1 //====- X86InstrSSE.td - Describe the X86 Instruction Set --*- 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 describes the X86 SSE instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
17 //===----------------------------------------------------------------------===//
18 // SSE specific DAG Nodes.
19 //===----------------------------------------------------------------------===//
21 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
22 SDTCisFP<0>, SDTCisInt<2> ]>;
23 def SDTX86VFCMP : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>,
24 SDTCisFP<1>, SDTCisVT<3, i8>]>;
26 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
27 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
28 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
29 [SDNPCommutative, SDNPAssociative]>;
30 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
31 [SDNPCommutative, SDNPAssociative]>;
32 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
33 [SDNPCommutative, SDNPAssociative]>;
34 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
35 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
36 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
37 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>;
38 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>;
39 def X86pshufb : SDNode<"X86ISD::PSHUFB",
40 SDTypeProfile<1, 2, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
42 def X86pextrb : SDNode<"X86ISD::PEXTRB",
43 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
44 def X86pextrw : SDNode<"X86ISD::PEXTRW",
45 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
46 def X86pinsrb : SDNode<"X86ISD::PINSRB",
47 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
48 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
49 def X86pinsrw : SDNode<"X86ISD::PINSRW",
50 SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
51 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
52 def X86insrtps : SDNode<"X86ISD::INSERTPS",
53 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
54 SDTCisVT<2, f32>, SDTCisPtrTy<3>]>>;
55 def X86vzmovl : SDNode<"X86ISD::VZEXT_MOVL",
56 SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>;
57 def X86vzload : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
58 [SDNPHasChain, SDNPMayLoad]>;
59 def X86vshl : SDNode<"X86ISD::VSHL", SDTIntShiftOp>;
60 def X86vshr : SDNode<"X86ISD::VSRL", SDTIntShiftOp>;
61 def X86cmpps : SDNode<"X86ISD::CMPPS", SDTX86VFCMP>;
62 def X86cmppd : SDNode<"X86ISD::CMPPD", SDTX86VFCMP>;
63 def X86pcmpeqb : SDNode<"X86ISD::PCMPEQB", SDTIntBinOp, [SDNPCommutative]>;
64 def X86pcmpeqw : SDNode<"X86ISD::PCMPEQW", SDTIntBinOp, [SDNPCommutative]>;
65 def X86pcmpeqd : SDNode<"X86ISD::PCMPEQD", SDTIntBinOp, [SDNPCommutative]>;
66 def X86pcmpeqq : SDNode<"X86ISD::PCMPEQQ", SDTIntBinOp, [SDNPCommutative]>;
67 def X86pcmpgtb : SDNode<"X86ISD::PCMPGTB", SDTIntBinOp>;
68 def X86pcmpgtw : SDNode<"X86ISD::PCMPGTW", SDTIntBinOp>;
69 def X86pcmpgtd : SDNode<"X86ISD::PCMPGTD", SDTIntBinOp>;
70 def X86pcmpgtq : SDNode<"X86ISD::PCMPGTQ", SDTIntBinOp>;
72 //===----------------------------------------------------------------------===//
73 // SSE Complex Patterns
74 //===----------------------------------------------------------------------===//
76 // These are 'extloads' from a scalar to the low element of a vector, zeroing
77 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
79 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
80 [SDNPHasChain, SDNPMayLoad]>;
81 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
82 [SDNPHasChain, SDNPMayLoad]>;
84 def ssmem : Operand<v4f32> {
85 let PrintMethod = "printf32mem";
86 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm, i8imm);
88 def sdmem : Operand<v2f64> {
89 let PrintMethod = "printf64mem";
90 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm, i8imm);
93 //===----------------------------------------------------------------------===//
94 // SSE pattern fragments
95 //===----------------------------------------------------------------------===//
97 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
98 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
99 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
100 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
102 // Like 'store', but always requires vector alignment.
103 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
104 (store node:$val, node:$ptr), [{
105 return cast<StoreSDNode>(N)->getAlignment() >= 16;
108 // Like 'load', but always requires vector alignment.
109 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
110 return cast<LoadSDNode>(N)->getAlignment() >= 16;
113 def alignedloadfsf32 : PatFrag<(ops node:$ptr), (f32 (alignedload node:$ptr))>;
114 def alignedloadfsf64 : PatFrag<(ops node:$ptr), (f64 (alignedload node:$ptr))>;
115 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
116 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
117 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
118 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
120 // Like 'load', but uses special alignment checks suitable for use in
121 // memory operands in most SSE instructions, which are required to
122 // be naturally aligned on some targets but not on others.
123 // FIXME: Actually implement support for targets that don't require the
124 // alignment. This probably wants a subtarget predicate.
125 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
126 return cast<LoadSDNode>(N)->getAlignment() >= 16;
129 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
130 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
131 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
132 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
133 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
134 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
135 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
137 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
139 // FIXME: 8 byte alignment for mmx reads is not required
140 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
141 return cast<LoadSDNode>(N)->getAlignment() >= 8;
144 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
145 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
146 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
147 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
149 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
150 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
151 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
152 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
153 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
154 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
156 def vzmovl_v2i64 : PatFrag<(ops node:$src),
157 (bitconvert (v2i64 (X86vzmovl
158 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
159 def vzmovl_v4i32 : PatFrag<(ops node:$src),
160 (bitconvert (v4i32 (X86vzmovl
161 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
163 def vzload_v2i64 : PatFrag<(ops node:$src),
164 (bitconvert (v2i64 (X86vzload node:$src)))>;
167 def fp32imm0 : PatLeaf<(f32 fpimm), [{
168 return N->isExactlyValue(+0.0);
171 def PSxLDQ_imm : SDNodeXForm<imm, [{
172 // Transformation function: imm >> 3
173 return getI32Imm(N->getZExtValue() >> 3);
176 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
178 def SHUFFLE_get_shuf_imm : SDNodeXForm<vector_shuffle, [{
179 return getI8Imm(X86::getShuffleSHUFImmediate(N));
182 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
184 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<vector_shuffle, [{
185 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
188 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
190 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<vector_shuffle, [{
191 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
194 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
195 (vector_shuffle node:$lhs, node:$rhs), [{
196 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
197 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
200 def movddup : PatFrag<(ops node:$lhs, node:$rhs),
201 (vector_shuffle node:$lhs, node:$rhs), [{
202 return X86::isMOVDDUPMask(cast<ShuffleVectorSDNode>(N));
205 def movhlps : PatFrag<(ops node:$lhs, node:$rhs),
206 (vector_shuffle node:$lhs, node:$rhs), [{
207 return X86::isMOVHLPSMask(cast<ShuffleVectorSDNode>(N));
