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 Pattern Fragments
16 //===----------------------------------------------------------------------===//
18 def load_mmx : PatFrag<(ops node:$ptr), (x86mmx (load node:$ptr))>;
19 def bc_mmx : PatFrag<(ops node:$in), (x86mmx (bitconvert node:$in))>;
21 //===----------------------------------------------------------------------===//
22 // SSE specific DAG Nodes.
23 //===----------------------------------------------------------------------===//
25 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
26 SDTCisFP<0>, SDTCisInt<2> ]>;
27 def SDTX86VFCMP : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>,
28 SDTCisFP<1>, SDTCisVT<3, i8>,
31 def X86umin : SDNode<"X86ISD::UMIN", SDTIntBinOp>;
32 def X86umax : SDNode<"X86ISD::UMAX", SDTIntBinOp>;
33 def X86smin : SDNode<"X86ISD::SMIN", SDTIntBinOp>;
34 def X86smax : SDNode<"X86ISD::SMAX", SDTIntBinOp>;
36 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
37 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
39 // Commutative and Associative FMIN and FMAX.
40 def X86fminc : SDNode<"X86ISD::FMINC", SDTFPBinOp,
41 [SDNPCommutative, SDNPAssociative]>;
42 def X86fmaxc : SDNode<"X86ISD::FMAXC", SDTFPBinOp,
43 [SDNPCommutative, SDNPAssociative]>;
45 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
46 [SDNPCommutative, SDNPAssociative]>;
47 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
48 [SDNPCommutative, SDNPAssociative]>;
49 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
50 [SDNPCommutative, SDNPAssociative]>;
51 def X86fandn : SDNode<"X86ISD::FANDN", SDTFPBinOp,
52 [SDNPCommutative, SDNPAssociative]>;
53 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
54 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
55 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
56 def X86fgetsign: SDNode<"X86ISD::FGETSIGNx86",SDTFPToIntOp>;
57 def X86fhadd : SDNode<"X86ISD::FHADD", SDTFPBinOp>;
58 def X86fhsub : SDNode<"X86ISD::FHSUB", SDTFPBinOp>;
59 def X86hadd : SDNode<"X86ISD::HADD", SDTIntBinOp>;
60 def X86hsub : SDNode<"X86ISD::HSUB", SDTIntBinOp>;
61 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>;
62 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>;
63 def X86cmps : SDNode<"X86ISD::FSETCC", SDTX86Cmps>;
64 //def X86cmpsd : SDNode<"X86ISD::FSETCCsd", SDTX86Cmpsd>;
65 def X86pshufb : SDNode<"X86ISD::PSHUFB",
66 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
68 def X86andnp : SDNode<"X86ISD::ANDNP",
69 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
71 def X86psign : SDNode<"X86ISD::PSIGN",
72 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
74 def X86pextrb : SDNode<"X86ISD::PEXTRB",
75 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
76 def X86pextrw : SDNode<"X86ISD::PEXTRW",
77 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
78 def X86pinsrb : SDNode<"X86ISD::PINSRB",
79 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
80 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
81 def X86pinsrw : SDNode<"X86ISD::PINSRW",
82 SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
83 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
84 def X86insrtps : SDNode<"X86ISD::INSERTPS",
85 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
86 SDTCisVT<2, v4f32>, SDTCisPtrTy<3>]>>;
87 def X86vzmovl : SDNode<"X86ISD::VZEXT_MOVL",
88 SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>;
90 def X86vzmovly : SDNode<"X86ISD::VZEXT_MOVL",
91 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
92 SDTCisInt<0>, SDTCisInt<1>,
93 SDTCisOpSmallerThanOp<1, 0> ]>>;
95 def X86vsmovl : SDNode<"X86ISD::VSEXT_MOVL",
