1 //===- X86InstrCompiler.td - Compiler Pseudos and Patterns -*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the various pseudo instructions used by the compiler,
11 // as well as Pat patterns used during instruction selection.
13 //===----------------------------------------------------------------------===//
15 //===----------------------------------------------------------------------===//
16 // Pattern Matching Support
18 def GetLo32XForm : SDNodeXForm<imm, [{
19 // Transformation function: get the low 32 bits.
20 return getI32Imm((unsigned)N->getZExtValue(), SDLoc(N));
23 def GetLo8XForm : SDNodeXForm<imm, [{
24 // Transformation function: get the low 8 bits.
25 return getI8Imm((uint8_t)N->getZExtValue(), SDLoc(N));
29 //===----------------------------------------------------------------------===//
30 // Random Pseudo Instructions.
32 // PIC base construction. This expands to code that looks like this:
35 let hasSideEffects = 0, isNotDuplicable = 1, Uses = [ESP] in
36 def MOVPC32r : Ii32<0xE8, Pseudo, (outs GR32:$reg), (ins i32imm:$label),
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.
45 let Defs = [ESP, EFLAGS], Uses = [ESP] in {
46 def ADJCALLSTACKDOWN32 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
50 def ADJCALLSTACKUP32 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
52 [(X86callseq_end timm:$amt1, timm:$amt2)]>,
55 def : Pat<(X86callseq_start timm:$amt1),
56 (ADJCALLSTACKDOWN32 i32imm:$amt1, 0)>, Requires<[NotLP64]>;
59 // ADJCALLSTACKDOWN/UP implicitly use/def RSP because they may be expanded into
60 // a stack adjustment and the codegen must know that they may modify the stack
61 // pointer before prolog-epilog rewriting occurs.
62 // Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
63 // sub / add which can clobber EFLAGS.
64 let Defs = [RSP, EFLAGS], Uses = [RSP] in {
65 def ADJCALLSTACKDOWN64 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
69 def ADJCALLSTACKUP64 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
71 [(X86callseq_end timm:$amt1, timm:$amt2)]>,
74 def : Pat<(X86callseq_start timm:$amt1),
75 (ADJCALLSTACKDOWN64 i32imm:$amt1, 0)>, Requires<[IsLP64]>;
78 // x86-64 va_start lowering magic.
79 let usesCustomInserter = 1, Defs = [EFLAGS] in {
80 def VASTART_SAVE_XMM_REGS : I<0, Pseudo,
83 i64imm:$regsavefi, i64imm:$offset,
85 "#VASTART_SAVE_XMM_REGS $al, $regsavefi, $offset",
86 [(X86vastart_save_xmm_regs GR8:$al,
91 // The VAARG_64 pseudo-instruction takes the address of the va_list,
92 // and places the address of the next argument into a register.
93 let Defs = [EFLAGS] in
94 def VAARG_64 : I<0, Pseudo,
96 (ins i8mem:$ap, i32imm:$size, i8imm:$mode, i32imm:$align),
97 "#VAARG_64 $dst, $ap, $size, $mode, $align",
99 (X86vaarg64 addr:$ap, imm:$size, imm:$mode, imm:$align)),
102 // Dynamic stack allocation yields a _chkstk or _alloca call for all Windows
103 // targets. These calls are needed to probe the stack when allocating more than
104 // 4k bytes in one go. Touching the stack at 4K increments is necessary to
105 // ensure that the guard pages used by the OS virtual memory manager are
106 // allocated in correct sequence.
107 // The main point of having separate instruction are extra unmodelled effects
108 // (compared to ordinary calls) like stack pointer change.
110 let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
111 def WIN_ALLOCA : I<0, Pseudo, (outs), (ins),
112 "# dynamic stack allocation",
115 // When using segmented stacks these are lowered into instructions which first
116 // check if the current stacklet has enough free memory. If it does, memory is
117 // allocated by bumping the stack pointer. Otherwise memory is allocated from
120 let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
121 def SEG_ALLOCA_32 : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$size),
122 "# variable sized alloca for segmented stacks",
124 (X86SegAlloca GR32:$size))]>,
127 let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
128 def SEG_ALLOCA_64 : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$size),
129 "# variable sized alloca for segmented stacks",
131 (X86SegAlloca GR64:$size))]>,
132 Requires<[In64BitMode]>;
135 // The MSVC runtime contains an _ftol2 routine for converting floating-point
136 // to integer values. It has a strange calling convention: the input is
137 // popped from the x87 stack, and the return value is given in EDX:EAX. ECX is
138 // used as a temporary register. No other registers (aside from flags) are
140 // Microsoft toolchains do not support 80-bit precision, so a WIN_FTOL_80
141 // variant is unnecessary.
143 let Defs = [EAX, EDX, ECX, EFLAGS], FPForm = SpecialFP in {
144 def WIN_FTOL_32 : I<0, Pseudo, (outs), (ins RFP32:$src),
146 [(X86WinFTOL RFP32:$src)]>,
147 Requires<[Not64BitMode]>;
149 def WIN_FTOL_64 : I<0, Pseudo, (outs), (ins RFP64:$src),
151 [(X86WinFTOL RFP64:$src)]>,
152 Requires<[Not64BitMode]>;
155 //===----------------------------------------------------------------------===//
156 // EH Pseudo Instructions
158 let SchedRW = [WriteSystem] in {
159 let isTerminator = 1, isReturn = 1, isBarrier = 1,
160 hasCtrlDep = 1, isCodeGenOnly = 1 in {
161 def EH_RETURN : I<0xC3, RawFrm, (outs), (ins GR32:$addr),
162 "ret\t#eh_return, addr: $addr",
163 [(X86ehret GR32:$addr)], IIC_RET>, Sched<[WriteJumpLd]>;
167 let isTerminator = 1, isReturn = 1, isBarrier = 1,
168 hasCtrlDep = 1, isCodeGenOnly = 1 in {
169 def EH_RETURN64 : I<0xC3, RawFrm, (outs), (ins GR64:$addr),
170 "ret\t#eh_return, addr: $addr",
171 [(X86ehret GR64:$addr)], IIC_RET>, Sched<[WriteJumpLd]>;
175 let hasSideEffects = 1, isBarrier = 1, isCodeGenOnly = 1,
176 usesCustomInserter = 1 in {
177 def EH_SjLj_SetJmp32 : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$buf),
179 [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
180 Requires<[Not64BitMode]>;
181 def EH_SjLj_SetJmp64 : I<0, Pseudo, (outs GR32:$dst), (ins i64mem:$buf),
183 [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
184 Requires<[In64BitMode]>;
185 let isTerminator = 1 in {
186 def EH_SjLj_LongJmp32 : I<0, Pseudo, (outs), (ins i32mem:$buf),
187 "#EH_SJLJ_LONGJMP32",
188 [(X86eh_sjlj_longjmp addr:$buf)]>,
189 Requires<[Not64BitMode]>;
190 def EH_SjLj_LongJmp64 : I<0, Pseudo, (outs), (ins i64mem:$buf),
191 "#EH_SJLJ_LONGJMP64",
192 [(X86eh_sjlj_longjmp addr:$buf)]>,
193 Requires<[In64BitMode]>;
198 let isBranch = 1, isTerminator = 1, isCodeGenOnly = 1 in {
199 def EH_SjLj_Setup : I<0, Pseudo, (outs), (ins brtarget:$dst),
200 "#EH_SjLj_Setup\t$dst", []>;
203 //===----------------------------------------------------------------------===//
204 // Pseudo instructions used by unwind info.
206 let isPseudo = 1 in {
207 def SEH_PushReg : I<0, Pseudo, (outs), (ins i32imm:$reg),
208 "#SEH_PushReg $reg", []>;
209 def SEH_SaveReg : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
210 "#SEH_SaveReg $reg, $dst", []>;
211 def SEH_SaveXMM : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
212 "#SEH_SaveXMM $reg, $dst", []>;
213 def SEH_StackAlloc : I<0, Pseudo, (outs), (ins i32imm:$size),
214 "#SEH_StackAlloc $size", []>;
215 def SEH_SetFrame : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$offset),
216 "#SEH_SetFrame $reg, $offset", []>;
217 def SEH_PushFrame : I<0, Pseudo, (outs), (ins i1imm:$mode),
218 "#SEH_PushFrame $mode", []>;
219 def SEH_EndPrologue : I<0, Pseudo, (outs), (ins),
220 "#SEH_EndPrologue", []>;
221 def SEH_Epilogue : I<0, Pseudo, (outs), (ins),
222 "#SEH_Epilogue", []>;
225 //===----------------------------------------------------------------------===//
226 // Pseudo instructions used by segmented stacks.
229 // This is lowered into a RET instruction by MCInstLower. We need
230 // this so that we don't have to have a MachineBasicBlock which ends
231 // with a RET and also has successors.
232 let isPseudo = 1 in {
233 def MORESTACK_RET: I<0, Pseudo, (outs), (ins),
236 // This instruction is lowered to a RET followed by a MOV. The two
237 // instructions are not generated on a higher level since then the
238 // verifier sees a MachineBasicBlock ending with a non-terminator.
239 def MORESTACK_RET_RESTORE_R10 : I<0, Pseudo, (outs), (ins),
243 //===----------------------------------------------------------------------===//
244 // Alias Instructions
245 //===----------------------------------------------------------------------===//
247 // Alias instruction mapping movr0 to xor.
248 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
249 let Defs = [EFLAGS], isReMaterializable = 1, isAsCheapAsAMove = 1,
251 def MOV32r0 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
252 [(set GR32:$dst, 0)], IIC_ALU_NONMEM>, Sched<[WriteZero]>;
254 // Other widths can also make use of the 32-bit xor, which may have a smaller
255 // encoding and avoid partial register updates.
256 def : Pat<(i8 0), (EXTRACT_SUBREG (MOV32r0), sub_8bit)>;
257 def : Pat<(i16 0), (EXTRACT_SUBREG (MOV32r0), sub_16bit)>;
258 def : Pat<(i64 0), (SUBREG_TO_REG (i64 0), (MOV32r0), sub_32bit)> {
259 let AddedComplexity = 20;
262 // Materialize i64 constant where top 32-bits are zero. This could theoretically
263 // use MOV32ri with a SUBREG_TO_REG to represent the zero-extension, however
264 // that would make it more difficult to rematerialize.
265 let AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1,
266 isCodeGenOnly = 1, hasSideEffects = 0 in
267 def MOV32ri64 : Ii32<0xb8, AddRegFrm, (outs GR32:$dst), (ins i64i32imm:$src),
268 "", [], IIC_ALU_NONMEM>, Sched<[WriteALU]>;
270 // This 64-bit pseudo-move can be used for both a 64-bit constant that is
271 // actually the zero-extension of a 32-bit constant, and for labels in the
272 // x86-64 small code model.
