• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1//===- X86InstrCompiler.td - Compiler Pseudos and Patterns -*- tablegen -*-===//
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 describes the various pseudo instructions used by the compiler,
11// as well as Pat patterns used during instruction selection.
12//
13//===----------------------------------------------------------------------===//
14
15//===----------------------------------------------------------------------===//
16// Pattern Matching Support
17
18def GetLo32XForm : SDNodeXForm<imm, [{
19  // Transformation function: get the low 32 bits.
20  return getI32Imm((unsigned)N->getZExtValue());
21}]>;
22
23def GetLo8XForm : SDNodeXForm<imm, [{
24  // Transformation function: get the low 8 bits.
25  return getI8Imm((uint8_t)N->getZExtValue());
26}]>;
27
28
29//===----------------------------------------------------------------------===//
30// Random Pseudo Instructions.
31
32// PIC base construction.  This expands to code that looks like this:
33//     call  $next_inst
34//     popl %destreg"
35let neverHasSideEffects = 1, isNotDuplicable = 1, Uses = [ESP] in
36  def MOVPC32r : Ii32<0xE8, Pseudo, (outs GR32:$reg), (ins i32imm:$label),
37                      "", []>;
38
39
40// ADJCALLSTACKDOWN/UP implicitly use/def ESP because they may be expanded into
41// a stack adjustment and the codegen must know that they may modify the stack
42// pointer before prolog-epilog rewriting occurs.
43// Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
44// sub / add which can clobber EFLAGS.
45let Defs = [ESP, EFLAGS], Uses = [ESP] in {
46def ADJCALLSTACKDOWN32 : I<0, Pseudo, (outs), (ins i32imm:$amt),
47                           "#ADJCALLSTACKDOWN",
48                           [(X86callseq_start timm:$amt)]>,
49                          Requires<[In32BitMode]>;
50def ADJCALLSTACKUP32   : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
51                           "#ADJCALLSTACKUP",
52                           [(X86callseq_end timm:$amt1, timm:$amt2)]>,
53                          Requires<[In32BitMode]>;
54}
55
56// ADJCALLSTACKDOWN/UP implicitly use/def RSP because they may be expanded into
57// a stack adjustment and the codegen must know that they may modify the stack
58// pointer before prolog-epilog rewriting occurs.
59// Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
60// sub / add which can clobber EFLAGS.
61let Defs = [RSP, EFLAGS], Uses = [RSP] in {
62def ADJCALLSTACKDOWN64 : I<0, Pseudo, (outs), (ins i32imm:$amt),
63                           "#ADJCALLSTACKDOWN",
64                           [(X86callseq_start timm:$amt)]>,
65                          Requires<[In64BitMode]>;
66def ADJCALLSTACKUP64   : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
67                           "#ADJCALLSTACKUP",
68                           [(X86callseq_end timm:$amt1, timm:$amt2)]>,
69                          Requires<[In64BitMode]>;
70}
71
72
73
74// x86-64 va_start lowering magic.
75let usesCustomInserter = 1 in {
76def VASTART_SAVE_XMM_REGS : I<0, Pseudo,
77                              (outs),
78                              (ins GR8:$al,
79                                   i64imm:$regsavefi, i64imm:$offset,
80                                   variable_ops),
81                              "#VASTART_SAVE_XMM_REGS $al, $regsavefi, $offset",
82                              [(X86vastart_save_xmm_regs GR8:$al,
83                                                         imm:$regsavefi,
84                                                         imm:$offset)]>;
85
86// The VAARG_64 pseudo-instruction takes the address of the va_list,
87// and places the address of the next argument into a register.
88let Defs = [EFLAGS] in
89def VAARG_64 : I<0, Pseudo,
90                 (outs GR64:$dst),
91                 (ins i8mem:$ap, i32imm:$size, i8imm:$mode, i32imm:$align),
92                 "#VAARG_64 $dst, $ap, $size, $mode, $align",
93                 [(set GR64:$dst,
94                    (X86vaarg64 addr:$ap, imm:$size, imm:$mode, imm:$align)),
95                  (implicit EFLAGS)]>;
96
97// Dynamic stack allocation yields a _chkstk or _alloca call for all Windows
98// targets.  These calls are needed to probe the stack when allocating more than
99// 4k bytes in one go. Touching the stack at 4K increments is necessary to
100// ensure that the guard pages used by the OS virtual memory manager are
101// allocated in correct sequence.
102// The main point of having separate instruction are extra unmodelled effects
103// (compared to ordinary calls) like stack pointer change.
104
105let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
106  def WIN_ALLOCA : I<0, Pseudo, (outs), (ins),
107                     "# dynamic stack allocation",
108                     [(X86WinAlloca)]>;
109
110// When using segmented stacks these are lowered into instructions which first
111// check if the current stacklet has enough free memory. If it does, memory is
112// allocated by bumping the stack pointer. Otherwise memory is allocated from
113// the heap.
114
115let Defs = [EAX, ESP, EFLAGS], Uses = [ESP, EAX] in
116def SEG_ALLOCA_32 : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$size),
117                      "# variable sized alloca for segmented stacks",
118                      [(set GR32:$dst,
119                         (X86SegAlloca GR32:$size))]>,
120                    Requires<[In32BitMode]>;
121
122let Defs = [RAX, RSP, EFLAGS], Uses = [RSP, RAX] in
123def SEG_ALLOCA_64 : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$size),
124                      "# variable sized alloca for segmented stacks",
125                      [(set GR64:$dst,
126                         (X86SegAlloca GR64:$size))]>,
127                    Requires<[In64BitMode]>;
128
129}
130
131
132
133//===----------------------------------------------------------------------===//
134// EH Pseudo Instructions
135//
136let isTerminator = 1, isReturn = 1, isBarrier = 1,
137    hasCtrlDep = 1, isCodeGenOnly = 1 in {
138def EH_RETURN   : I<0xC3, RawFrm, (outs), (ins GR32:$addr),
139                    "ret\t#eh_return, addr: $addr",
140                    [(X86ehret GR32:$addr)]>;
141
142}
143
144let isTerminator = 1, isReturn = 1, isBarrier = 1,
145    hasCtrlDep = 1, isCodeGenOnly = 1 in {
146def EH_RETURN64   : I<0xC3, RawFrm, (outs), (ins GR64:$addr),
147                     "ret\t#eh_return, addr: $addr",
148                     [(X86ehret GR64:$addr)]>;
149
150}
151
152//===----------------------------------------------------------------------===//
153// Alias Instructions
154//===----------------------------------------------------------------------===//
155
156// Alias instructions that map movr0 to xor.
157// FIXME: remove when we can teach regalloc that xor reg, reg is ok.
158// FIXME: Set encoding to pseudo.
159let Defs = [EFLAGS], isReMaterializable = 1, isAsCheapAsAMove = 1,
160    isCodeGenOnly = 1 in {
161def MOV8r0   : I<0x30, MRMInitReg, (outs GR8 :$dst), (ins), "",
162                 [(set GR8:$dst, 0)]>;
163
164// We want to rewrite MOV16r0 in terms of MOV32r0, because it's a smaller
165// encoding and avoids a partial-register update sometimes, but doing so
166// at isel time interferes with rematerialization in the current register
167// allocator. For now, this is rewritten when the instruction is lowered
168// to an MCInst.
169def MOV16r0   : I<0x31, MRMInitReg, (outs GR16:$dst), (ins),
170                 "",
171                 [(set GR16:$dst, 0)]>, OpSize;
172
173// FIXME: Set encoding to pseudo.
174def MOV32r0  : I<0x31, MRMInitReg, (outs GR32:$dst), (ins), "",
175                 [(set GR32:$dst, 0)]>;
176}
177
178// We want to rewrite MOV64r0 in terms of MOV32r0, because it's sometimes a
179// smaller encoding, but doing so at isel time interferes with rematerialization
180// in the current register allocator. For now, this is rewritten when the
181// instruction is lowered to an MCInst.
182// FIXME: AddedComplexity gives this a higher priority than MOV64ri32. Remove
183// when we have a better way to specify isel priority.
184let Defs = [EFLAGS], isCodeGenOnly=1,
185    AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1 in
186def MOV64r0   : I<0x31, MRMInitReg, (outs GR64:$dst), (ins), "",
187                 [(set GR64:$dst, 0)]>;
188
189// Materialize i64 constant where top 32-bits are zero. This could theoretically
190// use MOV32ri with a SUBREG_TO_REG to represent the zero-extension, however
191// that would make it more difficult to rematerialize.
192let AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1,
193    isCodeGenOnly = 1 in
194def MOV64ri64i32 : Ii32<0xB8, AddRegFrm, (outs GR64:$dst), (ins i64i32imm:$src),
195                        "", [(set GR64:$dst, i64immZExt32:$src)]>;
196
197// Use sbb to materialize carry bit.
198let Uses = [EFLAGS], Defs = [EFLAGS], isCodeGenOnly = 1 in {
199// FIXME: These are pseudo ops that should be replaced with Pat<> patterns.
200// However, Pat<> can't replicate the destination reg into the inputs of the
201// result.
202// FIXME: Change these to have encoding Pseudo when X86MCCodeEmitter replaces
203// X86CodeEmitter.
204def SETB_C8r : I<0x18, MRMInitReg, (outs GR8:$dst), (ins), "",
205                 [(set GR8:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
206def SETB_C16r : I<0x19, MRMInitReg, (outs GR16:$dst), (ins), "",
207                 [(set GR16:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>,
208                OpSize;
209def SETB_C32r : I<0x19, MRMInitReg, (outs GR32:$dst), (ins), "",
210                 [(set GR32:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
211def SETB_C64r : RI<0x19, MRMInitReg, (outs GR64:$dst), (ins), "",
212                 [(set GR64:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
213} // isCodeGenOnly
214
215
216def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
217          (SETB_C16r)>;
218def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
219          (SETB_C32r)>;
220def : Pat<(i64 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
221          (SETB_C64r)>;
222
223def : Pat<(i16 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
224          (SETB_C16r)>;
225def : Pat<(i32 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
226          (SETB_C32r)>;
227def : Pat<(i64 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
228          (SETB_C64r)>;
229
230// We canonicalize 'setb' to "(and (sbb reg,reg), 1)" on the hope that the and
231// will be eliminated and that the sbb can be extended up to a wider type.  When
232// this happens, it is great.  However, if we are left with an 8-bit sbb and an
233// and, we might as well just match it as a setb.
