adjust the stack if we don't want to clobber the
4541     // first frame index.
4542     // See X86ISelLowering.cpp - X86::hasCopyImplyingStackAdjustment.
4543 
4544     const TargetRegisterInfo *TRI = &getRegisterInfo();
4545     MachineBasicBlock::LivenessQueryResult LQR =
4546         MBB.computeRegisterLiveness(TRI, AX, MI);
4547     // We do not want to save and restore AX if we do not have to.
4548     // Moreover, if we do so whereas AX is dead, we would need to set
4549     // an undef flag on the use of AX, otherwise the verifier will
4550     // complain that we read an undef value.
4551     // We do not want to change the behavior of the machine verifier
4552     // as this is usually wrong to read an undef value.
4553     if (MachineBasicBlock::LQR_Unknown == LQR) {
4554       LivePhysRegs LPR(TRI);
4555       LPR.addLiveOuts(MBB);
4556       MachineBasicBlock::iterator I = MBB.end();
4557       while (I != MI) {
4558         --I;
4559         LPR.stepBackward(*I);
4560       }
4561       // AX contains the top most register in the aliasing hierarchy.
4562       // It may not be live, but one of its aliases may be.
4563       for (MCRegAliasIterator AI(AX, TRI, true);
4564            AI.isValid() && LQR != MachineBasicBlock::LQR_Live; ++AI)
4565         LQR = LPR.contains(*AI) ? MachineBasicBlock::LQR_Live
4566                                 : MachineBasicBlock::LQR_Dead;
4567     }
4568     bool AXDead = (Reg == AX) || (MachineBasicBlock::LQR_Dead == LQR);
4569     if (!AXDead)
4570       BuildMI(MBB, MI, DL, get(Push)).addReg(AX, getKillRegState(true));
4571     if (FromEFLAGS) {
4572       BuildMI(MBB, MI, DL, get(X86::SETOr), X86::AL);
4573       BuildMI(MBB, MI, DL, get(X86::LAHF));
4574       BuildMI(MBB, MI, DL, get(Mov), Reg).addReg(AX);
4575     }
4576     if (ToEFLAGS) {
4577       BuildMI(MBB, MI, DL, get(Mov), AX).addReg(Reg, getKillRegState(KillSrc));
4578       BuildMI(MBB, MI, DL, get(X86::ADD8ri), X86::AL)
4579           .addReg(X86::AL)
4580           .addImm(INT8_MAX);
4581       BuildMI(MBB, MI, DL, get(X86::SAHF));
4582     }
4583     if (!AXDead)
4584       BuildMI(MBB, MI, DL, get(Pop), AX);
4585     return;
4586   }
4587 
4588   DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
4589                << " to " << RI.getName(DestReg) << '\n');
4590   llvm_unreachable("Cannot emit physreg copy instruction");
4591 }
4592 
getLoadStoreMaskRegOpcode(const TargetRegisterClass * RC,bool load)4593 static unsigned getLoadStoreMaskRegOpcode(const TargetRegisterClass *RC,
4594                                           bool load) {
4595   switch (RC->getSize()) {
4596   default:
4597     llvm_unreachable("Unknown spill size");
4598   case 2:
4599     return load ? X86::KMOVWkm : X86::KMOVWmk;
4600   case 4:
4601     return load ? X86::KMOVDkm : X86::KMOVDmk;
4602   case 8:
4603     return load ? X86::KMOVQkm : X86::KMOVQmk;
4604   }
4605 }
4606 
getLoadStoreRegOpcode(unsigned Reg,const TargetRegisterClass * RC,bool isStackAligned,const X86Subtarget & STI,bool load)4607 static unsigned getLoadStoreRegOpcode(unsigned Reg,
4608                                       const TargetRegisterClass *RC,
4609                                       bool isStackAligned,
4610                                       const X86Subtarget &STI,
4611                                       bool load) {
4612   if (STI.hasAVX512()) {
4613     if (isMaskRegClass(RC))
4614       return getLoadStoreMaskRegOpcode(RC, load);
4615     if (RC->getSize() == 4 && X86::FR32XRegClass.hasSubClassEq(RC))
4616       return load ? X86::VMOVSSZrm : X86::VMOVSSZmr;
4617     if (RC->getSize() == 8 && X86::FR64XRegClass.hasSubClassEq(RC))
4618       return load ? X86::VMOVSDZrm : X86::VMOVSDZmr;
4619     if (X86::VR512RegClass.hasSubClassEq(RC))
4620       return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
4621   }
4622 
4623   bool HasAVX = STI.hasAVX();
4624   switch (RC->getSize()) {
4625   default:
4626     llvm_unreachable("Unknown spill size");
4627   case 1:
4628     assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
4629     if (STI.is64Bit())
4630       // Copying to or from a physical H register on x86-64 requires a NOREX
4631       // move.  Otherwise use a normal move.
4632       if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
4633         return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
4634     return load ? X86::MOV8rm : X86::MOV8mr;
4635   case 2:
4636     assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
4637     return load ? X86::MOV16rm : X86::MOV16mr;
4638   case 4:
4639     if (X86::GR32RegClass.hasSubClassEq(RC))
4640       return load ? X86::MOV32rm : X86::MOV32mr;
4641     if (X86::FR32RegClass.hasSubClassEq(RC))
4642       return load ?
4643         (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
4644         (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
4645     if (X86::RFP32RegClass.hasSubClassEq(RC))
4646       return load ? X86::LD_Fp32m : X86::ST_Fp32m;
4647     llvm_unreachable("Unknown 4-byte regclass");
4648   case 8:
4649     if (X86::GR64RegClass.hasSubClassEq(RC))
4650       return load ? X86::MOV64rm : X86::MOV64mr;
4651     if (X86::FR64RegClass.hasSubClassEq(RC))
4652       return load ?
4653         (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
4654         (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
4655     if (X86::VR64RegClass.hasSubClassEq(RC))
4656       return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
4657     if (X86::RFP64RegClass.hasSubClassEq(RC))
4658       return load ? X86::LD_Fp64m : X86::ST_Fp64m;
4659     llvm_unreachable("Unknown 8-byte regclass");
4660   case 10:
4661     assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
4662     return load ? X86::LD_Fp80m : X86::ST_FpP80m;
4663   case 16: {
4664     assert((X86::VR128RegClass.hasSubClassEq(RC) ||
4665             X86::VR128XRegClass.hasSubClassEq(RC))&& "Unknown 16-byte regclass");
4666     // If stack is realigned we can use aligned stores.
4667     if (X86::VR128RegClass.hasSubClassEq(RC)) {
4668       if (isStackAligned)
4669         return load ? (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm)
4670                     : (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
4671       else
4672         return load ? (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm)
4673                     : (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
4674     }
4675     assert(STI.hasVLX() && "Using extended register requires VLX");
4676     if (isStackAligned)
4677       return load ? X86::VMOVAPSZ128rm : X86::VMOVAPSZ128mr;
4678     else
4679       return load ? X86::VMOVUPSZ128rm : X86::VMOVUPSZ128mr;
4680   }
4681   case 32:
4682     assert((X86::VR256RegClass.hasSubClassEq(RC) ||
4683             X86::VR256XRegClass.hasSubClassEq(RC)) && "Unknown 32-byte regclass");
4684     // If stack is realigned we can use aligned stores.
4685     if (X86::VR256RegClass.hasSubClassEq(RC)) {
4686       if (isStackAligned)
4687         return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
4688       else
4689         return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
4690     }
4691     assert(STI.hasVLX() && "Using extended register requires VLX");
4692     if (isStackAligned)
4693       return load ? X86::VMOVAPSZ256rm : X86::VMOVAPSZ256mr;
4694     else
4695       return load ? X86::VMOVUPSZ256rm : X86::VMOVUPSZ256mr;
4696   case 64:
4697     assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
4698     assert(STI.hasVLX() && "Using 512-bit register requires AVX512");
4699     if (isStackAligned)
4700       return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
4701     else
4702       return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
4703   }
4704 }
4705 
getMemOpBaseRegImmOfs(MachineInstr & MemOp,unsigned & BaseReg,int64_t & Offset,const TargetRegisterInfo * TRI) const4706 bool X86InstrInfo::getMemOpBaseRegImmOfs(MachineInstr &MemOp, unsigned &BaseReg,
4707                                          int64_t &Offset,
4708                                          const TargetRegisterInfo *TRI) const {
4709   const MCInstrDesc &Desc = MemOp.getDesc();
4710   int MemRefBegin = X86II::getMemoryOperandNo(Desc.TSFlags);
4711   if (MemRefBegin < 0)
4712     return false;
4713 
4714   MemRefBegin += X86II::getOperandBias(Desc);
4715 
4716   MachineOperand &BaseMO = MemOp.getOperand(MemRefBegin + X86::AddrBaseReg);
4717   if (!BaseMO.isReg()) // Can be an MO_FrameIndex
4718     return false;
4719 
4720   BaseReg = BaseMO.getReg();
4721   if (MemOp.getOperand(MemRefBegin + X86::AddrScaleAmt).getImm() != 1)
4722     return false;
4723 
4724   if (MemOp.getOperand(MemRefBegin + X86::AddrIndexReg).getReg() !=
4725       X86::NoRegister)
4726     return false;
4727 
4728   const MachineOperand &DispMO = MemOp.getOperand(MemRefBegin + X86::AddrDisp);
4729 
4730   // Displacement can be symbolic
4731   if (!DispMO.isImm())
4732     return false;
4733 
4734   Offset = DispMO.getImm();
4735 
4736   return MemOp.getOperand(MemRefBegin + X86::AddrIndexReg).getReg() ==
4737          X86::NoRegister;
4738 }
4739 
getStoreRegOpcode(unsigned SrcReg,const TargetRegisterClass * RC,bool isStackAligned,const X86Subtarget & STI)4740 static unsigned getStoreRegOpcode(unsigned SrcReg,
4741                                   const TargetRegisterClass *RC,
4742                                   bool isStackAligned,
4743                                   const X86Subtarget &STI) {
4744   return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, STI, false);
4745 }
4746 
4747 
getLoadRegOpcode(unsigned DestReg,const TargetRegisterClass * RC,bool isStackAligned,const X86Subtarget & STI)4748 static unsigned getLoadRegOpcode(unsigned DestReg,
4749                                  const TargetRegisterClass *RC,
4750                                  bool isStackAligned,
4751                                  const X86Subtarget &STI) {
4752   return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, STI, true);
4753 }
4754 
storeRegToStackSlot(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,unsigned SrcReg,bool isKill,int FrameIdx,const TargetRegisterClass * RC,const TargetRegisterInfo * TRI) const4755 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
4756                                        MachineBasicBlock::iterator MI,
4757                                        unsigned SrcReg, bool isKill, int FrameIdx,
4758                                        const TargetRegisterClass *RC,
4759                                        const TargetRegisterInfo *TRI) const {
4760   const MachineFunction &MF = *MBB.getParent();
4761   assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
4762          "Stack slot too small for store");
4763   unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
4764   bool isAligned =
4765       (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
4766       RI.canRealignStack(MF);
4767   unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
4768   DebugLoc DL = MBB.findDebugLoc(MI);
4769   addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
4770     .addReg(SrcReg, getKillRegState(isKill));
4771 }
4772 
storeRegToAddr(MachineFunction & MF,unsigned SrcReg,bool isKill,SmallVectorImpl<MachineOperand> & Addr,const TargetRegisterClass * RC,MachineInstr::mmo_iterator MMOBegin,MachineInstr::mmo_iterator MMOEnd,SmallVectorImpl<MachineInstr * > & NewMIs) const4773 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
4774                                   bool isKill,
4775                                   SmallVectorImpl<MachineOperand> &Addr,
4776                                   const TargetRegisterClass *RC,
4777                                   MachineInstr::mmo_iterator MMOBegin,
4778                                   MachineInstr::mmo_iterator MMOEnd,
4779                                   SmallVectorImpl<MachineInstr*> &NewMIs) const {
4780   unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
4781   bool isAligned = MMOBegin != MMOEnd &&
4782                    (*MMOBegin)->getAlignment() >= Alignment;
4783   unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
4784   DebugLoc DL;
4785   MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
4786   for (unsigned i = 0, e = Addr.size(); i != e; ++i)
4787     MIB.addOperand(Addr[i]);
4788   MIB.addReg(SrcReg, getKillRegState(isKill));
4789   (*MIB).setMemRefs(MMOBegin, MMOEnd);
4790   NewMIs.push_back(MIB);
4791 }
4792 
4793 
loadRegFromStackSlot(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,unsigned DestReg,int FrameIdx,const TargetRegisterClass * RC,const TargetRegisterInfo * TRI) const4794 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
4795                                         MachineBasicBlock::iterator MI,
4796                                         unsigned DestReg, int FrameIdx,
4797                                         const TargetRegisterClass *RC,
4798                                         const TargetRegisterInfo *TRI) const {
4799   const MachineFunction &MF = *MBB.getParent();
4800   unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
4801   bool isAligned =
4802       (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
4803       RI.canRealignStack(MF);
4804   unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
4805   DebugLoc DL = MBB.findDebugLoc(MI);
4806   addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
4807 }
4808 
loadRegFromAddr(MachineFunction & MF,unsigned DestReg,SmallVectorImpl<MachineOperand> & Addr,const TargetRegisterClass * RC,MachineInstr::mmo_iterator MMOBegin,MachineInstr::mmo_iterator MMOEnd,SmallVectorImpl<MachineInstr * > & NewMIs) const4809 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
4810                                  SmallVectorImpl<MachineOperand> &Addr,
4811                                  const TargetRegisterClass *RC,
4812                                  MachineInstr::mmo_iterator MMOBegin,
4813                                  MachineInstr::mmo_iterator MMOEnd,
4814                                  SmallVectorImpl<MachineInstr*> &NewMIs) const {
4815   unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
4816   bool isAligned = MMOBegin != MMOEnd &&
4817                    (*MMOBegin)->getAlignment() >= Alignment;
4818   unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
4819   DebugLoc DL;
4820   MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
4821   for (unsigned i = 0, e = Addr.size(); i != e; ++i)
4822     MIB.addOperand(Addr[i]);
4823   (*MIB).setMemRefs(MMOBegin, MMOEnd);
4824   NewMIs.push_back(MIB);
4825 }
4826 
analyzeCompare(const MachineInstr & MI,unsigned & SrcReg,unsigned & SrcReg2,int & CmpMask,int & CmpValue) const4827 bool X86InstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
4828                                   unsigned &SrcReg2, int &CmpMask,
4829                                   int &CmpValue) const {
4830   switch (MI.getOpcode()) {
4831   default: break;
4832   case X86::CMP64ri32:
4833   case X86::CMP64ri8:
4834   case X86::CMP32ri:
4835   case X86::CMP32ri8:
4836   case X86::CMP16ri:
4837   case X86::CMP16ri8:
4838   case X86::CMP8ri:
4839     SrcReg = MI.getOperand(0).getReg();
4840     SrcReg2 = 0;
4841     CmpMask = ~0;
4842     CmpValue = MI.getOperand(1).getImm();
4843     return true;
4844   // A SUB can be used to perform comparison.
4845   case X86::SUB64rm:
4846   case X86::SUB32rm:
4847   case X86::SUB16rm:
4848   case X86::SUB8rm:
4849     SrcReg = MI.getOperand(1).getReg();
4850     SrcReg2 = 0;
4851     CmpMask = ~0;
4852     CmpValue = 0;
4853     return true;
4854   case X86::SUB64rr:
4855   case X86::SUB32rr:
4856   case X86::SUB16rr:
4857   case X86::SUB8rr:
4858     SrcReg = MI.getOperand(1).getReg();
4859     SrcReg2 = MI.getOperand(2).getReg();
4860     CmpMask = ~0;
4861     CmpValue = 0;
4862     return true;
4863   case X86::SUB64ri32:
4864   case X86::SUB64ri8:
4865   case X86::SUB32ri:
4866   case X86::SUB32ri8:
4867   case X86::SUB16ri:
4868   case X86::SUB16ri8:
4869   case X86::SUB8ri:
4870     SrcReg = MI.getOperand(1).getReg();
4871     SrcReg2 = 0;
4872     CmpMask = ~0;
4873     CmpValue = MI.getOperand(2).getImm();
4874     return true;
4875   case X86::CMP64rr:
4876   case X86::CMP32rr:
4877   case X86::CMP16rr:
4878   case X86::CMP8rr:
4879     SrcReg = MI.getOperand(0).getReg();
4880     SrcReg2 = MI.getOperand(1).getReg();
4881     CmpMask = ~0;
4882     CmpValue = 0;
4883     return true;
4884   case X86::TEST8rr:
4885   case X86::TEST16rr:
4886   case X86::TEST32rr:
4887   case X86::TEST64rr:
4888     SrcReg = MI.getOperand(0).getReg();
4889     if (MI.getOperand(1).getReg() != SrcReg)
4890       return false;
4891     // Compare against zero.
4892     SrcReg2 = 0;
4893     CmpMask = ~0;
4894     CmpValue = 0;
4895     return true;
4896   }
4897   return false;
4898 }
4899 
4900 /// Check whether the first instruction, whose only
4901 /// purpose is to update flags, can be made redundant.
4902 /// CMPrr can be made redundant by SUBrr if the operands are the same.
4903 /// This function can be extended later on.
4904 /// SrcReg, SrcRegs: register operands for FlagI.
