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 // (add OP, SETB) -> (adc OP, 0)
218 def : Pat<(add (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR8:$op),
219 (ADC8ri GR8:$op, 0)>;
220 def : Pat<(add (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR32:$op),
221 (ADC32ri8 GR32:$op, 0)>;
222 def : Pat<(add (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR64:$op),
223 (ADC64ri8 GR64:$op, 0)>;
225 // (sub OP, SETB) -> (sbb OP, 0)
226 def : Pat<(sub GR8:$op, (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
227 (SBB8ri GR8:$op, 0)>;
228 def : Pat<(sub GR32:$op, (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
229 (SBB32ri8 GR32:$op, 0)>;
230 def : Pat<(sub GR64:$op, (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
231 (SBB64ri8 GR64:$op, 0)>;
233 // (sub OP, SETCC_CARRY) -> (adc OP, 0)
234 def : Pat<(sub GR8:$op, (i8 (X86setcc_c X86_COND_B, EFLAGS))),
235 (ADC8ri GR8:$op, 0)>;
236 def : Pat<(sub GR32:$op, (i32 (X86setcc_c X86_COND_B, EFLAGS))),
237 (ADC32ri8 GR32:$op, 0)>;
238 def : Pat<(sub GR64:$op, (i64 (X86setcc_c X86_COND_B, EFLAGS))),
239 (ADC64ri8 GR64:$op, 0)>;
241 //===----------------------------------------------------------------------===//
242 // String Pseudo Instructions
244 let Defs = [ECX,EDI,ESI], Uses = [ECX,EDI,ESI], isCodeGenOnly = 1 in {
245 def REP_MOVSB : I<0xA4, RawFrm, (outs), (ins), "{rep;movsb|rep movsb}",
246 [(X86rep_movs i8)]>, REP;
247 def REP_MOVSW : I<0xA5, RawFrm, (outs), (ins), "{rep;movsw|rep movsw}",
248 [(X86rep_movs i16)]>, REP, OpSize;
249 def REP_MOVSD : I<0xA5, RawFrm, (outs), (ins), "{rep;movsl|rep movsd}",
250 [(X86rep_movs i32)]>, REP;
253 let Defs = [RCX,RDI,RSI], Uses = [RCX,RDI,RSI], isCodeGenOnly = 1 in
254 def REP_MOVSQ : RI<0xA5, RawFrm, (outs), (ins), "{rep;movsq|rep movsq}",
255 [(X86rep_movs i64)]>, REP;
258 // FIXME: Should use "(X86rep_stos AL)" as the pattern.
259 let Defs = [ECX,EDI], Uses = [AL,ECX,EDI], isCodeGenOnly = 1 in
260 def REP_STOSB : I<0xAA, RawFrm, (outs), (ins), "{rep;stosb|rep stosb}",
261 [(X86rep_stos i8)]>, REP;
262 let Defs = [ECX,EDI], Uses = [AX,ECX,EDI], isCodeGenOnly = 1 in
263 def REP_STOSW : I<0xAB, RawFrm, (outs), (ins), "{rep;stosw|rep stosw}",
264 [(X86rep_stos i16)]>, REP, OpSize;
265 let Defs = [ECX,EDI], Uses = [EAX,ECX,EDI], isCodeGenOnly = 1 in
266 def REP_STOSD : I<0xAB, RawFrm, (outs), (ins), "{rep;stosl|rep stosd}",
267 [(X86rep_stos i32)]>, REP;
269 let Defs = [RCX,RDI], Uses = [RAX,RCX,RDI], isCodeGenOnly = 1 in
270 def REP_STOSQ : RI<0xAB, RawFrm, (outs), (ins), "{rep;stosq|rep stosq}",
271 [(X86rep_stos i64)]>, REP;
274 //===----------------------------------------------------------------------===//
275 // Thread Local Storage Instructions
279 // All calls clobber the non-callee saved registers. ESP is marked as
280 // a use to prevent stack-pointer assignments that appear immediately
281 // before calls from potentially appearing dead.
282 let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0,
283 MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
284 XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
285 XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
287 def TLS_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
289 [(X86tlsaddr tls32addr:$sym)]>,
290 Requires<[In32BitMode]>;
292 // All calls clobber the non-callee saved registers. RSP is marked as
293 // a use to prevent stack-pointer assignments that appear immediately
294 // before calls from potentially appearing dead.
295 let Defs = [RAX, RCX, RDX, RSI, RDI, R8, R9, R10, R11,
296 FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0, ST1,
297 MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
298 XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
299 XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
301 def TLS_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
303 [(X86tlsaddr tls64addr:$sym)]>,
304 Requires<[In64BitMode]>;
306 // Darwin TLS Support
307 // For i386, the address of the thunk is passed on the stack, on return the
308 // address of the variable is in %eax. %ecx is trashed during the function
309 // call. All other registers are preserved.
310 let Defs = [EAX, ECX, EFLAGS],
312 usesCustomInserter = 1 in
313 def TLSCall_32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
315 [(X86TLSCall addr:$sym)]>,
316 Requires<[In32BitMode]>;
318 // For x86_64, the address of the thunk is passed in %rdi, on return
319 // the address of the variable is in %rax. All other registers are preserved.
320 let Defs = [RAX, EFLAGS],
322 usesCustomInserter = 1 in
323 def TLSCall_64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
325 [(X86TLSCall addr:$sym)]>,
326 Requires<[In64BitMode]>;
329 //===----------------------------------------------------------------------===//
330 // Conditional Move Pseudo Instructions
332 let Constraints = "$src1 = $dst" in {
335 let Uses = [EFLAGS] in {
337 // X86 doesn't have 8-bit conditional moves. Use a customInserter to
338 // emit control flow. An alternative to this is to mark i8 SELECT as Promote,
339 // however that requires promoting the operands, and can induce additional
340 // i8 register pressure. Note that CMOV_GR8 is conservatively considered to
341 // clobber EFLAGS, because if one of the operands is zero, the expansion
342 // could involve an xor.
343 let usesCustomInserter = 1, Constraints = "", Defs = [EFLAGS] in {
344 def CMOV_GR8 : I<0, Pseudo,
345 (outs GR8:$dst), (ins GR8:$src1, GR8:$src2, i8imm:$cond),
347 [(set GR8:$dst, (X86cmov GR8:$src1, GR8:$src2,
348 imm:$cond, EFLAGS))]>;
350 let Predicates = [NoCMov] in {
351 def CMOV_GR32 : I<0, Pseudo,
352 (outs GR32:$dst), (ins GR32:$src1, GR32:$src2, i8imm:$cond),
353 "#CMOV_GR32* PSEUDO!",
355 (X86cmov GR32:$src1, GR32:$src2, imm:$cond, EFLAGS))]>;
356 def CMOV_GR16 : I<0, Pseudo,
357 (outs GR16:$dst), (ins GR16:$src1, GR16:$src2, i8imm:$cond),
358 "#CMOV_GR16* PSEUDO!",
360 (X86cmov GR16:$src1, GR16:$src2, imm:$cond, EFLAGS))]>;
361 def CMOV_RFP32 : I<0, Pseudo,
363 (ins RFP32:$src1, RFP32:$src2, i8imm:$cond),
364 "#CMOV_RFP32 PSEUDO!",
366 (X86cmov RFP32:$src1, RFP32:$src2, imm:$cond,
368 def CMOV_RFP64 : I<0, Pseudo,
370 (ins RFP64:$src1, RFP64:$src2, i8imm:$cond),
371 "#CMOV_RFP64 PSEUDO!",
373 (X86cmov RFP64:$src1, RFP64:$src2, imm:$cond,
375 def CMOV_RFP80 : I<0, Pseudo,
377 (ins RFP80:$src1, RFP80:$src2, i8imm:$cond),
378 "#CMOV_RFP80 PSEUDO!",
380 (X86cmov RFP80:$src1, RFP80:$src2, imm:$cond,
382 } // Predicates = [NoCMov]
383 } // UsesCustomInserter = 1, Constraints = "", Defs = [EFLAGS]
386 } // Constraints = "$src1 = $dst" in
389 //===----------------------------------------------------------------------===//
390 // Atomic Instruction Pseudo Instructions
391 //===----------------------------------------------------------------------===//
393 // Atomic exchange, and, or, xor