273 def mov64imm32 : ComplexPattern<i64, 1, "SelectMOV64Imm32", [imm, X86Wrapper]>;
275 let AddedComplexity = 1 in
276 def : Pat<(i64 mov64imm32:$src),
277 (SUBREG_TO_REG (i64 0), (MOV32ri64 mov64imm32:$src), sub_32bit)>;
279 // Use sbb to materialize carry bit.
280 let Uses = [EFLAGS], Defs = [EFLAGS], isPseudo = 1, SchedRW = [WriteALU] in {
281 // FIXME: These are pseudo ops that should be replaced with Pat<> patterns.
282 // However, Pat<> can't replicate the destination reg into the inputs of the
284 def SETB_C8r : I<0, Pseudo, (outs GR8:$dst), (ins), "",
285 [(set GR8:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
286 def SETB_C16r : I<0, Pseudo, (outs GR16:$dst), (ins), "",
287 [(set GR16:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
288 def SETB_C32r : I<0, Pseudo, (outs GR32:$dst), (ins), "",
289 [(set GR32:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
290 def SETB_C64r : I<0, Pseudo, (outs GR64:$dst), (ins), "",
291 [(set GR64:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
295 def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
297 def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
299 def : Pat<(i64 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
302 def : Pat<(i16 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
304 def : Pat<(i32 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
306 def : Pat<(i64 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
309 // We canonicalize 'setb' to "(and (sbb reg,reg), 1)" on the hope that the and
310 // will be eliminated and that the sbb can be extended up to a wider type. When
311 // this happens, it is great. However, if we are left with an 8-bit sbb and an
312 // and, we might as well just match it as a setb.
313 def : Pat<(and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1),
316 // (add OP, SETB) -> (adc OP, 0)
317 def : Pat<(add (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR8:$op),
318 (ADC8ri GR8:$op, 0)>;
319 def : Pat<(add (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR32:$op),
320 (ADC32ri8 GR32:$op, 0)>;
321 def : Pat<(add (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR64:$op),
322 (ADC64ri8 GR64:$op, 0)>;
324 // (sub OP, SETB) -> (sbb OP, 0)
325 def : Pat<(sub GR8:$op, (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
326 (SBB8ri GR8:$op, 0)>;
327 def : Pat<(sub GR32:$op, (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
328 (SBB32ri8 GR32:$op, 0)>;
329 def : Pat<(sub GR64:$op, (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
330 (SBB64ri8 GR64:$op, 0)>;
332 // (sub OP, SETCC_CARRY) -> (adc OP, 0)
333 def : Pat<(sub GR8:$op, (i8 (X86setcc_c X86_COND_B, EFLAGS))),
334 (ADC8ri GR8:$op, 0)>;
335 def : Pat<(sub GR32:$op, (i32 (X86setcc_c X86_COND_B, EFLAGS))),
336 (ADC32ri8 GR32:$op, 0)>;
337 def : Pat<(sub GR64:$op, (i64 (X86setcc_c X86_COND_B, EFLAGS))),
338 (ADC64ri8 GR64:$op, 0)>;
340 //===----------------------------------------------------------------------===//
341 // String Pseudo Instructions
343 let SchedRW = [WriteMicrocoded] in {
344 let Defs = [ECX,EDI,ESI], Uses = [ECX,EDI,ESI], isCodeGenOnly = 1 in {
345 def REP_MOVSB_32 : I<0xA4, RawFrm, (outs), (ins), "{rep;movsb|rep movsb}",
346 [(X86rep_movs i8)], IIC_REP_MOVS>, REP,
347 Requires<[Not64BitMode]>;
348 def REP_MOVSW_32 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsw|rep movsw}",
349 [(X86rep_movs i16)], IIC_REP_MOVS>, REP, OpSize16,
350 Requires<[Not64BitMode]>;
351 def REP_MOVSD_32 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsl|rep movsd}",
352 [(X86rep_movs i32)], IIC_REP_MOVS>, REP, OpSize32,
353 Requires<[Not64BitMode]>;
356 let Defs = [RCX,RDI,RSI], Uses = [RCX,RDI,RSI], isCodeGenOnly = 1 in {
357 def REP_MOVSB_64 : I<0xA4, RawFrm, (outs), (ins), "{rep;movsb|rep movsb}",
358 [(X86rep_movs i8)], IIC_REP_MOVS>, REP,
359 Requires<[In64BitMode]>;
360 def REP_MOVSW_64 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsw|rep movsw}",
361 [(X86rep_movs i16)], IIC_REP_MOVS>, REP, OpSize16,
362 Requires<[In64BitMode]>;
363 def REP_MOVSD_64 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsl|rep movsd}",
364 [(X86rep_movs i32)], IIC_REP_MOVS>, REP, OpSize32,
365 Requires<[In64BitMode]>;
366 def REP_MOVSQ_64 : RI<0xA5, RawFrm, (outs), (ins), "{rep;movsq|rep movsq}",
367 [(X86rep_movs i64)], IIC_REP_MOVS>, REP,
368 Requires<[In64BitMode]>;
371 // FIXME: Should use "(X86rep_stos AL)" as the pattern.
372 let Defs = [ECX,EDI], isCodeGenOnly = 1 in {
373 let Uses = [AL,ECX,EDI] in
374 def REP_STOSB_32 : I<0xAA, RawFrm, (outs), (ins), "{rep;stosb|rep stosb}",
375 [(X86rep_stos i8)], IIC_REP_STOS>, REP,
376 Requires<[Not64BitMode]>;
377 let Uses = [AX,ECX,EDI] in
378 def REP_STOSW_32 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosw|rep stosw}",
379 [(X86rep_stos i16)], IIC_REP_STOS>, REP, OpSize16,
380 Requires<[Not64BitMode]>;
381 let Uses = [EAX,ECX,EDI] in
382 def REP_STOSD_32 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosl|rep stosd}",
383 [(X86rep_stos i32)], IIC_REP_STOS>, REP, OpSize32,
384 Requires<[Not64BitMode]>;
387 let Defs = [RCX,RDI], isCodeGenOnly = 1 in {
388 let Uses = [AL,RCX,RDI] in
389 def REP_STOSB_64 : I<0xAA, RawFrm, (outs), (ins), "{rep;stosb|rep stosb}",
390 [(X86rep_stos i8)], IIC_REP_STOS>, REP,
391 Requires<[In64BitMode]>;
392 let Uses = [AX,RCX,RDI] in
393 def REP_STOSW_64 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosw|rep stosw}",
394 [(X86rep_stos i16)], IIC_REP_STOS>, REP, OpSize16,
395 Requires<[In64BitMode]>;
396 let Uses = [RAX,RCX,RDI] in
397 def REP_STOSD_64 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosl|rep stosd}",
398 [(X86rep_stos i32)], IIC_REP_STOS>, REP, OpSize32,
399 Requires<[In64BitMode]>;
401 let Uses = [RAX,RCX,RDI] in
402 def REP_STOSQ_64 : RI<0xAB, RawFrm, (outs), (ins), "{rep;stosq|rep stosq}",
403 [(X86rep_stos i64)], IIC_REP_STOS>, REP,
404 Requires<[In64BitMode]>;
408 //===----------------------------------------------------------------------===//
409 // Thread Local Storage Instructions
413 // All calls clobber the non-callee saved registers. ESP is marked as
414 // a use to prevent stack-pointer assignments that appear immediately
415 // before calls from potentially appearing dead.
416 let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, FP7,
417 ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7,
418 MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
419 XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
420 XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
422 def TLS_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
424 [(X86tlsaddr tls32addr:$sym)]>,
425 Requires<[Not64BitMode]>;
426 def TLS_base_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
428 [(X86tlsbaseaddr tls32baseaddr:$sym)]>,
429 Requires<[Not64BitMode]>;
432 // All calls clobber the non-callee saved registers. RSP is marked as
433 // a use to prevent stack-pointer assignments that appear immediately
434 // before calls from potentially appearing dead.
435 let Defs = [RAX, RCX, RDX, RSI, RDI, R8, R9, R10, R11,
436 FP0, FP1, FP2, FP3, FP4, FP5, FP6, FP7,
437 ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7,
438 MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
439 XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
440 XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
442 def TLS_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
444 [(X86tlsaddr tls64addr:$sym)]>,
445 Requires<[In64BitMode]>;
446 def TLS_base_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
448 [(X86tlsbaseaddr tls64baseaddr:$sym)]>,
449 Requires<[In64BitMode]>;
452 // Darwin TLS Support
453 // For i386, the address of the thunk is passed on the stack, on return the
454 // address of the variable is in %eax. %ecx is trashed during the function
455 // call. All other registers are preserved.
456 let Defs = [EAX, ECX, EFLAGS],
458 usesCustomInserter = 1 in
459 def TLSCall_32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
461 [(X86TLSCall addr:$sym)]>,
462 Requires<[Not64BitMode]>;
464 // For x86_64, the address of the thunk is passed in %rdi, on return
465 // the address of the variable is in %rax. All other registers are preserved.
466 let Defs = [RAX, EFLAGS],
468 usesCustomInserter = 1 in
469 def TLSCall_64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
471 [(X86TLSCall addr:$sym)]>,
472 Requires<[In64BitMode]>;
475 //===----------------------------------------------------------------------===//
476 // Conditional Move Pseudo Instructions
478 // CMOV* - Used to implement the SELECT DAG operation. Expanded after
479 // instruction selection into a branch sequence.
480 multiclass CMOVrr_PSEUDO<RegisterClass RC, ValueType VT> {
481 def CMOV#NAME : I<0, Pseudo,
482 (outs RC:$dst), (ins RC:$t, RC:$f, i8imm:$cond),
483 "#CMOV_"#NAME#" PSEUDO!",
484 [(set RC:$dst, (VT (X86cmov RC:$t, RC:$f, imm:$cond,
488 let usesCustomInserter = 1, Uses = [EFLAGS] in {
489 // X86 doesn't have 8-bit conditional moves. Use a customInserter to
490 // emit control flow. An alternative to this is to mark i8 SELECT as Promote,
491 // however that requires promoting the operands, and can induce additional
492 // i8 register pressure.