234def : Pat<(and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1),
235          (SETBr)>;
236
237// (add OP, SETB) -> (adc OP, 0)
238def : Pat<(add (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR8:$op),
239          (ADC8ri GR8:$op, 0)>;
240def : Pat<(add (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR32:$op),
241          (ADC32ri8 GR32:$op, 0)>;
242def : Pat<(add (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR64:$op),
243          (ADC64ri8 GR64:$op, 0)>;
244
245// (sub OP, SETB) -> (sbb OP, 0)
246def : Pat<(sub GR8:$op, (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
247          (SBB8ri GR8:$op, 0)>;
248def : Pat<(sub GR32:$op, (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
249          (SBB32ri8 GR32:$op, 0)>;
250def : Pat<(sub GR64:$op, (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
251          (SBB64ri8 GR64:$op, 0)>;
252
253// (sub OP, SETCC_CARRY) -> (adc OP, 0)
254def : Pat<(sub GR8:$op, (i8 (X86setcc_c X86_COND_B, EFLAGS))),
255          (ADC8ri GR8:$op, 0)>;
256def : Pat<(sub GR32:$op, (i32 (X86setcc_c X86_COND_B, EFLAGS))),
257          (ADC32ri8 GR32:$op, 0)>;
258def : Pat<(sub GR64:$op, (i64 (X86setcc_c X86_COND_B, EFLAGS))),
259          (ADC64ri8 GR64:$op, 0)>;
260
261//===----------------------------------------------------------------------===//
262// String Pseudo Instructions
263//
264let Defs = [ECX,EDI,ESI], Uses = [ECX,EDI,ESI], isCodeGenOnly = 1 in {
265def REP_MOVSB : I<0xA4, RawFrm, (outs), (ins), "{rep;movsb|rep movsb}",
266                  [(X86rep_movs i8)]>, REP;
267def REP_MOVSW : I<0xA5, RawFrm, (outs), (ins), "{rep;movsw|rep movsw}",
268                  [(X86rep_movs i16)]>, REP, OpSize;
269def REP_MOVSD : I<0xA5, RawFrm, (outs), (ins), "{rep;movsl|rep movsd}",
270                  [(X86rep_movs i32)]>, REP;
271}
272
273let Defs = [RCX,RDI,RSI], Uses = [RCX,RDI,RSI], isCodeGenOnly = 1 in
274def REP_MOVSQ : RI<0xA5, RawFrm, (outs), (ins), "{rep;movsq|rep movsq}",
275                   [(X86rep_movs i64)]>, REP;
276
277
278// FIXME: Should use "(X86rep_stos AL)" as the pattern.
279let Defs = [ECX,EDI], Uses = [AL,ECX,EDI], isCodeGenOnly = 1 in
280def REP_STOSB : I<0xAA, RawFrm, (outs), (ins), "{rep;stosb|rep stosb}",
281                  [(X86rep_stos i8)]>, REP;
282let Defs = [ECX,EDI], Uses = [AX,ECX,EDI], isCodeGenOnly = 1 in
283def REP_STOSW : I<0xAB, RawFrm, (outs), (ins), "{rep;stosw|rep stosw}",
284                  [(X86rep_stos i16)]>, REP, OpSize;
285let Defs = [ECX,EDI], Uses = [EAX,ECX,EDI], isCodeGenOnly = 1 in
286def REP_STOSD : I<0xAB, RawFrm, (outs), (ins), "{rep;stosl|rep stosd}",
287                  [(X86rep_stos i32)]>, REP;
288
289let Defs = [RCX,RDI], Uses = [RAX,RCX,RDI], isCodeGenOnly = 1 in
290def REP_STOSQ : RI<0xAB, RawFrm, (outs), (ins), "{rep;stosq|rep stosq}",
291                   [(X86rep_stos i64)]>, REP;
292
293
294//===----------------------------------------------------------------------===//
295// Thread Local Storage Instructions
296//
297
298// ELF TLS Support
299// All calls clobber the non-callee saved registers. ESP is marked as
300// a use to prevent stack-pointer assignments that appear immediately
301// before calls from potentially appearing dead.
302let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0,
303            MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
304            XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
305            XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
306    Uses = [ESP] in
307def TLS_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
308                  "# TLS_addr32",
309                  [(X86tlsaddr tls32addr:$sym)]>,
310                  Requires<[In32BitMode]>;
311
312// All calls clobber the non-callee saved registers. RSP is marked as
313// a use to prevent stack-pointer assignments that appear immediately
314// before calls from potentially appearing dead.
315let Defs = [RAX, RCX, RDX, RSI, RDI, R8, R9, R10, R11,
316            FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0, ST1,
317            MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
318            XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
319            XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
320    Uses = [RSP] in
321def TLS_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
322                   "# TLS_addr64",
323                  [(X86tlsaddr tls64addr:$sym)]>,
324                  Requires<[In64BitMode]>;
325
326// Darwin TLS Support
327// For i386, the address of the thunk is passed on the stack, on return the
328// address of the variable is in %eax.  %ecx is trashed during the function
329// call.  All other registers are preserved.
330let Defs = [EAX, ECX, EFLAGS],
331    Uses = [ESP],
332    usesCustomInserter = 1 in
333def TLSCall_32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
334                "# TLSCall_32",
335                [(X86TLSCall addr:$sym)]>,
336                Requires<[In32BitMode]>;
337
338// For x86_64, the address of the thunk is passed in %rdi, on return
339// the address of the variable is in %rax.  All other registers are preserved.
340let Defs = [RAX, EFLAGS],
341    Uses = [RSP, RDI],
342    usesCustomInserter = 1 in
343def TLSCall_64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
344                  "# TLSCall_64",
345                  [(X86TLSCall addr:$sym)]>,
346                  Requires<[In64BitMode]>;
347
348
349//===----------------------------------------------------------------------===//
350// Conditional Move Pseudo Instructions
351
352// X86 doesn't have 8-bit conditional moves. Use a customInserter to
353// emit control flow. An alternative to this is to mark i8 SELECT as Promote,
354// however that requires promoting the operands, and can induce additional
355// i8 register pressure.
356let usesCustomInserter = 1, Uses = [EFLAGS] in {
357def CMOV_GR8 : I<0, Pseudo,
358                 (outs GR8:$dst), (ins GR8:$src1, GR8:$src2, i8imm:$cond),
359                 "#CMOV_GR8 PSEUDO!",
360                 [(set GR8:$dst, (X86cmov GR8:$src1, GR8:$src2,
361                                          imm:$cond, EFLAGS))]>;
362
363let Predicates = [NoCMov] in {
364def CMOV_GR32 : I<0, Pseudo,
365                    (outs GR32:$dst), (ins GR32:$src1, GR32:$src2, i8imm:$cond),
366                    "#CMOV_GR32* PSEUDO!",
367                    [(set GR32:$dst,
368                      (X86cmov GR32:$src1, GR32:$src2, imm:$cond, EFLAGS))]>;
369def CMOV_GR16 : I<0, Pseudo,
370                    (outs GR16:$dst), (ins GR16:$src1, GR16:$src2, i8imm:$cond),
371                    "#CMOV_GR16* PSEUDO!",
372                    [(set GR16:$dst,
373                      (X86cmov GR16:$src1, GR16:$src2, imm:$cond, EFLAGS))]>;
374def CMOV_RFP32 : I<0, Pseudo,
375                    (outs RFP32:$dst),
376                    (ins RFP32:$src1, RFP32:$src2, i8imm:$cond),
377                    "#CMOV_RFP32 PSEUDO!",
378                    [(set RFP32:$dst,
379                      (X86cmov RFP32:$src1, RFP32:$src2, imm:$cond,
380                                                  EFLAGS))]>;
381def CMOV_RFP64 : I<0, Pseudo,
382                    (outs RFP64:$dst),
383                    (ins RFP64:$src1, RFP64:$src2, i8imm:$cond),
384                    "#CMOV_RFP64 PSEUDO!",
385                    [(set RFP64:$dst,
386                      (X86cmov RFP64:$src1, RFP64:$src2, imm:$cond,
387                                                  EFLAGS))]>;
388def CMOV_RFP80 : I<0, Pseudo,
389                    (outs RFP80:$dst),
390                    (ins RFP80:$src1, RFP80:$src2, i8imm:$cond),
391                    "#CMOV_RFP80 PSEUDO!",
392                    [(set RFP80:$dst,
393                      (X86cmov RFP80:$src1, RFP80:$src2, imm:$cond,
394                                                  EFLAGS))]>;
395} // Predicates = [NoCMov]
396} // UsesCustomInserter = 1, Uses = [EFLAGS]
397
398
399//===----------------------------------------------------------------------===//
400// Atomic Instruction Pseudo Instructions
401//===----------------------------------------------------------------------===//
402
403// Atomic exchange, and, or, xor
404let Constraints = "$val = $dst", Defs = [EFLAGS],
405                  usesCustomInserter = 1 in {
406
407def ATOMAND8 : I<0, Pseudo, (outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
408               "#ATOMAND8 PSEUDO!",
409               [(set GR8:$dst, (atomic_load_and_8 addr:$ptr, GR8:$val))]>;
410def ATOMOR8 : I<0, Pseudo, (outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
411               "#ATOMOR8 PSEUDO!",
412               [(set GR8:$dst, (atomic_load_or_8 addr:$ptr, GR8:$val))]>;
413def ATOMXOR8 : I<0, Pseudo,(outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
414               "#ATOMXOR8 PSEUDO!",
415               [(set GR8:$dst, (atomic_load_xor_8 addr:$ptr, GR8:$val))]>;
416def ATOMNAND8 : I<0, Pseudo,(outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
417               "#ATOMNAND8 PSEUDO!",
418               [(set GR8:$dst, (atomic_load_nand_8 addr:$ptr, GR8:$val))]>;
419
420def ATOMAND16 : I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
421               "#ATOMAND16 PSEUDO!",
422               [(set GR16:$dst, (atomic_load_and_16 addr:$ptr, GR16:$val))]>;
423def ATOMOR16 : I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
424               "#ATOMOR16 PSEUDO!",
425               [(set GR16:$dst, (atomic_load_or_16 addr:$ptr, GR16:$val))]>;
426def ATOMXOR16 : I<0, Pseudo,(outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
427               "#ATOMXOR16 PSEUDO!",
428               [(set GR16:$dst, (atomic_load_xor_16 addr:$ptr, GR16:$val))]>;
429def ATOMNAND16 : I<0, Pseudo,(outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
430               "#ATOMNAND16 PSEUDO!",
431               [(set GR16:$dst, (atomic_load_nand_16 addr:$ptr, GR16:$val))]>;
432def ATOMMIN16: I<0, Pseudo, (outs GR16:$dst), (ins i16mem:$ptr, GR16:$val),
433               "#ATOMMIN16 PSEUDO!",
434               [(set GR16:$dst, (atomic_load_min_16 addr:$ptr, GR16:$val))]>;
435def ATOMMAX16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
436               "#ATOMMAX16 PSEUDO!",
437               [(set GR16:$dst, (atomic_load_max_16 addr:$ptr, GR16:$val))]>;
438def ATOMUMIN16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
439               "#ATOMUMIN16 PSEUDO!",
440               [(set GR16:$dst, (atomic_load_umin_16 addr:$ptr, GR16:$val))]>;
441def ATOMUMAX16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
442               "#ATOMUMAX16 PSEUDO!",
443               [(set GR16:$dst, (atomic_load_umax_16 addr:$ptr, GR16:$val))]>;
444
445
446def ATOMAND32 : I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
447               "#ATOMAND32 PSEUDO!",
448               [(set GR32:$dst, (atomic_load_and_32 addr:$ptr, GR32:$val))]>;
449def ATOMOR32 : I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
450               "#ATOMOR32 PSEUDO!",
451               [(set GR32:$dst, (atomic_load_or_32 addr:$ptr, GR32:$val))]>;
452def ATOMXOR32 : I<0, Pseudo,(outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
453               "#ATOMXOR32 PSEUDO!",
454               [(set GR32:$dst, (atomic_load_xor_32 addr:$ptr, GR32:$val))]>;
455def ATOMNAND32 : I<0, Pseudo,(outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
456               "#ATOMNAND32 PSEUDO!",
457               [(set GR32:$dst, (atomic_load_nand_32 addr:$ptr, GR32:$val))]>;
458def ATOMMIN32: I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$ptr, GR32:$val),
459               "#ATOMMIN32 PSEUDO!",
460               [(set GR32:$dst, (atomic_load_min_32 addr:$ptr, GR32:$val))]>;