210 def movhlps_undef : PatFrag<(ops node:$lhs, node:$rhs),
211 (vector_shuffle node:$lhs, node:$rhs), [{
212 return X86::isMOVHLPS_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
215 def movhp : PatFrag<(ops node:$lhs, node:$rhs),
216 (vector_shuffle node:$lhs, node:$rhs), [{
217 return X86::isMOVHPMask(cast<ShuffleVectorSDNode>(N));
220 def movlp : PatFrag<(ops node:$lhs, node:$rhs),
221 (vector_shuffle node:$lhs, node:$rhs), [{
222 return X86::isMOVLPMask(cast<ShuffleVectorSDNode>(N));
225 def movl : PatFrag<(ops node:$lhs, node:$rhs),
226 (vector_shuffle node:$lhs, node:$rhs), [{
227 return X86::isMOVLMask(cast<ShuffleVectorSDNode>(N));
230 def movshdup : PatFrag<(ops node:$lhs, node:$rhs),
231 (vector_shuffle node:$lhs, node:$rhs), [{
232 return X86::isMOVSHDUPMask(cast<ShuffleVectorSDNode>(N));
235 def movsldup : PatFrag<(ops node:$lhs, node:$rhs),
236 (vector_shuffle node:$lhs, node:$rhs), [{
237 return X86::isMOVSLDUPMask(cast<ShuffleVectorSDNode>(N));
240 def unpckl : PatFrag<(ops node:$lhs, node:$rhs),
241 (vector_shuffle node:$lhs, node:$rhs), [{
242 return X86::isUNPCKLMask(cast<ShuffleVectorSDNode>(N));
245 def unpckh : PatFrag<(ops node:$lhs, node:$rhs),
246 (vector_shuffle node:$lhs, node:$rhs), [{
247 return X86::isUNPCKHMask(cast<ShuffleVectorSDNode>(N));
250 def unpckl_undef : PatFrag<(ops node:$lhs, node:$rhs),
251 (vector_shuffle node:$lhs, node:$rhs), [{
252 return X86::isUNPCKL_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
255 def unpckh_undef : PatFrag<(ops node:$lhs, node:$rhs),
256 (vector_shuffle node:$lhs, node:$rhs), [{
257 return X86::isUNPCKH_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
260 def pshufd : PatFrag<(ops node:$lhs, node:$rhs),
261 (vector_shuffle node:$lhs, node:$rhs), [{
262 return X86::isPSHUFDMask(cast<ShuffleVectorSDNode>(N));
263 }], SHUFFLE_get_shuf_imm>;
265 def shufp : PatFrag<(ops node:$lhs, node:$rhs),
266 (vector_shuffle node:$lhs, node:$rhs), [{
267 return X86::isSHUFPMask(cast<ShuffleVectorSDNode>(N));
268 }], SHUFFLE_get_shuf_imm>;
270 def pshufhw : PatFrag<(ops node:$lhs, node:$rhs),
271 (vector_shuffle node:$lhs, node:$rhs), [{
272 return X86::isPSHUFHWMask(cast<ShuffleVectorSDNode>(N));
273 }], SHUFFLE_get_pshufhw_imm>;
275 def pshuflw : PatFrag<(ops node:$lhs, node:$rhs),
276 (vector_shuffle node:$lhs, node:$rhs), [{
277 return X86::isPSHUFLWMask(cast<ShuffleVectorSDNode>(N));
278 }], SHUFFLE_get_pshuflw_imm>;
280 //===----------------------------------------------------------------------===//
281 // SSE scalar FP Instructions
282 //===----------------------------------------------------------------------===//
284 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
285 // scheduler into a branch sequence.
286 // These are expanded by the scheduler.
287 let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
288 def CMOV_FR32 : I<0, Pseudo,
289 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
290 "#CMOV_FR32 PSEUDO!",
291 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
293 def CMOV_FR64 : I<0, Pseudo,
294 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
295 "#CMOV_FR64 PSEUDO!",
296 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
298 def CMOV_V4F32 : I<0, Pseudo,
299 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
300 "#CMOV_V4F32 PSEUDO!",
302 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
304 def CMOV_V2F64 : I<0, Pseudo,
305 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
306 "#CMOV_V2F64 PSEUDO!",
308 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
310 def CMOV_V2I64 : I<0, Pseudo,
311 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
312 "#CMOV_V2I64 PSEUDO!",
314 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
318 //===----------------------------------------------------------------------===//
320 //===----------------------------------------------------------------------===//
323 let neverHasSideEffects = 1 in
324 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
325 "movss\t{$src, $dst|$dst, $src}", []>;
326 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
327 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
328 "movss\t{$src, $dst|$dst, $src}",
329 [(set FR32:$dst, (loadf32 addr:$src))]>;
330 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
331 "movss\t{$src, $dst|$dst, $src}",
332 [(store FR32:$src, addr:$dst)]>;
334 // Conversion instructions
335 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
336 "cvttss2si\t{$src, $dst|$dst, $src}",
337 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
338 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
339 "cvttss2si\t{$src, $dst|$dst, $src}",
340 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
341 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
342 "cvtsi2ss\t{$src, $dst|$dst, $src}",
343 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
344 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
345 "cvtsi2ss\t{$src, $dst|$dst, $src}",
346 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
348 // Match intrinsics which expect XMM operand(s).
349 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
350 "cvtss2si\t{$src, $dst|$dst, $src}",
351 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
352 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
353 "cvtss2si\t{$src, $dst|$dst, $src}",
354 [(set GR32:$dst, (int_x86_sse_cvtss2si
355 (load addr:$src)))]>;
357 // Match intrinisics which expect MM and XMM operand(s).
358 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
359 "cvtps2pi\t{$src, $dst|$dst, $src}",
360 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
361 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
362 "cvtps2pi\t{$src, $dst|$dst, $src}",
363 [(set VR64:$dst, (int_x86_sse_cvtps2pi
364 (load addr:$src)))]>;
365 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
366 "cvttps2pi\t{$src, $dst|$dst, $src}",
367 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
368 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
369 "cvttps2pi\t{$src, $dst|$dst, $src}",
370 [(set VR64:$dst, (int_x86_sse_cvttps2pi
371 (load addr:$src)))]>;
372 let Constraints = "$src1 = $dst" in {
373 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
374 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
375 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
376 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
378 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
379 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
380 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
381 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
382 (load addr:$src2)))]>;
385 // Aliases for intrinsics
386 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
387 "cvttss2si\t{$src, $dst|$dst, $src}",
389 (int_x86_sse_cvttss2si VR128:$src))]>;
390 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
391 "cvttss2si\t{$src, $dst|$dst, $src}",
393 (int_x86_sse_cvttss2si(load addr:$src)))]>;
395 let Constraints = "$src1 = $dst" in {
396 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
397 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
398 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
399 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
401 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
402 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
403 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
404 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
405 (loadi32 addr:$src2)))]>;
408 // Comparison instructions
409 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
410 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
411 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
412 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
414 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
415 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
416 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
419 let Defs = [EFLAGS] in {
420 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
421 "ucomiss\t{$src2, $src1|$src1, $src2}",
422 [(X86cmp FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
423 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
424 "ucomiss\t{$src2, $src1|$src1, $src2}",
425 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
429 // Aliases to match intrinsics which expect XMM operand(s).