493 defm _GR8 : CMOVrr_PSEUDO<GR8, i8>;
495 let Predicates = [NoCMov] in {
496 defm _GR32 : CMOVrr_PSEUDO<GR32, i32>;
497 defm _GR16 : CMOVrr_PSEUDO<GR16, i16>;
498 } // Predicates = [NoCMov]
500 // fcmov doesn't handle all possible EFLAGS, provide a fallback if there is no
502 let Predicates = [FPStackf32] in
503 defm _RFP32 : CMOVrr_PSEUDO<RFP32, f32>;
505 let Predicates = [FPStackf64] in
506 defm _RFP64 : CMOVrr_PSEUDO<RFP64, f64>;
508 defm _RFP80 : CMOVrr_PSEUDO<RFP80, f80>;
510 defm _FR32 : CMOVrr_PSEUDO<FR32, f32>;
511 defm _FR64 : CMOVrr_PSEUDO<FR64, f64>;
512 defm _V4F32 : CMOVrr_PSEUDO<VR128, v4f32>;
513 defm _V2F64 : CMOVrr_PSEUDO<VR128, v2f64>;
514 defm _V2I64 : CMOVrr_PSEUDO<VR128, v2i64>;
515 defm _V8F32 : CMOVrr_PSEUDO<VR256, v8f32>;
516 defm _V4F64 : CMOVrr_PSEUDO<VR256, v4f64>;
517 defm _V4I64 : CMOVrr_PSEUDO<VR256, v4i64>;
518 defm _V8I64 : CMOVrr_PSEUDO<VR512, v8i64>;
519 defm _V8F64 : CMOVrr_PSEUDO<VR512, v8f64>;
520 defm _V16F32 : CMOVrr_PSEUDO<VR512, v16f32>;
521 defm _V8I1 : CMOVrr_PSEUDO<VK8, v8i1>;
522 defm _V16I1 : CMOVrr_PSEUDO<VK16, v16i1>;
523 defm _V32I1 : CMOVrr_PSEUDO<VK32, v32i1>;
524 defm _V64I1 : CMOVrr_PSEUDO<VK64, v64i1>;
525 } // usesCustomInserter = 1, Uses = [EFLAGS]
527 //===----------------------------------------------------------------------===//
528 // Normal-Instructions-With-Lock-Prefix Pseudo Instructions
529 //===----------------------------------------------------------------------===//
531 // FIXME: Use normal instructions and add lock prefix dynamically.
535 // TODO: Get this to fold the constant into the instruction.
536 let isCodeGenOnly = 1, Defs = [EFLAGS] in
537 def OR32mrLocked : I<0x09, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$zero),
538 "or{l}\t{$zero, $dst|$dst, $zero}",
539 [], IIC_ALU_MEM>, Requires<[Not64BitMode]>, LOCK,
540 Sched<[WriteALULd, WriteRMW]>;
542 let hasSideEffects = 1 in
543 def Int_MemBarrier : I<0, Pseudo, (outs), (ins),
545 [(X86MemBarrier)]>, Sched<[WriteLoad]>;
547 // RegOpc corresponds to the mr version of the instruction
548 // ImmOpc corresponds to the mi version of the instruction
549 // ImmOpc8 corresponds to the mi8 version of the instruction
550 // ImmMod corresponds to the instruction format of the mi and mi8 versions
551 multiclass LOCK_ArithBinOp<bits<8> RegOpc, bits<8> ImmOpc, bits<8> ImmOpc8,
552 Format ImmMod, string mnemonic> {
553 let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
554 SchedRW = [WriteALULd, WriteRMW] in {
556 def NAME#8mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
557 RegOpc{3}, RegOpc{2}, RegOpc{1}, 0 },
558 MRMDestMem, (outs), (ins i8mem:$dst, GR8:$src2),
559 !strconcat(mnemonic, "{b}\t",
560 "{$src2, $dst|$dst, $src2}"),
561 [], IIC_ALU_NONMEM>, LOCK;
562 def NAME#16mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
563 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
564 MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
565 !strconcat(mnemonic, "{w}\t",
566 "{$src2, $dst|$dst, $src2}"),
567 [], IIC_ALU_NONMEM>, OpSize16, LOCK;
568 def NAME#32mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
569 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
570 MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
571 !strconcat(mnemonic, "{l}\t",
572 "{$src2, $dst|$dst, $src2}"),
573 [], IIC_ALU_NONMEM>, OpSize32, LOCK;
574 def NAME#64mr : RI<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
575 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
576 MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
577 !strconcat(mnemonic, "{q}\t",
578 "{$src2, $dst|$dst, $src2}"),
579 [], IIC_ALU_NONMEM>, LOCK;
581 def NAME#8mi : Ii8<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
582 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 0 },
583 ImmMod, (outs), (ins i8mem :$dst, i8imm :$src2),
584 !strconcat(mnemonic, "{b}\t",
585 "{$src2, $dst|$dst, $src2}"),
586 [], IIC_ALU_MEM>, LOCK;
588 def NAME#16mi : Ii16<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
589 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
590 ImmMod, (outs), (ins i16mem :$dst, i16imm :$src2),
591 !strconcat(mnemonic, "{w}\t",
592 "{$src2, $dst|$dst, $src2}"),
593 [], IIC_ALU_MEM>, OpSize16, LOCK;
595 def NAME#32mi : Ii32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
596 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
597 ImmMod, (outs), (ins i32mem :$dst, i32imm :$src2),
598 !strconcat(mnemonic, "{l}\t",
599 "{$src2, $dst|$dst, $src2}"),
600 [], IIC_ALU_MEM>, OpSize32, LOCK;
602 def NAME#64mi32 : RIi32S<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
603 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
604 ImmMod, (outs), (ins i64mem :$dst, i64i32imm :$src2),
605 !strconcat(mnemonic, "{q}\t",
606 "{$src2, $dst|$dst, $src2}"),
607 [], IIC_ALU_MEM>, LOCK;
609 def NAME#16mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
610 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
611 ImmMod, (outs), (ins i16mem :$dst, i16i8imm :$src2),
612 !strconcat(mnemonic, "{w}\t",
613 "{$src2, $dst|$dst, $src2}"),
614 [], IIC_ALU_MEM>, OpSize16, LOCK;
615 def NAME#32mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
616 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
617 ImmMod, (outs), (ins i32mem :$dst, i32i8imm :$src2),
618 !strconcat(mnemonic, "{l}\t",
619 "{$src2, $dst|$dst, $src2}"),
620 [], IIC_ALU_MEM>, OpSize32, LOCK;
621 def NAME#64mi8 : RIi8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
622 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
623 ImmMod, (outs), (ins i64mem :$dst, i64i8imm :$src2),
624 !strconcat(mnemonic, "{q}\t",
625 "{$src2, $dst|$dst, $src2}"),
626 [], IIC_ALU_MEM>, LOCK;
632 defm LOCK_ADD : LOCK_ArithBinOp<0x00, 0x80, 0x83, MRM0m, "add">;
633 defm LOCK_SUB : LOCK_ArithBinOp<0x28, 0x80, 0x83, MRM5m, "sub">;
634 defm LOCK_OR : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM1m, "or">;
635 defm LOCK_AND : LOCK_ArithBinOp<0x20, 0x80, 0x83, MRM4m, "and">;
636 defm LOCK_XOR : LOCK_ArithBinOp<0x30, 0x80, 0x83, MRM6m, "xor">;
638 // Optimized codegen when the non-memory output is not used.
639 multiclass LOCK_ArithUnOp<bits<8> Opc8, bits<8> Opc, Format Form,
641 let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
642 SchedRW = [WriteALULd, WriteRMW] in {
644 def NAME#8m : I<Opc8, Form, (outs), (ins i8mem :$dst),
645 !strconcat(mnemonic, "{b}\t$dst"),
646 [], IIC_UNARY_MEM>, LOCK;
647 def NAME#16m : I<Opc, Form, (outs), (ins i16mem:$dst),
648 !strconcat(mnemonic, "{w}\t$dst"),
649 [], IIC_UNARY_MEM>, OpSize16, LOCK;
650 def NAME#32m : I<Opc, Form, (outs), (ins i32mem:$dst),
651 !strconcat(mnemonic, "{l}\t$dst"),
652 [], IIC_UNARY_MEM>, OpSize32, LOCK;
653 def NAME#64m : RI<Opc, Form, (outs), (ins i64mem:$dst),
654 !strconcat(mnemonic, "{q}\t$dst"),
655 [], IIC_UNARY_MEM>, LOCK;
659 defm LOCK_INC : LOCK_ArithUnOp<0xFE, 0xFF, MRM0m, "inc">;
660 defm LOCK_DEC : LOCK_ArithUnOp<0xFE, 0xFF, MRM1m, "dec">;
662 // Atomic compare and swap.
