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<[Not64BitMode]>;
50 def ADJCALLSTACKUP32 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
52 [(X86callseq_end timm:$amt1, timm:$amt2)]>,
53 Requires<[Not64BitMode]>;
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, Defs = [EFLAGS] 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,
87 // The VAARG_64 pseudo-instruction takes the address of the va_list,
88 // and places the address of the next argument into a register.
89 let Defs = [EFLAGS] in
90 def VAARG_64 : I<0, Pseudo,
92 (ins i8mem:$ap, i32imm:$size, i8imm:$mode, i32imm:$align),
93 "#VAARG_64 $dst, $ap, $size, $mode, $align",
95 (X86vaarg64 addr:$ap, imm:$size, imm:$mode, imm:$align)),
98 // Dynamic stack allocation yields a _chkstk or _alloca call for all Windows
99 // targets. These calls are needed to probe the stack when allocating more than
100 // 4k bytes in one go. Touching the stack at 4K increments is necessary to
101 // ensure that the guard pages used by the OS virtual memory manager are
102 // allocated in correct sequence.
103 // The main point of having separate instruction are extra unmodelled effects
104 // (compared to ordinary calls) like stack pointer change.
106 let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
107 def WIN_ALLOCA : I<0, Pseudo, (outs), (ins),
108 "# dynamic stack allocation",
111 // When using segmented stacks these are lowered into instructions which first
112 // check if the current stacklet has enough free memory. If it does, memory is
113 // allocated by bumping the stack pointer. Otherwise memory is allocated from
116 let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
117 def SEG_ALLOCA_32 : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$size),
118 "# variable sized alloca for segmented stacks",
120 (X86SegAlloca GR32:$size))]>,
121 Requires<[Not64BitMode]>;
123 let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
124 def SEG_ALLOCA_64 : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$size),
125 "# variable sized alloca for segmented stacks",
127 (X86SegAlloca GR64:$size))]>,
128 Requires<[In64BitMode]>;
131 // The MSVC runtime contains an _ftol2 routine for converting floating-point
132 // to integer values. It has a strange calling convention: the input is
133 // popped from the x87 stack, and the return value is given in EDX:EAX. ECX is
134 // used as a temporary register. No other registers (aside from flags) are
136 // Microsoft toolchains do not support 80-bit precision, so a WIN_FTOL_80
137 // variant is unnecessary.
139 let Defs = [EAX, EDX, ECX, EFLAGS], FPForm = SpecialFP in {
140 def WIN_FTOL_32 : I<0, Pseudo, (outs), (ins RFP32:$src),
142 [(X86WinFTOL RFP32:$src)]>,
143 Requires<[Not64BitMode]>;
145 def WIN_FTOL_64 : I<0, Pseudo, (outs), (ins RFP64:$src),
147 [(X86WinFTOL RFP64:$src)]>,
148 Requires<[Not64BitMode]>;
151 //===----------------------------------------------------------------------===//
152 // EH Pseudo Instructions
154 let SchedRW = [WriteSystem] in {
155 let isTerminator = 1, isReturn = 1, isBarrier = 1,
156 hasCtrlDep = 1, isCodeGenOnly = 1 in {
157 def EH_RETURN : I<0xC3, RawFrm, (outs), (ins GR32:$addr),
158 "ret\t#eh_return, addr: $addr",
159 [(X86ehret GR32:$addr)], IIC_RET>, Sched<[WriteJumpLd]>;
163 let isTerminator = 1, isReturn = 1, isBarrier = 1,
164 hasCtrlDep = 1, isCodeGenOnly = 1 in {
165 def EH_RETURN64 : I<0xC3, RawFrm, (outs), (ins GR64:$addr),
166 "ret\t#eh_return, addr: $addr",
167 [(X86ehret GR64:$addr)], IIC_RET>, Sched<[WriteJumpLd]>;
171 let hasSideEffects = 1, isBarrier = 1, isCodeGenOnly = 1,
172 usesCustomInserter = 1 in {
173 def EH_SjLj_SetJmp32 : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$buf),
175 [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
176 Requires<[Not64BitMode]>;
177 def EH_SjLj_SetJmp64 : I<0, Pseudo, (outs GR32:$dst), (ins i64mem:$buf),
179 [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
180 Requires<[In64BitMode]>;
181 let isTerminator = 1 in {
182 def EH_SjLj_LongJmp32 : I<0, Pseudo, (outs), (ins i32mem:$buf),
183 "#EH_SJLJ_LONGJMP32",
184 [(X86eh_sjlj_longjmp addr:$buf)]>,
185 Requires<[Not64BitMode]>;
186 def EH_SjLj_LongJmp64 : I<0, Pseudo, (outs), (ins i64mem:$buf),
187 "#EH_SJLJ_LONGJMP64",
188 [(X86eh_sjlj_longjmp addr:$buf)]>,
189 Requires<[In64BitMode]>;
194 let isBranch = 1, isTerminator = 1, isCodeGenOnly = 1 in {
195 def EH_SjLj_Setup : I<0, Pseudo, (outs), (ins brtarget:$dst),
196 "#EH_SjLj_Setup\t$dst", []>;
199 //===----------------------------------------------------------------------===//
200 // Pseudo instructions used by unwind info.
202 let isPseudo = 1 in {
203 def SEH_PushReg : I<0, Pseudo, (outs), (ins i32imm:$reg),
204 "#SEH_PushReg $reg", []>;
205 def SEH_SaveReg : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
206 "#SEH_SaveReg $reg, $dst", []>;
207 def SEH_SaveXMM : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
208 "#SEH_SaveXMM $reg, $dst", []>;
209 def SEH_StackAlloc : I<0, Pseudo, (outs), (ins i32imm:$size),
210 "#SEH_StackAlloc $size", []>;
211 def SEH_SetFrame : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$offset),
212 "#SEH_SetFrame $reg, $offset", []>;
213 def SEH_PushFrame : I<0, Pseudo, (outs), (ins i1imm:$mode),
214 "#SEH_PushFrame $mode", []>;
215 def SEH_EndPrologue : I<0, Pseudo, (outs), (ins),
216 "#SEH_EndPrologue", []>;
219 //===----------------------------------------------------------------------===//
220 // Pseudo instructions used by segmented stacks.
223 // This is lowered into a RET instruction by MCInstLower. We need
224 // this so that we don't have to have a MachineBasicBlock which ends
225 // with a RET and also has successors.
226 let isPseudo = 1 in {
227 def MORESTACK_RET: I<0, Pseudo, (outs), (ins),
230 // This instruction is lowered to a RET followed by a MOV. The two
231 // instructions are not generated on a higher level since then the
232 // verifier sees a MachineBasicBlock ending with a non-terminator.
233 def MORESTACK_RET_RESTORE_R10 : I<0, Pseudo, (outs), (ins),
237 //===----------------------------------------------------------------------===//
238 // Alias Instructions
239 //===----------------------------------------------------------------------===//
241 // Alias instruction mapping movr0 to xor.
242 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
243 let Defs = [EFLAGS], isReMaterializable = 1, isAsCheapAsAMove = 1,
245 def MOV32r0 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
246 [(set GR32:$dst, 0)], IIC_ALU_NONMEM>, Sched<[WriteZero]>;
248 // Other widths can also make use of the 32-bit xor, which may have a smaller
249 // encoding and avoid partial register updates.
250 def : Pat<(i8 0), (EXTRACT_SUBREG (MOV32r0), sub_8bit)>;
251 def : Pat<(i16 0), (EXTRACT_SUBREG (MOV32r0), sub_16bit)>;
252 def : Pat<(i64 0), (SUBREG_TO_REG (i64 0), (MOV32r0), sub_32bit)> {
253 let AddedComplexity = 20;
256 // Materialize i64 constant where top 32-bits are zero. This could theoretically
257 // use MOV32ri with a SUBREG_TO_REG to represent the zero-extension, however
258 // that would make it more difficult to rematerialize.
259 let AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1,
260 isCodeGenOnly = 1, neverHasSideEffects = 1 in
261 def MOV32ri64 : Ii32<0xb8, AddRegFrm, (outs GR32:$dst), (ins i64i32imm:$src),
262 "", [], IIC_ALU_NONMEM>, Sched<[WriteALU]>;
264 // This 64-bit pseudo-move can be used for both a 64-bit constant that is
265 // actually the zero-extension of a 32-bit constant, and for labels in the
266 // x86-64 small code model.
267 def mov64imm32 : ComplexPattern<i64, 1, "SelectMOV64Imm32", [imm, X86Wrapper]>;
269 let AddedComplexity = 1 in
270 def : Pat<(i64 mov64imm32:$src),
271 (SUBREG_TO_REG (i64 0), (MOV32ri64 mov64imm32:$src), sub_32bit)>;
273 // Use sbb to materialize carry bit.
274 let Uses = [EFLAGS], Defs = [EFLAGS], isPseudo = 1, SchedRW = [WriteALU] in {
275 // FIXME: These are pseudo ops that should be replaced with Pat<> patterns.
276 // However, Pat<> can't replicate the destination reg into the inputs of the
278 def SETB_C8r : I<0, Pseudo, (outs GR8:$dst), (ins), "",
279 [(set GR8:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
280 def SETB_C16r : I<0, Pseudo, (outs GR16:$dst), (ins), "",
281 [(set GR16:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
282 def SETB_C32r : I<0, Pseudo, (outs GR32:$dst), (ins), "",
283 [(set GR32:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
284 def SETB_C64r : I<0, Pseudo, (outs GR64:$dst), (ins), "",
285 [(set GR64:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
289 def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
291 def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
293 def : Pat<(i64 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
296 def : Pat<(i16 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
298 def : Pat<(i32 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
300 def : Pat<(i64 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
303 // We canonicalize 'setb' to "(and (sbb reg,reg), 1)" on the hope that the and
304 // will be eliminated and that the sbb can be extended up to a wider type. When
305 // this happens, it is great. However, if we are left with an 8-bit sbb and an
306 // and, we might as well just match it as a setb.
