xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86InstrCompiler.td (revision 96190b4fef3b4a0cc3ca0606b0c4e3e69a5e6717)
1//===- X86InstrCompiler.td - Compiler Pseudos and Patterns -*- tablegen -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file describes the various pseudo instructions used by the compiler,
10// as well as Pat patterns used during instruction selection.
11//
12//===----------------------------------------------------------------------===//
13
14//===----------------------------------------------------------------------===//
15// Pattern Matching Support
16
17def GetLo32XForm : SDNodeXForm<imm, [{
18  // Transformation function: get the low 32 bits.
19  return getI32Imm((uint32_t)N->getZExtValue(), SDLoc(N));
20}]>;
21
22
23//===----------------------------------------------------------------------===//
24// Random Pseudo Instructions.
25
26// PIC base construction.  This expands to code that looks like this:
27//     call  $next_inst
28//     popl %destreg"
29let hasSideEffects = 0, isNotDuplicable = 1, Uses = [ESP, SSP],
30    SchedRW = [WriteJump] in
31  def MOVPC32r : Ii32<0xE8, Pseudo, (outs GR32:$reg), (ins i32imm:$label),
32                      "", []>;
33
34// ADJCALLSTACKDOWN/UP implicitly use/def ESP because they may be expanded into
35// a stack adjustment and the codegen must know that they may modify the stack
36// pointer before prolog-epilog rewriting occurs.
37// Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
38// sub / add which can clobber EFLAGS.
39let Defs = [ESP, EFLAGS, SSP], Uses = [ESP, SSP], SchedRW = [WriteALU] in {
40def ADJCALLSTACKDOWN32 : I<0, Pseudo, (outs),
41                           (ins i32imm:$amt1, i32imm:$amt2, i32imm:$amt3),
42                           "#ADJCALLSTACKDOWN", []>, Requires<[NotLP64]>;
43def ADJCALLSTACKUP32   : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
44                           "#ADJCALLSTACKUP",
45                           [(X86callseq_end timm:$amt1, timm:$amt2)]>,
46                           Requires<[NotLP64]>;
47}
48def : Pat<(X86callseq_start timm:$amt1, timm:$amt2),
49       (ADJCALLSTACKDOWN32 i32imm:$amt1, i32imm:$amt2, 0)>, Requires<[NotLP64]>;
50
51
52// ADJCALLSTACKDOWN/UP implicitly use/def RSP because they may be expanded into
53// a stack adjustment and the codegen must know that they may modify the stack
54// pointer before prolog-epilog rewriting occurs.
55// Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
56// sub / add which can clobber EFLAGS.
57let Defs = [RSP, EFLAGS, SSP], Uses = [RSP, SSP], SchedRW = [WriteALU] in {
58def ADJCALLSTACKDOWN64 : I<0, Pseudo, (outs),
59                           (ins i32imm:$amt1, i32imm:$amt2, i32imm:$amt3),
60                           "#ADJCALLSTACKDOWN", []>, Requires<[IsLP64]>;
61def ADJCALLSTACKUP64   : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
62                           "#ADJCALLSTACKUP",
63                           [(X86callseq_end timm:$amt1, timm:$amt2)]>,
64                           Requires<[IsLP64]>;
65}
66def : Pat<(X86callseq_start timm:$amt1, timm:$amt2),
67        (ADJCALLSTACKDOWN64 i32imm:$amt1, i32imm:$amt2, 0)>, Requires<[IsLP64]>;
68
69let SchedRW = [WriteSystem] in {
70
71// x86-64 va_start lowering magic.
72let hasSideEffects = 1, mayStore = 1, Defs = [EFLAGS] in {
73def VASTART_SAVE_XMM_REGS : I<0, Pseudo,
74                              (outs),
75                              (ins GR8:$al, i8mem:$regsavefi, variable_ops),
76                              "#VASTART_SAVE_XMM_REGS $al, $regsavefi",
77                              [(X86vastart_save_xmm_regs GR8:$al, addr:$regsavefi),
78                               (implicit EFLAGS)]>;
79}
80
81let usesCustomInserter = 1, Defs = [EFLAGS] in {
82// The VAARG_64 and VAARG_X32 pseudo-instructions take the address of the
83// va_list, and place the address of the next argument into a register.
84let Defs = [EFLAGS] in {
85def VAARG_64 : I<0, Pseudo,
86                 (outs GR64:$dst),
87                 (ins i8mem:$ap, i32imm:$size, i8imm:$mode, i32imm:$align),
88                 "#VAARG_64 $dst, $ap, $size, $mode, $align",
89                 [(set GR64:$dst,
90                    (X86vaarg64 addr:$ap, timm:$size, timm:$mode, timm:$align)),
91                  (implicit EFLAGS)]>, Requires<[In64BitMode, IsLP64]>;
92def VAARG_X32 : I<0, Pseudo,
93                 (outs GR32:$dst),
94                 (ins i8mem:$ap, i32imm:$size, i8imm:$mode, i32imm:$align),
95                 "#VAARG_X32 $dst, $ap, $size, $mode, $align",
96                 [(set GR32:$dst,
97                    (X86vaargx32 addr:$ap, timm:$size, timm:$mode, timm:$align)),
98                  (implicit EFLAGS)]>, Requires<[In64BitMode, NotLP64]>;
99}
100
101// When using segmented stacks these are lowered into instructions which first
102// check if the current stacklet has enough free memory. If it does, memory is
103// allocated by bumping the stack pointer. Otherwise memory is allocated from
104// the heap.
105
106let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
107def SEG_ALLOCA_32 : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$size),
108                      "# variable sized alloca for segmented stacks",
109                      [(set GR32:$dst,
110                         (X86SegAlloca GR32:$size))]>,
111                    Requires<[NotLP64]>;
112
113let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
114def SEG_ALLOCA_64 : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$size),
115                      "# variable sized alloca for segmented stacks",
116                      [(set GR64:$dst,
117                         (X86SegAlloca GR64:$size))]>,
118                    Requires<[In64BitMode]>;
119
120// To protect against stack clash, dynamic allocation should perform a memory
121// probe at each page.
122
123let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
124def PROBED_ALLOCA_32 : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$size),
125                      "# variable sized alloca with probing",
126                      [(set GR32:$dst,
127                         (X86ProbedAlloca GR32:$size))]>,
128                    Requires<[NotLP64]>;
129
130let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
131def PROBED_ALLOCA_64 : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$size),
132                      "# variable sized alloca with probing",
133                      [(set GR64:$dst,
134                         (X86ProbedAlloca GR64:$size))]>,
135                    Requires<[In64BitMode]>;
136}
137
138let hasNoSchedulingInfo = 1 in
139def STACKALLOC_W_PROBING : I<0, Pseudo, (outs), (ins i64imm:$stacksize),
140                             "# fixed size alloca with probing",
141                             []>;
142
143// Dynamic stack allocation yields a _chkstk or _alloca call for all Windows
144// targets.  These calls are needed to probe the stack when allocating more than
145// 4k bytes in one go. Touching the stack at 4K increments is necessary to
146// ensure that the guard pages used by the OS virtual memory manager are
147// allocated in correct sequence.
148// The main point of having separate instruction are extra unmodelled effects
149// (compared to ordinary calls) like stack pointer change.
150
151let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
152def DYN_ALLOCA_32 : I<0, Pseudo, (outs), (ins GR32:$size),
153                     "# dynamic stack allocation",
154                     [(X86DynAlloca GR32:$size)]>,
155                     Requires<[NotLP64]>;
156
157let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
158def DYN_ALLOCA_64 : I<0, Pseudo, (outs), (ins GR64:$size),
159                     "# dynamic stack allocation",
160                     [(X86DynAlloca GR64:$size)]>,
161                     Requires<[In64BitMode]>;
162} // SchedRW
163
164// These instructions XOR the frame pointer into a GPR. They are used in some
165// stack protection schemes. These are post-RA pseudos because we only know the
166// frame register after register allocation.
167let Constraints = "$src = $dst", isMoveImm = 1, isPseudo = 1, Defs = [EFLAGS] in {
168  def XOR32_FP : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src),
169                  "xorl\t$$FP, $src", []>,
170                  Requires<[NotLP64]>, Sched<[WriteALU]>;
171  def XOR64_FP : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src),
172                  "xorq\t$$FP $src", []>,
173                  Requires<[In64BitMode]>, Sched<[WriteALU]>;
174}
175
176//===----------------------------------------------------------------------===//
177// EH Pseudo Instructions
178//
179let SchedRW = [WriteSystem] in {
180let isTerminator = 1, isReturn = 1, isBarrier = 1,
181    hasCtrlDep = 1, isCodeGenOnly = 1 in {
182def EH_RETURN   : I<0xC3, RawFrm, (outs), (ins GR32:$addr),
183                    "ret\t#eh_return, addr: $addr",
184                    [(X86ehret GR32:$addr)]>, Sched<[WriteJumpLd]>;
185
186}
187
188let isTerminator = 1, isReturn = 1, isBarrier = 1,
189    hasCtrlDep = 1, isCodeGenOnly = 1 in {
190def EH_RETURN64   : I<0xC3, RawFrm, (outs), (ins GR64:$addr),
191                     "ret\t#eh_return, addr: $addr",
192                     [(X86ehret GR64:$addr)]>, Sched<[WriteJumpLd]>;
193
194}
195
196let isTerminator = 1, hasSideEffects = 1, isBarrier = 1, hasCtrlDep = 1,
197    isCodeGenOnly = 1, isReturn = 1, isEHScopeReturn = 1 in {
198  def CLEANUPRET : I<0, Pseudo, (outs), (ins), "# CLEANUPRET", [(cleanupret)]>;
199
200  // CATCHRET needs a custom inserter for SEH.
201  let usesCustomInserter = 1 in
202    def CATCHRET : I<0, Pseudo, (outs), (ins brtarget32:$dst, brtarget32:$from),
203                     "# CATCHRET",
204                     [(catchret bb:$dst, bb:$from)]>;
205}
206
207let hasSideEffects = 1, isBarrier = 1, isCodeGenOnly = 1,
208    usesCustomInserter = 1 in {
209  def EH_SjLj_SetJmp32  : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$buf),
210                            "#EH_SJLJ_SETJMP32",
211                            [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
212                          Requires<[Not64BitMode]>;
213  def EH_SjLj_SetJmp64  : I<0, Pseudo, (outs GR32:$dst), (ins i64mem:$buf),
214                            "#EH_SJLJ_SETJMP64",
215                            [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
216                          Requires<[In64BitMode]>;
217  let isTerminator = 1 in {
218  def EH_SjLj_LongJmp32 : I<0, Pseudo, (outs), (ins i32mem:$buf),
219                            "#EH_SJLJ_LONGJMP32",
220                            [(X86eh_sjlj_longjmp addr:$buf)]>,
221                          Requires<[Not64BitMode]>;
222  def EH_SjLj_LongJmp64 : I<0, Pseudo, (outs), (ins i64mem:$buf),
223                            "#EH_SJLJ_LONGJMP64",
224                            [(X86eh_sjlj_longjmp addr:$buf)]>,
225                          Requires<[In64BitMode]>;
226  }
227}
228
229let isBranch = 1, isTerminator = 1, isCodeGenOnly = 1 in {
230  def EH_SjLj_Setup : I<0, Pseudo, (outs), (ins brtarget:$dst),
231                        "#EH_SjLj_Setup\t$dst", []>;
232}
233} // SchedRW
234
235//===----------------------------------------------------------------------===//
236// Pseudo instructions used by unwind info.
237//
238let isPseudo = 1, SchedRW = [WriteSystem] in {
239  def SEH_PushReg : I<0, Pseudo, (outs), (ins i32imm:$reg),
240                            "#SEH_PushReg $reg", []>;
241  def SEH_SaveReg : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
242                            "#SEH_SaveReg $reg, $dst", []>;
243  def SEH_SaveXMM : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
244                            "#SEH_SaveXMM $reg, $dst", []>;
245  def SEH_StackAlloc : I<0, Pseudo, (outs), (ins i32imm:$size),
246                            "#SEH_StackAlloc $size", []>;
247  def SEH_StackAlign : I<0, Pseudo, (outs), (ins i32imm:$align),
248                            "#SEH_StackAlign $align", []>;
249  def SEH_SetFrame : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$offset),
250                            "#SEH_SetFrame $reg, $offset", []>;
251  def SEH_PushFrame : I<0, Pseudo, (outs), (ins i1imm:$mode),
252                            "#SEH_PushFrame $mode", []>;
253  def SEH_EndPrologue : I<0, Pseudo, (outs), (ins),
254                            "#SEH_EndPrologue", []>;
255  def SEH_Epilogue : I<0, Pseudo, (outs), (ins),
256                            "#SEH_Epilogue", []>;
257}
258
259//===----------------------------------------------------------------------===//
260// Pseudo instructions used by KCFI.
261//===----------------------------------------------------------------------===//
262let
263  Defs = [R10, R11, EFLAGS] in {
264def KCFI_CHECK : PseudoI<
265  (outs), (ins GR64:$ptr, i32imm:$type), []>, Sched<[]>;
266}
267
268//===----------------------------------------------------------------------===//
269// Pseudo instructions used by address sanitizer.
270//===----------------------------------------------------------------------===//
271let
272  Defs = [R10, R11, EFLAGS] in {
273def ASAN_CHECK_MEMACCESS : PseudoI<
274  (outs), (ins GR64PLTSafe:$addr, i32imm:$accessinfo),
275  [(int_asan_check_memaccess GR64PLTSafe:$addr, (i32 timm:$accessinfo))]>,
276  Sched<[]>;
277}
278
279//===----------------------------------------------------------------------===//
280// Pseudo instructions used by segmented stacks.
281//
282
283// This is lowered into a RET instruction by MCInstLower.  We need
284// this so that we don't have to have a MachineBasicBlock which ends
285// with a RET and also has successors.
286let isPseudo = 1, SchedRW = [WriteJumpLd] in {
287def MORESTACK_RET: I<0, Pseudo, (outs), (ins), "", []>;
288
289// This instruction is lowered to a RET followed by a MOV.  The two
290// instructions are not generated on a higher level since then the
291// verifier sees a MachineBasicBlock ending with a non-terminator.
292def MORESTACK_RET_RESTORE_R10 : I<0, Pseudo, (outs), (ins), "", []>;
293}
294
295//===----------------------------------------------------------------------===//
296// Alias Instructions
297//===----------------------------------------------------------------------===//
298
299// Alias instruction mapping movr0 to xor.
300// FIXME: remove when we can teach regalloc that xor reg, reg is ok.
301let Defs = [EFLAGS], isReMaterializable = 1, isAsCheapAsAMove = 1,
302    isPseudo = 1, isMoveImm = 1, AddedComplexity = 10 in
303def MOV32r0  : I<0, Pseudo, (outs GR32:$dst), (ins), "",
304                 [(set GR32:$dst, 0)]>, Sched<[WriteZero]>;
305
306// Other widths can also make use of the 32-bit xor, which may have a smaller
307// encoding and avoid partial register updates.