461def ATOMMAX32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
462               "#ATOMMAX32 PSEUDO!",
463               [(set GR32:$dst, (atomic_load_max_32 addr:$ptr, GR32:$val))]>;
464def ATOMUMIN32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
465               "#ATOMUMIN32 PSEUDO!",
466               [(set GR32:$dst, (atomic_load_umin_32 addr:$ptr, GR32:$val))]>;
467def ATOMUMAX32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
468               "#ATOMUMAX32 PSEUDO!",
469               [(set GR32:$dst, (atomic_load_umax_32 addr:$ptr, GR32:$val))]>;
470
471
472
473def ATOMAND64 : I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
474               "#ATOMAND64 PSEUDO!",
475               [(set GR64:$dst, (atomic_load_and_64 addr:$ptr, GR64:$val))]>;
476def ATOMOR64 : I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
477               "#ATOMOR64 PSEUDO!",
478               [(set GR64:$dst, (atomic_load_or_64 addr:$ptr, GR64:$val))]>;
479def ATOMXOR64 : I<0, Pseudo,(outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
480               "#ATOMXOR64 PSEUDO!",
481               [(set GR64:$dst, (atomic_load_xor_64 addr:$ptr, GR64:$val))]>;
482def ATOMNAND64 : I<0, Pseudo,(outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
483               "#ATOMNAND64 PSEUDO!",
484               [(set GR64:$dst, (atomic_load_nand_64 addr:$ptr, GR64:$val))]>;
485def ATOMMIN64: I<0, Pseudo, (outs GR64:$dst), (ins i64mem:$ptr, GR64:$val),
486               "#ATOMMIN64 PSEUDO!",
487               [(set GR64:$dst, (atomic_load_min_64 addr:$ptr, GR64:$val))]>;
488def ATOMMAX64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
489               "#ATOMMAX64 PSEUDO!",
490               [(set GR64:$dst, (atomic_load_max_64 addr:$ptr, GR64:$val))]>;
491def ATOMUMIN64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
492               "#ATOMUMIN64 PSEUDO!",
493               [(set GR64:$dst, (atomic_load_umin_64 addr:$ptr, GR64:$val))]>;
494def ATOMUMAX64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
495               "#ATOMUMAX64 PSEUDO!",
496               [(set GR64:$dst, (atomic_load_umax_64 addr:$ptr, GR64:$val))]>;
497}
498
499let Constraints = "$val1 = $dst1, $val2 = $dst2",
500                  Defs = [EFLAGS, EAX, EBX, ECX, EDX],
501                  Uses = [EAX, EBX, ECX, EDX],
502                  mayLoad = 1, mayStore = 1,
503                  usesCustomInserter = 1 in {
504def ATOMAND6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
505                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
506               "#ATOMAND6432 PSEUDO!", []>;
507def ATOMOR6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
508                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
509               "#ATOMOR6432 PSEUDO!", []>;
510def ATOMXOR6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
511                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
512               "#ATOMXOR6432 PSEUDO!", []>;
513def ATOMNAND6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
514                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
515               "#ATOMNAND6432 PSEUDO!", []>;
516def ATOMADD6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
517                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
518               "#ATOMADD6432 PSEUDO!", []>;
519def ATOMSUB6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
520                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
521               "#ATOMSUB6432 PSEUDO!", []>;
522def ATOMSWAP6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
523                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
524               "#ATOMSWAP6432 PSEUDO!", []>;
525}
526
527//===----------------------------------------------------------------------===//
528// Normal-Instructions-With-Lock-Prefix Pseudo Instructions
529//===----------------------------------------------------------------------===//
530
531// FIXME: Use normal instructions and add lock prefix dynamically.
532
533// Memory barriers
534
535// TODO: Get this to fold the constant into the instruction.
536let isCodeGenOnly = 1 in
537def OR32mrLocked  : I<0x09, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$zero),
538                      "lock\n\t"
539                      "or{l}\t{$zero, $dst|$dst, $zero}",
540                      []>, Requires<[In32BitMode]>, LOCK;
541
542let hasSideEffects = 1 in
543def Int_MemBarrier : I<0, Pseudo, (outs), (ins),
544                     "#MEMBARRIER",
545                     [(X86MemBarrier)]>;
546
547// TODO: Get this to fold the constant into the instruction.
548let hasSideEffects = 1, Defs = [ESP], isCodeGenOnly = 1 in
549def Int_MemBarrierNoSSE64  : RI<0x09, MRM1r, (outs), (ins GR64:$zero),
550                           "lock\n\t"
551                           "or{q}\t{$zero, (%rsp)|(%rsp), $zero}",
552                           [(X86MemBarrierNoSSE GR64:$zero)]>,
553                           Requires<[In64BitMode]>, LOCK;
554
555
556// RegOpc corresponds to the mr version of the instruction
557// ImmOpc corresponds to the mi version of the instruction
558// ImmOpc8 corresponds to the mi8 version of the instruction
559// ImmMod corresponds to the instruction format of the mi and mi8 versions
560multiclass LOCK_ArithBinOp<bits<8> RegOpc, bits<8> ImmOpc, bits<8> ImmOpc8,
561                           Format ImmMod, string mnemonic> {
562let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1 in {
563
564def #NAME#8mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
565                   RegOpc{3}, RegOpc{2}, RegOpc{1}, 0 },
566                   MRMDestMem, (outs), (ins i8mem:$dst, GR8:$src2),
567                   !strconcat("lock\n\t", mnemonic, "{b}\t",
568                              "{$src2, $dst|$dst, $src2}"),
569                   []>, LOCK;
570def #NAME#16mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
571                    RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
572                    MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
573                    !strconcat("lock\n\t", mnemonic, "{w}\t",
574                               "{$src2, $dst|$dst, $src2}"),
575                    []>, OpSize, LOCK;
576def #NAME#32mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
577                    RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
578                    MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
579                    !strconcat("lock\n\t", mnemonic, "{l}\t",
580                               "{$src2, $dst|$dst, $src2}"),
581                    []>, LOCK;
582def #NAME#64mr : RI<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
583                     RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
584                     MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
585                     !strconcat("lock\n\t", mnemonic, "{q}\t",
586                                "{$src2, $dst|$dst, $src2}"),
587                     []>, LOCK;
588
589def #NAME#8mi : Ii8<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
590                     ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 0 },
591                     ImmMod, (outs), (ins i8mem :$dst, i8imm :$src2),
592                     !strconcat("lock\n\t", mnemonic, "{b}\t",
593                                "{$src2, $dst|$dst, $src2}"),
594                     []>, LOCK;
595
596def #NAME#16mi : Ii16<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
597                       ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
598                       ImmMod, (outs), (ins i16mem :$dst, i16imm :$src2),
599                       !strconcat("lock\n\t", mnemonic, "{w}\t",
600                                  "{$src2, $dst|$dst, $src2}"),
601                       []>, LOCK;
602
603def #NAME#32mi : Ii32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
604                       ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
605                       ImmMod, (outs), (ins i32mem :$dst, i32imm :$src2),
606                       !strconcat("lock\n\t", mnemonic, "{l}\t",
607                                  "{$src2, $dst|$dst, $src2}"),
608                       []>, LOCK;
609
610def #NAME#64mi32 : RIi32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
611                          ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
612                          ImmMod, (outs), (ins i64mem :$dst, i64i32imm :$src2),
613                          !strconcat("lock\n\t", mnemonic, "{q}\t",
614                                     "{$src2, $dst|$dst, $src2}"),
615                          []>, LOCK;
616
617def #NAME#16mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
618                       ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
619                       ImmMod, (outs), (ins i16mem :$dst, i16i8imm :$src2),
620                       !strconcat("lock\n\t", mnemonic, "{w}\t",
621                                  "{$src2, $dst|$dst, $src2}"),
622                       []>, LOCK;
623def #NAME#32mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
624                       ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
625                       ImmMod, (outs), (ins i32mem :$dst, i32i8imm :$src2),
626                       !strconcat("lock\n\t", mnemonic, "{l}\t",
627                                  "{$src2, $dst|$dst, $src2}"),
628                       []>, LOCK;
629def #NAME#64mi8 : RIi8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
630                        ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
631                        ImmMod, (outs), (ins i64mem :$dst, i64i8imm :$src2),
632                        !strconcat("lock\n\t", mnemonic, "{q}\t",
633                                   "{$src2, $dst|$dst, $src2}"),
634                        []>, LOCK;
635
636}
637
638}
639
640defm LOCK_ADD : LOCK_ArithBinOp<0x00, 0x80, 0x83, MRM0m, "add">;
641defm LOCK_SUB : LOCK_ArithBinOp<0x28, 0x80, 0x83, MRM5m, "sub">;
642defm LOCK_OR  : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM1m, "or">;
643defm LOCK_AND : LOCK_ArithBinOp<0x20, 0x80, 0x83, MRM4m, "and">;
644defm LOCK_XOR : LOCK_ArithBinOp<0x30, 0x80, 0x83, MRM6m, "xor">;
645
646// Optimized codegen when the non-memory output is not used.
647let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1 in {
648
649def LOCK_INC8m  : I<0xFE, MRM0m, (outs), (ins i8mem :$dst),
650                    "lock\n\t"
651                    "inc{b}\t$dst", []>, LOCK;
652def LOCK_INC16m : I<0xFF, MRM0m, (outs), (ins i16mem:$dst),
653                    "lock\n\t"
654                    "inc{w}\t$dst", []>, OpSize, LOCK;
655def LOCK_INC32m : I<0xFF, MRM0m, (outs), (ins i32mem:$dst),
656                    "lock\n\t"
657                    "inc{l}\t$dst", []>, LOCK;
658def LOCK_INC64m : RI<0xFF, MRM0m, (outs), (ins i64mem:$dst),
659                     "lock\n\t"
660                     "inc{q}\t$dst", []>, LOCK;
661
662def LOCK_DEC8m  : I<0xFE, MRM1m, (outs), (ins i8mem :$dst),
663                    "lock\n\t"
664                    "dec{b}\t$dst", []>, LOCK;
665def LOCK_DEC16m : I<0xFF, MRM1m, (outs), (ins i16mem:$dst),
666                    "lock\n\t"
667                    "dec{w}\t$dst", []>, OpSize, LOCK;
668def LOCK_DEC32m : I<0xFF, MRM1m, (outs), (ins i32mem:$dst),
669                    "lock\n\t"
670                    "dec{l}\t$dst", []>, LOCK;
671def LOCK_DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst),
672                      "lock\n\t"
673                      "dec{q}\t$dst", []>, LOCK;
674}
675
676// Atomic compare and swap.
