1//===- X86InstrCompiler.td - Compiler Pseudos and Patterns -*- tablegen -*-===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file describes the various pseudo instructions used by the compiler,
11// as well as Pat patterns used during instruction selection.
12//
13//===----------------------------------------------------------------------===//
14
15//===----------------------------------------------------------------------===//
16// Pattern Matching Support
17
18def GetLo32XForm : SDNodeXForm<imm, [{
19  // Transformation function: get the low 32 bits.
20  return getI32Imm((unsigned)N->getZExtValue(), SDLoc(N));
21}]>;
22
23def GetLo8XForm : SDNodeXForm<imm, [{
24  // Transformation function: get the low 8 bits.
25  return getI8Imm((uint8_t)N->getZExtValue(), SDLoc(N));
26}]>;
27
28
29//===----------------------------------------------------------------------===//
30// Random Pseudo Instructions.
31
32// PIC base construction.  This expands to code that looks like this:
33//     call  $next_inst
34//     popl %destreg"
35let hasSideEffects = 0, isNotDuplicable = 1, Uses = [ESP] in
36  def MOVPC32r : Ii32<0xE8, Pseudo, (outs GR32:$reg), (ins i32imm:$label),
37                      "", []>;
38
39
40// ADJCALLSTACKDOWN/UP implicitly use/def ESP because they may be expanded into
41// a stack adjustment and the codegen must know that they may modify the stack
42// pointer before prolog-epilog rewriting occurs.
43// Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
44// sub / add which can clobber EFLAGS.
45let Defs = [ESP, EFLAGS], Uses = [ESP] in {
46def ADJCALLSTACKDOWN32 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
47                           "#ADJCALLSTACKDOWN",
48                           []>,
49                          Requires<[NotLP64]>;
50def ADJCALLSTACKUP32   : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
51                           "#ADJCALLSTACKUP",
52                           [(X86callseq_end timm:$amt1, timm:$amt2)]>,
53                          Requires<[NotLP64]>;
54}
55def : Pat<(X86callseq_start timm:$amt1),
56          (ADJCALLSTACKDOWN32 i32imm:$amt1, 0)>, Requires<[NotLP64]>;
57
58
59// ADJCALLSTACKDOWN/UP implicitly use/def RSP because they may be expanded into
60// a stack adjustment and the codegen must know that they may modify the stack
61// pointer before prolog-epilog rewriting occurs.
62// Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
63// sub / add which can clobber EFLAGS.
64let Defs = [RSP, EFLAGS], Uses = [RSP] in {
65def ADJCALLSTACKDOWN64 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
66                           "#ADJCALLSTACKDOWN",
67                           []>,
68                          Requires<[IsLP64]>;
69def ADJCALLSTACKUP64   : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
70                           "#ADJCALLSTACKUP",
71                           [(X86callseq_end timm:$amt1, timm:$amt2)]>,
72                          Requires<[IsLP64]>;
73}
74def : Pat<(X86callseq_start timm:$amt1),
75          (ADJCALLSTACKDOWN64 i32imm:$amt1, 0)>, Requires<[IsLP64]>;
76
77
78// x86-64 va_start lowering magic.
79let usesCustomInserter = 1, Defs = [EFLAGS] in {
80def VASTART_SAVE_XMM_REGS : I<0, Pseudo,
81                              (outs),
82                              (ins GR8:$al,
83                                   i64imm:$regsavefi, i64imm:$offset,
84                                   variable_ops),
85                              "#VASTART_SAVE_XMM_REGS $al, $regsavefi, $offset",
86                              [(X86vastart_save_xmm_regs GR8:$al,
87                                                         imm:$regsavefi,
88                                                         imm:$offset),
89                               (implicit EFLAGS)]>;
90
91// The VAARG_64 pseudo-instruction takes the address of the va_list,
92// and places the address of the next argument into a register.
93let Defs = [EFLAGS] in
94def VAARG_64 : I<0, Pseudo,
95                 (outs GR64:$dst),
96                 (ins i8mem:$ap, i32imm:$size, i8imm:$mode, i32imm:$align),
97                 "#VAARG_64 $dst, $ap, $size, $mode, $align",
98                 [(set GR64:$dst,
99                    (X86vaarg64 addr:$ap, imm:$size, imm:$mode, imm:$align)),
100                  (implicit EFLAGS)]>;
101
102// Dynamic stack allocation yields a _chkstk or _alloca call for all Windows
103// targets.  These calls are needed to probe the stack when allocating more than
104// 4k bytes in one go. Touching the stack at 4K increments is necessary to
105// ensure that the guard pages used by the OS virtual memory manager are
106// allocated in correct sequence.
107// The main point of having separate instruction are extra unmodelled effects
108// (compared to ordinary calls) like stack pointer change.
109
110let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
111  def WIN_ALLOCA : I<0, Pseudo, (outs), (ins),
112                     "# dynamic stack allocation",
113                     [(X86WinAlloca)]>;
114
115// When using segmented stacks these are lowered into instructions which first
116// check if the current stacklet has enough free memory. If it does, memory is
117// allocated by bumping the stack pointer. Otherwise memory is allocated from
118// the heap.
119
120let Defs = [EAX, ESP, EFLAGS], Uses = [ESP] in
121def SEG_ALLOCA_32 : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$size),
122                      "# variable sized alloca for segmented stacks",
123                      [(set GR32:$dst,
124                         (X86SegAlloca GR32:$size))]>,
125                    Requires<[NotLP64]>;
126
127let Defs = [RAX, RSP, EFLAGS], Uses = [RSP] in
128def SEG_ALLOCA_64 : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$size),
129                      "# variable sized alloca for segmented stacks",
130                      [(set GR64:$dst,
131                         (X86SegAlloca GR64:$size))]>,
132                    Requires<[In64BitMode]>;
133}
134
135//===----------------------------------------------------------------------===//
136// EH Pseudo Instructions
137//
138let SchedRW = [WriteSystem] in {
139let isTerminator = 1, isReturn = 1, isBarrier = 1,
140    hasCtrlDep = 1, isCodeGenOnly = 1 in {
141def EH_RETURN   : I<0xC3, RawFrm, (outs), (ins GR32:$addr),
142                    "ret\t#eh_return, addr: $addr",
143                    [(X86ehret GR32:$addr)], IIC_RET>, Sched<[WriteJumpLd]>;
144
145}
146
147let isTerminator = 1, isReturn = 1, isBarrier = 1,
148    hasCtrlDep = 1, isCodeGenOnly = 1 in {
149def EH_RETURN64   : I<0xC3, RawFrm, (outs), (ins GR64:$addr),
150                     "ret\t#eh_return, addr: $addr",
151                     [(X86ehret GR64:$addr)], IIC_RET>, Sched<[WriteJumpLd]>;
152
153}
154
155let isTerminator = 1, hasSideEffects = 1, isBarrier = 1, hasCtrlDep = 1,
156    isCodeGenOnly = 1, isReturn = 1 in {
157  def CLEANUPRET : I<0, Pseudo, (outs), (ins), "# CLEANUPRET", [(cleanupret)]>;
158
159  // CATCHRET needs a custom inserter for SEH.
160  let usesCustomInserter = 1 in
161    def CATCHRET : I<0, Pseudo, (outs), (ins brtarget32:$dst, brtarget32:$from),
162                     "# CATCHRET",
163                     [(catchret bb:$dst, bb:$from)]>;
164}
165
166let hasSideEffects = 1, hasCtrlDep = 1, isCodeGenOnly = 1,
167    usesCustomInserter = 1 in
168def CATCHPAD : I<0, Pseudo, (outs), (ins), "# CATCHPAD", [(catchpad)]>;
169
170// This instruction is responsible for re-establishing stack pointers after an
171// exception has been caught and we are rejoining normal control flow in the
172// parent function or funclet. It generally sets ESP and EBP, and optionally
173// ESI. It is only needed for 32-bit WinEH, as the runtime restores CSRs for us
174// elsewhere.
175let hasSideEffects = 1, hasCtrlDep = 1, isCodeGenOnly = 1 in
176def EH_RESTORE : I<0, Pseudo, (outs), (ins), "# EH_RESTORE", []>;
177
178let hasSideEffects = 1, isBarrier = 1, isCodeGenOnly = 1,
179    usesCustomInserter = 1 in {
180  def EH_SjLj_SetJmp32  : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$buf),
181                            "#EH_SJLJ_SETJMP32",
182                            [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
183                          Requires<[Not64BitMode]>;
184  def EH_SjLj_SetJmp64  : I<0, Pseudo, (outs GR32:$dst), (ins i64mem:$buf),
185                            "#EH_SJLJ_SETJMP64",
186                            [(set GR32:$dst, (X86eh_sjlj_setjmp addr:$buf))]>,
187                          Requires<[In64BitMode]>;
188  let isTerminator = 1 in {
189  def EH_SjLj_LongJmp32 : I<0, Pseudo, (outs), (ins i32mem:$buf),
190                            "#EH_SJLJ_LONGJMP32",
191                            [(X86eh_sjlj_longjmp addr:$buf)]>,
192                          Requires<[Not64BitMode]>;
193  def EH_SjLj_LongJmp64 : I<0, Pseudo, (outs), (ins i64mem:$buf),
194                            "#EH_SJLJ_LONGJMP64",
195                            [(X86eh_sjlj_longjmp addr:$buf)]>,
196                          Requires<[In64BitMode]>;
197  }
198}
199} // SchedRW
200
201let isBranch = 1, isTerminator = 1, isCodeGenOnly = 1 in {
202  def EH_SjLj_Setup : I<0, Pseudo, (outs), (ins brtarget:$dst),
203                        "#EH_SjLj_Setup\t$dst", []>;
204}
205
206//===----------------------------------------------------------------------===//
207// Pseudo instructions used by unwind info.
208//
209let isPseudo = 1 in {
210  def SEH_PushReg : I<0, Pseudo, (outs), (ins i32imm:$reg),
211                            "#SEH_PushReg $reg", []>;
212  def SEH_SaveReg : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
213                            "#SEH_SaveReg $reg, $dst", []>;
214  def SEH_SaveXMM : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$dst),
215                            "#SEH_SaveXMM $reg, $dst", []>;
216  def SEH_StackAlloc : I<0, Pseudo, (outs), (ins i32imm:$size),
217                            "#SEH_StackAlloc $size", []>;
218  def SEH_SetFrame : I<0, Pseudo, (outs), (ins i32imm:$reg, i32imm:$offset),
219                            "#SEH_SetFrame $reg, $offset", []>;
220  def SEH_PushFrame : I<0, Pseudo, (outs), (ins i1imm:$mode),
221                            "#SEH_PushFrame $mode", []>;
222  def SEH_EndPrologue : I<0, Pseudo, (outs), (ins),
223                            "#SEH_EndPrologue", []>;
224  def SEH_Epilogue : I<0, Pseudo, (outs), (ins),
225                            "#SEH_Epilogue", []>;
226}
227
228//===----------------------------------------------------------------------===//
229// Pseudo instructions used by segmented stacks.
230//
231
232// This is lowered into a RET instruction by MCInstLower.  We need
233// this so that we don't have to have a MachineBasicBlock which ends
234// with a RET and also has successors.
235let isPseudo = 1 in {
236def MORESTACK_RET: I<0, Pseudo, (outs), (ins),
237                          "", []>;
238
239// This instruction is lowered to a RET followed by a MOV.  The two
240// instructions are not generated on a higher level since then the
241// verifier sees a MachineBasicBlock ending with a non-terminator.
242def MORESTACK_RET_RESTORE_R10 : I<0, Pseudo, (outs), (ins),
243                                  "", []>;
244}
245
246//===----------------------------------------------------------------------===//
247// Alias Instructions
248//===----------------------------------------------------------------------===//
249
250// Alias instruction mapping movr0 to xor.
251// FIXME: remove when we can teach regalloc that xor reg, reg is ok.
252let Defs = [EFLAGS], isReMaterializable = 1, isAsCheapAsAMove = 1,
253    isPseudo = 1, AddedComplexity = 20 in
254def MOV32r0  : I<0, Pseudo, (outs GR32:$dst), (ins), "",
255                 [(set GR32:$dst, 0)], IIC_ALU_NONMEM>, Sched<[WriteZero]>;
256
257// Other widths can also make use of the 32-bit xor, which may have a smaller
258// encoding and avoid partial register updates.
259def : Pat<(i8 0), (EXTRACT_SUBREG (MOV32r0), sub_8bit)>;
260def : Pat<(i16 0), (EXTRACT_SUBREG (MOV32r0), sub_16bit)>;
261def : Pat<(i64 0), (SUBREG_TO_REG (i64 0), (MOV32r0), sub_32bit)> {
262  let AddedComplexity = 20;
263}
264
265let Predicates = [OptForSize, NotSlowIncDec, Not64BitMode],
266    AddedComplexity = 15 in {
267  // Pseudo instructions for materializing 1 and -1 using XOR+INC/DEC,
268  // which only require 3 bytes compared to MOV32ri which requires 5.
269  let Defs = [EFLAGS], isReMaterializable = 1, isPseudo = 1 in {
270    def MOV32r1 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
271                        [(set GR32:$dst, 1)]>;
272    def MOV32r_1 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
273                        [(set GR32:$dst, -1)]>;
274  }
275
276  // MOV16ri is 4 bytes, so the instructions above are smaller.
277  def : Pat<(i16 1), (EXTRACT_SUBREG (MOV32r1), sub_16bit)>;
278  def : Pat<(i16 -1), (EXTRACT_SUBREG (MOV32r_1), sub_16bit)>;
279}
280
281let isReMaterializable = 1, isPseudo = 1, AddedComplexity = 10 in {
282// AddedComplexity higher than MOV64ri but lower than MOV32r0 and MOV32r1.
