xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86InstrCompiler.td (revision ba3c1f5972d7b90feb6e6da47905ff2757e0fe57)
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:$src2))]>,
778                          LOCK;
779}
780
781}
782
783defm LOCK_ADD : LOCK_ArithBinOp<0x00, 0x80, 0x83, MRM0m, X86lock_add, "add">;
784defm LOCK_SUB : LOCK_ArithBinOp<0x28, 0x80, 0x83, MRM5m, X86lock_sub, "sub">;
785defm LOCK_OR  : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM1m, X86lock_or , "or">;
786defm LOCK_AND : LOCK_ArithBinOp<0x20, 0x80, 0x83, MRM4m, X86lock_and, "and">;
787defm LOCK_XOR : LOCK_ArithBinOp<0x30, 0x80, 0x83, MRM6m, X86lock_xor, "xor">;
788
789def X86lock_add_nocf : PatFrag<(ops node:$lhs, node:$rhs),
790                               (X86lock_add node:$lhs, node:$rhs), [{
791  return hasNoCarryFlagUses(SDValue(N, 0));
792}]>;
793
794def X86lock_sub_nocf : PatFrag<(ops node:$lhs, node:$rhs),
795                               (X86lock_sub node:$lhs, node:$rhs), [{
796  return hasNoCarryFlagUses(SDValue(N, 0));
797}]>;
798
799let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
800    SchedRW = [WriteALURMW]  in {
801  let Predicates = [UseIncDec] in {
802    def LOCK_INC8m  : I<0xFE, MRM0m, (outs), (ins i8mem :$dst),
803                        "inc{b}\t$dst",
804                        [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i8 1)))]>,
805                        LOCK;
806    def LOCK_INC16m : I<0xFF, MRM0m, (outs), (ins i16mem:$dst),
807                        "inc{w}\t$dst",
808                        [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i16 1)))]>,
809                        OpSize16, LOCK;
810    def LOCK_INC32m : I<0xFF, MRM0m, (outs), (ins i32mem:$dst),
811                        "inc{l}\t$dst",
812                        [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i32 1)))]>,
813                        OpSize32, LOCK;
814
815    def LOCK_DEC8m  : I<0xFE, MRM1m, (outs), (ins i8mem :$dst),
816                        "dec{b}\t$dst",
817                        [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i8 1)))]>,
818                        LOCK;
819    def LOCK_DEC16m : I<0xFF, MRM1m, (outs), (ins i16mem:$dst),
820                        "dec{w}\t$dst",
821                        [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i16 1)))]>,
822                        OpSize16, LOCK;
823    def LOCK_DEC32m : I<0xFF, MRM1m, (outs), (ins i32mem:$dst),
824                        "dec{l}\t$dst",
825                        [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i32 1)))]>,
826                        OpSize32, LOCK;
827  }
828
829  let Predicates = [UseIncDec, In64BitMode] in {
830    def LOCK_INC64m : RI<0xFF, MRM0m, (outs), (ins i64mem:$dst),
831                         "inc{q}\t$dst",
832                         [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i64 1)))]>,
833                         LOCK;
834    def LOCK_DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst),
835                         "dec{q}\t$dst",
836                         [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i64 1)))]>,
837                         LOCK;
838  }
839}
840
841let Predicates = [UseIncDec] in {
842  // Additional patterns for -1 constant.
843  def : Pat<(X86lock_add addr:$dst, (i8  -1)), (LOCK_DEC8m  addr:$dst)>;
844  def : Pat<(X86lock_add addr:$dst, (i16 -1)), (LOCK_DEC16m addr:$dst)>;
845  def : Pat<(X86lock_add addr:$dst, (i32 -1)), (LOCK_DEC32m addr:$dst)>;
846  def : Pat<(X86lock_sub addr:$dst, (i8  -1)), (LOCK_INC8m  addr:$dst)>;
847  def : Pat<(X86lock_sub addr:$dst, (i16 -1)), (LOCK_INC16m addr:$dst)>;
848  def : Pat<(X86lock_sub addr:$dst, (i32 -1)), (LOCK_INC32m addr:$dst)>;
849}
850
851let Predicates = [UseIncDec, In64BitMode] in {
852  // Additional patterns for -1 constant.
853  def : Pat<(X86lock_add addr:$dst, (i64 -1)), (LOCK_DEC64m addr:$dst)>;
854  def : Pat<(X86lock_sub addr:$dst, (i64 -1)), (LOCK_INC64m addr:$dst)>;
855}
856
857// Atomic bit test.
858def X86LBTest : SDTypeProfile<1, 3, [SDTCisVT<0, i32>, SDTCisPtrTy<1>,
859                                     SDTCisVT<2, i8>, SDTCisVT<3, i32>]>;
860def x86bts : SDNode<"X86ISD::LBTS", X86LBTest,
861                    [SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPMemOperand]>;
862def x86btc : SDNode<"X86ISD::LBTC", X86LBTest,
863                    [SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPMemOperand]>;
864def x86btr : SDNode<"X86ISD::LBTR", X86LBTest,
865                    [SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPMemOperand]>;
866
867def X86LBTestRM : SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<1>,
868                                       SDTCisInt<2>]>;
869
870def x86_rm_bts : SDNode<"X86ISD::LBTS_RM", X86LBTestRM,
871                        [SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPMemOperand]>;
872def x86_rm_btc : SDNode<"X86ISD::LBTC_RM", X86LBTestRM,
873                        [SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPMemOperand]>;
874def x86_rm_btr : SDNode<"X86ISD::LBTR_RM", X86LBTestRM,
875                        [SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPMemOperand]>;
876
877
878multiclass ATOMIC_LOGIC_OP<Format Form, string s> {
879  let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
880      SchedRW = [WriteBitTestSetRegRMW]  in {
881    def 16m : Ii8<0xBA, Form, (outs), (ins i16mem:$src1, i8imm:$src2),
882                  !strconcat(s, "{w}\t{$src2, $src1|$src1, $src2}"),
883                  [(set EFLAGS, (!cast<SDNode>("x86" # s) addr:$src1, timm:$src2, (i32 16)))]>,
884              OpSize16, TB, LOCK;
885    def 32m : Ii8<0xBA, Form, (outs), (ins i32mem:$src1, i8imm:$src2),
886                  !strconcat(s, "{l}\t{$src2, $src1|$src1, $src2}"),
887                  [(set EFLAGS, (!cast<SDNode>("x86" # s) addr:$src1, timm:$src2, (i32 32)))]>,
888              OpSize32, TB, LOCK;
889    def 64m : RIi8<0xBA, Form, (outs), (ins i64mem:$src1, i8imm:$src2),
890                   !strconcat(s, "{q}\t{$src2, $src1|$src1, $src2}"),
891                   [(set EFLAGS, (!cast<SDNode>("x86" # s) addr:$src1, timm:$src2, (i32 64)))]>,
892              TB, LOCK;
893  }
894}
895
896multiclass ATOMIC_LOGIC_OP_RM<bits<8> Opc8, string s> {
897  let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
898      SchedRW = [WriteBitTestSetRegRMW]  in {
899    def 16rm : I<Opc8, MRMDestMem, (outs), (ins i16mem:$src1, GR16:$src2),
900                  !strconcat(s, "{w}\t{$src2, $src1|$src1, $src2}"),
901                  [(set EFLAGS, (!cast<SDNode>("x86_rm_" # s) addr:$src1, GR16:$src2))]>,
902               OpSize16, TB, LOCK;
903    def 32rm : I<Opc8, MRMDestMem, (outs), (ins i32mem:$src1, GR32:$src2),
904                  !strconcat(s, "{l}\t{$src2, $src1|$src1, $src2}"),
905                  [(set EFLAGS, (!cast<SDNode>("x86_rm_" # s) addr:$src1, GR32:$src2))]>,
906               OpSize32, TB, LOCK;
907    def 64rm : RI<Opc8, MRMDestMem, (outs), (ins i64mem:$src1, GR64:$src2),
908                   !strconcat(s, "{q}\t{$src2, $src1|$src1, $src2}"),
909                   [(set EFLAGS, (!cast<SDNode>("x86_rm_" # s) addr:$src1, GR64:$src2))]>,
910               TB, LOCK;
911  }
912}
913
914
915defm LOCK_BTS : ATOMIC_LOGIC_OP<MRM5m, "bts">;
916defm LOCK_BTC : ATOMIC_LOGIC_OP<MRM7m, "btc">;
917defm LOCK_BTR : ATOMIC_LOGIC_OP<MRM6m, "btr">;
918
919defm LOCK_BTS_RM : ATOMIC_LOGIC_OP_RM<0xAB, "bts">;
920defm LOCK_BTC_RM : ATOMIC_LOGIC_OP_RM<0xBB, "btc">;
921defm LOCK_BTR_RM : ATOMIC_LOGIC_OP_RM<0xB3, "btr">;
922
923// Atomic compare and swap.
924multiclass LCMPXCHG_BinOp<bits<8> Opc8, bits<8> Opc, Format Form,
925                          string mnemonic, SDPatternOperator frag> {
926let isCodeGenOnly = 1, SchedRW = [WriteCMPXCHGRMW] in {
927  let Defs = [AL, EFLAGS], Uses = [AL] in
928  def NAME#8  : I<Opc8, Form, (outs), (ins i8mem:$ptr, GR8:$swap),
929                  !strconcat(mnemonic, "{b}\t{$swap, $ptr|$ptr, $swap}"),
930                  [(frag addr:$ptr, GR8:$swap, 1)]>, TB, LOCK;
931  let Defs = [AX, EFLAGS], Uses = [AX] in
932  def NAME#16 : I<Opc, Form, (outs), (ins i16mem:$ptr, GR16:$swap),
933                  !strconcat(mnemonic, "{w}\t{$swap, $ptr|$ptr, $swap}"),
934                  [(frag addr:$ptr, GR16:$swap, 2)]>, TB, OpSize16, LOCK;
935  let Defs = [EAX, EFLAGS], Uses = [EAX] in
936  def NAME#32 : I<Opc, Form, (outs), (ins i32mem:$ptr, GR32:$swap),
937                  !strconcat(mnemonic, "{l}\t{$swap, $ptr|$ptr, $swap}"),
938                  [(frag addr:$ptr, GR32:$swap, 4)]>, TB, OpSize32, LOCK;
939  let Defs = [RAX, EFLAGS], Uses = [RAX] in
940  def NAME#64 : RI<Opc, Form, (outs), (ins i64mem:$ptr, GR64:$swap),
941                   !strconcat(mnemonic, "{q}\t{$swap, $ptr|$ptr, $swap}"),
942                   [(frag addr:$ptr, GR64:$swap, 8)]>, TB, LOCK;
943}
944}
945
946let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
947    Predicates = [HasCX8], SchedRW = [WriteCMPXCHGRMW],
948    isCodeGenOnly = 1, usesCustomInserter = 1 in {
949def LCMPXCHG8B : I<0xC7, MRM1m, (outs), (ins i64mem:$ptr),
950                   "cmpxchg8b\t$ptr",
951                   [(X86cas8 addr:$ptr)]>, TB, LOCK;
952}
953
954let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RBX, RCX, RDX],
955    Predicates = [HasCX16,In64BitMode], SchedRW = [WriteCMPXCHGRMW],
956    isCodeGenOnly = 1, mayLoad = 1, mayStore = 1, hasSideEffects = 0 in {
957def LCMPXCHG16B : RI<0xC7, MRM1m, (outs), (ins i128mem:$ptr),
958                     "cmpxchg16b\t$ptr",
959                     []>, TB, LOCK;
960}
961
962// This pseudo must be used when the frame uses RBX as
963// the base pointer. Indeed, in such situation RBX is a reserved
964// register and the register allocator will ignore any use/def of
965// it. In other words, the register will not fix the clobbering of
966// RBX that will happen when setting the arguments for the instrucion.
