1 //===-- X86InstrFragmentsSIMD.td - x86 SIMD ISA ------------*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file provides pattern fragments useful for SIMD instructions.
12 //===----------------------------------------------------------------------===//
14 //===----------------------------------------------------------------------===//
15 // MMX specific DAG Nodes.
16 //===----------------------------------------------------------------------===//
18 // Low word of MMX to GPR.
19 def MMX_X86movd2w : SDNode<"X86ISD::MMX_MOVD2W", SDTypeProfile<1, 1,
20 [SDTCisVT<0, i32>, SDTCisVT<1, x86mmx>]>>;
21 // GPR to low word of MMX.
22 def MMX_X86movw2d : SDNode<"X86ISD::MMX_MOVW2D", SDTypeProfile<1, 1,
23 [SDTCisVT<0, x86mmx>, SDTCisVT<1, i32>]>>;
25 //===----------------------------------------------------------------------===//
26 // MMX Pattern Fragments
27 //===----------------------------------------------------------------------===//
29 def load_mmx : PatFrag<(ops node:$ptr), (x86mmx (load node:$ptr))>;
30 def load_mvmmx : PatFrag<(ops node:$ptr),
31 (x86mmx (MMX_X86movw2d (load node:$ptr)))>;
32 def bc_mmx : PatFrag<(ops node:$in), (x86mmx (bitconvert node:$in))>;
34 //===----------------------------------------------------------------------===//
35 // SSE specific DAG Nodes.
36 //===----------------------------------------------------------------------===//
38 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
39 SDTCisFP<0>, SDTCisInt<2> ]>;
40 def SDTX86VFCMP : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>,
41 SDTCisFP<1>, SDTCisVT<3, i8>,
44 def X86umin : SDNode<"X86ISD::UMIN", SDTIntBinOp>;
45 def X86umax : SDNode<"X86ISD::UMAX", SDTIntBinOp>;
46 def X86smin : SDNode<"X86ISD::SMIN", SDTIntBinOp>;
47 def X86smax : SDNode<"X86ISD::SMAX", SDTIntBinOp>;
49 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
50 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
52 // Commutative and Associative FMIN and FMAX.
53 def X86fminc : SDNode<"X86ISD::FMINC", SDTFPBinOp,
54 [SDNPCommutative, SDNPAssociative]>;
55 def X86fmaxc : SDNode<"X86ISD::FMAXC", SDTFPBinOp,
56 [SDNPCommutative, SDNPAssociative]>;
58 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
59 [SDNPCommutative, SDNPAssociative]>;
60 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
61 [SDNPCommutative, SDNPAssociative]>;
62 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
63 [SDNPCommutative, SDNPAssociative]>;
64 def X86fandn : SDNode<"X86ISD::FANDN", SDTFPBinOp,
65 [SDNPCommutative, SDNPAssociative]>;
66 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
67 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
68 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
69 def X86fgetsign: SDNode<"X86ISD::FGETSIGNx86",SDTFPToIntOp>;
70 def X86fhadd : SDNode<"X86ISD::FHADD", SDTFPBinOp>;
71 def X86fhsub : SDNode<"X86ISD::FHSUB", SDTFPBinOp>;
72 def X86hadd : SDNode<"X86ISD::HADD", SDTIntBinOp>;
73 def X86hsub : SDNode<"X86ISD::HSUB", SDTIntBinOp>;
74 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>;
75 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>;
76 def X86cmps : SDNode<"X86ISD::FSETCC", SDTX86Cmps>;
77 //def X86cmpsd : SDNode<"X86ISD::FSETCCsd", SDTX86Cmpsd>;
78 def X86pshufb : SDNode<"X86ISD::PSHUFB",
79 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
81 def X86andnp : SDNode<"X86ISD::ANDNP",
82 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
84 def X86psign : SDNode<"X86ISD::PSIGN",
85 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
87 def X86pextrb : SDNode<"X86ISD::PEXTRB",
88 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
89 def X86pextrw : SDNode<"X86ISD::PEXTRW",
90 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
91 def X86pinsrb : SDNode<"X86ISD::PINSRB",
92 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
93 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
94 def X86pinsrw : SDNode<"X86ISD::PINSRW",
95 SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
96 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
97 def X86insertps : SDNode<"X86ISD::INSERTPS",
98 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
