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());
23 def GetLo8XForm : SDNodeXForm<imm, [{
24 // Transformation function: get the low 8 bits.
25 return getI8Imm((uint8_t)N->getZExtValue());
29 //===----------------------------------------------------------------------===//
30 // Random Pseudo Instructions.
32 // PIC base construction. This expands to code that looks like this:
35 let neverHasSideEffects = 1, 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:$amt),
48 [(X86callseq_start timm:$amt)]>,
49 Requires<[In32BitMode]>;
50 def ADJCALLSTACKUP32 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
52 [(X86callseq_end timm:$amt1, timm:$amt2)]>,
53 Requires<[In32BitMode]>;
56 // ADJCALLSTACKDOWN/UP implicitly use/def RSP because they may be expanded into
57 // a stack adjustment and the codegen must know that they may modify the stack
58 // pointer before prolog-epilog rewriting occurs.
59 // Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
60 // sub / add which can clobber EFLAGS.
61 let Defs = [RSP, EFLAGS], Uses = [RSP] in {
62 def ADJCALLSTACKDOWN64 : I<0, Pseudo, (outs), (ins i32imm:$amt),
64 [(X86callseq_start timm:$amt)]>,
65 Requires<[In64BitMode]>;
66 def ADJCALLSTACKUP64 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
68 [(X86callseq_end timm:$amt1, timm:$amt2)]>,
69 Requires<[In64BitMode]>;
74 // x86-64 va_start lowering magic.
75 let usesCustomInserter = 1 in {
76 def VASTART_SAVE_XMM_REGS : I<0, Pseudo,
79 i64imm:$regsavefi, i64imm:$offset,
81 "#VASTART_SAVE_XMM_REGS $al, $regsavefi, $offset",
82 [(X86vastart_save_xmm_regs GR8:$al,
86 // The VAARG_64 pseudo-instruction takes the address of the va_list,
87 // and places the address of the next argument into a register.
88 let Defs = [EFLAGS] in
89 def VAARG_64 : I<0, Pseudo,
91 (ins i8mem:$ap, i32imm:$size, i8imm:$mode, i32imm:$align),
92 "#VAARG_64 $dst, $ap, $size, $mode, $align",
94 (X86vaarg64 addr:$ap, imm:$size, imm:$mode, imm:$align)),
97 // Dynamic stack allocation yields a _chkstk or _alloca call for all Windows
98 // targets. These calls are needed to probe the stack when allocating more than
99 // 4k bytes in one go. Touching the stack at 4K increments is necessary to
100 // ensure that the guard pages used by the OS virtual memory manager are
101 // allocated in correct sequence.
102 // The main point of having separate instruction are extra unmodelled effects
103 // (compared to ordinary calls) like stack pointer change.
105 let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
106 def WIN_ALLOCA : I<0, Pseudo, (outs), (ins),
107 "# dynamic stack allocation",
110 // When using segmented stacks these are lowered into instructions which first
111 // check if the current stacklet has enough free memory. If it does, memory is
112 // allocated by bumping the stack pointer. Otherwise memory is allocated from
115 let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
116 def SEG_ALLOCA_32 : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$size),
117 "# variable sized alloca for segmented stacks",
119 (X86SegAlloca GR32:$size))]>,
120 Requires<[In32BitMode]>;
122 let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
123 def SEG_ALLOCA_64 : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$size),
124 "# variable sized alloca for segmented stacks",
126 (X86SegAlloca GR64:$size))]>,
127 Requires<[In64BitMode]>;
133 //===----------------------------------------------------------------------===//
134 // EH Pseudo Instructions
136 let isTerminator = 1, isReturn = 1, isBarrier = 1,
137 hasCtrlDep = 1, isCodeGenOnly = 1 in {
138 def EH_RETURN : I<0xC3, RawFrm, (outs), (ins GR32:$addr),
139 "ret\t#eh_return, addr: $addr",
140 [(X86ehret GR32:$addr)]>;
144 let isTerminator = 1, isReturn = 1, isBarrier = 1,
145 hasCtrlDep = 1, isCodeGenOnly = 1 in {
146 def EH_RETURN64 : I<0xC3, RawFrm, (outs), (ins GR64:$addr),
147 "ret\t#eh_return, addr: $addr",
148 [(X86ehret GR64:$addr)]>;
152 //===----------------------------------------------------------------------===//
153 // Pseudo instructions used by segmented stacks.
156 // This is lowered into a RET instruction by MCInstLower. We need
157 // this so that we don't have to have a MachineBasicBlock which ends
158 // with a RET and also has successors.
159 let isPseudo = 1 in {
160 def MORESTACK_RET: I<0, Pseudo, (outs), (ins),
163 // This instruction is lowered to a RET followed by a MOV. The two
164 // instructions are not generated on a higher level since then the
165 // verifier sees a MachineBasicBlock ending with a non-terminator.
166 def MORESTACK_RET_RESTORE_R10 : I<0, Pseudo, (outs), (ins),
170 //===----------------------------------------------------------------------===//
171 // Alias Instructions
172 //===----------------------------------------------------------------------===//
174 // Alias instructions that map movr0 to xor.
175 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
176 // FIXME: Set encoding to pseudo.
177 let Defs = [EFLAGS], isReMaterializable = 1, isAsCheapAsAMove = 1,
178 isCodeGenOnly = 1 in {
179 def MOV8r0 : I<0x30, MRMInitReg, (outs GR8 :$dst), (ins), "",
180 [(set GR8:$dst, 0)]>;
182 // We want to rewrite MOV16r0 in terms of MOV32r0, because it's a smaller
183 // encoding and avoids a partial-register update sometimes, but doing so
184 // at isel time interferes with rematerialization in the current register
185 // allocator. For now, this is rewritten when the instruction is lowered
187 def MOV16r0 : I<0x31, MRMInitReg, (outs GR16:$dst), (ins),
189 [(set GR16:$dst, 0)]>, OpSize;
191 // FIXME: Set encoding to pseudo.
192 def MOV32r0 : I<0x31, MRMInitReg, (outs GR32:$dst), (ins), "",
193 [(set GR32:$dst, 0)]>;
196 // We want to rewrite MOV64r0 in terms of MOV32r0, because it's sometimes a
197 // smaller encoding, but doing so at isel time interferes with rematerialization
198 // in the current register allocator. For now, this is rewritten when the
199 // instruction is lowered to an MCInst.
200 // FIXME: AddedComplexity gives this a higher priority than MOV64ri32. Remove
201 // when we have a better way to specify isel priority.
202 let Defs = [EFLAGS], isCodeGenOnly=1,
203 AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1 in
204 def MOV64r0 : I<0x31, MRMInitReg, (outs GR64:$dst), (ins), "",
205 [(set GR64:$dst, 0)]>;
207 // Materialize i64 constant where top 32-bits are zero. This could theoretically
208 // use MOV32ri with a SUBREG_TO_REG to represent the zero-extension, however
209 // that would make it more difficult to rematerialize.
210 let AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1,
212 def MOV64ri64i32 : Ii32<0xB8, AddRegFrm, (outs GR64:$dst), (ins i64i32imm:$src),
213 "", [(set GR64:$dst, i64immZExt32:$src)]>;
215 // Use sbb to materialize carry bit.
216 let Uses = [EFLAGS], Defs = [EFLAGS], isCodeGenOnly = 1 in {
217 // FIXME: These are pseudo ops that should be replaced with Pat<> patterns.
218 // However, Pat<> can't replicate the destination reg into the inputs of the
220 // FIXME: Change these to have encoding Pseudo when X86MCCodeEmitter replaces
222 def SETB_C8r : I<0x18, MRMInitReg, (outs GR8:$dst), (ins), "",
223 [(set GR8:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
224 def SETB_C16r : I<0x19, MRMInitReg, (outs GR16:$dst), (ins), "",
225 [(set GR16:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>,
227 def SETB_C32r : I<0x19, MRMInitReg, (outs GR32:$dst), (ins), "",
228 [(set GR32:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
229 def SETB_C64r : RI<0x19, MRMInitReg, (outs GR64:$dst), (ins), "",
230 [(set GR64:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
234 def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
236 def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
238 def : Pat<(i64 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
241 def : Pat<(i16 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
243 def : Pat<(i32 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
245 def : Pat<(i64 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
248 // We canonicalize 'setb' to "(and (sbb reg,reg), 1)" on the hope that the and
249 // will be eliminated and that the sbb can be extended up to a wider type. When
250 // this happens, it is great. However, if we are left with an 8-bit sbb and an
251 // and, we might as well just match it as a setb.
252 def : Pat<(and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1),
255 // (add OP, SETB) -> (adc OP, 0)
256 def : Pat<(add (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR8:$op),
257 (ADC8ri GR8:$op, 0)>;
258 def : Pat<(add (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR32:$op),
259 (ADC32ri8 GR32:$op, 0)>;
260 def : Pat<(add (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR64:$op),
261 (ADC64ri8 GR64:$op, 0)>;
263 // (sub OP, SETB) -> (sbb OP, 0)
264 def : Pat<(sub GR8:$op, (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
265 (SBB8ri GR8:$op, 0)>;
266 def : Pat<(sub GR32:$op, (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
267 (SBB32ri8 GR32:$op, 0)>;
268 def : Pat<(sub GR64:$op, (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
269 (SBB64ri8 GR64:$op, 0)>;
271 // (sub OP, SETCC_CARRY) -> (adc OP, 0)
272 def : Pat<(sub GR8:$op, (i8 (X86setcc_c X86_COND_B, EFLAGS))),
273 (ADC8ri GR8:$op, 0)>;
274 def : Pat<(sub GR32:$op, (i32 (X86setcc_c X86_COND_B, EFLAGS))),
275 (ADC32ri8 GR32:$op, 0)>;
276 def : Pat<(sub GR64:$op, (i64 (X86setcc_c X86_COND_B, EFLAGS))),
277 (ADC64ri8 GR64:$op, 0)>;
279 //===----------------------------------------------------------------------===//
280 // String Pseudo Instructions
282 let Defs = [ECX,EDI,ESI], Uses = [ECX,EDI,ESI], isCodeGenOnly = 1 in {
283 def REP_MOVSB : I<0xA4, RawFrm, (outs), (ins), "{rep;movsb|rep movsb}",
284 [(X86rep_movs i8)]>, REP;
285 def REP_MOVSW : I<0xA5, RawFrm, (outs), (ins), "{rep;movsw|rep movsw}",
286 [(X86rep_movs i16)]>, REP, OpSize;
287 def REP_MOVSD : I<0xA5, RawFrm, (outs), (ins), "{rep;movsl|rep movsd}",
288 [(X86rep_movs i32)]>, REP;
291 let Defs = [RCX,RDI,RSI], Uses = [RCX,RDI,RSI], isCodeGenOnly = 1 in
292 def REP_MOVSQ : RI<0xA5, RawFrm, (outs), (ins), "{rep;movsq|rep movsq}",
293 [(X86rep_movs i64)]>, REP;
296 // FIXME: Should use "(X86rep_stos AL)" as the pattern.