430 let Constraints = "$src1 = $dst" in {
431 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
432 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
433 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
434 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
435 VR128:$src, imm:$cc))]>;
436 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
437 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src, SSECC:$cc),
438 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
439 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
440 (load addr:$src), imm:$cc))]>;
443 let Defs = [EFLAGS] in {
444 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
445 "ucomiss\t{$src2, $src1|$src1, $src2}",
446 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
448 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
449 "ucomiss\t{$src2, $src1|$src1, $src2}",
450 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
453 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
454 "comiss\t{$src2, $src1|$src1, $src2}",
455 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
457 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
458 "comiss\t{$src2, $src1|$src1, $src2}",
459 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
463 // Aliases of packed SSE1 instructions for scalar use. These all have names that
466 // Alias instructions that map fld0 to pxor for sse.
467 let isReMaterializable = 1, isAsCheapAsAMove = 1 in
468 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
469 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
470 Requires<[HasSSE1]>, TB, OpSize;
472 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
474 let neverHasSideEffects = 1 in
475 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
476 "movaps\t{$src, $dst|$dst, $src}", []>;
478 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
480 let canFoldAsLoad = 1 in
481 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
482 "movaps\t{$src, $dst|$dst, $src}",
483 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
485 // Alias bitwise logical operations using SSE logical ops on packed FP values.
486 let Constraints = "$src1 = $dst" in {
487 let isCommutable = 1 in {
488 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst),
489 (ins FR32:$src1, FR32:$src2),
490 "andps\t{$src2, $dst|$dst, $src2}",
491 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
492 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst),
493 (ins FR32:$src1, FR32:$src2),
494 "orps\t{$src2, $dst|$dst, $src2}",
495 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
496 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst),
497 (ins FR32:$src1, FR32:$src2),
498 "xorps\t{$src2, $dst|$dst, $src2}",
499 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
502 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst),
503 (ins FR32:$src1, f128mem:$src2),
504 "andps\t{$src2, $dst|$dst, $src2}",
505 [(set FR32:$dst, (X86fand FR32:$src1,
506 (memopfsf32 addr:$src2)))]>;
507 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst),
508 (ins FR32:$src1, f128mem:$src2),
509 "orps\t{$src2, $dst|$dst, $src2}",
510 [(set FR32:$dst, (X86for FR32:$src1,
511 (memopfsf32 addr:$src2)))]>;
512 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst),
513 (ins FR32:$src1, f128mem:$src2),
514 "xorps\t{$src2, $dst|$dst, $src2}",
515 [(set FR32:$dst, (X86fxor FR32:$src1,
516 (memopfsf32 addr:$src2)))]>;
518 let neverHasSideEffects = 1 in {
519 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
520 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
521 "andnps\t{$src2, $dst|$dst, $src2}", []>;
523 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
524 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
525 "andnps\t{$src2, $dst|$dst, $src2}", []>;
529 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
531 /// In addition, we also have a special variant of the scalar form here to
532 /// represent the associated intrinsic operation. This form is unlike the
533 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
534 /// and leaves the top elements unmodified (therefore these cannot be commuted).
536 /// These three forms can each be reg+reg or reg+mem, so there are a total of
537 /// six "instructions".
539 let Constraints = "$src1 = $dst" in {
540 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
541 SDNode OpNode, Intrinsic F32Int,
542 bit Commutable = 0> {
543 // Scalar operation, reg+reg.
544 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
545 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
546 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
547 let isCommutable = Commutable;
550 // Scalar operation, reg+mem.
551 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
552 (ins FR32:$src1, f32mem:$src2),
553 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
554 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
556 // Vector operation, reg+reg.
557 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
558 (ins VR128:$src1, VR128:$src2),
559 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
560 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
561 let isCommutable = Commutable;
564 // Vector operation, reg+mem.
565 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
566 (ins VR128:$src1, f128mem:$src2),
567 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
568 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
570 // Intrinsic operation, reg+reg.
571 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
572 (ins VR128:$src1, VR128:$src2),
573 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
574 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]>;
576 // Intrinsic operation, reg+mem.
577 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
578 (ins VR128:$src1, ssmem:$src2),
579 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
580 [(set VR128:$dst, (F32Int VR128:$src1,
581 sse_load_f32:$src2))]>;
585 // Arithmetic instructions
586 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
587 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
588 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
589 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
591 /// sse1_fp_binop_rm - Other SSE1 binops
593 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
594 /// instructions for a full-vector intrinsic form. Operations that map
595 /// onto C operators don't use this form since they just use the plain
596 /// vector form instead of having a separate vector intrinsic form.
598 /// This provides a total of eight "instructions".
600 let Constraints = "$src1 = $dst" in {
601 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
605 bit Commutable = 0> {
607 // Scalar operation, reg+reg.
608 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
609 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
610 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
611 let isCommutable = Commutable;
614 // Scalar operation, reg+mem.
615 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
616 (ins FR32:$src1, f32mem:$src2),
617 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
618 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
620 // Vector operation, reg+reg.
621 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
622 (ins VR128:$src1, VR128:$src2),
623 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
624 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
625 let isCommutable = Commutable;
628 // Vector operation, reg+mem.
629 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
630 (ins VR128:$src1, f128mem:$src2),
631 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
632 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
634 // Intrinsic operation, reg+reg.
635 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
636 (ins VR128:$src1, VR128:$src2),
637 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
638 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
639 let isCommutable = Commutable;
642 // Intrinsic operation, reg+mem.
643 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
644 (ins VR128:$src1, ssmem:$src2),
645 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
646 [(set VR128:$dst, (F32Int VR128:$src1,
647 sse_load_f32:$src2))]>;
649 // Vector intrinsic operation, reg+reg.
650 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
651 (ins VR128:$src1, VR128:$src2),
652 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
653 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
654 let isCommutable = Commutable;
657 // Vector intrinsic operation, reg+mem.
658 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
659 (ins VR128:$src1, f128mem:$src2),
660 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
661 [(set VR128:$dst, (V4F32Int VR128:$src1, (memopv4f32 addr:$src2)))]>;
665 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
666 int_x86_sse_max_ss, int_x86_sse_max_ps>;
667 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
668 int_x86_sse_min_ss, int_x86_sse_min_ps>;
670 //===----------------------------------------------------------------------===//
671 // SSE packed FP Instructions
674 let neverHasSideEffects = 1 in
675 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
676 "movaps\t{$src, $dst|$dst, $src}", []>;
677 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
678 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
679 "movaps\t{$src, $dst|$dst, $src}",
680 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
682 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
683 "movaps\t{$src, $dst|$dst, $src}",
684 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
686 let neverHasSideEffects = 1 in
687 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
688 "movups\t{$src, $dst|$dst, $src}", []>;
689 let canFoldAsLoad = 1 in
690 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
691 "movups\t{$src, $dst|$dst, $src}",
692 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
693 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
694 "movups\t{$src, $dst|$dst, $src}",
695 [(store (v4f32 VR128:$src), addr:$dst)]>;
697 // Intrinsic forms of MOVUPS load and store
698 let canFoldAsLoad = 1 in
699 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
700 "movups\t{$src, $dst|$dst, $src}",
701 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
702 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
703 "movups\t{$src, $dst|$dst, $src}",
704 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
706 let Constraints = "$src1 = $dst" in {
707 let AddedComplexity = 20 in {
708 def MOVLPSrm : PSI<0x12, MRMSrcMem,
709 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
710 "movlps\t{$src2, $dst|$dst, $src2}",
713 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
714 def MOVHPSrm : PSI<0x16, MRMSrcMem,
715 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
716 "movhps\t{$src2, $dst|$dst, $src2}",
719 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
721 } // Constraints = "$src1 = $dst"
724 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
725 "movlps\t{$src, $dst|$dst, $src}",
726 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
727 (iPTR 0))), addr:$dst)]>;
729 // v2f64 extract element 1 is always custom lowered to unpack high to low
730 // and extract element 0 so the non-store version isn't too horrible.