677let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
678    isCodeGenOnly = 1 in
679def LCMPXCHG8B : I<0xC7, MRM1m, (outs), (ins i64mem:$ptr),
680               "lock\n\t"
681               "cmpxchg8b\t$ptr",
682               [(X86cas8 addr:$ptr)]>, TB, LOCK;
683
684let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RBX, RCX, RDX],
685    isCodeGenOnly = 1 in
686def LCMPXCHG16B : RI<0xC7, MRM1m, (outs), (ins i128mem:$ptr),
687                    "lock\n\t"
688                    "cmpxchg16b\t$ptr",
689                    [(X86cas16 addr:$ptr)]>, TB, LOCK,
690                    Requires<[HasCmpxchg16b]>;
691
692let Defs = [AL, EFLAGS], Uses = [AL], isCodeGenOnly = 1 in {
693def LCMPXCHG8 : I<0xB0, MRMDestMem, (outs), (ins i8mem:$ptr, GR8:$swap),
694               "lock\n\t"
695               "cmpxchg{b}\t{$swap, $ptr|$ptr, $swap}",
696               [(X86cas addr:$ptr, GR8:$swap, 1)]>, TB, LOCK;
697}
698
699let Defs = [AX, EFLAGS], Uses = [AX], isCodeGenOnly = 1 in {
700def LCMPXCHG16 : I<0xB1, MRMDestMem, (outs), (ins i16mem:$ptr, GR16:$swap),
701               "lock\n\t"
702               "cmpxchg{w}\t{$swap, $ptr|$ptr, $swap}",
703               [(X86cas addr:$ptr, GR16:$swap, 2)]>, TB, OpSize, LOCK;
704}
705
706let Defs = [EAX, EFLAGS], Uses = [EAX], isCodeGenOnly = 1 in {
707def LCMPXCHG32 : I<0xB1, MRMDestMem, (outs), (ins i32mem:$ptr, GR32:$swap),
708               "lock\n\t"
709               "cmpxchg{l}\t{$swap, $ptr|$ptr, $swap}",
710               [(X86cas addr:$ptr, GR32:$swap, 4)]>, TB, LOCK;
711}
712
713let Defs = [RAX, EFLAGS], Uses = [RAX], isCodeGenOnly = 1 in {
714def LCMPXCHG64 : RI<0xB1, MRMDestMem, (outs), (ins i64mem:$ptr, GR64:$swap),
715               "lock\n\t"
716               "cmpxchg{q}\t{$swap, $ptr|$ptr, $swap}",
717               [(X86cas addr:$ptr, GR64:$swap, 8)]>, TB, LOCK;
718}
719
720// Atomic exchange and add
721let Constraints = "$val = $dst", Defs = [EFLAGS], isCodeGenOnly = 1 in {
722def LXADD8  : I<0xC0, MRMSrcMem, (outs GR8:$dst), (ins GR8:$val, i8mem:$ptr),
723               "lock\n\t"
724               "xadd{b}\t{$val, $ptr|$ptr, $val}",
725               [(set GR8:$dst, (atomic_load_add_8 addr:$ptr, GR8:$val))]>,
726                TB, LOCK;
727def LXADD16 : I<0xC1, MRMSrcMem, (outs GR16:$dst), (ins GR16:$val, i16mem:$ptr),
728               "lock\n\t"
729               "xadd{w}\t{$val, $ptr|$ptr, $val}",
730               [(set GR16:$dst, (atomic_load_add_16 addr:$ptr, GR16:$val))]>,
731                TB, OpSize, LOCK;
732def LXADD32 : I<0xC1, MRMSrcMem, (outs GR32:$dst), (ins GR32:$val, i32mem:$ptr),
733               "lock\n\t"
734               "xadd{l}\t{$val, $ptr|$ptr, $val}",
735               [(set GR32:$dst, (atomic_load_add_32 addr:$ptr, GR32:$val))]>,
736                TB, LOCK;
737def LXADD64 : RI<0xC1, MRMSrcMem, (outs GR64:$dst), (ins GR64:$val,i64mem:$ptr),
738               "lock\n\t"
739               "xadd{q}\t{$val, $ptr|$ptr, $val}",
740               [(set GR64:$dst, (atomic_load_add_64 addr:$ptr, GR64:$val))]>,
741                TB, LOCK;
742}
743
744def ACQUIRE_MOV8rm  : I<0, Pseudo, (outs GR8 :$dst), (ins i8mem :$src),
745                      "#ACQUIRE_MOV PSEUDO!",
746                      [(set GR8:$dst,  (atomic_load_8  addr:$src))]>;
747def ACQUIRE_MOV16rm : I<0, Pseudo, (outs GR16:$dst), (ins i16mem:$src),
748                      "#ACQUIRE_MOV PSEUDO!",
749                      [(set GR16:$dst, (atomic_load_16 addr:$src))]>;
750def ACQUIRE_MOV32rm : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$src),
751                      "#ACQUIRE_MOV PSEUDO!",
752                      [(set GR32:$dst, (atomic_load_32 addr:$src))]>;
753def ACQUIRE_MOV64rm : I<0, Pseudo, (outs GR64:$dst), (ins i64mem:$src),
754                      "#ACQUIRE_MOV PSEUDO!",
755                      [(set GR64:$dst, (atomic_load_64 addr:$src))]>;
756
757def RELEASE_MOV8mr  : I<0, Pseudo, (outs), (ins i8mem :$dst, GR8 :$src),
758                        "#RELEASE_MOV PSEUDO!",
759                        [(atomic_store_8  addr:$dst, GR8 :$src)]>;
760def RELEASE_MOV16mr : I<0, Pseudo, (outs), (ins i16mem:$dst, GR16:$src),
761                        "#RELEASE_MOV PSEUDO!",
762                        [(atomic_store_16 addr:$dst, GR16:$src)]>;
763def RELEASE_MOV32mr : I<0, Pseudo, (outs), (ins i32mem:$dst, GR32:$src),
764                        "#RELEASE_MOV PSEUDO!",
765                        [(atomic_store_32 addr:$dst, GR32:$src)]>;
766def RELEASE_MOV64mr : I<0, Pseudo, (outs), (ins i64mem:$dst, GR64:$src),
767                        "#RELEASE_MOV PSEUDO!",
768                        [(atomic_store_64 addr:$dst, GR64:$src)]>;
769
770//===----------------------------------------------------------------------===//
771// Conditional Move Pseudo Instructions.
772//===----------------------------------------------------------------------===//
773
774
775// CMOV* - Used to implement the SSE SELECT DAG operation.  Expanded after
776// instruction selection into a branch sequence.
777let Uses = [EFLAGS], usesCustomInserter = 1 in {
778  def CMOV_FR32 : I<0, Pseudo,
779                    (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
780                    "#CMOV_FR32 PSEUDO!",
781                    [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
782                                                  EFLAGS))]>;
783  def CMOV_FR64 : I<0, Pseudo,
784                    (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
785                    "#CMOV_FR64 PSEUDO!",
786                    [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
787                                                  EFLAGS))]>;
788  def CMOV_V4F32 : I<0, Pseudo,
789                    (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
790                    "#CMOV_V4F32 PSEUDO!",
791                    [(set VR128:$dst,
792                      (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
793                                          EFLAGS)))]>;
794  def CMOV_V2F64 : I<0, Pseudo,
795                    (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
796                    "#CMOV_V2F64 PSEUDO!",
797                    [(set VR128:$dst,
798                      (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
799                                          EFLAGS)))]>;
800  def CMOV_V2I64 : I<0, Pseudo,
801                    (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
802                    "#CMOV_V2I64 PSEUDO!",
803                    [(set VR128:$dst,
804                      (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
805                                          EFLAGS)))]>;
806  def CMOV_V8F32 : I<0, Pseudo,
807                    (outs VR256:$dst), (ins VR256:$t, VR256:$f, i8imm:$cond),
808                    "#CMOV_V8F32 PSEUDO!",
809                    [(set VR256:$dst,
810                      (v8f32 (X86cmov VR256:$t, VR256:$f, imm:$cond,
811                                          EFLAGS)))]>;
812  def CMOV_V4F64 : I<0, Pseudo,
813                    (outs VR256:$dst), (ins VR256:$t, VR256:$f, i8imm:$cond),
814                    "#CMOV_V4F64 PSEUDO!",
815                    [(set VR256:$dst,
816                      (v4f64 (X86cmov VR256:$t, VR256:$f, imm:$cond,
817                                          EFLAGS)))]>;
818  def CMOV_V4I64 : I<0, Pseudo,
819                    (outs VR256:$dst), (ins VR256:$t, VR256:$f, i8imm:$cond),
820                    "#CMOV_V4I64 PSEUDO!",
821                    [(set VR256:$dst,
822                      (v4i64 (X86cmov VR256:$t, VR256:$f, imm:$cond,
823                                          EFLAGS)))]>;
824}
825
826
827//===----------------------------------------------------------------------===//
828// DAG Pattern Matching Rules
829//===----------------------------------------------------------------------===//
830
831// ConstantPool GlobalAddress, ExternalSymbol, and JumpTable
832def : Pat<(i32 (X86Wrapper tconstpool  :$dst)), (MOV32ri tconstpool  :$dst)>;
833def : Pat<(i32 (X86Wrapper tjumptable  :$dst)), (MOV32ri tjumptable  :$dst)>;
834def : Pat<(i32 (X86Wrapper tglobaltlsaddr:$dst)),(MOV32ri tglobaltlsaddr:$dst)>;
835def : Pat<(i32 (X86Wrapper tglobaladdr :$dst)), (MOV32ri tglobaladdr :$dst)>;
836def : Pat<(i32 (X86Wrapper texternalsym:$dst)), (MOV32ri texternalsym:$dst)>;
837def : Pat<(i32 (X86Wrapper tblockaddress:$dst)), (MOV32ri tblockaddress:$dst)>;
838
839def : Pat<(add GR32:$src1, (X86Wrapper tconstpool:$src2)),
840          (ADD32ri GR32:$src1, tconstpool:$src2)>;
841def : Pat<(add GR32:$src1, (X86Wrapper tjumptable:$src2)),
842          (ADD32ri GR32:$src1, tjumptable:$src2)>;
843def : Pat<(add GR32:$src1, (X86Wrapper tglobaladdr :$src2)),
844          (ADD32ri GR32:$src1, tglobaladdr:$src2)>;
845def : Pat<(add GR32:$src1, (X86Wrapper texternalsym:$src2)),
846          (ADD32ri GR32:$src1, texternalsym:$src2)>;
847def : Pat<(add GR32:$src1, (X86Wrapper tblockaddress:$src2)),
848          (ADD32ri GR32:$src1, tblockaddress:$src2)>;
849
850def : Pat<(store (i32 (X86Wrapper tglobaladdr:$src)), addr:$dst),
851          (MOV32mi addr:$dst, tglobaladdr:$src)>;
852def : Pat<(store (i32 (X86Wrapper texternalsym:$src)), addr:$dst),
853          (MOV32mi addr:$dst, texternalsym:$src)>;
854def : Pat<(store (i32 (X86Wrapper tblockaddress:$src)), addr:$dst),
855          (MOV32mi addr:$dst, tblockaddress:$src)>;
856
857
858
859// ConstantPool GlobalAddress, ExternalSymbol, and JumpTable when not in small
860// code model mode, should use 'movabs'.  FIXME: This is really a hack, the
861//  'movabs' predicate should handle this sort of thing.
