xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86InstrCompiler.td (revision cfd6422a5217410fbd66f7a7a8a64d9d85e61229)
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:$amt2, 0)>, Requires<[IsLP64]>;
68
69let SchedRW = [WriteSystem] in {
70
71// x86-64 va_start lowering magic.
72let usesCustomInserter = 1, Defs = [EFLAGS] in {
73def VASTART_SAVE_XMM_REGS : I<0, Pseudo,
74                              (outs),
75                              (ins GR8:$al,
76                                   i64imm:$regsavefi, i64imm:$offset,
77                                   variable_ops),
78                              "#VASTART_SAVE_XMM_REGS $al, $regsavefi, $offset",
79                              [(X86vastart_save_xmm_regs GR8:$al,
80                                                         imm:$regsavefi,
81                                                         imm:$offset),
82                               (implicit EFLAGS)]>;
83
84// The VAARG_64 pseudo-instruction takes the address of the va_list,
85// and places the address of the next argument into a register.
86let Defs = [EFLAGS] in
87def VAARG_64 : I<0, Pseudo,
88                 (outs GR64:$dst),
89                 (ins i8mem:$ap, i32imm:$size, i8imm:$mode, i32imm:$align),
90                 "#VAARG_64 $dst, $ap, $size, $mode, $align",
91                 [(set GR64:$dst,
92                    (X86vaarg64 addr:$ap, imm:$size, imm:$mode, imm:$align)),
93                  (implicit EFLAGS)]>;
94
95
96// When using segmented stacks these are lowered into instructions which first
97// check if the current stacklet has enough free memory. If it does, memory is
98// allocated by bumping the stack pointer. Otherwise memory is allocated from
99// the heap.
100
101let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
102def SEG_ALLOCA_32 : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$size),
103                      "# variable sized alloca for segmented stacks",
104                      [(set GR32:$dst,
105                         (X86SegAlloca GR32:$size))]>,
106                    Requires<[NotLP64]>;
107
108let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
109def SEG_ALLOCA_64 : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$size),
110                      "# variable sized alloca for segmented stacks",
111                      [(set GR64:$dst,
112                         (X86SegAlloca GR64:$size))]>,
113                    Requires<[In64BitMode]>;
114
115// To protect against stack clash, dynamic allocation should perform a memory
116// probe at each page.
117
118let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
119def PROBED_ALLOCA_32 : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$size),
120                      "# variable sized alloca with probing",
121                      [(set GR32:$dst,
122                         (X86ProbedAlloca GR32:$size))]>,
123                    Requires<[NotLP64]>;
124
125let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
126def PROBED_ALLOCA_64 : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$size),
127                      "# variable sized alloca with probing",
128                      [(set GR64:$dst,
129                         (X86ProbedAlloca GR64:$size))]>,
130                    Requires<[In64BitMode]>;
131}
132
133let hasNoSchedulingInfo = 1 in
134def STACKALLOC_W_PROBING : I<0, Pseudo, (outs), (ins i64imm:$stacksize),
135                             "# fixed size alloca with probing",
136                             []>;
137
138// Dynamic stack allocation yields a _chkstk or _alloca call for all Windows
139// targets.  These calls are needed to probe the stack when allocating more than
140// 4k bytes in one go. Touching the stack at 4K increments is necessary to
141// ensure that the guard pages used by the OS virtual memory manager are
142// allocated in correct sequence.
143// The main point of having separate instruction are extra unmodelled effects
144// (compared to ordinary calls) like stack pointer change.
145
146let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
147def WIN_ALLOCA_32 : I<0, Pseudo, (outs), (ins GR32:$size),
148                     "# dynamic stack allocation",
149                     [(X86WinAlloca GR32:$size)]>,
150                     Requires<[NotLP64]>;
151
152let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
153def WIN_ALLOCA_64 : I<0, Pseudo, (outs), (ins GR64:$size),
154                     "# dynamic stack allocation",
155                     [(X86WinAlloca GR64:$size)]>,
156                     Requires<[In64BitMode]>;
157} // SchedRW
158
159// These instructions XOR the frame pointer into a GPR. They are used in some
160// stack protection schemes. These are post-RA pseudos because we only know the
161// frame register after register allocation.
162let Constraints = "$src = $dst", isMoveImm = 1, isPseudo = 1, Defs = [EFLAGS] in {
163  def XOR32_FP : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src),
164                  "xorl\t$$FP, $src", []>,
165                  Requires<[NotLP64]>, Sched<[WriteALU]>;
166  def XOR64_FP : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src),
167                  "xorq\t$$FP $src", []>,
168                  Requires<[In64BitMode]>, Sched<[WriteALU]>;
169}
170
171//===----------------------------------------------------------------------===//
172// EH Pseudo Instructions
173//
174let SchedRW = [WriteSystem] in {
175let isTerminator = 1, isReturn = 1, isBarrier = 1,
176    hasCtrlDep = 1, isCodeGenOnly = 1 in {
177def EH_RETURN   : I<0xC3, RawFrm, (outs), (ins GR32:$addr),
178                    "ret\t#eh_return, addr: $addr",
179                    [(X86ehret GR32:$addr)]>, Sched<[WriteJumpLd]>;
180
181}
182
183let isTerminator = 1, isReturn = 1, isBarrier = 1,
184    hasCtrlDep = 1, isCodeGenOnly = 1 in {
185def EH_RETURN64   : I<0xC3, RawFrm, (outs), (ins GR64:$addr),
186                     "ret\t#eh_return, addr: $addr",
187                     [(X86ehret GR64:$addr)]>, Sched<[WriteJumpLd]>;
188
189}
190
191let isTerminator = 1, hasSideEffects = 1, isBarrier = 1, hasCtrlDep = 1,
192    isCodeGenOnly = 1, isReturn = 1, isEHScopeReturn = 1 in {
193  def CLEANUPRET : I<0, Pseudo, (outs), (ins), "# CLEANUPRET", [(cleanupret)]>;
194
195  // CATCHRET needs a custom inserter for SEH.
196  let usesCustomInserter = 1 in
197    def CATCHRET : I<0, Pseudo, (outs), (ins brtarget32:$dst, brtarget32:$from),
198                     "# CATCHRET",
199                     [(catchret bb:$dst, bb:$from)]>;
200}
201
202let hasSideEffects = 1, isBarrier = 1, isCodeGenOnly = 1,
203    usesCustomInserter = 1 in {
204  def EH_SjLj_SetJmp32  : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$buf),
205                            "#EH_SJLJ_SETJMP32",
206                            [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
207                          Requires<[Not64BitMode]>;
208  def EH_SjLj_SetJmp64  : I<0, Pseudo, (outs GR32:$dst), (ins i64mem:$buf),
209                            "#EH_SJLJ_SETJMP64",
210                            [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
211                          Requires<[In64BitMode]>;
212  let isTerminator = 1 in {
213  def EH_SjLj_LongJmp32 : I<0, Pseudo, (outs), (ins i32mem:$buf),
214                            "#EH_SJLJ_LONGJMP32",
215                            [(X86eh_sjlj_longjmp addr:$buf)]>,
216                          Requires<[Not64BitMode]>;
217  def EH_SjLj_LongJmp64 : I<0, Pseudo, (outs), (ins i64mem:$buf),
218                            "#EH_SJLJ_LONGJMP64",
219                            [(X86eh_sjlj_longjmp addr:$buf)]>,
220                          Requires<[In64BitMode]>;
221  }
222}
223
224let isBranch = 1, isTerminator = 1, isCodeGenOnly = 1 in {
225  def EH_SjLj_Setup : I<0, Pseudo, (outs), (ins brtarget:$dst),
226                        "#EH_SjLj_Setup\t$dst", []>;
227}
228} // SchedRW
229
230//===----------------------------------------------------------------------===//
231// Pseudo instructions used by unwind info.
232//
233let isPseudo = 1, SchedRW = [WriteSystem] in {
234  def SEH_PushReg : I<0, Pseudo, (outs), (ins i32imm:$reg),
235                            "#SEH_PushReg $reg", []>;
236  def SEH_SaveReg : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
237                            "#SEH_SaveReg $reg, $dst", []>;
238  def SEH_SaveXMM : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
239                            "#SEH_SaveXMM $reg, $dst", []>;
240  def SEH_StackAlloc : I<0, Pseudo, (outs), (ins i32imm:$size),
241                            "#SEH_StackAlloc $size", []>;
242  def SEH_StackAlign : I<0, Pseudo, (outs), (ins i32imm:$align),
243                            "#SEH_StackAlign $align", []>;
244  def SEH_SetFrame : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$offset),
245                            "#SEH_SetFrame $reg, $offset", []>;
246  def SEH_PushFrame : I<0, Pseudo, (outs), (ins i1imm:$mode),
247                            "#SEH_PushFrame $mode", []>;
248  def SEH_EndPrologue : I<0, Pseudo, (outs), (ins),
249                            "#SEH_EndPrologue", []>;
250  def SEH_Epilogue : I<0, Pseudo, (outs), (ins),
251                            "#SEH_Epilogue", []>;
252}
253
254//===----------------------------------------------------------------------===//
255// Pseudo instructions used by segmented stacks.
256//
257
258// This is lowered into a RET instruction by MCInstLower.  We need
259// this so that we don't have to have a MachineBasicBlock which ends
260// with a RET and also has successors.
261let isPseudo = 1, SchedRW = [WriteJumpLd] in {
262def MORESTACK_RET: I<0, Pseudo, (outs), (ins), "", []>;
263
264// This instruction is lowered to a RET followed by a MOV.  The two
265// instructions are not generated on a higher level since then the
266// verifier sees a MachineBasicBlock ending with a non-terminator.
267def MORESTACK_RET_RESTORE_R10 : I<0, Pseudo, (outs), (ins), "", []>;
268}
269
270//===----------------------------------------------------------------------===//
271// Alias Instructions
272//===----------------------------------------------------------------------===//
273
274// Alias instruction mapping movr0 to xor.
275// FIXME: remove when we can teach regalloc that xor reg, reg is ok.
276let Defs = [EFLAGS], isReMaterializable = 1, isAsCheapAsAMove = 1,
277    isPseudo = 1, isMoveImm = 1, AddedComplexity = 10 in
278def MOV32r0  : I<0, Pseudo, (outs GR32:$dst), (ins), "",
279                 [(set GR32:$dst, 0)]>, Sched<[WriteZero]>;
280
281// Other widths can also make use of the 32-bit xor, which may have a smaller
282// encoding and avoid partial register updates.
283let AddedComplexity = 10 in {
284def : Pat<(i8 0), (EXTRACT_SUBREG (MOV32r0), sub_8bit)>;
285def : Pat<(i16 0), (EXTRACT_SUBREG (MOV32r0), sub_16bit)>;
286def : Pat<(i64 0), (SUBREG_TO_REG (i64 0), (MOV32r0), sub_32bit)>;
287}
288
289let Predicates = [OptForSize, Not64BitMode],
290    AddedComplexity = 10 in {
291  let SchedRW = [WriteALU] in {
292  // Pseudo instructions for materializing 1 and -1 using XOR+INC/DEC,
293  // which only require 3 bytes compared to MOV32ri which requires 5.
294  let Defs = [EFLAGS], isReMaterializable = 1, isPseudo = 1 in {
295    def MOV32r1 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
296                        [(set GR32:$dst, 1)]>;
297    def MOV32r_1 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
298                        [(set GR32:$dst, -1)]>;
299  }
300  } // SchedRW
301
302  // MOV16ri is 4 bytes, so the instructions above are smaller.
303  def : Pat<(i16 1), (EXTRACT_SUBREG (MOV32r1), sub_16bit)>;
304  def : Pat<(i16 -1), (EXTRACT_SUBREG (MOV32r_1), sub_16bit)>;
305}
306
307let isReMaterializable = 1, isPseudo = 1, AddedComplexity = 5,
308    SchedRW = [WriteALU] in {
309// AddedComplexity higher than MOV64ri but lower than MOV32r0 and MOV32r1.
310def MOV32ImmSExti8 : I<0, Pseudo, (outs GR32:$dst), (ins i32i8imm:$src), "",
311                       [(set GR32:$dst, i32immSExt8:$src)]>,
312                       Requires<[OptForMinSize, NotWin64WithoutFP]>;
313def MOV64ImmSExti8 : I<0, Pseudo, (outs GR64:$dst), (ins i64i8imm:$src), "",
314                       [(set GR64:$dst, i64immSExt8:$src)]>,
315                       Requires<[OptForMinSize, NotWin64WithoutFP]>;
316}
317
318// Materialize i64 constant where top 32-bits are zero. This could theoretically
319// use MOV32ri with a SUBREG_TO_REG to represent the zero-extension, however
320// that would make it more difficult to rematerialize.
321let AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1,
322    isPseudo = 1, SchedRW = [WriteMove] in
323def MOV32ri64 : I<0, Pseudo, (outs GR64:$dst), (ins i64i32imm:$src), "",
324                  [(set GR64:$dst, i64immZExt32:$src)]>;
325
326// This 64-bit pseudo-move can also be used for labels in the x86-64 small code
327// model.
328def mov64imm32 : ComplexPattern<i64, 1, "selectMOV64Imm32", [X86Wrapper]>;
329def : Pat<(i64 mov64imm32:$src), (MOV32ri64 mov64imm32:$src)>;
330
331// Use sbb to materialize carry bit.
332let Uses = [EFLAGS], Defs = [EFLAGS], isPseudo = 1, SchedRW = [WriteADC],
333    hasSideEffects = 0 in {
334// FIXME: These are pseudo ops that should be replaced with Pat<> patterns.
335// However, Pat<> can't replicate the destination reg into the inputs of the
336// result.