283// FIXME: Add itinerary class and Schedule.
284def MOV32ImmSExti8 : I<0, Pseudo, (outs GR32:$dst), (ins i32i8imm:$src), "",
285                       [(set GR32:$dst, i32immSExt8:$src)]>,
286                     Requires<[OptForMinSize]>;
287def MOV64ImmSExti8 : I<0, Pseudo, (outs GR64:$dst), (ins i64i8imm:$src), "",
288                       [(set GR64:$dst, i64immSExt8:$src)]>,
289                     Requires<[OptForMinSize, NotWin64WithoutFP]>;
290}
291
292// Materialize i64 constant where top 32-bits are zero. This could theoretically
293// use MOV32ri with a SUBREG_TO_REG to represent the zero-extension, however
294// that would make it more difficult to rematerialize.
295let AddedComplexity = 1, isReMaterializable = 1, isAsCheapAsAMove = 1,
296    isCodeGenOnly = 1, hasSideEffects = 0 in
297def MOV32ri64 : Ii32<0xb8, AddRegFrm, (outs GR32:$dst), (ins i64i32imm:$src),
298                     "", [], IIC_ALU_NONMEM>, Sched<[WriteALU]>;
299
300// This 64-bit pseudo-move can be used for both a 64-bit constant that is
301// actually the zero-extension of a 32-bit constant and for labels in the
302// x86-64 small code model.
303def mov64imm32 : ComplexPattern<i64, 1, "selectMOV64Imm32", [imm, X86Wrapper]>;
304
305let AddedComplexity = 1 in
306def : Pat<(i64 mov64imm32:$src),
307          (SUBREG_TO_REG (i64 0), (MOV32ri64 mov64imm32:$src), sub_32bit)>;
308
309// Use sbb to materialize carry bit.
310let Uses = [EFLAGS], Defs = [EFLAGS], isPseudo = 1, SchedRW = [WriteALU] in {
311// FIXME: These are pseudo ops that should be replaced with Pat<> patterns.
312// However, Pat<> can't replicate the destination reg into the inputs of the
313// result.
314def SETB_C8r : I<0, Pseudo, (outs GR8:$dst), (ins), "",
315                 [(set GR8:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
316def SETB_C16r : I<0, Pseudo, (outs GR16:$dst), (ins), "",
317                 [(set GR16:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
318def SETB_C32r : I<0, Pseudo, (outs GR32:$dst), (ins), "",
319                 [(set GR32:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
320def SETB_C64r : I<0, Pseudo, (outs GR64:$dst), (ins), "",
321                 [(set GR64:$dst, (X86setcc_c X86_COND_B, EFLAGS))]>;
322} // isCodeGenOnly
323
324
325def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
326          (SETB_C16r)>;
327def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
328          (SETB_C32r)>;
329def : Pat<(i64 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
330          (SETB_C64r)>;
331
332def : Pat<(i16 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
333          (SETB_C16r)>;
334def : Pat<(i32 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
335          (SETB_C32r)>;
336def : Pat<(i64 (sext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
337          (SETB_C64r)>;
338
339// We canonicalize 'setb' to "(and (sbb reg,reg), 1)" on the hope that the and
340// will be eliminated and that the sbb can be extended up to a wider type.  When
341// this happens, it is great.  However, if we are left with an 8-bit sbb and an
342// and, we might as well just match it as a setb.
343def : Pat<(and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1),
344          (SETBr)>;
345
346// (add OP, SETB) -> (adc OP, 0)
347def : Pat<(add (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR8:$op),
348          (ADC8ri GR8:$op, 0)>;
349def : Pat<(add (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR32:$op),
350          (ADC32ri8 GR32:$op, 0)>;
351def : Pat<(add (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1), GR64:$op),
352          (ADC64ri8 GR64:$op, 0)>;
353
354// (sub OP, SETB) -> (sbb OP, 0)
355def : Pat<(sub GR8:$op, (and (i8 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
356          (SBB8ri GR8:$op, 0)>;
357def : Pat<(sub GR32:$op, (and (i32 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
358          (SBB32ri8 GR32:$op, 0)>;
359def : Pat<(sub GR64:$op, (and (i64 (X86setcc_c X86_COND_B, EFLAGS)), 1)),
360          (SBB64ri8 GR64:$op, 0)>;
361
362// (sub OP, SETCC_CARRY) -> (adc OP, 0)
363def : Pat<(sub GR8:$op, (i8 (X86setcc_c X86_COND_B, EFLAGS))),
364          (ADC8ri GR8:$op, 0)>;
365def : Pat<(sub GR32:$op, (i32 (X86setcc_c X86_COND_B, EFLAGS))),
366          (ADC32ri8 GR32:$op, 0)>;
367def : Pat<(sub GR64:$op, (i64 (X86setcc_c X86_COND_B, EFLAGS))),
368          (ADC64ri8 GR64:$op, 0)>;
369
370//===----------------------------------------------------------------------===//
371// String Pseudo Instructions
372//
373let SchedRW = [WriteMicrocoded] in {
374let Defs = [ECX,EDI,ESI], Uses = [ECX,EDI,ESI], isCodeGenOnly = 1 in {
375def REP_MOVSB_32 : I<0xA4, RawFrm, (outs), (ins), "{rep;movsb|rep movsb}",
376                    [(X86rep_movs i8)], IIC_REP_MOVS>, REP,
377                   Requires<[Not64BitMode]>;
378def REP_MOVSW_32 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsw|rep movsw}",
379                    [(X86rep_movs i16)], IIC_REP_MOVS>, REP, OpSize16,
380                   Requires<[Not64BitMode]>;
381def REP_MOVSD_32 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsl|rep movsd}",
382                    [(X86rep_movs i32)], IIC_REP_MOVS>, REP, OpSize32,
383                   Requires<[Not64BitMode]>;
384}
385
386let Defs = [RCX,RDI,RSI], Uses = [RCX,RDI,RSI], isCodeGenOnly = 1 in {
387def REP_MOVSB_64 : I<0xA4, RawFrm, (outs), (ins), "{rep;movsb|rep movsb}",
388                    [(X86rep_movs i8)], IIC_REP_MOVS>, REP,
389                   Requires<[In64BitMode]>;
390def REP_MOVSW_64 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsw|rep movsw}",
391                    [(X86rep_movs i16)], IIC_REP_MOVS>, REP, OpSize16,
392                   Requires<[In64BitMode]>;
393def REP_MOVSD_64 : I<0xA5, RawFrm, (outs), (ins), "{rep;movsl|rep movsd}",
394                    [(X86rep_movs i32)], IIC_REP_MOVS>, REP, OpSize32,
395                   Requires<[In64BitMode]>;
396def REP_MOVSQ_64 : RI<0xA5, RawFrm, (outs), (ins), "{rep;movsq|rep movsq}",
397                    [(X86rep_movs i64)], IIC_REP_MOVS>, REP,
398                   Requires<[In64BitMode]>;
399}
400
401// FIXME: Should use "(X86rep_stos AL)" as the pattern.
402let Defs = [ECX,EDI], isCodeGenOnly = 1 in {
403  let Uses = [AL,ECX,EDI] in
404  def REP_STOSB_32 : I<0xAA, RawFrm, (outs), (ins), "{rep;stosb|rep stosb}",
405                      [(X86rep_stos i8)], IIC_REP_STOS>, REP,
406                     Requires<[Not64BitMode]>;
407  let Uses = [AX,ECX,EDI] in
408  def REP_STOSW_32 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosw|rep stosw}",
409                      [(X86rep_stos i16)], IIC_REP_STOS>, REP, OpSize16,
410                     Requires<[Not64BitMode]>;
411  let Uses = [EAX,ECX,EDI] in
412  def REP_STOSD_32 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosl|rep stosd}",
413                      [(X86rep_stos i32)], IIC_REP_STOS>, REP, OpSize32,
414                     Requires<[Not64BitMode]>;
415}
416
417let Defs = [RCX,RDI], isCodeGenOnly = 1 in {
418  let Uses = [AL,RCX,RDI] in
419  def REP_STOSB_64 : I<0xAA, RawFrm, (outs), (ins), "{rep;stosb|rep stosb}",
420                      [(X86rep_stos i8)], IIC_REP_STOS>, REP,
421                     Requires<[In64BitMode]>;
422  let Uses = [AX,RCX,RDI] in
423  def REP_STOSW_64 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosw|rep stosw}",
424                      [(X86rep_stos i16)], IIC_REP_STOS>, REP, OpSize16,
425                     Requires<[In64BitMode]>;
426  let Uses = [RAX,RCX,RDI] in
427  def REP_STOSD_64 : I<0xAB, RawFrm, (outs), (ins), "{rep;stosl|rep stosd}",
428                      [(X86rep_stos i32)], IIC_REP_STOS>, REP, OpSize32,
429                     Requires<[In64BitMode]>;
430
431  let Uses = [RAX,RCX,RDI] in
432  def REP_STOSQ_64 : RI<0xAB, RawFrm, (outs), (ins), "{rep;stosq|rep stosq}",
433                      [(X86rep_stos i64)], IIC_REP_STOS>, REP,
434                     Requires<[In64BitMode]>;
435}
436} // SchedRW
437
438//===----------------------------------------------------------------------===//
439// Thread Local Storage Instructions
440//
441
442// ELF TLS Support
443// All calls clobber the non-callee saved registers. ESP is marked as
444// a use to prevent stack-pointer assignments that appear immediately
445// before calls from potentially appearing dead.
446let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, FP7,
447            ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7,
448            MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
449            XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
450            XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
451    Uses = [ESP] in {
452def TLS_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
453                  "# TLS_addr32",
454                  [(X86tlsaddr tls32addr:$sym)]>,
455                  Requires<[Not64BitMode]>;
456def TLS_base_addr32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
457                  "# TLS_base_addr32",
458                  [(X86tlsbaseaddr tls32baseaddr:$sym)]>,
459                  Requires<[Not64BitMode]>;
460}
461
462// All calls clobber the non-callee saved registers. RSP is marked as
463// a use to prevent stack-pointer assignments that appear immediately
464// before calls from potentially appearing dead.
465let Defs = [RAX, RCX, RDX, RSI, RDI, R8, R9, R10, R11,
466            FP0, FP1, FP2, FP3, FP4, FP5, FP6, FP7,
467            ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7,
468            MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7,
469            XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
470            XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15, EFLAGS],
471    Uses = [RSP] in {
472def TLS_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
473                   "# TLS_addr64",
474                  [(X86tlsaddr tls64addr:$sym)]>,
475                  Requires<[In64BitMode]>;
476def TLS_base_addr64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
477                   "# TLS_base_addr64",
478                  [(X86tlsbaseaddr tls64baseaddr:$sym)]>,
479                  Requires<[In64BitMode]>;
480}
481
482// Darwin TLS Support
483// For i386, the address of the thunk is passed on the stack, on return the
484// address of the variable is in %eax.  %ecx is trashed during the function
485// call.  All other registers are preserved.
486let Defs = [EAX, ECX, EFLAGS],
487    Uses = [ESP],
488    usesCustomInserter = 1 in
489def TLSCall_32 : I<0, Pseudo, (outs), (ins i32mem:$sym),
490                "# TLSCall_32",
491                [(X86TLSCall addr:$sym)]>,
492                Requires<[Not64BitMode]>;
493
494// For x86_64, the address of the thunk is passed in %rdi, on return
495// the address of the variable is in %rax.  All other registers are preserved.
496let Defs = [RAX, EFLAGS],
497    Uses = [RSP, RDI],
498    usesCustomInserter = 1 in
499def TLSCall_64 : I<0, Pseudo, (outs), (ins i64mem:$sym),
500                  "# TLSCall_64",
501                  [(X86TLSCall addr:$sym)]>,
502                  Requires<[In64BitMode]>;
503
504
505//===----------------------------------------------------------------------===//
506// Conditional Move Pseudo Instructions
507
508// CMOV* - Used to implement the SELECT DAG operation.  Expanded after
509// instruction selection into a branch sequence.
510multiclass CMOVrr_PSEUDO<RegisterClass RC, ValueType VT> {
511  def CMOV#NAME  : I<0, Pseudo,
512                    (outs RC:$dst), (ins RC:$t, RC:$f, i8imm:$cond),
513                    "#CMOV_"#NAME#" PSEUDO!",
514                    [(set RC:$dst, (VT (X86cmov RC:$t, RC:$f, imm:$cond,
515                                                EFLAGS)))]>;
516}
517
518let usesCustomInserter = 1, Uses = [EFLAGS] in {
519  // X86 doesn't have 8-bit conditional moves. Use a customInserter to
520  // emit control flow. An alternative to this is to mark i8 SELECT as Promote,
521  // however that requires promoting the operands, and can induce additional
522  // i8 register pressure.
523  defm _GR8 : CMOVrr_PSEUDO<GR8, i8>;
524
525  let Predicates = [NoCMov] in {
526    defm _GR32 : CMOVrr_PSEUDO<GR32, i32>;
527    defm _GR16 : CMOVrr_PSEUDO<GR16, i16>;
528  } // Predicates = [NoCMov]
529
530  // fcmov doesn't handle all possible EFLAGS, provide a fallback if there is no
531  // SSE1/SSE2.