663 multiclass LCMPXCHG_UnOp<bits<8> Opc, Format Form, string mnemonic,
664 SDPatternOperator frag, X86MemOperand x86memop,
665 InstrItinClass itin> {
666 let isCodeGenOnly = 1 in {
667 def NAME : I<Opc, Form, (outs), (ins x86memop:$ptr),
668 !strconcat(mnemonic, "\t$ptr"),
669 [(frag addr:$ptr)], itin>, TB, LOCK;
673 multiclass LCMPXCHG_BinOp<bits<8> Opc8, bits<8> Opc, Format Form,
674 string mnemonic, SDPatternOperator frag,
675 InstrItinClass itin8, InstrItinClass itin> {
676 let isCodeGenOnly = 1, SchedRW = [WriteALULd, WriteRMW] in {
677 let Defs = [AL, EFLAGS], Uses = [AL] in
678 def NAME#8 : I<Opc8, Form, (outs), (ins i8mem:$ptr, GR8:$swap),
679 !strconcat(mnemonic, "{b}\t{$swap, $ptr|$ptr, $swap}"),
680 [(frag addr:$ptr, GR8:$swap, 1)], itin8>, TB, LOCK;
681 let Defs = [AX, EFLAGS], Uses = [AX] in
682 def NAME#16 : I<Opc, Form, (outs), (ins i16mem:$ptr, GR16:$swap),
683 !strconcat(mnemonic, "{w}\t{$swap, $ptr|$ptr, $swap}"),
684 [(frag addr:$ptr, GR16:$swap, 2)], itin>, TB, OpSize16, LOCK;
685 let Defs = [EAX, EFLAGS], Uses = [EAX] in
686 def NAME#32 : I<Opc, Form, (outs), (ins i32mem:$ptr, GR32:$swap),
687 !strconcat(mnemonic, "{l}\t{$swap, $ptr|$ptr, $swap}"),
688 [(frag addr:$ptr, GR32:$swap, 4)], itin>, TB, OpSize32, LOCK;
689 let Defs = [RAX, EFLAGS], Uses = [RAX] in
690 def NAME#64 : RI<Opc, Form, (outs), (ins i64mem:$ptr, GR64:$swap),
691 !strconcat(mnemonic, "{q}\t{$swap, $ptr|$ptr, $swap}"),
692 [(frag addr:$ptr, GR64:$swap, 8)], itin>, TB, LOCK;
696 let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
697 SchedRW = [WriteALULd, WriteRMW] in {
698 defm LCMPXCHG8B : LCMPXCHG_UnOp<0xC7, MRM1m, "cmpxchg8b",
703 let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RBX, RCX, RDX],
704 Predicates = [HasCmpxchg16b], SchedRW = [WriteALULd, WriteRMW] in {
705 defm LCMPXCHG16B : LCMPXCHG_UnOp<0xC7, MRM1m, "cmpxchg16b",
707 IIC_CMPX_LOCK_16B>, REX_W;
710 defm LCMPXCHG : LCMPXCHG_BinOp<0xB0, 0xB1, MRMDestMem, "cmpxchg",
711 X86cas, IIC_CMPX_LOCK_8, IIC_CMPX_LOCK>;
713 // Atomic exchange and add
714 multiclass ATOMIC_LOAD_BINOP<bits<8> opc8, bits<8> opc, string mnemonic,
716 InstrItinClass itin8, InstrItinClass itin> {
717 let Constraints = "$val = $dst", Defs = [EFLAGS], isCodeGenOnly = 1,
718 SchedRW = [WriteALULd, WriteRMW] in {
719 def NAME#8 : I<opc8, MRMSrcMem, (outs GR8:$dst),
720 (ins GR8:$val, i8mem:$ptr),
721 !strconcat(mnemonic, "{b}\t{$val, $ptr|$ptr, $val}"),
723 (!cast<PatFrag>(frag # "_8") addr:$ptr, GR8:$val))],
725 def NAME#16 : I<opc, MRMSrcMem, (outs GR16:$dst),
726 (ins GR16:$val, i16mem:$ptr),
727 !strconcat(mnemonic, "{w}\t{$val, $ptr|$ptr, $val}"),
730 (!cast<PatFrag>(frag # "_16") addr:$ptr, GR16:$val))],
732 def NAME#32 : I<opc, MRMSrcMem, (outs GR32:$dst),
733 (ins GR32:$val, i32mem:$ptr),
734 !strconcat(mnemonic, "{l}\t{$val, $ptr|$ptr, $val}"),
737 (!cast<PatFrag>(frag # "_32") addr:$ptr, GR32:$val))],
739 def NAME#64 : RI<opc, MRMSrcMem, (outs GR64:$dst),
740 (ins GR64:$val, i64mem:$ptr),
741 !strconcat(mnemonic, "{q}\t{$val, $ptr|$ptr, $val}"),
744 (!cast<PatFrag>(frag # "_64") addr:$ptr, GR64:$val))],
749 defm LXADD : ATOMIC_LOAD_BINOP<0xc0, 0xc1, "xadd", "atomic_load_add",
750 IIC_XADD_LOCK_MEM8, IIC_XADD_LOCK_MEM>,
753 /* The following multiclass tries to make sure that in code like
754 * x.store (immediate op x.load(acquire), release)
755 * an operation directly on memory is generated instead of wasting a register.
756 * It is not automatic as atomic_store/load are only lowered to MOV instructions
757 * extremely late to prevent them from being accidentally reordered in the backend
758 * (see below the RELEASE_MOV* / ACQUIRE_MOV* pseudo-instructions)
760 multiclass RELEASE_BINOP_MI<string op> {
761 def NAME#8mi : I<0, Pseudo, (outs), (ins i8mem:$dst, i8imm:$src),
762 "#RELEASE_BINOP PSEUDO!",
763 [(atomic_store_8 addr:$dst, (!cast<PatFrag>(op)
764 (atomic_load_8 addr:$dst), (i8 imm:$src)))]>;
765 // NAME#16 is not generated as 16-bit arithmetic instructions are considered
766 // costly and avoided as far as possible by this backend anyway
767 def NAME#32mi : I<0, Pseudo, (outs), (ins i32mem:$dst, i32imm:$src),
768 "#RELEASE_BINOP PSEUDO!",
769 [(atomic_store_32 addr:$dst, (!cast<PatFrag>(op)
770 (atomic_load_32 addr:$dst), (i32 imm:$src)))]>;
771 def NAME#64mi32 : I<0, Pseudo, (outs), (ins i64mem:$dst, i64i32imm:$src),
772 "#RELEASE_BINOP PSEUDO!",
773 [(atomic_store_64 addr:$dst, (!cast<PatFrag>(op)
774 (atomic_load_64 addr:$dst), (i64immSExt32:$src)))]>;
776 defm RELEASE_ADD : RELEASE_BINOP_MI<"add">;
777 defm RELEASE_AND : RELEASE_BINOP_MI<"and">;
778 defm RELEASE_OR : RELEASE_BINOP_MI<"or">;
779 defm RELEASE_XOR : RELEASE_BINOP_MI<"xor">;
780 // Note: we don't deal with sub, because substractions of constants are
781 // optimized into additions before this code can run
783 multiclass RELEASE_UNOP<dag dag8, dag dag16, dag dag32, dag dag64> {
784 def NAME#8m : I<0, Pseudo, (outs), (ins i8mem:$dst),
785 "#RELEASE_UNOP PSEUDO!",
786 [(atomic_store_8 addr:$dst, dag8)]>;
787 def NAME#16m : I<0, Pseudo, (outs), (ins i16mem:$dst),
788 "#RELEASE_UNOP PSEUDO!",
789 [(atomic_store_16 addr:$dst, dag16)]>;
790 def NAME#32m : I<0, Pseudo, (outs), (ins i32mem:$dst),
791 "#RELEASE_UNOP PSEUDO!",
792 [(atomic_store_32 addr:$dst, dag32)]>;
793 def NAME#64m : I<0, Pseudo, (outs), (ins i64mem:$dst),
794 "#RELEASE_UNOP PSEUDO!",
795 [(atomic_store_64 addr:$dst, dag64)]>;
798 defm RELEASE_INC : RELEASE_UNOP<
799 (add (atomic_load_8 addr:$dst), (i8 1)),
800 (add (atomic_load_16 addr:$dst), (i16 1)),
801 (add (atomic_load_32 addr:$dst), (i32 1)),
802 (add (atomic_load_64 addr:$dst), (i64 1))>, Requires<[NotSlowIncDec]>;
803 defm RELEASE_DEC : RELEASE_UNOP<
804 (add (atomic_load_8 addr:$dst), (i8 -1)),
805 (add (atomic_load_16 addr:$dst), (i16 -1)),
806 (add (atomic_load_32 addr:$dst), (i32 -1)),
807 (add (atomic_load_64 addr:$dst), (i64 -1))>, Requires<[NotSlowIncDec]>;
809 TODO: These don't work because the type inference of TableGen fails.
810 TODO: find a way to fix it.
811 defm RELEASE_NEG : RELEASE_UNOP<
812 (ineg (atomic_load_8 addr:$dst)),
813 (ineg (atomic_load_16 addr:$dst)),
814 (ineg (atomic_load_32 addr:$dst)),
815 (ineg (atomic_load_64 addr:$dst))>;
816 defm RELEASE_NOT : RELEASE_UNOP<
817 (not (atomic_load_8 addr:$dst)),
818 (not (atomic_load_16 addr:$dst)),
819 (not (atomic_load_32 addr:$dst)),
820 (not (atomic_load_64 addr:$dst))>;
823 def RELEASE_MOV8mi : I<0, Pseudo, (outs), (ins i8mem:$dst, i8imm:$src),
824 "#RELEASE_MOV PSEUDO !",
825 [(atomic_store_8 addr:$dst, (i8 imm:$src))]>;
826 def RELEASE_MOV16mi : I<0, Pseudo, (outs), (ins i16mem:$dst, i16imm:$src),
827 "#RELEASE_MOV PSEUDO !",
828 [(atomic_store_16 addr:$dst, (i16 imm:$src))]>;
829 def RELEASE_MOV32mi : I<0, Pseudo, (outs), (ins i32mem:$dst, i32imm:$src),
830 "#RELEASE_MOV PSEUDO !",
831 [(atomic_store_32 addr:$dst, (i32 imm:$src))]>;
832 def RELEASE_MOV64mi32 : I<0, Pseudo, (outs), (ins i64mem:$dst, i64i32imm:$src),
833 "#RELEASE_MOV PSEUDO !",
834 [(atomic_store_64 addr:$dst, i64immSExt32:$src)]>;
836 def RELEASE_MOV8mr : I<0, Pseudo, (outs), (ins i8mem :$dst, GR8 :$src),
837 "#RELEASE_MOV PSEUDO!",
838 [(atomic_store_8 addr:$dst, GR8 :$src)]>;
839 def RELEASE_MOV16mr : I<0, Pseudo, (outs), (ins i16mem:$dst, GR16:$src),
840 "#RELEASE_MOV PSEUDO!",
841 [(atomic_store_16 addr:$dst, GR16:$src)]>;
842 def RELEASE_MOV32mr : I<0, Pseudo, (outs), (ins i32mem:$dst, GR32:$src),
843 "#RELEASE_MOV PSEUDO!",
844 [(atomic_store_32 addr:$dst, GR32:$src)]>;
845 def RELEASE_MOV64mr : I<0, Pseudo, (outs), (ins i64mem:$dst, GR64:$src),
846 "#RELEASE_MOV PSEUDO!",
847 [(atomic_store_64 addr:$dst, GR64:$src)]>;
849 def ACQUIRE_MOV8rm : I<0, Pseudo, (outs GR8 :$dst), (ins i8mem :$src),
850 "#ACQUIRE_MOV PSEUDO!",
851 [(set GR8:$dst, (atomic_load_8 addr:$src))]>;
852 def ACQUIRE_MOV16rm : I<0, Pseudo, (outs GR16:$dst), (ins i16mem:$src),
853 "#ACQUIRE_MOV PSEUDO!",
854 [(set GR16:$dst, (atomic_load_16 addr:$src))]>;
855 def ACQUIRE_MOV32rm : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$src),
856 "#ACQUIRE_MOV PSEUDO!",
857 [(set GR32:$dst, (atomic_load_32 addr:$src))]>;
858 def ACQUIRE_MOV64rm : I<0, Pseudo, (outs GR64:$dst), (ins i64mem:$src),
859 "#ACQUIRE_MOV PSEUDO!",
860 [(set GR64:$dst, (atomic_load_64 addr:$src))]>;
862 //===----------------------------------------------------------------------===//
863 // DAG Pattern Matching Rules
864 //===----------------------------------------------------------------------===//
866 // ConstantPool GlobalAddress, ExternalSymbol, and JumpTable
867 def : Pat<(i32 (X86Wrapper tconstpool :$dst)), (MOV32ri tconstpool :$dst)>;
868 def : Pat<(i32 (X86Wrapper tjumptable :$dst)), (MOV32ri tjumptable :$dst)>;
869 def : Pat<(i32 (X86Wrapper tglobaltlsaddr:$dst)),(MOV32ri tglobaltlsaddr:$dst)>;
870 def : Pat<(i32 (X86Wrapper tglobaladdr :$dst)), (MOV32ri tglobaladdr :$dst)>;
871 def : Pat<(i32 (X86Wrapper texternalsym:$dst)), (MOV32ri texternalsym:$dst)>;
872 def : Pat<(i32 (X86Wrapper tblockaddress:$dst)), (MOV32ri tblockaddress:$dst)>;
874 def : Pat<(add GR32:$src1, (X86Wrapper tconstpool:$src2)),
875 (ADD32ri GR32:$src1, tconstpool:$src2)>;
876 def : Pat<(add GR32:$src1, (X86Wrapper tjumptable:$src2)),
877 (ADD32ri GR32:$src1, tjumptable:$src2)>;
878 def : Pat<(add GR32:$src1, (X86Wrapper tglobaladdr :$src2)),
879 (ADD32ri GR32:$src1, tglobaladdr:$src2)>;
880 def : Pat<(add GR32:$src1, (X86Wrapper texternalsym:$src2)),
881 (ADD32ri GR32:$src1, texternalsym:$src2)>;
882 def : Pat<(add GR32:$src1, (X86Wrapper tblockaddress:$src2)),
883 (ADD32ri GR32:$src1, tblockaddress:$src2)>;
885 def : Pat<(store (i32 (X86Wrapper tglobaladdr:$src)), addr:$dst),
886 (MOV32mi addr:$dst, tglobaladdr:$src)>;
887 def : Pat<(store (i32 (X86Wrapper texternalsym:$src)), addr:$dst),
888 (MOV32mi addr:$dst, texternalsym:$src)>;
889 def : Pat<(store (i32 (X86Wrapper tblockaddress:$src)), addr:$dst),
890 (MOV32mi addr:$dst, tblockaddress:$src)>;
892 // ConstantPool GlobalAddress, ExternalSymbol, and JumpTable when not in small
893 // code model mode, should use 'movabs'. FIXME: This is really a hack, the
894 // 'movabs' predicate should handle this sort of thing.