97 [SDTCisVec<0>, SDTCisVec<1>, SDTCisInt<0>, SDTCisInt<1>,
98 SDTCisOpSmallerThanOp<1, 0>]>>;
100 def X86vzload : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
101 [SDNPHasChain, SDNPMayLoad, SDNPMemOperand]>;
103 def X86vzext : SDNode<"X86ISD::VZEXT",
104 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
105 SDTCisInt<0>, SDTCisInt<1>,
106 SDTCisOpSmallerThanOp<1, 0>]>>;
108 def X86vsext : SDNode<"X86ISD::VSEXT",
109 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
110 SDTCisInt<0>, SDTCisInt<1>,
111 SDTCisOpSmallerThanOp<1, 0>]>>;
113 def X86vtrunc : SDNode<"X86ISD::VTRUNC",
114 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
115 SDTCisInt<0>, SDTCisInt<1>,
116 SDTCisOpSmallerThanOp<0, 1>]>>;
117 def X86trunc : SDNode<"X86ISD::TRUNC",
118 SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisInt<1>,
119 SDTCisOpSmallerThanOp<0, 1>]>>;
121 def X86vtruncm : SDNode<"X86ISD::VTRUNCM",
122 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
123 SDTCisInt<0>, SDTCisInt<1>,
124 SDTCisVec<2>, SDTCisInt<2>,
125 SDTCisOpSmallerThanOp<0, 2>]>>;
126 def X86vfpext : SDNode<"X86ISD::VFPEXT",
127 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
128 SDTCisFP<0>, SDTCisFP<1>,
129 SDTCisOpSmallerThanOp<1, 0>]>>;
130 def X86vfpround: SDNode<"X86ISD::VFPROUND",
131 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
132 SDTCisFP<0>, SDTCisFP<1>,
133 SDTCisOpSmallerThanOp<0, 1>]>>;
135 def X86vshldq : SDNode<"X86ISD::VSHLDQ", SDTIntShiftOp>;
136 def X86vshrdq : SDNode<"X86ISD::VSRLDQ", SDTIntShiftOp>;
137 def X86cmpp : SDNode<"X86ISD::CMPP", SDTX86VFCMP>;
138 def X86pcmpeq : SDNode<"X86ISD::PCMPEQ", SDTIntBinOp, [SDNPCommutative]>;
139 def X86pcmpgt : SDNode<"X86ISD::PCMPGT", SDTIntBinOp>;
141 def X86IntCmpMask : SDTypeProfile<1, 2,
142 [SDTCisVec<0>, SDTCisSameAs<1, 2>, SDTCisInt<1>]>;
143 def X86pcmpeqm : SDNode<"X86ISD::PCMPEQM", X86IntCmpMask, [SDNPCommutative]>;
144 def X86pcmpgtm : SDNode<"X86ISD::PCMPGTM", X86IntCmpMask>;
147 SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisInt<0>, SDTCisVec<1>,
148 SDTCisSameAs<1, 2>, SDTCisVT<3, i8>]>;
149 def X86CmpMaskCCScalar :
150 SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>, SDTCisVT<3, i8>]>;
152 def X86cmpm : SDNode<"X86ISD::CMPM", X86CmpMaskCC>;
153 def X86cmpmu : SDNode<"X86ISD::CMPMU", X86CmpMaskCC>;
154 def X86cmpms : SDNode<"X86ISD::FSETCC", X86CmpMaskCCScalar>;
156 def X86vshl : SDNode<"X86ISD::VSHL",
157 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
159 def X86vsrl : SDNode<"X86ISD::VSRL",
160 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
162 def X86vsra : SDNode<"X86ISD::VSRA",
163 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
166 def X86vshli : SDNode<"X86ISD::VSHLI", SDTIntShiftOp>;
167 def X86vsrli : SDNode<"X86ISD::VSRLI", SDTIntShiftOp>;
168 def X86vsrai : SDNode<"X86ISD::VSRAI", SDTIntShiftOp>;
170 def SDTX86CmpPTest : SDTypeProfile<1, 2, [SDTCisVT<0, i32>,
172 SDTCisSameAs<2, 1>]>;
173 def X86subus : SDNode<"X86ISD::SUBUS", SDTIntBinOp>;
174 def X86ptest : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
175 def X86testp : SDNode<"X86ISD::TESTP", SDTX86CmpPTest>;
176 def X86kortest : SDNode<"X86ISD::KORTEST", SDTX86CmpPTest>;
177 def X86testm : SDNode<"X86ISD::TESTM", SDTypeProfile<1, 2, [SDTCisVec<0>,
179 SDTCisSameAs<2, 1>]>>;
180 def X86select : SDNode<"X86ISD::SELECT" , SDTSelect>;
182 def X86pmuludq : SDNode<"X86ISD::PMULUDQ",
183 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
184 SDTCisSameAs<1,2>]>>;
186 // Specific shuffle nodes - At some point ISD::VECTOR_SHUFFLE will always get
187 // translated into one of the target nodes below during lowering.