4905 /// ImmValue: immediate for FlagI if it takes an immediate.
isRedundantFlagInstr(MachineInstr & FlagI,unsigned SrcReg,unsigned SrcReg2,int ImmValue,MachineInstr & OI)4906 inline static bool isRedundantFlagInstr(MachineInstr &FlagI, unsigned SrcReg,
4907                                         unsigned SrcReg2, int ImmValue,
4908                                         MachineInstr &OI) {
4909   if (((FlagI.getOpcode() == X86::CMP64rr && OI.getOpcode() == X86::SUB64rr) ||
4910        (FlagI.getOpcode() == X86::CMP32rr && OI.getOpcode() == X86::SUB32rr) ||
4911        (FlagI.getOpcode() == X86::CMP16rr && OI.getOpcode() == X86::SUB16rr) ||
4912        (FlagI.getOpcode() == X86::CMP8rr && OI.getOpcode() == X86::SUB8rr)) &&
4913       ((OI.getOperand(1).getReg() == SrcReg &&
4914         OI.getOperand(2).getReg() == SrcReg2) ||
4915        (OI.getOperand(1).getReg() == SrcReg2 &&
4916         OI.getOperand(2).getReg() == SrcReg)))
4917     return true;
4918 
4919   if (((FlagI.getOpcode() == X86::CMP64ri32 &&
4920         OI.getOpcode() == X86::SUB64ri32) ||
4921        (FlagI.getOpcode() == X86::CMP64ri8 &&
4922         OI.getOpcode() == X86::SUB64ri8) ||
4923        (FlagI.getOpcode() == X86::CMP32ri && OI.getOpcode() == X86::SUB32ri) ||
4924        (FlagI.getOpcode() == X86::CMP32ri8 &&
4925         OI.getOpcode() == X86::SUB32ri8) ||
4926        (FlagI.getOpcode() == X86::CMP16ri && OI.getOpcode() == X86::SUB16ri) ||
4927        (FlagI.getOpcode() == X86::CMP16ri8 &&
4928         OI.getOpcode() == X86::SUB16ri8) ||
4929        (FlagI.getOpcode() == X86::CMP8ri && OI.getOpcode() == X86::SUB8ri)) &&
4930       OI.getOperand(1).getReg() == SrcReg &&
4931       OI.getOperand(2).getImm() == ImmValue)
4932     return true;
4933   return false;
4934 }
4935 
4936 /// Check whether the definition can be converted
4937 /// to remove a comparison against zero.
isDefConvertible(MachineInstr & MI)4938 inline static bool isDefConvertible(MachineInstr &MI) {
4939   switch (MI.getOpcode()) {
4940   default: return false;
4941 
4942   // The shift instructions only modify ZF if their shift count is non-zero.
4943   // N.B.: The processor truncates the shift count depending on the encoding.
4944   case X86::SAR8ri:    case X86::SAR16ri:  case X86::SAR32ri:case X86::SAR64ri:
4945   case X86::SHR8ri:    case X86::SHR16ri:  case X86::SHR32ri:case X86::SHR64ri:
4946      return getTruncatedShiftCount(MI, 2) != 0;
4947 
4948   // Some left shift instructions can be turned into LEA instructions but only
4949   // if their flags aren't used. Avoid transforming such instructions.
4950   case X86::SHL8ri:    case X86::SHL16ri:  case X86::SHL32ri:case X86::SHL64ri:{
4951     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
4952     if (isTruncatedShiftCountForLEA(ShAmt)) return false;
4953     return ShAmt != 0;
4954   }
4955 
4956   case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
4957   case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
4958      return getTruncatedShiftCount(MI, 3) != 0;
4959 
4960   case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
4961   case X86::SUB32ri8:  case X86::SUB16ri:  case X86::SUB16ri8:
4962   case X86::SUB8ri:    case X86::SUB64rr:  case X86::SUB32rr:
4963   case X86::SUB16rr:   case X86::SUB8rr:   case X86::SUB64rm:
4964   case X86::SUB32rm:   case X86::SUB16rm:  case X86::SUB8rm:
4965   case X86::DEC64r:    case X86::DEC32r:   case X86::DEC16r: case X86::DEC8r:
4966   case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
4967   case X86::ADD32ri8:  case X86::ADD16ri:  case X86::ADD16ri8:
4968   case X86::ADD8ri:    case X86::ADD64rr:  case X86::ADD32rr:
4969   case X86::ADD16rr:   case X86::ADD8rr:   case X86::ADD64rm:
4970   case X86::ADD32rm:   case X86::ADD16rm:  case X86::ADD8rm:
4971   case X86::INC64r:    case X86::INC32r:   case X86::INC16r: case X86::INC8r:
4972   case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
4973   case X86::AND32ri8:  case X86::AND16ri:  case X86::AND16ri8:
4974   case X86::AND8ri:    case X86::AND64rr:  case X86::AND32rr:
4975   case X86::AND16rr:   case X86::AND8rr:   case X86::AND64rm:
4976   case X86::AND32rm:   case X86::AND16rm:  case X86::AND8rm:
4977   case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
4978   case X86::XOR32ri8:  case X86::XOR16ri:  case X86::XOR16ri8:
4979   case X86::XOR8ri:    case X86::XOR64rr:  case X86::XOR32rr:
4980   case X86::XOR16rr:   case X86::XOR8rr:   case X86::XOR64rm:
4981   case X86::XOR32rm:   case X86::XOR16rm:  case X86::XOR8rm:
4982   case X86::OR64ri32:  case X86::OR64ri8:  case X86::OR32ri:
4983   case X86::OR32ri8:   case X86::OR16ri:   case X86::OR16ri8:
4984   case X86::OR8ri:     case X86::OR64rr:   case X86::OR32rr:
4985   case X86::OR16rr:    case X86::OR8rr:    case X86::OR64rm:
4986   case X86::OR32rm:    case X86::OR16rm:   case X86::OR8rm:
4987   case X86::NEG8r:     case X86::NEG16r:   case X86::NEG32r: case X86::NEG64r:
4988   case X86::SAR8r1:    case X86::SAR16r1:  case X86::SAR32r1:case X86::SAR64r1:
4989   case X86::SHR8r1:    case X86::SHR16r1:  case X86::SHR32r1:case X86::SHR64r1:
4990   case X86::SHL8r1:    case X86::SHL16r1:  case X86::SHL32r1:case X86::SHL64r1:
4991   case X86::ADC32ri:   case X86::ADC32ri8:
4992   case X86::ADC32rr:   case X86::ADC64ri32:
4993   case X86::ADC64ri8:  case X86::ADC64rr:
4994   case X86::SBB32ri:   case X86::SBB32ri8:
4995   case X86::SBB32rr:   case X86::SBB64ri32:
4996   case X86::SBB64ri8:  case X86::SBB64rr:
4997   case X86::ANDN32rr:  case X86::ANDN32rm:
4998   case X86::ANDN64rr:  case X86::ANDN64rm:
4999   case X86::BEXTR32rr: case X86::BEXTR64rr:
5000   case X86::BEXTR32rm: case X86::BEXTR64rm:
5001   case X86::BLSI32rr:  case X86::BLSI32rm:
5002   case X86::BLSI64rr:  case X86::BLSI64rm:
5003   case X86::BLSMSK32rr:case X86::BLSMSK32rm:
5004   case X86::BLSMSK64rr:case X86::BLSMSK64rm:
5005   case X86::BLSR32rr:  case X86::BLSR32rm:
5006   case X86::BLSR64rr:  case X86::BLSR64rm:
5007   case X86::BZHI32rr:  case X86::BZHI32rm:
5008   case X86::BZHI64rr:  case X86::BZHI64rm:
5009   case X86::LZCNT16rr: case X86::LZCNT16rm:
5010   case X86::LZCNT32rr: case X86::LZCNT32rm:
5011   case X86::LZCNT64rr: case X86::LZCNT64rm:
5012   case X86::POPCNT16rr:case X86::POPCNT16rm:
5013   case X86::POPCNT32rr:case X86::POPCNT32rm:
5014   case X86::POPCNT64rr:case X86::POPCNT64rm:
5015   case X86::TZCNT16rr: case X86::TZCNT16rm:
5016   case X86::TZCNT32rr: case X86::TZCNT32rm:
5017   case X86::TZCNT64rr: case X86::TZCNT64rm:
5018     return true;
5019   }
5020 }
5021 
5022 /// Check whether the use can be converted to remove a comparison against zero.
isUseDefConvertible(MachineInstr & MI)5023 static X86::CondCode isUseDefConvertible(MachineInstr &MI) {
5024   switch (MI.getOpcode()) {
5025   default: return X86::COND_INVALID;
5026   case X86::LZCNT16rr: case X86::LZCNT16rm:
5027   case X86::LZCNT32rr: case X86::LZCNT32rm:
5028   case X86::LZCNT64rr: case X86::LZCNT64rm:
5029     return X86::COND_B;
5030   case X86::POPCNT16rr:case X86::POPCNT16rm:
5031   case X86::POPCNT32rr:case X86::POPCNT32rm:
5032   case X86::POPCNT64rr:case X86::POPCNT64rm:
5033     return X86::COND_E;
5034   case X86::TZCNT16rr: case X86::TZCNT16rm:
5035   case X86::TZCNT32rr: case X86::TZCNT32rm:
5036   case X86::TZCNT64rr: case X86::TZCNT64rm:
5037     return X86::COND_B;
5038   }
5039 }
5040 
5041 /// Check if there exists an earlier instruction that
5042 /// operates on the same source operands and sets flags in the same way as
5043 /// Compare; remove Compare if possible.
optimizeCompareInstr(MachineInstr & CmpInstr,unsigned SrcReg,unsigned SrcReg2,int CmpMask,int CmpValue,const MachineRegisterInfo * MRI) const5044 bool X86InstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
5045                                         unsigned SrcReg2, int CmpMask,
5046                                         int CmpValue,
5047                                         const MachineRegisterInfo *MRI) const {
5048   // Check whether we can replace SUB with CMP.
5049   unsigned NewOpcode = 0;
5050   switch (CmpInstr.getOpcode()) {
5051   default: break;
5052   case X86::SUB64ri32:
5053   case X86::SUB64ri8:
5054   case X86::SUB32ri:
5055   case X86::SUB32ri8:
5056   case X86::SUB16ri:
5057   case X86::SUB16ri8:
5058   case X86::SUB8ri:
5059   case X86::SUB64rm:
5060   case X86::SUB32rm:
5061   case X86::SUB16rm:
5062   case X86::SUB8rm:
5063   case X86::SUB64rr:
5064   case X86::SUB32rr:
5065   case X86::SUB16rr:
5066   case X86::SUB8rr: {
5067     if (!MRI->use_nodbg_empty(CmpInstr.getOperand(0).getReg()))
5068       return false;
5069     // There is no use of the destination register, we can replace SUB with CMP.
5070     switch (CmpInstr.getOpcode()) {
5071     default: llvm_unreachable("Unreachable!");
5072     case X86::SUB64rm:   NewOpcode = X86::CMP64rm;   break;
5073     case X86::SUB32rm:   NewOpcode = X86::CMP32rm;   break;
5074     case X86::SUB16rm:   NewOpcode = X86::CMP16rm;   break;
5075     case X86::SUB8rm:    NewOpcode = X86::CMP8rm;    break;
5076     case X86::SUB64rr:   NewOpcode = X86::CMP64rr;   break;
5077     case X86::SUB32rr:   NewOpcode = X86::CMP32rr;   break;
5078     case X86::SUB16rr:   NewOpcode = X86::CMP16rr;   break;
5079     case X86::SUB8rr:    NewOpcode = X86::CMP8rr;    break;
5080     case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
5081     case X86::SUB64ri8:  NewOpcode = X86::CMP64ri8;  break;
5082     case X86::SUB32ri:   NewOpcode = X86::CMP32ri;   break;
5083     case X86::SUB32ri8:  NewOpcode = X86::CMP32ri8;  break;
5084     case X86::SUB16ri:   NewOpcode = X86::CMP16ri;   break;
5085     case X86::SUB16ri8:  NewOpcode = X86::CMP16ri8;  break;
5086     case X86::SUB8ri:    NewOpcode = X86::CMP8ri;    break;
5087     }
5088     CmpInstr.setDesc(get(NewOpcode));
5089     CmpInstr.RemoveOperand(0);
5090     // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
5091     if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
5092         NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
5093       return false;
5094   }
5095   }
5096 
5097   // Get the unique definition of SrcReg.
5098   MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
5099   if (!MI) return false;
5100 
5101   // CmpInstr is the first instruction of the BB.
5102   MachineBasicBlock::iterator I = CmpInstr, Def = MI;
5103 
5104   // If we are comparing against zero, check whether we can use MI to update
5105   // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
5106   bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
5107   if (IsCmpZero && MI->getParent() != CmpInstr.getParent())
5108     return false;
5109 
5110   // If we have a use of the source register between the def and our compare
5111   // instruction we can eliminate the compare iff the use sets EFLAGS in the
5112   // right way.
5113   bool ShouldUpdateCC = false;
5114   X86::CondCode NewCC = X86::COND_INVALID;
5115   if (IsCmpZero && !isDefConvertible(*MI)) {
5116     // Scan forward from the use until we hit the use we're looking for or the
5117     // compare instruction.
5118     for (MachineBasicBlock::iterator J = MI;; ++J) {
5119       // Do we have a convertible instruction?
5120       NewCC = isUseDefConvertible(*J);
5121       if (NewCC != X86::COND_INVALID && J->getOperand(1).isReg() &&
5122           J->getOperand(1).getReg() == SrcReg) {
5123         assert(J->definesRegister(X86::EFLAGS) && "Must be an EFLAGS def!");
5124         ShouldUpdateCC = true; // Update CC later on.
5125         // This is not a def of SrcReg, but still a def of EFLAGS. Keep going
5126         // with the new def.
5127         Def = J;
5128         MI = &*Def;
5129         break;
5130       }
5131 
5132       if (J == I)
5133         return false;
5134     }
5135   }
5136 
5137   // We are searching for an earlier instruction that can make CmpInstr
5138   // redundant and that instruction will be saved in Sub.
5139   MachineInstr *Sub = nullptr;
5140   const TargetRegisterInfo *TRI = &getRegisterInfo();
5141 
5142   // We iterate backward, starting from the instruction before CmpInstr and
5143   // stop when reaching the definition of a source register or done with the BB.
5144   // RI points to the instruction before CmpInstr.
5145   // If the definition is in this basic block, RE points to the definition;
5146   // otherwise, RE is the rend of the basic block.
5147   MachineBasicBlock::reverse_iterator
5148       RI = MachineBasicBlock::reverse_iterator(I),
5149       RE = CmpInstr.getParent() == MI->getParent()
5150                ? MachineBasicBlock::reverse_iterator(++Def) /* points to MI */
5151                : CmpInstr.getParent()->rend();
5152   MachineInstr *Movr0Inst = nullptr;
5153   for (; RI != RE; ++RI) {
5154     MachineInstr &Instr = *RI;
5155     // Check whether CmpInstr can be made redundant by the current instruction.
5156     if (!IsCmpZero &&
5157         isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
5158       Sub = &Instr;
5159       break;
5160     }
5161 
5162     if (Instr.modifiesRegister(X86::EFLAGS, TRI) ||
5163         Instr.readsRegister(X86::EFLAGS, TRI)) {
5164       // This instruction modifies or uses EFLAGS.
5165 
5166       // MOV32r0 etc. are implemented with xor which clobbers condition code.
5167       // They are safe to move up, if the definition to EFLAGS is dead and
5168       // earlier instructions do not read or write EFLAGS.
5169       if (!Movr0Inst && Instr.getOpcode() == X86::MOV32r0 &&
5170           Instr.registerDefIsDead(X86::EFLAGS, TRI)) {
5171         Movr0Inst = &Instr;
5172         continue;
5173       }
5174 
5175       // We can't remove CmpInstr.
5176       return false;
5177     }
5178   }
5179 
5180   // Return false if no candidates exist.
5181   if (!IsCmpZero && !Sub)
5182     return false;
5183 
5184   bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
5185                     Sub->getOperand(2).getReg() == SrcReg);
5186 
5187   // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
5188   // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
5189   // If we are done with the basic block, we need to check whether EFLAGS is
5190   // live-out.
5191   bool IsSafe = false;
5192   SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
5193   MachineBasicBlock::iterator E = CmpInstr.getParent()->end();
5194   for (++I; I != E; ++I) {
5195     const MachineInstr &Instr = *I;
5196     bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
5197     bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
5198     // We should check the usage if this instruction uses and updates EFLAGS.
5199     if (!UseEFLAGS && ModifyEFLAGS) {
5200       // It is safe to remove CmpInstr if EFLAGS is updated again.
5201       IsSafe = true;
5202       break;
5203     }
5204     if (!UseEFLAGS && !ModifyEFLAGS)
5205       continue;
5206 
5207     // EFLAGS is used by this instruction.
5208     X86::CondCode OldCC = X86::COND_INVALID;
5209     bool OpcIsSET = false;
5210     if (IsCmpZero || IsSwapped) {
5211       // We decode the condition code from opcode.
5212       if (Instr.isBranch())
5213         OldCC = getCondFromBranchOpc(Instr.getOpcode());
5214       else {
5215         OldCC = getCondFromSETOpc(Instr.getOpcode());
5216         if (OldCC != X86::COND_INVALID)
5217           OpcIsSET = true;
5218         else
5219           OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
5220       }
5221       if (OldCC == X86::COND_INVALID) return false;
5222     }
5223     if (IsCmpZero) {
5224       switch (OldCC) {
5225       default: break;
5226       case X86::COND_A: case X86::COND_AE:
5227       case X86::COND_B: case X86::COND_BE:
5228       case X86::COND_G: case X86::COND_GE:
5229       case X86::COND_L: case X86::COND_LE:
5230       case X86::COND_O: case X86::COND_NO:
5231         // CF and OF are used, we can't perform this optimization.