307 def : Pat<(and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1),
310 // (add OP, SETB) -> (adc OP, 0)
311 def : Pat<(add (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR8:$op),
312 (ADC8ri GR8:$op, 0)>;
313 def : Pat<(add (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR32:$op),
314 (ADC32ri8 GR32:$op, 0)>;
315 def : Pat<(add (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR64:$op),
316 (ADC64ri8 GR64:$op, 0)>;
318 // (sub OP, SETB) -> (sbb OP, 0)
319 def : Pat<(sub GR8:$op, (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
320 (SBB8ri GR8:$op, 0)>;
321 def : Pat<(sub GR32:$op, (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
322 (SBB32ri8 GR32:$op, 0)>;
323 def : Pat<(sub GR64:$op, (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
324 (SBB64ri8 GR64:$op, 0)>;
326 // (sub OP, SETCC_CARRY) -> (adc OP, 0)
327 def : Pat<(sub (add GR8:$op1, GR8:$op2), (i8 (X86setcc_c X86_COND_B, EFLAGS))),
328 (ADC8ri GR8:$op1, GR8:$op2)>;
329 def : Pat<(sub GR8:$op, (i8 (X86setcc_c X86_COND_B, EFLAGS))),
330 (ADC8ri GR8:$op, 0)>;
331 def : Pat<(sub (add GR32:$op1, GR32:$op2), (i32 (X86setcc_c X86_COND_B, EFLAGS))),
332 (ADC32ri8 GR32:$op1, GR32:$op2)>;
333 def : Pat<(sub GR32:$op, (i32 (X86setcc_c X86_COND_B, EFLAGS))),
334 (ADC32ri8 GR32:$op, 0)>;
335 def : Pat<(sub (add GR64:$op1, GR64:$op2), (i64 (X86setcc_c X86_COND_B, EFLAGS))),
336 (ADC64ri8 GR64:$op1, GR64:$op2)>;
337 def : Pat<(sub GR64:$op, (i64 (X86setcc_c X86_COND_B, EFLAGS))),
338 (ADC64ri8 GR64:$op, 0)>;
340 //===----------------------------------------------------------------------===//
341 // String Pseudo Instructions
343 let SchedRW = [WriteMicrocoded] in {
344 let Defs = [ECX,EDI,ESI], Uses = [ECX,EDI,ESI], isCodeGenOnly = 1 in {
345 def REP_MOVSB_32 : I<0xA4, RawFrm, (outs), (ins), "{rep;movsb|rep movsb}",
346 [(X86rep_movs i8)], IIC_REP_MOVS>, REP,
347 Requires<[Not64BitMode]>;
348 def REP_MOVSW_32 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsw|rep movsw}",
349 [(X86rep_movs i16)], IIC_REP_MOVS>, REP, OpSize16,
350 Requires<[Not64BitMode]>;
351 def REP_MOVSD_32 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsl|rep movsd}",
352 [(X86rep_movs i32)], IIC_REP_MOVS>, REP, OpSize32,
353 Requires<[Not64BitMode]>;
356 let Defs = [RCX,RDI,RSI], Uses = [RCX,RDI,RSI], isCodeGenOnly = 1 in {
357 def REP_MOVSB_64 : I<0xA4, RawFrm, (outs), (ins), "{rep;movsb|rep movsb}",
358 [(X86rep_movs i8)], IIC_REP_MOVS>, REP,
359 Requires<[In64BitMode]>;
360 def REP_MOVSW_64 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsw|rep movsw}",
361 [(X86rep_movs i16)], IIC_REP_MOVS>, REP, OpSize16,
362 Requires<[In64BitMode]>;
363 def REP_MOVSD_64 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsl|rep movsd}",
364 [(X86rep_movs i32)], IIC_REP_MOVS>, REP, OpSize32,
365 Requires<[In64BitMode]>;
366 def REP_MOVSQ_64 : RI<0xA5, RawFrm, (outs), (ins), "{rep;movsq|rep movsq}",
367 [(X86rep_movs i64)], IIC_REP_MOVS>, REP,
368 Requires<[In64BitMode]>;
371 // FIXME: Should use "(X86rep_stos AL)" as the pattern.
372 let Defs = [ECX,EDI], isCodeGenOnly = 1 in {
373 let Uses = [AL,ECX,EDI] in
374 def REP_STOSB_32 : I<0xAA, RawFrm, (outs), (ins), "{rep;stosb|rep stosb}",
375 [(X86rep_stos i8)], IIC_REP_STOS>, REP,
376 Requires<[Not64BitMode]>;
377 let Uses = [AX,ECX,EDI] in
378 def REP_STOSW_32 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosw|rep stosw}",
379 [(X86rep_stos i16)], IIC_REP_STOS>, REP, OpSize16,
380 Requires<[Not64BitMode]>;
381 let Uses = [EAX,ECX,EDI] in
382 def REP_STOSD_32 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosl|rep stosd}",
383 [(X86rep_stos i32)], IIC_REP_STOS>, REP, OpSize32,
384 Requires<[Not64BitMode]>;
387 let Defs = [RCX,RDI], isCodeGenOnly = 1 in {
388 let Uses = [AL,RCX,RDI] in
389 def REP_STOSB_64 : I<0xAA, RawFrm, (outs), (ins), "{rep;stosb|rep stosb}",
390 [(X86rep_stos i8)], IIC_REP_STOS>, REP,
391 Requires<[In64BitMode]>;
392 let Uses = [AX,RCX,RDI] in
393 def REP_STOSW_64 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosw|rep stosw}",
394 [(X86rep_stos i16)], IIC_REP_STOS>, REP, OpSize16,
395 Requires<[In64BitMode]>;
396 let Uses = [RAX,RCX,RDI] in
397 def REP_STOSD_64 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosl|rep stosd}",
398 [(X86rep_stos i32)], IIC_REP_STOS>, REP, OpSize32,
399 Requires<[In64BitMode]>;
401 let Uses = [RAX,RCX,RDI] in
402 def REP_STOSQ_64 : RI<0xAB, RawFrm, (outs), (ins), "{rep;stosq|rep stosq}",
403 [(X86rep_stos i64)], IIC_REP_STOS>, REP,
404 Requires<[In64BitMode]>;
408 //===----------------------------------------------------------------------===//
409 // Thread Local Storage Instructions
413 // All calls clobber the non-callee saved registers. ESP is marked as
414 // a use to prevent stack-pointer assignments that appear immediately
415 // before calls from potentially appearing dead.
416 let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0,
417 MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
418 XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
419 XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
421 def TLS_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
423 [(X86tlsaddr tls32addr:$sym)]>,
424 Requires<[Not64BitMode]>;
425 def TLS_base_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
427 [(X86tlsbaseaddr tls32baseaddr:$sym)]>,
428 Requires<[Not64BitMode]>;
431 // All calls clobber the non-callee saved registers. RSP is marked as
432 // a use to prevent stack-pointer assignments that appear immediately
433 // before calls from potentially appearing dead.
434 let Defs = [RAX, RCX, RDX, RSI, RDI, R8, R9, R10, R11,
435 FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0, ST1,
436 MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
437 XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
438 XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
440 def TLS_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
442 [(X86tlsaddr tls64addr:$sym)]>,
443 Requires<[In64BitMode]>;
444 def TLS_base_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
446 [(X86tlsbaseaddr tls64baseaddr:$sym)]>,
447 Requires<[In64BitMode]>;
450 // Darwin TLS Support
451 // For i386, the address of the thunk is passed on the stack, on return the
452 // address of the variable is in %eax. %ecx is trashed during the function
453 // call. All other registers are preserved.
454 let Defs = [EAX, ECX, EFLAGS],
456 usesCustomInserter = 1 in
457 def TLSCall_32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
459 [(X86TLSCall addr:$sym)]>,
460 Requires<[Not64BitMode]>;
462 // For x86_64, the address of the thunk is passed in %rdi, on return
463 // the address of the variable is in %rax. All other registers are preserved.
464 let Defs = [RAX, EFLAGS],
466 usesCustomInserter = 1 in
467 def TLSCall_64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
469 [(X86TLSCall addr:$sym)]>,
470 Requires<[In64BitMode]>;
473 //===----------------------------------------------------------------------===//
474 // Conditional Move Pseudo Instructions
476 // X86 doesn't have 8-bit conditional moves. Use a customInserter to
477 // emit control flow. An alternative to this is to mark i8 SELECT as Promote,
478 // however that requires promoting the operands, and can induce additional
479 // i8 register pressure.
480 let usesCustomInserter = 1, Uses = [EFLAGS] in {
481 def CMOV_GR8 : I<0, Pseudo,
482 (outs GR8:$dst), (ins GR8:$src1, GR8:$src2, i8imm:$cond),
484 [(set GR8:$dst, (X86cmov GR8:$src1, GR8:$src2,
485 imm:$cond, EFLAGS))]>;
487 let Predicates = [NoCMov] in {
488 def CMOV_GR32 : I<0, Pseudo,
489 (outs GR32:$dst), (ins GR32:$src1, GR32:$src2, i8imm:$cond),
490 "#CMOV_GR32* PSEUDO!",
492 (X86cmov GR32:$src1, GR32:$src2, imm:$cond, EFLAGS))]>;
493 def CMOV_GR16 : I<0, Pseudo,
494 (outs GR16:$dst), (ins GR16:$src1, GR16:$src2, i8imm:$cond),
495 "#CMOV_GR16* PSEUDO!",
497 (X86cmov GR16:$src1, GR16:$src2, imm:$cond, EFLAGS))]>;
498 } // Predicates = [NoCMov]
500 // fcmov doesn't handle all possible EFLAGS, provide a fallback if there is no
502 let Predicates = [FPStackf32] in
503 def CMOV_RFP32 : I<0, Pseudo,
505 (ins RFP32:$src1, RFP32:$src2, i8imm:$cond),
506 "#CMOV_RFP32 PSEUDO!",
508 (X86cmov RFP32:$src1, RFP32:$src2, imm:$cond,
510 // fcmov doesn't handle all possible EFLAGS, provide a fallback if there is no
512 let Predicates = [FPStackf64] in
513 def CMOV_RFP64 : I<0, Pseudo,
515 (ins RFP64:$src1, RFP64:$src2, i8imm:$cond),
516 "#CMOV_RFP64 PSEUDO!",
518 (X86cmov RFP64:$src1, RFP64:$src2, imm:$cond,
520 def CMOV_RFP80 : I<0, Pseudo,
522 (ins RFP80:$src1, RFP80:$src2, i8imm:$cond),
523 "#CMOV_RFP80 PSEUDO!",
525 (X86cmov RFP80:$src1, RFP80:$src2, imm:$cond,
527 } // UsesCustomInserter = 1, Uses = [EFLAGS]
530 //===----------------------------------------------------------------------===//
531 // Normal-Instructions-With-Lock-Prefix Pseudo Instructions
532 //===----------------------------------------------------------------------===//
534 // FIXME: Use normal instructions and add lock prefix dynamically.