188 // Note: this is a work in progress...
189 def SDTShuff1Op : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisSameAs<0,1>]>;
190 def SDTShuff2Op : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
192 def SDTShuff3Op : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
193 SDTCisSameAs<0,2>, SDTCisSameAs<0,3>]>;
195 def SDTShuff2OpI : SDTypeProfile<1, 2, [SDTCisVec<0>,
196 SDTCisSameAs<0,1>, SDTCisInt<2>]>;
197 def SDTShuff3OpI : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
198 SDTCisSameAs<0,2>, SDTCisInt<3>]>;
200 def SDTVBroadcast : SDTypeProfile<1, 1, [SDTCisVec<0>]>;
201 def SDTVBroadcastm : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>]>;
203 def SDTBlend : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
204 SDTCisSameAs<1,2>, SDTCisVT<3, i32>]>;
206 def SDTFma : SDTypeProfile<1, 3, [SDTCisSameAs<0,1>,
207 SDTCisSameAs<1,2>, SDTCisSameAs<1,3>]>;
209 def X86PAlignr : SDNode<"X86ISD::PALIGNR", SDTShuff3OpI>;
211 def X86PShufd : SDNode<"X86ISD::PSHUFD", SDTShuff2OpI>;
212 def X86PShufhw : SDNode<"X86ISD::PSHUFHW", SDTShuff2OpI>;
213 def X86PShuflw : SDNode<"X86ISD::PSHUFLW", SDTShuff2OpI>;
215 def X86Shufp : SDNode<"X86ISD::SHUFP", SDTShuff3OpI>;
217 def X86Movddup : SDNode<"X86ISD::MOVDDUP", SDTShuff1Op>;
218 def X86Movshdup : SDNode<"X86ISD::MOVSHDUP", SDTShuff1Op>;
219 def X86Movsldup : SDNode<"X86ISD::MOVSLDUP", SDTShuff1Op>;
221 def X86Movsd : SDNode<"X86ISD::MOVSD", SDTShuff2Op>;
222 def X86Movss : SDNode<"X86ISD::MOVSS", SDTShuff2Op>;
224 def X86Movlhps : SDNode<"X86ISD::MOVLHPS", SDTShuff2Op>;
225 def X86Movlhpd : SDNode<"X86ISD::MOVLHPD", SDTShuff2Op>;
226 def X86Movhlps : SDNode<"X86ISD::MOVHLPS", SDTShuff2Op>;
228 def X86Movlps : SDNode<"X86ISD::MOVLPS", SDTShuff2Op>;
229 def X86Movlpd : SDNode<"X86ISD::MOVLPD", SDTShuff2Op>;
231 def X86Unpckl : SDNode<"X86ISD::UNPCKL", SDTShuff2Op>;
232 def X86Unpckh : SDNode<"X86ISD::UNPCKH", SDTShuff2Op>;
234 def X86VPermilp : SDNode<"X86ISD::VPERMILP", SDTShuff2OpI>;
235 def X86VPermv : SDNode<"X86ISD::VPERMV", SDTShuff2Op>;
236 def X86VPermi : SDNode<"X86ISD::VPERMI", SDTShuff2OpI>;
237 def X86VPermv3 : SDNode<"X86ISD::VPERMV3", SDTShuff3Op>;
238 def X86VPermiv3 : SDNode<"X86ISD::VPERMIV3", SDTShuff3Op>;
240 def X86VPerm2x128 : SDNode<"X86ISD::VPERM2X128", SDTShuff3OpI>;
242 def X86VBroadcast : SDNode<"X86ISD::VBROADCAST", SDTVBroadcast>;
243 def X86VBroadcastm : SDNode<"X86ISD::VBROADCASTM", SDTVBroadcastm>;
244 def X86Vinsert : SDNode<"X86ISD::VINSERT", SDTypeProfile<1, 3,
245 [SDTCisSameAs<0, 1>, SDTCisPtrTy<3>]>, []>;
247 def X86Blendi : SDNode<"X86ISD::BLENDI", SDTBlend>;
248 def X86Fmadd : SDNode<"X86ISD::FMADD", SDTFma>;
249 def X86Fnmadd : SDNode<"X86ISD::FNMADD", SDTFma>;
250 def X86Fmsub : SDNode<"X86ISD::FMSUB", SDTFma>;
251 def X86Fnmsub : SDNode<"X86ISD::FNMSUB", SDTFma>;
252 def X86Fmaddsub : SDNode<"X86ISD::FMADDSUB", SDTFma>;
253 def X86Fmsubadd : SDNode<"X86ISD::FMSUBADD", SDTFma>;
