• 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] 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] 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// The MSVC runtime contains an _ftol2 routine for converting floating-point
131// to integer values. It has a strange calling convention: the input is
132// popped from the x87 stack, and the return value is given in EDX:EAX. No
133// other registers (aside from flags) are touched.
134// Microsoft toolchains do not support 80-bit precision, so a WIN_FTOL_80
135// variant is unnecessary.
136
137let Defs = [EAX, EDX, EFLAGS], FPForm = SpecialFP in {
138  def WIN_FTOL_32 : I<0, Pseudo, (outs), (ins RFP32:$src),
139                      "# win32 fptoui",
140                      [(X86WinFTOL RFP32:$src)]>,
141                    Requires<[In32BitMode]>;
142
143  def WIN_FTOL_64 : I<0, Pseudo, (outs), (ins RFP64:$src),
144                      "# win32 fptoui",
145                      [(X86WinFTOL RFP64:$src)]>,
146                    Requires<[In32BitMode]>;
147}
148
149//===----------------------------------------------------------------------===//
150// EH Pseudo Instructions
151//
152let isTerminator = 1, isReturn = 1, isBarrier = 1,
153    hasCtrlDep = 1, isCodeGenOnly = 1 in {
154def EH_RETURN   : I<0xC3, RawFrm, (outs), (ins GR32:$addr),
155                    "ret\t#eh_return, addr: $addr",
156                    [(X86ehret GR32:$addr)], IIC_RET>;
157
158}
159
160let isTerminator = 1, isReturn = 1, isBarrier = 1,
161    hasCtrlDep = 1, isCodeGenOnly = 1 in {
162def EH_RETURN64   : I<0xC3, RawFrm, (outs), (ins GR64:$addr),
163                     "ret\t#eh_return, addr: $addr",
164                     [(X86ehret GR64:$addr)], IIC_RET>;
165
166}
167
168//===----------------------------------------------------------------------===//
169// Pseudo instructions used by segmented stacks.
170//
171
172// This is lowered into a RET instruction by MCInstLower.  We need
173// this so that we don't have to have a MachineBasicBlock which ends
174// with a RET and also has successors.
175let isPseudo = 1 in {
176def MORESTACK_RET: I<0, Pseudo, (outs), (ins),
177                          "", []>;
178
179// This instruction is lowered to a RET followed by a MOV.  The two
180// instructions are not generated on a higher level since then the
181// verifier sees a MachineBasicBlock ending with a non-terminator.
182def MORESTACK_RET_RESTORE_R10 : I<0, Pseudo, (outs), (ins),
183                                  "", []>;
184}
185
186//===----------------------------------------------------------------------===//
187// Alias Instructions
188//===----------------------------------------------------------------------===//
189
190// Alias instructions that map movr0 to xor.
191// FIXME: remove when we can teach regalloc that xor reg, reg is ok.
192// FIXME: Set encoding to pseudo.
193let Defs = [EFLAGS], isReMaterializable = 1, isAsCheapAsAMove = 1,
194    isCodeGenOnly = 1 in {
195def MOV8r0   : I<0x30, MRMInitReg, (outs GR8 :$dst), (ins), "",
196                 [(set GR8:$dst, 0)], IIC_ALU_NONMEM>;
197
198// We want to rewrite MOV16r0 in terms of MOV32r0, because it's a smaller
199// encoding and avoids a partial-register update sometimes, but doing so
200// at isel time interferes with rematerialization in the current register
201// allocator. For now, this is rewritten when the instruction is lowered
202// to an MCInst.
203def MOV16r0   : I<0x31, MRMInitReg, (outs GR16:$dst), (ins),
204                 "",
205                 [(set GR16:$dst, 0)], IIC_ALU_NONMEM>, OpSize;
206
207// FIXME: Set encoding to pseudo.
208def MOV32r0  : I<0x31, MRMInitReg, (outs GR32:$dst), (ins), "",
209                 [(set GR32:$dst, 0)], IIC_ALU_NONMEM>;
210}
211
212// We want to rewrite MOV64r0 in terms of MOV32r0, because it's sometimes a
213// smaller encoding, but doing so at isel time interferes with rematerialization
214// in the current register allocator. For now, this is rewritten when the
215// instruction is lowered to an MCInst.
216// FIXME: AddedComplexity gives this a higher priority than MOV64ri32. Remove
217// when we have a better way to specify isel priority.
218let Defs = [EFLAGS], isCodeGenOnly=1,
219    AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1 in
220def MOV64r0   : I<0x31, MRMInitReg, (outs GR64:$dst), (ins), "",
221                 [(set GR64:$dst, 0)], IIC_ALU_NONMEM>;
222
223// Materialize i64 constant where top 32-bits are zero. This could theoretically
224// use MOV32ri with a SUBREG_TO_REG to represent the zero-extension, however
225// that would make it more difficult to rematerialize.
226let AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1,
227    isCodeGenOnly = 1 in
228def MOV64ri64i32 : Ii32<0xB8, AddRegFrm, (outs GR64:$dst), (ins i64i32imm:$src),
229                        "", [(set GR64:$dst, i64immZExt32:$src)],
230                        IIC_ALU_NONMEM>;
231
232// Use sbb to materialize carry bit.
233let Uses = [EFLAGS], Defs = [EFLAGS], isCodeGenOnly = 1 in {
234// FIXME: These are pseudo ops that should be replaced with Pat<> patterns.
235// However, Pat<> can't replicate the destination reg into the inputs of the
236// result.
237// FIXME: Change these to have encoding Pseudo when X86MCCodeEmitter replaces
238// X86CodeEmitter.
239def SETB_C8r : I<0x18, MRMInitReg, (outs GR8:$dst), (ins), "",
240                 [(set GR8:$dst, (X86setcc_c X86_COND_B, EFLAGS))],
241                 IIC_ALU_NONMEM>;
242def SETB_C16r : I<0x19, MRMInitReg, (outs GR16:$dst), (ins), "",
243                 [(set GR16:$dst, (X86setcc_c X86_COND_B, EFLAGS))],
244                 IIC_ALU_NONMEM>,
245                OpSize;
246def SETB_C32r : I<0x19, MRMInitReg, (outs GR32:$dst), (ins), "",
247                 [(set GR32:$dst, (X86setcc_c X86_COND_B, EFLAGS))],
248                 IIC_ALU_NONMEM>;
249def SETB_C64r : RI<0x19, MRMInitReg, (outs GR64:$dst), (ins), "",
250                 [(set GR64:$dst, (X86setcc_c X86_COND_B, EFLAGS))],
251                 IIC_ALU_NONMEM>;
252} // isCodeGenOnly
253
254
255def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
256          (SETB_C16r)>;
257def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
258          (SETB_C32r)>;
259def : Pat<(i64 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
260          (SETB_C64r)>;
261
262def : Pat<(i16 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
263          (SETB_C16r)>;
264def : Pat<(i32 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
265          (SETB_C32r)>;
266def : Pat<(i64 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
267          (SETB_C64r)>;
268
269// We canonicalize 'setb' to "(and (sbb reg,reg), 1)" on the hope that the and
270// will be eliminated and that the sbb can be extended up to a wider type.  When
271// this happens, it is great.  However, if we are left with an 8-bit sbb and an
272// and, we might as well just match it as a setb.
273def : Pat<(and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1),
274          (SETBr)>;
275
276// (add OP, SETB) -> (adc OP, 0)
277def : Pat<(add (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR8:$op),
278          (ADC8ri GR8:$op, 0)>;
279def : Pat<(add (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR32:$op),
280          (ADC32ri8 GR32:$op, 0)>;
281def : Pat<(add (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR64:$op),
282          (ADC64ri8 GR64:$op, 0)>;
283
284// (sub OP, SETB) -> (sbb OP, 0)
285def : Pat<(sub GR8:$op, (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
286          (SBB8ri GR8:$op, 0)>;
287def : Pat<(sub GR32:$op, (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
288          (SBB32ri8 GR32:$op, 0)>;
289def : Pat<(sub GR64:$op, (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
290          (SBB64ri8 GR64:$op, 0)>;
291
292// (sub OP, SETCC_CARRY) -> (adc OP, 0)
293def : Pat<(sub GR8:$op, (i8 (X86setcc_c X86_COND_B, EFLAGS))),
294          (ADC8ri GR8:$op, 0)>;
295def : Pat<(sub GR32:$op, (i32 (X86setcc_c X86_COND_B, EFLAGS))),
296          (ADC32ri8 GR32:$op, 0)>;
297def : Pat<(sub GR64:$op, (i64 (X86setcc_c X86_COND_B, EFLAGS))),
298          (ADC64ri8 GR64:$op, 0)>;
299
300//===----------------------------------------------------------------------===//
301// String Pseudo Instructions
302//
303let Defs = [ECX,EDI,ESI], Uses = [ECX,EDI,ESI], isCodeGenOnly = 1 in {
304def REP_MOVSB_32 : I<0xA4, RawFrm, (outs), (ins), "{rep;movsb|rep movsb}",
305                    [(X86rep_movs i8)], IIC_REP_MOVS>, REP,
306                   Requires<[In32BitMode]>;
307def REP_MOVSW_32 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsw|rep movsw}",
308                    [(X86rep_movs i16)], IIC_REP_MOVS>, REP, OpSize,
309                   Requires<[In32BitMode]>;
310def REP_MOVSD_32 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsl|rep movsd}",
311                    [(X86rep_movs i32)], IIC_REP_MOVS>, REP,
312                   Requires<[In32BitMode]>;
313}
314
315let Defs = [RCX,RDI,RSI], Uses = [RCX,RDI,RSI], isCodeGenOnly = 1 in {
316def REP_MOVSB_64 : I<0xA4, RawFrm, (outs), (ins), "{rep;movsb|rep movsb}",
317                    [(X86rep_movs i8)], IIC_REP_MOVS>, REP,
318                   Requires<[In64BitMode]>;
319def REP_MOVSW_64 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsw|rep movsw}",
320                    [(X86rep_movs i16)], IIC_REP_MOVS>, REP, OpSize,
321                   Requires<[In64BitMode]>;
322def REP_MOVSD_64 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsl|rep movsd}",
323                    [(X86rep_movs i32)], IIC_REP_MOVS>, REP,
324                   Requires<[In64BitMode]>;
325def REP_MOVSQ_64 : RI<0xA5, RawFrm, (outs), (ins), "{rep;movsq|rep movsq}",
326                    [(X86rep_movs i64)], IIC_REP_MOVS>, REP,
327                   Requires<[In64BitMode]>;
328}
329
330// FIXME: Should use "(X86rep_stos AL)" as the pattern.