394 let Constraints = "$val = $dst", Defs = [EFLAGS],
395 usesCustomInserter = 1 in {
397 def ATOMAND8 : I<0, Pseudo, (outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
399 [(set GR8:$dst, (atomic_load_and_8 addr:$ptr, GR8:$val))]>;
400 def ATOMOR8 : I<0, Pseudo, (outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
402 [(set GR8:$dst, (atomic_load_or_8 addr:$ptr, GR8:$val))]>;
403 def ATOMXOR8 : I<0, Pseudo,(outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
405 [(set GR8:$dst, (atomic_load_xor_8 addr:$ptr, GR8:$val))]>;
406 def ATOMNAND8 : I<0, Pseudo,(outs GR8:$dst),(ins i8mem:$ptr, GR8:$val),
407 "#ATOMNAND8 PSEUDO!",
408 [(set GR8:$dst, (atomic_load_nand_8 addr:$ptr, GR8:$val))]>;
410 def ATOMAND16 : I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
411 "#ATOMAND16 PSEUDO!",
412 [(set GR16:$dst, (atomic_load_and_16 addr:$ptr, GR16:$val))]>;
413 def ATOMOR16 : I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
415 [(set GR16:$dst, (atomic_load_or_16 addr:$ptr, GR16:$val))]>;
416 def ATOMXOR16 : I<0, Pseudo,(outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
417 "#ATOMXOR16 PSEUDO!",
418 [(set GR16:$dst, (atomic_load_xor_16 addr:$ptr, GR16:$val))]>;
419 def ATOMNAND16 : I<0, Pseudo,(outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
420 "#ATOMNAND16 PSEUDO!",
421 [(set GR16:$dst, (atomic_load_nand_16 addr:$ptr, GR16:$val))]>;
422 def ATOMMIN16: I<0, Pseudo, (outs GR16:$dst), (ins i16mem:$ptr, GR16:$val),
423 "#ATOMMIN16 PSEUDO!",
424 [(set GR16:$dst, (atomic_load_min_16 addr:$ptr, GR16:$val))]>;
425 def ATOMMAX16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
426 "#ATOMMAX16 PSEUDO!",
427 [(set GR16:$dst, (atomic_load_max_16 addr:$ptr, GR16:$val))]>;
428 def ATOMUMIN16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
429 "#ATOMUMIN16 PSEUDO!",
430 [(set GR16:$dst, (atomic_load_umin_16 addr:$ptr, GR16:$val))]>;
431 def ATOMUMAX16: I<0, Pseudo, (outs GR16:$dst),(ins i16mem:$ptr, GR16:$val),
432 "#ATOMUMAX16 PSEUDO!",
433 [(set GR16:$dst, (atomic_load_umax_16 addr:$ptr, GR16:$val))]>;
436 def ATOMAND32 : I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
437 "#ATOMAND32 PSEUDO!",
438 [(set GR32:$dst, (atomic_load_and_32 addr:$ptr, GR32:$val))]>;
439 def ATOMOR32 : I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
441 [(set GR32:$dst, (atomic_load_or_32 addr:$ptr, GR32:$val))]>;
442 def ATOMXOR32 : I<0, Pseudo,(outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
443 "#ATOMXOR32 PSEUDO!",
444 [(set GR32:$dst, (atomic_load_xor_32 addr:$ptr, GR32:$val))]>;
445 def ATOMNAND32 : I<0, Pseudo,(outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
446 "#ATOMNAND32 PSEUDO!",
447 [(set GR32:$dst, (atomic_load_nand_32 addr:$ptr, GR32:$val))]>;
448 def ATOMMIN32: I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$ptr, GR32:$val),
449 "#ATOMMIN32 PSEUDO!",
450 [(set GR32:$dst, (atomic_load_min_32 addr:$ptr, GR32:$val))]>;
451 def ATOMMAX32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
452 "#ATOMMAX32 PSEUDO!",
453 [(set GR32:$dst, (atomic_load_max_32 addr:$ptr, GR32:$val))]>;
454 def ATOMUMIN32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
455 "#ATOMUMIN32 PSEUDO!",
456 [(set GR32:$dst, (atomic_load_umin_32 addr:$ptr, GR32:$val))]>;
457 def ATOMUMAX32: I<0, Pseudo, (outs GR32:$dst),(ins i32mem:$ptr, GR32:$val),
458 "#ATOMUMAX32 PSEUDO!",
459 [(set GR32:$dst, (atomic_load_umax_32 addr:$ptr, GR32:$val))]>;
463 def ATOMAND64 : I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
464 "#ATOMAND64 PSEUDO!",
465 [(set GR64:$dst, (atomic_load_and_64 addr:$ptr, GR64:$val))]>;
466 def ATOMOR64 : I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
468 [(set GR64:$dst, (atomic_load_or_64 addr:$ptr, GR64:$val))]>;
469 def ATOMXOR64 : I<0, Pseudo,(outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
470 "#ATOMXOR64 PSEUDO!",
471 [(set GR64:$dst, (atomic_load_xor_64 addr:$ptr, GR64:$val))]>;
472 def ATOMNAND64 : I<0, Pseudo,(outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
473 "#ATOMNAND64 PSEUDO!",
474 [(set GR64:$dst, (atomic_load_nand_64 addr:$ptr, GR64:$val))]>;
475 def ATOMMIN64: I<0, Pseudo, (outs GR64:$dst), (ins i64mem:$ptr, GR64:$val),
476 "#ATOMMIN64 PSEUDO!",
477 [(set GR64:$dst, (atomic_load_min_64 addr:$ptr, GR64:$val))]>;
478 def ATOMMAX64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
479 "#ATOMMAX64 PSEUDO!",
480 [(set GR64:$dst, (atomic_load_max_64 addr:$ptr, GR64:$val))]>;
481 def ATOMUMIN64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
482 "#ATOMUMIN64 PSEUDO!",
483 [(set GR64:$dst, (atomic_load_umin_64 addr:$ptr, GR64:$val))]>;
484 def ATOMUMAX64: I<0, Pseudo, (outs GR64:$dst),(ins i64mem:$ptr, GR64:$val),
485 "#ATOMUMAX64 PSEUDO!",
486 [(set GR64:$dst, (atomic_load_umax_64 addr:$ptr, GR64:$val))]>;
489 let Constraints = "$val1 = $dst1, $val2 = $dst2",
490 Defs = [EFLAGS, EAX, EBX, ECX, EDX],
491 Uses = [EAX, EBX, ECX, EDX],
492 mayLoad = 1, mayStore = 1,
493 usesCustomInserter = 1 in {
494 def ATOMAND6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
495 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
496 "#ATOMAND6432 PSEUDO!", []>;
497 def ATOMOR6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
498 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
499 "#ATOMOR6432 PSEUDO!", []>;
500 def ATOMXOR6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
501 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
502 "#ATOMXOR6432 PSEUDO!", []>;
503 def ATOMNAND6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
504 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
505 "#ATOMNAND6432 PSEUDO!", []>;
506 def ATOMADD6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
507 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
508 "#ATOMADD6432 PSEUDO!", []>;
509 def ATOMSUB6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
510 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
511 "#ATOMSUB6432 PSEUDO!", []>;
512 def ATOMSWAP6432 : I<0, Pseudo, (outs GR32:$dst1, GR32:$dst2),
513 (ins i64mem:$ptr, GR32:$val1, GR32:$val2),
514 "#ATOMSWAP6432 PSEUDO!", []>;
517 //===----------------------------------------------------------------------===//
518 // Normal-Instructions-With-Lock-Prefix Pseudo Instructions
519 //===----------------------------------------------------------------------===//
521 // FIXME: Use normal instructions and add lock prefix dynamically.
525 // TODO: Get this to fold the constant into the instruction.
526 let isCodeGenOnly = 1 in
527 def OR32mrLocked : I<0x09, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$zero),
529 "or{l}\t{$zero, $dst|$dst, $zero}",
530 []>, Requires<[In32BitMode]>, LOCK;
532 let hasSideEffects = 1 in
533 def Int_MemBarrier : I<0, Pseudo, (outs), (ins),
535 [(X86MemBarrier)]>, Requires<[HasSSE2]>;
537 // TODO: Get this to fold the constant into the instruction.
