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:$align)),
97 // Dynamic stack allocation yields a _chkstk or _alloca call for all Windows
98 // targets. These calls are needed to probe the stack when allocating more than
99 // 4k bytes in one go. Touching the stack at 4K increments is necessary to
100 // ensure that the guard pages used by the OS virtual memory manager are
101 // allocated in correct sequence.
102 // The main point of having separate instruction are extra unmodelled effects
103 // (compared to ordinary calls) like stack pointer change.
105 let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
106 def WIN_ALLOCA : I<0, Pseudo, (outs), (ins),
107 "# dynamic stack allocation",
113 //===----------------------------------------------------------------------===//
114 // EH Pseudo Instructions
116 let isTerminator = 1, isReturn = 1, isBarrier = 1,
117 hasCtrlDep = 1, isCodeGenOnly = 1 in {
118 def EH_RETURN : I<0xC3, RawFrm, (outs), (ins GR32:$addr),
119 "ret\t#eh_return, addr: $addr",
120 [(X86ehret GR32:$addr)]>;
124 let isTerminator = 1, isReturn = 1, isBarrier = 1,
125 hasCtrlDep = 1, isCodeGenOnly = 1 in {
126 def EH_RETURN64 : I<0xC3, RawFrm, (outs), (ins GR64:$addr),
127 "ret\t#eh_return, addr: $addr",
128 [(X86ehret GR64:$addr)]>;
132 //===----------------------------------------------------------------------===//
133 // Alias Instructions
134 //===----------------------------------------------------------------------===//
136 // Alias instructions that map movr0 to xor.
137 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
138 // FIXME: Set encoding to pseudo.
139 let Defs = [EFLAGS], isReMaterializable = 1, isAsCheapAsAMove = 1,
140 isCodeGenOnly = 1 in {
141 def MOV8r0 : I<0x30, MRMInitReg, (outs GR8 :$dst), (ins), "",
142 [(set GR8:$dst, 0)]>;
144 // We want to rewrite MOV16r0 in terms of MOV32r0, because it's a smaller
145 // encoding and avoids a partial-register update sometimes, but doing so
146 // at isel time interferes with rematerialization in the current register
147 // allocator. For now, this is rewritten when the instruction is lowered
149 def MOV16r0 : I<0x31, MRMInitReg, (outs GR16:$dst), (ins),
151 [(set GR16:$dst, 0)]>, OpSize;
153 // FIXME: Set encoding to pseudo.
154 def MOV32r0 : I<0x31, MRMInitReg, (outs GR32:$dst), (ins), "",
155 [(set GR32:$dst, 0)]>;
158 // We want to rewrite MOV64r0 in terms of MOV32r0, because it's sometimes a
159 // smaller encoding, but doing so at isel time interferes with rematerialization
160 // in the current register allocator. For now, this is rewritten when the
161 // instruction is lowered to an MCInst.
162 // FIXME: AddedComplexity gives this a higher priority than MOV64ri32. Remove
163 // when we have a better way to specify isel priority.
164 let Defs = [EFLAGS], isCodeGenOnly=1,
165 AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1 in
166 def MOV64r0 : I<0x31, MRMInitReg, (outs GR64:$dst), (ins), "",
167 [(set GR64:$dst, 0)]>;
169 // Materialize i64 constant where top 32-bits are zero. This could theoretically
170 // use MOV32ri with a SUBREG_TO_REG to represent the zero-extension, however
171 // that would make it more difficult to rematerialize.
172 let AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1,
174 def MOV64ri64i32 : Ii32<0xB8, AddRegFrm, (outs GR64:$dst), (ins i64i32imm:$src),
175 "", [(set GR64:$dst, i64immZExt32:$src)]>;
177 // Use sbb to materialize carry bit.
178 let Uses = [EFLAGS], Defs = [EFLAGS], isCodeGenOnly = 1 in {
179 // FIXME: These are pseudo ops that should be replaced with Pat<> patterns.
180 // However, Pat<> can't replicate the destination reg into the inputs of the
182 // FIXME: Change these to have encoding Pseudo when X86MCCodeEmitter replaces
184 def SETB_C8r : I<0x18, MRMInitReg, (outs GR8:$dst), (ins), "",
185 [(set GR8:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
186 def SETB_C16r : I<0x19, MRMInitReg, (outs GR16:$dst), (ins), "",
187 [(set GR16:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>,
189 def SETB_C32r : I<0x19, MRMInitReg, (outs GR32:$dst), (ins), "",
190 [(set GR32:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
191 def SETB_C64r : RI<0x19, MRMInitReg, (outs GR64:$dst), (ins), "",
192 [(set GR64:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
196 def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
198 def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
200 def : Pat<(i64 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
203 def : Pat<(i16 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
205 def : Pat<(i32 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
207 def : Pat<(i64 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
210 // We canonicalize 'setb' to "(and (sbb reg,reg), 1)" on the hope that the and
211 // will be eliminated and that the sbb can be extended up to a wider type. When
212 // this happens, it is great. However, if we are left with an 8-bit sbb and an
213 // and, we might as well just match it as a setb.
214 def : Pat<(and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1),
217 //===----------------------------------------------------------------------===//
218 // String Pseudo Instructions
220 let Defs = [ECX,EDI,ESI], Uses = [ECX,EDI,ESI], isCodeGenOnly = 1 in {
221 def REP_MOVSB : I<0xA4, RawFrm, (outs), (ins), "{rep;movsb|rep movsb}",
222 [(X86rep_movs i8)]>, REP;
223 def REP_MOVSW : I<0xA5, RawFrm, (outs), (ins), "{rep;movsw|rep movsw}",
224 [(X86rep_movs i16)]>, REP, OpSize;
225 def REP_MOVSD : I<0xA5, RawFrm, (outs), (ins), "{rep;movsl|rep movsd}",
226 [(X86rep_movs i32)]>, REP;
229 let Defs = [RCX,RDI,RSI], Uses = [RCX,RDI,RSI], isCodeGenOnly = 1 in
230 def REP_MOVSQ : RI<0xA5, RawFrm, (outs), (ins), "{rep;movsq|rep movsq}",
231 [(X86rep_movs i64)]>, REP;
234 // FIXME: Should use "(X86rep_stos AL)" as the pattern.
235 let Defs = [ECX,EDI], Uses = [AL,ECX,EDI], isCodeGenOnly = 1 in
236 def REP_STOSB : I<0xAA, RawFrm, (outs), (ins), "{rep;stosb|rep stosb}",
237 [(X86rep_stos i8)]>, REP;
238 let Defs = [ECX,EDI], Uses = [AX,ECX,EDI], isCodeGenOnly = 1 in
239 def REP_STOSW : I<0xAB, RawFrm, (outs), (ins), "{rep;stosw|rep stosw}",
240 [(X86rep_stos i16)]>, REP, OpSize;
241 let Defs = [ECX,EDI], Uses = [EAX,ECX,EDI], isCodeGenOnly = 1 in
242 def REP_STOSD : I<0xAB, RawFrm, (outs), (ins), "{rep;stosl|rep stosd}",
243 [(X86rep_stos i32)]>, REP;
245 let Defs = [RCX,RDI], Uses = [RAX,RCX,RDI], isCodeGenOnly = 1 in
246 def REP_STOSQ : RI<0xAB, RawFrm, (outs), (ins), "{rep;stosq|rep stosq}",
247 [(X86rep_stos i64)]>, REP;
250 //===----------------------------------------------------------------------===//
251 // Thread Local Storage Instructions
255 // All calls clobber the non-callee saved registers. ESP is marked as
256 // a use to prevent stack-pointer assignments that appear immediately
257 // before calls from potentially appearing dead.
258 let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0,
259 MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
260 XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
261 XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
263 def TLS_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
265 [(X86tlsaddr tls32addr:$sym)]>,
266 Requires<[In32BitMode]>;
268 // All calls clobber the non-callee saved registers. RSP is marked as
269 // a use to prevent stack-pointer assignments that appear immediately
270 // before calls from potentially appearing dead.
271 let Defs = [RAX, RCX, RDX, RSI, RDI, R8, R9, R10, R11,
272 FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0, ST1,
273 MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
274 XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
275 XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
277 def TLS_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
279 [(X86tlsaddr tls64addr:$sym)]>,
280 Requires<[In64BitMode]>;
282 // Darwin TLS Support
283 // For i386, the address of the thunk is passed on the stack, on return the
284 // address of the variable is in %eax. %ecx is trashed during the function
285 // call. All other registers are preserved.
286 let Defs = [EAX, ECX, EFLAGS],
288 usesCustomInserter = 1 in
289 def TLSCall_32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
291 [(X86TLSCall addr:$sym)]>,
292 Requires<[In32BitMode]>;
294 // For x86_64, the address of the thunk is passed in %rdi, on return
295 // the address of the variable is in %rax. All other registers are preserved.