337def SETB_C32r : I<0, Pseudo, (outs GR32:$dst), (ins), "", []>;
338def SETB_C64r : I<0, Pseudo, (outs GR64:$dst), (ins), "", []>;
339} // isCodeGenOnly
340
341//===----------------------------------------------------------------------===//
342// String Pseudo Instructions
343//
344let SchedRW = [WriteMicrocoded] in {
345let Defs = [ECX,EDI,ESI], Uses = [ECX,EDI,ESI], isCodeGenOnly = 1 in {
346def REP_MOVSB_32 : I<0xA4, RawFrm, (outs), (ins),
347                    "{rep;movsb (%esi), %es:(%edi)|rep movsb es:[edi], [esi]}",
348                    [(X86rep_movs i8)]>, REP, AdSize32,
349                   Requires<[NotLP64]>;
350def REP_MOVSW_32 : I<0xA5, RawFrm, (outs), (ins),
351                    "{rep;movsw (%esi), %es:(%edi)|rep movsw es:[edi], [esi]}",
352                    [(X86rep_movs i16)]>, REP, AdSize32, OpSize16,
353                   Requires<[NotLP64]>;
354def REP_MOVSD_32 : I<0xA5, RawFrm, (outs), (ins),
355                    "{rep;movsl (%esi), %es:(%edi)|rep movsd es:[edi], [esi]}",
356                    [(X86rep_movs i32)]>, REP, AdSize32, OpSize32,
357                   Requires<[NotLP64]>;
358def REP_MOVSQ_32 : RI<0xA5, RawFrm, (outs), (ins),
359                    "{rep;movsq (%esi), %es:(%edi)|rep movsq es:[edi], [esi]}",
360                    [(X86rep_movs i64)]>, REP, AdSize32,
361                   Requires<[NotLP64, In64BitMode]>;
362}
363
364let Defs = [RCX,RDI,RSI], Uses = [RCX,RDI,RSI], isCodeGenOnly = 1 in {
365def REP_MOVSB_64 : I<0xA4, RawFrm, (outs), (ins),
366                    "{rep;movsb (%rsi), %es:(%rdi)|rep movsb es:[rdi], [rsi]}",
367                    [(X86rep_movs i8)]>, REP, AdSize64,
368                   Requires<[IsLP64]>;
369def REP_MOVSW_64 : I<0xA5, RawFrm, (outs), (ins),
370                    "{rep;movsw (%rsi), %es:(%rdi)|rep movsw es:[rdi], [rsi]}",
371                    [(X86rep_movs i16)]>, REP, AdSize64, OpSize16,
372                   Requires<[IsLP64]>;
373def REP_MOVSD_64 : I<0xA5, RawFrm, (outs), (ins),
374                    "{rep;movsl (%rsi), %es:(%rdi)|rep movsdi es:[rdi], [rsi]}",
375                    [(X86rep_movs i32)]>, REP, AdSize64, OpSize32,
376                   Requires<[IsLP64]>;
377def REP_MOVSQ_64 : RI<0xA5, RawFrm, (outs), (ins),
378                    "{rep;movsq (%rsi), %es:(%rdi)|rep movsq es:[rdi], [rsi]}",
379                    [(X86rep_movs i64)]>, REP, AdSize64,
380                   Requires<[IsLP64]>;
381}
382
383// FIXME: Should use "(X86rep_stos AL)" as the pattern.
384let Defs = [ECX,EDI], isCodeGenOnly = 1 in {
385  let Uses = [AL,ECX,EDI] in
386  def REP_STOSB_32 : I<0xAA, RawFrm, (outs), (ins),
387                       "{rep;stosb %al, %es:(%edi)|rep stosb es:[edi], al}",
388                      [(X86rep_stos i8)]>, REP, AdSize32,
389                     Requires<[NotLP64]>;
390  let Uses = [AX,ECX,EDI] in
391  def REP_STOSW_32 : I<0xAB, RawFrm, (outs), (ins),
392                      "{rep;stosw %ax, %es:(%edi)|rep stosw es:[edi], ax}",
393                      [(X86rep_stos i16)]>, REP, AdSize32, OpSize16,
394                     Requires<[NotLP64]>;
395  let Uses = [EAX,ECX,EDI] in
396  def REP_STOSD_32 : I<0xAB, RawFrm, (outs), (ins),
397                      "{rep;stosl %eax, %es:(%edi)|rep stosd es:[edi], eax}",
398                      [(X86rep_stos i32)]>, REP, AdSize32, OpSize32,
399                     Requires<[NotLP64]>;
400  let Uses = [RAX,RCX,RDI] in
401  def REP_STOSQ_32 : RI<0xAB, RawFrm, (outs), (ins),
402                        "{rep;stosq %rax, %es:(%edi)|rep stosq es:[edi], rax}",
403                        [(X86rep_stos i64)]>, REP, AdSize32,
404                        Requires<[NotLP64, In64BitMode]>;
405}
406
407let Defs = [RCX,RDI], isCodeGenOnly = 1 in {
408  let Uses = [AL,RCX,RDI] in
409  def REP_STOSB_64 : I<0xAA, RawFrm, (outs), (ins),
410                       "{rep;stosb %al, %es:(%rdi)|rep stosb es:[rdi], al}",
411                       [(X86rep_stos i8)]>, REP, AdSize64,
412                       Requires<[IsLP64]>;
413  let Uses = [AX,RCX,RDI] in
414  def REP_STOSW_64 : I<0xAB, RawFrm, (outs), (ins),
415                       "{rep;stosw %ax, %es:(%rdi)|rep stosw es:[rdi], ax}",
416                       [(X86rep_stos i16)]>, REP, AdSize64, OpSize16,
417                       Requires<[IsLP64]>;
418  let Uses = [RAX,RCX,RDI] in
419  def REP_STOSD_64 : I<0xAB, RawFrm, (outs), (ins),
420                      "{rep;stosl %eax, %es:(%rdi)|rep stosd es:[rdi], eax}",
421                       [(X86rep_stos i32)]>, REP, AdSize64, OpSize32,
422                       Requires<[IsLP64]>;
423
424  let Uses = [RAX,RCX,RDI] in
425  def REP_STOSQ_64 : RI<0xAB, RawFrm, (outs), (ins),
426                        "{rep;stosq %rax, %es:(%rdi)|rep stosq es:[rdi], rax}",
427                        [(X86rep_stos i64)]>, REP, AdSize64,
428                        Requires<[IsLP64]>;
429}
430} // SchedRW
431
432//===----------------------------------------------------------------------===//
433// Thread Local Storage Instructions
434//
435let SchedRW = [WriteSystem] in {
436
437// ELF TLS Support
438// All calls clobber the non-callee saved registers. ESP is marked as
439// a use to prevent stack-pointer assignments that appear immediately
440// before calls from potentially appearing dead.
441let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, FP7,
442            ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7,
443            MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
444            XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
445            XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS, DF],
446    usesCustomInserter = 1, Uses = [ESP, SSP] in {
447def TLS_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
448                  "# TLS_addr32",
449                  [(X86tlsaddr tls32addr:$sym)]>,
450                  Requires<[Not64BitMode]>;
451def TLS_base_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
452                  "# TLS_base_addr32",
453                  [(X86tlsbaseaddr tls32baseaddr:$sym)]>,
454                  Requires<[Not64BitMode]>;
455}
456
457// All calls clobber the non-callee saved registers. RSP is marked as
458// a use to prevent stack-pointer assignments that appear immediately
459// before calls from potentially appearing dead.
460let Defs = [RAX, RCX, RDX, RSI, RDI, R8, R9, R10, R11,
461            FP0, FP1, FP2, FP3, FP4, FP5, FP6, FP7,
462            ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7,
463            MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
464            XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
465            XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS, DF],
466    usesCustomInserter = 1, Uses = [RSP, SSP] in {
467def TLS_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
468                   "# TLS_addr64",
469                  [(X86tlsaddr tls64addr:$sym)]>,
470                  Requires<[In64BitMode]>;
471def TLS_base_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
472                   "# TLS_base_addr64",
473                  [(X86tlsbaseaddr tls64baseaddr:$sym)]>,
474                  Requires<[In64BitMode]>;
475}
476
477// Darwin TLS Support
478// For i386, the address of the thunk is passed on the stack, on return the
479// address of the variable is in %eax.  %ecx is trashed during the function
480// call.  All other registers are preserved.
481let Defs = [EAX, ECX, EFLAGS, DF],
482    Uses = [ESP, SSP],
483    usesCustomInserter = 1 in
484def TLSCall_32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
485                "# TLSCall_32",
486                [(X86TLSCall addr:$sym)]>,
487                Requires<[Not64BitMode]>;
488
489// For x86_64, the address of the thunk is passed in %rdi, but the
490// pseudo directly use the symbol, so do not add an implicit use of
491// %rdi. The lowering will do the right thing with RDI.
492// On return the address of the variable is in %rax.  All other
493// registers are preserved.
494let Defs = [RAX, EFLAGS, DF],
495    Uses = [RSP, SSP],
496    usesCustomInserter = 1 in
497def TLSCall_64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
498                  "# TLSCall_64",
499                  [(X86TLSCall addr:$sym)]>,
500                  Requires<[In64BitMode]>;
501} // SchedRW
502
503//===----------------------------------------------------------------------===//
504// Conditional Move Pseudo Instructions
505
506// CMOV* - Used to implement the SELECT DAG operation.  Expanded after
507// instruction selection into a branch sequence.
508multiclass CMOVrr_PSEUDO<RegisterClass RC, ValueType VT> {
509  def CMOV#NAME  : I<0, Pseudo,
510                    (outs RC:$dst), (ins RC:$t, RC:$f, i8imm:$cond),
511                    "#CMOV_"#NAME#" PSEUDO!",
512                    [(set RC:$dst, (VT (X86cmov RC:$t, RC:$f, timm:$cond,
513                                                EFLAGS)))]>;
514}
515
516let usesCustomInserter = 1, hasNoSchedulingInfo = 1, Uses = [EFLAGS] in {
517  // X86 doesn't have 8-bit conditional moves. Use a customInserter to
518  // emit control flow. An alternative to this is to mark i8 SELECT as Promote,
519  // however that requires promoting the operands, and can induce additional
520  // i8 register pressure.
521  defm _GR8 : CMOVrr_PSEUDO<GR8, i8>;
522
523  let Predicates = [NoCMov] in {
524    defm _GR32 : CMOVrr_PSEUDO<GR32, i32>;
525    defm _GR16 : CMOVrr_PSEUDO<GR16, i16>;
526  } // Predicates = [NoCMov]
527
528  // fcmov doesn't handle all possible EFLAGS, provide a fallback if there is no
529  // SSE1/SSE2.
530  let Predicates = [FPStackf32] in
531    defm _RFP32 : CMOVrr_PSEUDO<RFP32, f32>;
532
533  let Predicates = [FPStackf64] in
534    defm _RFP64 : CMOVrr_PSEUDO<RFP64, f64>;
535
536  defm _RFP80 : CMOVrr_PSEUDO<RFP80, f80>;
537
538  let Predicates = [HasMMX] in
539    defm _VR64   : CMOVrr_PSEUDO<VR64, x86mmx>;
540
541  let Predicates = [HasSSE1,NoAVX512] in
542    defm _FR32   : CMOVrr_PSEUDO<FR32, f32>;
543  let Predicates = [HasSSE2,NoAVX512] in
544    defm _FR64   : CMOVrr_PSEUDO<FR64, f64>;
545  let Predicates = [HasAVX512] in {
546    defm _FR32X  : CMOVrr_PSEUDO<FR32X, f32>;
547    defm _FR64X  : CMOVrr_PSEUDO<FR64X, f64>;
548  }
549  let Predicates = [NoVLX] in {
550    defm _VR128  : CMOVrr_PSEUDO<VR128, v2i64>;
551    defm _VR256  : CMOVrr_PSEUDO<VR256, v4i64>;
552  }
553  let Predicates = [HasVLX] in {
554    defm _VR128X : CMOVrr_PSEUDO<VR128X, v2i64>;
555    defm _VR256X : CMOVrr_PSEUDO<VR256X, v4i64>;
556  }
557  defm _VR512  : CMOVrr_PSEUDO<VR512, v8i64>;
558  defm _VK1    : CMOVrr_PSEUDO<VK1,  v1i1>;
559  defm _VK2    : CMOVrr_PSEUDO<VK2,  v2i1>;
560  defm _VK4    : CMOVrr_PSEUDO<VK4,  v4i1>;
561  defm _VK8    : CMOVrr_PSEUDO<VK8,  v8i1>;
562  defm _VK16   : CMOVrr_PSEUDO<VK16, v16i1>;
563  defm _VK32   : CMOVrr_PSEUDO<VK32, v32i1>;
564  defm _VK64   : CMOVrr_PSEUDO<VK64, v64i1>;
565} // usesCustomInserter = 1, hasNoSchedulingInfo = 1, Uses = [EFLAGS]
566
567def : Pat<(f128 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
568          (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
569
570let Predicates = [NoVLX] in {
571  def : Pat<(v16i8 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
572            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
573  def : Pat<(v8i16 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
574            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
575  def : Pat<(v4i32 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
576            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
577  def : Pat<(v4f32 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
578            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
579  def : Pat<(v2f64 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
580            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
581
582  def : Pat<(v32i8 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
583            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
584  def : Pat<(v16i16 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
585            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
586  def : Pat<(v8i32 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
587            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
588  def : Pat<(v8f32 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
589            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
590  def : Pat<(v4f64 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
591            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
592}
593let Predicates = [HasVLX] in {
594  def : Pat<(v16i8 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
595            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
596  def : Pat<(v8i16 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
597            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
598  def : Pat<(v4i32 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
599            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
600  def : Pat<(v4f32 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
601            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
602  def : Pat<(v2f64 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
603            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
604
605  def : Pat<(v32i8 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
606            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
607  def : Pat<(v16i16 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
608            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
609  def : Pat<(v8i32 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
610            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
611  def : Pat<(v8f32 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
612            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
613  def : Pat<(v4f64 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
614            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
615}
616
617def : Pat<(v64i8 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
618          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
619def : Pat<(v32i16 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
620          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
621def : Pat<(v16i32 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
622          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
623def : Pat<(v16f32 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
624          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
625def : Pat<(v8f64 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
626          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
627
628//===----------------------------------------------------------------------===//
629// Normal-Instructions-With-Lock-Prefix Pseudo Instructions
630//===----------------------------------------------------------------------===//
631
632// FIXME: Use normal instructions and add lock prefix dynamically.