532  let Predicates = [FPStackf32] in
533    defm _RFP32 : CMOVrr_PSEUDO<RFP32, f32>;
534
535  let Predicates = [FPStackf64] in
536    defm _RFP64 : CMOVrr_PSEUDO<RFP64, f64>;
537
538  defm _RFP80 : CMOVrr_PSEUDO<RFP80, f80>;
539
540  defm _FR32   : CMOVrr_PSEUDO<FR32, f32>;
541  defm _FR64   : CMOVrr_PSEUDO<FR64, f64>;
542  defm _FR128  : CMOVrr_PSEUDO<FR128, f128>;
543  defm _V4F32  : CMOVrr_PSEUDO<VR128, v4f32>;
544  defm _V2F64  : CMOVrr_PSEUDO<VR128, v2f64>;
545  defm _V2I64  : CMOVrr_PSEUDO<VR128, v2i64>;
546  defm _V8F32  : CMOVrr_PSEUDO<VR256, v8f32>;
547  defm _V4F64  : CMOVrr_PSEUDO<VR256, v4f64>;
548  defm _V4I64  : CMOVrr_PSEUDO<VR256, v4i64>;
549  defm _V8I64  : CMOVrr_PSEUDO<VR512, v8i64>;
550  defm _V8F64  : CMOVrr_PSEUDO<VR512, v8f64>;
551  defm _V16F32 : CMOVrr_PSEUDO<VR512, v16f32>;
552  defm _V8I1   : CMOVrr_PSEUDO<VK8,  v8i1>;
553  defm _V16I1  : CMOVrr_PSEUDO<VK16, v16i1>;
554  defm _V32I1  : CMOVrr_PSEUDO<VK32, v32i1>;
555  defm _V64I1  : CMOVrr_PSEUDO<VK64, v64i1>;
556} // usesCustomInserter = 1, Uses = [EFLAGS]
557
558//===----------------------------------------------------------------------===//
559// Normal-Instructions-With-Lock-Prefix Pseudo Instructions
560//===----------------------------------------------------------------------===//
561
562// FIXME: Use normal instructions and add lock prefix dynamically.
563
564// Memory barriers
565
566// TODO: Get this to fold the constant into the instruction.
567let isCodeGenOnly = 1, Defs = [EFLAGS] in
568def OR32mrLocked  : I<0x09, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$zero),
569                      "or{l}\t{$zero, $dst|$dst, $zero}",
570                      [], IIC_ALU_MEM>, Requires<[Not64BitMode]>, LOCK,
571                    Sched<[WriteALULd, WriteRMW]>;
572
573let hasSideEffects = 1 in
574def Int_MemBarrier : I<0, Pseudo, (outs), (ins),
575                     "#MEMBARRIER",
576                     [(X86MemBarrier)]>, Sched<[WriteLoad]>;
577
578// RegOpc corresponds to the mr version of the instruction
579// ImmOpc corresponds to the mi version of the instruction
580// ImmOpc8 corresponds to the mi8 version of the instruction
581// ImmMod corresponds to the instruction format of the mi and mi8 versions
582multiclass LOCK_ArithBinOp<bits<8> RegOpc, bits<8> ImmOpc, bits<8> ImmOpc8,
583                           Format ImmMod, string mnemonic> {
584let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
585    SchedRW = [WriteALULd, WriteRMW] in {
586
587def NAME#8mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
588                  RegOpc{3}, RegOpc{2}, RegOpc{1}, 0 },
589                  MRMDestMem, (outs), (ins i8mem:$dst, GR8:$src2),
590                  !strconcat(mnemonic, "{b}\t",
591                             "{$src2, $dst|$dst, $src2}"),
592                  [], IIC_ALU_NONMEM>, LOCK;
593def NAME#16mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
594                   RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
595                   MRMDestMem, (outs), (ins i16mem:$dst, GR16:$src2),
596                   !strconcat(mnemonic, "{w}\t",
597                              "{$src2, $dst|$dst, $src2}"),
598                   [], IIC_ALU_NONMEM>, OpSize16, LOCK;
599def NAME#32mr : I<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
600                   RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
601                   MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src2),
602                   !strconcat(mnemonic, "{l}\t",
603                              "{$src2, $dst|$dst, $src2}"),
604                   [], IIC_ALU_NONMEM>, OpSize32, LOCK;
605def NAME#64mr : RI<{RegOpc{7}, RegOpc{6}, RegOpc{5}, RegOpc{4},
606                    RegOpc{3}, RegOpc{2}, RegOpc{1}, 1 },
607                    MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src2),
608                    !strconcat(mnemonic, "{q}\t",
609                               "{$src2, $dst|$dst, $src2}"),
610                    [], IIC_ALU_NONMEM>, LOCK;
611
612def NAME#8mi : Ii8<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
613                    ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 0 },
614                    ImmMod, (outs), (ins i8mem :$dst, i8imm :$src2),
615                    !strconcat(mnemonic, "{b}\t",
616                               "{$src2, $dst|$dst, $src2}"),
617                    [], IIC_ALU_MEM>, LOCK;
618
619def NAME#16mi : Ii16<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
620                      ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
621                      ImmMod, (outs), (ins i16mem :$dst, i16imm :$src2),
622                      !strconcat(mnemonic, "{w}\t",
623                                 "{$src2, $dst|$dst, $src2}"),
624                      [], IIC_ALU_MEM>, OpSize16, LOCK;
625
626def NAME#32mi : Ii32<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
627                      ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
628                      ImmMod, (outs), (ins i32mem :$dst, i32imm :$src2),
629                      !strconcat(mnemonic, "{l}\t",
630                                 "{$src2, $dst|$dst, $src2}"),
631                      [], IIC_ALU_MEM>, OpSize32, LOCK;
632
633def NAME#64mi32 : RIi32S<{ImmOpc{7}, ImmOpc{6}, ImmOpc{5}, ImmOpc{4},
634                          ImmOpc{3}, ImmOpc{2}, ImmOpc{1}, 1 },
635                          ImmMod, (outs), (ins i64mem :$dst, i64i32imm :$src2),
636                          !strconcat(mnemonic, "{q}\t",
637                                     "{$src2, $dst|$dst, $src2}"),
638                          [], IIC_ALU_MEM>, LOCK;
639
640def NAME#16mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
641                      ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
642                      ImmMod, (outs), (ins i16mem :$dst, i16i8imm :$src2),
643                      !strconcat(mnemonic, "{w}\t",
644                                 "{$src2, $dst|$dst, $src2}"),
645                      [], IIC_ALU_MEM>, OpSize16, LOCK;
646def NAME#32mi8 : Ii8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
647                      ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
648                      ImmMod, (outs), (ins i32mem :$dst, i32i8imm :$src2),
649                      !strconcat(mnemonic, "{l}\t",
650                                 "{$src2, $dst|$dst, $src2}"),
651                      [], IIC_ALU_MEM>, OpSize32, LOCK;
652def NAME#64mi8 : RIi8<{ImmOpc8{7}, ImmOpc8{6}, ImmOpc8{5}, ImmOpc8{4},
653                       ImmOpc8{3}, ImmOpc8{2}, ImmOpc8{1}, 1 },
654                       ImmMod, (outs), (ins i64mem :$dst, i64i8imm :$src2),
655                       !strconcat(mnemonic, "{q}\t",
656                                  "{$src2, $dst|$dst, $src2}"),
657                       [], IIC_ALU_MEM>, LOCK;
658
659}
660
661}
662
663defm LOCK_ADD : LOCK_ArithBinOp<0x00, 0x80, 0x83, MRM0m, "add">;
664defm LOCK_SUB : LOCK_ArithBinOp<0x28, 0x80, 0x83, MRM5m, "sub">;
665defm LOCK_OR  : LOCK_ArithBinOp<0x08, 0x80, 0x83, MRM1m, "or">;
666defm LOCK_AND : LOCK_ArithBinOp<0x20, 0x80, 0x83, MRM4m, "and">;
667defm LOCK_XOR : LOCK_ArithBinOp<0x30, 0x80, 0x83, MRM6m, "xor">;
668
669// Optimized codegen when the non-memory output is not used.
670multiclass LOCK_ArithUnOp<bits<8> Opc8, bits<8> Opc, Format Form,
671                          string mnemonic> {
672let Defs = [EFLAGS], mayLoad = 1, mayStore = 1, isCodeGenOnly = 1,
673    SchedRW = [WriteALULd, WriteRMW] in {
674
675def NAME#8m  : I<Opc8, Form, (outs), (ins i8mem :$dst),
676                 !strconcat(mnemonic, "{b}\t$dst"),
677                 [], IIC_UNARY_MEM>, LOCK;
678def NAME#16m : I<Opc, Form, (outs), (ins i16mem:$dst),
679                 !strconcat(mnemonic, "{w}\t$dst"),
680                 [], IIC_UNARY_MEM>, OpSize16, LOCK;
681def NAME#32m : I<Opc, Form, (outs), (ins i32mem:$dst),
682                 !strconcat(mnemonic, "{l}\t$dst"),
683                 [], IIC_UNARY_MEM>, OpSize32, LOCK;
684def NAME#64m : RI<Opc, Form, (outs), (ins i64mem:$dst),
685                  !strconcat(mnemonic, "{q}\t$dst"),
686                  [], IIC_UNARY_MEM>, LOCK;
687}
688}
689
690defm LOCK_INC    : LOCK_ArithUnOp<0xFE, 0xFF, MRM0m, "inc">;
691defm LOCK_DEC    : LOCK_ArithUnOp<0xFE, 0xFF, MRM1m, "dec">;
692
693// Atomic compare and swap.
694multiclass LCMPXCHG_UnOp<bits<8> Opc, Format Form, string mnemonic,
695                         SDPatternOperator frag, X86MemOperand x86memop,
696                         InstrItinClass itin> {
697let isCodeGenOnly = 1 in {
698  def NAME : I<Opc, Form, (outs), (ins x86memop:$ptr),
699               !strconcat(mnemonic, "\t$ptr"),
700               [(frag addr:$ptr)], itin>, TB, LOCK;
701}
702}
703
704multiclass LCMPXCHG_BinOp<bits<8> Opc8, bits<8> Opc, Format Form,
705                          string mnemonic, SDPatternOperator frag,
706                          InstrItinClass itin8, InstrItinClass itin> {
707let isCodeGenOnly = 1, SchedRW = [WriteALULd, WriteRMW] in {
708  let Defs = [AL, EFLAGS], Uses = [AL] in
709  def NAME#8  : I<Opc8, Form, (outs), (ins i8mem:$ptr, GR8:$swap),
710                  !strconcat(mnemonic, "{b}\t{$swap, $ptr|$ptr, $swap}"),
711                  [(frag addr:$ptr, GR8:$swap, 1)], itin8>, TB, LOCK;
712  let Defs = [AX, EFLAGS], Uses = [AX] in
713  def NAME#16 : I<Opc, Form, (outs), (ins i16mem:$ptr, GR16:$swap),
714                  !strconcat(mnemonic, "{w}\t{$swap, $ptr|$ptr, $swap}"),
715                  [(frag addr:$ptr, GR16:$swap, 2)], itin>, TB, OpSize16, LOCK;
716  let Defs = [EAX, EFLAGS], Uses = [EAX] in
717  def NAME#32 : I<Opc, Form, (outs), (ins i32mem:$ptr, GR32:$swap),
718                  !strconcat(mnemonic, "{l}\t{$swap, $ptr|$ptr, $swap}"),
719                  [(frag addr:$ptr, GR32:$swap, 4)], itin>, TB, OpSize32, LOCK;
720  let Defs = [RAX, EFLAGS], Uses = [RAX] in
721  def NAME#64 : RI<Opc, Form, (outs), (ins i64mem:$ptr, GR64:$swap),
722                   !strconcat(mnemonic, "{q}\t{$swap, $ptr|$ptr, $swap}"),
723                   [(frag addr:$ptr, GR64:$swap, 8)], itin>, TB, LOCK;
724}
725}
726
727let Defs = [EAX, EDX, EFLAGS], Uses = [EAX, EBX, ECX, EDX],
728    SchedRW = [WriteALULd, WriteRMW] in {
729defm LCMPXCHG8B : LCMPXCHG_UnOp<0xC7, MRM1m, "cmpxchg8b",
730                                X86cas8, i64mem,
731                                IIC_CMPX_LOCK_8B>;
732}
733
734let Defs = [RAX, RDX, EFLAGS], Uses = [RAX, RBX, RCX, RDX],
735    Predicates = [HasCmpxchg16b], SchedRW = [WriteALULd, WriteRMW] in {
736defm LCMPXCHG16B : LCMPXCHG_UnOp<0xC7, MRM1m, "cmpxchg16b",
737                                 X86cas16, i128mem,
738                                 IIC_CMPX_LOCK_16B>, REX_W;
739}
740
741defm LCMPXCHG : LCMPXCHG_BinOp<0xB0, 0xB1, MRMDestMem, "cmpxchg",
742                               X86cas, IIC_CMPX_LOCK_8, IIC_CMPX_LOCK>;
743
744// Atomic exchange and add
745multiclass ATOMIC_LOAD_BINOP<bits<8> opc8, bits<8> opc, string mnemonic,
746                             string frag,
747                             InstrItinClass itin8, InstrItinClass itin> {
748  let Constraints = "$val = $dst", Defs = [EFLAGS], isCodeGenOnly = 1,
749      SchedRW = [WriteALULd, WriteRMW] in {
750    def NAME#8  : I<opc8, MRMSrcMem, (outs GR8:$dst),
751                    (ins GR8:$val, i8mem:$ptr),
752                    !strconcat(mnemonic, "{b}\t{$val, $ptr|$ptr, $val}"),
753                    [(set GR8:$dst,
754                          (!cast<PatFrag>(frag # "_8") addr:$ptr, GR8:$val))],
755                    itin8>;
756    def NAME#16 : I<opc, MRMSrcMem, (outs GR16:$dst),
757                    (ins GR16:$val, i16mem:$ptr),
758                    !strconcat(mnemonic, "{w}\t{$val, $ptr|$ptr, $val}"),
759                    [(set
760                       GR16:$dst,
761                       (!cast<PatFrag>(frag # "_16") addr:$ptr, GR16:$val))],
762                    itin>, OpSize16;
763    def NAME#32 : I<opc, MRMSrcMem, (outs GR32:$dst),
764                    (ins GR32:$val, i32mem:$ptr),
765                    !strconcat(mnemonic, "{l}\t{$val, $ptr|$ptr, $val}"),
766                    [(set
767                       GR32:$dst,
768                       (!cast<PatFrag>(frag # "_32") addr:$ptr, GR32:$val))],
769                    itin>, OpSize32;
770    def NAME#64 : RI<opc, MRMSrcMem, (outs GR64:$dst),
771                     (ins GR64:$val, i64mem:$ptr),
772                     !strconcat(mnemonic, "{q}\t{$val, $ptr|$ptr, $val}"),
773                     [(set
774                        GR64:$dst,
775                        (!cast<PatFrag>(frag # "_64") addr:$ptr, GR64:$val))],
776                     itin>;
777  }
778}
779
780defm LXADD : ATOMIC_LOAD_BINOP<0xc0, 0xc1, "xadd", "atomic_load_add",
781                               IIC_XADD_LOCK_MEM8, IIC_XADD_LOCK_MEM>,
782             TB, LOCK;
783
784/* The following multiclass tries to make sure that in code like
785 *    x.store (immediate op x.load(acquire), release)
786 * and
787 *    x.store (register op x.load(acquire), release)
788 * an operation directly on memory is generated instead of wasting a register.