296 let Defs = [RAX, EFLAGS],
298 usesCustomInserter = 1 in
299 def TLSCall_64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
301 [(X86TLSCall addr:$sym)]>,
302 Requires<[In64BitMode]>;
305 //===----------------------------------------------------------------------===//
306 // Conditional Move Pseudo Instructions
308 let Constraints = "$src1 = $dst" in {
311 let Uses = [EFLAGS] in {
313 // X86 doesn't have 8-bit conditional moves. Use a customInserter to
314 // emit control flow. An alternative to this is to mark i8 SELECT as Promote,
315 // however that requires promoting the operands, and can induce additional
316 // i8 register pressure. Note that CMOV_GR8 is conservatively considered to
317 // clobber EFLAGS, because if one of the operands is zero, the expansion
318 // could involve an xor.
319 let usesCustomInserter = 1, Constraints = "", Defs = [EFLAGS] in {
320 def CMOV_GR8 : I<0, Pseudo,
321 (outs GR8:$dst), (ins GR8:$src1, GR8:$src2, i8imm:$cond),
323 [(set GR8:$dst, (X86cmov GR8:$src1, GR8:$src2,
324 imm:$cond, EFLAGS))]>;
326 let Predicates = [NoCMov] in {
327 def CMOV_GR32 : I<0, Pseudo,
328 (outs GR32:$dst), (ins GR32:$src1, GR32:$src2, i8imm:$cond),
329 "#CMOV_GR32* PSEUDO!",
331 (X86cmov GR32:$src1, GR32:$src2, imm:$cond, EFLAGS))]>;
332 def CMOV_GR16 : I<0, Pseudo,
333 (outs GR16:$dst), (ins GR16:$src1, GR16:$src2, i8imm:$cond),
334 "#CMOV_GR16* PSEUDO!",
336 (X86cmov GR16:$src1, GR16:$src2, imm:$cond, EFLAGS))]>;
337 def CMOV_RFP32 : I<0, Pseudo,
339 (ins RFP32:$src1, RFP32:$src2, i8imm:$cond),
340 "#CMOV_RFP32 PSEUDO!",
342 (X86cmov RFP32:$src1, RFP32:$src2, imm:$cond,
344 def CMOV_RFP64 : I<0, Pseudo,
346 (ins RFP64:$src1, RFP64:$src2, i8imm:$cond),
347 "#CMOV_RFP64 PSEUDO!",
349 (X86cmov RFP64:$src1, RFP64:$src2, imm:$cond,
351 def CMOV_RFP80 : I<0, Pseudo,
353 (ins RFP80:$src1, RFP80:$src2, i8imm:$cond),
354 "#CMOV_RFP80 PSEUDO!",
356 (X86cmov RFP80:$src1, RFP80:$src2, imm:$cond,
358 } // Predicates = [NoCMov]
359 } // UsesCustomInserter = 1, Constraints = "", Defs = [EFLAGS]
362 } // Constraints = "$src1 = $dst" in
365 //===----------------------------------------------------------------------===//
366 // Atomic Instruction Pseudo Instructions
367 //===----------------------------------------------------------------------===//
369 // Atomic exchange, and, or, xor
370 let Constraints = "$val = $dst", Defs = [EFLAGS],
371 usesCustomInserter = 1 in {
373 def ATOMAND8 : I<0, Pseudo, (outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
375 [(set GR8:$dst, (atomic_load_and_8 addr:$ptr, GR8:$val))]>;
376 def ATOMOR8 : I<0, Pseudo, (outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
378 [(set GR8:$dst, (atomic_load_or_8 addr:$ptr, GR8:$val))]>;
379 def ATOMXOR8 : I<0, Pseudo,(outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
381 [(set GR8:$dst, (atomic_load_xor_8 addr:$ptr, GR8:$val))]>;
382 def ATOMNAND8 : I<0, Pseudo,(outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
383 "#ATOMNAND8 PSEUDO!",
384 [(set GR8:$dst, (atomic_load_nand_8 addr:$ptr, GR8:$val))]>;
386 def ATOMAND16 : I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
387 "#ATOMAND16 PSEUDO!",
388 [(set GR16:$dst, (atomic_load_and_16 addr:$ptr, GR16:$val))]>;
389 def ATOMOR16 : I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
391 [(set GR16:$dst, (atomic_load_or_16 addr:$ptr, GR16:$val))]>;
392 def ATOMXOR16 : I<0, Pseudo,(outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
393 "#ATOMXOR16 PSEUDO!",
394 [(set GR16:$dst, (atomic_load_xor_16 addr:$ptr, GR16:$val))]>;
395 def ATOMNAND16 : I<0, Pseudo,(outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
396 "#ATOMNAND16 PSEUDO!",
397 [(set GR16:$dst, (atomic_load_nand_16 addr:$ptr, GR16:$val))]>;
398 def ATOMMIN16: I<0, Pseudo, (outs GR16:$dst), (ins i16mem:$ptr, GR16:$val),
399 "#ATOMMIN16 PSEUDO!",
400 [(set GR16:$dst, (atomic_load_min_16 addr:$ptr, GR16:$val))]>;
401 def ATOMMAX16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
402 "#ATOMMAX16 PSEUDO!",
403 [(set GR16:$dst, (atomic_load_max_16 addr:$ptr, GR16:$val))]>;
404 def ATOMUMIN16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
405 "#ATOMUMIN16 PSEUDO!",
406 [(set GR16:$dst, (atomic_load_umin_16 addr:$ptr, GR16:$val))]>;
407 def ATOMUMAX16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
408 "#ATOMUMAX16 PSEUDO!",
409 [(set GR16:$dst, (atomic_load_umax_16 addr:$ptr, GR16:$val))]>;
412 def ATOMAND32 : I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
413 "#ATOMAND32 PSEUDO!",
414 [(set GR32:$dst, (atomic_load_and_32 addr:$ptr, GR32:$val))]>;
415 def ATOMOR32 : I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
417 [(set GR32:$dst, (atomic_load_or_32 addr:$ptr, GR32:$val))]>;
418 def ATOMXOR32 : I<0, Pseudo,(outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
419 "#ATOMXOR32 PSEUDO!",
420 [(set GR32:$dst, (atomic_load_xor_32 addr:$ptr, GR32:$val))]>;
421 def ATOMNAND32 : I<0, Pseudo,(outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
422 "#ATOMNAND32 PSEUDO!",
423 [(set GR32:$dst, (atomic_load_nand_32 addr:$ptr, GR32:$val))]>;
424 def ATOMMIN32: I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$ptr, GR32:$val),
425 "#ATOMMIN32 PSEUDO!",
426 [(set GR32:$dst, (atomic_load_min_32 addr:$ptr, GR32:$val))]>;
427 def ATOMMAX32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
428 "#ATOMMAX32 PSEUDO!",
429 [(set GR32:$dst, (atomic_load_max_32 addr:$ptr, GR32:$val))]>;
430 def ATOMUMIN32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
431 "#ATOMUMIN32 PSEUDO!",
432 [(set GR32:$dst, (atomic_load_umin_32 addr:$ptr, GR32:$val))]>;
433 def ATOMUMAX32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
434 "#ATOMUMAX32 PSEUDO!",
435 [(set GR32:$dst, (atomic_load_umax_32 addr:$ptr, GR32:$val))]>;
439 def ATOMAND64 : I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
440 "#ATOMAND64 PSEUDO!",
441 [(set GR64:$dst, (atomic_load_and_64 addr:$ptr, GR64:$val))]>;
442 def ATOMOR64 : I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
444 [(set GR64:$dst, (atomic_load_or_64 addr:$ptr, GR64:$val))]>;
445 def ATOMXOR64 : I<0, Pseudo,(outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
446 "#ATOMXOR64 PSEUDO!",
447 [(set GR64:$dst, (atomic_load_xor_64 addr:$ptr, GR64:$val))]>;
448 def ATOMNAND64 : I<0, Pseudo,(outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
449 "#ATOMNAND64 PSEUDO!",
450 [(set GR64:$dst, (atomic_load_nand_64 addr:$ptr, GR64:$val))]>;
451 def ATOMMIN64: I<0, Pseudo, (outs GR64:$dst), (ins i64mem:$ptr, GR64:$val),
452 "#ATOMMIN64 PSEUDO!",
453 [(set GR64:$dst, (atomic_load_min_64 addr:$ptr, GR64:$val))]>;
454 def ATOMMAX64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
455 "#ATOMMAX64 PSEUDO!",
456 [(set GR64:$dst, (atomic_load_max_64 addr:$ptr, GR64:$val))]>;
457 def ATOMUMIN64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
458 "#ATOMUMIN64 PSEUDO!",
459 [(set GR64:$dst, (atomic_load_umin_64 addr:$ptr, GR64:$val))]>;
460 def ATOMUMAX64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
461 "#ATOMUMAX64 PSEUDO!",
462 [(set GR64:$dst, (atomic_load_umax_64 addr:$ptr, GR64:$val))]>;
465 let Constraints = "$val1 = $dst1, $val2 = $dst2",
466 Defs = [EFLAGS, EAX, EBX, ECX, EDX],
467 Uses = [EAX, EBX, ECX, EDX],
468 mayLoad = 1, mayStore = 1,
469 usesCustomInserter = 1 in {
470 def ATOMAND6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
471 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
472 "#ATOMAND6432 PSEUDO!", []>;
473 def ATOMOR6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
474 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
475 "#ATOMOR6432 PSEUDO!", []>;
476 def ATOMXOR6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
477 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
478 "#ATOMXOR6432 PSEUDO!", []>;
479 def ATOMNAND6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
480 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
481 "#ATOMNAND6432 PSEUDO!", []>;
482 def ATOMADD6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
483 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
484 "#ATOMADD6432 PSEUDO!", []>;
485 def ATOMSUB6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
486 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
487 "#ATOMSUB6432 PSEUDO!", []>;
488 def ATOMSWAP6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
489 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
490 "#ATOMSWAP6432 PSEUDO!", []>;
493 //===----------------------------------------------------------------------===//
494 // Normal-Instructions-With-Lock-Prefix Pseudo Instructions
495 //===----------------------------------------------------------------------===//
497 // FIXME: Use normal instructions and add lock prefix dynamically.