967//
968// Unlike the actual related instruction, we mark that this one
969// defines RBX (instead of using RBX).
970// The rationale is that we will define RBX during the expansion of
971// the pseudo. The argument feeding RBX is rbx_input.
972//
973// The additional argument, $rbx_save, is a temporary register used to
974// save the value of RBX across the actual instruction.
975//
976// To make sure the register assigned to $rbx_save does not interfere with
977// the definition of the actual instruction, we use a definition $dst which
978// is tied to $rbx_save. That way, the live-range of $rbx_save spans across
979// the instruction and we are sure we will have a valid register to restore
980// the value of RBX.
981let Defs = [RAX, RDX, RBX, EFLAGS], Uses = [RAX, RCX, RDX],
982    Predicates = [HasCX16,In64BitMode], SchedRW = [WriteCMPXCHGRMW],
983    isCodeGenOnly = 1, isPseudo = 1,
984    mayLoad = 1, mayStore = 1, hasSideEffects = 0,
985    Constraints = "$rbx_save = $dst" in {
986def LCMPXCHG16B_SAVE_RBX :
987    I<0, Pseudo, (outs GR64:$dst),
988      (ins i128mem:$ptr, GR64:$rbx_input, GR64:$rbx_save), "", []>;
989}
990
991// Pseudo instruction that doesn't read/write RBX. Will be turned into either
992// LCMPXCHG16B_SAVE_RBX or LCMPXCHG16B via a custom inserter.
993let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RCX, RDX],
994    Predicates = [HasCX16,In64BitMode], SchedRW = [WriteCMPXCHGRMW],
995    isCodeGenOnly = 1, isPseudo = 1,
996    mayLoad = 1, mayStore = 1, hasSideEffects = 0,
997    usesCustomInserter = 1 in {
998def LCMPXCHG16B_NO_RBX :
999    I<0, Pseudo, (outs), (ins i128mem:$ptr, GR64:$rbx_input), "",
1000      [(X86cas16 addr:$ptr, GR64:$rbx_input)]>;
1001}
1002
1003// This pseudo must be used when the frame uses RBX/EBX as
1004// the base pointer.
1005// cf comment for LCMPXCHG16B_SAVE_RBX.
1006let Defs = [EBX], Uses = [ECX, EAX],
1007    Predicates = [HasMWAITX], SchedRW = [WriteSystem],
1008    isCodeGenOnly = 1, isPseudo = 1, Constraints = "$rbx_save = $dst" in {
1009def MWAITX_SAVE_RBX :
1010    I<0, Pseudo, (outs GR64:$dst),
1011      (ins GR32:$ebx_input, GR64:$rbx_save),
1012      "mwaitx",
1013      []>;
1014}
1015
1016// Pseudo mwaitx instruction to use for custom insertion.
1017let Predicates = [HasMWAITX], SchedRW = [WriteSystem],
1018    isCodeGenOnly = 1, isPseudo = 1,
1019    usesCustomInserter = 1 in {
1020def MWAITX :
1021    I<0, Pseudo, (outs), (ins GR32:$ecx, GR32:$eax, GR32:$ebx),
1022      "mwaitx",
1023      [(int_x86_mwaitx GR32:$ecx, GR32:$eax, GR32:$ebx)]>;
1024}
1025
1026
1027defm LCMPXCHG : LCMPXCHG_BinOp<0xB0, 0xB1, MRMDestMem, "cmpxchg", X86cas>;
1028
1029// Atomic exchange and add
1030multiclass ATOMIC_RMW_BINOP<bits<8> opc8, bits<8> opc, string mnemonic,
1031                            string frag> {
1032  let Constraints = "$val = $dst", Defs = [EFLAGS], mayLoad = 1, mayStore = 1,
1033      isCodeGenOnly = 1, SchedRW = [WriteALURMW] in {
1034    def NAME#8  : I<opc8, MRMSrcMem, (outs GR8:$dst),
1035                    (ins GR8:$val, i8mem:$ptr),
1036                    !strconcat(mnemonic, "{b}\t{$val, $ptr|$ptr, $val}"),
1037                    [(set GR8:$dst,
1038                          (!cast<PatFrag>(frag # "_8") addr:$ptr, GR8:$val))]>;
1039    def NAME#16 : I<opc, MRMSrcMem, (outs GR16:$dst),
1040                    (ins GR16:$val, i16mem:$ptr),
1041                    !strconcat(mnemonic, "{w}\t{$val, $ptr|$ptr, $val}"),
1042                    [(set
1043                       GR16:$dst,
1044                       (!cast<PatFrag>(frag # "_16") addr:$ptr, GR16:$val))]>,
1045                    OpSize16;
1046    def NAME#32 : I<opc, MRMSrcMem, (outs GR32:$dst),
1047                    (ins GR32:$val, i32mem:$ptr),
1048                    !strconcat(mnemonic, "{l}\t{$val, $ptr|$ptr, $val}"),
1049                    [(set
1050                       GR32:$dst,
1051                       (!cast<PatFrag>(frag # "_32") addr:$ptr, GR32:$val))]>,
1052                    OpSize32;
1053    def NAME#64 : RI<opc, MRMSrcMem, (outs GR64:$dst),
1054                     (ins GR64:$val, i64mem:$ptr),
1055                     !strconcat(mnemonic, "{q}\t{$val, $ptr|$ptr, $val}"),
1056                     [(set
1057                        GR64:$dst,
1058                        (!cast<PatFrag>(frag # "_64") addr:$ptr, GR64:$val))]>;
1059  }
1060}
1061
1062defm LXADD : ATOMIC_RMW_BINOP<0xc0, 0xc1, "xadd", "atomic_load_add">, TB, LOCK;
1063
1064/* The following multiclass tries to make sure that in code like
1065 *    x.store (immediate op x.load(acquire), release)
1066 * and
1067 *    x.store (register op x.load(acquire), release)
1068 * an operation directly on memory is generated instead of wasting a register.
1069 * It is not automatic as atomic_store/load are only lowered to MOV instructions
1070 * extremely late to prevent them from being accidentally reordered in the backend
1071 * (see below the RELEASE_MOV* / ACQUIRE_MOV* pseudo-instructions)
1072 */
1073multiclass RELEASE_BINOP_MI<string Name, SDNode op> {
1074  def : Pat<(atomic_store_8 addr:$dst,
1075             (op (atomic_load_8 addr:$dst), (i8 imm:$src))),
1076            (!cast<Instruction>(Name#"8mi") addr:$dst, imm:$src)>;
1077  def : Pat<(atomic_store_16 addr:$dst,
1078             (op (atomic_load_16 addr:$dst), (i16 imm:$src))),
1079            (!cast<Instruction>(Name#"16mi") addr:$dst, imm:$src)>;
1080  def : Pat<(atomic_store_32 addr:$dst,
1081             (op (atomic_load_32 addr:$dst), (i32 imm:$src))),
1082            (!cast<Instruction>(Name#"32mi") addr:$dst, imm:$src)>;
1083  def : Pat<(atomic_store_64 addr:$dst,
1084             (op (atomic_load_64 addr:$dst), (i64immSExt32:$src))),
1085            (!cast<Instruction>(Name#"64mi32") addr:$dst, (i64immSExt32:$src))>;
1086
1087  def : Pat<(atomic_store_8 addr:$dst,
1088             (op (atomic_load_8 addr:$dst), (i8 GR8:$src))),
1089            (!cast<Instruction>(Name#"8mr") addr:$dst, GR8:$src)>;
1090  def : Pat<(atomic_store_16 addr:$dst,
1091             (op (atomic_load_16 addr:$dst), (i16 GR16:$src))),
1092            (!cast<Instruction>(Name#"16mr") addr:$dst, GR16:$src)>;
1093  def : Pat<(atomic_store_32 addr:$dst,
1094             (op (atomic_load_32 addr:$dst), (i32 GR32:$src))),
1095            (!cast<Instruction>(Name#"32mr") addr:$dst, GR32:$src)>;
1096  def : Pat<(atomic_store_64 addr:$dst,
1097             (op (atomic_load_64 addr:$dst), (i64 GR64:$src))),
1098            (!cast<Instruction>(Name#"64mr") addr:$dst, GR64:$src)>;
1099}
1100defm : RELEASE_BINOP_MI<"ADD", add>;
1101defm : RELEASE_BINOP_MI<"AND", and>;
1102defm : RELEASE_BINOP_MI<"OR",  or>;
1103defm : RELEASE_BINOP_MI<"XOR", xor>;
1104defm : RELEASE_BINOP_MI<"SUB", sub>;
1105
1106// Atomic load + floating point patterns.
1107// FIXME: This could also handle SIMD operations with *ps and *pd instructions.