5232         return false;
5233       }
5234 
5235       // If we're updating the condition code check if we have to reverse the
5236       // condition.
5237       if (ShouldUpdateCC)
5238         switch (OldCC) {
5239         default:
5240           return false;
5241         case X86::COND_E:
5242           break;
5243         case X86::COND_NE:
5244           NewCC = GetOppositeBranchCondition(NewCC);
5245           break;
5246         }
5247     } else if (IsSwapped) {
5248       // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
5249       // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
5250       // We swap the condition code and synthesize the new opcode.
5251       NewCC = getSwappedCondition(OldCC);
5252       if (NewCC == X86::COND_INVALID) return false;
5253     }
5254 
5255     if ((ShouldUpdateCC || IsSwapped) && NewCC != OldCC) {
5256       // Synthesize the new opcode.
5257       bool HasMemoryOperand = Instr.hasOneMemOperand();
5258       unsigned NewOpc;
5259       if (Instr.isBranch())
5260         NewOpc = GetCondBranchFromCond(NewCC);
5261       else if(OpcIsSET)
5262         NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
5263       else {
5264         unsigned DstReg = Instr.getOperand(0).getReg();
5265         NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
5266                                  HasMemoryOperand);
5267       }
5268 
5269       // Push the MachineInstr to OpsToUpdate.
5270       // If it is safe to remove CmpInstr, the condition code of these
5271       // instructions will be modified.
5272       OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
5273     }
5274     if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
5275       // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
5276       IsSafe = true;
5277       break;
5278     }
5279   }
5280 
5281   // If EFLAGS is not killed nor re-defined, we should check whether it is
5282   // live-out. If it is live-out, do not optimize.
5283   if ((IsCmpZero || IsSwapped) && !IsSafe) {
5284     MachineBasicBlock *MBB = CmpInstr.getParent();
5285     for (MachineBasicBlock *Successor : MBB->successors())
5286       if (Successor->isLiveIn(X86::EFLAGS))
5287         return false;
5288   }
5289 
5290   // The instruction to be updated is either Sub or MI.
5291   Sub = IsCmpZero ? MI : Sub;
5292   // Move Movr0Inst to the appropriate place before Sub.
5293   if (Movr0Inst) {
5294     // Look backwards until we find a def that doesn't use the current EFLAGS.
5295     Def = Sub;
5296     MachineBasicBlock::reverse_iterator
5297       InsertI = MachineBasicBlock::reverse_iterator(++Def),
5298                 InsertE = Sub->getParent()->rend();
5299     for (; InsertI != InsertE; ++InsertI) {
5300       MachineInstr *Instr = &*InsertI;
5301       if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
5302           Instr->modifiesRegister(X86::EFLAGS, TRI)) {
5303         Sub->getParent()->remove(Movr0Inst);
5304         Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
5305                                    Movr0Inst);
5306         break;
5307       }
5308     }
5309     if (InsertI == InsertE)
5310       return false;
5311   }
5312 
5313   // Make sure Sub instruction defines EFLAGS and mark the def live.
5314   unsigned i = 0, e = Sub->getNumOperands();
5315   for (; i != e; ++i) {
5316     MachineOperand &MO = Sub->getOperand(i);
5317     if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
5318       MO.setIsDead(false);
5319       break;
5320     }
5321   }
5322   assert(i != e && "Unable to locate a def EFLAGS operand");
5323 
5324   CmpInstr.eraseFromParent();
5325 
5326   // Modify the condition code of instructions in OpsToUpdate.
5327   for (auto &Op : OpsToUpdate)
5328     Op.first->setDesc(get(Op.second));
5329   return true;
5330 }
5331 
5332 /// Try to remove the load by folding it to a register
5333 /// operand at the use. We fold the load instructions if load defines a virtual
5334 /// register, the virtual register is used once in the same BB, and the
5335 /// instructions in-between do not load or store, and have no side effects.
optimizeLoadInstr(MachineInstr & MI,const MachineRegisterInfo * MRI,unsigned & FoldAsLoadDefReg,MachineInstr * & DefMI) const5336 MachineInstr *X86InstrInfo::optimizeLoadInstr(MachineInstr &MI,
5337                                               const MachineRegisterInfo *MRI,
5338                                               unsigned &FoldAsLoadDefReg,
5339                                               MachineInstr *&DefMI) const {
5340   if (FoldAsLoadDefReg == 0)
5341     return nullptr;
5342   // To be conservative, if there exists another load, clear the load candidate.
5343   if (MI.mayLoad()) {
5344     FoldAsLoadDefReg = 0;
5345     return nullptr;
5346   }
5347 
5348   // Check whether we can move DefMI here.
5349   DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
5350   assert(DefMI);
5351   bool SawStore = false;
5352   if (!DefMI->isSafeToMove(nullptr, SawStore))
5353     return nullptr;
5354 
5355   // Collect information about virtual register operands of MI.
5356   unsigned SrcOperandId = 0;
5357   bool FoundSrcOperand = false;
5358   for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) {
5359     MachineOperand &MO = MI.getOperand(i);
5360     if (!MO.isReg())
5361       continue;
5362     unsigned Reg = MO.getReg();
5363     if (Reg != FoldAsLoadDefReg)
5364       continue;
5365     // Do not fold if we have a subreg use or a def or multiple uses.
5366     if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
5367       return nullptr;
5368 
5369     SrcOperandId = i;
5370     FoundSrcOperand = true;
5371   }
5372   if (!FoundSrcOperand)
5373     return nullptr;
5374 
5375   // Check whether we can fold the def into SrcOperandId.
5376   if (MachineInstr *FoldMI = foldMemoryOperand(MI, SrcOperandId, *DefMI)) {
5377     FoldAsLoadDefReg = 0;
5378     return FoldMI;
5379   }
5380 
5381   return nullptr;
5382 }
5383 
5384 /// Expand a single-def pseudo instruction to a two-addr
5385 /// instruction with two undef reads of the register being defined.
5386 /// This is used for mapping:
5387 ///   %xmm4 = V_SET0
5388 /// to:
5389 ///   %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
5390 ///
Expand2AddrUndef(MachineInstrBuilder & MIB,const MCInstrDesc & Desc)5391 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
5392                              const MCInstrDesc &Desc) {
5393   assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
5394   unsigned Reg = MIB->getOperand(0).getReg();
5395   MIB->setDesc(Desc);
5396 
5397   // MachineInstr::addOperand() will insert explicit operands before any
5398   // implicit operands.
5399   MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
5400   // But we don't trust that.
5401   assert(MIB->getOperand(1).getReg() == Reg &&
5402          MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
5403   return true;
5404 }
5405 
5406 /// Expand a single-def pseudo instruction to a two-addr
5407 /// instruction with two %k0 reads.
5408 /// This is used for mapping:
5409 ///   %k4 = K_SET1
5410 /// to:
5411 ///   %k4 = KXNORrr %k0, %k0
Expand2AddrKreg(MachineInstrBuilder & MIB,const MCInstrDesc & Desc,unsigned Reg)5412 static bool Expand2AddrKreg(MachineInstrBuilder &MIB,
5413                             const MCInstrDesc &Desc, unsigned Reg) {
5414   assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
5415   MIB->setDesc(Desc);
5416   MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
5417   return true;
5418 }
5419 
expandMOV32r1(MachineInstrBuilder & MIB,const TargetInstrInfo & TII,bool MinusOne)5420 static bool expandMOV32r1(MachineInstrBuilder &MIB, const TargetInstrInfo &TII,
5421                           bool MinusOne) {
5422   MachineBasicBlock &MBB = *MIB->getParent();
5423   DebugLoc DL = MIB->getDebugLoc();
5424   unsigned Reg = MIB->getOperand(0).getReg();
5425 
5426   // Insert the XOR.
5427   BuildMI(MBB, MIB.getInstr(), DL, TII.get(X86::XOR32rr), Reg)
5428       .addReg(Reg, RegState::Undef)
5429       .addReg(Reg, RegState::Undef);
5430 
5431   // Turn the pseudo into an INC or DEC.
5432   MIB->setDesc(TII.get(MinusOne ? X86::DEC32r : X86::INC32r));
5433   MIB.addReg(Reg);
5434 
5435   return true;
5436 }
5437 
ExpandMOVImmSExti8(MachineInstrBuilder & MIB) const5438 bool X86InstrInfo::ExpandMOVImmSExti8(MachineInstrBuilder &MIB) const {
5439   MachineBasicBlock &MBB = *MIB->getParent();
5440   DebugLoc DL = MIB->getDebugLoc();
5441   int64_t Imm = MIB->getOperand(1).getImm();
5442   assert(Imm != 0 && "Using push/pop for 0 is not efficient.");
5443   MachineBasicBlock::iterator I = MIB.getInstr();
5444 
5445   int StackAdjustment;
5446 
5447   if (Subtarget.is64Bit()) {
5448     assert(MIB->getOpcode() == X86::MOV64ImmSExti8 ||
5449            MIB->getOpcode() == X86::MOV32ImmSExti8);
5450 
5451     // Can't use push/pop lowering if the function might write to the red zone.
5452     X86MachineFunctionInfo *X86FI =
5453         MBB.getParent()->getInfo<X86MachineFunctionInfo>();
5454     if (X86FI->getUsesRedZone()) {
5455       MIB->setDesc(get(MIB->getOpcode() == X86::MOV32ImmSExti8 ? X86::MOV32ri
5456                                                                : X86::MOV64ri));
5457       return true;
5458     }
5459 
5460     // 64-bit mode doesn't have 32-bit push/pop, so use 64-bit operations and
5461     // widen the register if necessary.
5462     StackAdjustment = 8;
5463     BuildMI(MBB, I, DL, get(X86::PUSH64i8)).addImm(Imm);
5464     MIB->setDesc(get(X86::POP64r));
5465     MIB->getOperand(0)
5466         .setReg(getX86SubSuperRegister(MIB->getOperand(0).getReg(), 64));
5467   } else {
5468     assert(MIB->getOpcode() == X86::MOV32ImmSExti8);
5469     StackAdjustment = 4;
5470     BuildMI(MBB, I, DL, get(X86::PUSH32i8)).addImm(Imm);
5471     MIB->setDesc(get(X86::POP32r));
5472   }
5473 
5474   // Build CFI if necessary.
5475   MachineFunction &MF = *MBB.getParent();
5476   const X86FrameLowering *TFL = Subtarget.getFrameLowering();
5477   bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
5478   bool NeedsDwarfCFI =
5479       !IsWin64Prologue &&
5480       (MF.getMMI().hasDebugInfo() || MF.getFunction()->needsUnwindTableEntry());
5481   bool EmitCFI = !TFL->hasFP(MF) && NeedsDwarfCFI;
5482   if (EmitCFI) {
5483     TFL->BuildCFI(MBB, I, DL,
5484         MCCFIInstruction::createAdjustCfaOffset(nullptr, StackAdjustment));
5485     TFL->BuildCFI(MBB, std::next(I), DL,
5486         MCCFIInstruction::createAdjustCfaOffset(nullptr, -StackAdjustment));
5487   }
5488 
5489   return true;
5490 }
5491 
5492 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different
5493 // code sequence is needed for other targets.
expandLoadStackGuard(MachineInstrBuilder & MIB,const TargetInstrInfo & TII)5494 static void expandLoadStackGuard(MachineInstrBuilder &MIB,
5495                                  const TargetInstrInfo &TII) {
5496   MachineBasicBlock &MBB = *MIB->getParent();
5497   DebugLoc DL = MIB->getDebugLoc();
5498   unsigned Reg = MIB->getOperand(0).getReg();
5499   const GlobalValue *GV =
5500       cast<GlobalValue>((*MIB->memoperands_begin())->getValue());
5501   unsigned Flag = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant;
5502   MachineMemOperand *MMO = MBB.getParent()->getMachineMemOperand(
5503       MachinePointerInfo::getGOT(*MBB.getParent()), Flag, 8, 8);
5504   MachineBasicBlock::iterator I = MIB.getInstr();
5505 
5506   BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1)
5507       .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0)
5508       .addMemOperand(MMO);
5509   MIB->setDebugLoc(DL);
5510   MIB->setDesc(TII.get(X86::MOV64rm));
5511   MIB.addReg(Reg, RegState::Kill).addImm(1).addReg(0).addImm(0).addReg(0);
5512 }
5513 
expandPostRAPseudo(MachineInstr & MI) const5514 bool X86InstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
5515   bool HasAVX = Subtarget.hasAVX();
5516   MachineInstrBuilder MIB(*MI.getParent()->getParent(), MI);
5517   switch (MI.getOpcode()) {
5518   case X86::MOV32r0:
5519     return Expand2AddrUndef(MIB, get(X86::XOR32rr));
5520   case X86::MOV32r1:
5521     return expandMOV32r1(MIB, *this, /*MinusOne=*/ false);
5522   case X86::MOV32r_1:
5523     return expandMOV32r1(MIB, *this, /*MinusOne=*/ true);
5524   case X86::MOV32ImmSExti8:
5525   case X86::MOV64ImmSExti8:
5526     return ExpandMOVImmSExti8(MIB);
5527   case X86::SETB_C8r:
5528     return Expand2AddrUndef(MIB, get(X86::SBB8rr));
5529   case X86::SETB_C16r:
5530     return Expand2AddrUndef(MIB, get(X86::SBB16rr));
5531   case X86::SETB_C32r:
5532     return Expand2AddrUndef(MIB, get(X86::SBB32rr));
5533   case X86::SETB_C64r:
5534     return Expand2AddrUndef(MIB, get(X86::SBB64rr));
5535   case X86::V_SET0:
5536   case X86::FsFLD0SS:
5537   case X86::FsFLD0SD:
5538     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
5539   case X86::AVX_SET0:
5540     assert(HasAVX && "AVX not supported");
5541     return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
5542   case X86::AVX512_128_SET0:
5543     return Expand2AddrUndef(MIB, get(X86::VPXORDZ128rr));
5544   case X86::AVX512_256_SET0:
5545     return Expand2AddrUndef(MIB, get(X86::VPXORDZ256rr));
5546   case X86::AVX512_512_SET0:
5547     return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
5548   case X86::V_SETALLONES:
5549     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
5550   case X86::AVX2_SETALLONES:
5551     return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
5552   case X86::AVX512_512_SETALLONES: {
5553     unsigned Reg = MIB->getOperand(0).getReg();
5554     MIB->setDesc(get(X86::VPTERNLOGDZrri));
5555     // VPTERNLOGD needs 3 register inputs and an immediate.
5556     // 0xff will return 1s for any input.
5557     MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef)
5558        .addReg(Reg, RegState::Undef).addImm(0xff);
5559     return true;
5560   }
5561   case X86::TEST8ri_NOREX:
5562     MI.setDesc(get(X86::TEST8ri));
5563     return true;
5564   case X86::MOV32ri64:
5565     MI.setDesc(get(X86::MOV32ri));
5566     return true;
5567 
5568   // KNL does not recognize dependency-breaking idioms for mask registers,
5569   // so kxnor %k1, %k1, %k2 has a RAW dependence on %k1.
5570   // Using %k0 as the undef input register is a performance heuristic based
5571   // on the assumption that %k0 is used less frequently than the other mask
5572   // registers, since it is not usable as a write mask.
5573   // FIXME: A more advanced approach would be to choose the best input mask
5574   // register based on context.
5575   case X86::KSET0B:
5576   case X86::KSET0W: return Expand2AddrKreg(MIB, get(X86::KXORWrr), X86::K0);
5577   case X86::KSET0D: return Expand2AddrKreg(MIB, get(X86::KXORDrr), X86::K0);
5578   case X86::KSET0Q: return Expand2AddrKreg(MIB, get(X86::KXORQrr), X86::K0);
5579   case X86::KSET1B:
5580   case X86::KSET1W: return Expand2AddrKreg(MIB, get(X86::KXNORWrr), X86::K0);
5581   case X86::KSET1D: return Expand2AddrKreg(MIB, get(X86::KXNORDrr), X86::K0);
5582   case X86::KSET1Q: return Expand2AddrKreg(MIB, get(X86::KXNORQrr), X86::K0);
5583   case TargetOpcode::LOAD_STACK_GUARD:
5584     expandLoadStackGuard(MIB, *this);
5585     return true;
5586   }
5587   return false;
5588 }
5589 
addOperands(MachineInstrBuilder & MIB,ArrayRef<MachineOperand> MOs,int PtrOffset=0)5590 static void addOperands(MachineInstrBuilder &MIB, ArrayRef<MachineOperand> MOs,
5591                         int PtrOffset = 0) {
5592   unsigned NumAddrOps = MOs.size();
5593 
5594   if (NumAddrOps < 4) {
5595     // FrameIndex only - add an immediate offset (whether its zero or not).
5596     for (unsigned i = 0; i != NumAddrOps; ++i)
5597       MIB.addOperand(MOs[i]);
5598     addOffset(MIB, PtrOffset);
5599   } else {
5600     // General Memory Addressing - we need to add any offset to an existing
5601     // offset.
5602     assert(MOs.size() == 5 && "Unexpected memory operand list length");
5603     for (unsigned i = 0; i != NumAddrOps; ++i) {
5604       const MachineOperand &MO = MOs[i];
5605       if (i == 3 && PtrOffset != 0) {
5606         MIB.addDisp(MO, PtrOffset);
5607       } else {
5608         MIB.addOperand(MO);
5609       }
5610     }
5611   }
5612 }
5613 
FuseTwoAddrInst(MachineFunction & MF,unsigned Opcode,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,MachineInstr & MI,const TargetInstrInfo & TII)5614 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
5615                                      ArrayRef<MachineOperand> MOs,
5616                                      MachineBasicBlock::iterator InsertPt,
5617                                      MachineInstr &MI,
5618                                      const TargetInstrInfo &TII) {
5619   // Create the base instruction with the memory operand as the first part.