331let Defs = [ECX,EDI], isCodeGenOnly = 1 in {
332  let Uses = [AL,ECX,EDI] in
333  def REP_STOSB_32 : I<0xAA, RawFrm, (outs), (ins), "{rep;stosb|rep stosb}",
334                      [(X86rep_stos i8)], IIC_REP_STOS>, REP,
335                     Requires<[In32BitMode]>;
336  let Uses = [AX,ECX,EDI] in
337  def REP_STOSW_32 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosw|rep stosw}",
338                      [(X86rep_stos i16)], IIC_REP_STOS>, REP, OpSize,
339                     Requires<[In32BitMode]>;
340  let Uses = [EAX,ECX,EDI] in
341  def REP_STOSD_32 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosl|rep stosd}",
342                      [(X86rep_stos i32)], IIC_REP_STOS>, REP,
343                     Requires<[In32BitMode]>;
344}
345
346let Defs = [RCX,RDI], isCodeGenOnly = 1 in {
347  let Uses = [AL,RCX,RDI] in
348  def REP_STOSB_64 : I<0xAA, RawFrm, (outs), (ins), "{rep;stosb|rep stosb}",
349                      [(X86rep_stos i8)], IIC_REP_STOS>, REP,
350                     Requires<[In64BitMode]>;
351  let Uses = [AX,RCX,RDI] in
352  def REP_STOSW_64 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosw|rep stosw}",
353                      [(X86rep_stos i16)], IIC_REP_STOS>, REP, OpSize,
354                     Requires<[In64BitMode]>;
355  let Uses = [RAX,RCX,RDI] in
356  def REP_STOSD_64 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosl|rep stosd}",
357                      [(X86rep_stos i32)], IIC_REP_STOS>, REP,
358                     Requires<[In64BitMode]>;
359
360  let Uses = [RAX,RCX,RDI] in
361  def REP_STOSQ_64 : RI<0xAB, RawFrm, (outs), (ins), "{rep;stosq|rep stosq}",
362                      [(X86rep_stos i64)], IIC_REP_STOS>, REP,
363                     Requires<[In64BitMode]>;
364}
365
366//===----------------------------------------------------------------------===//
367// Thread Local Storage Instructions
368//
369
370// ELF TLS Support
371// All calls clobber the non-callee saved registers. ESP is marked as
372// a use to prevent stack-pointer assignments that appear immediately
373// before calls from potentially appearing dead.
374let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0,
375            MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
376            XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
377            XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
378    Uses = [ESP] in {
379def TLS_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
380                  "# TLS_addr32",
381                  [(X86tlsaddr tls32addr:$sym)]>,
382                  Requires<[In32BitMode]>;
383def TLS_base_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
384                  "# TLS_base_addr32",
385                  [(X86tlsbaseaddr tls32baseaddr:$sym)]>,
386                  Requires<[In32BitMode]>;
387}
388
389// All calls clobber the non-callee saved registers. RSP is marked as
390// a use to prevent stack-pointer assignments that appear immediately
391// before calls from potentially appearing dead.
392let Defs = [RAX, RCX, RDX, RSI, RDI, R8, R9, R10, R11,
393            FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0, ST1,
394            MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
395            XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
396            XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
397    Uses = [RSP] in {
398def TLS_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
399                   "# TLS_addr64",
400                  [(X86tlsaddr tls64addr:$sym)]>,
401                  Requires<[In64BitMode]>;
402def TLS_base_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
403                   "# TLS_base_addr64",
404                  [(X86tlsbaseaddr tls64baseaddr:$sym)]>,
405                  Requires<[In64BitMode]>;
406}
407
408// Darwin TLS Support
409// For i386, the address of the thunk is passed on the stack, on return the
410// address of the variable is in %eax.  %ecx is trashed during the function
411// call.  All other registers are preserved.
412let Defs = [EAX, ECX, EFLAGS],
413    Uses = [ESP],
414    usesCustomInserter = 1 in
415def TLSCall_32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
416                "# TLSCall_32",
417                [(X86TLSCall addr:$sym)]>,
418                Requires<[In32BitMode]>;
419
420// For x86_64, the address of the thunk is passed in %rdi, on return
421// the address of the variable is in %rax.  All other registers are preserved.
422let Defs = [RAX, EFLAGS],
423    Uses = [RSP, RDI],
424    usesCustomInserter = 1 in
425def TLSCall_64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
426                  "# TLSCall_64",
427                  [(X86TLSCall addr:$sym)]>,
428                  Requires<[In64BitMode]>;
429
430
431//===----------------------------------------------------------------------===//
432// Conditional Move Pseudo Instructions
433
434// X86 doesn't have 8-bit conditional moves. Use a customInserter to
435// emit control flow. An alternative to this is to mark i8 SELECT as Promote,
436// however that requires promoting the operands, and can induce additional
437// i8 register pressure.
438let usesCustomInserter = 1, Uses = [EFLAGS] in {
439def CMOV_GR8 : I<0, Pseudo,
440                 (outs GR8:$dst), (ins GR8:$src1, GR8:$src2, i8imm:$cond),
441                 "#CMOV_GR8 PSEUDO!",
442                 [(set GR8:$dst, (X86cmov GR8:$src1, GR8:$src2,
443                                          imm:$cond, EFLAGS))]>;
444
445let Predicates = [NoCMov] in {
446def CMOV_GR32 : I<0, Pseudo,
447                    (outs GR32:$dst), (ins GR32:$src1, GR32:$src2, i8imm:$cond),
448                    "#CMOV_GR32* PSEUDO!",
449                    [(set GR32:$dst,
450                      (X86cmov GR32:$src1, GR32:$src2, imm:$cond, EFLAGS))]>;
451def CMOV_GR16 : I<0, Pseudo,
452                    (outs GR16:$dst), (ins GR16:$src1, GR16:$src2, i8imm:$cond),
453                    "#CMOV_GR16* PSEUDO!",
454                    [(set GR16:$dst,
455                      (X86cmov GR16:$src1, GR16:$src2, imm:$cond, EFLAGS))]>;
456def CMOV_RFP32 : I<0, Pseudo,
457                    (outs RFP32:$dst),
458                    (ins RFP32:$src1, RFP32:$src2, i8imm:$cond),
459                    "#CMOV_RFP32 PSEUDO!",
460                    [(set RFP32:$dst,
461                      (X86cmov RFP32:$src1, RFP32:$src2, imm:$cond,
462                                                  EFLAGS))]>;
463def CMOV_RFP64 : I<0, Pseudo,
464                    (outs RFP64:$dst),
465                    (ins RFP64:$src1, RFP64:$src2, i8imm:$cond),
466                    "#CMOV_RFP64 PSEUDO!",
467                    [(set RFP64:$dst,
468                      (X86cmov RFP64:$src1, RFP64:$src2, imm:$cond,
469                                                  EFLAGS))]>;
470def CMOV_RFP80 : I<0, Pseudo,
471                    (outs RFP80:$dst),
472                    (ins RFP80:$src1, RFP80:$src2, i8imm:$cond),
473                    "#CMOV_RFP80 PSEUDO!",
474                    [(set RFP80:$dst,
475                      (X86cmov RFP80:$src1, RFP80:$src2, imm:$cond,
476                                                  EFLAGS))]>;
477} // Predicates = [NoCMov]
478} // UsesCustomInserter = 1, Uses = [EFLAGS]
479
480
481//===----------------------------------------------------------------------===//
482// Atomic Instruction Pseudo Instructions
483//===----------------------------------------------------------------------===//
484
485// Atomic exchange, and, or, xor
486let Constraints = "$val = $dst", Defs = [EFLAGS],
487                  usesCustomInserter = 1 in {
488
489def ATOMAND8 : I<0, Pseudo, (outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
490               "#ATOMAND8 PSEUDO!",
491               [(set GR8:$dst, (atomic_load_and_8 addr:$ptr, GR8:$val))]>;
492def ATOMOR8 : I<0, Pseudo, (outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
493               "#ATOMOR8 PSEUDO!",
494               [(set GR8:$dst, (atomic_load_or_8 addr:$ptr, GR8:$val))]>;
495def ATOMXOR8 : I<0, Pseudo,(outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
496               "#ATOMXOR8 PSEUDO!",
497               [(set GR8:$dst, (atomic_load_xor_8 addr:$ptr, GR8:$val))]>;
498def ATOMNAND8 : I<0, Pseudo,(outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
499               "#ATOMNAND8 PSEUDO!",
500               [(set GR8:$dst, (atomic_load_nand_8 addr:$ptr, GR8:$val))]>;
501
502def ATOMAND16 : I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
503               "#ATOMAND16 PSEUDO!",
504               [(set GR16:$dst, (atomic_load_and_16 addr:$ptr, GR16:$val))]>;
505def ATOMOR16 : I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
506               "#ATOMOR16 PSEUDO!",
507               [(set GR16:$dst, (atomic_load_or_16 addr:$ptr, GR16:$val))]>;
508def ATOMXOR16 : I<0, Pseudo,(outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
509               "#ATOMXOR16 PSEUDO!",
510               [(set GR16:$dst, (atomic_load_xor_16 addr:$ptr, GR16:$val))]>;
511def ATOMNAND16 : I<0, Pseudo,(outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
512               "#ATOMNAND16 PSEUDO!",
513               [(set GR16:$dst, (atomic_load_nand_16 addr:$ptr, GR16:$val))]>;
514def ATOMMIN16: I<0, Pseudo, (outs GR16:$dst), (ins i16mem:$ptr, GR16:$val),
515               "#ATOMMIN16 PSEUDO!",
516               [(set GR16:$dst, (atomic_load_min_16 addr:$ptr, GR16:$val))]>;
517def ATOMMAX16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
518               "#ATOMMAX16 PSEUDO!",
519               [(set GR16:$dst, (atomic_load_max_16 addr:$ptr, GR16:$val))]>;
520def ATOMUMIN16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
521               "#ATOMUMIN16 PSEUDO!",
522               [(set GR16:$dst, (atomic_load_umin_16 addr:$ptr, GR16:$val))]>;
523def ATOMUMAX16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
524               "#ATOMUMAX16 PSEUDO!",
525               [(set GR16:$dst, (atomic_load_umax_16 addr:$ptr, GR16:$val))]>;
526
527
528def ATOMAND32 : I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
529               "#ATOMAND32 PSEUDO!",
530               [(set GR32:$dst, (atomic_load_and_32 addr:$ptr, GR32:$val))]>;
531def ATOMOR32 : I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
532               "#ATOMOR32 PSEUDO!",
533               [(set GR32:$dst, (atomic_load_or_32 addr:$ptr, GR32:$val))]>;
534def ATOMXOR32 : I<0, Pseudo,(outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
535               "#ATOMXOR32 PSEUDO!",
536               [(set GR32:$dst, (atomic_load_xor_32 addr:$ptr, GR32:$val))]>;
537def ATOMNAND32 : I<0, Pseudo,(outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
538               "#ATOMNAND32 PSEUDO!",
539               [(set GR32:$dst, (atomic_load_nand_32 addr:$ptr, GR32:$val))]>;
540def ATOMMIN32: I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$ptr, GR32:$val),
541               "#ATOMMIN32 PSEUDO!",
542               [(set GR32:$dst, (atomic_load_min_32 addr:$ptr, GR32:$val))]>;
543def ATOMMAX32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
544               "#ATOMMAX32 PSEUDO!",
545               [(set GR32:$dst, (atomic_load_max_32 addr:$ptr, GR32:$val))]>;
546def ATOMUMIN32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
547               "#ATOMUMIN32 PSEUDO!",
548               [(set GR32:$dst, (atomic_load_umin_32 addr:$ptr, GR32:$val))]>;
549def ATOMUMAX32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
550               "#ATOMUMAX32 PSEUDO!",
551               [(set GR32:$dst, (atomic_load_umax_32 addr:$ptr, GR32:$val))]>;
552
553
554
555def ATOMAND64 : I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
556               "#ATOMAND64 PSEUDO!",
557               [(set GR64:$dst, (atomic_load_and_64 addr:$ptr, GR64:$val))]>;
558def ATOMOR64 : I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
559               "#ATOMOR64 PSEUDO!",
560               [(set GR64:$dst, (atomic_load_or_64 addr:$ptr, GR64:$val))]>;
561def ATOMXOR64 : I<0, Pseudo,(outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
562               "#ATOMXOR64 PSEUDO!",
563               [(set GR64:$dst, (atomic_load_xor_64 addr:$ptr, GR64:$val))]>;
564def ATOMNAND64 : I<0, Pseudo,(outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
565               "#ATOMNAND64 PSEUDO!",
566               [(set GR64:$dst, (atomic_load_nand_64 addr:$ptr, GR64:$val))]>;
567def ATOMMIN64: I<0, Pseudo, (outs GR64:$dst), (ins i64mem:$ptr, GR64:$val),
568               "#ATOMMIN64 PSEUDO!",
569               [(set GR64:$dst, (atomic_load_min_64 addr:$ptr, GR64:$val))]>;
570def ATOMMAX64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
571               "#ATOMMAX64 PSEUDO!",
572               [(set GR64:$dst, (atomic_load_max_64 addr:$ptr, GR64:$val))]>;
573def ATOMUMIN64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
574               "#ATOMUMIN64 PSEUDO!",
575               [(set GR64:$dst, (atomic_load_umin_64 addr:$ptr, GR64:$val))]>;
576def ATOMUMAX64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
577               "#ATOMUMAX64 PSEUDO!",
578               [(set GR64:$dst, (atomic_load_umax_64 addr:$ptr, GR64:$val))]>;
579}
580
581let Constraints = "$val1 = $dst1, $val2 = $dst2",
582                  Defs = [EFLAGS, EAX, EBX, ECX, EDX],
583                  Uses = [EAX, EBX, ECX, EDX],
584                  mayLoad = 1, mayStore = 1,
585                  usesCustomInserter = 1 in {
586def ATOMAND6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
587                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
588               "#ATOMAND6432 PSEUDO!", []>;
589def ATOMOR6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
590                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
591               "#ATOMOR6432 PSEUDO!", []>;
592def ATOMXOR6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
593                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
594               "#ATOMXOR6432 PSEUDO!", []>;
595def ATOMNAND6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
596                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
597               "#ATOMNAND6432 PSEUDO!", []>;
598def ATOMADD6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
599                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
600               "#ATOMADD6432 PSEUDO!", []>;
601def ATOMSUB6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
602                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
603               "#ATOMSUB6432 PSEUDO!", []>;
604def ATOMSWAP6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
605                               (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
606               "#ATOMSWAP6432 PSEUDO!", []>;
607}
608
609//===----------------------------------------------------------------------===//
610// Normal-Instructions-With-Lock-Prefix Pseudo Instructions
611//===----------------------------------------------------------------------===//
612
613// FIXME: Use normal instructions and add lock prefix dynamically.