99 SDTCisVT<2, v4f32>, SDTCisVT<3, i8>]>>;
100 def X86vzmovl : SDNode<"X86ISD::VZEXT_MOVL",
101 SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>;
103 def X86vzload : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
104 [SDNPHasChain, SDNPMayLoad, SDNPMemOperand]>;
106 def X86vzext : SDNode<"X86ISD::VZEXT",
107 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
108 SDTCisInt<0>, SDTCisInt<1>,
109 SDTCisOpSmallerThanOp<1, 0>]>>;
111 def X86vsext : SDNode<"X86ISD::VSEXT",
112 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
113 SDTCisInt<0>, SDTCisInt<1>,
114 SDTCisOpSmallerThanOp<1, 0>]>>;
116 def X86vtrunc : SDNode<"X86ISD::VTRUNC",
117 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
118 SDTCisInt<0>, SDTCisInt<1>,
119 SDTCisOpSmallerThanOp<0, 1>]>>;
120 def X86trunc : SDNode<"X86ISD::TRUNC",
121 SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisInt<1>,
122 SDTCisOpSmallerThanOp<0, 1>]>>;
124 def X86vtruncm : SDNode<"X86ISD::VTRUNCM",
125 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
126 SDTCisInt<0>, SDTCisInt<1>,
127 SDTCisVec<2>, SDTCisInt<2>,
128 SDTCisOpSmallerThanOp<0, 2>]>>;
129 def X86vfpext : SDNode<"X86ISD::VFPEXT",
130 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
131 SDTCisFP<0>, SDTCisFP<1>,
132 SDTCisOpSmallerThanOp<1, 0>]>>;
133 def X86vfpround: SDNode<"X86ISD::VFPROUND",
134 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
135 SDTCisFP<0>, SDTCisFP<1>,
136 SDTCisOpSmallerThanOp<0, 1>]>>;
138 def X86vshldq : SDNode<"X86ISD::VSHLDQ", SDTIntShiftOp>;
139 def X86vshrdq : SDNode<"X86ISD::VSRLDQ", SDTIntShiftOp>;
140 def X86cmpp : SDNode<"X86ISD::CMPP", SDTX86VFCMP>;
141 def X86pcmpeq : SDNode<"X86ISD::PCMPEQ", SDTIntBinOp, [SDNPCommutative]>;
142 def X86pcmpgt : SDNode<"X86ISD::PCMPGT", SDTIntBinOp>;
144 def X86IntCmpMask : SDTypeProfile<1, 2,
145 [SDTCisVec<0>, SDTCisSameAs<1, 2>, SDTCisInt<1>]>;
146 def X86pcmpeqm : SDNode<"X86ISD::PCMPEQM", X86IntCmpMask, [SDNPCommutative]>;
147 def X86pcmpgtm : SDNode<"X86ISD::PCMPGTM", X86IntCmpMask>;
150 SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisInt<0>, SDTCisVec<1>,
151 SDTCisSameAs<1, 2>, SDTCisVT<3, i8>]>;
152 def X86CmpMaskCCScalar :
153 SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>, SDTCisVT<3, i8>]>;
155 def X86cmpm : SDNode<"X86ISD::CMPM", X86CmpMaskCC>;
156 def X86cmpmu : SDNode<"X86ISD::CMPMU", X86CmpMaskCC>;
157 def X86cmpms : SDNode<"X86ISD::FSETCC", X86CmpMaskCCScalar>;
159 def X86vshl : SDNode<"X86ISD::VSHL",
160 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
162 def X86vsrl : SDNode<"X86ISD::VSRL",
163 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
165 def X86vsra : SDNode<"X86ISD::VSRA",
166 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
169 def X86vshli : SDNode<"X86ISD::VSHLI", SDTIntShiftOp>;
170 def X86vsrli : SDNode<"X86ISD::VSRLI", SDTIntShiftOp>;
171 def X86vsrai : SDNode<"X86ISD::VSRAI", SDTIntShiftOp>;
173 def SDTX86CmpPTest : SDTypeProfile<1, 2, [SDTCisVT<0, i32>,
175 SDTCisSameAs<2, 1>]>;
176 def X86subus : SDNode<"X86ISD::SUBUS", SDTIntBinOp>;
177 def X86ptest : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
178 def X86testp : SDNode<"X86ISD::TESTP", SDTX86CmpPTest>;
179 def X86kortest : SDNode<"X86ISD::KORTEST", SDTX86CmpPTest>;
180 def X86testm : SDNode<"X86ISD::TESTM", SDTypeProfile<1, 2, [SDTCisVec<0>,
181 SDTCisVec<1>, SDTCisSameAs<2, 1>,
182 SDTCVecEltisVT<0, i1>,
183 SDTCisSameNumEltsAs<0, 1>]>>;
184 def X86testnm : SDNode<"X86ISD::TESTNM", SDTypeProfile<1, 2, [SDTCisVec<0>,
185 SDTCisVec<1>, SDTCisSameAs<2, 1>,
186 SDTCVecEltisVT<0, i1>,
187 SDTCisSameNumEltsAs<0, 1>]>>;
188 def X86select : SDNode<"X86ISD::SELECT" , SDTSelect>;
190 def X86pmuludq : SDNode<"X86ISD::PMULUDQ",
191 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
192 SDTCisSameAs<1,2>]>>;
193 def X86pmuldq : SDNode<"X86ISD::PMULDQ",
194 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
195 SDTCisSameAs<1,2>]>>;
197 // Specific shuffle nodes - At some point ISD::VECTOR_SHUFFLE will always get
198 // translated into one of the target nodes below during lowering.