538 let hasSideEffects = 1, Defs = [ESP], isCodeGenOnly = 1 in
539 def Int_MemBarrierNoSSE64 : RI<0x09, MRM1r, (outs), (ins GR64:$zero),
541 "or{q}\t{$zero, (%rsp)|(%rsp), $zero}",
542 [(X86MemBarrierNoSSE GR64:$zero)]>,
543 Requires<[In64BitMode]>, LOCK;
546 // RegOpc corresponds to the mr version of the instruction
547 // ImmOpc corresponds to the mi version of the instruction
548 // ImmOpc8 corresponds to the mi8 version of the instruction
549 // ImmMod corresponds to the instruction format of the mi and mi8 versions
550 multiclass LOCK_ArithBinOp<bits<8> RegOpc, bits<8> ImmOpc, bits<8> ImmOpc8,
551 Format ImmMod, string mnemonic> {
552 let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1 in {
554 def #NAME#8mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
555 RegOpc{3}, RegOpc{2}, RegOpc{1}, 0 },
556 MRMDestMem, (outs), (ins i8mem:$dst, GR8:$src2),
557 !strconcat("lock\n\t", mnemonic, "{b}\t",
558 "{$src2, $dst|$dst, $src2}"),
560 def #NAME#16mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
561 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
562 MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
563 !strconcat("lock\n\t", mnemonic, "{w}\t",
564 "{$src2, $dst|$dst, $src2}"),
566 def #NAME#32mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
567 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
568 MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
569 !strconcat("lock\n\t", mnemonic, "{l}\t",
570 "{$src2, $dst|$dst, $src2}"),
572 def #NAME#64mr : RI<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
573 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
574 MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
575 !strconcat("lock\n\t", mnemonic, "{q}\t",
576 "{$src2, $dst|$dst, $src2}"),
579 def #NAME#8mi : Ii8<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
580 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 0 },
581 ImmMod, (outs), (ins i8mem :$dst, i8imm :$src2),
582 !strconcat("lock\n\t", mnemonic, "{b}\t",
583 "{$src2, $dst|$dst, $src2}"),
586 def #NAME#16mi : Ii16<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
587 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
588 ImmMod, (outs), (ins i16mem :$dst, i16imm :$src2),
589 !strconcat("lock\n\t", mnemonic, "{w}\t",
590 "{$src2, $dst|$dst, $src2}"),
593 def #NAME#32mi : Ii32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
594 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
595 ImmMod, (outs), (ins i32mem :$dst, i32imm :$src2),
596 !strconcat("lock\n\t", mnemonic, "{l}\t",
597 "{$src2, $dst|$dst, $src2}"),
600 def #NAME#64mi32 : RIi32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
601 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
602 ImmMod, (outs), (ins i64mem :$dst, i64i32imm :$src2),
603 !strconcat("lock\n\t", mnemonic, "{q}\t",
604 "{$src2, $dst|$dst, $src2}"),
607 def #NAME#16mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
608 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
609 ImmMod, (outs), (ins i16mem :$dst, i16i8imm :$src2),
610 !strconcat("lock\n\t", mnemonic, "{w}\t",
611 "{$src2, $dst|$dst, $src2}"),
613 def #NAME#32mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
614 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
615 ImmMod, (outs), (ins i32mem :$dst, i32i8imm :$src2),
616 !strconcat("lock\n\t", mnemonic, "{l}\t",
617 "{$src2, $dst|$dst, $src2}"),
619 def #NAME#64mi8 : RIi8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
620 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
621 ImmMod, (outs), (ins i64mem :$dst, i64i8imm :$src2),
622 !strconcat("lock\n\t", mnemonic, "{q}\t",
623 "{$src2, $dst|$dst, $src2}"),
630 defm LOCK_ADD : LOCK_ArithBinOp<0x00, 0x80, 0x83, MRM0m, "add">;
631 defm LOCK_SUB : LOCK_ArithBinOp<0x28, 0x80, 0x83, MRM5m, "sub">;
632 defm LOCK_OR : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM1m, "or">;
633 defm LOCK_AND : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM4m, "and">;
634 defm LOCK_XOR : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM6m, "xor">;
636 // Optimized codegen when the non-memory output is not used.
637 let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1 in {
639 def LOCK_INC8m : I<0xFE, MRM0m, (outs), (ins i8mem :$dst),
641 "inc{b}\t$dst", []>, LOCK;
642 def LOCK_INC16m : I<0xFF, MRM0m, (outs), (ins i16mem:$dst),
644 "inc{w}\t$dst", []>, OpSize, LOCK;
645 def LOCK_INC32m : I<0xFF, MRM0m, (outs), (ins i32mem:$dst),
647 "inc{l}\t$dst", []>, LOCK;
648 def LOCK_INC64m : RI<0xFF, MRM0m, (outs), (ins i64mem:$dst),
650 "inc{q}\t$dst", []>, LOCK;
652 def LOCK_DEC8m : I<0xFE, MRM1m, (outs), (ins i8mem :$dst),
654 "dec{b}\t$dst", []>, LOCK;
655 def LOCK_DEC16m : I<0xFF, MRM1m, (outs), (ins i16mem:$dst),
657 "dec{w}\t$dst", []>, OpSize, LOCK;
658 def LOCK_DEC32m : I<0xFF, MRM1m, (outs), (ins i32mem:$dst),
660 "dec{l}\t$dst", []>, LOCK;
661 def LOCK_DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst),
663 "dec{q}\t$dst", []>, LOCK;
666 // Atomic compare and swap.
667 let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
668 isCodeGenOnly = 1 in {
669 def LCMPXCHG8B : I<0xC7, MRM1m, (outs), (ins i64mem:$ptr),
672 [(X86cas8 addr:$ptr)]>, TB, LOCK;
674 let Defs = [AL, EFLAGS], Uses = [AL], isCodeGenOnly = 1 in {
675 def LCMPXCHG8 : I<0xB0, MRMDestMem, (outs), (ins i8mem:$ptr, GR8:$swap),
677 "cmpxchg{b}\t{$swap, $ptr|$ptr, $swap}",
678 [(X86cas addr:$ptr, GR8:$swap, 1)]>, TB, LOCK;
681 let Defs = [AX, EFLAGS], Uses = [AX], isCodeGenOnly = 1 in {
682 def LCMPXCHG16 : I<0xB1, MRMDestMem, (outs), (ins i16mem:$ptr, GR16:$swap),
684 "cmpxchg{w}\t{$swap, $ptr|$ptr, $swap}",
685 [(X86cas addr:$ptr, GR16:$swap, 2)]>, TB, OpSize, LOCK;
688 let Defs = [EAX, EFLAGS], Uses = [EAX], isCodeGenOnly = 1 in {
689 def LCMPXCHG32 : I<0xB1, MRMDestMem, (outs), (ins i32mem:$ptr, GR32:$swap),
691 "cmpxchg{l}\t{$swap, $ptr|$ptr, $swap}",
692 [(X86cas addr:$ptr, GR32:$swap, 4)]>, TB, LOCK;
695 let Defs = [RAX, EFLAGS], Uses = [RAX], isCodeGenOnly = 1 in {
696 def LCMPXCHG64 : RI<0xB1, MRMDestMem, (outs), (ins i64mem:$ptr, GR64:$swap),
698 "cmpxchgq\t$swap,$ptr",
699 [(X86cas addr:$ptr, GR64:$swap, 8)]>, TB, LOCK;
702 // Atomic exchange and add
703 let Constraints = "$val = $dst", Defs = [EFLAGS], isCodeGenOnly = 1 in {
704 def LXADD8 : I<0xC0, MRMSrcMem, (outs GR8:$dst), (ins GR8:$val, i8mem:$ptr),
706 "xadd{b}\t{$val, $ptr|$ptr, $val}",
707 [(set GR8:$dst, (atomic_load_add_8 addr:$ptr, GR8:$val))]>,
709 def LXADD16 : I<0xC1, MRMSrcMem, (outs GR16:$dst), (ins GR16:$val, i16mem:$ptr),
711 "xadd{w}\t{$val, $ptr|$ptr, $val}",
712 [(set GR16:$dst, (atomic_load_add_16 addr:$ptr, GR16:$val))]>,
714 def LXADD32 : I<0xC1, MRMSrcMem, (outs GR32:$dst), (ins GR32:$val, i32mem:$ptr),
716 "xadd{l}\t{$val, $ptr|$ptr, $val}",
717 [(set GR32:$dst, (atomic_load_add_32 addr:$ptr, GR32:$val))]>,
719 def LXADD64 : RI<0xC1, MRMSrcMem, (outs GR64:$dst), (ins GR64:$val,i64mem:$ptr),
722 [(set GR64:$dst, (atomic_load_add_64 addr:$ptr, GR64:$val))]>,
726 //===----------------------------------------------------------------------===//
727 // Conditional Move Pseudo Instructions.
728 //===----------------------------------------------------------------------===//
731 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded after
732 // instruction selection into a branch sequence.