297 let Defs = [ECX,EDI], Uses = [AL,ECX,EDI], isCodeGenOnly = 1 in
298 def REP_STOSB : I<0xAA, RawFrm, (outs), (ins), "{rep;stosb|rep stosb}",
299 [(X86rep_stos i8)]>, REP;
300 let Defs = [ECX,EDI], Uses = [AX,ECX,EDI], isCodeGenOnly = 1 in
301 def REP_STOSW : I<0xAB, RawFrm, (outs), (ins), "{rep;stosw|rep stosw}",
302 [(X86rep_stos i16)]>, REP, OpSize;
303 let Defs = [ECX,EDI], Uses = [EAX,ECX,EDI], isCodeGenOnly = 1 in
304 def REP_STOSD : I<0xAB, RawFrm, (outs), (ins), "{rep;stosl|rep stosd}",
305 [(X86rep_stos i32)]>, REP;
307 let Defs = [RCX,RDI], Uses = [RAX,RCX,RDI], isCodeGenOnly = 1 in
308 def REP_STOSQ : RI<0xAB, RawFrm, (outs), (ins), "{rep;stosq|rep stosq}",
309 [(X86rep_stos i64)]>, REP;
312 //===----------------------------------------------------------------------===//
313 // Thread Local Storage Instructions
317 // All calls clobber the non-callee saved registers. ESP is marked as
318 // a use to prevent stack-pointer assignments that appear immediately
319 // before calls from potentially appearing dead.
320 let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0,
321 MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
322 XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
323 XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
325 def TLS_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
327 [(X86tlsaddr tls32addr:$sym)]>,
328 Requires<[In32BitMode]>;
330 // All calls clobber the non-callee saved registers. RSP is marked as
331 // a use to prevent stack-pointer assignments that appear immediately
332 // before calls from potentially appearing dead.
333 let Defs = [RAX, RCX, RDX, RSI, RDI, R8, R9, R10, R11,
334 FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0, ST1,
335 MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
336 XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
337 XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
339 def TLS_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
341 [(X86tlsaddr tls64addr:$sym)]>,
342 Requires<[In64BitMode]>;
344 // Darwin TLS Support
345 // For i386, the address of the thunk is passed on the stack, on return the
346 // address of the variable is in %eax. %ecx is trashed during the function
347 // call. All other registers are preserved.
348 let Defs = [EAX, ECX, EFLAGS],
350 usesCustomInserter = 1 in
351 def TLSCall_32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
353 [(X86TLSCall addr:$sym)]>,
354 Requires<[In32BitMode]>;
356 // For x86_64, the address of the thunk is passed in %rdi, on return
357 // the address of the variable is in %rax. All other registers are preserved.
358 let Defs = [RAX, EFLAGS],
360 usesCustomInserter = 1 in
361 def TLSCall_64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
363 [(X86TLSCall addr:$sym)]>,
364 Requires<[In64BitMode]>;
367 //===----------------------------------------------------------------------===//
368 // Conditional Move Pseudo Instructions
370 // X86 doesn't have 8-bit conditional moves. Use a customInserter to
371 // emit control flow. An alternative to this is to mark i8 SELECT as Promote,
372 // however that requires promoting the operands, and can induce additional
373 // i8 register pressure.
374 let usesCustomInserter = 1, Uses = [EFLAGS] in {
375 def CMOV_GR8 : I<0, Pseudo,
376 (outs GR8:$dst), (ins GR8:$src1, GR8:$src2, i8imm:$cond),
378 [(set GR8:$dst, (X86cmov GR8:$src1, GR8:$src2,
379 imm:$cond, EFLAGS))]>;
381 let Predicates = [NoCMov] in {
382 def CMOV_GR32 : I<0, Pseudo,
383 (outs GR32:$dst), (ins GR32:$src1, GR32:$src2, i8imm:$cond),
384 "#CMOV_GR32* PSEUDO!",
386 (X86cmov GR32:$src1, GR32:$src2, imm:$cond, EFLAGS))]>;
387 def CMOV_GR16 : I<0, Pseudo,
388 (outs GR16:$dst), (ins GR16:$src1, GR16:$src2, i8imm:$cond),
389 "#CMOV_GR16* PSEUDO!",
391 (X86cmov GR16:$src1, GR16:$src2, imm:$cond, EFLAGS))]>;
392 def CMOV_RFP32 : I<0, Pseudo,
394 (ins RFP32:$src1, RFP32:$src2, i8imm:$cond),
395 "#CMOV_RFP32 PSEUDO!",
397 (X86cmov RFP32:$src1, RFP32:$src2, imm:$cond,
399 def CMOV_RFP64 : I<0, Pseudo,
401 (ins RFP64:$src1, RFP64:$src2, i8imm:$cond),
402 "#CMOV_RFP64 PSEUDO!",
404 (X86cmov RFP64:$src1, RFP64:$src2, imm:$cond,
406 def CMOV_RFP80 : I<0, Pseudo,
408 (ins RFP80:$src1, RFP80:$src2, i8imm:$cond),
409 "#CMOV_RFP80 PSEUDO!",
411 (X86cmov RFP80:$src1, RFP80:$src2, imm:$cond,
413 } // Predicates = [NoCMov]
414 } // UsesCustomInserter = 1, Uses = [EFLAGS]
417 //===----------------------------------------------------------------------===//
418 // Atomic Instruction Pseudo Instructions
419 //===----------------------------------------------------------------------===//
421 // Atomic exchange, and, or, xor
422 let Constraints = "$val = $dst", Defs = [EFLAGS],
423 usesCustomInserter = 1 in {
425 def ATOMAND8 : I<0, Pseudo, (outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
427 [(set GR8:$dst, (atomic_load_and_8 addr:$ptr, GR8:$val))]>;
428 def ATOMOR8 : I<0, Pseudo, (outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
430 [(set GR8:$dst, (atomic_load_or_8 addr:$ptr, GR8:$val))]>;
431 def ATOMXOR8 : I<0, Pseudo,(outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
433 [(set GR8:$dst, (atomic_load_xor_8 addr:$ptr, GR8:$val))]>;
434 def ATOMNAND8 : I<0, Pseudo,(outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
435 "#ATOMNAND8 PSEUDO!",
436 [(set GR8:$dst, (atomic_load_nand_8 addr:$ptr, GR8:$val))]>;
438 def ATOMAND16 : I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
439 "#ATOMAND16 PSEUDO!",
440 [(set GR16:$dst, (atomic_load_and_16 addr:$ptr, GR16:$val))]>;
441 def ATOMOR16 : I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
443 [(set GR16:$dst, (atomic_load_or_16 addr:$ptr, GR16:$val))]>;
444 def ATOMXOR16 : I<0, Pseudo,(outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
445 "#ATOMXOR16 PSEUDO!",
446 [(set GR16:$dst, (atomic_load_xor_16 addr:$ptr, GR16:$val))]>;
447 def ATOMNAND16 : I<0, Pseudo,(outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
448 "#ATOMNAND16 PSEUDO!",
449 [(set GR16:$dst, (atomic_load_nand_16 addr:$ptr, GR16:$val))]>;
450 def ATOMMIN16: I<0, Pseudo, (outs GR16:$dst), (ins i16mem:$ptr, GR16:$val),
451 "#ATOMMIN16 PSEUDO!",
452 [(set GR16:$dst, (atomic_load_min_16 addr:$ptr, GR16:$val))]>;
453 def ATOMMAX16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
454 "#ATOMMAX16 PSEUDO!",
455 [(set GR16:$dst, (atomic_load_max_16 addr:$ptr, GR16:$val))]>;
456 def ATOMUMIN16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
457 "#ATOMUMIN16 PSEUDO!",
458 [(set GR16:$dst, (atomic_load_umin_16 addr:$ptr, GR16:$val))]>;
459 def ATOMUMAX16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
460 "#ATOMUMAX16 PSEUDO!",
461 [(set GR16:$dst, (atomic_load_umax_16 addr:$ptr, GR16:$val))]>;
464 def ATOMAND32 : I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
465 "#ATOMAND32 PSEUDO!",
466 [(set GR32:$dst, (atomic_load_and_32 addr:$ptr, GR32:$val))]>;
467 def ATOMOR32 : I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
469 [(set GR32:$dst, (atomic_load_or_32 addr:$ptr, GR32:$val))]>;
470 def ATOMXOR32 : I<0, Pseudo,(outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
471 "#ATOMXOR32 PSEUDO!",
472 [(set GR32:$dst, (atomic_load_xor_32 addr:$ptr, GR32:$val))]>;
473 def ATOMNAND32 : I<0, Pseudo,(outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
474 "#ATOMNAND32 PSEUDO!",
475 [(set GR32:$dst, (atomic_load_nand_32 addr:$ptr, GR32:$val))]>;
476 def ATOMMIN32: I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$ptr, GR32:$val),
477 "#ATOMMIN32 PSEUDO!",
478 [(set GR32:$dst, (atomic_load_min_32 addr:$ptr, GR32:$val))]>;
479 def ATOMMAX32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
480 "#ATOMMAX32 PSEUDO!",
481 [(set GR32:$dst, (atomic_load_max_32 addr:$ptr, GR32:$val))]>;
482 def ATOMUMIN32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
483 "#ATOMUMIN32 PSEUDO!",
484 [(set GR32:$dst, (atomic_load_umin_32 addr:$ptr, GR32:$val))]>;
485 def ATOMUMAX32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
486 "#ATOMUMAX32 PSEUDO!",
487 [(set GR32:$dst, (atomic_load_umax_32 addr:$ptr, GR32:$val))]>;
491 def ATOMAND64 : I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
492 "#ATOMAND64 PSEUDO!",
493 [(set GR64:$dst, (atomic_load_and_64 addr:$ptr, GR64:$val))]>;