199 // Note: this is a work in progress...
200 def SDTShuff1Op : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisSameAs<0,1>]>;
201 def SDTShuff2Op : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
203 def SDTShuff3Op : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
204 SDTCisSameAs<0,2>, SDTCisSameAs<0,3>]>;
206 def SDTShuff2OpM : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
208 def SDTShuff2OpI : SDTypeProfile<1, 2, [SDTCisVec<0>,
209 SDTCisSameAs<0,1>, SDTCisInt<2>]>;
210 def SDTShuff3OpI : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
211 SDTCisSameAs<0,2>, SDTCisInt<3>]>;
213 def SDTVBroadcast : SDTypeProfile<1, 1, [SDTCisVec<0>]>;
214 def SDTVBroadcastm : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>]>;
216 def SDTBlend : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
217 SDTCisSameAs<1,2>, SDTCisVT<3, i8>]>;
219 def SDTFPBinOpRound : SDTypeProfile<1, 3, [ // fadd_round, fmul_round, etc.
220 SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>, SDTCisFP<0>, SDTCisInt<3>]>;
222 def SDTFma : SDTypeProfile<1, 3, [SDTCisSameAs<0,1>,
223 SDTCisSameAs<1,2>, SDTCisSameAs<1,3>]>;
224 def SDTFmaRound : SDTypeProfile<1, 4, [SDTCisSameAs<0,1>,
225 SDTCisSameAs<1,2>, SDTCisSameAs<1,3>, SDTCisInt<4>]>;
226 def STDFp1SrcRm : SDTypeProfile<1, 2, [SDTCisSameAs<0,1>,
227 SDTCisVec<0>, SDTCisInt<2>]>;
228 def STDFp2SrcRm : SDTypeProfile<1, 3, [SDTCisSameAs<0,1>,
229 SDTCisVec<0>, SDTCisInt<3>]>;
230 def STDFp3SrcRm : SDTypeProfile<1, 4, [SDTCisSameAs<0,1>,
231 SDTCisVec<0>, SDTCisInt<3>, SDTCisInt<4>]>;
233 def X86PAlignr : SDNode<"X86ISD::PALIGNR", SDTShuff3OpI>;
234 def X86VAlign : SDNode<"X86ISD::VALIGN", SDTShuff3OpI>;
236 def X86PShufd : SDNode<"X86ISD::PSHUFD", SDTShuff2OpI>;
237 def X86PShufhw : SDNode<"X86ISD::PSHUFHW", SDTShuff2OpI>;
238 def X86PShuflw : SDNode<"X86ISD::PSHUFLW", SDTShuff2OpI>;
240 def X86Shufp : SDNode<"X86ISD::SHUFP", SDTShuff3OpI>;
242 def X86Movddup : SDNode<"X86ISD::MOVDDUP", SDTShuff1Op>;
243 def X86Movshdup : SDNode<"X86ISD::MOVSHDUP", SDTShuff1Op>;
244 def X86Movsldup : SDNode<"X86ISD::MOVSLDUP", SDTShuff1Op>;
246 def X86Movsd : SDNode<"X86ISD::MOVSD", SDTShuff2Op>;
247 def X86Movss : SDNode<"X86ISD::MOVSS", SDTShuff2Op>;
249 def X86Movlhps : SDNode<"X86ISD::MOVLHPS", SDTShuff2Op>;
250 def X86Movlhpd : SDNode<"X86ISD::MOVLHPD", SDTShuff2Op>;
251 def X86Movhlps : SDNode<"X86ISD::MOVHLPS", SDTShuff2Op>;
253 def X86Movlps : SDNode<"X86ISD::MOVLPS", SDTShuff2Op>;
254 def X86Movlpd : SDNode<"X86ISD::MOVLPD", SDTShuff2Op>;
256 def SDTPack : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>, SDTCisSameAs<2, 1>]>;
257 def X86Packss : SDNode<"X86ISD::PACKSS", SDTPack>;
258 def X86Packus : SDNode<"X86ISD::PACKUS", SDTPack>;
260 def X86Unpckl : SDNode<"X86ISD::UNPCKL", SDTShuff2Op>;
261 def X86Unpckh : SDNode<"X86ISD::UNPCKH", SDTShuff2Op>;
263 def X86VPermilpv : SDNode<"X86ISD::VPERMILPV", SDTShuff2OpM>;
264 def X86VPermilpi : SDNode<"X86ISD::VPERMILPI", SDTShuff2OpI>;
265 def X86VPermv : SDNode<"X86ISD::VPERMV", SDTShuff2Op>;
266 def X86VPermi : SDNode<"X86ISD::VPERMI", SDTShuff2OpI>;
267 def X86VPermv3 : SDNode<"X86ISD::VPERMV3", SDTShuff3Op>;
