xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86InstrCompiler.td (revision 18054d0220cfc8df9c9568c437bd6fbb59d53c3c)
1//===- X86InstrCompiler.td - Compiler Pseudos and Patterns -*- tablegen -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file describes the various pseudo instructions used by the compiler,
10// as well as Pat patterns used during instruction selection.
11//
12//===----------------------------------------------------------------------===//
13
14//===----------------------------------------------------------------------===//
15// Pattern Matching Support
16
17def GetLo32XForm : SDNodeXForm<imm, [{
18  // Transformation function: get the low 32 bits.
19  return getI32Imm((uint32_t)N->getZExtValue(), SDLoc(N));
20}]>;
21
22
23//===----------------------------------------------------------------------===//
24// Random Pseudo Instructions.
25
26// PIC base construction.  This expands to code that looks like this:
27//     call  $next_inst
28//     popl %destreg"
29let hasSideEffects = 0, isNotDuplicable = 1, Uses = [ESP, SSP],
30    SchedRW = [WriteJump] in
31  def MOVPC32r : Ii32<0xE8, Pseudo, (outs GR32:$reg), (ins i32imm:$label),
32                      "", []>;
33
34// ADJCALLSTACKDOWN/UP implicitly use/def ESP because they may be expanded into
35// a stack adjustment and the codegen must know that they may modify the stack
36// pointer before prolog-epilog rewriting occurs.
37// Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
38// sub / add which can clobber EFLAGS.
39let Defs = [ESP, EFLAGS, SSP], Uses = [ESP, SSP], SchedRW = [WriteALU] in {
40def ADJCALLSTACKDOWN32 : I<0, Pseudo, (outs),
41                           (ins i32imm:$amt1, i32imm:$amt2, i32imm:$amt3),
42                           "#ADJCALLSTACKDOWN", []>, Requires<[NotLP64]>;
43def ADJCALLSTACKUP32   : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
44                           "#ADJCALLSTACKUP",
45                           [(X86callseq_end timm:$amt1, timm:$amt2)]>,
46                           Requires<[NotLP64]>;
47}
48def : Pat<(X86callseq_start timm:$amt1, timm:$amt2),
49       (ADJCALLSTACKDOWN32 i32imm:$amt1, i32imm:$amt2, 0)>, Requires<[NotLP64]>;
50
51
52// ADJCALLSTACKDOWN/UP implicitly use/def RSP because they may be expanded into
53// a stack adjustment and the codegen must know that they may modify the stack
54// pointer before prolog-epilog rewriting occurs.
55// Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
56// sub / add which can clobber EFLAGS.
57let Defs = [RSP, EFLAGS, SSP], Uses = [RSP, SSP], SchedRW = [WriteALU] in {
58def ADJCALLSTACKDOWN64 : I<0, Pseudo, (outs),
59                           (ins i32imm:$amt1, i32imm:$amt2, i32imm:$amt3),
60                           "#ADJCALLSTACKDOWN", []>, Requires<[IsLP64]>;
61def ADJCALLSTACKUP64   : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
62                           "#ADJCALLSTACKUP",
63                           [(X86callseq_end timm:$amt1, timm:$amt2)]>,
64                           Requires<[IsLP64]>;
65}
66def : Pat<(X86callseq_start timm:$amt1, timm:$amt2),
67        (ADJCALLSTACKDOWN64 i32imm:$amt1, i32imm:$amt2, 0)>, Requires<[IsLP64]>;
68
69let SchedRW = [WriteSystem] in {
70
71// x86-64 va_start lowering magic.
72let hasSideEffects = 1, mayStore = 1, Defs = [EFLAGS] in {
73def VASTART_SAVE_XMM_REGS : I<0, Pseudo,
74                              (outs),
75                              (ins GR8:$al, i8mem:$regsavefi, variable_ops),
76                              "#VASTART_SAVE_XMM_REGS $al, $regsavefi",
77                              [(X86vastart_save_xmm_regs GR8:$al, addr:$regsavefi),
78                               (implicit EFLAGS)]>;
79}
80
81let usesCustomInserter = 1, Defs = [EFLAGS] in {
82// The VAARG_64 and VAARG_X32 pseudo-instructions take the address of the
83// va_list, and place the address of the next argument into a register.
84let Defs = [EFLAGS] in {
85def VAARG_64 : I<0, Pseudo,
86                 (outs GR64:$dst),
87                 (ins i8mem:$ap, i32imm:$size, i8imm:$mode, i32imm:$align),
88                 "#VAARG_64 $dst, $ap, $size, $mode, $align",
89                 [(set GR64:$dst,
90                    (X86vaarg64 addr:$ap, timm:$size, timm:$mode, timm:$align)),
91                  (implicit EFLAGS)]>, Requires<[In64BitMode, IsLP64]>;
92def VAARG_X32 : I<0, Pseudo,
93                 (outs GR32:$dst),
94                 (ins i8mem:$ap, i32imm:$size, i8imm:$mode, i32imm:$align),
95                 "#VAARG_X32 $dst, $ap, $size, $mode, $align",
96                 [(set GR32:$dst,
97                    (X86vaargx32 addr:$ap, timm:$size, timm:$mode, timm:$align)),
98                  (implicit EFLAGS)]>, Requires<[In64BitMode, NotLP64]>;
99}
100
101// When using segmented stacks these are lowered into instructions which first
102// check if the current stacklet has enough free memory. If it does, memory is
103// allocated by bumping the stack pointer. Otherwise memory is allocated from
104// the heap.
105
106let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
107def SEG_ALLOCA_32 : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$size),
108                      "# variable sized alloca for segmented stacks",
109                      [(set GR32:$dst,
110                         (X86SegAlloca GR32:$size))]>,
111                    Requires<[NotLP64]>;
112
113let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
114def SEG_ALLOCA_64 : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$size),
115                      "# variable sized alloca for segmented stacks",
116                      [(set GR64:$dst,
117                         (X86SegAlloca GR64:$size))]>,
118                    Requires<[In64BitMode]>;
119
120// To protect against stack clash, dynamic allocation should perform a memory
121// probe at each page.
122
123let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
124def PROBED_ALLOCA_32 : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$size),
125                      "# variable sized alloca with probing",
126                      [(set GR32:$dst,
127                         (X86ProbedAlloca GR32:$size))]>,
128                    Requires<[NotLP64]>;
129
130let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
131def PROBED_ALLOCA_64 : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$size),
132                      "# variable sized alloca with probing",
133                      [(set GR64:$dst,
134                         (X86ProbedAlloca GR64:$size))]>,
135                    Requires<[In64BitMode]>;
136}
137
138let hasNoSchedulingInfo = 1 in
139def STACKALLOC_W_PROBING : I<0, Pseudo, (outs), (ins i64imm:$stacksize),
140                             "# fixed size alloca with probing",
141                             []>;
142
143// Dynamic stack allocation yields a _chkstk or _alloca call for all Windows
144// targets.  These calls are needed to probe the stack when allocating more than
145// 4k bytes in one go. Touching the stack at 4K increments is necessary to
146// ensure that the guard pages used by the OS virtual memory manager are
147// allocated in correct sequence.
148// The main point of having separate instruction are extra unmodelled effects
149// (compared to ordinary calls) like stack pointer change.
150
151let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
152def DYN_ALLOCA_32 : I<0, Pseudo, (outs), (ins GR32:$size),
153                     "# dynamic stack allocation",
154                     [(X86DynAlloca GR32:$size)]>,
155                     Requires<[NotLP64]>;
156
157let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
158def DYN_ALLOCA_64 : I<0, Pseudo, (outs), (ins GR64:$size),
159                     "# dynamic stack allocation",
160                     [(X86DynAlloca GR64:$size)]>,
161                     Requires<[In64BitMode]>;
162} // SchedRW
163
164// These instructions XOR the frame pointer into a GPR. They are used in some
165// stack protection schemes. These are post-RA pseudos because we only know the
166// frame register after register allocation.
167let Constraints = "$src = $dst", isMoveImm = 1, isPseudo = 1, Defs = [EFLAGS] in {
168  def XOR32_FP : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src),
169                  "xorl\t$$FP, $src", []>,
170                  Requires<[NotLP64]>, Sched<[WriteALU]>;
171  def XOR64_FP : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src),
172                  "xorq\t$$FP $src", []>,
173                  Requires<[In64BitMode]>, Sched<[WriteALU]>;
174}
175
176//===----------------------------------------------------------------------===//
177// EH Pseudo Instructions
178//
179let SchedRW = [WriteSystem] in {
180let isTerminator = 1, isReturn = 1, isBarrier = 1,
181    hasCtrlDep = 1, isCodeGenOnly = 1 in {
182def EH_RETURN   : I<0xC3, RawFrm, (outs), (ins GR32:$addr),
183                    "ret\t#eh_return, addr: $addr",
184                    [(X86ehret GR32:$addr)]>, Sched<[WriteJumpLd]>;
185
186}
187
188let isTerminator = 1, isReturn = 1, isBarrier = 1,
189    hasCtrlDep = 1, isCodeGenOnly = 1 in {
190def EH_RETURN64   : I<0xC3, RawFrm, (outs), (ins GR64:$addr),
191                     "ret\t#eh_return, addr: $addr",
192                     [(X86ehret GR64:$addr)]>, Sched<[WriteJumpLd]>;
193
194}
195
196let isTerminator = 1, hasSideEffects = 1, isBarrier = 1, hasCtrlDep = 1,
197    isCodeGenOnly = 1, isReturn = 1, isEHScopeReturn = 1 in {
198  def CLEANUPRET : I<0, Pseudo, (outs), (ins), "# CLEANUPRET", [(cleanupret)]>;
199
200  // CATCHRET needs a custom inserter for SEH.
201  let usesCustomInserter = 1 in
202    def CATCHRET : I<0, Pseudo, (outs), (ins brtarget32:$dst, brtarget32:$from),
203                     "# CATCHRET",
204                     [(catchret bb:$dst, bb:$from)]>;
205}
206
207let hasSideEffects = 1, isBarrier = 1, isCodeGenOnly = 1,
208    usesCustomInserter = 1 in {
209  def EH_SjLj_SetJmp32  : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$buf),
210                            "#EH_SJLJ_SETJMP32",
211                            [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
212                          Requires<[Not64BitMode]>;
213  def EH_SjLj_SetJmp64  : I<0, Pseudo, (outs GR32:$dst), (ins i64mem:$buf),
214                            "#EH_SJLJ_SETJMP64",
215                            [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
216                          Requires<[In64BitMode]>;
217  let isTerminator = 1 in {
218  def EH_SjLj_LongJmp32 : I<0, Pseudo, (outs), (ins i32mem:$buf),
219                            "#EH_SJLJ_LONGJMP32",
220                            [(X86eh_sjlj_longjmp addr:$buf)]>,
221                          Requires<[Not64BitMode]>;
222  def EH_SjLj_LongJmp64 : I<0, Pseudo, (outs), (ins i64mem:$buf),
223                            "#EH_SJLJ_LONGJMP64",
224                            [(X86eh_sjlj_longjmp addr:$buf)]>,
225                          Requires<[In64BitMode]>;
226  }
227}
228
229let isBranch = 1, isTerminator = 1, isCodeGenOnly = 1 in {
230  def EH_SjLj_Setup : I<0, Pseudo, (outs), (ins brtarget:$dst),
231                        "#EH_SjLj_Setup\t$dst", []>;
232}
233} // SchedRW
234
235//===----------------------------------------------------------------------===//
236// Pseudo instructions used by unwind info.
237//
238let isPseudo = 1, SchedRW = [WriteSystem] in {
239  def SEH_PushReg : I<0, Pseudo, (outs), (ins i32imm:$reg),
240                            "#SEH_PushReg $reg", []>;
241  def SEH_SaveReg : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
242                            "#SEH_SaveReg $reg, $dst", []>;
243  def SEH_SaveXMM : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
244                            "#SEH_SaveXMM $reg, $dst", []>;
245  def SEH_StackAlloc : I<0, Pseudo, (outs), (ins i32imm:$size),
246                            "#SEH_StackAlloc $size", []>;
247  def SEH_StackAlign : I<0, Pseudo, (outs), (ins i32imm:$align),
248                            "#SEH_StackAlign $align", []>;
249  def SEH_SetFrame : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$offset),
250                            "#SEH_SetFrame $reg, $offset", []>;
251  def SEH_PushFrame : I<0, Pseudo, (outs), (ins i1imm:$mode),
252                            "#SEH_PushFrame $mode", []>;
253  def SEH_EndPrologue : I<0, Pseudo, (outs), (ins),
254                            "#SEH_EndPrologue", []>;
255  def SEH_Epilogue : I<0, Pseudo, (outs), (ins),
256                            "#SEH_Epilogue", []>;
257}
258
259//===----------------------------------------------------------------------===//
260// Pseudo instructions used by address sanitizer.
261//===----------------------------------------------------------------------===//
262let
263  Defs = [R10, R11, EFLAGS] in {
264def ASAN_CHECK_MEMACCESS : PseudoI<
265  (outs), (ins GR64PLTSafe:$addr, i32imm:$accessinfo),
266  [(int_asan_check_memaccess GR64PLTSafe:$addr, (i32 timm:$accessinfo))]>,
267  Sched<[]>;
268}
269
270//===----------------------------------------------------------------------===//
271// Pseudo instructions used by segmented stacks.
272//
273
274// This is lowered into a RET instruction by MCInstLower.  We need
275// this so that we don't have to have a MachineBasicBlock which ends
276// with a RET and also has successors.
277let isPseudo = 1, SchedRW = [WriteJumpLd] in {
278def MORESTACK_RET: I<0, Pseudo, (outs), (ins), "", []>;
279
280// This instruction is lowered to a RET followed by a MOV.  The two
281// instructions are not generated on a higher level since then the
282// verifier sees a MachineBasicBlock ending with a non-terminator.
283def MORESTACK_RET_RESTORE_R10 : I<0, Pseudo, (outs), (ins), "", []>;
284}
285
286//===----------------------------------------------------------------------===//
287// Alias Instructions
288//===----------------------------------------------------------------------===//
289
290// Alias instruction mapping movr0 to xor.
291// FIXME: remove when we can teach regalloc that xor reg, reg is ok.
292let Defs = [EFLAGS], isReMaterializable = 1, isAsCheapAsAMove = 1,
293    isPseudo = 1, isMoveImm = 1, AddedComplexity = 10 in
294def MOV32r0  : I<0, Pseudo, (outs GR32:$dst), (ins), "",
295                 [(set GR32:$dst, 0)]>, Sched<[WriteZero]>;
296
297// Other widths can also make use of the 32-bit xor, which may have a smaller
298// encoding and avoid partial register updates.