614
615// Memory barriers
616
617// TODO: Get this to fold the constant into the instruction.
618let isCodeGenOnly = 1, Defs = [EFLAGS] in
619def OR32mrLocked  : I<0x09, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$zero),
620                      "lock\n\t"
621                      "or{l}\t{$zero, $dst|$dst, $zero}",
622                      [], IIC_ALU_MEM>, Requires<[In32BitMode]>, LOCK;
623
624let hasSideEffects = 1 in
625def Int_MemBarrier : I<0, Pseudo, (outs), (ins),
626                     "#MEMBARRIER",
627                     [(X86MemBarrier)]>;
628
629// RegOpc corresponds to the mr version of the instruction
630// ImmOpc corresponds to the mi version of the instruction
631// ImmOpc8 corresponds to the mi8 version of the instruction
632// ImmMod corresponds to the instruction format of the mi and mi8 versions
633multiclass LOCK_ArithBinOp<bits<8> RegOpc, bits<8> ImmOpc, bits<8> ImmOpc8,
634                           Format ImmMod, string mnemonic> {
635let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1 in {
636
637def #NAME#8mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
638                   RegOpc{3}, RegOpc{2}, RegOpc{1}, 0 },
639                   MRMDestMem, (outs), (ins i8mem:$dst, GR8:$src2),
640                   !strconcat("lock\n\t", mnemonic, "{b}\t",
641                              "{$src2, $dst|$dst, $src2}"),
642                   [], IIC_ALU_NONMEM>, LOCK;
643def #NAME#16mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
644                    RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
645                    MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
646                    !strconcat("lock\n\t", mnemonic, "{w}\t",
647                               "{$src2, $dst|$dst, $src2}"),
648                    [], IIC_ALU_NONMEM>, OpSize, LOCK;
649def #NAME#32mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
650                    RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
651                    MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
652                    !strconcat("lock\n\t", mnemonic, "{l}\t",
653                               "{$src2, $dst|$dst, $src2}"),
654                    [], IIC_ALU_NONMEM>, LOCK;
655def #NAME#64mr : RI<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
656                     RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
657                     MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
658                     !strconcat("lock\n\t", mnemonic, "{q}\t",
659                                "{$src2, $dst|$dst, $src2}"),
660                     [], IIC_ALU_NONMEM>, LOCK;
661
662def #NAME#8mi : Ii8<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
663                     ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 0 },
664                     ImmMod, (outs), (ins i8mem :$dst, i8imm :$src2),
665                     !strconcat("lock\n\t", mnemonic, "{b}\t",
666                                "{$src2, $dst|$dst, $src2}"),
667                     [], IIC_ALU_MEM>, LOCK;
668
669def #NAME#16mi : Ii16<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
670                       ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
671                       ImmMod, (outs), (ins i16mem :$dst, i16imm :$src2),
672                       !strconcat("lock\n\t", mnemonic, "{w}\t",
673                                  "{$src2, $dst|$dst, $src2}"),
674                       [], IIC_ALU_MEM>, LOCK;
675
676def #NAME#32mi : Ii32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
677                       ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
678                       ImmMod, (outs), (ins i32mem :$dst, i32imm :$src2),
679                       !strconcat("lock\n\t", mnemonic, "{l}\t",
680                                  "{$src2, $dst|$dst, $src2}"),
681                       [], IIC_ALU_MEM>, LOCK;
682
683def #NAME#64mi32 : RIi32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
684                          ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
685                          ImmMod, (outs), (ins i64mem :$dst, i64i32imm :$src2),
686                          !strconcat("lock\n\t", mnemonic, "{q}\t",
687                                     "{$src2, $dst|$dst, $src2}"),
688                          [], IIC_ALU_MEM>, LOCK;
689
690def #NAME#16mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
691                       ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
692                       ImmMod, (outs), (ins i16mem :$dst, i16i8imm :$src2),
693                       !strconcat("lock\n\t", mnemonic, "{w}\t",
694                                  "{$src2, $dst|$dst, $src2}"),
695                       [], IIC_ALU_MEM>, LOCK;
696def #NAME#32mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
697                       ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
698                       ImmMod, (outs), (ins i32mem :$dst, i32i8imm :$src2),
699                       !strconcat("lock\n\t", mnemonic, "{l}\t",
700                                  "{$src2, $dst|$dst, $src2}"),
701                       [], IIC_ALU_MEM>, LOCK;
702def #NAME#64mi8 : RIi8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
703                        ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
704                        ImmMod, (outs), (ins i64mem :$dst, i64i8imm :$src2),
705                        !strconcat("lock\n\t", mnemonic, "{q}\t",
706                                   "{$src2, $dst|$dst, $src2}"),
707                        [], IIC_ALU_MEM>, LOCK;
708
709}
710
711}
712
713defm LOCK_ADD : LOCK_ArithBinOp<0x00, 0x80, 0x83, MRM0m, "add">;
714defm LOCK_SUB : LOCK_ArithBinOp<0x28, 0x80, 0x83, MRM5m, "sub">;
715defm LOCK_OR  : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM1m, "or">;
716defm LOCK_AND : LOCK_ArithBinOp<0x20, 0x80, 0x83, MRM4m, "and">;
717defm LOCK_XOR : LOCK_ArithBinOp<0x30, 0x80, 0x83, MRM6m, "xor">;
718
719// Optimized codegen when the non-memory output is not used.
720let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1 in {
721
722def LOCK_INC8m  : I<0xFE, MRM0m, (outs), (ins i8mem :$dst),
723                    "lock\n\t"
724                    "inc{b}\t$dst", [], IIC_UNARY_MEM>, LOCK;
725def LOCK_INC16m : I<0xFF, MRM0m, (outs), (ins i16mem:$dst),
726                    "lock\n\t"
727                    "inc{w}\t$dst", [], IIC_UNARY_MEM>, OpSize, LOCK;
728def LOCK_INC32m : I<0xFF, MRM0m, (outs), (ins i32mem:$dst),
729                    "lock\n\t"
730                    "inc{l}\t$dst", [], IIC_UNARY_MEM>, LOCK;
731def LOCK_INC64m : RI<0xFF, MRM0m, (outs), (ins i64mem:$dst),
732                     "lock\n\t"
733                     "inc{q}\t$dst", [], IIC_UNARY_MEM>, LOCK;
734
735def LOCK_DEC8m  : I<0xFE, MRM1m, (outs), (ins i8mem :$dst),
736                    "lock\n\t"
737                    "dec{b}\t$dst", [], IIC_UNARY_MEM>, LOCK;
738def LOCK_DEC16m : I<0xFF, MRM1m, (outs), (ins i16mem:$dst),
739                    "lock\n\t"
740                    "dec{w}\t$dst", [], IIC_UNARY_MEM>, OpSize, LOCK;
741def LOCK_DEC32m : I<0xFF, MRM1m, (outs), (ins i32mem:$dst),
742                    "lock\n\t"
743                    "dec{l}\t$dst", [], IIC_UNARY_MEM>, LOCK;
744def LOCK_DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst),
745                      "lock\n\t"
746                      "dec{q}\t$dst", [], IIC_UNARY_MEM>, LOCK;
747}
748
749// Atomic compare and swap.