308let AddedComplexity = 10 in {
309def : Pat<(i8 0), (EXTRACT_SUBREG (MOV32r0), sub_8bit)>;
310def : Pat<(i16 0), (EXTRACT_SUBREG (MOV32r0), sub_16bit)>;
311def : Pat<(i64 0), (SUBREG_TO_REG (i64 0), (MOV32r0), sub_32bit)>;
312}
313
314let Predicates = [OptForSize, Not64BitMode],
315    AddedComplexity = 10 in {
316  let SchedRW = [WriteALU] in {
317  // Pseudo instructions for materializing 1 and -1 using XOR+INC/DEC,
318  // which only require 3 bytes compared to MOV32ri which requires 5.
319  let Defs = [EFLAGS], isReMaterializable = 1, isPseudo = 1 in {
320    def MOV32r1 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
321                        [(set GR32:$dst, 1)]>;
322    def MOV32r_1 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
323                        [(set GR32:$dst, -1)]>;
324  }
325  } // SchedRW
326
327  // MOV16ri is 4 bytes, so the instructions above are smaller.
328  def : Pat<(i16 1), (EXTRACT_SUBREG (MOV32r1), sub_16bit)>;
329  def : Pat<(i16 -1), (EXTRACT_SUBREG (MOV32r_1), sub_16bit)>;
330}
331
332let isReMaterializable = 1, isPseudo = 1, AddedComplexity = 5,
333    SchedRW = [WriteALU] in {
334// AddedComplexity higher than MOV64ri but lower than MOV32r0 and MOV32r1.
335def MOV32ImmSExti8 : I<0, Pseudo, (outs GR32:$dst), (ins i32i8imm:$src), "",
336                       [(set GR32:$dst, i32immSExt8:$src)]>,
337                       Requires<[OptForMinSize, NotWin64WithoutFP]>;
338def MOV64ImmSExti8 : I<0, Pseudo, (outs GR64:$dst), (ins i64i8imm:$src), "",
339                       [(set GR64:$dst, i64immSExt8:$src)]>,
340                       Requires<[OptForMinSize, NotWin64WithoutFP]>;
341}
342
343// Materialize i64 constant where top 32-bits are zero. This could theoretically
344// use MOV32ri with a SUBREG_TO_REG to represent the zero-extension, however
345// that would make it more difficult to rematerialize.
346let AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1,
347    isPseudo = 1, SchedRW = [WriteMove] in
348def MOV32ri64 : I<0, Pseudo, (outs GR64:$dst), (ins i64i32imm:$src), "",
349                  [(set GR64:$dst, i64immZExt32:$src)]>;
350
351// This 64-bit pseudo-move can also be used for labels in the x86-64 small code
352// model.
353def mov64imm32 : ComplexPattern<i64, 1, "selectMOV64Imm32", [X86Wrapper]>;
354def : Pat<(i64 mov64imm32:$src), (MOV32ri64 mov64imm32:$src)>;
355
356// Use sbb to materialize carry bit.
357let Uses = [EFLAGS], Defs = [EFLAGS], isPseudo = 1, SchedRW = [WriteADC],
358    hasSideEffects = 0 in {
359// FIXME: These are pseudo ops that should be replaced with Pat<> patterns.
360// However, Pat<> can't replicate the destination reg into the inputs of the
361// result.
362def SETB_C32r : I<0, Pseudo, (outs GR32:$dst), (ins), "", []>;
363def SETB_C64r : I<0, Pseudo, (outs GR64:$dst), (ins), "", []>;
364} // isCodeGenOnly
365
366//===----------------------------------------------------------------------===//
367// String Pseudo Instructions
368//
369let SchedRW = [WriteMicrocoded] in {
370let Defs = [ECX,EDI,ESI], Uses = [ECX,EDI,ESI], isCodeGenOnly = 1 in {
371def REP_MOVSB_32 : I<0xA4, RawFrm, (outs), (ins),
372                    "{rep;movsb (%esi), %es:(%edi)|rep movsb es:[edi], [esi]}",
373                    [(X86rep_movs i8)]>, REP, AdSize32,
374                   Requires<[NotLP64]>;
375def REP_MOVSW_32 : I<0xA5, RawFrm, (outs), (ins),
376                    "{rep;movsw (%esi), %es:(%edi)|rep movsw es:[edi], [esi]}",
377                    [(X86rep_movs i16)]>, REP, AdSize32, OpSize16,
378                   Requires<[NotLP64]>;
379def REP_MOVSD_32 : I<0xA5, RawFrm, (outs), (ins),
380                    "{rep;movsl (%esi), %es:(%edi)|rep movsd es:[edi], [esi]}",
381                    [(X86rep_movs i32)]>, REP, AdSize32, OpSize32,
382                   Requires<[NotLP64]>;
383def REP_MOVSQ_32 : RI<0xA5, RawFrm, (outs), (ins),
384                    "{rep;movsq (%esi), %es:(%edi)|rep movsq es:[edi], [esi]}",
385                    [(X86rep_movs i64)]>, REP, AdSize32,
386                   Requires<[NotLP64, In64BitMode]>;
387}
388
389let Defs = [RCX,RDI,RSI], Uses = [RCX,RDI,RSI], isCodeGenOnly = 1 in {
390def REP_MOVSB_64 : I<0xA4, RawFrm, (outs), (ins),
391                    "{rep;movsb (%rsi), %es:(%rdi)|rep movsb es:[rdi], [rsi]}",
392                    [(X86rep_movs i8)]>, REP, AdSize64,
393                   Requires<[IsLP64]>;
394def REP_MOVSW_64 : I<0xA5, RawFrm, (outs), (ins),
395                    "{rep;movsw (%rsi), %es:(%rdi)|rep movsw es:[rdi], [rsi]}",
396                    [(X86rep_movs i16)]>, REP, AdSize64, OpSize16,
397                   Requires<[IsLP64]>;
398def REP_MOVSD_64 : I<0xA5, RawFrm, (outs), (ins),
399                    "{rep;movsl (%rsi), %es:(%rdi)|rep movsdi es:[rdi], [rsi]}",
400                    [(X86rep_movs i32)]>, REP, AdSize64, OpSize32,
401                   Requires<[IsLP64]>;
402def REP_MOVSQ_64 : RI<0xA5, RawFrm, (outs), (ins),
403                    "{rep;movsq (%rsi), %es:(%rdi)|rep movsq es:[rdi], [rsi]}",
404                    [(X86rep_movs i64)]>, REP, AdSize64,
405                   Requires<[IsLP64]>;
406}
407
408// FIXME: Should use "(X86rep_stos AL)" as the pattern.
409let Defs = [ECX,EDI], isCodeGenOnly = 1 in {
410  let Uses = [AL,ECX,EDI] in
411  def REP_STOSB_32 : I<0xAA, RawFrm, (outs), (ins),
412                       "{rep;stosb %al, %es:(%edi)|rep stosb es:[edi], al}",
413                      [(X86rep_stos i8)]>, REP, AdSize32,
414                     Requires<[NotLP64]>;
415  let Uses = [AX,ECX,EDI] in
416  def REP_STOSW_32 : I<0xAB, RawFrm, (outs), (ins),
417                      "{rep;stosw %ax, %es:(%edi)|rep stosw es:[edi], ax}",
418                      [(X86rep_stos i16)]>, REP, AdSize32, OpSize16,
419                     Requires<[NotLP64]>;
420  let Uses = [EAX,ECX,EDI] in
421  def REP_STOSD_32 : I<0xAB, RawFrm, (outs), (ins),
422                      "{rep;stosl %eax, %es:(%edi)|rep stosd es:[edi], eax}",
423                      [(X86rep_stos i32)]>, REP, AdSize32, OpSize32,
424                     Requires<[NotLP64]>;
425  let Uses = [RAX,RCX,RDI] in
426  def REP_STOSQ_32 : RI<0xAB, RawFrm, (outs), (ins),
427                        "{rep;stosq %rax, %es:(%edi)|rep stosq es:[edi], rax}",
428                        [(X86rep_stos i64)]>, REP, AdSize32,
429                        Requires<[NotLP64, In64BitMode]>;
430}
431
432let Defs = [RCX,RDI], isCodeGenOnly = 1 in {
433  let Uses = [AL,RCX,RDI] in
434  def REP_STOSB_64 : I<0xAA, RawFrm, (outs), (ins),
435                       "{rep;stosb %al, %es:(%rdi)|rep stosb es:[rdi], al}",
436                       [(X86rep_stos i8)]>, REP, AdSize64,
437                       Requires<[IsLP64]>;
438  let Uses = [AX,RCX,RDI] in
439  def REP_STOSW_64 : I<0xAB, RawFrm, (outs), (ins),
440                       "{rep;stosw %ax, %es:(%rdi)|rep stosw es:[rdi], ax}",
441                       [(X86rep_stos i16)]>, REP, AdSize64, OpSize16,
442                       Requires<[IsLP64]>;
443  let Uses = [RAX,RCX,RDI] in
444  def REP_STOSD_64 : I<0xAB, RawFrm, (outs), (ins),
445                      "{rep;stosl %eax, %es:(%rdi)|rep stosd es:[rdi], eax}",
446                       [(X86rep_stos i32)]>, REP, AdSize64, OpSize32,
447                       Requires<[IsLP64]>;
448
449  let Uses = [RAX,RCX,RDI] in
450  def REP_STOSQ_64 : RI<0xAB, RawFrm, (outs), (ins),
451                        "{rep;stosq %rax, %es:(%rdi)|rep stosq es:[rdi], rax}",
452                        [(X86rep_stos i64)]>, REP, AdSize64,
453                        Requires<[IsLP64]>;
454}
455} // SchedRW
456
457//===----------------------------------------------------------------------===//
458// Thread Local Storage Instructions
459//
460let SchedRW = [WriteSystem] in {
461
462// ELF TLS Support
463// All calls clobber the non-callee saved registers. ESP is marked as
464// a use to prevent stack-pointer assignments that appear immediately
465// before calls from potentially appearing dead.
466let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, FP7,
467            ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7,
468            MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
469            XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
470            XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS, DF],
471    usesCustomInserter = 1, Uses = [ESP, SSP] in {
472def TLS_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
473                  "# TLS_addr32",
474                  [(X86tlsaddr tls32addr:$sym)]>,
475                  Requires<[Not64BitMode]>;
476def TLS_base_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
477                  "# TLS_base_addr32",
478                  [(X86tlsbaseaddr tls32baseaddr:$sym)]>,
479                  Requires<[Not64BitMode]>;
480}
481
482// All calls clobber the non-callee saved registers. RSP is marked as
483// a use to prevent stack-pointer assignments that appear immediately
484// before calls from potentially appearing dead.
485let Defs = [RAX, RCX, RDX, RSI, RDI, R8, R9, R10, R11,
486            FP0, FP1, FP2, FP3, FP4, FP5, FP6, FP7,
487            ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7,
488            MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
489            XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
490            XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS, DF],
491    usesCustomInserter = 1, Uses = [RSP, SSP] in {
492def TLS_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
493                   "# TLS_addr64",
494                  [(X86tlsaddr tls64addr:$sym)]>,
495                  Requires<[In64BitMode, IsLP64]>;
496def TLS_base_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
497                   "# TLS_base_addr64",
498                  [(X86tlsbaseaddr tls64baseaddr:$sym)]>,
499                  Requires<[In64BitMode, IsLP64]>;
500def TLS_addrX32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
501                   "# TLS_addrX32",
502                  [(X86tlsaddr tls32addr:$sym)]>,
503                  Requires<[In64BitMode, NotLP64]>;
504def TLS_base_addrX32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
505                   "# TLS_base_addrX32",
506                  [(X86tlsbaseaddr tls32baseaddr:$sym)]>,
507                  Requires<[In64BitMode, NotLP64]>;
508}
509
510// Darwin TLS Support
511// For i386, the address of the thunk is passed on the stack, on return the
512// address of the variable is in %eax.  %ecx is trashed during the function
513// call.  All other registers are preserved.
514let Defs = [EAX, ECX, EFLAGS, DF],
515    Uses = [ESP, SSP],
516    usesCustomInserter = 1 in
517def TLSCall_32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
518                "# TLSCall_32",
519                [(X86TLSCall addr:$sym)]>,
520                Requires<[Not64BitMode]>;
521
522// For x86_64, the address of the thunk is passed in %rdi, but the
523// pseudo directly use the symbol, so do not add an implicit use of
524// %rdi. The lowering will do the right thing with RDI.
525// On return the address of the variable is in %rax.  All other
526// registers are preserved.
527let Defs = [RAX, EFLAGS, DF],
528    Uses = [RSP, SSP],
529    usesCustomInserter = 1 in
530def TLSCall_64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
531                  "# TLSCall_64",
532                  [(X86TLSCall addr:$sym)]>,
533                  Requires<[In64BitMode]>;
534} // SchedRW
535
536//===----------------------------------------------------------------------===//
537// Conditional Move Pseudo Instructions
538
539// CMOV* - Used to implement the SELECT DAG operation.  Expanded after
540// instruction selection into a branch sequence.
541multiclass CMOVrr_PSEUDO<RegisterClass RC, ValueType VT> {
542  def CMOV#NAME  : I<0, Pseudo,
543                    (outs RC:$dst), (ins RC:$t, RC:$f, i8imm:$cond),
544                    "#CMOV_"#NAME#" PSEUDO!",
545                    [(set RC:$dst, (VT (X86cmov RC:$t, RC:$f, timm:$cond,
546                                                EFLAGS)))]>;
547}
548
549let usesCustomInserter = 1, hasNoSchedulingInfo = 1, Uses = [EFLAGS] in {
550  // X86 doesn't have 8-bit conditional moves. Use a customInserter to
551  // emit control flow. An alternative to this is to mark i8 SELECT as Promote,
552  // however that requires promoting the operands, and can induce additional
553  // i8 register pressure.
554  defm _GR8 : CMOVrr_PSEUDO<GR8, i8>;
555
556  let Predicates = [NoCMOV] in {
557    defm _GR32 : CMOVrr_PSEUDO<GR32, i32>;
558    defm _GR16 : CMOVrr_PSEUDO<GR16, i16>;
559  } // Predicates = [NoCMOV]
560
561  // fcmov doesn't handle all possible EFLAGS, provide a fallback if there is no
562  // SSE1/SSE2.