1108multiclass ATOMIC_LOAD_FP_BINOP_MI<string Name, SDNode op> {
1109  def : Pat<(op FR32:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
1110            (!cast<Instruction>(Name#"SSrm") FR32:$src1, addr:$src2)>,
1111            Requires<[UseSSE1]>;
1112  def : Pat<(op FR32:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
1113            (!cast<Instruction>("V"#Name#"SSrm") FR32:$src1, addr:$src2)>,
1114            Requires<[UseAVX]>;
1115  def : Pat<(op FR32X:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
1116            (!cast<Instruction>("V"#Name#"SSZrm") FR32X:$src1, addr:$src2)>,
1117            Requires<[HasAVX512]>;
1118
1119  def : Pat<(op FR64:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
1120            (!cast<Instruction>(Name#"SDrm") FR64:$src1, addr:$src2)>,
1121            Requires<[UseSSE1]>;
1122  def : Pat<(op FR64:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
1123            (!cast<Instruction>("V"#Name#"SDrm") FR64:$src1, addr:$src2)>,
1124            Requires<[UseAVX]>;
1125  def : Pat<(op FR64X:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
1126            (!cast<Instruction>("V"#Name#"SDZrm") FR64X:$src1, addr:$src2)>,
1127            Requires<[HasAVX512]>;
1128}
1129defm : ATOMIC_LOAD_FP_BINOP_MI<"ADD", fadd>;
1130// FIXME: Add fsub, fmul, fdiv, ...
1131
1132multiclass RELEASE_UNOP<string Name, dag dag8, dag dag16, dag dag32,
1133                        dag dag64> {
1134  def : Pat<(atomic_store_8 addr:$dst, dag8),
1135            (!cast<Instruction>(Name#8m) addr:$dst)>;
1136  def : Pat<(atomic_store_16 addr:$dst, dag16),
1137            (!cast<Instruction>(Name#16m) addr:$dst)>;
1138  def : Pat<(atomic_store_32 addr:$dst, dag32),
1139            (!cast<Instruction>(Name#32m) addr:$dst)>;
1140  def : Pat<(atomic_store_64 addr:$dst, dag64),
1141            (!cast<Instruction>(Name#64m) addr:$dst)>;
1142}
1143
1144let Predicates = [UseIncDec] in {
1145  defm : RELEASE_UNOP<"INC",
1146      (add (atomic_load_8  addr:$dst), (i8 1)),
1147      (add (atomic_load_16 addr:$dst), (i16 1)),
1148      (add (atomic_load_32 addr:$dst), (i32 1)),
1149      (add (atomic_load_64 addr:$dst), (i64 1))>;
1150  defm : RELEASE_UNOP<"DEC",
1151      (add (atomic_load_8  addr:$dst), (i8 -1)),
1152      (add (atomic_load_16 addr:$dst), (i16 -1)),
1153      (add (atomic_load_32 addr:$dst), (i32 -1)),
1154      (add (atomic_load_64 addr:$dst), (i64 -1))>;
1155}
1156
1157defm : RELEASE_UNOP<"NEG",
1158    (ineg (i8 (atomic_load_8  addr:$dst))),
1159    (ineg (i16 (atomic_load_16 addr:$dst))),
1160    (ineg (i32 (atomic_load_32 addr:$dst))),
1161    (ineg (i64 (atomic_load_64 addr:$dst)))>;
1162defm : RELEASE_UNOP<"NOT",
1163    (not (i8 (atomic_load_8  addr:$dst))),
1164    (not (i16 (atomic_load_16 addr:$dst))),
1165    (not (i32 (atomic_load_32 addr:$dst))),
1166    (not (i64 (atomic_load_64 addr:$dst)))>;
1167
1168def : Pat<(atomic_store_8 addr:$dst, (i8 imm:$src)),
1169          (MOV8mi addr:$dst, imm:$src)>;
1170def : Pat<(atomic_store_16 addr:$dst, (i16 imm:$src)),
1171          (MOV16mi addr:$dst, imm:$src)>;
1172def : Pat<(atomic_store_32 addr:$dst, (i32 imm:$src)),
1173          (MOV32mi addr:$dst, imm:$src)>;
1174def : Pat<(atomic_store_64 addr:$dst, (i64immSExt32:$src)),
1175          (MOV64mi32 addr:$dst, i64immSExt32:$src)>;
1176
1177def : Pat<(atomic_store_8 addr:$dst, GR8:$src),
1178          (MOV8mr addr:$dst, GR8:$src)>;
1179def : Pat<(atomic_store_16 addr:$dst, GR16:$src),
1180          (MOV16mr addr:$dst, GR16:$src)>;
1181def : Pat<(atomic_store_32 addr:$dst, GR32:$src),
1182          (MOV32mr addr:$dst, GR32:$src)>;
1183def : Pat<(atomic_store_64 addr:$dst, GR64:$src),
1184          (MOV64mr addr:$dst, GR64:$src)>;
1185
1186def : Pat<(i8  (atomic_load_8 addr:$src)),  (MOV8rm addr:$src)>;
1187def : Pat<(i16 (atomic_load_16 addr:$src)), (MOV16rm addr:$src)>;
1188def : Pat<(i32 (atomic_load_32 addr:$src)), (MOV32rm addr:$src)>;
1189def : Pat<(i64 (atomic_load_64 addr:$src)), (MOV64rm addr:$src)>;
1190
1191// Floating point loads/stores.
1192def : Pat<(atomic_store_32 addr:$dst, (i32 (bitconvert (f32 FR32:$src)))),
1193          (MOVSSmr addr:$dst, FR32:$src)>, Requires<[UseSSE1]>;
1194def : Pat<(atomic_store_32 addr:$dst, (i32 (bitconvert (f32 FR32:$src)))),
1195          (VMOVSSmr addr:$dst, FR32:$src)>, Requires<[UseAVX]>;
1196def : Pat<(atomic_store_32 addr:$dst, (i32 (bitconvert (f32 FR32:$src)))),
1197          (VMOVSSZmr addr:$dst, FR32:$src)>, Requires<[HasAVX512]>;
1198
1199def : Pat<(atomic_store_64 addr:$dst, (i64 (bitconvert (f64 FR64:$src)))),
1200          (MOVSDmr addr:$dst, FR64:$src)>, Requires<[UseSSE2]>;
1201def : Pat<(atomic_store_64 addr:$dst, (i64 (bitconvert (f64 FR64:$src)))),
1202          (VMOVSDmr addr:$dst, FR64:$src)>, Requires<[UseAVX]>;
1203def : Pat<(atomic_store_64 addr:$dst, (i64 (bitconvert (f64 FR64:$src)))),
1204          (VMOVSDmr addr:$dst, FR64:$src)>, Requires<[HasAVX512]>;
1205
1206def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1207          (MOVSSrm_alt addr:$src)>, Requires<[UseSSE1]>;
1208def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1209          (VMOVSSrm_alt addr:$src)>, Requires<[UseAVX]>;
1210def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1211          (VMOVSSZrm_alt addr:$src)>, Requires<[HasAVX512]>;
1212
1213def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1214          (MOVSDrm_alt addr:$src)>, Requires<[UseSSE2]>;
1215def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1216          (VMOVSDrm_alt addr:$src)>, Requires<[UseAVX]>;
1217def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1218          (VMOVSDZrm_alt addr:$src)>, Requires<[HasAVX512]>;
1219
1220//===----------------------------------------------------------------------===//
1221// DAG Pattern Matching Rules
1222//===----------------------------------------------------------------------===//
1223
1224// Use AND/OR to store 0/-1 in memory when optimizing for minsize. This saves
1225// binary size compared to a regular MOV, but it introduces an unnecessary
1226// load, so is not suitable for regular or optsize functions.
1227let Predicates = [OptForMinSize] in {
1228def : Pat<(simple_store (i16 0), addr:$dst), (AND16mi8 addr:$dst, 0)>;
1229def : Pat<(simple_store (i32 0), addr:$dst), (AND32mi8 addr:$dst, 0)>;
1230def : Pat<(simple_store (i64 0), addr:$dst), (AND64mi8 addr:$dst, 0)>;
1231def : Pat<(simple_store (i16 -1), addr:$dst), (OR16mi8 addr:$dst, -1)>;
1232def : Pat<(simple_store (i32 -1), addr:$dst), (OR32mi8 addr:$dst, -1)>;
1233def : Pat<(simple_store (i64 -1), addr:$dst), (OR64mi8 addr:$dst, -1)>;
1234}
1235
1236// In kernel code model, we can get the address of a label
1237// into a register with 'movq'.  FIXME: This is a hack, the 'imm' predicate of
1238// the MOV64ri32 should accept these.
1239def : Pat<(i64 (X86Wrapper tconstpool  :$dst)),
1240          (MOV64ri32 tconstpool  :$dst)>, Requires<[KernelCode]>;
1241def : Pat<(i64 (X86Wrapper tjumptable  :$dst)),
1242          (MOV64ri32 tjumptable  :$dst)>, Requires<[KernelCode]>;
1243def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
1244          (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
1245def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
1246          (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
1247def : Pat<(i64 (X86Wrapper mcsym:$dst)),
1248          (MOV64ri32 mcsym:$dst)>, Requires<[KernelCode]>;
1249def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
1250          (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
1251
1252// If we have small model and -static mode, it is safe to store global addresses
1253// directly as immediates.  FIXME: This is really a hack, the 'imm' predicate
1254// for MOV64mi32 should handle this sort of thing.
1255def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
1256          (MOV64mi32 addr:$dst, tconstpool:$src)>,
1257          Requires<[NearData, IsNotPIC]>;
1258def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
1259          (MOV64mi32 addr:$dst, tjumptable:$src)>,
1260          Requires<[NearData, IsNotPIC]>;
1261def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
1262          (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
1263          Requires<[NearData, IsNotPIC]>;
1264def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
1265          (MOV64mi32 addr:$dst, texternalsym:$src)>,
1266          Requires<[NearData, IsNotPIC]>;
1267def : Pat<(store (i64 (X86Wrapper mcsym:$src)), addr:$dst),
1268          (MOV64mi32 addr:$dst, mcsym:$src)>,
1269          Requires<[NearData, IsNotPIC]>;
1270def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
1271          (MOV64mi32 addr:$dst, tblockaddress:$src)>,
1272          Requires<[NearData, IsNotPIC]>;
1273
1274def : Pat<(i32 (X86RecoverFrameAlloc mcsym:$dst)), (MOV32ri mcsym:$dst)>;
1275def : Pat<(i64 (X86RecoverFrameAlloc mcsym:$dst)), (MOV64ri mcsym:$dst)>;
1276
1277// Calls
1278
1279// tls has some funny stuff here...
1280// This corresponds to movabs $foo@tpoff, %rax
1281def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
1282          (MOV64ri32 tglobaltlsaddr :$dst)>;
1283// This corresponds to add $foo@tpoff, %rax
1284def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
1285          (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
1286
1287
1288// Direct PC relative function call for small code model. 32-bit displacement
1289// sign extended to 64-bit.