494 def ATOMOR64 : I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
496 [(set GR64:$dst, (atomic_load_or_64 addr:$ptr, GR64:$val))]>;
497 def ATOMXOR64 : I<0, Pseudo,(outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
498 "#ATOMXOR64 PSEUDO!",
499 [(set GR64:$dst, (atomic_load_xor_64 addr:$ptr, GR64:$val))]>;
500 def ATOMNAND64 : I<0, Pseudo,(outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
501 "#ATOMNAND64 PSEUDO!",
502 [(set GR64:$dst, (atomic_load_nand_64 addr:$ptr, GR64:$val))]>;
503 def ATOMMIN64: I<0, Pseudo, (outs GR64:$dst), (ins i64mem:$ptr, GR64:$val),
504 "#ATOMMIN64 PSEUDO!",
505 [(set GR64:$dst, (atomic_load_min_64 addr:$ptr, GR64:$val))]>;
506 def ATOMMAX64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
507 "#ATOMMAX64 PSEUDO!",
508 [(set GR64:$dst, (atomic_load_max_64 addr:$ptr, GR64:$val))]>;
509 def ATOMUMIN64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
510 "#ATOMUMIN64 PSEUDO!",
511 [(set GR64:$dst, (atomic_load_umin_64 addr:$ptr, GR64:$val))]>;
512 def ATOMUMAX64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
513 "#ATOMUMAX64 PSEUDO!",
514 [(set GR64:$dst, (atomic_load_umax_64 addr:$ptr, GR64:$val))]>;
517 let Constraints = "$val1 = $dst1, $val2 = $dst2",
518 Defs = [EFLAGS, EAX, EBX, ECX, EDX],
519 Uses = [EAX, EBX, ECX, EDX],
520 mayLoad = 1, mayStore = 1,
521 usesCustomInserter = 1 in {
522 def ATOMAND6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
523 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
524 "#ATOMAND6432 PSEUDO!", []>;
525 def ATOMOR6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
526 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
527 "#ATOMOR6432 PSEUDO!", []>;
528 def ATOMXOR6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
529 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
530 "#ATOMXOR6432 PSEUDO!", []>;
531 def ATOMNAND6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
532 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
533 "#ATOMNAND6432 PSEUDO!", []>;
534 def ATOMADD6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
535 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
536 "#ATOMADD6432 PSEUDO!", []>;
537 def ATOMSUB6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
538 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
539 "#ATOMSUB6432 PSEUDO!", []>;
540 def ATOMSWAP6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
541 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
542 "#ATOMSWAP6432 PSEUDO!", []>;
545 //===----------------------------------------------------------------------===//
546 // Normal-Instructions-With-Lock-Prefix Pseudo Instructions
547 //===----------------------------------------------------------------------===//
549 // FIXME: Use normal instructions and add lock prefix dynamically.
553 // TODO: Get this to fold the constant into the instruction.
554 let isCodeGenOnly = 1 in
555 def OR32mrLocked : I<0x09, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$zero),
557 "or{l}\t{$zero, $dst|$dst, $zero}",
558 []>, Requires<[In32BitMode]>, LOCK;
560 let hasSideEffects = 1 in
561 def Int_MemBarrier : I<0, Pseudo, (outs), (ins),
565 // TODO: Get this to fold the constant into the instruction.
566 let hasSideEffects = 1, Defs = [ESP], isCodeGenOnly = 1 in
567 def Int_MemBarrierNoSSE64 : RI<0x09, MRM1r, (outs), (ins GR64:$zero),
569 "or{q}\t{$zero, (%rsp)|(%rsp), $zero}",
570 [(X86MemBarrierNoSSE GR64:$zero)]>,
571 Requires<[In64BitMode]>, LOCK;
574 // RegOpc corresponds to the mr version of the instruction
575 // ImmOpc corresponds to the mi version of the instruction
576 // ImmOpc8 corresponds to the mi8 version of the instruction
577 // ImmMod corresponds to the instruction format of the mi and mi8 versions
578 multiclass LOCK_ArithBinOp<bits<8> RegOpc, bits<8> ImmOpc, bits<8> ImmOpc8,
579 Format ImmMod, string mnemonic> {
580 let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1 in {
582 def #NAME#8mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
583 RegOpc{3}, RegOpc{2}, RegOpc{1}, 0 },
584 MRMDestMem, (outs), (ins i8mem:$dst, GR8:$src2),
585 !strconcat("lock\n\t", mnemonic, "{b}\t",
586 "{$src2, $dst|$dst, $src2}"),
588 def #NAME#16mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
589 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
590 MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
591 !strconcat("lock\n\t", mnemonic, "{w}\t",
592 "{$src2, $dst|$dst, $src2}"),
594 def #NAME#32mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
595 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
596 MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
597 !strconcat("lock\n\t", mnemonic, "{l}\t",
598 "{$src2, $dst|$dst, $src2}"),
600 def #NAME#64mr : RI<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
601 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
602 MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
603 !strconcat("lock\n\t", mnemonic, "{q}\t",
604 "{$src2, $dst|$dst, $src2}"),
607 def #NAME#8mi : Ii8<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
608 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 0 },
609 ImmMod, (outs), (ins i8mem :$dst, i8imm :$src2),
610 !strconcat("lock\n\t", mnemonic, "{b}\t",
611 "{$src2, $dst|$dst, $src2}"),
614 def #NAME#16mi : Ii16<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
615 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
616 ImmMod, (outs), (ins i16mem :$dst, i16imm :$src2),
617 !strconcat("lock\n\t", mnemonic, "{w}\t",
618 "{$src2, $dst|$dst, $src2}"),
621 def #NAME#32mi : Ii32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
622 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
623 ImmMod, (outs), (ins i32mem :$dst, i32imm :$src2),
624 !strconcat("lock\n\t", mnemonic, "{l}\t",
625 "{$src2, $dst|$dst, $src2}"),
628 def #NAME#64mi32 : RIi32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
629 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
630 ImmMod, (outs), (ins i64mem :$dst, i64i32imm :$src2),
631 !strconcat("lock\n\t", mnemonic, "{q}\t",
632 "{$src2, $dst|$dst, $src2}"),
635 def #NAME#16mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
636 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
637 ImmMod, (outs), (ins i16mem :$dst, i16i8imm :$src2),
638 !strconcat("lock\n\t", mnemonic, "{w}\t",
639 "{$src2, $dst|$dst, $src2}"),
641 def #NAME#32mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
642 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
643 ImmMod, (outs), (ins i32mem :$dst, i32i8imm :$src2),
644 !strconcat("lock\n\t", mnemonic, "{l}\t",
645 "{$src2, $dst|$dst, $src2}"),
647 def #NAME#64mi8 : RIi8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
648 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
649 ImmMod, (outs), (ins i64mem :$dst, i64i8imm :$src2),
650 !strconcat("lock\n\t", mnemonic, "{q}\t",
651 "{$src2, $dst|$dst, $src2}"),
658 defm LOCK_ADD : LOCK_ArithBinOp<0x00, 0x80, 0x83, MRM0m, "add">;
659 defm LOCK_SUB : LOCK_ArithBinOp<0x28, 0x80, 0x83, MRM5m, "sub">;
660 defm LOCK_OR : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM1m, "or">;
661 defm LOCK_AND : LOCK_ArithBinOp<0x20, 0x80, 0x83, MRM4m, "and">;
662 defm LOCK_XOR : LOCK_ArithBinOp<0x30, 0x80, 0x83, MRM6m, "xor">;
664 // Optimized codegen when the non-memory output is not used.
665 let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1 in {
667 def LOCK_INC8m : I<0xFE, MRM0m, (outs), (ins i8mem :$dst),
669 "inc{b}\t$dst", []>, LOCK;
670 def LOCK_INC16m : I<0xFF, MRM0m, (outs), (ins i16mem:$dst),
672 "inc{w}\t$dst", []>, OpSize, LOCK;
673 def LOCK_INC32m : I<0xFF, MRM0m, (outs), (ins i32mem:$dst),
675 "inc{l}\t$dst", []>, LOCK;
676 def LOCK_INC64m : RI<0xFF, MRM0m, (outs), (ins i64mem:$dst),
678 "inc{q}\t$dst", []>, LOCK;
680 def LOCK_DEC8m : I<0xFE, MRM1m, (outs), (ins i8mem :$dst),
682 "dec{b}\t$dst", []>, LOCK;
683 def LOCK_DEC16m : I<0xFF, MRM1m, (outs), (ins i16mem:$dst),
685 "dec{w}\t$dst", []>, OpSize, LOCK;
686 def LOCK_DEC32m : I<0xFF, MRM1m, (outs), (ins i32mem:$dst),
688 "dec{l}\t$dst", []>, LOCK;
689 def LOCK_DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst),
691 "dec{q}\t$dst", []>, LOCK;
694 // Atomic compare and swap.