5620   // Omit the implicit operands, something BuildMI can't do.
5621   MachineInstr *NewMI =
5622       MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true);
5623   MachineInstrBuilder MIB(MF, NewMI);
5624   addOperands(MIB, MOs);
5625 
5626   // Loop over the rest of the ri operands, converting them over.
5627   unsigned NumOps = MI.getDesc().getNumOperands() - 2;
5628   for (unsigned i = 0; i != NumOps; ++i) {
5629     MachineOperand &MO = MI.getOperand(i + 2);
5630     MIB.addOperand(MO);
5631   }
5632   for (unsigned i = NumOps + 2, e = MI.getNumOperands(); i != e; ++i) {
5633     MachineOperand &MO = MI.getOperand(i);
5634     MIB.addOperand(MO);
5635   }
5636 
5637   MachineBasicBlock *MBB = InsertPt->getParent();
5638   MBB->insert(InsertPt, NewMI);
5639 
5640   return MIB;
5641 }
5642 
FuseInst(MachineFunction & MF,unsigned Opcode,unsigned OpNo,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,MachineInstr & MI,const TargetInstrInfo & TII,int PtrOffset=0)5643 static MachineInstr *FuseInst(MachineFunction &MF, unsigned Opcode,
5644                               unsigned OpNo, ArrayRef<MachineOperand> MOs,
5645                               MachineBasicBlock::iterator InsertPt,
5646                               MachineInstr &MI, const TargetInstrInfo &TII,
5647                               int PtrOffset = 0) {
5648   // Omit the implicit operands, something BuildMI can't do.
5649   MachineInstr *NewMI =
5650       MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true);
5651   MachineInstrBuilder MIB(MF, NewMI);
5652 
5653   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
5654     MachineOperand &MO = MI.getOperand(i);
5655     if (i == OpNo) {
5656       assert(MO.isReg() && "Expected to fold into reg operand!");
5657       addOperands(MIB, MOs, PtrOffset);
5658     } else {
5659       MIB.addOperand(MO);
5660     }
5661   }
5662 
5663   MachineBasicBlock *MBB = InsertPt->getParent();
5664   MBB->insert(InsertPt, NewMI);
5665 
5666   return MIB;
5667 }
5668 
MakeM0Inst(const TargetInstrInfo & TII,unsigned Opcode,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,MachineInstr & MI)5669 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
5670                                 ArrayRef<MachineOperand> MOs,
5671                                 MachineBasicBlock::iterator InsertPt,
5672                                 MachineInstr &MI) {
5673   MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt,
5674                                     MI.getDebugLoc(), TII.get(Opcode));
5675   addOperands(MIB, MOs);
5676   return MIB.addImm(0);
5677 }
5678 
foldMemoryOperandCustom(MachineFunction & MF,MachineInstr & MI,unsigned OpNum,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,unsigned Size,unsigned Align) const5679 MachineInstr *X86InstrInfo::foldMemoryOperandCustom(
5680     MachineFunction &MF, MachineInstr &MI, unsigned OpNum,
5681     ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
5682     unsigned Size, unsigned Align) const {
5683   switch (MI.getOpcode()) {
5684   case X86::INSERTPSrr:
5685   case X86::VINSERTPSrr:
5686     // Attempt to convert the load of inserted vector into a fold load
5687     // of a single float.
5688     if (OpNum == 2) {
5689       unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm();
5690       unsigned ZMask = Imm & 15;
5691       unsigned DstIdx = (Imm >> 4) & 3;
5692       unsigned SrcIdx = (Imm >> 6) & 3;
5693 
5694       unsigned RCSize = getRegClass(MI.getDesc(), OpNum, &RI, MF)->getSize();
5695       if (Size <= RCSize && 4 <= Align) {
5696         int PtrOffset = SrcIdx * 4;
5697         unsigned NewImm = (DstIdx << 4) | ZMask;
5698         unsigned NewOpCode =
5699             (MI.getOpcode() == X86::VINSERTPSrr ? X86::VINSERTPSrm
5700                                                 : X86::INSERTPSrm);
5701         MachineInstr *NewMI =
5702             FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, PtrOffset);
5703         NewMI->getOperand(NewMI->getNumOperands() - 1).setImm(NewImm);
5704         return NewMI;
5705       }
5706     }
5707     break;
5708   case X86::MOVHLPSrr:
5709   case X86::VMOVHLPSrr:
5710     // Move the upper 64-bits of the second operand to the lower 64-bits.
5711     // To fold the load, adjust the pointer to the upper and use (V)MOVLPS.
5712     // TODO: In most cases AVX doesn't have a 8-byte alignment requirement.
5713     if (OpNum == 2) {
5714       unsigned RCSize = getRegClass(MI.getDesc(), OpNum, &RI, MF)->getSize();
5715       if (Size <= RCSize && 8 <= Align) {
5716         unsigned NewOpCode =
5717             (MI.getOpcode() == X86::VMOVHLPSrr ? X86::VMOVLPSrm
5718                                                : X86::MOVLPSrm);
5719         MachineInstr *NewMI =
5720             FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, 8);
5721         return NewMI;
5722       }
5723     }
5724     break;
5725   };
5726 
5727   return nullptr;
5728 }
5729 
foldMemoryOperandImpl(MachineFunction & MF,MachineInstr & MI,unsigned OpNum,ArrayRef<MachineOperand> MOs,MachineBasicBlock::iterator InsertPt,unsigned Size,unsigned Align,bool AllowCommute) const5730 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
5731     MachineFunction &MF, MachineInstr &MI, unsigned OpNum,
5732     ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
5733     unsigned Size, unsigned Align, bool AllowCommute) const {
5734   const DenseMap<unsigned,
5735                  std::pair<uint16_t, uint16_t> > *OpcodeTablePtr = nullptr;
5736   bool isCallRegIndirect = Subtarget.callRegIndirect();
5737   bool isTwoAddrFold = false;
5738 
5739   // For CPUs that favor the register form of a call or push,
5740   // do not fold loads into calls or pushes, unless optimizing for size
5741   // aggressively.
5742   if (isCallRegIndirect && !MF.getFunction()->optForMinSize() &&
5743       (MI.getOpcode() == X86::CALL32r || MI.getOpcode() == X86::CALL64r ||
5744        MI.getOpcode() == X86::PUSH16r || MI.getOpcode() == X86::PUSH32r ||
5745        MI.getOpcode() == X86::PUSH64r))
5746     return nullptr;
5747 
5748   unsigned NumOps = MI.getDesc().getNumOperands();
5749   bool isTwoAddr =
5750       NumOps > 1 && MI.getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
5751 
5752   // FIXME: AsmPrinter doesn't know how to handle
5753   // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
5754   if (MI.getOpcode() == X86::ADD32ri &&
5755       MI.getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
5756     return nullptr;
5757 
5758   MachineInstr *NewMI = nullptr;
5759 
5760   // Attempt to fold any custom cases we have.
5761   if (MachineInstr *CustomMI =
5762           foldMemoryOperandCustom(MF, MI, OpNum, MOs, InsertPt, Size, Align))
5763     return CustomMI;
5764 
5765   // Folding a memory location into the two-address part of a two-address
5766   // instruction is different than folding it other places.  It requires
5767   // replacing the *two* registers with the memory location.
5768   if (isTwoAddr && NumOps >= 2 && OpNum < 2 && MI.getOperand(0).isReg() &&
5769       MI.getOperand(1).isReg() &&
5770       MI.getOperand(0).getReg() == MI.getOperand(1).getReg()) {
5771     OpcodeTablePtr = &RegOp2MemOpTable2Addr;
5772     isTwoAddrFold = true;
5773   } else if (OpNum == 0) {
5774     if (MI.getOpcode() == X86::MOV32r0) {
5775       NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, InsertPt, MI);
5776       if (NewMI)
5777         return NewMI;
5778     }
5779 
5780     OpcodeTablePtr = &RegOp2MemOpTable0;
5781   } else if (OpNum == 1) {
5782     OpcodeTablePtr = &RegOp2MemOpTable1;
5783   } else if (OpNum == 2) {
5784     OpcodeTablePtr = &RegOp2MemOpTable2;
5785   } else if (OpNum == 3) {
5786     OpcodeTablePtr = &RegOp2MemOpTable3;
5787   } else if (OpNum == 4) {
5788     OpcodeTablePtr = &RegOp2MemOpTable4;
5789   }
5790 
5791   // If table selected...
5792   if (OpcodeTablePtr) {
5793     // Find the Opcode to fuse
5794     auto I = OpcodeTablePtr->find(MI.getOpcode());
5795     if (I != OpcodeTablePtr->end()) {
5796       unsigned Opcode = I->second.first;
5797       unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
5798       if (Align < MinAlign)
5799         return nullptr;
5800       bool NarrowToMOV32rm = false;
5801       if (Size) {
5802         unsigned RCSize = getRegClass(MI.getDesc(), OpNum, &RI, MF)->getSize();
5803         if (Size < RCSize) {
5804           // Check if it's safe to fold the load. If the size of the object is
5805           // narrower than the load width, then it's not.
5806           if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
5807             return nullptr;
5808           // If this is a 64-bit load, but the spill slot is 32, then we can do
5809           // a 32-bit load which is implicitly zero-extended. This likely is
5810           // due to live interval analysis remat'ing a load from stack slot.
5811           if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
5812             return nullptr;
5813           Opcode = X86::MOV32rm;
5814           NarrowToMOV32rm = true;
5815         }
5816       }
5817 
5818       if (isTwoAddrFold)
5819         NewMI = FuseTwoAddrInst(MF, Opcode, MOs, InsertPt, MI, *this);
5820       else
5821         NewMI = FuseInst(MF, Opcode, OpNum, MOs, InsertPt, MI, *this);
5822 
5823       if (NarrowToMOV32rm) {
5824         // If this is the special case where we use a MOV32rm to load a 32-bit
5825         // value and zero-extend the top bits. Change the destination register
5826         // to a 32-bit one.
5827         unsigned DstReg = NewMI->getOperand(0).getReg();
5828         if (TargetRegisterInfo::isPhysicalRegister(DstReg))
5829           NewMI->getOperand(0).setReg(RI.getSubReg(DstReg, X86::sub_32bit));
5830         else
5831           NewMI->getOperand(0).setSubReg(X86::sub_32bit);
5832       }
5833       return NewMI;
5834     }
5835   }
5836 
5837   // If the instruction and target operand are commutable, commute the
5838   // instruction and try again.
5839   if (AllowCommute) {
5840     unsigned CommuteOpIdx1 = OpNum, CommuteOpIdx2 = CommuteAnyOperandIndex;
5841     if (findCommutedOpIndices(MI, CommuteOpIdx1, CommuteOpIdx2)) {
5842       bool HasDef = MI.getDesc().getNumDefs();
5843       unsigned Reg0 = HasDef ? MI.getOperand(0).getReg() : 0;
5844       unsigned Reg1 = MI.getOperand(CommuteOpIdx1).getReg();
5845       unsigned Reg2 = MI.getOperand(CommuteOpIdx2).getReg();
5846       bool Tied1 =
5847           0 == MI.getDesc().getOperandConstraint(CommuteOpIdx1, MCOI::TIED_TO);
5848       bool Tied2 =
5849           0 == MI.getDesc().getOperandConstraint(CommuteOpIdx2, MCOI::TIED_TO);
5850 
5851       // If either of the commutable operands are tied to the destination
5852       // then we can not commute + fold.
5853       if ((HasDef && Reg0 == Reg1 && Tied1) ||
5854           (HasDef && Reg0 == Reg2 && Tied2))
5855         return nullptr;
5856 
5857       MachineInstr *CommutedMI =
5858           commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2);
5859       if (!CommutedMI) {
5860         // Unable to commute.
5861         return nullptr;
5862       }
5863       if (CommutedMI != &MI) {
5864         // New instruction. We can't fold from this.
5865         CommutedMI->eraseFromParent();
5866         return nullptr;
5867       }
5868 
5869       // Attempt to fold with the commuted version of the instruction.
5870       NewMI = foldMemoryOperandImpl(MF, MI, CommuteOpIdx2, MOs, InsertPt,
5871                                     Size, Align, /*AllowCommute=*/false);
5872       if (NewMI)
5873         return NewMI;
5874 
5875       // Folding failed again - undo the commute before returning.
5876       MachineInstr *UncommutedMI =
5877           commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2);
5878       if (!UncommutedMI) {
5879         // Unable to commute.
5880         return nullptr;
5881       }
5882       if (UncommutedMI != &MI) {
5883         // New instruction. It doesn't need to be kept.
5884         UncommutedMI->eraseFromParent();
5885         return nullptr;
5886       }
5887 
5888       // Return here to prevent duplicate fuse failure report.
5889       return nullptr;
5890     }
5891   }
5892 
5893   // No fusion
5894   if (PrintFailedFusing && !MI.isCopy())
5895     dbgs() << "We failed to fuse operand " << OpNum << " in " << MI;
5896   return nullptr;
5897 }
5898 
5899 /// Return true for all instructions that only update
5900 /// the first 32 or 64-bits of the destination register and leave the rest
5901 /// unmodified. This can be used to avoid folding loads if the instructions
5902 /// only update part of the destination register, and the non-updated part is
5903 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
5904 /// instructions breaks the partial register dependency and it can improve
5905 /// performance. e.g.:
5906 ///
5907 ///   movss (%rdi), %xmm0
5908 ///   cvtss2sd %xmm0, %xmm0
5909 ///
5910 /// Instead of
5911 ///   cvtss2sd (%rdi), %xmm0
5912 ///
5913 /// FIXME: This should be turned into a TSFlags.
5914 ///
hasPartialRegUpdate(unsigned Opcode)5915 static bool hasPartialRegUpdate(unsigned Opcode) {
5916   switch (Opcode) {
5917   case X86::CVTSI2SSrr:
5918   case X86::CVTSI2SSrm:
5919   case X86::CVTSI2SS64rr:
5920   case X86::CVTSI2SS64rm:
5921   case X86::CVTSI2SDrr:
5922   case X86::CVTSI2SDrm:
5923   case X86::CVTSI2SD64rr:
5924   case X86::CVTSI2SD64rm:
5925   case X86::CVTSD2SSrr:
5926   case X86::CVTSD2SSrm:
5927   case X86::Int_CVTSD2SSrr:
5928   case X86::Int_CVTSD2SSrm:
5929   case X86::CVTSS2SDrr:
5930   case X86::CVTSS2SDrm:
5931   case X86::Int_CVTSS2SDrr:
5932   case X86::Int_CVTSS2SDrm:
5933   case X86::MOVHPDrm:
5934   case X86::MOVHPSrm:
5935   case X86::MOVLPDrm:
5936   case X86::MOVLPSrm:
5937   case X86::RCPSSr:
5938   case X86::RCPSSm:
5939   case X86::RCPSSr_Int:
5940   case X86::RCPSSm_Int:
5941   case X86::ROUNDSDr:
5942   case X86::ROUNDSDm:
5943   case X86::ROUNDSDr_Int:
5944   case X86::ROUNDSSr:
5945   case X86::ROUNDSSm:
5946   case X86::ROUNDSSr_Int:
5947   case X86::RSQRTSSr:
5948   case X86::RSQRTSSm:
5949   case X86::RSQRTSSr_Int:
5950   case X86::RSQRTSSm_Int:
5951   case X86::SQRTSSr:
5952   case X86::SQRTSSm:
5953   case X86::SQRTSSr_Int:
5954   case X86::SQRTSSm_Int:
5955   case X86::SQRTSDr:
5956   case X86::SQRTSDm:
5957   case X86::SQRTSDr_Int:
5958   case X86::SQRTSDm_Int:
5959     return true;
5960   }
5961 
5962   return false;
5963 }
5964 
5965 /// Inform the ExeDepsFix pass how many idle
5966 /// instructions we would like before a partial register update.
getPartialRegUpdateClearance(const MachineInstr & MI,unsigned OpNum,const TargetRegisterInfo * TRI) const5967 unsigned X86InstrInfo::getPartialRegUpdateClearance(
5968     const MachineInstr &MI, unsigned OpNum,
5969     const TargetRegisterInfo *TRI) const {
5970   if (OpNum != 0 || !hasPartialRegUpdate(MI.getOpcode()))
5971     return 0;
5972 
5973   // If MI is marked as reading Reg, the partial register update is wanted.
5974   const MachineOperand &MO = MI.getOperand(0);
5975   unsigned Reg = MO.getReg();
5976   if (TargetRegisterInfo::isVirtualRegister(Reg)) {
5977     if (MO.readsReg() || MI.readsVirtualRegister(Reg))
5978       return 0;
5979   } else {
5980     if (MI.readsRegister(Reg, TRI))
5981       return 0;
5982   }
5983 
5984   // If any instructions in the clearance range are reading Reg, insert a
5985   // dependency breaking instruction, which is inexpensive and is likely to
5986   // be hidden in other instruction's cycles.