733 let Uses = [EFLAGS], usesCustomInserter = 1 in {
734 def CMOV_FR32 : I<0, Pseudo,
735 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
736 "#CMOV_FR32 PSEUDO!",
737 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
739 def CMOV_FR64 : I<0, Pseudo,
740 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
741 "#CMOV_FR64 PSEUDO!",
742 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
744 def CMOV_V4F32 : I<0, Pseudo,
745 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
746 "#CMOV_V4F32 PSEUDO!",
748 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
750 def CMOV_V2F64 : I<0, Pseudo,
751 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
752 "#CMOV_V2F64 PSEUDO!",
754 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
756 def CMOV_V2I64 : I<0, Pseudo,
757 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
758 "#CMOV_V2I64 PSEUDO!",
760 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
765 //===----------------------------------------------------------------------===//
766 // DAG Pattern Matching Rules
767 //===----------------------------------------------------------------------===//
769 // ConstantPool GlobalAddress, ExternalSymbol, and JumpTable
770 def : Pat<(i32 (X86Wrapper tconstpool :$dst)), (MOV32ri tconstpool :$dst)>;
771 def : Pat<(i32 (X86Wrapper tjumptable :$dst)), (MOV32ri tjumptable :$dst)>;
772 def : Pat<(i32 (X86Wrapper tglobaltlsaddr:$dst)),(MOV32ri tglobaltlsaddr:$dst)>;
773 def : Pat<(i32 (X86Wrapper tglobaladdr :$dst)), (MOV32ri tglobaladdr :$dst)>;
774 def : Pat<(i32 (X86Wrapper texternalsym:$dst)), (MOV32ri texternalsym:$dst)>;
775 def : Pat<(i32 (X86Wrapper tblockaddress:$dst)), (MOV32ri tblockaddress:$dst)>;
777 def : Pat<(add GR32:$src1, (X86Wrapper tconstpool:$src2)),
778 (ADD32ri GR32:$src1, tconstpool:$src2)>;
779 def : Pat<(add GR32:$src1, (X86Wrapper tjumptable:$src2)),
780 (ADD32ri GR32:$src1, tjumptable:$src2)>;
781 def : Pat<(add GR32:$src1, (X86Wrapper tglobaladdr :$src2)),
782 (ADD32ri GR32:$src1, tglobaladdr:$src2)>;
783 def : Pat<(add GR32:$src1, (X86Wrapper texternalsym:$src2)),
784 (ADD32ri GR32:$src1, texternalsym:$src2)>;
785 def : Pat<(add GR32:$src1, (X86Wrapper tblockaddress:$src2)),
786 (ADD32ri GR32:$src1, tblockaddress:$src2)>;
788 def : Pat<(store (i32 (X86Wrapper tglobaladdr:$src)), addr:$dst),
789 (MOV32mi addr:$dst, tglobaladdr:$src)>;
790 def : Pat<(store (i32 (X86Wrapper texternalsym:$src)), addr:$dst),
791 (MOV32mi addr:$dst, texternalsym:$src)>;
792 def : Pat<(store (i32 (X86Wrapper tblockaddress:$src)), addr:$dst),
793 (MOV32mi addr:$dst, tblockaddress:$src)>;
797 // ConstantPool GlobalAddress, ExternalSymbol, and JumpTable when not in small
798 // code model mode, should use 'movabs'. FIXME: This is really a hack, the
799 // 'movabs' predicate should handle this sort of thing.
800 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
801 (MOV64ri tconstpool :$dst)>, Requires<[FarData]>;
802 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
803 (MOV64ri tjumptable :$dst)>, Requires<[FarData]>;
804 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
805 (MOV64ri tglobaladdr :$dst)>, Requires<[FarData]>;
806 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
807 (MOV64ri texternalsym:$dst)>, Requires<[FarData]>;
808 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
809 (MOV64ri tblockaddress:$dst)>, Requires<[FarData]>;
811 // In static codegen with small code model, we can get the address of a label
812 // into a register with 'movl'. FIXME: This is a hack, the 'imm' predicate of
813 // the MOV64ri64i32 should accept these.
814 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
815 (MOV64ri64i32 tconstpool :$dst)>, Requires<[SmallCode]>;
816 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
817 (MOV64ri64i32 tjumptable :$dst)>, Requires<[SmallCode]>;
818 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
819 (MOV64ri64i32 tglobaladdr :$dst)>, Requires<[SmallCode]>;
820 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
821 (MOV64ri64i32 texternalsym:$dst)>, Requires<[SmallCode]>;
822 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
823 (MOV64ri64i32 tblockaddress:$dst)>, Requires<[SmallCode]>;
825 // In kernel code model, we can get the address of a label
826 // into a register with 'movq'. FIXME: This is a hack, the 'imm' predicate of
827 // the MOV64ri32 should accept these.
828 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
829 (MOV64ri32 tconstpool :$dst)>, Requires<[KernelCode]>;
830 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
831 (MOV64ri32 tjumptable :$dst)>, Requires<[KernelCode]>;
832 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
833 (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
834 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
835 (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
836 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
837 (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
839 // If we have small model and -static mode, it is safe to store global addresses
840 // directly as immediates. FIXME: This is really a hack, the 'imm' predicate
841 // for MOV64mi32 should handle this sort of thing.
842 def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
843 (MOV64mi32 addr:$dst, tconstpool:$src)>,
844 Requires<[NearData, IsStatic]>;
845 def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
846 (MOV64mi32 addr:$dst, tjumptable:$src)>,
847 Requires<[NearData, IsStatic]>;
848 def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
849 (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
850 Requires<[NearData, IsStatic]>;
851 def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
852 (MOV64mi32 addr:$dst, texternalsym:$src)>,
853 Requires<[NearData, IsStatic]>;
854 def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
855 (MOV64mi32 addr:$dst, tblockaddress:$src)>,
856 Requires<[NearData, IsStatic]>;
862 // tls has some funny stuff here...
863 // This corresponds to movabs $foo@tpoff, %rax
864 def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
865 (MOV64ri tglobaltlsaddr :$dst)>;
866 // This corresponds to add $foo@tpoff, %rax
867 def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
868 (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
869 // This corresponds to mov foo@tpoff(%rbx), %eax
870 def : Pat<(load (i64 (X86Wrapper tglobaltlsaddr :$dst))),
871 (MOV64rm tglobaltlsaddr :$dst)>;
874 // Direct PC relative function call for small code model. 32-bit displacement
875 // sign extended to 64-bit.
876 def : Pat<(X86call (i64 tglobaladdr:$dst)),
877 (CALL64pcrel32 tglobaladdr:$dst)>, Requires<[NotWin64]>;
878 def : Pat<(X86call (i64 texternalsym:$dst)),
879 (CALL64pcrel32 texternalsym:$dst)>, Requires<[NotWin64]>;
881 def : Pat<(X86call (i64 tglobaladdr:$dst)),
882 (WINCALL64pcrel32 tglobaladdr:$dst)>, Requires<[IsWin64]>;
883 def : Pat<(X86call (i64 texternalsym:$dst)),
884 (WINCALL64pcrel32 texternalsym:$dst)>, Requires<[IsWin64]>;
887 def : Pat<(X86tcret GR32_TC:$dst, imm:$off),
888 (TCRETURNri GR32_TC:$dst, imm:$off)>,
889 Requires<[In32BitMode]>;
891 // FIXME: This is disabled for 32-bit PIC mode because the global base
892 // register which is part of the address mode may be assigned a
893 // callee-saved register.
894 def : Pat<(X86tcret (load addr:$dst), imm:$off),
895 (TCRETURNmi addr:$dst, imm:$off)>,
896 Requires<[In32BitMode, IsNotPIC]>;
898 def : Pat<(X86tcret (i32 tglobaladdr:$dst), imm:$off),
899 (TCRETURNdi texternalsym:$dst, imm:$off)>,
900 Requires<[In32BitMode]>;
902 def : Pat<(X86tcret (i32 texternalsym:$dst), imm:$off),
903 (TCRETURNdi texternalsym:$dst, imm:$off)>,
904 Requires<[In32BitMode]>;
906 def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
907 (TCRETURNri64 ptr_rc_tailcall:$dst, imm:$off)>,
908 Requires<[In64BitMode]>;
910 def : Pat<(X86tcret (load addr:$dst), imm:$off),
911 (TCRETURNmi64 addr:$dst, imm:$off)>,
912 Requires<[In64BitMode]>;
914 def : Pat<(X86tcret (i64 tglobaladdr:$dst), imm:$off),
915 (TCRETURNdi64 tglobaladdr:$dst, imm:$off)>,
916 Requires<[In64BitMode]>;
918 def : Pat<(X86tcret (i64 texternalsym:$dst), imm:$off),
919 (TCRETURNdi64 texternalsym:$dst, imm:$off)>,
920 Requires<[In64BitMode]>;
922 // Normal calls, with various flavors of addresses.