862def : Pat<(i64 (X86Wrapper tconstpool  :$dst)),
863          (MOV64ri tconstpool  :$dst)>, Requires<[FarData]>;
864def : Pat<(i64 (X86Wrapper tjumptable  :$dst)),
865          (MOV64ri tjumptable  :$dst)>, Requires<[FarData]>;
866def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
867          (MOV64ri tglobaladdr :$dst)>, Requires<[FarData]>;
868def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
869          (MOV64ri texternalsym:$dst)>, Requires<[FarData]>;
870def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
871          (MOV64ri tblockaddress:$dst)>, Requires<[FarData]>;
872
873// In static codegen with small code model, we can get the address of a label
874// into a register with 'movl'.  FIXME: This is a hack, the 'imm' predicate of
875// the MOV64ri64i32 should accept these.
876def : Pat<(i64 (X86Wrapper tconstpool  :$dst)),
877          (MOV64ri64i32 tconstpool  :$dst)>, Requires<[SmallCode]>;
878def : Pat<(i64 (X86Wrapper tjumptable  :$dst)),
879          (MOV64ri64i32 tjumptable  :$dst)>, Requires<[SmallCode]>;
880def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
881          (MOV64ri64i32 tglobaladdr :$dst)>, Requires<[SmallCode]>;
882def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
883          (MOV64ri64i32 texternalsym:$dst)>, Requires<[SmallCode]>;
884def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
885          (MOV64ri64i32 tblockaddress:$dst)>, Requires<[SmallCode]>;
886
887// In kernel code model, we can get the address of a label
888// into a register with 'movq'.  FIXME: This is a hack, the 'imm' predicate of
889// the MOV64ri32 should accept these.
890def : Pat<(i64 (X86Wrapper tconstpool  :$dst)),
891          (MOV64ri32 tconstpool  :$dst)>, Requires<[KernelCode]>;
892def : Pat<(i64 (X86Wrapper tjumptable  :$dst)),
893          (MOV64ri32 tjumptable  :$dst)>, Requires<[KernelCode]>;
894def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
895          (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
896def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
897          (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
898def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
899          (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
900
901// If we have small model and -static mode, it is safe to store global addresses
902// directly as immediates.  FIXME: This is really a hack, the 'imm' predicate
903// for MOV64mi32 should handle this sort of thing.
904def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
905          (MOV64mi32 addr:$dst, tconstpool:$src)>,
906          Requires<[NearData, IsStatic]>;
907def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
908          (MOV64mi32 addr:$dst, tjumptable:$src)>,
909          Requires<[NearData, IsStatic]>;
910def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
911          (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
912          Requires<[NearData, IsStatic]>;
913def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
914          (MOV64mi32 addr:$dst, texternalsym:$src)>,
915          Requires<[NearData, IsStatic]>;
916def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
917          (MOV64mi32 addr:$dst, tblockaddress:$src)>,
918          Requires<[NearData, IsStatic]>;
919
920
921
922// Calls
923
924// tls has some funny stuff here...
925// This corresponds to movabs $foo@tpoff, %rax
926def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
927          (MOV64ri tglobaltlsaddr :$dst)>;
928// This corresponds to add $foo@tpoff, %rax
929def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
930          (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
931// This corresponds to mov foo@tpoff(%rbx), %eax
932def : Pat<(load (i64 (X86Wrapper tglobaltlsaddr :$dst))),
933          (MOV64rm tglobaltlsaddr :$dst)>;
934
935
936// Direct PC relative function call for small code model. 32-bit displacement
937// sign extended to 64-bit.
938def : Pat<(X86call (i64 tglobaladdr:$dst)),
939          (CALL64pcrel32 tglobaladdr:$dst)>, Requires<[NotWin64]>;
940def : Pat<(X86call (i64 texternalsym:$dst)),
941          (CALL64pcrel32 texternalsym:$dst)>, Requires<[NotWin64]>;
942
943def : Pat<(X86call (i64 tglobaladdr:$dst)),
944          (WINCALL64pcrel32 tglobaladdr:$dst)>, Requires<[IsWin64]>;
945def : Pat<(X86call (i64 texternalsym:$dst)),
946          (WINCALL64pcrel32 texternalsym:$dst)>, Requires<[IsWin64]>;
947
948// tailcall stuff
949def : Pat<(X86tcret GR32_TC:$dst, imm:$off),
950          (TCRETURNri GR32_TC:$dst, imm:$off)>,
951          Requires<[In32BitMode]>;
952
953// FIXME: This is disabled for 32-bit PIC mode because the global base
954// register which is part of the address mode may be assigned a
955// callee-saved register.
956def : Pat<(X86tcret (load addr:$dst), imm:$off),
957          (TCRETURNmi addr:$dst, imm:$off)>,
958          Requires<[In32BitMode, IsNotPIC]>;
959
960def : Pat<(X86tcret (i32 tglobaladdr:$dst), imm:$off),
961          (TCRETURNdi texternalsym:$dst, imm:$off)>,
962          Requires<[In32BitMode]>;
963
964def : Pat<(X86tcret (i32 texternalsym:$dst), imm:$off),
965          (TCRETURNdi texternalsym:$dst, imm:$off)>,
966          Requires<[In32BitMode]>;
967
968def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
969          (TCRETURNri64 ptr_rc_tailcall:$dst, imm:$off)>,
970          Requires<[In64BitMode]>;
971
972def : Pat<(X86tcret (load addr:$dst), imm:$off),
973          (TCRETURNmi64 addr:$dst, imm:$off)>,
974          Requires<[In64BitMode]>;
975
976def : Pat<(X86tcret (i64 tglobaladdr:$dst), imm:$off),
977          (TCRETURNdi64 tglobaladdr:$dst, imm:$off)>,
978          Requires<[In64BitMode]>;
979
980def : Pat<(X86tcret (i64 texternalsym:$dst), imm:$off),
981          (TCRETURNdi64 texternalsym:$dst, imm:$off)>,
982          Requires<[In64BitMode]>;
983
984// Normal calls, with various flavors of addresses.
985def : Pat<(X86call (i32 tglobaladdr:$dst)),
986          (CALLpcrel32 tglobaladdr:$dst)>;
987def : Pat<(X86call (i32 texternalsym:$dst)),
988          (CALLpcrel32 texternalsym:$dst)>;
989def : Pat<(X86call (i32 imm:$dst)),
990          (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
991
992// Comparisons.
993
994// TEST R,R is smaller than CMP R,0
995def : Pat<(X86cmp GR8:$src1, 0),
996          (TEST8rr GR8:$src1, GR8:$src1)>;
997def : Pat<(X86cmp GR16:$src1, 0),
998          (TEST16rr GR16:$src1, GR16:$src1)>;
999def : Pat<(X86cmp GR32:$src1, 0),
1000          (TEST32rr GR32:$src1, GR32:$src1)>;
1001def : Pat<(X86cmp GR64:$src1, 0),
1002          (TEST64rr GR64:$src1, GR64:$src1)>;
1003
1004// Conditional moves with folded loads with operands swapped and conditions
1005// inverted.
1006multiclass CMOVmr<PatLeaf InvertedCond, Instruction Inst16, Instruction Inst32,
1007                  Instruction Inst64> {
1008  def : Pat<(X86cmov (loadi16 addr:$src1), GR16:$src2, InvertedCond, EFLAGS),
1009            (Inst16 GR16:$src2, addr:$src1)>;
1010  def : Pat<(X86cmov (loadi32 addr:$src1), GR32:$src2, InvertedCond, EFLAGS),
1011            (Inst32 GR32:$src2, addr:$src1)>;
1012  def : Pat<(X86cmov (loadi64 addr:$src1), GR64:$src2, InvertedCond, EFLAGS),
1013            (Inst64 GR64:$src2, addr:$src1)>;
1014}
1015
1016defm : CMOVmr<X86_COND_B , CMOVAE16rm, CMOVAE32rm, CMOVAE64rm>;
1017defm : CMOVmr<X86_COND_AE, CMOVB16rm , CMOVB32rm , CMOVB64rm>;
1018defm : CMOVmr<X86_COND_E , CMOVNE16rm, CMOVNE32rm, CMOVNE64rm>;
1019defm : CMOVmr<X86_COND_NE, CMOVE16rm , CMOVE32rm , CMOVE64rm>;
1020defm : CMOVmr<X86_COND_BE, CMOVA16rm , CMOVA32rm , CMOVA64rm>;
1021defm : CMOVmr<X86_COND_A , CMOVBE16rm, CMOVBE32rm, CMOVBE64rm>;
1022defm : CMOVmr<X86_COND_L , CMOVGE16rm, CMOVGE32rm, CMOVGE64rm>;
1023defm : CMOVmr<X86_COND_GE, CMOVL16rm , CMOVL32rm , CMOVL64rm>;
1024defm : CMOVmr<X86_COND_LE, CMOVG16rm , CMOVG32rm , CMOVG64rm>;
1025defm : CMOVmr<X86_COND_G , CMOVLE16rm, CMOVLE32rm, CMOVLE64rm>;
1026defm : CMOVmr<X86_COND_P , CMOVNP16rm, CMOVNP32rm, CMOVNP64rm>;
1027defm : CMOVmr<X86_COND_NP, CMOVP16rm , CMOVP32rm , CMOVP64rm>;
1028defm : CMOVmr<X86_COND_S , CMOVNS16rm, CMOVNS32rm, CMOVNS64rm>;
1029defm : CMOVmr<X86_COND_NS, CMOVS16rm , CMOVS32rm , CMOVS64rm>;
1030defm : CMOVmr<X86_COND_O , CMOVNO16rm, CMOVNO32rm, CMOVNO64rm>;
1031defm : CMOVmr<X86_COND_NO, CMOVO16rm , CMOVO32rm , CMOVO64rm>;
1032
1033// zextload bool -> zextload byte
1034def : Pat<(zextloadi8i1  addr:$src), (MOV8rm     addr:$src)>;
1035def : Pat<(zextloadi16i1 addr:$src), (MOVZX16rm8 addr:$src)>;
1036def : Pat<(zextloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1037def : Pat<(zextloadi64i1 addr:$src), (MOVZX64rm8 addr:$src)>;
1038
1039// extload bool -> extload byte
1040// When extloading from 16-bit and smaller memory locations into 64-bit
1041// registers, use zero-extending loads so that the entire 64-bit register is
1042// defined, avoiding partial-register updates.