1290def : Pat<(X86call (i64 tglobaladdr:$dst)),
1291          (CALL64pcrel32 tglobaladdr:$dst)>;
1292def : Pat<(X86call (i64 texternalsym:$dst)),
1293          (CALL64pcrel32 texternalsym:$dst)>;
1294
1295def : Pat<(X86call_rvmarker (i64 tglobaladdr:$rvfunc), (i64 texternalsym:$dst)),
1296          (CALL64pcrel32_RVMARKER tglobaladdr:$rvfunc, texternalsym:$dst)>;
1297def : Pat<(X86call_rvmarker (i64 tglobaladdr:$rvfunc), (i64 tglobaladdr:$dst)),
1298          (CALL64pcrel32_RVMARKER tglobaladdr:$rvfunc, tglobaladdr:$dst)>;
1299
1300
1301// Tailcall stuff. The TCRETURN instructions execute after the epilog, so they
1302// can never use callee-saved registers. That is the purpose of the GR64_TC
1303// register classes.
1304//
1305// The only volatile register that is never used by the calling convention is
1306// %r11. This happens when calling a vararg function with 6 arguments.
1307//
1308// Match an X86tcret that uses less than 7 volatile registers.
1309def X86tcret_6regs : PatFrag<(ops node:$ptr, node:$off),
1310                             (X86tcret node:$ptr, node:$off), [{
1311  // X86tcret args: (*chain, ptr, imm, regs..., glue)
1312  unsigned NumRegs = 0;
1313  for (unsigned i = 3, e = N->getNumOperands(); i != e; ++i)
1314    if (isa<RegisterSDNode>(N->getOperand(i)) && ++NumRegs > 6)
1315      return false;
1316  return true;
1317}]>;
1318
1319def X86tcret_1reg : PatFrag<(ops node:$ptr, node:$off),
1320                             (X86tcret node:$ptr, node:$off), [{
1321  // X86tcret args: (*chain, ptr, imm, regs..., glue)
1322  unsigned NumRegs = 1;
1323  const SDValue& BasePtr = cast<LoadSDNode>(N->getOperand(1))->getBasePtr();
1324  if (isa<FrameIndexSDNode>(BasePtr))
1325    NumRegs = 3;
1326  else if (BasePtr->getNumOperands() && isa<GlobalAddressSDNode>(BasePtr->getOperand(0)))
1327    NumRegs = 3;
1328  for (unsigned i = 3, e = N->getNumOperands(); i != e; ++i)
1329    if (isa<RegisterSDNode>(N->getOperand(i)) && ( NumRegs-- == 0))
1330      return false;
1331  return true;
1332}]>;
1333
1334def : Pat<(X86tcret ptr_rc_tailcall:$dst, timm:$off),
1335          (TCRETURNri ptr_rc_tailcall:$dst, timm:$off)>,
1336          Requires<[Not64BitMode, NotUseIndirectThunkCalls]>;
1337
1338// FIXME: This is disabled for 32-bit PIC mode because the global base
1339// register which is part of the address mode may be assigned a
1340// callee-saved register.
1341// Similar to X86tcret_6regs, here we only have 1 register left
1342def : Pat<(X86tcret_1reg (load addr:$dst), timm:$off),
1343          (TCRETURNmi addr:$dst, timm:$off)>,
1344          Requires<[Not64BitMode, IsNotPIC, NotUseIndirectThunkCalls]>;
1345
1346def : Pat<(X86tcret (i32 tglobaladdr:$dst), timm:$off),
1347          (TCRETURNdi tglobaladdr:$dst, timm:$off)>,
1348          Requires<[NotLP64]>;
1349
1350def : Pat<(X86tcret (i32 texternalsym:$dst), timm:$off),
1351          (TCRETURNdi texternalsym:$dst, timm:$off)>,
1352          Requires<[NotLP64]>;
1353
1354def : Pat<(X86tcret ptr_rc_tailcall:$dst, timm:$off),
1355          (TCRETURNri64 ptr_rc_tailcall:$dst, timm:$off)>,
1356          Requires<[In64BitMode, NotUseIndirectThunkCalls]>;
1357
1358// Don't fold loads into X86tcret requiring more than 6 regs.
1359// There wouldn't be enough scratch registers for base+index.
1360def : Pat<(X86tcret_6regs (load addr:$dst), timm:$off),
1361          (TCRETURNmi64 addr:$dst, timm:$off)>,
1362          Requires<[In64BitMode, NotUseIndirectThunkCalls]>;
1363
1364def : Pat<(X86tcret ptr_rc_tailcall:$dst, timm:$off),
1365          (INDIRECT_THUNK_TCRETURN64 ptr_rc_tailcall:$dst, timm:$off)>,
1366          Requires<[In64BitMode, UseIndirectThunkCalls]>;
1367
1368def : Pat<(X86tcret ptr_rc_tailcall:$dst, timm:$off),
1369          (INDIRECT_THUNK_TCRETURN32 ptr_rc_tailcall:$dst, timm:$off)>,
1370          Requires<[Not64BitMode, UseIndirectThunkCalls]>;
1371
1372def : Pat<(X86tcret (i64 tglobaladdr:$dst), timm:$off),
1373          (TCRETURNdi64 tglobaladdr:$dst, timm:$off)>,
1374          Requires<[IsLP64]>;
1375
1376def : Pat<(X86tcret (i64 texternalsym:$dst), timm:$off),
1377          (TCRETURNdi64 texternalsym:$dst, timm:$off)>,
1378          Requires<[IsLP64]>;
1379
1380// Normal calls, with various flavors of addresses.
1381def : Pat<(X86call (i32 tglobaladdr:$dst)),
1382          (CALLpcrel32 tglobaladdr:$dst)>;
1383def : Pat<(X86call (i32 texternalsym:$dst)),
1384          (CALLpcrel32 texternalsym:$dst)>;
1385def : Pat<(X86call (i32 imm:$dst)),
1386          (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
1387
1388// Comparisons.
1389
1390// TEST R,R is smaller than CMP R,0
1391def : Pat<(X86cmp GR8:$src1, 0),
1392          (TEST8rr GR8:$src1, GR8:$src1)>;
1393def : Pat<(X86cmp GR16:$src1, 0),
1394          (TEST16rr GR16:$src1, GR16:$src1)>;
1395def : Pat<(X86cmp GR32:$src1, 0),
1396          (TEST32rr GR32:$src1, GR32:$src1)>;
1397def : Pat<(X86cmp GR64:$src1, 0),
1398          (TEST64rr GR64:$src1, GR64:$src1)>;
1399
1400// zextload bool -> zextload byte
1401// i1 stored in one byte in zero-extended form.
1402// Upper bits cleanup should be executed before Store.
1403def : Pat<(zextloadi8i1  addr:$src), (MOV8rm addr:$src)>;
1404def : Pat<(zextloadi16i1 addr:$src),
1405          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1406def : Pat<(zextloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1407def : Pat<(zextloadi64i1 addr:$src),
1408          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1409
1410// extload bool -> extload byte
1411// When extloading from 16-bit and smaller memory locations into 64-bit
1412// registers, use zero-extending loads so that the entire 64-bit register is
1413// defined, avoiding partial-register updates.
1414
1415def : Pat<(extloadi8i1 addr:$src),   (MOV8rm      addr:$src)>;
1416def : Pat<(extloadi16i1 addr:$src),
1417          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1418def : Pat<(extloadi32i1 addr:$src),  (MOVZX32rm8  addr:$src)>;
1419def : Pat<(extloadi16i8 addr:$src),
1420          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1421def : Pat<(extloadi32i8 addr:$src),  (MOVZX32rm8  addr:$src)>;
1422def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
1423
1424// For other extloads, use subregs, since the high contents of the register are
1425// defined after an extload.
1426// NOTE: The extloadi64i32 pattern needs to be first as it will try to form
1427// 32-bit loads for 4 byte aligned i8/i16 loads.
1428def : Pat<(extloadi64i32 addr:$src),
1429          (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src), sub_32bit)>;
1430def : Pat<(extloadi64i1 addr:$src),
1431          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1432def : Pat<(extloadi64i8 addr:$src),
1433          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1434def : Pat<(extloadi64i16 addr:$src),
1435          (SUBREG_TO_REG (i64 0), (MOVZX32rm16 addr:$src), sub_32bit)>;
1436
1437// anyext. Define these to do an explicit zero-extend to
1438// avoid partial-register updates.
1439def : Pat<(i16 (anyext GR8 :$src)), (EXTRACT_SUBREG
1440                                     (MOVZX32rr8 GR8 :$src), sub_16bit)>;
1441def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8  GR8 :$src)>;
1442
1443// Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
1444def : Pat<(i32 (anyext GR16:$src)),
1445          (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
1446
1447def : Pat<(i64 (anyext GR8 :$src)),
1448          (SUBREG_TO_REG (i64 0), (MOVZX32rr8  GR8  :$src), sub_32bit)>;
1449def : Pat<(i64 (anyext GR16:$src)),
1450          (SUBREG_TO_REG (i64 0), (MOVZX32rr16 GR16 :$src), sub_32bit)>;
1451def : Pat<(i64 (anyext GR32:$src)),
1452          (INSERT_SUBREG (i64 (IMPLICIT_DEF)), GR32:$src, sub_32bit)>;
1453
1454// If this is an anyext of the remainder of an 8-bit sdivrem, use a MOVSX
1455// instead of a MOVZX. The sdivrem lowering will emit emit a MOVSX to move
1456// %ah to the lower byte of a register. By using a MOVSX here we allow a
1457// post-isel peephole to merge the two MOVSX instructions into one.
1458def anyext_sdiv : PatFrag<(ops node:$lhs), (anyext node:$lhs),[{
1459  return (N->getOperand(0).getOpcode() == ISD::SDIVREM &&
1460          N->getOperand(0).getResNo() == 1);
1461}]>;
1462def : Pat<(i32 (anyext_sdiv GR8:$src)), (MOVSX32rr8 GR8:$src)>;
1463
1464// Any instruction that defines a 32-bit result leaves the high half of the
1465// register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
1466// be copying from a truncate. AssertSext/AssertZext/AssertAlign aren't saying
1467// anything about the upper 32 bits, they're probably just qualifying a
1468// CopyFromReg. FREEZE may be coming from a a truncate. Any other 32-bit
1469// operation will zero-extend up to 64 bits.