750let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
751    isCodeGenOnly = 1 in
752def LCMPXCHG8B : I<0xC7, MRM1m, (outs), (ins i64mem:$ptr),
753               "lock\n\t"
754               "cmpxchg8b\t$ptr",
755               [(X86cas8 addr:$ptr)], IIC_CMPX_LOCK_8B>, TB, LOCK;
756
757let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RBX, RCX, RDX],
758    isCodeGenOnly = 1 in
759def LCMPXCHG16B : RI<0xC7, MRM1m, (outs), (ins i128mem:$ptr),
760                    "lock\n\t"
761                    "cmpxchg16b\t$ptr",
762                    [(X86cas16 addr:$ptr)], IIC_CMPX_LOCK_16B>, TB, LOCK,
763                    Requires<[HasCmpxchg16b]>;
764
765let Defs = [AL, EFLAGS], Uses = [AL], isCodeGenOnly = 1 in {
766def LCMPXCHG8 : I<0xB0, MRMDestMem, (outs), (ins i8mem:$ptr, GR8:$swap),
767               "lock\n\t"
768               "cmpxchg{b}\t{$swap, $ptr|$ptr, $swap}",
769               [(X86cas addr:$ptr, GR8:$swap, 1)], IIC_CMPX_LOCK_8>, TB, LOCK;
770}
771
772let Defs = [AX, EFLAGS], Uses = [AX], isCodeGenOnly = 1 in {
773def LCMPXCHG16 : I<0xB1, MRMDestMem, (outs), (ins i16mem:$ptr, GR16:$swap),
774               "lock\n\t"
775               "cmpxchg{w}\t{$swap, $ptr|$ptr, $swap}",
776               [(X86cas addr:$ptr, GR16:$swap, 2)], IIC_CMPX_LOCK>, TB, OpSize, LOCK;
777}
778
779let Defs = [EAX, EFLAGS], Uses = [EAX], isCodeGenOnly = 1 in {
780def LCMPXCHG32 : I<0xB1, MRMDestMem, (outs), (ins i32mem:$ptr, GR32:$swap),
781               "lock\n\t"
782               "cmpxchg{l}\t{$swap, $ptr|$ptr, $swap}",
783               [(X86cas addr:$ptr, GR32:$swap, 4)], IIC_CMPX_LOCK>, TB, LOCK;
784}
785
786let Defs = [RAX, EFLAGS], Uses = [RAX], isCodeGenOnly = 1 in {
787def LCMPXCHG64 : RI<0xB1, MRMDestMem, (outs), (ins i64mem:$ptr, GR64:$swap),
788               "lock\n\t"
789               "cmpxchg{q}\t{$swap, $ptr|$ptr, $swap}",
790               [(X86cas addr:$ptr, GR64:$swap, 8)], IIC_CMPX_LOCK>, TB, LOCK;
791}
792
793// Atomic exchange and add
794let Constraints = "$val = $dst", Defs = [EFLAGS], isCodeGenOnly = 1 in {
795def LXADD8  : I<0xC0, MRMSrcMem, (outs GR8:$dst), (ins GR8:$val, i8mem:$ptr),
796               "lock\n\t"
797               "xadd{b}\t{$val, $ptr|$ptr, $val}",
798               [(set GR8:$dst, (atomic_load_add_8 addr:$ptr, GR8:$val))],
799                IIC_XADD_LOCK_MEM8>,
800                TB, LOCK;
801def LXADD16 : I<0xC1, MRMSrcMem, (outs GR16:$dst), (ins GR16:$val, i16mem:$ptr),
802               "lock\n\t"
803               "xadd{w}\t{$val, $ptr|$ptr, $val}",
804               [(set GR16:$dst, (atomic_load_add_16 addr:$ptr, GR16:$val))],
805                IIC_XADD_LOCK_MEM>,
806                TB, OpSize, LOCK;
807def LXADD32 : I<0xC1, MRMSrcMem, (outs GR32:$dst), (ins GR32:$val, i32mem:$ptr),
808               "lock\n\t"
809               "xadd{l}\t{$val, $ptr|$ptr, $val}",
810               [(set GR32:$dst, (atomic_load_add_32 addr:$ptr, GR32:$val))],
811                IIC_XADD_LOCK_MEM>,
812                TB, LOCK;
813def LXADD64 : RI<0xC1, MRMSrcMem, (outs GR64:$dst), (ins GR64:$val,i64mem:$ptr),
814               "lock\n\t"
815               "xadd{q}\t{$val, $ptr|$ptr, $val}",
816               [(set GR64:$dst, (atomic_load_add_64 addr:$ptr, GR64:$val))],
817                IIC_XADD_LOCK_MEM>,
818                TB, LOCK;
819}
820
821def ACQUIRE_MOV8rm  : I<0, Pseudo, (outs GR8 :$dst), (ins i8mem :$src),
822                      "#ACQUIRE_MOV PSEUDO!",
823                      [(set GR8:$dst,  (atomic_load_8  addr:$src))]>;
824def ACQUIRE_MOV16rm : I<0, Pseudo, (outs GR16:$dst), (ins i16mem:$src),
825                      "#ACQUIRE_MOV PSEUDO!",
826                      [(set GR16:$dst, (atomic_load_16 addr:$src))]>;
827def ACQUIRE_MOV32rm : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$src),
828                      "#ACQUIRE_MOV PSEUDO!",
829                      [(set GR32:$dst, (atomic_load_32 addr:$src))]>;
830def ACQUIRE_MOV64rm : I<0, Pseudo, (outs GR64:$dst), (ins i64mem:$src),
831                      "#ACQUIRE_MOV PSEUDO!",
832                      [(set GR64:$dst, (atomic_load_64 addr:$src))]>;
833
834def RELEASE_MOV8mr  : I<0, Pseudo, (outs), (ins i8mem :$dst, GR8 :$src),
835                        "#RELEASE_MOV PSEUDO!",
836                        [(atomic_store_8  addr:$dst, GR8 :$src)]>;
837def RELEASE_MOV16mr : I<0, Pseudo, (outs), (ins i16mem:$dst, GR16:$src),
838                        "#RELEASE_MOV PSEUDO!",
839                        [(atomic_store_16 addr:$dst, GR16:$src)]>;
840def RELEASE_MOV32mr : I<0, Pseudo, (outs), (ins i32mem:$dst, GR32:$src),
841                        "#RELEASE_MOV PSEUDO!",
842                        [(atomic_store_32 addr:$dst, GR32:$src)]>;
843def RELEASE_MOV64mr : I<0, Pseudo, (outs), (ins i64mem:$dst, GR64:$src),
844                        "#RELEASE_MOV PSEUDO!",
845                        [(atomic_store_64 addr:$dst, GR64:$src)]>;
846
847//===----------------------------------------------------------------------===//
848// Conditional Move Pseudo Instructions.
849//===----------------------------------------------------------------------===//
850
851
852// CMOV* - Used to implement the SSE SELECT DAG operation.  Expanded after
853// instruction selection into a branch sequence.
854let Uses = [EFLAGS], usesCustomInserter = 1 in {
855  def CMOV_FR32 : I<0, Pseudo,
856                    (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
857                    "#CMOV_FR32 PSEUDO!",
858                    [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
859                                                  EFLAGS))]>;
860  def CMOV_FR64 : I<0, Pseudo,
861                    (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
862                    "#CMOV_FR64 PSEUDO!",
863                    [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
864                                                  EFLAGS))]>;
865  def CMOV_V4F32 : I<0, Pseudo,
866                    (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
867                    "#CMOV_V4F32 PSEUDO!",
868                    [(set VR128:$dst,
869                      (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
870                                          EFLAGS)))]>;
871  def CMOV_V2F64 : I<0, Pseudo,
872                    (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
873                    "#CMOV_V2F64 PSEUDO!",
874                    [(set VR128:$dst,
875                      (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
876                                          EFLAGS)))]>;
877  def CMOV_V2I64 : I<0, Pseudo,
878                    (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
879                    "#CMOV_V2I64 PSEUDO!",
880                    [(set VR128:$dst,
881                      (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
882                                          EFLAGS)))]>;
883  def CMOV_V8F32 : I<0, Pseudo,
884                    (outs VR256:$dst), (ins VR256:$t, VR256:$f, i8imm:$cond),
885                    "#CMOV_V8F32 PSEUDO!",
886                    [(set VR256:$dst,
887                      (v8f32 (X86cmov VR256:$t, VR256:$f, imm:$cond,
888                                          EFLAGS)))]>;
889  def CMOV_V4F64 : I<0, Pseudo,
890                    (outs VR256:$dst), (ins VR256:$t, VR256:$f, i8imm:$cond),
891                    "#CMOV_V4F64 PSEUDO!",
892                    [(set VR256:$dst,
893                      (v4f64 (X86cmov VR256:$t, VR256:$f, imm:$cond,
894                                          EFLAGS)))]>;
895  def CMOV_V4I64 : I<0, Pseudo,
896                    (outs VR256:$dst), (ins VR256:$t, VR256:$f, i8imm:$cond),
897                    "#CMOV_V4I64 PSEUDO!",
898                    [(set VR256:$dst,
899                      (v4i64 (X86cmov VR256:$t, VR256:$f, imm:$cond,
900                                          EFLAGS)))]>;
901}
902
903
904//===----------------------------------------------------------------------===//
905// DAG Pattern Matching Rules
906//===----------------------------------------------------------------------===//
907
908// ConstantPool GlobalAddress, ExternalSymbol, and JumpTable
909def : Pat<(i32 (X86Wrapper tconstpool  :$dst)), (MOV32ri tconstpool  :$dst)>;
910def : Pat<(i32 (X86Wrapper tjumptable  :$dst)), (MOV32ri tjumptable  :$dst)>;
911def : Pat<(i32 (X86Wrapper tglobaltlsaddr:$dst)),(MOV32ri tglobaltlsaddr:$dst)>;
912def : Pat<(i32 (X86Wrapper tglobaladdr :$dst)), (MOV32ri tglobaladdr :$dst)>;
913def : Pat<(i32 (X86Wrapper texternalsym:$dst)), (MOV32ri texternalsym:$dst)>;
914def : Pat<(i32 (X86Wrapper tblockaddress:$dst)), (MOV32ri tblockaddress:$dst)>;
915
916def : Pat<(add GR32:$src1, (X86Wrapper tconstpool:$src2)),
917          (ADD32ri GR32:$src1, tconstpool:$src2)>;
918def : Pat<(add GR32:$src1, (X86Wrapper tjumptable:$src2)),
919          (ADD32ri GR32:$src1, tjumptable:$src2)>;
920def : Pat<(add GR32:$src1, (X86Wrapper tglobaladdr :$src2)),
921          (ADD32ri GR32:$src1, tglobaladdr:$src2)>;
922def : Pat<(add GR32:$src1, (X86Wrapper texternalsym:$src2)),
923          (ADD32ri GR32:$src1, texternalsym:$src2)>;
924def : Pat<(add GR32:$src1, (X86Wrapper tblockaddress:$src2)),
925          (ADD32ri GR32:$src1, tblockaddress:$src2)>;
926
927def : Pat<(store (i32 (X86Wrapper tglobaladdr:$src)), addr:$dst),
928          (MOV32mi addr:$dst, tglobaladdr:$src)>;
929def : Pat<(store (i32 (X86Wrapper texternalsym:$src)), addr:$dst),
930          (MOV32mi addr:$dst, texternalsym:$src)>;
931def : Pat<(store (i32 (X86Wrapper tblockaddress:$src)), addr:$dst),
932          (MOV32mi addr:$dst, tblockaddress:$src)>;
933
934
935
936// ConstantPool GlobalAddress, ExternalSymbol, and JumpTable when not in small
937// code model mode, should use 'movabs'.  FIXME: This is really a hack, the
938//  'movabs' predicate should handle this sort of thing.