1043
1044def : Pat<(extloadi8i1 addr:$src),   (MOV8rm      addr:$src)>;
1045def : Pat<(extloadi16i1 addr:$src),  (MOVZX16rm8  addr:$src)>;
1046def : Pat<(extloadi32i1 addr:$src),  (MOVZX32rm8  addr:$src)>;
1047def : Pat<(extloadi16i8 addr:$src),  (MOVZX16rm8  addr:$src)>;
1048def : Pat<(extloadi32i8 addr:$src),  (MOVZX32rm8  addr:$src)>;
1049def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
1050
1051def : Pat<(extloadi64i1 addr:$src),  (MOVZX64rm8  addr:$src)>;
1052def : Pat<(extloadi64i8 addr:$src),  (MOVZX64rm8  addr:$src)>;
1053def : Pat<(extloadi64i16 addr:$src), (MOVZX64rm16 addr:$src)>;
1054// For other extloads, use subregs, since the high contents of the register are
1055// defined after an extload.
1056def : Pat<(extloadi64i32 addr:$src),
1057          (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src),
1058                         sub_32bit)>;
1059
1060// anyext. Define these to do an explicit zero-extend to
1061// avoid partial-register updates.
1062def : Pat<(i16 (anyext GR8 :$src)), (EXTRACT_SUBREG
1063                                     (MOVZX32rr8 GR8 :$src), sub_16bit)>;
1064def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8  GR8 :$src)>;
1065
1066// Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
1067def : Pat<(i32 (anyext GR16:$src)),
1068          (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
1069
1070def : Pat<(i64 (anyext GR8 :$src)), (MOVZX64rr8  GR8  :$src)>;
1071def : Pat<(i64 (anyext GR16:$src)), (MOVZX64rr16 GR16 :$src)>;
1072def : Pat<(i64 (anyext GR32:$src)),
1073          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1074
1075
1076// Any instruction that defines a 32-bit result leaves the high half of the
1077// register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
1078// be copying from a truncate. And x86's cmov doesn't do anything if the
1079// condition is false. But any other 32-bit operation will zero-extend
1080// up to 64 bits.
1081def def32 : PatLeaf<(i32 GR32:$src), [{
1082  return N->getOpcode() != ISD::TRUNCATE &&
1083         N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
1084         N->getOpcode() != ISD::CopyFromReg &&
1085         N->getOpcode() != X86ISD::CMOV;
1086}]>;
1087
1088// In the case of a 32-bit def that is known to implicitly zero-extend,
1089// we can use a SUBREG_TO_REG.
1090def : Pat<(i64 (zext def32:$src)),
1091          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1092
1093//===----------------------------------------------------------------------===//
1094// Pattern match OR as ADD
1095//===----------------------------------------------------------------------===//
1096
1097// If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
1098// 3-addressified into an LEA instruction to avoid copies.  However, we also
1099// want to finally emit these instructions as an or at the end of the code
1100// generator to make the generated code easier to read.  To do this, we select
1101// into "disjoint bits" pseudo ops.
1102
1103// Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.
1104def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
1105  if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1106    return CurDAG->MaskedValueIsZero(N->getOperand(0), CN->getAPIntValue());
1107
1108  unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
1109  APInt Mask = APInt::getAllOnesValue(BitWidth);
1110  APInt KnownZero0, KnownOne0;
1111  CurDAG->ComputeMaskedBits(N->getOperand(0), Mask, KnownZero0, KnownOne0, 0);
1112  APInt KnownZero1, KnownOne1;
1113  CurDAG->ComputeMaskedBits(N->getOperand(1), Mask, KnownZero1, KnownOne1, 0);
1114  return (~KnownZero0 & ~KnownZero1) == 0;
1115}]>;
1116
1117
1118// (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1119let AddedComplexity = 5 in { // Try this before the selecting to OR
1120
1121let isConvertibleToThreeAddress = 1,
1122    Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1123let isCommutable = 1 in {
1124def ADD16rr_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1125                    "", // orw/addw REG, REG
1126                    [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1127def ADD32rr_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1128                    "", // orl/addl REG, REG
1129                    [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1130def ADD64rr_DB  : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1131                    "", // orq/addq REG, REG
1132                    [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1133} // isCommutable
1134
1135// NOTE: These are order specific, we want the ri8 forms to be listed
1136// first so that they are slightly preferred to the ri forms.
1137
1138def ADD16ri8_DB : I<0, Pseudo,
1139                    (outs GR16:$dst), (ins GR16:$src1, i16i8imm:$src2),
1140                    "", // orw/addw REG, imm8
1141                    [(set GR16:$dst,(or_is_add GR16:$src1,i16immSExt8:$src2))]>;
1142def ADD16ri_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1143                    "", // orw/addw REG, imm
1144                    [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1145
1146def ADD32ri8_DB : I<0, Pseudo,
1147                    (outs GR32:$dst), (ins GR32:$src1, i32i8imm:$src2),
1148                    "", // orl/addl REG, imm8
1149                    [(set GR32:$dst,(or_is_add GR32:$src1,i32immSExt8:$src2))]>;
1150def ADD32ri_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1151                    "", // orl/addl REG, imm
1152                    [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1153
1154
1155def ADD64ri8_DB : I<0, Pseudo,
1156                    (outs GR64:$dst), (ins GR64:$src1, i64i8imm:$src2),
1157                    "", // orq/addq REG, imm8
1158                    [(set GR64:$dst, (or_is_add GR64:$src1,
1159                                                i64immSExt8:$src2))]>;
1160def ADD64ri32_DB : I<0, Pseudo,
1161                     (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1162                      "", // orq/addq REG, imm
1163                      [(set GR64:$dst, (or_is_add GR64:$src1,
1164                                                  i64immSExt32:$src2))]>;
1165}
1166} // AddedComplexity
1167
1168
1169//===----------------------------------------------------------------------===//
1170// Some peepholes
1171//===----------------------------------------------------------------------===//
1172
1173// Odd encoding trick: -128 fits into an 8-bit immediate field while
1174// +128 doesn't, so in this special case use a sub instead of an add.
1175def : Pat<(add GR16:$src1, 128),
1176          (SUB16ri8 GR16:$src1, -128)>;
1177def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1178          (SUB16mi8 addr:$dst, -128)>;
1179
1180def : Pat<(add GR32:$src1, 128),
1181          (SUB32ri8 GR32:$src1, -128)>;
1182def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1183          (SUB32mi8 addr:$dst, -128)>;
1184
1185def : Pat<(add GR64:$src1, 128),
1186          (SUB64ri8 GR64:$src1, -128)>;
1187def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1188          (SUB64mi8 addr:$dst, -128)>;
1189
1190// The same trick applies for 32-bit immediate fields in 64-bit
1191// instructions.
1192def : Pat<(add GR64:$src1, 0x0000000080000000),
1193          (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1194def : Pat<(store (add (loadi64 addr:$dst), 0x00000000800000000), addr:$dst),
1195          (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1196
1197// To avoid needing to materialize an immediate in a register, use a 32-bit and
1198// with implicit zero-extension instead of a 64-bit and if the immediate has at
1199// least 32 bits of leading zeros. If in addition the last 32 bits can be
1200// represented with a sign extension of a 8 bit constant, use that.
1201
1202def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
1203          (SUBREG_TO_REG
1204            (i64 0),
1205            (AND32ri8
1206              (EXTRACT_SUBREG GR64:$src, sub_32bit),
1207              (i32 (GetLo8XForm imm:$imm))),
1208            sub_32bit)>;
1209
1210def : Pat<(and GR64:$src, i64immZExt32:$imm),
1211          (SUBREG_TO_REG
1212            (i64 0),
1213            (AND32ri
1214              (EXTRACT_SUBREG GR64:$src, sub_32bit),
1215              (i32 (GetLo32XForm imm:$imm))),
1216            sub_32bit)>;
1217
1218
1219// r & (2^16-1) ==> movz
1220def : Pat<(and GR32:$src1, 0xffff),
1221          (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1222// r & (2^8-1) ==> movz
1223def : Pat<(and GR32:$src1, 0xff),
1224          (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src1,
1225                                                             GR32_ABCD)),
1226                                      sub_8bit))>,
1227      Requires<[In32BitMode]>;
1228// r & (2^8-1) ==> movz
1229def : Pat<(and GR16:$src1, 0xff),
1230           (EXTRACT_SUBREG (MOVZX32rr8 (EXTRACT_SUBREG
1231            (i16 (COPY_TO_REGCLASS GR16:$src1, GR16_ABCD)), sub_8bit)),
1232             sub_16bit)>,
1233      Requires<[In32BitMode]>;
1234
1235// r & (2^32-1) ==> movz
1236def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1237          (MOVZX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1238// r & (2^16-1) ==> movz
1239def : Pat<(and GR64:$src, 0xffff),
1240          (MOVZX64rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit)))>;
1241// r & (2^8-1) ==> movz
1242def : Pat<(and GR64:$src, 0xff),
1243          (MOVZX64rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit)))>;
1244// r & (2^8-1) ==> movz
1245def : Pat<(and GR32:$src1, 0xff),
1246           (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>,
1247      Requires<[In64BitMode]>;
1248// r & (2^8-1) ==> movz
1249def : Pat<(and GR16:$src1, 0xff),
1250           (EXTRACT_SUBREG (MOVZX32rr8 (i8
1251            (EXTRACT_SUBREG GR16:$src1, sub_8bit))), sub_16bit)>,
1252      Requires<[In64BitMode]>;
1253
1254
1255// sext_inreg patterns
1256def : Pat<(sext_inreg GR32:$src, i16),
1257          (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1258def : Pat<(sext_inreg GR32:$src, i8),
1259          (MOVSX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1260                                                             GR32_ABCD)),
1261                                      sub_8bit))>,
1262      Requires<[In32BitMode]>;
1263
1264def : Pat<(sext_inreg GR16:$src, i8),
1265           (EXTRACT_SUBREG (i32 (MOVSX32rr8 (EXTRACT_SUBREG
1266            (i32 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)), sub_8bit))),
1267             sub_16bit)>,
1268      Requires<[In32BitMode]>;
1269
1270def : Pat<(sext_inreg GR64:$src, i32),
1271          (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1272def : Pat<(sext_inreg GR64:$src, i16),
1273          (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1274def : Pat<(sext_inreg GR64:$src, i8),
1275          (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1276def : Pat<(sext_inreg GR32:$src, i8),
1277          (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>,
1278      Requires<[In64BitMode]>;
1279def : Pat<(sext_inreg GR16:$src, i8),
1280           (EXTRACT_SUBREG (MOVSX32rr8
1281            (EXTRACT_SUBREG GR16:$src, sub_8bit)), sub_16bit)>,
1282      Requires<[In64BitMode]>;
1283
1284// sext, sext_load, zext, zext_load
1285def: Pat<(i16 (sext GR8:$src)),
1286          (EXTRACT_SUBREG (MOVSX32rr8 GR8:$src), sub_16bit)>;
1287def: Pat<(sextloadi16i8 addr:$src),
1288          (EXTRACT_SUBREG (MOVSX32rm8 addr:$src), sub_16bit)>;
1289def: Pat<(i16 (zext GR8:$src)),
1290          (EXTRACT_SUBREG (MOVZX32rr8 GR8:$src), sub_16bit)>;
1291def: Pat<(zextloadi16i8 addr:$src),
1292          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1293
1294// trunc patterns
1295def : Pat<(i16 (trunc GR32:$src)),
1296          (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1297def : Pat<(i8 (trunc GR32:$src)),
1298          (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1299                          sub_8bit)>,
1300      Requires<[In32BitMode]>;
1301def : Pat<(i8 (trunc GR16:$src)),
1302          (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1303                          sub_8bit)>,
1304      Requires<[In32BitMode]>;
1305def : Pat<(i32 (trunc GR64:$src)),
1306          (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1307def : Pat<(i16 (trunc GR64:$src)),
1308          (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1309def : Pat<(i8 (trunc GR64:$src)),
1310          (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1311def : Pat<(i8 (trunc GR32:$src)),
1312          (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1313      Requires<[In64BitMode]>;
1314def : Pat<(i8 (trunc GR16:$src)),
1315          (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1316      Requires<[In64BitMode]>;
1317
1318// h-register tricks
1319def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1320          (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1321                          sub_8bit_hi)>,
1322      Requires<[In32BitMode]>;
1323def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1324          (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1325                          sub_8bit_hi)>,
1326      Requires<[In32BitMode]>;
1327def : Pat<(srl GR16:$src, (i8 8)),
1328          (EXTRACT_SUBREG
1329            (MOVZX32rr8
1330              (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1331                              sub_8bit_hi)),
1332            sub_16bit)>,
1333      Requires<[In32BitMode]>;
1334def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1335          (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1336                                                             GR16_ABCD)),
1337                                      sub_8bit_hi))>,
1338      Requires<[In32BitMode]>;
1339def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1340          (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1341                                                             GR16_ABCD)),
1342                                      sub_8bit_hi))>,
1343      Requires<[In32BitMode]>;
1344def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1345          (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1346                                                             GR32_ABCD)),
1347                                      sub_8bit_hi))>,
1348      Requires<[In32BitMode]>;
1349def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1350          (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1351                                                             GR32_ABCD)),
1352                                      sub_8bit_hi))>,
1353      Requires<[In32BitMode]>;
1354
1355// h-register tricks.