895 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
896 (MOV64ri tconstpool :$dst)>, Requires<[FarData]>;
897 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
898 (MOV64ri tjumptable :$dst)>, Requires<[FarData]>;
899 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
900 (MOV64ri tglobaladdr :$dst)>, Requires<[FarData]>;
901 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
902 (MOV64ri texternalsym:$dst)>, Requires<[FarData]>;
903 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
904 (MOV64ri tblockaddress:$dst)>, Requires<[FarData]>;
906 // In kernel code model, we can get the address of a label
907 // into a register with 'movq'. FIXME: This is a hack, the 'imm' predicate of
908 // the MOV64ri32 should accept these.
909 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
910 (MOV64ri32 tconstpool :$dst)>, Requires<[KernelCode]>;
911 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
912 (MOV64ri32 tjumptable :$dst)>, Requires<[KernelCode]>;
913 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
914 (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
915 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
916 (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
917 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
918 (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
920 // If we have small model and -static mode, it is safe to store global addresses
921 // directly as immediates. FIXME: This is really a hack, the 'imm' predicate
922 // for MOV64mi32 should handle this sort of thing.
923 def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
924 (MOV64mi32 addr:$dst, tconstpool:$src)>,
925 Requires<[NearData, IsStatic]>;
926 def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
927 (MOV64mi32 addr:$dst, tjumptable:$src)>,
928 Requires<[NearData, IsStatic]>;
929 def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
930 (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
931 Requires<[NearData, IsStatic]>;
932 def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
933 (MOV64mi32 addr:$dst, texternalsym:$src)>,
934 Requires<[NearData, IsStatic]>;
935 def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
936 (MOV64mi32 addr:$dst, tblockaddress:$src)>,
937 Requires<[NearData, IsStatic]>;
939 def : Pat<(i32 (X86RecoverFrameAlloc texternalsym:$dst)), (MOV32ri texternalsym:$dst)>;
940 def : Pat<(i64 (X86RecoverFrameAlloc texternalsym:$dst)), (MOV64ri texternalsym:$dst)>;
944 // tls has some funny stuff here...
945 // This corresponds to movabs $foo@tpoff, %rax
946 def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
947 (MOV64ri32 tglobaltlsaddr :$dst)>;
948 // This corresponds to add $foo@tpoff, %rax
949 def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
950 (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
953 // Direct PC relative function call for small code model. 32-bit displacement
954 // sign extended to 64-bit.
955 def : Pat<(X86call (i64 tglobaladdr:$dst)),
956 (CALL64pcrel32 tglobaladdr:$dst)>;
957 def : Pat<(X86call (i64 texternalsym:$dst)),
958 (CALL64pcrel32 texternalsym:$dst)>;
960 // Tailcall stuff. The TCRETURN instructions execute after the epilog, so they
961 // can never use callee-saved registers. That is the purpose of the GR64_TC
964 // The only volatile register that is never used by the calling convention is
965 // %r11. This happens when calling a vararg function with 6 arguments.
967 // Match an X86tcret that uses less than 7 volatile registers.
968 def X86tcret_6regs : PatFrag<(ops node:$ptr, node:$off),
969 (X86tcret node:$ptr, node:$off), [{
970 // X86tcret args: (*chain, ptr, imm, regs..., glue)
971 unsigned NumRegs = 0;
972 for (unsigned i = 3, e = N->getNumOperands(); i != e; ++i)
973 if (isa<RegisterSDNode>(N->getOperand(i)) && ++NumRegs > 6)
978 def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
979 (TCRETURNri ptr_rc_tailcall:$dst, imm:$off)>,
980 Requires<[Not64BitMode]>;
982 // FIXME: This is disabled for 32-bit PIC mode because the global base
983 // register which is part of the address mode may be assigned a
984 // callee-saved register.
985 def : Pat<(X86tcret (load addr:$dst), imm:$off),
986 (TCRETURNmi addr:$dst, imm:$off)>,
987 Requires<[Not64BitMode, IsNotPIC]>;
989 def : Pat<(X86tcret (i32 tglobaladdr:$dst), imm:$off),
990 (TCRETURNdi tglobaladdr:$dst, imm:$off)>,
993 def : Pat<(X86tcret (i32 texternalsym:$dst), imm:$off),
994 (TCRETURNdi texternalsym:$dst, imm:$off)>,
997 def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
998 (TCRETURNri64 ptr_rc_tailcall:$dst, imm:$off)>,
999 Requires<[In64BitMode]>;
1001 // Don't fold loads into X86tcret requiring more than 6 regs.
1002 // There wouldn't be enough scratch registers for base+index.
1003 def : Pat<(X86tcret_6regs (load addr:$dst), imm:$off),
1004 (TCRETURNmi64 addr:$dst, imm:$off)>,
1005 Requires<[In64BitMode]>;
1007 def : Pat<(X86tcret (i64 tglobaladdr:$dst), imm:$off),
1008 (TCRETURNdi64 tglobaladdr:$dst, imm:$off)>,
1011 def : Pat<(X86tcret (i64 texternalsym:$dst), imm:$off),
1012 (TCRETURNdi64 texternalsym:$dst, imm:$off)>,
1015 // Normal calls, with various flavors of addresses.
1016 def : Pat<(X86call (i32 tglobaladdr:$dst)),
1017 (CALLpcrel32 tglobaladdr:$dst)>;
1018 def : Pat<(X86call (i32 texternalsym:$dst)),
1019 (CALLpcrel32 texternalsym:$dst)>;
1020 def : Pat<(X86call (i32 imm:$dst)),
1021 (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
1025 // TEST R,R is smaller than CMP R,0
1026 def : Pat<(X86cmp GR8:$src1, 0),
1027 (TEST8rr GR8:$src1, GR8:$src1)>;
1028 def : Pat<(X86cmp GR16:$src1, 0),
1029 (TEST16rr GR16:$src1, GR16:$src1)>;
1030 def : Pat<(X86cmp GR32:$src1, 0),
1031 (TEST32rr GR32:$src1, GR32:$src1)>;
1032 def : Pat<(X86cmp GR64:$src1, 0),
1033 (TEST64rr GR64:$src1, GR64:$src1)>;
1035 // Conditional moves with folded loads with operands swapped and conditions
1037 multiclass CMOVmr<PatLeaf InvertedCond, Instruction Inst16, Instruction Inst32,
1038 Instruction Inst64> {
1039 let Predicates = [HasCMov] in {
1040 def : Pat<(X86cmov (loadi16 addr:$src1), GR16:$src2, InvertedCond, EFLAGS),
1041 (Inst16 GR16:$src2, addr:$src1)>;
1042 def : Pat<(X86cmov (loadi32 addr:$src1), GR32:$src2, InvertedCond, EFLAGS),
1043 (Inst32 GR32:$src2, addr:$src1)>;
1044 def : Pat<(X86cmov (loadi64 addr:$src1), GR64:$src2, InvertedCond, EFLAGS),
1045 (Inst64 GR64:$src2, addr:$src1)>;
1049 defm : CMOVmr<X86_COND_B , CMOVAE16rm, CMOVAE32rm, CMOVAE64rm>;
1050 defm : CMOVmr<X86_COND_AE, CMOVB16rm , CMOVB32rm , CMOVB64rm>;
1051 defm : CMOVmr<X86_COND_E , CMOVNE16rm, CMOVNE32rm, CMOVNE64rm>;
1052 defm : CMOVmr<X86_COND_NE, CMOVE16rm , CMOVE32rm , CMOVE64rm>;
1053 defm : CMOVmr<X86_COND_BE, CMOVA16rm , CMOVA32rm , CMOVA64rm>;
1054 defm : CMOVmr<X86_COND_A , CMOVBE16rm, CMOVBE32rm, CMOVBE64rm>;
1055 defm : CMOVmr<X86_COND_L , CMOVGE16rm, CMOVGE32rm, CMOVGE64rm>;
1056 defm : CMOVmr<X86_COND_GE, CMOVL16rm , CMOVL32rm , CMOVL64rm>;
1057 defm : CMOVmr<X86_COND_LE, CMOVG16rm , CMOVG32rm , CMOVG64rm>;
1058 defm : CMOVmr<X86_COND_G , CMOVLE16rm, CMOVLE32rm, CMOVLE64rm>;
1059 defm : CMOVmr<X86_COND_P , CMOVNP16rm, CMOVNP32rm, CMOVNP64rm>;
1060 defm : CMOVmr<X86_COND_NP, CMOVP16rm , CMOVP32rm , CMOVP64rm>;
1061 defm : CMOVmr<X86_COND_S , CMOVNS16rm, CMOVNS32rm, CMOVNS64rm>;
1062 defm : CMOVmr<X86_COND_NS, CMOVS16rm , CMOVS32rm , CMOVS64rm>;
1063 defm : CMOVmr<X86_COND_O , CMOVNO16rm, CMOVNO32rm, CMOVNO64rm>;
1064 defm : CMOVmr<X86_COND_NO, CMOVO16rm , CMOVO32rm , CMOVO64rm>;
1066 // zextload bool -> zextload byte
1067 def : Pat<(zextloadi8i1 addr:$src), (AND8ri (MOV8rm addr:$src), (i8 1))>;
1068 def : Pat<(zextloadi16i1 addr:$src), (AND16ri (MOVZX16rm8 addr:$src), (i16 1))>;
1069 def : Pat<(zextloadi32i1 addr:$src), (AND32ri (MOVZX32rm8 addr:$src), (i32 1))>;
1070 def : Pat<(zextloadi64i1 addr:$src),
1071 (SUBREG_TO_REG (i64 0),
1072 (AND32ri (MOVZX32rm8 addr:$src), (i32 1)), sub_32bit)>;
1074 // extload bool -> extload byte
1075 // When extloading from 16-bit and smaller memory locations into 64-bit
1076 // registers, use zero-extending loads so that the entire 64-bit register is
1077 // defined, avoiding partial-register updates.