255 def SDT_PCMPISTRI : SDTypeProfile<2, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>,
256 SDTCisVT<2, v16i8>, SDTCisVT<3, v16i8>,
258 def SDT_PCMPESTRI : SDTypeProfile<2, 5, [SDTCisVT<0, i32>, SDTCisVT<1, i32>,
259 SDTCisVT<2, v16i8>, SDTCisVT<3, i32>,
260 SDTCisVT<4, v16i8>, SDTCisVT<5, i32>,
263 def X86pcmpistri : SDNode<"X86ISD::PCMPISTRI", SDT_PCMPISTRI>;
264 def X86pcmpestri : SDNode<"X86ISD::PCMPESTRI", SDT_PCMPESTRI>;
266 //===----------------------------------------------------------------------===//
267 // SSE Complex Patterns
268 //===----------------------------------------------------------------------===//
270 // These are 'extloads' from a scalar to the low element of a vector, zeroing
271 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
273 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
274 [SDNPHasChain, SDNPMayLoad, SDNPMemOperand,
276 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
277 [SDNPHasChain, SDNPMayLoad, SDNPMemOperand,
280 def ssmem : Operand<v4f32> {
281 let PrintMethod = "printf32mem";
282 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
283 let ParserMatchClass = X86Mem32AsmOperand;
284 let OperandType = "OPERAND_MEMORY";
286 def sdmem : Operand<v2f64> {
287 let PrintMethod = "printf64mem";
288 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
289 let ParserMatchClass = X86Mem64AsmOperand;
290 let OperandType = "OPERAND_MEMORY";
293 //===----------------------------------------------------------------------===//
294 // SSE pattern fragments
295 //===----------------------------------------------------------------------===//
297 // 128-bit load pattern fragments
298 // NOTE: all 128-bit integer vector loads are promoted to v2i64
299 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
300 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
301 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
303 // 256-bit load pattern fragments
304 // NOTE: all 256-bit integer vector loads are promoted to v4i64
305 def loadv8f32 : PatFrag<(ops node:$ptr), (v8f32 (load node:$ptr))>;
306 def loadv4f64 : PatFrag<(ops node:$ptr), (v4f64 (load node:$ptr))>;
307 def loadv4i64 : PatFrag<(ops node:$ptr), (v4i64 (load node:$ptr))>;
309 // 512-bit load pattern fragments
310 def loadv16f32 : PatFrag<(ops node:$ptr), (v16f32 (load node:$ptr))>;
311 def loadv8f64 : PatFrag<(ops node:$ptr), (v8f64 (load node:$ptr))>;
312 def loadv16i32 : PatFrag<(ops node:$ptr), (v16i32 (load node:$ptr))>;
313 def loadv8i64 : PatFrag<(ops node:$ptr), (v8i64 (load node:$ptr))>;
315 // 128-/256-/512-bit extload pattern fragments
316 def extloadv2f32 : PatFrag<(ops node:$ptr), (v2f64 (extloadvf32 node:$ptr))>;
317 def extloadv4f32 : PatFrag<(ops node:$ptr), (v4f64 (extloadvf32 node:$ptr))>;
318 def extloadv8f32 : PatFrag<(ops node:$ptr), (v8f64 (extloadvf32 node:$ptr))>;
320 // Like 'store', but always requires 128-bit vector alignment.