789 * It is not automatic as atomic_store/load are only lowered to MOV instructions
790 * extremely late to prevent them from being accidentally reordered in the backend
791 * (see below the RELEASE_MOV* / ACQUIRE_MOV* pseudo-instructions)
792 */
793multiclass RELEASE_BINOP_MI<SDNode op> {
794    def NAME#8mi : I<0, Pseudo, (outs), (ins i8mem:$dst, i8imm:$src),
795        "#BINOP "#NAME#"8mi PSEUDO!",
796        [(atomic_store_8 addr:$dst, (op
797            (atomic_load_8 addr:$dst), (i8 imm:$src)))]>;
798    def NAME#8mr : I<0, Pseudo, (outs), (ins i8mem:$dst, GR8:$src),
799        "#BINOP "#NAME#"8mr PSEUDO!",
800        [(atomic_store_8 addr:$dst, (op
801            (atomic_load_8 addr:$dst), GR8:$src))]>;
802    // NAME#16 is not generated as 16-bit arithmetic instructions are considered
803    // costly and avoided as far as possible by this backend anyway
804    def NAME#32mi : I<0, Pseudo, (outs), (ins i32mem:$dst, i32imm:$src),
805        "#BINOP "#NAME#"32mi PSEUDO!",
806        [(atomic_store_32 addr:$dst, (op
807            (atomic_load_32 addr:$dst), (i32 imm:$src)))]>;
808    def NAME#32mr : I<0, Pseudo, (outs), (ins i32mem:$dst, GR32:$src),
809        "#BINOP "#NAME#"32mr PSEUDO!",
810        [(atomic_store_32 addr:$dst, (op
811            (atomic_load_32 addr:$dst), GR32:$src))]>;
812    def NAME#64mi32 : I<0, Pseudo, (outs), (ins i64mem:$dst, i64i32imm:$src),
813        "#BINOP "#NAME#"64mi32 PSEUDO!",
814        [(atomic_store_64 addr:$dst, (op
815            (atomic_load_64 addr:$dst), (i64immSExt32:$src)))]>;
816    def NAME#64mr : I<0, Pseudo, (outs), (ins i64mem:$dst, GR64:$src),
817        "#BINOP "#NAME#"64mr PSEUDO!",
818        [(atomic_store_64 addr:$dst, (op
819            (atomic_load_64 addr:$dst), GR64:$src))]>;
820}
821let Defs = [EFLAGS] in {
822  defm RELEASE_ADD : RELEASE_BINOP_MI<add>;
823  defm RELEASE_AND : RELEASE_BINOP_MI<and>;
824  defm RELEASE_OR  : RELEASE_BINOP_MI<or>;
825  defm RELEASE_XOR : RELEASE_BINOP_MI<xor>;
826  // Note: we don't deal with sub, because substractions of constants are
827  //       optimized into additions before this code can run.
828}
829
830// Same as above, but for floating-point.
831// FIXME: imm version.
832// FIXME: Version that doesn't clobber $src, using AVX's VADDSS.
833// FIXME: This could also handle SIMD operations with *ps and *pd instructions.
834let usesCustomInserter = 1 in {
835multiclass RELEASE_FP_BINOP_MI<SDNode op> {
836    def NAME#32mr : I<0, Pseudo, (outs), (ins i32mem:$dst, FR32:$src),
837        "#BINOP "#NAME#"32mr PSEUDO!",
838        [(atomic_store_32 addr:$dst,
839	   (i32 (bitconvert (op
840             (f32 (bitconvert (i32 (atomic_load_32 addr:$dst)))),
841	      FR32:$src))))]>, Requires<[HasSSE1]>;
842    def NAME#64mr : I<0, Pseudo, (outs), (ins i64mem:$dst, FR64:$src),
843        "#BINOP "#NAME#"64mr PSEUDO!",
844        [(atomic_store_64 addr:$dst,
845	   (i64 (bitconvert (op
846             (f64 (bitconvert (i64 (atomic_load_64 addr:$dst)))),
847	      FR64:$src))))]>, Requires<[HasSSE2]>;
848}
849defm RELEASE_FADD : RELEASE_FP_BINOP_MI<fadd>;
850// FIXME: Add fsub, fmul, fdiv, ...
851}
852
853multiclass RELEASE_UNOP<dag dag8, dag dag16, dag dag32, dag dag64> {
854    def NAME#8m : I<0, Pseudo, (outs), (ins i8mem:$dst),
855        "#UNOP "#NAME#"8m PSEUDO!",
856        [(atomic_store_8 addr:$dst, dag8)]>;
857    def NAME#16m : I<0, Pseudo, (outs), (ins i16mem:$dst),
858        "#UNOP "#NAME#"16m PSEUDO!",
859        [(atomic_store_16 addr:$dst, dag16)]>;
860    def NAME#32m : I<0, Pseudo, (outs), (ins i32mem:$dst),
861        "#UNOP "#NAME#"32m PSEUDO!",
862        [(atomic_store_32 addr:$dst, dag32)]>;
863    def NAME#64m : I<0, Pseudo, (outs), (ins i64mem:$dst),
864        "#UNOP "#NAME#"64m PSEUDO!",
865        [(atomic_store_64 addr:$dst, dag64)]>;
866}
867
868let Defs = [EFLAGS] in {
869  defm RELEASE_INC : RELEASE_UNOP<
870      (add (atomic_load_8  addr:$dst), (i8 1)),
871      (add (atomic_load_16 addr:$dst), (i16 1)),
872      (add (atomic_load_32 addr:$dst), (i32 1)),
873      (add (atomic_load_64 addr:$dst), (i64 1))>, Requires<[NotSlowIncDec]>;
874  defm RELEASE_DEC : RELEASE_UNOP<
875      (add (atomic_load_8  addr:$dst), (i8 -1)),
876      (add (atomic_load_16 addr:$dst), (i16 -1)),
877      (add (atomic_load_32 addr:$dst), (i32 -1)),
878      (add (atomic_load_64 addr:$dst), (i64 -1))>, Requires<[NotSlowIncDec]>;
879}
880/*
881TODO: These don't work because the type inference of TableGen fails.
882TODO: find a way to fix it.
883let Defs = [EFLAGS] in {
884  defm RELEASE_NEG : RELEASE_UNOP<
885      (ineg (atomic_load_8  addr:$dst)),
886      (ineg (atomic_load_16 addr:$dst)),
887      (ineg (atomic_load_32 addr:$dst)),
888      (ineg (atomic_load_64 addr:$dst))>;
889}
890// NOT doesn't set flags.
891defm RELEASE_NOT : RELEASE_UNOP<
892    (not (atomic_load_8  addr:$dst)),
893    (not (atomic_load_16 addr:$dst)),
894    (not (atomic_load_32 addr:$dst)),
895    (not (atomic_load_64 addr:$dst))>;
896*/
897
898def RELEASE_MOV8mi : I<0, Pseudo, (outs), (ins i8mem:$dst, i8imm:$src),
899			"#RELEASE_MOV8mi PSEUDO!",
900			[(atomic_store_8 addr:$dst, (i8 imm:$src))]>;
901def RELEASE_MOV16mi : I<0, Pseudo, (outs), (ins i16mem:$dst, i16imm:$src),
902			"#RELEASE_MOV16mi PSEUDO!",
903			[(atomic_store_16 addr:$dst, (i16 imm:$src))]>;
904def RELEASE_MOV32mi : I<0, Pseudo, (outs), (ins i32mem:$dst, i32imm:$src),
905			"#RELEASE_MOV32mi PSEUDO!",
906			[(atomic_store_32 addr:$dst, (i32 imm:$src))]>;
907def RELEASE_MOV64mi32 : I<0, Pseudo, (outs), (ins i64mem:$dst, i64i32imm:$src),
908			"#RELEASE_MOV64mi32 PSEUDO!",
909			[(atomic_store_64 addr:$dst, i64immSExt32:$src)]>;
910
911def RELEASE_MOV8mr  : I<0, Pseudo, (outs), (ins i8mem :$dst, GR8 :$src),
912                        "#RELEASE_MOV8mr PSEUDO!",
913                        [(atomic_store_8  addr:$dst, GR8 :$src)]>;
914def RELEASE_MOV16mr : I<0, Pseudo, (outs), (ins i16mem:$dst, GR16:$src),
915                        "#RELEASE_MOV16mr PSEUDO!",
916                        [(atomic_store_16 addr:$dst, GR16:$src)]>;
917def RELEASE_MOV32mr : I<0, Pseudo, (outs), (ins i32mem:$dst, GR32:$src),
918                        "#RELEASE_MOV32mr PSEUDO!",
919                        [(atomic_store_32 addr:$dst, GR32:$src)]>;
920def RELEASE_MOV64mr : I<0, Pseudo, (outs), (ins i64mem:$dst, GR64:$src),
921                        "#RELEASE_MOV64mr PSEUDO!",
922                        [(atomic_store_64 addr:$dst, GR64:$src)]>;
923
924def ACQUIRE_MOV8rm  : I<0, Pseudo, (outs GR8 :$dst), (ins i8mem :$src),
925                      "#ACQUIRE_MOV8rm PSEUDO!",
926                      [(set GR8:$dst,  (atomic_load_8  addr:$src))]>;
927def ACQUIRE_MOV16rm : I<0, Pseudo, (outs GR16:$dst), (ins i16mem:$src),
928                      "#ACQUIRE_MOV16rm PSEUDO!",
929                      [(set GR16:$dst, (atomic_load_16 addr:$src))]>;
930def ACQUIRE_MOV32rm : I<0, Pseudo, (outs GR32:$dst), (ins i32mem:$src),
931                      "#ACQUIRE_MOV32rm PSEUDO!",
932                      [(set GR32:$dst, (atomic_load_32 addr:$src))]>;
933def ACQUIRE_MOV64rm : I<0, Pseudo, (outs GR64:$dst), (ins i64mem:$src),
934                      "#ACQUIRE_MOV64rm PSEUDO!",
935                      [(set GR64:$dst, (atomic_load_64 addr:$src))]>;
936
937//===----------------------------------------------------------------------===//
938// DAG Pattern Matching Rules
939//===----------------------------------------------------------------------===//
940
941// ConstantPool GlobalAddress, ExternalSymbol, and JumpTable
942def : Pat<(i32 (X86Wrapper tconstpool  :$dst)), (MOV32ri tconstpool  :$dst)>;
943def : Pat<(i32 (X86Wrapper tjumptable  :$dst)), (MOV32ri tjumptable  :$dst)>;
944def : Pat<(i32 (X86Wrapper tglobaltlsaddr:$dst)),(MOV32ri tglobaltlsaddr:$dst)>;
945def : Pat<(i32 (X86Wrapper tglobaladdr :$dst)), (MOV32ri tglobaladdr :$dst)>;
946def : Pat<(i32 (X86Wrapper texternalsym:$dst)), (MOV32ri texternalsym:$dst)>;
947def : Pat<(i32 (X86Wrapper mcsym:$dst)), (MOV32ri mcsym:$dst)>;
948def : Pat<(i32 (X86Wrapper tblockaddress:$dst)), (MOV32ri tblockaddress:$dst)>;
949
950def : Pat<(add GR32:$src1, (X86Wrapper tconstpool:$src2)),
951          (ADD32ri GR32:$src1, tconstpool:$src2)>;
952def : Pat<(add GR32:$src1, (X86Wrapper tjumptable:$src2)),
953          (ADD32ri GR32:$src1, tjumptable:$src2)>;
954def : Pat<(add GR32:$src1, (X86Wrapper tglobaladdr :$src2)),
955          (ADD32ri GR32:$src1, tglobaladdr:$src2)>;
956def : Pat<(add GR32:$src1, (X86Wrapper texternalsym:$src2)),
957          (ADD32ri GR32:$src1, texternalsym:$src2)>;
958def : Pat<(add GR32:$src1, (X86Wrapper mcsym:$src2)),
959          (ADD32ri GR32:$src1, mcsym:$src2)>;
960def : Pat<(add GR32:$src1, (X86Wrapper tblockaddress:$src2)),
961          (ADD32ri GR32:$src1, tblockaddress:$src2)>;
962
963def : Pat<(store (i32 (X86Wrapper tglobaladdr:$src)), addr:$dst),
964          (MOV32mi addr:$dst, tglobaladdr:$src)>;
965def : Pat<(store (i32 (X86Wrapper texternalsym:$src)), addr:$dst),
966          (MOV32mi addr:$dst, texternalsym:$src)>;
967def : Pat<(store (i32 (X86Wrapper mcsym:$src)), addr:$dst),
968          (MOV32mi addr:$dst, mcsym:$src)>;
969def : Pat<(store (i32 (X86Wrapper tblockaddress:$src)), addr:$dst),
970          (MOV32mi addr:$dst, tblockaddress:$src)>;
971
972// ConstantPool GlobalAddress, ExternalSymbol, and JumpTable when not in small
973// code model mode, should use 'movabs'.  FIXME: This is really a hack, the
974//  'movabs' predicate should handle this sort of thing.