731 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
732 "movhps\t{$src, $dst|$dst, $src}",
733 [(store (f64 (vector_extract
734 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
735 (undef)), (iPTR 0))), addr:$dst)]>;
737 let Constraints = "$src1 = $dst" in {
738 let AddedComplexity = 20 in {
739 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
740 "movlhps\t{$src2, $dst|$dst, $src2}",
742 (v4f32 (movhp VR128:$src1, VR128:$src2)))]>;
744 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
745 "movhlps\t{$src2, $dst|$dst, $src2}",
747 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
749 } // Constraints = "$src1 = $dst"
751 let AddedComplexity = 20 in
752 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
753 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
760 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
762 /// In addition, we also have a special variant of the scalar form here to
763 /// represent the associated intrinsic operation. This form is unlike the
764 /// plain scalar form, in that it takes an entire vector (instead of a
765 /// scalar) and leaves the top elements undefined.
767 /// And, we have a special variant form for a full-vector intrinsic form.
769 /// These four forms can each have a reg or a mem operand, so there are a
770 /// total of eight "instructions".
772 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
776 bit Commutable = 0> {
777 // Scalar operation, reg.
778 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
779 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
780 [(set FR32:$dst, (OpNode FR32:$src))]> {
781 let isCommutable = Commutable;
784 // Scalar operation, mem.
785 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
786 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
787 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
789 // Vector operation, reg.
790 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
791 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
792 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
793 let isCommutable = Commutable;
796 // Vector operation, mem.
797 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
798 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
799 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
801 // Intrinsic operation, reg.
802 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
803 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
804 [(set VR128:$dst, (F32Int VR128:$src))]> {
805 let isCommutable = Commutable;
808 // Intrinsic operation, mem.
809 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
810 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
811 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
813 // Vector intrinsic operation, reg
814 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
815 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
816 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
817 let isCommutable = Commutable;
820 // Vector intrinsic operation, mem
821 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
822 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
823 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
827 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
828 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
830 // Reciprocal approximations. Note that these typically require refinement
831 // in order to obtain suitable precision.
832 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
833 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
834 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
835 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
838 let Constraints = "$src1 = $dst" in {
839 let isCommutable = 1 in {
840 def ANDPSrr : PSI<0x54, MRMSrcReg,
841 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
842 "andps\t{$src2, $dst|$dst, $src2}",
843 [(set VR128:$dst, (v2i64
844 (and VR128:$src1, VR128:$src2)))]>;
845 def ORPSrr : PSI<0x56, MRMSrcReg,
846 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
847 "orps\t{$src2, $dst|$dst, $src2}",
848 [(set VR128:$dst, (v2i64
849 (or VR128:$src1, VR128:$src2)))]>;
850 def XORPSrr : PSI<0x57, MRMSrcReg,
851 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
852 "xorps\t{$src2, $dst|$dst, $src2}",
853 [(set VR128:$dst, (v2i64
854 (xor VR128:$src1, VR128:$src2)))]>;
857 def ANDPSrm : PSI<0x54, MRMSrcMem,
858 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
859 "andps\t{$src2, $dst|$dst, $src2}",
860 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
861 (memopv2i64 addr:$src2)))]>;
862 def ORPSrm : PSI<0x56, MRMSrcMem,
863 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
864 "orps\t{$src2, $dst|$dst, $src2}",
865 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
866 (memopv2i64 addr:$src2)))]>;
867 def XORPSrm : PSI<0x57, MRMSrcMem,
868 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
869 "xorps\t{$src2, $dst|$dst, $src2}",
870 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
871 (memopv2i64 addr:$src2)))]>;
872 def ANDNPSrr : PSI<0x55, MRMSrcReg,
873 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
874 "andnps\t{$src2, $dst|$dst, $src2}",
876 (v2i64 (and (xor VR128:$src1,
877 (bc_v2i64 (v4i32 immAllOnesV))),
879 def ANDNPSrm : PSI<0x55, MRMSrcMem,
880 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
881 "andnps\t{$src2, $dst|$dst, $src2}",
883 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
884 (bc_v2i64 (v4i32 immAllOnesV))),
885 (memopv2i64 addr:$src2))))]>;
888 let Constraints = "$src1 = $dst" in {
889 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
890 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
891 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
892 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
893 VR128:$src, imm:$cc))]>;
894 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
895 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
896 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
897 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
898 (memop addr:$src), imm:$cc))]>;
900 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
901 (CMPPSrri VR128:$src1, VR128:$src2, imm:$cc)>;
902 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
903 (CMPPSrmi VR128:$src1, addr:$src2, imm:$cc)>;
905 // Shuffle and unpack instructions
906 let Constraints = "$src1 = $dst" in {
907 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
908 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
909 (outs VR128:$dst), (ins VR128:$src1,
910 VR128:$src2, i8imm:$src3),
911 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
913 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
914 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
915 (outs VR128:$dst), (ins VR128:$src1,
916 f128mem:$src2, i8imm:$src3),
917 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
920 VR128:$src1, (memopv4f32 addr:$src2))))]>;
922 let AddedComplexity = 10 in {
923 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
924 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
925 "unpckhps\t{$src2, $dst|$dst, $src2}",
927 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
928 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
929 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
930 "unpckhps\t{$src2, $dst|$dst, $src2}",
932 (v4f32 (unpckh VR128:$src1,
933 (memopv4f32 addr:$src2))))]>;
935 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
936 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
937 "unpcklps\t{$src2, $dst|$dst, $src2}",
939 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
940 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
941 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
942 "unpcklps\t{$src2, $dst|$dst, $src2}",
944 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
946 } // Constraints = "$src1 = $dst"
949 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
950 "movmskps\t{$src, $dst|$dst, $src}",
951 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
952 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
953 "movmskpd\t{$src, $dst|$dst, $src}",
954 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
956 // Prefetch intrinsic.
957 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
958 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
959 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
960 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
961 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
962 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
963 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
964 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
966 // Non-temporal stores
967 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
968 "movntps\t{$src, $dst|$dst, $src}",
969 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
971 // Load, store, and memory fence
972 def SFENCE : PSI<0xAE, MRM7m, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
975 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
976 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
977 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
978 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
980 // Alias instructions that map zero vector to pxor / xorp* for sse.
981 // We set canFoldAsLoad because this can be converted to a constant-pool
982 // load of an all-zeros value if folding it would be beneficial.
983 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1 in
984 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
986 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
988 let Predicates = [HasSSE1] in {
989 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
990 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
991 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
992 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
993 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
996 // FR32 to 128-bit vector conversion.