501 // TODO: Get this to fold the constant into the instruction.
502 let isCodeGenOnly = 1 in
503 def OR32mrLocked : I<0x09, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$zero),
505 "or{l}\t{$zero, $dst|$dst, $zero}",
506 []>, Requires<[In32BitMode]>, LOCK;
508 let hasSideEffects = 1 in
509 def Int_MemBarrier : I<0, Pseudo, (outs), (ins),
511 [(X86MemBarrier)]>, Requires<[HasSSE2]>;
513 // TODO: Get this to fold the constant into the instruction.
514 let hasSideEffects = 1, Defs = [ESP], isCodeGenOnly = 1 in
515 def Int_MemBarrierNoSSE64 : RI<0x09, MRM1r, (outs), (ins GR64:$zero),
517 "or{q}\t{$zero, (%rsp)|(%rsp), $zero}",
518 [(X86MemBarrierNoSSE GR64:$zero)]>,
519 Requires<[In64BitMode]>, LOCK;
522 // Optimized codegen when the non-memory output is not used.
523 let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1 in {
524 def LOCK_ADD8mr : I<0x00, MRMDestMem, (outs), (ins i8mem:$dst, GR8:$src2),
526 "add{b}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
527 def LOCK_ADD16mr : I<0x01, MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
529 "add{w}\t{$src2, $dst|$dst, $src2}", []>, OpSize, LOCK;
530 def LOCK_ADD32mr : I<0x01, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
532 "add{l}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
533 def LOCK_ADD64mr : RI<0x01, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
535 "add{q}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
537 def LOCK_ADD8mi : Ii8<0x80, MRM0m, (outs), (ins i8mem :$dst, i8imm :$src2),
539 "add{b}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
540 def LOCK_ADD16mi : Ii16<0x81, MRM0m, (outs), (ins i16mem:$dst, i16imm:$src2),
542 "add{w}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
543 def LOCK_ADD32mi : Ii32<0x81, MRM0m, (outs), (ins i32mem:$dst, i32imm:$src2),
545 "add{l}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
546 def LOCK_ADD64mi32 : RIi32<0x81, MRM0m, (outs),
547 (ins i64mem:$dst, i64i32imm :$src2),
549 "add{q}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
551 def LOCK_ADD16mi8 : Ii8<0x83, MRM0m, (outs), (ins i16mem:$dst, i16i8imm :$src2),
553 "add{w}\t{$src2, $dst|$dst, $src2}", []>, OpSize, LOCK;
554 def LOCK_ADD32mi8 : Ii8<0x83, MRM0m, (outs), (ins i32mem:$dst, i32i8imm :$src2),
556 "add{l}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
557 def LOCK_ADD64mi8 : RIi8<0x83, MRM0m, (outs),
558 (ins i64mem:$dst, i64i8imm :$src2),
560 "add{q}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
562 def LOCK_SUB8mr : I<0x28, MRMDestMem, (outs), (ins i8mem :$dst, GR8 :$src2),
564 "sub{b}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
565 def LOCK_SUB16mr : I<0x29, MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
567 "sub{w}\t{$src2, $dst|$dst, $src2}", []>, OpSize, LOCK;
568 def LOCK_SUB32mr : I<0x29, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
570 "sub{l}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
571 def LOCK_SUB64mr : RI<0x29, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
573 "sub{q}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
576 def LOCK_SUB8mi : Ii8<0x80, MRM5m, (outs), (ins i8mem :$dst, i8imm:$src2),
578 "sub{b}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
579 def LOCK_SUB16mi : Ii16<0x81, MRM5m, (outs), (ins i16mem:$dst, i16imm:$src2),
581 "sub{w}\t{$src2, $dst|$dst, $src2}", []>, OpSize, LOCK;
582 def LOCK_SUB32mi : Ii32<0x81, MRM5m, (outs), (ins i32mem:$dst, i32imm:$src2),
584 "sub{l}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
585 def LOCK_SUB64mi32 : RIi32<0x81, MRM5m, (outs),
586 (ins i64mem:$dst, i64i32imm:$src2),
588 "sub{q}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
591 def LOCK_SUB16mi8 : Ii8<0x83, MRM5m, (outs), (ins i16mem:$dst, i16i8imm :$src2),
593 "sub{w}\t{$src2, $dst|$dst, $src2}", []>, OpSize, LOCK;
594 def LOCK_SUB32mi8 : Ii8<0x83, MRM5m, (outs), (ins i32mem:$dst, i32i8imm :$src2),
596 "sub{l}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
597 def LOCK_SUB64mi8 : RIi8<0x83, MRM5m, (outs),
598 (ins i64mem:$dst, i64i8imm :$src2),
600 "sub{q}\t{$src2, $dst|$dst, $src2}", []>, LOCK;
602 def LOCK_INC8m : I<0xFE, MRM0m, (outs), (ins i8mem :$dst),
604 "inc{b}\t$dst", []>, LOCK;
605 def LOCK_INC16m : I<0xFF, MRM0m, (outs), (ins i16mem:$dst),
607 "inc{w}\t$dst", []>, OpSize, LOCK;
608 def LOCK_INC32m : I<0xFF, MRM0m, (outs), (ins i32mem:$dst),
610 "inc{l}\t$dst", []>, LOCK;
611 def LOCK_INC64m : RI<0xFF, MRM0m, (outs), (ins i64mem:$dst),
613 "inc{q}\t$dst", []>, LOCK;
615 def LOCK_DEC8m : I<0xFE, MRM1m, (outs), (ins i8mem :$dst),
617 "dec{b}\t$dst", []>, LOCK;
618 def LOCK_DEC16m : I<0xFF, MRM1m, (outs), (ins i16mem:$dst),
620 "dec{w}\t$dst", []>, OpSize, LOCK;
621 def LOCK_DEC32m : I<0xFF, MRM1m, (outs), (ins i32mem:$dst),
623 "dec{l}\t$dst", []>, LOCK;
624 def LOCK_DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst),
626 "dec{q}\t$dst", []>, LOCK;
629 // Atomic compare and swap.