975def : Pat<(i64 (X86Wrapper tconstpool  :$dst)),
976          (MOV64ri tconstpool  :$dst)>, Requires<[FarData]>;
977def : Pat<(i64 (X86Wrapper tjumptable  :$dst)),
978          (MOV64ri tjumptable  :$dst)>, Requires<[FarData]>;
979def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
980          (MOV64ri tglobaladdr :$dst)>, Requires<[FarData]>;
981def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
982          (MOV64ri texternalsym:$dst)>, Requires<[FarData]>;
983def : Pat<(i64 (X86Wrapper mcsym:$dst)),
984          (MOV64ri mcsym:$dst)>, Requires<[FarData]>;
985def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
986          (MOV64ri tblockaddress:$dst)>, Requires<[FarData]>;
987
988// In kernel code model, we can get the address of a label
989// into a register with 'movq'.  FIXME: This is a hack, the 'imm' predicate of
990// the MOV64ri32 should accept these.
991def : Pat<(i64 (X86Wrapper tconstpool  :$dst)),
992          (MOV64ri32 tconstpool  :$dst)>, Requires<[KernelCode]>;
993def : Pat<(i64 (X86Wrapper tjumptable  :$dst)),
994          (MOV64ri32 tjumptable  :$dst)>, Requires<[KernelCode]>;
995def : Pat<(i64 (X86Wrapper tglobaladdr :$dst)),
996          (MOV64ri32 tglobaladdr :$dst)>, Requires<[KernelCode]>;
997def : Pat<(i64 (X86Wrapper texternalsym:$dst)),
998          (MOV64ri32 texternalsym:$dst)>, Requires<[KernelCode]>;
999def : Pat<(i64 (X86Wrapper mcsym:$dst)),
1000          (MOV64ri32 mcsym:$dst)>, Requires<[KernelCode]>;
1001def : Pat<(i64 (X86Wrapper tblockaddress:$dst)),
1002          (MOV64ri32 tblockaddress:$dst)>, Requires<[KernelCode]>;
1003
1004// If we have small model and -static mode, it is safe to store global addresses
1005// directly as immediates.  FIXME: This is really a hack, the 'imm' predicate
1006// for MOV64mi32 should handle this sort of thing.
1007def : Pat<(store (i64 (X86Wrapper tconstpool:$src)), addr:$dst),
1008          (MOV64mi32 addr:$dst, tconstpool:$src)>,
1009          Requires<[NearData, IsStatic]>;
1010def : Pat<(store (i64 (X86Wrapper tjumptable:$src)), addr:$dst),
1011          (MOV64mi32 addr:$dst, tjumptable:$src)>,
1012          Requires<[NearData, IsStatic]>;
1013def : Pat<(store (i64 (X86Wrapper tglobaladdr:$src)), addr:$dst),
1014          (MOV64mi32 addr:$dst, tglobaladdr:$src)>,
1015          Requires<[NearData, IsStatic]>;
1016def : Pat<(store (i64 (X86Wrapper texternalsym:$src)), addr:$dst),
1017          (MOV64mi32 addr:$dst, texternalsym:$src)>,
1018          Requires<[NearData, IsStatic]>;
1019def : Pat<(store (i64 (X86Wrapper mcsym:$src)), addr:$dst),
1020          (MOV64mi32 addr:$dst, mcsym:$src)>,
1021          Requires<[NearData, IsStatic]>;
1022def : Pat<(store (i64 (X86Wrapper tblockaddress:$src)), addr:$dst),
1023          (MOV64mi32 addr:$dst, tblockaddress:$src)>,
1024          Requires<[NearData, IsStatic]>;
1025
1026def : Pat<(i32 (X86RecoverFrameAlloc mcsym:$dst)), (MOV32ri mcsym:$dst)>;
1027def : Pat<(i64 (X86RecoverFrameAlloc mcsym:$dst)), (MOV64ri mcsym:$dst)>;
1028
1029// Calls
1030
1031// tls has some funny stuff here...
1032// This corresponds to movabs $foo@tpoff, %rax
1033def : Pat<(i64 (X86Wrapper tglobaltlsaddr :$dst)),
1034          (MOV64ri32 tglobaltlsaddr :$dst)>;
1035// This corresponds to add $foo@tpoff, %rax
1036def : Pat<(add GR64:$src1, (X86Wrapper tglobaltlsaddr :$dst)),
1037          (ADD64ri32 GR64:$src1, tglobaltlsaddr :$dst)>;
1038
1039
1040// Direct PC relative function call for small code model. 32-bit displacement
1041// sign extended to 64-bit.
1042def : Pat<(X86call (i64 tglobaladdr:$dst)),
1043          (CALL64pcrel32 tglobaladdr:$dst)>;
1044def : Pat<(X86call (i64 texternalsym:$dst)),
1045          (CALL64pcrel32 texternalsym:$dst)>;
1046
1047// Tailcall stuff. The TCRETURN instructions execute after the epilog, so they
1048// can never use callee-saved registers. That is the purpose of the GR64_TC
1049// register classes.
1050//
1051// The only volatile register that is never used by the calling convention is
1052// %r11. This happens when calling a vararg function with 6 arguments.
1053//
1054// Match an X86tcret that uses less than 7 volatile registers.
1055def X86tcret_6regs : PatFrag<(ops node:$ptr, node:$off),
1056                             (X86tcret node:$ptr, node:$off), [{
1057  // X86tcret args: (*chain, ptr, imm, regs..., glue)
1058  unsigned NumRegs = 0;
1059  for (unsigned i = 3, e = N->getNumOperands(); i != e; ++i)
1060    if (isa<RegisterSDNode>(N->getOperand(i)) && ++NumRegs > 6)
1061      return false;
1062  return true;
1063}]>;
1064
1065def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
1066          (TCRETURNri ptr_rc_tailcall:$dst, imm:$off)>,
1067          Requires<[Not64BitMode]>;
1068
1069// FIXME: This is disabled for 32-bit PIC mode because the global base
1070// register which is part of the address mode may be assigned a
1071// callee-saved register.
1072def : Pat<(X86tcret (load addr:$dst), imm:$off),
1073          (TCRETURNmi addr:$dst, imm:$off)>,
1074          Requires<[Not64BitMode, IsNotPIC]>;
1075
1076def : Pat<(X86tcret (i32 tglobaladdr:$dst), imm:$off),
1077          (TCRETURNdi tglobaladdr:$dst, imm:$off)>,
1078          Requires<[NotLP64]>;
1079
1080def : Pat<(X86tcret (i32 texternalsym:$dst), imm:$off),
1081          (TCRETURNdi texternalsym:$dst, imm:$off)>,
1082          Requires<[NotLP64]>;
1083
1084def : Pat<(X86tcret ptr_rc_tailcall:$dst, imm:$off),
1085          (TCRETURNri64 ptr_rc_tailcall:$dst, imm:$off)>,
1086          Requires<[In64BitMode]>;
1087
1088// Don't fold loads into X86tcret requiring more than 6 regs.
1089// There wouldn't be enough scratch registers for base+index.
1090def : Pat<(X86tcret_6regs (load addr:$dst), imm:$off),
1091          (TCRETURNmi64 addr:$dst, imm:$off)>,
1092          Requires<[In64BitMode]>;
1093
1094def : Pat<(X86tcret (i64 tglobaladdr:$dst), imm:$off),
1095          (TCRETURNdi64 tglobaladdr:$dst, imm:$off)>,
1096          Requires<[IsLP64]>;
1097
1098def : Pat<(X86tcret (i64 texternalsym:$dst), imm:$off),
1099          (TCRETURNdi64 texternalsym:$dst, imm:$off)>,
1100          Requires<[IsLP64]>;
1101
1102// Normal calls, with various flavors of addresses.
1103def : Pat<(X86call (i32 tglobaladdr:$dst)),
1104          (CALLpcrel32 tglobaladdr:$dst)>;
1105def : Pat<(X86call (i32 texternalsym:$dst)),
1106          (CALLpcrel32 texternalsym:$dst)>;
1107def : Pat<(X86call (i32 imm:$dst)),
1108          (CALLpcrel32 imm:$dst)>, Requires<[CallImmAddr]>;
1109
1110// Comparisons.
1111
1112// TEST R,R is smaller than CMP R,0
1113def : Pat<(X86cmp GR8:$src1, 0),
1114          (TEST8rr GR8:$src1, GR8:$src1)>;
1115def : Pat<(X86cmp GR16:$src1, 0),
1116          (TEST16rr GR16:$src1, GR16:$src1)>;
1117def : Pat<(X86cmp GR32:$src1, 0),
1118          (TEST32rr GR32:$src1, GR32:$src1)>;
1119def : Pat<(X86cmp GR64:$src1, 0),
1120          (TEST64rr GR64:$src1, GR64:$src1)>;
1121
1122// Conditional moves with folded loads with operands swapped and conditions
1123// inverted.
1124multiclass CMOVmr<PatLeaf InvertedCond, Instruction Inst16, Instruction Inst32,
1125                  Instruction Inst64> {
1126  let Predicates = [HasCMov] in {
1127    def : Pat<(X86cmov (loadi16 addr:$src1), GR16:$src2, InvertedCond, EFLAGS),
1128              (Inst16 GR16:$src2, addr:$src1)>;
1129    def : Pat<(X86cmov (loadi32 addr:$src1), GR32:$src2, InvertedCond, EFLAGS),
1130              (Inst32 GR32:$src2, addr:$src1)>;
1131    def : Pat<(X86cmov (loadi64 addr:$src1), GR64:$src2, InvertedCond, EFLAGS),
1132              (Inst64 GR64:$src2, addr:$src1)>;
1133  }
1134}
1135
1136defm : CMOVmr<X86_COND_B , CMOVAE16rm, CMOVAE32rm, CMOVAE64rm>;
1137defm : CMOVmr<X86_COND_AE, CMOVB16rm , CMOVB32rm , CMOVB64rm>;
1138defm : CMOVmr<X86_COND_E , CMOVNE16rm, CMOVNE32rm, CMOVNE64rm>;
1139defm : CMOVmr<X86_COND_NE, CMOVE16rm , CMOVE32rm , CMOVE64rm>;
1140defm : CMOVmr<X86_COND_BE, CMOVA16rm , CMOVA32rm , CMOVA64rm>;
1141defm : CMOVmr<X86_COND_A , CMOVBE16rm, CMOVBE32rm, CMOVBE64rm>;
1142defm : CMOVmr<X86_COND_L , CMOVGE16rm, CMOVGE32rm, CMOVGE64rm>;
1143defm : CMOVmr<X86_COND_GE, CMOVL16rm , CMOVL32rm , CMOVL64rm>;
1144defm : CMOVmr<X86_COND_LE, CMOVG16rm , CMOVG32rm , CMOVG64rm>;
1145defm : CMOVmr<X86_COND_G , CMOVLE16rm, CMOVLE32rm, CMOVLE64rm>;
1146defm : CMOVmr<X86_COND_P , CMOVNP16rm, CMOVNP32rm, CMOVNP64rm>;
1147defm : CMOVmr<X86_COND_NP, CMOVP16rm , CMOVP32rm , CMOVP64rm>;
1148defm : CMOVmr<X86_COND_S , CMOVNS16rm, CMOVNS32rm, CMOVNS64rm>;
1149defm : CMOVmr<X86_COND_NS, CMOVS16rm , CMOVS32rm , CMOVS64rm>;
1150defm : CMOVmr<X86_COND_O , CMOVNO16rm, CMOVNO32rm, CMOVNO64rm>;
1151defm : CMOVmr<X86_COND_NO, CMOVO16rm , CMOVO32rm , CMOVO64rm>;
1152
1153// zextload bool -> zextload byte
1154def : Pat<(zextloadi8i1  addr:$src), (AND8ri (MOV8rm addr:$src), (i8 1))>;
1155def : Pat<(zextloadi16i1 addr:$src), (AND16ri8 (MOVZX16rm8 addr:$src), (i16 1))>;
1156def : Pat<(zextloadi32i1 addr:$src), (AND32ri8 (MOVZX32rm8 addr:$src), (i32 1))>;
1157def : Pat<(zextloadi64i1 addr:$src),
1158          (SUBREG_TO_REG (i64 0),
1159           (AND32ri8 (MOVZX32rm8 addr:$src), (i32 1)), sub_32bit)>;
1160
1161// extload bool -> extload byte
1162// When extloading from 16-bit and smaller memory locations into 64-bit
1163// registers, use zero-extending loads so that the entire 64-bit register is
1164// defined, avoiding partial-register updates.