923 def : Pat<(X86call (i32 tglobaladdr:$dst)),
924 (CALLpcrel32 tglobaladdr:$dst)>;
925 def : Pat<(X86call (i32 texternalsym:$dst)),
926 (CALLpcrel32 texternalsym:$dst)>;
927 def : Pat<(X86call (i32 imm:$dst)),
928 (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
932 // TEST R,R is smaller than CMP R,0
933 def : Pat<(X86cmp GR8:$src1, 0),
934 (TEST8rr GR8:$src1, GR8:$src1)>;
935 def : Pat<(X86cmp GR16:$src1, 0),
936 (TEST16rr GR16:$src1, GR16:$src1)>;
937 def : Pat<(X86cmp GR32:$src1, 0),
938 (TEST32rr GR32:$src1, GR32:$src1)>;
939 def : Pat<(X86cmp GR64:$src1, 0),
940 (TEST64rr GR64:$src1, GR64:$src1)>;
942 // Conditional moves with folded loads with operands swapped and conditions
944 multiclass CMOVmr<PatLeaf InvertedCond, Instruction Inst16, Instruction Inst32,
945 Instruction Inst64> {
946 def : Pat<(X86cmov (loadi16 addr:$src1), GR16:$src2, InvertedCond, EFLAGS),
947 (Inst16 GR16:$src2, addr:$src1)>;
948 def : Pat<(X86cmov (loadi32 addr:$src1), GR32:$src2, InvertedCond, EFLAGS),
949 (Inst32 GR32:$src2, addr:$src1)>;
950 def : Pat<(X86cmov (loadi64 addr:$src1), GR64:$src2, InvertedCond, EFLAGS),
951 (Inst64 GR64:$src2, addr:$src1)>;
954 defm : CMOVmr<X86_COND_B , CMOVAE16rm, CMOVAE32rm, CMOVAE64rm>;
955 defm : CMOVmr<X86_COND_AE, CMOVB16rm , CMOVB32rm , CMOVB64rm>;
956 defm : CMOVmr<X86_COND_E , CMOVNE16rm, CMOVNE32rm, CMOVNE64rm>;
957 defm : CMOVmr<X86_COND_NE, CMOVE16rm , CMOVE32rm , CMOVE64rm>;
958 defm : CMOVmr<X86_COND_BE, CMOVA16rm , CMOVA32rm , CMOVA64rm>;
959 defm : CMOVmr<X86_COND_A , CMOVBE16rm, CMOVBE32rm, CMOVBE64rm>;
960 defm : CMOVmr<X86_COND_L , CMOVGE16rm, CMOVGE32rm, CMOVGE64rm>;
961 defm : CMOVmr<X86_COND_GE, CMOVL16rm , CMOVL32rm , CMOVL64rm>;
962 defm : CMOVmr<X86_COND_LE, CMOVG16rm , CMOVG32rm , CMOVG64rm>;
963 defm : CMOVmr<X86_COND_G , CMOVLE16rm, CMOVLE32rm, CMOVLE64rm>;
964 defm : CMOVmr<X86_COND_P , CMOVNP16rm, CMOVNP32rm, CMOVNP64rm>;
965 defm : CMOVmr<X86_COND_NP, CMOVP16rm , CMOVP32rm , CMOVP64rm>;
966 defm : CMOVmr<X86_COND_S , CMOVNS16rm, CMOVNS32rm, CMOVNS64rm>;
967 defm : CMOVmr<X86_COND_NS, CMOVS16rm , CMOVS32rm , CMOVS64rm>;
968 defm : CMOVmr<X86_COND_O , CMOVNO16rm, CMOVNO32rm, CMOVNO64rm>;
969 defm : CMOVmr<X86_COND_NO, CMOVO16rm , CMOVO32rm , CMOVO64rm>;
971 // zextload bool -> zextload byte
972 def : Pat<(zextloadi8i1 addr:$src), (MOV8rm addr:$src)>;
973 def : Pat<(zextloadi16i1 addr:$src), (MOVZX16rm8 addr:$src)>;
974 def : Pat<(zextloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
975 def : Pat<(zextloadi64i1 addr:$src), (MOVZX64rm8 addr:$src)>;
977 // extload bool -> extload byte
978 // When extloading from 16-bit and smaller memory locations into 64-bit
979 // registers, use zero-extending loads so that the entire 64-bit register is
980 // defined, avoiding partial-register updates.
982 def : Pat<(extloadi8i1 addr:$src), (MOV8rm addr:$src)>;
983 def : Pat<(extloadi16i1 addr:$src), (MOVZX16rm8 addr:$src)>;
984 def : Pat<(extloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
985 def : Pat<(extloadi16i8 addr:$src), (MOVZX16rm8 addr:$src)>;
986 def : Pat<(extloadi32i8 addr:$src), (MOVZX32rm8 addr:$src)>;
987 def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
989 def : Pat<(extloadi64i1 addr:$src), (MOVZX64rm8 addr:$src)>;
990 def : Pat<(extloadi64i8 addr:$src), (MOVZX64rm8 addr:$src)>;
991 def : Pat<(extloadi64i16 addr:$src), (MOVZX64rm16 addr:$src)>;
992 // For other extloads, use subregs, since the high contents of the register are
993 // defined after an extload.
994 def : Pat<(extloadi64i32 addr:$src),
995 (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src),
998 // anyext. Define these to do an explicit zero-extend to
999 // avoid partial-register updates.
1000 def : Pat<(i16 (anyext GR8 :$src)), (EXTRACT_SUBREG
1001 (MOVZX32rr8 GR8 :$src), sub_16bit)>;
1002 def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8 GR8 :$src)>;
1004 // Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
1005 def : Pat<(i32 (anyext GR16:$src)),
1006 (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
1008 def : Pat<(i64 (anyext GR8 :$src)), (MOVZX64rr8 GR8 :$src)>;
1009 def : Pat<(i64 (anyext GR16:$src)), (MOVZX64rr16 GR16 :$src)>;
1010 def : Pat<(i64 (anyext GR32:$src)),
1011 (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1014 // Any instruction that defines a 32-bit result leaves the high half of the
1015 // register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
1016 // be copying from a truncate. And x86's cmov doesn't do anything if the
1017 // condition is false. But any other 32-bit operation will zero-extend
1019 def def32 : PatLeaf<(i32 GR32:$src), [{
1020 return N->getOpcode() != ISD::TRUNCATE &&
1021 N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
1022 N->getOpcode() != ISD::CopyFromReg &&
1023 N->getOpcode() != X86ISD::CMOV;
1026 // In the case of a 32-bit def that is known to implicitly zero-extend,
1027 // we can use a SUBREG_TO_REG.
1028 def : Pat<(i64 (zext def32:$src)),
1029 (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1031 //===----------------------------------------------------------------------===//
1032 // Pattern match OR as ADD
1033 //===----------------------------------------------------------------------===//
1035 // If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
1036 // 3-addressified into an LEA instruction to avoid copies. However, we also
1037 // want to finally emit these instructions as an or at the end of the code
1038 // generator to make the generated code easier to read. To do this, we select
1039 // into "disjoint bits" pseudo ops.
1041 // Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.
1042 def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
1043 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1044 return CurDAG->MaskedValueIsZero(N->getOperand(0), CN->getAPIntValue());
1046 unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits();
1047 APInt Mask = APInt::getAllOnesValue(BitWidth);
1048 APInt KnownZero0, KnownOne0;
1049 CurDAG->ComputeMaskedBits(N->getOperand(0), Mask, KnownZero0, KnownOne0, 0);
1050 APInt KnownZero1, KnownOne1;
1051 CurDAG->ComputeMaskedBits(N->getOperand(1), Mask, KnownZero1, KnownOne1, 0);
1052 return (~KnownZero0 & ~KnownZero1) == 0;
1056 // (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1057 let AddedComplexity = 5 in { // Try this before the selecting to OR
1059 let isConvertibleToThreeAddress = 1,
1060 Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1061 let isCommutable = 1 in {
1062 def ADD16rr_DB : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1063 "", // orw/addw REG, REG
1064 [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1065 def ADD32rr_DB : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1066 "", // orl/addl REG, REG
1067 [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1068 def ADD64rr_DB : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1069 "", // orq/addq REG, REG
1070 [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1073 // NOTE: These are order specific, we want the ri8 forms to be listed
1074 // first so that they are slightly preferred to the ri forms.