299let AddedComplexity = 10 in {
300def : Pat<(i8 0), (EXTRACT_SUBREG (MOV32r0), sub_8bit)>;
301def : Pat<(i16 0), (EXTRACT_SUBREG (MOV32r0), sub_16bit)>;
302def : Pat<(i64 0), (SUBREG_TO_REG (i64 0), (MOV32r0), sub_32bit)>;
303}
304
305let Predicates = [OptForSize, Not64BitMode],
306    AddedComplexity = 10 in {
307  let SchedRW = [WriteALU] in {
308  // Pseudo instructions for materializing 1 and -1 using XOR+INC/DEC,
309  // which only require 3 bytes compared to MOV32ri which requires 5.
310  let Defs = [EFLAGS], isReMaterializable = 1, isPseudo = 1 in {
311    def MOV32r1 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
312                        [(set GR32:$dst, 1)]>;
313    def MOV32r_1 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
314                        [(set GR32:$dst, -1)]>;
315  }
316  } // SchedRW
317
318  // MOV16ri is 4 bytes, so the instructions above are smaller.
319  def : Pat<(i16 1), (EXTRACT_SUBREG (MOV32r1), sub_16bit)>;
320  def : Pat<(i16 -1), (EXTRACT_SUBREG (MOV32r_1), sub_16bit)>;
321}
322
323let isReMaterializable = 1, isPseudo = 1, AddedComplexity = 5,
324    SchedRW = [WriteALU] in {
325// AddedComplexity higher than MOV64ri but lower than MOV32r0 and MOV32r1.
326def MOV32ImmSExti8 : I<0, Pseudo, (outs GR32:$dst), (ins i32i8imm:$src), "",
327                       [(set GR32:$dst, i32immSExt8:$src)]>,
328                       Requires<[OptForMinSize, NotWin64WithoutFP]>;
329def MOV64ImmSExti8 : I<0, Pseudo, (outs GR64:$dst), (ins i64i8imm:$src), "",
330                       [(set GR64:$dst, i64immSExt8:$src)]>,
331                       Requires<[OptForMinSize, NotWin64WithoutFP]>;
332}
333
334// Materialize i64 constant where top 32-bits are zero. This could theoretically
335// use MOV32ri with a SUBREG_TO_REG to represent the zero-extension, however
336// that would make it more difficult to rematerialize.
337let AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1,
338    isPseudo = 1, SchedRW = [WriteMove] in
339def MOV32ri64 : I<0, Pseudo, (outs GR64:$dst), (ins i64i32imm:$src), "",
340                  [(set GR64:$dst, i64immZExt32:$src)]>;
341
342// This 64-bit pseudo-move can also be used for labels in the x86-64 small code
343// model.
344def mov64imm32 : ComplexPattern<i64, 1, "selectMOV64Imm32", [X86Wrapper]>;
345def : Pat<(i64 mov64imm32:$src), (MOV32ri64 mov64imm32:$src)>;
346
347// Use sbb to materialize carry bit.
348let Uses = [EFLAGS], Defs = [EFLAGS], isPseudo = 1, SchedRW = [WriteADC],
349    hasSideEffects = 0 in {
350// FIXME: These are pseudo ops that should be replaced with Pat<> patterns.
351// However, Pat<> can't replicate the destination reg into the inputs of the
352// result.
353def SETB_C32r : I<0, Pseudo, (outs GR32:$dst), (ins), "", []>;
354def SETB_C64r : I<0, Pseudo, (outs GR64:$dst), (ins), "", []>;
355} // isCodeGenOnly
356
357//===----------------------------------------------------------------------===//
358// String Pseudo Instructions
359//
360let SchedRW = [WriteMicrocoded] in {
361let Defs = [ECX,EDI,ESI], Uses = [ECX,EDI,ESI], isCodeGenOnly = 1 in {
362def REP_MOVSB_32 : I<0xA4, RawFrm, (outs), (ins),
363                    "{rep;movsb (%esi), %es:(%edi)|rep movsb es:[edi], [esi]}",
364                    [(X86rep_movs i8)]>, REP, AdSize32,
365                   Requires<[NotLP64]>;
366def REP_MOVSW_32 : I<0xA5, RawFrm, (outs), (ins),
367                    "{rep;movsw (%esi), %es:(%edi)|rep movsw es:[edi], [esi]}",
368                    [(X86rep_movs i16)]>, REP, AdSize32, OpSize16,
369                   Requires<[NotLP64]>;
370def REP_MOVSD_32 : I<0xA5, RawFrm, (outs), (ins),
371                    "{rep;movsl (%esi), %es:(%edi)|rep movsd es:[edi], [esi]}",
372                    [(X86rep_movs i32)]>, REP, AdSize32, OpSize32,
373                   Requires<[NotLP64]>;
374def REP_MOVSQ_32 : RI<0xA5, RawFrm, (outs), (ins),
375                    "{rep;movsq (%esi), %es:(%edi)|rep movsq es:[edi], [esi]}",
376                    [(X86rep_movs i64)]>, REP, AdSize32,
377                   Requires<[NotLP64, In64BitMode]>;
378}
379
380let Defs = [RCX,RDI,RSI], Uses = [RCX,RDI,RSI], isCodeGenOnly = 1 in {
381def REP_MOVSB_64 : I<0xA4, RawFrm, (outs), (ins),
382                    "{rep;movsb (%rsi), %es:(%rdi)|rep movsb es:[rdi], [rsi]}",
383                    [(X86rep_movs i8)]>, REP, AdSize64,
384                   Requires<[IsLP64]>;
385def REP_MOVSW_64 : I<0xA5, RawFrm, (outs), (ins),
386                    "{rep;movsw (%rsi), %es:(%rdi)|rep movsw es:[rdi], [rsi]}",
387                    [(X86rep_movs i16)]>, REP, AdSize64, OpSize16,
388                   Requires<[IsLP64]>;
389def REP_MOVSD_64 : I<0xA5, RawFrm, (outs), (ins),
390                    "{rep;movsl (%rsi), %es:(%rdi)|rep movsdi es:[rdi], [rsi]}",
391                    [(X86rep_movs i32)]>, REP, AdSize64, OpSize32,
392                   Requires<[IsLP64]>;
393def REP_MOVSQ_64 : RI<0xA5, RawFrm, (outs), (ins),
394                    "{rep;movsq (%rsi), %es:(%rdi)|rep movsq es:[rdi], [rsi]}",
395                    [(X86rep_movs i64)]>, REP, AdSize64,
396                   Requires<[IsLP64]>;
397}
398
399// FIXME: Should use "(X86rep_stos AL)" as the pattern.
400let Defs = [ECX,EDI], isCodeGenOnly = 1 in {
401  let Uses = [AL,ECX,EDI] in
402  def REP_STOSB_32 : I<0xAA, RawFrm, (outs), (ins),
403                       "{rep;stosb %al, %es:(%edi)|rep stosb es:[edi], al}",
404                      [(X86rep_stos i8)]>, REP, AdSize32,
405                     Requires<[NotLP64]>;
406  let Uses = [AX,ECX,EDI] in
407  def REP_STOSW_32 : I<0xAB, RawFrm, (outs), (ins),
408                      "{rep;stosw %ax, %es:(%edi)|rep stosw es:[edi], ax}",
409                      [(X86rep_stos i16)]>, REP, AdSize32, OpSize16,
410                     Requires<[NotLP64]>;
411  let Uses = [EAX,ECX,EDI] in
412  def REP_STOSD_32 : I<0xAB, RawFrm, (outs), (ins),
413                      "{rep;stosl %eax, %es:(%edi)|rep stosd es:[edi], eax}",
414                      [(X86rep_stos i32)]>, REP, AdSize32, OpSize32,
415                     Requires<[NotLP64]>;
416  let Uses = [RAX,RCX,RDI] in
417  def REP_STOSQ_32 : RI<0xAB, RawFrm, (outs), (ins),
418                        "{rep;stosq %rax, %es:(%edi)|rep stosq es:[edi], rax}",
419                        [(X86rep_stos i64)]>, REP, AdSize32,
420                        Requires<[NotLP64, In64BitMode]>;
421}
422
423let Defs = [RCX,RDI], isCodeGenOnly = 1 in {
424  let Uses = [AL,RCX,RDI] in
425  def REP_STOSB_64 : I<0xAA, RawFrm, (outs), (ins),
426                       "{rep;stosb %al, %es:(%rdi)|rep stosb es:[rdi], al}",
427                       [(X86rep_stos i8)]>, REP, AdSize64,
428                       Requires<[IsLP64]>;
429  let Uses = [AX,RCX,RDI] in
430  def REP_STOSW_64 : I<0xAB, RawFrm, (outs), (ins),
431                       "{rep;stosw %ax, %es:(%rdi)|rep stosw es:[rdi], ax}",
432                       [(X86rep_stos i16)]>, REP, AdSize64, OpSize16,
433                       Requires<[IsLP64]>;
434  let Uses = [RAX,RCX,RDI] in
435  def REP_STOSD_64 : I<0xAB, RawFrm, (outs), (ins),
436                      "{rep;stosl %eax, %es:(%rdi)|rep stosd es:[rdi], eax}",
437                       [(X86rep_stos i32)]>, REP, AdSize64, OpSize32,
438                       Requires<[IsLP64]>;
439
440  let Uses = [RAX,RCX,RDI] in
441  def REP_STOSQ_64 : RI<0xAB, RawFrm, (outs), (ins),
442                        "{rep;stosq %rax, %es:(%rdi)|rep stosq es:[rdi], rax}",
443                        [(X86rep_stos i64)]>, REP, AdSize64,
444                        Requires<[IsLP64]>;
445}
446} // SchedRW
447
448//===----------------------------------------------------------------------===//
449// Thread Local Storage Instructions
450//
451let SchedRW = [WriteSystem] in {
452
453// ELF TLS Support
454// All calls clobber the non-callee saved registers. ESP is marked as
455// a use to prevent stack-pointer assignments that appear immediately
456// before calls from potentially appearing dead.
457let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, FP7,
458            ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7,
459            MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
460            XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
461            XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS, DF],
462    usesCustomInserter = 1, Uses = [ESP, SSP] in {
463def TLS_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
464                  "# TLS_addr32",
465                  [(X86tlsaddr tls32addr:$sym)]>,
466                  Requires<[Not64BitMode]>;
467def TLS_base_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
468                  "# TLS_base_addr32",
469                  [(X86tlsbaseaddr tls32baseaddr:$sym)]>,
470                  Requires<[Not64BitMode]>;
471}
472
473// All calls clobber the non-callee saved registers. RSP is marked as
474// a use to prevent stack-pointer assignments that appear immediately
475// before calls from potentially appearing dead.
476let Defs = [RAX, RCX, RDX, RSI, RDI, R8, R9, R10, R11,
477            FP0, FP1, FP2, FP3, FP4, FP5, FP6, FP7,
478            ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7,
479            MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
480            XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
481            XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS, DF],
482    usesCustomInserter = 1, Uses = [RSP, SSP] in {
483def TLS_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
484                   "# TLS_addr64",
485                  [(X86tlsaddr tls64addr:$sym)]>,
486                  Requires<[In64BitMode, IsLP64]>;
487def TLS_base_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
488                   "# TLS_base_addr64",
489                  [(X86tlsbaseaddr tls64baseaddr:$sym)]>,
490                  Requires<[In64BitMode, IsLP64]>;
491def TLS_addrX32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
492                   "# TLS_addrX32",
493                  [(X86tlsaddr tls32addr:$sym)]>,
494                  Requires<[In64BitMode, NotLP64]>;
495def TLS_base_addrX32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
496                   "# TLS_base_addrX32",
497                  [(X86tlsbaseaddr tls32baseaddr:$sym)]>,
498                  Requires<[In64BitMode, NotLP64]>;
499}
500
501// Darwin TLS Support
502// For i386, the address of the thunk is passed on the stack, on return the
503// address of the variable is in %eax.  %ecx is trashed during the function
504// call.  All other registers are preserved.
505let Defs = [EAX, ECX, EFLAGS, DF],
506    Uses = [ESP, SSP],
507    usesCustomInserter = 1 in
508def TLSCall_32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
509                "# TLSCall_32",
510                [(X86TLSCall addr:$sym)]>,
511                Requires<[Not64BitMode]>;
512
513// For x86_64, the address of the thunk is passed in %rdi, but the
514// pseudo directly use the symbol, so do not add an implicit use of
515// %rdi. The lowering will do the right thing with RDI.
516// On return the address of the variable is in %rax.  All other
517// registers are preserved.
518let Defs = [RAX, EFLAGS, DF],
519    Uses = [RSP, SSP],
520    usesCustomInserter = 1 in
521def TLSCall_64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
522                  "# TLSCall_64",
523                  [(X86TLSCall addr:$sym)]>,
524                  Requires<[In64BitMode]>;
525} // SchedRW
526
527//===----------------------------------------------------------------------===//
528// Conditional Move Pseudo Instructions
529
530// CMOV* - Used to implement the SELECT DAG operation.  Expanded after
531// instruction selection into a branch sequence.
532multiclass CMOVrr_PSEUDO<RegisterClass RC, ValueType VT> {
533  def CMOV#NAME  : I<0, Pseudo,
534                    (outs RC:$dst), (ins RC:$t, RC:$f, i8imm:$cond),
535                    "#CMOV_"#NAME#" PSEUDO!",
536                    [(set RC:$dst, (VT (X86cmov RC:$t, RC:$f, timm:$cond,
537                                                EFLAGS)))]>;
538}
539
540let usesCustomInserter = 1, hasNoSchedulingInfo = 1, Uses = [EFLAGS] in {
541  // X86 doesn't have 8-bit conditional moves. Use a customInserter to
542  // emit control flow. An alternative to this is to mark i8 SELECT as Promote,
543  // however that requires promoting the operands, and can induce additional
544  // i8 register pressure.
545  defm _GR8 : CMOVrr_PSEUDO<GR8, i8>;
546
547  let Predicates = [NoCMov] in {
548    defm _GR32 : CMOVrr_PSEUDO<GR32, i32>;
549    defm _GR16 : CMOVrr_PSEUDO<GR16, i16>;
550  } // Predicates = [NoCMov]
551
552  // fcmov doesn't handle all possible EFLAGS, provide a fallback if there is no
553  // SSE1/SSE2.