1165
1166def : Pat<(extloadi8i1 addr:$src),   (MOV8rm      addr:$src)>;
1167def : Pat<(extloadi16i1 addr:$src),  (MOVZX16rm8  addr:$src)>;
1168def : Pat<(extloadi32i1 addr:$src),  (MOVZX32rm8  addr:$src)>;
1169def : Pat<(extloadi16i8 addr:$src),  (MOVZX16rm8  addr:$src)>;
1170def : Pat<(extloadi32i8 addr:$src),  (MOVZX32rm8  addr:$src)>;
1171def : Pat<(extloadi32i16 addr:$src), (MOVZX32rm16 addr:$src)>;
1172
1173// For other extloads, use subregs, since the high contents of the register are
1174// defined after an extload.
1175def : Pat<(extloadi64i1 addr:$src),
1176          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1177def : Pat<(extloadi64i8 addr:$src),
1178          (SUBREG_TO_REG (i64 0), (MOVZX32rm8 addr:$src), sub_32bit)>;
1179def : Pat<(extloadi64i16 addr:$src),
1180          (SUBREG_TO_REG (i64 0), (MOVZX32rm16 addr:$src), sub_32bit)>;
1181def : Pat<(extloadi64i32 addr:$src),
1182          (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src), sub_32bit)>;
1183
1184// anyext. Define these to do an explicit zero-extend to
1185// avoid partial-register updates.
1186def : Pat<(i16 (anyext GR8 :$src)), (EXTRACT_SUBREG
1187                                     (MOVZX32rr8 GR8 :$src), sub_16bit)>;
1188def : Pat<(i32 (anyext GR8 :$src)), (MOVZX32rr8  GR8 :$src)>;
1189
1190// Except for i16 -> i32 since isel expect i16 ops to be promoted to i32.
1191def : Pat<(i32 (anyext GR16:$src)),
1192          (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR16:$src, sub_16bit)>;
1193
1194def : Pat<(i64 (anyext GR8 :$src)),
1195          (SUBREG_TO_REG (i64 0), (MOVZX32rr8  GR8  :$src), sub_32bit)>;
1196def : Pat<(i64 (anyext GR16:$src)),
1197          (SUBREG_TO_REG (i64 0), (MOVZX32rr16 GR16 :$src), sub_32bit)>;
1198def : Pat<(i64 (anyext GR32:$src)),
1199          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1200
1201
1202// Any instruction that defines a 32-bit result leaves the high half of the
1203// register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
1204// be copying from a truncate. And x86's cmov doesn't do anything if the
1205// condition is false. But any other 32-bit operation will zero-extend
1206// up to 64 bits.
1207def def32 : PatLeaf<(i32 GR32:$src), [{
1208  return N->getOpcode() != ISD::TRUNCATE &&
1209         N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
1210         N->getOpcode() != ISD::CopyFromReg &&
1211         N->getOpcode() != ISD::AssertSext &&
1212         N->getOpcode() != X86ISD::CMOV;
1213}]>;
1214
1215// In the case of a 32-bit def that is known to implicitly zero-extend,
1216// we can use a SUBREG_TO_REG.
1217def : Pat<(i64 (zext def32:$src)),
1218          (SUBREG_TO_REG (i64 0), GR32:$src, sub_32bit)>;
1219
1220//===----------------------------------------------------------------------===//
1221// Pattern match OR as ADD
1222//===----------------------------------------------------------------------===//
1223
1224// If safe, we prefer to pattern match OR as ADD at isel time. ADD can be
1225// 3-addressified into an LEA instruction to avoid copies.  However, we also
1226// want to finally emit these instructions as an or at the end of the code
1227// generator to make the generated code easier to read.  To do this, we select
1228// into "disjoint bits" pseudo ops.
1229
1230// Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.
1231def or_is_add : PatFrag<(ops node:$lhs, node:$rhs), (or node:$lhs, node:$rhs),[{
1232  if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1233    return CurDAG->MaskedValueIsZero(N->getOperand(0), CN->getAPIntValue());
1234
1235  APInt KnownZero0, KnownOne0;
1236  CurDAG->computeKnownBits(N->getOperand(0), KnownZero0, KnownOne0, 0);
1237  APInt KnownZero1, KnownOne1;
1238  CurDAG->computeKnownBits(N->getOperand(1), KnownZero1, KnownOne1, 0);
1239  return (~KnownZero0 & ~KnownZero1) == 0;
1240}]>;
1241
1242
1243// (or x1, x2) -> (add x1, x2) if two operands are known not to share bits.
1244// Try this before the selecting to OR.
1245let AddedComplexity = 5, SchedRW = [WriteALU] in {
1246
1247let isConvertibleToThreeAddress = 1,
1248    Constraints = "$src1 = $dst", Defs = [EFLAGS] in {
1249let isCommutable = 1 in {
1250def ADD16rr_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
1251                    "", // orw/addw REG, REG
1252                    [(set GR16:$dst, (or_is_add GR16:$src1, GR16:$src2))]>;
1253def ADD32rr_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, GR32:$src2),
1254                    "", // orl/addl REG, REG
1255                    [(set GR32:$dst, (or_is_add GR32:$src1, GR32:$src2))]>;
1256def ADD64rr_DB  : I<0, Pseudo, (outs GR64:$dst), (ins GR64:$src1, GR64:$src2),
1257                    "", // orq/addq REG, REG
1258                    [(set GR64:$dst, (or_is_add GR64:$src1, GR64:$src2))]>;
1259} // isCommutable
1260
1261// NOTE: These are order specific, we want the ri8 forms to be listed
1262// first so that they are slightly preferred to the ri forms.
1263
1264def ADD16ri8_DB : I<0, Pseudo,
1265                    (outs GR16:$dst), (ins GR16:$src1, i16i8imm:$src2),
1266                    "", // orw/addw REG, imm8
1267                    [(set GR16:$dst,(or_is_add GR16:$src1,i16immSExt8:$src2))]>;
1268def ADD16ri_DB  : I<0, Pseudo, (outs GR16:$dst), (ins GR16:$src1, i16imm:$src2),
1269                    "", // orw/addw REG, imm
1270                    [(set GR16:$dst, (or_is_add GR16:$src1, imm:$src2))]>;
1271
1272def ADD32ri8_DB : I<0, Pseudo,
1273                    (outs GR32:$dst), (ins GR32:$src1, i32i8imm:$src2),
1274                    "", // orl/addl REG, imm8
1275                    [(set GR32:$dst,(or_is_add GR32:$src1,i32immSExt8:$src2))]>;
1276def ADD32ri_DB  : I<0, Pseudo, (outs GR32:$dst), (ins GR32:$src1, i32imm:$src2),
1277                    "", // orl/addl REG, imm
1278                    [(set GR32:$dst, (or_is_add GR32:$src1, imm:$src2))]>;
1279
1280
1281def ADD64ri8_DB : I<0, Pseudo,
1282                    (outs GR64:$dst), (ins GR64:$src1, i64i8imm:$src2),
1283                    "", // orq/addq REG, imm8
1284                    [(set GR64:$dst, (or_is_add GR64:$src1,
1285                                                i64immSExt8:$src2))]>;
1286def ADD64ri32_DB : I<0, Pseudo,
1287                     (outs GR64:$dst), (ins GR64:$src1, i64i32imm:$src2),
1288                      "", // orq/addq REG, imm
1289                      [(set GR64:$dst, (or_is_add GR64:$src1,
1290                                                  i64immSExt32:$src2))]>;
1291}
1292} // AddedComplexity, SchedRW
1293
1294
1295//===----------------------------------------------------------------------===//
1296// Some peepholes
1297//===----------------------------------------------------------------------===//
1298
1299// Odd encoding trick: -128 fits into an 8-bit immediate field while
1300// +128 doesn't, so in this special case use a sub instead of an add.
1301def : Pat<(add GR16:$src1, 128),
1302          (SUB16ri8 GR16:$src1, -128)>;
1303def : Pat<(store (add (loadi16 addr:$dst), 128), addr:$dst),
1304          (SUB16mi8 addr:$dst, -128)>;
1305
1306def : Pat<(add GR32:$src1, 128),
1307          (SUB32ri8 GR32:$src1, -128)>;
1308def : Pat<(store (add (loadi32 addr:$dst), 128), addr:$dst),
1309          (SUB32mi8 addr:$dst, -128)>;
1310
1311def : Pat<(add GR64:$src1, 128),
1312          (SUB64ri8 GR64:$src1, -128)>;
1313def : Pat<(store (add (loadi64 addr:$dst), 128), addr:$dst),
1314          (SUB64mi8 addr:$dst, -128)>;
1315
1316// The same trick applies for 32-bit immediate fields in 64-bit
1317// instructions.
1318def : Pat<(add GR64:$src1, 0x0000000080000000),
1319          (SUB64ri32 GR64:$src1, 0xffffffff80000000)>;
1320def : Pat<(store (add (loadi64 addr:$dst), 0x00000000800000000), addr:$dst),
1321          (SUB64mi32 addr:$dst, 0xffffffff80000000)>;
1322
1323// To avoid needing to materialize an immediate in a register, use a 32-bit and
1324// with implicit zero-extension instead of a 64-bit and if the immediate has at
1325// least 32 bits of leading zeros. If in addition the last 32 bits can be
1326// represented with a sign extension of a 8 bit constant, use that.
1327// This can also reduce instruction size by eliminating the need for the REX
1328// prefix.
1329
1330// AddedComplexity is needed to give priority over i64immSExt8 and i64immSExt32.
1331let AddedComplexity = 1 in {
1332def : Pat<(and GR64:$src, i64immZExt32SExt8:$imm),
1333          (SUBREG_TO_REG
1334            (i64 0),
1335            (AND32ri8
1336              (EXTRACT_SUBREG GR64:$src, sub_32bit),
1337              (i32 (GetLo8XForm imm:$imm))),
1338            sub_32bit)>;
1339
1340def : Pat<(and GR64:$src, i64immZExt32:$imm),
1341          (SUBREG_TO_REG
1342            (i64 0),
1343            (AND32ri
1344              (EXTRACT_SUBREG GR64:$src, sub_32bit),
1345              (i32 (GetLo32XForm imm:$imm))),
1346            sub_32bit)>;
1347} // AddedComplexity = 1
1348
1349
1350// AddedComplexity is needed due to the increased complexity on the
1351// i64immZExt32SExt8 and i64immZExt32 patterns above. Applying this to all
1352// the MOVZX patterns keeps thems together in DAGIsel tables.