1076 def ADD16ri8_DB : I<0, Pseudo,
1077 (outs GR16:$dst), (ins GR16:$src1, i16i8imm:$src2),
1078 "", // orw/addw REG, imm8
1079 [(set GR16:$dst,(or_is_add GR16:$src1,i16immSExt8:$src2))]>;
1080 def ADD16ri_DB : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1081 "", // orw/addw REG, imm
1082 [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1084 def ADD32ri8_DB : I<0, Pseudo,
1085 (outs GR32:$dst), (ins GR32:$src1, i32i8imm:$src2),
1086 "", // orl/addl REG, imm8
1087 [(set GR32:$dst,(or_is_add GR32:$src1,i32immSExt8:$src2))]>;
1088 def ADD32ri_DB : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1089 "", // orl/addl REG, imm
1090 [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1093 def ADD64ri8_DB : I<0, Pseudo,
1094 (outs GR64:$dst), (ins GR64:$src1, i64i8imm:$src2),
1095 "", // orq/addq REG, imm8
1096 [(set GR64:$dst, (or_is_add GR64:$src1,
1097 i64immSExt8:$src2))]>;
1098 def ADD64ri32_DB : I<0, Pseudo,
1099 (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1100 "", // orq/addq REG, imm
1101 [(set GR64:$dst, (or_is_add GR64:$src1,
1102 i64immSExt32:$src2))]>;
1104 } // AddedComplexity
1107 //===----------------------------------------------------------------------===//
1109 //===----------------------------------------------------------------------===//
1111 // Odd encoding trick: -128 fits into an 8-bit immediate field while
1112 // +128 doesn't, so in this special case use a sub instead of an add.
1113 def : Pat<(add GR16:$src1, 128),
1114 (SUB16ri8 GR16:$src1, -128)>;
1115 def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1116 (SUB16mi8 addr:$dst, -128)>;
1118 def : Pat<(add GR32:$src1, 128),
1119 (SUB32ri8 GR32:$src1, -128)>;
1120 def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1121 (SUB32mi8 addr:$dst, -128)>;
1123 def : Pat<(add GR64:$src1, 128),
1124 (SUB64ri8 GR64:$src1, -128)>;
1125 def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1126 (SUB64mi8 addr:$dst, -128)>;
1128 // The same trick applies for 32-bit immediate fields in 64-bit
1130 def : Pat<(add GR64:$src1, 0x0000000080000000),
1131 (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1132 def : Pat<(store (add (loadi64 addr:$dst), 0x00000000800000000), addr:$dst),
1133 (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1135 // To avoid needing to materialize an immediate in a register, use a 32-bit and
1136 // with implicit zero-extension instead of a 64-bit and if the immediate has at
1137 // least 32 bits of leading zeros. If in addition the last 32 bits can be
1138 // represented with a sign extension of a 8 bit constant, use that.
1140 def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
1144 (EXTRACT_SUBREG GR64:$src, sub_32bit),
1145 (i32 (GetLo8XForm imm:$imm))),
1148 def : Pat<(and GR64:$src, i64immZExt32:$imm),
1152 (EXTRACT_SUBREG GR64:$src, sub_32bit),
1153 (i32 (GetLo32XForm imm:$imm))),
1157 // r & (2^16-1) ==> movz
1158 def : Pat<(and GR32:$src1, 0xffff),
1159 (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1160 // r & (2^8-1) ==> movz
1161 def : Pat<(and GR32:$src1, 0xff),
1162 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src1,
1165 Requires<[In32BitMode]>;
1166 // r & (2^8-1) ==> movz
1167 def : Pat<(and GR16:$src1, 0xff),
1168 (EXTRACT_SUBREG (MOVZX32rr8 (EXTRACT_SUBREG
1169 (i16 (COPY_TO_REGCLASS GR16:$src1, GR16_ABCD)), sub_8bit)),
1171 Requires<[In32BitMode]>;
1173 // r & (2^32-1) ==> movz
1174 def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1175 (MOVZX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1176 // r & (2^16-1) ==> movz
1177 def : Pat<(and GR64:$src, 0xffff),
1178 (MOVZX64rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit)))>;
1179 // r & (2^8-1) ==> movz
1180 def : Pat<(and GR64:$src, 0xff),
1181 (MOVZX64rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit)))>;
1182 // r & (2^8-1) ==> movz
1183 def : Pat<(and GR32:$src1, 0xff),
1184 (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>,
1185 Requires<[In64BitMode]>;
1186 // r & (2^8-1) ==> movz
1187 def : Pat<(and GR16:$src1, 0xff),
1188 (EXTRACT_SUBREG (MOVZX32rr8 (i8
1189 (EXTRACT_SUBREG GR16:$src1, sub_8bit))), sub_16bit)>,
1190 Requires<[In64BitMode]>;
1193 // sext_inreg patterns
1194 def : Pat<(sext_inreg GR32:$src, i16),
1195 (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1196 def : Pat<(sext_inreg GR32:$src, i8),
1197 (MOVSX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1200 Requires<[In32BitMode]>;
1202 def : Pat<(sext_inreg GR16:$src, i8),
1203 (EXTRACT_SUBREG (i32 (MOVSX32rr8 (EXTRACT_SUBREG
1204 (i32 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)), sub_8bit))),
1206 Requires<[In32BitMode]>;
1208 def : Pat<(sext_inreg GR64:$src, i32),
1209 (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1210 def : Pat<(sext_inreg GR64:$src, i16),
1211 (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1212 def : Pat<(sext_inreg GR64:$src, i8),
1213 (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1214 def : Pat<(sext_inreg GR32:$src, i8),
1215 (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>,
1216 Requires<[In64BitMode]>;
1217 def : Pat<(sext_inreg GR16:$src, i8),
1218 (EXTRACT_SUBREG (MOVSX32rr8
1219 (EXTRACT_SUBREG GR16:$src, sub_8bit)), sub_16bit)>,
1220 Requires<[In64BitMode]>;
1222 // sext, sext_load, zext, zext_load
1223 def: Pat<(i16 (sext GR8:$src)),
1224 (EXTRACT_SUBREG (MOVSX32rr8 GR8:$src), sub_16bit)>;
1225 def: Pat<(sextloadi16i8 addr:$src),
1226 (EXTRACT_SUBREG (MOVSX32rm8 addr:$src), sub_16bit)>;
1227 def: Pat<(i16 (zext GR8:$src)),
1228 (EXTRACT_SUBREG (MOVZX32rr8 GR8:$src), sub_16bit)>;
1229 def: Pat<(zextloadi16i8 addr:$src),
1230 (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1233 def : Pat<(i16 (trunc GR32:$src)),
1234 (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1235 def : Pat<(i8 (trunc GR32:$src)),
1236 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1238 Requires<[In32BitMode]>;
1239 def : Pat<(i8 (trunc GR16:$src)),
1240 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1242 Requires<[In32BitMode]>;
1243 def : Pat<(i32 (trunc GR64:$src)),
1244 (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1245 def : Pat<(i16 (trunc GR64:$src)),
1246 (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1247 def : Pat<(i8 (trunc GR64:$src)),
1248 (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1249 def : Pat<(i8 (trunc GR32:$src)),
1250 (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1251 Requires<[In64BitMode]>;
1252 def : Pat<(i8 (trunc GR16:$src)),
1253 (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1254 Requires<[In64BitMode]>;
1256 // h-register tricks
1257 def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1258 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1260 Requires<[In32BitMode]>;
1261 def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1262 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1264 Requires<[In32BitMode]>;
1265 def : Pat<(srl GR16:$src, (i8 8)),
1268 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1271 Requires<[In32BitMode]>;
1272 def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1273 (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1276 Requires<[In32BitMode]>;
1277 def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1278 (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1281 Requires<[In32BitMode]>;
1282 def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1283 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1286 Requires<[In32BitMode]>;
1287 def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1288 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1291 Requires<[In32BitMode]>;
1293 // h-register tricks.
1294 // For now, be conservative on x86-64 and use an h-register extract only if the
1295 // value is immediately zero-extended or stored, which are somewhat common
1296 // cases. This uses a bunch of code to prevent a register requiring a REX prefix
1297 // from being allocated in the same instruction as the h register, as there's
1298 // currently no way to describe this requirement to the register allocator.
1300 // h-register extract and zero-extend.
1301 def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1305 (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1308 def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1310 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1312 Requires<[In64BitMode]>;
1313 def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1314 (MOVZX32_NOREXrr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1317 Requires<[In64BitMode]>;
1318 def : Pat<(srl GR16:$src, (i8 8)),
1321 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1324 Requires<[In64BitMode]>;
1325 def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1327 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1329 Requires<[In64BitMode]>;
1330 def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1332 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1334 Requires<[In64BitMode]>;
1335 def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1339 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1342 def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1346 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1350 // h-register extract and store.