554  let Predicates = [FPStackf32] in
555    defm _RFP32 : CMOVrr_PSEUDO<RFP32, f32>;
556
557  let Predicates = [FPStackf64] in
558    defm _RFP64 : CMOVrr_PSEUDO<RFP64, f64>;
559
560  defm _RFP80 : CMOVrr_PSEUDO<RFP80, f80>;
561
562  let Predicates = [HasMMX] in
563    defm _VR64   : CMOVrr_PSEUDO<VR64, x86mmx>;
564
565  defm _FR16X    : CMOVrr_PSEUDO<FR16X, f16>;
566  let Predicates = [HasSSE1,NoAVX512] in
567    defm _FR32   : CMOVrr_PSEUDO<FR32, f32>;
568  let Predicates = [HasSSE2,NoAVX512] in
569    defm _FR64   : CMOVrr_PSEUDO<FR64, f64>;
570  let Predicates = [HasAVX512] in {
571    defm _FR32X  : CMOVrr_PSEUDO<FR32X, f32>;
572    defm _FR64X  : CMOVrr_PSEUDO<FR64X, f64>;
573  }
574  let Predicates = [NoVLX] in {
575    defm _VR128  : CMOVrr_PSEUDO<VR128, v2i64>;
576    defm _VR256  : CMOVrr_PSEUDO<VR256, v4i64>;
577  }
578  let Predicates = [HasVLX] in {
579    defm _VR128X : CMOVrr_PSEUDO<VR128X, v2i64>;
580    defm _VR256X : CMOVrr_PSEUDO<VR256X, v4i64>;
581  }
582  defm _VR512  : CMOVrr_PSEUDO<VR512, v8i64>;
583  defm _VK1    : CMOVrr_PSEUDO<VK1,  v1i1>;
584  defm _VK2    : CMOVrr_PSEUDO<VK2,  v2i1>;
585  defm _VK4    : CMOVrr_PSEUDO<VK4,  v4i1>;
586  defm _VK8    : CMOVrr_PSEUDO<VK8,  v8i1>;
587  defm _VK16   : CMOVrr_PSEUDO<VK16, v16i1>;
588  defm _VK32   : CMOVrr_PSEUDO<VK32, v32i1>;
589  defm _VK64   : CMOVrr_PSEUDO<VK64, v64i1>;
590} // usesCustomInserter = 1, hasNoSchedulingInfo = 1, Uses = [EFLAGS]
591
592def : Pat<(f128 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
593          (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
594
595let Predicates = [NoVLX] in {
596  def : Pat<(v16i8 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
597            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
598  def : Pat<(v8i16 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
599            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
600  def : Pat<(v4i32 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
601            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
602  def : Pat<(v4f32 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
603            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
604  def : Pat<(v2f64 (X86cmov VR128:$t, VR128:$f, timm:$cond, EFLAGS)),
605            (CMOV_VR128 VR128:$t, VR128:$f, timm:$cond)>;
606
607  def : Pat<(v32i8 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
608            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
609  def : Pat<(v16i16 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
610            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
611  def : Pat<(v8i32 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
612            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
613  def : Pat<(v8f32 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
614            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
615  def : Pat<(v4f64 (X86cmov VR256:$t, VR256:$f, timm:$cond, EFLAGS)),
616            (CMOV_VR256 VR256:$t, VR256:$f, timm:$cond)>;
617}
618let Predicates = [HasVLX] in {
619  def : Pat<(v16i8 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
620            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
621  def : Pat<(v8i16 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
622            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
623  def : Pat<(v8f16 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
624            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
625  def : Pat<(v4i32 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
626            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
627  def : Pat<(v4f32 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
628            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
629  def : Pat<(v2f64 (X86cmov VR128X:$t, VR128X:$f, timm:$cond, EFLAGS)),
630            (CMOV_VR128X VR128X:$t, VR128X:$f, timm:$cond)>;
631
632  def : Pat<(v32i8 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
633            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
634  def : Pat<(v16i16 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
635            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
636  def : Pat<(v16f16 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
637            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
638  def : Pat<(v8i32 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
639            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
640  def : Pat<(v8f32 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
641            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
642  def : Pat<(v4f64 (X86cmov VR256X:$t, VR256X:$f, timm:$cond, EFLAGS)),
643            (CMOV_VR256X VR256X:$t, VR256X:$f, timm:$cond)>;
644}
645
646def : Pat<(v64i8 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
647          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
648def : Pat<(v32i16 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
649          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
650def : Pat<(v32f16 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
651          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
652def : Pat<(v16i32 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
653          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
654def : Pat<(v16f32 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
655          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
656def : Pat<(v8f64 (X86cmov VR512:$t, VR512:$f, timm:$cond, EFLAGS)),
657          (CMOV_VR512 VR512:$t, VR512:$f, timm:$cond)>;
658
659//===----------------------------------------------------------------------===//
660// Normal-Instructions-With-Lock-Prefix Pseudo Instructions
661//===----------------------------------------------------------------------===//
662
663// FIXME: Use normal instructions and add lock prefix dynamically.
664
665// Memory barriers
666
667let isCodeGenOnly = 1, Defs = [EFLAGS] in
668def OR32mi8Locked  : Ii8<0x83, MRM1m, (outs), (ins i32mem:$dst, i32i8imm:$zero),
669                         "or{l}\t{$zero, $dst|$dst, $zero}", []>,
670                         Requires<[Not64BitMode]>, OpSize32, LOCK,
671                         Sched<[WriteALURMW]>;
672
673let hasSideEffects = 1 in
674def Int_MemBarrier : I<0, Pseudo, (outs), (ins),
675                     "#MEMBARRIER",
676                     [(X86MemBarrier)]>, Sched<[WriteLoad]>;
677
678// RegOpc corresponds to the mr version of the instruction
679// ImmOpc corresponds to the mi version of the instruction
680// ImmOpc8 corresponds to the mi8 version of the instruction
681// ImmMod corresponds to the instruction format of the mi and mi8 versions
682multiclass LOCK_ArithBinOp<bits<8> RegOpc, bits<8> ImmOpc, bits<8> ImmOpc8,
683                           Format ImmMod, SDNode Op, string mnemonic> {
684let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
685    SchedRW = [WriteALURMW] in {
686
687def NAME#8mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
688                  RegOpc{3}, RegOpc{2}, RegOpc{1}, 0 },
689                  MRMDestMem, (outs), (ins i8mem:$dst, GR8:$src2),
690                  !strconcat(mnemonic, "{b}\t",
691                             "{$src2, $dst|$dst, $src2}"),
692                  [(set EFLAGS, (Op addr:$dst, GR8:$src2))]>, LOCK;
693
694def NAME#16mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
695                   RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
696                   MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
697                   !strconcat(mnemonic, "{w}\t",
698                              "{$src2, $dst|$dst, $src2}"),
699                   [(set EFLAGS, (Op addr:$dst, GR16:$src2))]>,
700                   OpSize16, LOCK;
701
702def NAME#32mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
703                   RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
704                   MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
705                   !strconcat(mnemonic, "{l}\t",
706                              "{$src2, $dst|$dst, $src2}"),
707                   [(set EFLAGS, (Op addr:$dst, GR32:$src2))]>,
708                   OpSize32, LOCK;
709
710def NAME#64mr : RI<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
711                    RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
712                    MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
713                    !strconcat(mnemonic, "{q}\t",
714                               "{$src2, $dst|$dst, $src2}"),
715                    [(set EFLAGS, (Op addr:$dst, GR64:$src2))]>, LOCK;
716
717// NOTE: These are order specific, we want the mi8 forms to be listed
718// first so that they are slightly preferred to the mi forms.
719def NAME#16mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
720                      ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
721                      ImmMod, (outs), (ins i16mem :$dst, i16i8imm :$src2),
722                      !strconcat(mnemonic, "{w}\t",
723                                 "{$src2, $dst|$dst, $src2}"),
724                      [(set EFLAGS, (Op addr:$dst, i16immSExt8:$src2))]>,
725                      OpSize16, LOCK;
726
727def NAME#32mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
728                      ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
729                      ImmMod, (outs), (ins i32mem :$dst, i32i8imm :$src2),
730                      !strconcat(mnemonic, "{l}\t",
731                                 "{$src2, $dst|$dst, $src2}"),
732                      [(set EFLAGS, (Op addr:$dst, i32immSExt8:$src2))]>,
733                      OpSize32, LOCK;
734
735def NAME#64mi8 : RIi8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
736                       ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
737                       ImmMod, (outs), (ins i64mem :$dst, i64i8imm :$src2),
738                       !strconcat(mnemonic, "{q}\t",
739                                  "{$src2, $dst|$dst, $src2}"),
740                       [(set EFLAGS, (Op addr:$dst, i64immSExt8:$src2))]>,
741                       LOCK;
742
743def NAME#8mi : Ii8<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
744                    ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 0 },
745                    ImmMod, (outs), (ins i8mem :$dst, i8imm :$src2),
746                    !strconcat(mnemonic, "{b}\t",
747                               "{$src2, $dst|$dst, $src2}"),
748                    [(set EFLAGS, (Op addr:$dst, (i8 imm:$src2)))]>, LOCK;
749
750def NAME#16mi : Ii16<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
751                      ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
752                      ImmMod, (outs), (ins i16mem :$dst, i16imm :$src2),
753                      !strconcat(mnemonic, "{w}\t",
754                                 "{$src2, $dst|$dst, $src2}"),
755                      [(set EFLAGS, (Op addr:$dst, (i16 imm:$src2)))]>,
756                      OpSize16, LOCK;
757
758def NAME#32mi : Ii32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
759                      ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
760                      ImmMod, (outs), (ins i32mem :$dst, i32imm :$src2),
761                      !strconcat(mnemonic, "{l}\t",
762                                 "{$src2, $dst|$dst, $src2}"),
763                      [(set EFLAGS, (Op addr:$dst, (i32 imm:$src2)))]>,
764                      OpSize32, LOCK;
765
766def NAME#64mi32 : RIi32S<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
767                          ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
768                          ImmMod, (outs), (ins i64mem :$dst, i64i32imm :$src2),
769                          !strconcat(mnemonic, "{q}\t",
770                                     "{$src2, $dst|$dst, $src2}"),
771                          [(set EFLAGS, (Op addr:$dst, i64immSExt32:$src2))]>,
772                          LOCK;
773}
774
775}
776
777defm LOCK_ADD : LOCK_ArithBinOp<0x00, 0x80, 0x83, MRM0m, X86lock_add, "add">;
778defm LOCK_SUB : LOCK_ArithBinOp<0x28, 0x80, 0x83, MRM5m, X86lock_sub, "sub">;
779defm LOCK_OR  : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM1m, X86lock_or , "or">;
780defm LOCK_AND : LOCK_ArithBinOp<0x20, 0x80, 0x83, MRM4m, X86lock_and, "and">;
781defm LOCK_XOR : LOCK_ArithBinOp<0x30, 0x80, 0x83, MRM6m, X86lock_xor, "xor">;
782
783def X86lock_add_nocf : PatFrag<(ops node:$lhs, node:$rhs),
784                               (X86lock_add node:$lhs, node:$rhs), [{
785  return hasNoCarryFlagUses(SDValue(N, 0));
786}]>;
787
788def X86lock_sub_nocf : PatFrag<(ops node:$lhs, node:$rhs),
789                               (X86lock_sub node:$lhs, node:$rhs), [{
790  return hasNoCarryFlagUses(SDValue(N, 0));
791}]>;
792
793let Predicates = [UseIncDec] in {
794  let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
795      SchedRW = [WriteALURMW]  in {
796    def LOCK_INC8m  : I<0xFE, MRM0m, (outs), (ins i8mem :$dst),
797                        "inc{b}\t$dst",
798                        [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i8 1)))]>,
799                        LOCK;
800    def LOCK_INC16m : I<0xFF, MRM0m, (outs), (ins i16mem:$dst),
801                        "inc{w}\t$dst",
802                        [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i16 1)))]>,
803                        OpSize16, LOCK;
804    def LOCK_INC32m : I<0xFF, MRM0m, (outs), (ins i32mem:$dst),
805                        "inc{l}\t$dst",
806                        [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i32 1)))]>,
807                        OpSize32, LOCK;
808    def LOCK_INC64m : RI<0xFF, MRM0m, (outs), (ins i64mem:$dst),
809                         "inc{q}\t$dst",
810                         [(set EFLAGS, (X86lock_add_nocf addr:$dst, (i64 1)))]>,
811                         LOCK;
812
813    def LOCK_DEC8m  : I<0xFE, MRM1m, (outs), (ins i8mem :$dst),
814                        "dec{b}\t$dst",
815                        [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i8 1)))]>,
816                        LOCK;
817    def LOCK_DEC16m : I<0xFF, MRM1m, (outs), (ins i16mem:$dst),
818                        "dec{w}\t$dst",
819                        [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i16 1)))]>,
820                        OpSize16, LOCK;
821    def LOCK_DEC32m : I<0xFF, MRM1m, (outs), (ins i32mem:$dst),
822                        "dec{l}\t$dst",
823                        [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i32 1)))]>,
824                        OpSize32, LOCK;
825    def LOCK_DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst),
826                         "dec{q}\t$dst",
827                         [(set EFLAGS, (X86lock_sub_nocf addr:$dst, (i64 1)))]>,
828                         LOCK;
829  }
830
831  // Additional patterns for -1 constant.
832  def : Pat<(X86lock_add addr:$dst, (i8  -1)), (LOCK_DEC8m  addr:$dst)>;
833  def : Pat<(X86lock_add addr:$dst, (i16 -1)), (LOCK_DEC16m addr:$dst)>;
834  def : Pat<(X86lock_add addr:$dst, (i32 -1)), (LOCK_DEC32m addr:$dst)>;
835  def : Pat<(X86lock_add addr:$dst, (i64 -1)), (LOCK_DEC64m addr:$dst)>;
836  def : Pat<(X86lock_sub addr:$dst, (i8  -1)), (LOCK_INC8m  addr:$dst)>;
837  def : Pat<(X86lock_sub addr:$dst, (i16 -1)), (LOCK_INC16m addr:$dst)>;
838  def : Pat<(X86lock_sub addr:$dst, (i32 -1)), (LOCK_INC32m addr:$dst)>;
839  def : Pat<(X86lock_sub addr:$dst, (i64 -1)), (LOCK_INC64m addr:$dst)>;
840}
841
842// Atomic compare and swap.