997 let isAsCheapAsAMove = 1 in
998 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
999 "movss\t{$src, $dst|$dst, $src}",
1001 (v4f32 (scalar_to_vector FR32:$src)))]>;
1002 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1003 "movss\t{$src, $dst|$dst, $src}",
1005 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
1007 // FIXME: may not be able to eliminate this movss with coalescing the src and
1008 // dest register classes are different. We really want to write this pattern
1010 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1011 // (f32 FR32:$src)>;
1012 let isAsCheapAsAMove = 1 in
1013 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
1014 "movss\t{$src, $dst|$dst, $src}",
1015 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
1017 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
1018 "movss\t{$src, $dst|$dst, $src}",
1019 [(store (f32 (vector_extract (v4f32 VR128:$src),
1020 (iPTR 0))), addr:$dst)]>;
1023 // Move to lower bits of a VR128, leaving upper bits alone.
1024 // Three operand (but two address) aliases.
1025 let Constraints = "$src1 = $dst" in {
1026 let neverHasSideEffects = 1 in
1027 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
1028 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
1029 "movss\t{$src2, $dst|$dst, $src2}", []>;
1031 let AddedComplexity = 15 in
1032 def MOVLPSrr : SSI<0x10, MRMSrcReg,
1033 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1034 "movss\t{$src2, $dst|$dst, $src2}",
1036 (v4f32 (movl VR128:$src1, VR128:$src2)))]>;
1039 // Move to lower bits of a VR128 and zeroing upper bits.
1040 // Loading from memory automatically zeroing upper bits.
1041 let AddedComplexity = 20 in
1042 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1043 "movss\t{$src, $dst|$dst, $src}",
1044 [(set VR128:$dst, (v4f32 (X86vzmovl (v4f32 (scalar_to_vector
1045 (loadf32 addr:$src))))))]>;
1047 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
1048 (MOVZSS2PSrm addr:$src)>;
1050 //===----------------------------------------------------------------------===//
1051 // SSE2 Instructions
1052 //===----------------------------------------------------------------------===//
1054 // Move Instructions
1055 let neverHasSideEffects = 1 in
1056 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1057 "movsd\t{$src, $dst|$dst, $src}", []>;
1058 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1059 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1060 "movsd\t{$src, $dst|$dst, $src}",
1061 [(set FR64:$dst, (loadf64 addr:$src))]>;
1062 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1063 "movsd\t{$src, $dst|$dst, $src}",
1064 [(store FR64:$src, addr:$dst)]>;
1066 // Conversion instructions
1067 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1068 "cvttsd2si\t{$src, $dst|$dst, $src}",
1069 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1070 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1071 "cvttsd2si\t{$src, $dst|$dst, $src}",
1072 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1073 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1074 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1075 [(set FR32:$dst, (fround FR64:$src))]>;
1076 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1077 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1078 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1079 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1080 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1081 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1082 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1083 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1084 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1086 // SSE2 instructions with XS prefix
1087 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1088 "cvtss2sd\t{$src, $dst|$dst, $src}",
1089 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1090 Requires<[HasSSE2]>;
1091 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1092 "cvtss2sd\t{$src, $dst|$dst, $src}",
1093 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1094 Requires<[HasSSE2]>;
1096 // Match intrinsics which expect XMM operand(s).
1097 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1098 "cvtsd2si\t{$src, $dst|$dst, $src}",
1099 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1100 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1101 "cvtsd2si\t{$src, $dst|$dst, $src}",
1102 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1103 (load addr:$src)))]>;
1105 // Match intrinisics which expect MM and XMM operand(s).
1106 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1107 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1108 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1109 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1110 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1111 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1112 (memop addr:$src)))]>;
1113 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1114 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1115 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1116 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1117 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1118 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1119 (memop addr:$src)))]>;
1120 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1121 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1122 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1123 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1124 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1125 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1126 (load addr:$src)))]>;
1128 // Aliases for intrinsics
1129 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1130 "cvttsd2si\t{$src, $dst|$dst, $src}",
1132 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1133 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1134 "cvttsd2si\t{$src, $dst|$dst, $src}",
1135 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1136 (load addr:$src)))]>;
1138 // Comparison instructions
1139 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1140 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1141 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1142 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1144 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1145 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1146 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1149 let Defs = [EFLAGS] in {
1150 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1151 "ucomisd\t{$src2, $src1|$src1, $src2}",
1152 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1153 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1154 "ucomisd\t{$src2, $src1|$src1, $src2}",
1155 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1156 (implicit EFLAGS)]>;
1157 } // Defs = [EFLAGS]
1159 // Aliases to match intrinsics which expect XMM operand(s).
1160 let Constraints = "$src1 = $dst" in {
1161 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1162 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1163 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1164 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1165 VR128:$src, imm:$cc))]>;
1166 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1167 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1168 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1169 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1170 (load addr:$src), imm:$cc))]>;
1173 let Defs = [EFLAGS] in {
1174 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1175 "ucomisd\t{$src2, $src1|$src1, $src2}",
1176 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1177 (implicit EFLAGS)]>;
1178 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1179 "ucomisd\t{$src2, $src1|$src1, $src2}",
1180 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1181 (implicit EFLAGS)]>;
1183 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1184 "comisd\t{$src2, $src1|$src1, $src2}",
1185 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1186 (implicit EFLAGS)]>;
1187 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1188 "comisd\t{$src2, $src1|$src1, $src2}",
1189 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1190 (implicit EFLAGS)]>;
1191 } // Defs = [EFLAGS]
1193 // Aliases of packed SSE2 instructions for scalar use. These all have names that
1196 // Alias instructions that map fld0 to pxor for sse.
1197 let isReMaterializable = 1, isAsCheapAsAMove = 1 in
1198 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1199 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1200 Requires<[HasSSE2]>, TB, OpSize;
1202 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1204 let neverHasSideEffects = 1 in
1205 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1206 "movapd\t{$src, $dst|$dst, $src}", []>;
1208 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1210 let canFoldAsLoad = 1 in
1211 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1212 "movapd\t{$src, $dst|$dst, $src}",
1213 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1215 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1216 let Constraints = "$src1 = $dst" in {
1217 let isCommutable = 1 in {
1218 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1219 (ins FR64:$src1, FR64:$src2),
1220 "andpd\t{$src2, $dst|$dst, $src2}",
1221 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1222 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1223 (ins FR64:$src1, FR64:$src2),
1224 "orpd\t{$src2, $dst|$dst, $src2}",
1225 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1226 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1227 (ins FR64:$src1, FR64:$src2),
1228 "xorpd\t{$src2, $dst|$dst, $src2}",
1229 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1232 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1233 (ins FR64:$src1, f128mem:$src2),
1234 "andpd\t{$src2, $dst|$dst, $src2}",
1235 [(set FR64:$dst, (X86fand FR64:$src1,
1236 (memopfsf64 addr:$src2)))]>;
1237 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1238 (ins FR64:$src1, f128mem:$src2),
1239 "orpd\t{$src2, $dst|$dst, $src2}",
1240 [(set FR64:$dst, (X86for FR64:$src1,
1241 (memopfsf64 addr:$src2)))]>;
1242 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1243 (ins FR64:$src1, f128mem:$src2),
1244 "xorpd\t{$src2, $dst|$dst, $src2}",
1245 [(set FR64:$dst, (X86fxor FR64:$src1,
1246 (memopfsf64 addr:$src2)))]>;
1248 let neverHasSideEffects = 1 in {
1249 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1250 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1251 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1253 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1254 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1255 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1259 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1261 /// In addition, we also have a special variant of the scalar form here to
1262 /// represent the associated intrinsic operation. This form is unlike the
1263 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1264 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1266 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1267 /// six "instructions".