321 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
322 (store node:$val, node:$ptr), [{
323 return cast<StoreSDNode>(N)->getAlignment() >= 16;
326 // Like 'store', but always requires 256-bit vector alignment.
327 def alignedstore256 : PatFrag<(ops node:$val, node:$ptr),
328 (store node:$val, node:$ptr), [{
329 return cast<StoreSDNode>(N)->getAlignment() >= 32;
332 // Like 'store', but always requires 512-bit vector alignment.
333 def alignedstore512 : PatFrag<(ops node:$val, node:$ptr),
334 (store node:$val, node:$ptr), [{
335 return cast<StoreSDNode>(N)->getAlignment() >= 64;
338 // Like 'load', but always requires 128-bit vector alignment.
339 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
340 return cast<LoadSDNode>(N)->getAlignment() >= 16;
343 // Like 'X86vzload', but always requires 128-bit vector alignment.
344 def alignedX86vzload : PatFrag<(ops node:$ptr), (X86vzload node:$ptr), [{
345 return cast<MemSDNode>(N)->getAlignment() >= 16;
348 // Like 'load', but always requires 256-bit vector alignment.
349 def alignedload256 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
350 return cast<LoadSDNode>(N)->getAlignment() >= 32;
353 // Like 'load', but always requires 512-bit vector alignment.
354 def alignedload512 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
355 return cast<LoadSDNode>(N)->getAlignment() >= 64;
358 def alignedloadfsf32 : PatFrag<(ops node:$ptr),
359 (f32 (alignedload node:$ptr))>;
360 def alignedloadfsf64 : PatFrag<(ops node:$ptr),
361 (f64 (alignedload node:$ptr))>;
363 // 128-bit aligned load pattern fragments
364 // NOTE: all 128-bit integer vector loads are promoted to v2i64
365 def alignedloadv4f32 : PatFrag<(ops node:$ptr),
366 (v4f32 (alignedload node:$ptr))>;
367 def alignedloadv2f64 : PatFrag<(ops node:$ptr),
368 (v2f64 (alignedload node:$ptr))>;
369 def alignedloadv2i64 : PatFrag<(ops node:$ptr),
370 (v2i64 (alignedload node:$ptr))>;
372 // 256-bit aligned load pattern fragments
373 // NOTE: all 256-bit integer vector loads are promoted to v4i64
374 def alignedloadv8f32 : PatFrag<(ops node:$ptr),
375 (v8f32 (alignedload256 node:$ptr))>;
376 def alignedloadv4f64 : PatFrag<(ops node:$ptr),
377 (v4f64 (alignedload256 node:$ptr))>;
378 def alignedloadv4i64 : PatFrag<(ops node:$ptr),
379 (v4i64 (alignedload256 node:$ptr))>;
381 // 512-bit aligned load pattern fragments
382 def alignedloadv16f32 : PatFrag<(ops node:$ptr),
383 (v16f32 (alignedload512 node:$ptr))>;
384 def alignedloadv16i32 : PatFrag<(ops node:$ptr),
385 (v16i32 (alignedload512 node:$ptr))>;
386 def alignedloadv8f64 : PatFrag<(ops node:$ptr),
387 (v8f64 (alignedload512 node:$ptr))>;
388 def alignedloadv8i64 : PatFrag<(ops node:$ptr),
389 (v8i64 (alignedload512 node:$ptr))>;
391 // Like 'load', but uses special alignment checks suitable for use in
392 // memory operands in most SSE instructions, which are required to
393 // be naturally aligned on some targets but not on others. If the subtarget
394 // allows unaligned accesses, match any load, though this may require
395 // setting a feature bit in the processor (on startup, for example).