563  let Predicates = [FPStackf32] in
564    defm _RFP32 : CMOVrr_PSEUDO<RFP32, f32>;
565
566  let Predicates = [FPStackf64] in
567    defm _RFP64 : CMOVrr_PSEUDO<RFP64, f64>;
568
569  defm _RFP80 : CMOVrr_PSEUDO<RFP80, f80>;
570
571  let Predicates = [HasMMX] in
572    defm _VR64   : CMOVrr_PSEUDO<VR64, x86mmx>;
573
574  let Predicates = [HasSSE1,NoAVX512] in
575    defm _FR32   : CMOVrr_PSEUDO<FR32, f32>;
576  let Predicates = [HasSSE2,NoAVX512] in {
577    defm _FR16   : CMOVrr_PSEUDO<FR16, f16>;
578    defm _FR64   : CMOVrr_PSEUDO<FR64, f64>;
579  }
580  let Predicates = [HasAVX512] in {
581    defm _FR16X  : CMOVrr_PSEUDO<FR16X, f16>;
582    defm _FR32X  : CMOVrr_PSEUDO<FR32X, f32>;
583    defm _FR64X  : CMOVrr_PSEUDO<FR64X, f64>;
584  }
585  let Predicates = [NoVLX] in {
586    defm _VR128  : CMOVrr_PSEUDO<VR128, v2i64>;
587    defm _VR256  : CMOVrr_PSEUDO<VR256, v4i64>;
588  }
589  let Predicates = [HasVLX] in {
590    defm _VR128X : CMOVrr_PSEUDO<VR128X, v2i64>;
591    defm _VR256X : CMOVrr_PSEUDO<VR256X, v4i64>;
592  }
593  defm _VR512  : CMOVrr_PSEUDO<VR512, v8i64>;
594  defm _VK1    : CMOVrr_PSEUDO<VK1,  v1i1>;
595  defm _VK2    : CMOVrr_PSEUDO<VK2,  v2i1>;
596  defm _VK4    : CMOVrr_PSEUDO<VK4,  v4i1>;
597  defm _VK8    : CMOVrr_PSEUDO<VK8,  v8i1>;
598  defm _VK16   : CMOVrr_PSEUDO<VK16, v16i1>;
599  defm _VK32   : CMOVrr_PSEUDO<VK32, v32i1>;
600  defm _VK64   : CMOVrr_PSEUDO<VK64, v64i1>;
601} // usesCustomInserter = 1, hasNoSchedulingInfo = 1, Uses = [EFLAGS]
602
603def : Pat<(f128 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
604          (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
605
606let Predicates = [NoVLX] in {
607  def : Pat<(v16i8 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
608            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
609  def : Pat<(v8i16 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
610            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
611  def : Pat<(v4i32 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
612            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
613  def : Pat<(v4f32 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
614            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
615  def : Pat<(v2f64 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
616            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
617
618  def : Pat<(v32i8 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
619            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
620  def : Pat<(v16i16 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
621            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
622  def : Pat<(v8i32 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
623            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
624  def : Pat<(v8f32 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
625            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
626  def : Pat<(v4f64 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
627            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
628}
629let Predicates = [HasVLX] in {
630  def : Pat<(v16i8 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
631            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
632  def : Pat<(v8i16 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
633            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
634  def : Pat<(v8f16 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
635            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
636  def : Pat<(v4i32 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
637            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
638  def : Pat<(v4f32 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
639            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
640  def : Pat<(v2f64 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
641            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
642
643  def : Pat<(v32i8 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
644            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
645  def : Pat<(v16i16 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
646            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
647  def : Pat<(v16f16 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
648            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
649  def : Pat<(v8i32 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
650            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
651  def : Pat<(v8f32 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
652            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
653  def : Pat<(v4f64 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
654            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
655}
656
657def : Pat<(v64i8 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
658          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
659def : Pat<(v32i16 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
660          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
661def : Pat<(v32f16 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
662          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
663def : Pat<(v16i32 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
664          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
665def : Pat<(v16f32 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
666          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
667def : Pat<(v8f64 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
668          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
669
670//===----------------------------------------------------------------------===//
671// Normal-Instructions-With-Lock-Prefix Pseudo Instructions
672//===----------------------------------------------------------------------===//
673
674// FIXME: Use normal instructions and add lock prefix dynamically.
675
676// Memory barriers
677
678let isCodeGenOnly = 1, Defs = [EFLAGS] in
679def OR32mi8Locked  : Ii8<0x83, MRM1m, (outs), (ins i32mem:$dst, i32i8imm:$zero),
680                         "or{l}\t{$zero, $dst|$dst, $zero}", []>,
681                         Requires<[Not64BitMode]>, OpSize32, LOCK,
682                         Sched<[WriteALURMW]>;
683
684// RegOpc corresponds to the mr version of the instruction
685// ImmOpc corresponds to the mi version of the instruction
686// ImmOpc8 corresponds to the mi8 version of the instruction
687// ImmMod corresponds to the instruction format of the mi and mi8 versions
688multiclass LOCK_ArithBinOp<bits<8> RegOpc, bits<8> ImmOpc, bits<8> ImmOpc8,
689                           Format ImmMod, SDNode Op, string mnemonic> {
690let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
691    SchedRW = [WriteALURMW] in {
692
693def NAME#8mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
694                  RegOpc{3}, RegOpc{2}, RegOpc{1}, 0 },
695                  MRMDestMem, (outs), (ins i8mem:$dst, GR8:$src2),
696                  !strconcat(mnemonic, "{b}\t",
697                             "{$src2, $dst|$dst, $src2}"),
698                  [(set EFLAGS, (Op addr:$dst, GR8:$src2))]>, LOCK;
699
700def NAME#16mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
701                   RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
702                   MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
703                   !strconcat(mnemonic, "{w}\t",
704                              "{$src2, $dst|$dst, $src2}"),
705                   [(set EFLAGS, (Op addr:$dst, GR16:$src2))]>,
706                   OpSize16, LOCK;
707
708def NAME#32mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
709                   RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
710                   MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
711                   !strconcat(mnemonic, "{l}\t",
712                              "{$src2, $dst|$dst, $src2}"),
713                   [(set EFLAGS, (Op addr:$dst, GR32:$src2))]>,
714                   OpSize32, LOCK;
715
716def NAME#64mr : RI<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
717                    RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
718                    MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
719                    !strconcat(mnemonic, "{q}\t",
720                               "{$src2, $dst|$dst, $src2}"),
721                    [(set EFLAGS, (Op addr:$dst, GR64:$src2))]>, LOCK;
722
723// NOTE: These are order specific, we want the mi8 forms to be listed
724// first so that they are slightly preferred to the mi forms.
725def NAME#16mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
726                      ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
727                      ImmMod, (outs), (ins i16mem :$dst, i16i8imm :$src2),
728                      !strconcat(mnemonic, "{w}\t",
729                                 "{$src2, $dst|$dst, $src2}"),
730                      [(set EFLAGS, (Op addr:$dst, i16immSExt8:$src2))]>,
731                      OpSize16, LOCK;
732
733def NAME#32mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
734                      ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
735                      ImmMod, (outs), (ins i32mem :$dst, i32i8imm :$src2),
736                      !strconcat(mnemonic, "{l}\t",
737                                 "{$src2, $dst|$dst, $src2}"),
738                      [(set EFLAGS, (Op addr:$dst, i32immSExt8:$src2))]>,
739                      OpSize32, LOCK;
740
741def NAME#64mi8 : RIi8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
742                       ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
743                       ImmMod, (outs), (ins i64mem :$dst, i64i8imm :$src2),
744                       !strconcat(mnemonic, "{q}\t",
745                                  "{$src2, $dst|$dst, $src2}"),
746                       [(set EFLAGS, (Op addr:$dst, i64immSExt8:$src2))]>,
747                       LOCK;
748
749def NAME#8mi : Ii8<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
750                    ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 0 },
751                    ImmMod, (outs), (ins i8mem :$dst, i8imm :$src2),
752                    !strconcat(mnemonic, "{b}\t",
753                               "{$src2, $dst|$dst, $src2}"),
754                    [(set EFLAGS, (Op addr:$dst, (i8 imm:$src2)))]>, LOCK;
755
756def NAME#16mi : Ii16<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
757                      ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
758                      ImmMod, (outs), (ins i16mem :$dst, i16imm :$src2),
759                      !strconcat(mnemonic, "{w}\t",
760                                 "{$src2, $dst|$dst, $src2}"),
761                      [(set EFLAGS, (Op addr:$dst, (i16 imm:$src2)))]>,
762                      OpSize16, LOCK;
763
764def NAME#32mi : Ii32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
765                      ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
766                      ImmMod, (outs), (ins i32mem :$dst, i32imm :$src2),
767                      !strconcat(mnemonic, "{l}\t",
768                                 "{$src2, $dst|$dst, $src2}"),
769                      [(set EFLAGS, (Op addr:$dst, (i32 imm:$src2)))]>,
770                      OpSize32, LOCK;
771
772def NAME#64mi32 : RIi32S<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
773                          ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
774                          ImmMod, (outs), (ins i64mem :$dst, i64i32imm :$src2),
775                          !strconcat(mnemonic, "{q}\t",
776                                     "{$src2, $dst|$dst, $src2}"),
777                          [(set EFLAGS, (Op addr:$dst, i64immSExt32:$src2))]>,
778                          LOCK;
779}
780
781}
782
783defm LOCK_ADD : LOCK_ArithBinOp<0x00, 0x80, 0x83, MRM0m, X86lock_add, "add">;
784defm LOCK_SUB : LOCK_ArithBinOp<0x28, 0x80, 0x83, MRM5m, X86lock_sub, "sub">;
785defm LOCK_OR  : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM1m, X86lock_or , "or">;
786defm LOCK_AND : LOCK_ArithBinOp<0x20, 0x80, 0x83, MRM4m, X86lock_and, "and">;
787defm LOCK_XOR : LOCK_ArithBinOp<0x30, 0x80, 0x83, MRM6m, X86lock_xor, "xor">;
788
789let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
790    SchedRW = [WriteALURMW]  in {
791  let Predicates = [UseIncDec] in {
792    def LOCK_INC8m  : I<0xFE, MRM0m, (outs), (ins i8mem :$dst),
793                        "inc{b}\t$dst",
794                        [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i8 1)))]>,
795                        LOCK;
796    def LOCK_INC16m : I<0xFF, MRM0m, (outs), (ins i16mem:$dst),
797                        "inc{w}\t$dst",
798                        [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i16 1)))]>,
799                        OpSize16, LOCK;
800    def LOCK_INC32m : I<0xFF, MRM0m, (outs), (ins i32mem:$dst),
801                        "inc{l}\t$dst",
802                        [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i32 1)))]>,
803                        OpSize32, LOCK;
804
805    def LOCK_DEC8m  : I<0xFE, MRM1m, (outs), (ins i8mem :$dst),
806                        "dec{b}\t$dst",
807                        [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i8 1)))]>,
808                        LOCK;
809    def LOCK_DEC16m : I<0xFF, MRM1m, (outs), (ins i16mem:$dst),
810                        "dec{w}\t$dst",
811                        [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i16 1)))]>,
812                        OpSize16, LOCK;
813    def LOCK_DEC32m : I<0xFF, MRM1m, (outs), (ins i32mem:$dst),
814                        "dec{l}\t$dst",
815                        [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i32 1)))]>,
816                        OpSize32, LOCK;
817  }
818
819  let Predicates = [UseIncDec, In64BitMode] in {
820    def LOCK_INC64m : RI<0xFF, MRM0m, (outs), (ins i64mem:$dst),
821                         "inc{q}\t$dst",
822                         [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i64 1)))]>,
823                         LOCK;
824    def LOCK_DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst),
825                         "dec{q}\t$dst",
826                         [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i64 1)))]>,
827                         LOCK;
828  }
829}
830
831let Predicates = [UseIncDec] in {
832  // Additional patterns for -1 constant.
833  def : Pat<(X86lock_add addr:$dst, (i8  -1)), (LOCK_DEC8m  addr:$dst)>;
834  def : Pat<(X86lock_add addr:$dst, (i16 -1)), (LOCK_DEC16m addr:$dst)>;
835  def : Pat<(X86lock_add addr:$dst, (i32 -1)), (LOCK_DEC32m addr:$dst)>;
836  def : Pat<(X86lock_sub addr:$dst, (i8  -1)), (LOCK_INC8m  addr:$dst)>;
837  def : Pat<(X86lock_sub addr:$dst, (i16 -1)), (LOCK_INC16m addr:$dst)>;
838  def : Pat<(X86lock_sub addr:$dst, (i32 -1)), (LOCK_INC32m addr:$dst)>;
839}
840
841let Predicates = [UseIncDec, In64BitMode] in {
842  // Additional patterns for -1 constant.
843  def : Pat<(X86lock_add addr:$dst, (i64 -1)), (LOCK_DEC64m addr:$dst)>;
844  def : Pat<(X86lock_sub addr:$dst, (i64 -1)), (LOCK_INC64m addr:$dst)>;
845}
846
847// Atomic bit test.
848def X86LBTest : SDTypeProfile<1, 3, [SDTCisVT<0, i32>, SDTCisPtrTy<1>,
849                                     SDTCisVT<2, i8>, SDTCisVT<3, i32>]>;
850def x86bts : SDNode<"X86ISD::LBTS", X86LBTest,
851                    [SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPMemOperand]>;
852def x86btc : SDNode<"X86ISD::LBTC", X86LBTest,
853                    [SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPMemOperand]>;
854def x86btr : SDNode<"X86ISD::LBTR", X86LBTest,
855                    [SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPMemOperand]>;
856
857def X86LBTestRM : SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<1>,
858                                       SDTCisInt<2>]>;
859
860def x86_rm_bts : SDNode<"X86ISD::LBTS_RM", X86LBTestRM,
861                        [SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPMemOperand]>;
862def x86_rm_btc : SDNode<"X86ISD::LBTC_RM", X86LBTestRM,
863                        [SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPMemOperand]>;
864def x86_rm_btr : SDNode<"X86ISD::LBTR_RM", X86LBTestRM,
865                        [SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPMemOperand]>;
866
867
868multiclass ATOMIC_LOGIC_OP<Format Form, string s> {
869  let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
870      SchedRW = [WriteBitTestSetRegRMW]  in {
871    def 16m : Ii8<0xBA, Form, (outs), (ins i16mem:$src1, i8imm:$src2),
872                  !strconcat(s, "{w}\t{$src2, $src1|$src1, $src2}"),
873                  [(set EFLAGS, (!cast<SDNode>("x86" # s) addr:$src1, timm:$src2, (i32 16)))]>,
874              OpSize16, TB, LOCK;
875    def 32m : Ii8<0xBA, Form, (outs), (ins i32mem:$src1, i8imm:$src2),
876                  !strconcat(s, "{l}\t{$src2, $src1|$src1, $src2}"),
877                  [(set EFLAGS, (!cast<SDNode>("x86" # s) addr:$src1, timm:$src2, (i32 32)))]>,
878              OpSize32, TB, LOCK;
879    def 64m : RIi8<0xBA, Form, (outs), (ins i64mem:$src1, i8imm:$src2),
880                   !strconcat(s, "{q}\t{$src2, $src1|$src1, $src2}"),
881                   [(set EFLAGS, (!cast<SDNode>("x86" # s) addr:$src1, timm:$src2, (i32 64)))]>,
882              TB, LOCK;
883  }
884}
885
886multiclass ATOMIC_LOGIC_OP_RM<bits<8> Opc8, string s> {
887  let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
888      SchedRW = [WriteBitTestSetRegRMW]  in {
889    def 16rm : I<Opc8, MRMDestMem, (outs), (ins i16mem:$src1, GR16:$src2),
890                  !strconcat(s, "{w}\t{$src2, $src1|$src1, $src2}"),
891                  [(set EFLAGS, (!cast<SDNode>("x86_rm_" # s) addr:$src1, GR16:$src2))]>,
892               OpSize16, TB, LOCK;
893    def 32rm : I<Opc8, MRMDestMem, (outs), (ins i32mem:$src1, GR32:$src2),
894                  !strconcat(s, "{l}\t{$src2, $src1|$src1, $src2}"),
895                  [(set EFLAGS, (!cast<SDNode>("x86_rm_" # s) addr:$src1, GR32:$src2))]>,
896               OpSize32, TB, LOCK;
897    def 64rm : RI<Opc8, MRMDestMem, (outs), (ins i64mem:$src1, GR64:$src2),
898                   !strconcat(s, "{q}\t{$src2, $src1|$src1, $src2}"),
899                   [(set EFLAGS, (!cast<SDNode>("x86_rm_" # s) addr:$src1, GR64:$src2))]>,
900               TB, LOCK;
901  }
902}
903
904
905defm LOCK_BTS : ATOMIC_LOGIC_OP<MRM5m, "bts">;
906defm LOCK_BTC : ATOMIC_LOGIC_OP<MRM7m, "btc">;
907defm LOCK_BTR : ATOMIC_LOGIC_OP<MRM6m, "btr">;
908
909defm LOCK_BTS_RM : ATOMIC_LOGIC_OP_RM<0xAB, "bts">;
910defm LOCK_BTC_RM : ATOMIC_LOGIC_OP_RM<0xBB, "btc">;
911defm LOCK_BTR_RM : ATOMIC_LOGIC_OP_RM<0xB3, "btr">;
912
913// Atomic compare and swap.