1351 def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1354 (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1356 def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1359 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1361 Requires<[In64BitMode]>;
1362 def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1365 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1367 Requires<[In64BitMode]>;
1370 // (shl x, 1) ==> (add x, x)
1371 def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr GR8 :$src1, GR8 :$src1)>;
1372 def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1373 def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1374 def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1376 // (shl x (and y, 31)) ==> (shl x, y)
1377 def : Pat<(shl GR8:$src1, (and CL, 31)),
1378 (SHL8rCL GR8:$src1)>;
1379 def : Pat<(shl GR16:$src1, (and CL, 31)),
1380 (SHL16rCL GR16:$src1)>;
1381 def : Pat<(shl GR32:$src1, (and CL, 31)),
1382 (SHL32rCL GR32:$src1)>;
1383 def : Pat<(store (shl (loadi8 addr:$dst), (and CL, 31)), addr:$dst),
1384 (SHL8mCL addr:$dst)>;
1385 def : Pat<(store (shl (loadi16 addr:$dst), (and CL, 31)), addr:$dst),
1386 (SHL16mCL addr:$dst)>;
1387 def : Pat<(store (shl (loadi32 addr:$dst), (and CL, 31)), addr:$dst),
1388 (SHL32mCL addr:$dst)>;
1390 def : Pat<(srl GR8:$src1, (and CL, 31)),
1391 (SHR8rCL GR8:$src1)>;
1392 def : Pat<(srl GR16:$src1, (and CL, 31)),
1393 (SHR16rCL GR16:$src1)>;
1394 def : Pat<(srl GR32:$src1, (and CL, 31)),
1395 (SHR32rCL GR32:$src1)>;
1396 def : Pat<(store (srl (loadi8 addr:$dst), (and CL, 31)), addr:$dst),
1397 (SHR8mCL addr:$dst)>;
1398 def : Pat<(store (srl (loadi16 addr:$dst), (and CL, 31)), addr:$dst),
1399 (SHR16mCL addr:$dst)>;
1400 def : Pat<(store (srl (loadi32 addr:$dst), (and CL, 31)), addr:$dst),
1401 (SHR32mCL addr:$dst)>;
1403 def : Pat<(sra GR8:$src1, (and CL, 31)),
1404 (SAR8rCL GR8:$src1)>;
1405 def : Pat<(sra GR16:$src1, (and CL, 31)),
1406 (SAR16rCL GR16:$src1)>;
1407 def : Pat<(sra GR32:$src1, (and CL, 31)),
1408 (SAR32rCL GR32:$src1)>;
1409 def : Pat<(store (sra (loadi8 addr:$dst), (and CL, 31)), addr:$dst),
1410 (SAR8mCL addr:$dst)>;
1411 def : Pat<(store (sra (loadi16 addr:$dst), (and CL, 31)), addr:$dst),
1412 (SAR16mCL addr:$dst)>;
1413 def : Pat<(store (sra (loadi32 addr:$dst), (and CL, 31)), addr:$dst),
1414 (SAR32mCL addr:$dst)>;
1416 // (shl x (and y, 63)) ==> (shl x, y)
1417 def : Pat<(shl GR64:$src1, (and CL, 63)),
1418 (SHL64rCL GR64:$src1)>;
1419 def : Pat<(store (shl (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1420 (SHL64mCL addr:$dst)>;
1422 def : Pat<(srl GR64:$src1, (and CL, 63)),
1423 (SHR64rCL GR64:$src1)>;
1424 def : Pat<(store (srl (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1425 (SHR64mCL addr:$dst)>;
1427 def : Pat<(sra GR64:$src1, (and CL, 63)),
1428 (SAR64rCL GR64:$src1)>;
1429 def : Pat<(store (sra (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1430 (SAR64mCL addr:$dst)>;
1433 // (anyext (setcc_carry)) -> (setcc_carry)
1434 def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1436 def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1438 def : Pat<(i32 (anyext (i16 (X86setcc_c X86_COND_B, EFLAGS)))),
1444 //===----------------------------------------------------------------------===//
1445 // EFLAGS-defining Patterns
1446 //===----------------------------------------------------------------------===//
1449 def : Pat<(add GR8 :$src1, GR8 :$src2), (ADD8rr GR8 :$src1, GR8 :$src2)>;
1450 def : Pat<(add GR16:$src1, GR16:$src2), (ADD16rr GR16:$src1, GR16:$src2)>;
1451 def : Pat<(add GR32:$src1, GR32:$src2), (ADD32rr GR32:$src1, GR32:$src2)>;
1454 def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
1455 (ADD8rm GR8:$src1, addr:$src2)>;
1456 def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
1457 (ADD16rm GR16:$src1, addr:$src2)>;
1458 def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
1459 (ADD32rm GR32:$src1, addr:$src2)>;
1462 def : Pat<(add GR8 :$src1, imm:$src2), (ADD8ri GR8:$src1 , imm:$src2)>;
1463 def : Pat<(add GR16:$src1, imm:$src2), (ADD16ri GR16:$src1, imm:$src2)>;
1464 def : Pat<(add GR32:$src1, imm:$src2), (ADD32ri GR32:$src1, imm:$src2)>;
1465 def : Pat<(add GR16:$src1, i16immSExt8:$src2),
1466 (ADD16ri8 GR16:$src1, i16immSExt8:$src2)>;
1467 def : Pat<(add GR32:$src1, i32immSExt8:$src2),
1468 (ADD32ri8 GR32:$src1, i32immSExt8:$src2)>;
1471 def : Pat<(sub GR8 :$src1, GR8 :$src2), (SUB8rr GR8 :$src1, GR8 :$src2)>;
1472 def : Pat<(sub GR16:$src1, GR16:$src2), (SUB16rr GR16:$src1, GR16:$src2)>;
1473 def : Pat<(sub GR32:$src1, GR32:$src2), (SUB32rr GR32:$src1, GR32:$src2)>;
1476 def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
1477 (SUB8rm GR8:$src1, addr:$src2)>;
1478 def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
1479 (SUB16rm GR16:$src1, addr:$src2)>;
1480 def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
1481 (SUB32rm GR32:$src1, addr:$src2)>;
1484 def : Pat<(sub GR8:$src1, imm:$src2),
1485 (SUB8ri GR8:$src1, imm:$src2)>;
1486 def : Pat<(sub GR16:$src1, imm:$src2),
1487 (SUB16ri GR16:$src1, imm:$src2)>;
1488 def : Pat<(sub GR32:$src1, imm:$src2),
1489 (SUB32ri GR32:$src1, imm:$src2)>;
1490 def : Pat<(sub GR16:$src1, i16immSExt8:$src2),
1491 (SUB16ri8 GR16:$src1, i16immSExt8:$src2)>;
1492 def : Pat<(sub GR32:$src1, i32immSExt8:$src2),
1493 (SUB32ri8 GR32:$src1, i32immSExt8:$src2)>;
1496 def : Pat<(mul GR16:$src1, GR16:$src2),
1497 (IMUL16rr GR16:$src1, GR16:$src2)>;
1498 def : Pat<(mul GR32:$src1, GR32:$src2),
1499 (IMUL32rr GR32:$src1, GR32:$src2)>;
1502 def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
1503 (IMUL16rm GR16:$src1, addr:$src2)>;
1504 def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
1505 (IMUL32rm GR32:$src1, addr:$src2)>;
1508 def : Pat<(mul GR16:$src1, imm:$src2),
1509 (IMUL16rri GR16:$src1, imm:$src2)>;
1510 def : Pat<(mul GR32:$src1, imm:$src2),
1511 (IMUL32rri GR32:$src1, imm:$src2)>;
1512 def : Pat<(mul GR16:$src1, i16immSExt8:$src2),
1513 (IMUL16rri8 GR16:$src1, i16immSExt8:$src2)>;
1514 def : Pat<(mul GR32:$src1, i32immSExt8:$src2),
1515 (IMUL32rri8 GR32:$src1, i32immSExt8:$src2)>;
1517 // reg = mul mem, imm
1518 def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
1519 (IMUL16rmi addr:$src1, imm:$src2)>;
1520 def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
1521 (IMUL32rmi addr:$src1, imm:$src2)>;
1522 def : Pat<(mul (loadi16 addr:$src1), i16immSExt8:$src2),
1523 (IMUL16rmi8 addr:$src1, i16immSExt8:$src2)>;
1524 def : Pat<(mul (loadi32 addr:$src1), i32immSExt8:$src2),
1525 (IMUL32rmi8 addr:$src1, i32immSExt8:$src2)>;
1527 // Patterns for nodes that do not produce flags, for instructions that do.