633
634// Memory barriers
635
636let isCodeGenOnly = 1, Defs = [EFLAGS] in
637def OR32mi8Locked  : Ii8<0x83, MRM1m, (outs), (ins i32mem:$dst, i32i8imm:$zero),
638                         "or{l}\t{$zero, $dst|$dst, $zero}", []>,
639                         Requires<[Not64BitMode]>, OpSize32, LOCK,
640                         Sched<[WriteALURMW]>;
641
642let hasSideEffects = 1 in
643def Int_MemBarrier : I<0, Pseudo, (outs), (ins),
644                     "#MEMBARRIER",
645                     [(X86MemBarrier)]>, Sched<[WriteLoad]>;
646
647// RegOpc corresponds to the mr version of the instruction
648// ImmOpc corresponds to the mi version of the instruction
649// ImmOpc8 corresponds to the mi8 version of the instruction
650// ImmMod corresponds to the instruction format of the mi and mi8 versions
651multiclass LOCK_ArithBinOp<bits<8> RegOpc, bits<8> ImmOpc, bits<8> ImmOpc8,
652                           Format ImmMod, SDNode Op, string mnemonic> {
653let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
654    SchedRW = [WriteALURMW] in {
655
656def NAME#8mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
657                  RegOpc{3}, RegOpc{2}, RegOpc{1}, 0 },
658                  MRMDestMem, (outs), (ins i8mem:$dst, GR8:$src2),
659                  !strconcat(mnemonic, "{b}\t",
660                             "{$src2, $dst|$dst, $src2}"),
661                  [(set EFLAGS, (Op addr:$dst, GR8:$src2))]>, LOCK;
662
663def NAME#16mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
664                   RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
665                   MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
666                   !strconcat(mnemonic, "{w}\t",
667                              "{$src2, $dst|$dst, $src2}"),
668                   [(set EFLAGS, (Op addr:$dst, GR16:$src2))]>,
669                   OpSize16, LOCK;
670
671def NAME#32mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
672                   RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
673                   MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
674                   !strconcat(mnemonic, "{l}\t",
675                              "{$src2, $dst|$dst, $src2}"),
676                   [(set EFLAGS, (Op addr:$dst, GR32:$src2))]>,
677                   OpSize32, LOCK;
678
679def NAME#64mr : RI<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
680                    RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
681                    MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
682                    !strconcat(mnemonic, "{q}\t",
683                               "{$src2, $dst|$dst, $src2}"),
684                    [(set EFLAGS, (Op addr:$dst, GR64:$src2))]>, LOCK;
685
686// NOTE: These are order specific, we want the mi8 forms to be listed
687// first so that they are slightly preferred to the mi forms.
688def NAME#16mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
689                      ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
690                      ImmMod, (outs), (ins i16mem :$dst, i16i8imm :$src2),
691                      !strconcat(mnemonic, "{w}\t",
692                                 "{$src2, $dst|$dst, $src2}"),
693                      [(set EFLAGS, (Op addr:$dst, i16immSExt8:$src2))]>,
694                      OpSize16, LOCK;
695
696def NAME#32mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
697                      ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
698                      ImmMod, (outs), (ins i32mem :$dst, i32i8imm :$src2),
699                      !strconcat(mnemonic, "{l}\t",
700                                 "{$src2, $dst|$dst, $src2}"),
701                      [(set EFLAGS, (Op addr:$dst, i32immSExt8:$src2))]>,
702                      OpSize32, LOCK;
703
704def NAME#64mi8 : RIi8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
705                       ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
706                       ImmMod, (outs), (ins i64mem :$dst, i64i8imm :$src2),
707                       !strconcat(mnemonic, "{q}\t",
708                                  "{$src2, $dst|$dst, $src2}"),
709                       [(set EFLAGS, (Op addr:$dst, i64immSExt8:$src2))]>,
710                       LOCK;
711
712def NAME#8mi : Ii8<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
713                    ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 0 },
714                    ImmMod, (outs), (ins i8mem :$dst, i8imm :$src2),
715                    !strconcat(mnemonic, "{b}\t",
716                               "{$src2, $dst|$dst, $src2}"),
717                    [(set EFLAGS, (Op addr:$dst, (i8 imm:$src2)))]>, LOCK;
718
719def NAME#16mi : Ii16<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
720                      ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
721                      ImmMod, (outs), (ins i16mem :$dst, i16imm :$src2),
722                      !strconcat(mnemonic, "{w}\t",
723                                 "{$src2, $dst|$dst, $src2}"),
724                      [(set EFLAGS, (Op addr:$dst, (i16 imm:$src2)))]>,
725                      OpSize16, LOCK;
726
727def NAME#32mi : Ii32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
728                      ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
729                      ImmMod, (outs), (ins i32mem :$dst, i32imm :$src2),
730                      !strconcat(mnemonic, "{l}\t",
731                                 "{$src2, $dst|$dst, $src2}"),
732                      [(set EFLAGS, (Op addr:$dst, (i32 imm:$src2)))]>,
733                      OpSize32, LOCK;
734
735def NAME#64mi32 : RIi32S<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
736                          ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
737                          ImmMod, (outs), (ins i64mem :$dst, i64i32imm :$src2),
738                          !strconcat(mnemonic, "{q}\t",
739                                     "{$src2, $dst|$dst, $src2}"),
740                          [(set EFLAGS, (Op addr:$dst, i64immSExt32:$src2))]>,
741                          LOCK;
742}
743
744}
745
746defm LOCK_ADD : LOCK_ArithBinOp<0x00, 0x80, 0x83, MRM0m, X86lock_add, "add">;
747defm LOCK_SUB : LOCK_ArithBinOp<0x28, 0x80, 0x83, MRM5m, X86lock_sub, "sub">;
748defm LOCK_OR  : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM1m, X86lock_or , "or">;
749defm LOCK_AND : LOCK_ArithBinOp<0x20, 0x80, 0x83, MRM4m, X86lock_and, "and">;
750defm LOCK_XOR : LOCK_ArithBinOp<0x30, 0x80, 0x83, MRM6m, X86lock_xor, "xor">;
751
752def X86lock_add_nocf : PatFrag<(ops node:$lhs, node:$rhs),
753                               (X86lock_add node:$lhs, node:$rhs), [{
754  return hasNoCarryFlagUses(SDValue(N, 0));
755}]>;
756
757def X86lock_sub_nocf : PatFrag<(ops node:$lhs, node:$rhs),
758                               (X86lock_sub node:$lhs, node:$rhs), [{
759  return hasNoCarryFlagUses(SDValue(N, 0));
760}]>;
761
762let Predicates = [UseIncDec] in {
763  let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
764      SchedRW = [WriteALURMW]  in {
765    def LOCK_INC8m  : I<0xFE, MRM0m, (outs), (ins i8mem :$dst),
766                        "inc{b}\t$dst",
767                        [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i8 1)))]>,
768                        LOCK;
769    def LOCK_INC16m : I<0xFF, MRM0m, (outs), (ins i16mem:$dst),
770                        "inc{w}\t$dst",
771                        [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i16 1)))]>,
772                        OpSize16, LOCK;
773    def LOCK_INC32m : I<0xFF, MRM0m, (outs), (ins i32mem:$dst),
774                        "inc{l}\t$dst",
775                        [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i32 1)))]>,
776                        OpSize32, LOCK;
777    def LOCK_INC64m : RI<0xFF, MRM0m, (outs), (ins i64mem:$dst),
778                         "inc{q}\t$dst",
779                         [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i64 1)))]>,
780                         LOCK;
781
782    def LOCK_DEC8m  : I<0xFE, MRM1m, (outs), (ins i8mem :$dst),
783                        "dec{b}\t$dst",
784                        [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i8 1)))]>,
785                        LOCK;
786    def LOCK_DEC16m : I<0xFF, MRM1m, (outs), (ins i16mem:$dst),
787                        "dec{w}\t$dst",
788                        [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i16 1)))]>,
789                        OpSize16, LOCK;
790    def LOCK_DEC32m : I<0xFF, MRM1m, (outs), (ins i32mem:$dst),
791                        "dec{l}\t$dst",
792                        [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i32 1)))]>,
793                        OpSize32, LOCK;
794    def LOCK_DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst),
795                         "dec{q}\t$dst",
796                         [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i64 1)))]>,
797                         LOCK;
798  }
799
800  // Additional patterns for -1 constant.
801  def : Pat<(X86lock_add addr:$dst, (i8  -1)), (LOCK_DEC8m  addr:$dst)>;
802  def : Pat<(X86lock_add addr:$dst, (i16 -1)), (LOCK_DEC16m addr:$dst)>;
803  def : Pat<(X86lock_add addr:$dst, (i32 -1)), (LOCK_DEC32m addr:$dst)>;
804  def : Pat<(X86lock_add addr:$dst, (i64 -1)), (LOCK_DEC64m addr:$dst)>;
805  def : Pat<(X86lock_sub addr:$dst, (i8  -1)), (LOCK_INC8m  addr:$dst)>;
806  def : Pat<(X86lock_sub addr:$dst, (i16 -1)), (LOCK_INC16m addr:$dst)>;
807  def : Pat<(X86lock_sub addr:$dst, (i32 -1)), (LOCK_INC32m addr:$dst)>;
808  def : Pat<(X86lock_sub addr:$dst, (i64 -1)), (LOCK_INC64m addr:$dst)>;
809}
810
811// Atomic compare and swap.
812multiclass LCMPXCHG_UnOp<bits<8> Opc, Format Form, string mnemonic,
813                         SDPatternOperator frag, X86MemOperand x86memop> {
814let isCodeGenOnly = 1, usesCustomInserter = 1 in {
815  def NAME : I<Opc, Form, (outs), (ins x86memop:$ptr),
816               !strconcat(mnemonic, "\t$ptr"),
817               [(frag addr:$ptr)]>, TB, LOCK;
818}
819}
820
821multiclass LCMPXCHG_BinOp<bits<8> Opc8, bits<8> Opc, Format Form,
822                          string mnemonic, SDPatternOperator frag> {
823let isCodeGenOnly = 1, SchedRW = [WriteCMPXCHGRMW] in {
824  let Defs = [AL, EFLAGS], Uses = [AL] in
825  def NAME#8  : I<Opc8, Form, (outs), (ins i8mem:$ptr, GR8:$swap),
826                  !strconcat(mnemonic, "{b}\t{$swap, $ptr|$ptr, $swap}"),
827                  [(frag addr:$ptr, GR8:$swap, 1)]>, TB, LOCK;
828  let Defs = [AX, EFLAGS], Uses = [AX] in
829  def NAME#16 : I<Opc, Form, (outs), (ins i16mem:$ptr, GR16:$swap),
830                  !strconcat(mnemonic, "{w}\t{$swap, $ptr|$ptr, $swap}"),
831                  [(frag addr:$ptr, GR16:$swap, 2)]>, TB, OpSize16, LOCK;
832  let Defs = [EAX, EFLAGS], Uses = [EAX] in
833  def NAME#32 : I<Opc, Form, (outs), (ins i32mem:$ptr, GR32:$swap),
834                  !strconcat(mnemonic, "{l}\t{$swap, $ptr|$ptr, $swap}"),
835                  [(frag addr:$ptr, GR32:$swap, 4)]>, TB, OpSize32, LOCK;
836  let Defs = [RAX, EFLAGS], Uses = [RAX] in
837  def NAME#64 : RI<Opc, Form, (outs), (ins i64mem:$ptr, GR64:$swap),
838                   !strconcat(mnemonic, "{q}\t{$swap, $ptr|$ptr, $swap}"),
839                   [(frag addr:$ptr, GR64:$swap, 8)]>, TB, LOCK;
840}
841}
842
843let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
844    Predicates = [HasCmpxchg8b], SchedRW = [WriteCMPXCHGRMW] in {
845defm LCMPXCHG8B : LCMPXCHG_UnOp<0xC7, MRM1m, "cmpxchg8b", X86cas8, i64mem>;
846}
847
848// This pseudo must be used when the frame uses RBX as
849// the base pointer. Indeed, in such situation RBX is a reserved
850// register and the register allocator will ignore any use/def of
851// it. In other words, the register will not fix the clobbering of
852// RBX that will happen when setting the arguments for the instrucion.
853//
854// Unlike the actual related instruction, we mark that this one
855// defines EBX (instead of using EBX).
856// The rationale is that we will define RBX during the expansion of
857// the pseudo. The argument feeding EBX is ebx_input.
858//
859// The additional argument, $ebx_save, is a temporary register used to
860// save the value of RBX across the actual instruction.
861//
862// To make sure the register assigned to $ebx_save does not interfere with
863// the definition of the actual instruction, we use a definition $dst which
864// is tied to $rbx_save. That way, the live-range of $rbx_save spans across
865// the instruction and we are sure we will have a valid register to restore
866// the value of RBX.