1269 let Constraints = "$src1 = $dst" in {
1270 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1271 SDNode OpNode, Intrinsic F64Int,
1272 bit Commutable = 0> {
1273 // Scalar operation, reg+reg.
1274 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1275 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1276 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1277 let isCommutable = Commutable;
1280 // Scalar operation, reg+mem.
1281 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1282 (ins FR64:$src1, f64mem:$src2),
1283 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1284 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1286 // Vector operation, reg+reg.
1287 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1288 (ins VR128:$src1, VR128:$src2),
1289 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1290 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1291 let isCommutable = Commutable;
1294 // Vector operation, reg+mem.
1295 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1296 (ins VR128:$src1, f128mem:$src2),
1297 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1298 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1300 // Intrinsic operation, reg+reg.
1301 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1302 (ins VR128:$src1, VR128:$src2),
1303 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1304 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]>;
1306 // Intrinsic operation, reg+mem.
1307 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1308 (ins VR128:$src1, sdmem:$src2),
1309 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1310 [(set VR128:$dst, (F64Int VR128:$src1,
1311 sse_load_f64:$src2))]>;
1315 // Arithmetic instructions
1316 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1317 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1318 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1319 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1321 /// sse2_fp_binop_rm - Other SSE2 binops
1323 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1324 /// instructions for a full-vector intrinsic form. Operations that map
1325 /// onto C operators don't use this form since they just use the plain
1326 /// vector form instead of having a separate vector intrinsic form.
1328 /// This provides a total of eight "instructions".
1330 let Constraints = "$src1 = $dst" in {
1331 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1335 bit Commutable = 0> {
1337 // Scalar operation, reg+reg.
1338 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1339 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1340 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1341 let isCommutable = Commutable;
1344 // Scalar operation, reg+mem.
1345 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1346 (ins FR64:$src1, f64mem:$src2),
1347 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1348 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1350 // Vector operation, reg+reg.
1351 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1352 (ins VR128:$src1, VR128:$src2),
1353 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1354 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1355 let isCommutable = Commutable;
1358 // Vector operation, reg+mem.
1359 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1360 (ins VR128:$src1, f128mem:$src2),
1361 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1362 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1364 // Intrinsic operation, reg+reg.
1365 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1366 (ins VR128:$src1, VR128:$src2),
1367 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1368 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1369 let isCommutable = Commutable;
1372 // Intrinsic operation, reg+mem.
1373 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1374 (ins VR128:$src1, sdmem:$src2),
1375 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1376 [(set VR128:$dst, (F64Int VR128:$src1,
1377 sse_load_f64:$src2))]>;
1379 // Vector intrinsic operation, reg+reg.
1380 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1381 (ins VR128:$src1, VR128:$src2),
1382 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1383 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1384 let isCommutable = Commutable;
1387 // Vector intrinsic operation, reg+mem.
1388 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1389 (ins VR128:$src1, f128mem:$src2),
1390 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1391 [(set VR128:$dst, (V2F64Int VR128:$src1,
1392 (memopv2f64 addr:$src2)))]>;
1396 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1397 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1398 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1399 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1401 //===----------------------------------------------------------------------===//
1402 // SSE packed FP Instructions
1404 // Move Instructions
1405 let neverHasSideEffects = 1 in
1406 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1407 "movapd\t{$src, $dst|$dst, $src}", []>;
1408 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1409 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1410 "movapd\t{$src, $dst|$dst, $src}",
1411 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1413 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1414 "movapd\t{$src, $dst|$dst, $src}",
1415 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1417 let neverHasSideEffects = 1 in
1418 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1419 "movupd\t{$src, $dst|$dst, $src}", []>;
1420 let canFoldAsLoad = 1 in
1421 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1422 "movupd\t{$src, $dst|$dst, $src}",
1423 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1424 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1425 "movupd\t{$src, $dst|$dst, $src}",
1426 [(store (v2f64 VR128:$src), addr:$dst)]>;
1428 // Intrinsic forms of MOVUPD load and store
1429 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1430 "movupd\t{$src, $dst|$dst, $src}",
1431 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1432 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1433 "movupd\t{$src, $dst|$dst, $src}",
1434 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1436 let Constraints = "$src1 = $dst" in {
1437 let AddedComplexity = 20 in {
1438 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1439 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1440 "movlpd\t{$src2, $dst|$dst, $src2}",
1442 (v2f64 (movlp VR128:$src1,
1443 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1444 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1445 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1446 "movhpd\t{$src2, $dst|$dst, $src2}",
1448 (v2f64 (movhp VR128:$src1,
1449 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1450 } // AddedComplexity
1451 } // Constraints = "$src1 = $dst"
1453 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1454 "movlpd\t{$src, $dst|$dst, $src}",
1455 [(store (f64 (vector_extract (v2f64 VR128:$src),
1456 (iPTR 0))), addr:$dst)]>;
1458 // v2f64 extract element 1 is always custom lowered to unpack high to low
1459 // and extract element 0 so the non-store version isn't too horrible.
1460 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1461 "movhpd\t{$src, $dst|$dst, $src}",
1462 [(store (f64 (vector_extract
1463 (v2f64 (unpckh VR128:$src, (undef))),
1464 (iPTR 0))), addr:$dst)]>;
1466 // SSE2 instructions without OpSize prefix
1467 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1468 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1469 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1470 TB, Requires<[HasSSE2]>;
1471 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1472 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1473 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1474 (bitconvert (memopv2i64 addr:$src))))]>,
1475 TB, Requires<[HasSSE2]>;
1477 // SSE2 instructions with XS prefix
1478 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1479 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1480 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1481 XS, Requires<[HasSSE2]>;
1482 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1483 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1484 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1485 (bitconvert (memopv2i64 addr:$src))))]>,
1486 XS, Requires<[HasSSE2]>;
1488 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1489 "cvtps2dq\t{$src, $dst|$dst, $src}",
1490 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1491 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1492 "cvtps2dq\t{$src, $dst|$dst, $src}",
1493 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1494 (memop addr:$src)))]>;
1495 // SSE2 packed instructions with XS prefix
1496 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1497 "cvttps2dq\t{$src, $dst|$dst, $src}",
1498 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1499 XS, Requires<[HasSSE2]>;
1500 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1501 "cvttps2dq\t{$src, $dst|$dst, $src}",
1502 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1503 (memop addr:$src)))]>,
1504 XS, Requires<[HasSSE2]>;
1506 // SSE2 packed instructions with XD prefix
1507 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1508 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1509 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1510 XD, Requires<[HasSSE2]>;
1511 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1512 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1513 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1514 (memop addr:$src)))]>,
1515 XD, Requires<[HasSSE2]>;
1517 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1518 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1519 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1520 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1521 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1522 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1523 (memop addr:$src)))]>;
1525 // SSE2 instructions without OpSize prefix
1526 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1527 "cvtps2pd\t{$src, $dst|$dst, $src}",
1528 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1529 TB, Requires<[HasSSE2]>;
1530 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1531 "cvtps2pd\t{$src, $dst|$dst, $src}",
1532 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1533 (load addr:$src)))]>,
1534 TB, Requires<[HasSSE2]>;
1536 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1537 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1538 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1539 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1540 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1541 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1542 (memop addr:$src)))]>;
1544 // Match intrinsics which expect XMM operand(s).