630 let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
631 isCodeGenOnly = 1 in {
632 def LCMPXCHG8B : I<0xC7, MRM1m, (outs), (ins i64mem:$ptr),
635 [(X86cas8 addr:$ptr)]>, TB, LOCK;
637 let Defs = [AL, EFLAGS], Uses = [AL], isCodeGenOnly = 1 in {
638 def LCMPXCHG8 : I<0xB0, MRMDestMem, (outs), (ins i8mem:$ptr, GR8:$swap),
640 "cmpxchg{b}\t{$swap, $ptr|$ptr, $swap}",
641 [(X86cas addr:$ptr, GR8:$swap, 1)]>, TB, LOCK;
644 let Defs = [AX, EFLAGS], Uses = [AX], isCodeGenOnly = 1 in {
645 def LCMPXCHG16 : I<0xB1, MRMDestMem, (outs), (ins i16mem:$ptr, GR16:$swap),
647 "cmpxchg{w}\t{$swap, $ptr|$ptr, $swap}",
648 [(X86cas addr:$ptr, GR16:$swap, 2)]>, TB, OpSize, LOCK;
651 let Defs = [EAX, EFLAGS], Uses = [EAX], isCodeGenOnly = 1 in {
652 def LCMPXCHG32 : I<0xB1, MRMDestMem, (outs), (ins i32mem:$ptr, GR32:$swap),
654 "cmpxchg{l}\t{$swap, $ptr|$ptr, $swap}",
655 [(X86cas addr:$ptr, GR32:$swap, 4)]>, TB, LOCK;
658 let Defs = [RAX, EFLAGS], Uses = [RAX], isCodeGenOnly = 1 in {
659 def LCMPXCHG64 : RI<0xB1, MRMDestMem, (outs), (ins i64mem:$ptr, GR64:$swap),
661 "cmpxchgq\t$swap,$ptr",
662 [(X86cas addr:$ptr, GR64:$swap, 8)]>, TB, LOCK;
665 // Atomic exchange and add
666 let Constraints = "$val = $dst", Defs = [EFLAGS], isCodeGenOnly = 1 in {
667 def LXADD8 : I<0xC0, MRMSrcMem, (outs GR8:$dst), (ins GR8:$val, i8mem:$ptr),
669 "xadd{b}\t{$val, $ptr|$ptr, $val}",
670 [(set GR8:$dst, (atomic_load_add_8 addr:$ptr, GR8:$val))]>,
672 def LXADD16 : I<0xC1, MRMSrcMem, (outs GR16:$dst), (ins GR16:$val, i16mem:$ptr),
674 "xadd{w}\t{$val, $ptr|$ptr, $val}",
675 [(set GR16:$dst, (atomic_load_add_16 addr:$ptr, GR16:$val))]>,
677 def LXADD32 : I<0xC1, MRMSrcMem, (outs GR32:$dst), (ins GR32:$val, i32mem:$ptr),
679 "xadd{l}\t{$val, $ptr|$ptr, $val}",
680 [(set GR32:$dst, (atomic_load_add_32 addr:$ptr, GR32:$val))]>,
682 def LXADD64 : RI<0xC1, MRMSrcMem, (outs GR64:$dst), (ins GR64:$val,i64mem:$ptr),
685 [(set GR64:$dst, (atomic_load_add_64 addr:$ptr, GR64:$val))]>,
689 //===----------------------------------------------------------------------===//
690 // Conditional Move Pseudo Instructions.
691 //===----------------------------------------------------------------------===//
694 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded after
695 // instruction selection into a branch sequence.
696 let Uses = [EFLAGS], usesCustomInserter = 1 in {
697 def CMOV_FR32 : I<0, Pseudo,
698 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
699 "#CMOV_FR32 PSEUDO!",
700 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
702 def CMOV_FR64 : I<0, Pseudo,
703 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
704 "#CMOV_FR64 PSEUDO!",
705 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
707 def CMOV_V4F32 : I<0, Pseudo,
708 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
709 "#CMOV_V4F32 PSEUDO!",
711 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
713 def CMOV_V2F64 : I<0, Pseudo,
714 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
715 "#CMOV_V2F64 PSEUDO!",
717 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
719 def CMOV_V2I64 : I<0, Pseudo,
720 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
721 "#CMOV_V2I64 PSEUDO!",
723 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
728 //===----------------------------------------------------------------------===//
729 // DAG Pattern Matching Rules
730 //===----------------------------------------------------------------------===//
732 // ConstantPool GlobalAddress, ExternalSymbol, and JumpTable
733 def : Pat<(i32 (X86Wrapper tconstpool :$dst)), (MOV32ri tconstpool :$dst)>;
734 def : Pat<(i32 (X86Wrapper tjumptable :$dst)), (MOV32ri tjumptable :$dst)>;
735 def : Pat<(i32 (X86Wrapper tglobaltlsaddr:$dst)),(MOV32ri tglobaltlsaddr:$dst)>;
736 def : Pat<(i32 (X86Wrapper tglobaladdr :$dst)), (MOV32ri tglobaladdr :$dst)>;
737 def : Pat<(i32 (X86Wrapper texternalsym:$dst)), (MOV32ri texternalsym:$dst)>;
738 def : Pat<(i32 (X86Wrapper tblockaddress:$dst)), (MOV32ri tblockaddress:$dst)>;
740 def : Pat<(add GR32:$src1, (X86Wrapper tconstpool:$src2)),
741 (ADD32ri GR32:$src1, tconstpool:$src2)>;
742 def : Pat<(add GR32:$src1, (X86Wrapper tjumptable:$src2)),
743 (ADD32ri GR32:$src1, tjumptable:$src2)>;
744 def : Pat<(add GR32:$src1, (X86Wrapper tglobaladdr :$src2)),
745 (ADD32ri GR32:$src1, tglobaladdr:$src2)>;
746 def : Pat<(add GR32:$src1, (X86Wrapper texternalsym:$src2)),
747 (ADD32ri GR32:$src1, texternalsym:$src2)>;
748 def : Pat<(add GR32:$src1, (X86Wrapper tblockaddress:$src2)),
749 (ADD32ri GR32:$src1, tblockaddress:$src2)>;
751 def : Pat<(store (i32 (X86Wrapper tglobaladdr:$src)), addr:$dst),
752 (MOV32mi addr:$dst, tglobaladdr:$src)>;
753 def : Pat<(store (i32 (X86Wrapper texternalsym:$src)), addr:$dst),
754 (MOV32mi addr:$dst, texternalsym:$src)>;
755 def : Pat<(store (i32 (X86Wrapper tblockaddress:$src)), addr:$dst),
756 (MOV32mi addr:$dst, tblockaddress:$src)>;
760 // ConstantPool GlobalAddress, ExternalSymbol, and JumpTable when not in small
761 // code model mode, should use 'movabs'. FIXME: This is really a hack, the
762 // 'movabs' predicate should handle this sort of thing.
763 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
764 (MOV64ri tconstpool :$dst)>, Requires<[FarData]>;
765 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
766 (MOV64ri tjumptable :$dst)>, Requires<[FarData]>;
767 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
768 (MOV64ri tglobaladdr :$dst)>, Requires<[FarData]>;
769 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
770 (MOV64ri texternalsym:$dst)>, Requires<[FarData]>;
771 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
772 (MOV64ri tblockaddress:$dst)>, Requires<[FarData]>;
774 // In static codegen with small code model, we can get the address of a label
775 // into a register with 'movl'. FIXME: This is a hack, the 'imm' predicate of
776 // the MOV64ri64i32 should accept these.
777 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
778 (MOV64ri64i32 tconstpool :$dst)>, Requires<[SmallCode]>;
779 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
780 (MOV64ri64i32 tjumptable :$dst)>, Requires<[SmallCode]>;
781 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
782 (MOV64ri64i32 tglobaladdr :$dst)>, Requires<[SmallCode]>;
783 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
784 (MOV64ri64i32 texternalsym:$dst)>, Requires<[SmallCode]>;
785 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
786 (MOV64ri64i32 tblockaddress:$dst)>, Requires<[SmallCode]>;
788 // In kernel code model, we can get the address of a label
789 // into a register with 'movq'. FIXME: This is a hack, the 'imm' predicate of
790 // the MOV64ri32 should accept these.
791 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
792 (MOV64ri32 tconstpool :$dst)>, Requires<[KernelCode]>;
793 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
794 (MOV64ri32 tjumptable :$dst)>, Requires<[KernelCode]>;
795 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
796 (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
797 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
798 (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
799 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
800 (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
802 // If we have small model and -static mode, it is safe to store global addresses
803 // directly as immediates. FIXME: This is really a hack, the 'imm' predicate
804 // for MOV64mi32 should handle this sort of thing.
805 def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
806 (MOV64mi32 addr:$dst, tconstpool:$src)>,
807 Requires<[NearData, IsStatic]>;
808 def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
809 (MOV64mi32 addr:$dst, tjumptable:$src)>,
810 Requires<[NearData, IsStatic]>;
811 def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
812 (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
813 Requires<[NearData, IsStatic]>;
814 def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
815 (MOV64mi32 addr:$dst, texternalsym:$src)>,
816 Requires<[NearData, IsStatic]>;
817 def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
818 (MOV64mi32 addr:$dst, tblockaddress:$src)>,
819 Requires<[NearData, IsStatic]>;
825 // tls has some funny stuff here...
826 // This corresponds to movabs $foo@tpoff, %rax
827 def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
828 (MOV64ri tglobaltlsaddr :$dst)>;
829 // This corresponds to add $foo@tpoff, %rax
830 def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
831 (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
832 // This corresponds to mov foo@tpoff(%rbx), %eax
833 def : Pat<(load (i64 (X86Wrapper tglobaltlsaddr :$dst))),
834 (MOV64rm tglobaltlsaddr :$dst)>;
837 // Direct PC relative function call for small code model. 32-bit displacement
838 // sign extended to 64-bit.