867let Defs = [EAX, EDX, EBX, EFLAGS], Uses = [EAX, ECX, EDX],
868    Predicates = [HasCmpxchg8b], SchedRW = [WriteCMPXCHGRMW],
869    isCodeGenOnly = 1, isPseudo = 1, Constraints = "$ebx_save = $dst",
870    usesCustomInserter = 1 in {
871def LCMPXCHG8B_SAVE_EBX :
872    I<0, Pseudo, (outs GR32:$dst),
873      (ins i64mem:$ptr, GR32:$ebx_input, GR32:$ebx_save),
874      !strconcat("cmpxchg8b", "\t$ptr"),
875      [(set GR32:$dst, (X86cas8save_ebx addr:$ptr, GR32:$ebx_input,
876                                        GR32:$ebx_save))]>;
877}
878
879
880let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RBX, RCX, RDX],
881    Predicates = [HasCmpxchg16b,In64BitMode], SchedRW = [WriteCMPXCHGRMW] in {
882defm LCMPXCHG16B : LCMPXCHG_UnOp<0xC7, MRM1m, "cmpxchg16b",
883                                 X86cas16, i128mem>, REX_W;
884}
885
886// Same as LCMPXCHG8B_SAVE_RBX but for the 16 Bytes variant.
887let Defs = [RAX, RDX, RBX, EFLAGS], Uses = [RAX, RCX, RDX],
888    Predicates = [HasCmpxchg16b,In64BitMode], SchedRW = [WriteCMPXCHGRMW],
889    isCodeGenOnly = 1, isPseudo = 1, Constraints = "$rbx_save = $dst",
890    usesCustomInserter = 1 in {
891def LCMPXCHG16B_SAVE_RBX :
892    I<0, Pseudo, (outs GR64:$dst),
893      (ins i128mem:$ptr, GR64:$rbx_input, GR64:$rbx_save),
894      !strconcat("cmpxchg16b", "\t$ptr"),
895      [(set GR64:$dst, (X86cas16save_rbx addr:$ptr, GR64:$rbx_input,
896                                                    GR64:$rbx_save))]>;
897}
898
899defm LCMPXCHG : LCMPXCHG_BinOp<0xB0, 0xB1, MRMDestMem, "cmpxchg", X86cas>;
900
901// Atomic exchange and add
902multiclass ATOMIC_LOAD_BINOP<bits<8> opc8, bits<8> opc, string mnemonic,
903                             string frag> {
904  let Constraints = "$val = $dst", Defs = [EFLAGS], isCodeGenOnly = 1,
905      SchedRW = [WriteALURMW] in {
906    def NAME#8  : I<opc8, MRMSrcMem, (outs GR8:$dst),
907                    (ins GR8:$val, i8mem:$ptr),
908                    !strconcat(mnemonic, "{b}\t{$val, $ptr|$ptr, $val}"),
909                    [(set GR8:$dst,
910                          (!cast<PatFrag>(frag # "_8") addr:$ptr, GR8:$val))]>;
911    def NAME#16 : I<opc, MRMSrcMem, (outs GR16:$dst),
912                    (ins GR16:$val, i16mem:$ptr),
913                    !strconcat(mnemonic, "{w}\t{$val, $ptr|$ptr, $val}"),
914                    [(set
915                       GR16:$dst,
916                       (!cast<PatFrag>(frag # "_16") addr:$ptr, GR16:$val))]>,
917                    OpSize16;
918    def NAME#32 : I<opc, MRMSrcMem, (outs GR32:$dst),
919                    (ins GR32:$val, i32mem:$ptr),
920                    !strconcat(mnemonic, "{l}\t{$val, $ptr|$ptr, $val}"),
921                    [(set
922                       GR32:$dst,
923                       (!cast<PatFrag>(frag # "_32") addr:$ptr, GR32:$val))]>,
924                    OpSize32;
925    def NAME#64 : RI<opc, MRMSrcMem, (outs GR64:$dst),
926                     (ins GR64:$val, i64mem:$ptr),
927                     !strconcat(mnemonic, "{q}\t{$val, $ptr|$ptr, $val}"),
928                     [(set
929                        GR64:$dst,
930                        (!cast<PatFrag>(frag # "_64") addr:$ptr, GR64:$val))]>;
931  }
932}
933
934defm LXADD : ATOMIC_LOAD_BINOP<0xc0, 0xc1, "xadd", "atomic_load_add">, TB, LOCK;
935
936/* The following multiclass tries to make sure that in code like
937 *    x.store (immediate op x.load(acquire), release)
938 * and
939 *    x.store (register op x.load(acquire), release)
940 * an operation directly on memory is generated instead of wasting a register.
941 * It is not automatic as atomic_store/load are only lowered to MOV instructions
942 * extremely late to prevent them from being accidentally reordered in the backend
943 * (see below the RELEASE_MOV* / ACQUIRE_MOV* pseudo-instructions)
944 */
945multiclass RELEASE_BINOP_MI<string Name, SDNode op> {
946  def : Pat<(atomic_store_8 addr:$dst,
947             (op (atomic_load_8 addr:$dst), (i8 imm:$src))),
948            (!cast<Instruction>(Name#"8mi") addr:$dst, imm:$src)>;
949  def : Pat<(atomic_store_16 addr:$dst,
950             (op (atomic_load_16 addr:$dst), (i16 imm:$src))),
951            (!cast<Instruction>(Name#"16mi") addr:$dst, imm:$src)>;
952  def : Pat<(atomic_store_32 addr:$dst,
953             (op (atomic_load_32 addr:$dst), (i32 imm:$src))),
954            (!cast<Instruction>(Name#"32mi") addr:$dst, imm:$src)>;
955  def : Pat<(atomic_store_64 addr:$dst,
956             (op (atomic_load_64 addr:$dst), (i64immSExt32:$src))),
957            (!cast<Instruction>(Name#"64mi32") addr:$dst, (i64immSExt32:$src))>;
958
959  def : Pat<(atomic_store_8 addr:$dst,
960             (op (atomic_load_8 addr:$dst), (i8 GR8:$src))),
961            (!cast<Instruction>(Name#"8mr") addr:$dst, GR8:$src)>;
962  def : Pat<(atomic_store_16 addr:$dst,
963             (op (atomic_load_16 addr:$dst), (i16 GR16:$src))),
964            (!cast<Instruction>(Name#"16mr") addr:$dst, GR16:$src)>;
965  def : Pat<(atomic_store_32 addr:$dst,
966             (op (atomic_load_32 addr:$dst), (i32 GR32:$src))),
967            (!cast<Instruction>(Name#"32mr") addr:$dst, GR32:$src)>;
968  def : Pat<(atomic_store_64 addr:$dst,
969             (op (atomic_load_64 addr:$dst), (i64 GR64:$src))),
970            (!cast<Instruction>(Name#"64mr") addr:$dst, GR64:$src)>;
971}
972defm : RELEASE_BINOP_MI<"ADD", add>;
973defm : RELEASE_BINOP_MI<"AND", and>;
974defm : RELEASE_BINOP_MI<"OR",  or>;
975defm : RELEASE_BINOP_MI<"XOR", xor>;
976defm : RELEASE_BINOP_MI<"SUB", sub>;
977
978// Atomic load + floating point patterns.
979// FIXME: This could also handle SIMD operations with *ps and *pd instructions.
980multiclass ATOMIC_LOAD_FP_BINOP_MI<string Name, SDNode op> {
981  def : Pat<(op FR32:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
982            (!cast<Instruction>(Name#"SSrm") FR32:$src1, addr:$src2)>,
983            Requires<[UseSSE1]>;
984  def : Pat<(op FR32:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
985            (!cast<Instruction>("V"#Name#"SSrm") FR32:$src1, addr:$src2)>,
986            Requires<[UseAVX]>;
987  def : Pat<(op FR32X:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
988            (!cast<Instruction>("V"#Name#"SSZrm") FR32X:$src1, addr:$src2)>,
989            Requires<[HasAVX512]>;
990
991  def : Pat<(op FR64:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
992            (!cast<Instruction>(Name#"SDrm") FR64:$src1, addr:$src2)>,
993            Requires<[UseSSE1]>;
994  def : Pat<(op FR64:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
995            (!cast<Instruction>("V"#Name#"SDrm") FR64:$src1, addr:$src2)>,
996            Requires<[UseAVX]>;
997  def : Pat<(op FR64X:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
998            (!cast<Instruction>("V"#Name#"SDZrm") FR64X:$src1, addr:$src2)>,
999            Requires<[HasAVX512]>;
1000}
1001defm : ATOMIC_LOAD_FP_BINOP_MI<"ADD", fadd>;
1002// FIXME: Add fsub, fmul, fdiv, ...
1003
1004multiclass RELEASE_UNOP<string Name, dag dag8, dag dag16, dag dag32,
1005                        dag dag64> {
1006  def : Pat<(atomic_store_8 addr:$dst, dag8),
1007            (!cast<Instruction>(Name#8m) addr:$dst)>;
1008  def : Pat<(atomic_store_16 addr:$dst, dag16),
1009            (!cast<Instruction>(Name#16m) addr:$dst)>;
1010  def : Pat<(atomic_store_32 addr:$dst, dag32),
1011            (!cast<Instruction>(Name#32m) addr:$dst)>;
1012  def : Pat<(atomic_store_64 addr:$dst, dag64),
1013            (!cast<Instruction>(Name#64m) addr:$dst)>;
1014}
1015
1016let Predicates = [UseIncDec] in {
1017  defm : RELEASE_UNOP<"INC",
1018      (add (atomic_load_8  addr:$dst), (i8 1)),
1019      (add (atomic_load_16 addr:$dst), (i16 1)),
1020      (add (atomic_load_32 addr:$dst), (i32 1)),
1021      (add (atomic_load_64 addr:$dst), (i64 1))>;
1022  defm : RELEASE_UNOP<"DEC",
1023      (add (atomic_load_8  addr:$dst), (i8 -1)),
1024      (add (atomic_load_16 addr:$dst), (i16 -1)),
1025      (add (atomic_load_32 addr:$dst), (i32 -1)),
1026      (add (atomic_load_64 addr:$dst), (i64 -1))>;
1027}
1028
1029defm : RELEASE_UNOP<"NEG",
1030    (ineg (i8 (atomic_load_8  addr:$dst))),
1031    (ineg (i16 (atomic_load_16 addr:$dst))),
1032    (ineg (i32 (atomic_load_32 addr:$dst))),
1033    (ineg (i64 (atomic_load_64 addr:$dst)))>;
1034defm : RELEASE_UNOP<"NOT",
1035    (not (i8 (atomic_load_8  addr:$dst))),
1036    (not (i16 (atomic_load_16 addr:$dst))),
1037    (not (i32 (atomic_load_32 addr:$dst))),
1038    (not (i64 (atomic_load_64 addr:$dst)))>;
1039
1040def : Pat<(atomic_store_8 addr:$dst, (i8 imm:$src)),
1041          (MOV8mi addr:$dst, imm:$src)>;
1042def : Pat<(atomic_store_16 addr:$dst, (i16 imm:$src)),
1043          (MOV16mi addr:$dst, imm:$src)>;
1044def : Pat<(atomic_store_32 addr:$dst, (i32 imm:$src)),
1045          (MOV32mi addr:$dst, imm:$src)>;
1046def : Pat<(atomic_store_64 addr:$dst, (i64immSExt32:$src)),
1047          (MOV64mi32 addr:$dst, i64immSExt32:$src)>;
1048
1049def : Pat<(atomic_store_8 addr:$dst, GR8:$src),
1050          (MOV8mr addr:$dst, GR8:$src)>;
1051def : Pat<(atomic_store_16 addr:$dst, GR16:$src),
1052          (MOV16mr addr:$dst, GR16:$src)>;
1053def : Pat<(atomic_store_32 addr:$dst, GR32:$src),
1054          (MOV32mr addr:$dst, GR32:$src)>;
1055def : Pat<(atomic_store_64 addr:$dst, GR64:$src),
1056          (MOV64mr addr:$dst, GR64:$src)>;
1057
1058def : Pat<(i8  (atomic_load_8 addr:$src)),  (MOV8rm addr:$src)>;
1059def : Pat<(i16 (atomic_load_16 addr:$src)), (MOV16rm addr:$src)>;
1060def : Pat<(i32 (atomic_load_32 addr:$src)), (MOV32rm addr:$src)>;
1061def : Pat<(i64 (atomic_load_64 addr:$src)), (MOV64rm addr:$src)>;
1062
1063// Floating point loads/stores.
1064def : Pat<(atomic_store_32 addr:$dst, (i32 (bitconvert (f32 FR32:$src)))),
1065          (MOVSSmr addr:$dst, FR32:$src)>, Requires<[UseSSE1]>;
1066def : Pat<(atomic_store_32 addr:$dst, (i32 (bitconvert (f32 FR32:$src)))),
1067          (VMOVSSmr addr:$dst, FR32:$src)>, Requires<[UseAVX]>;
1068def : Pat<(atomic_store_32 addr:$dst, (i32 (bitconvert (f32 FR32:$src)))),
1069          (VMOVSSZmr addr:$dst, FR32:$src)>, Requires<[HasAVX512]>;
1070
1071def : Pat<(atomic_store_64 addr:$dst, (i64 (bitconvert (f64 FR64:$src)))),
1072          (MOVSDmr addr:$dst, FR64:$src)>, Requires<[UseSSE2]>;
1073def : Pat<(atomic_store_64 addr:$dst, (i64 (bitconvert (f64 FR64:$src)))),
1074          (VMOVSDmr addr:$dst, FR64:$src)>, Requires<[UseAVX]>;
1075def : Pat<(atomic_store_64 addr:$dst, (i64 (bitconvert (f64 FR64:$src)))),
1076          (VMOVSDmr addr:$dst, FR64:$src)>, Requires<[HasAVX512]>;
1077
1078def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1079          (MOVSSrm_alt addr:$src)>, Requires<[UseSSE1]>;
1080def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1081          (VMOVSSrm_alt addr:$src)>, Requires<[UseAVX]>;
1082def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1083          (VMOVSSZrm_alt addr:$src)>, Requires<[HasAVX512]>;
1084
1085def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1086          (MOVSDrm_alt addr:$src)>, Requires<[UseSSE2]>;
1087def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1088          (VMOVSDrm_alt addr:$src)>, Requires<[UseAVX]>;
1089def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1090          (VMOVSDZrm_alt addr:$src)>, Requires<[HasAVX512]>;
1091
1092//===----------------------------------------------------------------------===//
1093// DAG Pattern Matching Rules
1094//===----------------------------------------------------------------------===//
1095
1096// Use AND/OR to store 0/-1 in memory when optimizing for minsize. This saves
1097// binary size compared to a regular MOV, but it introduces an unnecessary
1098// load, so is not suitable for regular or optsize functions.