1545 // Aliases for intrinsics
1546 let Constraints = "$src1 = $dst" in {
1547 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1548 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1549 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1550 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1552 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1553 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1554 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1555 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1556 (loadi32 addr:$src2)))]>;
1557 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1558 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1559 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1560 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1562 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1563 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1564 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1565 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1566 (load addr:$src2)))]>;
1567 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1568 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1569 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1570 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1571 VR128:$src2))]>, XS,
1572 Requires<[HasSSE2]>;
1573 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1574 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1575 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1576 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1577 (load addr:$src2)))]>, XS,
1578 Requires<[HasSSE2]>;
1583 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1585 /// In addition, we also have a special variant of the scalar form here to
1586 /// represent the associated intrinsic operation. This form is unlike the
1587 /// plain scalar form, in that it takes an entire vector (instead of a
1588 /// scalar) and leaves the top elements undefined.
1590 /// And, we have a special variant form for a full-vector intrinsic form.
1592 /// These four forms can each have a reg or a mem operand, so there are a
1593 /// total of eight "instructions".
1595 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1599 bit Commutable = 0> {
1600 // Scalar operation, reg.
1601 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1602 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1603 [(set FR64:$dst, (OpNode FR64:$src))]> {
1604 let isCommutable = Commutable;
1607 // Scalar operation, mem.
1608 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1609 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1610 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1612 // Vector operation, reg.
1613 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1614 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1615 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1616 let isCommutable = Commutable;
1619 // Vector operation, mem.
1620 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1621 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1622 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1624 // Intrinsic operation, reg.
1625 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1626 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1627 [(set VR128:$dst, (F64Int VR128:$src))]> {
1628 let isCommutable = Commutable;
1631 // Intrinsic operation, mem.
1632 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1633 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1634 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1636 // Vector intrinsic operation, reg
1637 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1638 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1639 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1640 let isCommutable = Commutable;
1643 // Vector intrinsic operation, mem
1644 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1645 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1646 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1650 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1651 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1653 // There is no f64 version of the reciprocal approximation instructions.
1656 let Constraints = "$src1 = $dst" in {
1657 let isCommutable = 1 in {
1658 def ANDPDrr : PDI<0x54, MRMSrcReg,
1659 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1660 "andpd\t{$src2, $dst|$dst, $src2}",
1662 (and (bc_v2i64 (v2f64 VR128:$src1)),
1663 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1664 def ORPDrr : PDI<0x56, MRMSrcReg,
1665 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1666 "orpd\t{$src2, $dst|$dst, $src2}",
1668 (or (bc_v2i64 (v2f64 VR128:$src1)),
1669 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1670 def XORPDrr : PDI<0x57, MRMSrcReg,
1671 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1672 "xorpd\t{$src2, $dst|$dst, $src2}",
1674 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1675 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1678 def ANDPDrm : PDI<0x54, MRMSrcMem,
1679 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1680 "andpd\t{$src2, $dst|$dst, $src2}",
1682 (and (bc_v2i64 (v2f64 VR128:$src1)),
1683 (memopv2i64 addr:$src2)))]>;
1684 def ORPDrm : PDI<0x56, MRMSrcMem,
1685 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1686 "orpd\t{$src2, $dst|$dst, $src2}",
1688 (or (bc_v2i64 (v2f64 VR128:$src1)),
1689 (memopv2i64 addr:$src2)))]>;
1690 def XORPDrm : PDI<0x57, MRMSrcMem,
1691 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1692 "xorpd\t{$src2, $dst|$dst, $src2}",
1694 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1695 (memopv2i64 addr:$src2)))]>;
1696 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1697 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1698 "andnpd\t{$src2, $dst|$dst, $src2}",
1700 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1701 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1702 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1703 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1704 "andnpd\t{$src2, $dst|$dst, $src2}",
1706 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1707 (memopv2i64 addr:$src2)))]>;
1710 let Constraints = "$src1 = $dst" in {
1711 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1712 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1713 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1714 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1715 VR128:$src, imm:$cc))]>;
1716 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1717 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1718 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1719 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1720 (memop addr:$src), imm:$cc))]>;
1722 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1723 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1724 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1725 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1727 // Shuffle and unpack instructions
1728 let Constraints = "$src1 = $dst" in {
1729 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1730 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1731 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1733 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1734 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1735 (outs VR128:$dst), (ins VR128:$src1,
1736 f128mem:$src2, i8imm:$src3),
1737 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1740 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1742 let AddedComplexity = 10 in {
1743 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1744 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1745 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1747 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1748 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1749 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1750 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1752 (v2f64 (unpckh VR128:$src1,
1753 (memopv2f64 addr:$src2))))]>;
1755 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1756 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1757 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1759 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1760 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1761 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1762 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1764 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1765 } // AddedComplexity
1766 } // Constraints = "$src1 = $dst"
1769 //===----------------------------------------------------------------------===//
1770 // SSE integer instructions
1772 // Move Instructions
1773 let neverHasSideEffects = 1 in
1774 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1775 "movdqa\t{$src, $dst|$dst, $src}", []>;
1776 let canFoldAsLoad = 1, mayLoad = 1 in
1777 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1778 "movdqa\t{$src, $dst|$dst, $src}",
1779 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1781 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1782 "movdqa\t{$src, $dst|$dst, $src}",
1783 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1784 let canFoldAsLoad = 1, mayLoad = 1 in
1785 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1786 "movdqu\t{$src, $dst|$dst, $src}",
1787 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1788 XS, Requires<[HasSSE2]>;
1790 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1791 "movdqu\t{$src, $dst|$dst, $src}",
1792 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1793 XS, Requires<[HasSSE2]>;
1795 // Intrinsic forms of MOVDQU load and store
1796 let canFoldAsLoad = 1 in
1797 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1798 "movdqu\t{$src, $dst|$dst, $src}",
1799 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1800 XS, Requires<[HasSSE2]>;
1801 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1802 "movdqu\t{$src, $dst|$dst, $src}",
1803 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1804 XS, Requires<[HasSSE2]>;
1806 let Constraints = "$src1 = $dst" in {
1808 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1809 bit Commutable = 0> {
1810 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1811 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1812 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1813 let isCommutable = Commutable;
1815 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1816 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1817 [(set VR128:$dst, (IntId VR128:$src1,
1818 (bitconvert (memopv2i64 addr:$src2))))]>;
1821 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1823 Intrinsic IntId, Intrinsic IntId2> {
1824 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1825 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1826 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1827 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1828 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1829 [(set VR128:$dst, (IntId VR128:$src1,
1830 (bitconvert (memopv2i64 addr:$src2))))]>;
1831 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1832 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1833 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1836 /// PDI_binop_rm - Simple SSE2 binary operator.
