1 //===- X86InstrCompiler.td - Compiler Pseudos and Patterns -*- tablegen -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file describes the various pseudo instructions used by the compiler,
10 // as well as Pat patterns used during instruction selection.
12 //===----------------------------------------------------------------------===//
14 //===----------------------------------------------------------------------===//
15 // Pattern Matching Support
17 def GetLo32XForm : SDNodeXForm<imm, [{
18 // Transformation function: get the low 32 bits.
19 return getI32Imm((uint32_t)N->getZExtValue(), SDLoc(N));
23 //===----------------------------------------------------------------------===//
24 // Random Pseudo Instructions.
26 // PIC base construction. This expands to code that looks like this:
29 let hasSideEffects = 0, isNotDuplicable = 1, Uses = [ESP, SSP],
30 SchedRW = [WriteJump] in
31 def MOVPC32r : Ii32<0xE8, Pseudo, (outs GR32:$reg), (ins i32imm:$label),
34 // ADJCALLSTACKDOWN/UP implicitly use/def ESP because they may be expanded into
35 // a stack adjustment and the codegen must know that they may modify the stack
36 // pointer before prolog-epilog rewriting occurs.
37 // Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
38 // sub / add which can clobber EFLAGS.
39 let Defs = [ESP, EFLAGS, SSP], Uses = [ESP, SSP], SchedRW = [WriteALU] in {
40 def ADJCALLSTACKDOWN32 : I<0, Pseudo, (outs),
41 (ins i32imm:$amt1, i32imm:$amt2, i32imm:$amt3),
42 "#ADJCALLSTACKDOWN", []>, Requires<[NotLP64]>;
43 def ADJCALLSTACKUP32 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
45 [(X86callseq_end timm:$amt1, timm:$amt2)]>,
48 def : Pat<(X86callseq_start timm:$amt1, timm:$amt2),
49 (ADJCALLSTACKDOWN32 i32imm:$amt1, i32imm:$amt2, 0)>, Requires<[NotLP64]>;
52 // ADJCALLSTACKDOWN/UP implicitly use/def RSP because they may be expanded into
53 // a stack adjustment and the codegen must know that they may modify the stack
54 // pointer before prolog-epilog rewriting occurs.
55 // Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
56 // sub / add which can clobber EFLAGS.
57 let Defs = [RSP, EFLAGS, SSP], Uses = [RSP, SSP], SchedRW = [WriteALU] in {
58 def ADJCALLSTACKDOWN64 : I<0, Pseudo, (outs),
59 (ins i32imm:$amt1, i32imm:$amt2, i32imm:$amt3),
60 "#ADJCALLSTACKDOWN", []>, Requires<[IsLP64]>;
61 def ADJCALLSTACKUP64 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
63 [(X86callseq_end timm:$amt1, timm:$amt2)]>,
66 def : Pat<(X86callseq_start timm:$amt1, timm:$amt2),
67 (ADJCALLSTACKDOWN64 i32imm:$amt1, i32imm:$amt2, 0)>, Requires<[IsLP64]>;
69 let SchedRW = [WriteSystem] in {
71 // x86-64 va_start lowering magic.
72 let usesCustomInserter = 1, Defs = [EFLAGS] in {
73 def VASTART_SAVE_XMM_REGS : I<0, Pseudo,
76 i64imm:$regsavefi, i64imm:$offset,
78 "#VASTART_SAVE_XMM_REGS $al, $regsavefi, $offset",
79 [(X86vastart_save_xmm_regs GR8:$al,
84 // The VAARG_64 pseudo-instruction takes the address of the va_list,
85 // and places the address of the next argument into a register.
86 let Defs = [EFLAGS] in
87 def VAARG_64 : I<0, Pseudo,
89 (ins i8mem:$ap, i32imm:$size, i8imm:$mode, i32imm:$align),
90 "#VAARG_64 $dst, $ap, $size, $mode, $align",
92 (X86vaarg64 addr:$ap, imm:$size, imm:$mode, imm:$align)),
96 // When using segmented stacks these are lowered into instructions which first
97 // check if the current stacklet has enough free memory. If it does, memory is
98 // allocated by bumping the stack pointer. Otherwise memory is allocated from
101 let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
102 def SEG_ALLOCA_32 : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$size),
103 "# variable sized alloca for segmented stacks",
105 (X86SegAlloca GR32:$size))]>,
108 let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
109 def SEG_ALLOCA_64 : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$size),
110 "# variable sized alloca for segmented stacks",
112 (X86SegAlloca GR64:$size))]>,
113 Requires<[In64BitMode]>;
116 // Dynamic stack allocation yields a _chkstk or _alloca call for all Windows
117 // targets. These calls are needed to probe the stack when allocating more than
118 // 4k bytes in one go. Touching the stack at 4K increments is necessary to
119 // ensure that the guard pages used by the OS virtual memory manager are
120 // allocated in correct sequence.
121 // The main point of having separate instruction are extra unmodelled effects
122 // (compared to ordinary calls) like stack pointer change.
124 let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
125 def WIN_ALLOCA_32 : I<0, Pseudo, (outs), (ins GR32:$size),
126 "# dynamic stack allocation",
127 [(X86WinAlloca GR32:$size)]>,
130 let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
131 def WIN_ALLOCA_64 : I<0, Pseudo, (outs), (ins GR64:$size),
132 "# dynamic stack allocation",
133 [(X86WinAlloca GR64:$size)]>,
134 Requires<[In64BitMode]>;
137 // These instructions XOR the frame pointer into a GPR. They are used in some
138 // stack protection schemes. These are post-RA pseudos because we only know the
139 // frame register after register allocation.
140 let Constraints = "$src = $dst", isMoveImm = 1, isPseudo = 1, Defs = [EFLAGS] in {
141 def XOR32_FP : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src),
142 "xorl\t$$FP, $src", []>,
143 Requires<[NotLP64]>, Sched<[WriteALU]>;
144 def XOR64_FP : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src),
145 "xorq\t$$FP $src", []>,
146 Requires<[In64BitMode]>, Sched<[WriteALU]>;
149 //===----------------------------------------------------------------------===//
150 // EH Pseudo Instructions
152 let SchedRW = [WriteSystem] in {
153 let isTerminator = 1, isReturn = 1, isBarrier = 1,
154 hasCtrlDep = 1, isCodeGenOnly = 1 in {
155 def EH_RETURN : I<0xC3, RawFrm, (outs), (ins GR32:$addr),
156 "ret\t#eh_return, addr: $addr",
157 [(X86ehret GR32:$addr)]>, Sched<[WriteJumpLd]>;
161 let isTerminator = 1, isReturn = 1, isBarrier = 1,
162 hasCtrlDep = 1, isCodeGenOnly = 1 in {
163 def EH_RETURN64 : I<0xC3, RawFrm, (outs), (ins GR64:$addr),
164 "ret\t#eh_return, addr: $addr",
165 [(X86ehret GR64:$addr)]>, Sched<[WriteJumpLd]>;
169 let isTerminator = 1, hasSideEffects = 1, isBarrier = 1, hasCtrlDep = 1,
170 isCodeGenOnly = 1, isReturn = 1, isEHScopeReturn = 1 in {
171 def CLEANUPRET : I<0, Pseudo, (outs), (ins), "# CLEANUPRET", [(cleanupret)]>;
173 // CATCHRET needs a custom inserter for SEH.
174 let usesCustomInserter = 1 in
175 def CATCHRET : I<0, Pseudo, (outs), (ins brtarget32:$dst, brtarget32:$from),
177 [(catchret bb:$dst, bb:$from)]>;
180 let hasSideEffects = 1, hasCtrlDep = 1, isCodeGenOnly = 1,
181 usesCustomInserter = 1 in
182 def CATCHPAD : I<0, Pseudo, (outs), (ins), "# CATCHPAD", [(catchpad)]>;
184 // This instruction is responsible for re-establishing stack pointers after an
185 // exception has been caught and we are rejoining normal control flow in the
186 // parent function or funclet. It generally sets ESP and EBP, and optionally
187 // ESI. It is only needed for 32-bit WinEH, as the runtime restores CSRs for us
189 let hasSideEffects = 1, hasCtrlDep = 1, isCodeGenOnly = 1 in
190 def EH_RESTORE : I<0, Pseudo, (outs), (ins), "# EH_RESTORE", []>;
192 let hasSideEffects = 1, isBarrier = 1, isCodeGenOnly = 1,
193 usesCustomInserter = 1 in {
194 def EH_SjLj_SetJmp32 : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$buf),
196 [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
197 Requires<[Not64BitMode]>;
198 def EH_SjLj_SetJmp64 : I<0, Pseudo, (outs GR32:$dst), (ins i64mem:$buf),
200 [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
201 Requires<[In64BitMode]>;
202 let isTerminator = 1 in {
203 def EH_SjLj_LongJmp32 : I<0, Pseudo, (outs), (ins i32mem:$buf),
204 "#EH_SJLJ_LONGJMP32",
205 [(X86eh_sjlj_longjmp addr:$buf)]>,
206 Requires<[Not64BitMode]>;
207 def EH_SjLj_LongJmp64 : I<0, Pseudo, (outs), (ins i64mem:$buf),
208 "#EH_SJLJ_LONGJMP64",
209 [(X86eh_sjlj_longjmp addr:$buf)]>,
210 Requires<[In64BitMode]>;
214 let isBranch = 1, isTerminator = 1, isCodeGenOnly = 1 in {
215 def EH_SjLj_Setup : I<0, Pseudo, (outs), (ins brtarget:$dst),
216 "#EH_SjLj_Setup\t$dst", []>;
220 //===----------------------------------------------------------------------===//
221 // Pseudo instructions used by unwind info.
223 let isPseudo = 1, SchedRW = [WriteSystem] in {
224 def SEH_PushReg : I<0, Pseudo, (outs), (ins i32imm:$reg),
225 "#SEH_PushReg $reg", []>;
226 def SEH_SaveReg : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
227 "#SEH_SaveReg $reg, $dst", []>;
228 def SEH_SaveXMM : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
229 "#SEH_SaveXMM $reg, $dst", []>;
230 def SEH_StackAlloc : I<0, Pseudo, (outs), (ins i32imm:$size),
231 "#SEH_StackAlloc $size", []>;
232 def SEH_StackAlign : I<0, Pseudo, (outs), (ins i32imm:$align),
233 "#SEH_StackAlign $align", []>;
234 def SEH_SetFrame : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$offset),
235 "#SEH_SetFrame $reg, $offset", []>;
236 def SEH_PushFrame : I<0, Pseudo, (outs), (ins i1imm:$mode),
237 "#SEH_PushFrame $mode", []>;
238 def SEH_EndPrologue : I<0, Pseudo, (outs), (ins),
239 "#SEH_EndPrologue", []>;
240 def SEH_Epilogue : I<0, Pseudo, (outs), (ins),
241 "#SEH_Epilogue", []>;
244 //===----------------------------------------------------------------------===//
245 // Pseudo instructions used by segmented stacks.
248 // This is lowered into a RET instruction by MCInstLower. We need
249 // this so that we don't have to have a MachineBasicBlock which ends
250 // with a RET and also has successors.
251 let isPseudo = 1, SchedRW = [WriteJumpLd] in {
252 def MORESTACK_RET: I<0, Pseudo, (outs), (ins), "", []>;
254 // This instruction is lowered to a RET followed by a MOV. The two
255 // instructions are not generated on a higher level since then the
256 // verifier sees a MachineBasicBlock ending with a non-terminator.
257 def MORESTACK_RET_RESTORE_R10 : I<0, Pseudo, (outs), (ins), "", []>;
260 //===----------------------------------------------------------------------===//
261 // Alias Instructions
262 //===----------------------------------------------------------------------===//
264 // Alias instruction mapping movr0 to xor.
265 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
266 let Defs = [EFLAGS], isReMaterializable = 1, isAsCheapAsAMove = 1,
267 isPseudo = 1, isMoveImm = 1, AddedComplexity = 10 in
268 def MOV32r0 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
269 [(set GR32:$dst, 0)]>, Sched<[WriteZero]>;
271 // Other widths can also make use of the 32-bit xor, which may have a smaller
272 // encoding and avoid partial register updates.
273 let AddedComplexity = 10 in {
274 def : Pat<(i8 0), (EXTRACT_SUBREG (MOV32r0), sub_8bit)>;
275 def : Pat<(i16 0), (EXTRACT_SUBREG (MOV32r0), sub_16bit)>;
276 def : Pat<(i64 0), (SUBREG_TO_REG (i64 0), (MOV32r0), sub_32bit)>;
279 let Predicates = [OptForSize, Not64BitMode],
280 AddedComplexity = 10 in {
281 let SchedRW = [WriteALU] in {
282 // Pseudo instructions for materializing 1 and -1 using XOR+INC/DEC,
283 // which only require 3 bytes compared to MOV32ri which requires 5.
284 let Defs = [EFLAGS], isReMaterializable = 1, isPseudo = 1 in {
285 def MOV32r1 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
286 [(set GR32:$dst, 1)]>;
287 def MOV32r_1 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
288 [(set GR32:$dst, -1)]>;
292 // MOV16ri is 4 bytes, so the instructions above are smaller.