939def : Pat<(i64 (X86Wrapper tconstpool  :$dst)),
940          (MOV64ri tconstpool  :$dst)>, Requires<[FarData]>;
941def : Pat<(i64 (X86Wrapper tjumptable  :$dst)),
942          (MOV64ri tjumptable  :$dst)>, Requires<[FarData]>;
943def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
944          (MOV64ri tglobaladdr :$dst)>, Requires<[FarData]>;
945def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
946          (MOV64ri texternalsym:$dst)>, Requires<[FarData]>;
947def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
948          (MOV64ri tblockaddress:$dst)>, Requires<[FarData]>;
949
950// In static codegen with small code model, we can get the address of a label
951// into a register with 'movl'.  FIXME: This is a hack, the 'imm' predicate of
952// the MOV64ri64i32 should accept these.
953def : Pat<(i64 (X86Wrapper tconstpool  :$dst)),
954          (MOV64ri64i32 tconstpool  :$dst)>, Requires<[SmallCode]>;
955def : Pat<(i64 (X86Wrapper tjumptable  :$dst)),
956          (MOV64ri64i32 tjumptable  :$dst)>, Requires<[SmallCode]>;
957def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
958          (MOV64ri64i32 tglobaladdr :$dst)>, Requires<[SmallCode]>;
959def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
960          (MOV64ri64i32 texternalsym:$dst)>, Requires<[SmallCode]>;
961def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
962          (MOV64ri64i32 tblockaddress:$dst)>, Requires<[SmallCode]>;
963
964// In kernel code model, we can get the address of a label
965// into a register with 'movq'.  FIXME: This is a hack, the 'imm' predicate of
966// the MOV64ri32 should accept these.
967def : Pat<(i64 (X86Wrapper tconstpool  :$dst)),
968          (MOV64ri32 tconstpool  :$dst)>, Requires<[KernelCode]>;
969def : Pat<(i64 (X86Wrapper tjumptable  :$dst)),
970          (MOV64ri32 tjumptable  :$dst)>, Requires<[KernelCode]>;
971def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
972          (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
973def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
974          (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
975def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
976          (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
977
978// If we have small model and -static mode, it is safe to store global addresses
979// directly as immediates.  FIXME: This is really a hack, the 'imm' predicate
980// for MOV64mi32 should handle this sort of thing.
981def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
982          (MOV64mi32 addr:$dst, tconstpool:$src)>,
983          Requires<[NearData, IsStatic]>;
984def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
985          (MOV64mi32 addr:$dst, tjumptable:$src)>,
986          Requires<[NearData, IsStatic]>;
987def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
988          (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
989          Requires<[NearData, IsStatic]>;
990def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
991          (MOV64mi32 addr:$dst, texternalsym:$src)>,
992          Requires<[NearData, IsStatic]>;
993def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
994          (MOV64mi32 addr:$dst, tblockaddress:$src)>,
995          Requires<[NearData, IsStatic]>;
996
997
998
999// Calls
1000
1001// tls has some funny stuff here...
1002// This corresponds to movabs $foo@tpoff, %rax
1003def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
1004          (MOV64ri tglobaltlsaddr :$dst)>;
1005// This corresponds to add $foo@tpoff, %rax
1006def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
1007          (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
1008// This corresponds to mov foo@tpoff(%rbx), %eax
1009def : Pat<(load (i64 (X86Wrapper tglobaltlsaddr :$dst))),
1010          (MOV64rm tglobaltlsaddr :$dst)>;
1011
1012
1013// Direct PC relative function call for small code model. 32-bit displacement
1014// sign extended to 64-bit.
1015def : Pat<(X86call (i64 tglobaladdr:$dst)),
1016          (CALL64pcrel32 tglobaladdr:$dst)>;
1017def : Pat<(X86call (i64 texternalsym:$dst)),
1018          (CALL64pcrel32 texternalsym:$dst)>;
1019
1020// tailcall stuff
1021def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
1022          (TCRETURNri ptr_rc_tailcall:$dst, imm:$off)>,
1023          Requires<[In32BitMode]>;
1024
1025// FIXME: This is disabled for 32-bit PIC mode because the global base
1026// register which is part of the address mode may be assigned a
1027// callee-saved register.
1028def : Pat<(X86tcret (load addr:$dst), imm:$off),
1029          (TCRETURNmi addr:$dst, imm:$off)>,
1030          Requires<[In32BitMode, IsNotPIC]>;
1031
1032def : Pat<(X86tcret (i32 tglobaladdr:$dst), imm:$off),
1033          (TCRETURNdi texternalsym:$dst, imm:$off)>,
1034          Requires<[In32BitMode]>;
1035
1036def : Pat<(X86tcret (i32 texternalsym:$dst), imm:$off),
1037          (TCRETURNdi texternalsym:$dst, imm:$off)>,
1038          Requires<[In32BitMode]>;
1039
1040def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
1041          (TCRETURNri64 ptr_rc_tailcall:$dst, imm:$off)>,
1042          Requires<[In64BitMode]>;
1043
1044def : Pat<(X86tcret (load addr:$dst), imm:$off),
1045          (TCRETURNmi64 addr:$dst, imm:$off)>,
1046          Requires<[In64BitMode]>;
1047
1048def : Pat<(X86tcret (i64 tglobaladdr:$dst), imm:$off),
1049          (TCRETURNdi64 tglobaladdr:$dst, imm:$off)>,
1050          Requires<[In64BitMode]>;
1051
1052def : Pat<(X86tcret (i64 texternalsym:$dst), imm:$off),
1053          (TCRETURNdi64 texternalsym:$dst, imm:$off)>,
1054          Requires<[In64BitMode]>;
1055
1056// Normal calls, with various flavors of addresses.
1057def : Pat<(X86call (i32 tglobaladdr:$dst)),
1058          (CALLpcrel32 tglobaladdr:$dst)>;
1059def : Pat<(X86call (i32 texternalsym:$dst)),
1060          (CALLpcrel32 texternalsym:$dst)>;
1061def : Pat<(X86call (i32 imm:$dst)),
1062          (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
1063
1064// Comparisons.
1065
1066// TEST R,R is smaller than CMP R,0
1067def : Pat<(X86cmp GR8:$src1, 0),
1068          (TEST8rr GR8:$src1, GR8:$src1)>;
1069def : Pat<(X86cmp GR16:$src1, 0),
1070          (TEST16rr GR16:$src1, GR16:$src1)>;
1071def : Pat<(X86cmp GR32:$src1, 0),
1072          (TEST32rr GR32:$src1, GR32:$src1)>;
1073def : Pat<(X86cmp GR64:$src1, 0),
1074          (TEST64rr GR64:$src1, GR64:$src1)>;
1075
1076// Conditional moves with folded loads with operands swapped and conditions
1077// inverted.
1078multiclass CMOVmr<PatLeaf InvertedCond, Instruction Inst16, Instruction Inst32,
1079                  Instruction Inst64> {
1080  def : Pat<(X86cmov (loadi16 addr:$src1), GR16:$src2, InvertedCond, EFLAGS),
1081            (Inst16 GR16:$src2, addr:$src1)>;
1082  def : Pat<(X86cmov (loadi32 addr:$src1), GR32:$src2, InvertedCond, EFLAGS),
1083            (Inst32 GR32:$src2, addr:$src1)>;
1084  def : Pat<(X86cmov (loadi64 addr:$src1), GR64:$src2, InvertedCond, EFLAGS),
1085            (Inst64 GR64:$src2, addr:$src1)>;
1086}
1087
1088defm : CMOVmr<X86_COND_B , CMOVAE16rm, CMOVAE32rm, CMOVAE64rm>;
1089defm : CMOVmr<X86_COND_AE, CMOVB16rm , CMOVB32rm , CMOVB64rm>;
1090defm : CMOVmr<X86_COND_E , CMOVNE16rm, CMOVNE32rm, CMOVNE64rm>;
1091defm : CMOVmr<X86_COND_NE, CMOVE16rm , CMOVE32rm , CMOVE64rm>;
1092defm : CMOVmr<X86_COND_BE, CMOVA16rm , CMOVA32rm , CMOVA64rm>;
1093defm : CMOVmr<X86_COND_A , CMOVBE16rm, CMOVBE32rm, CMOVBE64rm>;
1094defm : CMOVmr<X86_COND_L , CMOVGE16rm, CMOVGE32rm, CMOVGE64rm>;
1095defm : CMOVmr<X86_COND_GE, CMOVL16rm , CMOVL32rm , CMOVL64rm>;
1096defm : CMOVmr<X86_COND_LE, CMOVG16rm , CMOVG32rm , CMOVG64rm>;
1097defm : CMOVmr<X86_COND_G , CMOVLE16rm, CMOVLE32rm, CMOVLE64rm>;
1098defm : CMOVmr<X86_COND_P , CMOVNP16rm, CMOVNP32rm, CMOVNP64rm>;
1099defm : CMOVmr<X86_COND_NP, CMOVP16rm , CMOVP32rm , CMOVP64rm>;
1100defm : CMOVmr<X86_COND_S , CMOVNS16rm, CMOVNS32rm, CMOVNS64rm>;
1101defm : CMOVmr<X86_COND_NS, CMOVS16rm , CMOVS32rm , CMOVS64rm>;
1102defm : CMOVmr<X86_COND_O , CMOVNO16rm, CMOVNO32rm, CMOVNO64rm>;
1103defm : CMOVmr<X86_COND_NO, CMOVO16rm , CMOVO32rm , CMOVO64rm>;
1104
1105// zextload bool -> zextload byte
1106def : Pat<(zextloadi8i1  addr:$src), (MOV8rm     addr:$src)>;
1107def : Pat<(zextloadi16i1 addr:$src), (MOVZX16rm8 addr:$src)>;
1108def : Pat<(zextloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1109def : Pat<(zextloadi64i1 addr:$src), (MOVZX64rm8 addr:$src)>;
1110
1111// extload bool -> extload byte
1112// When extloading from 16-bit and smaller memory locations into 64-bit
1113// registers, use zero-extending loads so that the entire 64-bit register is
1114// defined, avoiding partial-register updates.
1115
1116def : Pat<(extloadi8i1 addr:$src),   (MOV8rm      addr:$src)>;
1117def : Pat<(extloadi16i1 addr:$src),  (MOVZX16rm8  addr:$src)>;
1118def : Pat<(extloadi32i1 addr:$src),  (MOVZX32rm8  addr:$src)>;
1119def : Pat<(extloadi16i8 addr:$src),  (MOVZX16rm8  addr:$src)>;
1120def : Pat<(extloadi32i8 addr:$src),  (MOVZX32rm8  addr:$src)>;
1121def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
1122
1123def : Pat<(extloadi64i1 addr:$src),  (MOVZX64rm8  addr:$src)>;
1124def : Pat<(extloadi64i8 addr:$src),  (MOVZX64rm8  addr:$src)>;
1125def : Pat<(extloadi64i16 addr:$src), (MOVZX64rm16 addr:$src)>;
1126// For other extloads, use subregs, since the high contents of the register are
1127// defined after an extload.
1128def : Pat<(extloadi64i32 addr:$src),
1129          (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src),
1130                         sub_32bit)>;
1131
1132// anyext. Define these to do an explicit zero-extend to
1133// avoid partial-register updates.