914multiclass LCMPXCHG_BinOp<bits<8> Opc8, bits<8> Opc, Format Form,
915                          string mnemonic, SDPatternOperator frag> {
916let isCodeGenOnly = 1, SchedRW = [WriteCMPXCHGRMW] in {
917  let Defs = [AL, EFLAGS], Uses = [AL] in
918  def NAME#8  : I<Opc8, Form, (outs), (ins i8mem:$ptr, GR8:$swap),
919                  !strconcat(mnemonic, "{b}\t{$swap, $ptr|$ptr, $swap}"),
920                  [(frag addr:$ptr, GR8:$swap, 1)]>, TB, LOCK;
921  let Defs = [AX, EFLAGS], Uses = [AX] in
922  def NAME#16 : I<Opc, Form, (outs), (ins i16mem:$ptr, GR16:$swap),
923                  !strconcat(mnemonic, "{w}\t{$swap, $ptr|$ptr, $swap}"),
924                  [(frag addr:$ptr, GR16:$swap, 2)]>, TB, OpSize16, LOCK;
925  let Defs = [EAX, EFLAGS], Uses = [EAX] in
926  def NAME#32 : I<Opc, Form, (outs), (ins i32mem:$ptr, GR32:$swap),
927                  !strconcat(mnemonic, "{l}\t{$swap, $ptr|$ptr, $swap}"),
928                  [(frag addr:$ptr, GR32:$swap, 4)]>, TB, OpSize32, LOCK;
929  let Defs = [RAX, EFLAGS], Uses = [RAX] in
930  def NAME#64 : RI<Opc, Form, (outs), (ins i64mem:$ptr, GR64:$swap),
931                   !strconcat(mnemonic, "{q}\t{$swap, $ptr|$ptr, $swap}"),
932                   [(frag addr:$ptr, GR64:$swap, 8)]>, TB, LOCK;
933}
934}
935
936let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
937    Predicates = [HasCX8], SchedRW = [WriteCMPXCHGRMW],
938    isCodeGenOnly = 1, usesCustomInserter = 1 in {
939def LCMPXCHG8B : I<0xC7, MRM1m, (outs), (ins i64mem:$ptr),
940                   "cmpxchg8b\t$ptr",
941                   [(X86cas8 addr:$ptr)]>, TB, LOCK;
942}
943
944let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RBX, RCX, RDX],
945    Predicates = [HasCX16,In64BitMode], SchedRW = [WriteCMPXCHGRMW],
946    isCodeGenOnly = 1, mayLoad = 1, mayStore = 1, hasSideEffects = 0 in {
947def LCMPXCHG16B : RI<0xC7, MRM1m, (outs), (ins i128mem:$ptr),
948                     "cmpxchg16b\t$ptr",
949                     []>, TB, LOCK;
950}
951
952// This pseudo must be used when the frame uses RBX as
953// the base pointer. Indeed, in such situation RBX is a reserved
954// register and the register allocator will ignore any use/def of
955// it. In other words, the register will not fix the clobbering of
956// RBX that will happen when setting the arguments for the instrucion.
957//
958// Unlike the actual related instruction, we mark that this one
959// defines RBX (instead of using RBX).
960// The rationale is that we will define RBX during the expansion of
961// the pseudo. The argument feeding RBX is rbx_input.
962//
963// The additional argument, $rbx_save, is a temporary register used to
964// save the value of RBX across the actual instruction.
965//
966// To make sure the register assigned to $rbx_save does not interfere with
967// the definition of the actual instruction, we use a definition $dst which
968// is tied to $rbx_save. That way, the live-range of $rbx_save spans across
969// the instruction and we are sure we will have a valid register to restore
970// the value of RBX.
971let Defs = [RAX, RDX, RBX, EFLAGS], Uses = [RAX, RCX, RDX],
972    Predicates = [HasCX16,In64BitMode], SchedRW = [WriteCMPXCHGRMW],
973    isCodeGenOnly = 1, isPseudo = 1,
974    mayLoad = 1, mayStore = 1, hasSideEffects = 0,
975    Constraints = "$rbx_save = $dst" in {
976def LCMPXCHG16B_SAVE_RBX :
977    I<0, Pseudo, (outs GR64:$dst),
978      (ins i128mem:$ptr, GR64:$rbx_input, GR64:$rbx_save), "", []>;
979}
980
981// Pseudo instruction that doesn't read/write RBX. Will be turned into either
982// LCMPXCHG16B_SAVE_RBX or LCMPXCHG16B via a custom inserter.
983let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RCX, RDX],
984    Predicates = [HasCX16,In64BitMode], SchedRW = [WriteCMPXCHGRMW],
985    isCodeGenOnly = 1, isPseudo = 1,
986    mayLoad = 1, mayStore = 1, hasSideEffects = 0,
987    usesCustomInserter = 1 in {
988def LCMPXCHG16B_NO_RBX :
989    I<0, Pseudo, (outs), (ins i128mem:$ptr, GR64:$rbx_input), "",
990      [(X86cas16 addr:$ptr, GR64:$rbx_input)]>;
991}
992
993// This pseudo must be used when the frame uses RBX/EBX as
994// the base pointer.
995// cf comment for LCMPXCHG16B_SAVE_RBX.
996let Defs = [EBX], Uses = [ECX, EAX],
997    Predicates = [HasMWAITX], SchedRW = [WriteSystem],
998    isCodeGenOnly = 1, isPseudo = 1, Constraints = "$rbx_save = $dst" in {
999def MWAITX_SAVE_RBX :
1000    I<0, Pseudo, (outs GR64:$dst),
1001      (ins GR32:$ebx_input, GR64:$rbx_save),
1002      "mwaitx",
1003      []>;
1004}
1005
1006// Pseudo mwaitx instruction to use for custom insertion.
1007let Predicates = [HasMWAITX], SchedRW = [WriteSystem],
1008    isCodeGenOnly = 1, isPseudo = 1,
1009    usesCustomInserter = 1 in {
1010def MWAITX :
1011    I<0, Pseudo, (outs), (ins GR32:$ecx, GR32:$eax, GR32:$ebx),
1012      "mwaitx",
1013      [(int_x86_mwaitx GR32:$ecx, GR32:$eax, GR32:$ebx)]>;
1014}
1015
1016
1017defm LCMPXCHG : LCMPXCHG_BinOp<0xB0, 0xB1, MRMDestMem, "cmpxchg", X86cas>;
1018
1019// Atomic exchange and add
1020multiclass ATOMIC_RMW_BINOP<bits<8> opc8, bits<8> opc, string mnemonic,
1021                            string frag> {
1022  let Constraints = "$val = $dst", Defs = [EFLAGS], mayLoad = 1, mayStore = 1,
1023      isCodeGenOnly = 1, SchedRW = [WriteALURMW] in {
1024    def NAME#8  : I<opc8, MRMSrcMem, (outs GR8:$dst),
1025                    (ins GR8:$val, i8mem:$ptr),
1026                    !strconcat(mnemonic, "{b}\t{$val, $ptr|$ptr, $val}"),
1027                    [(set GR8:$dst,
1028                          (!cast<PatFrag>(frag # "_8") addr:$ptr, GR8:$val))]>;
1029    def NAME#16 : I<opc, MRMSrcMem, (outs GR16:$dst),
1030                    (ins GR16:$val, i16mem:$ptr),
1031                    !strconcat(mnemonic, "{w}\t{$val, $ptr|$ptr, $val}"),
1032                    [(set
1033                       GR16:$dst,
1034                       (!cast<PatFrag>(frag # "_16") addr:$ptr, GR16:$val))]>,
1035                    OpSize16;
1036    def NAME#32 : I<opc, MRMSrcMem, (outs GR32:$dst),
1037                    (ins GR32:$val, i32mem:$ptr),
1038                    !strconcat(mnemonic, "{l}\t{$val, $ptr|$ptr, $val}"),
1039                    [(set
1040                       GR32:$dst,
1041                       (!cast<PatFrag>(frag # "_32") addr:$ptr, GR32:$val))]>,
1042                    OpSize32;
1043    def NAME#64 : RI<opc, MRMSrcMem, (outs GR64:$dst),
1044                     (ins GR64:$val, i64mem:$ptr),
1045                     !strconcat(mnemonic, "{q}\t{$val, $ptr|$ptr, $val}"),
1046                     [(set
1047                        GR64:$dst,
1048                        (!cast<PatFrag>(frag # "_64") addr:$ptr, GR64:$val))]>;
1049  }
1050}
1051
1052defm LXADD : ATOMIC_RMW_BINOP<0xc0, 0xc1, "xadd", "atomic_load_add">, TB, LOCK;
1053
1054/* The following multiclass tries to make sure that in code like
1055 *    x.store (immediate op x.load(acquire), release)
1056 * and
1057 *    x.store (register op x.load(acquire), release)
1058 * an operation directly on memory is generated instead of wasting a register.
1059 * It is not automatic as atomic_store/load are only lowered to MOV instructions
1060 * extremely late to prevent them from being accidentally reordered in the backend
1061 * (see below the RELEASE_MOV* / ACQUIRE_MOV* pseudo-instructions)
1062 */
1063multiclass RELEASE_BINOP_MI<string Name, SDNode op> {
1064  def : Pat<(atomic_store_8 (op (atomic_load_8 addr:$dst), (i8 imm:$src)),
1065                            addr:$dst),
1066            (!cast<Instruction>(Name#"8mi") addr:$dst, imm:$src)>;
1067  def : Pat<(atomic_store_16 (op (atomic_load_16 addr:$dst), (i16 imm:$src)),
1068                             addr:$dst),
1069            (!cast<Instruction>(Name#"16mi") addr:$dst, imm:$src)>;
1070  def : Pat<(atomic_store_32 (op (atomic_load_32 addr:$dst), (i32 imm:$src)),
1071                             addr:$dst),
1072            (!cast<Instruction>(Name#"32mi") addr:$dst, imm:$src)>;
1073  def : Pat<(atomic_store_64 (op (atomic_load_64 addr:$dst), (i64immSExt32:$src)),
1074                             addr:$dst),
1075            (!cast<Instruction>(Name#"64mi32") addr:$dst, (i64immSExt32:$src))>;
1076  def : Pat<(atomic_store_8 (op (atomic_load_8 addr:$dst), (i8 GR8:$src)), addr:$dst),
1077            (!cast<Instruction>(Name#"8mr") addr:$dst, GR8:$src)>;
1078  def : Pat<(atomic_store_16 (op (atomic_load_16 addr:$dst), (i16 GR16:$src)),
1079                             addr:$dst),
1080            (!cast<Instruction>(Name#"16mr") addr:$dst, GR16:$src)>;
1081  def : Pat<(atomic_store_32 (op (atomic_load_32 addr:$dst), (i32 GR32:$src)),
1082                             addr:$dst),
1083            (!cast<Instruction>(Name#"32mr") addr:$dst, GR32:$src)>;
1084  def : Pat<(atomic_store_64 (op (atomic_load_64 addr:$dst), (i64 GR64:$src)),
1085                             addr:$dst),
1086            (!cast<Instruction>(Name#"64mr") addr:$dst, GR64:$src)>;
1087}
1088defm : RELEASE_BINOP_MI<"ADD", add>;
1089defm : RELEASE_BINOP_MI<"AND", and>;
1090defm : RELEASE_BINOP_MI<"OR",  or>;
1091defm : RELEASE_BINOP_MI<"XOR", xor>;
1092defm : RELEASE_BINOP_MI<"SUB", sub>;
1093
1094// Atomic load + floating point patterns.
1095// FIXME: This could also handle SIMD operations with *ps and *pd instructions.
1096multiclass ATOMIC_LOAD_FP_BINOP_MI<string Name, SDNode op> {
1097  def : Pat<(op FR32:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
1098            (!cast<Instruction>(Name#"SSrm") FR32:$src1, addr:$src2)>,
1099            Requires<[UseSSE1]>;
1100  def : Pat<(op FR32:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
1101            (!cast<Instruction>("V"#Name#"SSrm") FR32:$src1, addr:$src2)>,
1102            Requires<[UseAVX]>;
1103  def : Pat<(op FR32X:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
1104            (!cast<Instruction>("V"#Name#"SSZrm") FR32X:$src1, addr:$src2)>,
1105            Requires<[HasAVX512]>;
1106
1107  def : Pat<(op FR64:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
1108            (!cast<Instruction>(Name#"SDrm") FR64:$src1, addr:$src2)>,
1109            Requires<[UseSSE1]>;
1110  def : Pat<(op FR64:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
1111            (!cast<Instruction>("V"#Name#"SDrm") FR64:$src1, addr:$src2)>,
1112            Requires<[UseAVX]>;
1113  def : Pat<(op FR64X:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
1114            (!cast<Instruction>("V"#Name#"SDZrm") FR64X:$src1, addr:$src2)>,
1115            Requires<[HasAVX512]>;
1116}
1117defm : ATOMIC_LOAD_FP_BINOP_MI<"ADD", fadd>;
1118// FIXME: Add fsub, fmul, fdiv, ...