293 def : Pat<(i16 1), (EXTRACT_SUBREG (MOV32r1), sub_16bit)>;
294 def : Pat<(i16 -1), (EXTRACT_SUBREG (MOV32r_1), sub_16bit)>;
297 let isReMaterializable = 1, isPseudo = 1, AddedComplexity = 5,
298 SchedRW = [WriteALU] in {
299 // AddedComplexity higher than MOV64ri but lower than MOV32r0 and MOV32r1.
300 def MOV32ImmSExti8 : I<0, Pseudo, (outs GR32:$dst), (ins i32i8imm:$src), "",
301 [(set GR32:$dst, i32immSExt8:$src)]>,
302 Requires<[OptForMinSize, NotWin64WithoutFP]>;
303 def MOV64ImmSExti8 : I<0, Pseudo, (outs GR64:$dst), (ins i64i8imm:$src), "",
304 [(set GR64:$dst, i64immSExt8:$src)]>,
305 Requires<[OptForMinSize, NotWin64WithoutFP]>;
308 // Materialize i64 constant where top 32-bits are zero. This could theoretically
309 // use MOV32ri with a SUBREG_TO_REG to represent the zero-extension, however
310 // that would make it more difficult to rematerialize.
311 let isReMaterializable = 1, isAsCheapAsAMove = 1,
312 isPseudo = 1, hasSideEffects = 0, SchedRW = [WriteMove] in
313 def MOV32ri64 : I<0, Pseudo, (outs GR64:$dst), (ins i64i32imm:$src), "", []>;
315 // This 64-bit pseudo-move can be used for both a 64-bit constant that is
316 // actually the zero-extension of a 32-bit constant and for labels in the
317 // x86-64 small code model.
318 def mov64imm32 : ComplexPattern<i64, 1, "selectMOV64Imm32", [imm, X86Wrapper]>;
320 def : Pat<(i64 mov64imm32:$src), (MOV32ri64 mov64imm32:$src)>;
322 // Use sbb to materialize carry bit.
323 let Uses = [EFLAGS], Defs = [EFLAGS], isPseudo = 1, SchedRW = [WriteALU] in {
324 // FIXME: These are pseudo ops that should be replaced with Pat<> patterns.
325 // However, Pat<> can't replicate the destination reg into the inputs of the
327 def SETB_C8r : I<0, Pseudo, (outs GR8:$dst), (ins), "",
328 [(set GR8:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
329 def SETB_C16r : I<0, Pseudo, (outs GR16:$dst), (ins), "",
330 [(set GR16:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
331 def SETB_C32r : I<0, Pseudo, (outs GR32:$dst), (ins), "",
332 [(set GR32:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
333 def SETB_C64r : I<0, Pseudo, (outs GR64:$dst), (ins), "",
334 [(set GR64:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
338 def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
340 def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
342 def : Pat<(i64 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
345 def : Pat<(i16 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
347 def : Pat<(i32 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
349 def : Pat<(i64 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
352 // We canonicalize 'setb' to "(and (sbb reg,reg), 1)" on the hope that the and
353 // will be eliminated and that the sbb can be extended up to a wider type. When
354 // this happens, it is great. However, if we are left with an 8-bit sbb and an
355 // and, we might as well just match it as a setb.
356 def : Pat<(and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1),
359 // Patterns to give priority when both inputs are zero so that we don't use
360 // an immediate for the RHS.
361 // TODO: Should we use a 32-bit sbb for 8/16 to push the extract_subreg out?
362 def : Pat<(X86sbb_flag (i8 0), (i8 0), EFLAGS),
363 (SBB8rr (EXTRACT_SUBREG (MOV32r0), sub_8bit),
364 (EXTRACT_SUBREG (MOV32r0), sub_8bit))>;
365 def : Pat<(X86sbb_flag (i16 0), (i16 0), EFLAGS),
366 (SBB16rr (EXTRACT_SUBREG (MOV32r0), sub_16bit),
367 (EXTRACT_SUBREG (MOV32r0), sub_16bit))>;
368 def : Pat<(X86sbb_flag (i32 0), (i32 0), EFLAGS),
369 (SBB32rr (MOV32r0), (MOV32r0))>;
370 def : Pat<(X86sbb_flag (i64 0), (i64 0), EFLAGS),
371 (SBB64rr (SUBREG_TO_REG (i64 0), (MOV32r0), sub_32bit),
372 (SUBREG_TO_REG (i64 0), (MOV32r0), sub_32bit))>;
374 //===----------------------------------------------------------------------===//
375 // String Pseudo Instructions
377 let SchedRW = [WriteMicrocoded] in {
378 let Defs = [ECX,EDI,ESI], Uses = [ECX,EDI,ESI], isCodeGenOnly = 1 in {
379 def REP_MOVSB_32 : I<0xA4, RawFrm, (outs), (ins),
380 "{rep;movsb (%esi), %es:(%edi)|rep movsb es:[edi], [esi]}",
381 [(X86rep_movs i8)]>, REP, AdSize32,
383 def REP_MOVSW_32 : I<0xA5, RawFrm, (outs), (ins),
384 "{rep;movsw (%esi), %es:(%edi)|rep movsw es:[edi], [esi]}",
385 [(X86rep_movs i16)]>, REP, AdSize32, OpSize16,
387 def REP_MOVSD_32 : I<0xA5, RawFrm, (outs), (ins),
388 "{rep;movsl (%esi), %es:(%edi)|rep movsd es:[edi], [esi]}",
389 [(X86rep_movs i32)]>, REP, AdSize32, OpSize32,
391 def REP_MOVSQ_32 : RI<0xA5, RawFrm, (outs), (ins),
392 "{rep;movsq (%esi), %es:(%edi)|rep movsq es:[edi], [esi]}",
393 [(X86rep_movs i64)]>, REP, AdSize32,
394 Requires<[NotLP64, In64BitMode]>;
397 let Defs = [RCX,RDI,RSI], Uses = [RCX,RDI,RSI], isCodeGenOnly = 1 in {
398 def REP_MOVSB_64 : I<0xA4, RawFrm, (outs), (ins),
399 "{rep;movsb (%rsi), %es:(%rdi)|rep movsb es:[rdi], [rsi]}",
400 [(X86rep_movs i8)]>, REP, AdSize64,
402 def REP_MOVSW_64 : I<0xA5, RawFrm, (outs), (ins),
403 "{rep;movsw (%rsi), %es:(%rdi)|rep movsw es:[rdi], [rsi]}",
404 [(X86rep_movs i16)]>, REP, AdSize64, OpSize16,
406 def REP_MOVSD_64 : I<0xA5, RawFrm, (outs), (ins),
407 "{rep;movsl (%rsi), %es:(%rdi)|rep movsdi es:[rdi], [rsi]}",
408 [(X86rep_movs i32)]>, REP, AdSize64, OpSize32,
410 def REP_MOVSQ_64 : RI<0xA5, RawFrm, (outs), (ins),
411 "{rep;movsq (%rsi), %es:(%rdi)|rep movsq es:[rdi], [rsi]}",
412 [(X86rep_movs i64)]>, REP, AdSize64,
416 // FIXME: Should use "(X86rep_stos AL)" as the pattern.
417 let Defs = [ECX,EDI], isCodeGenOnly = 1 in {
418 let Uses = [AL,ECX,EDI] in
419 def REP_STOSB_32 : I<0xAA, RawFrm, (outs), (ins),
420 "{rep;stosb %al, %es:(%edi)|rep stosb es:[edi], al}",
421 [(X86rep_stos i8)]>, REP, AdSize32,
423 let Uses = [AX,ECX,EDI] in
424 def REP_STOSW_32 : I<0xAB, RawFrm, (outs), (ins),
425 "{rep;stosw %ax, %es:(%edi)|rep stosw es:[edi], ax}",
426 [(X86rep_stos i16)]>, REP, AdSize32, OpSize16,
428 let Uses = [EAX,ECX,EDI] in
429 def REP_STOSD_32 : I<0xAB, RawFrm, (outs), (ins),
430 "{rep;stosl %eax, %es:(%edi)|rep stosd es:[edi], eax}",
431 [(X86rep_stos i32)]>, REP, AdSize32, OpSize32,
433 let Uses = [RAX,RCX,RDI] in
434 def REP_STOSQ_32 : RI<0xAB, RawFrm, (outs), (ins),
435 "{rep;stosq %rax, %es:(%edi)|rep stosq es:[edi], rax}",
436 [(X86rep_stos i64)]>, REP, AdSize32,
437 Requires<[NotLP64, In64BitMode]>;
440 let Defs = [RCX,RDI], isCodeGenOnly = 1 in {
441 let Uses = [AL,RCX,RDI] in
442 def REP_STOSB_64 : I<0xAA, RawFrm, (outs), (ins),
443 "{rep;stosb %al, %es:(%rdi)|rep stosb es:[rdi], al}",
444 [(X86rep_stos i8)]>, REP, AdSize64,
446 let Uses = [AX,RCX,RDI] in
447 def REP_STOSW_64 : I<0xAB, RawFrm, (outs), (ins),
448 "{rep;stosw %ax, %es:(%rdi)|rep stosw es:[rdi], ax}",
449 [(X86rep_stos i16)]>, REP, AdSize64, OpSize16,
451 let Uses = [RAX,RCX,RDI] in
452 def REP_STOSD_64 : I<0xAB, RawFrm, (outs), (ins),
453 "{rep;stosl %eax, %es:(%rdi)|rep stosd es:[rdi], eax}",
454 [(X86rep_stos i32)]>, REP, AdSize64, OpSize32,
457 let Uses = [RAX,RCX,RDI] in
458 def REP_STOSQ_64 : RI<0xAB, RawFrm, (outs), (ins),
459 "{rep;stosq %rax, %es:(%rdi)|rep stosq es:[rdi], rax}",
460 [(X86rep_stos i64)]>, REP, AdSize64,
465 //===----------------------------------------------------------------------===//
466 // Thread Local Storage Instructions
468 let SchedRW = [WriteSystem] in {
471 // All calls clobber the non-callee saved registers. ESP is marked as
472 // a use to prevent stack-pointer assignments that appear immediately
473 // before calls from potentially appearing dead.
474 let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, FP7,
475 ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7,
476 MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
477 XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
478 XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS, DF],
479 usesCustomInserter = 1, Uses = [ESP, SSP] in {
480 def TLS_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
482 [(X86tlsaddr tls32addr:$sym)]>,
483 Requires<[Not64BitMode]>;
484 def TLS_base_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
486 [(X86tlsbaseaddr tls32baseaddr:$sym)]>,
487 Requires<[Not64BitMode]>;
490 // All calls clobber the non-callee saved registers. RSP is marked as
491 // a use to prevent stack-pointer assignments that appear immediately
492 // before calls from potentially appearing dead.
493 let Defs = [RAX, RCX, RDX, RSI, RDI, R8, R9, R10, R11,
494 FP0, FP1, FP2, FP3, FP4, FP5, FP6, FP7,
495 ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7,
496 MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
497 XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
498 XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS, DF],
499 usesCustomInserter = 1, Uses = [RSP, SSP] in {
500 def TLS_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
502 [(X86tlsaddr tls64addr:$sym)]>,
503 Requires<[In64BitMode]>;
504 def TLS_base_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
506 [(X86tlsbaseaddr tls64baseaddr:$sym)]>,
507 Requires<[In64BitMode]>;
510 // Darwin TLS Support
511 // For i386, the address of the thunk is passed on the stack, on return the
512 // address of the variable is in %eax. %ecx is trashed during the function
513 // call. All other registers are preserved.
514 let Defs = [EAX, ECX, EFLAGS, DF],
516 usesCustomInserter = 1 in
517 def TLSCall_32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
519 [(X86TLSCall addr:$sym)]>,
520 Requires<[Not64BitMode]>;
522 // For x86_64, the address of the thunk is passed in %rdi, but the
523 // pseudo directly use the symbol, so do not add an implicit use of
524 // %rdi. The lowering will do the right thing with RDI.
525 // On return the address of the variable is in %rax. All other
526 // registers are preserved.
527 let Defs = [RAX, EFLAGS, DF],
529 usesCustomInserter = 1 in
530 def TLSCall_64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
532 [(X86TLSCall addr:$sym)]>,
533 Requires<[In64BitMode]>;
536 //===----------------------------------------------------------------------===//
537 // Conditional Move Pseudo Instructions
539 // CMOV* - Used to implement the SELECT DAG operation. Expanded after
540 // instruction selection into a branch sequence.
541 multiclass CMOVrr_PSEUDO<RegisterClass RC, ValueType VT> {
542 def CMOV#NAME : I<0, Pseudo,
543 (outs RC:$dst), (ins RC:$t, RC:$f, i8imm:$cond),
544 "#CMOV_"#NAME#" PSEUDO!",
545 [(set RC:$dst, (VT (X86cmov RC:$t, RC:$f, timm:$cond,
549 let usesCustomInserter = 1, hasNoSchedulingInfo = 1, Uses = [EFLAGS] in {
550 // X86 doesn't have 8-bit conditional moves. Use a customInserter to
551 // emit control flow. An alternative to this is to mark i8 SELECT as Promote,
552 // however that requires promoting the operands, and can induce additional
553 // i8 register pressure.