1099let Predicates = [OptForMinSize] in {
1100def : Pat<(simple_store (i16 0), addr:$dst), (AND16mi8 addr:$dst, 0)>;
1101def : Pat<(simple_store (i32 0), addr:$dst), (AND32mi8 addr:$dst, 0)>;
1102def : Pat<(simple_store (i64 0), addr:$dst), (AND64mi8 addr:$dst, 0)>;
1103def : Pat<(simple_store (i16 -1), addr:$dst), (OR16mi8 addr:$dst, -1)>;
1104def : Pat<(simple_store (i32 -1), addr:$dst), (OR32mi8 addr:$dst, -1)>;
1105def : Pat<(simple_store (i64 -1), addr:$dst), (OR64mi8 addr:$dst, -1)>;
1106}
1107
1108// In kernel code model, we can get the address of a label
1109// into a register with 'movq'.  FIXME: This is a hack, the 'imm' predicate of
1110// the MOV64ri32 should accept these.
1111def : Pat<(i64 (X86Wrapper tconstpool  :$dst)),
1112          (MOV64ri32 tconstpool  :$dst)>, Requires<[KernelCode]>;
1113def : Pat<(i64 (X86Wrapper tjumptable  :$dst)),
1114          (MOV64ri32 tjumptable  :$dst)>, Requires<[KernelCode]>;
1115def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
1116          (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
1117def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
1118          (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
1119def : Pat<(i64 (X86Wrapper mcsym:$dst)),
1120          (MOV64ri32 mcsym:$dst)>, Requires<[KernelCode]>;
1121def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
1122          (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
1123
1124// If we have small model and -static mode, it is safe to store global addresses
1125// directly as immediates.  FIXME: This is really a hack, the 'imm' predicate
1126// for MOV64mi32 should handle this sort of thing.
1127def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
1128          (MOV64mi32 addr:$dst, tconstpool:$src)>,
1129          Requires<[NearData, IsNotPIC]>;
1130def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
1131          (MOV64mi32 addr:$dst, tjumptable:$src)>,
1132          Requires<[NearData, IsNotPIC]>;
1133def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
1134          (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
1135          Requires<[NearData, IsNotPIC]>;
1136def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
1137          (MOV64mi32 addr:$dst, texternalsym:$src)>,
1138          Requires<[NearData, IsNotPIC]>;
1139def : Pat<(store (i64 (X86Wrapper mcsym:$src)), addr:$dst),
1140          (MOV64mi32 addr:$dst, mcsym:$src)>,
1141          Requires<[NearData, IsNotPIC]>;
1142def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
1143          (MOV64mi32 addr:$dst, tblockaddress:$src)>,
1144          Requires<[NearData, IsNotPIC]>;
1145
1146def : Pat<(i32 (X86RecoverFrameAlloc mcsym:$dst)), (MOV32ri mcsym:$dst)>;
1147def : Pat<(i64 (X86RecoverFrameAlloc mcsym:$dst)), (MOV64ri mcsym:$dst)>;
1148
1149// Calls
1150
1151// tls has some funny stuff here...
1152// This corresponds to movabs $foo@tpoff, %rax
1153def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
1154          (MOV64ri32 tglobaltlsaddr :$dst)>;
1155// This corresponds to add $foo@tpoff, %rax
1156def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
1157          (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
1158
1159
1160// Direct PC relative function call for small code model. 32-bit displacement
1161// sign extended to 64-bit.
1162def : Pat<(X86call (i64 tglobaladdr:$dst)),
1163          (CALL64pcrel32 tglobaladdr:$dst)>;
1164def : Pat<(X86call (i64 texternalsym:$dst)),
1165          (CALL64pcrel32 texternalsym:$dst)>;
1166
1167// Tailcall stuff. The TCRETURN instructions execute after the epilog, so they
1168// can never use callee-saved registers. That is the purpose of the GR64_TC
1169// register classes.
1170//
1171// The only volatile register that is never used by the calling convention is
1172// %r11. This happens when calling a vararg function with 6 arguments.
1173//
1174// Match an X86tcret that uses less than 7 volatile registers.
1175def X86tcret_6regs : PatFrag<(ops node:$ptr, node:$off),
1176                             (X86tcret node:$ptr, node:$off), [{
1177  // X86tcret args: (*chain, ptr, imm, regs..., glue)
1178  unsigned NumRegs = 0;
1179  for (unsigned i = 3, e = N->getNumOperands(); i != e; ++i)
1180    if (isa<RegisterSDNode>(N->getOperand(i)) && ++NumRegs > 6)
1181      return false;
1182  return true;
1183}]>;
1184
1185def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
1186          (TCRETURNri ptr_rc_tailcall:$dst, imm:$off)>,
1187          Requires<[Not64BitMode, NotUseIndirectThunkCalls]>;
1188
1189// FIXME: This is disabled for 32-bit PIC mode because the global base
1190// register which is part of the address mode may be assigned a
1191// callee-saved register.
1192def : Pat<(X86tcret (load addr:$dst), imm:$off),
1193          (TCRETURNmi addr:$dst, imm:$off)>,
1194          Requires<[Not64BitMode, IsNotPIC, NotUseIndirectThunkCalls]>;
1195
1196def : Pat<(X86tcret (i32 tglobaladdr:$dst), imm:$off),
1197          (TCRETURNdi tglobaladdr:$dst, imm:$off)>,
1198          Requires<[NotLP64]>;
1199
1200def : Pat<(X86tcret (i32 texternalsym:$dst), imm:$off),
1201          (TCRETURNdi texternalsym:$dst, imm:$off)>,
1202          Requires<[NotLP64]>;
1203
1204def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
1205          (TCRETURNri64 ptr_rc_tailcall:$dst, imm:$off)>,
1206          Requires<[In64BitMode, NotUseIndirectThunkCalls]>;
1207
1208// Don't fold loads into X86tcret requiring more than 6 regs.
1209// There wouldn't be enough scratch registers for base+index.
1210def : Pat<(X86tcret_6regs (load addr:$dst), imm:$off),
1211          (TCRETURNmi64 addr:$dst, imm:$off)>,
1212          Requires<[In64BitMode, NotUseIndirectThunkCalls]>;
1213
1214def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
1215          (INDIRECT_THUNK_TCRETURN64 ptr_rc_tailcall:$dst, imm:$off)>,
1216          Requires<[In64BitMode, UseIndirectThunkCalls]>;
1217
1218def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
1219          (INDIRECT_THUNK_TCRETURN32 ptr_rc_tailcall:$dst, imm:$off)>,
1220          Requires<[Not64BitMode, UseIndirectThunkCalls]>;
1221
1222def : Pat<(X86tcret (i64 tglobaladdr:$dst), imm:$off),
1223          (TCRETURNdi64 tglobaladdr:$dst, imm:$off)>,
1224          Requires<[IsLP64]>;
1225
1226def : Pat<(X86tcret (i64 texternalsym:$dst), imm:$off),
1227          (TCRETURNdi64 texternalsym:$dst, imm:$off)>,
1228          Requires<[IsLP64]>;
1229
1230// Normal calls, with various flavors of addresses.
1231def : Pat<(X86call (i32 tglobaladdr:$dst)),
1232          (CALLpcrel32 tglobaladdr:$dst)>;
1233def : Pat<(X86call (i32 texternalsym:$dst)),
1234          (CALLpcrel32 texternalsym:$dst)>;
1235def : Pat<(X86call (i32 imm:$dst)),
1236          (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
1237
1238// Comparisons.
1239
1240// TEST R,R is smaller than CMP R,0
1241def : Pat<(X86cmp GR8:$src1, 0),
1242          (TEST8rr GR8:$src1, GR8:$src1)>;
1243def : Pat<(X86cmp GR16:$src1, 0),
1244          (TEST16rr GR16:$src1, GR16:$src1)>;
1245def : Pat<(X86cmp GR32:$src1, 0),
1246          (TEST32rr GR32:$src1, GR32:$src1)>;
1247def : Pat<(X86cmp GR64:$src1, 0),
1248          (TEST64rr GR64:$src1, GR64:$src1)>;
1249
1250// zextload bool -> zextload byte
1251// i1 stored in one byte in zero-extended form.
1252// Upper bits cleanup should be executed before Store.
1253def : Pat<(zextloadi8i1  addr:$src), (MOV8rm addr:$src)>;
1254def : Pat<(zextloadi16i1 addr:$src),
1255          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1256def : Pat<(zextloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1257def : Pat<(zextloadi64i1 addr:$src),
1258          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1259
1260// extload bool -> extload byte
1261// When extloading from 16-bit and smaller memory locations into 64-bit
1262// registers, use zero-extending loads so that the entire 64-bit register is
1263// defined, avoiding partial-register updates.
1264
1265def : Pat<(extloadi8i1 addr:$src),   (MOV8rm      addr:$src)>;
1266def : Pat<(extloadi16i1 addr:$src),
1267          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1268def : Pat<(extloadi32i1 addr:$src),  (MOVZX32rm8  addr:$src)>;
1269def : Pat<(extloadi16i8 addr:$src),
1270          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1271def : Pat<(extloadi32i8 addr:$src),  (MOVZX32rm8  addr:$src)>;
1272def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
1273
1274// For other extloads, use subregs, since the high contents of the register are
1275// defined after an extload.
1276// NOTE: The extloadi64i32 pattern needs to be first as it will try to form
1277// 32-bit loads for 4 byte aligned i8/i16 loads.
1278def : Pat<(extloadi64i32 addr:$src),
1279          (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src), sub_32bit)>;
1280def : Pat<(extloadi64i1 addr:$src),
1281          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1282def : Pat<(extloadi64i8 addr:$src),
1283          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1284def : Pat<(extloadi64i16 addr:$src),
1285          (SUBREG_TO_REG (i64 0), (MOVZX32rm16 addr:$src), sub_32bit)>;
1286
1287// anyext. Define these to do an explicit zero-extend to
1288// avoid partial-register updates.
1289def : Pat<(i16 (anyext GR8 :$src)), (EXTRACT_SUBREG
1290                                     (MOVZX32rr8 GR8 :$src), sub_16bit)>;
1291def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8  GR8 :$src)>;
1292
1293// Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
1294def : Pat<(i32 (anyext GR16:$src)),
1295          (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
1296
1297def : Pat<(i64 (anyext GR8 :$src)),
1298          (SUBREG_TO_REG (i64 0), (MOVZX32rr8  GR8  :$src), sub_32bit)>;
1299def : Pat<(i64 (anyext GR16:$src)),
1300          (SUBREG_TO_REG (i64 0), (MOVZX32rr16 GR16 :$src), sub_32bit)>;
1301def : Pat<(i64 (anyext GR32:$src)),
1302          (INSERT_SUBREG (i64 (IMPLICIT_DEF)), GR32:$src, sub_32bit)>;
1303
1304// If this is an anyext of the remainder of an 8-bit sdivrem, use a MOVSX
1305// instead of a MOVZX. The sdivrem lowering will emit emit a MOVSX to move
1306// %ah to the lower byte of a register. By using a MOVSX here we allow a
1307// post-isel peephole to merge the two MOVSX instructions into one.
1308def anyext_sdiv : PatFrag<(ops node:$lhs), (anyext node:$lhs),[{
1309  return (N->getOperand(0).getOpcode() == ISD::SDIVREM &&
1310          N->getOperand(0).getResNo() == 1);
1311}]>;
1312def : Pat<(i32 (anyext_sdiv GR8:$src)), (MOVSX32rr8 GR8:$src)>;
1313
1314// Any instruction that defines a 32-bit result leaves the high half of the
1315// register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
1316// be copying from a truncate. Any other 32-bit operation will zero-extend
1317// up to 64 bits. AssertSext/AssertZext aren't saying anything about the upper
1318// 32 bits, they're probably just qualifying a CopyFromReg.
1319def def32 : PatLeaf<(i32 GR32:$src), [{
1320  return N->getOpcode() != ISD::TRUNCATE &&
1321         N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
1322         N->getOpcode() != ISD::CopyFromReg &&
1323         N->getOpcode() != ISD::AssertSext &&
1324         N->getOpcode() != ISD::AssertZext;
1325}]>;
1326
1327// In the case of a 32-bit def that is known to implicitly zero-extend,
1328// we can use a SUBREG_TO_REG.
1329def : Pat<(i64 (zext def32:$src)),
1330          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1331def : Pat<(i64 (and (anyext def32:$src), 0x00000000FFFFFFFF)),
1332          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1333
1334//===----------------------------------------------------------------------===//
1335// Pattern match OR as ADD
1336//===----------------------------------------------------------------------===//
1337
1338// If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
1339// 3-addressified into an LEA instruction to avoid copies.  However, we also
1340// want to finally emit these instructions as an or at the end of the code
1341// generator to make the generated code easier to read.  To do this, we select
1342// into "disjoint bits" pseudo ops.