1119
1120multiclass RELEASE_UNOP<string Name, dag dag8, dag dag16, dag dag32,
1121                        dag dag64> {
1122  def : Pat<(atomic_store_8 dag8, addr:$dst),
1123            (!cast<Instruction>(Name#8m) addr:$dst)>;
1124  def : Pat<(atomic_store_16 dag16, addr:$dst),
1125            (!cast<Instruction>(Name#16m) addr:$dst)>;
1126  def : Pat<(atomic_store_32 dag32, addr:$dst),
1127            (!cast<Instruction>(Name#32m) addr:$dst)>;
1128  def : Pat<(atomic_store_64 dag64, addr:$dst),
1129            (!cast<Instruction>(Name#64m) addr:$dst)>;
1130}
1131
1132let Predicates = [UseIncDec] in {
1133  defm : RELEASE_UNOP<"INC",
1134      (add (atomic_load_8  addr:$dst), (i8 1)),
1135      (add (atomic_load_16 addr:$dst), (i16 1)),
1136      (add (atomic_load_32 addr:$dst), (i32 1)),
1137      (add (atomic_load_64 addr:$dst), (i64 1))>;
1138  defm : RELEASE_UNOP<"DEC",
1139      (add (atomic_load_8  addr:$dst), (i8 -1)),
1140      (add (atomic_load_16 addr:$dst), (i16 -1)),
1141      (add (atomic_load_32 addr:$dst), (i32 -1)),
1142      (add (atomic_load_64 addr:$dst), (i64 -1))>;
1143}
1144
1145defm : RELEASE_UNOP<"NEG",
1146    (ineg (i8 (atomic_load_8  addr:$dst))),
1147    (ineg (i16 (atomic_load_16 addr:$dst))),
1148    (ineg (i32 (atomic_load_32 addr:$dst))),
1149    (ineg (i64 (atomic_load_64 addr:$dst)))>;
1150defm : RELEASE_UNOP<"NOT",
1151    (not (i8 (atomic_load_8  addr:$dst))),
1152    (not (i16 (atomic_load_16 addr:$dst))),
1153    (not (i32 (atomic_load_32 addr:$dst))),
1154    (not (i64 (atomic_load_64 addr:$dst)))>;
1155
1156def : Pat<(atomic_store_8 (i8 imm:$src), addr:$dst),
1157          (MOV8mi addr:$dst, imm:$src)>;
1158def : Pat<(atomic_store_16 (i16 imm:$src), addr:$dst),
1159          (MOV16mi addr:$dst, imm:$src)>;
1160def : Pat<(atomic_store_32 (i32 imm:$src), addr:$dst),
1161          (MOV32mi addr:$dst, imm:$src)>;
1162def : Pat<(atomic_store_64 (i64immSExt32:$src), addr:$dst),
1163          (MOV64mi32 addr:$dst, i64immSExt32:$src)>;
1164
1165def : Pat<(atomic_store_8 GR8:$src, addr:$dst),
1166          (MOV8mr addr:$dst, GR8:$src)>;
1167def : Pat<(atomic_store_16 GR16:$src, addr:$dst),
1168          (MOV16mr addr:$dst, GR16:$src)>;
1169def : Pat<(atomic_store_32 GR32:$src, addr:$dst),
1170          (MOV32mr addr:$dst, GR32:$src)>;
1171def : Pat<(atomic_store_64 GR64:$src, addr:$dst),
1172          (MOV64mr addr:$dst, GR64:$src)>;
1173
1174def : Pat<(i8  (atomic_load_8 addr:$src)),  (MOV8rm addr:$src)>;
1175def : Pat<(i16 (atomic_load_16 addr:$src)), (MOV16rm addr:$src)>;
1176def : Pat<(i32 (atomic_load_32 addr:$src)), (MOV32rm addr:$src)>;
1177def : Pat<(i64 (atomic_load_64 addr:$src)), (MOV64rm addr:$src)>;
1178
1179// Floating point loads/stores.
1180def : Pat<(atomic_store_32 (i32 (bitconvert (f32 FR32:$src))), addr:$dst),
1181          (MOVSSmr addr:$dst, FR32:$src)>, Requires<[UseSSE1]>;
1182def : Pat<(atomic_store_32 (i32 (bitconvert (f32 FR32:$src))), addr:$dst),
1183          (VMOVSSmr addr:$dst, FR32:$src)>, Requires<[UseAVX]>;
1184def : Pat<(atomic_store_32 (i32 (bitconvert (f32 FR32:$src))), addr:$dst),
1185          (VMOVSSZmr addr:$dst, FR32:$src)>, Requires<[HasAVX512]>;
1186
1187def : Pat<(atomic_store_64 (i64 (bitconvert (f64 FR64:$src))), addr:$dst),
1188          (MOVSDmr addr:$dst, FR64:$src)>, Requires<[UseSSE2]>;
1189def : Pat<(atomic_store_64 (i64 (bitconvert (f64 FR64:$src))), addr:$dst),
1190          (VMOVSDmr addr:$dst, FR64:$src)>, Requires<[UseAVX]>;
1191def : Pat<(atomic_store_64 (i64 (bitconvert (f64 FR64:$src))), addr:$dst),
1192          (VMOVSDmr addr:$dst, FR64:$src)>, Requires<[HasAVX512]>;
1193
1194def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1195          (MOVSSrm_alt addr:$src)>, Requires<[UseSSE1]>;
1196def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1197          (VMOVSSrm_alt addr:$src)>, Requires<[UseAVX]>;
1198def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1199          (VMOVSSZrm_alt addr:$src)>, Requires<[HasAVX512]>;
1200
1201def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1202          (MOVSDrm_alt addr:$src)>, Requires<[UseSSE2]>;
1203def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1204          (VMOVSDrm_alt addr:$src)>, Requires<[UseAVX]>;
1205def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1206          (VMOVSDZrm_alt addr:$src)>, Requires<[HasAVX512]>;
1207
1208//===----------------------------------------------------------------------===//
1209// DAG Pattern Matching Rules
1210//===----------------------------------------------------------------------===//
1211
1212// Use AND/OR to store 0/-1 in memory when optimizing for minsize. This saves
1213// binary size compared to a regular MOV, but it introduces an unnecessary
1214// load, so is not suitable for regular or optsize functions.
1215let Predicates = [OptForMinSize] in {
1216def : Pat<(simple_store (i16 0), addr:$dst), (AND16mi addr:$dst, 0)>;
1217def : Pat<(simple_store (i32 0), addr:$dst), (AND32mi addr:$dst, 0)>;
1218def : Pat<(simple_store (i64 0), addr:$dst), (AND64mi32 addr:$dst, 0)>;
1219def : Pat<(simple_store (i16 -1), addr:$dst), (OR16mi addr:$dst, -1)>;
1220def : Pat<(simple_store (i32 -1), addr:$dst), (OR32mi addr:$dst, -1)>;
1221def : Pat<(simple_store (i64 -1), addr:$dst), (OR64mi32 addr:$dst, -1)>;
1222}
1223
1224// In kernel code model, we can get the address of a label
1225// into a register with 'movq'.  FIXME: This is a hack, the 'imm' predicate of
1226// the MOV64ri32 should accept these.
1227def : Pat<(i64 (X86Wrapper tconstpool  :$dst)),
1228          (MOV64ri32 tconstpool  :$dst)>, Requires<[KernelCode]>;
1229def : Pat<(i64 (X86Wrapper tjumptable  :$dst)),
1230          (MOV64ri32 tjumptable  :$dst)>, Requires<[KernelCode]>;
1231def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
1232          (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
1233def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
1234          (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
1235def : Pat<(i64 (X86Wrapper mcsym:$dst)),
1236          (MOV64ri32 mcsym:$dst)>, Requires<[KernelCode]>;
1237def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
1238          (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
1239
1240// If we have small model and -static mode, it is safe to store global addresses
1241// directly as immediates.  FIXME: This is really a hack, the 'imm' predicate
1242// for MOV64mi32 should handle this sort of thing.
1243def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
1244          (MOV64mi32 addr:$dst, tconstpool:$src)>,
1245          Requires<[NearData, IsNotPIC]>;
1246def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
1247          (MOV64mi32 addr:$dst, tjumptable:$src)>,
1248          Requires<[NearData, IsNotPIC]>;
1249def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
1250          (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
1251          Requires<[NearData, IsNotPIC]>;
1252def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
1253          (MOV64mi32 addr:$dst, texternalsym:$src)>,
1254          Requires<[NearData, IsNotPIC]>;
1255def : Pat<(store (i64 (X86Wrapper mcsym:$src)), addr:$dst),
1256          (MOV64mi32 addr:$dst, mcsym:$src)>,
1257          Requires<[NearData, IsNotPIC]>;
1258def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
1259          (MOV64mi32 addr:$dst, tblockaddress:$src)>,
1260          Requires<[NearData, IsNotPIC]>;
1261
1262def : Pat<(i32 (X86RecoverFrameAlloc mcsym:$dst)), (MOV32ri mcsym:$dst)>;
1263def : Pat<(i64 (X86RecoverFrameAlloc mcsym:$dst)), (MOV64ri mcsym:$dst)>;
1264
1265// Calls
1266
1267// tls has some funny stuff here...
1268// This corresponds to movabs $foo@tpoff, %rax
1269def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
1270          (MOV64ri32 tglobaltlsaddr :$dst)>;
1271// This corresponds to add $foo@tpoff, %rax
1272def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
1273          (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
1274
1275
1276// Direct PC relative function call for small code model. 32-bit displacement
1277// sign extended to 64-bit.
1278def : Pat<(X86call (i64 tglobaladdr:$dst)),
1279          (CALL64pcrel32 tglobaladdr:$dst)>;
1280def : Pat<(X86call (i64 texternalsym:$dst)),
1281          (CALL64pcrel32 texternalsym:$dst)>;
1282
1283def : Pat<(X86call_rvmarker (i64 tglobaladdr:$rvfunc), (i64 texternalsym:$dst)),
1284          (CALL64pcrel32_RVMARKER tglobaladdr:$rvfunc, texternalsym:$dst)>;
1285def : Pat<(X86call_rvmarker (i64 tglobaladdr:$rvfunc), (i64 tglobaladdr:$dst)),
1286          (CALL64pcrel32_RVMARKER tglobaladdr:$rvfunc, tglobaladdr:$dst)>;
1287
1288
1289// Tailcall stuff. The TCRETURN instructions execute after the epilog, so they
1290// can never use callee-saved registers. That is the purpose of the GR64_TC
1291// register classes.
1292//
1293// The only volatile register that is never used by the calling convention is
1294// %r11. This happens when calling a vararg function with 6 arguments.
1295//
1296// Match an X86tcret that uses less than 7 volatile registers.
1297def : Pat<(X86tcret ptr_rc_tailcall:$dst, timm:$off),
1298          (TCRETURNri ptr_rc_tailcall:$dst, timm:$off)>,
1299          Requires<[Not64BitMode, NotUseIndirectThunkCalls]>;
1300
1301// FIXME: This is disabled for 32-bit PIC mode because the global base
1302// register which is part of the address mode may be assigned a
1303// callee-saved register.
1304// Similar to X86tcret_6regs, here we only have 1 register left
1305def : Pat<(X86tcret_1reg (load addr:$dst), timm:$off),
1306          (TCRETURNmi addr:$dst, timm:$off)>,
1307          Requires<[Not64BitMode, IsNotPIC, NotUseIndirectThunkCalls]>;
1308
1309def : Pat<(X86tcret (i32 tglobaladdr:$dst), timm:$off),
1310          (TCRETURNdi tglobaladdr:$dst, timm:$off)>,
1311          Requires<[NotLP64]>;
1312
1313def : Pat<(X86tcret (i32 texternalsym:$dst), timm:$off),
1314          (TCRETURNdi texternalsym:$dst, timm:$off)>,
1315          Requires<[NotLP64]>;
1316
1317def : Pat<(X86tcret ptr_rc_tailcall:$dst, timm:$off),
1318          (TCRETURNri64 ptr_rc_tailcall:$dst, timm:$off)>,
1319          Requires<[In64BitMode, NotUseIndirectThunkCalls]>;
1320
1321// Don't fold loads into X86tcret requiring more than 6 regs.
1322// There wouldn't be enough scratch registers for base+index.
1323def : Pat<(X86tcret_6regs (load addr:$dst), timm:$off),
1324          (TCRETURNmi64 addr:$dst, timm:$off)>,
1325          Requires<[In64BitMode, NotUseIndirectThunkCalls]>;
1326
1327def : Pat<(X86tcret ptr_rc_tailcall:$dst, timm:$off),
1328          (INDIRECT_THUNK_TCRETURN64 ptr_rc_tailcall:$dst, timm:$off)>,
1329          Requires<[In64BitMode, UseIndirectThunkCalls]>;
1330
1331def : Pat<(X86tcret ptr_rc_tailcall:$dst, timm:$off),
1332          (INDIRECT_THUNK_TCRETURN32 ptr_rc_tailcall:$dst, timm:$off)>,
1333          Requires<[Not64BitMode, UseIndirectThunkCalls]>;
1334
1335def : Pat<(X86tcret (i64 tglobaladdr:$dst), timm:$off),
1336          (TCRETURNdi64 tglobaladdr:$dst, timm:$off)>,
1337          Requires<[IsLP64]>;
1338
1339def : Pat<(X86tcret (i64 texternalsym:$dst), timm:$off),
1340          (TCRETURNdi64 texternalsym:$dst, timm:$off)>,
1341          Requires<[IsLP64]>;
1342
1343// Normal calls, with various flavors of addresses.
1344def : Pat<(X86call (i32 tglobaladdr:$dst)),
1345          (CALLpcrel32 tglobaladdr:$dst)>;
1346def : Pat<(X86call (i32 texternalsym:$dst)),
1347          (CALLpcrel32 texternalsym:$dst)>;
1348def : Pat<(X86call (i32 imm:$dst)),
1349          (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
1350
1351// Comparisons.
1352
1353// TEST R,R is smaller than CMP R,0
1354def : Pat<(X86cmp GR8:$src1, 0),
1355          (TEST8rr GR8:$src1, GR8:$src1)>;
1356def : Pat<(X86cmp GR16:$src1, 0),
1357          (TEST16rr GR16:$src1, GR16:$src1)>;
1358def : Pat<(X86cmp GR32:$src1, 0),
1359          (TEST32rr GR32:$src1, GR32:$src1)>;
1360def : Pat<(X86cmp GR64:$src1, 0),
1361          (TEST64rr GR64:$src1, GR64:$src1)>;
1362
1363// zextload bool -> zextload byte
1364// i1 stored in one byte in zero-extended form.
1365// Upper bits cleanup should be executed before Store.
1366def : Pat<(zextloadi8i1  addr:$src), (MOV8rm addr:$src)>;
1367def : Pat<(zextloadi16i1 addr:$src),
1368          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1369def : Pat<(zextloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1370def : Pat<(zextloadi64i1 addr:$src),
1371          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1372
1373// extload bool -> extload byte
1374// When extloading from 16-bit and smaller memory locations into 64-bit
1375// registers, use zero-extending loads so that the entire 64-bit register is
1376// defined, avoiding partial-register updates.