554 defm _GR8 : CMOVrr_PSEUDO<GR8, i8>;
556 let Predicates = [NoCMov] in {
557 defm _GR32 : CMOVrr_PSEUDO<GR32, i32>;
558 defm _GR16 : CMOVrr_PSEUDO<GR16, i16>;
559 } // Predicates = [NoCMov]
561 // fcmov doesn't handle all possible EFLAGS, provide a fallback if there is no
563 let Predicates = [FPStackf32] in
564 defm _RFP32 : CMOVrr_PSEUDO<RFP32, f32>;
566 let Predicates = [FPStackf64] in
567 defm _RFP64 : CMOVrr_PSEUDO<RFP64, f64>;
569 defm _RFP80 : CMOVrr_PSEUDO<RFP80, f80>;
571 let Predicates = [NoAVX512] in {
572 defm _FR32 : CMOVrr_PSEUDO<FR32, f32>;
573 defm _FR64 : CMOVrr_PSEUDO<FR64, f64>;
575 let Predicates = [HasAVX512] in {
576 defm _FR32X : CMOVrr_PSEUDO<FR32X, f32>;
577 defm _FR64X : CMOVrr_PSEUDO<FR64X, f64>;
579 let Predicates = [NoVLX] in {
580 defm _VR128 : CMOVrr_PSEUDO<VR128, v2i64>;
581 defm _VR256 : CMOVrr_PSEUDO<VR256, v4i64>;
583 let Predicates = [HasVLX] in {
584 defm _VR128X : CMOVrr_PSEUDO<VR128X, v2i64>;
585 defm _VR256X : CMOVrr_PSEUDO<VR256X, v4i64>;
587 defm _VR512 : CMOVrr_PSEUDO<VR512, v8i64>;
588 defm _VK2 : CMOVrr_PSEUDO<VK2, v2i1>;
589 defm _VK4 : CMOVrr_PSEUDO<VK4, v4i1>;
590 defm _VK8 : CMOVrr_PSEUDO<VK8, v8i1>;
591 defm _VK16 : CMOVrr_PSEUDO<VK16, v16i1>;
592 defm _VK32 : CMOVrr_PSEUDO<VK32, v32i1>;
593 defm _VK64 : CMOVrr_PSEUDO<VK64, v64i1>;
594 } // usesCustomInserter = 1, hasNoSchedulingInfo = 1, Uses = [EFLAGS]
596 def : Pat<(f128 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
597 (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
599 let Predicates = [NoVLX] in {
600 def : Pat<(v16i8 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
601 (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
602 def : Pat<(v8i16 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
603 (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
604 def : Pat<(v4i32 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
605 (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
606 def : Pat<(v4f32 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
607 (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
608 def : Pat<(v2f64 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
609 (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
611 def : Pat<(v32i8 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
612 (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
613 def : Pat<(v16i16 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
614 (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
615 def : Pat<(v8i32 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
616 (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
617 def : Pat<(v8f32 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
618 (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
619 def : Pat<(v4f64 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
620 (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
622 let Predicates = [HasVLX] in {
623 def : Pat<(v16i8 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
624 (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
625 def : Pat<(v8i16 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
626 (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
627 def : Pat<(v4i32 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
628 (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
629 def : Pat<(v4f32 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
630 (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
631 def : Pat<(v2f64 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
632 (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
634 def : Pat<(v32i8 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
635 (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
636 def : Pat<(v16i16 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
637 (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
638 def : Pat<(v8i32 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
639 (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
640 def : Pat<(v8f32 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
641 (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
642 def : Pat<(v4f64 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
643 (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
646 def : Pat<(v64i8 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
647 (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
648 def : Pat<(v32i16 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
649 (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
650 def : Pat<(v16i32 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
651 (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
652 def : Pat<(v16f32 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
653 (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
654 def : Pat<(v8f64 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
655 (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
657 //===----------------------------------------------------------------------===//
658 // Normal-Instructions-With-Lock-Prefix Pseudo Instructions
659 //===----------------------------------------------------------------------===//
661 // FIXME: Use normal instructions and add lock prefix dynamically.
665 let isCodeGenOnly = 1, Defs = [EFLAGS] in
666 def OR32mi8Locked : Ii8<0x83, MRM1m, (outs), (ins i32mem:$dst, i32i8imm:$zero),
667 "or{l}\t{$zero, $dst|$dst, $zero}", []>,
668 Requires<[Not64BitMode]>, OpSize32, LOCK,
669 Sched<[WriteALURMW]>;
671 let hasSideEffects = 1 in
672 def Int_MemBarrier : I<0, Pseudo, (outs), (ins),
674 [(X86MemBarrier)]>, Sched<[WriteLoad]>;
676 // RegOpc corresponds to the mr version of the instruction
677 // ImmOpc corresponds to the mi version of the instruction
678 // ImmOpc8 corresponds to the mi8 version of the instruction
679 // ImmMod corresponds to the instruction format of the mi and mi8 versions
680 multiclass LOCK_ArithBinOp<bits<8> RegOpc, bits<8> ImmOpc, bits<8> ImmOpc8,
681 Format ImmMod, SDNode Op, string mnemonic> {
682 let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
683 SchedRW = [WriteALURMW] in {
685 def NAME#8mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
686 RegOpc{3}, RegOpc{2}, RegOpc{1}, 0 },
687 MRMDestMem, (outs), (ins i8mem:$dst, GR8:$src2),
688 !strconcat(mnemonic, "{b}\t",
689 "{$src2, $dst|$dst, $src2}"),
690 [(set EFLAGS, (Op addr:$dst, GR8:$src2))]>, LOCK;
692 def NAME#16mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
693 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
694 MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
695 !strconcat(mnemonic, "{w}\t",
696 "{$src2, $dst|$dst, $src2}"),
697 [(set EFLAGS, (Op addr:$dst, GR16:$src2))]>,
700 def NAME#32mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
701 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
702 MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
703 !strconcat(mnemonic, "{l}\t",
704 "{$src2, $dst|$dst, $src2}"),
705 [(set EFLAGS, (Op addr:$dst, GR32:$src2))]>,
708 def NAME#64mr : RI<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
709 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
710 MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
711 !strconcat(mnemonic, "{q}\t",
712 "{$src2, $dst|$dst, $src2}"),
713 [(set EFLAGS, (Op addr:$dst, GR64:$src2))]>, LOCK;
715 // NOTE: These are order specific, we want the mi8 forms to be listed
716 // first so that they are slightly preferred to the mi forms.
717 def NAME#16mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
718 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
719 ImmMod, (outs), (ins i16mem :$dst, i16i8imm :$src2),
720 !strconcat(mnemonic, "{w}\t",
721 "{$src2, $dst|$dst, $src2}"),
722 [(set EFLAGS, (Op addr:$dst, i16immSExt8:$src2))]>,
725 def NAME#32mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
726 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
727 ImmMod, (outs), (ins i32mem :$dst, i32i8imm :$src2),
728 !strconcat(mnemonic, "{l}\t",
729 "{$src2, $dst|$dst, $src2}"),
730 [(set EFLAGS, (Op addr:$dst, i32immSExt8:$src2))]>,
733 def NAME#64mi8 : RIi8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
734 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
735 ImmMod, (outs), (ins i64mem :$dst, i64i8imm :$src2),
736 !strconcat(mnemonic, "{q}\t",
737 "{$src2, $dst|$dst, $src2}"),
738 [(set EFLAGS, (Op addr:$dst, i64immSExt8:$src2))]>,
741 def NAME#8mi : Ii8<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
742 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 0 },
743 ImmMod, (outs), (ins i8mem :$dst, i8imm :$src2),
744 !strconcat(mnemonic, "{b}\t",
745 "{$src2, $dst|$dst, $src2}"),
746 [(set EFLAGS, (Op addr:$dst, (i8 imm:$src2)))]>, LOCK;
748 def NAME#16mi : Ii16<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
749 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
750 ImmMod, (outs), (ins i16mem :$dst, i16imm :$src2),
751 !strconcat(mnemonic, "{w}\t",
752 "{$src2, $dst|$dst, $src2}"),
753 [(set EFLAGS, (Op addr:$dst, (i16 imm:$src2)))]>,
756 def NAME#32mi : Ii32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
757 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
758 ImmMod, (outs), (ins i32mem :$dst, i32imm :$src2),
759 !strconcat(mnemonic, "{l}\t",
760 "{$src2, $dst|$dst, $src2}"),
761 [(set EFLAGS, (Op addr:$dst, (i32 imm:$src2)))]>,
764 def NAME#64mi32 : RIi32S<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
765 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
766 ImmMod, (outs), (ins i64mem :$dst, i64i32imm :$src2),
767 !strconcat(mnemonic, "{q}\t",
768 "{$src2, $dst|$dst, $src2}"),
769 [(set EFLAGS, (Op addr:$dst, i64immSExt32:$src2))]>,
775 defm LOCK_ADD : LOCK_ArithBinOp<0x00, 0x80, 0x83, MRM0m, X86lock_add, "add">;
776 defm LOCK_SUB : LOCK_ArithBinOp<0x28, 0x80, 0x83, MRM5m, X86lock_sub, "sub">;
777 defm LOCK_OR : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM1m, X86lock_or , "or">;
778 defm LOCK_AND : LOCK_ArithBinOp<0x20, 0x80, 0x83, MRM4m, X86lock_and, "and">;
779 defm LOCK_XOR : LOCK_ArithBinOp<0x30, 0x80, 0x83, MRM6m, X86lock_xor, "xor">;
781 def X86lock_add_nocf : PatFrag<(ops node:$lhs, node:$rhs),
782 (X86lock_add node:$lhs, node:$rhs), [{
783 return hasNoCarryFlagUses(SDValue(N, 0));
786 def X86lock_sub_nocf : PatFrag<(ops node:$lhs, node:$rhs),
787 (X86lock_sub node:$lhs, node:$rhs), [{
788 return hasNoCarryFlagUses(SDValue(N, 0));
791 let Predicates = [UseIncDec] in {
792 let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
793 SchedRW = [WriteALURMW] in {
794 def LOCK_INC8m : I<0xFE, MRM0m, (outs), (ins i8mem :$dst),
796 [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i8 1)))]>,
798 def LOCK_INC16m : I<0xFF, MRM0m, (outs), (ins i16mem:$dst),
800 [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i16 1)))]>,
802 def LOCK_INC32m : I<0xFF, MRM0m, (outs), (ins i32mem:$dst),
804 [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i32 1)))]>,
806 def LOCK_INC64m : RI<0xFF, MRM0m, (outs), (ins i64mem:$dst),
808 [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i64 1)))]>,
811 def LOCK_DEC8m : I<0xFE, MRM1m, (outs), (ins i8mem :$dst),
813 [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i8 1)))]>,
815 def LOCK_DEC16m : I<0xFF, MRM1m, (outs), (ins i16mem:$dst),
817 [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i16 1)))]>,
819 def LOCK_DEC32m : I<0xFF, MRM1m, (outs), (ins i32mem:$dst),
821 [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i32 1)))]>,
823 def LOCK_DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst),
825 [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i64 1)))]>,
829 // Additional patterns for -1 constant.
830 def : Pat<(X86lock_add addr:$dst, (i8 -1)), (LOCK_DEC8m addr:$dst)>;
831 def : Pat<(X86lock_add addr:$dst, (i16 -1)), (LOCK_DEC16m addr:$dst)>;
832 def : Pat<(X86lock_add addr:$dst, (i32 -1)), (LOCK_DEC32m addr:$dst)>;
833 def : Pat<(X86lock_add addr:$dst, (i64 -1)), (LOCK_DEC64m addr:$dst)>;
834 def : Pat<(X86lock_sub addr:$dst, (i8 -1)), (LOCK_INC8m addr:$dst)>;
835 def : Pat<(X86lock_sub addr:$dst, (i16 -1)), (LOCK_INC16m addr:$dst)>;
836 def : Pat<(X86lock_sub addr:$dst, (i32 -1)), (LOCK_INC32m addr:$dst)>;
837 def : Pat<(X86lock_sub addr:$dst, (i64 -1)), (LOCK_INC64m addr:$dst)>;
840 // Atomic compare and swap.
841 multiclass LCMPXCHG_UnOp<bits<8> Opc, Format Form, string mnemonic,
842 SDPatternOperator frag, X86MemOperand x86memop> {
843 let isCodeGenOnly = 1, usesCustomInserter = 1 in {
844 def NAME : I<Opc, Form, (outs), (ins x86memop:$ptr),
845 !strconcat(mnemonic, "\t$ptr"),
846 [(frag addr:$ptr)]>, TB, LOCK;
850 multiclass LCMPXCHG_BinOp<bits<8> Opc8, bits<8> Opc, Format Form,
851 string mnemonic, SDPatternOperator frag> {
852 let isCodeGenOnly = 1, SchedRW = [WriteCMPXCHGRMW] in {
853 let Defs = [AL, EFLAGS], Uses = [AL] in
854 def NAME#8 : I<Opc8, Form, (outs), (ins i8mem:$ptr, GR8:$swap),
855 !strconcat(mnemonic, "{b}\t{$swap, $ptr|$ptr, $swap}"),
856 [(frag addr:$ptr, GR8:$swap, 1)]>, TB, LOCK;
857 let Defs = [AX, EFLAGS], Uses = [AX] in
858 def NAME#16 : I<Opc, Form, (outs), (ins i16mem:$ptr, GR16:$swap),
859 !strconcat(mnemonic, "{w}\t{$swap, $ptr|$ptr, $swap}"),
860 [(frag addr:$ptr, GR16:$swap, 2)]>, TB, OpSize16, LOCK;
861 let Defs = [EAX, EFLAGS], Uses = [EAX] in
862 def NAME#32 : I<Opc, Form, (outs), (ins i32mem:$ptr, GR32:$swap),
863 !strconcat(mnemonic, "{l}\t{$swap, $ptr|$ptr, $swap}"),
864 [(frag addr:$ptr, GR32:$swap, 4)]>, TB, OpSize32, LOCK;
865 let Defs = [RAX, EFLAGS], Uses = [RAX] in
866 def NAME#64 : RI<Opc, Form, (outs), (ins i64mem:$ptr, GR64:$swap),
867 !strconcat(mnemonic, "{q}\t{$swap, $ptr|$ptr, $swap}"),
868 [(frag addr:$ptr, GR64:$swap, 8)]>, TB, LOCK;
872 let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
873 Predicates = [HasCmpxchg8b], SchedRW = [WriteCMPXCHGRMW] in {
874 defm LCMPXCHG8B : LCMPXCHG_UnOp<0xC7, MRM1m, "cmpxchg8b", X86cas8, i64mem>;
877 // This pseudo must be used when the frame uses RBX as
878 // the base pointer. Indeed, in such situation RBX is a reserved
879 // register and the register allocator will ignore any use/def of
880 // it. In other words, the register will not fix the clobbering of
881 // RBX that will happen when setting the arguments for the instrucion.