1470def def32 : PatLeaf<(i32 GR32:$src), [{
1471  return N->getOpcode() != ISD::TRUNCATE &&
1472         N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
1473         N->getOpcode() != ISD::CopyFromReg &&
1474         N->getOpcode() != ISD::AssertSext &&
1475         N->getOpcode() != ISD::AssertZext &&
1476         N->getOpcode() != ISD::AssertAlign &&
1477         N->getOpcode() != ISD::FREEZE;
1478}]>;
1479
1480// In the case of a 32-bit def that is known to implicitly zero-extend,
1481// we can use a SUBREG_TO_REG.
1482def : Pat<(i64 (zext def32:$src)),
1483          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1484def : Pat<(i64 (and (anyext def32:$src), 0x00000000FFFFFFFF)),
1485          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1486
1487//===----------------------------------------------------------------------===//
1488// Pattern match OR as ADD
1489//===----------------------------------------------------------------------===//
1490
1491// If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
1492// 3-addressified into an LEA instruction to avoid copies.  However, we also
1493// want to finally emit these instructions as an or at the end of the code
1494// generator to make the generated code easier to read.  To do this, we select
1495// into "disjoint bits" pseudo ops.
1496
1497// Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.
1498def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
1499  if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1500    return CurDAG->MaskedValueIsZero(N->getOperand(0), CN->getAPIntValue());
1501
1502  KnownBits Known0 = CurDAG->computeKnownBits(N->getOperand(0), 0);
1503  KnownBits Known1 = CurDAG->computeKnownBits(N->getOperand(1), 0);
1504  return (~Known0.Zero & ~Known1.Zero) == 0;
1505}]>;
1506
1507
1508// (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1509// Try this before the selecting to OR.
1510let SchedRW = [WriteALU] in {
1511
1512let isConvertibleToThreeAddress = 1, isPseudo = 1,
1513    Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1514let isCommutable = 1 in {
1515def ADD8rr_DB   : I<0, Pseudo, (outs GR8:$dst), (ins GR8:$src1, GR8:$src2),
1516                    "", // orb/addb REG, REG
1517                    [(set GR8:$dst, (or_is_add GR8:$src1, GR8:$src2))]>;
1518def ADD16rr_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1519                    "", // orw/addw REG, REG
1520                    [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1521def ADD32rr_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1522                    "", // orl/addl REG, REG
1523                    [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1524def ADD64rr_DB  : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1525                    "", // orq/addq REG, REG
1526                    [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1527} // isCommutable
1528
1529// NOTE: These are order specific, we want the ri8 forms to be listed
1530// first so that they are slightly preferred to the ri forms.
1531
1532def ADD8ri_DB :   I<0, Pseudo,
1533                    (outs GR8:$dst), (ins GR8:$src1, i8imm:$src2),
1534                    "", // orb/addb REG, imm8
1535                    [(set GR8:$dst, (or_is_add GR8:$src1, imm:$src2))]>;
1536def ADD16ri8_DB : I<0, Pseudo,
1537                    (outs GR16:$dst), (ins GR16:$src1, i16i8imm:$src2),
1538                    "", // orw/addw REG, imm8
1539                    [(set GR16:$dst,(or_is_add GR16:$src1,i16immSExt8:$src2))]>;
1540def ADD16ri_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1541                    "", // orw/addw REG, imm
1542                    [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1543
1544def ADD32ri8_DB : I<0, Pseudo,
1545                    (outs GR32:$dst), (ins GR32:$src1, i32i8imm:$src2),
1546                    "", // orl/addl REG, imm8
1547                    [(set GR32:$dst,(or_is_add GR32:$src1,i32immSExt8:$src2))]>;
1548def ADD32ri_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1549                    "", // orl/addl REG, imm
1550                    [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1551
1552
1553def ADD64ri8_DB : I<0, Pseudo,
1554                    (outs GR64:$dst), (ins GR64:$src1, i64i8imm:$src2),
1555                    "", // orq/addq REG, imm8
1556                    [(set GR64:$dst, (or_is_add GR64:$src1,
1557                                                i64immSExt8:$src2))]>;
1558def ADD64ri32_DB : I<0, Pseudo,
1559                     (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1560                     "", // orq/addq REG, imm
1561                     [(set GR64:$dst, (or_is_add GR64:$src1,
1562                                                 i64immSExt32:$src2))]>;
1563}
1564} // AddedComplexity, SchedRW
1565
1566//===----------------------------------------------------------------------===//
1567// Pattern match XOR as ADD
1568//===----------------------------------------------------------------------===//
1569
1570// Prefer to pattern match XOR with min_signed_value as ADD at isel time.
1571// ADD can be 3-addressified into an LEA instruction to avoid copies.
1572let AddedComplexity = 5 in {
1573def : Pat<(xor GR8:$src1, -128),
1574          (ADD8ri GR8:$src1, -128)>;
1575def : Pat<(xor GR16:$src1, -32768),
1576          (ADD16ri GR16:$src1, -32768)>;
1577def : Pat<(xor GR32:$src1, -2147483648),
1578          (ADD32ri GR32:$src1, -2147483648)>;
1579}
1580
1581//===----------------------------------------------------------------------===//
1582// Some peepholes
1583//===----------------------------------------------------------------------===//
1584
1585// Odd encoding trick: -128 fits into an 8-bit immediate field while
1586// +128 doesn't, so in this special case use a sub instead of an add.
1587def : Pat<(add GR16:$src1, 128),
1588          (SUB16ri8 GR16:$src1, -128)>;
1589def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1590          (SUB16mi8 addr:$dst, -128)>;
1591
1592def : Pat<(add GR32:$src1, 128),
1593          (SUB32ri8 GR32:$src1, -128)>;
1594def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1595          (SUB32mi8 addr:$dst, -128)>;
1596
1597def : Pat<(add GR64:$src1, 128),
1598          (SUB64ri8 GR64:$src1, -128)>;
1599def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1600          (SUB64mi8 addr:$dst, -128)>;
1601
1602def : Pat<(X86add_flag_nocf GR16:$src1, 128),
1603          (SUB16ri8 GR16:$src1, -128)>;
1604def : Pat<(X86add_flag_nocf GR32:$src1, 128),
1605          (SUB32ri8 GR32:$src1, -128)>;
1606def : Pat<(X86add_flag_nocf GR64:$src1, 128),
1607          (SUB64ri8 GR64:$src1, -128)>;
1608
1609// The same trick applies for 32-bit immediate fields in 64-bit
1610// instructions.
1611def : Pat<(add GR64:$src1, 0x0000000080000000),
1612          (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1613def : Pat<(store (add (loadi64 addr:$dst), 0x0000000080000000), addr:$dst),
1614          (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1615
1616def : Pat<(X86add_flag_nocf GR64:$src1, 0x0000000080000000),
1617          (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1618
1619// To avoid needing to materialize an immediate in a register, use a 32-bit and
1620// with implicit zero-extension instead of a 64-bit and if the immediate has at
1621// least 32 bits of leading zeros. If in addition the last 32 bits can be
1622// represented with a sign extension of a 8 bit constant, use that.
1623// This can also reduce instruction size by eliminating the need for the REX
1624// prefix.
1625
1626// AddedComplexity is needed to give priority over i64immSExt8 and i64immSExt32.
1627let AddedComplexity = 1 in {
1628def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
1629          (SUBREG_TO_REG
1630            (i64 0),
1631            (AND32ri8
1632              (EXTRACT_SUBREG GR64:$src, sub_32bit),
1633              (i32 (GetLo32XForm imm:$imm))),
1634            sub_32bit)>;
1635
1636def : Pat<(and GR64:$src, i64immZExt32:$imm),
1637          (SUBREG_TO_REG
1638            (i64 0),
1639            (AND32ri
1640              (EXTRACT_SUBREG GR64:$src, sub_32bit),
1641              (i32 (GetLo32XForm imm:$imm))),
1642            sub_32bit)>;
1643} // AddedComplexity = 1
1644
1645
1646// AddedComplexity is needed due to the increased complexity on the
1647// i64immZExt32SExt8 and i64immZExt32 patterns above. Applying this to all
1648// the MOVZX patterns keeps thems together in DAGIsel tables.
1649let AddedComplexity = 1 in {
1650// r & (2^16-1) ==> movz
1651def : Pat<(and GR32:$src1, 0xffff),
1652          (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1653// r & (2^8-1) ==> movz
1654def : Pat<(and GR32:$src1, 0xff),
1655          (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>;
1656// r & (2^8-1) ==> movz
1657def : Pat<(and GR16:$src1, 0xff),
1658           (EXTRACT_SUBREG (MOVZX32rr8 (EXTRACT_SUBREG GR16:$src1, sub_8bit)),
1659             sub_16bit)>;
1660
1661// r & (2^32-1) ==> movz
1662def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1663          (SUBREG_TO_REG (i64 0),
1664                         (MOV32rr (EXTRACT_SUBREG GR64:$src, sub_32bit)),
1665                         sub_32bit)>;
1666// r & (2^16-1) ==> movz
1667def : Pat<(and GR64:$src, 0xffff),
1668          (SUBREG_TO_REG (i64 0),
1669                      (MOVZX32rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit))),
1670                      sub_32bit)>;
1671// r & (2^8-1) ==> movz
1672def : Pat<(and GR64:$src, 0xff),
1673          (SUBREG_TO_REG (i64 0),
1674                         (MOVZX32rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit))),
1675                         sub_32bit)>;
1676} // AddedComplexity = 1
1677
1678
1679// Try to use BTS/BTR/BTC for single bit operations on the upper 32-bits.
1680
1681def BTRXForm : SDNodeXForm<imm, [{
1682  // Transformation function: Find the lowest 0.
1683  return getI64Imm((uint8_t)N->getAPIntValue().countTrailingOnes(), SDLoc(N));
1684}]>;
1685
1686def BTCBTSXForm : SDNodeXForm<imm, [{
1687  // Transformation function: Find the lowest 1.
1688  return getI64Imm((uint8_t)N->getAPIntValue().countTrailingZeros(), SDLoc(N));
1689}]>;
1690
1691def BTRMask64 : ImmLeaf<i64, [{
1692  return !isUInt<32>(Imm) && !isInt<32>(Imm) && isPowerOf2_64(~Imm);
1693}]>;
1694
1695def BTCBTSMask64 : ImmLeaf<i64, [{
1696  return !isInt<32>(Imm) && isPowerOf2_64(Imm);
1697}]>;
1698
1699// For now only do this for optsize.