268 def X86VPermiv3 : SDNode<"X86ISD::VPERMIV3", SDTShuff3Op>;
270 def X86VPerm2x128 : SDNode<"X86ISD::VPERM2X128", SDTShuff3OpI>;
272 def X86VBroadcast : SDNode<"X86ISD::VBROADCAST", SDTVBroadcast>;
273 def X86VBroadcastm : SDNode<"X86ISD::VBROADCASTM", SDTVBroadcastm>;
274 def X86Vinsert : SDNode<"X86ISD::VINSERT", SDTypeProfile<1, 3,
275 [SDTCisSameAs<0, 1>, SDTCisPtrTy<3>]>, []>;
276 def X86Vextract : SDNode<"X86ISD::VEXTRACT", SDTypeProfile<1, 2,
277 [SDTCisVec<1>, SDTCisPtrTy<2>]>, []>;
279 def X86Blendi : SDNode<"X86ISD::BLENDI", SDTBlend>;
281 def X86Addsub : SDNode<"X86ISD::ADDSUB", SDTFPBinOp>;
283 def X86faddRnd : SDNode<"X86ISD::FADD_RND", SDTFPBinOpRound>;
284 def X86fsubRnd : SDNode<"X86ISD::FSUB_RND", SDTFPBinOpRound>;
285 def X86fmulRnd : SDNode<"X86ISD::FMUL_RND", SDTFPBinOpRound>;
286 def X86fdivRnd : SDNode<"X86ISD::FDIV_RND", SDTFPBinOpRound>;
287 def X86fmaxRnd : SDNode<"X86ISD::FMAX", SDTFPBinOpRound>;
288 def X86fminRnd : SDNode<"X86ISD::FMIN", SDTFPBinOpRound>;
290 def X86Fmadd : SDNode<"X86ISD::FMADD", SDTFma>;
291 def X86Fnmadd : SDNode<"X86ISD::FNMADD", SDTFma>;
292 def X86Fmsub : SDNode<"X86ISD::FMSUB", SDTFma>;
293 def X86Fnmsub : SDNode<"X86ISD::FNMSUB", SDTFma>;
294 def X86Fmaddsub : SDNode<"X86ISD::FMADDSUB", SDTFma>;
295 def X86Fmsubadd : SDNode<"X86ISD::FMSUBADD", SDTFma>;
297 def X86FmaddRnd : SDNode<"X86ISD::FMADD_RND", SDTFmaRound>;
298 def X86FnmaddRnd : SDNode<"X86ISD::FNMADD_RND", SDTFmaRound>;
299 def X86FmsubRnd : SDNode<"X86ISD::FMSUB_RND", SDTFmaRound>;
300 def X86FnmsubRnd : SDNode<"X86ISD::FNMSUB_RND", SDTFmaRound>;
301 def X86FmaddsubRnd : SDNode<"X86ISD::FMADDSUB_RND", SDTFmaRound>;
302 def X86FmsubaddRnd : SDNode<"X86ISD::FMSUBADD_RND", SDTFmaRound>;
304 def X86rsqrt28 : SDNode<"X86ISD::RSQRT28", STDFp1SrcRm>;
305 def X86rcp28 : SDNode<"X86ISD::RCP28", STDFp1SrcRm>;
306 def X86exp2 : SDNode<"X86ISD::EXP2", STDFp1SrcRm>;
308 def X86rsqrt28s : SDNode<"X86ISD::RSQRT28", STDFp2SrcRm>;
309 def X86rcp28s : SDNode<"X86ISD::RCP28", STDFp2SrcRm>;
310 def X86RndScale : SDNode<"X86ISD::RNDSCALE", STDFp3SrcRm>;
312 def SDT_PCMPISTRI : SDTypeProfile<2, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>,
313 SDTCisVT<2, v16i8>, SDTCisVT<3, v16i8>,
315 def SDT_PCMPESTRI : SDTypeProfile<2, 5, [SDTCisVT<0, i32>, SDTCisVT<1, i32>,
316 SDTCisVT<2, v16i8>, SDTCisVT<3, i32>,
317 SDTCisVT<4, v16i8>, SDTCisVT<5, i32>,
320 def X86pcmpistri : SDNode<"X86ISD::PCMPISTRI", SDT_PCMPISTRI>;
321 def X86pcmpestri : SDNode<"X86ISD::PCMPESTRI", SDT_PCMPESTRI>;
323 def X86compress: SDNode<"X86ISD::COMPRESS", SDTypeProfile<1, 3,
324 [SDTCisSameAs<0, 2>, SDTCisSameAs<0, 3>,
325 SDTCisVec<3>, SDTCisVec<1>, SDTCisInt<1>]>, []>;
326 def X86expand : SDNode<"X86ISD::EXPAND", SDTypeProfile<1, 3,
328 SDTCisVec<3>, SDTCisVec<1>, SDTCisInt<1>]>, []>;
330 //===----------------------------------------------------------------------===//
331 // SSE Complex Patterns
332 //===----------------------------------------------------------------------===//
334 // These are 'extloads' from a scalar to the low element of a vector, zeroing
335 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
337 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