538 // TODO: Get this to fold the constant into the instruction.
539 let isCodeGenOnly = 1, Defs = [EFLAGS] in
540 def OR32mrLocked : I<0x09, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$zero),
541 "or{l}\t{$zero, $dst|$dst, $zero}",
542 [], IIC_ALU_MEM>, Requires<[Not64BitMode]>, LOCK,
543 Sched<[WriteALULd, WriteRMW]>;
545 let hasSideEffects = 1 in
546 def Int_MemBarrier : I<0, Pseudo, (outs), (ins),
548 [(X86MemBarrier)]>, Sched<[WriteLoad]>;
550 // RegOpc corresponds to the mr version of the instruction
551 // ImmOpc corresponds to the mi version of the instruction
552 // ImmOpc8 corresponds to the mi8 version of the instruction
553 // ImmMod corresponds to the instruction format of the mi and mi8 versions
554 multiclass LOCK_ArithBinOp<bits<8> RegOpc, bits<8> ImmOpc, bits<8> ImmOpc8,
555 Format ImmMod, string mnemonic> {
556 let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
557 SchedRW = [WriteALULd, WriteRMW] in {
559 def NAME#8mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
560 RegOpc{3}, RegOpc{2}, RegOpc{1}, 0 },
561 MRMDestMem, (outs), (ins i8mem:$dst, GR8:$src2),
562 !strconcat(mnemonic, "{b}\t",
563 "{$src2, $dst|$dst, $src2}"),
564 [], IIC_ALU_NONMEM>, LOCK;
565 def NAME#16mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
566 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
567 MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
568 !strconcat(mnemonic, "{w}\t",
569 "{$src2, $dst|$dst, $src2}"),
570 [], IIC_ALU_NONMEM>, OpSize16, LOCK;
571 def NAME#32mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
572 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
573 MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
574 !strconcat(mnemonic, "{l}\t",
575 "{$src2, $dst|$dst, $src2}"),
576 [], IIC_ALU_NONMEM>, OpSize32, LOCK;
577 def NAME#64mr : RI<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
578 RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
579 MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
580 !strconcat(mnemonic, "{q}\t",
581 "{$src2, $dst|$dst, $src2}"),
582 [], IIC_ALU_NONMEM>, LOCK;
584 def NAME#8mi : Ii8<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
585 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 0 },
586 ImmMod, (outs), (ins i8mem :$dst, i8imm :$src2),
587 !strconcat(mnemonic, "{b}\t",
588 "{$src2, $dst|$dst, $src2}"),
589 [], IIC_ALU_MEM>, LOCK;
591 def NAME#16mi : Ii16<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
592 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
593 ImmMod, (outs), (ins i16mem :$dst, i16imm :$src2),
594 !strconcat(mnemonic, "{w}\t",
595 "{$src2, $dst|$dst, $src2}"),
596 [], IIC_ALU_MEM>, OpSize16, LOCK;
598 def NAME#32mi : Ii32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
599 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
600 ImmMod, (outs), (ins i32mem :$dst, i32imm :$src2),
601 !strconcat(mnemonic, "{l}\t",
602 "{$src2, $dst|$dst, $src2}"),
603 [], IIC_ALU_MEM>, OpSize32, LOCK;
605 def NAME#64mi32 : RIi32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
606 ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
607 ImmMod, (outs), (ins i64mem :$dst, i64i32imm :$src2),
608 !strconcat(mnemonic, "{q}\t",
609 "{$src2, $dst|$dst, $src2}"),
610 [], IIC_ALU_MEM>, LOCK;
612 def NAME#16mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
613 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
614 ImmMod, (outs), (ins i16mem :$dst, i16i8imm :$src2),
615 !strconcat(mnemonic, "{w}\t",
616 "{$src2, $dst|$dst, $src2}"),
617 [], IIC_ALU_MEM>, OpSize16, LOCK;
618 def NAME#32mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
619 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
620 ImmMod, (outs), (ins i32mem :$dst, i32i8imm :$src2),
621 !strconcat(mnemonic, "{l}\t",
622 "{$src2, $dst|$dst, $src2}"),
623 [], IIC_ALU_MEM>, OpSize32, LOCK;
624 def NAME#64mi8 : RIi8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
625 ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
626 ImmMod, (outs), (ins i64mem :$dst, i64i8imm :$src2),
627 !strconcat(mnemonic, "{q}\t",
628 "{$src2, $dst|$dst, $src2}"),
629 [], IIC_ALU_MEM>, LOCK;
635 defm LOCK_ADD : LOCK_ArithBinOp<0x00, 0x80, 0x83, MRM0m, "add">;
636 defm LOCK_SUB : LOCK_ArithBinOp<0x28, 0x80, 0x83, MRM5m, "sub">;
637 defm LOCK_OR : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM1m, "or">;
638 defm LOCK_AND : LOCK_ArithBinOp<0x20, 0x80, 0x83, MRM4m, "and">;
639 defm LOCK_XOR : LOCK_ArithBinOp<0x30, 0x80, 0x83, MRM6m, "xor">;
641 // Optimized codegen when the non-memory output is not used.
642 multiclass LOCK_ArithUnOp<bits<8> Opc8, bits<8> Opc, Format Form,
644 let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
645 SchedRW = [WriteALULd, WriteRMW] in {
647 def NAME#8m : I<Opc8, Form, (outs), (ins i8mem :$dst),
648 !strconcat(mnemonic, "{b}\t$dst"),
649 [], IIC_UNARY_MEM>, LOCK;
650 def NAME#16m : I<Opc, Form, (outs), (ins i16mem:$dst),
651 !strconcat(mnemonic, "{w}\t$dst"),
652 [], IIC_UNARY_MEM>, OpSize16, LOCK;
653 def NAME#32m : I<Opc, Form, (outs), (ins i32mem:$dst),
654 !strconcat(mnemonic, "{l}\t$dst"),
655 [], IIC_UNARY_MEM>, OpSize32, LOCK;
656 def NAME#64m : RI<Opc, Form, (outs), (ins i64mem:$dst),
657 !strconcat(mnemonic, "{q}\t$dst"),
658 [], IIC_UNARY_MEM>, LOCK;
662 defm LOCK_INC : LOCK_ArithUnOp<0xFE, 0xFF, MRM0m, "inc">;
663 defm LOCK_DEC : LOCK_ArithUnOp<0xFE, 0xFF, MRM1m, "dec">;
665 // Atomic compare and swap.