839 def : Pat<(X86call (i64 tglobaladdr:$dst)),
840 (CALL64pcrel32 tglobaladdr:$dst)>, Requires<[NotWin64]>;
841 def : Pat<(X86call (i64 texternalsym:$dst)),
842 (CALL64pcrel32 texternalsym:$dst)>, Requires<[NotWin64]>;
844 def : Pat<(X86call (i64 tglobaladdr:$dst)),
845 (WINCALL64pcrel32 tglobaladdr:$dst)>, Requires<[IsWin64]>;
846 def : Pat<(X86call (i64 texternalsym:$dst)),
847 (WINCALL64pcrel32 texternalsym:$dst)>, Requires<[IsWin64]>;
850 def : Pat<(X86tcret GR32_TC:$dst, imm:$off),
851 (TCRETURNri GR32_TC:$dst, imm:$off)>,
852 Requires<[In32BitMode]>;
854 // FIXME: This is disabled for 32-bit PIC mode because the global base
855 // register which is part of the address mode may be assigned a
856 // callee-saved register.
857 def : Pat<(X86tcret (load addr:$dst), imm:$off),
858 (TCRETURNmi addr:$dst, imm:$off)>,
859 Requires<[In32BitMode, IsNotPIC]>;
861 def : Pat<(X86tcret (i32 tglobaladdr:$dst), imm:$off),
862 (TCRETURNdi texternalsym:$dst, imm:$off)>,
863 Requires<[In32BitMode]>;
865 def : Pat<(X86tcret (i32 texternalsym:$dst), imm:$off),
866 (TCRETURNdi texternalsym:$dst, imm:$off)>,
867 Requires<[In32BitMode]>;
869 def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
870 (TCRETURNri64 ptr_rc_tailcall:$dst, imm:$off)>,
871 Requires<[In64BitMode]>;
873 def : Pat<(X86tcret (load addr:$dst), imm:$off),
874 (TCRETURNmi64 addr:$dst, imm:$off)>,
875 Requires<[In64BitMode]>;
877 def : Pat<(X86tcret (i64 tglobaladdr:$dst), imm:$off),
878 (TCRETURNdi64 tglobaladdr:$dst, imm:$off)>,
879 Requires<[In64BitMode]>;
881 def : Pat<(X86tcret (i64 texternalsym:$dst), imm:$off),
882 (TCRETURNdi64 texternalsym:$dst, imm:$off)>,
883 Requires<[In64BitMode]>;
885 // Normal calls, with various flavors of addresses.
886 def : Pat<(X86call (i32 tglobaladdr:$dst)),
887 (CALLpcrel32 tglobaladdr:$dst)>;
888 def : Pat<(X86call (i32 texternalsym:$dst)),
889 (CALLpcrel32 texternalsym:$dst)>;
890 def : Pat<(X86call (i32 imm:$dst)),
891 (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
895 // TEST R,R is smaller than CMP R,0
896 def : Pat<(X86cmp GR8:$src1, 0),
897 (TEST8rr GR8:$src1, GR8:$src1)>;
898 def : Pat<(X86cmp GR16:$src1, 0),
899 (TEST16rr GR16:$src1, GR16:$src1)>;
900 def : Pat<(X86cmp GR32:$src1, 0),
901 (TEST32rr GR32:$src1, GR32:$src1)>;
902 def : Pat<(X86cmp GR64:$src1, 0),
903 (TEST64rr GR64:$src1, GR64:$src1)>;
905 // Conditional moves with folded loads with operands swapped and conditions
907 multiclass CMOVmr<PatLeaf InvertedCond, Instruction Inst16, Instruction Inst32,
908 Instruction Inst64> {
909 def : Pat<(X86cmov (loadi16 addr:$src1), GR16:$src2, InvertedCond, EFLAGS),
910 (Inst16 GR16:$src2, addr:$src1)>;
911 def : Pat<(X86cmov (loadi32 addr:$src1), GR32:$src2, InvertedCond, EFLAGS),
912 (Inst32 GR32:$src2, addr:$src1)>;
913 def : Pat<(X86cmov (loadi64 addr:$src1), GR64:$src2, InvertedCond, EFLAGS),
914 (Inst64 GR64:$src2, addr:$src1)>;
917 defm : CMOVmr<X86_COND_B , CMOVAE16rm, CMOVAE32rm, CMOVAE64rm>;
918 defm : CMOVmr<X86_COND_AE, CMOVB16rm , CMOVB32rm , CMOVB64rm>;
919 defm : CMOVmr<X86_COND_E , CMOVNE16rm, CMOVNE32rm, CMOVNE64rm>;
920 defm : CMOVmr<X86_COND_NE, CMOVE16rm , CMOVE32rm , CMOVE64rm>;
921 defm : CMOVmr<X86_COND_BE, CMOVA16rm , CMOVA32rm , CMOVA64rm>;
922 defm : CMOVmr<X86_COND_A , CMOVBE16rm, CMOVBE32rm, CMOVBE64rm>;
923 defm : CMOVmr<X86_COND_L , CMOVGE16rm, CMOVGE32rm, CMOVGE64rm>;
924 defm : CMOVmr<X86_COND_GE, CMOVL16rm , CMOVL32rm , CMOVL64rm>;
925 defm : CMOVmr<X86_COND_LE, CMOVG16rm , CMOVG32rm , CMOVG64rm>;
926 defm : CMOVmr<X86_COND_G , CMOVLE16rm, CMOVLE32rm, CMOVLE64rm>;
927 defm : CMOVmr<X86_COND_P , CMOVNP16rm, CMOVNP32rm, CMOVNP64rm>;
928 defm : CMOVmr<X86_COND_NP, CMOVP16rm , CMOVP32rm , CMOVP64rm>;
929 defm : CMOVmr<X86_COND_S , CMOVNS16rm, CMOVNS32rm, CMOVNS64rm>;
930 defm : CMOVmr<X86_COND_NS, CMOVS16rm , CMOVS32rm , CMOVS64rm>;
931 defm : CMOVmr<X86_COND_O , CMOVNO16rm, CMOVNO32rm, CMOVNO64rm>;
932 defm : CMOVmr<X86_COND_NO, CMOVO16rm , CMOVO32rm , CMOVO64rm>;
934 // zextload bool -> zextload byte
935 def : Pat<(zextloadi8i1 addr:$src), (MOV8rm addr:$src)>;
936 def : Pat<(zextloadi16i1 addr:$src), (MOVZX16rm8 addr:$src)>;
937 def : Pat<(zextloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
938 def : Pat<(zextloadi64i1 addr:$src), (MOVZX64rm8 addr:$src)>;
940 // extload bool -> extload byte
941 // When extloading from 16-bit and smaller memory locations into 64-bit
942 // registers, use zero-extending loads so that the entire 64-bit register is
943 // defined, avoiding partial-register updates.
945 def : Pat<(extloadi8i1 addr:$src), (MOV8rm addr:$src)>;
946 def : Pat<(extloadi16i1 addr:$src), (MOVZX16rm8 addr:$src)>;
947 def : Pat<(extloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
948 def : Pat<(extloadi16i8 addr:$src), (MOVZX16rm8 addr:$src)>;
949 def : Pat<(extloadi32i8 addr:$src), (MOVZX32rm8 addr:$src)>;
950 def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
952 def : Pat<(extloadi64i1 addr:$src), (MOVZX64rm8 addr:$src)>;
953 def : Pat<(extloadi64i8 addr:$src), (MOVZX64rm8 addr:$src)>;
954 def : Pat<(extloadi64i16 addr:$src), (MOVZX64rm16 addr:$src)>;
955 // For other extloads, use subregs, since the high contents of the register are
956 // defined after an extload.
957 def : Pat<(extloadi64i32 addr:$src),
958 (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src),
961 // anyext. Define these to do an explicit zero-extend to
962 // avoid partial-register updates.
963 def : Pat<(i16 (anyext GR8 :$src)), (MOVZX16rr8 GR8 :$src)>;
964 def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8 GR8 :$src)>;
966 // Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
967 def : Pat<(i32 (anyext GR16:$src)),
968 (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
970 def : Pat<(i64 (anyext GR8 :$src)), (MOVZX64rr8 GR8 :$src)>;
971 def : Pat<(i64 (anyext GR16:$src)), (MOVZX64rr16 GR16 :$src)>;
972 def : Pat<(i64 (anyext GR32:$src)),
973 (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
976 // Any instruction that defines a 32-bit result leaves the high half of the
977 // register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
978 // be copying from a truncate. And x86's cmov doesn't do anything if the
979 // condition is false. But any other 32-bit operation will zero-extend
981 def def32 : PatLeaf<(i32 GR32:$src), [{
982 return N->getOpcode() != ISD::TRUNCATE &&
983 N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
984 N->getOpcode() != ISD::CopyFromReg &&
985 N->getOpcode() != X86ISD::CMOV;
988 // In the case of a 32-bit def that is known to implicitly zero-extend,
989 // we can use a SUBREG_TO_REG.