695 let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
697 def LCMPXCHG8B : I<0xC7, MRM1m, (outs), (ins i64mem:$ptr),
700 [(X86cas8 addr:$ptr)]>, TB, LOCK;
702 let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RBX, RCX, RDX],
704 def LCMPXCHG16B : RI<0xC7, MRM1m, (outs), (ins i128mem:$ptr),
707 [(X86cas16 addr:$ptr)]>, TB, LOCK,
708 Requires<[HasCmpxchg16b]>;
710 let Defs = [AL, EFLAGS], Uses = [AL], isCodeGenOnly = 1 in {
711 def LCMPXCHG8 : I<0xB0, MRMDestMem, (outs), (ins i8mem:$ptr, GR8:$swap),
713 "cmpxchg{b}\t{$swap, $ptr|$ptr, $swap}",
714 [(X86cas addr:$ptr, GR8:$swap, 1)]>, TB, LOCK;
717 let Defs = [AX, EFLAGS], Uses = [AX], isCodeGenOnly = 1 in {
718 def LCMPXCHG16 : I<0xB1, MRMDestMem, (outs), (ins i16mem:$ptr, GR16:$swap),
720 "cmpxchg{w}\t{$swap, $ptr|$ptr, $swap}",
721 [(X86cas addr:$ptr, GR16:$swap, 2)]>, TB, OpSize, LOCK;
724 let Defs = [EAX, EFLAGS], Uses = [EAX], isCodeGenOnly = 1 in {
725 def LCMPXCHG32 : I<0xB1, MRMDestMem, (outs), (ins i32mem:$ptr, GR32:$swap),
727 "cmpxchg{l}\t{$swap, $ptr|$ptr, $swap}",
728 [(X86cas addr:$ptr, GR32:$swap, 4)]>, TB, LOCK;
731 let Defs = [RAX, EFLAGS], Uses = [RAX], isCodeGenOnly = 1 in {
732 def LCMPXCHG64 : RI<0xB1, MRMDestMem, (outs), (ins i64mem:$ptr, GR64:$swap),
734 "cmpxchg{q}\t{$swap, $ptr|$ptr, $swap}",
735 [(X86cas addr:$ptr, GR64:$swap, 8)]>, TB, LOCK;
738 // Atomic exchange and add
739 let Constraints = "$val = $dst", Defs = [EFLAGS], isCodeGenOnly = 1 in {
740 def LXADD8 : I<0xC0, MRMSrcMem, (outs GR8:$dst), (ins GR8:$val, i8mem:$ptr),
742 "xadd{b}\t{$val, $ptr|$ptr, $val}",
743 [(set GR8:$dst, (atomic_load_add_8 addr:$ptr, GR8:$val))]>,
745 def LXADD16 : I<0xC1, MRMSrcMem, (outs GR16:$dst), (ins GR16:$val, i16mem:$ptr),
747 "xadd{w}\t{$val, $ptr|$ptr, $val}",
748 [(set GR16:$dst, (atomic_load_add_16 addr:$ptr, GR16:$val))]>,
750 def LXADD32 : I<0xC1, MRMSrcMem, (outs GR32:$dst), (ins GR32:$val, i32mem:$ptr),
752 "xadd{l}\t{$val, $ptr|$ptr, $val}",
753 [(set GR32:$dst, (atomic_load_add_32 addr:$ptr, GR32:$val))]>,
755 def LXADD64 : RI<0xC1, MRMSrcMem, (outs GR64:$dst), (ins GR64:$val,i64mem:$ptr),
757 "xadd{q}\t{$val, $ptr|$ptr, $val}",
758 [(set GR64:$dst, (atomic_load_add_64 addr:$ptr, GR64:$val))]>,
762 def ACQUIRE_MOV8rm : I<0, Pseudo, (outs GR8 :$dst), (ins i8mem :$src),
763 "#ACQUIRE_MOV PSEUDO!",
764 [(set GR8:$dst, (atomic_load_8 addr:$src))]>;
765 def ACQUIRE_MOV16rm : I<0, Pseudo, (outs GR16:$dst), (ins i16mem:$src),
766 "#ACQUIRE_MOV PSEUDO!",
767 [(set GR16:$dst, (atomic_load_16 addr:$src))]>;
768 def ACQUIRE_MOV32rm : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$src),
769 "#ACQUIRE_MOV PSEUDO!",
770 [(set GR32:$dst, (atomic_load_32 addr:$src))]>;
771 def ACQUIRE_MOV64rm : I<0, Pseudo, (outs GR64:$dst), (ins i64mem:$src),
772 "#ACQUIRE_MOV PSEUDO!",
773 [(set GR64:$dst, (atomic_load_64 addr:$src))]>;
775 def RELEASE_MOV8mr : I<0, Pseudo, (outs), (ins i8mem :$dst, GR8 :$src),
776 "#RELEASE_MOV PSEUDO!",
777 [(atomic_store_8 addr:$dst, GR8 :$src)]>;
778 def RELEASE_MOV16mr : I<0, Pseudo, (outs), (ins i16mem:$dst, GR16:$src),
779 "#RELEASE_MOV PSEUDO!",
780 [(atomic_store_16 addr:$dst, GR16:$src)]>;
781 def RELEASE_MOV32mr : I<0, Pseudo, (outs), (ins i32mem:$dst, GR32:$src),
782 "#RELEASE_MOV PSEUDO!",
783 [(atomic_store_32 addr:$dst, GR32:$src)]>;
784 def RELEASE_MOV64mr : I<0, Pseudo, (outs), (ins i64mem:$dst, GR64:$src),
785 "#RELEASE_MOV PSEUDO!",
786 [(atomic_store_64 addr:$dst, GR64:$src)]>;
788 //===----------------------------------------------------------------------===//
789 // Conditional Move Pseudo Instructions.
790 //===----------------------------------------------------------------------===//
793 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded after
794 // instruction selection into a branch sequence.
795 let Uses = [EFLAGS], usesCustomInserter = 1 in {
796 def CMOV_FR32 : I<0, Pseudo,
797 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
798 "#CMOV_FR32 PSEUDO!",
799 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
801 def CMOV_FR64 : I<0, Pseudo,
802 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
803 "#CMOV_FR64 PSEUDO!",
804 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
806 def CMOV_V4F32 : I<0, Pseudo,
807 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
808 "#CMOV_V4F32 PSEUDO!",
810 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
812 def CMOV_V2F64 : I<0, Pseudo,
813 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
814 "#CMOV_V2F64 PSEUDO!",
816 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
818 def CMOV_V2I64 : I<0, Pseudo,
819 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
820 "#CMOV_V2I64 PSEUDO!",
822 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
824 def CMOV_V8F32 : I<0, Pseudo,
825 (outs VR256:$dst), (ins VR256:$t, VR256:$f, i8imm:$cond),
826 "#CMOV_V8F32 PSEUDO!",
828 (v8f32 (X86cmov VR256:$t, VR256:$f, imm:$cond,
830 def CMOV_V4F64 : I<0, Pseudo,
831 (outs VR256:$dst), (ins VR256:$t, VR256:$f, i8imm:$cond),
832 "#CMOV_V4F64 PSEUDO!",
834 (v4f64 (X86cmov VR256:$t, VR256:$f, imm:$cond,
836 def CMOV_V4I64 : I<0, Pseudo,
837 (outs VR256:$dst), (ins VR256:$t, VR256:$f, i8imm:$cond),
838 "#CMOV_V4I64 PSEUDO!",
840 (v4i64 (X86cmov VR256:$t, VR256:$f, imm:$cond,
845 //===----------------------------------------------------------------------===//
846 // DAG Pattern Matching Rules
847 //===----------------------------------------------------------------------===//
849 // ConstantPool GlobalAddress, ExternalSymbol, and JumpTable
850 def : Pat<(i32 (X86Wrapper tconstpool :$dst)), (MOV32ri tconstpool :$dst)>;
851 def : Pat<(i32 (X86Wrapper tjumptable :$dst)), (MOV32ri tjumptable :$dst)>;
852 def : Pat<(i32 (X86Wrapper tglobaltlsaddr:$dst)),(MOV32ri tglobaltlsaddr:$dst)>;
853 def : Pat<(i32 (X86Wrapper tglobaladdr :$dst)), (MOV32ri tglobaladdr :$dst)>;
854 def : Pat<(i32 (X86Wrapper texternalsym:$dst)), (MOV32ri texternalsym:$dst)>;
855 def : Pat<(i32 (X86Wrapper tblockaddress:$dst)), (MOV32ri tblockaddress:$dst)>;
857 def : Pat<(add GR32:$src1, (X86Wrapper tconstpool:$src2)),
858 (ADD32ri GR32:$src1, tconstpool:$src2)>;
859 def : Pat<(add GR32:$src1, (X86Wrapper tjumptable:$src2)),
860 (ADD32ri GR32:$src1, tjumptable:$src2)>;
861 def : Pat<(add GR32:$src1, (X86Wrapper tglobaladdr :$src2)),
862 (ADD32ri GR32:$src1, tglobaladdr:$src2)>;
863 def : Pat<(add GR32:$src1, (X86Wrapper texternalsym:$src2)),
864 (ADD32ri GR32:$src1, texternalsym:$src2)>;
865 def : Pat<(add GR32:$src1, (X86Wrapper tblockaddress:$src2)),
866 (ADD32ri GR32:$src1, tblockaddress:$src2)>;
868 def : Pat<(store (i32 (X86Wrapper tglobaladdr:$src)), addr:$dst),
869 (MOV32mi addr:$dst, tglobaladdr:$src)>;
870 def : Pat<(store (i32 (X86Wrapper texternalsym:$src)), addr:$dst),
871 (MOV32mi addr:$dst, texternalsym:$src)>;
872 def : Pat<(store (i32 (X86Wrapper tblockaddress:$src)), addr:$dst),
873 (MOV32mi addr:$dst, tblockaddress:$src)>;
877 // ConstantPool GlobalAddress, ExternalSymbol, and JumpTable when not in small
878 // code model mode, should use 'movabs'. FIXME: This is really a hack, the
879 // 'movabs' predicate should handle this sort of thing.
880 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
881 (MOV64ri tconstpool :$dst)>, Requires<[FarData]>;
882 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
883 (MOV64ri tjumptable :$dst)>, Requires<[FarData]>;
884 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
885 (MOV64ri tglobaladdr :$dst)>, Requires<[FarData]>;
886 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
887 (MOV64ri texternalsym:$dst)>, Requires<[FarData]>;
888 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
889 (MOV64ri tblockaddress:$dst)>, Requires<[FarData]>;
891 // In static codegen with small code model, we can get the address of a label
892 // into a register with 'movl'. FIXME: This is a hack, the 'imm' predicate of
893 // the MOV64ri64i32 should accept these.