883 // Unlike the actual related instuction, we mark that this one
884 // defines EBX (instead of using EBX).
885 // The rationale is that we will define RBX during the expansion of
886 // the pseudo. The argument feeding EBX is ebx_input.
888 // The additional argument, $ebx_save, is a temporary register used to
889 // save the value of RBX across the actual instruction.
891 // To make sure the register assigned to $ebx_save does not interfere with
892 // the definition of the actual instruction, we use a definition $dst which
893 // is tied to $rbx_save. That way, the live-range of $rbx_save spans across
894 // the instruction and we are sure we will have a valid register to restore
896 let Defs = [EAX, EDX, EBX, EFLAGS], Uses = [EAX, ECX, EDX],
897 Predicates = [HasCmpxchg8b], SchedRW = [WriteCMPXCHGRMW],
898 isCodeGenOnly = 1, isPseudo = 1, Constraints = "$ebx_save = $dst",
899 usesCustomInserter = 1 in {
900 def LCMPXCHG8B_SAVE_EBX :
901 I<0, Pseudo, (outs GR32:$dst),
902 (ins i64mem:$ptr, GR32:$ebx_input, GR32:$ebx_save),
903 !strconcat("cmpxchg8b", "\t$ptr"),
904 [(set GR32:$dst, (X86cas8save_ebx addr:$ptr, GR32:$ebx_input,
909 let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RBX, RCX, RDX],
910 Predicates = [HasCmpxchg16b,In64BitMode], SchedRW = [WriteCMPXCHGRMW] in {
911 defm LCMPXCHG16B : LCMPXCHG_UnOp<0xC7, MRM1m, "cmpxchg16b",
912 X86cas16, i128mem>, REX_W;
915 // Same as LCMPXCHG8B_SAVE_RBX but for the 16 Bytes variant.
916 let Defs = [RAX, RDX, RBX, EFLAGS], Uses = [RAX, RCX, RDX],
917 Predicates = [HasCmpxchg16b,In64BitMode], SchedRW = [WriteCMPXCHGRMW],
918 isCodeGenOnly = 1, isPseudo = 1, Constraints = "$rbx_save = $dst",
919 usesCustomInserter = 1 in {
920 def LCMPXCHG16B_SAVE_RBX :
921 I<0, Pseudo, (outs GR64:$dst),
922 (ins i128mem:$ptr, GR64:$rbx_input, GR64:$rbx_save),
923 !strconcat("cmpxchg16b", "\t$ptr"),
924 [(set GR64:$dst, (X86cas16save_rbx addr:$ptr, GR64:$rbx_input,
928 defm LCMPXCHG : LCMPXCHG_BinOp<0xB0, 0xB1, MRMDestMem, "cmpxchg", X86cas>;
930 // Atomic exchange and add
931 multiclass ATOMIC_LOAD_BINOP<bits<8> opc8, bits<8> opc, string mnemonic,
933 let Constraints = "$val = $dst", Defs = [EFLAGS], isCodeGenOnly = 1,
934 SchedRW = [WriteALURMW] in {
935 def NAME#8 : I<opc8, MRMSrcMem, (outs GR8:$dst),
936 (ins GR8:$val, i8mem:$ptr),
937 !strconcat(mnemonic, "{b}\t{$val, $ptr|$ptr, $val}"),
939 (!cast<PatFrag>(frag # "_8") addr:$ptr, GR8:$val))]>;
940 def NAME#16 : I<opc, MRMSrcMem, (outs GR16:$dst),
941 (ins GR16:$val, i16mem:$ptr),
942 !strconcat(mnemonic, "{w}\t{$val, $ptr|$ptr, $val}"),
945 (!cast<PatFrag>(frag # "_16") addr:$ptr, GR16:$val))]>,
947 def NAME#32 : I<opc, MRMSrcMem, (outs GR32:$dst),
948 (ins GR32:$val, i32mem:$ptr),
949 !strconcat(mnemonic, "{l}\t{$val, $ptr|$ptr, $val}"),
952 (!cast<PatFrag>(frag # "_32") addr:$ptr, GR32:$val))]>,
954 def NAME#64 : RI<opc, MRMSrcMem, (outs GR64:$dst),
955 (ins GR64:$val, i64mem:$ptr),
956 !strconcat(mnemonic, "{q}\t{$val, $ptr|$ptr, $val}"),
959 (!cast<PatFrag>(frag # "_64") addr:$ptr, GR64:$val))]>;
963 defm LXADD : ATOMIC_LOAD_BINOP<0xc0, 0xc1, "xadd", "atomic_load_add">, TB, LOCK;
965 /* The following multiclass tries to make sure that in code like
966 * x.store (immediate op x.load(acquire), release)
968 * x.store (register op x.load(acquire), release)
969 * an operation directly on memory is generated instead of wasting a register.
970 * It is not automatic as atomic_store/load are only lowered to MOV instructions
971 * extremely late to prevent them from being accidentally reordered in the backend
972 * (see below the RELEASE_MOV* / ACQUIRE_MOV* pseudo-instructions)
974 multiclass RELEASE_BINOP_MI<string Name, SDNode op> {
975 def : Pat<(atomic_store_8 addr:$dst,
976 (op (atomic_load_8 addr:$dst), (i8 imm:$src))),
977 (!cast<Instruction>(Name#"8mi") addr:$dst, imm:$src)>;
978 def : Pat<(atomic_store_16 addr:$dst,
979 (op (atomic_load_16 addr:$dst), (i16 imm:$src))),
980 (!cast<Instruction>(Name#"16mi") addr:$dst, imm:$src)>;
981 def : Pat<(atomic_store_32 addr:$dst,
982 (op (atomic_load_32 addr:$dst), (i32 imm:$src))),
983 (!cast<Instruction>(Name#"32mi") addr:$dst, imm:$src)>;
984 def : Pat<(atomic_store_64 addr:$dst,
985 (op (atomic_load_64 addr:$dst), (i64immSExt32:$src))),
986 (!cast<Instruction>(Name#"64mi32") addr:$dst, (i64immSExt32:$src))>;
988 def : Pat<(atomic_store_8 addr:$dst,
989 (op (atomic_load_8 addr:$dst), (i8 GR8:$src))),
990 (!cast<Instruction>(Name#"8mr") addr:$dst, GR8:$src)>;
991 def : Pat<(atomic_store_16 addr:$dst,
992 (op (atomic_load_16 addr:$dst), (i16 GR16:$src))),
993 (!cast<Instruction>(Name#"16mr") addr:$dst, GR16:$src)>;
994 def : Pat<(atomic_store_32 addr:$dst,
995 (op (atomic_load_32 addr:$dst), (i32 GR32:$src))),
996 (!cast<Instruction>(Name#"32mr") addr:$dst, GR32:$src)>;
997 def : Pat<(atomic_store_64 addr:$dst,
998 (op (atomic_load_64 addr:$dst), (i64 GR64:$src))),
999 (!cast<Instruction>(Name#"64mr") addr:$dst, GR64:$src)>;
1001 defm : RELEASE_BINOP_MI<"ADD", add>;
1002 defm : RELEASE_BINOP_MI<"AND", and>;
1003 defm : RELEASE_BINOP_MI<"OR", or>;
1004 defm : RELEASE_BINOP_MI<"XOR", xor>;
1005 defm : RELEASE_BINOP_MI<"SUB", sub>;
1007 // Atomic load + floating point patterns.
1008 // FIXME: This could also handle SIMD operations with *ps and *pd instructions.
1009 multiclass ATOMIC_LOAD_FP_BINOP_MI<string Name, SDNode op> {
1010 def : Pat<(op FR32:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
1011 (!cast<Instruction>(Name#"SSrm") FR32:$src1, addr:$src2)>,
1012 Requires<[UseSSE1]>;
1013 def : Pat<(op FR32:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
1014 (!cast<Instruction>("V"#Name#"SSrm") FR32:$src1, addr:$src2)>,
1016 def : Pat<(op FR32X:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
1017 (!cast<Instruction>("V"#Name#"SSZrm") FR32X:$src1, addr:$src2)>,
1018 Requires<[HasAVX512]>;
1020 def : Pat<(op FR64:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
1021 (!cast<Instruction>(Name#"SDrm") FR64:$src1, addr:$src2)>,
1022 Requires<[UseSSE1]>;
1023 def : Pat<(op FR64:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
1024 (!cast<Instruction>("V"#Name#"SDrm") FR64:$src1, addr:$src2)>,
1026 def : Pat<(op FR64X:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
1027 (!cast<Instruction>("V"#Name#"SDZrm") FR64X:$src1, addr:$src2)>,
1028 Requires<[HasAVX512]>;
1030 defm : ATOMIC_LOAD_FP_BINOP_MI<"ADD", fadd>;
1031 // FIXME: Add fsub, fmul, fdiv, ...
1033 multiclass RELEASE_UNOP<string Name, dag dag8, dag dag16, dag dag32,
1035 def : Pat<(atomic_store_8 addr:$dst, dag8),
1036 (!cast<Instruction>(Name#8m) addr:$dst)>;
1037 def : Pat<(atomic_store_16 addr:$dst, dag16),
1038 (!cast<Instruction>(Name#16m) addr:$dst)>;
1039 def : Pat<(atomic_store_32 addr:$dst, dag32),
1040 (!cast<Instruction>(Name#32m) addr:$dst)>;
1041 def : Pat<(atomic_store_64 addr:$dst, dag64),
1042 (!cast<Instruction>(Name#64m) addr:$dst)>;
1045 let Predicates = [UseIncDec] in {
1046 defm : RELEASE_UNOP<"INC",
1047 (add (atomic_load_8 addr:$dst), (i8 1)),
1048 (add (atomic_load_16 addr:$dst), (i16 1)),
1049 (add (atomic_load_32 addr:$dst), (i32 1)),
1050 (add (atomic_load_64 addr:$dst), (i64 1))>;
1051 defm : RELEASE_UNOP<"DEC",
1052 (add (atomic_load_8 addr:$dst), (i8 -1)),
1053 (add (atomic_load_16 addr:$dst), (i16 -1)),
1054 (add (atomic_load_32 addr:$dst), (i32 -1)),
1055 (add (atomic_load_64 addr:$dst), (i64 -1))>;
1058 defm : RELEASE_UNOP<"NEG",
1059 (ineg (i8 (atomic_load_8 addr:$dst))),
1060 (ineg (i16 (atomic_load_16 addr:$dst))),
1061 (ineg (i32 (atomic_load_32 addr:$dst))),
1062 (ineg (i64 (atomic_load_64 addr:$dst)))>;
1063 defm : RELEASE_UNOP<"NOT",
1064 (not (i8 (atomic_load_8 addr:$dst))),
1065 (not (i16 (atomic_load_16 addr:$dst))),
1066 (not (i32 (atomic_load_32 addr:$dst))),
1067 (not (i64 (atomic_load_64 addr:$dst)))>;
1069 def : Pat<(atomic_store_8 addr:$dst, (i8 imm:$src)),
1070 (MOV8mi addr:$dst, imm:$src)>;
1071 def : Pat<(atomic_store_16 addr:$dst, (i16 imm:$src)),
1072 (MOV16mi addr:$dst, imm:$src)>;
1073 def : Pat<(atomic_store_32 addr:$dst, (i32 imm:$src)),
1074 (MOV32mi addr:$dst, imm:$src)>;
1075 def : Pat<(atomic_store_64 addr:$dst, (i64immSExt32:$src)),
1076 (MOV64mi32 addr:$dst, i64immSExt32:$src)>;
1078 def : Pat<(atomic_store_8 addr:$dst, GR8:$src),
1079 (MOV8mr addr:$dst, GR8:$src)>;
1080 def : Pat<(atomic_store_16 addr:$dst, GR16:$src),
1081 (MOV16mr addr:$dst, GR16:$src)>;
1082 def : Pat<(atomic_store_32 addr:$dst, GR32:$src),
1083 (MOV32mr addr:$dst, GR32:$src)>;
1084 def : Pat<(atomic_store_64 addr:$dst, GR64:$src),
1085 (MOV64mr addr:$dst, GR64:$src)>;
1087 def : Pat<(i8 (atomic_load_8 addr:$src)), (MOV8rm addr:$src)>;
1088 def : Pat<(i16 (atomic_load_16 addr:$src)), (MOV16rm addr:$src)>;
1089 def : Pat<(i32 (atomic_load_32 addr:$src)), (MOV32rm addr:$src)>;
1090 def : Pat<(i64 (atomic_load_64 addr:$src)), (MOV64rm addr:$src)>;
1092 // Floating point loads/stores.