1134def : Pat<(i16 (anyext GR8 :$src)), (EXTRACT_SUBREG
1135                                     (MOVZX32rr8 GR8 :$src), sub_16bit)>;
1136def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8  GR8 :$src)>;
1137
1138// Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
1139def : Pat<(i32 (anyext GR16:$src)),
1140          (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
1141
1142def : Pat<(i64 (anyext GR8 :$src)), (MOVZX64rr8  GR8  :$src)>;
1143def : Pat<(i64 (anyext GR16:$src)), (MOVZX64rr16 GR16 :$src)>;
1144def : Pat<(i64 (anyext GR32:$src)),
1145          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1146
1147
1148// Any instruction that defines a 32-bit result leaves the high half of the
1149// register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
1150// be copying from a truncate. And x86's cmov doesn't do anything if the
1151// condition is false. But any other 32-bit operation will zero-extend
1152// up to 64 bits.
1153def def32 : PatLeaf<(i32 GR32:$src), [{
1154  return N->getOpcode() != ISD::TRUNCATE &&
1155         N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
1156         N->getOpcode() != ISD::CopyFromReg &&
1157         N->getOpcode() != X86ISD::CMOV;
1158}]>;
1159
1160// In the case of a 32-bit def that is known to implicitly zero-extend,
1161// we can use a SUBREG_TO_REG.
1162def : Pat<(i64 (zext def32:$src)),
1163          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1164
1165//===----------------------------------------------------------------------===//
1166// Pattern match OR as ADD
1167//===----------------------------------------------------------------------===//
1168
1169// If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
1170// 3-addressified into an LEA instruction to avoid copies.  However, we also
1171// want to finally emit these instructions as an or at the end of the code
1172// generator to make the generated code easier to read.  To do this, we select
1173// into "disjoint bits" pseudo ops.
1174
1175// Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.
1176def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
1177  if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1178    return CurDAG->MaskedValueIsZero(N->getOperand(0), CN->getAPIntValue());
1179
1180  APInt KnownZero0, KnownOne0;
1181  CurDAG->ComputeMaskedBits(N->getOperand(0), KnownZero0, KnownOne0, 0);
1182  APInt KnownZero1, KnownOne1;
1183  CurDAG->ComputeMaskedBits(N->getOperand(1), KnownZero1, KnownOne1, 0);
1184  return (~KnownZero0 & ~KnownZero1) == 0;
1185}]>;
1186
1187
1188// (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1189let AddedComplexity = 5 in { // Try this before the selecting to OR
1190
1191let isConvertibleToThreeAddress = 1,
1192    Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1193let isCommutable = 1 in {
1194def ADD16rr_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1195                    "", // orw/addw REG, REG
1196                    [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1197def ADD32rr_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1198                    "", // orl/addl REG, REG
1199                    [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1200def ADD64rr_DB  : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1201                    "", // orq/addq REG, REG
1202                    [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1203} // isCommutable
1204
1205// NOTE: These are order specific, we want the ri8 forms to be listed
1206// first so that they are slightly preferred to the ri forms.
1207
1208def ADD16ri8_DB : I<0, Pseudo,
1209                    (outs GR16:$dst), (ins GR16:$src1, i16i8imm:$src2),
1210                    "", // orw/addw REG, imm8
1211                    [(set GR16:$dst,(or_is_add GR16:$src1,i16immSExt8:$src2))]>;
1212def ADD16ri_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1213                    "", // orw/addw REG, imm
1214                    [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1215
1216def ADD32ri8_DB : I<0, Pseudo,
1217                    (outs GR32:$dst), (ins GR32:$src1, i32i8imm:$src2),
1218                    "", // orl/addl REG, imm8
1219                    [(set GR32:$dst,(or_is_add GR32:$src1,i32immSExt8:$src2))]>;
1220def ADD32ri_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1221                    "", // orl/addl REG, imm
1222                    [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1223
1224
1225def ADD64ri8_DB : I<0, Pseudo,
1226                    (outs GR64:$dst), (ins GR64:$src1, i64i8imm:$src2),
1227                    "", // orq/addq REG, imm8
1228                    [(set GR64:$dst, (or_is_add GR64:$src1,
1229                                                i64immSExt8:$src2))]>;
1230def ADD64ri32_DB : I<0, Pseudo,
1231                     (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1232                      "", // orq/addq REG, imm
1233                      [(set GR64:$dst, (or_is_add GR64:$src1,
1234                                                  i64immSExt32:$src2))]>;
1235}
1236} // AddedComplexity
1237
1238
1239//===----------------------------------------------------------------------===//
1240// Some peepholes
1241//===----------------------------------------------------------------------===//
1242
1243// Odd encoding trick: -128 fits into an 8-bit immediate field while
1244// +128 doesn't, so in this special case use a sub instead of an add.
1245def : Pat<(add GR16:$src1, 128),
1246          (SUB16ri8 GR16:$src1, -128)>;
1247def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1248          (SUB16mi8 addr:$dst, -128)>;
1249
1250def : Pat<(add GR32:$src1, 128),
1251          (SUB32ri8 GR32:$src1, -128)>;
1252def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1253          (SUB32mi8 addr:$dst, -128)>;
1254
1255def : Pat<(add GR64:$src1, 128),
1256          (SUB64ri8 GR64:$src1, -128)>;
1257def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1258          (SUB64mi8 addr:$dst, -128)>;
1259
1260// The same trick applies for 32-bit immediate fields in 64-bit
1261// instructions.
1262def : Pat<(add GR64:$src1, 0x0000000080000000),
1263          (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1264def : Pat<(store (add (loadi64 addr:$dst), 0x00000000800000000), addr:$dst),
1265          (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1266
1267// To avoid needing to materialize an immediate in a register, use a 32-bit and
1268// with implicit zero-extension instead of a 64-bit and if the immediate has at
1269// least 32 bits of leading zeros. If in addition the last 32 bits can be
1270// represented with a sign extension of a 8 bit constant, use that.
1271
1272def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
1273          (SUBREG_TO_REG
1274            (i64 0),
1275            (AND32ri8
1276              (EXTRACT_SUBREG GR64:$src, sub_32bit),
1277              (i32 (GetLo8XForm imm:$imm))),
1278            sub_32bit)>;
1279
1280def : Pat<(and GR64:$src, i64immZExt32:$imm),
1281          (SUBREG_TO_REG
1282            (i64 0),
1283            (AND32ri
1284              (EXTRACT_SUBREG GR64:$src, sub_32bit),
1285              (i32 (GetLo32XForm imm:$imm))),
1286            sub_32bit)>;
1287
1288
1289// r & (2^16-1) ==> movz
1290def : Pat<(and GR32:$src1, 0xffff),
1291          (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1292// r & (2^8-1) ==> movz
1293def : Pat<(and GR32:$src1, 0xff),
1294          (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src1,
1295                                                             GR32_ABCD)),
1296                                      sub_8bit))>,
1297      Requires<[In32BitMode]>;
1298// r & (2^8-1) ==> movz
1299def : Pat<(and GR16:$src1, 0xff),
1300           (EXTRACT_SUBREG (MOVZX32rr8 (EXTRACT_SUBREG
1301            (i16 (COPY_TO_REGCLASS GR16:$src1, GR16_ABCD)), sub_8bit)),
1302             sub_16bit)>,
1303      Requires<[In32BitMode]>;
1304
1305// r & (2^32-1) ==> movz
1306def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1307          (MOVZX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1308// r & (2^16-1) ==> movz
1309def : Pat<(and GR64:$src, 0xffff),
1310          (MOVZX64rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit)))>;
1311// r & (2^8-1) ==> movz
1312def : Pat<(and GR64:$src, 0xff),
1313          (MOVZX64rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit)))>;
1314// r & (2^8-1) ==> movz
1315def : Pat<(and GR32:$src1, 0xff),
1316           (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>,
1317      Requires<[In64BitMode]>;
1318// r & (2^8-1) ==> movz
1319def : Pat<(and GR16:$src1, 0xff),
1320           (EXTRACT_SUBREG (MOVZX32rr8 (i8
1321            (EXTRACT_SUBREG GR16:$src1, sub_8bit))), sub_16bit)>,
1322      Requires<[In64BitMode]>;
1323
1324
1325// sext_inreg patterns
1326def : Pat<(sext_inreg GR32:$src, i16),
1327          (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1328def : Pat<(sext_inreg GR32:$src, i8),
1329          (MOVSX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1330                                                             GR32_ABCD)),
1331                                      sub_8bit))>,
1332      Requires<[In32BitMode]>;
1333
1334def : Pat<(sext_inreg GR16:$src, i8),
1335           (EXTRACT_SUBREG (i32 (MOVSX32rr8 (EXTRACT_SUBREG
1336            (i32 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)), sub_8bit))),
1337             sub_16bit)>,
1338      Requires<[In32BitMode]>;
1339
1340def : Pat<(sext_inreg GR64:$src, i32),
1341          (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1342def : Pat<(sext_inreg GR64:$src, i16),
1343          (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1344def : Pat<(sext_inreg GR64:$src, i8),
1345          (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1346def : Pat<(sext_inreg GR32:$src, i8),
1347          (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>,
1348      Requires<[In64BitMode]>;
1349def : Pat<(sext_inreg GR16:$src, i8),
1350           (EXTRACT_SUBREG (MOVSX32rr8
1351            (EXTRACT_SUBREG GR16:$src, sub_8bit)), sub_16bit)>,
1352      Requires<[In64BitMode]>;
1353
1354// sext, sext_load, zext, zext_load
1355def: Pat<(i16 (sext GR8:$src)),
1356          (EXTRACT_SUBREG (MOVSX32rr8 GR8:$src), sub_16bit)>;
1357def: Pat<(sextloadi16i8 addr:$src),
1358          (EXTRACT_SUBREG (MOVSX32rm8 addr:$src), sub_16bit)>;
1359def: Pat<(i16 (zext GR8:$src)),
1360          (EXTRACT_SUBREG (MOVZX32rr8 GR8:$src), sub_16bit)>;
1361def: Pat<(zextloadi16i8 addr:$src),
1362          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1363
1364// trunc patterns
1365def : Pat<(i16 (trunc GR32:$src)),
1366          (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1367def : Pat<(i8 (trunc GR32:$src)),
1368          (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1369                          sub_8bit)>,
1370      Requires<[In32BitMode]>;
1371def : Pat<(i8 (trunc GR16:$src)),
1372          (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1373                          sub_8bit)>,
1374      Requires<[In32BitMode]>;
1375def : Pat<(i32 (trunc GR64:$src)),
1376          (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1377def : Pat<(i16 (trunc GR64:$src)),
1378          (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1379def : Pat<(i8 (trunc GR64:$src)),
1380          (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1381def : Pat<(i8 (trunc GR32:$src)),
1382          (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1383      Requires<[In64BitMode]>;
1384def : Pat<(i8 (trunc GR16:$src)),
1385          (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1386      Requires<[In64BitMode]>;
1387
1388// h-register tricks
1389def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1390          (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1391                          sub_8bit_hi)>,
1392      Requires<[In32BitMode]>;
1393def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1394          (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1395                          sub_8bit_hi)>,
1396      Requires<[In32BitMode]>;
1397def : Pat<(srl GR16:$src, (i8 8)),
1398          (EXTRACT_SUBREG
1399            (MOVZX32rr8
1400              (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1401                              sub_8bit_hi)),
1402            sub_16bit)>,
1403      Requires<[In32BitMode]>;
1404def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1405          (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1406                                                             GR16_ABCD)),
1407                                      sub_8bit_hi))>,
1408      Requires<[In32BitMode]>;
1409def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1410          (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1411                                                             GR16_ABCD)),
1412                                      sub_8bit_hi))>,
1413      Requires<[In32BitMode]>;
1414def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1415          (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1416                                                             GR32_ABCD)),
1417                                      sub_8bit_hi))>,
1418      Requires<[In32BitMode]>;
1419def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1420          (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1421                                                             GR32_ABCD)),
1422                                      sub_8bit_hi))>,
1423      Requires<[In32BitMode]>;
1424
1425// h-register tricks.