843multiclass LCMPXCHG_BinOp<bits<8> Opc8, bits<8> Opc, Format Form,
844                          string mnemonic, SDPatternOperator frag> {
845let isCodeGenOnly = 1, SchedRW = [WriteCMPXCHGRMW] in {
846  let Defs = [AL, EFLAGS], Uses = [AL] in
847  def NAME#8  : I<Opc8, Form, (outs), (ins i8mem:$ptr, GR8:$swap),
848                  !strconcat(mnemonic, "{b}\t{$swap, $ptr|$ptr, $swap}"),
849                  [(frag addr:$ptr, GR8:$swap, 1)]>, TB, LOCK;
850  let Defs = [AX, EFLAGS], Uses = [AX] in
851  def NAME#16 : I<Opc, Form, (outs), (ins i16mem:$ptr, GR16:$swap),
852                  !strconcat(mnemonic, "{w}\t{$swap, $ptr|$ptr, $swap}"),
853                  [(frag addr:$ptr, GR16:$swap, 2)]>, TB, OpSize16, LOCK;
854  let Defs = [EAX, EFLAGS], Uses = [EAX] in
855  def NAME#32 : I<Opc, Form, (outs), (ins i32mem:$ptr, GR32:$swap),
856                  !strconcat(mnemonic, "{l}\t{$swap, $ptr|$ptr, $swap}"),
857                  [(frag addr:$ptr, GR32:$swap, 4)]>, TB, OpSize32, LOCK;
858  let Defs = [RAX, EFLAGS], Uses = [RAX] in
859  def NAME#64 : RI<Opc, Form, (outs), (ins i64mem:$ptr, GR64:$swap),
860                   !strconcat(mnemonic, "{q}\t{$swap, $ptr|$ptr, $swap}"),
861                   [(frag addr:$ptr, GR64:$swap, 8)]>, TB, LOCK;
862}
863}
864
865let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
866    Predicates = [HasCmpxchg8b], SchedRW = [WriteCMPXCHGRMW],
867    isCodeGenOnly = 1, usesCustomInserter = 1 in {
868def LCMPXCHG8B : I<0xC7, MRM1m, (outs), (ins i64mem:$ptr),
869                   "cmpxchg8b\t$ptr",
870                   [(X86cas8 addr:$ptr)]>, TB, LOCK;
871}
872
873let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RBX, RCX, RDX],
874    Predicates = [HasCmpxchg16b,In64BitMode], SchedRW = [WriteCMPXCHGRMW],
875    isCodeGenOnly = 1, mayLoad = 1, mayStore = 1, hasSideEffects = 0 in {
876def LCMPXCHG16B : RI<0xC7, MRM1m, (outs), (ins i128mem:$ptr),
877                     "cmpxchg16b\t$ptr",
878                     []>, TB, LOCK;
879}
880
881// This pseudo must be used when the frame uses RBX as
882// the base pointer. Indeed, in such situation RBX is a reserved
883// register and the register allocator will ignore any use/def of
884// it. In other words, the register will not fix the clobbering of
885// RBX that will happen when setting the arguments for the instrucion.
886//
887// Unlike the actual related instruction, we mark that this one
888// defines RBX (instead of using RBX).
889// The rationale is that we will define RBX during the expansion of
890// the pseudo. The argument feeding RBX is rbx_input.
891//
892// The additional argument, $rbx_save, is a temporary register used to
893// save the value of RBX across the actual instruction.
894//
895// To make sure the register assigned to $rbx_save does not interfere with
896// the definition of the actual instruction, we use a definition $dst which
897// is tied to $rbx_save. That way, the live-range of $rbx_save spans across
898// the instruction and we are sure we will have a valid register to restore
899// the value of RBX.
900let Defs = [RAX, RDX, RBX, EFLAGS], Uses = [RAX, RCX, RDX],
901    Predicates = [HasCmpxchg16b,In64BitMode], SchedRW = [WriteCMPXCHGRMW],
902    isCodeGenOnly = 1, isPseudo = 1,
903    mayLoad = 1, mayStore = 1, hasSideEffects = 0,
904    Constraints = "$rbx_save = $dst" in {
905def LCMPXCHG16B_SAVE_RBX :
906    I<0, Pseudo, (outs GR64:$dst),
907      (ins i128mem:$ptr, GR64:$rbx_input, GR64:$rbx_save), "", []>;
908}
909
910// Pseudo instruction that doesn't read/write RBX. Will be turned into either
911// LCMPXCHG16B_SAVE_RBX or LCMPXCHG16B via a custom inserter.
912let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RCX, RDX],
913    Predicates = [HasCmpxchg16b,In64BitMode], SchedRW = [WriteCMPXCHGRMW],
914    isCodeGenOnly = 1, isPseudo = 1,
915    mayLoad = 1, mayStore = 1, hasSideEffects = 0,
916    usesCustomInserter = 1 in {
917def LCMPXCHG16B_NO_RBX :
918    I<0, Pseudo, (outs), (ins i128mem:$ptr, GR64:$rbx_input), "",
919      [(X86cas16 addr:$ptr, GR64:$rbx_input)]>;
920}
921
922// This pseudo must be used when the frame uses RBX/EBX as
923// the base pointer.
924// cf comment for LCMPXCHG16B_SAVE_RBX.
925let Defs = [EBX], Uses = [ECX, EAX],
926    Predicates = [HasMWAITX], SchedRW = [WriteSystem],
927    isCodeGenOnly = 1, isPseudo = 1, Constraints = "$rbx_save = $dst" in {
928def MWAITX_SAVE_RBX :
929    I<0, Pseudo, (outs GR64:$dst),
930      (ins GR32:$ebx_input, GR64:$rbx_save),
931      "mwaitx",
932      []>;
933}
934
935// Pseudo mwaitx instruction to use for custom insertion.
936let Predicates = [HasMWAITX], SchedRW = [WriteSystem],
937    isCodeGenOnly = 1, isPseudo = 1,
938    usesCustomInserter = 1 in {
939def MWAITX :
940    I<0, Pseudo, (outs), (ins GR32:$ecx, GR32:$eax, GR32:$ebx),
941      "mwaitx",
942      [(int_x86_mwaitx GR32:$ecx, GR32:$eax, GR32:$ebx)]>;
943}
944
945
946defm LCMPXCHG : LCMPXCHG_BinOp<0xB0, 0xB1, MRMDestMem, "cmpxchg", X86cas>;
947
948// Atomic exchange and add
949multiclass ATOMIC_RMW_BINOP<bits<8> opc8, bits<8> opc, string mnemonic,
950                            string frag> {
951  let Constraints = "$val = $dst", Defs = [EFLAGS], mayLoad = 1, mayStore = 1,
952      isCodeGenOnly = 1, SchedRW = [WriteALURMW] in {
953    def NAME#8  : I<opc8, MRMSrcMem, (outs GR8:$dst),
954                    (ins GR8:$val, i8mem:$ptr),
955                    !strconcat(mnemonic, "{b}\t{$val, $ptr|$ptr, $val}"),
956                    [(set GR8:$dst,
957                          (!cast<PatFrag>(frag # "_8") addr:$ptr, GR8:$val))]>;
958    def NAME#16 : I<opc, MRMSrcMem, (outs GR16:$dst),
959                    (ins GR16:$val, i16mem:$ptr),
960                    !strconcat(mnemonic, "{w}\t{$val, $ptr|$ptr, $val}"),
961                    [(set
962                       GR16:$dst,
963                       (!cast<PatFrag>(frag # "_16") addr:$ptr, GR16:$val))]>,
964                    OpSize16;
965    def NAME#32 : I<opc, MRMSrcMem, (outs GR32:$dst),
966                    (ins GR32:$val, i32mem:$ptr),
967                    !strconcat(mnemonic, "{l}\t{$val, $ptr|$ptr, $val}"),
968                    [(set
969                       GR32:$dst,
970                       (!cast<PatFrag>(frag # "_32") addr:$ptr, GR32:$val))]>,
971                    OpSize32;
972    def NAME#64 : RI<opc, MRMSrcMem, (outs GR64:$dst),
973                     (ins GR64:$val, i64mem:$ptr),
974                     !strconcat(mnemonic, "{q}\t{$val, $ptr|$ptr, $val}"),
975                     [(set
976                        GR64:$dst,
977                        (!cast<PatFrag>(frag # "_64") addr:$ptr, GR64:$val))]>;
978  }
979}
980
981defm LXADD : ATOMIC_RMW_BINOP<0xc0, 0xc1, "xadd", "atomic_load_add">, TB, LOCK;
982
983/* The following multiclass tries to make sure that in code like
984 *    x.store (immediate op x.load(acquire), release)
985 * and
986 *    x.store (register op x.load(acquire), release)
987 * an operation directly on memory is generated instead of wasting a register.
988 * It is not automatic as atomic_store/load are only lowered to MOV instructions
989 * extremely late to prevent them from being accidentally reordered in the backend
990 * (see below the RELEASE_MOV* / ACQUIRE_MOV* pseudo-instructions)
991 */
992multiclass RELEASE_BINOP_MI<string Name, SDNode op> {
993  def : Pat<(atomic_store_8 addr:$dst,
994             (op (atomic_load_8 addr:$dst), (i8 imm:$src))),
995            (!cast<Instruction>(Name#"8mi") addr:$dst, imm:$src)>;
996  def : Pat<(atomic_store_16 addr:$dst,
997             (op (atomic_load_16 addr:$dst), (i16 imm:$src))),
998            (!cast<Instruction>(Name#"16mi") addr:$dst, imm:$src)>;
999  def : Pat<(atomic_store_32 addr:$dst,
1000             (op (atomic_load_32 addr:$dst), (i32 imm:$src))),
1001            (!cast<Instruction>(Name#"32mi") addr:$dst, imm:$src)>;
1002  def : Pat<(atomic_store_64 addr:$dst,
1003             (op (atomic_load_64 addr:$dst), (i64immSExt32:$src))),
1004            (!cast<Instruction>(Name#"64mi32") addr:$dst, (i64immSExt32:$src))>;
1005
1006  def : Pat<(atomic_store_8 addr:$dst,
1007             (op (atomic_load_8 addr:$dst), (i8 GR8:$src))),
1008            (!cast<Instruction>(Name#"8mr") addr:$dst, GR8:$src)>;
1009  def : Pat<(atomic_store_16 addr:$dst,
1010             (op (atomic_load_16 addr:$dst), (i16 GR16:$src))),
1011            (!cast<Instruction>(Name#"16mr") addr:$dst, GR16:$src)>;
1012  def : Pat<(atomic_store_32 addr:$dst,
1013             (op (atomic_load_32 addr:$dst), (i32 GR32:$src))),
1014            (!cast<Instruction>(Name#"32mr") addr:$dst, GR32:$src)>;
1015  def : Pat<(atomic_store_64 addr:$dst,
1016             (op (atomic_load_64 addr:$dst), (i64 GR64:$src))),
1017            (!cast<Instruction>(Name#"64mr") addr:$dst, GR64:$src)>;
1018}
1019defm : RELEASE_BINOP_MI<"ADD", add>;
1020defm : RELEASE_BINOP_MI<"AND", and>;
1021defm : RELEASE_BINOP_MI<"OR",  or>;
1022defm : RELEASE_BINOP_MI<"XOR", xor>;
1023defm : RELEASE_BINOP_MI<"SUB", sub>;
1024
1025// Atomic load + floating point patterns.
1026// FIXME: This could also handle SIMD operations with *ps and *pd instructions.
1027multiclass ATOMIC_LOAD_FP_BINOP_MI<string Name, SDNode op> {
1028  def : Pat<(op FR32:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
1029            (!cast<Instruction>(Name#"SSrm") FR32:$src1, addr:$src2)>,
1030            Requires<[UseSSE1]>;
1031  def : Pat<(op FR32:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
1032            (!cast<Instruction>("V"#Name#"SSrm") FR32:$src1, addr:$src2)>,
1033            Requires<[UseAVX]>;
1034  def : Pat<(op FR32X:$src1, (bitconvert (i32 (atomic_load_32 addr:$src2)))),
1035            (!cast<Instruction>("V"#Name#"SSZrm") FR32X:$src1, addr:$src2)>,
1036            Requires<[HasAVX512]>;
1037
1038  def : Pat<(op FR64:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
1039            (!cast<Instruction>(Name#"SDrm") FR64:$src1, addr:$src2)>,
1040            Requires<[UseSSE1]>;
1041  def : Pat<(op FR64:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
1042            (!cast<Instruction>("V"#Name#"SDrm") FR64:$src1, addr:$src2)>,
1043            Requires<[UseAVX]>;
1044  def : Pat<(op FR64X:$src1, (bitconvert (i64 (atomic_load_64 addr:$src2)))),
1045            (!cast<Instruction>("V"#Name#"SDZrm") FR64X:$src1, addr:$src2)>,
1046            Requires<[HasAVX512]>;
1047}
1048defm : ATOMIC_LOAD_FP_BINOP_MI<"ADD", fadd>;
1049// FIXME: Add fsub, fmul, fdiv, ...
1050
1051multiclass RELEASE_UNOP<string Name, dag dag8, dag dag16, dag dag32,
1052                        dag dag64> {
1053  def : Pat<(atomic_store_8 addr:$dst, dag8),
1054            (!cast<Instruction>(Name#8m) addr:$dst)>;
1055  def : Pat<(atomic_store_16 addr:$dst, dag16),
1056            (!cast<Instruction>(Name#16m) addr:$dst)>;
1057  def : Pat<(atomic_store_32 addr:$dst, dag32),
1058            (!cast<Instruction>(Name#32m) addr:$dst)>;
1059  def : Pat<(atomic_store_64 addr:$dst, dag64),
1060            (!cast<Instruction>(Name#64m) addr:$dst)>;
1061}
1062
1063let Predicates = [UseIncDec] in {
1064  defm : RELEASE_UNOP<"INC",
1065      (add (atomic_load_8  addr:$dst), (i8 1)),
1066      (add (atomic_load_16 addr:$dst), (i16 1)),
1067      (add (atomic_load_32 addr:$dst), (i32 1)),
1068      (add (atomic_load_64 addr:$dst), (i64 1))>;
1069  defm : RELEASE_UNOP<"DEC",
1070      (add (atomic_load_8  addr:$dst), (i8 -1)),
1071      (add (atomic_load_16 addr:$dst), (i16 -1)),
1072      (add (atomic_load_32 addr:$dst), (i32 -1)),
1073      (add (atomic_load_64 addr:$dst), (i64 -1))>;
1074}
1075
1076defm : RELEASE_UNOP<"NEG",
1077    (ineg (i8 (atomic_load_8  addr:$dst))),
1078    (ineg (i16 (atomic_load_16 addr:$dst))),
1079    (ineg (i32 (atomic_load_32 addr:$dst))),
1080    (ineg (i64 (atomic_load_64 addr:$dst)))>;
1081defm : RELEASE_UNOP<"NOT",
1082    (not (i8 (atomic_load_8  addr:$dst))),
1083    (not (i16 (atomic_load_16 addr:$dst))),
1084    (not (i32 (atomic_load_32 addr:$dst))),
1085    (not (i64 (atomic_load_64 addr:$dst)))>;
1086
1087def : Pat<(atomic_store_8 addr:$dst, (i8 imm:$src)),
1088          (MOV8mi addr:$dst, imm:$src)>;
1089def : Pat<(atomic_store_16 addr:$dst, (i16 imm:$src)),
1090          (MOV16mi addr:$dst, imm:$src)>;
1091def : Pat<(atomic_store_32 addr:$dst, (i32 imm:$src)),
1092          (MOV32mi addr:$dst, imm:$src)>;
1093def : Pat<(atomic_store_64 addr:$dst, (i64immSExt32:$src)),
1094          (MOV64mi32 addr:$dst, i64immSExt32:$src)>;
1095
1096def : Pat<(atomic_store_8 addr:$dst, GR8:$src),
1097          (MOV8mr addr:$dst, GR8:$src)>;
1098def : Pat<(atomic_store_16 addr:$dst, GR16:$src),
1099          (MOV16mr addr:$dst, GR16:$src)>;
1100def : Pat<(atomic_store_32 addr:$dst, GR32:$src),
1101          (MOV32mr addr:$dst, GR32:$src)>;
1102def : Pat<(atomic_store_64 addr:$dst, GR64:$src),
1103          (MOV64mr addr:$dst, GR64:$src)>;
1104
1105def : Pat<(i8  (atomic_load_8 addr:$src)),  (MOV8rm addr:$src)>;
1106def : Pat<(i16 (atomic_load_16 addr:$src)), (MOV16rm addr:$src)>;
1107def : Pat<(i32 (atomic_load_32 addr:$src)), (MOV32rm addr:$src)>;
1108def : Pat<(i64 (atomic_load_64 addr:$src)), (MOV64rm addr:$src)>;
1109
1110// Floating point loads/stores.