1093 def : Pat<(atomic_store_32 addr:$dst, (i32 (bitconvert (f32 FR32:$src)))),
1094 (MOVSSmr addr:$dst, FR32:$src)>, Requires<[UseSSE1]>;
1095 def : Pat<(atomic_store_32 addr:$dst, (i32 (bitconvert (f32 FR32:$src)))),
1096 (VMOVSSmr addr:$dst, FR32:$src)>, Requires<[UseAVX]>;
1097 def : Pat<(atomic_store_32 addr:$dst, (i32 (bitconvert (f32 FR32:$src)))),
1098 (VMOVSSZmr addr:$dst, FR32:$src)>, Requires<[HasAVX512]>;
1100 def : Pat<(atomic_store_64 addr:$dst, (i64 (bitconvert (f64 FR64:$src)))),
1101 (MOVSDmr addr:$dst, FR64:$src)>, Requires<[UseSSE2]>;
1102 def : Pat<(atomic_store_64 addr:$dst, (i64 (bitconvert (f64 FR64:$src)))),
1103 (VMOVSDmr addr:$dst, FR64:$src)>, Requires<[UseAVX]>;
1104 def : Pat<(atomic_store_64 addr:$dst, (i64 (bitconvert (f64 FR64:$src)))),
1105 (VMOVSDmr addr:$dst, FR64:$src)>, Requires<[HasAVX512]>;
1107 def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1108 (MOVSSrm_alt addr:$src)>, Requires<[UseSSE1]>;
1109 def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1110 (VMOVSSrm_alt addr:$src)>, Requires<[UseAVX]>;
1111 def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1112 (VMOVSSZrm_alt addr:$src)>, Requires<[HasAVX512]>;
1114 def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1115 (MOVSDrm_alt addr:$src)>, Requires<[UseSSE2]>;
1116 def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1117 (VMOVSDrm_alt addr:$src)>, Requires<[UseAVX]>;
1118 def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1119 (VMOVSDZrm_alt addr:$src)>, Requires<[HasAVX512]>;
1121 //===----------------------------------------------------------------------===//
1122 // DAG Pattern Matching Rules
1123 //===----------------------------------------------------------------------===//
1125 // Use AND/OR to store 0/-1 in memory when optimizing for minsize. This saves
1126 // binary size compared to a regular MOV, but it introduces an unnecessary
1127 // load, so is not suitable for regular or optsize functions.
1128 let Predicates = [OptForMinSize] in {
1129 def : Pat<(simple_store (i16 0), addr:$dst), (AND16mi8 addr:$dst, 0)>;
1130 def : Pat<(simple_store (i32 0), addr:$dst), (AND32mi8 addr:$dst, 0)>;
1131 def : Pat<(simple_store (i64 0), addr:$dst), (AND64mi8 addr:$dst, 0)>;
1132 def : Pat<(simple_store (i16 -1), addr:$dst), (OR16mi8 addr:$dst, -1)>;
1133 def : Pat<(simple_store (i32 -1), addr:$dst), (OR32mi8 addr:$dst, -1)>;
1134 def : Pat<(simple_store (i64 -1), addr:$dst), (OR64mi8 addr:$dst, -1)>;
1137 // In kernel code model, we can get the address of a label
1138 // into a register with 'movq'. FIXME: This is a hack, the 'imm' predicate of
1139 // the MOV64ri32 should accept these.
1140 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
1141 (MOV64ri32 tconstpool :$dst)>, Requires<[KernelCode]>;
1142 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
1143 (MOV64ri32 tjumptable :$dst)>, Requires<[KernelCode]>;
1144 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
1145 (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
1146 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
1147 (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
1148 def : Pat<(i64 (X86Wrapper mcsym:$dst)),
1149 (MOV64ri32 mcsym:$dst)>, Requires<[KernelCode]>;
1150 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
1151 (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
1153 // If we have small model and -static mode, it is safe to store global addresses
1154 // directly as immediates. FIXME: This is really a hack, the 'imm' predicate
1155 // for MOV64mi32 should handle this sort of thing.
1156 def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
1157 (MOV64mi32 addr:$dst, tconstpool:$src)>,
1158 Requires<[NearData, IsNotPIC]>;
1159 def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
1160 (MOV64mi32 addr:$dst, tjumptable:$src)>,
1161 Requires<[NearData, IsNotPIC]>;
1162 def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
1163 (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
1164 Requires<[NearData, IsNotPIC]>;
1165 def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
1166 (MOV64mi32 addr:$dst, texternalsym:$src)>,
1167 Requires<[NearData, IsNotPIC]>;
1168 def : Pat<(store (i64 (X86Wrapper mcsym:$src)), addr:$dst),
1169 (MOV64mi32 addr:$dst, mcsym:$src)>,
1170 Requires<[NearData, IsNotPIC]>;
1171 def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
1172 (MOV64mi32 addr:$dst, tblockaddress:$src)>,
1173 Requires<[NearData, IsNotPIC]>;
1175 def : Pat<(i32 (X86RecoverFrameAlloc mcsym:$dst)), (MOV32ri mcsym:$dst)>;
1176 def : Pat<(i64 (X86RecoverFrameAlloc mcsym:$dst)), (MOV64ri mcsym:$dst)>;
1180 // tls has some funny stuff here...
1181 // This corresponds to movabs $foo@tpoff, %rax
1182 def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
1183 (MOV64ri32 tglobaltlsaddr :$dst)>;
1184 // This corresponds to add $foo@tpoff, %rax
1185 def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
1186 (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
1189 // Direct PC relative function call for small code model. 32-bit displacement
1190 // sign extended to 64-bit.
1191 def : Pat<(X86call (i64 tglobaladdr:$dst)),
1192 (CALL64pcrel32 tglobaladdr:$dst)>;
1193 def : Pat<(X86call (i64 texternalsym:$dst)),
1194 (CALL64pcrel32 texternalsym:$dst)>;
1196 // Tailcall stuff. The TCRETURN instructions execute after the epilog, so they
1197 // can never use callee-saved registers. That is the purpose of the GR64_TC
1198 // register classes.
1200 // The only volatile register that is never used by the calling convention is
1201 // %r11. This happens when calling a vararg function with 6 arguments.
1203 // Match an X86tcret that uses less than 7 volatile registers.
1204 def X86tcret_6regs : PatFrag<(ops node:$ptr, node:$off),
1205 (X86tcret node:$ptr, node:$off), [{
1206 // X86tcret args: (*chain, ptr, imm, regs..., glue)
1207 unsigned NumRegs = 0;
1208 for (unsigned i = 3, e = N->getNumOperands(); i != e; ++i)
1209 if (isa<RegisterSDNode>(N->getOperand(i)) && ++NumRegs > 6)
1214 def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
1215 (TCRETURNri ptr_rc_tailcall:$dst, imm:$off)>,
1216 Requires<[Not64BitMode, NotUseRetpolineIndirectCalls]>;
1218 // FIXME: This is disabled for 32-bit PIC mode because the global base
1219 // register which is part of the address mode may be assigned a
1220 // callee-saved register.
1221 def : Pat<(X86tcret (load addr:$dst), imm:$off),
1222 (TCRETURNmi addr:$dst, imm:$off)>,
1223 Requires<[Not64BitMode, IsNotPIC, NotUseRetpolineIndirectCalls]>;
1225 def : Pat<(X86tcret (i32 tglobaladdr:$dst), imm:$off),
1226 (TCRETURNdi tglobaladdr:$dst, imm:$off)>,
1227 Requires<[NotLP64]>;
1229 def : Pat<(X86tcret (i32 texternalsym:$dst), imm:$off),
1230 (TCRETURNdi texternalsym:$dst, imm:$off)>,
1231 Requires<[NotLP64]>;
1233 def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
1234 (TCRETURNri64 ptr_rc_tailcall:$dst, imm:$off)>,
1235 Requires<[In64BitMode, NotUseRetpolineIndirectCalls]>;
1237 // Don't fold loads into X86tcret requiring more than 6 regs.
1238 // There wouldn't be enough scratch registers for base+index.
1239 def : Pat<(X86tcret_6regs (load addr:$dst), imm:$off),
1240 (TCRETURNmi64 addr:$dst, imm:$off)>,
1241 Requires<[In64BitMode, NotUseRetpolineIndirectCalls]>;
1243 def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
1244 (RETPOLINE_TCRETURN64 ptr_rc_tailcall:$dst, imm:$off)>,
1245 Requires<[In64BitMode, UseRetpolineIndirectCalls]>;
1247 def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
1248 (RETPOLINE_TCRETURN32 ptr_rc_tailcall:$dst, imm:$off)>,
1249 Requires<[Not64BitMode, UseRetpolineIndirectCalls]>;
1251 def : Pat<(X86tcret (i64 tglobaladdr:$dst), imm:$off),
1252 (TCRETURNdi64 tglobaladdr:$dst, imm:$off)>,
1255 def : Pat<(X86tcret (i64 texternalsym:$dst), imm:$off),
1256 (TCRETURNdi64 texternalsym:$dst, imm:$off)>,
1259 // Normal calls, with various flavors of addresses.
1260 def : Pat<(X86call (i32 tglobaladdr:$dst)),
1261 (CALLpcrel32 tglobaladdr:$dst)>;
1262 def : Pat<(X86call (i32 texternalsym:$dst)),
1263 (CALLpcrel32 texternalsym:$dst)>;
1264 def : Pat<(X86call (i32 imm:$dst)),
1265 (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
1269 // TEST R,R is smaller than CMP R,0
1270 def : Pat<(X86cmp GR8:$src1, 0),
1271 (TEST8rr GR8:$src1, GR8:$src1)>;
1272 def : Pat<(X86cmp GR16:$src1, 0),
1273 (TEST16rr GR16:$src1, GR16:$src1)>;
1274 def : Pat<(X86cmp GR32:$src1, 0),
1275 (TEST32rr GR32:$src1, GR32:$src1)>;
1276 def : Pat<(X86cmp GR64:$src1, 0),
1277 (TEST64rr GR64:$src1, GR64:$src1)>;
1279 // zextload bool -> zextload byte
1280 // i1 stored in one byte in zero-extended form.
1281 // Upper bits cleanup should be executed before Store.
1282 def : Pat<(zextloadi8i1 addr:$src), (MOV8rm addr:$src)>;
1283 def : Pat<(zextloadi16i1 addr:$src),
1284 (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1285 def : Pat<(zextloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1286 def : Pat<(zextloadi64i1 addr:$src),
1287 (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1289 // extload bool -> extload byte
1290 // When extloading from 16-bit and smaller memory locations into 64-bit
1291 // registers, use zero-extending loads so that the entire 64-bit register is
1292 // defined, avoiding partial-register updates.
1294 def : Pat<(extloadi8i1 addr:$src), (MOV8rm addr:$src)>;
1295 def : Pat<(extloadi16i1 addr:$src),
1296 (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1297 def : Pat<(extloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1298 def : Pat<(extloadi16i8 addr:$src),
1299 (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1300 def : Pat<(extloadi32i8 addr:$src), (MOVZX32rm8 addr:$src)>;
1301 def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
1303 // For other extloads, use subregs, since the high contents of the register are
1304 // defined after an extload.
1305 // NOTE: The extloadi64i32 pattern needs to be first as it will try to form
1306 // 32-bit loads for 4 byte aligned i8/i16 loads.
1307 def : Pat<(extloadi64i32 addr:$src),
1308 (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src), sub_32bit)>;
1309 def : Pat<(extloadi64i1 addr:$src),
1310 (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1311 def : Pat<(extloadi64i8 addr:$src),
1312 (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1313 def : Pat<(extloadi64i16 addr:$src),
1314 (SUBREG_TO_REG (i64 0), (MOVZX32rm16 addr:$src), sub_32bit)>;
1316 // anyext. Define these to do an explicit zero-extend to
1317 // avoid partial-register updates.
1318 def : Pat<(i16 (anyext GR8 :$src)), (EXTRACT_SUBREG
1319 (MOVZX32rr8 GR8 :$src), sub_16bit)>;
1320 def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8 GR8 :$src)>;
1322 // Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
1323 def : Pat<(i32 (anyext GR16:$src)),
1324 (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
1326 def : Pat<(i64 (anyext GR8 :$src)),
1327 (SUBREG_TO_REG (i64 0), (MOVZX32rr8 GR8 :$src), sub_32bit)>;
1328 def : Pat<(i64 (anyext GR16:$src)),
1329 (SUBREG_TO_REG (i64 0), (MOVZX32rr16 GR16 :$src), sub_32bit)>;
1330 def : Pat<(i64 (anyext GR32:$src)),
1331 (INSERT_SUBREG (i64 (IMPLICIT_DEF)), GR32:$src, sub_32bit)>;
1333 // If this is an anyext of the remainder of an 8-bit sdivrem, use a MOVSX
1334 // instead of a MOVZX. The sdivrem lowering will emit emit a MOVSX to move
1335 // %ah to the lower byte of a register. By using a MOVSX here we allow a
1336 // post-isel peephole to merge the two MOVSX instructions into one.