894 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
895 (MOV64ri64i32 tconstpool :$dst)>, Requires<[SmallCode]>;
896 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
897 (MOV64ri64i32 tjumptable :$dst)>, Requires<[SmallCode]>;
898 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
899 (MOV64ri64i32 tglobaladdr :$dst)>, Requires<[SmallCode]>;
900 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
901 (MOV64ri64i32 texternalsym:$dst)>, Requires<[SmallCode]>;
902 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
903 (MOV64ri64i32 tblockaddress:$dst)>, Requires<[SmallCode]>;
905 // In kernel code model, we can get the address of a label
906 // into a register with 'movq'. FIXME: This is a hack, the 'imm' predicate of
907 // the MOV64ri32 should accept these.
908 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
909 (MOV64ri32 tconstpool :$dst)>, Requires<[KernelCode]>;
910 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
911 (MOV64ri32 tjumptable :$dst)>, Requires<[KernelCode]>;
912 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
913 (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
914 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
915 (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
916 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
917 (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
919 // If we have small model and -static mode, it is safe to store global addresses
920 // directly as immediates. FIXME: This is really a hack, the 'imm' predicate
921 // for MOV64mi32 should handle this sort of thing.
922 def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
923 (MOV64mi32 addr:$dst, tconstpool:$src)>,
924 Requires<[NearData, IsStatic]>;
925 def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
926 (MOV64mi32 addr:$dst, tjumptable:$src)>,
927 Requires<[NearData, IsStatic]>;
928 def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
929 (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
930 Requires<[NearData, IsStatic]>;
931 def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
932 (MOV64mi32 addr:$dst, texternalsym:$src)>,
933 Requires<[NearData, IsStatic]>;
934 def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
935 (MOV64mi32 addr:$dst, tblockaddress:$src)>,
936 Requires<[NearData, IsStatic]>;
942 // tls has some funny stuff here...
943 // This corresponds to movabs $foo@tpoff, %rax
944 def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
945 (MOV64ri tglobaltlsaddr :$dst)>;
946 // This corresponds to add $foo@tpoff, %rax
947 def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
948 (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
949 // This corresponds to mov foo@tpoff(%rbx), %eax
950 def : Pat<(load (i64 (X86Wrapper tglobaltlsaddr :$dst))),
951 (MOV64rm tglobaltlsaddr :$dst)>;
954 // Direct PC relative function call for small code model. 32-bit displacement
955 // sign extended to 64-bit.
956 def : Pat<(X86call (i64 tglobaladdr:$dst)),
957 (CALL64pcrel32 tglobaladdr:$dst)>, Requires<[NotWin64]>;
958 def : Pat<(X86call (i64 texternalsym:$dst)),
959 (CALL64pcrel32 texternalsym:$dst)>, Requires<[NotWin64]>;
961 def : Pat<(X86call (i64 tglobaladdr:$dst)),
962 (WINCALL64pcrel32 tglobaladdr:$dst)>, Requires<[IsWin64]>;
963 def : Pat<(X86call (i64 texternalsym:$dst)),
964 (WINCALL64pcrel32 texternalsym:$dst)>, Requires<[IsWin64]>;
967 def : Pat<(X86tcret GR32_TC:$dst, imm:$off),
968 (TCRETURNri GR32_TC:$dst, imm:$off)>,
969 Requires<[In32BitMode]>;
971 // FIXME: This is disabled for 32-bit PIC mode because the global base
972 // register which is part of the address mode may be assigned a
973 // callee-saved register.
974 def : Pat<(X86tcret (load addr:$dst), imm:$off),
975 (TCRETURNmi addr:$dst, imm:$off)>,
976 Requires<[In32BitMode, IsNotPIC]>;
978 def : Pat<(X86tcret (i32 tglobaladdr:$dst), imm:$off),
979 (TCRETURNdi texternalsym:$dst, imm:$off)>,
980 Requires<[In32BitMode]>;
982 def : Pat<(X86tcret (i32 texternalsym:$dst), imm:$off),
983 (TCRETURNdi texternalsym:$dst, imm:$off)>,
984 Requires<[In32BitMode]>;
986 def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
987 (TCRETURNri64 ptr_rc_tailcall:$dst, imm:$off)>,
988 Requires<[In64BitMode]>;
990 def : Pat<(X86tcret (load addr:$dst), imm:$off),
991 (TCRETURNmi64 addr:$dst, imm:$off)>,
992 Requires<[In64BitMode]>;
994 def : Pat<(X86tcret (i64 tglobaladdr:$dst), imm:$off),
995 (TCRETURNdi64 tglobaladdr:$dst, imm:$off)>,
996 Requires<[In64BitMode]>;
998 def : Pat<(X86tcret (i64 texternalsym:$dst), imm:$off),
999 (TCRETURNdi64 texternalsym:$dst, imm:$off)>,
1000 Requires<[In64BitMode]>;
1002 // Normal calls, with various flavors of addresses.
1003 def : Pat<(X86call (i32 tglobaladdr:$dst)),
1004 (CALLpcrel32 tglobaladdr:$dst)>;
1005 def : Pat<(X86call (i32 texternalsym:$dst)),
1006 (CALLpcrel32 texternalsym:$dst)>;
1007 def : Pat<(X86call (i32 imm:$dst)),
1008 (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
1012 // TEST R,R is smaller than CMP R,0
1013 def : Pat<(X86cmp GR8:$src1, 0),
1014 (TEST8rr GR8:$src1, GR8:$src1)>;
1015 def : Pat<(X86cmp GR16:$src1, 0),
1016 (TEST16rr GR16:$src1, GR16:$src1)>;
1017 def : Pat<(X86cmp GR32:$src1, 0),
1018 (TEST32rr GR32:$src1, GR32:$src1)>;
1019 def : Pat<(X86cmp GR64:$src1, 0),
1020 (TEST64rr GR64:$src1, GR64:$src1)>;
1022 // Conditional moves with folded loads with operands swapped and conditions
1024 multiclass CMOVmr<PatLeaf InvertedCond, Instruction Inst16, Instruction Inst32,
1025 Instruction Inst64> {
1026 def : Pat<(X86cmov (loadi16 addr:$src1), GR16:$src2, InvertedCond, EFLAGS),
1027 (Inst16 GR16:$src2, addr:$src1)>;
1028 def : Pat<(X86cmov (loadi32 addr:$src1), GR32:$src2, InvertedCond, EFLAGS),
1029 (Inst32 GR32:$src2, addr:$src1)>;
1030 def : Pat<(X86cmov (loadi64 addr:$src1), GR64:$src2, InvertedCond, EFLAGS),
1031 (Inst64 GR64:$src2, addr:$src1)>;
1034 defm : CMOVmr<X86_COND_B , CMOVAE16rm, CMOVAE32rm, CMOVAE64rm>;
1035 defm : CMOVmr<X86_COND_AE, CMOVB16rm , CMOVB32rm , CMOVB64rm>;
1036 defm : CMOVmr<X86_COND_E , CMOVNE16rm, CMOVNE32rm, CMOVNE64rm>;
1037 defm : CMOVmr<X86_COND_NE, CMOVE16rm , CMOVE32rm , CMOVE64rm>;
1038 defm : CMOVmr<X86_COND_BE, CMOVA16rm , CMOVA32rm , CMOVA64rm>;
1039 defm : CMOVmr<X86_COND_A , CMOVBE16rm, CMOVBE32rm, CMOVBE64rm>;
1040 defm : CMOVmr<X86_COND_L , CMOVGE16rm, CMOVGE32rm, CMOVGE64rm>;
1041 defm : CMOVmr<X86_COND_GE, CMOVL16rm , CMOVL32rm , CMOVL64rm>;
1042 defm : CMOVmr<X86_COND_LE, CMOVG16rm , CMOVG32rm , CMOVG64rm>;
1043 defm : CMOVmr<X86_COND_G , CMOVLE16rm, CMOVLE32rm, CMOVLE64rm>;
1044 defm : CMOVmr<X86_COND_P , CMOVNP16rm, CMOVNP32rm, CMOVNP64rm>;
1045 defm : CMOVmr<X86_COND_NP, CMOVP16rm , CMOVP32rm , CMOVP64rm>;
1046 defm : CMOVmr<X86_COND_S , CMOVNS16rm, CMOVNS32rm, CMOVNS64rm>;
1047 defm : CMOVmr<X86_COND_NS, CMOVS16rm , CMOVS32rm , CMOVS64rm>;
1048 defm : CMOVmr<X86_COND_O , CMOVNO16rm, CMOVNO32rm, CMOVNO64rm>;
1049 defm : CMOVmr<X86_COND_NO, CMOVO16rm , CMOVO32rm , CMOVO64rm>;
1051 // zextload bool -> zextload byte
1052 def : Pat<(zextloadi8i1 addr:$src), (MOV8rm addr:$src)>;
1053 def : Pat<(zextloadi16i1 addr:$src), (MOVZX16rm8 addr:$src)>;
1054 def : Pat<(zextloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1055 def : Pat<(zextloadi64i1 addr:$src), (MOVZX64rm8 addr:$src)>;
1057 // extload bool -> extload byte
1058 // When extloading from 16-bit and smaller memory locations into 64-bit
1059 // registers, use zero-extending loads so that the entire 64-bit register is
1060 // defined, avoiding partial-register updates.
1062 def : Pat<(extloadi8i1 addr:$src), (MOV8rm addr:$src)>;
1063 def : Pat<(extloadi16i1 addr:$src), (MOVZX16rm8 addr:$src)>;
1064 def : Pat<(extloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1065 def : Pat<(extloadi16i8 addr:$src), (MOVZX16rm8 addr:$src)>;
1066 def : Pat<(extloadi32i8 addr:$src), (MOVZX32rm8 addr:$src)>;
1067 def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
1069 def : Pat<(extloadi64i1 addr:$src), (MOVZX64rm8 addr:$src)>;
1070 def : Pat<(extloadi64i8 addr:$src), (MOVZX64rm8 addr:$src)>;
1071 def : Pat<(extloadi64i16 addr:$src), (MOVZX64rm16 addr:$src)>;
1072 // For other extloads, use subregs, since the high contents of the register are
1073 // defined after an extload.