666 multiclass LCMPXCHG_UnOp<bits<8> Opc, Format Form, string mnemonic,
667 SDPatternOperator frag, X86MemOperand x86memop,
668 InstrItinClass itin> {
669 let isCodeGenOnly = 1 in {
670 def NAME : I<Opc, Form, (outs), (ins x86memop:$ptr),
671 !strconcat(mnemonic, "\t$ptr"),
672 [(frag addr:$ptr)], itin>, TB, LOCK;
676 multiclass LCMPXCHG_BinOp<bits<8> Opc8, bits<8> Opc, Format Form,
677 string mnemonic, SDPatternOperator frag,
678 InstrItinClass itin8, InstrItinClass itin> {
679 let isCodeGenOnly = 1, SchedRW = [WriteALULd, WriteRMW] in {
680 let Defs = [AL, EFLAGS], Uses = [AL] in
681 def NAME#8 : I<Opc8, Form, (outs), (ins i8mem:$ptr, GR8:$swap),
682 !strconcat(mnemonic, "{b}\t{$swap, $ptr|$ptr, $swap}"),
683 [(frag addr:$ptr, GR8:$swap, 1)], itin8>, TB, LOCK;
684 let Defs = [AX, EFLAGS], Uses = [AX] in
685 def NAME#16 : I<Opc, Form, (outs), (ins i16mem:$ptr, GR16:$swap),
686 !strconcat(mnemonic, "{w}\t{$swap, $ptr|$ptr, $swap}"),
687 [(frag addr:$ptr, GR16:$swap, 2)], itin>, TB, OpSize16, LOCK;
688 let Defs = [EAX, EFLAGS], Uses = [EAX] in
689 def NAME#32 : I<Opc, Form, (outs), (ins i32mem:$ptr, GR32:$swap),
690 !strconcat(mnemonic, "{l}\t{$swap, $ptr|$ptr, $swap}"),
691 [(frag addr:$ptr, GR32:$swap, 4)], itin>, TB, OpSize32, LOCK;
692 let Defs = [RAX, EFLAGS], Uses = [RAX] in
693 def NAME#64 : RI<Opc, Form, (outs), (ins i64mem:$ptr, GR64:$swap),
694 !strconcat(mnemonic, "{q}\t{$swap, $ptr|$ptr, $swap}"),
695 [(frag addr:$ptr, GR64:$swap, 8)], itin>, TB, LOCK;
699 let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
700 SchedRW = [WriteALULd, WriteRMW] in {
701 defm LCMPXCHG8B : LCMPXCHG_UnOp<0xC7, MRM1m, "cmpxchg8b",
706 let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RBX, RCX, RDX],
707 Predicates = [HasCmpxchg16b], SchedRW = [WriteALULd, WriteRMW] in {
708 defm LCMPXCHG16B : LCMPXCHG_UnOp<0xC7, MRM1m, "cmpxchg16b",
710 IIC_CMPX_LOCK_16B>, REX_W;
713 defm LCMPXCHG : LCMPXCHG_BinOp<0xB0, 0xB1, MRMDestMem, "cmpxchg",
714 X86cas, IIC_CMPX_LOCK_8, IIC_CMPX_LOCK>;
716 // Atomic exchange and add
717 multiclass ATOMIC_LOAD_BINOP<bits<8> opc8, bits<8> opc, string mnemonic,
719 InstrItinClass itin8, InstrItinClass itin> {
720 let Constraints = "$val = $dst", Defs = [EFLAGS], isCodeGenOnly = 1,
721 SchedRW = [WriteALULd, WriteRMW] in {
722 def NAME#8 : I<opc8, MRMSrcMem, (outs GR8:$dst),
723 (ins GR8:$val, i8mem:$ptr),
724 !strconcat(mnemonic, "{b}\t{$val, $ptr|$ptr, $val}"),
726 (!cast<PatFrag>(frag # "_8") addr:$ptr, GR8:$val))],
728 def NAME#16 : I<opc, MRMSrcMem, (outs GR16:$dst),
729 (ins GR16:$val, i16mem:$ptr),
730 !strconcat(mnemonic, "{w}\t{$val, $ptr|$ptr, $val}"),
733 (!cast<PatFrag>(frag # "_16") addr:$ptr, GR16:$val))],
735 def NAME#32 : I<opc, MRMSrcMem, (outs GR32:$dst),
736 (ins GR32:$val, i32mem:$ptr),
737 !strconcat(mnemonic, "{l}\t{$val, $ptr|$ptr, $val}"),
740 (!cast<PatFrag>(frag # "_32") addr:$ptr, GR32:$val))],
742 def NAME#64 : RI<opc, MRMSrcMem, (outs GR64:$dst),
743 (ins GR64:$val, i64mem:$ptr),
744 !strconcat(mnemonic, "{q}\t{$val, $ptr|$ptr, $val}"),
747 (!cast<PatFrag>(frag # "_64") addr:$ptr, GR64:$val))],
752 defm LXADD : ATOMIC_LOAD_BINOP<0xc0, 0xc1, "xadd", "atomic_load_add",
753 IIC_XADD_LOCK_MEM8, IIC_XADD_LOCK_MEM>,
756 def ACQUIRE_MOV8rm : I<0, Pseudo, (outs GR8 :$dst), (ins i8mem :$src),
757 "#ACQUIRE_MOV PSEUDO!",
758 [(set GR8:$dst, (atomic_load_8 addr:$src))]>;
759 def ACQUIRE_MOV16rm : I<0, Pseudo, (outs GR16:$dst), (ins i16mem:$src),
760 "#ACQUIRE_MOV PSEUDO!",
761 [(set GR16:$dst, (atomic_load_16 addr:$src))]>;
762 def ACQUIRE_MOV32rm : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$src),
763 "#ACQUIRE_MOV PSEUDO!",
764 [(set GR32:$dst, (atomic_load_32 addr:$src))]>;
765 def ACQUIRE_MOV64rm : I<0, Pseudo, (outs GR64:$dst), (ins i64mem:$src),
766 "#ACQUIRE_MOV PSEUDO!",
767 [(set GR64:$dst, (atomic_load_64 addr:$src))]>;
769 def RELEASE_MOV8mr : I<0, Pseudo, (outs), (ins i8mem :$dst, GR8 :$src),
770 "#RELEASE_MOV PSEUDO!",
771 [(atomic_store_8 addr:$dst, GR8 :$src)]>;
772 def RELEASE_MOV16mr : I<0, Pseudo, (outs), (ins i16mem:$dst, GR16:$src),
773 "#RELEASE_MOV PSEUDO!",
774 [(atomic_store_16 addr:$dst, GR16:$src)]>;
775 def RELEASE_MOV32mr : I<0, Pseudo, (outs), (ins i32mem:$dst, GR32:$src),
776 "#RELEASE_MOV PSEUDO!",
777 [(atomic_store_32 addr:$dst, GR32:$src)]>;
778 def RELEASE_MOV64mr : I<0, Pseudo, (outs), (ins i64mem:$dst, GR64:$src),
779 "#RELEASE_MOV PSEUDO!",
780 [(atomic_store_64 addr:$dst, GR64:$src)]>;
782 //===----------------------------------------------------------------------===//
783 // Conditional Move Pseudo Instructions.
784 //===----------------------------------------------------------------------===//
787 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded after
788 // instruction selection into a branch sequence.
789 let Uses = [EFLAGS], usesCustomInserter = 1 in {
790 def CMOV_FR32 : I<0, Pseudo,
791 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
792 "#CMOV_FR32 PSEUDO!",
793 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
795 def CMOV_FR64 : I<0, Pseudo,
796 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
797 "#CMOV_FR64 PSEUDO!",
798 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
800 def CMOV_V4F32 : I<0, Pseudo,
801 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
802 "#CMOV_V4F32 PSEUDO!",
804 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
806 def CMOV_V2F64 : I<0, Pseudo,
807 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
808 "#CMOV_V2F64 PSEUDO!",
810 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
812 def CMOV_V2I64 : I<0, Pseudo,
813 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
814 "#CMOV_V2I64 PSEUDO!",
816 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
818 def CMOV_V8F32 : I<0, Pseudo,
819 (outs VR256:$dst), (ins VR256:$t, VR256:$f, i8imm:$cond),
820 "#CMOV_V8F32 PSEUDO!",
822 (v8f32 (X86cmov VR256:$t, VR256:$f, imm:$cond,
824 def CMOV_V4F64 : I<0, Pseudo,
825 (outs VR256:$dst), (ins VR256:$t, VR256:$f, i8imm:$cond),
826 "#CMOV_V4F64 PSEUDO!",
828 (v4f64 (X86cmov VR256:$t, VR256:$f, imm:$cond,
830 def CMOV_V4I64 : I<0, Pseudo,
831 (outs VR256:$dst), (ins VR256:$t, VR256:$f, i8imm:$cond),
832 "#CMOV_V4I64 PSEUDO!",
834 (v4i64 (X86cmov VR256:$t, VR256:$f, imm:$cond,
836 def CMOV_V8I64 : I<0, Pseudo,
837 (outs VR512:$dst), (ins VR512:$t, VR512:$f, i8imm:$cond),
838 "#CMOV_V8I64 PSEUDO!",
840 (v8i64 (X86cmov VR512:$t, VR512:$f, imm:$cond,
842 def CMOV_V8F64 : I<0, Pseudo,
843 (outs VR512:$dst), (ins VR512:$t, VR512:$f, i8imm:$cond),
844 "#CMOV_V8F64 PSEUDO!",
846 (v8f64 (X86cmov VR512:$t, VR512:$f, imm:$cond,
848 def CMOV_V16F32 : I<0, Pseudo,
849 (outs VR512:$dst), (ins VR512:$t, VR512:$f, i8imm:$cond),
850 "#CMOV_V16F32 PSEUDO!",
852 (v16f32 (X86cmov VR512:$t, VR512:$f, imm:$cond,
857 //===----------------------------------------------------------------------===//
858 // DAG Pattern Matching Rules
859 //===----------------------------------------------------------------------===//
861 // ConstantPool GlobalAddress, ExternalSymbol, and JumpTable
862 def : Pat<(i32 (X86Wrapper tconstpool :$dst)), (MOV32ri tconstpool :$dst)>;
863 def : Pat<(i32 (X86Wrapper tjumptable :$dst)), (MOV32ri tjumptable :$dst)>;
864 def : Pat<(i32 (X86Wrapper tglobaltlsaddr:$dst)),(MOV32ri tglobaltlsaddr:$dst)>;
865 def : Pat<(i32 (X86Wrapper tglobaladdr :$dst)), (MOV32ri tglobaladdr :$dst)>;
866 def : Pat<(i32 (X86Wrapper texternalsym:$dst)), (MOV32ri texternalsym:$dst)>;
867 def : Pat<(i32 (X86Wrapper tblockaddress:$dst)), (MOV32ri tblockaddress:$dst)>;
869 def : Pat<(add GR32:$src1, (X86Wrapper tconstpool:$src2)),
870 (ADD32ri GR32:$src1, tconstpool:$src2)>;
871 def : Pat<(add GR32:$src1, (X86Wrapper tjumptable:$src2)),
872 (ADD32ri GR32:$src1, tjumptable:$src2)>;
873 def : Pat<(add GR32:$src1, (X86Wrapper tglobaladdr :$src2)),
874 (ADD32ri GR32:$src1, tglobaladdr:$src2)>;
875 def : Pat<(add GR32:$src1, (X86Wrapper texternalsym:$src2)),
876 (ADD32ri GR32:$src1, texternalsym:$src2)>;
877 def : Pat<(add GR32:$src1, (X86Wrapper tblockaddress:$src2)),
878 (ADD32ri GR32:$src1, tblockaddress:$src2)>;
880 def : Pat<(store (i32 (X86Wrapper tglobaladdr:$src)), addr:$dst),
881 (MOV32mi addr:$dst, tglobaladdr:$src)>;
882 def : Pat<(store (i32 (X86Wrapper texternalsym:$src)), addr:$dst),
883 (MOV32mi addr:$dst, texternalsym:$src)>;
884 def : Pat<(store (i32 (X86Wrapper tblockaddress:$src)), addr:$dst),
885 (MOV32mi addr:$dst, tblockaddress:$src)>;
887 // ConstantPool GlobalAddress, ExternalSymbol, and JumpTable when not in small
888 // code model mode, should use 'movabs'. FIXME: This is really a hack, the
889 // 'movabs' predicate should handle this sort of thing.