1426// For now, be conservative on x86-64 and use an h-register extract only if the
1427// value is immediately zero-extended or stored, which are somewhat common
1428// cases. This uses a bunch of code to prevent a register requiring a REX prefix
1429// from being allocated in the same instruction as the h register, as there's
1430// currently no way to describe this requirement to the register allocator.
1431
1432// h-register extract and zero-extend.
1433def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1434          (SUBREG_TO_REG
1435            (i64 0),
1436            (MOVZX32_NOREXrr8
1437              (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1438                              sub_8bit_hi)),
1439            sub_32bit)>;
1440def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1441          (MOVZX32_NOREXrr8
1442            (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1443                            sub_8bit_hi))>,
1444      Requires<[In64BitMode]>;
1445def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1446          (MOVZX32_NOREXrr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1447                                                                   GR32_ABCD)),
1448                                             sub_8bit_hi))>,
1449      Requires<[In64BitMode]>;
1450def : Pat<(srl GR16:$src, (i8 8)),
1451          (EXTRACT_SUBREG
1452            (MOVZX32_NOREXrr8
1453              (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1454                              sub_8bit_hi)),
1455            sub_16bit)>,
1456      Requires<[In64BitMode]>;
1457def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1458          (MOVZX32_NOREXrr8
1459            (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1460                            sub_8bit_hi))>,
1461      Requires<[In64BitMode]>;
1462def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1463          (MOVZX32_NOREXrr8
1464            (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1465                            sub_8bit_hi))>,
1466      Requires<[In64BitMode]>;
1467def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1468          (SUBREG_TO_REG
1469            (i64 0),
1470            (MOVZX32_NOREXrr8
1471              (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1472                              sub_8bit_hi)),
1473            sub_32bit)>;
1474def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1475          (SUBREG_TO_REG
1476            (i64 0),
1477            (MOVZX32_NOREXrr8
1478              (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1479                              sub_8bit_hi)),
1480            sub_32bit)>;
1481
1482// h-register extract and store.
1483def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1484          (MOV8mr_NOREX
1485            addr:$dst,
1486            (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1487                            sub_8bit_hi))>;
1488def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1489          (MOV8mr_NOREX
1490            addr:$dst,
1491            (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1492                            sub_8bit_hi))>,
1493      Requires<[In64BitMode]>;
1494def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1495          (MOV8mr_NOREX
1496            addr:$dst,
1497            (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1498                            sub_8bit_hi))>,
1499      Requires<[In64BitMode]>;
1500
1501
1502// (shl x, 1) ==> (add x, x)
1503// Note that if x is undef (immediate or otherwise), we could theoretically
1504// end up with the two uses of x getting different values, producing a result
1505// where the least significant bit is not 0. However, the probability of this
1506// happening is considered low enough that this is officially not a
1507// "real problem".
1508def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr  GR8 :$src1, GR8 :$src1)>;
1509def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1510def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1511def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1512
1513// Helper imms that check if a mask doesn't change significant shift bits.
1514def immShift32 : ImmLeaf<i8, [{ return CountTrailingOnes_32(Imm) >= 5; }]>;
1515def immShift64 : ImmLeaf<i8, [{ return CountTrailingOnes_32(Imm) >= 6; }]>;
1516
1517// (shl x (and y, 31)) ==> (shl x, y)
1518def : Pat<(shl GR8:$src1, (and CL, immShift32)),
1519          (SHL8rCL GR8:$src1)>;
1520def : Pat<(shl GR16:$src1, (and CL, immShift32)),
1521          (SHL16rCL GR16:$src1)>;
1522def : Pat<(shl GR32:$src1, (and CL, immShift32)),
1523          (SHL32rCL GR32:$src1)>;
1524def : Pat<(store (shl (loadi8 addr:$dst), (and CL, immShift32)), addr:$dst),
1525          (SHL8mCL addr:$dst)>;
1526def : Pat<(store (shl (loadi16 addr:$dst), (and CL, immShift32)), addr:$dst),
1527          (SHL16mCL addr:$dst)>;
1528def : Pat<(store (shl (loadi32 addr:$dst), (and CL, immShift32)), addr:$dst),
1529          (SHL32mCL addr:$dst)>;
1530
1531def : Pat<(srl GR8:$src1, (and CL, immShift32)),
1532          (SHR8rCL GR8:$src1)>;
1533def : Pat<(srl GR16:$src1, (and CL, immShift32)),
1534          (SHR16rCL GR16:$src1)>;
1535def : Pat<(srl GR32:$src1, (and CL, immShift32)),
1536          (SHR32rCL GR32:$src1)>;
1537def : Pat<(store (srl (loadi8 addr:$dst), (and CL, immShift32)), addr:$dst),
1538          (SHR8mCL addr:$dst)>;
1539def : Pat<(store (srl (loadi16 addr:$dst), (and CL, immShift32)), addr:$dst),
1540          (SHR16mCL addr:$dst)>;
1541def : Pat<(store (srl (loadi32 addr:$dst), (and CL, immShift32)), addr:$dst),
1542          (SHR32mCL addr:$dst)>;
1543
1544def : Pat<(sra GR8:$src1, (and CL, immShift32)),
1545          (SAR8rCL GR8:$src1)>;
1546def : Pat<(sra GR16:$src1, (and CL, immShift32)),
1547          (SAR16rCL GR16:$src1)>;
1548def : Pat<(sra GR32:$src1, (and CL, immShift32)),
1549          (SAR32rCL GR32:$src1)>;
1550def : Pat<(store (sra (loadi8 addr:$dst), (and CL, immShift32)), addr:$dst),
1551          (SAR8mCL addr:$dst)>;
1552def : Pat<(store (sra (loadi16 addr:$dst), (and CL, immShift32)), addr:$dst),
1553          (SAR16mCL addr:$dst)>;
1554def : Pat<(store (sra (loadi32 addr:$dst), (and CL, immShift32)), addr:$dst),
1555          (SAR32mCL addr:$dst)>;
1556
1557// (shl x (and y, 63)) ==> (shl x, y)
1558def : Pat<(shl GR64:$src1, (and CL, immShift64)),
1559          (SHL64rCL GR64:$src1)>;
1560def : Pat<(store (shl (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1561          (SHL64mCL addr:$dst)>;
1562
1563def : Pat<(srl GR64:$src1, (and CL, immShift64)),
1564          (SHR64rCL GR64:$src1)>;
1565def : Pat<(store (srl (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1566          (SHR64mCL addr:$dst)>;
1567
1568def : Pat<(sra GR64:$src1, (and CL, immShift64)),
1569          (SAR64rCL GR64:$src1)>;
1570def : Pat<(store (sra (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1571          (SAR64mCL addr:$dst)>;
1572
1573
1574// (anyext (setcc_carry)) -> (setcc_carry)
1575def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1576          (SETB_C16r)>;
1577def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1578          (SETB_C32r)>;
1579def : Pat<(i32 (anyext (i16 (X86setcc_c X86_COND_B, EFLAGS)))),
1580          (SETB_C32r)>;
1581
1582
1583
1584
1585//===----------------------------------------------------------------------===//
1586// EFLAGS-defining Patterns
1587//===----------------------------------------------------------------------===//
1588
1589// add reg, reg
1590def : Pat<(add GR8 :$src1, GR8 :$src2), (ADD8rr  GR8 :$src1, GR8 :$src2)>;
1591def : Pat<(add GR16:$src1, GR16:$src2), (ADD16rr GR16:$src1, GR16:$src2)>;
1592def : Pat<(add GR32:$src1, GR32:$src2), (ADD32rr GR32:$src1, GR32:$src2)>;
1593
1594// add reg, mem
1595def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
1596          (ADD8rm GR8:$src1, addr:$src2)>;
1597def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
1598          (ADD16rm GR16:$src1, addr:$src2)>;
1599def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
1600          (ADD32rm GR32:$src1, addr:$src2)>;
1601
1602// add reg, imm
1603def : Pat<(add GR8 :$src1, imm:$src2), (ADD8ri  GR8:$src1 , imm:$src2)>;
1604def : Pat<(add GR16:$src1, imm:$src2), (ADD16ri GR16:$src1, imm:$src2)>;
1605def : Pat<(add GR32:$src1, imm:$src2), (ADD32ri GR32:$src1, imm:$src2)>;
1606def : Pat<(add GR16:$src1, i16immSExt8:$src2),
1607          (ADD16ri8 GR16:$src1, i16immSExt8:$src2)>;
1608def : Pat<(add GR32:$src1, i32immSExt8:$src2),
1609          (ADD32ri8 GR32:$src1, i32immSExt8:$src2)>;
1610
1611// sub reg, reg
1612def : Pat<(sub GR8 :$src1, GR8 :$src2), (SUB8rr  GR8 :$src1, GR8 :$src2)>;
1613def : Pat<(sub GR16:$src1, GR16:$src2), (SUB16rr GR16:$src1, GR16:$src2)>;
1614def : Pat<(sub GR32:$src1, GR32:$src2), (SUB32rr GR32:$src1, GR32:$src2)>;
1615
1616// sub reg, mem
1617def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
1618          (SUB8rm GR8:$src1, addr:$src2)>;
1619def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
1620          (SUB16rm GR16:$src1, addr:$src2)>;
1621def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
1622          (SUB32rm GR32:$src1, addr:$src2)>;
1623
1624// sub reg, imm
1625def : Pat<(sub GR8:$src1, imm:$src2),
1626          (SUB8ri GR8:$src1, imm:$src2)>;
1627def : Pat<(sub GR16:$src1, imm:$src2),
1628          (SUB16ri GR16:$src1, imm:$src2)>;
1629def : Pat<(sub GR32:$src1, imm:$src2),
1630          (SUB32ri GR32:$src1, imm:$src2)>;
1631def : Pat<(sub GR16:$src1, i16immSExt8:$src2),
1632          (SUB16ri8 GR16:$src1, i16immSExt8:$src2)>;
1633def : Pat<(sub GR32:$src1, i32immSExt8:$src2),
1634          (SUB32ri8 GR32:$src1, i32immSExt8:$src2)>;
1635
1636// sub 0, reg
1637def : Pat<(X86sub_flag 0, GR8 :$src), (NEG8r  GR8 :$src)>;
1638def : Pat<(X86sub_flag 0, GR16:$src), (NEG16r GR16:$src)>;
1639def : Pat<(X86sub_flag 0, GR32:$src), (NEG32r GR32:$src)>;
1640def : Pat<(X86sub_flag 0, GR64:$src), (NEG64r GR64:$src)>;
1641
1642// mul reg, reg
1643def : Pat<(mul GR16:$src1, GR16:$src2),
1644          (IMUL16rr GR16:$src1, GR16:$src2)>;
1645def : Pat<(mul GR32:$src1, GR32:$src2),
1646          (IMUL32rr GR32:$src1, GR32:$src2)>;
1647
1648// mul reg, mem
1649def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
1650          (IMUL16rm GR16:$src1, addr:$src2)>;
1651def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
1652          (IMUL32rm GR32:$src1, addr:$src2)>;
1653
1654// mul reg, imm
1655def : Pat<(mul GR16:$src1, imm:$src2),
1656          (IMUL16rri GR16:$src1, imm:$src2)>;
1657def : Pat<(mul GR32:$src1, imm:$src2),
1658          (IMUL32rri GR32:$src1, imm:$src2)>;
1659def : Pat<(mul GR16:$src1, i16immSExt8:$src2),
1660          (IMUL16rri8 GR16:$src1, i16immSExt8:$src2)>;
1661def : Pat<(mul GR32:$src1, i32immSExt8:$src2),
1662          (IMUL32rri8 GR32:$src1, i32immSExt8:$src2)>;
1663
1664// reg = mul mem, imm
1665def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
1666          (IMUL16rmi addr:$src1, imm:$src2)>;
1667def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
1668          (IMUL32rmi addr:$src1, imm:$src2)>;
1669def : Pat<(mul (loadi16 addr:$src1), i16immSExt8:$src2),
1670          (IMUL16rmi8 addr:$src1, i16immSExt8:$src2)>;
1671def : Pat<(mul (loadi32 addr:$src1), i32immSExt8:$src2),
1672          (IMUL32rmi8 addr:$src1, i32immSExt8:$src2)>;
1673
1674// Patterns for nodes that do not produce flags, for instructions that do.