1074 def : Pat<(extloadi64i32 addr:$src),
1075 (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src),
1078 // anyext. Define these to do an explicit zero-extend to
1079 // avoid partial-register updates.
1080 def : Pat<(i16 (anyext GR8 :$src)), (EXTRACT_SUBREG
1081 (MOVZX32rr8 GR8 :$src), sub_16bit)>;
1082 def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8 GR8 :$src)>;
1084 // Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
1085 def : Pat<(i32 (anyext GR16:$src)),
1086 (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
1088 def : Pat<(i64 (anyext GR8 :$src)), (MOVZX64rr8 GR8 :$src)>;
1089 def : Pat<(i64 (anyext GR16:$src)), (MOVZX64rr16 GR16 :$src)>;
1090 def : Pat<(i64 (anyext GR32:$src)),
1091 (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1094 // Any instruction that defines a 32-bit result leaves the high half of the
1095 // register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
1096 // be copying from a truncate. And x86's cmov doesn't do anything if the
1097 // condition is false. But any other 32-bit operation will zero-extend
1099 def def32 : PatLeaf<(i32 GR32:$src), [{
1100 return N->getOpcode() != ISD::TRUNCATE &&
1101 N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
1102 N->getOpcode() != ISD::CopyFromReg &&
1103 N->getOpcode() != X86ISD::CMOV;
1106 // In the case of a 32-bit def that is known to implicitly zero-extend,
1107 // we can use a SUBREG_TO_REG.
1108 def : Pat<(i64 (zext def32:$src)),
1109 (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1111 //===----------------------------------------------------------------------===//
1112 // Pattern match OR as ADD
1113 //===----------------------------------------------------------------------===//
1115 // If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
1116 // 3-addressified into an LEA instruction to avoid copies. However, we also
1117 // want to finally emit these instructions as an or at the end of the code
1118 // generator to make the generated code easier to read. To do this, we select
1119 // into "disjoint bits" pseudo ops.
1121 // Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.
1122 def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
1123 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1124 return CurDAG->MaskedValueIsZero(N->getOperand(0), CN->getAPIntValue());
1126 unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
1127 APInt Mask = APInt::getAllOnesValue(BitWidth);
1128 APInt KnownZero0, KnownOne0;
1129 CurDAG->ComputeMaskedBits(N->getOperand(0), Mask, KnownZero0, KnownOne0, 0);
1130 APInt KnownZero1, KnownOne1;
1131 CurDAG->ComputeMaskedBits(N->getOperand(1), Mask, KnownZero1, KnownOne1, 0);
1132 return (~KnownZero0 & ~KnownZero1) == 0;
1136 // (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1137 let AddedComplexity = 5 in { // Try this before the selecting to OR
1139 let isConvertibleToThreeAddress = 1,
1140 Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1141 let isCommutable = 1 in {
1142 def ADD16rr_DB : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1143 "", // orw/addw REG, REG
1144 [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1145 def ADD32rr_DB : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1146 "", // orl/addl REG, REG
1147 [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1148 def ADD64rr_DB : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1149 "", // orq/addq REG, REG
1150 [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1153 // NOTE: These are order specific, we want the ri8 forms to be listed
1154 // first so that they are slightly preferred to the ri forms.
1156 def ADD16ri8_DB : I<0, Pseudo,
1157 (outs GR16:$dst), (ins GR16:$src1, i16i8imm:$src2),
1158 "", // orw/addw REG, imm8
1159 [(set GR16:$dst,(or_is_add GR16:$src1,i16immSExt8:$src2))]>;
1160 def ADD16ri_DB : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1161 "", // orw/addw REG, imm
1162 [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1164 def ADD32ri8_DB : I<0, Pseudo,
1165 (outs GR32:$dst), (ins GR32:$src1, i32i8imm:$src2),
1166 "", // orl/addl REG, imm8
1167 [(set GR32:$dst,(or_is_add GR32:$src1,i32immSExt8:$src2))]>;
1168 def ADD32ri_DB : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1169 "", // orl/addl REG, imm
1170 [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1173 def ADD64ri8_DB : I<0, Pseudo,
1174 (outs GR64:$dst), (ins GR64:$src1, i64i8imm:$src2),
1175 "", // orq/addq REG, imm8
1176 [(set GR64:$dst, (or_is_add GR64:$src1,
1177 i64immSExt8:$src2))]>;
1178 def ADD64ri32_DB : I<0, Pseudo,
1179 (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1180 "", // orq/addq REG, imm
1181 [(set GR64:$dst, (or_is_add GR64:$src1,
1182 i64immSExt32:$src2))]>;
1184 } // AddedComplexity
1187 //===----------------------------------------------------------------------===//
1189 //===----------------------------------------------------------------------===//
1191 // Odd encoding trick: -128 fits into an 8-bit immediate field while
1192 // +128 doesn't, so in this special case use a sub instead of an add.
1193 def : Pat<(add GR16:$src1, 128),
1194 (SUB16ri8 GR16:$src1, -128)>;
1195 def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1196 (SUB16mi8 addr:$dst, -128)>;
1198 def : Pat<(add GR32:$src1, 128),
1199 (SUB32ri8 GR32:$src1, -128)>;
1200 def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1201 (SUB32mi8 addr:$dst, -128)>;
1203 def : Pat<(add GR64:$src1, 128),
1204 (SUB64ri8 GR64:$src1, -128)>;
1205 def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1206 (SUB64mi8 addr:$dst, -128)>;
1208 // The same trick applies for 32-bit immediate fields in 64-bit
1210 def : Pat<(add GR64:$src1, 0x0000000080000000),
1211 (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1212 def : Pat<(store (add (loadi64 addr:$dst), 0x00000000800000000), addr:$dst),
1213 (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1215 // To avoid needing to materialize an immediate in a register, use a 32-bit and
1216 // with implicit zero-extension instead of a 64-bit and if the immediate has at
1217 // least 32 bits of leading zeros. If in addition the last 32 bits can be
1218 // represented with a sign extension of a 8 bit constant, use that.
1220 def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
1224 (EXTRACT_SUBREG GR64:$src, sub_32bit),
1225 (i32 (GetLo8XForm imm:$imm))),
1228 def : Pat<(and GR64:$src, i64immZExt32:$imm),
1232 (EXTRACT_SUBREG GR64:$src, sub_32bit),
1233 (i32 (GetLo32XForm imm:$imm))),
1237 // r & (2^16-1) ==> movz
1238 def : Pat<(and GR32:$src1, 0xffff),
1239 (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1240 // r & (2^8-1) ==> movz
1241 def : Pat<(and GR32:$src1, 0xff),
1242 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src1,
1245 Requires<[In32BitMode]>;
1246 // r & (2^8-1) ==> movz
1247 def : Pat<(and GR16:$src1, 0xff),
1248 (EXTRACT_SUBREG (MOVZX32rr8 (EXTRACT_SUBREG
1249 (i16 (COPY_TO_REGCLASS GR16:$src1, GR16_ABCD)), sub_8bit)),
1251 Requires<[In32BitMode]>;
1253 // r & (2^32-1) ==> movz
1254 def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1255 (MOVZX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1256 // r & (2^16-1) ==> movz
1257 def : Pat<(and GR64:$src, 0xffff),
1258 (MOVZX64rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit)))>;
1259 // r & (2^8-1) ==> movz
1260 def : Pat<(and GR64:$src, 0xff),
1261 (MOVZX64rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit)))>;
1262 // r & (2^8-1) ==> movz
1263 def : Pat<(and GR32:$src1, 0xff),
1264 (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>,
1265 Requires<[In64BitMode]>;
1266 // r & (2^8-1) ==> movz
1267 def : Pat<(and GR16:$src1, 0xff),
1268 (EXTRACT_SUBREG (MOVZX32rr8 (i8
1269 (EXTRACT_SUBREG GR16:$src1, sub_8bit))), sub_16bit)>,
1270 Requires<[In64BitMode]>;
1273 // sext_inreg patterns
1274 def : Pat<(sext_inreg GR32:$src, i16),
1275 (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1276 def : Pat<(sext_inreg GR32:$src, i8),
1277 (MOVSX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1280 Requires<[In32BitMode]>;
1282 def : Pat<(sext_inreg GR16:$src, i8),
1283 (EXTRACT_SUBREG (i32 (MOVSX32rr8 (EXTRACT_SUBREG
1284 (i32 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)), sub_8bit))),
1286 Requires<[In32BitMode]>;
1288 def : Pat<(sext_inreg GR64:$src, i32),
1289 (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1290 def : Pat<(sext_inreg GR64:$src, i16),
1291 (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1292 def : Pat<(sext_inreg GR64:$src, i8),
1293 (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1294 def : Pat<(sext_inreg GR32:$src, i8),
1295 (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>,
1296 Requires<[In64BitMode]>;
1297 def : Pat<(sext_inreg GR16:$src, i8),
1298 (EXTRACT_SUBREG (MOVSX32rr8
1299 (EXTRACT_SUBREG GR16:$src, sub_8bit)), sub_16bit)>,
1300 Requires<[In64BitMode]>;
1302 // sext, sext_load, zext, zext_load
1303 def: Pat<(i16 (sext GR8:$src)),
1304 (EXTRACT_SUBREG (MOVSX32rr8 GR8:$src), sub_16bit)>;
1305 def: Pat<(sextloadi16i8 addr:$src),
1306 (EXTRACT_SUBREG (MOVSX32rm8 addr:$src), sub_16bit)>;
1307 def: Pat<(i16 (zext GR8:$src)),
1308 (EXTRACT_SUBREG (MOVZX32rr8 GR8:$src), sub_16bit)>;
1309 def: Pat<(zextloadi16i8 addr:$src),
1310 (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1313 def : Pat<(i16 (trunc GR32:$src)),
1314 (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1315 def : Pat<(i8 (trunc GR32:$src)),
1316 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1318 Requires<[In32BitMode]>;
1319 def : Pat<(i8 (trunc GR16:$src)),
1320 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1322 Requires<[In32BitMode]>;
1323 def : Pat<(i32 (trunc GR64:$src)),
1324 (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1325 def : Pat<(i16 (trunc GR64:$src)),
1326 (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1327 def : Pat<(i8 (trunc GR64:$src)),
1328 (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1329 def : Pat<(i8 (trunc GR32:$src)),
1330 (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1331 Requires<[In64BitMode]>;
1332 def : Pat<(i8 (trunc GR16:$src)),
1333 (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1334 Requires<[In64BitMode]>;
1336 // h-register tricks
1337 def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1338 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1340 Requires<[In32BitMode]>;
1341 def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1342 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1344 Requires<[In32BitMode]>;
1345 def : Pat<(srl GR16:$src, (i8 8)),
1348 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1351 Requires<[In32BitMode]>;
1352 def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1353 (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1356 Requires<[In32BitMode]>;
1357 def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1358 (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1361 Requires<[In32BitMode]>;
1362 def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1363 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1366 Requires<[In32BitMode]>;
1367 def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1368 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1371 Requires<[In32BitMode]>;
1373 // h-register tricks.