890 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
891 (MOV64ri tconstpool :$dst)>, Requires<[FarData]>;
892 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
893 (MOV64ri tjumptable :$dst)>, Requires<[FarData]>;
894 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
895 (MOV64ri tglobaladdr :$dst)>, Requires<[FarData]>;
896 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
897 (MOV64ri texternalsym:$dst)>, Requires<[FarData]>;
898 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
899 (MOV64ri tblockaddress:$dst)>, Requires<[FarData]>;
901 // In kernel code model, we can get the address of a label
902 // into a register with 'movq'. FIXME: This is a hack, the 'imm' predicate of
903 // the MOV64ri32 should accept these.
904 def : Pat<(i64 (X86Wrapper tconstpool :$dst)),
905 (MOV64ri32 tconstpool :$dst)>, Requires<[KernelCode]>;
906 def : Pat<(i64 (X86Wrapper tjumptable :$dst)),
907 (MOV64ri32 tjumptable :$dst)>, Requires<[KernelCode]>;
908 def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
909 (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
910 def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
911 (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
912 def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
913 (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
915 // If we have small model and -static mode, it is safe to store global addresses
916 // directly as immediates. FIXME: This is really a hack, the 'imm' predicate
917 // for MOV64mi32 should handle this sort of thing.
918 def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
919 (MOV64mi32 addr:$dst, tconstpool:$src)>,
920 Requires<[NearData, IsStatic]>;
921 def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
922 (MOV64mi32 addr:$dst, tjumptable:$src)>,
923 Requires<[NearData, IsStatic]>;
924 def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
925 (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
926 Requires<[NearData, IsStatic]>;
927 def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
928 (MOV64mi32 addr:$dst, texternalsym:$src)>,
929 Requires<[NearData, IsStatic]>;
930 def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
931 (MOV64mi32 addr:$dst, tblockaddress:$src)>,
932 Requires<[NearData, IsStatic]>;
936 // tls has some funny stuff here...
937 // This corresponds to movabs $foo@tpoff, %rax
938 def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
939 (MOV64ri32 tglobaltlsaddr :$dst)>;
940 // This corresponds to add $foo@tpoff, %rax
941 def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
942 (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
945 // Direct PC relative function call for small code model. 32-bit displacement
946 // sign extended to 64-bit.
947 def : Pat<(X86call (i64 tglobaladdr:$dst)),
948 (CALL64pcrel32 tglobaladdr:$dst)>;
949 def : Pat<(X86call (i64 texternalsym:$dst)),
950 (CALL64pcrel32 texternalsym:$dst)>;
952 // Tailcall stuff. The TCRETURN instructions execute after the epilog, so they
953 // can never use callee-saved registers. That is the purpose of the GR64_TC
956 // The only volatile register that is never used by the calling convention is
957 // %r11. This happens when calling a vararg function with 6 arguments.
959 // Match an X86tcret that uses less than 7 volatile registers.
960 def X86tcret_6regs : PatFrag<(ops node:$ptr, node:$off),
961 (X86tcret node:$ptr, node:$off), [{
962 // X86tcret args: (*chain, ptr, imm, regs..., glue)
963 unsigned NumRegs = 0;
964 for (unsigned i = 3, e = N->getNumOperands(); i != e; ++i)
965 if (isa<RegisterSDNode>(N->getOperand(i)) && ++NumRegs > 6)
970 def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
971 (TCRETURNri ptr_rc_tailcall:$dst, imm:$off)>,
972 Requires<[Not64BitMode]>;
974 // FIXME: This is disabled for 32-bit PIC mode because the global base
975 // register which is part of the address mode may be assigned a
976 // callee-saved register.
977 def : Pat<(X86tcret (load addr:$dst), imm:$off),
978 (TCRETURNmi addr:$dst, imm:$off)>,
979 Requires<[Not64BitMode, IsNotPIC]>;
981 def : Pat<(X86tcret (i32 tglobaladdr:$dst), imm:$off),
982 (TCRETURNdi texternalsym:$dst, imm:$off)>,
983 Requires<[Not64BitMode]>;
985 def : Pat<(X86tcret (i32 texternalsym:$dst), imm:$off),
986 (TCRETURNdi texternalsym:$dst, imm:$off)>,
987 Requires<[Not64BitMode]>;
989 def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
990 (TCRETURNri64 ptr_rc_tailcall:$dst, imm:$off)>,
991 Requires<[In64BitMode]>;
993 // Don't fold loads into X86tcret requiring more than 6 regs.
994 // There wouldn't be enough scratch registers for base+index.
995 def : Pat<(X86tcret_6regs (load addr:$dst), imm:$off),
996 (TCRETURNmi64 addr:$dst, imm:$off)>,
997 Requires<[In64BitMode]>;
999 def : Pat<(X86tcret (i64 tglobaladdr:$dst), imm:$off),
1000 (TCRETURNdi64 tglobaladdr:$dst, imm:$off)>,
1001 Requires<[In64BitMode]>;
1003 def : Pat<(X86tcret (i64 texternalsym:$dst), imm:$off),
1004 (TCRETURNdi64 texternalsym:$dst, imm:$off)>,
1005 Requires<[In64BitMode]>;
1007 // Normal calls, with various flavors of addresses.
1008 def : Pat<(X86call (i32 tglobaladdr:$dst)),
1009 (CALLpcrel32 tglobaladdr:$dst)>;
1010 def : Pat<(X86call (i32 texternalsym:$dst)),
1011 (CALLpcrel32 texternalsym:$dst)>;
1012 def : Pat<(X86call (i32 imm:$dst)),
1013 (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
1017 // TEST R,R is smaller than CMP R,0
1018 def : Pat<(X86cmp GR8:$src1, 0),
1019 (TEST8rr GR8:$src1, GR8:$src1)>;
1020 def : Pat<(X86cmp GR16:$src1, 0),
1021 (TEST16rr GR16:$src1, GR16:$src1)>;
1022 def : Pat<(X86cmp GR32:$src1, 0),
1023 (TEST32rr GR32:$src1, GR32:$src1)>;
1024 def : Pat<(X86cmp GR64:$src1, 0),
1025 (TEST64rr GR64:$src1, GR64:$src1)>;
1027 // Conditional moves with folded loads with operands swapped and conditions
1029 multiclass CMOVmr<PatLeaf InvertedCond, Instruction Inst16, Instruction Inst32,
1030 Instruction Inst64> {
1031 let Predicates = [HasCMov] in {
1032 def : Pat<(X86cmov (loadi16 addr:$src1), GR16:$src2, InvertedCond, EFLAGS),
1033 (Inst16 GR16:$src2, addr:$src1)>;
1034 def : Pat<(X86cmov (loadi32 addr:$src1), GR32:$src2, InvertedCond, EFLAGS),
1035 (Inst32 GR32:$src2, addr:$src1)>;
1036 def : Pat<(X86cmov (loadi64 addr:$src1), GR64:$src2, InvertedCond, EFLAGS),
1037 (Inst64 GR64:$src2, addr:$src1)>;
1041 defm : CMOVmr<X86_COND_B , CMOVAE16rm, CMOVAE32rm, CMOVAE64rm>;
1042 defm : CMOVmr<X86_COND_AE, CMOVB16rm , CMOVB32rm , CMOVB64rm>;
1043 defm : CMOVmr<X86_COND_E , CMOVNE16rm, CMOVNE32rm, CMOVNE64rm>;
1044 defm : CMOVmr<X86_COND_NE, CMOVE16rm , CMOVE32rm , CMOVE64rm>;
1045 defm : CMOVmr<X86_COND_BE, CMOVA16rm , CMOVA32rm , CMOVA64rm>;
1046 defm : CMOVmr<X86_COND_A , CMOVBE16rm, CMOVBE32rm, CMOVBE64rm>;
1047 defm : CMOVmr<X86_COND_L , CMOVGE16rm, CMOVGE32rm, CMOVGE64rm>;
1048 defm : CMOVmr<X86_COND_GE, CMOVL16rm , CMOVL32rm , CMOVL64rm>;
1049 defm : CMOVmr<X86_COND_LE, CMOVG16rm , CMOVG32rm , CMOVG64rm>;
1050 defm : CMOVmr<X86_COND_G , CMOVLE16rm, CMOVLE32rm, CMOVLE64rm>;
1051 defm : CMOVmr<X86_COND_P , CMOVNP16rm, CMOVNP32rm, CMOVNP64rm>;
1052 defm : CMOVmr<X86_COND_NP, CMOVP16rm , CMOVP32rm , CMOVP64rm>;
1053 defm : CMOVmr<X86_COND_S , CMOVNS16rm, CMOVNS32rm, CMOVNS64rm>;
1054 defm : CMOVmr<X86_COND_NS, CMOVS16rm , CMOVS32rm , CMOVS64rm>;
1055 defm : CMOVmr<X86_COND_O , CMOVNO16rm, CMOVNO32rm, CMOVNO64rm>;
1056 defm : CMOVmr<X86_COND_NO, CMOVO16rm , CMOVO32rm , CMOVO64rm>;
1058 // zextload bool -> zextload byte
1059 def : Pat<(zextloadi8i1 addr:$src), (MOV8rm addr:$src)>;
1060 def : Pat<(zextloadi16i1 addr:$src), (MOVZX16rm8 addr:$src)>;
1061 def : Pat<(zextloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1062 def : Pat<(zextloadi64i1 addr:$src),
1063 (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1065 // extload bool -> extload byte
1066 // When extloading from 16-bit and smaller memory locations into 64-bit
1067 // registers, use zero-extending loads so that the entire 64-bit register is
1068 // defined, avoiding partial-register updates.
1070 def : Pat<(extloadi8i1 addr:$src), (MOV8rm addr:$src)>;
1071 def : Pat<(extloadi16i1 addr:$src), (MOVZX16rm8 addr:$src)>;
1072 def : Pat<(extloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1073 def : Pat<(extloadi16i8 addr:$src), (MOVZX16rm8 addr:$src)>;
1074 def : Pat<(extloadi32i8 addr:$src), (MOVZX32rm8 addr:$src)>;
1075 def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
1077 // For other extloads, use subregs, since the high contents of the register are
1078 // defined after an extload.