1356// For now, be conservative on x86-64 and use an h-register extract only if the
1357// value is immediately zero-extended or stored, which are somewhat common
1358// cases. This uses a bunch of code to prevent a register requiring a REX prefix
1359// from being allocated in the same instruction as the h register, as there's
1360// currently no way to describe this requirement to the register allocator.
1361
1362// h-register extract and zero-extend.
1363def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1364          (SUBREG_TO_REG
1365            (i64 0),
1366            (MOVZX32_NOREXrr8
1367              (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1368                              sub_8bit_hi)),
1369            sub_32bit)>;
1370def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1371          (MOVZX32_NOREXrr8
1372            (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1373                            sub_8bit_hi))>,
1374      Requires<[In64BitMode]>;
1375def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1376          (MOVZX32_NOREXrr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1377                                                                   GR32_ABCD)),
1378                                             sub_8bit_hi))>,
1379      Requires<[In64BitMode]>;
1380def : Pat<(srl GR16:$src, (i8 8)),
1381          (EXTRACT_SUBREG
1382            (MOVZX32_NOREXrr8
1383              (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1384                              sub_8bit_hi)),
1385            sub_16bit)>,
1386      Requires<[In64BitMode]>;
1387def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1388          (MOVZX32_NOREXrr8
1389            (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1390                            sub_8bit_hi))>,
1391      Requires<[In64BitMode]>;
1392def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1393          (MOVZX32_NOREXrr8
1394            (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1395                            sub_8bit_hi))>,
1396      Requires<[In64BitMode]>;
1397def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1398          (SUBREG_TO_REG
1399            (i64 0),
1400            (MOVZX32_NOREXrr8
1401              (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1402                              sub_8bit_hi)),
1403            sub_32bit)>;
1404def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1405          (SUBREG_TO_REG
1406            (i64 0),
1407            (MOVZX32_NOREXrr8
1408              (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1409                              sub_8bit_hi)),
1410            sub_32bit)>;
1411
1412// h-register extract and store.
1413def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1414          (MOV8mr_NOREX
1415            addr:$dst,
1416            (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1417                            sub_8bit_hi))>;
1418def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1419          (MOV8mr_NOREX
1420            addr:$dst,
1421            (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1422                            sub_8bit_hi))>,
1423      Requires<[In64BitMode]>;
1424def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1425          (MOV8mr_NOREX
1426            addr:$dst,
1427            (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1428                            sub_8bit_hi))>,
1429      Requires<[In64BitMode]>;
1430
1431
1432// (shl x, 1) ==> (add x, x)
1433// Note that if x is undef (immediate or otherwise), we could theoretically
1434// end up with the two uses of x getting different values, producing a result
1435// where the least significant bit is not 0. However, the probability of this
1436// happening is considered low enough that this is officially not a
1437// "real problem".
1438def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr  GR8 :$src1, GR8 :$src1)>;
1439def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1440def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1441def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1442
1443// (shl x (and y, 31)) ==> (shl x, y)
1444def : Pat<(shl GR8:$src1, (and CL, 31)),
1445          (SHL8rCL GR8:$src1)>;
1446def : Pat<(shl GR16:$src1, (and CL, 31)),
1447          (SHL16rCL GR16:$src1)>;
1448def : Pat<(shl GR32:$src1, (and CL, 31)),
1449          (SHL32rCL GR32:$src1)>;
1450def : Pat<(store (shl (loadi8 addr:$dst), (and CL, 31)), addr:$dst),
1451          (SHL8mCL addr:$dst)>;
1452def : Pat<(store (shl (loadi16 addr:$dst), (and CL, 31)), addr:$dst),
1453          (SHL16mCL addr:$dst)>;
1454def : Pat<(store (shl (loadi32 addr:$dst), (and CL, 31)), addr:$dst),
1455          (SHL32mCL addr:$dst)>;
1456
1457def : Pat<(srl GR8:$src1, (and CL, 31)),
1458          (SHR8rCL GR8:$src1)>;
1459def : Pat<(srl GR16:$src1, (and CL, 31)),
1460          (SHR16rCL GR16:$src1)>;
1461def : Pat<(srl GR32:$src1, (and CL, 31)),
1462          (SHR32rCL GR32:$src1)>;
1463def : Pat<(store (srl (loadi8 addr:$dst), (and CL, 31)), addr:$dst),
1464          (SHR8mCL addr:$dst)>;
1465def : Pat<(store (srl (loadi16 addr:$dst), (and CL, 31)), addr:$dst),
1466          (SHR16mCL addr:$dst)>;
1467def : Pat<(store (srl (loadi32 addr:$dst), (and CL, 31)), addr:$dst),
1468          (SHR32mCL addr:$dst)>;
1469
1470def : Pat<(sra GR8:$src1, (and CL, 31)),
1471          (SAR8rCL GR8:$src1)>;
1472def : Pat<(sra GR16:$src1, (and CL, 31)),
1473          (SAR16rCL GR16:$src1)>;
1474def : Pat<(sra GR32:$src1, (and CL, 31)),
1475          (SAR32rCL GR32:$src1)>;
1476def : Pat<(store (sra (loadi8 addr:$dst), (and CL, 31)), addr:$dst),
1477          (SAR8mCL addr:$dst)>;
1478def : Pat<(store (sra (loadi16 addr:$dst), (and CL, 31)), addr:$dst),
1479          (SAR16mCL addr:$dst)>;
1480def : Pat<(store (sra (loadi32 addr:$dst), (and CL, 31)), addr:$dst),
1481          (SAR32mCL addr:$dst)>;
1482
1483// (shl x (and y, 63)) ==> (shl x, y)
1484def : Pat<(shl GR64:$src1, (and CL, 63)),
1485          (SHL64rCL GR64:$src1)>;
1486def : Pat<(store (shl (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1487          (SHL64mCL addr:$dst)>;
1488
1489def : Pat<(srl GR64:$src1, (and CL, 63)),
1490          (SHR64rCL GR64:$src1)>;
1491def : Pat<(store (srl (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1492          (SHR64mCL addr:$dst)>;
1493
1494def : Pat<(sra GR64:$src1, (and CL, 63)),
1495          (SAR64rCL GR64:$src1)>;
1496def : Pat<(store (sra (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1497          (SAR64mCL addr:$dst)>;
1498
1499
1500// (anyext (setcc_carry)) -> (setcc_carry)
1501def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1502          (SETB_C16r)>;
1503def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1504          (SETB_C32r)>;
1505def : Pat<(i32 (anyext (i16 (X86setcc_c X86_COND_B, EFLAGS)))),
1506          (SETB_C32r)>;
1507
1508
1509
1510
1511//===----------------------------------------------------------------------===//
1512// EFLAGS-defining Patterns
1513//===----------------------------------------------------------------------===//
1514
1515// add reg, reg
1516def : Pat<(add GR8 :$src1, GR8 :$src2), (ADD8rr  GR8 :$src1, GR8 :$src2)>;
1517def : Pat<(add GR16:$src1, GR16:$src2), (ADD16rr GR16:$src1, GR16:$src2)>;
1518def : Pat<(add GR32:$src1, GR32:$src2), (ADD32rr GR32:$src1, GR32:$src2)>;
1519
1520// add reg, mem
1521def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
1522          (ADD8rm GR8:$src1, addr:$src2)>;
1523def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
1524          (ADD16rm GR16:$src1, addr:$src2)>;
1525def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
1526          (ADD32rm GR32:$src1, addr:$src2)>;
1527
1528// add reg, imm
1529def : Pat<(add GR8 :$src1, imm:$src2), (ADD8ri  GR8:$src1 , imm:$src2)>;
1530def : Pat<(add GR16:$src1, imm:$src2), (ADD16ri GR16:$src1, imm:$src2)>;
1531def : Pat<(add GR32:$src1, imm:$src2), (ADD32ri GR32:$src1, imm:$src2)>;
1532def : Pat<(add GR16:$src1, i16immSExt8:$src2),
1533          (ADD16ri8 GR16:$src1, i16immSExt8:$src2)>;
1534def : Pat<(add GR32:$src1, i32immSExt8:$src2),
1535          (ADD32ri8 GR32:$src1, i32immSExt8:$src2)>;
1536
1537// sub reg, reg
1538def : Pat<(sub GR8 :$src1, GR8 :$src2), (SUB8rr  GR8 :$src1, GR8 :$src2)>;
1539def : Pat<(sub GR16:$src1, GR16:$src2), (SUB16rr GR16:$src1, GR16:$src2)>;
1540def : Pat<(sub GR32:$src1, GR32:$src2), (SUB32rr GR32:$src1, GR32:$src2)>;
1541
1542// sub reg, mem
1543def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
1544          (SUB8rm GR8:$src1, addr:$src2)>;
1545def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
1546          (SUB16rm GR16:$src1, addr:$src2)>;
1547def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
1548          (SUB32rm GR32:$src1, addr:$src2)>;
1549
1550// sub reg, imm
1551def : Pat<(sub GR8:$src1, imm:$src2),
1552          (SUB8ri GR8:$src1, imm:$src2)>;
1553def : Pat<(sub GR16:$src1, imm:$src2),
1554          (SUB16ri GR16:$src1, imm:$src2)>;
1555def : Pat<(sub GR32:$src1, imm:$src2),
1556          (SUB32ri GR32:$src1, imm:$src2)>;
1557def : Pat<(sub GR16:$src1, i16immSExt8:$src2),
1558          (SUB16ri8 GR16:$src1, i16immSExt8:$src2)>;
1559def : Pat<(sub GR32:$src1, i32immSExt8:$src2),
1560          (SUB32ri8 GR32:$src1, i32immSExt8:$src2)>;
1561
1562// mul reg, reg
1563def : Pat<(mul GR16:$src1, GR16:$src2),
1564          (IMUL16rr GR16:$src1, GR16:$src2)>;
1565def : Pat<(mul GR32:$src1, GR32:$src2),
1566          (IMUL32rr GR32:$src1, GR32:$src2)>;
1567
1568// mul reg, mem
1569def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
1570          (IMUL16rm GR16:$src1, addr:$src2)>;
1571def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
1572          (IMUL32rm GR32:$src1, addr:$src2)>;
1573
1574// mul reg, imm
1575def : Pat<(mul GR16:$src1, imm:$src2),
1576          (IMUL16rri GR16:$src1, imm:$src2)>;
1577def : Pat<(mul GR32:$src1, imm:$src2),
1578          (IMUL32rri GR32:$src1, imm:$src2)>;
1579def : Pat<(mul GR16:$src1, i16immSExt8:$src2),
1580          (IMUL16rri8 GR16:$src1, i16immSExt8:$src2)>;
1581def : Pat<(mul GR32:$src1, i32immSExt8:$src2),
1582          (IMUL32rri8 GR32:$src1, i32immSExt8:$src2)>;
1583
1584// reg = mul mem, imm
1585def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
1586          (IMUL16rmi addr:$src1, imm:$src2)>;
1587def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
1588          (IMUL32rmi addr:$src1, imm:$src2)>;
1589def : Pat<(mul (loadi16 addr:$src1), i16immSExt8:$src2),
1590          (IMUL16rmi8 addr:$src1, i16immSExt8:$src2)>;
1591def : Pat<(mul (loadi32 addr:$src1), i32immSExt8:$src2),
1592          (IMUL32rmi8 addr:$src1, i32immSExt8:$src2)>;
1593
1594// Patterns for nodes that do not produce flags, for instructions that do.