1343
1344// Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.
1345def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
1346  if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1347    return CurDAG->MaskedValueIsZero(N->getOperand(0), CN->getAPIntValue());
1348
1349  KnownBits Known0 = CurDAG->computeKnownBits(N->getOperand(0), 0);
1350  KnownBits Known1 = CurDAG->computeKnownBits(N->getOperand(1), 0);
1351  return (~Known0.Zero & ~Known1.Zero) == 0;
1352}]>;
1353
1354
1355// (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1356// Try this before the selecting to OR.
1357let SchedRW = [WriteALU] in {
1358
1359let isConvertibleToThreeAddress = 1, isPseudo = 1,
1360    Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1361let isCommutable = 1 in {
1362def ADD8rr_DB   : I<0, Pseudo, (outs GR8:$dst), (ins GR8:$src1, GR8:$src2),
1363                    "", // orb/addb REG, REG
1364                    [(set GR8:$dst, (or_is_add GR8:$src1, GR8:$src2))]>;
1365def ADD16rr_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1366                    "", // orw/addw REG, REG
1367                    [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1368def ADD32rr_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1369                    "", // orl/addl REG, REG
1370                    [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1371def ADD64rr_DB  : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1372                    "", // orq/addq REG, REG
1373                    [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1374} // isCommutable
1375
1376// NOTE: These are order specific, we want the ri8 forms to be listed
1377// first so that they are slightly preferred to the ri forms.
1378
1379def ADD8ri_DB :   I<0, Pseudo,
1380                    (outs GR8:$dst), (ins GR8:$src1, i8imm:$src2),
1381                    "", // orb/addb REG, imm8
1382                    [(set GR8:$dst, (or_is_add GR8:$src1, imm:$src2))]>;
1383def ADD16ri8_DB : I<0, Pseudo,
1384                    (outs GR16:$dst), (ins GR16:$src1, i16i8imm:$src2),
1385                    "", // orw/addw REG, imm8
1386                    [(set GR16:$dst,(or_is_add GR16:$src1,i16immSExt8:$src2))]>;
1387def ADD16ri_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1388                    "", // orw/addw REG, imm
1389                    [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1390
1391def ADD32ri8_DB : I<0, Pseudo,
1392                    (outs GR32:$dst), (ins GR32:$src1, i32i8imm:$src2),
1393                    "", // orl/addl REG, imm8
1394                    [(set GR32:$dst,(or_is_add GR32:$src1,i32immSExt8:$src2))]>;
1395def ADD32ri_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1396                    "", // orl/addl REG, imm
1397                    [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1398
1399
1400def ADD64ri8_DB : I<0, Pseudo,
1401                    (outs GR64:$dst), (ins GR64:$src1, i64i8imm:$src2),
1402                    "", // orq/addq REG, imm8
1403                    [(set GR64:$dst, (or_is_add GR64:$src1,
1404                                                i64immSExt8:$src2))]>;
1405def ADD64ri32_DB : I<0, Pseudo,
1406                     (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1407                     "", // orq/addq REG, imm
1408                     [(set GR64:$dst, (or_is_add GR64:$src1,
1409                                                 i64immSExt32:$src2))]>;
1410}
1411} // AddedComplexity, SchedRW
1412
1413//===----------------------------------------------------------------------===//
1414// Pattern match SUB as XOR
1415//===----------------------------------------------------------------------===//
1416
1417// An immediate in the LHS of a subtract can't be encoded in the instruction.
1418// If there is no possibility of a borrow we can use an XOR instead of a SUB
1419// to enable the immediate to be folded.
1420// TODO: Move this to a DAG combine?
1421
1422def sub_is_xor : PatFrag<(ops node:$lhs, node:$rhs), (sub node:$lhs, node:$rhs),[{
1423  if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1424    KnownBits Known = CurDAG->computeKnownBits(N->getOperand(1));
1425
1426    // If all possible ones in the RHS are set in the LHS then there can't be
1427    // a borrow and we can use xor.
1428    return (~Known.Zero).isSubsetOf(CN->getAPIntValue());
1429  }
1430
1431  return false;
1432}]>;
1433
1434let AddedComplexity = 5 in {
1435def : Pat<(sub_is_xor imm:$src2, GR8:$src1),
1436          (XOR8ri GR8:$src1, imm:$src2)>;
1437def : Pat<(sub_is_xor i16immSExt8:$src2, GR16:$src1),
1438          (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1439def : Pat<(sub_is_xor imm:$src2, GR16:$src1),
1440          (XOR16ri GR16:$src1, imm:$src2)>;
1441def : Pat<(sub_is_xor i32immSExt8:$src2, GR32:$src1),
1442          (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1443def : Pat<(sub_is_xor imm:$src2, GR32:$src1),
1444          (XOR32ri GR32:$src1, imm:$src2)>;
1445def : Pat<(sub_is_xor i64immSExt8:$src2, GR64:$src1),
1446          (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1447def : Pat<(sub_is_xor i64immSExt32:$src2, GR64:$src1),
1448          (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1449}
1450
1451//===----------------------------------------------------------------------===//
1452// Some peepholes
1453//===----------------------------------------------------------------------===//
1454
1455// Odd encoding trick: -128 fits into an 8-bit immediate field while
1456// +128 doesn't, so in this special case use a sub instead of an add.
1457def : Pat<(add GR16:$src1, 128),
1458          (SUB16ri8 GR16:$src1, -128)>;
1459def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1460          (SUB16mi8 addr:$dst, -128)>;
1461
1462def : Pat<(add GR32:$src1, 128),
1463          (SUB32ri8 GR32:$src1, -128)>;
1464def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1465          (SUB32mi8 addr:$dst, -128)>;
1466
1467def : Pat<(add GR64:$src1, 128),
1468          (SUB64ri8 GR64:$src1, -128)>;
1469def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1470          (SUB64mi8 addr:$dst, -128)>;
1471
1472def : Pat<(X86add_flag_nocf GR16:$src1, 128),
1473          (SUB16ri8 GR16:$src1, -128)>;
1474def : Pat<(X86add_flag_nocf GR32:$src1, 128),
1475          (SUB32ri8 GR32:$src1, -128)>;
1476def : Pat<(X86add_flag_nocf GR64:$src1, 128),
1477          (SUB64ri8 GR64:$src1, -128)>;
1478
1479// The same trick applies for 32-bit immediate fields in 64-bit
1480// instructions.
1481def : Pat<(add GR64:$src1, 0x0000000080000000),
1482          (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1483def : Pat<(store (add (loadi64 addr:$dst), 0x0000000080000000), addr:$dst),
1484          (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1485
1486def : Pat<(X86add_flag_nocf GR64:$src1, 0x0000000080000000),
1487          (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1488
1489// To avoid needing to materialize an immediate in a register, use a 32-bit and
1490// with implicit zero-extension instead of a 64-bit and if the immediate has at
1491// least 32 bits of leading zeros. If in addition the last 32 bits can be
1492// represented with a sign extension of a 8 bit constant, use that.
1493// This can also reduce instruction size by eliminating the need for the REX
1494// prefix.
1495
1496// AddedComplexity is needed to give priority over i64immSExt8 and i64immSExt32.
1497let AddedComplexity = 1 in {
1498def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
1499          (SUBREG_TO_REG
1500            (i64 0),
1501            (AND32ri8
1502              (EXTRACT_SUBREG GR64:$src, sub_32bit),
1503              (i32 (GetLo32XForm imm:$imm))),
1504            sub_32bit)>;
1505
1506def : Pat<(and GR64:$src, i64immZExt32:$imm),
1507          (SUBREG_TO_REG
1508            (i64 0),
1509            (AND32ri
1510              (EXTRACT_SUBREG GR64:$src, sub_32bit),
1511              (i32 (GetLo32XForm imm:$imm))),
1512            sub_32bit)>;
1513} // AddedComplexity = 1
1514
1515
1516// AddedComplexity is needed due to the increased complexity on the
1517// i64immZExt32SExt8 and i64immZExt32 patterns above. Applying this to all
1518// the MOVZX patterns keeps thems together in DAGIsel tables.
1519let AddedComplexity = 1 in {
1520// r & (2^16-1) ==> movz
1521def : Pat<(and GR32:$src1, 0xffff),
1522          (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1523// r & (2^8-1) ==> movz
1524def : Pat<(and GR32:$src1, 0xff),
1525          (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>;
1526// r & (2^8-1) ==> movz
1527def : Pat<(and GR16:$src1, 0xff),
1528           (EXTRACT_SUBREG (MOVZX32rr8 (EXTRACT_SUBREG GR16:$src1, sub_8bit)),
1529             sub_16bit)>;
1530
1531// r & (2^32-1) ==> movz
1532def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1533          (SUBREG_TO_REG (i64 0),
1534                         (MOV32rr (EXTRACT_SUBREG GR64:$src, sub_32bit)),
1535                         sub_32bit)>;
1536// r & (2^16-1) ==> movz
1537def : Pat<(and GR64:$src, 0xffff),
1538          (SUBREG_TO_REG (i64 0),
1539                      (MOVZX32rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit))),
1540                      sub_32bit)>;
1541// r & (2^8-1) ==> movz
1542def : Pat<(and GR64:$src, 0xff),
1543          (SUBREG_TO_REG (i64 0),
1544                         (MOVZX32rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit))),
1545                         sub_32bit)>;
1546} // AddedComplexity = 1
1547
1548
1549// Try to use BTS/BTR/BTC for single bit operations on the upper 32-bits.
1550
1551def BTRXForm : SDNodeXForm<imm, [{
1552  // Transformation function: Find the lowest 0.
1553  return getI64Imm((uint8_t)N->getAPIntValue().countTrailingOnes(), SDLoc(N));
1554}]>;
1555
1556def BTCBTSXForm : SDNodeXForm<imm, [{
1557  // Transformation function: Find the lowest 1.
1558  return getI64Imm((uint8_t)N->getAPIntValue().countTrailingZeros(), SDLoc(N));
1559}]>;
1560
1561def BTRMask64 : ImmLeaf<i64, [{
1562  return !isUInt<32>(Imm) && !isInt<32>(Imm) && isPowerOf2_64(~Imm);
1563}]>;
1564
1565def BTCBTSMask64 : ImmLeaf<i64, [{
1566  return !isInt<32>(Imm) && isPowerOf2_64(Imm);
1567}]>;
1568
1569// For now only do this for optsize.
1570let AddedComplexity = 1, Predicates=[OptForSize] in {
1571  def : Pat<(and GR64:$src1, BTRMask64:$mask),
1572            (BTR64ri8 GR64:$src1, (BTRXForm imm:$mask))>;
1573  def : Pat<(or GR64:$src1, BTCBTSMask64:$mask),
1574            (BTS64ri8 GR64:$src1, (BTCBTSXForm imm:$mask))>;
1575  def : Pat<(xor GR64:$src1, BTCBTSMask64:$mask),
1576            (BTC64ri8 GR64:$src1, (BTCBTSXForm imm:$mask))>;
1577}
1578
1579
1580// sext_inreg patterns
1581def : Pat<(sext_inreg GR32:$src, i16),
1582          (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1583def : Pat<(sext_inreg GR32:$src, i8),
1584          (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>;
1585
1586def : Pat<(sext_inreg GR16:$src, i8),
1587           (EXTRACT_SUBREG (MOVSX32rr8 (EXTRACT_SUBREG GR16:$src, sub_8bit)),
1588             sub_16bit)>;
1589
1590def : Pat<(sext_inreg GR64:$src, i32),
1591          (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1592def : Pat<(sext_inreg GR64:$src, i16),
1593          (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1594def : Pat<(sext_inreg GR64:$src, i8),
1595          (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1596
1597// sext, sext_load, zext, zext_load
1598def: Pat<(i16 (sext GR8:$src)),
1599          (EXTRACT_SUBREG (MOVSX32rr8 GR8:$src), sub_16bit)>;
1600def: Pat<(sextloadi16i8 addr:$src),
1601          (EXTRACT_SUBREG (MOVSX32rm8 addr:$src), sub_16bit)>;
1602def: Pat<(i16 (zext GR8:$src)),
1603          (EXTRACT_SUBREG (MOVZX32rr8 GR8:$src), sub_16bit)>;
1604def: Pat<(zextloadi16i8 addr:$src),
1605          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1606
1607// trunc patterns
1608def : Pat<(i16 (trunc GR32:$src)),
1609          (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1610def : Pat<(i8 (trunc GR32:$src)),
1611          (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1612                          sub_8bit)>,
1613      Requires<[Not64BitMode]>;
1614def : Pat<(i8 (trunc GR16:$src)),
1615          (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1616                          sub_8bit)>,
1617      Requires<[Not64BitMode]>;
1618def : Pat<(i32 (trunc GR64:$src)),
1619          (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1620def : Pat<(i16 (trunc GR64:$src)),
1621          (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1622def : Pat<(i8 (trunc GR64:$src)),
1623          (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1624def : Pat<(i8 (trunc GR32:$src)),
1625          (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1626      Requires<[In64BitMode]>;
1627def : Pat<(i8 (trunc GR16:$src)),
1628          (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1629      Requires<[In64BitMode]>;
1630
1631def immff00_ffff  : ImmLeaf<i32, [{
1632  return Imm >= 0xff00 && Imm <= 0xffff;
1633}]>;
1634
1635// h-register tricks
1636def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1637          (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)>,
1638      Requires<[Not64BitMode]>;
1639def : Pat<(i8 (trunc (srl_su (i32 (anyext GR16:$src)), (i8 8)))),
1640          (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)>,
1641      Requires<[Not64BitMode]>;
1642def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1643          (EXTRACT_SUBREG GR32:$src, sub_8bit_hi)>,
1644      Requires<[Not64BitMode]>;
1645def : Pat<(srl GR16:$src, (i8 8)),
1646          (EXTRACT_SUBREG
1647            (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1648            sub_16bit)>;
1649def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1650          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>;
1651def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1652          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>;
1653def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1654          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1655def : Pat<(srl (and_su GR32:$src, immff00_ffff), (i8 8)),
1656          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1657
1658// h-register tricks.