1079 def : Pat<(extloadi8i1 addr:$src), (MOV8rm addr:$src)>;
1080 def : Pat<(extloadi16i1 addr:$src), (MOVZX16rm8 addr:$src)>;
1081 def : Pat<(extloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1082 def : Pat<(extloadi16i8 addr:$src), (MOVZX16rm8 addr:$src)>;
1083 def : Pat<(extloadi32i8 addr:$src), (MOVZX32rm8 addr:$src)>;
1084 def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
1086 // For other extloads, use subregs, since the high contents of the register are
1087 // defined after an extload.
1088 def : Pat<(extloadi64i1 addr:$src),
1089 (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1090 def : Pat<(extloadi64i8 addr:$src),
1091 (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1092 def : Pat<(extloadi64i16 addr:$src),
1093 (SUBREG_TO_REG (i64 0), (MOVZX32rm16 addr:$src), sub_32bit)>;
1094 def : Pat<(extloadi64i32 addr:$src),
1095 (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src), sub_32bit)>;
1097 // anyext. Define these to do an explicit zero-extend to
1098 // avoid partial-register updates.
1099 def : Pat<(i16 (anyext GR8 :$src)), (EXTRACT_SUBREG
1100 (MOVZX32rr8 GR8 :$src), sub_16bit)>;
1101 def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8 GR8 :$src)>;
1103 // Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
1104 def : Pat<(i32 (anyext GR16:$src)),
1105 (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
1107 def : Pat<(i64 (anyext GR8 :$src)),
1108 (SUBREG_TO_REG (i64 0), (MOVZX32rr8 GR8 :$src), sub_32bit)>;
1109 def : Pat<(i64 (anyext GR16:$src)),
1110 (SUBREG_TO_REG (i64 0), (MOVZX32rr16 GR16 :$src), sub_32bit)>;
1111 def : Pat<(i64 (anyext GR32:$src)),
1112 (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1115 // Any instruction that defines a 32-bit result leaves the high half of the
1116 // register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
1117 // be copying from a truncate. And x86's cmov doesn't do anything if the
1118 // condition is false. But any other 32-bit operation will zero-extend
1120 def def32 : PatLeaf<(i32 GR32:$src), [{
1121 return N->getOpcode() != ISD::TRUNCATE &&
1122 N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
1123 N->getOpcode() != ISD::CopyFromReg &&
1124 N->getOpcode() != ISD::AssertSext &&
1125 N->getOpcode() != X86ISD::CMOV;
1128 // In the case of a 32-bit def that is known to implicitly zero-extend,
1129 // we can use a SUBREG_TO_REG.
1130 def : Pat<(i64 (zext def32:$src)),
1131 (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1133 //===----------------------------------------------------------------------===//
1134 // Pattern match OR as ADD
1135 //===----------------------------------------------------------------------===//
1137 // If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
1138 // 3-addressified into an LEA instruction to avoid copies. However, we also
1139 // want to finally emit these instructions as an or at the end of the code
1140 // generator to make the generated code easier to read. To do this, we select
1141 // into "disjoint bits" pseudo ops.
1143 // Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.
1144 def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
1145 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1146 return CurDAG->MaskedValueIsZero(N->getOperand(0), CN->getAPIntValue());
1148 APInt KnownZero0, KnownOne0;
1149 CurDAG->computeKnownBits(N->getOperand(0), KnownZero0, KnownOne0, 0);
1150 APInt KnownZero1, KnownOne1;
1151 CurDAG->computeKnownBits(N->getOperand(1), KnownZero1, KnownOne1, 0);
1152 return (~KnownZero0 & ~KnownZero1) == 0;
1156 // (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1157 // Try this before the selecting to OR.
1158 let AddedComplexity = 5, SchedRW = [WriteALU] in {
1160 let isConvertibleToThreeAddress = 1,
1161 Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1162 let isCommutable = 1 in {
1163 def ADD16rr_DB : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1164 "", // orw/addw REG, REG
1165 [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1166 def ADD32rr_DB : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1167 "", // orl/addl REG, REG
1168 [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1169 def ADD64rr_DB : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1170 "", // orq/addq REG, REG
1171 [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1174 // NOTE: These are order specific, we want the ri8 forms to be listed
1175 // first so that they are slightly preferred to the ri forms.
1177 def ADD16ri8_DB : I<0, Pseudo,
1178 (outs GR16:$dst), (ins GR16:$src1, i16i8imm:$src2),
1179 "", // orw/addw REG, imm8
1180 [(set GR16:$dst,(or_is_add GR16:$src1,i16immSExt8:$src2))]>;
1181 def ADD16ri_DB : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1182 "", // orw/addw REG, imm
1183 [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1185 def ADD32ri8_DB : I<0, Pseudo,
1186 (outs GR32:$dst), (ins GR32:$src1, i32i8imm:$src2),
1187 "", // orl/addl REG, imm8
1188 [(set GR32:$dst,(or_is_add GR32:$src1,i32immSExt8:$src2))]>;
1189 def ADD32ri_DB : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1190 "", // orl/addl REG, imm
1191 [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1194 def ADD64ri8_DB : I<0, Pseudo,
1195 (outs GR64:$dst), (ins GR64:$src1, i64i8imm:$src2),
1196 "", // orq/addq REG, imm8
1197 [(set GR64:$dst, (or_is_add GR64:$src1,
1198 i64immSExt8:$src2))]>;
1199 def ADD64ri32_DB : I<0, Pseudo,
1200 (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1201 "", // orq/addq REG, imm
1202 [(set GR64:$dst, (or_is_add GR64:$src1,
1203 i64immSExt32:$src2))]>;
1205 } // AddedComplexity, SchedRW
1208 //===----------------------------------------------------------------------===//
1210 //===----------------------------------------------------------------------===//
1212 // Odd encoding trick: -128 fits into an 8-bit immediate field while
1213 // +128 doesn't, so in this special case use a sub instead of an add.
1214 def : Pat<(add GR16:$src1, 128),
1215 (SUB16ri8 GR16:$src1, -128)>;
1216 def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1217 (SUB16mi8 addr:$dst, -128)>;
1219 def : Pat<(add GR32:$src1, 128),
1220 (SUB32ri8 GR32:$src1, -128)>;
1221 def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1222 (SUB32mi8 addr:$dst, -128)>;
1224 def : Pat<(add GR64:$src1, 128),
1225 (SUB64ri8 GR64:$src1, -128)>;
1226 def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1227 (SUB64mi8 addr:$dst, -128)>;
1229 // The same trick applies for 32-bit immediate fields in 64-bit
1231 def : Pat<(add GR64:$src1, 0x0000000080000000),
1232 (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1233 def : Pat<(store (add (loadi64 addr:$dst), 0x00000000800000000), addr:$dst),
1234 (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1236 // To avoid needing to materialize an immediate in a register, use a 32-bit and
1237 // with implicit zero-extension instead of a 64-bit and if the immediate has at
1238 // least 32 bits of leading zeros. If in addition the last 32 bits can be
1239 // represented with a sign extension of a 8 bit constant, use that.
1240 // This can also reduce instruction size by eliminating the need for the REX
1243 // AddedComplexity is needed to give priority over i64immSExt8 and i64immSExt32.
1244 let AddedComplexity = 1 in {
1245 def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
1249 (EXTRACT_SUBREG GR64:$src, sub_32bit),
1250 (i32 (GetLo8XForm imm:$imm))),
1253 def : Pat<(and GR64:$src, i64immZExt32:$imm),
1257 (EXTRACT_SUBREG GR64:$src, sub_32bit),
1258 (i32 (GetLo32XForm imm:$imm))),
1260 } // AddedComplexity = 1
1263 // AddedComplexity is needed due to the increased complexity on the
1264 // i64immZExt32SExt8 and i64immZExt32 patterns above. Applying this to all
1265 // the MOVZX patterns keeps thems together in DAGIsel tables.
1266 let AddedComplexity = 1 in {
1267 // r & (2^16-1) ==> movz
1268 def : Pat<(and GR32:$src1, 0xffff),
1269 (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1270 // r & (2^8-1) ==> movz
1271 def : Pat<(and GR32:$src1, 0xff),
1272 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src1,
1275 Requires<[Not64BitMode]>;
1276 // r & (2^8-1) ==> movz
1277 def : Pat<(and GR16:$src1, 0xff),
1278 (EXTRACT_SUBREG (MOVZX32rr8 (EXTRACT_SUBREG
1279 (i16 (COPY_TO_REGCLASS GR16:$src1, GR16_ABCD)), sub_8bit)),
1281 Requires<[Not64BitMode]>;
1283 // r & (2^32-1) ==> movz
1284 def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1285 (SUBREG_TO_REG (i64 0),
1286 (MOV32rr (EXTRACT_SUBREG GR64:$src, sub_32bit)),
1288 // r & (2^16-1) ==> movz
1289 let AddedComplexity = 1 in // Give priority over i64immZExt32.