338 [SDNPHasChain, SDNPMayLoad, SDNPMemOperand,
340 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
341 [SDNPHasChain, SDNPMayLoad, SDNPMemOperand,
344 def ssmem : Operand<v4f32> {
345 let PrintMethod = "printf32mem";
346 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
347 let ParserMatchClass = X86Mem32AsmOperand;
348 let OperandType = "OPERAND_MEMORY";
350 def sdmem : Operand<v2f64> {
351 let PrintMethod = "printf64mem";
352 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
353 let ParserMatchClass = X86Mem64AsmOperand;
354 let OperandType = "OPERAND_MEMORY";
357 //===----------------------------------------------------------------------===//
358 // SSE pattern fragments
359 //===----------------------------------------------------------------------===//
361 // 128-bit load pattern fragments
362 // NOTE: all 128-bit integer vector loads are promoted to v2i64
363 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
364 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
365 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
367 // 256-bit load pattern fragments
368 // NOTE: all 256-bit integer vector loads are promoted to v4i64
369 def loadv8f32 : PatFrag<(ops node:$ptr), (v8f32 (load node:$ptr))>;
370 def loadv4f64 : PatFrag<(ops node:$ptr), (v4f64 (load node:$ptr))>;
371 def loadv4i64 : PatFrag<(ops node:$ptr), (v4i64 (load node:$ptr))>;
373 // 512-bit load pattern fragments
374 def loadv16f32 : PatFrag<(ops node:$ptr), (v16f32 (load node:$ptr))>;
375 def loadv8f64 : PatFrag<(ops node:$ptr), (v8f64 (load node:$ptr))>;
376 def loadv64i8 : PatFrag<(ops node:$ptr), (v64i8 (load node:$ptr))>;
377 def loadv32i16 : PatFrag<(ops node:$ptr), (v32i16 (load node:$ptr))>;
378 def loadv16i32 : PatFrag<(ops node:$ptr), (v16i32 (load node:$ptr))>;
379 def loadv8i64 : PatFrag<(ops node:$ptr), (v8i64 (load node:$ptr))>;
381 // 128-/256-/512-bit extload pattern fragments
382 def extloadv2f32 : PatFrag<(ops node:$ptr), (v2f64 (extloadvf32 node:$ptr))>;
383 def extloadv4f32 : PatFrag<(ops node:$ptr), (v4f64 (extloadvf32 node:$ptr))>;
384 def extloadv8f32 : PatFrag<(ops node:$ptr), (v8f64 (extloadvf32 node:$ptr))>;
386 // These are needed to match a scalar load that is used in a vector-only
387 // math instruction such as the FP logical ops: andps, andnps, orps, xorps.
388 // The memory operand is required to be a 128-bit load, so it must be converted
389 // from a vector to a scalar.
390 def loadf32_128 : PatFrag<(ops node:$ptr),
391 (f32 (vector_extract (loadv4f32 node:$ptr), (iPTR 0)))>;
392 def loadf64_128 : PatFrag<(ops node:$ptr),
393 (f64 (vector_extract (loadv2f64 node:$ptr), (iPTR 0)))>;
395 // Like 'store', but always requires 128-bit vector alignment.
396 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
397 (store node:$val, node:$ptr), [{
398 return cast<StoreSDNode>(N)->getAlignment() >= 16;
401 // Like 'store', but always requires 256-bit vector alignment.
402 def alignedstore256 : PatFrag<(ops node:$val, node:$ptr),
403 (store node:$val, node:$ptr), [{
404 return cast<StoreSDNode>(N)->getAlignment() >= 32;
407 // Like 'store', but always requires 512-bit vector alignment.