1353let AddedComplexity = 1 in {
1354// r & (2^16-1) ==> movz
1355def : Pat<(and GR32:$src1, 0xffff),
1356          (MOVZX32rr16 (EXTRACT_SUBREG GR32:$src1, sub_16bit))>;
1357// r & (2^8-1) ==> movz
1358def : Pat<(and GR32:$src1, 0xff),
1359          (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src1,
1360                                                             GR32_ABCD)),
1361                                      sub_8bit))>,
1362      Requires<[Not64BitMode]>;
1363// r & (2^8-1) ==> movz
1364def : Pat<(and GR16:$src1, 0xff),
1365           (EXTRACT_SUBREG (MOVZX32rr8 (EXTRACT_SUBREG
1366            (i16 (COPY_TO_REGCLASS GR16:$src1, GR16_ABCD)), sub_8bit)),
1367             sub_16bit)>,
1368      Requires<[Not64BitMode]>;
1369
1370// r & (2^32-1) ==> movz
1371def : Pat<(and GR64:$src, 0x00000000FFFFFFFF),
1372          (SUBREG_TO_REG (i64 0),
1373                         (MOV32rr (EXTRACT_SUBREG GR64:$src, sub_32bit)),
1374                         sub_32bit)>;
1375// r & (2^16-1) ==> movz
1376def : Pat<(and GR64:$src, 0xffff),
1377          (SUBREG_TO_REG (i64 0),
1378                      (MOVZX32rr16 (i16 (EXTRACT_SUBREG GR64:$src, sub_16bit))),
1379                      sub_32bit)>;
1380// r & (2^8-1) ==> movz
1381def : Pat<(and GR64:$src, 0xff),
1382          (SUBREG_TO_REG (i64 0),
1383                         (MOVZX32rr8 (i8 (EXTRACT_SUBREG GR64:$src, sub_8bit))),
1384                         sub_32bit)>;
1385// r & (2^8-1) ==> movz
1386def : Pat<(and GR32:$src1, 0xff),
1387           (MOVZX32rr8 (EXTRACT_SUBREG GR32:$src1, sub_8bit))>,
1388      Requires<[In64BitMode]>;
1389// r & (2^8-1) ==> movz
1390def : Pat<(and GR16:$src1, 0xff),
1391           (EXTRACT_SUBREG (MOVZX32rr8 (i8
1392            (EXTRACT_SUBREG GR16:$src1, sub_8bit))), sub_16bit)>,
1393      Requires<[In64BitMode]>;
1394} // AddedComplexity = 1
1395
1396
1397// sext_inreg patterns
1398def : Pat<(sext_inreg GR32:$src, i16),
1399          (MOVSX32rr16 (EXTRACT_SUBREG GR32:$src, sub_16bit))>;
1400def : Pat<(sext_inreg GR32:$src, i8),
1401          (MOVSX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1402                                                             GR32_ABCD)),
1403                                      sub_8bit))>,
1404      Requires<[Not64BitMode]>;
1405
1406def : Pat<(sext_inreg GR16:$src, i8),
1407           (EXTRACT_SUBREG (i32 (MOVSX32rr8 (EXTRACT_SUBREG
1408            (i32 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)), sub_8bit))),
1409             sub_16bit)>,
1410      Requires<[Not64BitMode]>;
1411
1412def : Pat<(sext_inreg GR64:$src, i32),
1413          (MOVSX64rr32 (EXTRACT_SUBREG GR64:$src, sub_32bit))>;
1414def : Pat<(sext_inreg GR64:$src, i16),
1415          (MOVSX64rr16 (EXTRACT_SUBREG GR64:$src, sub_16bit))>;
1416def : Pat<(sext_inreg GR64:$src, i8),
1417          (MOVSX64rr8 (EXTRACT_SUBREG GR64:$src, sub_8bit))>;
1418def : Pat<(sext_inreg GR32:$src, i8),
1419          (MOVSX32rr8 (EXTRACT_SUBREG GR32:$src, sub_8bit))>,
1420      Requires<[In64BitMode]>;
1421def : Pat<(sext_inreg GR16:$src, i8),
1422           (EXTRACT_SUBREG (MOVSX32rr8
1423            (EXTRACT_SUBREG GR16:$src, sub_8bit)), sub_16bit)>,
1424      Requires<[In64BitMode]>;
1425
1426// sext, sext_load, zext, zext_load
1427def: Pat<(i16 (sext GR8:$src)),
1428          (EXTRACT_SUBREG (MOVSX32rr8 GR8:$src), sub_16bit)>;
1429def: Pat<(sextloadi16i8 addr:$src),
1430          (EXTRACT_SUBREG (MOVSX32rm8 addr:$src), sub_16bit)>;
1431def: Pat<(i16 (zext GR8:$src)),
1432          (EXTRACT_SUBREG (MOVZX32rr8 GR8:$src), sub_16bit)>;
1433def: Pat<(zextloadi16i8 addr:$src),
1434          (EXTRACT_SUBREG (MOVZX32rm8 addr:$src), sub_16bit)>;
1435
1436// trunc patterns
1437def : Pat<(i16 (trunc GR32:$src)),
1438          (EXTRACT_SUBREG GR32:$src, sub_16bit)>;
1439def : Pat<(i8 (trunc GR32:$src)),
1440          (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1441                          sub_8bit)>,
1442      Requires<[Not64BitMode]>;
1443def : Pat<(i8 (trunc GR16:$src)),
1444          (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1445                          sub_8bit)>,
1446      Requires<[Not64BitMode]>;
1447def : Pat<(i32 (trunc GR64:$src)),
1448          (EXTRACT_SUBREG GR64:$src, sub_32bit)>;
1449def : Pat<(i16 (trunc GR64:$src)),
1450          (EXTRACT_SUBREG GR64:$src, sub_16bit)>;
1451def : Pat<(i8 (trunc GR64:$src)),
1452          (EXTRACT_SUBREG GR64:$src, sub_8bit)>;
1453def : Pat<(i8 (trunc GR32:$src)),
1454          (EXTRACT_SUBREG GR32:$src, sub_8bit)>,
1455      Requires<[In64BitMode]>;
1456def : Pat<(i8 (trunc GR16:$src)),
1457          (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
1458      Requires<[In64BitMode]>;
1459
1460// h-register tricks
1461def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
1462          (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1463                          sub_8bit_hi)>,
1464      Requires<[Not64BitMode]>;
1465def : Pat<(i8 (trunc (srl_su GR32:$src, (i8 8)))),
1466          (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1467                          sub_8bit_hi)>,
1468      Requires<[Not64BitMode]>;
1469def : Pat<(srl GR16:$src, (i8 8)),
1470          (EXTRACT_SUBREG
1471            (MOVZX32rr8
1472              (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1473                              sub_8bit_hi)),
1474            sub_16bit)>,
1475      Requires<[Not64BitMode]>;
1476def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1477          (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1478                                                             GR16_ABCD)),
1479                                      sub_8bit_hi))>,
1480      Requires<[Not64BitMode]>;
1481def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1482          (MOVZX32rr8 (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src,
1483                                                             GR16_ABCD)),
1484                                      sub_8bit_hi))>,
1485      Requires<[Not64BitMode]>;
1486def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1487          (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1488                                                             GR32_ABCD)),
1489                                      sub_8bit_hi))>,
1490      Requires<[Not64BitMode]>;
1491def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1492          (MOVZX32rr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1493                                                             GR32_ABCD)),
1494                                      sub_8bit_hi))>,
1495      Requires<[Not64BitMode]>;
1496
1497// h-register tricks.
1498// For now, be conservative on x86-64 and use an h-register extract only if the
1499// value is immediately zero-extended or stored, which are somewhat common
1500// cases. This uses a bunch of code to prevent a register requiring a REX prefix
1501// from being allocated in the same instruction as the h register, as there's
1502// currently no way to describe this requirement to the register allocator.
1503
1504// h-register extract and zero-extend.
1505def : Pat<(and (srl_su GR64:$src, (i8 8)), (i64 255)),
1506          (SUBREG_TO_REG
1507            (i64 0),
1508            (MOVZX32_NOREXrr8
1509              (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1510                              sub_8bit_hi)),
1511            sub_32bit)>;
1512def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
1513          (MOVZX32_NOREXrr8
1514            (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1515                            sub_8bit_hi))>,
1516      Requires<[In64BitMode]>;
1517def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
1518          (MOVZX32_NOREXrr8 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src,
1519                                                                   GR32_ABCD)),
1520                                             sub_8bit_hi))>,
1521      Requires<[In64BitMode]>;
1522def : Pat<(srl GR16:$src, (i8 8)),
1523          (EXTRACT_SUBREG
1524            (MOVZX32_NOREXrr8
1525              (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1526                              sub_8bit_hi)),
1527            sub_16bit)>,
1528      Requires<[In64BitMode]>;
1529def : Pat<(i32 (zext (srl_su GR16:$src, (i8 8)))),
1530          (MOVZX32_NOREXrr8
1531            (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1532                            sub_8bit_hi))>,
1533      Requires<[In64BitMode]>;
1534def : Pat<(i32 (anyext (srl_su GR16:$src, (i8 8)))),
1535          (MOVZX32_NOREXrr8
1536            (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1537                            sub_8bit_hi))>,
1538      Requires<[In64BitMode]>;
1539def : Pat<(i64 (zext (srl_su GR16:$src, (i8 8)))),
1540          (SUBREG_TO_REG
1541            (i64 0),
1542            (MOVZX32_NOREXrr8
1543              (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1544                              sub_8bit_hi)),
1545            sub_32bit)>;
1546def : Pat<(i64 (anyext (srl_su GR16:$src, (i8 8)))),
1547          (SUBREG_TO_REG
1548            (i64 0),
1549            (MOVZX32_NOREXrr8
1550              (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1551                              sub_8bit_hi)),
1552            sub_32bit)>;
1553
1554// h-register extract and store.
1555def : Pat<(store (i8 (trunc_su (srl_su GR64:$src, (i8 8)))), addr:$dst),
1556          (MOV8mr_NOREX
1557            addr:$dst,
1558            (EXTRACT_SUBREG (i64 (COPY_TO_REGCLASS GR64:$src, GR64_ABCD)),
1559                            sub_8bit_hi))>;
1560def : Pat<(store (i8 (trunc_su (srl_su GR32:$src, (i8 8)))), addr:$dst),
1561          (MOV8mr_NOREX
1562            addr:$dst,
1563            (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS GR32:$src, GR32_ABCD)),
1564                            sub_8bit_hi))>,
1565      Requires<[In64BitMode]>;
1566def : Pat<(store (i8 (trunc_su (srl_su GR16:$src, (i8 8)))), addr:$dst),
1567          (MOV8mr_NOREX
1568            addr:$dst,
1569            (EXTRACT_SUBREG (i16 (COPY_TO_REGCLASS GR16:$src, GR16_ABCD)),
1570                            sub_8bit_hi))>,
1571      Requires<[In64BitMode]>;
1572
1573
1574// (shl x, 1) ==> (add x, x)
1575// Note that if x is undef (immediate or otherwise), we could theoretically
1576// end up with the two uses of x getting different values, producing a result
1577// where the least significant bit is not 0. However, the probability of this
1578// happening is considered low enough that this is officially not a
1579// "real problem".
1580def : Pat<(shl GR8 :$src1, (i8 1)), (ADD8rr  GR8 :$src1, GR8 :$src1)>;
1581def : Pat<(shl GR16:$src1, (i8 1)), (ADD16rr GR16:$src1, GR16:$src1)>;
1582def : Pat<(shl GR32:$src1, (i8 1)), (ADD32rr GR32:$src1, GR32:$src1)>;
1583def : Pat<(shl GR64:$src1, (i8 1)), (ADD64rr GR64:$src1, GR64:$src1)>;
1584
1585// Helper imms that check if a mask doesn't change significant shift bits.
1586def immShift32 : ImmLeaf<i8, [{
1587  return countTrailingOnes<uint64_t>(Imm) >= 5;
1588}]>;
1589def immShift64 : ImmLeaf<i8, [{
1590  return countTrailingOnes<uint64_t>(Imm) >= 6;
1591}]>;
1592
1593// Shift amount is implicitly masked.
1594multiclass MaskedShiftAmountPats<SDNode frag, string name> {
1595  // (shift x (and y, 31)) ==> (shift x, y)
1596  def : Pat<(frag GR8:$src1, (and CL, immShift32)),
1597            (!cast<Instruction>(name # "8rCL") GR8:$src1)>;
1598  def : Pat<(frag GR16:$src1, (and CL, immShift32)),
1599            (!cast<Instruction>(name # "16rCL") GR16:$src1)>;
1600  def : Pat<(frag GR32:$src1, (and CL, immShift32)),
1601            (!cast<Instruction>(name # "32rCL") GR32:$src1)>;
1602  def : Pat<(store (frag (loadi8 addr:$dst), (and CL, immShift32)), addr:$dst),
1603            (!cast<Instruction>(name # "8mCL") addr:$dst)>;
1604  def : Pat<(store (frag (loadi16 addr:$dst), (and CL, immShift32)), addr:$dst),
1605            (!cast<Instruction>(name # "16mCL") addr:$dst)>;
1606  def : Pat<(store (frag (loadi32 addr:$dst), (and CL, immShift32)), addr:$dst),
1607            (!cast<Instruction>(name # "32mCL") addr:$dst)>;
1608
1609  // (shift x (and y, 63)) ==> (shift x, y)
1610  def : Pat<(frag GR64:$src1, (and CL, immShift64)),
1611            (!cast<Instruction>(name # "64rCL") GR64:$src1)>;
1612  def : Pat<(store (frag (loadi64 addr:$dst), (and CL, 63)), addr:$dst),
1613            (!cast<Instruction>(name # "64mCL") addr:$dst)>;
1614}
1615
1616defm : MaskedShiftAmountPats<shl, "SHL">;
1617defm : MaskedShiftAmountPats<srl, "SHR">;
1618defm : MaskedShiftAmountPats<sra, "SAR">;
1619defm : MaskedShiftAmountPats<rotl, "ROL">;
1620defm : MaskedShiftAmountPats<rotr, "ROR">;
1621
1622// (anyext (setcc_carry)) -> (setcc_carry)
1623def : Pat<(i16 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1624          (SETB_C16r)>;
1625def : Pat<(i32 (anyext (i8 (X86setcc_c X86_COND_B, EFLAGS)))),
1626          (SETB_C32r)>;
1627def : Pat<(i32 (anyext (i16 (X86setcc_c X86_COND_B, EFLAGS)))),
1628          (SETB_C32r)>;
1629
1630
1631
1632
1633//===----------------------------------------------------------------------===//
1634// EFLAGS-defining Patterns
1635//===----------------------------------------------------------------------===//
1636
1637// add reg, reg
1638def : Pat<(add GR8 :$src1, GR8 :$src2), (ADD8rr  GR8 :$src1, GR8 :$src2)>;
1639def : Pat<(add GR16:$src1, GR16:$src2), (ADD16rr GR16:$src1, GR16:$src2)>;
1640def : Pat<(add GR32:$src1, GR32:$src2), (ADD32rr GR32:$src1, GR32:$src2)>;
1641
1642// add reg, mem
1643def : Pat<(add GR8:$src1, (loadi8 addr:$src2)),
1644          (ADD8rm GR8:$src1, addr:$src2)>;
1645def : Pat<(add GR16:$src1, (loadi16 addr:$src2)),
1646          (ADD16rm GR16:$src1, addr:$src2)>;
1647def : Pat<(add GR32:$src1, (loadi32 addr:$src2)),
1648          (ADD32rm GR32:$src1, addr:$src2)>;
1649
1650// add reg, imm
1651def : Pat<(add GR8 :$src1, imm:$src2), (ADD8ri  GR8:$src1 , imm:$src2)>;
1652def : Pat<(add GR16:$src1, imm:$src2), (ADD16ri GR16:$src1, imm:$src2)>;
1653def : Pat<(add GR32:$src1, imm:$src2), (ADD32ri GR32:$src1, imm:$src2)>;
1654def : Pat<(add GR16:$src1, i16immSExt8:$src2),
1655          (ADD16ri8 GR16:$src1, i16immSExt8:$src2)>;
1656def : Pat<(add GR32:$src1, i32immSExt8:$src2),
1657          (ADD32ri8 GR32:$src1, i32immSExt8:$src2)>;
1658
1659// sub reg, reg
1660def : Pat<(sub GR8 :$src1, GR8 :$src2), (SUB8rr  GR8 :$src1, GR8 :$src2)>;
1661def : Pat<(sub GR16:$src1, GR16:$src2), (SUB16rr GR16:$src1, GR16:$src2)>;
1662def : Pat<(sub GR32:$src1, GR32:$src2), (SUB32rr GR32:$src1, GR32:$src2)>;
1663
1664// sub reg, mem
1665def : Pat<(sub GR8:$src1, (loadi8 addr:$src2)),
1666          (SUB8rm GR8:$src1, addr:$src2)>;
1667def : Pat<(sub GR16:$src1, (loadi16 addr:$src2)),
1668          (SUB16rm GR16:$src1, addr:$src2)>;
1669def : Pat<(sub GR32:$src1, (loadi32 addr:$src2)),
1670          (SUB32rm GR32:$src1, addr:$src2)>;
1671
1672// sub reg, imm
1673def : Pat<(sub GR8:$src1, imm:$src2),
1674          (SUB8ri GR8:$src1, imm:$src2)>;
1675def : Pat<(sub GR16:$src1, imm:$src2),
1676          (SUB16ri GR16:$src1, imm:$src2)>;
1677def : Pat<(sub GR32:$src1, imm:$src2),
1678          (SUB32ri GR32:$src1, imm:$src2)>;
1679def : Pat<(sub GR16:$src1, i16immSExt8:$src2),
1680          (SUB16ri8 GR16:$src1, i16immSExt8:$src2)>;
1681def : Pat<(sub GR32:$src1, i32immSExt8:$src2),
1682          (SUB32ri8 GR32:$src1, i32immSExt8:$src2)>;
1683
1684// sub 0, reg
1685def : Pat<(X86sub_flag 0, GR8 :$src), (NEG8r  GR8 :$src)>;
1686def : Pat<(X86sub_flag 0, GR16:$src), (NEG16r GR16:$src)>;
1687def : Pat<(X86sub_flag 0, GR32:$src), (NEG32r GR32:$src)>;
1688def : Pat<(X86sub_flag 0, GR64:$src), (NEG64r GR64:$src)>;
1689
1690// mul reg, reg
1691def : Pat<(mul GR16:$src1, GR16:$src2),
1692          (IMUL16rr GR16:$src1, GR16:$src2)>;
1693def : Pat<(mul GR32:$src1, GR32:$src2),
1694          (IMUL32rr GR32:$src1, GR32:$src2)>;
1695
1696// mul reg, mem
1697def : Pat<(mul GR16:$src1, (loadi16 addr:$src2)),
1698          (IMUL16rm GR16:$src1, addr:$src2)>;
1699def : Pat<(mul GR32:$src1, (loadi32 addr:$src2)),
1700          (IMUL32rm GR32:$src1, addr:$src2)>;
1701
1702// mul reg, imm
1703def : Pat<(mul GR16:$src1, imm:$src2),
1704          (IMUL16rri GR16:$src1, imm:$src2)>;
1705def : Pat<(mul GR32:$src1, imm:$src2),
1706          (IMUL32rri GR32:$src1, imm:$src2)>;
1707def : Pat<(mul GR16:$src1, i16immSExt8:$src2),
1708          (IMUL16rri8 GR16:$src1, i16immSExt8:$src2)>;
1709def : Pat<(mul GR32:$src1, i32immSExt8:$src2),
1710          (IMUL32rri8 GR32:$src1, i32immSExt8:$src2)>;
1711
1712// reg = mul mem, imm
1713def : Pat<(mul (loadi16 addr:$src1), imm:$src2),
1714          (IMUL16rmi addr:$src1, imm:$src2)>;
1715def : Pat<(mul (loadi32 addr:$src1), imm:$src2),
1716          (IMUL32rmi addr:$src1, imm:$src2)>;
1717def : Pat<(mul (loadi16 addr:$src1), i16immSExt8:$src2),
1718          (IMUL16rmi8 addr:$src1, i16immSExt8:$src2)>;
1719def : Pat<(mul (loadi32 addr:$src1), i32immSExt8:$src2),
1720          (IMUL32rmi8 addr:$src1, i32immSExt8:$src2)>;
1721
1722// Patterns for nodes that do not produce flags, for instructions that do.