1595
1596// addition
1597def : Pat<(add GR64:$src1, GR64:$src2),
1598          (ADD64rr GR64:$src1, GR64:$src2)>;
1599def : Pat<(add GR64:$src1, i64immSExt8:$src2),
1600          (ADD64ri8 GR64:$src1, i64immSExt8:$src2)>;
1601def : Pat<(add GR64:$src1, i64immSExt32:$src2),
1602          (ADD64ri32 GR64:$src1, i64immSExt32:$src2)>;
1603def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
1604          (ADD64rm GR64:$src1, addr:$src2)>;
1605
1606// subtraction
1607def : Pat<(sub GR64:$src1, GR64:$src2),
1608          (SUB64rr GR64:$src1, GR64:$src2)>;
1609def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
1610          (SUB64rm GR64:$src1, addr:$src2)>;
1611def : Pat<(sub GR64:$src1, i64immSExt8:$src2),
1612          (SUB64ri8 GR64:$src1, i64immSExt8:$src2)>;
1613def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
1614          (SUB64ri32 GR64:$src1, i64immSExt32:$src2)>;
1615
1616// Multiply
1617def : Pat<(mul GR64:$src1, GR64:$src2),
1618          (IMUL64rr GR64:$src1, GR64:$src2)>;
1619def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
1620          (IMUL64rm GR64:$src1, addr:$src2)>;
1621def : Pat<(mul GR64:$src1, i64immSExt8:$src2),
1622          (IMUL64rri8 GR64:$src1, i64immSExt8:$src2)>;
1623def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
1624          (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
1625def : Pat<(mul (loadi64 addr:$src1), i64immSExt8:$src2),
1626          (IMUL64rmi8 addr:$src1, i64immSExt8:$src2)>;
1627def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
1628          (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
1629
1630// Increment reg.
1631def : Pat<(add GR8 :$src, 1), (INC8r     GR8 :$src)>;
1632def : Pat<(add GR16:$src, 1), (INC16r    GR16:$src)>, Requires<[In32BitMode]>;
1633def : Pat<(add GR16:$src, 1), (INC64_16r GR16:$src)>, Requires<[In64BitMode]>;
1634def : Pat<(add GR32:$src, 1), (INC32r    GR32:$src)>, Requires<[In32BitMode]>;
1635def : Pat<(add GR32:$src, 1), (INC64_32r GR32:$src)>, Requires<[In64BitMode]>;
1636def : Pat<(add GR64:$src, 1), (INC64r    GR64:$src)>;
1637
1638// Decrement reg.
1639def : Pat<(add GR8 :$src, -1), (DEC8r     GR8 :$src)>;
1640def : Pat<(add GR16:$src, -1), (DEC16r    GR16:$src)>, Requires<[In32BitMode]>;
1641def : Pat<(add GR16:$src, -1), (DEC64_16r GR16:$src)>, Requires<[In64BitMode]>;
1642def : Pat<(add GR32:$src, -1), (DEC32r    GR32:$src)>, Requires<[In32BitMode]>;
1643def : Pat<(add GR32:$src, -1), (DEC64_32r GR32:$src)>, Requires<[In64BitMode]>;
1644def : Pat<(add GR64:$src, -1), (DEC64r    GR64:$src)>;
1645
1646// or reg/reg.
1647def : Pat<(or GR8 :$src1, GR8 :$src2), (OR8rr  GR8 :$src1, GR8 :$src2)>;
1648def : Pat<(or GR16:$src1, GR16:$src2), (OR16rr GR16:$src1, GR16:$src2)>;
1649def : Pat<(or GR32:$src1, GR32:$src2), (OR32rr GR32:$src1, GR32:$src2)>;
1650def : Pat<(or GR64:$src1, GR64:$src2), (OR64rr GR64:$src1, GR64:$src2)>;
1651
1652// or reg/mem
1653def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
1654          (OR8rm GR8:$src1, addr:$src2)>;
1655def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
1656          (OR16rm GR16:$src1, addr:$src2)>;
1657def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
1658          (OR32rm GR32:$src1, addr:$src2)>;
1659def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
1660          (OR64rm GR64:$src1, addr:$src2)>;
1661
1662// or reg/imm
1663def : Pat<(or GR8:$src1 , imm:$src2), (OR8ri  GR8 :$src1, imm:$src2)>;
1664def : Pat<(or GR16:$src1, imm:$src2), (OR16ri GR16:$src1, imm:$src2)>;
1665def : Pat<(or GR32:$src1, imm:$src2), (OR32ri GR32:$src1, imm:$src2)>;
1666def : Pat<(or GR16:$src1, i16immSExt8:$src2),
1667          (OR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1668def : Pat<(or GR32:$src1, i32immSExt8:$src2),
1669          (OR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1670def : Pat<(or GR64:$src1, i64immSExt8:$src2),
1671          (OR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1672def : Pat<(or GR64:$src1, i64immSExt32:$src2),
1673          (OR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1674
1675// xor reg/reg
1676def : Pat<(xor GR8 :$src1, GR8 :$src2), (XOR8rr  GR8 :$src1, GR8 :$src2)>;
1677def : Pat<(xor GR16:$src1, GR16:$src2), (XOR16rr GR16:$src1, GR16:$src2)>;
1678def : Pat<(xor GR32:$src1, GR32:$src2), (XOR32rr GR32:$src1, GR32:$src2)>;
1679def : Pat<(xor GR64:$src1, GR64:$src2), (XOR64rr GR64:$src1, GR64:$src2)>;
1680
1681// xor reg/mem
1682def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
1683          (XOR8rm GR8:$src1, addr:$src2)>;
1684def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
1685          (XOR16rm GR16:$src1, addr:$src2)>;
1686def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
1687          (XOR32rm GR32:$src1, addr:$src2)>;
1688def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
1689          (XOR64rm GR64:$src1, addr:$src2)>;
1690
1691// xor reg/imm
1692def : Pat<(xor GR8:$src1, imm:$src2),
1693          (XOR8ri GR8:$src1, imm:$src2)>;
1694def : Pat<(xor GR16:$src1, imm:$src2),
1695          (XOR16ri GR16:$src1, imm:$src2)>;
1696def : Pat<(xor GR32:$src1, imm:$src2),
1697          (XOR32ri GR32:$src1, imm:$src2)>;
1698def : Pat<(xor GR16:$src1, i16immSExt8:$src2),
1699          (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1700def : Pat<(xor GR32:$src1, i32immSExt8:$src2),
1701          (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1702def : Pat<(xor GR64:$src1, i64immSExt8:$src2),
1703          (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1704def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
1705          (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1706
1707// and reg/reg
1708def : Pat<(and GR8 :$src1, GR8 :$src2), (AND8rr  GR8 :$src1, GR8 :$src2)>;
1709def : Pat<(and GR16:$src1, GR16:$src2), (AND16rr GR16:$src1, GR16:$src2)>;
1710def : Pat<(and GR32:$src1, GR32:$src2), (AND32rr GR32:$src1, GR32:$src2)>;
1711def : Pat<(and GR64:$src1, GR64:$src2), (AND64rr GR64:$src1, GR64:$src2)>;
1712
1713// and reg/mem
1714def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
1715          (AND8rm GR8:$src1, addr:$src2)>;
1716def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
1717          (AND16rm GR16:$src1, addr:$src2)>;
1718def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
1719          (AND32rm GR32:$src1, addr:$src2)>;
1720def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
1721          (AND64rm GR64:$src1, addr:$src2)>;
1722
1723// and reg/imm
1724def : Pat<(and GR8:$src1, imm:$src2),
1725          (AND8ri GR8:$src1, imm:$src2)>;
1726def : Pat<(and GR16:$src1, imm:$src2),
1727          (AND16ri GR16:$src1, imm:$src2)>;
1728def : Pat<(and GR32:$src1, imm:$src2),
1729          (AND32ri GR32:$src1, imm:$src2)>;
1730def : Pat<(and GR16:$src1, i16immSExt8:$src2),
1731          (AND16ri8 GR16:$src1, i16immSExt8:$src2)>;
1732def : Pat<(and GR32:$src1, i32immSExt8:$src2),
1733          (AND32ri8 GR32:$src1, i32immSExt8:$src2)>;
1734def : Pat<(and GR64:$src1, i64immSExt8:$src2),
1735          (AND64ri8 GR64:$src1, i64immSExt8:$src2)>;
1736def : Pat<(and GR64:$src1, i64immSExt32:$src2),
1737          (AND64ri32 GR64:$src1, i64immSExt32:$src2)>;
1738