1530 def : Pat<(add GR64:$src1, GR64:$src2),
1531 (ADD64rr GR64:$src1, GR64:$src2)>;
1532 def : Pat<(add GR64:$src1, i64immSExt8:$src2),
1533 (ADD64ri8 GR64:$src1, i64immSExt8:$src2)>;
1534 def : Pat<(add GR64:$src1, i64immSExt32:$src2),
1535 (ADD64ri32 GR64:$src1, i64immSExt32:$src2)>;
1536 def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
1537 (ADD64rm GR64:$src1, addr:$src2)>;
1540 def : Pat<(sub GR64:$src1, GR64:$src2),
1541 (SUB64rr GR64:$src1, GR64:$src2)>;
1542 def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
1543 (SUB64rm GR64:$src1, addr:$src2)>;
1544 def : Pat<(sub GR64:$src1, i64immSExt8:$src2),
1545 (SUB64ri8 GR64:$src1, i64immSExt8:$src2)>;
1546 def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
1547 (SUB64ri32 GR64:$src1, i64immSExt32:$src2)>;
1550 def : Pat<(mul GR64:$src1, GR64:$src2),
1551 (IMUL64rr GR64:$src1, GR64:$src2)>;
1552 def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
1553 (IMUL64rm GR64:$src1, addr:$src2)>;
1554 def : Pat<(mul GR64:$src1, i64immSExt8:$src2),
1555 (IMUL64rri8 GR64:$src1, i64immSExt8:$src2)>;
1556 def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
1557 (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
1558 def : Pat<(mul (loadi64 addr:$src1), i64immSExt8:$src2),
1559 (IMUL64rmi8 addr:$src1, i64immSExt8:$src2)>;
1560 def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
1561 (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
1564 def : Pat<(add GR8 :$src, 1), (INC8r GR8 :$src)>;
1565 def : Pat<(add GR16:$src, 1), (INC16r GR16:$src)>, Requires<[In32BitMode]>;
1566 def : Pat<(add GR16:$src, 1), (INC64_16r GR16:$src)>, Requires<[In64BitMode]>;
1567 def : Pat<(add GR32:$src, 1), (INC32r GR32:$src)>, Requires<[In32BitMode]>;
1568 def : Pat<(add GR32:$src, 1), (INC64_32r GR32:$src)>, Requires<[In64BitMode]>;
1569 def : Pat<(add GR64:$src, 1), (INC64r GR64:$src)>;
1572 def : Pat<(add GR8 :$src, -1), (DEC8r GR8 :$src)>;
1573 def : Pat<(add GR16:$src, -1), (DEC16r GR16:$src)>, Requires<[In32BitMode]>;
1574 def : Pat<(add GR16:$src, -1), (DEC64_16r GR16:$src)>, Requires<[In64BitMode]>;
1575 def : Pat<(add GR32:$src, -1), (DEC32r GR32:$src)>, Requires<[In32BitMode]>;
1576 def : Pat<(add GR32:$src, -1), (DEC64_32r GR32:$src)>, Requires<[In64BitMode]>;
1577 def : Pat<(add GR64:$src, -1), (DEC64r GR64:$src)>;
1580 def : Pat<(or GR8 :$src1, GR8 :$src2), (OR8rr GR8 :$src1, GR8 :$src2)>;
1581 def : Pat<(or GR16:$src1, GR16:$src2), (OR16rr GR16:$src1, GR16:$src2)>;
1582 def : Pat<(or GR32:$src1, GR32:$src2), (OR32rr GR32:$src1, GR32:$src2)>;
1583 def : Pat<(or GR64:$src1, GR64:$src2), (OR64rr GR64:$src1, GR64:$src2)>;
1586 def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
1587 (OR8rm GR8:$src1, addr:$src2)>;
1588 def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
1589 (OR16rm GR16:$src1, addr:$src2)>;
1590 def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
1591 (OR32rm GR32:$src1, addr:$src2)>;
1592 def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
1593 (OR64rm GR64:$src1, addr:$src2)>;
1596 def : Pat<(or GR8:$src1 , imm:$src2), (OR8ri GR8 :$src1, imm:$src2)>;
1597 def : Pat<(or GR16:$src1, imm:$src2), (OR16ri GR16:$src1, imm:$src2)>;
1598 def : Pat<(or GR32:$src1, imm:$src2), (OR32ri GR32:$src1, imm:$src2)>;
1599 def : Pat<(or GR16:$src1, i16immSExt8:$src2),
1600 (OR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1601 def : Pat<(or GR32:$src1, i32immSExt8:$src2),
1602 (OR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1603 def : Pat<(or GR64:$src1, i64immSExt8:$src2),
1604 (OR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1605 def : Pat<(or GR64:$src1, i64immSExt32:$src2),
1606 (OR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1609 def : Pat<(xor GR8 :$src1, GR8 :$src2), (XOR8rr GR8 :$src1, GR8 :$src2)>;
1610 def : Pat<(xor GR16:$src1, GR16:$src2), (XOR16rr GR16:$src1, GR16:$src2)>;
1611 def : Pat<(xor GR32:$src1, GR32:$src2), (XOR32rr GR32:$src1, GR32:$src2)>;
1612 def : Pat<(xor GR64:$src1, GR64:$src2), (XOR64rr GR64:$src1, GR64:$src2)>;
1615 def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
1616 (XOR8rm GR8:$src1, addr:$src2)>;
1617 def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
1618 (XOR16rm GR16:$src1, addr:$src2)>;
1619 def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
1620 (XOR32rm GR32:$src1, addr:$src2)>;
1621 def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
1622 (XOR64rm GR64:$src1, addr:$src2)>;
1625 def : Pat<(xor GR8:$src1, imm:$src2),
1626 (XOR8ri GR8:$src1, imm:$src2)>;
1627 def : Pat<(xor GR16:$src1, imm:$src2),
1628 (XOR16ri GR16:$src1, imm:$src2)>;
1629 def : Pat<(xor GR32:$src1, imm:$src2),
1630 (XOR32ri GR32:$src1, imm:$src2)>;
1631 def : Pat<(xor GR16:$src1, i16immSExt8:$src2),
1632 (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1633 def : Pat<(xor GR32:$src1, i32immSExt8:$src2),
1634 (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1635 def : Pat<(xor GR64:$src1, i64immSExt8:$src2),
1636 (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1637 def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
1638 (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1641 def : Pat<(and GR8 :$src1, GR8 :$src2), (AND8rr GR8 :$src1, GR8 :$src2)>;
1642 def : Pat<(and GR16:$src1, GR16:$src2), (AND16rr GR16:$src1, GR16:$src2)>;
1643 def : Pat<(and GR32:$src1, GR32:$src2), (AND32rr GR32:$src1, GR32:$src2)>;
1644 def : Pat<(and GR64:$src1, GR64:$src2), (AND64rr GR64:$src1, GR64:$src2)>;
1647 def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
1648 (AND8rm GR8:$src1, addr:$src2)>;
1649 def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
1650 (AND16rm GR16:$src1, addr:$src2)>;
1651 def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
1652 (AND32rm GR32:$src1, addr:$src2)>;
1653 def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
1654 (AND64rm GR64:$src1, addr:$src2)>;
1657 def : Pat<(and GR8:$src1, imm:$src2),
1658 (AND8ri GR8:$src1, imm:$src2)>;
1659 def : Pat<(and GR16:$src1, imm:$src2),
1660 (AND16ri GR16:$src1, imm:$src2)>;
1661 def : Pat<(and GR32:$src1, imm:$src2),
1662 (AND32ri GR32:$src1, imm:$src2)>;
1663 def : Pat<(and GR16:$src1, i16immSExt8:$src2),
1664 (AND16ri8 GR16:$src1, i16immSExt8:$src2)>;
1665 def : Pat<(and GR32:$src1, i32immSExt8:$src2),
1666 (AND32ri8 GR32:$src1, i32immSExt8:$src2)>;
1667 def : Pat<(and GR64:$src1, i64immSExt8:$src2),
1668 (AND64ri8 GR64:$src1, i64immSExt8:$src2)>;
1669 def : Pat<(and GR64:$src1, i64immSExt32:$src2),
1670 (AND64ri32 GR64:$src1, i64immSExt32:$src2)>;