1700let AddedComplexity = 1, Predicates=[OptForSize] in {
1701  def : Pat<(and GR64:$src1, BTRMask64:$mask),
1702            (BTR64ri8 GR64:$src1, (BTRXForm imm:$mask))>;
1703  def : Pat<(or GR64:$src1, BTCBTSMask64:$mask),
1704            (BTS64ri8 GR64:$src1, (BTCBTSXForm imm:$mask))>;
1705  def : Pat<(xor GR64:$src1, BTCBTSMask64:$mask),
1706            (BTC64ri8 GR64:$src1, (BTCBTSXForm imm:$mask))>;
1707}
1708
1709
1710// sext_inreg patterns
1711def : Pat<(sext_inreg GR32:$src, i16),
1712          (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1713def : Pat<(sext_inreg GR32:$src, i8),
1714          (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>;
1715
1716def : Pat<(sext_inreg GR16:$src, i8),
1717           (EXTRACT_SUBREG (MOVSX32rr8 (EXTRACT_SUBREG GR16:$src, sub_8bit)),
1718             sub_16bit)>;
1719
1720def : Pat<(sext_inreg GR64:$src, i32),
1721          (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1722def : Pat<(sext_inreg GR64:$src, i16),
1723          (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1724def : Pat<(sext_inreg GR64:$src, i8),
1725          (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1726
1727// sext, sext_load, zext, zext_load
1728def: Pat<(i16 (sext GR8:$src)),
1729          (EXTRACT_SUBREG (MOVSX32rr8 GR8:$src), sub_16bit)>;
1730def: Pat<(sextloadi16i8 addr:$src),
1731          (EXTRACT_SUBREG (MOVSX32rm8 addr:$src), sub_16bit)>;
1732def: Pat<(i16 (zext GR8:$src)),
1733          (EXTRACT_SUBREG (MOVZX32rr8 GR8:$src), sub_16bit)>;
1734def: Pat<(zextloadi16i8 addr:$src),
1735          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1736
1737// trunc patterns
1738def : Pat<(i16 (trunc GR32:$src)),
1739          (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1740def : Pat<(i8 (trunc GR32:$src)),
1741          (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1742                          sub_8bit)>,
1743      Requires<[Not64BitMode]>;
1744def : Pat<(i8 (trunc GR16:$src)),
1745          (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1746                          sub_8bit)>,
1747      Requires<[Not64BitMode]>;
1748def : Pat<(i32 (trunc GR64:$src)),
1749          (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1750def : Pat<(i16 (trunc GR64:$src)),
1751          (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1752def : Pat<(i8 (trunc GR64:$src)),
1753          (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1754def : Pat<(i8 (trunc GR32:$src)),
1755          (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1756      Requires<[In64BitMode]>;
1757def : Pat<(i8 (trunc GR16:$src)),
1758          (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1759      Requires<[In64BitMode]>;
1760
1761def immff00_ffff  : ImmLeaf<i32, [{
1762  return Imm >= 0xff00 && Imm <= 0xffff;
1763}]>;
1764
1765// h-register tricks
1766def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1767          (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)>,
1768      Requires<[Not64BitMode]>;
1769def : Pat<(i8 (trunc (srl_su (i32 (anyext GR16:$src)), (i8 8)))),
1770          (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)>,
1771      Requires<[Not64BitMode]>;
1772def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1773          (EXTRACT_SUBREG GR32:$src, sub_8bit_hi)>,
1774      Requires<[Not64BitMode]>;
1775def : Pat<(srl GR16:$src, (i8 8)),
1776          (EXTRACT_SUBREG
1777            (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1778            sub_16bit)>;
1779def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1780          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>;
1781def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1782          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>;
1783def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1784          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1785def : Pat<(srl (and_su GR32:$src, immff00_ffff), (i8 8)),
1786          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1787
1788// h-register tricks.
1789// For now, be conservative on x86-64 and use an h-register extract only if the
1790// value is immediately zero-extended or stored, which are somewhat common
1791// cases. This uses a bunch of code to prevent a register requiring a REX prefix
1792// from being allocated in the same instruction as the h register, as there's
1793// currently no way to describe this requirement to the register allocator.
1794
1795// h-register extract and zero-extend.
1796def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1797          (SUBREG_TO_REG
1798            (i64 0),
1799            (MOVZX32rr8_NOREX
1800              (EXTRACT_SUBREG GR64:$src, sub_8bit_hi)),
1801            sub_32bit)>;
1802def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1803          (SUBREG_TO_REG
1804            (i64 0),
1805            (MOVZX32rr8_NOREX
1806              (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1807            sub_32bit)>;
1808def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1809          (SUBREG_TO_REG
1810            (i64 0),
1811            (MOVZX32rr8_NOREX
1812              (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1813            sub_32bit)>;
1814
1815// h-register extract and store.
1816def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1817          (MOV8mr_NOREX
1818            addr:$dst,
1819            (EXTRACT_SUBREG GR64:$src, sub_8bit_hi))>;
1820def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1821          (MOV8mr_NOREX
1822            addr:$dst,
1823            (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>,
1824      Requires<[In64BitMode]>;
1825def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1826          (MOV8mr_NOREX
1827            addr:$dst,
1828            (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>,
1829      Requires<[In64BitMode]>;
1830
1831// Special pattern to catch the last step of __builtin_parity handling. Our
1832// goal is to use an xor of an h-register with the corresponding l-register.
1833// The above patterns would handle this on non 64-bit targets, but for 64-bit
1834// we need to be more careful. We're using a NOREX instruction here in case
1835// register allocation fails to keep the two registers together. So we need to
1836// make sure we can't accidentally mix R8-R15 with an h-register.
1837def : Pat<(X86xor_flag (i8 (trunc GR32:$src)),
1838                       (i8 (trunc (srl_su GR32:$src, (i8 8))))),
1839          (XOR8rr_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit),
1840                        (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1841
1842// (shl x, 1) ==> (add x, x)
1843// Note that if x is undef (immediate or otherwise), we could theoretically
1844// end up with the two uses of x getting different values, producing a result
1845// where the least significant bit is not 0. However, the probability of this
1846// happening is considered low enough that this is officially not a
1847// "real problem".
1848def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr  GR8 :$src1, GR8 :$src1)>;
1849def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1850def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1851def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1852
1853def shiftMask8 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1854  return isUnneededShiftMask(N, 3);
1855}]>;
1856
1857def shiftMask16 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1858  return isUnneededShiftMask(N, 4);
1859}]>;
1860
1861def shiftMask32 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1862  return isUnneededShiftMask(N, 5);
1863}]>;
1864
1865def shiftMask64 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1866  return isUnneededShiftMask(N, 6);
1867}]>;
1868
1869
1870// Shift amount is implicitly masked.
1871multiclass MaskedShiftAmountPats<SDNode frag, string name> {
1872  // (shift x (and y, 31)) ==> (shift x, y)
1873  def : Pat<(frag GR8:$src1, (shiftMask32 CL)),
1874            (!cast<Instruction>(name # "8rCL") GR8:$src1)>;
1875  def : Pat<(frag GR16:$src1, (shiftMask32 CL)),
1876            (!cast<Instruction>(name # "16rCL") GR16:$src1)>;
1877  def : Pat<(frag GR32:$src1, (shiftMask32 CL)),
1878            (!cast<Instruction>(name # "32rCL") GR32:$src1)>;
1879  def : Pat<(store (frag (loadi8 addr:$dst), (shiftMask32 CL)), addr:$dst),
1880            (!cast<Instruction>(name # "8mCL") addr:$dst)>;
1881  def : Pat<(store (frag (loadi16 addr:$dst), (shiftMask32 CL)), addr:$dst),
1882            (!cast<Instruction>(name # "16mCL") addr:$dst)>;
1883  def : Pat<(store (frag (loadi32 addr:$dst), (shiftMask32 CL)), addr:$dst),
1884            (!cast<Instruction>(name # "32mCL") addr:$dst)>;
1885
1886  // (shift x (and y, 63)) ==> (shift x, y)
1887  def : Pat<(frag GR64:$src1, (shiftMask64 CL)),
1888            (!cast<Instruction>(name # "64rCL") GR64:$src1)>;
1889  def : Pat<(store (frag (loadi64 addr:$dst), (shiftMask64 CL)), addr:$dst),
1890            (!cast<Instruction>(name # "64mCL") addr:$dst)>;
1891}
1892
1893defm : MaskedShiftAmountPats<shl, "SHL">;
1894defm : MaskedShiftAmountPats<srl, "SHR">;
1895defm : MaskedShiftAmountPats<sra, "SAR">;
1896
1897// ROL/ROR instructions allow a stronger mask optimization than shift for 8- and
1898// 16-bit. We can remove a mask of any (bitwidth - 1) on the rotation amount
1899// because over-rotating produces the same result. This is noted in the Intel
1900// docs with: "tempCOUNT <- (COUNT & COUNTMASK) MOD SIZE". Masking the rotation
1901// amount could affect EFLAGS results, but that does not matter because we are
1902// not tracking flags for these nodes.
1903multiclass MaskedRotateAmountPats<SDNode frag, string name> {
1904  // (rot x (and y, BitWidth - 1)) ==> (rot x, y)
1905  def : Pat<(frag GR8:$src1, (shiftMask8 CL)),
1906  (!cast<Instruction>(name # "8rCL") GR8:$src1)>;
1907  def : Pat<(frag GR16:$src1, (shiftMask16 CL)),
1908  (!cast<Instruction>(name # "16rCL") GR16:$src1)>;
1909  def : Pat<(frag GR32:$src1, (shiftMask32 CL)),
1910  (!cast<Instruction>(name # "32rCL") GR32:$src1)>;
1911  def : Pat<(store (frag (loadi8 addr:$dst), (shiftMask8 CL)), addr:$dst),
1912  (!cast<Instruction>(name # "8mCL") addr:$dst)>;
1913  def : Pat<(store (frag (loadi16 addr:$dst), (shiftMask16 CL)), addr:$dst),
1914  (!cast<Instruction>(name # "16mCL") addr:$dst)>;
1915  def : Pat<(store (frag (loadi32 addr:$dst), (shiftMask32 CL)), addr:$dst),
1916  (!cast<Instruction>(name # "32mCL") addr:$dst)>;
1917
1918  // (rot x (and y, 63)) ==> (rot x, y)
1919  def : Pat<(frag GR64:$src1, (shiftMask64 CL)),
1920  (!cast<Instruction>(name # "64rCL") GR64:$src1)>;
1921  def : Pat<(store (frag (loadi64 addr:$dst), (shiftMask64 CL)), addr:$dst),
1922  (!cast<Instruction>(name # "64mCL") addr:$dst)>;
1923}
1924
1925
1926defm : MaskedRotateAmountPats<rotl, "ROL">;
1927defm : MaskedRotateAmountPats<rotr, "ROR">;
1928
1929// Double "funnel" shift amount is implicitly masked.