1675
1676// addition
1677def : Pat<(add GR64:$src1, GR64:$src2),
1678          (ADD64rr GR64:$src1, GR64:$src2)>;
1679def : Pat<(add GR64:$src1, i64immSExt8:$src2),
1680          (ADD64ri8 GR64:$src1, i64immSExt8:$src2)>;
1681def : Pat<(add GR64:$src1, i64immSExt32:$src2),
1682          (ADD64ri32 GR64:$src1, i64immSExt32:$src2)>;
1683def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
1684          (ADD64rm GR64:$src1, addr:$src2)>;
1685
1686// subtraction
1687def : Pat<(sub GR64:$src1, GR64:$src2),
1688          (SUB64rr GR64:$src1, GR64:$src2)>;
1689def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
1690          (SUB64rm GR64:$src1, addr:$src2)>;
1691def : Pat<(sub GR64:$src1, i64immSExt8:$src2),
1692          (SUB64ri8 GR64:$src1, i64immSExt8:$src2)>;
1693def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
1694          (SUB64ri32 GR64:$src1, i64immSExt32:$src2)>;
1695
1696// Multiply
1697def : Pat<(mul GR64:$src1, GR64:$src2),
1698          (IMUL64rr GR64:$src1, GR64:$src2)>;
1699def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
1700          (IMUL64rm GR64:$src1, addr:$src2)>;
1701def : Pat<(mul GR64:$src1, i64immSExt8:$src2),
1702          (IMUL64rri8 GR64:$src1, i64immSExt8:$src2)>;
1703def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
1704          (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
1705def : Pat<(mul (loadi64 addr:$src1), i64immSExt8:$src2),
1706          (IMUL64rmi8 addr:$src1, i64immSExt8:$src2)>;
1707def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
1708          (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
1709
1710// Increment reg.
1711def : Pat<(add GR8 :$src, 1), (INC8r     GR8 :$src)>;
1712def : Pat<(add GR16:$src, 1), (INC16r    GR16:$src)>, Requires<[In32BitMode]>;
1713def : Pat<(add GR16:$src, 1), (INC64_16r GR16:$src)>, Requires<[In64BitMode]>;
1714def : Pat<(add GR32:$src, 1), (INC32r    GR32:$src)>, Requires<[In32BitMode]>;
1715def : Pat<(add GR32:$src, 1), (INC64_32r GR32:$src)>, Requires<[In64BitMode]>;
1716def : Pat<(add GR64:$src, 1), (INC64r    GR64:$src)>;
1717
1718// Decrement reg.
1719def : Pat<(add GR8 :$src, -1), (DEC8r     GR8 :$src)>;
1720def : Pat<(add GR16:$src, -1), (DEC16r    GR16:$src)>, Requires<[In32BitMode]>;
1721def : Pat<(add GR16:$src, -1), (DEC64_16r GR16:$src)>, Requires<[In64BitMode]>;
1722def : Pat<(add GR32:$src, -1), (DEC32r    GR32:$src)>, Requires<[In32BitMode]>;
1723def : Pat<(add GR32:$src, -1), (DEC64_32r GR32:$src)>, Requires<[In64BitMode]>;
1724def : Pat<(add GR64:$src, -1), (DEC64r    GR64:$src)>;
1725
1726// or reg/reg.
1727def : Pat<(or GR8 :$src1, GR8 :$src2), (OR8rr  GR8 :$src1, GR8 :$src2)>;
1728def : Pat<(or GR16:$src1, GR16:$src2), (OR16rr GR16:$src1, GR16:$src2)>;
1729def : Pat<(or GR32:$src1, GR32:$src2), (OR32rr GR32:$src1, GR32:$src2)>;
1730def : Pat<(or GR64:$src1, GR64:$src2), (OR64rr GR64:$src1, GR64:$src2)>;
1731
1732// or reg/mem
1733def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
1734          (OR8rm GR8:$src1, addr:$src2)>;
1735def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
1736          (OR16rm GR16:$src1, addr:$src2)>;
1737def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
1738          (OR32rm GR32:$src1, addr:$src2)>;
1739def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
1740          (OR64rm GR64:$src1, addr:$src2)>;
1741
1742// or reg/imm
1743def : Pat<(or GR8:$src1 , imm:$src2), (OR8ri  GR8 :$src1, imm:$src2)>;
1744def : Pat<(or GR16:$src1, imm:$src2), (OR16ri GR16:$src1, imm:$src2)>;
1745def : Pat<(or GR32:$src1, imm:$src2), (OR32ri GR32:$src1, imm:$src2)>;
1746def : Pat<(or GR16:$src1, i16immSExt8:$src2),
1747          (OR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1748def : Pat<(or GR32:$src1, i32immSExt8:$src2),
1749          (OR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1750def : Pat<(or GR64:$src1, i64immSExt8:$src2),
1751          (OR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1752def : Pat<(or GR64:$src1, i64immSExt32:$src2),
1753          (OR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1754
1755// xor reg/reg
1756def : Pat<(xor GR8 :$src1, GR8 :$src2), (XOR8rr  GR8 :$src1, GR8 :$src2)>;
1757def : Pat<(xor GR16:$src1, GR16:$src2), (XOR16rr GR16:$src1, GR16:$src2)>;
1758def : Pat<(xor GR32:$src1, GR32:$src2), (XOR32rr GR32:$src1, GR32:$src2)>;
1759def : Pat<(xor GR64:$src1, GR64:$src2), (XOR64rr GR64:$src1, GR64:$src2)>;
1760
1761// xor reg/mem
1762def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
1763          (XOR8rm GR8:$src1, addr:$src2)>;
1764def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
1765          (XOR16rm GR16:$src1, addr:$src2)>;
1766def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
1767          (XOR32rm GR32:$src1, addr:$src2)>;
1768def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
1769          (XOR64rm GR64:$src1, addr:$src2)>;
1770
1771// xor reg/imm
1772def : Pat<(xor GR8:$src1, imm:$src2),
1773          (XOR8ri GR8:$src1, imm:$src2)>;
1774def : Pat<(xor GR16:$src1, imm:$src2),
1775          (XOR16ri GR16:$src1, imm:$src2)>;
1776def : Pat<(xor GR32:$src1, imm:$src2),
1777          (XOR32ri GR32:$src1, imm:$src2)>;
1778def : Pat<(xor GR16:$src1, i16immSExt8:$src2),
1779          (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1780def : Pat<(xor GR32:$src1, i32immSExt8:$src2),
1781          (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1782def : Pat<(xor GR64:$src1, i64immSExt8:$src2),
1783          (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1784def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
1785          (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1786
1787// and reg/reg
1788def : Pat<(and GR8 :$src1, GR8 :$src2), (AND8rr  GR8 :$src1, GR8 :$src2)>;
1789def : Pat<(and GR16:$src1, GR16:$src2), (AND16rr GR16:$src1, GR16:$src2)>;
1790def : Pat<(and GR32:$src1, GR32:$src2), (AND32rr GR32:$src1, GR32:$src2)>;
1791def : Pat<(and GR64:$src1, GR64:$src2), (AND64rr GR64:$src1, GR64:$src2)>;
1792
1793// and reg/mem
1794def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
1795          (AND8rm GR8:$src1, addr:$src2)>;
1796def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
1797          (AND16rm GR16:$src1, addr:$src2)>;
1798def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
1799          (AND32rm GR32:$src1, addr:$src2)>;
1800def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
1801          (AND64rm GR64:$src1, addr:$src2)>;
1802
1803// and reg/imm
1804def : Pat<(and GR8:$src1, imm:$src2),
1805          (AND8ri GR8:$src1, imm:$src2)>;
1806def : Pat<(and GR16:$src1, imm:$src2),
1807          (AND16ri GR16:$src1, imm:$src2)>;
1808def : Pat<(and GR32:$src1, imm:$src2),
1809          (AND32ri GR32:$src1, imm:$src2)>;
1810def : Pat<(and GR16:$src1, i16immSExt8:$src2),
1811          (AND16ri8 GR16:$src1, i16immSExt8:$src2)>;
1812def : Pat<(and GR32:$src1, i32immSExt8:$src2),
1813          (AND32ri8 GR32:$src1, i32immSExt8:$src2)>;
1814def : Pat<(and GR64:$src1, i64immSExt8:$src2),
1815          (AND64ri8 GR64:$src1, i64immSExt8:$src2)>;
1816def : Pat<(and GR64:$src1, i64immSExt32:$src2),
1817          (AND64ri32 GR64:$src1, i64immSExt32:$src2)>;
1818
1819// Bit scan instruction patterns to match explicit zero-undef behavior.
1820def : Pat<(cttz_zero_undef GR16:$src), (BSF16rr GR16:$src)>;
1821def : Pat<(cttz_zero_undef GR32:$src), (BSF32rr GR32:$src)>;
1822def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr GR64:$src)>;
1823def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm addr:$src)>;
1824def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm addr:$src)>;
1825def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm addr:$src)>;
1826