396 // Opteron 10h and later implement such a feature.
397 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
398 return Subtarget->hasVectorUAMem()
399 || cast<LoadSDNode>(N)->getAlignment() >= 16;
402 def memop4 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
403 return Subtarget->hasVectorUAMem()
404 || cast<LoadSDNode>(N)->getAlignment() >= 4;
407 def memop8 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
408 return Subtarget->hasVectorUAMem()
409 || cast<LoadSDNode>(N)->getAlignment() >= 8;
412 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
413 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
415 // 128-bit memop pattern fragments
416 // NOTE: all 128-bit integer vector loads are promoted to v2i64
417 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
418 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
419 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
421 // 256-bit memop pattern fragments
422 // NOTE: all 256-bit integer vector loads are promoted to v4i64
423 def memopv8f32 : PatFrag<(ops node:$ptr), (v8f32 (memop node:$ptr))>;
424 def memopv4f64 : PatFrag<(ops node:$ptr), (v4f64 (memop node:$ptr))>;
425 def memopv4i64 : PatFrag<(ops node:$ptr), (v4i64 (memop node:$ptr))>;
427 // 512-bit memop pattern fragments
428 def memopv16f32 : PatFrag<(ops node:$ptr), (v16f32 (memop4 node:$ptr))>;
429 def memopv8f64 : PatFrag<(ops node:$ptr), (v8f64 (memop8 node:$ptr))>;
430 def memopv16i32 : PatFrag<(ops node:$ptr), (v16i32 (memop4 node:$ptr))>;
431 def memopv8i64 : PatFrag<(ops node:$ptr), (v8i64 (memop8 node:$ptr))>;
433 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
435 // FIXME: 8 byte alignment for mmx reads is not required
436 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
437 return cast<LoadSDNode>(N)->getAlignment() >= 8;
440 def memopmmx : PatFrag<(ops node:$ptr), (x86mmx (memop64 node:$ptr))>;
443 // Like 'store', but requires the non-temporal bit to be set
444 def nontemporalstore : PatFrag<(ops node:$val, node:$ptr),
445 (st node:$val, node:$ptr), [{
446 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
447 return ST->isNonTemporal();
451 def alignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
452 (st node:$val, node:$ptr), [{
453 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
454 return ST->isNonTemporal() && !ST->isTruncatingStore() &&
455 ST->getAddressingMode() == ISD::UNINDEXED &&
456 ST->getAlignment() >= 16;
460 def unalignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
461 (st node:$val, node:$ptr), [{
462 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
463 return ST->isNonTemporal() &&
464 ST->getAlignment() < 16;
468 // 128-bit bitconvert pattern fragments
469 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
470 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
471 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
472 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
473 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
474 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
476 // 256-bit bitconvert pattern fragments
477 def bc_v32i8 : PatFrag<(ops node:$in), (v32i8 (bitconvert node:$in))>;
478 def bc_v16i16 : PatFrag<(ops node:$in), (v16i16 (bitconvert node:$in))>;