1930// (fshl/fshr x (and y, 31)) ==> (fshl/fshr x, y) (NOTE: modulo32)
1931def : Pat<(X86fshl GR16:$src1, GR16:$src2, (shiftMask32 CL)),
1932          (SHLD16rrCL GR16:$src1, GR16:$src2)>;
1933def : Pat<(X86fshr GR16:$src2, GR16:$src1, (shiftMask32 CL)),
1934          (SHRD16rrCL GR16:$src1, GR16:$src2)>;
1935
1936// (fshl/fshr x (and y, 31)) ==> (fshl/fshr x, y)
1937def : Pat<(fshl GR32:$src1, GR32:$src2, (shiftMask32 CL)),
1938          (SHLD32rrCL GR32:$src1, GR32:$src2)>;
1939def : Pat<(fshr GR32:$src2, GR32:$src1, (shiftMask32 CL)),
1940          (SHRD32rrCL GR32:$src1, GR32:$src2)>;
1941
1942// (fshl/fshr x (and y, 63)) ==> (fshl/fshr x, y)
1943def : Pat<(fshl GR64:$src1, GR64:$src2, (shiftMask64 CL)),
1944          (SHLD64rrCL GR64:$src1, GR64:$src2)>;
1945def : Pat<(fshr GR64:$src2, GR64:$src1, (shiftMask64 CL)),
1946          (SHRD64rrCL GR64:$src1, GR64:$src2)>;
1947
1948let Predicates = [HasBMI2] in {
1949  let AddedComplexity = 1 in {
1950    def : Pat<(sra GR32:$src1, (shiftMask32 GR8:$src2)),
1951              (SARX32rr GR32:$src1,
1952                        (INSERT_SUBREG
1953                          (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1954    def : Pat<(sra GR64:$src1, (shiftMask64 GR8:$src2)),
1955              (SARX64rr GR64:$src1,
1956                        (INSERT_SUBREG
1957                          (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1958
1959    def : Pat<(srl GR32:$src1, (shiftMask32 GR8:$src2)),
1960              (SHRX32rr GR32:$src1,
1961                        (INSERT_SUBREG
1962                          (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1963    def : Pat<(srl GR64:$src1, (shiftMask64 GR8:$src2)),
1964              (SHRX64rr GR64:$src1,
1965                        (INSERT_SUBREG
1966                          (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1967
1968    def : Pat<(shl GR32:$src1, (shiftMask32 GR8:$src2)),
1969              (SHLX32rr GR32:$src1,
1970                        (INSERT_SUBREG
1971                          (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1972    def : Pat<(shl GR64:$src1, (shiftMask64 GR8:$src2)),
1973              (SHLX64rr GR64:$src1,
1974                        (INSERT_SUBREG
1975                          (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1976  }
1977
1978  def : Pat<(sra (loadi32 addr:$src1), (shiftMask32 GR8:$src2)),
1979            (SARX32rm addr:$src1,
1980                      (INSERT_SUBREG
1981                        (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1982  def : Pat<(sra (loadi64 addr:$src1), (shiftMask64 GR8:$src2)),
1983            (SARX64rm addr:$src1,
1984                      (INSERT_SUBREG
1985                        (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1986
1987  def : Pat<(srl (loadi32 addr:$src1), (shiftMask32 GR8:$src2)),
1988            (SHRX32rm addr:$src1,
1989                      (INSERT_SUBREG
1990                        (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1991  def : Pat<(srl (loadi64 addr:$src1), (shiftMask64 GR8:$src2)),
1992            (SHRX64rm addr:$src1,
1993                      (INSERT_SUBREG
1994                        (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1995
1996  def : Pat<(shl (loadi32 addr:$src1), (shiftMask32 GR8:$src2)),
1997            (SHLX32rm addr:$src1,
1998                      (INSERT_SUBREG
1999                        (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
2000  def : Pat<(shl (loadi64 addr:$src1), (shiftMask64 GR8:$src2)),
2001            (SHLX64rm addr:$src1,
2002                      (INSERT_SUBREG
2003                        (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
2004}
2005
2006// Use BTR/BTS/BTC for clearing/setting/toggling a bit in a variable location.
2007multiclass one_bit_patterns<RegisterClass RC, ValueType VT, Instruction BTR,
2008                            Instruction BTS, Instruction BTC,
2009                            PatFrag ShiftMask> {
2010  def : Pat<(and RC:$src1, (rotl -2, GR8:$src2)),
2011            (BTR RC:$src1,
2012                 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
2013  def : Pat<(or RC:$src1, (shl 1, GR8:$src2)),
2014            (BTS RC:$src1,
2015                 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
2016  def : Pat<(xor RC:$src1, (shl 1, GR8:$src2)),
2017            (BTC RC:$src1,
2018                 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
2019
2020  // Similar to above, but removing unneeded masking of the shift amount.
2021  def : Pat<(and RC:$src1, (rotl -2, (ShiftMask GR8:$src2))),
2022            (BTR RC:$src1,
2023                 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
2024  def : Pat<(or RC:$src1, (shl 1, (ShiftMask GR8:$src2))),
2025            (BTS RC:$src1,
2026                (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
2027  def : Pat<(xor RC:$src1, (shl 1, (ShiftMask GR8:$src2))),
2028            (BTC RC:$src1,
2029                (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
2030}
2031
2032defm : one_bit_patterns<GR16, i16, BTR16rr, BTS16rr, BTC16rr, shiftMask16>;
2033defm : one_bit_patterns<GR32, i32, BTR32rr, BTS32rr, BTC32rr, shiftMask32>;
2034defm : one_bit_patterns<GR64, i64, BTR64rr, BTS64rr, BTC64rr, shiftMask64>;
2035
2036//===----------------------------------------------------------------------===//
2037// EFLAGS-defining Patterns
2038//===----------------------------------------------------------------------===//
2039
2040// add reg, reg
2041def : Pat<(add GR8 :$src1, GR8 :$src2), (ADD8rr  GR8 :$src1, GR8 :$src2)>;
2042def : Pat<(add GR16:$src1, GR16:$src2), (ADD16rr GR16:$src1, GR16:$src2)>;
2043def : Pat<(add GR32:$src1, GR32:$src2), (ADD32rr GR32:$src1, GR32:$src2)>;
2044def : Pat<(add GR64:$src1, GR64:$src2), (ADD64rr GR64:$src1, GR64:$src2)>;
2045
2046// add reg, mem
2047def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
2048          (ADD8rm GR8:$src1, addr:$src2)>;
2049def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
2050          (ADD16rm GR16:$src1, addr:$src2)>;
2051def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
2052          (ADD32rm GR32:$src1, addr:$src2)>;
2053def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
2054          (ADD64rm GR64:$src1, addr:$src2)>;
2055
2056// add reg, imm
2057def : Pat<(add GR8 :$src1, imm:$src2), (ADD8ri  GR8:$src1 , imm:$src2)>;
2058def : Pat<(add GR16:$src1, imm:$src2), (ADD16ri GR16:$src1, imm:$src2)>;
2059def : Pat<(add GR32:$src1, imm:$src2), (ADD32ri GR32:$src1, imm:$src2)>;
2060def : Pat<(add GR16:$src1, i16immSExt8:$src2),
2061          (ADD16ri8 GR16:$src1, i16immSExt8:$src2)>;
2062def : Pat<(add GR32:$src1, i32immSExt8:$src2),
2063          (ADD32ri8 GR32:$src1, i32immSExt8:$src2)>;
2064def : Pat<(add GR64:$src1, i64immSExt8:$src2),
2065          (ADD64ri8 GR64:$src1, i64immSExt8:$src2)>;
2066def : Pat<(add GR64:$src1, i64immSExt32:$src2),
2067          (ADD64ri32 GR64:$src1, i64immSExt32:$src2)>;
2068
2069// sub reg, reg
2070def : Pat<(sub GR8 :$src1, GR8 :$src2), (SUB8rr  GR8 :$src1, GR8 :$src2)>;
2071def : Pat<(sub GR16:$src1, GR16:$src2), (SUB16rr GR16:$src1, GR16:$src2)>;
2072def : Pat<(sub GR32:$src1, GR32:$src2), (SUB32rr GR32:$src1, GR32:$src2)>;
2073def : Pat<(sub GR64:$src1, GR64:$src2), (SUB64rr GR64:$src1, GR64:$src2)>;
2074
2075// sub reg, mem
2076def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
2077          (SUB8rm GR8:$src1, addr:$src2)>;
2078def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
2079          (SUB16rm GR16:$src1, addr:$src2)>;
2080def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
2081          (SUB32rm GR32:$src1, addr:$src2)>;
2082def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
2083          (SUB64rm GR64:$src1, addr:$src2)>;
2084
2085// sub reg, imm
2086def : Pat<(sub GR8:$src1, imm:$src2),
2087          (SUB8ri GR8:$src1, imm:$src2)>;
2088def : Pat<(sub GR16:$src1, imm:$src2),
2089          (SUB16ri GR16:$src1, imm:$src2)>;
2090def : Pat<(sub GR32:$src1, imm:$src2),
2091          (SUB32ri GR32:$src1, imm:$src2)>;
2092def : Pat<(sub GR16:$src1, i16immSExt8:$src2),
2093          (SUB16ri8 GR16:$src1, i16immSExt8:$src2)>;
2094def : Pat<(sub GR32:$src1, i32immSExt8:$src2),
2095          (SUB32ri8 GR32:$src1, i32immSExt8:$src2)>;
2096def : Pat<(sub GR64:$src1, i64immSExt8:$src2),
2097          (SUB64ri8 GR64:$src1, i64immSExt8:$src2)>;
2098def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
2099          (SUB64ri32 GR64:$src1, i64immSExt32:$src2)>;
2100
2101// sub 0, reg
2102def : Pat<(X86sub_flag 0, GR8 :$src), (NEG8r  GR8 :$src)>;
2103def : Pat<(X86sub_flag 0, GR16:$src), (NEG16r GR16:$src)>;
2104def : Pat<(X86sub_flag 0, GR32:$src), (NEG32r GR32:$src)>;
2105def : Pat<(X86sub_flag 0, GR64:$src), (NEG64r GR64:$src)>;
2106
2107// mul reg, reg
2108def : Pat<(mul GR16:$src1, GR16:$src2),
2109          (IMUL16rr GR16:$src1, GR16:$src2)>;
2110def : Pat<(mul GR32:$src1, GR32:$src2),
2111          (IMUL32rr GR32:$src1, GR32:$src2)>;
2112def : Pat<(mul GR64:$src1, GR64:$src2),
2113          (IMUL64rr GR64:$src1, GR64:$src2)>;
2114
2115// mul reg, mem
2116def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
2117          (IMUL16rm GR16:$src1, addr:$src2)>;
2118def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
2119          (IMUL32rm GR32:$src1, addr:$src2)>;
2120def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
2121          (IMUL64rm GR64:$src1, addr:$src2)>;
2122
2123// mul reg, imm
2124def : Pat<(mul GR16:$src1, imm:$src2),
2125          (IMUL16rri GR16:$src1, imm:$src2)>;
2126def : Pat<(mul GR32:$src1, imm:$src2),
2127          (IMUL32rri GR32:$src1, imm:$src2)>;
2128def : Pat<(mul GR16:$src1, i16immSExt8:$src2),
2129          (IMUL16rri8 GR16:$src1, i16immSExt8:$src2)>;
2130def : Pat<(mul GR32:$src1, i32immSExt8:$src2),
2131          (IMUL32rri8 GR32:$src1, i32immSExt8:$src2)>;
2132def : Pat<(mul GR64:$src1, i64immSExt8:$src2),
2133          (IMUL64rri8 GR64:$src1, i64immSExt8:$src2)>;
2134def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
2135          (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
2136
2137// reg = mul mem, imm
2138def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
2139          (IMUL16rmi addr:$src1, imm:$src2)>;
2140def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
2141          (IMUL32rmi addr:$src1, imm:$src2)>;
2142def : Pat<(mul (loadi16 addr:$src1), i16immSExt8:$src2),
2143          (IMUL16rmi8 addr:$src1, i16immSExt8:$src2)>;
2144def : Pat<(mul (loadi32 addr:$src1), i32immSExt8:$src2),
2145          (IMUL32rmi8 addr:$src1, i32immSExt8:$src2)>;
2146def : Pat<(mul (loadi64 addr:$src1), i64immSExt8:$src2),
2147          (IMUL64rmi8 addr:$src1, i64immSExt8:$src2)>;
2148def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
2149          (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
2150
2151// Increment/Decrement reg.