408 def alignedstore512 : PatFrag<(ops node:$val, node:$ptr),
409 (store node:$val, node:$ptr), [{
410 return cast<StoreSDNode>(N)->getAlignment() >= 64;
413 // Like 'load', but always requires 128-bit vector alignment.
414 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
415 return cast<LoadSDNode>(N)->getAlignment() >= 16;
418 // Like 'X86vzload', but always requires 128-bit vector alignment.
419 def alignedX86vzload : PatFrag<(ops node:$ptr), (X86vzload node:$ptr), [{
420 return cast<MemSDNode>(N)->getAlignment() >= 16;
423 // Like 'load', but always requires 256-bit vector alignment.
424 def alignedload256 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
425 return cast<LoadSDNode>(N)->getAlignment() >= 32;
428 // Like 'load', but always requires 512-bit vector alignment.
429 def alignedload512 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
430 return cast<LoadSDNode>(N)->getAlignment() >= 64;
433 def alignedloadfsf32 : PatFrag<(ops node:$ptr),
434 (f32 (alignedload node:$ptr))>;
435 def alignedloadfsf64 : PatFrag<(ops node:$ptr),
436 (f64 (alignedload node:$ptr))>;
438 // 128-bit aligned load pattern fragments
439 // NOTE: all 128-bit integer vector loads are promoted to v2i64
440 def alignedloadv4f32 : PatFrag<(ops node:$ptr),
441 (v4f32 (alignedload node:$ptr))>;
442 def alignedloadv2f64 : PatFrag<(ops node:$ptr),
443 (v2f64 (alignedload node:$ptr))>;
444 def alignedloadv2i64 : PatFrag<(ops node:$ptr),
445 (v2i64 (alignedload node:$ptr))>;
447 // 256-bit aligned load pattern fragments
448 // NOTE: all 256-bit integer vector loads are promoted to v4i64
449 def alignedloadv8f32 : PatFrag<(ops node:$ptr),
450 (v8f32 (alignedload256 node:$ptr))>;
451 def alignedloadv4f64 : PatFrag<(ops node:$ptr),
452 (v4f64 (alignedload256 node:$ptr))>;
453 def alignedloadv4i64 : PatFrag<(ops node:$ptr),
454 (v4i64 (alignedload256 node:$ptr))>;
456 // 512-bit aligned load pattern fragments
457 def alignedloadv16f32 : PatFrag<(ops node:$ptr),
458 (v16f32 (alignedload512 node:$ptr))>;
459 def alignedloadv16i32 : PatFrag<(ops node:$ptr),
460 (v16i32 (alignedload512 node:$ptr))>;
461 def alignedloadv8f64 : PatFrag<(ops node:$ptr),
462 (v8f64 (alignedload512 node:$ptr))>;
463 def alignedloadv8i64 : PatFrag<(ops node:$ptr),
464 (v8i64 (alignedload512 node:$ptr))>;
466 // Like 'load', but uses special alignment checks suitable for use in
467 // memory operands in most SSE instructions, which are required to
468 // be naturally aligned on some targets but not on others. If the subtarget
469 // allows unaligned accesses, match any load, though this may require
470 // setting a feature bit in the processor (on startup, for example).
471 // Opteron 10h and later implement such a feature.
472 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
473 return Subtarget->hasSSEUnalignedMem()
474 || cast<LoadSDNode>(N)->getAlignment() >= 16;
477 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
478 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
480 // 128-bit memop pattern fragments
481 // NOTE: all 128-bit integer vector loads are promoted to v2i64
482 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
483 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
484 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
486 // These are needed to match a scalar memop that is used in a vector-only
487 // math instruction such as the FP logical ops: andps, andnps, orps, xorps.
488 // The memory operand is required to be a 128-bit load, so it must be converted
489 // from a vector to a scalar.