1111def : Pat<(atomic_store_32 addr:$dst, (i32 (bitconvert (f32 FR32:$src)))),
1112          (MOVSSmr addr:$dst, FR32:$src)>, Requires<[UseSSE1]>;
1113def : Pat<(atomic_store_32 addr:$dst, (i32 (bitconvert (f32 FR32:$src)))),
1114          (VMOVSSmr addr:$dst, FR32:$src)>, Requires<[UseAVX]>;
1115def : Pat<(atomic_store_32 addr:$dst, (i32 (bitconvert (f32 FR32:$src)))),
1116          (VMOVSSZmr addr:$dst, FR32:$src)>, Requires<[HasAVX512]>;
1117
1118def : Pat<(atomic_store_64 addr:$dst, (i64 (bitconvert (f64 FR64:$src)))),
1119          (MOVSDmr addr:$dst, FR64:$src)>, Requires<[UseSSE2]>;
1120def : Pat<(atomic_store_64 addr:$dst, (i64 (bitconvert (f64 FR64:$src)))),
1121          (VMOVSDmr addr:$dst, FR64:$src)>, Requires<[UseAVX]>;
1122def : Pat<(atomic_store_64 addr:$dst, (i64 (bitconvert (f64 FR64:$src)))),
1123          (VMOVSDmr addr:$dst, FR64:$src)>, Requires<[HasAVX512]>;
1124
1125def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1126          (MOVSSrm_alt addr:$src)>, Requires<[UseSSE1]>;
1127def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1128          (VMOVSSrm_alt addr:$src)>, Requires<[UseAVX]>;
1129def : Pat<(f32 (bitconvert (i32 (atomic_load_32 addr:$src)))),
1130          (VMOVSSZrm_alt addr:$src)>, Requires<[HasAVX512]>;
1131
1132def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1133          (MOVSDrm_alt addr:$src)>, Requires<[UseSSE2]>;
1134def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1135          (VMOVSDrm_alt addr:$src)>, Requires<[UseAVX]>;
1136def : Pat<(f64 (bitconvert (i64 (atomic_load_64 addr:$src)))),
1137          (VMOVSDZrm_alt addr:$src)>, Requires<[HasAVX512]>;
1138
1139//===----------------------------------------------------------------------===//
1140// DAG Pattern Matching Rules
1141//===----------------------------------------------------------------------===//
1142
1143// Use AND/OR to store 0/-1 in memory when optimizing for minsize. This saves
1144// binary size compared to a regular MOV, but it introduces an unnecessary
1145// load, so is not suitable for regular or optsize functions.
1146let Predicates = [OptForMinSize] in {
1147def : Pat<(simple_store (i16 0), addr:$dst), (AND16mi8 addr:$dst, 0)>;
1148def : Pat<(simple_store (i32 0), addr:$dst), (AND32mi8 addr:$dst, 0)>;
1149def : Pat<(simple_store (i64 0), addr:$dst), (AND64mi8 addr:$dst, 0)>;
1150def : Pat<(simple_store (i16 -1), addr:$dst), (OR16mi8 addr:$dst, -1)>;
1151def : Pat<(simple_store (i32 -1), addr:$dst), (OR32mi8 addr:$dst, -1)>;
1152def : Pat<(simple_store (i64 -1), addr:$dst), (OR64mi8 addr:$dst, -1)>;
1153}
1154
1155// In kernel code model, we can get the address of a label
1156// into a register with 'movq'.  FIXME: This is a hack, the 'imm' predicate of
1157// the MOV64ri32 should accept these.
1158def : Pat<(i64 (X86Wrapper tconstpool  :$dst)),
1159          (MOV64ri32 tconstpool  :$dst)>, Requires<[KernelCode]>;
1160def : Pat<(i64 (X86Wrapper tjumptable  :$dst)),
1161          (MOV64ri32 tjumptable  :$dst)>, Requires<[KernelCode]>;
1162def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
1163          (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
1164def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
1165          (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
1166def : Pat<(i64 (X86Wrapper mcsym:$dst)),
1167          (MOV64ri32 mcsym:$dst)>, Requires<[KernelCode]>;
1168def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
1169          (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
1170
1171// If we have small model and -static mode, it is safe to store global addresses
1172// directly as immediates.  FIXME: This is really a hack, the 'imm' predicate
1173// for MOV64mi32 should handle this sort of thing.
1174def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
1175          (MOV64mi32 addr:$dst, tconstpool:$src)>,
1176          Requires<[NearData, IsNotPIC]>;
1177def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
1178          (MOV64mi32 addr:$dst, tjumptable:$src)>,
1179          Requires<[NearData, IsNotPIC]>;
1180def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
1181          (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
1182          Requires<[NearData, IsNotPIC]>;
1183def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
1184          (MOV64mi32 addr:$dst, texternalsym:$src)>,
1185          Requires<[NearData, IsNotPIC]>;
1186def : Pat<(store (i64 (X86Wrapper mcsym:$src)), addr:$dst),
1187          (MOV64mi32 addr:$dst, mcsym:$src)>,
1188          Requires<[NearData, IsNotPIC]>;
1189def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
1190          (MOV64mi32 addr:$dst, tblockaddress:$src)>,
1191          Requires<[NearData, IsNotPIC]>;
1192
1193def : Pat<(i32 (X86RecoverFrameAlloc mcsym:$dst)), (MOV32ri mcsym:$dst)>;
1194def : Pat<(i64 (X86RecoverFrameAlloc mcsym:$dst)), (MOV64ri mcsym:$dst)>;
1195
1196// Calls
1197
1198// tls has some funny stuff here...
1199// This corresponds to movabs $foo@tpoff, %rax
1200def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
1201          (MOV64ri32 tglobaltlsaddr :$dst)>;
1202// This corresponds to add $foo@tpoff, %rax
1203def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
1204          (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
1205
1206
1207// Direct PC relative function call for small code model. 32-bit displacement
1208// sign extended to 64-bit.
1209def : Pat<(X86call (i64 tglobaladdr:$dst)),
1210          (CALL64pcrel32 tglobaladdr:$dst)>;
1211def : Pat<(X86call (i64 texternalsym:$dst)),
1212          (CALL64pcrel32 texternalsym:$dst)>;
1213
1214def : Pat<(X86call_rvmarker (i64 tglobaladdr:$rvfunc), (i64 texternalsym:$dst)),
1215          (CALL64pcrel32_RVMARKER tglobaladdr:$rvfunc, texternalsym:$dst)>;
1216def : Pat<(X86call_rvmarker (i64 tglobaladdr:$rvfunc), (i64 tglobaladdr:$dst)),
1217          (CALL64pcrel32_RVMARKER tglobaladdr:$rvfunc, tglobaladdr:$dst)>;
1218
1219
1220// Tailcall stuff. The TCRETURN instructions execute after the epilog, so they
1221// can never use callee-saved registers. That is the purpose of the GR64_TC
1222// register classes.
1223//
1224// The only volatile register that is never used by the calling convention is
1225// %r11. This happens when calling a vararg function with 6 arguments.
1226//
1227// Match an X86tcret that uses less than 7 volatile registers.
1228def X86tcret_6regs : PatFrag<(ops node:$ptr, node:$off),
1229                             (X86tcret node:$ptr, node:$off), [{
1230  // X86tcret args: (*chain, ptr, imm, regs..., glue)
1231  unsigned NumRegs = 0;
1232  for (unsigned i = 3, e = N->getNumOperands(); i != e; ++i)
1233    if (isa<RegisterSDNode>(N->getOperand(i)) && ++NumRegs > 6)
1234      return false;
1235  return true;
1236}]>;
1237
1238def : Pat<(X86tcret ptr_rc_tailcall:$dst, timm:$off),
1239          (TCRETURNri ptr_rc_tailcall:$dst, timm:$off)>,
1240          Requires<[Not64BitMode, NotUseIndirectThunkCalls]>;
1241
1242// FIXME: This is disabled for 32-bit PIC mode because the global base
1243// register which is part of the address mode may be assigned a
1244// callee-saved register.
1245def : Pat<(X86tcret (load addr:$dst), timm:$off),
1246          (TCRETURNmi addr:$dst, timm:$off)>,
1247          Requires<[Not64BitMode, IsNotPIC, NotUseIndirectThunkCalls]>;
1248
1249def : Pat<(X86tcret (i32 tglobaladdr:$dst), timm:$off),
1250          (TCRETURNdi tglobaladdr:$dst, timm:$off)>,
1251          Requires<[NotLP64]>;
1252
1253def : Pat<(X86tcret (i32 texternalsym:$dst), timm:$off),
1254          (TCRETURNdi texternalsym:$dst, timm:$off)>,
1255          Requires<[NotLP64]>;
1256
1257def : Pat<(X86tcret ptr_rc_tailcall:$dst, timm:$off),
1258          (TCRETURNri64 ptr_rc_tailcall:$dst, timm:$off)>,
1259          Requires<[In64BitMode, NotUseIndirectThunkCalls]>;
1260
1261// Don't fold loads into X86tcret requiring more than 6 regs.
1262// There wouldn't be enough scratch registers for base+index.
1263def : Pat<(X86tcret_6regs (load addr:$dst), timm:$off),
1264          (TCRETURNmi64 addr:$dst, timm:$off)>,
1265          Requires<[In64BitMode, NotUseIndirectThunkCalls]>;
1266
1267def : Pat<(X86tcret ptr_rc_tailcall:$dst, timm:$off),
1268          (INDIRECT_THUNK_TCRETURN64 ptr_rc_tailcall:$dst, timm:$off)>,
1269          Requires<[In64BitMode, UseIndirectThunkCalls]>;
1270
1271def : Pat<(X86tcret ptr_rc_tailcall:$dst, timm:$off),
1272          (INDIRECT_THUNK_TCRETURN32 ptr_rc_tailcall:$dst, timm:$off)>,
1273          Requires<[Not64BitMode, UseIndirectThunkCalls]>;
1274
1275def : Pat<(X86tcret (i64 tglobaladdr:$dst), timm:$off),
1276          (TCRETURNdi64 tglobaladdr:$dst, timm:$off)>,
1277          Requires<[IsLP64]>;
1278
1279def : Pat<(X86tcret (i64 texternalsym:$dst), timm:$off),
1280          (TCRETURNdi64 texternalsym:$dst, timm:$off)>,
1281          Requires<[IsLP64]>;
1282
1283// Normal calls, with various flavors of addresses.
1284def : Pat<(X86call (i32 tglobaladdr:$dst)),
1285          (CALLpcrel32 tglobaladdr:$dst)>;
1286def : Pat<(X86call (i32 texternalsym:$dst)),
1287          (CALLpcrel32 texternalsym:$dst)>;
1288def : Pat<(X86call (i32 imm:$dst)),
1289          (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
1290
1291// Comparisons.
1292
1293// TEST R,R is smaller than CMP R,0
1294def : Pat<(X86cmp GR8:$src1, 0),
1295          (TEST8rr GR8:$src1, GR8:$src1)>;
1296def : Pat<(X86cmp GR16:$src1, 0),
1297          (TEST16rr GR16:$src1, GR16:$src1)>;
1298def : Pat<(X86cmp GR32:$src1, 0),
1299          (TEST32rr GR32:$src1, GR32:$src1)>;
1300def : Pat<(X86cmp GR64:$src1, 0),
1301          (TEST64rr GR64:$src1, GR64:$src1)>;
1302
1303// zextload bool -> zextload byte
1304// i1 stored in one byte in zero-extended form.
1305// Upper bits cleanup should be executed before Store.
1306def : Pat<(zextloadi8i1  addr:$src), (MOV8rm addr:$src)>;
1307def : Pat<(zextloadi16i1 addr:$src),
1308          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1309def : Pat<(zextloadi32i1 addr:$src), (MOVZX32rm8 addr:$src)>;
1310def : Pat<(zextloadi64i1 addr:$src),
1311          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1312
1313// extload bool -> extload byte
1314// When extloading from 16-bit and smaller memory locations into 64-bit
1315// registers, use zero-extending loads so that the entire 64-bit register is
1316// defined, avoiding partial-register updates.
1317
1318def : Pat<(extloadi8i1 addr:$src),   (MOV8rm      addr:$src)>;
1319def : Pat<(extloadi16i1 addr:$src),
1320          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1321def : Pat<(extloadi32i1 addr:$src),  (MOVZX32rm8  addr:$src)>;
1322def : Pat<(extloadi16i8 addr:$src),
1323          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1324def : Pat<(extloadi32i8 addr:$src),  (MOVZX32rm8  addr:$src)>;
1325def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
1326
1327// For other extloads, use subregs, since the high contents of the register are
1328// defined after an extload.
1329// NOTE: The extloadi64i32 pattern needs to be first as it will try to form
1330// 32-bit loads for 4 byte aligned i8/i16 loads.
1331def : Pat<(extloadi64i32 addr:$src),
1332          (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src), sub_32bit)>;
1333def : Pat<(extloadi64i1 addr:$src),
1334          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1335def : Pat<(extloadi64i8 addr:$src),
1336          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1337def : Pat<(extloadi64i16 addr:$src),
1338          (SUBREG_TO_REG (i64 0), (MOVZX32rm16 addr:$src), sub_32bit)>;
1339
1340// anyext. Define these to do an explicit zero-extend to
1341// avoid partial-register updates.