1374 // For now, be conservative on x86-64 and use an h-register extract only if the
1375 // value is immediately zero-extended or stored, which are somewhat common
1376 // cases. This uses a bunch of code to prevent a register requiring a REX prefix
1377 // from being allocated in the same instruction as the h register, as there's
1378 // currently no way to describe this requirement to the register allocator.
1380 // h-register extract and zero-extend.
1381 def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1385 (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1388 def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1390 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1392 Requires<[In64BitMode]>;
1393 def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1394 (MOVZX32_NOREXrr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1397 Requires<[In64BitMode]>;
1398 def : Pat<(srl GR16:$src, (i8 8)),
1401 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1404 Requires<[In64BitMode]>;
1405 def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1407 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1409 Requires<[In64BitMode]>;
1410 def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1412 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1414 Requires<[In64BitMode]>;
1415 def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1419 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1422 def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1426 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1430 // h-register extract and store.
1431 def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1434 (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1436 def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1439 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1441 Requires<[In64BitMode]>;
1442 def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1445 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1447 Requires<[In64BitMode]>;
1450 // (shl x, 1) ==> (add x, x)
1451 // Note that if x is undef (immediate or otherwise), we could theoretically
1452 // end up with the two uses of x getting different values, producing a result
1453 // where the least significant bit is not 0. However, the probability of this
1454 // happening is considered low enough that this is officially not a
1456 def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr GR8 :$src1, GR8 :$src1)>;
1457 def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1458 def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1459 def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1461 // Helper imms that check if a mask doesn't change significant shift bits.
1462 def immShift32 : ImmLeaf<i8, [{ return CountTrailingOnes_32(Imm) >= 5; }]>;
1463 def immShift64 : ImmLeaf<i8, [{ return CountTrailingOnes_32(Imm) >= 6; }]>;
1465 // (shl x (and y, 31)) ==> (shl x, y)
1466 def : Pat<(shl GR8:$src1, (and CL, immShift32)),
1467 (SHL8rCL GR8:$src1)>;
1468 def : Pat<(shl GR16:$src1, (and CL, immShift32)),
1469 (SHL16rCL GR16:$src1)>;
1470 def : Pat<(shl GR32:$src1, (and CL, immShift32)),
1471 (SHL32rCL GR32:$src1)>;
1472 def : Pat<(store (shl (loadi8 addr:$dst), (and CL, immShift32)), addr:$dst),
1473 (SHL8mCL addr:$dst)>;
1474 def : Pat<(store (shl (loadi16 addr:$dst), (and CL, immShift32)), addr:$dst),
1475 (SHL16mCL addr:$dst)>;
1476 def : Pat<(store (shl (loadi32 addr:$dst), (and CL, immShift32)), addr:$dst),
1477 (SHL32mCL addr:$dst)>;
1479 def : Pat<(srl GR8:$src1, (and CL, immShift32)),
1480 (SHR8rCL GR8:$src1)>;
1481 def : Pat<(srl GR16:$src1, (and CL, immShift32)),
1482 (SHR16rCL GR16:$src1)>;
1483 def : Pat<(srl GR32:$src1, (and CL, immShift32)),
1484 (SHR32rCL GR32:$src1)>;
1485 def : Pat<(store (srl (loadi8 addr:$dst), (and CL, immShift32)), addr:$dst),
1486 (SHR8mCL addr:$dst)>;
1487 def : Pat<(store (srl (loadi16 addr:$dst), (and CL, immShift32)), addr:$dst),
1488 (SHR16mCL addr:$dst)>;
1489 def : Pat<(store (srl (loadi32 addr:$dst), (and CL, immShift32)), addr:$dst),
1490 (SHR32mCL addr:$dst)>;
1492 def : Pat<(sra GR8:$src1, (and CL, immShift32)),
1493 (SAR8rCL GR8:$src1)>;
1494 def : Pat<(sra GR16:$src1, (and CL, immShift32)),
1495 (SAR16rCL GR16:$src1)>;
1496 def : Pat<(sra GR32:$src1, (and CL, immShift32)),
1497 (SAR32rCL GR32:$src1)>;
1498 def : Pat<(store (sra (loadi8 addr:$dst), (and CL, immShift32)), addr:$dst),
1499 (SAR8mCL addr:$dst)>;
1500 def : Pat<(store (sra (loadi16 addr:$dst), (and CL, immShift32)), addr:$dst),
1501 (SAR16mCL addr:$dst)>;
1502 def : Pat<(store (sra (loadi32 addr:$dst), (and CL, immShift32)), addr:$dst),
1503 (SAR32mCL addr:$dst)>;
1505 // (shl x (and y, 63)) ==> (shl x, y)
1506 def : Pat<(shl GR64:$src1, (and CL, immShift64)),
1507 (SHL64rCL GR64:$src1)>;
1508 def : Pat<(store (shl (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1509 (SHL64mCL addr:$dst)>;
1511 def : Pat<(srl GR64:$src1, (and CL, immShift64)),
1512 (SHR64rCL GR64:$src1)>;
1513 def : Pat<(store (srl (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1514 (SHR64mCL addr:$dst)>;
1516 def : Pat<(sra GR64:$src1, (and CL, immShift64)),
1517 (SAR64rCL GR64:$src1)>;
1518 def : Pat<(store (sra (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1519 (SAR64mCL addr:$dst)>;
1522 // (anyext (setcc_carry)) -> (setcc_carry)
1523 def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1525 def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1527 def : Pat<(i32 (anyext (i16 (X86setcc_c X86_COND_B, EFLAGS)))),
1533 //===----------------------------------------------------------------------===//
1534 // EFLAGS-defining Patterns
1535 //===----------------------------------------------------------------------===//
1538 def : Pat<(add GR8 :$src1, GR8 :$src2), (ADD8rr GR8 :$src1, GR8 :$src2)>;
1539 def : Pat<(add GR16:$src1, GR16:$src2), (ADD16rr GR16:$src1, GR16:$src2)>;
1540 def : Pat<(add GR32:$src1, GR32:$src2), (ADD32rr GR32:$src1, GR32:$src2)>;
1543 def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
1544 (ADD8rm GR8:$src1, addr:$src2)>;
1545 def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
1546 (ADD16rm GR16:$src1, addr:$src2)>;
1547 def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
1548 (ADD32rm GR32:$src1, addr:$src2)>;
1551 def : Pat<(add GR8 :$src1, imm:$src2), (ADD8ri GR8:$src1 , imm:$src2)>;
1552 def : Pat<(add GR16:$src1, imm:$src2), (ADD16ri GR16:$src1, imm:$src2)>;
1553 def : Pat<(add GR32:$src1, imm:$src2), (ADD32ri GR32:$src1, imm:$src2)>;
1554 def : Pat<(add GR16:$src1, i16immSExt8:$src2),
1555 (ADD16ri8 GR16:$src1, i16immSExt8:$src2)>;
1556 def : Pat<(add GR32:$src1, i32immSExt8:$src2),
1557 (ADD32ri8 GR32:$src1, i32immSExt8:$src2)>;
1560 def : Pat<(sub GR8 :$src1, GR8 :$src2), (SUB8rr GR8 :$src1, GR8 :$src2)>;
1561 def : Pat<(sub GR16:$src1, GR16:$src2), (SUB16rr GR16:$src1, GR16:$src2)>;
1562 def : Pat<(sub GR32:$src1, GR32:$src2), (SUB32rr GR32:$src1, GR32:$src2)>;
1565 def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
1566 (SUB8rm GR8:$src1, addr:$src2)>;
1567 def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
1568 (SUB16rm GR16:$src1, addr:$src2)>;
1569 def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
1570 (SUB32rm GR32:$src1, addr:$src2)>;
1573 def : Pat<(sub GR8:$src1, imm:$src2),
1574 (SUB8ri GR8:$src1, imm:$src2)>;
1575 def : Pat<(sub GR16:$src1, imm:$src2),
1576 (SUB16ri GR16:$src1, imm:$src2)>;
1577 def : Pat<(sub GR32:$src1, imm:$src2),
1578 (SUB32ri GR32:$src1, imm:$src2)>;
1579 def : Pat<(sub GR16:$src1, i16immSExt8:$src2),
1580 (SUB16ri8 GR16:$src1, i16immSExt8:$src2)>;
1581 def : Pat<(sub GR32:$src1, i32immSExt8:$src2),
1582 (SUB32ri8 GR32:$src1, i32immSExt8:$src2)>;
1585 def : Pat<(mul GR16:$src1, GR16:$src2),
1586 (IMUL16rr GR16:$src1, GR16:$src2)>;
1587 def : Pat<(mul GR32:$src1, GR32:$src2),
1588 (IMUL32rr GR32:$src1, GR32:$src2)>;
1591 def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
1592 (IMUL16rm GR16:$src1, addr:$src2)>;
1593 def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
1594 (IMUL32rm GR32:$src1, addr:$src2)>;
1597 def : Pat<(mul GR16:$src1, imm:$src2),
1598 (IMUL16rri GR16:$src1, imm:$src2)>;
1599 def : Pat<(mul GR32:$src1, imm:$src2),
1600 (IMUL32rri GR32:$src1, imm:$src2)>;
1601 def : Pat<(mul GR16:$src1, i16immSExt8:$src2),
1602 (IMUL16rri8 GR16:$src1, i16immSExt8:$src2)>;
1603 def : Pat<(mul GR32:$src1, i32immSExt8:$src2),
1604 (IMUL32rri8 GR32:$src1, i32immSExt8:$src2)>;
1606 // reg = mul mem, imm
1607 def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
1608 (IMUL16rmi addr:$src1, imm:$src2)>;
1609 def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
1610 (IMUL32rmi addr:$src1, imm:$src2)>;
1611 def : Pat<(mul (loadi16 addr:$src1), i16immSExt8:$src2),
1612 (IMUL16rmi8 addr:$src1, i16immSExt8:$src2)>;
1613 def : Pat<(mul (loadi32 addr:$src1), i32immSExt8:$src2),
1614 (IMUL32rmi8 addr:$src1, i32immSExt8:$src2)>;
1616 // Patterns for nodes that do not produce flags, for instructions that do.