1723
1724// addition
1725def : Pat<(add GR64:$src1, GR64:$src2),
1726          (ADD64rr GR64:$src1, GR64:$src2)>;
1727def : Pat<(add GR64:$src1, i64immSExt8:$src2),
1728          (ADD64ri8 GR64:$src1, i64immSExt8:$src2)>;
1729def : Pat<(add GR64:$src1, i64immSExt32:$src2),
1730          (ADD64ri32 GR64:$src1, i64immSExt32:$src2)>;
1731def : Pat<(add GR64:$src1, (loadi64 addr:$src2)),
1732          (ADD64rm GR64:$src1, addr:$src2)>;
1733
1734// subtraction
1735def : Pat<(sub GR64:$src1, GR64:$src2),
1736          (SUB64rr GR64:$src1, GR64:$src2)>;
1737def : Pat<(sub GR64:$src1, (loadi64 addr:$src2)),
1738          (SUB64rm GR64:$src1, addr:$src2)>;
1739def : Pat<(sub GR64:$src1, i64immSExt8:$src2),
1740          (SUB64ri8 GR64:$src1, i64immSExt8:$src2)>;
1741def : Pat<(sub GR64:$src1, i64immSExt32:$src2),
1742          (SUB64ri32 GR64:$src1, i64immSExt32:$src2)>;
1743
1744// Multiply
1745def : Pat<(mul GR64:$src1, GR64:$src2),
1746          (IMUL64rr GR64:$src1, GR64:$src2)>;
1747def : Pat<(mul GR64:$src1, (loadi64 addr:$src2)),
1748          (IMUL64rm GR64:$src1, addr:$src2)>;
1749def : Pat<(mul GR64:$src1, i64immSExt8:$src2),
1750          (IMUL64rri8 GR64:$src1, i64immSExt8:$src2)>;
1751def : Pat<(mul GR64:$src1, i64immSExt32:$src2),
1752          (IMUL64rri32 GR64:$src1, i64immSExt32:$src2)>;
1753def : Pat<(mul (loadi64 addr:$src1), i64immSExt8:$src2),
1754          (IMUL64rmi8 addr:$src1, i64immSExt8:$src2)>;
1755def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
1756          (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
1757
1758// Increment/Decrement reg.
1759// Do not make INC/DEC if it is slow
1760let Predicates = [NotSlowIncDec] in {
1761  def : Pat<(add GR8:$src, 1),   (INC8r GR8:$src)>;
1762  def : Pat<(add GR16:$src, 1),  (INC16r GR16:$src)>;
1763  def : Pat<(add GR32:$src, 1),  (INC32r GR32:$src)>;
1764  def : Pat<(add GR64:$src, 1),  (INC64r GR64:$src)>;
1765  def : Pat<(add GR8:$src, -1),  (DEC8r GR8:$src)>;
1766  def : Pat<(add GR16:$src, -1), (DEC16r GR16:$src)>;
1767  def : Pat<(add GR32:$src, -1), (DEC32r GR32:$src)>;
1768  def : Pat<(add GR64:$src, -1), (DEC64r GR64:$src)>;
1769}
1770
1771// or reg/reg.
1772def : Pat<(or GR8 :$src1, GR8 :$src2), (OR8rr  GR8 :$src1, GR8 :$src2)>;
1773def : Pat<(or GR16:$src1, GR16:$src2), (OR16rr GR16:$src1, GR16:$src2)>;
1774def : Pat<(or GR32:$src1, GR32:$src2), (OR32rr GR32:$src1, GR32:$src2)>;
1775def : Pat<(or GR64:$src1, GR64:$src2), (OR64rr GR64:$src1, GR64:$src2)>;
1776
1777// or reg/mem
1778def : Pat<(or GR8:$src1, (loadi8 addr:$src2)),
1779          (OR8rm GR8:$src1, addr:$src2)>;
1780def : Pat<(or GR16:$src1, (loadi16 addr:$src2)),
1781          (OR16rm GR16:$src1, addr:$src2)>;
1782def : Pat<(or GR32:$src1, (loadi32 addr:$src2)),
1783          (OR32rm GR32:$src1, addr:$src2)>;
1784def : Pat<(or GR64:$src1, (loadi64 addr:$src2)),
1785          (OR64rm GR64:$src1, addr:$src2)>;
1786
1787// or reg/imm
1788def : Pat<(or GR8:$src1 , imm:$src2), (OR8ri  GR8 :$src1, imm:$src2)>;
1789def : Pat<(or GR16:$src1, imm:$src2), (OR16ri GR16:$src1, imm:$src2)>;
1790def : Pat<(or GR32:$src1, imm:$src2), (OR32ri GR32:$src1, imm:$src2)>;
1791def : Pat<(or GR16:$src1, i16immSExt8:$src2),
1792          (OR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1793def : Pat<(or GR32:$src1, i32immSExt8:$src2),
1794          (OR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1795def : Pat<(or GR64:$src1, i64immSExt8:$src2),
1796          (OR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1797def : Pat<(or GR64:$src1, i64immSExt32:$src2),
1798          (OR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1799
1800// xor reg/reg
1801def : Pat<(xor GR8 :$src1, GR8 :$src2), (XOR8rr  GR8 :$src1, GR8 :$src2)>;
1802def : Pat<(xor GR16:$src1, GR16:$src2), (XOR16rr GR16:$src1, GR16:$src2)>;
1803def : Pat<(xor GR32:$src1, GR32:$src2), (XOR32rr GR32:$src1, GR32:$src2)>;
1804def : Pat<(xor GR64:$src1, GR64:$src2), (XOR64rr GR64:$src1, GR64:$src2)>;
1805
1806// xor reg/mem
1807def : Pat<(xor GR8:$src1, (loadi8 addr:$src2)),
1808          (XOR8rm GR8:$src1, addr:$src2)>;
1809def : Pat<(xor GR16:$src1, (loadi16 addr:$src2)),
1810          (XOR16rm GR16:$src1, addr:$src2)>;
1811def : Pat<(xor GR32:$src1, (loadi32 addr:$src2)),
1812          (XOR32rm GR32:$src1, addr:$src2)>;
1813def : Pat<(xor GR64:$src1, (loadi64 addr:$src2)),
1814          (XOR64rm GR64:$src1, addr:$src2)>;
1815
1816// xor reg/imm
1817def : Pat<(xor GR8:$src1, imm:$src2),
1818          (XOR8ri GR8:$src1, imm:$src2)>;
1819def : Pat<(xor GR16:$src1, imm:$src2),
1820          (XOR16ri GR16:$src1, imm:$src2)>;
1821def : Pat<(xor GR32:$src1, imm:$src2),
1822          (XOR32ri GR32:$src1, imm:$src2)>;
1823def : Pat<(xor GR16:$src1, i16immSExt8:$src2),
1824          (XOR16ri8 GR16:$src1, i16immSExt8:$src2)>;
1825def : Pat<(xor GR32:$src1, i32immSExt8:$src2),
1826          (XOR32ri8 GR32:$src1, i32immSExt8:$src2)>;
1827def : Pat<(xor GR64:$src1, i64immSExt8:$src2),
1828          (XOR64ri8 GR64:$src1, i64immSExt8:$src2)>;
1829def : Pat<(xor GR64:$src1, i64immSExt32:$src2),
1830          (XOR64ri32 GR64:$src1, i64immSExt32:$src2)>;
1831
1832// and reg/reg
1833def : Pat<(and GR8 :$src1, GR8 :$src2), (AND8rr  GR8 :$src1, GR8 :$src2)>;
1834def : Pat<(and GR16:$src1, GR16:$src2), (AND16rr GR16:$src1, GR16:$src2)>;
1835def : Pat<(and GR32:$src1, GR32:$src2), (AND32rr GR32:$src1, GR32:$src2)>;
1836def : Pat<(and GR64:$src1, GR64:$src2), (AND64rr GR64:$src1, GR64:$src2)>;
1837
1838// and reg/mem
1839def : Pat<(and GR8:$src1, (loadi8 addr:$src2)),
1840          (AND8rm GR8:$src1, addr:$src2)>;
1841def : Pat<(and GR16:$src1, (loadi16 addr:$src2)),
1842          (AND16rm GR16:$src1, addr:$src2)>;
1843def : Pat<(and GR32:$src1, (loadi32 addr:$src2)),
1844          (AND32rm GR32:$src1, addr:$src2)>;
1845def : Pat<(and GR64:$src1, (loadi64 addr:$src2)),
1846          (AND64rm GR64:$src1, addr:$src2)>;
1847
1848// and reg/imm
1849def : Pat<(and GR8:$src1, imm:$src2),
1850          (AND8ri GR8:$src1, imm:$src2)>;
1851def : Pat<(and GR16:$src1, imm:$src2),
1852          (AND16ri GR16:$src1, imm:$src2)>;
1853def : Pat<(and GR32:$src1, imm:$src2),
1854          (AND32ri GR32:$src1, imm:$src2)>;
1855def : Pat<(and GR16:$src1, i16immSExt8:$src2),
1856          (AND16ri8 GR16:$src1, i16immSExt8:$src2)>;
1857def : Pat<(and GR32:$src1, i32immSExt8:$src2),
1858          (AND32ri8 GR32:$src1, i32immSExt8:$src2)>;
1859def : Pat<(and GR64:$src1, i64immSExt8:$src2),
1860          (AND64ri8 GR64:$src1, i64immSExt8:$src2)>;
1861def : Pat<(and GR64:$src1, i64immSExt32:$src2),
1862          (AND64ri32 GR64:$src1, i64immSExt32:$src2)>;
1863
1864// Bit scan instruction patterns to match explicit zero-undef behavior.
1865def : Pat<(cttz_zero_undef GR16:$src), (BSF16rr GR16:$src)>;
1866def : Pat<(cttz_zero_undef GR32:$src), (BSF32rr GR32:$src)>;
1867def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr GR64:$src)>;
1868def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm addr:$src)>;
1869def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm addr:$src)>;
1870def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm addr:$src)>;
1871
1872// When HasMOVBE is enabled it is possible to get a non-legalized
1873// register-register 16 bit bswap. This maps it to a ROL instruction.
1874let Predicates = [HasMOVBE] in {
1875 def : Pat<(bswap GR16:$src), (ROL16ri GR16:$src, (i8 8))>;
1876}
1877