1342def : Pat<(i16 (anyext GR8 :$src)), (EXTRACT_SUBREG
1343                                     (MOVZX32rr8 GR8 :$src), sub_16bit)>;
1344def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8  GR8 :$src)>;
1345
1346// Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
1347def : Pat<(i32 (anyext GR16:$src)),
1348          (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
1349
1350def : Pat<(i64 (anyext GR8 :$src)),
1351          (SUBREG_TO_REG (i64 0), (MOVZX32rr8  GR8  :$src), sub_32bit)>;
1352def : Pat<(i64 (anyext GR16:$src)),
1353          (SUBREG_TO_REG (i64 0), (MOVZX32rr16 GR16 :$src), sub_32bit)>;
1354def : Pat<(i64 (anyext GR32:$src)),
1355          (INSERT_SUBREG (i64 (IMPLICIT_DEF)), GR32:$src, sub_32bit)>;
1356
1357// If this is an anyext of the remainder of an 8-bit sdivrem, use a MOVSX
1358// instead of a MOVZX. The sdivrem lowering will emit emit a MOVSX to move
1359// %ah to the lower byte of a register. By using a MOVSX here we allow a
1360// post-isel peephole to merge the two MOVSX instructions into one.
1361def anyext_sdiv : PatFrag<(ops node:$lhs), (anyext node:$lhs),[{
1362  return (N->getOperand(0).getOpcode() == ISD::SDIVREM &&
1363          N->getOperand(0).getResNo() == 1);
1364}]>;
1365def : Pat<(i32 (anyext_sdiv GR8:$src)), (MOVSX32rr8 GR8:$src)>;
1366
1367// Any instruction that defines a 32-bit result leaves the high half of the
1368// register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
1369// be copying from a truncate. AssertSext/AssertZext/AssertAlign aren't saying
1370// anything about the upper 32 bits, they're probably just qualifying a
1371// CopyFromReg. FREEZE may be coming from a a truncate. Any other 32-bit
1372// operation will zero-extend up to 64 bits.
1373def def32 : PatLeaf<(i32 GR32:$src), [{
1374  return N->getOpcode() != ISD::TRUNCATE &&
1375         N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
1376         N->getOpcode() != ISD::CopyFromReg &&
1377         N->getOpcode() != ISD::AssertSext &&
1378         N->getOpcode() != ISD::AssertZext &&
1379         N->getOpcode() != ISD::AssertAlign &&
1380         N->getOpcode() != ISD::FREEZE;
1381}]>;
1382
1383// In the case of a 32-bit def that is known to implicitly zero-extend,
1384// we can use a SUBREG_TO_REG.
1385def : Pat<(i64 (zext def32:$src)),
1386          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1387def : Pat<(i64 (and (anyext def32:$src), 0x00000000FFFFFFFF)),
1388          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1389
1390//===----------------------------------------------------------------------===//
1391// Pattern match OR as ADD
1392//===----------------------------------------------------------------------===//
1393
1394// If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
1395// 3-addressified into an LEA instruction to avoid copies.  However, we also
1396// want to finally emit these instructions as an or at the end of the code
1397// generator to make the generated code easier to read.  To do this, we select
1398// into "disjoint bits" pseudo ops.
1399
1400// Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.
1401def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
1402  if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1403    return CurDAG->MaskedValueIsZero(N->getOperand(0), CN->getAPIntValue());
1404
1405  KnownBits Known0 = CurDAG->computeKnownBits(N->getOperand(0), 0);
1406  KnownBits Known1 = CurDAG->computeKnownBits(N->getOperand(1), 0);
1407  return (~Known0.Zero & ~Known1.Zero) == 0;
1408}]>;
1409
1410
1411// (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1412// Try this before the selecting to OR.
1413let SchedRW = [WriteALU] in {
1414
1415let isConvertibleToThreeAddress = 1, isPseudo = 1,
1416    Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1417let isCommutable = 1 in {
1418def ADD8rr_DB   : I<0, Pseudo, (outs GR8:$dst), (ins GR8:$src1, GR8:$src2),
1419                    "", // orb/addb REG, REG
1420                    [(set GR8:$dst, (or_is_add GR8:$src1, GR8:$src2))]>;
1421def ADD16rr_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1422                    "", // orw/addw REG, REG
1423                    [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1424def ADD32rr_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1425                    "", // orl/addl REG, REG
1426                    [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1427def ADD64rr_DB  : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1428                    "", // orq/addq REG, REG
1429                    [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1430} // isCommutable
1431
1432// NOTE: These are order specific, we want the ri8 forms to be listed
1433// first so that they are slightly preferred to the ri forms.
1434
1435def ADD8ri_DB :   I<0, Pseudo,
1436                    (outs GR8:$dst), (ins GR8:$src1, i8imm:$src2),
1437                    "", // orb/addb REG, imm8
1438                    [(set GR8:$dst, (or_is_add GR8:$src1, imm:$src2))]>;
1439def ADD16ri8_DB : I<0, Pseudo,
1440                    (outs GR16:$dst), (ins GR16:$src1, i16i8imm:$src2),
1441                    "", // orw/addw REG, imm8
1442                    [(set GR16:$dst,(or_is_add GR16:$src1,i16immSExt8:$src2))]>;
1443def ADD16ri_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1444                    "", // orw/addw REG, imm
1445                    [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1446
1447def ADD32ri8_DB : I<0, Pseudo,
1448                    (outs GR32:$dst), (ins GR32:$src1, i32i8imm:$src2),
1449                    "", // orl/addl REG, imm8
1450                    [(set GR32:$dst,(or_is_add GR32:$src1,i32immSExt8:$src2))]>;
1451def ADD32ri_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1452                    "", // orl/addl REG, imm
1453                    [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1454
1455
1456def ADD64ri8_DB : I<0, Pseudo,
1457                    (outs GR64:$dst), (ins GR64:$src1, i64i8imm:$src2),
1458                    "", // orq/addq REG, imm8
1459                    [(set GR64:$dst, (or_is_add GR64:$src1,
1460                                                i64immSExt8:$src2))]>;
1461def ADD64ri32_DB : I<0, Pseudo,
1462                     (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1463                     "", // orq/addq REG, imm
1464                     [(set GR64:$dst, (or_is_add GR64:$src1,
1465                                                 i64immSExt32:$src2))]>;
1466}
1467} // AddedComplexity, SchedRW
1468
1469//===----------------------------------------------------------------------===//
1470// Pattern match SUB as XOR
1471//===----------------------------------------------------------------------===//
1472
1473// An immediate in the LHS of a subtract can't be encoded in the instruction.
1474// If there is no possibility of a borrow we can use an XOR instead of a SUB
1475// to enable the immediate to be folded.
1476// TODO: Move this to a DAG combine?
1477
1478def sub_is_xor : PatFrag<(ops node:$lhs, node:$rhs), (sub node:$lhs, node:$rhs),[{
1479  if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1480    KnownBits Known = CurDAG->computeKnownBits(N->getOperand(1));
1481
1482    // If all possible ones in the RHS are set in the LHS then there can't be
1483    // a borrow and we can use xor.
1484    return (~Known.Zero).isSubsetOf(CN->getAPIntValue());
1485  }
1486
1487  return false;
1488}]>;
1489
1490let AddedComplexity = 5 in {
1491def : Pat<(sub_is_xor imm:$src2, GR8:$src1),
1492          (XOR8ri GR8:$src1, imm:$src2)>;
1493def : Pat<(sub_is_xor i16immSExt8:$src2, GR16:$src1),
1494          (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1495def : Pat<(sub_is_xor imm:$src2, GR16:$src1),
1496          (XOR16ri GR16:$src1, imm:$src2)>;
1497def : Pat<(sub_is_xor i32immSExt8:$src2, GR32:$src1),
1498          (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1499def : Pat<(sub_is_xor imm:$src2, GR32:$src1),
1500          (XOR32ri GR32:$src1, imm:$src2)>;
1501def : Pat<(sub_is_xor i64immSExt8:$src2, GR64:$src1),
1502          (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1503def : Pat<(sub_is_xor i64immSExt32:$src2, GR64:$src1),
1504          (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1505}
1506
1507//===----------------------------------------------------------------------===//
1508// Some peepholes
1509//===----------------------------------------------------------------------===//
1510
1511// Odd encoding trick: -128 fits into an 8-bit immediate field while
1512// +128 doesn't, so in this special case use a sub instead of an add.
1513def : Pat<(add GR16:$src1, 128),
1514          (SUB16ri8 GR16:$src1, -128)>;
1515def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1516          (SUB16mi8 addr:$dst, -128)>;
1517
1518def : Pat<(add GR32:$src1, 128),
1519          (SUB32ri8 GR32:$src1, -128)>;
1520def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1521          (SUB32mi8 addr:$dst, -128)>;
1522
1523def : Pat<(add GR64:$src1, 128),
1524          (SUB64ri8 GR64:$src1, -128)>;
1525def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1526          (SUB64mi8 addr:$dst, -128)>;
1527
1528def : Pat<(X86add_flag_nocf GR16:$src1, 128),
1529          (SUB16ri8 GR16:$src1, -128)>;
1530def : Pat<(X86add_flag_nocf GR32:$src1, 128),
1531          (SUB32ri8 GR32:$src1, -128)>;
1532def : Pat<(X86add_flag_nocf GR64:$src1, 128),
1533          (SUB64ri8 GR64:$src1, -128)>;
1534
1535// The same trick applies for 32-bit immediate fields in 64-bit
1536// instructions.
1537def : Pat<(add GR64:$src1, 0x0000000080000000),
1538          (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1539def : Pat<(store (add (loadi64 addr:$dst), 0x0000000080000000), addr:$dst),
1540          (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1541
1542def : Pat<(X86add_flag_nocf GR64:$src1, 0x0000000080000000),
1543          (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1544
1545// To avoid needing to materialize an immediate in a register, use a 32-bit and
1546// with implicit zero-extension instead of a 64-bit and if the immediate has at
1547// least 32 bits of leading zeros. If in addition the last 32 bits can be
1548// represented with a sign extension of a 8 bit constant, use that.
1549// This can also reduce instruction size by eliminating the need for the REX
1550// prefix.
1551
1552// AddedComplexity is needed to give priority over i64immSExt8 and i64immSExt32.
1553let AddedComplexity = 1 in {
1554def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
1555          (SUBREG_TO_REG
1556            (i64 0),
1557            (AND32ri8
1558              (EXTRACT_SUBREG GR64:$src, sub_32bit),
1559              (i32 (GetLo32XForm imm:$imm))),
1560            sub_32bit)>;
1561
1562def : Pat<(and GR64:$src, i64immZExt32:$imm),
1563          (SUBREG_TO_REG
1564            (i64 0),
1565            (AND32ri
1566              (EXTRACT_SUBREG GR64:$src, sub_32bit),
1567              (i32 (GetLo32XForm imm:$imm))),
1568            sub_32bit)>;
1569} // AddedComplexity = 1
1570
1571
1572// AddedComplexity is needed due to the increased complexity on the
1573// i64immZExt32SExt8 and i64immZExt32 patterns above. Applying this to all
1574// the MOVZX patterns keeps thems together in DAGIsel tables.
1575let AddedComplexity = 1 in {
1576// r & (2^16-1) ==> movz
1577def : Pat<(and GR32:$src1, 0xffff),
1578          (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1579// r & (2^8-1) ==> movz
1580def : Pat<(and GR32:$src1, 0xff),
1581          (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>;
1582// r & (2^8-1) ==> movz
1583def : Pat<(and GR16:$src1, 0xff),
1584           (EXTRACT_SUBREG (MOVZX32rr8 (EXTRACT_SUBREG GR16:$src1, sub_8bit)),
1585             sub_16bit)>;
1586
1587// r & (2^32-1) ==> movz
1588def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1589          (SUBREG_TO_REG (i64 0),
1590                         (MOV32rr (EXTRACT_SUBREG GR64:$src, sub_32bit)),
1591                         sub_32bit)>;
1592// r & (2^16-1) ==> movz
1593def : Pat<(and GR64:$src, 0xffff),
1594          (SUBREG_TO_REG (i64 0),
1595                      (MOVZX32rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit))),
1596                      sub_32bit)>;
1597// r & (2^8-1) ==> movz
1598def : Pat<(and GR64:$src, 0xff),
1599          (SUBREG_TO_REG (i64 0),
1600                         (MOVZX32rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit))),
1601                         sub_32bit)>;
1602} // AddedComplexity = 1
1603
1604
1605// Try to use BTS/BTR/BTC for single bit operations on the upper 32-bits.
1606
1607def BTRXForm : SDNodeXForm<imm, [{
1608  // Transformation function: Find the lowest 0.
1609  return getI64Imm((uint8_t)N->getAPIntValue().countTrailingOnes(), SDLoc(N));
1610}]>;
1611
1612def BTCBTSXForm : SDNodeXForm<imm, [{
1613  // Transformation function: Find the lowest 1.
1614  return getI64Imm((uint8_t)N->getAPIntValue().countTrailingZeros(), SDLoc(N));
1615}]>;
1616
1617def BTRMask64 : ImmLeaf<i64, [{
1618  return !isUInt<32>(Imm) && !isInt<32>(Imm) && isPowerOf2_64(~Imm);
1619}]>;
1620
1621def BTCBTSMask64 : ImmLeaf<i64, [{
1622  return !isInt<32>(Imm) && isPowerOf2_64(Imm);
1623}]>;
1624
1625// For now only do this for optsize.
1626let AddedComplexity = 1, Predicates=[OptForSize] in {
1627  def : Pat<(and GR64:$src1, BTRMask64:$mask),
1628            (BTR64ri8 GR64:$src1, (BTRXForm imm:$mask))>;
1629  def : Pat<(or GR64:$src1, BTCBTSMask64:$mask),
1630            (BTS64ri8 GR64:$src1, (BTCBTSXForm imm:$mask))>;
1631  def : Pat<(xor GR64:$src1, BTCBTSMask64:$mask),
1632            (BTC64ri8 GR64:$src1, (BTCBTSXForm imm:$mask))>;
1633}
1634
1635
1636// sext_inreg patterns
1637def : Pat<(sext_inreg GR32:$src, i16),
1638          (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1639def : Pat<(sext_inreg GR32:$src, i8),
1640          (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>;
1641
1642def : Pat<(sext_inreg GR16:$src, i8),
1643           (EXTRACT_SUBREG (MOVSX32rr8 (EXTRACT_SUBREG GR16:$src, sub_8bit)),
1644             sub_16bit)>;
1645
1646def : Pat<(sext_inreg GR64:$src, i32),
1647          (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1648def : Pat<(sext_inreg GR64:$src, i16),
1649          (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1650def : Pat<(sext_inreg GR64:$src, i8),
1651          (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1652
1653// sext, sext_load, zext, zext_load
1654def: Pat<(i16 (sext GR8:$src)),
1655          (EXTRACT_SUBREG (MOVSX32rr8 GR8:$src), sub_16bit)>;
1656def: Pat<(sextloadi16i8 addr:$src),
1657          (EXTRACT_SUBREG (MOVSX32rm8 addr:$src), sub_16bit)>;
1658def: Pat<(i16 (zext GR8:$src)),
1659          (EXTRACT_SUBREG (MOVZX32rr8 GR8:$src), sub_16bit)>;
1660def: Pat<(zextloadi16i8 addr:$src),
1661          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1662
1663// trunc patterns
1664def : Pat<(i16 (trunc GR32:$src)),
1665          (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1666def : Pat<(i8 (trunc GR32:$src)),
1667          (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1668                          sub_8bit)>,
1669      Requires<[Not64BitMode]>;
1670def : Pat<(i8 (trunc GR16:$src)),
1671          (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1672                          sub_8bit)>,
1673      Requires<[Not64BitMode]>;
1674def : Pat<(i32 (trunc GR64:$src)),
1675          (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1676def : Pat<(i16 (trunc GR64:$src)),
1677          (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1678def : Pat<(i8 (trunc GR64:$src)),
1679          (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1680def : Pat<(i8 (trunc GR32:$src)),
1681          (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1682      Requires<[In64BitMode]>;
1683def : Pat<(i8 (trunc GR16:$src)),
1684          (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1685      Requires<[In64BitMode]>;
1686
1687def immff00_ffff  : ImmLeaf<i32, [{
1688  return Imm >= 0xff00 && Imm <= 0xffff;
1689}]>;
1690
1691// h-register tricks
1692def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1693          (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)>,
1694      Requires<[Not64BitMode]>;
1695def : Pat<(i8 (trunc (srl_su (i32 (anyext GR16:$src)), (i8 8)))),
1696          (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)>,
1697      Requires<[Not64BitMode]>;
1698def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1699          (EXTRACT_SUBREG GR32:$src, sub_8bit_hi)>,
1700      Requires<[Not64BitMode]>;
1701def : Pat<(srl GR16:$src, (i8 8)),
1702          (EXTRACT_SUBREG
1703            (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1704            sub_16bit)>;
1705def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1706          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>;
1707def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1708          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>;
1709def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1710          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1711def : Pat<(srl (and_su GR32:$src, immff00_ffff), (i8 8)),
1712          (MOVZX32rr8_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1713
1714// h-register tricks.