1079 def : Pat<(extloadi64i1 addr:$src),
1080 (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1081 def : Pat<(extloadi64i8 addr:$src),
1082 (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1083 def : Pat<(extloadi64i16 addr:$src),
1084 (SUBREG_TO_REG (i64 0), (MOVZX32rm16 addr:$src), sub_32bit)>;
1085 def : Pat<(extloadi64i32 addr:$src),
1086 (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src), sub_32bit)>;
1088 // anyext. Define these to do an explicit zero-extend to
1089 // avoid partial-register updates.
1090 def : Pat<(i16 (anyext GR8 :$src)), (EXTRACT_SUBREG
1091 (MOVZX32rr8 GR8 :$src), sub_16bit)>;
1092 def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8 GR8 :$src)>;
1094 // Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
1095 def : Pat<(i32 (anyext GR16:$src)),
1096 (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
1098 def : Pat<(i64 (anyext GR8 :$src)),
1099 (SUBREG_TO_REG (i64 0), (MOVZX32rr8 GR8 :$src), sub_32bit)>;
1100 def : Pat<(i64 (anyext GR16:$src)),
1101 (SUBREG_TO_REG (i64 0), (MOVZX32rr16 GR16 :$src), sub_32bit)>;
1102 def : Pat<(i64 (anyext GR32:$src)),
1103 (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1106 // Any instruction that defines a 32-bit result leaves the high half of the
1107 // register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
1108 // be copying from a truncate. And x86's cmov doesn't do anything if the
1109 // condition is false. But any other 32-bit operation will zero-extend
1111 def def32 : PatLeaf<(i32 GR32:$src), [{
1112 return N->getOpcode() != ISD::TRUNCATE &&
1113 N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
1114 N->getOpcode() != ISD::CopyFromReg &&
1115 N->getOpcode() != X86ISD::CMOV;
1118 // In the case of a 32-bit def that is known to implicitly zero-extend,
1119 // we can use a SUBREG_TO_REG.
1120 def : Pat<(i64 (zext def32:$src)),
1121 (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1123 //===----------------------------------------------------------------------===//
1124 // Pattern match OR as ADD
1125 //===----------------------------------------------------------------------===//
1127 // If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
1128 // 3-addressified into an LEA instruction to avoid copies. However, we also
1129 // want to finally emit these instructions as an or at the end of the code
1130 // generator to make the generated code easier to read. To do this, we select
1131 // into "disjoint bits" pseudo ops.
1133 // Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.
1134 def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
1135 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1136 return CurDAG->MaskedValueIsZero(N->getOperand(0), CN->getAPIntValue());
1138 APInt KnownZero0, KnownOne0;
1139 CurDAG->computeKnownBits(N->getOperand(0), KnownZero0, KnownOne0, 0);
1140 APInt KnownZero1, KnownOne1;
1141 CurDAG->computeKnownBits(N->getOperand(1), KnownZero1, KnownOne1, 0);
1142 return (~KnownZero0 & ~KnownZero1) == 0;
1146 // (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1147 // Try this before the selecting to OR.
1148 let AddedComplexity = 5, SchedRW = [WriteALU] in {
1150 let isConvertibleToThreeAddress = 1,
1151 Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1152 let isCommutable = 1 in {
1153 def ADD16rr_DB : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1154 "", // orw/addw REG, REG
1155 [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1156 def ADD32rr_DB : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1157 "", // orl/addl REG, REG
1158 [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1159 def ADD64rr_DB : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1160 "", // orq/addq REG, REG
1161 [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1164 // NOTE: These are order specific, we want the ri8 forms to be listed
1165 // first so that they are slightly preferred to the ri forms.
1167 def ADD16ri8_DB : I<0, Pseudo,
1168 (outs GR16:$dst), (ins GR16:$src1, i16i8imm:$src2),
1169 "", // orw/addw REG, imm8
1170 [(set GR16:$dst,(or_is_add GR16:$src1,i16immSExt8:$src2))]>;
1171 def ADD16ri_DB : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1172 "", // orw/addw REG, imm
1173 [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1175 def ADD32ri8_DB : I<0, Pseudo,
1176 (outs GR32:$dst), (ins GR32:$src1, i32i8imm:$src2),
1177 "", // orl/addl REG, imm8
1178 [(set GR32:$dst,(or_is_add GR32:$src1,i32immSExt8:$src2))]>;
1179 def ADD32ri_DB : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1180 "", // orl/addl REG, imm
1181 [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1184 def ADD64ri8_DB : I<0, Pseudo,
1185 (outs GR64:$dst), (ins GR64:$src1, i64i8imm:$src2),
1186 "", // orq/addq REG, imm8
1187 [(set GR64:$dst, (or_is_add GR64:$src1,
1188 i64immSExt8:$src2))]>;
1189 def ADD64ri32_DB : I<0, Pseudo,
1190 (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1191 "", // orq/addq REG, imm
1192 [(set GR64:$dst, (or_is_add GR64:$src1,
1193 i64immSExt32:$src2))]>;
1195 } // AddedComplexity, SchedRW
1198 //===----------------------------------------------------------------------===//
1200 //===----------------------------------------------------------------------===//
1202 // Odd encoding trick: -128 fits into an 8-bit immediate field while
1203 // +128 doesn't, so in this special case use a sub instead of an add.
1204 def : Pat<(add GR16:$src1, 128),
1205 (SUB16ri8 GR16:$src1, -128)>;
1206 def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1207 (SUB16mi8 addr:$dst, -128)>;
1209 def : Pat<(add GR32:$src1, 128),
1210 (SUB32ri8 GR32:$src1, -128)>;
1211 def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1212 (SUB32mi8 addr:$dst, -128)>;
1214 def : Pat<(add GR64:$src1, 128),
1215 (SUB64ri8 GR64:$src1, -128)>;
1216 def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1217 (SUB64mi8 addr:$dst, -128)>;
1219 // The same trick applies for 32-bit immediate fields in 64-bit
1221 def : Pat<(add GR64:$src1, 0x0000000080000000),
1222 (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1223 def : Pat<(store (add (loadi64 addr:$dst), 0x00000000800000000), addr:$dst),
1224 (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1226 // To avoid needing to materialize an immediate in a register, use a 32-bit and
1227 // with implicit zero-extension instead of a 64-bit and if the immediate has at
1228 // least 32 bits of leading zeros. If in addition the last 32 bits can be
1229 // represented with a sign extension of a 8 bit constant, use that.
1231 def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
1235 (EXTRACT_SUBREG GR64:$src, sub_32bit),
1236 (i32 (GetLo8XForm imm:$imm))),
1239 def : Pat<(and GR64:$src, i64immZExt32:$imm),
1243 (EXTRACT_SUBREG GR64:$src, sub_32bit),
1244 (i32 (GetLo32XForm imm:$imm))),
1248 // r & (2^16-1) ==> movz
1249 def : Pat<(and GR32:$src1, 0xffff),
1250 (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1251 // r & (2^8-1) ==> movz
1252 def : Pat<(and GR32:$src1, 0xff),
1253 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src1,
1256 Requires<[Not64BitMode]>;
1257 // r & (2^8-1) ==> movz
1258 def : Pat<(and GR16:$src1, 0xff),
1259 (EXTRACT_SUBREG (MOVZX32rr8 (EXTRACT_SUBREG
1260 (i16 (COPY_TO_REGCLASS GR16:$src1, GR16_ABCD)), sub_8bit)),
1262 Requires<[Not64BitMode]>;
1264 // r & (2^32-1) ==> movz
1265 def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1266 (SUBREG_TO_REG (i64 0),
1267 (MOV32rr (EXTRACT_SUBREG GR64:$src, sub_32bit)),
1269 // r & (2^16-1) ==> movz
1270 def : Pat<(and GR64:$src, 0xffff),
1271 (SUBREG_TO_REG (i64 0),
1272 (MOVZX32rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit))),
1274 // r & (2^8-1) ==> movz
1275 def : Pat<(and GR64:$src, 0xff),
1276 (SUBREG_TO_REG (i64 0),
1277 (MOVZX32rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit))),
1279 // r & (2^8-1) ==> movz
1280 def : Pat<(and GR32:$src1, 0xff),
1281 (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>,
1282 Requires<[In64BitMode]>;
1283 // r & (2^8-1) ==> movz
1284 def : Pat<(and GR16:$src1, 0xff),
1285 (EXTRACT_SUBREG (MOVZX32rr8 (i8
1286 (EXTRACT_SUBREG GR16:$src1, sub_8bit))), sub_16bit)>,
1287 Requires<[In64BitMode]>;
1290 // sext_inreg patterns
1291 def : Pat<(sext_inreg GR32:$src, i16),
1292 (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1293 def : Pat<(sext_inreg GR32:$src, i8),
1294 (MOVSX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1297 Requires<[Not64BitMode]>;
1299 def : Pat<(sext_inreg GR16:$src, i8),
1300 (EXTRACT_SUBREG (i32 (MOVSX32rr8 (EXTRACT_SUBREG
1301 (i32 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)), sub_8bit))),
1303 Requires<[Not64BitMode]>;
1305 def : Pat<(sext_inreg GR64:$src, i32),
1306 (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1307 def : Pat<(sext_inreg GR64:$src, i16),
1308 (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1309 def : Pat<(sext_inreg GR64:$src, i8),
1310 (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1311 def : Pat<(sext_inreg GR32:$src, i8),
1312 (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>,
1313 Requires<[In64BitMode]>;
1314 def : Pat<(sext_inreg GR16:$src, i8),
1315 (EXTRACT_SUBREG (MOVSX32rr8
1316 (EXTRACT_SUBREG GR16:$src, sub_8bit)), sub_16bit)>,
1317 Requires<[In64BitMode]>;
1319 // sext, sext_load, zext, zext_load
1320 def: Pat<(i16 (sext GR8:$src)),
1321 (EXTRACT_SUBREG (MOVSX32rr8 GR8:$src), sub_16bit)>;
1322 def: Pat<(sextloadi16i8 addr:$src),
1323 (EXTRACT_SUBREG (MOVSX32rm8 addr:$src), sub_16bit)>;
1324 def: Pat<(i16 (zext GR8:$src)),
1325 (EXTRACT_SUBREG (MOVZX32rr8 GR8:$src), sub_16bit)>;
1326 def: Pat<(zextloadi16i8 addr:$src),
1327 (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1330 def : Pat<(i16 (trunc GR32:$src)),
1331 (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1332 def : Pat<(i8 (trunc GR32:$src)),
1333 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1335 Requires<[Not64BitMode]>;
1336 def : Pat<(i8 (trunc GR16:$src)),
1337 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1339 Requires<[Not64BitMode]>;
1340 def : Pat<(i32 (trunc GR64:$src)),
1341 (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1342 def : Pat<(i16 (trunc GR64:$src)),
1343 (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1344 def : Pat<(i8 (trunc GR64:$src)),
1345 (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1346 def : Pat<(i8 (trunc GR32:$src)),
1347 (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1348 Requires<[In64BitMode]>;
1349 def : Pat<(i8 (trunc GR16:$src)),
1350 (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1351 Requires<[In64BitMode]>;
1353 // h-register tricks
1354 def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1355 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1357 Requires<[Not64BitMode]>;
1358 def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1359 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1361 Requires<[Not64BitMode]>;
1362 def : Pat<(srl GR16:$src, (i8 8)),
1365 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1368 Requires<[Not64BitMode]>;
1369 def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1370 (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1373 Requires<[Not64BitMode]>;
1374 def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1375 (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1378 Requires<[Not64BitMode]>;
1379 def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1380 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1383 Requires<[Not64BitMode]>;
1384 def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1385 (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1388 Requires<[Not64BitMode]>;
1390 // h-register tricks.