1337 def anyext_sdiv : PatFrag<(ops node:$lhs), (anyext node:$lhs),[{
1338 return (N->getOperand(0).getOpcode() == ISD::SDIVREM &&
1339 N->getOperand(0).getResNo() == 1);
1341 def : Pat<(i32 (anyext_sdiv GR8:$src)), (MOVSX32rr8 GR8:$src)>;
1343 // Any instruction that defines a 32-bit result leaves the high half of the
1344 // register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
1345 // be copying from a truncate. Any other 32-bit operation will zero-extend
1346 // up to 64 bits. AssertSext/AssertZext aren't saying anything about the upper
1347 // 32 bits, they're probably just qualifying a CopyFromReg.
1348 def def32 : PatLeaf<(i32 GR32:$src), [{
1349 return N->getOpcode() != ISD::TRUNCATE &&
1350 N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
1351 N->getOpcode() != ISD::CopyFromReg &&
1352 N->getOpcode() != ISD::AssertSext &&
1353 N->getOpcode() != ISD::AssertZext;
1356 // In the case of a 32-bit def that is known to implicitly zero-extend,
1357 // we can use a SUBREG_TO_REG.
1358 def : Pat<(i64 (zext def32:$src)),
1359 (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1360 def : Pat<(i64 (and (anyext def32:$src), 0x00000000FFFFFFFF)),
1361 (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1363 //===----------------------------------------------------------------------===//
1364 // Pattern match OR as ADD
1365 //===----------------------------------------------------------------------===//
1367 // If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
1368 // 3-addressified into an LEA instruction to avoid copies. However, we also
1369 // want to finally emit these instructions as an or at the end of the code
1370 // generator to make the generated code easier to read. To do this, we select
1371 // into "disjoint bits" pseudo ops.
1373 // Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.
1374 def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
1375 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1376 return CurDAG->MaskedValueIsZero(N->getOperand(0), CN->getAPIntValue());
1378 KnownBits Known0 = CurDAG->computeKnownBits(N->getOperand(0), 0);
1379 KnownBits Known1 = CurDAG->computeKnownBits(N->getOperand(1), 0);
1380 return (~Known0.Zero & ~Known1.Zero) == 0;
1384 // (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1385 // Try this before the selecting to OR.
1386 let SchedRW = [WriteALU] in {
1388 let isConvertibleToThreeAddress = 1, isPseudo = 1,
1389 Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1390 let isCommutable = 1 in {
1391 def ADD8rr_DB : I<0, Pseudo, (outs GR8:$dst), (ins GR8:$src1, GR8:$src2),
1392 "", // orb/addb REG, REG
1393 [(set GR8:$dst, (or_is_add GR8:$src1, GR8:$src2))]>;
1394 def ADD16rr_DB : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1395 "", // orw/addw REG, REG
1396 [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1397 def ADD32rr_DB : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1398 "", // orl/addl REG, REG
1399 [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1400 def ADD64rr_DB : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1401 "", // orq/addq REG, REG
1402 [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1405 // NOTE: These are order specific, we want the ri8 forms to be listed
1406 // first so that they are slightly preferred to the ri forms.
1408 def ADD8ri_DB : I<0, Pseudo,
1409 (outs GR8:$dst), (ins GR8:$src1, i8imm:$src2),
1410 "", // orb/addb REG, imm8
1411 [(set GR8:$dst, (or_is_add GR8:$src1, imm:$src2))]>;
1412 def ADD16ri8_DB : I<0, Pseudo,
1413 (outs GR16:$dst), (ins GR16:$src1, i16i8imm:$src2),
1414 "", // orw/addw REG, imm8
1415 [(set GR16:$dst,(or_is_add GR16:$src1,i16immSExt8:$src2))]>;
1416 def ADD16ri_DB : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1417 "", // orw/addw REG, imm
1418 [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1420 def ADD32ri8_DB : I<0, Pseudo,
1421 (outs GR32:$dst), (ins GR32:$src1, i32i8imm:$src2),
1422 "", // orl/addl REG, imm8
1423 [(set GR32:$dst,(or_is_add GR32:$src1,i32immSExt8:$src2))]>;
1424 def ADD32ri_DB : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1425 "", // orl/addl REG, imm
1426 [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1429 def ADD64ri8_DB : I<0, Pseudo,
1430 (outs GR64:$dst), (ins GR64:$src1, i64i8imm:$src2),
1431 "", // orq/addq REG, imm8
1432 [(set GR64:$dst, (or_is_add GR64:$src1,
1433 i64immSExt8:$src2))]>;
1434 def ADD64ri32_DB : I<0, Pseudo,
1435 (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1436 "", // orq/addq REG, imm
1437 [(set GR64:$dst, (or_is_add GR64:$src1,
1438 i64immSExt32:$src2))]>;
1440 } // AddedComplexity, SchedRW
1442 //===----------------------------------------------------------------------===//
1443 // Pattern match SUB as XOR
1444 //===----------------------------------------------------------------------===//
1446 // An immediate in the LHS of a subtract can't be encoded in the instruction.
1447 // If there is no possibility of a borrow we can use an XOR instead of a SUB
1448 // to enable the immediate to be folded.
1449 // TODO: Move this to a DAG combine?
1451 def sub_is_xor : PatFrag<(ops node:$lhs, node:$rhs), (sub node:$lhs, node:$rhs),[{
1452 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1453 KnownBits Known = CurDAG->computeKnownBits(N->getOperand(1));
1455 // If all possible ones in the RHS are set in the LHS then there can't be
1456 // a borrow and we can use xor.
1457 return (~Known.Zero).isSubsetOf(CN->getAPIntValue());
1463 let AddedComplexity = 5 in {
1464 def : Pat<(sub_is_xor imm:$src2, GR8:$src1),
1465 (XOR8ri GR8:$src1, imm:$src2)>;
1466 def : Pat<(sub_is_xor i16immSExt8:$src2, GR16:$src1),
1467 (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1468 def : Pat<(sub_is_xor imm:$src2, GR16:$src1),
1469 (XOR16ri GR16:$src1, imm:$src2)>;
1470 def : Pat<(sub_is_xor i32immSExt8:$src2, GR32:$src1),
1471 (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1472 def : Pat<(sub_is_xor imm:$src2, GR32:$src1),
1473 (XOR32ri GR32:$src1, imm:$src2)>;
1474 def : Pat<(sub_is_xor i64immSExt8:$src2, GR64:$src1),
1475 (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1476 def : Pat<(sub_is_xor i64immSExt32:$src2, GR64:$src1),
1477 (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1480 //===----------------------------------------------------------------------===//
1482 //===----------------------------------------------------------------------===//
1484 // Odd encoding trick: -128 fits into an 8-bit immediate field while
1485 // +128 doesn't, so in this special case use a sub instead of an add.
1486 def : Pat<(add GR16:$src1, 128),
1487 (SUB16ri8 GR16:$src1, -128)>;
1488 def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1489 (SUB16mi8 addr:$dst, -128)>;
1491 def : Pat<(add GR32:$src1, 128),
1492 (SUB32ri8 GR32:$src1, -128)>;
1493 def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1494 (SUB32mi8 addr:$dst, -128)>;
1496 def : Pat<(add GR64:$src1, 128),
1497 (SUB64ri8 GR64:$src1, -128)>;
1498 def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1499 (SUB64mi8 addr:$dst, -128)>;
1501 def : Pat<(X86add_flag_nocf GR16:$src1, 128),
1502 (SUB16ri8 GR16:$src1, -128)>;
1503 def : Pat<(X86add_flag_nocf GR32:$src1, 128),
1504 (SUB32ri8 GR32:$src1, -128)>;
1505 def : Pat<(X86add_flag_nocf GR64:$src1, 128),
1506 (SUB64ri8 GR64:$src1, -128)>;
1508 // The same trick applies for 32-bit immediate fields in 64-bit
1510 def : Pat<(add GR64:$src1, 0x0000000080000000),
1511 (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1512 def : Pat<(store (add (loadi64 addr:$dst), 0x0000000080000000), addr:$dst),
1513 (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1515 def : Pat<(X86add_flag_nocf GR64:$src1, 0x0000000080000000),
1516 (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1518 // To avoid needing to materialize an immediate in a register, use a 32-bit and
1519 // with implicit zero-extension instead of a 64-bit and if the immediate has at
1520 // least 32 bits of leading zeros. If in addition the last 32 bits can be
1521 // represented with a sign extension of a 8 bit constant, use that.
1522 // This can also reduce instruction size by eliminating the need for the REX
1525 // AddedComplexity is needed to give priority over i64immSExt8 and i64immSExt32.
1526 let AddedComplexity = 1 in {
1527 def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
1531 (EXTRACT_SUBREG GR64:$src, sub_32bit),
1532 (i32 (GetLo32XForm imm:$imm))),
1535 def : Pat<(and GR64:$src, i64immZExt32:$imm),
1539 (EXTRACT_SUBREG GR64:$src, sub_32bit),
1540 (i32 (GetLo32XForm imm:$imm))),
1542 } // AddedComplexity = 1
1545 // AddedComplexity is needed due to the increased complexity on the
1546 // i64immZExt32SExt8 and i64immZExt32 patterns above. Applying this to all
1547 // the MOVZX patterns keeps thems together in DAGIsel tables.
1548 let AddedComplexity = 1 in {
1549 // r & (2^16-1) ==> movz
1550 def : Pat<(and GR32:$src1, 0xffff),
1551 (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1552 // r & (2^8-1) ==> movz
1553 def : Pat<(and GR32:$src1, 0xff),
1554 (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>;
1555 // r & (2^8-1) ==> movz
1556 def : Pat<(and GR16:$src1, 0xff),
1557 (EXTRACT_SUBREG (MOVZX32rr8 (EXTRACT_SUBREG GR16:$src1, sub_8bit)),
1560 // r & (2^32-1) ==> movz
1561 def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1562 (SUBREG_TO_REG (i64 0),
1563 (MOV32rr (EXTRACT_SUBREG GR64:$src, sub_32bit)),
1565 // r & (2^16-1) ==> movz
1566 def : Pat<(and GR64:$src, 0xffff),
1567 (SUBREG_TO_REG (i64 0),
1568 (MOVZX32rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit))),
1570 // r & (2^8-1) ==> movz
1571 def : Pat<(and GR64:$src, 0xff),
1572 (SUBREG_TO_REG (i64 0),
1573 (MOVZX32rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit))),
1575 } // AddedComplexity = 1
1578 // Try to use BTS/BTR/BTC for single bit operations on the upper 32-bits.
1580 def BTRXForm : SDNodeXForm<imm, [{
1581 // Transformation function: Find the lowest 0.
1582 return getI64Imm((uint8_t)N->getAPIntValue().countTrailingOnes(), SDLoc(N));
1585 def BTCBTSXForm : SDNodeXForm<imm, [{
1586 // Transformation function: Find the lowest 1.
1587 return getI64Imm((uint8_t)N->getAPIntValue().countTrailingZeros(), SDLoc(N));
1590 def BTRMask64 : ImmLeaf<i64, [{
1591 return !isUInt<32>(Imm) && !isInt<32>(Imm) && isPowerOf2_64(~Imm);
1594 def BTCBTSMask64 : ImmLeaf<i64, [{
1595 return !isInt<32>(Imm) && isPowerOf2_64(Imm);
1598 // For now only do this for optsize.
1599 let AddedComplexity = 1, Predicates=[OptForSize] in {
1600 def : Pat<(and GR64:$src1, BTRMask64:$mask),
1601 (BTR64ri8 GR64:$src1, (BTRXForm imm:$mask))>;
1602 def : Pat<(or GR64:$src1, BTCBTSMask64:$mask),
1603 (BTS64ri8 GR64:$src1, (BTCBTSXForm imm:$mask))>;
1604 def : Pat<(xor GR64:$src1, BTCBTSMask64:$mask),
1605 (BTC64ri8 GR64:$src1, (BTCBTSXForm imm:$mask))>;
1609 // sext_inreg patterns
1610 def : Pat<(sext_inreg GR32:$src, i16),
1611 (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1612 def : Pat<(sext_inreg GR32:$src, i8),
1613 (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>;
1615 def : Pat<(sext_inreg GR16:$src, i8),
1616 (EXTRACT_SUBREG (MOVSX32rr8 (EXTRACT_SUBREG GR16:$src, sub_8bit)),
1619 def : Pat<(sext_inreg GR64:$src, i32),
1620 (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1621 def : Pat<(sext_inreg GR64:$src, i16),
1622 (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1623 def : Pat<(sext_inreg GR64:$src, i8),
1624 (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1626 // sext, sext_load, zext, zext_load
1627 def: Pat<(i16 (sext GR8:$src)),
1628 (EXTRACT_SUBREG (MOVSX32rr8 GR8:$src), sub_16bit)>;
1629 def: Pat<(sextloadi16i8 addr:$src),
1630 (EXTRACT_SUBREG (MOVSX32rm8 addr:$src), sub_16bit)>;
1631 def: Pat<(i16 (zext GR8:$src)),
1632 (EXTRACT_SUBREG (MOVZX32rr8 GR8:$src), sub_16bit)>;
1633 def: Pat<(zextloadi16i8 addr:$src),
1634 (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1637 def : Pat<(i16 (trunc GR32:$src)),
1638 (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1639 def : Pat<(i8 (trunc GR32:$src)),
1640 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1642 Requires<[Not64BitMode]>;
1643 def : Pat<(i8 (trunc GR16:$src)),
1644 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1646 Requires<[Not64BitMode]>;
1647 def : Pat<(i32 (trunc GR64:$src)),
1648 (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1649 def : Pat<(i16 (trunc GR64:$src)),
1650 (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1651 def : Pat<(i8 (trunc GR64:$src)),
1652 (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1653 def : Pat<(i8 (trunc GR32:$src)),
1654 (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1655 Requires<[In64BitMode]>;
1656 def : Pat<(i8 (trunc GR16:$src)),
1657 (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1658 Requires<[In64BitMode]>;
1660 def immff00_ffff : ImmLeaf<i32, [{
1661 return Imm >= 0xff00 && Imm <= 0xffff;
1664 // h-register tricks
1665 def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1666 (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)>,
1667 Requires<[Not64BitMode]>;
1668 def : Pat<(i8 (trunc (srl_su (i32 (anyext GR16:$src)), (i8 8)))),
1669 (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)>,
1670 Requires<[Not64BitMode]>;
1671 def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1672 (EXTRACT_SUBREG GR32:$src, sub_8bit_hi)>,
1673 Requires<[Not64BitMode]>;
1674 def : Pat<(srl GR16:$src, (i8 8)),
1676 (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1678 def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1679 (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>;
1680 def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1681 (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>;
1682 def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1683 (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1684 def : Pat<(srl (and_su GR32:$src, immff00_ffff), (i8 8)),
1685 (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1687 // h-register tricks.