5987   return PartialRegUpdateClearance;
5988 }
5989 
5990 // Return true for any instruction the copies the high bits of the first source
5991 // operand into the unused high bits of the destination operand.
hasUndefRegUpdate(unsigned Opcode)5992 static bool hasUndefRegUpdate(unsigned Opcode) {
5993   switch (Opcode) {
5994   case X86::VCVTSI2SSrr:
5995   case X86::VCVTSI2SSrm:
5996   case X86::Int_VCVTSI2SSrr:
5997   case X86::Int_VCVTSI2SSrm:
5998   case X86::VCVTSI2SS64rr:
5999   case X86::VCVTSI2SS64rm:
6000   case X86::Int_VCVTSI2SS64rr:
6001   case X86::Int_VCVTSI2SS64rm:
6002   case X86::VCVTSI2SDrr:
6003   case X86::VCVTSI2SDrm:
6004   case X86::Int_VCVTSI2SDrr:
6005   case X86::Int_VCVTSI2SDrm:
6006   case X86::VCVTSI2SD64rr:
6007   case X86::VCVTSI2SD64rm:
6008   case X86::Int_VCVTSI2SD64rr:
6009   case X86::Int_VCVTSI2SD64rm:
6010   case X86::VCVTSD2SSrr:
6011   case X86::VCVTSD2SSrm:
6012   case X86::Int_VCVTSD2SSrr:
6013   case X86::Int_VCVTSD2SSrm:
6014   case X86::VCVTSS2SDrr:
6015   case X86::VCVTSS2SDrm:
6016   case X86::Int_VCVTSS2SDrr:
6017   case X86::Int_VCVTSS2SDrm:
6018   case X86::VRCPSSr:
6019   case X86::VRCPSSm:
6020   case X86::VRCPSSm_Int:
6021   case X86::VROUNDSDr:
6022   case X86::VROUNDSDm:
6023   case X86::VROUNDSDr_Int:
6024   case X86::VROUNDSSr:
6025   case X86::VROUNDSSm:
6026   case X86::VROUNDSSr_Int:
6027   case X86::VRSQRTSSr:
6028   case X86::VRSQRTSSm:
6029   case X86::VRSQRTSSm_Int:
6030   case X86::VSQRTSSr:
6031   case X86::VSQRTSSm:
6032   case X86::VSQRTSSm_Int:
6033   case X86::VSQRTSDr:
6034   case X86::VSQRTSDm:
6035   case X86::VSQRTSDm_Int:
6036     // AVX-512
6037   case X86::VCVTSD2SSZrr:
6038   case X86::VCVTSD2SSZrm:
6039   case X86::VCVTSS2SDZrr:
6040   case X86::VCVTSS2SDZrm:
6041     return true;
6042   }
6043 
6044   return false;
6045 }
6046 
6047 /// Inform the ExeDepsFix pass how many idle instructions we would like before
6048 /// certain undef register reads.
6049 ///
6050 /// This catches the VCVTSI2SD family of instructions:
6051 ///
6052 /// vcvtsi2sdq %rax, %xmm0<undef>, %xmm14
6053 ///
6054 /// We should to be careful *not* to catch VXOR idioms which are presumably
6055 /// handled specially in the pipeline:
6056 ///
6057 /// vxorps %xmm1<undef>, %xmm1<undef>, %xmm1
6058 ///
6059 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the
6060 /// high bits that are passed-through are not live.
6061 unsigned
getUndefRegClearance(const MachineInstr & MI,unsigned & OpNum,const TargetRegisterInfo * TRI) const6062 X86InstrInfo::getUndefRegClearance(const MachineInstr &MI, unsigned &OpNum,
6063                                    const TargetRegisterInfo *TRI) const {
6064   if (!hasUndefRegUpdate(MI.getOpcode()))
6065     return 0;
6066 
6067   // Set the OpNum parameter to the first source operand.
6068   OpNum = 1;
6069 
6070   const MachineOperand &MO = MI.getOperand(OpNum);
6071   if (MO.isUndef() && TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
6072     return UndefRegClearance;
6073   }
6074   return 0;
6075 }
6076 
breakPartialRegDependency(MachineInstr & MI,unsigned OpNum,const TargetRegisterInfo * TRI) const6077 void X86InstrInfo::breakPartialRegDependency(
6078     MachineInstr &MI, unsigned OpNum, const TargetRegisterInfo *TRI) const {
6079   unsigned Reg = MI.getOperand(OpNum).getReg();
6080   // If MI kills this register, the false dependence is already broken.
6081   if (MI.killsRegister(Reg, TRI))
6082     return;
6083 
6084   if (X86::VR128RegClass.contains(Reg)) {
6085     // These instructions are all floating point domain, so xorps is the best
6086     // choice.
6087     unsigned Opc = Subtarget.hasAVX() ? X86::VXORPSrr : X86::XORPSrr;
6088     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(Opc), Reg)
6089         .addReg(Reg, RegState::Undef)
6090         .addReg(Reg, RegState::Undef);
6091     MI.addRegisterKilled(Reg, TRI, true);
6092   } else if (X86::VR256RegClass.contains(Reg)) {
6093     // Use vxorps to clear the full ymm register.
6094     // It wants to read and write the xmm sub-register.
6095     unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
6096     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::VXORPSrr), XReg)
6097         .addReg(XReg, RegState::Undef)
6098         .addReg(XReg, RegState::Undef)
6099         .addReg(Reg, RegState::ImplicitDefine);
6100     MI.addRegisterKilled(Reg, TRI, true);
6101   }
6102 }
6103 
6104 MachineInstr *
foldMemoryOperandImpl(MachineFunction & MF,MachineInstr & MI,ArrayRef<unsigned> Ops,MachineBasicBlock::iterator InsertPt,int FrameIndex,LiveIntervals * LIS) const6105 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI,
6106                                     ArrayRef<unsigned> Ops,
6107                                     MachineBasicBlock::iterator InsertPt,
6108                                     int FrameIndex, LiveIntervals *LIS) const {
6109   // Check switch flag
6110   if (NoFusing)
6111     return nullptr;
6112 
6113   // Unless optimizing for size, don't fold to avoid partial
6114   // register update stalls
6115   if (!MF.getFunction()->optForSize() && hasPartialRegUpdate(MI.getOpcode()))
6116     return nullptr;
6117 
6118   const MachineFrameInfo *MFI = MF.getFrameInfo();
6119   unsigned Size = MFI->getObjectSize(FrameIndex);
6120   unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
6121   // If the function stack isn't realigned we don't want to fold instructions
6122   // that need increased alignment.
6123   if (!RI.needsStackRealignment(MF))
6124     Alignment =
6125         std::min(Alignment, Subtarget.getFrameLowering()->getStackAlignment());
6126   if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
6127     unsigned NewOpc = 0;
6128     unsigned RCSize = 0;
6129     switch (MI.getOpcode()) {
6130     default: return nullptr;
6131     case X86::TEST8rr:  NewOpc = X86::CMP8ri; RCSize = 1; break;
6132     case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
6133     case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
6134     case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
6135     }
6136     // Check if it's safe to fold the load. If the size of the object is
6137     // narrower than the load width, then it's not.
6138     if (Size < RCSize)
6139       return nullptr;
6140     // Change to CMPXXri r, 0 first.
6141     MI.setDesc(get(NewOpc));
6142     MI.getOperand(1).ChangeToImmediate(0);
6143   } else if (Ops.size() != 1)
6144     return nullptr;
6145 
6146   return foldMemoryOperandImpl(MF, MI, Ops[0],
6147                                MachineOperand::CreateFI(FrameIndex), InsertPt,
6148                                Size, Alignment, /*AllowCommute=*/true);
6149 }
6150 
6151 /// Check if \p LoadMI is a partial register load that we can't fold into \p MI
6152 /// because the latter uses contents that wouldn't be defined in the folded
6153 /// version.  For instance, this transformation isn't legal:
6154 ///   movss (%rdi), %xmm0
6155 ///   addps %xmm0, %xmm0
6156 /// ->
6157 ///   addps (%rdi), %xmm0
6158 ///
6159 /// But this one is:
6160 ///   movss (%rdi), %xmm0
6161 ///   addss %xmm0, %xmm0
6162 /// ->
6163 ///   addss (%rdi), %xmm0
6164 ///
isNonFoldablePartialRegisterLoad(const MachineInstr & LoadMI,const MachineInstr & UserMI,const MachineFunction & MF)6165 static bool isNonFoldablePartialRegisterLoad(const MachineInstr &LoadMI,
6166                                              const MachineInstr &UserMI,
6167                                              const MachineFunction &MF) {
6168   unsigned Opc = LoadMI.getOpcode();
6169   unsigned UserOpc = UserMI.getOpcode();
6170   unsigned RegSize =
6171       MF.getRegInfo().getRegClass(LoadMI.getOperand(0).getReg())->getSize();
6172 
6173   if ((Opc == X86::MOVSSrm || Opc == X86::VMOVSSrm) && RegSize > 4) {
6174     // These instructions only load 32 bits, we can't fold them if the
6175     // destination register is wider than 32 bits (4 bytes), and its user
6176     // instruction isn't scalar (SS).
6177     switch (UserOpc) {
6178     case X86::ADDSSrr_Int: case X86::VADDSSrr_Int:
6179     case X86::DIVSSrr_Int: case X86::VDIVSSrr_Int:
6180     case X86::MULSSrr_Int: case X86::VMULSSrr_Int:
6181     case X86::SUBSSrr_Int: case X86::VSUBSSrr_Int:
6182     case X86::VFMADDSSr132r_Int: case X86::VFNMADDSSr132r_Int:
6183     case X86::VFMADDSSr213r_Int: case X86::VFNMADDSSr213r_Int:
6184     case X86::VFMADDSSr231r_Int: case X86::VFNMADDSSr231r_Int:
6185     case X86::VFMSUBSSr132r_Int: case X86::VFNMSUBSSr132r_Int:
6186     case X86::VFMSUBSSr213r_Int: case X86::VFNMSUBSSr213r_Int:
6187     case X86::VFMSUBSSr231r_Int: case X86::VFNMSUBSSr231r_Int:
6188       return false;
6189     default:
6190       return true;
6191     }
6192   }
6193 
6194   if ((Opc == X86::MOVSDrm || Opc == X86::VMOVSDrm) && RegSize > 8) {
6195     // These instructions only load 64 bits, we can't fold them if the
6196     // destination register is wider than 64 bits (8 bytes), and its user
6197     // instruction isn't scalar (SD).
6198     switch (UserOpc) {
6199     case X86::ADDSDrr_Int: case X86::VADDSDrr_Int:
6200     case X86::DIVSDrr_Int: case X86::VDIVSDrr_Int:
6201     case X86::MULSDrr_Int: case X86::VMULSDrr_Int:
6202     case X86::SUBSDrr_Int: case X86::VSUBSDrr_Int:
6203     case X86::VFMADDSDr132r_Int: case X86::VFNMADDSDr132r_Int:
6204     case X86::VFMADDSDr213r_Int: case X86::VFNMADDSDr213r_Int:
6205     case X86::VFMADDSDr231r_Int: case X86::VFNMADDSDr231r_Int:
6206     case X86::VFMSUBSDr132r_Int: case X86::VFNMSUBSDr132r_Int:
6207     case X86::VFMSUBSDr213r_Int: case X86::VFNMSUBSDr213r_Int:
6208     case X86::VFMSUBSDr231r_Int: case X86::VFNMSUBSDr231r_Int:
6209       return false;
6210     default:
6211       return true;
6212     }
6213   }
6214 
6215   return false;
6216 }
6217 
foldMemoryOperandImpl(MachineFunction & MF,MachineInstr & MI,ArrayRef<unsigned> Ops,MachineBasicBlock::iterator InsertPt,MachineInstr & LoadMI,LiveIntervals * LIS) const6218 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
6219     MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
6220     MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI,
6221     LiveIntervals *LIS) const {
6222   // If loading from a FrameIndex, fold directly from the FrameIndex.
6223   unsigned NumOps = LoadMI.getDesc().getNumOperands();
6224   int FrameIndex;
6225   if (isLoadFromStackSlot(LoadMI, FrameIndex)) {
6226     if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF))
6227       return nullptr;
6228     return foldMemoryOperandImpl(MF, MI, Ops, InsertPt, FrameIndex, LIS);
6229   }
6230 
6231   // Check switch flag
6232   if (NoFusing) return nullptr;
6233 
6234   // Avoid partial register update stalls unless optimizing for size.
6235   if (!MF.getFunction()->optForSize() && hasPartialRegUpdate(MI.getOpcode()))
6236     return nullptr;
6237 
6238   // Determine the alignment of the load.
6239   unsigned Alignment = 0;
6240   if (LoadMI.hasOneMemOperand())
6241     Alignment = (*LoadMI.memoperands_begin())->getAlignment();
6242   else
6243     switch (LoadMI.getOpcode()) {
6244     case X86::AVX512_512_SET0:
6245     case X86::AVX512_512_SETALLONES:
6246       Alignment = 64;
6247       break;
6248     case X86::AVX2_SETALLONES:
6249     case X86::AVX_SET0:
6250     case X86::AVX512_256_SET0:
6251       Alignment = 32;
6252       break;
6253     case X86::V_SET0:
6254     case X86::V_SETALLONES:
6255     case X86::AVX512_128_SET0:
6256       Alignment = 16;
6257       break;
6258     case X86::FsFLD0SD:
6259       Alignment = 8;
6260       break;
6261     case X86::FsFLD0SS:
6262       Alignment = 4;
6263       break;
6264     default:
6265       return nullptr;
6266     }
6267   if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
6268     unsigned NewOpc = 0;
6269     switch (MI.getOpcode()) {
6270     default: return nullptr;
6271     case X86::TEST8rr:  NewOpc = X86::CMP8ri; break;
6272     case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
6273     case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
6274     case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
6275     }
6276     // Change to CMPXXri r, 0 first.
6277     MI.setDesc(get(NewOpc));
6278     MI.getOperand(1).ChangeToImmediate(0);
6279   } else if (Ops.size() != 1)
6280     return nullptr;
6281 
6282   // Make sure the subregisters match.
6283   // Otherwise we risk changing the size of the load.
6284   if (LoadMI.getOperand(0).getSubReg() != MI.getOperand(Ops[0]).getSubReg())
6285     return nullptr;
6286 
6287   SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
6288   switch (LoadMI.getOpcode()) {
6289   case X86::V_SET0:
6290   case X86::V_SETALLONES:
6291   case X86::AVX2_SETALLONES:
6292   case X86::AVX_SET0:
6293   case X86::AVX512_128_SET0:
6294   case X86::AVX512_256_SET0:
6295   case X86::AVX512_512_SET0:
6296   case X86::AVX512_512_SETALLONES:
6297   case X86::FsFLD0SD:
6298   case X86::FsFLD0SS: {
6299     // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
6300     // Create a constant-pool entry and operands to load from it.
6301 
6302     // Medium and large mode can't fold loads this way.
6303     if (MF.getTarget().getCodeModel() != CodeModel::Small &&
6304         MF.getTarget().getCodeModel() != CodeModel::Kernel)
6305       return nullptr;
6306 
6307     // x86-32 PIC requires a PIC base register for constant pools.
6308     unsigned PICBase = 0;
6309     if (MF.getTarget().isPositionIndependent()) {
6310       if (Subtarget.is64Bit())
6311         PICBase = X86::RIP;
6312       else
6313         // FIXME: PICBase = getGlobalBaseReg(&MF);
6314         // This doesn't work for several reasons.
6315         // 1. GlobalBaseReg may have been spilled.
6316         // 2. It may not be live at MI.
6317         return nullptr;
6318     }
6319 
6320     // Create a constant-pool entry.
6321     MachineConstantPool &MCP = *MF.getConstantPool();
6322     Type *Ty;
6323     unsigned Opc = LoadMI.getOpcode();
6324     if (Opc == X86::FsFLD0SS)
6325       Ty = Type::getFloatTy(MF.getFunction()->getContext());
6326     else if (Opc == X86::FsFLD0SD)
6327       Ty = Type::getDoubleTy(MF.getFunction()->getContext());
6328     else if (Opc == X86::AVX512_512_SET0 || Opc == X86::AVX512_512_SETALLONES)
6329       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()),16);
6330     else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0 ||
6331              Opc == X86::AVX512_256_SET0)
6332       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
6333     else
6334       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
6335 
6336     bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES ||
6337                       Opc == X86::AVX512_512_SETALLONES);
6338     const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
6339                                     Constant::getNullValue(Ty);
6340     unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
6341 
6342     // Create operands to load from the constant pool entry.
6343     MOs.push_back(MachineOperand::CreateReg(PICBase, false));
6344     MOs.push_back(MachineOperand::CreateImm(1));
6345     MOs.push_back(MachineOperand::CreateReg(0, false));
6346     MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
6347     MOs.push_back(MachineOperand::CreateReg(0, false));
6348     break;
6349   }
6350   default: {
6351     if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF))
6352       return nullptr;
6353 
6354     // Folding a normal load. Just copy the load's address operands.
6355     MOs.append(LoadMI.operands_begin() + NumOps - X86::AddrNumOperands,
6356                LoadMI.operands_begin() + NumOps);
6357     break;
6358   }
6359   }
6360   return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, InsertPt,
6361                                /*Size=*/0, Alignment, /*AllowCommute=*/true);
6362 }
6363 
unfoldMemoryOperand(MachineFunction & MF,MachineInstr & MI,unsigned Reg,bool UnfoldLoad,bool UnfoldStore,SmallVectorImpl<MachineInstr * > & NewMIs) const6364 bool X86InstrInfo::unfoldMemoryOperand(
6365     MachineFunction &MF, MachineInstr &MI, unsigned Reg, bool UnfoldLoad,
6366     bool UnfoldStore, SmallVectorImpl<MachineInstr *> &NewMIs) const {
6367   auto I = MemOp2RegOpTable.find(MI.getOpcode());
6368   if (I == MemOp2RegOpTable.end())
6369     return false;
6370   unsigned Opc = I->second.first;
6371   unsigned Index = I->second.second & TB_INDEX_MASK;
6372   bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
6373   bool FoldedStore = I->second.second & TB_FOLDED_STORE;
6374   if (UnfoldLoad && !FoldedLoad)
6375     return false;
6376   UnfoldLoad &= FoldedLoad;
6377   if (UnfoldStore && !FoldedStore)
6378     return false;
6379   UnfoldStore &= FoldedStore;
6380 
6381   const MCInstrDesc &MCID = get(Opc);
6382   const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
6383   // TODO: Check if 32-byte or greater accesses are slow too?