490 def memopfsf32_128 : PatFrag<(ops node:$ptr),
491 (f32 (vector_extract (memopv4f32 node:$ptr), (iPTR 0)))>;
492 def memopfsf64_128 : PatFrag<(ops node:$ptr),
493 (f64 (vector_extract (memopv2f64 node:$ptr), (iPTR 0)))>;
496 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
498 // FIXME: 8 byte alignment for mmx reads is not required
499 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
500 return cast<LoadSDNode>(N)->getAlignment() >= 8;
503 def memopmmx : PatFrag<(ops node:$ptr), (x86mmx (memop64 node:$ptr))>;
506 // Like 'store', but requires the non-temporal bit to be set
507 def nontemporalstore : PatFrag<(ops node:$val, node:$ptr),
508 (st node:$val, node:$ptr), [{
509 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
510 return ST->isNonTemporal();
514 def alignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
515 (st node:$val, node:$ptr), [{
516 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
517 return ST->isNonTemporal() && !ST->isTruncatingStore() &&
518 ST->getAddressingMode() == ISD::UNINDEXED &&
519 ST->getAlignment() >= 16;
523 def unalignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
524 (st node:$val, node:$ptr), [{
525 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
526 return ST->isNonTemporal() &&
527 ST->getAlignment() < 16;
531 def mgatherv8i32 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
532 (masked_gather node:$src1, node:$src2, node:$src3) , [{
533 if (MaskedGatherSDNode *Mgt = dyn_cast<MaskedGatherSDNode>(N))
534 return (Mgt->getIndex().getValueType() == MVT::v8i32 ||
535 Mgt->getBasePtr().getValueType() == MVT::v8i32);
539 def mgatherv8i64 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
540 (masked_gather node:$src1, node:$src2, node:$src3) , [{
541 if (MaskedGatherSDNode *Mgt = dyn_cast<MaskedGatherSDNode>(N))
542 return (Mgt->getIndex().getValueType() == MVT::v8i64 ||
543 Mgt->getBasePtr().getValueType() == MVT::v8i64);
546 def mgatherv16i32 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
547 (masked_gather node:$src1, node:$src2, node:$src3) , [{
548 if (MaskedGatherSDNode *Mgt = dyn_cast<MaskedGatherSDNode>(N))
549 return (Mgt->getIndex().getValueType() == MVT::v16i32 ||
550 Mgt->getBasePtr().getValueType() == MVT::v16i32);
554 def mscatterv8i32 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
555 (masked_scatter node:$src1, node:$src2, node:$src3) , [{
556 if (MaskedScatterSDNode *Sc = dyn_cast<MaskedScatterSDNode>(N))
557 return (Sc->getIndex().getValueType() == MVT::v8i32 ||
558 Sc->getBasePtr().getValueType() == MVT::v8i32);
562 def mscatterv8i64 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
563 (masked_scatter node:$src1, node:$src2, node:$src3) , [{
564 if (MaskedScatterSDNode *Sc = dyn_cast<MaskedScatterSDNode>(N))
565 return (Sc->getIndex().getValueType() == MVT::v8i64 ||
566 Sc->getBasePtr().getValueType() == MVT::v8i64);
569 def mscatterv16i32 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
570 (masked_scatter node:$src1, node:$src2, node:$src3) , [{
571 if (MaskedScatterSDNode *Sc = dyn_cast<MaskedScatterSDNode>(N))
572 return (Sc->getIndex().getValueType() == MVT::v16i32 ||
573 Sc->getBasePtr().getValueType() == MVT::v16i32);
577 // 128-bit bitconvert pattern fragments
578 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
579 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
580 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
581 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
582 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
583 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
585 // 256-bit bitconvert pattern fragments
586 def bc_v32i8 : PatFrag<(ops node:$in), (v32i8 (bitconvert node:$in))>;
587 def bc_v16i16 : PatFrag<(ops node:$in), (v16i16 (bitconvert node:$in))>;
588 def bc_v8i32 : PatFrag<(ops node:$in), (v8i32 (bitconvert node:$in))>;
589 def bc_v4i64 : PatFrag<(ops node:$in), (v4i64 (bitconvert node:$in))>;
590 def bc_v8f32 : PatFrag<(ops node:$in), (v8f32 (bitconvert node:$in))>;
592 // 512-bit bitconvert pattern fragments
593 def bc_v16i32 : PatFrag<(ops node:$in), (v16i32 (bitconvert node:$in))>;
594 def bc_v8i64 : PatFrag<(ops node:$in), (v8i64 (bitconvert node:$in))>;
595 def bc_v8f64 : PatFrag<(ops node:$in), (v8f64 (bitconvert node:$in))>;
596 def bc_v16f32 : PatFrag<(ops node:$in), (v16f32 (bitconvert node:$in))>;
598 def vzmovl_v2i64 : PatFrag<(ops node:$src),
599 (bitconvert (v2i64 (X86vzmovl
600 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
601 def vzmovl_v4i32 : PatFrag<(ops node:$src),
602 (bitconvert (v4i32 (X86vzmovl
603 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
605 def vzload_v2i64 : PatFrag<(ops node:$src),
606 (bitconvert (v2i64 (X86vzload node:$src)))>;
609 def fp32imm0 : PatLeaf<(f32 fpimm), [{
610 return N->isExactlyValue(+0.0);
613 def I8Imm : SDNodeXForm<imm, [{
614 // Transformation function: get the low 8 bits.