1290 def : Pat<(and GR64:$src, 0xffff),
1291 (SUBREG_TO_REG (i64 0),
1292 (MOVZX32rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit))),
1294 // r & (2^8-1) ==> movz
1295 def : Pat<(and GR64:$src, 0xff),
1296 (SUBREG_TO_REG (i64 0),
1297 (MOVZX32rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit))),
1299 // r & (2^8-1) ==> movz
1300 def : Pat<(and GR32:$src1, 0xff),
1301 (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>,
1302 Requires<[In64BitMode]>;
1303 // r & (2^8-1) ==> movz
1304 def : Pat<(and GR16:$src1, 0xff),
1305 (EXTRACT_SUBREG (MOVZX32rr8 (i8
1306 (EXTRACT_SUBREG GR16:$src1, sub_8bit))), sub_16bit)>,
1307 Requires<[In64BitMode]>;
1308 } // AddedComplexity = 1
1311 // sext_inreg patterns
1312 def : Pat<(sext_inreg GR32:$src, i16),
1313 (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1314 def : Pat<(sext_inreg GR32:$src, i8),
1315 (MOVSX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1318 Requires<[Not64BitMode]>;
1320 def : Pat<(sext_inreg GR16:$src, i8),
1321 (EXTRACT_SUBREG (i32 (MOVSX32rr8 (EXTRACT_SUBREG
1322 (i32 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)), sub_8bit))),
1324 Requires<[Not64BitMode]>;
1326 def : Pat<(sext_inreg GR64:$src, i32),
1327 (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1328 def : Pat<(sext_inreg GR64:$src, i16),
1329 (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1330 def : Pat<(sext_inreg GR64:$src, i8),
1331 (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1332 def : Pat<(sext_inreg GR32:$src, i8),
1333 (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>,
1334 Requires<[In64BitMode]>;
1335 def : Pat<(sext_inreg GR16:$src, i8),
1336 (EXTRACT_SUBREG (MOVSX32rr8
1337 (EXTRACT_SUBREG GR16:$src, sub_8bit)), sub_16bit)>,
1338 Requires<[In64BitMode]>;
1340 // sext, sext_load, zext, zext_load
1341 def: Pat<(i16 (sext GR8:$src)),
1342 (EXTRACT_SUBREG (MOVSX32rr8 GR8:$src), sub_16bit)>;
1343 def: Pat<(sextloadi16i8 addr:$src),
1344 (EXTRACT_SUBREG (MOVSX32rm8 addr:$src), sub_16bit)>;
1345 def: Pat<(i16 (zext GR8:$src)),
1346 (EXTRACT_SUBREG (MOVZX32rr8 GR8:$src), sub_16bit)>;
1347 def: Pat<(zextloadi16i8 addr:$src),
1348 (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1351 def : Pat<(i16 (trunc GR32:$src)),
1352 (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1353 def : Pat<(i8 (trunc GR32:$src)),
1354 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1356 Requires<[Not64BitMode]>;
1357 def : Pat<(i8 (trunc GR16:$src)),
1358 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1360 Requires<[Not64BitMode]>;
1361 def : Pat<(i32 (trunc GR64:$src)),
1362 (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1363 def : Pat<(i16 (trunc GR64:$src)),
1364 (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1365 def : Pat<(i8 (trunc GR64:$src)),
1366 (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1367 def : Pat<(i8 (trunc GR32:$src)),
1368 (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1369 Requires<[In64BitMode]>;
1370 def : Pat<(i8 (trunc GR16:$src)),
1371 (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1372 Requires<[In64BitMode]>;
1374 // h-register tricks
1375 def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1376 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1378 Requires<[Not64BitMode]>;
1379 def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1380 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1382 Requires<[Not64BitMode]>;
1383 def : Pat<(srl GR16:$src, (i8 8)),
1386 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1389 Requires<[Not64BitMode]>;
1390 def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1391 (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1394 Requires<[Not64BitMode]>;
1395 def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1396 (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1399 Requires<[Not64BitMode]>;
1400 def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1401 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1404 Requires<[Not64BitMode]>;
1405 def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1406 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1409 Requires<[Not64BitMode]>;
1411 // h-register tricks.
1412 // For now, be conservative on x86-64 and use an h-register extract only if the
1413 // value is immediately zero-extended or stored, which are somewhat common
1414 // cases. This uses a bunch of code to prevent a register requiring a REX prefix
1415 // from being allocated in the same instruction as the h register, as there's
1416 // currently no way to describe this requirement to the register allocator.
1418 // h-register extract and zero-extend.
1419 def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1423 (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1426 def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1428 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1430 Requires<[In64BitMode]>;
1431 def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1432 (MOVZX32_NOREXrr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1435 Requires<[In64BitMode]>;
1436 def : Pat<(srl GR16:$src, (i8 8)),
1439 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1442 Requires<[In64BitMode]>;
1443 def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1445 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1447 Requires<[In64BitMode]>;
1448 def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1450 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1452 Requires<[In64BitMode]>;
1453 def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1457 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1460 def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1464 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1468 // h-register extract and store.
1469 def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1472 (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1474 def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1477 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1479 Requires<[In64BitMode]>;
1480 def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1483 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1485 Requires<[In64BitMode]>;
1488 // (shl x, 1) ==> (add x, x)
1489 // Note that if x is undef (immediate or otherwise), we could theoretically
1490 // end up with the two uses of x getting different values, producing a result
1491 // where the least significant bit is not 0. However, the probability of this
1492 // happening is considered low enough that this is officially not a
1494 def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr GR8 :$src1, GR8 :$src1)>;
1495 def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1496 def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1497 def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1499 // Helper imms that check if a mask doesn't change significant shift bits.
1500 def immShift32 : ImmLeaf<i8, [{
1501 return countTrailingOnes<uint64_t>(Imm) >= 5;
1503 def immShift64 : ImmLeaf<i8, [{
1504 return countTrailingOnes<uint64_t>(Imm) >= 6;
1507 // Shift amount is implicitly masked.
1508 multiclass MaskedShiftAmountPats<SDNode frag, string name> {
1509 // (shift x (and y, 31)) ==> (shift x, y)
1510 def : Pat<(frag GR8:$src1, (and CL, immShift32)),
1511 (!cast<Instruction>(name # "8rCL") GR8:$src1)>;
1512 def : Pat<(frag GR16:$src1, (and CL, immShift32)),
1513 (!cast<Instruction>(name # "16rCL") GR16:$src1)>;
1514 def : Pat<(frag GR32:$src1, (and CL, immShift32)),
1515 (!cast<Instruction>(name # "32rCL") GR32:$src1)>;
1516 def : Pat<(store (frag (loadi8 addr:$dst), (and CL, immShift32)), addr:$dst),
1517 (!cast<Instruction>(name # "8mCL") addr:$dst)>;
1518 def : Pat<(store (frag (loadi16 addr:$dst), (and CL, immShift32)), addr:$dst),
1519 (!cast<Instruction>(name # "16mCL") addr:$dst)>;
1520 def : Pat<(store (frag (loadi32 addr:$dst), (and CL, immShift32)), addr:$dst),
1521 (!cast<Instruction>(name # "32mCL") addr:$dst)>;
1523 // (shift x (and y, 63)) ==> (shift x, y)
1524 def : Pat<(frag GR64:$src1, (and CL, immShift64)),
1525 (!cast<Instruction>(name # "64rCL") GR64:$src1)>;
1526 def : Pat<(store (frag (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1527 (!cast<Instruction>(name # "64mCL") addr:$dst)>;
1530 defm : MaskedShiftAmountPats<shl, "SHL">;
1531 defm : MaskedShiftAmountPats<srl, "SHR">;
1532 defm : MaskedShiftAmountPats<sra, "SAR">;
1533 defm : MaskedShiftAmountPats<rotl, "ROL">;
1534 defm : MaskedShiftAmountPats<rotr, "ROR">;
1536 // (anyext (setcc_carry)) -> (setcc_carry)
1537 def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1539 def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1541 def : Pat<(i32 (anyext (i16 (X86setcc_c X86_COND_B, EFLAGS)))),
1547 //===----------------------------------------------------------------------===//
1548 // EFLAGS-defining Patterns
1549 //===----------------------------------------------------------------------===//
1552 def : Pat<(add GR8 :$src1, GR8 :$src2), (ADD8rr GR8 :$src1, GR8 :$src2)>;
1553 def : Pat<(add GR16:$src1, GR16:$src2), (ADD16rr GR16:$src1, GR16:$src2)>;
1554 def : Pat<(add GR32:$src1, GR32:$src2), (ADD32rr GR32:$src1, GR32:$src2)>;
1557 def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
1558 (ADD8rm GR8:$src1, addr:$src2)>;
1559 def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
1560 (ADD16rm GR16:$src1, addr:$src2)>;
1561 def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
1562 (ADD32rm GR32:$src1, addr:$src2)>;
1565 def : Pat<(add GR8 :$src1, imm:$src2), (ADD8ri GR8:$src1 , imm:$src2)>;
1566 def : Pat<(add GR16:$src1, imm:$src2), (ADD16ri GR16:$src1, imm:$src2)>;
1567 def : Pat<(add GR32:$src1, imm:$src2), (ADD32ri GR32:$src1, imm:$src2)>;
1568 def : Pat<(add GR16:$src1, i16immSExt8:$src2),
1569 (ADD16ri8 GR16:$src1, i16immSExt8:$src2)>;
1570 def : Pat<(add GR32:$src1, i32immSExt8:$src2),
1571 (ADD32ri8 GR32:$src1, i32immSExt8:$src2)>;
1574 def : Pat<(sub GR8 :$src1, GR8 :$src2), (SUB8rr GR8 :$src1, GR8 :$src2)>;
1575 def : Pat<(sub GR16:$src1, GR16:$src2), (SUB16rr GR16:$src1, GR16:$src2)>;
1576 def : Pat<(sub GR32:$src1, GR32:$src2), (SUB32rr GR32:$src1, GR32:$src2)>;
1579 def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
1580 (SUB8rm GR8:$src1, addr:$src2)>;
1581 def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
1582 (SUB16rm GR16:$src1, addr:$src2)>;
1583 def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
1584 (SUB32rm GR32:$src1, addr:$src2)>;
1587 def : Pat<(sub GR8:$src1, imm:$src2),
1588 (SUB8ri GR8:$src1, imm:$src2)>;
1589 def : Pat<(sub GR16:$src1, imm:$src2),
1590 (SUB16ri GR16:$src1, imm:$src2)>;
1591 def : Pat<(sub GR32:$src1, imm:$src2),
1592 (SUB32ri GR32:$src1, imm:$src2)>;
1593 def : Pat<(sub GR16:$src1, i16immSExt8:$src2),
1594 (SUB16ri8 GR16:$src1, i16immSExt8:$src2)>;
1595 def : Pat<(sub GR32:$src1, i32immSExt8:$src2),
1596 (SUB32ri8 GR32:$src1, i32immSExt8:$src2)>;
1599 def : Pat<(X86sub_flag 0, GR8 :$src), (NEG8r GR8 :$src)>;
1600 def : Pat<(X86sub_flag 0, GR16:$src), (NEG16r GR16:$src)>;
1601 def : Pat<(X86sub_flag 0, GR32:$src), (NEG32r GR32:$src)>;
1602 def : Pat<(X86sub_flag 0, GR64:$src), (NEG64r GR64:$src)>;
1605 def : Pat<(mul GR16:$src1, GR16:$src2),
1606 (IMUL16rr GR16:$src1, GR16:$src2)>;
1607 def : Pat<(mul GR32:$src1, GR32:$src2),
1608 (IMUL32rr GR32:$src1, GR32:$src2)>;
1611 def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
1612 (IMUL16rm GR16:$src1, addr:$src2)>;
1613 def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
1614 (IMUL32rm GR32:$src1, addr:$src2)>;
1617 def : Pat<(mul GR16:$src1, imm:$src2),
1618 (IMUL16rri GR16:$src1, imm:$src2)>;
1619 def : Pat<(mul GR32:$src1, imm:$src2),
1620 (IMUL32rri GR32:$src1, imm:$src2)>;
1621 def : Pat<(mul GR16:$src1, i16immSExt8:$src2),
1622 (IMUL16rri8 GR16:$src1, i16immSExt8:$src2)>;
1623 def : Pat<(mul GR32:$src1, i32immSExt8:$src2),
1624 (IMUL32rri8 GR32:$src1, i32immSExt8:$src2)>;
1626 // reg = mul mem, imm
1627 def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
1628 (IMUL16rmi addr:$src1, imm:$src2)>;
1629 def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
1630 (IMUL32rmi addr:$src1, imm:$src2)>;
1631 def : Pat<(mul (loadi16 addr:$src1), i16immSExt8:$src2),
1632 (IMUL16rmi8 addr:$src1, i16immSExt8:$src2)>;
1633 def : Pat<(mul (loadi32 addr:$src1), i32immSExt8:$src2),
1634 (IMUL32rmi8 addr:$src1, i32immSExt8:$src2)>;
1636 // Patterns for nodes that do not produce flags, for instructions that do.