1377
1378def : Pat<(extloadi8i1 addr:$src),   (MOV8rm      addr:$src)>;
1379def : Pat<(extloadi16i1 addr:$src),
1380          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1381def : Pat<(extloadi32i1 addr:$src),  (MOVZX32rm8  addr:$src)>;
1382def : Pat<(extloadi16i8 addr:$src),
1383          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1384def : Pat<(extloadi32i8 addr:$src),  (MOVZX32rm8  addr:$src)>;
1385def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
1386
1387// For other extloads, use subregs, since the high contents of the register are
1388// defined after an extload.
1389// NOTE: The extloadi64i32 pattern needs to be first as it will try to form
1390// 32-bit loads for 4 byte aligned i8/i16 loads.
1391def : Pat<(extloadi64i32 addr:$src),
1392          (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src), sub_32bit)>;
1393def : Pat<(extloadi64i1 addr:$src),
1394          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1395def : Pat<(extloadi64i8 addr:$src),
1396          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1397def : Pat<(extloadi64i16 addr:$src),
1398          (SUBREG_TO_REG (i64 0), (MOVZX32rm16 addr:$src), sub_32bit)>;
1399
1400// anyext. Define these to do an explicit zero-extend to
1401// avoid partial-register updates.
1402def : Pat<(i16 (anyext GR8 :$src)), (EXTRACT_SUBREG
1403                                     (MOVZX32rr8 GR8 :$src), sub_16bit)>;
1404def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8  GR8 :$src)>;
1405
1406// Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
1407def : Pat<(i32 (anyext GR16:$src)),
1408          (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
1409
1410def : Pat<(i64 (anyext GR8 :$src)),
1411          (SUBREG_TO_REG (i64 0), (MOVZX32rr8  GR8  :$src), sub_32bit)>;
1412def : Pat<(i64 (anyext GR16:$src)),
1413          (SUBREG_TO_REG (i64 0), (MOVZX32rr16 GR16 :$src), sub_32bit)>;
1414def : Pat<(i64 (anyext GR32:$src)),
1415          (INSERT_SUBREG (i64 (IMPLICIT_DEF)), GR32:$src, sub_32bit)>;
1416
1417def : Pat<(i32 (anyext_sdiv GR8:$src)), (MOVSX32rr8 GR8:$src)>;
1418
1419// In the case of a 32-bit def that is known to implicitly zero-extend,
1420// we can use a SUBREG_TO_REG.
1421def : Pat<(i64 (zext def32:$src)),
1422          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1423def : Pat<(i64 (and (anyext def32:$src), 0x00000000FFFFFFFF)),
1424          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1425
1426//===----------------------------------------------------------------------===//
1427// Pattern match OR as ADD
1428//===----------------------------------------------------------------------===//
1429
1430// If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
1431// 3-addressified into an LEA instruction to avoid copies.  However, we also
1432// want to finally emit these instructions as an or at the end of the code
1433// generator to make the generated code easier to read.  To do this, we select
1434// into "disjoint bits" pseudo ops.
1435
1436// (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1437// Try this before the selecting to OR.
1438let SchedRW = [WriteALU] in {
1439
1440let isConvertibleToThreeAddress = 1, isPseudo = 1,
1441    Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1442let isCommutable = 1 in {
1443def ADD8rr_DB   : I<0, Pseudo, (outs GR8:$dst), (ins GR8:$src1, GR8:$src2),
1444                    "", // orb/addb REG, REG
1445                    [(set GR8:$dst, (or_is_add GR8:$src1, GR8:$src2))]>;
1446def ADD16rr_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1447                    "", // orw/addw REG, REG
1448                    [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1449def ADD32rr_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1450                    "", // orl/addl REG, REG
1451                    [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1452def ADD64rr_DB  : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1453                    "", // orq/addq REG, REG
1454                    [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1455} // isCommutable
1456
1457def ADD8ri_DB :   I<0, Pseudo,
1458                    (outs GR8:$dst), (ins GR8:$src1, i8imm:$src2),
1459                    "", // orb/addb REG, imm8
1460                    [(set GR8:$dst, (or_is_add GR8:$src1, imm:$src2))]>;
1461def ADD16ri_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1462                    "", // orw/addw REG, imm
1463                    [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1464def ADD32ri_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1465                    "", // orl/addl REG, imm
1466                    [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1467def ADD64ri32_DB : I<0, Pseudo,
1468                     (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1469                     "", // orq/addq REG, imm
1470                     [(set GR64:$dst, (or_is_add GR64:$src1,
1471                                                 i64immSExt32:$src2))]>;
1472}
1473} // AddedComplexity, SchedRW
1474
1475//===----------------------------------------------------------------------===//
1476// Pattern match XOR as ADD
1477//===----------------------------------------------------------------------===//
1478
1479// Prefer to pattern match XOR with min_signed_value as ADD at isel time.
1480// ADD can be 3-addressified into an LEA instruction to avoid copies.
1481let AddedComplexity = 5 in {
1482def : Pat<(xor GR8:$src1, -128),
1483          (ADD8ri GR8:$src1, -128)>;
1484def : Pat<(xor GR16:$src1, -32768),
1485          (ADD16ri GR16:$src1, -32768)>;
1486def : Pat<(xor GR32:$src1, -2147483648),
1487          (ADD32ri GR32:$src1, -2147483648)>;
1488}
1489
1490//===----------------------------------------------------------------------===//
1491// Some peepholes
1492//===----------------------------------------------------------------------===//
1493
1494// Odd encoding trick: -128 fits into an 8-bit immediate field while
1495// +128 doesn't, so in this special case use a sub instead of an add.
1496def : Pat<(add GR16:$src1, 128),
1497          (SUB16ri GR16:$src1, -128)>;
1498def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1499          (SUB16mi addr:$dst, -128)>;
1500
1501def : Pat<(add GR32:$src1, 128),
1502          (SUB32ri GR32:$src1, -128)>;
1503def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1504          (SUB32mi addr:$dst, -128)>;
1505
1506def : Pat<(add GR64:$src1, 128),
1507          (SUB64ri32 GR64:$src1, -128)>;
1508def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1509          (SUB64mi32 addr:$dst, -128)>;
1510
1511def : Pat<(X86add_flag_nocf GR16:$src1, 128),
1512          (SUB16ri GR16:$src1, -128)>;
1513def : Pat<(X86add_flag_nocf GR32:$src1, 128),
1514          (SUB32ri GR32:$src1, -128)>;
1515def : Pat<(X86add_flag_nocf GR64:$src1, 128),
1516          (SUB64ri32 GR64:$src1, -128)>;
1517
1518// Depositing value to 8/16 bit subreg:
1519def : Pat<(or (and GR64:$dst, -256),
1520              (i64 (zextloadi8 addr:$src))),
1521          (INSERT_SUBREG (i64 (COPY $dst)), (MOV8rm  i8mem:$src), sub_8bit)>;
1522
1523def : Pat<(or (and GR32:$dst, -256),
1524              (i32 (zextloadi8 addr:$src))),
1525          (INSERT_SUBREG (i32 (COPY $dst)), (MOV8rm  i8mem:$src), sub_8bit)>;
1526
1527def : Pat<(or (and GR64:$dst, -65536),
1528              (i64 (zextloadi16 addr:$src))),
1529          (INSERT_SUBREG (i64 (COPY $dst)), (MOV16rm  i16mem:$src), sub_16bit)>;
1530
1531def : Pat<(or (and GR32:$dst, -65536),
1532              (i32 (zextloadi16 addr:$src))),
1533          (INSERT_SUBREG (i32 (COPY $dst)), (MOV16rm  i16mem:$src), sub_16bit)>;
1534
1535// The same trick applies for 32-bit immediate fields in 64-bit
1536// instructions.
1537def : Pat<(add GR64:$src1, 0x0000000080000000),
1538          (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1539def : Pat<(store (add (loadi64 addr:$dst), 0x0000000080000000), addr:$dst),
1540          (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1541def : Pat<(X86add_flag_nocf GR64:$src1, 0x0000000080000000),
1542          (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1543
1544// To avoid needing to materialize an immediate in a register, use a 32-bit and
1545// with implicit zero-extension instead of a 64-bit and if the immediate has at
1546// least 32 bits of leading zeros. If in addition the last 32 bits can be
1547// represented with a sign extension of a 8 bit constant, use that.
1548// This can also reduce instruction size by eliminating the need for the REX
1549// prefix.
1550
1551// AddedComplexity is needed to give priority over i64immSExt8 and i64immSExt32.
1552let AddedComplexity = 1 in {
1553  let Predicates = [NoNDD] in {
1554    def : Pat<(and GR64:$src, i64immZExt32:$imm),
1555              (SUBREG_TO_REG
1556                (i64 0),
1557                (AND32ri
1558                  (EXTRACT_SUBREG GR64:$src, sub_32bit),
1559                  (i32 (GetLo32XForm imm:$imm))),
1560                sub_32bit)>;
1561  }
1562  let Predicates = [HasNDD] in {
1563    def : Pat<(and GR64:$src, i64immZExt32:$imm),
1564              (SUBREG_TO_REG
1565                (i64 0),
1566                (AND32ri_ND
1567                  (EXTRACT_SUBREG GR64:$src, sub_32bit),
1568                  (i32 (GetLo32XForm imm:$imm))),
1569                sub_32bit)>;
1570  }
1571} // AddedComplexity = 1
1572
1573
1574// AddedComplexity is needed due to the increased complexity on the
1575// i64immZExt32SExt8 and i64immZExt32 patterns above. Applying this to all
1576// the MOVZX patterns keeps thems together in DAGIsel tables.
1577let AddedComplexity = 1 in {
1578// r & (2^16-1) ==> movz
1579def : Pat<(and GR32:$src1, 0xffff),
1580          (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1581// r & (2^8-1) ==> movz
1582def : Pat<(and GR32:$src1, 0xff),
1583          (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>;
1584// r & (2^8-1) ==> movz
1585def : Pat<(and GR16:$src1, 0xff),
1586           (EXTRACT_SUBREG (MOVZX32rr8 (EXTRACT_SUBREG GR16:$src1, sub_8bit)),
1587             sub_16bit)>;
1588
1589// r & (2^32-1) ==> movz
1590def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1591          (SUBREG_TO_REG (i64 0),
1592                         (MOV32rr (EXTRACT_SUBREG GR64:$src, sub_32bit)),
1593                         sub_32bit)>;
1594// r & (2^16-1) ==> movz
1595def : Pat<(and GR64:$src, 0xffff),
1596          (SUBREG_TO_REG (i64 0),
1597                      (MOVZX32rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit))),
1598                      sub_32bit)>;
1599// r & (2^8-1) ==> movz
1600def : Pat<(and GR64:$src, 0xff),
1601          (SUBREG_TO_REG (i64 0),
1602                         (MOVZX32rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit))),
1603                         sub_32bit)>;
1604} // AddedComplexity = 1
1605
1606
1607// Try to use BTS/BTR/BTC for single bit operations on the upper 32-bits.
1608
1609def BTRXForm : SDNodeXForm<imm, [{
1610  // Transformation function: Find the lowest 0.
1611  return getI64Imm((uint8_t)N->getAPIntValue().countr_one(), SDLoc(N));
1612}]>;
1613
1614def BTCBTSXForm : SDNodeXForm<imm, [{
1615  // Transformation function: Find the lowest 1.
1616  return getI64Imm((uint8_t)N->getAPIntValue().countr_zero(), SDLoc(N));
1617}]>;
1618
1619def BTRMask64 : ImmLeaf<i64, [{
1620  return !isUInt<32>(Imm) && !isInt<32>(Imm) && isPowerOf2_64(~Imm);
1621}]>;
1622
1623def BTCBTSMask64 : ImmLeaf<i64, [{
1624  return !isInt<32>(Imm) && isPowerOf2_64(Imm);
1625}]>;
1626
1627// For now only do this for optsize.
1628let AddedComplexity = 1, Predicates=[OptForSize] in {
1629  def : Pat<(and GR64:$src1, BTRMask64:$mask),
1630            (BTR64ri8 GR64:$src1, (BTRXForm imm:$mask))>;
1631  def : Pat<(or GR64:$src1, BTCBTSMask64:$mask),
1632            (BTS64ri8 GR64:$src1, (BTCBTSXForm imm:$mask))>;
1633  def : Pat<(xor GR64:$src1, BTCBTSMask64:$mask),
1634            (BTC64ri8 GR64:$src1, (BTCBTSXForm imm:$mask))>;
1635}
1636
1637
1638// sext_inreg patterns
1639def : Pat<(sext_inreg GR32:$src, i16),
1640          (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1641def : Pat<(sext_inreg GR32:$src, i8),
1642          (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>;
1643
1644def : Pat<(sext_inreg GR16:$src, i8),
1645           (EXTRACT_SUBREG (MOVSX32rr8 (EXTRACT_SUBREG GR16:$src, sub_8bit)),
1646             sub_16bit)>;
1647
1648def : Pat<(sext_inreg GR64:$src, i32),
1649          (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1650def : Pat<(sext_inreg GR64:$src, i16),
1651          (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1652def : Pat<(sext_inreg GR64:$src, i8),
1653          (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1654
1655// sext, sext_load, zext, zext_load
1656def: Pat<(i16 (sext GR8:$src)),
1657          (EXTRACT_SUBREG (MOVSX32rr8 GR8:$src), sub_16bit)>;
1658def: Pat<(sextloadi16i8 addr:$src),
1659          (EXTRACT_SUBREG (MOVSX32rm8 addr:$src), sub_16bit)>;
1660def: Pat<(i16 (zext GR8:$src)),
1661          (EXTRACT_SUBREG (MOVZX32rr8 GR8:$src), sub_16bit)>;
1662def: Pat<(zextloadi16i8 addr:$src),
1663          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1664
1665// trunc patterns
1666def : Pat<(i16 (trunc GR32:$src)),
1667          (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1668def : Pat<(i8 (trunc GR32:$src)),
1669          (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1670                          sub_8bit)>,
1671      Requires<[Not64BitMode]>;
1672def : Pat<(i8 (trunc GR16:$src)),
1673          (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1674                          sub_8bit)>,
1675      Requires<[Not64BitMode]>;
1676def : Pat<(i32 (trunc GR64:$src)),
1677          (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1678def : Pat<(i16 (trunc GR64:$src)),
1679          (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1680def : Pat<(i8 (trunc GR64:$src)),
1681          (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1682def : Pat<(i8 (trunc GR32:$src)),
1683          (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1684      Requires<[In64BitMode]>;
1685def : Pat<(i8 (trunc GR16:$src)),
1686          (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1687      Requires<[In64BitMode]>;
1688
1689def immff00_ffff  : ImmLeaf<i32, [{
1690  return Imm >= 0xff00 && Imm <= 0xffff;
1691}]>;
1692
1693// h-register tricks
1694def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1695          (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)>,
1696      Requires<[Not64BitMode]>;
1697def : Pat<(i8 (trunc (srl_su (i32 (anyext GR16:$src)), (i8 8)))),
1698          (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)>,
1699      Requires<[Not64BitMode]>;
1700def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1701          (EXTRACT_SUBREG GR32:$src, sub_8bit_hi)>,
1702      Requires<[Not64BitMode]>;
1703def : Pat<(srl GR16:$src, (i8 8)),
1704          (EXTRACT_SUBREG
1705            (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1706            sub_16bit)>;
1707def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1708          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>;
1709def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1710          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>;
1711def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1712          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1713def : Pat<(srl (and_su GR32:$src, immff00_ffff), (i8 8)),
1714          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1715
1716// h-register tricks.