615 return getI8Imm((uint8_t)N->getZExtValue(), SDLoc(N));
618 def FROUND_NO_EXC : ImmLeaf<i32, [{ return Imm == 8; }]>;
619 def FROUND_CURRENT : ImmLeaf<i32, [{
620 return Imm == X86::STATIC_ROUNDING::CUR_DIRECTION;
623 // BYTE_imm - Transform bit immediates into byte immediates.
624 def BYTE_imm : SDNodeXForm<imm, [{
625 // Transformation function: imm >> 3
626 return getI32Imm(N->getZExtValue() >> 3, SDLoc(N));
629 // EXTRACT_get_vextract128_imm xform function: convert extract_subvector index
630 // to VEXTRACTF128/VEXTRACTI128 imm.
631 def EXTRACT_get_vextract128_imm : SDNodeXForm<extract_subvector, [{
632 return getI8Imm(X86::getExtractVEXTRACT128Immediate(N), SDLoc(N));
635 // INSERT_get_vinsert128_imm xform function: convert insert_subvector index to
636 // VINSERTF128/VINSERTI128 imm.
637 def INSERT_get_vinsert128_imm : SDNodeXForm<insert_subvector, [{
638 return getI8Imm(X86::getInsertVINSERT128Immediate(N), SDLoc(N));
641 // EXTRACT_get_vextract256_imm xform function: convert extract_subvector index
642 // to VEXTRACTF64x4 imm.
643 def EXTRACT_get_vextract256_imm : SDNodeXForm<extract_subvector, [{
644 return getI8Imm(X86::getExtractVEXTRACT256Immediate(N), SDLoc(N));
647 // INSERT_get_vinsert256_imm xform function: convert insert_subvector index to
649 def INSERT_get_vinsert256_imm : SDNodeXForm<insert_subvector, [{
650 return getI8Imm(X86::getInsertVINSERT256Immediate(N), SDLoc(N));
653 def vextract128_extract : PatFrag<(ops node:$bigvec, node:$index),
654 (extract_subvector node:$bigvec,
656 return X86::isVEXTRACT128Index(N);
657 }], EXTRACT_get_vextract128_imm>;
659 def vinsert128_insert : PatFrag<(ops node:$bigvec, node:$smallvec,
661 (insert_subvector node:$bigvec, node:$smallvec,
663 return X86::isVINSERT128Index(N);
664 }], INSERT_get_vinsert128_imm>;
667 def vextract256_extract : PatFrag<(ops node:$bigvec, node:$index),
668 (extract_subvector node:$bigvec,
670 return X86::isVEXTRACT256Index(N);
671 }], EXTRACT_get_vextract256_imm>;
673 def vinsert256_insert : PatFrag<(ops node:$bigvec, node:$smallvec,
675 (insert_subvector node:$bigvec, node:$smallvec,
677 return X86::isVINSERT256Index(N);
678 }], INSERT_get_vinsert256_imm>;
680 def masked_load_aligned128 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
681 (masked_load node:$src1, node:$src2, node:$src3), [{
682 if (auto *Load = dyn_cast<MaskedLoadSDNode>(N))
683 return Load->getAlignment() >= 16;
687 def masked_load_aligned256 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
688 (masked_load node:$src1, node:$src2, node:$src3), [{
689 if (auto *Load = dyn_cast<MaskedLoadSDNode>(N))
690 return Load->getAlignment() >= 32;
694 def masked_load_aligned512 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
695 (masked_load node:$src1, node:$src2, node:$src3), [{
696 if (auto *Load = dyn_cast<MaskedLoadSDNode>(N))
697 return Load->getAlignment() >= 64;
701 def masked_load_unaligned : PatFrag<(ops node:$src1, node:$src2, node:$src3),
702 (masked_load node:$src1, node:$src2, node:$src3), [{
703 return isa<MaskedLoadSDNode>(N);
706 def masked_store_aligned128 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
707 (masked_store node:$src1, node:$src2, node:$src3), [{
708 if (auto *Store = dyn_cast<MaskedStoreSDNode>(N))
709 return Store->getAlignment() >= 16;
713 def masked_store_aligned256 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
714 (masked_store node:$src1, node:$src2, node:$src3), [{
715 if (auto *Store = dyn_cast<MaskedStoreSDNode>(N))
716 return Store->getAlignment() >= 32;
720 def masked_store_aligned512 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
721 (masked_store node:$src1, node:$src2, node:$src3), [{
722 if (auto *Store = dyn_cast<MaskedStoreSDNode>(N))
723 return Store->getAlignment() >= 64;
727 def masked_store_unaligned : PatFrag<(ops node:$src1, node:$src2, node:$src3),
728 (masked_store node:$src1, node:$src2, node:$src3), [{
729 return isa<MaskedStoreSDNode>(N);