1619 def : Pat<(add GR64:$src1, GR64:$src2),
1620 (ADD64rr GR64:$src1, GR64:$src2)>;
1621 def : Pat<(add GR64:$src1, i64immSExt8:$src2),
1622 (ADD64ri8 GR64:$src1, i64immSExt8:$src2)>;
1623 def : Pat<(add GR64:$src1, i64immSExt32:$src2),
1624 (ADD64ri32 GR64:$src1, i64immSExt32:$src2)>;
1625 def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
1626 (ADD64rm GR64:$src1, addr:$src2)>;
1629 def : Pat<(sub GR64:$src1, GR64:$src2),
1630 (SUB64rr GR64:$src1, GR64:$src2)>;
1631 def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
1632 (SUB64rm GR64:$src1, addr:$src2)>;
1633 def : Pat<(sub GR64:$src1, i64immSExt8:$src2),
1634 (SUB64ri8 GR64:$src1, i64immSExt8:$src2)>;
1635 def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
1636 (SUB64ri32 GR64:$src1, i64immSExt32:$src2)>;
1639 def : Pat<(mul GR64:$src1, GR64:$src2),
1640 (IMUL64rr GR64:$src1, GR64:$src2)>;
1641 def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
1642 (IMUL64rm GR64:$src1, addr:$src2)>;
1643 def : Pat<(mul GR64:$src1, i64immSExt8:$src2),
1644 (IMUL64rri8 GR64:$src1, i64immSExt8:$src2)>;
1645 def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
1646 (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
1647 def : Pat<(mul (loadi64 addr:$src1), i64immSExt8:$src2),
1648 (IMUL64rmi8 addr:$src1, i64immSExt8:$src2)>;
1649 def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
1650 (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
1653 def : Pat<(add GR8 :$src, 1), (INC8r GR8 :$src)>;
1654 def : Pat<(add GR16:$src, 1), (INC16r GR16:$src)>, Requires<[In32BitMode]>;
1655 def : Pat<(add GR16:$src, 1), (INC64_16r GR16:$src)>, Requires<[In64BitMode]>;
1656 def : Pat<(add GR32:$src, 1), (INC32r GR32:$src)>, Requires<[In32BitMode]>;
1657 def : Pat<(add GR32:$src, 1), (INC64_32r GR32:$src)>, Requires<[In64BitMode]>;
1658 def : Pat<(add GR64:$src, 1), (INC64r GR64:$src)>;
1661 def : Pat<(add GR8 :$src, -1), (DEC8r GR8 :$src)>;
1662 def : Pat<(add GR16:$src, -1), (DEC16r GR16:$src)>, Requires<[In32BitMode]>;
1663 def : Pat<(add GR16:$src, -1), (DEC64_16r GR16:$src)>, Requires<[In64BitMode]>;
1664 def : Pat<(add GR32:$src, -1), (DEC32r GR32:$src)>, Requires<[In32BitMode]>;
1665 def : Pat<(add GR32:$src, -1), (DEC64_32r GR32:$src)>, Requires<[In64BitMode]>;
1666 def : Pat<(add GR64:$src, -1), (DEC64r GR64:$src)>;
1669 def : Pat<(or GR8 :$src1, GR8 :$src2), (OR8rr GR8 :$src1, GR8 :$src2)>;
1670 def : Pat<(or GR16:$src1, GR16:$src2), (OR16rr GR16:$src1, GR16:$src2)>;
1671 def : Pat<(or GR32:$src1, GR32:$src2), (OR32rr GR32:$src1, GR32:$src2)>;
1672 def : Pat<(or GR64:$src1, GR64:$src2), (OR64rr GR64:$src1, GR64:$src2)>;
1675 def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
1676 (OR8rm GR8:$src1, addr:$src2)>;
1677 def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
1678 (OR16rm GR16:$src1, addr:$src2)>;
1679 def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
1680 (OR32rm GR32:$src1, addr:$src2)>;
1681 def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
1682 (OR64rm GR64:$src1, addr:$src2)>;
1685 def : Pat<(or GR8:$src1 , imm:$src2), (OR8ri GR8 :$src1, imm:$src2)>;
1686 def : Pat<(or GR16:$src1, imm:$src2), (OR16ri GR16:$src1, imm:$src2)>;
1687 def : Pat<(or GR32:$src1, imm:$src2), (OR32ri GR32:$src1, imm:$src2)>;
1688 def : Pat<(or GR16:$src1, i16immSExt8:$src2),
1689 (OR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1690 def : Pat<(or GR32:$src1, i32immSExt8:$src2),
1691 (OR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1692 def : Pat<(or GR64:$src1, i64immSExt8:$src2),
1693 (OR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1694 def : Pat<(or GR64:$src1, i64immSExt32:$src2),
1695 (OR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1698 def : Pat<(xor GR8 :$src1, GR8 :$src2), (XOR8rr GR8 :$src1, GR8 :$src2)>;
1699 def : Pat<(xor GR16:$src1, GR16:$src2), (XOR16rr GR16:$src1, GR16:$src2)>;
1700 def : Pat<(xor GR32:$src1, GR32:$src2), (XOR32rr GR32:$src1, GR32:$src2)>;
1701 def : Pat<(xor GR64:$src1, GR64:$src2), (XOR64rr GR64:$src1, GR64:$src2)>;
1704 def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
1705 (XOR8rm GR8:$src1, addr:$src2)>;
1706 def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
1707 (XOR16rm GR16:$src1, addr:$src2)>;
1708 def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
1709 (XOR32rm GR32:$src1, addr:$src2)>;
1710 def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
1711 (XOR64rm GR64:$src1, addr:$src2)>;
1714 def : Pat<(xor GR8:$src1, imm:$src2),
1715 (XOR8ri GR8:$src1, imm:$src2)>;
1716 def : Pat<(xor GR16:$src1, imm:$src2),
1717 (XOR16ri GR16:$src1, imm:$src2)>;
1718 def : Pat<(xor GR32:$src1, imm:$src2),
1719 (XOR32ri GR32:$src1, imm:$src2)>;
1720 def : Pat<(xor GR16:$src1, i16immSExt8:$src2),
1721 (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1722 def : Pat<(xor GR32:$src1, i32immSExt8:$src2),
1723 (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1724 def : Pat<(xor GR64:$src1, i64immSExt8:$src2),
1725 (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1726 def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
1727 (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1730 def : Pat<(and GR8 :$src1, GR8 :$src2), (AND8rr GR8 :$src1, GR8 :$src2)>;
1731 def : Pat<(and GR16:$src1, GR16:$src2), (AND16rr GR16:$src1, GR16:$src2)>;
1732 def : Pat<(and GR32:$src1, GR32:$src2), (AND32rr GR32:$src1, GR32:$src2)>;
1733 def : Pat<(and GR64:$src1, GR64:$src2), (AND64rr GR64:$src1, GR64:$src2)>;
1736 def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
1737 (AND8rm GR8:$src1, addr:$src2)>;
1738 def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
1739 (AND16rm GR16:$src1, addr:$src2)>;
1740 def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
1741 (AND32rm GR32:$src1, addr:$src2)>;
1742 def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
1743 (AND64rm GR64:$src1, addr:$src2)>;
1746 def : Pat<(and GR8:$src1, imm:$src2),
1747 (AND8ri GR8:$src1, imm:$src2)>;
1748 def : Pat<(and GR16:$src1, imm:$src2),
1749 (AND16ri GR16:$src1, imm:$src2)>;
1750 def : Pat<(and GR32:$src1, imm:$src2),
1751 (AND32ri GR32:$src1, imm:$src2)>;
1752 def : Pat<(and GR16:$src1, i16immSExt8:$src2),
1753 (AND16ri8 GR16:$src1, i16immSExt8:$src2)>;
1754 def : Pat<(and GR32:$src1, i32immSExt8:$src2),
1755 (AND32ri8 GR32:$src1, i32immSExt8:$src2)>;
1756 def : Pat<(and GR64:$src1, i64immSExt8:$src2),
1757 (AND64ri8 GR64:$src1, i64immSExt8:$src2)>;
1758 def : Pat<(and GR64:$src1, i64immSExt32:$src2),
1759 (AND64ri32 GR64:$src1, i64immSExt32:$src2)>;
1761 // Bit scan instruction patterns to match explicit zero-undef behavior.
1762 def : Pat<(cttz_zero_undef GR16:$src), (BSF16rr GR16:$src)>;
1763 def : Pat<(cttz_zero_undef GR32:$src), (BSF32rr GR32:$src)>;
1764 def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr GR64:$src)>;
1765 def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm addr:$src)>;
1766 def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm addr:$src)>;
1767 def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm addr:$src)>;