1391 // For now, be conservative on x86-64 and use an h-register extract only if the
1392 // value is immediately zero-extended or stored, which are somewhat common
1393 // cases. This uses a bunch of code to prevent a register requiring a REX prefix
1394 // from being allocated in the same instruction as the h register, as there's
1395 // currently no way to describe this requirement to the register allocator.
1397 // h-register extract and zero-extend.
1398 def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1402 (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1405 def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1407 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1409 Requires<[In64BitMode]>;
1410 def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1411 (MOVZX32_NOREXrr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1414 Requires<[In64BitMode]>;
1415 def : Pat<(srl GR16:$src, (i8 8)),
1418 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1421 Requires<[In64BitMode]>;
1422 def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1424 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1426 Requires<[In64BitMode]>;
1427 def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1429 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1431 Requires<[In64BitMode]>;
1432 def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1436 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1439 def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1443 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1447 // h-register extract and store.
1448 def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1451 (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1453 def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1456 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1458 Requires<[In64BitMode]>;
1459 def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1462 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1464 Requires<[In64BitMode]>;
1467 // (shl x, 1) ==> (add x, x)
1468 // Note that if x is undef (immediate or otherwise), we could theoretically
1469 // end up with the two uses of x getting different values, producing a result
1470 // where the least significant bit is not 0. However, the probability of this
1471 // happening is considered low enough that this is officially not a
1473 def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr GR8 :$src1, GR8 :$src1)>;
1474 def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1475 def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1476 def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1478 // Helper imms that check if a mask doesn't change significant shift bits.
1479 def immShift32 : ImmLeaf<i8, [{ return CountTrailingOnes_32(Imm) >= 5; }]>;
1480 def immShift64 : ImmLeaf<i8, [{ return CountTrailingOnes_32(Imm) >= 6; }]>;
1482 // Shift amount is implicitly masked.
1483 multiclass MaskedShiftAmountPats<SDNode frag, string name> {
1484 // (shift x (and y, 31)) ==> (shift x, y)
1485 def : Pat<(frag GR8:$src1, (and CL, immShift32)),
1486 (!cast<Instruction>(name # "8rCL") GR8:$src1)>;
1487 def : Pat<(frag GR16:$src1, (and CL, immShift32)),
1488 (!cast<Instruction>(name # "16rCL") GR16:$src1)>;
1489 def : Pat<(frag GR32:$src1, (and CL, immShift32)),
1490 (!cast<Instruction>(name # "32rCL") GR32:$src1)>;
1491 def : Pat<(store (frag (loadi8 addr:$dst), (and CL, immShift32)), addr:$dst),
1492 (!cast<Instruction>(name # "8mCL") addr:$dst)>;
1493 def : Pat<(store (frag (loadi16 addr:$dst), (and CL, immShift32)), addr:$dst),
1494 (!cast<Instruction>(name # "16mCL") addr:$dst)>;
1495 def : Pat<(store (frag (loadi32 addr:$dst), (and CL, immShift32)), addr:$dst),
1496 (!cast<Instruction>(name # "32mCL") addr:$dst)>;
1498 // (shift x (and y, 63)) ==> (shift x, y)
1499 def : Pat<(frag GR64:$src1, (and CL, immShift64)),
1500 (!cast<Instruction>(name # "64rCL") GR64:$src1)>;
1501 def : Pat<(store (frag (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1502 (!cast<Instruction>(name # "64mCL") addr:$dst)>;
1505 defm : MaskedShiftAmountPats<shl, "SHL">;
1506 defm : MaskedShiftAmountPats<srl, "SHR">;
1507 defm : MaskedShiftAmountPats<sra, "SAR">;
1508 defm : MaskedShiftAmountPats<rotl, "ROL">;
1509 defm : MaskedShiftAmountPats<rotr, "ROR">;
1511 // (anyext (setcc_carry)) -> (setcc_carry)
1512 def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1514 def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1516 def : Pat<(i32 (anyext (i16 (X86setcc_c X86_COND_B, EFLAGS)))),
1522 //===----------------------------------------------------------------------===//
1523 // EFLAGS-defining Patterns
1524 //===----------------------------------------------------------------------===//
1527 def : Pat<(add GR8 :$src1, GR8 :$src2), (ADD8rr GR8 :$src1, GR8 :$src2)>;
1528 def : Pat<(add GR16:$src1, GR16:$src2), (ADD16rr GR16:$src1, GR16:$src2)>;
1529 def : Pat<(add GR32:$src1, GR32:$src2), (ADD32rr GR32:$src1, GR32:$src2)>;
1532 def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
1533 (ADD8rm GR8:$src1, addr:$src2)>;
1534 def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
1535 (ADD16rm GR16:$src1, addr:$src2)>;
1536 def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
1537 (ADD32rm GR32:$src1, addr:$src2)>;
1540 def : Pat<(add GR8 :$src1, imm:$src2), (ADD8ri GR8:$src1 , imm:$src2)>;
1541 def : Pat<(add GR16:$src1, imm:$src2), (ADD16ri GR16:$src1, imm:$src2)>;
1542 def : Pat<(add GR32:$src1, imm:$src2), (ADD32ri GR32:$src1, imm:$src2)>;
1543 def : Pat<(add GR16:$src1, i16immSExt8:$src2),
1544 (ADD16ri8 GR16:$src1, i16immSExt8:$src2)>;
1545 def : Pat<(add GR32:$src1, i32immSExt8:$src2),
1546 (ADD32ri8 GR32:$src1, i32immSExt8:$src2)>;
1549 def : Pat<(sub GR8 :$src1, GR8 :$src2), (SUB8rr GR8 :$src1, GR8 :$src2)>;
1550 def : Pat<(sub GR16:$src1, GR16:$src2), (SUB16rr GR16:$src1, GR16:$src2)>;
1551 def : Pat<(sub GR32:$src1, GR32:$src2), (SUB32rr GR32:$src1, GR32:$src2)>;
1554 def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
1555 (SUB8rm GR8:$src1, addr:$src2)>;
1556 def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
1557 (SUB16rm GR16:$src1, addr:$src2)>;
1558 def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
1559 (SUB32rm GR32:$src1, addr:$src2)>;
1562 def : Pat<(sub GR8:$src1, imm:$src2),
1563 (SUB8ri GR8:$src1, imm:$src2)>;
1564 def : Pat<(sub GR16:$src1, imm:$src2),
1565 (SUB16ri GR16:$src1, imm:$src2)>;
1566 def : Pat<(sub GR32:$src1, imm:$src2),
1567 (SUB32ri GR32:$src1, imm:$src2)>;
1568 def : Pat<(sub GR16:$src1, i16immSExt8:$src2),
1569 (SUB16ri8 GR16:$src1, i16immSExt8:$src2)>;
1570 def : Pat<(sub GR32:$src1, i32immSExt8:$src2),
1571 (SUB32ri8 GR32:$src1, i32immSExt8:$src2)>;
1574 def : Pat<(X86sub_flag 0, GR8 :$src), (NEG8r GR8 :$src)>;
1575 def : Pat<(X86sub_flag 0, GR16:$src), (NEG16r GR16:$src)>;
1576 def : Pat<(X86sub_flag 0, GR32:$src), (NEG32r GR32:$src)>;
1577 def : Pat<(X86sub_flag 0, GR64:$src), (NEG64r GR64:$src)>;
1580 def : Pat<(mul GR16:$src1, GR16:$src2),
1581 (IMUL16rr GR16:$src1, GR16:$src2)>;
1582 def : Pat<(mul GR32:$src1, GR32:$src2),
1583 (IMUL32rr GR32:$src1, GR32:$src2)>;
1586 def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
1587 (IMUL16rm GR16:$src1, addr:$src2)>;
1588 def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
1589 (IMUL32rm GR32:$src1, addr:$src2)>;
1592 def : Pat<(mul GR16:$src1, imm:$src2),
1593 (IMUL16rri GR16:$src1, imm:$src2)>;
1594 def : Pat<(mul GR32:$src1, imm:$src2),
1595 (IMUL32rri GR32:$src1, imm:$src2)>;
1596 def : Pat<(mul GR16:$src1, i16immSExt8:$src2),
1597 (IMUL16rri8 GR16:$src1, i16immSExt8:$src2)>;
1598 def : Pat<(mul GR32:$src1, i32immSExt8:$src2),
1599 (IMUL32rri8 GR32:$src1, i32immSExt8:$src2)>;
1601 // reg = mul mem, imm
1602 def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
1603 (IMUL16rmi addr:$src1, imm:$src2)>;
1604 def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
1605 (IMUL32rmi addr:$src1, imm:$src2)>;
1606 def : Pat<(mul (loadi16 addr:$src1), i16immSExt8:$src2),
1607 (IMUL16rmi8 addr:$src1, i16immSExt8:$src2)>;
1608 def : Pat<(mul (loadi32 addr:$src1), i32immSExt8:$src2),
1609 (IMUL32rmi8 addr:$src1, i32immSExt8:$src2)>;
1611 // Patterns for nodes that do not produce flags, for instructions that do.