1639 def : Pat<(add GR64:$src1, GR64:$src2),
1640 (ADD64rr GR64:$src1, GR64:$src2)>;
1641 def : Pat<(add GR64:$src1, i64immSExt8:$src2),
1642 (ADD64ri8 GR64:$src1, i64immSExt8:$src2)>;
1643 def : Pat<(add GR64:$src1, i64immSExt32:$src2),
1644 (ADD64ri32 GR64:$src1, i64immSExt32:$src2)>;
1645 def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
1646 (ADD64rm GR64:$src1, addr:$src2)>;
1649 def : Pat<(sub GR64:$src1, GR64:$src2),
1650 (SUB64rr GR64:$src1, GR64:$src2)>;
1651 def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
1652 (SUB64rm GR64:$src1, addr:$src2)>;
1653 def : Pat<(sub GR64:$src1, i64immSExt8:$src2),
1654 (SUB64ri8 GR64:$src1, i64immSExt8:$src2)>;
1655 def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
1656 (SUB64ri32 GR64:$src1, i64immSExt32:$src2)>;
1659 def : Pat<(mul GR64:$src1, GR64:$src2),
1660 (IMUL64rr GR64:$src1, GR64:$src2)>;
1661 def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
1662 (IMUL64rm GR64:$src1, addr:$src2)>;
1663 def : Pat<(mul GR64:$src1, i64immSExt8:$src2),
1664 (IMUL64rri8 GR64:$src1, i64immSExt8:$src2)>;
1665 def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
1666 (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
1667 def : Pat<(mul (loadi64 addr:$src1), i64immSExt8:$src2),
1668 (IMUL64rmi8 addr:$src1, i64immSExt8:$src2)>;
1669 def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
1670 (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
1672 // Increment/Decrement reg.
1673 // Do not make INC/DEC if it is slow
1674 let Predicates = [NotSlowIncDec] in {
1675 def : Pat<(add GR8:$src, 1), (INC8r GR8:$src)>;
1676 def : Pat<(add GR16:$src, 1), (INC16r GR16:$src)>;
1677 def : Pat<(add GR32:$src, 1), (INC32r GR32:$src)>;
1678 def : Pat<(add GR64:$src, 1), (INC64r GR64:$src)>;
1679 def : Pat<(add GR8:$src, -1), (DEC8r GR8:$src)>;
1680 def : Pat<(add GR16:$src, -1), (DEC16r GR16:$src)>;
1681 def : Pat<(add GR32:$src, -1), (DEC32r GR32:$src)>;
1682 def : Pat<(add GR64:$src, -1), (DEC64r GR64:$src)>;
1686 def : Pat<(or GR8 :$src1, GR8 :$src2), (OR8rr GR8 :$src1, GR8 :$src2)>;
1687 def : Pat<(or GR16:$src1, GR16:$src2), (OR16rr GR16:$src1, GR16:$src2)>;
1688 def : Pat<(or GR32:$src1, GR32:$src2), (OR32rr GR32:$src1, GR32:$src2)>;
1689 def : Pat<(or GR64:$src1, GR64:$src2), (OR64rr GR64:$src1, GR64:$src2)>;
1692 def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
1693 (OR8rm GR8:$src1, addr:$src2)>;
1694 def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
1695 (OR16rm GR16:$src1, addr:$src2)>;
1696 def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
1697 (OR32rm GR32:$src1, addr:$src2)>;
1698 def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
1699 (OR64rm GR64:$src1, addr:$src2)>;
1702 def : Pat<(or GR8:$src1 , imm:$src2), (OR8ri GR8 :$src1, imm:$src2)>;
1703 def : Pat<(or GR16:$src1, imm:$src2), (OR16ri GR16:$src1, imm:$src2)>;
1704 def : Pat<(or GR32:$src1, imm:$src2), (OR32ri GR32:$src1, imm:$src2)>;
1705 def : Pat<(or GR16:$src1, i16immSExt8:$src2),
1706 (OR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1707 def : Pat<(or GR32:$src1, i32immSExt8:$src2),
1708 (OR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1709 def : Pat<(or GR64:$src1, i64immSExt8:$src2),
1710 (OR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1711 def : Pat<(or GR64:$src1, i64immSExt32:$src2),
1712 (OR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1715 def : Pat<(xor GR8 :$src1, GR8 :$src2), (XOR8rr GR8 :$src1, GR8 :$src2)>;
1716 def : Pat<(xor GR16:$src1, GR16:$src2), (XOR16rr GR16:$src1, GR16:$src2)>;
1717 def : Pat<(xor GR32:$src1, GR32:$src2), (XOR32rr GR32:$src1, GR32:$src2)>;
1718 def : Pat<(xor GR64:$src1, GR64:$src2), (XOR64rr GR64:$src1, GR64:$src2)>;
1721 def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
1722 (XOR8rm GR8:$src1, addr:$src2)>;
1723 def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
1724 (XOR16rm GR16:$src1, addr:$src2)>;
1725 def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
1726 (XOR32rm GR32:$src1, addr:$src2)>;
1727 def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
1728 (XOR64rm GR64:$src1, addr:$src2)>;
1731 def : Pat<(xor GR8:$src1, imm:$src2),
1732 (XOR8ri GR8:$src1, imm:$src2)>;
1733 def : Pat<(xor GR16:$src1, imm:$src2),
1734 (XOR16ri GR16:$src1, imm:$src2)>;
1735 def : Pat<(xor GR32:$src1, imm:$src2),
1736 (XOR32ri GR32:$src1, imm:$src2)>;
1737 def : Pat<(xor GR16:$src1, i16immSExt8:$src2),
1738 (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1739 def : Pat<(xor GR32:$src1, i32immSExt8:$src2),
1740 (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1741 def : Pat<(xor GR64:$src1, i64immSExt8:$src2),
1742 (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1743 def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
1744 (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1747 def : Pat<(and GR8 :$src1, GR8 :$src2), (AND8rr GR8 :$src1, GR8 :$src2)>;
1748 def : Pat<(and GR16:$src1, GR16:$src2), (AND16rr GR16:$src1, GR16:$src2)>;
1749 def : Pat<(and GR32:$src1, GR32:$src2), (AND32rr GR32:$src1, GR32:$src2)>;
1750 def : Pat<(and GR64:$src1, GR64:$src2), (AND64rr GR64:$src1, GR64:$src2)>;
1753 def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
1754 (AND8rm GR8:$src1, addr:$src2)>;
1755 def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
1756 (AND16rm GR16:$src1, addr:$src2)>;
1757 def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
1758 (AND32rm GR32:$src1, addr:$src2)>;
1759 def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
1760 (AND64rm GR64:$src1, addr:$src2)>;
1763 def : Pat<(and GR8:$src1, imm:$src2),
1764 (AND8ri GR8:$src1, imm:$src2)>;
1765 def : Pat<(and GR16:$src1, imm:$src2),
1766 (AND16ri GR16:$src1, imm:$src2)>;
1767 def : Pat<(and GR32:$src1, imm:$src2),
1768 (AND32ri GR32:$src1, imm:$src2)>;
1769 def : Pat<(and GR16:$src1, i16immSExt8:$src2),
1770 (AND16ri8 GR16:$src1, i16immSExt8:$src2)>;
1771 def : Pat<(and GR32:$src1, i32immSExt8:$src2),
1772 (AND32ri8 GR32:$src1, i32immSExt8:$src2)>;
1773 def : Pat<(and GR64:$src1, i64immSExt8:$src2),
1774 (AND64ri8 GR64:$src1, i64immSExt8:$src2)>;
1775 def : Pat<(and GR64:$src1, i64immSExt32:$src2),
1776 (AND64ri32 GR64:$src1, i64immSExt32:$src2)>;
1778 // Bit scan instruction patterns to match explicit zero-undef behavior.
1779 def : Pat<(cttz_zero_undef GR16:$src), (BSF16rr GR16:$src)>;
1780 def : Pat<(cttz_zero_undef GR32:$src), (BSF32rr GR32:$src)>;
1781 def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr GR64:$src)>;
1782 def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm addr:$src)>;
1783 def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm addr:$src)>;
1784 def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm addr:$src)>;
1786 // When HasMOVBE is enabled it is possible to get a non-legalized
1787 // register-register 16 bit bswap. This maps it to a ROL instruction.
1788 let Predicates = [HasMOVBE] in {
1789 def : Pat<(bswap GR16:$src), (ROL16ri GR16:$src, (i8 8))>;