6384   if (!MI.hasOneMemOperand() && RC == &X86::VR128RegClass &&
6385       Subtarget.isUnalignedMem16Slow())
6386     // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
6387     // conservatively assume the address is unaligned. That's bad for
6388     // performance.
6389     return false;
6390   SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
6391   SmallVector<MachineOperand,2> BeforeOps;
6392   SmallVector<MachineOperand,2> AfterOps;
6393   SmallVector<MachineOperand,4> ImpOps;
6394   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
6395     MachineOperand &Op = MI.getOperand(i);
6396     if (i >= Index && i < Index + X86::AddrNumOperands)
6397       AddrOps.push_back(Op);
6398     else if (Op.isReg() && Op.isImplicit())
6399       ImpOps.push_back(Op);
6400     else if (i < Index)
6401       BeforeOps.push_back(Op);
6402     else if (i > Index)
6403       AfterOps.push_back(Op);
6404   }
6405 
6406   // Emit the load instruction.
6407   if (UnfoldLoad) {
6408     std::pair<MachineInstr::mmo_iterator, MachineInstr::mmo_iterator> MMOs =
6409         MF.extractLoadMemRefs(MI.memoperands_begin(), MI.memoperands_end());
6410     loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
6411     if (UnfoldStore) {
6412       // Address operands cannot be marked isKill.
6413       for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
6414         MachineOperand &MO = NewMIs[0]->getOperand(i);
6415         if (MO.isReg())
6416           MO.setIsKill(false);
6417       }
6418     }
6419   }
6420 
6421   // Emit the data processing instruction.
6422   MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI.getDebugLoc(), true);
6423   MachineInstrBuilder MIB(MF, DataMI);
6424 
6425   if (FoldedStore)
6426     MIB.addReg(Reg, RegState::Define);
6427   for (MachineOperand &BeforeOp : BeforeOps)
6428     MIB.addOperand(BeforeOp);
6429   if (FoldedLoad)
6430     MIB.addReg(Reg);
6431   for (MachineOperand &AfterOp : AfterOps)
6432     MIB.addOperand(AfterOp);
6433   for (MachineOperand &ImpOp : ImpOps) {
6434     MIB.addReg(ImpOp.getReg(),
6435                getDefRegState(ImpOp.isDef()) |
6436                RegState::Implicit |
6437                getKillRegState(ImpOp.isKill()) |
6438                getDeadRegState(ImpOp.isDead()) |
6439                getUndefRegState(ImpOp.isUndef()));
6440   }
6441   // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
6442   switch (DataMI->getOpcode()) {
6443   default: break;
6444   case X86::CMP64ri32:
6445   case X86::CMP64ri8:
6446   case X86::CMP32ri:
6447   case X86::CMP32ri8:
6448   case X86::CMP16ri:
6449   case X86::CMP16ri8:
6450   case X86::CMP8ri: {
6451     MachineOperand &MO0 = DataMI->getOperand(0);
6452     MachineOperand &MO1 = DataMI->getOperand(1);
6453     if (MO1.getImm() == 0) {
6454       unsigned NewOpc;
6455       switch (DataMI->getOpcode()) {
6456       default: llvm_unreachable("Unreachable!");
6457       case X86::CMP64ri8:
6458       case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
6459       case X86::CMP32ri8:
6460       case X86::CMP32ri:   NewOpc = X86::TEST32rr; break;
6461       case X86::CMP16ri8:
6462       case X86::CMP16ri:   NewOpc = X86::TEST16rr; break;
6463       case X86::CMP8ri:    NewOpc = X86::TEST8rr; break;
6464       }
6465       DataMI->setDesc(get(NewOpc));
6466       MO1.ChangeToRegister(MO0.getReg(), false);
6467     }
6468   }
6469   }
6470   NewMIs.push_back(DataMI);
6471 
6472   // Emit the store instruction.
6473   if (UnfoldStore) {
6474     const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
6475     std::pair<MachineInstr::mmo_iterator, MachineInstr::mmo_iterator> MMOs =
6476         MF.extractStoreMemRefs(MI.memoperands_begin(), MI.memoperands_end());
6477     storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
6478   }
6479 
6480   return true;
6481 }
6482 
6483 bool
unfoldMemoryOperand(SelectionDAG & DAG,SDNode * N,SmallVectorImpl<SDNode * > & NewNodes) const6484 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
6485                                   SmallVectorImpl<SDNode*> &NewNodes) const {
6486   if (!N->isMachineOpcode())
6487     return false;
6488 
6489   auto I = MemOp2RegOpTable.find(N->getMachineOpcode());
6490   if (I == MemOp2RegOpTable.end())
6491     return false;
6492   unsigned Opc = I->second.first;
6493   unsigned Index = I->second.second & TB_INDEX_MASK;
6494   bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
6495   bool FoldedStore = I->second.second & TB_FOLDED_STORE;
6496   const MCInstrDesc &MCID = get(Opc);
6497   MachineFunction &MF = DAG.getMachineFunction();
6498   const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
6499   unsigned NumDefs = MCID.NumDefs;
6500   std::vector<SDValue> AddrOps;
6501   std::vector<SDValue> BeforeOps;
6502   std::vector<SDValue> AfterOps;
6503   SDLoc dl(N);
6504   unsigned NumOps = N->getNumOperands();
6505   for (unsigned i = 0; i != NumOps-1; ++i) {
6506     SDValue Op = N->getOperand(i);
6507     if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
6508       AddrOps.push_back(Op);
6509     else if (i < Index-NumDefs)
6510       BeforeOps.push_back(Op);
6511     else if (i > Index-NumDefs)
6512       AfterOps.push_back(Op);
6513   }
6514   SDValue Chain = N->getOperand(NumOps-1);
6515   AddrOps.push_back(Chain);
6516 
6517   // Emit the load instruction.
6518   SDNode *Load = nullptr;
6519   if (FoldedLoad) {
6520     EVT VT = *RC->vt_begin();
6521     std::pair<MachineInstr::mmo_iterator,
6522               MachineInstr::mmo_iterator> MMOs =
6523       MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
6524                             cast<MachineSDNode>(N)->memoperands_end());
6525     if (!(*MMOs.first) &&
6526         RC == &X86::VR128RegClass &&
6527         Subtarget.isUnalignedMem16Slow())
6528       // Do not introduce a slow unaligned load.
6529       return false;
6530     // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
6531     // memory access is slow above.
6532     unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
6533     bool isAligned = (*MMOs.first) &&
6534                      (*MMOs.first)->getAlignment() >= Alignment;
6535     Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, Subtarget), dl,
6536                               VT, MVT::Other, AddrOps);
6537     NewNodes.push_back(Load);
6538 
6539     // Preserve memory reference information.
6540     cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
6541   }
6542 
6543   // Emit the data processing instruction.
6544   std::vector<EVT> VTs;
6545   const TargetRegisterClass *DstRC = nullptr;
6546   if (MCID.getNumDefs() > 0) {
6547     DstRC = getRegClass(MCID, 0, &RI, MF);
6548     VTs.push_back(*DstRC->vt_begin());
6549   }
6550   for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
6551     EVT VT = N->getValueType(i);
6552     if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
6553       VTs.push_back(VT);
6554   }
6555   if (Load)
6556     BeforeOps.push_back(SDValue(Load, 0));
6557   BeforeOps.insert(BeforeOps.end(), AfterOps.begin(), AfterOps.end());
6558   SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
6559   NewNodes.push_back(NewNode);
6560 
6561   // Emit the store instruction.
6562   if (FoldedStore) {
6563     AddrOps.pop_back();
6564     AddrOps.push_back(SDValue(NewNode, 0));
6565     AddrOps.push_back(Chain);
6566     std::pair<MachineInstr::mmo_iterator,
6567               MachineInstr::mmo_iterator> MMOs =
6568       MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
6569                              cast<MachineSDNode>(N)->memoperands_end());
6570     if (!(*MMOs.first) &&
6571         RC == &X86::VR128RegClass &&
6572         Subtarget.isUnalignedMem16Slow())
6573       // Do not introduce a slow unaligned store.
6574       return false;
6575     // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
6576     // memory access is slow above.
6577     unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
6578     bool isAligned = (*MMOs.first) &&
6579                      (*MMOs.first)->getAlignment() >= Alignment;
6580     SDNode *Store =
6581         DAG.getMachineNode(getStoreRegOpcode(0, DstRC, isAligned, Subtarget),
6582                            dl, MVT::Other, AddrOps);
6583     NewNodes.push_back(Store);
6584 
6585     // Preserve memory reference information.
6586     cast<MachineSDNode>(Store)->setMemRefs(MMOs.first, MMOs.second);
6587   }
6588 
6589   return true;
6590 }
6591 
getOpcodeAfterMemoryUnfold(unsigned Opc,bool UnfoldLoad,bool UnfoldStore,unsigned * LoadRegIndex) const6592 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
6593                                       bool UnfoldLoad, bool UnfoldStore,
6594                                       unsigned *LoadRegIndex) const {
6595   auto I = MemOp2RegOpTable.find(Opc);
6596   if (I == MemOp2RegOpTable.end())
6597     return 0;
6598   bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
6599   bool FoldedStore = I->second.second & TB_FOLDED_STORE;
6600   if (UnfoldLoad && !FoldedLoad)
6601     return 0;
6602   if (UnfoldStore && !FoldedStore)
6603     return 0;
6604   if (LoadRegIndex)
6605     *LoadRegIndex = I->second.second & TB_INDEX_MASK;
6606   return I->second.first;
6607 }
6608 
6609 bool
areLoadsFromSameBasePtr(SDNode * Load1,SDNode * Load2,int64_t & Offset1,int64_t & Offset2) const6610 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
6611                                      int64_t &Offset1, int64_t &Offset2) const {
6612   if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
6613     return false;
6614   unsigned Opc1 = Load1->getMachineOpcode();
6615   unsigned Opc2 = Load2->getMachineOpcode();
6616   switch (Opc1) {
6617   default: return false;
6618   case X86::MOV8rm:
6619   case X86::MOV16rm:
6620   case X86::MOV32rm:
6621   case X86::MOV64rm:
6622   case X86::LD_Fp32m:
6623   case X86::LD_Fp64m:
6624   case X86::LD_Fp80m:
6625   case X86::MOVSSrm:
6626   case X86::MOVSDrm:
6627   case X86::MMX_MOVD64rm:
6628   case X86::MMX_MOVQ64rm:
6629   case X86::FsMOVAPSrm:
6630   case X86::FsMOVAPDrm:
6631   case X86::MOVAPSrm:
6632   case X86::MOVUPSrm:
6633   case X86::MOVAPDrm:
6634   case X86::MOVDQArm:
6635   case X86::MOVDQUrm:
6636   // AVX load instructions
6637   case X86::VMOVSSrm:
6638   case X86::VMOVSDrm:
6639   case X86::FsVMOVAPSrm:
6640   case X86::FsVMOVAPDrm:
6641   case X86::VMOVAPSrm:
6642   case X86::VMOVUPSrm:
6643   case X86::VMOVAPDrm:
6644   case X86::VMOVDQArm:
6645   case X86::VMOVDQUrm:
6646   case X86::VMOVAPSYrm:
6647   case X86::VMOVUPSYrm:
6648   case X86::VMOVAPDYrm:
6649   case X86::VMOVDQAYrm:
6650   case X86::VMOVDQUYrm:
6651     break;
6652   }
6653   switch (Opc2) {
6654   default: return false;
6655   case X86::MOV8rm:
6656   case X86::MOV16rm:
6657   case X86::MOV32rm:
6658   case X86::MOV64rm:
6659   case X86::LD_Fp32m:
6660   case X86::LD_Fp64m:
6661   case X86::LD_Fp80m:
6662   case X86::MOVSSrm:
6663   case X86::MOVSDrm:
6664   case X86::MMX_MOVD64rm:
6665   case X86::MMX_MOVQ64rm:
6666   case X86::FsMOVAPSrm:
6667   case X86::FsMOVAPDrm:
6668   case X86::MOVAPSrm:
6669   case X86::MOVUPSrm:
6670   case X86::MOVAPDrm:
6671   case X86::MOVDQArm:
6672   case X86::MOVDQUrm:
6673   // AVX load instructions
6674   case X86::VMOVSSrm:
6675   case X86::VMOVSDrm:
6676   case X86::FsVMOVAPSrm:
6677   case X86::FsVMOVAPDrm:
6678   case X86::VMOVAPSrm:
6679   case X86::VMOVUPSrm:
6680   case X86::VMOVAPDrm:
6681   case X86::VMOVDQArm:
6682   case X86::VMOVDQUrm:
6683   case X86::VMOVAPSYrm:
6684   case X86::VMOVUPSYrm:
6685   case X86::VMOVAPDYrm:
6686   case X86::VMOVDQAYrm:
6687   case X86::VMOVDQUYrm:
6688     break;
6689   }
6690 
6691   // Check if chain operands and base addresses match.
6692   if (Load1->getOperand(0) != Load2->getOperand(0) ||
6693       Load1->getOperand(5) != Load2->getOperand(5))
6694     return false;
6695   // Segment operands should match as well.
6696   if (Load1->getOperand(4) != Load2->getOperand(4))
6697     return false;
6698   // Scale should be 1, Index should be Reg0.
6699   if (Load1->getOperand(1) == Load2->getOperand(1) &&
6700       Load1->getOperand(2) == Load2->getOperand(2)) {
6701     if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
6702       return false;
6703 
6704     // Now let's examine the displacements.
6705     if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
6706         isa<ConstantSDNode>(Load2->getOperand(3))) {
6707       Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
6708       Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
6709       return true;
6710     }
6711   }
6712   return false;
6713 }
6714 
shouldScheduleLoadsNear(SDNode * Load1,SDNode * Load2,int64_t Offset1,int64_t Offset2,unsigned NumLoads) const6715 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
6716                                            int64_t Offset1, int64_t Offset2,
6717                                            unsigned NumLoads) const {
6718   assert(Offset2 > Offset1);
6719   if ((Offset2 - Offset1) / 8 > 64)
6720     return false;
6721 
6722   unsigned Opc1 = Load1->getMachineOpcode();
6723   unsigned Opc2 = Load2->getMachineOpcode();
6724   if (Opc1 != Opc2)
6725     return false;  // FIXME: overly conservative?
6726 
6727   switch (Opc1) {
6728   default: break;
6729   case X86::LD_Fp32m:
6730   case X86::LD_Fp64m:
6731   case X86::LD_Fp80m:
6732   case X86::MMX_MOVD64rm:
6733   case X86::MMX_MOVQ64rm:
6734     return false;
6735   }
6736 
6737   EVT VT = Load1->getValueType(0);
6738   switch (VT.getSimpleVT().SimpleTy) {
6739   default:
6740     // XMM registers. In 64-bit mode we can be a bit more aggressive since we
6741     // have 16 of them to play with.
6742     if (Subtarget.is64Bit()) {
6743       if (NumLoads >= 3)
6744         return false;
6745     } else if (NumLoads) {
6746       return false;
6747     }
6748     break;
6749   case MVT::i8:
6750   case MVT::i16:
6751   case MVT::i32:
6752   case MVT::i64:
6753   case MVT::f32:
6754   case MVT::f64:
6755     if (NumLoads)
6756       return false;
6757     break;
6758   }
6759 
6760   return true;
6761 }
6762 
shouldScheduleAdjacent(MachineInstr & First,MachineInstr & Second) const6763 bool X86InstrInfo::shouldScheduleAdjacent(MachineInstr &First,
6764                                           MachineInstr &Second) const {
6765   // Check if this processor supports macro-fusion. Since this is a minor
6766   // heuristic, we haven't specifically reserved a feature. hasAVX is a decent
6767   // proxy for SandyBridge+.