990 def : Pat<(i64 (zext def32:$src)),
991 (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
993 //===----------------------------------------------------------------------===//
994 // Pattern match OR as ADD
995 //===----------------------------------------------------------------------===//
997 // If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
998 // 3-addressified into an LEA instruction to avoid copies. However, we also
999 // want to finally emit these instructions as an or at the end of the code
1000 // generator to make the generated code easier to read. To do this, we select
1001 // into "disjoint bits" pseudo ops.
1003 // Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.
1004 def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
1005 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1006 return CurDAG->MaskedValueIsZero(N->getOperand(0), CN->getAPIntValue());
1008 unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
1009 APInt Mask = APInt::getAllOnesValue(BitWidth);
1010 APInt KnownZero0, KnownOne0;
1011 CurDAG->ComputeMaskedBits(N->getOperand(0), Mask, KnownZero0, KnownOne0, 0);
1012 APInt KnownZero1, KnownOne1;
1013 CurDAG->ComputeMaskedBits(N->getOperand(1), Mask, KnownZero1, KnownOne1, 0);
1014 return (~KnownZero0 & ~KnownZero1) == 0;
1018 // (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1019 let AddedComplexity = 5 in { // Try this before the selecting to OR
1021 let isConvertibleToThreeAddress = 1,
1022 Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1023 let isCommutable = 1 in {
1024 def ADD16rr_DB : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1025 "", // orw/addw REG, REG
1026 [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1027 def ADD32rr_DB : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1028 "", // orl/addl REG, REG
1029 [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1030 def ADD64rr_DB : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1031 "", // orq/addq REG, REG
1032 [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1035 // NOTE: These are order specific, we want the ri8 forms to be listed
1036 // first so that they are slightly preferred to the ri forms.
1038 def ADD16ri8_DB : I<0, Pseudo,
1039 (outs GR16:$dst), (ins GR16:$src1, i16i8imm:$src2),
1040 "", // orw/addw REG, imm8
1041 [(set GR16:$dst,(or_is_add GR16:$src1,i16immSExt8:$src2))]>;
1042 def ADD16ri_DB : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1043 "", // orw/addw REG, imm
1044 [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1046 def ADD32ri8_DB : I<0, Pseudo,
1047 (outs GR32:$dst), (ins GR32:$src1, i32i8imm:$src2),
1048 "", // orl/addl REG, imm8
1049 [(set GR32:$dst,(or_is_add GR32:$src1,i32immSExt8:$src2))]>;
1050 def ADD32ri_DB : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1051 "", // orl/addl REG, imm
1052 [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1055 def ADD64ri8_DB : I<0, Pseudo,
1056 (outs GR64:$dst), (ins GR64:$src1, i64i8imm:$src2),
1057 "", // orq/addq REG, imm8
1058 [(set GR64:$dst, (or_is_add GR64:$src1,
1059 i64immSExt8:$src2))]>;
1060 def ADD64ri32_DB : I<0, Pseudo,
1061 (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1062 "", // orq/addq REG, imm
1063 [(set GR64:$dst, (or_is_add GR64:$src1,
1064 i64immSExt32:$src2))]>;
1066 } // AddedComplexity
1069 //===----------------------------------------------------------------------===//
1071 //===----------------------------------------------------------------------===//
1073 // Odd encoding trick: -128 fits into an 8-bit immediate field while
1074 // +128 doesn't, so in this special case use a sub instead of an add.
1075 def : Pat<(add GR16:$src1, 128),
1076 (SUB16ri8 GR16:$src1, -128)>;
1077 def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1078 (SUB16mi8 addr:$dst, -128)>;
1080 def : Pat<(add GR32:$src1, 128),
1081 (SUB32ri8 GR32:$src1, -128)>;
1082 def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1083 (SUB32mi8 addr:$dst, -128)>;
1085 def : Pat<(add GR64:$src1, 128),
1086 (SUB64ri8 GR64:$src1, -128)>;
1087 def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1088 (SUB64mi8 addr:$dst, -128)>;
1090 // The same trick applies for 32-bit immediate fields in 64-bit
1092 def : Pat<(add GR64:$src1, 0x0000000080000000),
1093 (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1094 def : Pat<(store (add (loadi64 addr:$dst), 0x00000000800000000), addr:$dst),
1095 (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1097 // To avoid needing to materialize an immediate in a register, use a 32-bit and
1098 // with implicit zero-extension instead of a 64-bit and if the immediate has at
1099 // least 32 bits of leading zeros. If in addition the last 32 bits can be
1100 // represented with a sign extension of a 8 bit constant, use that.
1102 def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
1106 (EXTRACT_SUBREG GR64:$src, sub_32bit),
1107 (i32 (GetLo8XForm imm:$imm))),
1110 def : Pat<(and GR64:$src, i64immZExt32:$imm),
1114 (EXTRACT_SUBREG GR64:$src, sub_32bit),
1115 (i32 (GetLo32XForm imm:$imm))),
1119 // r & (2^16-1) ==> movz
1120 def : Pat<(and GR32:$src1, 0xffff),
1121 (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1122 // r & (2^8-1) ==> movz
1123 def : Pat<(and GR32:$src1, 0xff),
1124 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src1,
1127 Requires<[In32BitMode]>;
1128 // r & (2^8-1) ==> movz
1129 def : Pat<(and GR16:$src1, 0xff),
1130 (MOVZX16rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src1,
1133 Requires<[In32BitMode]>;
1135 // r & (2^32-1) ==> movz
1136 def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1137 (MOVZX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1138 // r & (2^16-1) ==> movz
1139 def : Pat<(and GR64:$src, 0xffff),
1140 (MOVZX64rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit)))>;
1141 // r & (2^8-1) ==> movz
1142 def : Pat<(and GR64:$src, 0xff),
1143 (MOVZX64rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit)))>;
1144 // r & (2^8-1) ==> movz
1145 def : Pat<(and GR32:$src1, 0xff),
1146 (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>,
1147 Requires<[In64BitMode]>;
1148 // r & (2^8-1) ==> movz
1149 def : Pat<(and GR16:$src1, 0xff),
1150 (MOVZX16rr8 (i8 (EXTRACT_SUBREG GR16:$src1, sub_8bit)))>,
1151 Requires<[In64BitMode]>;
1154 // sext_inreg patterns
1155 def : Pat<(sext_inreg GR32:$src, i16),
1156 (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1157 def : Pat<(sext_inreg GR32:$src, i8),
1158 (MOVSX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1161 Requires<[In32BitMode]>;
1162 def : Pat<(sext_inreg GR16:$src, i8),
1163 (MOVSX16rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1166 Requires<[In32BitMode]>;
1168 def : Pat<(sext_inreg GR64:$src, i32),
1169 (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1170 def : Pat<(sext_inreg GR64:$src, i16),
1171 (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1172 def : Pat<(sext_inreg GR64:$src, i8),
1173 (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1174 def : Pat<(sext_inreg GR32:$src, i8),
1175 (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>,
1176 Requires<[In64BitMode]>;
1177 def : Pat<(sext_inreg GR16:$src, i8),
1178 (MOVSX16rr8 (i8 (EXTRACT_SUBREG GR16:$src, sub_8bit)))>,
1179 Requires<[In64BitMode]>;
1183 def : Pat<(i16 (trunc GR32:$src)),
1184 (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1185 def : Pat<(i8 (trunc GR32:$src)),
1186 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1188 Requires<[In32BitMode]>;
1189 def : Pat<(i8 (trunc GR16:$src)),
1190 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1192 Requires<[In32BitMode]>;
1193 def : Pat<(i32 (trunc GR64:$src)),
1194 (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1195 def : Pat<(i16 (trunc GR64:$src)),
1196 (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1197 def : Pat<(i8 (trunc GR64:$src)),
1198 (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1199 def : Pat<(i8 (trunc GR32:$src)),
1200 (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1201 Requires<[In64BitMode]>;
1202 def : Pat<(i8 (trunc GR16:$src)),
1203 (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1204 Requires<[In64BitMode]>;
1206 // h-register tricks
1207 def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1208 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1210 Requires<[In32BitMode]>;
1211 def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1212 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1214 Requires<[In32BitMode]>;
1215 def : Pat<(srl GR16:$src, (i8 8)),
1218 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1221 Requires<[In32BitMode]>;
1222 def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1223 (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1226 Requires<[In32BitMode]>;
1227 def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1228 (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1231 Requires<[In32BitMode]>;
1232 def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1233 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1236 Requires<[In32BitMode]>;
1237 def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1238 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1241 Requires<[In32BitMode]>;
1243 // h-register tricks.
1244 // For now, be conservative on x86-64 and use an h-register extract only if the
1245 // value is immediately zero-extended or stored, which are somewhat common
1246 // cases. This uses a bunch of code to prevent a register requiring a REX prefix
1247 // from being allocated in the same instruction as the h register, as there's
1248 // currently no way to describe this requirement to the register allocator.
1250 // h-register extract and zero-extend.