2152// Do not make INC/DEC if it is slow
2153let Predicates = [UseIncDec] in {
2154  def : Pat<(add GR8:$src, 1),   (INC8r GR8:$src)>;
2155  def : Pat<(add GR16:$src, 1),  (INC16r GR16:$src)>;
2156  def : Pat<(add GR32:$src, 1),  (INC32r GR32:$src)>;
2157  def : Pat<(add GR64:$src, 1),  (INC64r GR64:$src)>;
2158  def : Pat<(add GR8:$src, -1),  (DEC8r GR8:$src)>;
2159  def : Pat<(add GR16:$src, -1), (DEC16r GR16:$src)>;
2160  def : Pat<(add GR32:$src, -1), (DEC32r GR32:$src)>;
2161  def : Pat<(add GR64:$src, -1), (DEC64r GR64:$src)>;
2162
2163  def : Pat<(X86add_flag_nocf GR8:$src, -1),  (DEC8r GR8:$src)>;
2164  def : Pat<(X86add_flag_nocf GR16:$src, -1), (DEC16r GR16:$src)>;
2165  def : Pat<(X86add_flag_nocf GR32:$src, -1), (DEC32r GR32:$src)>;
2166  def : Pat<(X86add_flag_nocf GR64:$src, -1), (DEC64r GR64:$src)>;
2167  def : Pat<(X86sub_flag_nocf GR8:$src, -1),  (INC8r GR8:$src)>;
2168  def : Pat<(X86sub_flag_nocf GR16:$src, -1), (INC16r GR16:$src)>;
2169  def : Pat<(X86sub_flag_nocf GR32:$src, -1), (INC32r GR32:$src)>;
2170  def : Pat<(X86sub_flag_nocf GR64:$src, -1), (INC64r GR64:$src)>;
2171}
2172
2173// or reg/reg.
2174def : Pat<(or GR8 :$src1, GR8 :$src2), (OR8rr  GR8 :$src1, GR8 :$src2)>;
2175def : Pat<(or GR16:$src1, GR16:$src2), (OR16rr GR16:$src1, GR16:$src2)>;
2176def : Pat<(or GR32:$src1, GR32:$src2), (OR32rr GR32:$src1, GR32:$src2)>;
2177def : Pat<(or GR64:$src1, GR64:$src2), (OR64rr GR64:$src1, GR64:$src2)>;
2178
2179// or reg/mem
2180def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
2181          (OR8rm GR8:$src1, addr:$src2)>;
2182def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
2183          (OR16rm GR16:$src1, addr:$src2)>;
2184def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
2185          (OR32rm GR32:$src1, addr:$src2)>;
2186def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
2187          (OR64rm GR64:$src1, addr:$src2)>;
2188
2189// or reg/imm
2190def : Pat<(or GR8:$src1 , imm:$src2), (OR8ri  GR8 :$src1, imm:$src2)>;
2191def : Pat<(or GR16:$src1, imm:$src2), (OR16ri GR16:$src1, imm:$src2)>;
2192def : Pat<(or GR32:$src1, imm:$src2), (OR32ri GR32:$src1, imm:$src2)>;
2193def : Pat<(or GR16:$src1, i16immSExt8:$src2),
2194          (OR16ri8 GR16:$src1, i16immSExt8:$src2)>;
2195def : Pat<(or GR32:$src1, i32immSExt8:$src2),
2196          (OR32ri8 GR32:$src1, i32immSExt8:$src2)>;
2197def : Pat<(or GR64:$src1, i64immSExt8:$src2),
2198          (OR64ri8 GR64:$src1, i64immSExt8:$src2)>;
2199def : Pat<(or GR64:$src1, i64immSExt32:$src2),
2200          (OR64ri32 GR64:$src1, i64immSExt32:$src2)>;
2201
2202// xor reg/reg
2203def : Pat<(xor GR8 :$src1, GR8 :$src2), (XOR8rr  GR8 :$src1, GR8 :$src2)>;
2204def : Pat<(xor GR16:$src1, GR16:$src2), (XOR16rr GR16:$src1, GR16:$src2)>;
2205def : Pat<(xor GR32:$src1, GR32:$src2), (XOR32rr GR32:$src1, GR32:$src2)>;
2206def : Pat<(xor GR64:$src1, GR64:$src2), (XOR64rr GR64:$src1, GR64:$src2)>;
2207
2208// xor reg/mem
2209def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
2210          (XOR8rm GR8:$src1, addr:$src2)>;
2211def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
2212          (XOR16rm GR16:$src1, addr:$src2)>;
2213def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
2214          (XOR32rm GR32:$src1, addr:$src2)>;
2215def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
2216          (XOR64rm GR64:$src1, addr:$src2)>;
2217
2218// xor reg/imm
2219def : Pat<(xor GR8:$src1, imm:$src2),
2220          (XOR8ri GR8:$src1, imm:$src2)>;
2221def : Pat<(xor GR16:$src1, imm:$src2),
2222          (XOR16ri GR16:$src1, imm:$src2)>;
2223def : Pat<(xor GR32:$src1, imm:$src2),
2224          (XOR32ri GR32:$src1, imm:$src2)>;
2225def : Pat<(xor GR16:$src1, i16immSExt8:$src2),
2226          (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
2227def : Pat<(xor GR32:$src1, i32immSExt8:$src2),
2228          (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
2229def : Pat<(xor GR64:$src1, i64immSExt8:$src2),
2230          (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
2231def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
2232          (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
2233
2234// and reg/reg
2235def : Pat<(and GR8 :$src1, GR8 :$src2), (AND8rr  GR8 :$src1, GR8 :$src2)>;
2236def : Pat<(and GR16:$src1, GR16:$src2), (AND16rr GR16:$src1, GR16:$src2)>;
2237def : Pat<(and GR32:$src1, GR32:$src2), (AND32rr GR32:$src1, GR32:$src2)>;
2238def : Pat<(and GR64:$src1, GR64:$src2), (AND64rr GR64:$src1, GR64:$src2)>;
2239
2240// and reg/mem
2241def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
2242          (AND8rm GR8:$src1, addr:$src2)>;
2243def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
2244          (AND16rm GR16:$src1, addr:$src2)>;
2245def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
2246          (AND32rm GR32:$src1, addr:$src2)>;
2247def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
2248          (AND64rm GR64:$src1, addr:$src2)>;
2249
2250// and reg/imm
2251def : Pat<(and GR8:$src1, imm:$src2),
2252          (AND8ri GR8:$src1, imm:$src2)>;
2253def : Pat<(and GR16:$src1, imm:$src2),
2254          (AND16ri GR16:$src1, imm:$src2)>;
2255def : Pat<(and GR32:$src1, imm:$src2),
2256          (AND32ri GR32:$src1, imm:$src2)>;
2257def : Pat<(and GR16:$src1, i16immSExt8:$src2),
2258          (AND16ri8 GR16:$src1, i16immSExt8:$src2)>;
2259def : Pat<(and GR32:$src1, i32immSExt8:$src2),
2260          (AND32ri8 GR32:$src1, i32immSExt8:$src2)>;
2261def : Pat<(and GR64:$src1, i64immSExt8:$src2),
2262          (AND64ri8 GR64:$src1, i64immSExt8:$src2)>;
2263def : Pat<(and GR64:$src1, i64immSExt32:$src2),
2264          (AND64ri32 GR64:$src1, i64immSExt32:$src2)>;
2265
2266// Bit scan instruction patterns to match explicit zero-undef behavior.
2267def : Pat<(cttz_zero_undef GR16:$src), (BSF16rr GR16:$src)>;
2268def : Pat<(cttz_zero_undef GR32:$src), (BSF32rr GR32:$src)>;
2269def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr GR64:$src)>;
2270def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm addr:$src)>;
2271def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm addr:$src)>;
2272def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm addr:$src)>;
2273
2274// When HasMOVBE is enabled it is possible to get a non-legalized
2275// register-register 16 bit bswap. This maps it to a ROL instruction.
2276let Predicates = [HasMOVBE] in {
2277 def : Pat<(bswap GR16:$src), (ROL16ri GR16:$src, (i8 8))>;
2278}
2279