6768   if (!Subtarget.hasAVX())
6769     return false;
6770 
6771   enum {
6772     FuseTest,
6773     FuseCmp,
6774     FuseInc
6775   } FuseKind;
6776 
6777   switch (Second.getOpcode()) {
6778   default:
6779     return false;
6780   case X86::JE_1:
6781   case X86::JNE_1:
6782   case X86::JL_1:
6783   case X86::JLE_1:
6784   case X86::JG_1:
6785   case X86::JGE_1:
6786     FuseKind = FuseInc;
6787     break;
6788   case X86::JB_1:
6789   case X86::JBE_1:
6790   case X86::JA_1:
6791   case X86::JAE_1:
6792     FuseKind = FuseCmp;
6793     break;
6794   case X86::JS_1:
6795   case X86::JNS_1:
6796   case X86::JP_1:
6797   case X86::JNP_1:
6798   case X86::JO_1:
6799   case X86::JNO_1:
6800     FuseKind = FuseTest;
6801     break;
6802   }
6803   switch (First.getOpcode()) {
6804   default:
6805     return false;
6806   case X86::TEST8rr:
6807   case X86::TEST16rr:
6808   case X86::TEST32rr:
6809   case X86::TEST64rr:
6810   case X86::TEST8ri:
6811   case X86::TEST16ri:
6812   case X86::TEST32ri:
6813   case X86::TEST32i32:
6814   case X86::TEST64i32:
6815   case X86::TEST64ri32:
6816   case X86::TEST8rm:
6817   case X86::TEST16rm:
6818   case X86::TEST32rm:
6819   case X86::TEST64rm:
6820   case X86::TEST8ri_NOREX:
6821   case X86::AND16i16:
6822   case X86::AND16ri:
6823   case X86::AND16ri8:
6824   case X86::AND16rm:
6825   case X86::AND16rr:
6826   case X86::AND32i32:
6827   case X86::AND32ri:
6828   case X86::AND32ri8:
6829   case X86::AND32rm:
6830   case X86::AND32rr:
6831   case X86::AND64i32:
6832   case X86::AND64ri32:
6833   case X86::AND64ri8:
6834   case X86::AND64rm:
6835   case X86::AND64rr:
6836   case X86::AND8i8:
6837   case X86::AND8ri:
6838   case X86::AND8rm:
6839   case X86::AND8rr:
6840     return true;
6841   case X86::CMP16i16:
6842   case X86::CMP16ri:
6843   case X86::CMP16ri8:
6844   case X86::CMP16rm:
6845   case X86::CMP16rr:
6846   case X86::CMP32i32:
6847   case X86::CMP32ri:
6848   case X86::CMP32ri8:
6849   case X86::CMP32rm:
6850   case X86::CMP32rr:
6851   case X86::CMP64i32:
6852   case X86::CMP64ri32:
6853   case X86::CMP64ri8:
6854   case X86::CMP64rm:
6855   case X86::CMP64rr:
6856   case X86::CMP8i8:
6857   case X86::CMP8ri:
6858   case X86::CMP8rm:
6859   case X86::CMP8rr:
6860   case X86::ADD16i16:
6861   case X86::ADD16ri:
6862   case X86::ADD16ri8:
6863   case X86::ADD16ri8_DB:
6864   case X86::ADD16ri_DB:
6865   case X86::ADD16rm:
6866   case X86::ADD16rr:
6867   case X86::ADD16rr_DB:
6868   case X86::ADD32i32:
6869   case X86::ADD32ri:
6870   case X86::ADD32ri8:
6871   case X86::ADD32ri8_DB:
6872   case X86::ADD32ri_DB:
6873   case X86::ADD32rm:
6874   case X86::ADD32rr:
6875   case X86::ADD32rr_DB:
6876   case X86::ADD64i32:
6877   case X86::ADD64ri32:
6878   case X86::ADD64ri32_DB:
6879   case X86::ADD64ri8:
6880   case X86::ADD64ri8_DB:
6881   case X86::ADD64rm:
6882   case X86::ADD64rr:
6883   case X86::ADD64rr_DB:
6884   case X86::ADD8i8:
6885   case X86::ADD8mi:
6886   case X86::ADD8mr:
6887   case X86::ADD8ri:
6888   case X86::ADD8rm:
6889   case X86::ADD8rr:
6890   case X86::SUB16i16:
6891   case X86::SUB16ri:
6892   case X86::SUB16ri8:
6893   case X86::SUB16rm:
6894   case X86::SUB16rr:
6895   case X86::SUB32i32:
6896   case X86::SUB32ri:
6897   case X86::SUB32ri8:
6898   case X86::SUB32rm:
6899   case X86::SUB32rr:
6900   case X86::SUB64i32:
6901   case X86::SUB64ri32:
6902   case X86::SUB64ri8:
6903   case X86::SUB64rm:
6904   case X86::SUB64rr:
6905   case X86::SUB8i8:
6906   case X86::SUB8ri:
6907   case X86::SUB8rm:
6908   case X86::SUB8rr:
6909     return FuseKind == FuseCmp || FuseKind == FuseInc;
6910   case X86::INC16r:
6911   case X86::INC32r:
6912   case X86::INC64r:
6913   case X86::INC8r:
6914   case X86::DEC16r:
6915   case X86::DEC32r:
6916   case X86::DEC64r:
6917   case X86::DEC8r:
6918     return FuseKind == FuseInc;
6919   }
6920 }
6921 
6922 bool X86InstrInfo::
ReverseBranchCondition(SmallVectorImpl<MachineOperand> & Cond) const6923 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
6924   assert(Cond.size() == 1 && "Invalid X86 branch condition!");
6925   X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
6926   Cond[0].setImm(GetOppositeBranchCondition(CC));
6927   return false;
6928 }
6929 
6930 bool X86InstrInfo::
isSafeToMoveRegClassDefs(const TargetRegisterClass * RC) const6931 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
6932   // FIXME: Return false for x87 stack register classes for now. We can't
6933   // allow any loads of these registers before FpGet_ST0_80.
6934   return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
6935            RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
6936 }
6937 
6938 /// Return a virtual register initialized with the
6939 /// the global base register value. Output instructions required to
6940 /// initialize the register in the function entry block, if necessary.
6941 ///
6942 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
6943 ///
getGlobalBaseReg(MachineFunction * MF) const6944 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
6945   assert(!Subtarget.is64Bit() &&
6946          "X86-64 PIC uses RIP relative addressing");
6947 
6948   X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
6949   unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
6950   if (GlobalBaseReg != 0)
6951     return GlobalBaseReg;
6952 
6953   // Create the register. The code to initialize it is inserted
6954   // later, by the CGBR pass (below).
6955   MachineRegisterInfo &RegInfo = MF->getRegInfo();
6956   GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
6957   X86FI->setGlobalBaseReg(GlobalBaseReg);
6958   return GlobalBaseReg;
6959 }
6960 
6961 // These are the replaceable SSE instructions. Some of these have Int variants
6962 // that we don't include here. We don't want to replace instructions selected
6963 // by intrinsics.
6964 static const uint16_t ReplaceableInstrs[][3] = {
6965   //PackedSingle     PackedDouble    PackedInt
6966   { X86::MOVAPSmr,   X86::MOVAPDmr,  X86::MOVDQAmr  },
6967   { X86::MOVAPSrm,   X86::MOVAPDrm,  X86::MOVDQArm  },
6968   { X86::MOVAPSrr,   X86::MOVAPDrr,  X86::MOVDQArr  },
6969   { X86::MOVUPSmr,   X86::MOVUPDmr,  X86::MOVDQUmr  },
6970   { X86::MOVUPSrm,   X86::MOVUPDrm,  X86::MOVDQUrm  },
6971   { X86::MOVLPSmr,   X86::MOVLPDmr,  X86::MOVPQI2QImr  },
6972   { X86::MOVNTPSmr,  X86::MOVNTPDmr, X86::MOVNTDQmr },
6973   { X86::ANDNPSrm,   X86::ANDNPDrm,  X86::PANDNrm   },
6974   { X86::ANDNPSrr,   X86::ANDNPDrr,  X86::PANDNrr   },
6975   { X86::ANDPSrm,    X86::ANDPDrm,   X86::PANDrm    },
6976   { X86::ANDPSrr,    X86::ANDPDrr,   X86::PANDrr    },
6977   { X86::ORPSrm,     X86::ORPDrm,    X86::PORrm     },
6978   { X86::ORPSrr,     X86::ORPDrr,    X86::PORrr     },
6979   { X86::XORPSrm,    X86::XORPDrm,   X86::PXORrm    },
6980   { X86::XORPSrr,    X86::XORPDrr,   X86::PXORrr    },
6981   // AVX 128-bit support
6982   { X86::VMOVAPSmr,  X86::VMOVAPDmr,  X86::VMOVDQAmr  },
6983   { X86::VMOVAPSrm,  X86::VMOVAPDrm,  X86::VMOVDQArm  },
6984   { X86::VMOVAPSrr,  X86::VMOVAPDrr,  X86::VMOVDQArr  },
6985   { X86::VMOVUPSmr,  X86::VMOVUPDmr,  X86::VMOVDQUmr  },
6986   { X86::VMOVUPSrm,  X86::VMOVUPDrm,  X86::VMOVDQUrm  },
6987   { X86::VMOVLPSmr,  X86::VMOVLPDmr,  X86::VMOVPQI2QImr  },
6988   { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
6989   { X86::VANDNPSrm,  X86::VANDNPDrm,  X86::VPANDNrm   },
6990   { X86::VANDNPSrr,  X86::VANDNPDrr,  X86::VPANDNrr   },
6991   { X86::VANDPSrm,   X86::VANDPDrm,   X86::VPANDrm    },
6992   { X86::VANDPSrr,   X86::VANDPDrr,   X86::VPANDrr    },
6993   { X86::VORPSrm,    X86::VORPDrm,    X86::VPORrm     },
6994   { X86::VORPSrr,    X86::VORPDrr,    X86::VPORrr     },
6995   { X86::VXORPSrm,   X86::VXORPDrm,   X86::VPXORrm    },
6996   { X86::VXORPSrr,   X86::VXORPDrr,   X86::VPXORrr    },
6997   // AVX 256-bit support
6998   { X86::VMOVAPSYmr,   X86::VMOVAPDYmr,   X86::VMOVDQAYmr  },
6999   { X86::VMOVAPSYrm,   X86::VMOVAPDYrm,   X86::VMOVDQAYrm  },
7000   { X86::VMOVAPSYrr,   X86::VMOVAPDYrr,   X86::VMOVDQAYrr  },
7001   { X86::VMOVUPSYmr,   X86::VMOVUPDYmr,   X86::VMOVDQUYmr  },
7002   { X86::VMOVUPSYrm,   X86::VMOVUPDYrm,   X86::VMOVDQUYrm  },
7003   { X86::VMOVNTPSYmr,  X86::VMOVNTPDYmr,  X86::VMOVNTDQYmr }
7004 };
7005 
7006 static const uint16_t ReplaceableInstrsAVX2[][3] = {
7007   //PackedSingle       PackedDouble       PackedInt
7008   { X86::VANDNPSYrm,   X86::VANDNPDYrm,   X86::VPANDNYrm   },
7009   { X86::VANDNPSYrr,   X86::VANDNPDYrr,   X86::VPANDNYrr   },
7010   { X86::VANDPSYrm,    X86::VANDPDYrm,    X86::VPANDYrm    },
7011   { X86::VANDPSYrr,    X86::VANDPDYrr,    X86::VPANDYrr    },
7012   { X86::VORPSYrm,     X86::VORPDYrm,     X86::VPORYrm     },
7013   { X86::VORPSYrr,     X86::VORPDYrr,     X86::VPORYrr     },
7014   { X86::VXORPSYrm,    X86::VXORPDYrm,    X86::VPXORYrm    },
7015   { X86::VXORPSYrr,    X86::VXORPDYrr,    X86::VPXORYrr    },
7016   { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
7017   { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
7018   { X86::VINSERTF128rm,  X86::VINSERTF128rm,  X86::VINSERTI128rm },
7019   { X86::VINSERTF128rr,  X86::VINSERTF128rr,  X86::VINSERTI128rr },
7020   { X86::VPERM2F128rm,   X86::VPERM2F128rm,   X86::VPERM2I128rm },
7021   { X86::VPERM2F128rr,   X86::VPERM2F128rr,   X86::VPERM2I128rr },
7022   { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
7023   { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
7024   { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
7025   { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
7026   { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
7027   { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm}
7028 };
7029 
7030 // FIXME: Some shuffle and unpack instructions have equivalents in different
7031 // domains, but they require a bit more work than just switching opcodes.
7032 
lookup(unsigned opcode,unsigned domain)7033 static const uint16_t *lookup(unsigned opcode, unsigned domain) {
7034   for (const uint16_t (&Row)[3] : ReplaceableInstrs)
7035     if (Row[domain-1] == opcode)
7036       return Row;
7037   return nullptr;
7038 }
7039 
lookupAVX2(unsigned opcode,unsigned domain)7040 static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
7041   for (const uint16_t (&Row)[3] : ReplaceableInstrsAVX2)
7042     if (Row[domain-1] == opcode)
7043       return Row;
7044   return nullptr;
7045 }
7046 
7047 std::pair<uint16_t, uint16_t>
getExecutionDomain(const MachineInstr & MI) const7048 X86InstrInfo::getExecutionDomain(const MachineInstr &MI) const {
7049   uint16_t domain = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
7050   bool hasAVX2 = Subtarget.hasAVX2();
7051   uint16_t validDomains = 0;
7052   if (domain && lookup(MI.getOpcode(), domain))
7053     validDomains = 0xe;
7054   else if (domain && lookupAVX2(MI.getOpcode(), domain))
7055     validDomains = hasAVX2 ? 0xe : 0x6;
7056   return std::make_pair(domain, validDomains);
7057 }
7058 
setExecutionDomain(MachineInstr & MI,unsigned Domain) const7059 void X86InstrInfo::setExecutionDomain(MachineInstr &MI, unsigned Domain) const {
7060   assert(Domain>0 && Domain<4 && "Invalid execution domain");
7061   uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
7062   assert(dom && "Not an SSE instruction");
7063   const uint16_t *table = lookup(MI.getOpcode(), dom);
7064   if (!table) { // try the other table
7065     assert((Subtarget.hasAVX2() || Domain < 3) &&
7066            "256-bit vector operations only available in AVX2");
7067     table = lookupAVX2(MI.getOpcode(), dom);
7068   }
7069   assert(table && "Cannot change domain");
7070   MI.setDesc(get(table[Domain - 1]));
7071 }
7072 
7073 /// Return the noop instruction to use for a noop.
getNoopForMachoTarget(MCInst & NopInst) const7074 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
7075   NopInst.setOpcode(X86::NOOP);
7076 }
7077 
7078 // This code must remain in sync with getJumpInstrTableEntryBound in this class!
7079 // In particular, getJumpInstrTableEntryBound must always return an upper bound
7080 // on the encoding lengths of the instructions generated by
7081 // getUnconditionalBranch and getTrap.
getUnconditionalBranch(MCInst & Branch,const MCSymbolRefExpr * BranchTarget) const7082 void X86InstrInfo::getUnconditionalBranch(
7083     MCInst &Branch, const MCSymbolRefExpr *BranchTarget) const {
7084   Branch.setOpcode(X86::JMP_1);
7085   Branch.addOperand(MCOperand::createExpr(BranchTarget));
7086 }
7087 
7088 // This code must remain in sync with getJumpInstrTableEntryBound in this class!
7089 // In particular, getJumpInstrTableEntryBound must always return an upper bound
7090 // on the encoding lengths of the instructions generated by
7091 // getUnconditionalBranch and getTrap.
getTrap(MCInst & MI) const7092 void X86InstrInfo::getTrap(MCInst &MI) const {
7093   MI.setOpcode(X86::TRAP);
7094 }
7095 
7096 // See getTrap and getUnconditionalBranch for conditions on the value returned
7097 // by this function.
getJumpInstrTableEntryBound() const7098 unsigned X86InstrInfo::getJumpInstrTableEntryBound() const {
7099   // 5 bytes suffice: JMP_4 Symbol@PLT is uses 1 byte (E9) for the JMP_4 and 4
7100   // bytes for the symbol offset. And TRAP is ud2, which is two bytes (0F 0B).
7101   return 5;
7102 }
7103 
isHighLatencyDef(int opc) const7104 bool X86InstrInfo::isHighLatencyDef(int opc) const {
7105   switch (opc) {
7106   default: return false;
7107   case X86::DIVSDrm:
7108   case X86::DIVSDrm_Int:
7109   case X86::DIVSDrr:
7110   case X86::DIVSDrr_Int:
7111   case X86::DIVSSrm:
7112   case X86::DIVSSrm_Int:
7113   case X86::DIVSSrr:
7114   case X86::DIVSSrr_Int:
7115   case X86::SQRTPDm:
7116   case X86::SQRTPDr:
7117   case X86::SQRTPSm:
7118   case X86::SQRTPSr:
7119   case X86::SQRTSDm:
7120   case X86::SQRTSDm_Int:
7121   case X86::SQRTSDr:
7122   case X86::SQRTSDr_Int:
7123   case X86::SQRTSSm:
7124   case X86::SQRTSSm_Int:
7125   case X86::SQRTSSr:
7126   case X86::SQRTSSr_Int:
7127   // AVX instructions with high latency
7128   case X86::VDIVSDrm:
7129   case X86::VDIVSDrm_Int:
7130   case X86::VDIVSDrr:
7131   case X86::VDIVSDrr_Int:
7132   case X86::VDIVSSrm:
7133   case X86::VDIVSSrm_Int:
7134   case X86::VDIVSSrr:
7135   case X86::VDIVSSrr_Int:
7136   case X86::VSQRTPDm:
7137   case X86::VSQRTPDr:
7138   case X86::VSQRTPSm:
7139   case X86::VSQRTPSr:
7140   case X86::VSQRTSDm:
7141   case X86::VSQRTSDm_Int:
7142   case X86::VSQRTSDr:
7143   case X86::VSQRTSSm:
7144   case X86::VSQRTSSm_Int:
7145   case X86::VSQRTSSr:
7146   case X86::VSQRTPDZm:
7147   case X86::VSQRTPDZr:
7148   case X86::VSQRTPSZm:
7149   case X86::VSQRTPSZr:
7150   case X86::VSQRTSDZm:
7151   case X86::VSQRTSDZm_Int:
7152   case X86::VSQRTSDZr:
7153   case X86::VSQRTSSZm_Int:
7154   case X86::VSQRTSSZr:
7155   case X86::VSQRTSSZm:
7156   case X86::VDIVSDZrm:
7157   case X86::VDIVSDZrr:
7158   case X86::VDIVSSZrm:
7159   case X86::VDIVSSZrr:
7160 
7161   case X86::VGATHERQPSZrm:
7162   case X86::VGATHERQPDZrm:
7163   case X86::VGATHERDPDZrm:
7164   case X86::VGATHERDPSZrm:
7165   case X86::VPGATHERQDZrm:
7166   case X86::VPGATHERQQZrm:
7167   case X86::VPGATHERDDZrm:
7168   case X86::VPGATHERDQZrm:
7169   case X86::VSCATTERQPDZmr:
7170   case X86::VSCATTERQPSZmr:
7171   case X86::VSCATTERDPDZmr:
7172   case X86::VSCATTERDPSZmr:
7173   case X86::VPSCATTERQDZmr:
7174   case X86::VPSCATTERQQZmr:
7175   case X86::VPSCATTERDDZmr:
7176   case X86::VPSCATTERDQZmr:
7177     return true;
7178   }
7179 }
7180 
hasHighOperandLatency(const TargetSchedModel & SchedModel,const MachineRegisterInfo * MRI,const MachineInstr & DefMI,unsigned DefIdx,const MachineInstr & UseMI,unsigned UseIdx) const7181 bool X86InstrInfo::hasHighOperandLatency(const TargetSchedModel &SchedModel,
7182                                          const MachineRegisterInfo *MRI,
7183                                          const MachineInstr &DefMI,
7184                                          unsigned DefIdx,
7185                                          const MachineInstr &UseMI,
7186                                          unsigned UseIdx) const {
7187   return isHighLatencyDef(DefMI.getOpcode());
7188 }
7189 
hasReassociableOperands(const MachineInstr & Inst,const MachineBasicBlock * MBB) const7190 bool X86InstrInfo::hasReassociableOperands(const MachineInstr &Inst,
7191                                            const MachineBasicBlock *MBB) const {
7192   assert((Inst.getNumOperands() == 3 || Inst.getNumOperands() == 4) &&
7193          "Reassociation needs binary operators");
7194 
7195   // Integer binary math/logic instructions have a third source operand:
7196   // the EFLAGS register. That operand must be both defined here and never
7197   // used; ie, it must be dead. If the EFLAGS operand is live, then we can
7198   // not change anything because rearranging the operands could affect other
7199   // instructions that depend on the exact status flags (zero, sign, etc.)