1688 // For now, be conservative on x86-64 and use an h-register extract only if the
1689 // value is immediately zero-extended or stored, which are somewhat common
1690 // cases. This uses a bunch of code to prevent a register requiring a REX prefix
1691 // from being allocated in the same instruction as the h register, as there's
1692 // currently no way to describe this requirement to the register allocator.
1694 // h-register extract and zero-extend.
1695 def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1699 (EXTRACT_SUBREG GR64:$src, sub_8bit_hi)),
1701 def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1705 (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1707 def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1711 (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1714 // h-register extract and store.
1715 def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1718 (EXTRACT_SUBREG GR64:$src, sub_8bit_hi))>;
1719 def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1722 (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>,
1723 Requires<[In64BitMode]>;
1724 def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1727 (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>,
1728 Requires<[In64BitMode]>;
1731 // (shl x, 1) ==> (add x, x)
1732 // Note that if x is undef (immediate or otherwise), we could theoretically
1733 // end up with the two uses of x getting different values, producing a result
1734 // where the least significant bit is not 0. However, the probability of this
1735 // happening is considered low enough that this is officially not a
1737 def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr GR8 :$src1, GR8 :$src1)>;
1738 def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1739 def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1740 def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1742 def shiftMask8 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1743 return isUnneededShiftMask(N, 3);
1746 def shiftMask16 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1747 return isUnneededShiftMask(N, 4);
1750 def shiftMask32 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1751 return isUnneededShiftMask(N, 5);
1754 def shiftMask64 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1755 return isUnneededShiftMask(N, 6);
1759 // Shift amount is implicitly masked.
1760 multiclass MaskedShiftAmountPats<SDNode frag, string name> {
1761 // (shift x (and y, 31)) ==> (shift x, y)
1762 def : Pat<(frag GR8:$src1, (shiftMask32 CL)),
1763 (!cast<Instruction>(name # "8rCL") GR8:$src1)>;
1764 def : Pat<(frag GR16:$src1, (shiftMask32 CL)),
1765 (!cast<Instruction>(name # "16rCL") GR16:$src1)>;
1766 def : Pat<(frag GR32:$src1, (shiftMask32 CL)),
1767 (!cast<Instruction>(name # "32rCL") GR32:$src1)>;
1768 def : Pat<(store (frag (loadi8 addr:$dst), (shiftMask32 CL)), addr:$dst),
1769 (!cast<Instruction>(name # "8mCL") addr:$dst)>;
1770 def : Pat<(store (frag (loadi16 addr:$dst), (shiftMask32 CL)), addr:$dst),
1771 (!cast<Instruction>(name # "16mCL") addr:$dst)>;
1772 def : Pat<(store (frag (loadi32 addr:$dst), (shiftMask32 CL)), addr:$dst),
1773 (!cast<Instruction>(name # "32mCL") addr:$dst)>;
1775 // (shift x (and y, 63)) ==> (shift x, y)
1776 def : Pat<(frag GR64:$src1, (shiftMask64 CL)),
1777 (!cast<Instruction>(name # "64rCL") GR64:$src1)>;
1778 def : Pat<(store (frag (loadi64 addr:$dst), (shiftMask64 CL)), addr:$dst),
1779 (!cast<Instruction>(name # "64mCL") addr:$dst)>;
1782 defm : MaskedShiftAmountPats<shl, "SHL">;
1783 defm : MaskedShiftAmountPats<srl, "SHR">;
1784 defm : MaskedShiftAmountPats<sra, "SAR">;
1786 // ROL/ROR instructions allow a stronger mask optimization than shift for 8- and
1787 // 16-bit. We can remove a mask of any (bitwidth - 1) on the rotation amount
1788 // because over-rotating produces the same result. This is noted in the Intel
1789 // docs with: "tempCOUNT <- (COUNT & COUNTMASK) MOD SIZE". Masking the rotation
1790 // amount could affect EFLAGS results, but that does not matter because we are
1791 // not tracking flags for these nodes.
1792 multiclass MaskedRotateAmountPats<SDNode frag, string name> {
1793 // (rot x (and y, BitWidth - 1)) ==> (rot x, y)
1794 def : Pat<(frag GR8:$src1, (shiftMask8 CL)),
1795 (!cast<Instruction>(name # "8rCL") GR8:$src1)>;
1796 def : Pat<(frag GR16:$src1, (shiftMask16 CL)),
1797 (!cast<Instruction>(name # "16rCL") GR16:$src1)>;
1798 def : Pat<(frag GR32:$src1, (shiftMask32 CL)),
1799 (!cast<Instruction>(name # "32rCL") GR32:$src1)>;
1800 def : Pat<(store (frag (loadi8 addr:$dst), (shiftMask8 CL)), addr:$dst),
1801 (!cast<Instruction>(name # "8mCL") addr:$dst)>;
1802 def : Pat<(store (frag (loadi16 addr:$dst), (shiftMask16 CL)), addr:$dst),
1803 (!cast<Instruction>(name # "16mCL") addr:$dst)>;
1804 def : Pat<(store (frag (loadi32 addr:$dst), (shiftMask32 CL)), addr:$dst),
1805 (!cast<Instruction>(name # "32mCL") addr:$dst)>;
1807 // (rot x (and y, 63)) ==> (rot x, y)
1808 def : Pat<(frag GR64:$src1, (shiftMask64 CL)),
1809 (!cast<Instruction>(name # "64rCL") GR64:$src1)>;
1810 def : Pat<(store (frag (loadi64 addr:$dst), (shiftMask64 CL)), addr:$dst),
1811 (!cast<Instruction>(name # "64mCL") addr:$dst)>;
1815 defm : MaskedRotateAmountPats<rotl, "ROL">;
1816 defm : MaskedRotateAmountPats<rotr, "ROR">;
1818 // Double shift amount is implicitly masked.
1819 multiclass MaskedDoubleShiftAmountPats<SDNode frag, string name> {
1820 // (shift x (and y, 31)) ==> (shift x, y)
1821 def : Pat<(frag GR16:$src1, GR16:$src2, (shiftMask32 CL)),
1822 (!cast<Instruction>(name # "16rrCL") GR16:$src1, GR16:$src2)>;
1823 def : Pat<(frag GR32:$src1, GR32:$src2, (shiftMask32 CL)),
1824 (!cast<Instruction>(name # "32rrCL") GR32:$src1, GR32:$src2)>;
1826 // (shift x (and y, 63)) ==> (shift x, y)
1827 def : Pat<(frag GR64:$src1, GR64:$src2, (shiftMask32 CL)),
1828 (!cast<Instruction>(name # "64rrCL") GR64:$src1, GR64:$src2)>;
1831 defm : MaskedDoubleShiftAmountPats<X86shld, "SHLD">;
1832 defm : MaskedDoubleShiftAmountPats<X86shrd, "SHRD">;
1834 let Predicates = [HasBMI2] in {
1835 let AddedComplexity = 1 in {
1836 def : Pat<(sra GR32:$src1, (shiftMask32 GR8:$src2)),
1837 (SARX32rr GR32:$src1,
1839 (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1840 def : Pat<(sra GR64:$src1, (shiftMask64 GR8:$src2)),
1841 (SARX64rr GR64:$src1,
1843 (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1845 def : Pat<(srl GR32:$src1, (shiftMask32 GR8:$src2)),
1846 (SHRX32rr GR32:$src1,
1848 (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1849 def : Pat<(srl GR64:$src1, (shiftMask64 GR8:$src2)),
1850 (SHRX64rr GR64:$src1,
1852 (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1854 def : Pat<(shl GR32:$src1, (shiftMask32 GR8:$src2)),
1855 (SHLX32rr GR32:$src1,
1857 (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1858 def : Pat<(shl GR64:$src1, (shiftMask64 GR8:$src2)),
1859 (SHLX64rr GR64:$src1,
1861 (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1864 def : Pat<(sra (loadi32 addr:$src1), (shiftMask32 GR8:$src2)),
1865 (SARX32rm addr:$src1,
1867 (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1868 def : Pat<(sra (loadi64 addr:$src1), (shiftMask64 GR8:$src2)),
1869 (SARX64rm addr:$src1,
1871 (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1873 def : Pat<(srl (loadi32 addr:$src1), (shiftMask32 GR8:$src2)),
1874 (SHRX32rm addr:$src1,
1876 (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1877 def : Pat<(srl (loadi64 addr:$src1), (shiftMask64 GR8:$src2)),
1878 (SHRX64rm addr:$src1,
1880 (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1882 def : Pat<(shl (loadi32 addr:$src1), (shiftMask32 GR8:$src2)),
1883 (SHLX32rm addr:$src1,
1885 (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1886 def : Pat<(shl (loadi64 addr:$src1), (shiftMask64 GR8:$src2)),
1887 (SHLX64rm addr:$src1,
1889 (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1892 // Use BTR/BTS/BTC for clearing/setting/toggling a bit in a variable location.
1893 multiclass one_bit_patterns<RegisterClass RC, ValueType VT, Instruction BTR,
1894 Instruction BTS, Instruction BTC,
1895 PatFrag ShiftMask> {
1896 def : Pat<(and RC:$src1, (rotl -2, GR8:$src2)),
1898 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1899 def : Pat<(or RC:$src1, (shl 1, GR8:$src2)),
1901 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1902 def : Pat<(xor RC:$src1, (shl 1, GR8:$src2)),
1904 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1906 // Similar to above, but removing unneeded masking of the shift amount.