1614 def : Pat<(add GR64:$src1, GR64:$src2),
1615 (ADD64rr GR64:$src1, GR64:$src2)>;
1616 def : Pat<(add GR64:$src1, i64immSExt8:$src2),
1617 (ADD64ri8 GR64:$src1, i64immSExt8:$src2)>;
1618 def : Pat<(add GR64:$src1, i64immSExt32:$src2),
1619 (ADD64ri32 GR64:$src1, i64immSExt32:$src2)>;
1620 def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
1621 (ADD64rm GR64:$src1, addr:$src2)>;
1624 def : Pat<(sub GR64:$src1, GR64:$src2),
1625 (SUB64rr GR64:$src1, GR64:$src2)>;
1626 def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
1627 (SUB64rm GR64:$src1, addr:$src2)>;
1628 def : Pat<(sub GR64:$src1, i64immSExt8:$src2),
1629 (SUB64ri8 GR64:$src1, i64immSExt8:$src2)>;
1630 def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
1631 (SUB64ri32 GR64:$src1, i64immSExt32:$src2)>;
1634 def : Pat<(mul GR64:$src1, GR64:$src2),
1635 (IMUL64rr GR64:$src1, GR64:$src2)>;
1636 def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
1637 (IMUL64rm GR64:$src1, addr:$src2)>;
1638 def : Pat<(mul GR64:$src1, i64immSExt8:$src2),
1639 (IMUL64rri8 GR64:$src1, i64immSExt8:$src2)>;
1640 def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
1641 (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
1642 def : Pat<(mul (loadi64 addr:$src1), i64immSExt8:$src2),
1643 (IMUL64rmi8 addr:$src1, i64immSExt8:$src2)>;
1644 def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
1645 (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
1648 // Do not make INC if it is slow
1649 def : Pat<(add GR8:$src, 1),
1650 (INC8r GR8:$src)>, Requires<[NotSlowIncDec]>;
1651 def : Pat<(add GR16:$src, 1),
1652 (INC16r GR16:$src)>, Requires<[NotSlowIncDec, Not64BitMode]>;
1653 def : Pat<(add GR16:$src, 1),
1654 (INC64_16r GR16:$src)>, Requires<[NotSlowIncDec, In64BitMode]>;
1655 def : Pat<(add GR32:$src, 1),
1656 (INC32r GR32:$src)>, Requires<[NotSlowIncDec, Not64BitMode]>;
1657 def : Pat<(add GR32:$src, 1),
1658 (INC64_32r GR32:$src)>, Requires<[NotSlowIncDec, In64BitMode]>;
1659 def : Pat<(add GR64:$src, 1),
1660 (INC64r GR64:$src)>, Requires<[NotSlowIncDec]>;
1663 // Do not make DEC if it is slow
1664 def : Pat<(add GR8:$src, -1),
1665 (DEC8r GR8:$src)>, Requires<[NotSlowIncDec]>;
1666 def : Pat<(add GR16:$src, -1),
1667 (DEC16r GR16:$src)>, Requires<[NotSlowIncDec, Not64BitMode]>;
1668 def : Pat<(add GR16:$src, -1),
1669 (DEC64_16r GR16:$src)>, Requires<[NotSlowIncDec, In64BitMode]>;
1670 def : Pat<(add GR32:$src, -1),
1671 (DEC32r GR32:$src)>, Requires<[NotSlowIncDec, Not64BitMode]>;
1672 def : Pat<(add GR32:$src, -1),
1673 (DEC64_32r GR32:$src)>, Requires<[NotSlowIncDec, In64BitMode]>;
1674 def : Pat<(add GR64:$src, -1),
1675 (DEC64r GR64:$src)>, Requires<[NotSlowIncDec]>;
1678 def : Pat<(or GR8 :$src1, GR8 :$src2), (OR8rr GR8 :$src1, GR8 :$src2)>;
1679 def : Pat<(or GR16:$src1, GR16:$src2), (OR16rr GR16:$src1, GR16:$src2)>;
1680 def : Pat<(or GR32:$src1, GR32:$src2), (OR32rr GR32:$src1, GR32:$src2)>;
1681 def : Pat<(or GR64:$src1, GR64:$src2), (OR64rr GR64:$src1, GR64:$src2)>;
1684 def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
1685 (OR8rm GR8:$src1, addr:$src2)>;
1686 def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
1687 (OR16rm GR16:$src1, addr:$src2)>;
1688 def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
1689 (OR32rm GR32:$src1, addr:$src2)>;
1690 def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
1691 (OR64rm GR64:$src1, addr:$src2)>;
1694 def : Pat<(or GR8:$src1 , imm:$src2), (OR8ri GR8 :$src1, imm:$src2)>;
1695 def : Pat<(or GR16:$src1, imm:$src2), (OR16ri GR16:$src1, imm:$src2)>;
1696 def : Pat<(or GR32:$src1, imm:$src2), (OR32ri GR32:$src1, imm:$src2)>;
1697 def : Pat<(or GR16:$src1, i16immSExt8:$src2),
1698 (OR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1699 def : Pat<(or GR32:$src1, i32immSExt8:$src2),
1700 (OR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1701 def : Pat<(or GR64:$src1, i64immSExt8:$src2),
1702 (OR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1703 def : Pat<(or GR64:$src1, i64immSExt32:$src2),
1704 (OR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1707 def : Pat<(xor GR8 :$src1, GR8 :$src2), (XOR8rr GR8 :$src1, GR8 :$src2)>;
1708 def : Pat<(xor GR16:$src1, GR16:$src2), (XOR16rr GR16:$src1, GR16:$src2)>;
1709 def : Pat<(xor GR32:$src1, GR32:$src2), (XOR32rr GR32:$src1, GR32:$src2)>;
1710 def : Pat<(xor GR64:$src1, GR64:$src2), (XOR64rr GR64:$src1, GR64:$src2)>;
1713 def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
1714 (XOR8rm GR8:$src1, addr:$src2)>;
1715 def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
1716 (XOR16rm GR16:$src1, addr:$src2)>;
1717 def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
1718 (XOR32rm GR32:$src1, addr:$src2)>;
1719 def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
1720 (XOR64rm GR64:$src1, addr:$src2)>;
1723 def : Pat<(xor GR8:$src1, imm:$src2),
1724 (XOR8ri GR8:$src1, imm:$src2)>;
1725 def : Pat<(xor GR16:$src1, imm:$src2),
1726 (XOR16ri GR16:$src1, imm:$src2)>;
1727 def : Pat<(xor GR32:$src1, imm:$src2),
1728 (XOR32ri GR32:$src1, imm:$src2)>;
1729 def : Pat<(xor GR16:$src1, i16immSExt8:$src2),
1730 (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1731 def : Pat<(xor GR32:$src1, i32immSExt8:$src2),
1732 (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1733 def : Pat<(xor GR64:$src1, i64immSExt8:$src2),
1734 (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1735 def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
1736 (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1739 def : Pat<(and GR8 :$src1, GR8 :$src2), (AND8rr GR8 :$src1, GR8 :$src2)>;
1740 def : Pat<(and GR16:$src1, GR16:$src2), (AND16rr GR16:$src1, GR16:$src2)>;
1741 def : Pat<(and GR32:$src1, GR32:$src2), (AND32rr GR32:$src1, GR32:$src2)>;
1742 def : Pat<(and GR64:$src1, GR64:$src2), (AND64rr GR64:$src1, GR64:$src2)>;
1745 def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
1746 (AND8rm GR8:$src1, addr:$src2)>;
1747 def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
1748 (AND16rm GR16:$src1, addr:$src2)>;
1749 def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
1750 (AND32rm GR32:$src1, addr:$src2)>;
1751 def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
1752 (AND64rm GR64:$src1, addr:$src2)>;
1755 def : Pat<(and GR8:$src1, imm:$src2),
1756 (AND8ri GR8:$src1, imm:$src2)>;
1757 def : Pat<(and GR16:$src1, imm:$src2),
1758 (AND16ri GR16:$src1, imm:$src2)>;
1759 def : Pat<(and GR32:$src1, imm:$src2),
1760 (AND32ri GR32:$src1, imm:$src2)>;
1761 def : Pat<(and GR16:$src1, i16immSExt8:$src2),
1762 (AND16ri8 GR16:$src1, i16immSExt8:$src2)>;
1763 def : Pat<(and GR32:$src1, i32immSExt8:$src2),
1764 (AND32ri8 GR32:$src1, i32immSExt8:$src2)>;
1765 def : Pat<(and GR64:$src1, i64immSExt8:$src2),
1766 (AND64ri8 GR64:$src1, i64immSExt8:$src2)>;
1767 def : Pat<(and GR64:$src1, i64immSExt32:$src2),
1768 (AND64ri32 GR64:$src1, i64immSExt32:$src2)>;
1770 // Bit scan instruction patterns to match explicit zero-undef behavior.
1771 def : Pat<(cttz_zero_undef GR16:$src), (BSF16rr GR16:$src)>;
1772 def : Pat<(cttz_zero_undef GR32:$src), (BSF32rr GR32:$src)>;
1773 def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr GR64:$src)>;
1774 def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm addr:$src)>;
1775 def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm addr:$src)>;
1776 def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm addr:$src)>;
1778 // When HasMOVBE is enabled it is possible to get a non-legalized
1779 // register-register 16 bit bswap. This maps it to a ROL instruction.
1780 let Predicates = [HasMOVBE] in {
1781 def : Pat<(bswap GR16:$src), (ROL16ri GR16:$src, (i8 8))>;