7200   // that are set by using these particular operands with this operation.
7201   if (Inst.getNumOperands() == 4) {
7202     assert(Inst.getOperand(3).isReg() &&
7203            Inst.getOperand(3).getReg() == X86::EFLAGS &&
7204            "Unexpected operand in reassociable instruction");
7205     if (!Inst.getOperand(3).isDead())
7206       return false;
7207   }
7208 
7209   return TargetInstrInfo::hasReassociableOperands(Inst, MBB);
7210 }
7211 
7212 // TODO: There are many more machine instruction opcodes to match:
7213 //       1. Other data types (integer, vectors)
7214 //       2. Other math / logic operations (xor, or)
7215 //       3. Other forms of the same operation (intrinsics and other variants)
isAssociativeAndCommutative(const MachineInstr & Inst) const7216 bool X86InstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst) const {
7217   switch (Inst.getOpcode()) {
7218   case X86::AND8rr:
7219   case X86::AND16rr:
7220   case X86::AND32rr:
7221   case X86::AND64rr:
7222   case X86::OR8rr:
7223   case X86::OR16rr:
7224   case X86::OR32rr:
7225   case X86::OR64rr:
7226   case X86::XOR8rr:
7227   case X86::XOR16rr:
7228   case X86::XOR32rr:
7229   case X86::XOR64rr:
7230   case X86::IMUL16rr:
7231   case X86::IMUL32rr:
7232   case X86::IMUL64rr:
7233   case X86::PANDrr:
7234   case X86::PORrr:
7235   case X86::PXORrr:
7236   case X86::VPANDrr:
7237   case X86::VPANDYrr:
7238   case X86::VPORrr:
7239   case X86::VPORYrr:
7240   case X86::VPXORrr:
7241   case X86::VPXORYrr:
7242   // Normal min/max instructions are not commutative because of NaN and signed
7243   // zero semantics, but these are. Thus, there's no need to check for global
7244   // relaxed math; the instructions themselves have the properties we need.
7245   case X86::MAXCPDrr:
7246   case X86::MAXCPSrr:
7247   case X86::MAXCSDrr:
7248   case X86::MAXCSSrr:
7249   case X86::MINCPDrr:
7250   case X86::MINCPSrr:
7251   case X86::MINCSDrr:
7252   case X86::MINCSSrr:
7253   case X86::VMAXCPDrr:
7254   case X86::VMAXCPSrr:
7255   case X86::VMAXCPDYrr:
7256   case X86::VMAXCPSYrr:
7257   case X86::VMAXCSDrr:
7258   case X86::VMAXCSSrr:
7259   case X86::VMINCPDrr:
7260   case X86::VMINCPSrr:
7261   case X86::VMINCPDYrr:
7262   case X86::VMINCPSYrr:
7263   case X86::VMINCSDrr:
7264   case X86::VMINCSSrr:
7265     return true;
7266   case X86::ADDPDrr:
7267   case X86::ADDPSrr:
7268   case X86::ADDSDrr:
7269   case X86::ADDSSrr:
7270   case X86::MULPDrr:
7271   case X86::MULPSrr:
7272   case X86::MULSDrr:
7273   case X86::MULSSrr:
7274   case X86::VADDPDrr:
7275   case X86::VADDPSrr:
7276   case X86::VADDPDYrr:
7277   case X86::VADDPSYrr:
7278   case X86::VADDSDrr:
7279   case X86::VADDSSrr:
7280   case X86::VMULPDrr:
7281   case X86::VMULPSrr:
7282   case X86::VMULPDYrr:
7283   case X86::VMULPSYrr:
7284   case X86::VMULSDrr:
7285   case X86::VMULSSrr:
7286     return Inst.getParent()->getParent()->getTarget().Options.UnsafeFPMath;
7287   default:
7288     return false;
7289   }
7290 }
7291 
7292 /// This is an architecture-specific helper function of reassociateOps.
7293 /// Set special operand attributes for new instructions after reassociation.
setSpecialOperandAttr(MachineInstr & OldMI1,MachineInstr & OldMI2,MachineInstr & NewMI1,MachineInstr & NewMI2) const7294 void X86InstrInfo::setSpecialOperandAttr(MachineInstr &OldMI1,
7295                                          MachineInstr &OldMI2,
7296                                          MachineInstr &NewMI1,
7297                                          MachineInstr &NewMI2) const {
7298   // Integer instructions define an implicit EFLAGS source register operand as
7299   // the third source (fourth total) operand.
7300   if (OldMI1.getNumOperands() != 4 || OldMI2.getNumOperands() != 4)
7301     return;
7302 
7303   assert(NewMI1.getNumOperands() == 4 && NewMI2.getNumOperands() == 4 &&
7304          "Unexpected instruction type for reassociation");
7305 
7306   MachineOperand &OldOp1 = OldMI1.getOperand(3);
7307   MachineOperand &OldOp2 = OldMI2.getOperand(3);
7308   MachineOperand &NewOp1 = NewMI1.getOperand(3);
7309   MachineOperand &NewOp2 = NewMI2.getOperand(3);
7310 
7311   assert(OldOp1.isReg() && OldOp1.getReg() == X86::EFLAGS && OldOp1.isDead() &&
7312          "Must have dead EFLAGS operand in reassociable instruction");
7313   assert(OldOp2.isReg() && OldOp2.getReg() == X86::EFLAGS && OldOp2.isDead() &&
7314          "Must have dead EFLAGS operand in reassociable instruction");
7315 
7316   (void)OldOp1;
7317   (void)OldOp2;
7318 
7319   assert(NewOp1.isReg() && NewOp1.getReg() == X86::EFLAGS &&
7320          "Unexpected operand in reassociable instruction");
7321   assert(NewOp2.isReg() && NewOp2.getReg() == X86::EFLAGS &&
7322          "Unexpected operand in reassociable instruction");
7323 
7324   // Mark the new EFLAGS operands as dead to be helpful to subsequent iterations
7325   // of this pass or other passes. The EFLAGS operands must be dead in these new
7326   // instructions because the EFLAGS operands in the original instructions must
7327   // be dead in order for reassociation to occur.
7328   NewOp1.setIsDead();
7329   NewOp2.setIsDead();
7330 }
7331 
7332 std::pair<unsigned, unsigned>
decomposeMachineOperandsTargetFlags(unsigned TF) const7333 X86InstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
7334   return std::make_pair(TF, 0u);
7335 }
7336 
7337 ArrayRef<std::pair<unsigned, const char *>>
getSerializableDirectMachineOperandTargetFlags() const7338 X86InstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
7339   using namespace X86II;
7340   static const std::pair<unsigned, const char *> TargetFlags[] = {
7341       {MO_GOT_ABSOLUTE_ADDRESS, "x86-got-absolute-address"},
7342       {MO_PIC_BASE_OFFSET, "x86-pic-base-offset"},
7343       {MO_GOT, "x86-got"},
7344       {MO_GOTOFF, "x86-gotoff"},
7345       {MO_GOTPCREL, "x86-gotpcrel"},
7346       {MO_PLT, "x86-plt"},
7347       {MO_TLSGD, "x86-tlsgd"},
7348       {MO_TLSLD, "x86-tlsld"},
7349       {MO_TLSLDM, "x86-tlsldm"},
7350       {MO_GOTTPOFF, "x86-gottpoff"},
7351       {MO_INDNTPOFF, "x86-indntpoff"},
7352       {MO_TPOFF, "x86-tpoff"},
7353       {MO_DTPOFF, "x86-dtpoff"},
7354       {MO_NTPOFF, "x86-ntpoff"},
7355       {MO_GOTNTPOFF, "x86-gotntpoff"},
7356       {MO_DLLIMPORT, "x86-dllimport"},
7357       {MO_DARWIN_NONLAZY, "x86-darwin-nonlazy"},
7358       {MO_DARWIN_NONLAZY_PIC_BASE, "x86-darwin-nonlazy-pic-base"},
7359       {MO_TLVP, "x86-tlvp"},
7360       {MO_TLVP_PIC_BASE, "x86-tlvp-pic-base"},
7361       {MO_SECREL, "x86-secrel"}};
7362   return makeArrayRef(TargetFlags);
7363 }
7364 
7365 namespace {
7366   /// Create Global Base Reg pass. This initializes the PIC
7367   /// global base register for x86-32.
7368   struct CGBR : public MachineFunctionPass {
7369     static char ID;
CGBR__anon6956071e0411::CGBR7370     CGBR() : MachineFunctionPass(ID) {}
7371 
runOnMachineFunction__anon6956071e0411::CGBR7372     bool runOnMachineFunction(MachineFunction &MF) override {
7373       const X86TargetMachine *TM =
7374         static_cast<const X86TargetMachine *>(&MF.getTarget());
7375       const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
7376 
7377       // Don't do anything if this is 64-bit as 64-bit PIC
7378       // uses RIP relative addressing.
7379       if (STI.is64Bit())
7380         return false;
7381 
7382       // Only emit a global base reg in PIC mode.
7383       if (!TM->isPositionIndependent())
7384         return false;
7385 
7386       X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
7387       unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
7388 
7389       // If we didn't need a GlobalBaseReg, don't insert code.
7390       if (GlobalBaseReg == 0)
7391         return false;
7392 
7393       // Insert the set of GlobalBaseReg into the first MBB of the function
7394       MachineBasicBlock &FirstMBB = MF.front();
7395       MachineBasicBlock::iterator MBBI = FirstMBB.begin();
7396       DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
7397       MachineRegisterInfo &RegInfo = MF.getRegInfo();
7398       const X86InstrInfo *TII = STI.getInstrInfo();
7399 
7400       unsigned PC;
7401       if (STI.isPICStyleGOT())
7402         PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
7403       else
7404         PC = GlobalBaseReg;
7405 
7406       // Operand of MovePCtoStack is completely ignored by asm printer. It's
7407       // only used in JIT code emission as displacement to pc.
7408       BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
7409 
7410       // If we're using vanilla 'GOT' PIC style, we should use relative addressing
7411       // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
7412       if (STI.isPICStyleGOT()) {
7413         // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
7414         BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
7415           .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
7416                                         X86II::MO_GOT_ABSOLUTE_ADDRESS);
7417       }
7418 
7419       return true;
7420     }
7421 
getPassName__anon6956071e0411::CGBR7422     const char *getPassName() const override {
7423       return "X86 PIC Global Base Reg Initialization";
7424     }
7425 
getAnalysisUsage__anon6956071e0411::CGBR7426     void getAnalysisUsage(AnalysisUsage &AU) const override {
7427       AU.setPreservesCFG();
7428       MachineFunctionPass::getAnalysisUsage(AU);
7429     }
7430   };
7431 }
7432 
7433 char CGBR::ID = 0;
7434 FunctionPass*
createX86GlobalBaseRegPass()7435 llvm::createX86GlobalBaseRegPass() { return new CGBR(); }
7436 
7437 namespace {
7438   struct LDTLSCleanup : public MachineFunctionPass {
7439     static char ID;
LDTLSCleanup__anon6956071e0511::LDTLSCleanup7440     LDTLSCleanup() : MachineFunctionPass(ID) {}
7441 
runOnMachineFunction__anon6956071e0511::LDTLSCleanup7442     bool runOnMachineFunction(MachineFunction &MF) override {
7443       if (skipFunction(*MF.getFunction()))
7444         return false;
7445 
7446       X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>();
7447       if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
7448         // No point folding accesses if there isn't at least two.
7449         return false;
7450       }
7451 
7452       MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
7453       return VisitNode(DT->getRootNode(), 0);
7454     }
7455 
7456     // Visit the dominator subtree rooted at Node in pre-order.
7457     // If TLSBaseAddrReg is non-null, then use that to replace any
7458     // TLS_base_addr instructions. Otherwise, create the register
7459     // when the first such instruction is seen, and then use it
7460     // as we encounter more instructions.
VisitNode__anon6956071e0511::LDTLSCleanup7461     bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
7462       MachineBasicBlock *BB = Node->getBlock();
7463       bool Changed = false;
7464 
7465       // Traverse the current block.
7466       for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
7467            ++I) {
7468         switch (I->getOpcode()) {
7469           case X86::TLS_base_addr32:
7470           case X86::TLS_base_addr64:
7471             if (TLSBaseAddrReg)
7472               I = ReplaceTLSBaseAddrCall(*I, TLSBaseAddrReg);
7473             else
7474               I = SetRegister(*I, &TLSBaseAddrReg);
7475             Changed = true;
7476             break;
7477           default:
7478             break;
7479         }
7480       }
7481 
7482       // Visit the children of this block in the dominator tree.
7483       for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
7484            I != E; ++I) {
7485         Changed |= VisitNode(*I, TLSBaseAddrReg);
7486       }
7487 
7488       return Changed;
7489     }
7490 
7491     // Replace the TLS_base_addr instruction I with a copy from
7492     // TLSBaseAddrReg, returning the new instruction.
ReplaceTLSBaseAddrCall__anon6956071e0511::LDTLSCleanup7493     MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr &I,
7494                                          unsigned TLSBaseAddrReg) {
7495       MachineFunction *MF = I.getParent()->getParent();
7496       const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
7497       const bool is64Bit = STI.is64Bit();
7498       const X86InstrInfo *TII = STI.getInstrInfo();
7499 
7500       // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
7501       MachineInstr *Copy =
7502           BuildMI(*I.getParent(), I, I.getDebugLoc(),
7503                   TII->get(TargetOpcode::COPY), is64Bit ? X86::RAX : X86::EAX)
7504               .addReg(TLSBaseAddrReg);
7505 
7506       // Erase the TLS_base_addr instruction.
7507       I.eraseFromParent();
7508 
7509       return Copy;
7510     }
7511 
7512     // Create a virtal register in *TLSBaseAddrReg, and populate it by
7513     // inserting a copy instruction after I. Returns the new instruction.
SetRegister__anon6956071e0511::LDTLSCleanup7514     MachineInstr *SetRegister(MachineInstr &I, unsigned *TLSBaseAddrReg) {
7515       MachineFunction *MF = I.getParent()->getParent();
7516       const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
7517       const bool is64Bit = STI.is64Bit();
7518       const X86InstrInfo *TII = STI.getInstrInfo();
7519 
7520       // Create a virtual register for the TLS base address.
7521       MachineRegisterInfo &RegInfo = MF->getRegInfo();
7522       *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
7523                                                       ? &X86::GR64RegClass
7524                                                       : &X86::GR32RegClass);
7525 
7526       // Insert a copy from RAX/EAX to TLSBaseAddrReg.
7527       MachineInstr *Next = I.getNextNode();
7528       MachineInstr *Copy =
7529           BuildMI(*I.getParent(), Next, I.getDebugLoc(),
7530                   TII->get(TargetOpcode::COPY), *TLSBaseAddrReg)
7531               .addReg(is64Bit ? X86::RAX : X86::EAX);
7532 
7533       return Copy;
7534     }
7535 
getPassName__anon6956071e0511::LDTLSCleanup7536     const char *getPassName() const override {
7537       return "Local Dynamic TLS Access Clean-up";
7538     }
7539 
getAnalysisUsage__anon6956071e0511::LDTLSCleanup7540     void getAnalysisUsage(AnalysisUsage &AU) const override {
7541       AU.setPreservesCFG();
7542       AU.addRequired<MachineDominatorTree>();
7543       MachineFunctionPass::getAnalysisUsage(AU);
7544     }
7545   };
7546 }
7547 
7548 char LDTLSCleanup::ID = 0;
7549 FunctionPass*
createCleanupLocalDynamicTLSPass()7550 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }
7551