1907 def : Pat<(and RC:$src1, (rotl -2, (ShiftMask GR8:$src2))),
1909 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1910 def : Pat<(or RC:$src1, (shl 1, (ShiftMask GR8:$src2))),
1912 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1913 def : Pat<(xor RC:$src1, (shl 1, (ShiftMask GR8:$src2))),
1915 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1918 defm : one_bit_patterns<GR16, i16, BTR16rr, BTS16rr, BTC16rr, shiftMask16>;
1919 defm : one_bit_patterns<GR32, i32, BTR32rr, BTS32rr, BTC32rr, shiftMask32>;
1920 defm : one_bit_patterns<GR64, i64, BTR64rr, BTS64rr, BTC64rr, shiftMask64>;
1923 // (anyext (setcc_carry)) -> (setcc_carry)
1924 def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1926 def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1928 def : Pat<(i32 (anyext (i16 (X86setcc_c X86_COND_B, EFLAGS)))),
1931 //===----------------------------------------------------------------------===//
1932 // EFLAGS-defining Patterns
1933 //===----------------------------------------------------------------------===//
1936 def : Pat<(add GR8 :$src1, GR8 :$src2), (ADD8rr GR8 :$src1, GR8 :$src2)>;
1937 def : Pat<(add GR16:$src1, GR16:$src2), (ADD16rr GR16:$src1, GR16:$src2)>;
1938 def : Pat<(add GR32:$src1, GR32:$src2), (ADD32rr GR32:$src1, GR32:$src2)>;
1939 def : Pat<(add GR64:$src1, GR64:$src2), (ADD64rr GR64:$src1, GR64:$src2)>;
1942 def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
1943 (ADD8rm GR8:$src1, addr:$src2)>;
1944 def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
1945 (ADD16rm GR16:$src1, addr:$src2)>;
1946 def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
1947 (ADD32rm GR32:$src1, addr:$src2)>;
1948 def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
1949 (ADD64rm GR64:$src1, addr:$src2)>;
1952 def : Pat<(add GR8 :$src1, imm:$src2), (ADD8ri GR8:$src1 , imm:$src2)>;
1953 def : Pat<(add GR16:$src1, imm:$src2), (ADD16ri GR16:$src1, imm:$src2)>;
1954 def : Pat<(add GR32:$src1, imm:$src2), (ADD32ri GR32:$src1, imm:$src2)>;
1955 def : Pat<(add GR16:$src1, i16immSExt8:$src2),
1956 (ADD16ri8 GR16:$src1, i16immSExt8:$src2)>;
1957 def : Pat<(add GR32:$src1, i32immSExt8:$src2),
1958 (ADD32ri8 GR32:$src1, i32immSExt8:$src2)>;
1959 def : Pat<(add GR64:$src1, i64immSExt8:$src2),
1960 (ADD64ri8 GR64:$src1, i64immSExt8:$src2)>;
1961 def : Pat<(add GR64:$src1, i64immSExt32:$src2),
1962 (ADD64ri32 GR64:$src1, i64immSExt32:$src2)>;
1965 def : Pat<(sub GR8 :$src1, GR8 :$src2), (SUB8rr GR8 :$src1, GR8 :$src2)>;
1966 def : Pat<(sub GR16:$src1, GR16:$src2), (SUB16rr GR16:$src1, GR16:$src2)>;
1967 def : Pat<(sub GR32:$src1, GR32:$src2), (SUB32rr GR32:$src1, GR32:$src2)>;
1968 def : Pat<(sub GR64:$src1, GR64:$src2), (SUB64rr GR64:$src1, GR64:$src2)>;
1971 def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
1972 (SUB8rm GR8:$src1, addr:$src2)>;
1973 def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
1974 (SUB16rm GR16:$src1, addr:$src2)>;
1975 def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
1976 (SUB32rm GR32:$src1, addr:$src2)>;
1977 def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
1978 (SUB64rm GR64:$src1, addr:$src2)>;
1981 def : Pat<(sub GR8:$src1, imm:$src2),
1982 (SUB8ri GR8:$src1, imm:$src2)>;
1983 def : Pat<(sub GR16:$src1, imm:$src2),
1984 (SUB16ri GR16:$src1, imm:$src2)>;
1985 def : Pat<(sub GR32:$src1, imm:$src2),
1986 (SUB32ri GR32:$src1, imm:$src2)>;
1987 def : Pat<(sub GR16:$src1, i16immSExt8:$src2),
1988 (SUB16ri8 GR16:$src1, i16immSExt8:$src2)>;
1989 def : Pat<(sub GR32:$src1, i32immSExt8:$src2),
1990 (SUB32ri8 GR32:$src1, i32immSExt8:$src2)>;
1991 def : Pat<(sub GR64:$src1, i64immSExt8:$src2),
1992 (SUB64ri8 GR64:$src1, i64immSExt8:$src2)>;
1993 def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
1994 (SUB64ri32 GR64:$src1, i64immSExt32:$src2)>;
1997 def : Pat<(X86sub_flag 0, GR8 :$src), (NEG8r GR8 :$src)>;
1998 def : Pat<(X86sub_flag 0, GR16:$src), (NEG16r GR16:$src)>;
1999 def : Pat<(X86sub_flag 0, GR32:$src), (NEG32r GR32:$src)>;
2000 def : Pat<(X86sub_flag 0, GR64:$src), (NEG64r GR64:$src)>;
2002 // sub reg, relocImm
2003 def : Pat<(X86sub_flag GR64:$src1, i64relocImmSExt8_su:$src2),
2004 (SUB64ri8 GR64:$src1, i64relocImmSExt8_su:$src2)>;
2007 def : Pat<(mul GR16:$src1, GR16:$src2),
2008 (IMUL16rr GR16:$src1, GR16:$src2)>;
2009 def : Pat<(mul GR32:$src1, GR32:$src2),
2010 (IMUL32rr GR32:$src1, GR32:$src2)>;
2011 def : Pat<(mul GR64:$src1, GR64:$src2),
2012 (IMUL64rr GR64:$src1, GR64:$src2)>;
2015 def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
2016 (IMUL16rm GR16:$src1, addr:$src2)>;
2017 def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
2018 (IMUL32rm GR32:$src1, addr:$src2)>;
2019 def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
2020 (IMUL64rm GR64:$src1, addr:$src2)>;
2023 def : Pat<(mul GR16:$src1, imm:$src2),
2024 (IMUL16rri GR16:$src1, imm:$src2)>;
2025 def : Pat<(mul GR32:$src1, imm:$src2),
2026 (IMUL32rri GR32:$src1, imm:$src2)>;
2027 def : Pat<(mul GR16:$src1, i16immSExt8:$src2),
2028 (IMUL16rri8 GR16:$src1, i16immSExt8:$src2)>;
2029 def : Pat<(mul GR32:$src1, i32immSExt8:$src2),
2030 (IMUL32rri8 GR32:$src1, i32immSExt8:$src2)>;
2031 def : Pat<(mul GR64:$src1, i64immSExt8:$src2),
2032 (IMUL64rri8 GR64:$src1, i64immSExt8:$src2)>;
2033 def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
2034 (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
2036 // reg = mul mem, imm
2037 def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
2038 (IMUL16rmi addr:$src1, imm:$src2)>;
2039 def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
2040 (IMUL32rmi addr:$src1, imm:$src2)>;
2041 def : Pat<(mul (loadi16 addr:$src1), i16immSExt8:$src2),
2042 (IMUL16rmi8 addr:$src1, i16immSExt8:$src2)>;
2043 def : Pat<(mul (loadi32 addr:$src1), i32immSExt8:$src2),
2044 (IMUL32rmi8 addr:$src1, i32immSExt8:$src2)>;
2045 def : Pat<(mul (loadi64 addr:$src1), i64immSExt8:$src2),
2046 (IMUL64rmi8 addr:$src1, i64immSExt8:$src2)>;
2047 def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
2048 (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
2050 // Increment/Decrement reg.
2051 // Do not make INC/DEC if it is slow
2052 let Predicates = [UseIncDec] in {
2053 def : Pat<(add GR8:$src, 1), (INC8r GR8:$src)>;
2054 def : Pat<(add GR16:$src, 1), (INC16r GR16:$src)>;
2055 def : Pat<(add GR32:$src, 1), (INC32r GR32:$src)>;
2056 def : Pat<(add GR64:$src, 1), (INC64r GR64:$src)>;
2057 def : Pat<(add GR8:$src, -1), (DEC8r GR8:$src)>;
2058 def : Pat<(add GR16:$src, -1), (DEC16r GR16:$src)>;
2059 def : Pat<(add GR32:$src, -1), (DEC32r GR32:$src)>;
2060 def : Pat<(add GR64:$src, -1), (DEC64r GR64:$src)>;
2062 def : Pat<(X86add_flag_nocf GR8:$src, -1), (DEC8r GR8:$src)>;
2063 def : Pat<(X86add_flag_nocf GR16:$src, -1), (DEC16r GR16:$src)>;
2064 def : Pat<(X86add_flag_nocf GR32:$src, -1), (DEC32r GR32:$src)>;
2065 def : Pat<(X86add_flag_nocf GR64:$src, -1), (DEC64r GR64:$src)>;
2066 def : Pat<(X86sub_flag_nocf GR8:$src, -1), (INC8r GR8:$src)>;
2067 def : Pat<(X86sub_flag_nocf GR16:$src, -1), (INC16r GR16:$src)>;
2068 def : Pat<(X86sub_flag_nocf GR32:$src, -1), (INC32r GR32:$src)>;
2069 def : Pat<(X86sub_flag_nocf GR64:$src, -1), (INC64r GR64:$src)>;
2073 def : Pat<(or GR8 :$src1, GR8 :$src2), (OR8rr GR8 :$src1, GR8 :$src2)>;
2074 def : Pat<(or GR16:$src1, GR16:$src2), (OR16rr GR16:$src1, GR16:$src2)>;
2075 def : Pat<(or GR32:$src1, GR32:$src2), (OR32rr GR32:$src1, GR32:$src2)>;
2076 def : Pat<(or GR64:$src1, GR64:$src2), (OR64rr GR64:$src1, GR64:$src2)>;
2079 def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
2080 (OR8rm GR8:$src1, addr:$src2)>;
2081 def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
2082 (OR16rm GR16:$src1, addr:$src2)>;
2083 def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
2084 (OR32rm GR32:$src1, addr:$src2)>;
2085 def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
2086 (OR64rm GR64:$src1, addr:$src2)>;
2089 def : Pat<(or GR8:$src1 , imm:$src2), (OR8ri GR8 :$src1, imm:$src2)>;
2090 def : Pat<(or GR16:$src1, imm:$src2), (OR16ri GR16:$src1, imm:$src2)>;
2091 def : Pat<(or GR32:$src1, imm:$src2), (OR32ri GR32:$src1, imm:$src2)>;
2092 def : Pat<(or GR16:$src1, i16immSExt8:$src2),
2093 (OR16ri8 GR16:$src1, i16immSExt8:$src2)>;
2094 def : Pat<(or GR32:$src1, i32immSExt8:$src2),
2095 (OR32ri8 GR32:$src1, i32immSExt8:$src2)>;
2096 def : Pat<(or GR64:$src1, i64immSExt8:$src2),
2097 (OR64ri8 GR64:$src1, i64immSExt8:$src2)>;
2098 def : Pat<(or GR64:$src1, i64immSExt32:$src2),
2099 (OR64ri32 GR64:$src1, i64immSExt32:$src2)>;
2102 def : Pat<(xor GR8 :$src1, GR8 :$src2), (XOR8rr GR8 :$src1, GR8 :$src2)>;
2103 def : Pat<(xor GR16:$src1, GR16:$src2), (XOR16rr GR16:$src1, GR16:$src2)>;
2104 def : Pat<(xor GR32:$src1, GR32:$src2), (XOR32rr GR32:$src1, GR32:$src2)>;
2105 def : Pat<(xor GR64:$src1, GR64:$src2), (XOR64rr GR64:$src1, GR64:$src2)>;
2108 def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
2109 (XOR8rm GR8:$src1, addr:$src2)>;
2110 def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
2111 (XOR16rm GR16:$src1, addr:$src2)>;
2112 def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
2113 (XOR32rm GR32:$src1, addr:$src2)>;
2114 def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
2115 (XOR64rm GR64:$src1, addr:$src2)>;
2118 def : Pat<(xor GR8:$src1, imm:$src2),
2119 (XOR8ri GR8:$src1, imm:$src2)>;
2120 def : Pat<(xor GR16:$src1, imm:$src2),
2121 (XOR16ri GR16:$src1, imm:$src2)>;
2122 def : Pat<(xor GR32:$src1, imm:$src2),
2123 (XOR32ri GR32:$src1, imm:$src2)>;
2124 def : Pat<(xor GR16:$src1, i16immSExt8:$src2),
2125 (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
2126 def : Pat<(xor GR32:$src1, i32immSExt8:$src2),
2127 (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
2128 def : Pat<(xor GR64:$src1, i64immSExt8:$src2),
2129 (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
2130 def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
2131 (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
2134 def : Pat<(and GR8 :$src1, GR8 :$src2), (AND8rr GR8 :$src1, GR8 :$src2)>;
2135 def : Pat<(and GR16:$src1, GR16:$src2), (AND16rr GR16:$src1, GR16:$src2)>;
2136 def : Pat<(and GR32:$src1, GR32:$src2), (AND32rr GR32:$src1, GR32:$src2)>;
2137 def : Pat<(and GR64:$src1, GR64:$src2), (AND64rr GR64:$src1, GR64:$src2)>;
2140 def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
2141 (AND8rm GR8:$src1, addr:$src2)>;
2142 def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
2143 (AND16rm GR16:$src1, addr:$src2)>;
2144 def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
2145 (AND32rm GR32:$src1, addr:$src2)>;
2146 def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
2147 (AND64rm GR64:$src1, addr:$src2)>;
2150 def : Pat<(and GR8:$src1, imm:$src2),
2151 (AND8ri GR8:$src1, imm:$src2)>;
2152 def : Pat<(and GR16:$src1, imm:$src2),
2153 (AND16ri GR16:$src1, imm:$src2)>;
2154 def : Pat<(and GR32:$src1, imm:$src2),
2155 (AND32ri GR32:$src1, imm:$src2)>;
2156 def : Pat<(and GR16:$src1, i16immSExt8:$src2),
2157 (AND16ri8 GR16:$src1, i16immSExt8:$src2)>;
2158 def : Pat<(and GR32:$src1, i32immSExt8:$src2),
2159 (AND32ri8 GR32:$src1, i32immSExt8:$src2)>;
2160 def : Pat<(and GR64:$src1, i64immSExt8:$src2),
2161 (AND64ri8 GR64:$src1, i64immSExt8:$src2)>;
2162 def : Pat<(and GR64:$src1, i64immSExt32:$src2),
2163 (AND64ri32 GR64:$src1, i64immSExt32:$src2)>;
2165 // Bit scan instruction patterns to match explicit zero-undef behavior.
2166 def : Pat<(cttz_zero_undef GR16:$src), (BSF16rr GR16:$src)>;
2167 def : Pat<(cttz_zero_undef GR32:$src), (BSF32rr GR32:$src)>;
2168 def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr GR64:$src)>;
2169 def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm addr:$src)>;
2170 def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm addr:$src)>;
2171 def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm addr:$src)>;
2173 // When HasMOVBE is enabled it is possible to get a non-legalized
2174 // register-register 16 bit bswap. This maps it to a ROL instruction.
2175 let Predicates = [HasMOVBE] in {
2176 def : Pat<(bswap GR16:$src), (ROL16ri GR16:$src, (i8 8))>;