1251 def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1255 (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1258 def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1260 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1262 Requires<[In64BitMode]>;
1263 def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1264 (MOVZX32_NOREXrr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1267 Requires<[In64BitMode]>;
1268 def : Pat<(srl GR16:$src, (i8 8)),
1271 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1274 Requires<[In64BitMode]>;
1275 def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1277 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1279 Requires<[In64BitMode]>;
1280 def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1282 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1284 Requires<[In64BitMode]>;
1285 def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1289 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1292 def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1296 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1300 // h-register extract and store.
1301 def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1304 (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1306 def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1309 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1311 Requires<[In64BitMode]>;
1312 def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1315 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1317 Requires<[In64BitMode]>;
1320 // (shl x, 1) ==> (add x, x)
1321 def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr GR8 :$src1, GR8 :$src1)>;
1322 def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1323 def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1324 def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1326 // (shl x (and y, 31)) ==> (shl x, y)
1327 def : Pat<(shl GR8:$src1, (and CL, 31)),
1328 (SHL8rCL GR8:$src1)>;
1329 def : Pat<(shl GR16:$src1, (and CL, 31)),
1330 (SHL16rCL GR16:$src1)>;
1331 def : Pat<(shl GR32:$src1, (and CL, 31)),
1332 (SHL32rCL GR32:$src1)>;
1333 def : Pat<(store (shl (loadi8 addr:$dst), (and CL, 31)), addr:$dst),
1334 (SHL8mCL addr:$dst)>;
1335 def : Pat<(store (shl (loadi16 addr:$dst), (and CL, 31)), addr:$dst),
1336 (SHL16mCL addr:$dst)>;
1337 def : Pat<(store (shl (loadi32 addr:$dst), (and CL, 31)), addr:$dst),
1338 (SHL32mCL addr:$dst)>;
1340 def : Pat<(srl GR8:$src1, (and CL, 31)),
1341 (SHR8rCL GR8:$src1)>;
1342 def : Pat<(srl GR16:$src1, (and CL, 31)),
1343 (SHR16rCL GR16:$src1)>;
1344 def : Pat<(srl GR32:$src1, (and CL, 31)),
1345 (SHR32rCL GR32:$src1)>;
1346 def : Pat<(store (srl (loadi8 addr:$dst), (and CL, 31)), addr:$dst),
1347 (SHR8mCL addr:$dst)>;
1348 def : Pat<(store (srl (loadi16 addr:$dst), (and CL, 31)), addr:$dst),
1349 (SHR16mCL addr:$dst)>;
1350 def : Pat<(store (srl (loadi32 addr:$dst), (and CL, 31)), addr:$dst),
1351 (SHR32mCL addr:$dst)>;
1353 def : Pat<(sra GR8:$src1, (and CL, 31)),
1354 (SAR8rCL GR8:$src1)>;
1355 def : Pat<(sra GR16:$src1, (and CL, 31)),
1356 (SAR16rCL GR16:$src1)>;
1357 def : Pat<(sra GR32:$src1, (and CL, 31)),
1358 (SAR32rCL GR32:$src1)>;
1359 def : Pat<(store (sra (loadi8 addr:$dst), (and CL, 31)), addr:$dst),
1360 (SAR8mCL addr:$dst)>;
1361 def : Pat<(store (sra (loadi16 addr:$dst), (and CL, 31)), addr:$dst),
1362 (SAR16mCL addr:$dst)>;
1363 def : Pat<(store (sra (loadi32 addr:$dst), (and CL, 31)), addr:$dst),
1364 (SAR32mCL addr:$dst)>;
1366 // (shl x (and y, 63)) ==> (shl x, y)
1367 def : Pat<(shl GR64:$src1, (and CL, 63)),
1368 (SHL64rCL GR64:$src1)>;
1369 def : Pat<(store (shl (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1370 (SHL64mCL addr:$dst)>;
1372 def : Pat<(srl GR64:$src1, (and CL, 63)),
1373 (SHR64rCL GR64:$src1)>;
1374 def : Pat<(store (srl (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1375 (SHR64mCL addr:$dst)>;
1377 def : Pat<(sra GR64:$src1, (and CL, 63)),
1378 (SAR64rCL GR64:$src1)>;
1379 def : Pat<(store (sra (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1380 (SAR64mCL addr:$dst)>;
1383 // (anyext (setcc_carry)) -> (setcc_carry)
1384 def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1386 def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1388 def : Pat<(i32 (anyext (i16 (X86setcc_c X86_COND_B, EFLAGS)))),
1394 //===----------------------------------------------------------------------===//
1395 // EFLAGS-defining Patterns
1396 //===----------------------------------------------------------------------===//
1399 def : Pat<(add GR8 :$src1, GR8 :$src2), (ADD8rr GR8 :$src1, GR8 :$src2)>;
1400 def : Pat<(add GR16:$src1, GR16:$src2), (ADD16rr GR16:$src1, GR16:$src2)>;
1401 def : Pat<(add GR32:$src1, GR32:$src2), (ADD32rr GR32:$src1, GR32:$src2)>;
1404 def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
1405 (ADD8rm GR8:$src1, addr:$src2)>;
1406 def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
1407 (ADD16rm GR16:$src1, addr:$src2)>;
1408 def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
1409 (ADD32rm GR32:$src1, addr:$src2)>;
1412 def : Pat<(add GR8 :$src1, imm:$src2), (ADD8ri GR8:$src1 , imm:$src2)>;
1413 def : Pat<(add GR16:$src1, imm:$src2), (ADD16ri GR16:$src1, imm:$src2)>;
1414 def : Pat<(add GR32:$src1, imm:$src2), (ADD32ri GR32:$src1, imm:$src2)>;
1415 def : Pat<(add GR16:$src1, i16immSExt8:$src2),
1416 (ADD16ri8 GR16:$src1, i16immSExt8:$src2)>;
1417 def : Pat<(add GR32:$src1, i32immSExt8:$src2),
1418 (ADD32ri8 GR32:$src1, i32immSExt8:$src2)>;
1421 def : Pat<(sub GR8 :$src1, GR8 :$src2), (SUB8rr GR8 :$src1, GR8 :$src2)>;
1422 def : Pat<(sub GR16:$src1, GR16:$src2), (SUB16rr GR16:$src1, GR16:$src2)>;
1423 def : Pat<(sub GR32:$src1, GR32:$src2), (SUB32rr GR32:$src1, GR32:$src2)>;
1426 def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
1427 (SUB8rm GR8:$src1, addr:$src2)>;
1428 def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
1429 (SUB16rm GR16:$src1, addr:$src2)>;
1430 def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
1431 (SUB32rm GR32:$src1, addr:$src2)>;
1434 def : Pat<(sub GR8:$src1, imm:$src2),
1435 (SUB8ri GR8:$src1, imm:$src2)>;
1436 def : Pat<(sub GR16:$src1, imm:$src2),
1437 (SUB16ri GR16:$src1, imm:$src2)>;
1438 def : Pat<(sub GR32:$src1, imm:$src2),
1439 (SUB32ri GR32:$src1, imm:$src2)>;
1440 def : Pat<(sub GR16:$src1, i16immSExt8:$src2),
1441 (SUB16ri8 GR16:$src1, i16immSExt8:$src2)>;
1442 def : Pat<(sub GR32:$src1, i32immSExt8:$src2),
1443 (SUB32ri8 GR32:$src1, i32immSExt8:$src2)>;
1446 def : Pat<(mul GR16:$src1, GR16:$src2),
1447 (IMUL16rr GR16:$src1, GR16:$src2)>;
1448 def : Pat<(mul GR32:$src1, GR32:$src2),
1449 (IMUL32rr GR32:$src1, GR32:$src2)>;
1452 def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
1453 (IMUL16rm GR16:$src1, addr:$src2)>;
1454 def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
1455 (IMUL32rm GR32:$src1, addr:$src2)>;
1458 def : Pat<(mul GR16:$src1, imm:$src2),
1459 (IMUL16rri GR16:$src1, imm:$src2)>;
1460 def : Pat<(mul GR32:$src1, imm:$src2),
1461 (IMUL32rri GR32:$src1, imm:$src2)>;
1462 def : Pat<(mul GR16:$src1, i16immSExt8:$src2),
1463 (IMUL16rri8 GR16:$src1, i16immSExt8:$src2)>;
1464 def : Pat<(mul GR32:$src1, i32immSExt8:$src2),
1465 (IMUL32rri8 GR32:$src1, i32immSExt8:$src2)>;
1467 // reg = mul mem, imm
1468 def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
1469 (IMUL16rmi addr:$src1, imm:$src2)>;
1470 def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
1471 (IMUL32rmi addr:$src1, imm:$src2)>;
1472 def : Pat<(mul (loadi16 addr:$src1), i16immSExt8:$src2),
1473 (IMUL16rmi8 addr:$src1, i16immSExt8:$src2)>;
1474 def : Pat<(mul (loadi32 addr:$src1), i32immSExt8:$src2),
1475 (IMUL32rmi8 addr:$src1, i32immSExt8:$src2)>;
1477 // Optimize multiply by 2 with EFLAGS result.