1659// For now, be conservative on x86-64 and use an h-register extract only if the
1660// value is immediately zero-extended or stored, which are somewhat common
1661// cases. This uses a bunch of code to prevent a register requiring a REX prefix
1662// from being allocated in the same instruction as the h register, as there's
1663// currently no way to describe this requirement to the register allocator.
1664
1665// h-register extract and zero-extend.
1666def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1667          (SUBREG_TO_REG
1668            (i64 0),
1669            (MOVZX32rr8_NOREX
1670              (EXTRACT_SUBREG GR64:$src, sub_8bit_hi)),
1671            sub_32bit)>;
1672def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1673          (SUBREG_TO_REG
1674            (i64 0),
1675            (MOVZX32rr8_NOREX
1676              (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1677            sub_32bit)>;
1678def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1679          (SUBREG_TO_REG
1680            (i64 0),
1681            (MOVZX32rr8_NOREX
1682              (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1683            sub_32bit)>;
1684
1685// h-register extract and store.
1686def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1687          (MOV8mr_NOREX
1688            addr:$dst,
1689            (EXTRACT_SUBREG GR64:$src, sub_8bit_hi))>;
1690def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1691          (MOV8mr_NOREX
1692            addr:$dst,
1693            (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>,
1694      Requires<[In64BitMode]>;
1695def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1696          (MOV8mr_NOREX
1697            addr:$dst,
1698            (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>,
1699      Requires<[In64BitMode]>;
1700
1701
1702// (shl x, 1) ==> (add x, x)
1703// Note that if x is undef (immediate or otherwise), we could theoretically
1704// end up with the two uses of x getting different values, producing a result
1705// where the least significant bit is not 0. However, the probability of this
1706// happening is considered low enough that this is officially not a
1707// "real problem".
1708def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr  GR8 :$src1, GR8 :$src1)>;
1709def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1710def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1711def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1712
1713def shiftMask8 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1714  return isUnneededShiftMask(N, 3);
1715}]>;
1716
1717def shiftMask16 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1718  return isUnneededShiftMask(N, 4);
1719}]>;
1720
1721def shiftMask32 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1722  return isUnneededShiftMask(N, 5);
1723}]>;
1724
1725def shiftMask64 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1726  return isUnneededShiftMask(N, 6);
1727}]>;
1728
1729
1730// Shift amount is implicitly masked.
1731multiclass MaskedShiftAmountPats<SDNode frag, string name> {
1732  // (shift x (and y, 31)) ==> (shift x, y)
1733  def : Pat<(frag GR8:$src1, (shiftMask32 CL)),
1734            (!cast<Instruction>(name # "8rCL") GR8:$src1)>;
1735  def : Pat<(frag GR16:$src1, (shiftMask32 CL)),
1736            (!cast<Instruction>(name # "16rCL") GR16:$src1)>;
1737  def : Pat<(frag GR32:$src1, (shiftMask32 CL)),
1738            (!cast<Instruction>(name # "32rCL") GR32:$src1)>;
1739  def : Pat<(store (frag (loadi8 addr:$dst), (shiftMask32 CL)), addr:$dst),
1740            (!cast<Instruction>(name # "8mCL") addr:$dst)>;
1741  def : Pat<(store (frag (loadi16 addr:$dst), (shiftMask32 CL)), addr:$dst),
1742            (!cast<Instruction>(name # "16mCL") addr:$dst)>;
1743  def : Pat<(store (frag (loadi32 addr:$dst), (shiftMask32 CL)), addr:$dst),
1744            (!cast<Instruction>(name # "32mCL") addr:$dst)>;
1745
1746  // (shift x (and y, 63)) ==> (shift x, y)
1747  def : Pat<(frag GR64:$src1, (shiftMask64 CL)),
1748            (!cast<Instruction>(name # "64rCL") GR64:$src1)>;
1749  def : Pat<(store (frag (loadi64 addr:$dst), (shiftMask64 CL)), addr:$dst),
1750            (!cast<Instruction>(name # "64mCL") addr:$dst)>;
1751}
1752
1753defm : MaskedShiftAmountPats<shl, "SHL">;
1754defm : MaskedShiftAmountPats<srl, "SHR">;
1755defm : MaskedShiftAmountPats<sra, "SAR">;
1756
1757// ROL/ROR instructions allow a stronger mask optimization than shift for 8- and
1758// 16-bit. We can remove a mask of any (bitwidth - 1) on the rotation amount
1759// because over-rotating produces the same result. This is noted in the Intel
1760// docs with: "tempCOUNT <- (COUNT & COUNTMASK) MOD SIZE". Masking the rotation
1761// amount could affect EFLAGS results, but that does not matter because we are
1762// not tracking flags for these nodes.
1763multiclass MaskedRotateAmountPats<SDNode frag, string name> {
1764  // (rot x (and y, BitWidth - 1)) ==> (rot x, y)
1765  def : Pat<(frag GR8:$src1, (shiftMask8 CL)),
1766  (!cast<Instruction>(name # "8rCL") GR8:$src1)>;
1767  def : Pat<(frag GR16:$src1, (shiftMask16 CL)),
1768  (!cast<Instruction>(name # "16rCL") GR16:$src1)>;
1769  def : Pat<(frag GR32:$src1, (shiftMask32 CL)),
1770  (!cast<Instruction>(name # "32rCL") GR32:$src1)>;
1771  def : Pat<(store (frag (loadi8 addr:$dst), (shiftMask8 CL)), addr:$dst),
1772  (!cast<Instruction>(name # "8mCL") addr:$dst)>;
1773  def : Pat<(store (frag (loadi16 addr:$dst), (shiftMask16 CL)), addr:$dst),
1774  (!cast<Instruction>(name # "16mCL") addr:$dst)>;
1775  def : Pat<(store (frag (loadi32 addr:$dst), (shiftMask32 CL)), addr:$dst),
1776  (!cast<Instruction>(name # "32mCL") addr:$dst)>;
1777
1778  // (rot x (and y, 63)) ==> (rot x, y)
1779  def : Pat<(frag GR64:$src1, (shiftMask64 CL)),
1780  (!cast<Instruction>(name # "64rCL") GR64:$src1)>;
1781  def : Pat<(store (frag (loadi64 addr:$dst), (shiftMask64 CL)), addr:$dst),
1782  (!cast<Instruction>(name # "64mCL") addr:$dst)>;
1783}
1784
1785
1786defm : MaskedRotateAmountPats<rotl, "ROL">;
1787defm : MaskedRotateAmountPats<rotr, "ROR">;
1788
1789// Double "funnel" shift amount is implicitly masked.
1790// (fshl/fshr x (and y, 31)) ==> (fshl/fshr x, y) (NOTE: modulo32)
1791def : Pat<(X86fshl GR16:$src1, GR16:$src2, (shiftMask32 CL)),
1792          (SHLD16rrCL GR16:$src1, GR16:$src2)>;
1793def : Pat<(X86fshr GR16:$src2, GR16:$src1, (shiftMask32 CL)),
1794          (SHRD16rrCL GR16:$src1, GR16:$src2)>;
1795
1796// (fshl/fshr x (and y, 31)) ==> (fshl/fshr x, y)
1797def : Pat<(fshl GR32:$src1, GR32:$src2, (shiftMask32 CL)),
1798          (SHLD32rrCL GR32:$src1, GR32:$src2)>;
1799def : Pat<(fshr GR32:$src2, GR32:$src1, (shiftMask32 CL)),
1800          (SHRD32rrCL GR32:$src1, GR32:$src2)>;
1801
1802// (fshl/fshr x (and y, 63)) ==> (fshl/fshr x, y)
1803def : Pat<(fshl GR64:$src1, GR64:$src2, (shiftMask64 CL)),
1804          (SHLD64rrCL GR64:$src1, GR64:$src2)>;
1805def : Pat<(fshr GR64:$src2, GR64:$src1, (shiftMask64 CL)),
1806          (SHRD64rrCL GR64:$src1, GR64:$src2)>;
1807
1808let Predicates = [HasBMI2] in {
1809  let AddedComplexity = 1 in {
1810    def : Pat<(sra GR32:$src1, (shiftMask32 GR8:$src2)),
1811              (SARX32rr GR32:$src1,
1812                        (INSERT_SUBREG
1813                          (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1814    def : Pat<(sra GR64:$src1, (shiftMask64 GR8:$src2)),
1815              (SARX64rr GR64:$src1,
1816                        (INSERT_SUBREG
1817                          (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1818
1819    def : Pat<(srl GR32:$src1, (shiftMask32 GR8:$src2)),
1820              (SHRX32rr GR32:$src1,
1821                        (INSERT_SUBREG
1822                          (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1823    def : Pat<(srl GR64:$src1, (shiftMask64 GR8:$src2)),
1824              (SHRX64rr GR64:$src1,
1825                        (INSERT_SUBREG
1826                          (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1827
1828    def : Pat<(shl GR32:$src1, (shiftMask32 GR8:$src2)),
1829              (SHLX32rr GR32:$src1,
1830                        (INSERT_SUBREG
1831                          (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1832    def : Pat<(shl GR64:$src1, (shiftMask64 GR8:$src2)),
1833              (SHLX64rr GR64:$src1,
1834                        (INSERT_SUBREG
1835                          (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1836  }
1837
1838  def : Pat<(sra (loadi32 addr:$src1), (shiftMask32 GR8:$src2)),
1839            (SARX32rm addr:$src1,
1840                      (INSERT_SUBREG
1841                        (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1842  def : Pat<(sra (loadi64 addr:$src1), (shiftMask64 GR8:$src2)),
1843            (SARX64rm addr:$src1,
1844                      (INSERT_SUBREG
1845                        (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1846
1847  def : Pat<(srl (loadi32 addr:$src1), (shiftMask32 GR8:$src2)),
1848            (SHRX32rm addr:$src1,
1849                      (INSERT_SUBREG
1850                        (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1851  def : Pat<(srl (loadi64 addr:$src1), (shiftMask64 GR8:$src2)),
1852            (SHRX64rm addr:$src1,
1853                      (INSERT_SUBREG
1854                        (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1855
1856  def : Pat<(shl (loadi32 addr:$src1), (shiftMask32 GR8:$src2)),
1857            (SHLX32rm addr:$src1,
1858                      (INSERT_SUBREG
1859                        (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1860  def : Pat<(shl (loadi64 addr:$src1), (shiftMask64 GR8:$src2)),
1861            (SHLX64rm addr:$src1,
1862                      (INSERT_SUBREG
1863                        (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1864}
1865
1866// Use BTR/BTS/BTC for clearing/setting/toggling a bit in a variable location.
1867multiclass one_bit_patterns<RegisterClass RC, ValueType VT, Instruction BTR,
1868                            Instruction BTS, Instruction BTC,
1869                            PatFrag ShiftMask> {
1870  def : Pat<(and RC:$src1, (rotl -2, GR8:$src2)),
1871            (BTR RC:$src1,
1872                 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1873  def : Pat<(or RC:$src1, (shl 1, GR8:$src2)),
1874            (BTS RC:$src1,
1875                 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1876  def : Pat<(xor RC:$src1, (shl 1, GR8:$src2)),
1877            (BTC RC:$src1,
1878                 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1879
1880  // Similar to above, but removing unneeded masking of the shift amount.