1715// For now, be conservative on x86-64 and use an h-register extract only if the
1716// value is immediately zero-extended or stored, which are somewhat common
1717// cases. This uses a bunch of code to prevent a register requiring a REX prefix
1718// from being allocated in the same instruction as the h register, as there's
1719// currently no way to describe this requirement to the register allocator.
1720
1721// h-register extract and zero-extend.
1722def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1723          (SUBREG_TO_REG
1724            (i64 0),
1725            (MOVZX32rr8_NOREX
1726              (EXTRACT_SUBREG GR64:$src, sub_8bit_hi)),
1727            sub_32bit)>;
1728def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1729          (SUBREG_TO_REG
1730            (i64 0),
1731            (MOVZX32rr8_NOREX
1732              (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1733            sub_32bit)>;
1734def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1735          (SUBREG_TO_REG
1736            (i64 0),
1737            (MOVZX32rr8_NOREX
1738              (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)),
1739            sub_32bit)>;
1740
1741// h-register extract and store.
1742def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1743          (MOV8mr_NOREX
1744            addr:$dst,
1745            (EXTRACT_SUBREG GR64:$src, sub_8bit_hi))>;
1746def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1747          (MOV8mr_NOREX
1748            addr:$dst,
1749            (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>,
1750      Requires<[In64BitMode]>;
1751def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1752          (MOV8mr_NOREX
1753            addr:$dst,
1754            (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>,
1755      Requires<[In64BitMode]>;
1756
1757// Special pattern to catch the last step of __builtin_parity handling. Our
1758// goal is to use an xor of an h-register with the corresponding l-register.
1759// The above patterns would handle this on non 64-bit targets, but for 64-bit
1760// we need to be more careful. We're using a NOREX instruction here in case
1761// register allocation fails to keep the two registers together. So we need to
1762// make sure we can't accidentally mix R8-R15 with an h-register.
1763def : Pat<(X86xor_flag (i8 (trunc GR32:$src)),
1764                       (i8 (trunc (srl_su GR32:$src, (i8 8))))),
1765          (XOR8rr_NOREX (EXTRACT_SUBREG GR32:$src, sub_8bit),
1766                        (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
1767
1768// (shl x, 1) ==> (add x, x)
1769// Note that if x is undef (immediate or otherwise), we could theoretically
1770// end up with the two uses of x getting different values, producing a result
1771// where the least significant bit is not 0. However, the probability of this
1772// happening is considered low enough that this is officially not a
1773// "real problem".
1774def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr  GR8 :$src1, GR8 :$src1)>;
1775def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1776def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1777def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1778
1779def shiftMask8 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1780  return isUnneededShiftMask(N, 3);
1781}]>;
1782
1783def shiftMask16 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1784  return isUnneededShiftMask(N, 4);
1785}]>;
1786
1787def shiftMask32 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1788  return isUnneededShiftMask(N, 5);
1789}]>;
1790
1791def shiftMask64 : PatFrag<(ops node:$lhs), (and node:$lhs, imm), [{
1792  return isUnneededShiftMask(N, 6);
1793}]>;
1794
1795
1796// Shift amount is implicitly masked.
1797multiclass MaskedShiftAmountPats<SDNode frag, string name> {
1798  // (shift x (and y, 31)) ==> (shift x, y)
1799  def : Pat<(frag GR8:$src1, (shiftMask32 CL)),
1800            (!cast<Instruction>(name # "8rCL") GR8:$src1)>;
1801  def : Pat<(frag GR16:$src1, (shiftMask32 CL)),
1802            (!cast<Instruction>(name # "16rCL") GR16:$src1)>;
1803  def : Pat<(frag GR32:$src1, (shiftMask32 CL)),
1804            (!cast<Instruction>(name # "32rCL") GR32:$src1)>;
1805  def : Pat<(store (frag (loadi8 addr:$dst), (shiftMask32 CL)), addr:$dst),
1806            (!cast<Instruction>(name # "8mCL") addr:$dst)>;
1807  def : Pat<(store (frag (loadi16 addr:$dst), (shiftMask32 CL)), addr:$dst),
1808            (!cast<Instruction>(name # "16mCL") addr:$dst)>;
1809  def : Pat<(store (frag (loadi32 addr:$dst), (shiftMask32 CL)), addr:$dst),
1810            (!cast<Instruction>(name # "32mCL") addr:$dst)>;
1811
1812  // (shift x (and y, 63)) ==> (shift x, y)
1813  def : Pat<(frag GR64:$src1, (shiftMask64 CL)),
1814            (!cast<Instruction>(name # "64rCL") GR64:$src1)>;
1815  def : Pat<(store (frag (loadi64 addr:$dst), (shiftMask64 CL)), addr:$dst),
1816            (!cast<Instruction>(name # "64mCL") addr:$dst)>;
1817}
1818
1819defm : MaskedShiftAmountPats<shl, "SHL">;
1820defm : MaskedShiftAmountPats<srl, "SHR">;
1821defm : MaskedShiftAmountPats<sra, "SAR">;
1822
1823// ROL/ROR instructions allow a stronger mask optimization than shift for 8- and
1824// 16-bit. We can remove a mask of any (bitwidth - 1) on the rotation amount
1825// because over-rotating produces the same result. This is noted in the Intel
1826// docs with: "tempCOUNT <- (COUNT & COUNTMASK) MOD SIZE". Masking the rotation
1827// amount could affect EFLAGS results, but that does not matter because we are
1828// not tracking flags for these nodes.
1829multiclass MaskedRotateAmountPats<SDNode frag, string name> {
1830  // (rot x (and y, BitWidth - 1)) ==> (rot x, y)
1831  def : Pat<(frag GR8:$src1, (shiftMask8 CL)),
1832  (!cast<Instruction>(name # "8rCL") GR8:$src1)>;
1833  def : Pat<(frag GR16:$src1, (shiftMask16 CL)),
1834  (!cast<Instruction>(name # "16rCL") GR16:$src1)>;
1835  def : Pat<(frag GR32:$src1, (shiftMask32 CL)),
1836  (!cast<Instruction>(name # "32rCL") GR32:$src1)>;
1837  def : Pat<(store (frag (loadi8 addr:$dst), (shiftMask8 CL)), addr:$dst),
1838  (!cast<Instruction>(name # "8mCL") addr:$dst)>;
1839  def : Pat<(store (frag (loadi16 addr:$dst), (shiftMask16 CL)), addr:$dst),
1840  (!cast<Instruction>(name # "16mCL") addr:$dst)>;
1841  def : Pat<(store (frag (loadi32 addr:$dst), (shiftMask32 CL)), addr:$dst),
1842  (!cast<Instruction>(name # "32mCL") addr:$dst)>;
1843
1844  // (rot x (and y, 63)) ==> (rot x, y)
1845  def : Pat<(frag GR64:$src1, (shiftMask64 CL)),
1846  (!cast<Instruction>(name # "64rCL") GR64:$src1)>;
1847  def : Pat<(store (frag (loadi64 addr:$dst), (shiftMask64 CL)), addr:$dst),
1848  (!cast<Instruction>(name # "64mCL") addr:$dst)>;
1849}
1850
1851
1852defm : MaskedRotateAmountPats<rotl, "ROL">;
1853defm : MaskedRotateAmountPats<rotr, "ROR">;
1854
1855// Double "funnel" shift amount is implicitly masked.
1856// (fshl/fshr x (and y, 31)) ==> (fshl/fshr x, y) (NOTE: modulo32)
1857def : Pat<(X86fshl GR16:$src1, GR16:$src2, (shiftMask32 CL)),
1858          (SHLD16rrCL GR16:$src1, GR16:$src2)>;
1859def : Pat<(X86fshr GR16:$src2, GR16:$src1, (shiftMask32 CL)),
1860          (SHRD16rrCL GR16:$src1, GR16:$src2)>;
1861
1862// (fshl/fshr x (and y, 31)) ==> (fshl/fshr x, y)
1863def : Pat<(fshl GR32:$src1, GR32:$src2, (shiftMask32 CL)),
1864          (SHLD32rrCL GR32:$src1, GR32:$src2)>;
1865def : Pat<(fshr GR32:$src2, GR32:$src1, (shiftMask32 CL)),
1866          (SHRD32rrCL GR32:$src1, GR32:$src2)>;
1867
1868// (fshl/fshr x (and y, 63)) ==> (fshl/fshr x, y)
1869def : Pat<(fshl GR64:$src1, GR64:$src2, (shiftMask64 CL)),
1870          (SHLD64rrCL GR64:$src1, GR64:$src2)>;
1871def : Pat<(fshr GR64:$src2, GR64:$src1, (shiftMask64 CL)),
1872          (SHRD64rrCL GR64:$src1, GR64:$src2)>;
1873
1874let Predicates = [HasBMI2] in {
1875  let AddedComplexity = 1 in {
1876    def : Pat<(sra GR32:$src1, (shiftMask32 GR8:$src2)),
1877              (SARX32rr GR32:$src1,
1878                        (INSERT_SUBREG
1879                          (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1880    def : Pat<(sra GR64:$src1, (shiftMask64 GR8:$src2)),
1881              (SARX64rr GR64:$src1,
1882                        (INSERT_SUBREG
1883                          (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1884
1885    def : Pat<(srl GR32:$src1, (shiftMask32 GR8:$src2)),
1886              (SHRX32rr GR32:$src1,
1887                        (INSERT_SUBREG
1888                          (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1889    def : Pat<(srl GR64:$src1, (shiftMask64 GR8:$src2)),
1890              (SHRX64rr GR64:$src1,
1891                        (INSERT_SUBREG
1892                          (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1893
1894    def : Pat<(shl GR32:$src1, (shiftMask32 GR8:$src2)),
1895              (SHLX32rr GR32:$src1,
1896                        (INSERT_SUBREG
1897                          (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1898    def : Pat<(shl GR64:$src1, (shiftMask64 GR8:$src2)),
1899              (SHLX64rr GR64:$src1,
1900                        (INSERT_SUBREG
1901                          (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1902  }
1903
1904  def : Pat<(sra (loadi32 addr:$src1), (shiftMask32 GR8:$src2)),
1905            (SARX32rm addr:$src1,
1906                      (INSERT_SUBREG
1907                        (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1908  def : Pat<(sra (loadi64 addr:$src1), (shiftMask64 GR8:$src2)),
1909            (SARX64rm addr:$src1,
1910                      (INSERT_SUBREG
1911                        (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1912
1913  def : Pat<(srl (loadi32 addr:$src1), (shiftMask32 GR8:$src2)),
1914            (SHRX32rm addr:$src1,
1915                      (INSERT_SUBREG
1916                        (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1917  def : Pat<(srl (loadi64 addr:$src1), (shiftMask64 GR8:$src2)),
1918            (SHRX64rm addr:$src1,
1919                      (INSERT_SUBREG
1920                        (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1921
1922  def : Pat<(shl (loadi32 addr:$src1), (shiftMask32 GR8:$src2)),
1923            (SHLX32rm addr:$src1,
1924                      (INSERT_SUBREG
1925                        (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1926  def : Pat<(shl (loadi64 addr:$src1), (shiftMask64 GR8:$src2)),
1927            (SHLX64rm addr:$src1,
1928                      (INSERT_SUBREG
1929                        (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1930}
1931
1932// Use BTR/BTS/BTC for clearing/setting/toggling a bit in a variable location.
1933multiclass one_bit_patterns<RegisterClass RC, ValueType VT, Instruction BTR,
1934                            Instruction BTS, Instruction BTC,
1935                            PatFrag ShiftMask> {
1936  def : Pat<(and RC:$src1, (rotl -2, GR8:$src2)),
1937            (BTR RC:$src1,
1938                 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1939  def : Pat<(or RC:$src1, (shl 1, GR8:$src2)),
1940            (BTS RC:$src1,
1941                 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1942  def : Pat<(xor RC:$src1, (shl 1, GR8:$src2)),
1943            (BTC RC:$src1,
1944                 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1945
1946  // Similar to above, but removing unneeded masking of the shift amount.