1717// For now, be conservative on x86-64 and use an h-register extract only if the
1718// value is immediately zero-extended or stored, which are somewhat common
1719// cases. This uses a bunch of code to prevent a register requiring a REX prefix
1720// from being allocated in the same instruction as the h register, as there's
1721// currently no way to describe this requirement to the register allocator.
1722
1723// h-register extract and zero-extend.
1724def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1725          (SUBREG_TO_REG
1726            (i64 0),
1727            (MOVZX32rr8_NOREX
1728              (EXTRACT_SUBREG GR64:$src, sub_8bit_hi)),
1729            sub_32bit)>;
1730def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1731          (SUBREG_TO_REG
1732            (i64 0),
1733            (MOVZX32rr8_NOREX
1734              (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1735            sub_32bit)>;
1736def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1737          (SUBREG_TO_REG
1738            (i64 0),
1739            (MOVZX32rr8_NOREX
1740              (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1741            sub_32bit)>;
1742
1743// h-register extract and store.
1744def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1745          (MOV8mr_NOREX
1746            addr:$dst,
1747            (EXTRACT_SUBREG GR64:$src, sub_8bit_hi))>;
1748def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1749          (MOV8mr_NOREX
1750            addr:$dst,
1751            (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>,
1752      Requires<[In64BitMode]>;
1753def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1754          (MOV8mr_NOREX
1755            addr:$dst,
1756            (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>,
1757      Requires<[In64BitMode]>;
1758
1759// Special pattern to catch the last step of __builtin_parity handling. Our
1760// goal is to use an xor of an h-register with the corresponding l-register.
1761// The above patterns would handle this on non 64-bit targets, but for 64-bit
1762// we need to be more careful. We're using a NOREX instruction here in case
1763// register allocation fails to keep the two registers together. So we need to
1764// make sure we can't accidentally mix R8-R15 with an h-register.
1765def : Pat<(X86xor_flag (i8 (trunc GR32:$src)),
1766                       (i8 (trunc (srl_su GR32:$src, (i8 8))))),
1767          (XOR8rr_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit),
1768                        (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1769
1770// (shl x, 1) ==> (add x, x)
1771// Note that if x is undef (immediate or otherwise), we could theoretically
1772// end up with the two uses of x getting different values, producing a result
1773// where the least significant bit is not 0. However, the probability of this
1774// happening is considered low enough that this is officially not a
1775// "real problem".
1776let Predicates = [NoNDD] in {
1777  def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr  GR8 :$src1, GR8 :$src1)>;
1778  def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1779  def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1780  def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1781}
1782let Predicates = [HasNDD] in {
1783  def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr_ND  GR8 :$src1, GR8 :$src1)>;
1784  def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr_ND GR16:$src1, GR16:$src1)>;
1785  def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr_ND GR32:$src1, GR32:$src1)>;
1786  def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr_ND GR64:$src1, GR64:$src1)>;
1787}
1788
1789// Shift amount is implicitly masked.
1790multiclass MaskedShiftAmountPats<SDNode frag> {
1791  // (shift x (and y, 31)) ==> (shift x, y)
1792  // (shift x (and y, 63)) ==> (shift x, y)
1793  let Predicates = [NoNDD] in {
1794    def : Pat<(frag GR8:$src1, (shiftMask32 CL)),
1795              (!cast<Instruction>(NAME # "8rCL") GR8:$src1)>;
1796    def : Pat<(frag GR16:$src1, (shiftMask32 CL)),
1797              (!cast<Instruction>(NAME # "16rCL") GR16:$src1)>;
1798    def : Pat<(frag GR32:$src1, (shiftMask32 CL)),
1799              (!cast<Instruction>(NAME # "32rCL") GR32:$src1)>;
1800    def : Pat<(frag GR64:$src1, (shiftMask64 CL)),
1801              (!cast<Instruction>(NAME # "64rCL") GR64:$src1)>;
1802  }
1803  let Predicates = [HasNDD] in {
1804    def : Pat<(frag GR8:$src1, (shiftMask32 CL)),
1805              (!cast<Instruction>(NAME # "8rCL_ND") GR8:$src1)>;
1806    def : Pat<(frag GR16:$src1, (shiftMask32 CL)),
1807              (!cast<Instruction>(NAME # "16rCL_ND") GR16:$src1)>;
1808    def : Pat<(frag GR32:$src1, (shiftMask32 CL)),
1809              (!cast<Instruction>(NAME # "32rCL_ND") GR32:$src1)>;
1810    def : Pat<(frag GR64:$src1, (shiftMask64 CL)),
1811              (!cast<Instruction>(NAME # "64rCL_ND") GR64:$src1)>;
1812  }
1813
1814  def : Pat<(store (frag (loadi8 addr:$dst), (shiftMask32 CL)), addr:$dst),
1815            (!cast<Instruction>(NAME # "8mCL") addr:$dst)>;
1816  def : Pat<(store (frag (loadi16 addr:$dst), (shiftMask32 CL)), addr:$dst),
1817            (!cast<Instruction>(NAME # "16mCL") addr:$dst)>;
1818  def : Pat<(store (frag (loadi32 addr:$dst), (shiftMask32 CL)), addr:$dst),
1819            (!cast<Instruction>(NAME # "32mCL") addr:$dst)>;
1820  def : Pat<(store (frag (loadi64 addr:$dst), (shiftMask64 CL)), addr:$dst),
1821            (!cast<Instruction>(NAME # "64mCL") addr:$dst)>;
1822
1823  let Predicates = [HasNDD] in {
1824    def : Pat<(frag (loadi8 addr:$src), (shiftMask32 CL)),
1825              (!cast<Instruction>(NAME # "8mCL_ND") addr:$src)>;
1826    def : Pat<(frag (loadi16 addr:$src), (shiftMask32 CL)),
1827              (!cast<Instruction>(NAME # "16mCL_ND") addr:$src)>;
1828    def : Pat<(frag (loadi32 addr:$src), (shiftMask32 CL)),
1829              (!cast<Instruction>(NAME # "32mCL_ND") addr:$src)>;
1830    def : Pat<(frag (loadi64 addr:$src), (shiftMask64 CL)),
1831              (!cast<Instruction>(NAME # "64mCL_ND") addr:$src)>;
1832  }
1833}
1834
1835defm SHL : MaskedShiftAmountPats<shl>;
1836defm SHR : MaskedShiftAmountPats<srl>;
1837defm SAR : MaskedShiftAmountPats<sra>;
1838
1839// ROL/ROR instructions allow a stronger mask optimization than shift for 8- and
1840// 16-bit. We can remove a mask of any (bitwidth - 1) on the rotation amount
1841// because over-rotating produces the same result. This is noted in the Intel
1842// docs with: "tempCOUNT <- (COUNT & COUNTMASK) MOD SIZE". Masking the rotation
1843// amount could affect EFLAGS results, but that does not matter because we are
1844// not tracking flags for these nodes.
1845multiclass MaskedRotateAmountPats<SDNode frag> {
1846  // (rot x (and y, BitWidth - 1)) ==> (rot x, y)
1847  let Predicates = [NoNDD] in {
1848    def : Pat<(frag GR8:$src1, (shiftMask8 CL)),
1849              (!cast<Instruction>(NAME # "8rCL") GR8:$src1)>;
1850    def : Pat<(frag GR16:$src1, (shiftMask16 CL)),
1851              (!cast<Instruction>(NAME # "16rCL") GR16:$src1)>;
1852    def : Pat<(frag GR32:$src1, (shiftMask32 CL)),
1853              (!cast<Instruction>(NAME # "32rCL") GR32:$src1)>;
1854    def : Pat<(frag GR64:$src1, (shiftMask64 CL)),
1855              (!cast<Instruction>(NAME # "64rCL") GR64:$src1)>;
1856  }
1857  let Predicates = [HasNDD] in {
1858    def : Pat<(frag GR8:$src1, (shiftMask8 CL)),
1859              (!cast<Instruction>(NAME # "8rCL_ND") GR8:$src1)>;
1860    def : Pat<(frag GR16:$src1, (shiftMask16 CL)),
1861              (!cast<Instruction>(NAME # "16rCL_ND") GR16:$src1)>;
1862    def : Pat<(frag GR32:$src1, (shiftMask32 CL)),
1863              (!cast<Instruction>(NAME # "32rCL_ND") GR32:$src1)>;
1864    def : Pat<(frag GR64:$src1, (shiftMask64 CL)),
1865              (!cast<Instruction>(NAME # "64rCL_ND") GR64:$src1)>;
1866  }
1867
1868  def : Pat<(store (frag (loadi8 addr:$dst), (shiftMask8 CL)), addr:$dst),
1869            (!cast<Instruction>(NAME # "8mCL") addr:$dst)>;
1870  def : Pat<(store (frag (loadi16 addr:$dst), (shiftMask16 CL)), addr:$dst),
1871            (!cast<Instruction>(NAME # "16mCL") addr:$dst)>;
1872  def : Pat<(store (frag (loadi32 addr:$dst), (shiftMask32 CL)), addr:$dst),
1873            (!cast<Instruction>(NAME # "32mCL") addr:$dst)>;
1874  def : Pat<(store (frag (loadi64 addr:$dst), (shiftMask64 CL)), addr:$dst),
1875            (!cast<Instruction>(NAME # "64mCL") addr:$dst)>;
1876
1877  let Predicates = [HasNDD] in {
1878    def : Pat<(frag (loadi8 addr:$src), (shiftMask8 CL)),
1879              (!cast<Instruction>(NAME # "8mCL_ND") addr:$src)>;
1880    def : Pat<(frag (loadi16 addr:$src), (shiftMask16 CL)),
1881              (!cast<Instruction>(NAME # "16mCL_ND") addr:$src)>;
1882    def : Pat<(frag (loadi32 addr:$src), (shiftMask32 CL)),
1883              (!cast<Instruction>(NAME # "32mCL_ND") addr:$src)>;
1884    def : Pat<(frag (loadi64 addr:$src), (shiftMask64 CL)),
1885              (!cast<Instruction>(NAME # "64mCL_ND") addr:$src)>;
1886  }
1887}
1888
1889defm ROL : MaskedRotateAmountPats<rotl>;
1890defm ROR : MaskedRotateAmountPats<rotr>;
1891
1892multiclass MaskedShlrdAmountPats<string suffix, Predicate p> {
1893  let Predicates = [p] in {
1894    // Double "funnel" shift amount is implicitly masked.
1895    // (fshl/fshr x (and y, 31)) ==> (fshl/fshr x, y) (NOTE: modulo32)
1896    def : Pat<(X86fshl GR16:$src1, GR16:$src2, (shiftMask32 CL)),
1897              (!cast<Instruction>(SHLD16rrCL#suffix) GR16:$src1, GR16:$src2)>;
1898    def : Pat<(X86fshr GR16:$src2, GR16:$src1, (shiftMask32 CL)),
1899              (!cast<Instruction>(SHRD16rrCL#suffix) GR16:$src1, GR16:$src2)>;
1900
1901    // (fshl/fshr x (and y, 31)) ==> (fshl/fshr x, y)
1902    def : Pat<(fshl GR32:$src1, GR32:$src2, (shiftMask32 CL)),
1903              (!cast<Instruction>(SHLD32rrCL#suffix) GR32:$src1, GR32:$src2)>;
1904    def : Pat<(fshr GR32:$src2, GR32:$src1, (shiftMask32 CL)),
1905              (!cast<Instruction>(SHRD32rrCL#suffix) GR32:$src1, GR32:$src2)>;
1906
1907    // (fshl/fshr x (and y, 63)) ==> (fshl/fshr x, y)
1908    def : Pat<(fshl GR64:$src1, GR64:$src2, (shiftMask64 CL)),
1909              (!cast<Instruction>(SHLD64rrCL#suffix) GR64:$src1, GR64:$src2)>;
1910    def : Pat<(fshr GR64:$src2, GR64:$src1, (shiftMask64 CL)),
1911              (!cast<Instruction>(SHRD64rrCL#suffix) GR64:$src1, GR64:$src2)>;
1912  }
1913}
1914
1915defm : MaskedShlrdAmountPats<"", NoNDD>;
1916defm : MaskedShlrdAmountPats<"_ND", HasNDD>;
1917
1918// Use BTR/BTS/BTC for clearing/setting/toggling a bit in a variable location.
1919multiclass OneBitPats<RegisterClass rc, ValueType vt, Instruction btr,
1920                      Instruction bts, Instruction btc, PatFrag mask> {
1921  def : Pat<(and rc:$src1, (rotl -2, GR8:$src2)),
1922            (btr rc:$src1,
1923                 (INSERT_SUBREG (vt (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1924  def : Pat<(or rc:$src1, (shl 1, GR8:$src2)),
1925            (bts rc:$src1,
1926                 (INSERT_SUBREG (vt (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1927  def : Pat<(xor rc:$src1, (shl 1, GR8:$src2)),
1928            (btc rc:$src1,
1929                 (INSERT_SUBREG (vt (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1930
1931  // Similar to above, but removing unneeded masking of the shift amount.