1881  def : Pat<(and RC:$src1, (rotl -2, (ShiftMask GR8:$src2))),
1882            (BTR RC:$src1,
1883                 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1884  def : Pat<(or RC:$src1, (shl 1, (ShiftMask GR8:$src2))),
1885            (BTS RC:$src1,
1886                (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1887  def : Pat<(xor RC:$src1, (shl 1, (ShiftMask GR8:$src2))),
1888            (BTC RC:$src1,
1889                (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1890}
1891
1892defm : one_bit_patterns<GR16, i16, BTR16rr, BTS16rr, BTC16rr, shiftMask16>;
1893defm : one_bit_patterns<GR32, i32, BTR32rr, BTS32rr, BTC32rr, shiftMask32>;
1894defm : one_bit_patterns<GR64, i64, BTR64rr, BTS64rr, BTC64rr, shiftMask64>;
1895
1896//===----------------------------------------------------------------------===//
1897// EFLAGS-defining Patterns
1898//===----------------------------------------------------------------------===//
1899
1900// add reg, reg
1901def : Pat<(add GR8 :$src1, GR8 :$src2), (ADD8rr  GR8 :$src1, GR8 :$src2)>;
1902def : Pat<(add GR16:$src1, GR16:$src2), (ADD16rr GR16:$src1, GR16:$src2)>;
1903def : Pat<(add GR32:$src1, GR32:$src2), (ADD32rr GR32:$src1, GR32:$src2)>;
1904def : Pat<(add GR64:$src1, GR64:$src2), (ADD64rr GR64:$src1, GR64:$src2)>;
1905
1906// add reg, mem
1907def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
1908          (ADD8rm GR8:$src1, addr:$src2)>;
1909def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
1910          (ADD16rm GR16:$src1, addr:$src2)>;
1911def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
1912          (ADD32rm GR32:$src1, addr:$src2)>;
1913def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
1914          (ADD64rm GR64:$src1, addr:$src2)>;
1915
1916// add reg, imm
1917def : Pat<(add GR8 :$src1, imm:$src2), (ADD8ri  GR8:$src1 , imm:$src2)>;
1918def : Pat<(add GR16:$src1, imm:$src2), (ADD16ri GR16:$src1, imm:$src2)>;
1919def : Pat<(add GR32:$src1, imm:$src2), (ADD32ri GR32:$src1, imm:$src2)>;
1920def : Pat<(add GR16:$src1, i16immSExt8:$src2),
1921          (ADD16ri8 GR16:$src1, i16immSExt8:$src2)>;
1922def : Pat<(add GR32:$src1, i32immSExt8:$src2),
1923          (ADD32ri8 GR32:$src1, i32immSExt8:$src2)>;
1924def : Pat<(add GR64:$src1, i64immSExt8:$src2),
1925          (ADD64ri8 GR64:$src1, i64immSExt8:$src2)>;
1926def : Pat<(add GR64:$src1, i64immSExt32:$src2),
1927          (ADD64ri32 GR64:$src1, i64immSExt32:$src2)>;
1928
1929// sub reg, reg
1930def : Pat<(sub GR8 :$src1, GR8 :$src2), (SUB8rr  GR8 :$src1, GR8 :$src2)>;
1931def : Pat<(sub GR16:$src1, GR16:$src2), (SUB16rr GR16:$src1, GR16:$src2)>;
1932def : Pat<(sub GR32:$src1, GR32:$src2), (SUB32rr GR32:$src1, GR32:$src2)>;
1933def : Pat<(sub GR64:$src1, GR64:$src2), (SUB64rr GR64:$src1, GR64:$src2)>;
1934
1935// sub reg, mem
1936def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
1937          (SUB8rm GR8:$src1, addr:$src2)>;
1938def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
1939          (SUB16rm GR16:$src1, addr:$src2)>;
1940def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
1941          (SUB32rm GR32:$src1, addr:$src2)>;
1942def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
1943          (SUB64rm GR64:$src1, addr:$src2)>;
1944
1945// sub reg, imm
1946def : Pat<(sub GR8:$src1, imm:$src2),
1947          (SUB8ri GR8:$src1, imm:$src2)>;
1948def : Pat<(sub GR16:$src1, imm:$src2),
1949          (SUB16ri GR16:$src1, imm:$src2)>;
1950def : Pat<(sub GR32:$src1, imm:$src2),
1951          (SUB32ri GR32:$src1, imm:$src2)>;
1952def : Pat<(sub GR16:$src1, i16immSExt8:$src2),
1953          (SUB16ri8 GR16:$src1, i16immSExt8:$src2)>;
1954def : Pat<(sub GR32:$src1, i32immSExt8:$src2),
1955          (SUB32ri8 GR32:$src1, i32immSExt8:$src2)>;
1956def : Pat<(sub GR64:$src1, i64immSExt8:$src2),
1957          (SUB64ri8 GR64:$src1, i64immSExt8:$src2)>;
1958def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
1959          (SUB64ri32 GR64:$src1, i64immSExt32:$src2)>;
1960
1961// sub 0, reg
1962def : Pat<(X86sub_flag 0, GR8 :$src), (NEG8r  GR8 :$src)>;
1963def : Pat<(X86sub_flag 0, GR16:$src), (NEG16r GR16:$src)>;
1964def : Pat<(X86sub_flag 0, GR32:$src), (NEG32r GR32:$src)>;
1965def : Pat<(X86sub_flag 0, GR64:$src), (NEG64r GR64:$src)>;
1966
1967// mul reg, reg
1968def : Pat<(mul GR16:$src1, GR16:$src2),
1969          (IMUL16rr GR16:$src1, GR16:$src2)>;
1970def : Pat<(mul GR32:$src1, GR32:$src2),
1971          (IMUL32rr GR32:$src1, GR32:$src2)>;
1972def : Pat<(mul GR64:$src1, GR64:$src2),
1973          (IMUL64rr GR64:$src1, GR64:$src2)>;
1974
1975// mul reg, mem
1976def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
1977          (IMUL16rm GR16:$src1, addr:$src2)>;
1978def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
1979          (IMUL32rm GR32:$src1, addr:$src2)>;
1980def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
1981          (IMUL64rm GR64:$src1, addr:$src2)>;
1982
1983// mul reg, imm
1984def : Pat<(mul GR16:$src1, imm:$src2),
1985          (IMUL16rri GR16:$src1, imm:$src2)>;
1986def : Pat<(mul GR32:$src1, imm:$src2),
1987          (IMUL32rri GR32:$src1, imm:$src2)>;
1988def : Pat<(mul GR16:$src1, i16immSExt8:$src2),
1989          (IMUL16rri8 GR16:$src1, i16immSExt8:$src2)>;
1990def : Pat<(mul GR32:$src1, i32immSExt8:$src2),
1991          (IMUL32rri8 GR32:$src1, i32immSExt8:$src2)>;
1992def : Pat<(mul GR64:$src1, i64immSExt8:$src2),
1993          (IMUL64rri8 GR64:$src1, i64immSExt8:$src2)>;
1994def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
1995          (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
1996
1997// reg = mul mem, imm
1998def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
1999          (IMUL16rmi addr:$src1, imm:$src2)>;
2000def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
2001          (IMUL32rmi addr:$src1, imm:$src2)>;
2002def : Pat<(mul (loadi16 addr:$src1), i16immSExt8:$src2),
2003          (IMUL16rmi8 addr:$src1, i16immSExt8:$src2)>;
2004def : Pat<(mul (loadi32 addr:$src1), i32immSExt8:$src2),
2005          (IMUL32rmi8 addr:$src1, i32immSExt8:$src2)>;
2006def : Pat<(mul (loadi64 addr:$src1), i64immSExt8:$src2),
2007          (IMUL64rmi8 addr:$src1, i64immSExt8:$src2)>;
2008def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
2009          (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
2010
2011// Increment/Decrement reg.
2012// Do not make INC/DEC if it is slow
2013let Predicates = [UseIncDec] in {
2014  def : Pat<(add GR8:$src, 1),   (INC8r GR8:$src)>;
2015  def : Pat<(add GR16:$src, 1),  (INC16r GR16:$src)>;
2016  def : Pat<(add GR32:$src, 1),  (INC32r GR32:$src)>;
2017  def : Pat<(add GR64:$src, 1),  (INC64r GR64:$src)>;
2018  def : Pat<(add GR8:$src, -1),  (DEC8r GR8:$src)>;
2019  def : Pat<(add GR16:$src, -1), (DEC16r GR16:$src)>;
2020  def : Pat<(add GR32:$src, -1), (DEC32r GR32:$src)>;
2021  def : Pat<(add GR64:$src, -1), (DEC64r GR64:$src)>;
2022
2023  def : Pat<(X86add_flag_nocf GR8:$src, -1),  (DEC8r GR8:$src)>;
2024  def : Pat<(X86add_flag_nocf GR16:$src, -1), (DEC16r GR16:$src)>;
2025  def : Pat<(X86add_flag_nocf GR32:$src, -1), (DEC32r GR32:$src)>;
2026  def : Pat<(X86add_flag_nocf GR64:$src, -1), (DEC64r GR64:$src)>;
2027  def : Pat<(X86sub_flag_nocf GR8:$src, -1),  (INC8r GR8:$src)>;
2028  def : Pat<(X86sub_flag_nocf GR16:$src, -1), (INC16r GR16:$src)>;
2029  def : Pat<(X86sub_flag_nocf GR32:$src, -1), (INC32r GR32:$src)>;
2030  def : Pat<(X86sub_flag_nocf GR64:$src, -1), (INC64r GR64:$src)>;
2031}
2032
2033// or reg/reg.
2034def : Pat<(or GR8 :$src1, GR8 :$src2), (OR8rr  GR8 :$src1, GR8 :$src2)>;
2035def : Pat<(or GR16:$src1, GR16:$src2), (OR16rr GR16:$src1, GR16:$src2)>;
2036def : Pat<(or GR32:$src1, GR32:$src2), (OR32rr GR32:$src1, GR32:$src2)>;
2037def : Pat<(or GR64:$src1, GR64:$src2), (OR64rr GR64:$src1, GR64:$src2)>;
2038
2039// or reg/mem
2040def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
2041          (OR8rm GR8:$src1, addr:$src2)>;
2042def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
2043          (OR16rm GR16:$src1, addr:$src2)>;
2044def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
2045          (OR32rm GR32:$src1, addr:$src2)>;
2046def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
2047          (OR64rm GR64:$src1, addr:$src2)>;
2048
2049// or reg/imm
2050def : Pat<(or GR8:$src1 , imm:$src2), (OR8ri  GR8 :$src1, imm:$src2)>;
2051def : Pat<(or GR16:$src1, imm:$src2), (OR16ri GR16:$src1, imm:$src2)>;
2052def : Pat<(or GR32:$src1, imm:$src2), (OR32ri GR32:$src1, imm:$src2)>;
2053def : Pat<(or GR16:$src1, i16immSExt8:$src2),
2054          (OR16ri8 GR16:$src1, i16immSExt8:$src2)>;
2055def : Pat<(or GR32:$src1, i32immSExt8:$src2),
2056          (OR32ri8 GR32:$src1, i32immSExt8:$src2)>;
2057def : Pat<(or GR64:$src1, i64immSExt8:$src2),
2058          (OR64ri8 GR64:$src1, i64immSExt8:$src2)>;
2059def : Pat<(or GR64:$src1, i64immSExt32:$src2),
2060          (OR64ri32 GR64:$src1, i64immSExt32:$src2)>;
2061
2062// xor reg/reg
2063def : Pat<(xor GR8 :$src1, GR8 :$src2), (XOR8rr  GR8 :$src1, GR8 :$src2)>;
2064def : Pat<(xor GR16:$src1, GR16:$src2), (XOR16rr GR16:$src1, GR16:$src2)>;
2065def : Pat<(xor GR32:$src1, GR32:$src2), (XOR32rr GR32:$src1, GR32:$src2)>;
2066def : Pat<(xor GR64:$src1, GR64:$src2), (XOR64rr GR64:$src1, GR64:$src2)>;
2067
2068// xor reg/mem
2069def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
2070          (XOR8rm GR8:$src1, addr:$src2)>;
2071def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
2072          (XOR16rm GR16:$src1, addr:$src2)>;
2073def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
2074          (XOR32rm GR32:$src1, addr:$src2)>;
2075def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
2076          (XOR64rm GR64:$src1, addr:$src2)>;
2077
2078// xor reg/imm
2079def : Pat<(xor GR8:$src1, imm:$src2),
2080          (XOR8ri GR8:$src1, imm:$src2)>;
2081def : Pat<(xor GR16:$src1, imm:$src2),
2082          (XOR16ri GR16:$src1, imm:$src2)>;
2083def : Pat<(xor GR32:$src1, imm:$src2),
2084          (XOR32ri GR32:$src1, imm:$src2)>;
2085def : Pat<(xor GR16:$src1, i16immSExt8:$src2),
2086          (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
2087def : Pat<(xor GR32:$src1, i32immSExt8:$src2),
2088          (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
2089def : Pat<(xor GR64:$src1, i64immSExt8:$src2),
2090          (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
2091def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
2092          (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
2093
2094// and reg/reg
2095def : Pat<(and GR8 :$src1, GR8 :$src2), (AND8rr  GR8 :$src1, GR8 :$src2)>;
2096def : Pat<(and GR16:$src1, GR16:$src2), (AND16rr GR16:$src1, GR16:$src2)>;
2097def : Pat<(and GR32:$src1, GR32:$src2), (AND32rr GR32:$src1, GR32:$src2)>;
2098def : Pat<(and GR64:$src1, GR64:$src2), (AND64rr GR64:$src1, GR64:$src2)>;
2099
2100// and reg/mem
2101def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
2102          (AND8rm GR8:$src1, addr:$src2)>;
2103def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
2104          (AND16rm GR16:$src1, addr:$src2)>;
2105def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
2106          (AND32rm GR32:$src1, addr:$src2)>;
2107def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
2108          (AND64rm GR64:$src1, addr:$src2)>;
2109
2110// and reg/imm
2111def : Pat<(and GR8:$src1, imm:$src2),
2112          (AND8ri GR8:$src1, imm:$src2)>;
2113def : Pat<(and GR16:$src1, imm:$src2),
2114          (AND16ri GR16:$src1, imm:$src2)>;
2115def : Pat<(and GR32:$src1, imm:$src2),
2116          (AND32ri GR32:$src1, imm:$src2)>;
2117def : Pat<(and GR16:$src1, i16immSExt8:$src2),
2118          (AND16ri8 GR16:$src1, i16immSExt8:$src2)>;
2119def : Pat<(and GR32:$src1, i32immSExt8:$src2),
2120          (AND32ri8 GR32:$src1, i32immSExt8:$src2)>;
2121def : Pat<(and GR64:$src1, i64immSExt8:$src2),
2122          (AND64ri8 GR64:$src1, i64immSExt8:$src2)>;
2123def : Pat<(and GR64:$src1, i64immSExt32:$src2),
2124          (AND64ri32 GR64:$src1, i64immSExt32:$src2)>;
2125
2126// Bit scan instruction patterns to match explicit zero-undef behavior.
2127def : Pat<(cttz_zero_undef GR16:$src), (BSF16rr GR16:$src)>;
2128def : Pat<(cttz_zero_undef GR32:$src), (BSF32rr GR32:$src)>;
2129def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr GR64:$src)>;
2130def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm addr:$src)>;
2131def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm addr:$src)>;
2132def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm addr:$src)>;
2133
2134// When HasMOVBE is enabled it is possible to get a non-legalized
2135// register-register 16 bit bswap. This maps it to a ROL instruction.
2136let Predicates = [HasMOVBE] in {
2137 def : Pat<(bswap GR16:$src), (ROL16ri GR16:$src, (i8 8))>;
2138}
2139