479 def bc_v8i32 : PatFrag<(ops node:$in), (v8i32 (bitconvert node:$in))>;
480 def bc_v4i64 : PatFrag<(ops node:$in), (v4i64 (bitconvert node:$in))>;
481 def bc_v8f32 : PatFrag<(ops node:$in), (v8f32 (bitconvert node:$in))>;
483 // 512-bit bitconvert pattern fragments
484 def bc_v16i32 : PatFrag<(ops node:$in), (v16i32 (bitconvert node:$in))>;
485 def bc_v8i64 : PatFrag<(ops node:$in), (v8i64 (bitconvert node:$in))>;
486 def bc_v8f64 : PatFrag<(ops node:$in), (v8f64 (bitconvert node:$in))>;
487 def bc_v16f32 : PatFrag<(ops node:$in), (v16f32 (bitconvert node:$in))>;
489 def vzmovl_v2i64 : PatFrag<(ops node:$src),
490 (bitconvert (v2i64 (X86vzmovl
491 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
492 def vzmovl_v4i32 : PatFrag<(ops node:$src),
493 (bitconvert (v4i32 (X86vzmovl
494 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
496 def vzload_v2i64 : PatFrag<(ops node:$src),
497 (bitconvert (v2i64 (X86vzload node:$src)))>;
500 def fp32imm0 : PatLeaf<(f32 fpimm), [{
501 return N->isExactlyValue(+0.0);
504 def I8Imm : SDNodeXForm<imm, [{
505 // Transformation function: get the low 8 bits.
506 return getI8Imm((uint8_t)N->getZExtValue());
509 def FROUND_NO_EXC : ImmLeaf<i32, [{ return Imm == 8; }]>;
510 def FROUND_CURRENT : ImmLeaf<i32, [{ return Imm == 4; }]>;
512 // BYTE_imm - Transform bit immediates into byte immediates.
513 def BYTE_imm : SDNodeXForm<imm, [{
514 // Transformation function: imm >> 3
515 return getI32Imm(N->getZExtValue() >> 3);
518 // EXTRACT_get_vextract128_imm xform function: convert extract_subvector index
519 // to VEXTRACTF128/VEXTRACTI128 imm.
520 def EXTRACT_get_vextract128_imm : SDNodeXForm<extract_subvector, [{
521 return getI8Imm(X86::getExtractVEXTRACT128Immediate(N));
524 // INSERT_get_vinsert128_imm xform function: convert insert_subvector index to
525 // VINSERTF128/VINSERTI128 imm.
526 def INSERT_get_vinsert128_imm : SDNodeXForm<insert_subvector, [{
527 return getI8Imm(X86::getInsertVINSERT128Immediate(N));
530 // EXTRACT_get_vextract256_imm xform function: convert extract_subvector index
531 // to VEXTRACTF64x4 imm.
532 def EXTRACT_get_vextract256_imm : SDNodeXForm<extract_subvector, [{
533 return getI8Imm(X86::getExtractVEXTRACT256Immediate(N));
536 // INSERT_get_vinsert256_imm xform function: convert insert_subvector index to
538 def INSERT_get_vinsert256_imm : SDNodeXForm<insert_subvector, [{
539 return getI8Imm(X86::getInsertVINSERT256Immediate(N));
542 def vextract128_extract : PatFrag<(ops node:$bigvec, node:$index),
543 (extract_subvector node:$bigvec,
545 return X86::isVEXTRACT128Index(N);
546 }], EXTRACT_get_vextract128_imm>;
548 def vinsert128_insert : PatFrag<(ops node:$bigvec, node:$smallvec,
550 (insert_subvector node:$bigvec, node:$smallvec,
552 return X86::isVINSERT128Index(N);
553 }], INSERT_get_vinsert128_imm>;
556 def vextract256_extract : PatFrag<(ops node:$bigvec, node:$index),
557 (extract_subvector node:$bigvec,
559 return X86::isVEXTRACT256Index(N);
560 }], EXTRACT_get_vextract256_imm>;
562 def vinsert256_insert : PatFrag<(ops node:$bigvec, node:$smallvec,
564 (insert_subvector node:$bigvec, node:$smallvec,
566 return X86::isVINSERT256Index(N);
567 }], INSERT_get_vinsert256_imm>;