1932  def : Pat<(and rc:$src1, (rotl -2, (mask GR8:$src2))),
1933            (btr rc:$src1,
1934                 (INSERT_SUBREG (vt (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1935  def : Pat<(or rc:$src1, (shl 1, (mask GR8:$src2))),
1936            (bts rc:$src1,
1937                (INSERT_SUBREG (vt (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1938  def : Pat<(xor rc:$src1, (shl 1, (mask GR8:$src2))),
1939            (btc rc:$src1,
1940                (INSERT_SUBREG (vt (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1941}
1942
1943defm : OneBitPats<GR16, i16, BTR16rr, BTS16rr, BTC16rr, shiftMask16>;
1944defm : OneBitPats<GR32, i32, BTR32rr, BTS32rr, BTC32rr, shiftMask32>;
1945defm : OneBitPats<GR64, i64, BTR64rr, BTS64rr, BTC64rr, shiftMask64>;
1946
1947//===----------------------------------------------------------------------===//
1948// EFLAGS-defining Patterns
1949//===----------------------------------------------------------------------===//
1950
1951multiclass EFLAGSDefiningPats<string suffix, Predicate p> {
1952  let Predicates = [p] in {
1953    // add reg, reg
1954    def : Pat<(add GR8 :$src1, GR8 :$src2), (!cast<Instruction>(ADD8rr#suffix) GR8 :$src1, GR8 :$src2)>;
1955    def : Pat<(add GR16:$src1, GR16:$src2), (!cast<Instruction>(ADD16rr#suffix) GR16:$src1, GR16:$src2)>;
1956    def : Pat<(add GR32:$src1, GR32:$src2), (!cast<Instruction>(ADD32rr#suffix) GR32:$src1, GR32:$src2)>;
1957    def : Pat<(add GR64:$src1, GR64:$src2), (!cast<Instruction>(ADD64rr#suffix) GR64:$src1, GR64:$src2)>;
1958
1959    // add reg, mem
1960    def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
1961              (!cast<Instruction>(ADD8rm#suffix) GR8:$src1, addr:$src2)>;
1962    def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
1963              (!cast<Instruction>(ADD16rm#suffix) GR16:$src1, addr:$src2)>;
1964    def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
1965              (!cast<Instruction>(ADD32rm#suffix) GR32:$src1, addr:$src2)>;
1966    def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
1967              (!cast<Instruction>(ADD64rm#suffix) GR64:$src1, addr:$src2)>;
1968
1969    // add reg, imm
1970    def : Pat<(add GR8 :$src1, imm:$src2), (!cast<Instruction>(ADD8ri#suffix) GR8:$src1 , imm:$src2)>;
1971    def : Pat<(add GR16:$src1, imm:$src2), (!cast<Instruction>(ADD16ri#suffix) GR16:$src1, imm:$src2)>;
1972    def : Pat<(add GR32:$src1, imm:$src2), (!cast<Instruction>(ADD32ri#suffix) GR32:$src1, imm:$src2)>;
1973    def : Pat<(add GR64:$src1, i64immSExt32:$src2), (!cast<Instruction>(ADD64ri32#suffix) GR64:$src1, i64immSExt32:$src2)>;
1974
1975    // sub reg, reg
1976    def : Pat<(sub GR8 :$src1, GR8 :$src2), (!cast<Instruction>(SUB8rr#suffix)  GR8 :$src1, GR8 :$src2)>;
1977    def : Pat<(sub GR16:$src1, GR16:$src2), (!cast<Instruction>(SUB16rr#suffix) GR16:$src1, GR16:$src2)>;
1978    def : Pat<(sub GR32:$src1, GR32:$src2), (!cast<Instruction>(SUB32rr#suffix) GR32:$src1, GR32:$src2)>;
1979    def : Pat<(sub GR64:$src1, GR64:$src2), (!cast<Instruction>(SUB64rr#suffix) GR64:$src1, GR64:$src2)>;
1980
1981    // sub reg, mem
1982    def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
1983              (!cast<Instruction>(SUB8rm#suffix) GR8:$src1, addr:$src2)>;
1984    def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
1985              (!cast<Instruction>(SUB16rm#suffix) GR16:$src1, addr:$src2)>;
1986    def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
1987              (!cast<Instruction>(SUB32rm#suffix) GR32:$src1, addr:$src2)>;
1988    def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
1989              (!cast<Instruction>(SUB64rm#suffix) GR64:$src1, addr:$src2)>;
1990
1991    // sub reg, imm
1992    def : Pat<(sub GR8:$src1, imm:$src2),
1993              (!cast<Instruction>(SUB8ri#suffix) GR8:$src1, imm:$src2)>;
1994    def : Pat<(sub GR16:$src1, imm:$src2),
1995              (!cast<Instruction>(SUB16ri#suffix) GR16:$src1, imm:$src2)>;
1996    def : Pat<(sub GR32:$src1, imm:$src2),
1997              (!cast<Instruction>(SUB32ri#suffix) GR32:$src1, imm:$src2)>;
1998    def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
1999              (!cast<Instruction>(SUB64ri32#suffix) GR64:$src1, i64immSExt32:$src2)>;
2000
2001    // sub 0, reg
2002    def : Pat<(X86sub_flag 0, GR8 :$src), (!cast<Instruction>(NEG8r#suffix)  GR8 :$src)>;
2003    def : Pat<(X86sub_flag 0, GR16:$src), (!cast<Instruction>(NEG16r#suffix) GR16:$src)>;
2004    def : Pat<(X86sub_flag 0, GR32:$src), (!cast<Instruction>(NEG32r#suffix) GR32:$src)>;
2005    def : Pat<(X86sub_flag 0, GR64:$src), (!cast<Instruction>(NEG64r#suffix) GR64:$src)>;
2006
2007    // mul reg, reg
2008    def : Pat<(mul GR16:$src1, GR16:$src2),
2009              (!cast<Instruction>(IMUL16rr#suffix) GR16:$src1, GR16:$src2)>;
2010    def : Pat<(mul GR32:$src1, GR32:$src2),
2011              (!cast<Instruction>(IMUL32rr#suffix) GR32:$src1, GR32:$src2)>;
2012    def : Pat<(mul GR64:$src1, GR64:$src2),
2013              (!cast<Instruction>(IMUL64rr#suffix) GR64:$src1, GR64:$src2)>;
2014
2015    // mul reg, mem
2016    def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
2017              (!cast<Instruction>(IMUL16rm#suffix) GR16:$src1, addr:$src2)>;
2018    def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
2019              (!cast<Instruction>(IMUL32rm#suffix) GR32:$src1, addr:$src2)>;
2020    def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
2021              (!cast<Instruction>(IMUL64rm#suffix) GR64:$src1, addr:$src2)>;
2022
2023    // or reg/reg.
2024    def : Pat<(or GR8 :$src1, GR8 :$src2), (!cast<Instruction>(OR8rr#suffix)  GR8 :$src1, GR8 :$src2)>;
2025    def : Pat<(or GR16:$src1, GR16:$src2), (!cast<Instruction>(OR16rr#suffix) GR16:$src1, GR16:$src2)>;
2026    def : Pat<(or GR32:$src1, GR32:$src2), (!cast<Instruction>(OR32rr#suffix) GR32:$src1, GR32:$src2)>;
2027    def : Pat<(or GR64:$src1, GR64:$src2), (!cast<Instruction>(OR64rr#suffix) GR64:$src1, GR64:$src2)>;
2028
2029    // or reg/mem
2030    def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
2031              (!cast<Instruction>(OR8rm#suffix) GR8:$src1, addr:$src2)>;
2032    def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
2033              (!cast<Instruction>(OR16rm#suffix) GR16:$src1, addr:$src2)>;
2034    def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
2035              (!cast<Instruction>(OR32rm#suffix) GR32:$src1, addr:$src2)>;
2036    def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
2037              (!cast<Instruction>(OR64rm#suffix) GR64:$src1, addr:$src2)>;
2038
2039    // or reg/imm
2040    def : Pat<(or GR8:$src1 , imm:$src2), (!cast<Instruction>(OR8ri#suffix)  GR8 :$src1, imm:$src2)>;
2041    def : Pat<(or GR16:$src1, imm:$src2), (!cast<Instruction>(OR16ri#suffix) GR16:$src1, imm:$src2)>;
2042    def : Pat<(or GR32:$src1, imm:$src2), (!cast<Instruction>(OR32ri#suffix) GR32:$src1, imm:$src2)>;
2043    def : Pat<(or GR64:$src1, i64immSExt32:$src2),
2044              (!cast<Instruction>(OR64ri32#suffix) GR64:$src1, i64immSExt32:$src2)>;
2045
2046    // xor reg/reg
2047    def : Pat<(xor GR8 :$src1, GR8 :$src2), (!cast<Instruction>(XOR8rr#suffix)  GR8 :$src1, GR8 :$src2)>;
2048    def : Pat<(xor GR16:$src1, GR16:$src2), (!cast<Instruction>(XOR16rr#suffix) GR16:$src1, GR16:$src2)>;
2049    def : Pat<(xor GR32:$src1, GR32:$src2), (!cast<Instruction>(XOR32rr#suffix) GR32:$src1, GR32:$src2)>;
2050    def : Pat<(xor GR64:$src1, GR64:$src2), (!cast<Instruction>(XOR64rr#suffix) GR64:$src1, GR64:$src2)>;
2051
2052    // xor reg/mem
2053    def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
2054              (!cast<Instruction>(XOR8rm#suffix) GR8:$src1, addr:$src2)>;
2055    def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
2056              (!cast<Instruction>(XOR16rm#suffix) GR16:$src1, addr:$src2)>;
2057    def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
2058              (!cast<Instruction>(XOR32rm#suffix) GR32:$src1, addr:$src2)>;
2059    def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
2060              (!cast<Instruction>(XOR64rm#suffix) GR64:$src1, addr:$src2)>;
2061
2062    // xor reg/imm
2063    def : Pat<(xor GR8:$src1, imm:$src2),
2064              (!cast<Instruction>(XOR8ri#suffix) GR8:$src1, imm:$src2)>;
2065    def : Pat<(xor GR16:$src1, imm:$src2),
2066              (!cast<Instruction>(XOR16ri#suffix) GR16:$src1, imm:$src2)>;
2067    def : Pat<(xor GR32:$src1, imm:$src2),
2068              (!cast<Instruction>(XOR32ri#suffix) GR32:$src1, imm:$src2)>;
2069    def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
2070              (!cast<Instruction>(XOR64ri32#suffix) GR64:$src1, i64immSExt32:$src2)>;
2071
2072    // and reg/reg
2073    def : Pat<(and GR8 :$src1, GR8 :$src2), (!cast<Instruction>(AND8rr#suffix)  GR8 :$src1, GR8 :$src2)>;
2074    def : Pat<(and GR16:$src1, GR16:$src2), (!cast<Instruction>(AND16rr#suffix) GR16:$src1, GR16:$src2)>;
2075    def : Pat<(and GR32:$src1, GR32:$src2), (!cast<Instruction>(AND32rr#suffix) GR32:$src1, GR32:$src2)>;
2076    def : Pat<(and GR64:$src1, GR64:$src2), (!cast<Instruction>(AND64rr#suffix) GR64:$src1, GR64:$src2)>;
2077
2078    // and reg/mem
2079    def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
2080              (!cast<Instruction>(AND8rm#suffix) GR8:$src1, addr:$src2)>;
2081    def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
2082              (!cast<Instruction>(AND16rm#suffix) GR16:$src1, addr:$src2)>;
2083    def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
2084              (!cast<Instruction>(AND32rm#suffix) GR32:$src1, addr:$src2)>;
2085    def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
2086              (!cast<Instruction>(AND64rm#suffix) GR64:$src1, addr:$src2)>;
2087
2088    // and reg/imm
2089    def : Pat<(and GR8:$src1, imm:$src2),
2090              (!cast<Instruction>(AND8ri#suffix) GR8:$src1, imm:$src2)>;
2091    def : Pat<(and GR16:$src1, imm:$src2),
2092              (!cast<Instruction>(AND16ri#suffix) GR16:$src1, imm:$src2)>;
2093    def : Pat<(and GR32:$src1, imm:$src2),
2094              (!cast<Instruction>(AND32ri#suffix) GR32:$src1, imm:$src2)>;
2095    def : Pat<(and GR64:$src1, i64immSExt32:$src2),
2096              (!cast<Instruction>(AND64ri32#suffix) GR64:$src1, i64immSExt32:$src2)>;
2097  }
2098
2099  // Increment/Decrement reg.
2100  // Do not make INC/DEC if it is slow
2101  let Predicates = [UseIncDec, p] in {
2102    def : Pat<(add GR8:$src, 1),   (!cast<Instruction>(INC8r#suffix) GR8:$src)>;
2103    def : Pat<(add GR16:$src, 1),  (!cast<Instruction>(INC16r#suffix) GR16:$src)>;
2104    def : Pat<(add GR32:$src, 1),  (!cast<Instruction>(INC32r#suffix) GR32:$src)>;
2105    def : Pat<(add GR64:$src, 1),  (!cast<Instruction>(INC64r#suffix) GR64:$src)>;
2106    def : Pat<(add GR8:$src, -1),  (!cast<Instruction>(DEC8r#suffix) GR8:$src)>;
2107    def : Pat<(add GR16:$src, -1), (!cast<Instruction>(DEC16r#suffix) GR16:$src)>;
2108    def : Pat<(add GR32:$src, -1), (!cast<Instruction>(DEC32r#suffix) GR32:$src)>;
2109    def : Pat<(add GR64:$src, -1), (!cast<Instruction>(DEC64r#suffix) GR64:$src)>;
2110
2111    def : Pat<(X86add_flag_nocf GR8:$src, -1),  (!cast<Instruction>(DEC8r#suffix) GR8:$src)>;
2112    def : Pat<(X86add_flag_nocf GR16:$src, -1), (!cast<Instruction>(DEC16r#suffix) GR16:$src)>;
2113    def : Pat<(X86add_flag_nocf GR32:$src, -1), (!cast<Instruction>(DEC32r#suffix) GR32:$src)>;
2114    def : Pat<(X86add_flag_nocf GR64:$src, -1), (!cast<Instruction>(DEC64r#suffix) GR64:$src)>;
2115    def : Pat<(X86sub_flag_nocf GR8:$src, -1),  (!cast<Instruction>(INC8r#suffix) GR8:$src)>;
2116    def : Pat<(X86sub_flag_nocf GR16:$src, -1), (!cast<Instruction>(INC16r#suffix) GR16:$src)>;
2117    def : Pat<(X86sub_flag_nocf GR32:$src, -1), (!cast<Instruction>(INC32r#suffix) GR32:$src)>;
2118    def : Pat<(X86sub_flag_nocf GR64:$src, -1), (!cast<Instruction>(INC64r#suffix) GR64:$src)>;
2119  }
2120}
2121
2122defm : EFLAGSDefiningPats<"", NoNDD>;
2123defm : EFLAGSDefiningPats<"_ND", HasNDD>;
2124
2125// mul reg, imm
2126def : Pat<(mul GR16:$src1, imm:$src2),
2127          (IMUL16rri GR16:$src1, imm:$src2)>;
2128def : Pat<(mul GR32:$src1, imm:$src2),
2129          (IMUL32rri GR32:$src1, imm:$src2)>;
2130def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
2131          (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
2132
2133// reg = mul mem, imm
2134def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
2135          (IMUL16rmi addr:$src1, imm:$src2)>;
2136def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
2137          (IMUL32rmi addr:$src1, imm:$src2)>;
2138def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
2139          (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
2140
2141// Bit scan instruction patterns to match explicit zero-undef behavior.
2142def : Pat<(cttz_zero_undef GR16:$src), (BSF16rr GR16:$src)>;
2143def : Pat<(cttz_zero_undef GR32:$src), (BSF32rr GR32:$src)>;
2144def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr GR64:$src)>;
2145def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm addr:$src)>;
2146def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm addr:$src)>;
2147def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm addr:$src)>;
2148
2149// When HasMOVBE is enabled it is possible to get a non-legalized
2150// register-register 16 bit bswap. This maps it to a ROL instruction.
2151let Predicates = [HasMOVBE] in {
2152 def : Pat<(bswap GR16:$src), (ROL16ri GR16:$src, (i8 8))>;
2153}
2154