1947  def : Pat<(and RC:$src1, (rotl -2, (ShiftMask GR8:$src2))),
1948            (BTR RC:$src1,
1949                 (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1950  def : Pat<(or RC:$src1, (shl 1, (ShiftMask GR8:$src2))),
1951            (BTS RC:$src1,
1952                (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1953  def : Pat<(xor RC:$src1, (shl 1, (ShiftMask GR8:$src2))),
1954            (BTC RC:$src1,
1955                (INSERT_SUBREG (VT (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
1956}
1957
1958defm : one_bit_patterns<GR16, i16, BTR16rr, BTS16rr, BTC16rr, shiftMask16>;
1959defm : one_bit_patterns<GR32, i32, BTR32rr, BTS32rr, BTC32rr, shiftMask32>;
1960defm : one_bit_patterns<GR64, i64, BTR64rr, BTS64rr, BTC64rr, shiftMask64>;
1961
1962//===----------------------------------------------------------------------===//
1963// EFLAGS-defining Patterns
1964//===----------------------------------------------------------------------===//
1965
1966// add reg, reg
1967def : Pat<(add GR8 :$src1, GR8 :$src2), (ADD8rr  GR8 :$src1, GR8 :$src2)>;
1968def : Pat<(add GR16:$src1, GR16:$src2), (ADD16rr GR16:$src1, GR16:$src2)>;
1969def : Pat<(add GR32:$src1, GR32:$src2), (ADD32rr GR32:$src1, GR32:$src2)>;
1970def : Pat<(add GR64:$src1, GR64:$src2), (ADD64rr GR64:$src1, GR64:$src2)>;
1971
1972// add reg, mem
1973def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
1974          (ADD8rm GR8:$src1, addr:$src2)>;
1975def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
1976          (ADD16rm GR16:$src1, addr:$src2)>;
1977def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
1978          (ADD32rm GR32:$src1, addr:$src2)>;
1979def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
1980          (ADD64rm GR64:$src1, addr:$src2)>;
1981
1982// add reg, imm
1983def : Pat<(add GR8 :$src1, imm:$src2), (ADD8ri  GR8:$src1 , imm:$src2)>;
1984def : Pat<(add GR16:$src1, imm:$src2), (ADD16ri GR16:$src1, imm:$src2)>;
1985def : Pat<(add GR32:$src1, imm:$src2), (ADD32ri GR32:$src1, imm:$src2)>;
1986def : Pat<(add GR16:$src1, i16immSExt8:$src2),
1987          (ADD16ri8 GR16:$src1, i16immSExt8:$src2)>;
1988def : Pat<(add GR32:$src1, i32immSExt8:$src2),
1989          (ADD32ri8 GR32:$src1, i32immSExt8:$src2)>;
1990def : Pat<(add GR64:$src1, i64immSExt8:$src2),
1991          (ADD64ri8 GR64:$src1, i64immSExt8:$src2)>;
1992def : Pat<(add GR64:$src1, i64immSExt32:$src2),
1993          (ADD64ri32 GR64:$src1, i64immSExt32:$src2)>;
1994
1995// sub reg, reg
1996def : Pat<(sub GR8 :$src1, GR8 :$src2), (SUB8rr  GR8 :$src1, GR8 :$src2)>;
1997def : Pat<(sub GR16:$src1, GR16:$src2), (SUB16rr GR16:$src1, GR16:$src2)>;
1998def : Pat<(sub GR32:$src1, GR32:$src2), (SUB32rr GR32:$src1, GR32:$src2)>;
1999def : Pat<(sub GR64:$src1, GR64:$src2), (SUB64rr GR64:$src1, GR64:$src2)>;
2000
2001// sub reg, mem
2002def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
2003          (SUB8rm GR8:$src1, addr:$src2)>;
2004def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
2005          (SUB16rm GR16:$src1, addr:$src2)>;
2006def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
2007          (SUB32rm GR32:$src1, addr:$src2)>;
2008def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
2009          (SUB64rm GR64:$src1, addr:$src2)>;
2010
2011// sub reg, imm
2012def : Pat<(sub GR8:$src1, imm:$src2),
2013          (SUB8ri GR8:$src1, imm:$src2)>;
2014def : Pat<(sub GR16:$src1, imm:$src2),
2015          (SUB16ri GR16:$src1, imm:$src2)>;
2016def : Pat<(sub GR32:$src1, imm:$src2),
2017          (SUB32ri GR32:$src1, imm:$src2)>;
2018def : Pat<(sub GR16:$src1, i16immSExt8:$src2),
2019          (SUB16ri8 GR16:$src1, i16immSExt8:$src2)>;
2020def : Pat<(sub GR32:$src1, i32immSExt8:$src2),
2021          (SUB32ri8 GR32:$src1, i32immSExt8:$src2)>;
2022def : Pat<(sub GR64:$src1, i64immSExt8:$src2),
2023          (SUB64ri8 GR64:$src1, i64immSExt8:$src2)>;
2024def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
2025          (SUB64ri32 GR64:$src1, i64immSExt32:$src2)>;
2026
2027// sub 0, reg
2028def : Pat<(X86sub_flag 0, GR8 :$src), (NEG8r  GR8 :$src)>;
2029def : Pat<(X86sub_flag 0, GR16:$src), (NEG16r GR16:$src)>;
2030def : Pat<(X86sub_flag 0, GR32:$src), (NEG32r GR32:$src)>;
2031def : Pat<(X86sub_flag 0, GR64:$src), (NEG64r GR64:$src)>;
2032
2033// mul reg, reg
2034def : Pat<(mul GR16:$src1, GR16:$src2),
2035          (IMUL16rr GR16:$src1, GR16:$src2)>;
2036def : Pat<(mul GR32:$src1, GR32:$src2),
2037          (IMUL32rr GR32:$src1, GR32:$src2)>;
2038def : Pat<(mul GR64:$src1, GR64:$src2),
2039          (IMUL64rr GR64:$src1, GR64:$src2)>;
2040
2041// mul reg, mem
2042def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
2043          (IMUL16rm GR16:$src1, addr:$src2)>;
2044def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
2045          (IMUL32rm GR32:$src1, addr:$src2)>;
2046def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
2047          (IMUL64rm GR64:$src1, addr:$src2)>;
2048
2049// mul reg, imm
2050def : Pat<(mul GR16:$src1, imm:$src2),
2051          (IMUL16rri GR16:$src1, imm:$src2)>;
2052def : Pat<(mul GR32:$src1, imm:$src2),
2053          (IMUL32rri GR32:$src1, imm:$src2)>;
2054def : Pat<(mul GR16:$src1, i16immSExt8:$src2),
2055          (IMUL16rri8 GR16:$src1, i16immSExt8:$src2)>;
2056def : Pat<(mul GR32:$src1, i32immSExt8:$src2),
2057          (IMUL32rri8 GR32:$src1, i32immSExt8:$src2)>;
2058def : Pat<(mul GR64:$src1, i64immSExt8:$src2),
2059          (IMUL64rri8 GR64:$src1, i64immSExt8:$src2)>;
2060def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
2061          (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
2062
2063// reg = mul mem, imm
2064def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
2065          (IMUL16rmi addr:$src1, imm:$src2)>;
2066def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
2067          (IMUL32rmi addr:$src1, imm:$src2)>;
2068def : Pat<(mul (loadi16 addr:$src1), i16immSExt8:$src2),
2069          (IMUL16rmi8 addr:$src1, i16immSExt8:$src2)>;
2070def : Pat<(mul (loadi32 addr:$src1), i32immSExt8:$src2),
2071          (IMUL32rmi8 addr:$src1, i32immSExt8:$src2)>;
2072def : Pat<(mul (loadi64 addr:$src1), i64immSExt8:$src2),
2073          (IMUL64rmi8 addr:$src1, i64immSExt8:$src2)>;
2074def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
2075          (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
2076
2077// Increment/Decrement reg.
2078// Do not make INC/DEC if it is slow
2079let Predicates = [UseIncDec] in {
2080  def : Pat<(add GR8:$src, 1),   (INC8r GR8:$src)>;
2081  def : Pat<(add GR16:$src, 1),  (INC16r GR16:$src)>;
2082  def : Pat<(add GR32:$src, 1),  (INC32r GR32:$src)>;
2083  def : Pat<(add GR64:$src, 1),  (INC64r GR64:$src)>;
2084  def : Pat<(add GR8:$src, -1),  (DEC8r GR8:$src)>;
2085  def : Pat<(add GR16:$src, -1), (DEC16r GR16:$src)>;
2086  def : Pat<(add GR32:$src, -1), (DEC32r GR32:$src)>;
2087  def : Pat<(add GR64:$src, -1), (DEC64r GR64:$src)>;
2088
2089  def : Pat<(X86add_flag_nocf GR8:$src, -1),  (DEC8r GR8:$src)>;
2090  def : Pat<(X86add_flag_nocf GR16:$src, -1), (DEC16r GR16:$src)>;
2091  def : Pat<(X86add_flag_nocf GR32:$src, -1), (DEC32r GR32:$src)>;
2092  def : Pat<(X86add_flag_nocf GR64:$src, -1), (DEC64r GR64:$src)>;
2093  def : Pat<(X86sub_flag_nocf GR8:$src, -1),  (INC8r GR8:$src)>;
2094  def : Pat<(X86sub_flag_nocf GR16:$src, -1), (INC16r GR16:$src)>;
2095  def : Pat<(X86sub_flag_nocf GR32:$src, -1), (INC32r GR32:$src)>;
2096  def : Pat<(X86sub_flag_nocf GR64:$src, -1), (INC64r GR64:$src)>;
2097}
2098
2099// or reg/reg.
2100def : Pat<(or GR8 :$src1, GR8 :$src2), (OR8rr  GR8 :$src1, GR8 :$src2)>;
2101def : Pat<(or GR16:$src1, GR16:$src2), (OR16rr GR16:$src1, GR16:$src2)>;
2102def : Pat<(or GR32:$src1, GR32:$src2), (OR32rr GR32:$src1, GR32:$src2)>;
2103def : Pat<(or GR64:$src1, GR64:$src2), (OR64rr GR64:$src1, GR64:$src2)>;
2104
2105// or reg/mem
2106def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
2107          (OR8rm GR8:$src1, addr:$src2)>;
2108def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
2109          (OR16rm GR16:$src1, addr:$src2)>;
2110def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
2111          (OR32rm GR32:$src1, addr:$src2)>;
2112def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
2113          (OR64rm GR64:$src1, addr:$src2)>;
2114
2115// or reg/imm
2116def : Pat<(or GR8:$src1 , imm:$src2), (OR8ri  GR8 :$src1, imm:$src2)>;
2117def : Pat<(or GR16:$src1, imm:$src2), (OR16ri GR16:$src1, imm:$src2)>;
2118def : Pat<(or GR32:$src1, imm:$src2), (OR32ri GR32:$src1, imm:$src2)>;
2119def : Pat<(or GR16:$src1, i16immSExt8:$src2),
2120          (OR16ri8 GR16:$src1, i16immSExt8:$src2)>;
2121def : Pat<(or GR32:$src1, i32immSExt8:$src2),
2122          (OR32ri8 GR32:$src1, i32immSExt8:$src2)>;
2123def : Pat<(or GR64:$src1, i64immSExt8:$src2),
2124          (OR64ri8 GR64:$src1, i64immSExt8:$src2)>;
2125def : Pat<(or GR64:$src1, i64immSExt32:$src2),
2126          (OR64ri32 GR64:$src1, i64immSExt32:$src2)>;
2127
2128// xor reg/reg
2129def : Pat<(xor GR8 :$src1, GR8 :$src2), (XOR8rr  GR8 :$src1, GR8 :$src2)>;
2130def : Pat<(xor GR16:$src1, GR16:$src2), (XOR16rr GR16:$src1, GR16:$src2)>;
2131def : Pat<(xor GR32:$src1, GR32:$src2), (XOR32rr GR32:$src1, GR32:$src2)>;
2132def : Pat<(xor GR64:$src1, GR64:$src2), (XOR64rr GR64:$src1, GR64:$src2)>;
2133
2134// xor reg/mem
2135def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
2136          (XOR8rm GR8:$src1, addr:$src2)>;
2137def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
2138          (XOR16rm GR16:$src1, addr:$src2)>;
2139def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
2140          (XOR32rm GR32:$src1, addr:$src2)>;
2141def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
2142          (XOR64rm GR64:$src1, addr:$src2)>;
2143
2144// xor reg/imm
2145def : Pat<(xor GR8:$src1, imm:$src2),
2146          (XOR8ri GR8:$src1, imm:$src2)>;
2147def : Pat<(xor GR16:$src1, imm:$src2),
2148          (XOR16ri GR16:$src1, imm:$src2)>;
2149def : Pat<(xor GR32:$src1, imm:$src2),
2150          (XOR32ri GR32:$src1, imm:$src2)>;
2151def : Pat<(xor GR16:$src1, i16immSExt8:$src2),
2152          (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
2153def : Pat<(xor GR32:$src1, i32immSExt8:$src2),
2154          (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
2155def : Pat<(xor GR64:$src1, i64immSExt8:$src2),
2156          (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
2157def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
2158          (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
2159
2160// and reg/reg
2161def : Pat<(and GR8 :$src1, GR8 :$src2), (AND8rr  GR8 :$src1, GR8 :$src2)>;
2162def : Pat<(and GR16:$src1, GR16:$src2), (AND16rr GR16:$src1, GR16:$src2)>;
2163def : Pat<(and GR32:$src1, GR32:$src2), (AND32rr GR32:$src1, GR32:$src2)>;
2164def : Pat<(and GR64:$src1, GR64:$src2), (AND64rr GR64:$src1, GR64:$src2)>;
2165
2166// and reg/mem
2167def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
2168          (AND8rm GR8:$src1, addr:$src2)>;
2169def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
2170          (AND16rm GR16:$src1, addr:$src2)>;
2171def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
2172          (AND32rm GR32:$src1, addr:$src2)>;
2173def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
2174          (AND64rm GR64:$src1, addr:$src2)>;
2175
2176// and reg/imm
2177def : Pat<(and GR8:$src1, imm:$src2),
2178          (AND8ri GR8:$src1, imm:$src2)>;
2179def : Pat<(and GR16:$src1, imm:$src2),
2180          (AND16ri GR16:$src1, imm:$src2)>;
2181def : Pat<(and GR32:$src1, imm:$src2),
2182          (AND32ri GR32:$src1, imm:$src2)>;
2183def : Pat<(and GR16:$src1, i16immSExt8:$src2),
2184          (AND16ri8 GR16:$src1, i16immSExt8:$src2)>;
2185def : Pat<(and GR32:$src1, i32immSExt8:$src2),
2186          (AND32ri8 GR32:$src1, i32immSExt8:$src2)>;
2187def : Pat<(and GR64:$src1, i64immSExt8:$src2),
2188          (AND64ri8 GR64:$src1, i64immSExt8:$src2)>;
2189def : Pat<(and GR64:$src1, i64immSExt32:$src2),
2190          (AND64ri32 GR64:$src1, i64immSExt32:$src2)>;
2191
2192// Bit scan instruction patterns to match explicit zero-undef behavior.
2193def : Pat<(cttz_zero_undef GR16:$src), (BSF16rr GR16:$src)>;
2194def : Pat<(cttz_zero_undef GR32:$src), (BSF32rr GR32:$src)>;
2195def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr GR64:$src)>;
2196def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm addr:$src)>;
2197def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm addr:$src)>;
2198def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm addr:$src)>;
2199
2200// When HasMOVBE is enabled it is possible to get a non-legalized
2201// register-register 16 bit bswap. This maps it to a ROL instruction.
2202let Predicates = [HasMOVBE] in {
2203 def : Pat<(bswap GR16:$src), (ROL16ri GR16:$src, (i8 8))>;
2204}
2205