1478 let AddedComplexity = 2 in {
1479 def : Pat<(X86smul_flag GR16:$src1, 2), (ADD16rr GR16:$src1, GR16:$src1)>;
1480 def : Pat<(X86smul_flag GR32:$src1, 2), (ADD32rr GR32:$src1, GR32:$src1)>;
1483 // Patterns for nodes that do not produce flags, for instructions that do.
1486 def : Pat<(add GR64:$src1, GR64:$src2),
1487 (ADD64rr GR64:$src1, GR64:$src2)>;
1488 def : Pat<(add GR64:$src1, i64immSExt8:$src2),
1489 (ADD64ri8 GR64:$src1, i64immSExt8:$src2)>;
1490 def : Pat<(add GR64:$src1, i64immSExt32:$src2),
1491 (ADD64ri32 GR64:$src1, i64immSExt32:$src2)>;
1492 def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
1493 (ADD64rm GR64:$src1, addr:$src2)>;
1496 def : Pat<(sub GR64:$src1, GR64:$src2),
1497 (SUB64rr GR64:$src1, GR64:$src2)>;
1498 def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
1499 (SUB64rm GR64:$src1, addr:$src2)>;
1500 def : Pat<(sub GR64:$src1, i64immSExt8:$src2),
1501 (SUB64ri8 GR64:$src1, i64immSExt8:$src2)>;
1502 def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
1503 (SUB64ri32 GR64:$src1, i64immSExt32:$src2)>;
1506 def : Pat<(mul GR64:$src1, GR64:$src2),
1507 (IMUL64rr GR64:$src1, GR64:$src2)>;
1508 def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
1509 (IMUL64rm GR64:$src1, addr:$src2)>;
1510 def : Pat<(mul GR64:$src1, i64immSExt8:$src2),
1511 (IMUL64rri8 GR64:$src1, i64immSExt8:$src2)>;
1512 def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
1513 (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
1514 def : Pat<(mul (loadi64 addr:$src1), i64immSExt8:$src2),
1515 (IMUL64rmi8 addr:$src1, i64immSExt8:$src2)>;
1516 def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
1517 (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
1520 def : Pat<(add GR8 :$src, 1), (INC8r GR8 :$src)>;
1521 def : Pat<(add GR16:$src, 1), (INC16r GR16:$src)>, Requires<[In32BitMode]>;
1522 def : Pat<(add GR16:$src, 1), (INC64_16r GR16:$src)>, Requires<[In64BitMode]>;
1523 def : Pat<(add GR32:$src, 1), (INC32r GR32:$src)>, Requires<[In32BitMode]>;
1524 def : Pat<(add GR32:$src, 1), (INC64_32r GR32:$src)>, Requires<[In64BitMode]>;
1525 def : Pat<(add GR64:$src, 1), (INC64r GR64:$src)>;
1528 def : Pat<(add GR8 :$src, -1), (DEC8r GR8 :$src)>;
1529 def : Pat<(add GR16:$src, -1), (DEC16r GR16:$src)>, Requires<[In32BitMode]>;
1530 def : Pat<(add GR16:$src, -1), (DEC64_16r GR16:$src)>, Requires<[In64BitMode]>;
1531 def : Pat<(add GR32:$src, -1), (DEC32r GR32:$src)>, Requires<[In32BitMode]>;
1532 def : Pat<(add GR32:$src, -1), (DEC64_32r GR32:$src)>, Requires<[In64BitMode]>;
1533 def : Pat<(add GR64:$src, -1), (DEC64r GR64:$src)>;
1536 def : Pat<(or GR8 :$src1, GR8 :$src2), (OR8rr GR8 :$src1, GR8 :$src2)>;
1537 def : Pat<(or GR16:$src1, GR16:$src2), (OR16rr GR16:$src1, GR16:$src2)>;
1538 def : Pat<(or GR32:$src1, GR32:$src2), (OR32rr GR32:$src1, GR32:$src2)>;
1539 def : Pat<(or GR64:$src1, GR64:$src2), (OR64rr GR64:$src1, GR64:$src2)>;
1542 def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
1543 (OR8rm GR8:$src1, addr:$src2)>;
1544 def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
1545 (OR16rm GR16:$src1, addr:$src2)>;
1546 def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
1547 (OR32rm GR32:$src1, addr:$src2)>;
1548 def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
1549 (OR64rm GR64:$src1, addr:$src2)>;
1552 def : Pat<(or GR8:$src1 , imm:$src2), (OR8ri GR8 :$src1, imm:$src2)>;
1553 def : Pat<(or GR16:$src1, imm:$src2), (OR16ri GR16:$src1, imm:$src2)>;
1554 def : Pat<(or GR32:$src1, imm:$src2), (OR32ri GR32:$src1, imm:$src2)>;
1555 def : Pat<(or GR16:$src1, i16immSExt8:$src2),
1556 (OR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1557 def : Pat<(or GR32:$src1, i32immSExt8:$src2),
1558 (OR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1559 def : Pat<(or GR64:$src1, i64immSExt8:$src2),
1560 (OR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1561 def : Pat<(or GR64:$src1, i64immSExt32:$src2),
1562 (OR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1565 def : Pat<(xor GR8 :$src1, GR8 :$src2), (XOR8rr GR8 :$src1, GR8 :$src2)>;
1566 def : Pat<(xor GR16:$src1, GR16:$src2), (XOR16rr GR16:$src1, GR16:$src2)>;
1567 def : Pat<(xor GR32:$src1, GR32:$src2), (XOR32rr GR32:$src1, GR32:$src2)>;
1568 def : Pat<(xor GR64:$src1, GR64:$src2), (XOR64rr GR64:$src1, GR64:$src2)>;
1571 def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
1572 (XOR8rm GR8:$src1, addr:$src2)>;
1573 def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
1574 (XOR16rm GR16:$src1, addr:$src2)>;
1575 def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
1576 (XOR32rm GR32:$src1, addr:$src2)>;
1577 def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
1578 (XOR64rm GR64:$src1, addr:$src2)>;
1581 def : Pat<(xor GR8:$src1, imm:$src2),
1582 (XOR8ri GR8:$src1, imm:$src2)>;
1583 def : Pat<(xor GR16:$src1, imm:$src2),
1584 (XOR16ri GR16:$src1, imm:$src2)>;
1585 def : Pat<(xor GR32:$src1, imm:$src2),
1586 (XOR32ri GR32:$src1, imm:$src2)>;
1587 def : Pat<(xor GR16:$src1, i16immSExt8:$src2),
1588 (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1589 def : Pat<(xor GR32:$src1, i32immSExt8:$src2),
1590 (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1591 def : Pat<(xor GR64:$src1, i64immSExt8:$src2),
1592 (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1593 def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
1594 (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1597 def : Pat<(and GR8 :$src1, GR8 :$src2), (AND8rr GR8 :$src1, GR8 :$src2)>;
1598 def : Pat<(and GR16:$src1, GR16:$src2), (AND16rr GR16:$src1, GR16:$src2)>;
1599 def : Pat<(and GR32:$src1, GR32:$src2), (AND32rr GR32:$src1, GR32:$src2)>;
1600 def : Pat<(and GR64:$src1, GR64:$src2), (AND64rr GR64:$src1, GR64:$src2)>;
1603 def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
1604 (AND8rm GR8:$src1, addr:$src2)>;
1605 def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
1606 (AND16rm GR16:$src1, addr:$src2)>;
1607 def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
1608 (AND32rm GR32:$src1, addr:$src2)>;
1609 def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
1610 (AND64rm GR64:$src1, addr:$src2)>;
1613 def : Pat<(and GR8:$src1, imm:$src2),
1614 (AND8ri GR8:$src1, imm:$src2)>;
1615 def : Pat<(and GR16:$src1, imm:$src2),
1616 (AND16ri GR16:$src1, imm:$src2)>;
1617 def : Pat<(and GR32:$src1, imm:$src2),
1618 (AND32ri GR32:$src1, imm:$src2)>;
1619 def : Pat<(and GR16:$src1, i16immSExt8:$src2),
1620 (AND16ri8 GR16:$src1, i16immSExt8:$src2)>;
1621 def : Pat<(and GR32:$src1, i32immSExt8:$src2),
1622 (AND32ri8 GR32:$src1, i32immSExt8:$src2)>;
1623 def : Pat<(and GR64:$src1, i64immSExt8:$src2),
1624 (AND64ri8 GR64:$src1, i64immSExt8:$src2)>;
1625 def : Pat<(and GR64:$src1, i64immSExt32:$src2),
1626 (AND64ri32 GR64:$src1, i64immSExt32:$src2)>;