1837 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1838 ValueType OpVT, bit Commutable = 0> {
1839 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1840 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1841 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1842 let isCommutable = Commutable;
1844 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1845 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1846 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1847 (bitconvert (memopv2i64 addr:$src2)))))]>;
1850 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1852 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1853 /// to collapse (bitconvert VT to VT) into its operand.
1855 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1856 bit Commutable = 0> {
1857 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1858 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1859 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1860 let isCommutable = Commutable;
1862 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1863 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1864 [(set VR128:$dst, (OpNode VR128:$src1,(memopv2i64 addr:$src2)))]>;
1867 } // Constraints = "$src1 = $dst"
1869 // 128-bit Integer Arithmetic
1871 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1872 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1873 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1874 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1876 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1877 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1878 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1879 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1881 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1882 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1883 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1884 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1886 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1887 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1888 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1889 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1891 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1893 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1894 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1895 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1897 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1899 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1900 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1903 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1904 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1905 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1906 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1907 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1910 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
1911 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
1912 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
1913 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
1914 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
1915 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
1917 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
1918 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
1919 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
1920 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
1921 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
1922 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
1924 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
1925 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
1926 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
1927 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
1929 // 128-bit logical shifts.
1930 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1931 def PSLLDQri : PDIi8<0x73, MRM7r,
1932 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1933 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1934 def PSRLDQri : PDIi8<0x73, MRM3r,
1935 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1936 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1937 // PSRADQri doesn't exist in SSE[1-3].
1940 let Predicates = [HasSSE2] in {
1941 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1942 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1943 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1944 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1945 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
1946 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
1947 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
1948 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
1949 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1950 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1952 // Shift up / down and insert zero's.
1953 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
1954 (v2i64 (PSLLDQri VR128:$src, (PSxLDQ_imm imm:$amt)))>;
1955 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
1956 (v2i64 (PSRLDQri VR128:$src, (PSxLDQ_imm imm:$amt)))>;
1960 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1961 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1962 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1964 let Constraints = "$src1 = $dst" in {
1965 def PANDNrr : PDI<0xDF, MRMSrcReg,
1966 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1967 "pandn\t{$src2, $dst|$dst, $src2}",
1968 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1971 def PANDNrm : PDI<0xDF, MRMSrcMem,
1972 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1973 "pandn\t{$src2, $dst|$dst, $src2}",
1974 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1975 (memopv2i64 addr:$src2))))]>;
1978 // SSE2 Integer comparison
1979 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1980 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1981 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1982 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1983 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1984 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1986 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
1987 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
1988 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
1989 (PCMPEQBrm VR128:$src1, addr:$src2)>;
1990 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
1991 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
1992 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
1993 (PCMPEQWrm VR128:$src1, addr:$src2)>;
1994 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
1995 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
1996 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
1997 (PCMPEQDrm VR128:$src1, addr:$src2)>;
1999 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2000 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2001 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2002 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2003 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2004 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2005 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2006 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2007 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2008 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2009 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2010 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2013 // Pack instructions
2014 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2015 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2016 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2018 // Shuffle and unpack instructions
2019 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2020 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2021 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2022 [(set VR128:$dst, (v4i32 (pshufd:$src2
2023 VR128:$src1, (undef))))]>;
2024 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2025 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2026 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2027 [(set VR128:$dst, (v4i32 (pshufd:$src2
2028 (bc_v4i32(memopv2i64 addr:$src1)),
2031 // SSE2 with ImmT == Imm8 and XS prefix.
2032 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2033 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2034 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2035 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2037 XS, Requires<[HasSSE2]>;
2038 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2039 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2040 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2041 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2042 (bc_v8i16 (memopv2i64 addr:$src1)),
2044 XS, Requires<[HasSSE2]>;
2046 // SSE2 with ImmT == Imm8 and XD prefix.
2047 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2048 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2049 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2050 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2052 XD, Requires<[HasSSE2]>;
2053 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2054 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2055 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2056 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2057 (bc_v8i16 (memopv2i64 addr:$src1)),
2059 XD, Requires<[HasSSE2]>;
2062 let Constraints = "$src1 = $dst" in {
2063 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2064 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2065 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2067 (v16i8 (unpckl VR128:$src1, VR128:$src2)))]>;
2068 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2069 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2070 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2072 (unpckl VR128:$src1,
2073 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2074 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2075 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2076 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2078 (v8i16 (unpckl VR128:$src1, VR128:$src2)))]>;
2079 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2080 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2081 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2083 (unpckl VR128:$src1,
2084 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2085 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2086 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2087 "punpckldq\t{$src2, $dst|$dst, $src2}",
2089 (v4i32 (unpckl VR128:$src1, VR128:$src2)))]>;
2090 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2091 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2092 "punpckldq\t{$src2, $dst|$dst, $src2}",
2094 (unpckl VR128:$src1,
2095 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2096 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2097 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2098 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2100 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2101 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2102 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2103 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2105 (v2i64 (unpckl VR128:$src1,
2106 (memopv2i64 addr:$src2))))]>;
2108 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2109 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2110 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2112 (v16i8 (unpckh VR128:$src1, VR128:$src2)))]>;
2113 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2114 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2115 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2117 (unpckh VR128:$src1,
2118 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2119 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2120 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2121 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2123 (v8i16 (unpckh VR128:$src1, VR128:$src2)))]>;
2124 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2125 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2126 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2128 (unpckh VR128:$src1,
2129 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2130 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2131 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2132 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2134 (v4i32 (unpckh VR128:$src1, VR128:$src2)))]>;
2135 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2136 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2137 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2139 (unpckh VR128:$src1,
2140 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2141 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2142 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2143 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2145 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2146 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2147 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2148 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2150 (v2i64 (unpckh VR128:$src1,
2151 (memopv2i64 addr:$src2))))]>;
2155 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2156 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2157 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2158 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2160 let Constraints = "$src1 = $dst" in {
2161 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2162 (outs VR128:$dst), (ins VR128:$src1,
2163 GR32:$src2, i32i8imm:$src3),
2164 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2166 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2167 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2168 (outs VR128:$dst), (ins VR128:$src1,
2169 i16mem:$src2, i32i8imm:$src3),
2170 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2172 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2177 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2178 "pmovmskb\t{$src, $dst|$dst, $src}",
2179 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2181 // Conditional store
2183 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2184 "maskmovdqu\t{$mask, $src|$src, $mask}",
2185 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2188 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2189 "maskmovdqu\t{$mask, $src|$src, $mask}",
2190 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2192 // Non-temporal stores
2193 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2194 "movntpd\t{$src, $dst|$dst, $src}",
2195 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2196 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2197 "movntdq\t{$src, $dst|$dst, $src}",
2198 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2199 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2200 "movnti\t{$src, $dst|$dst, $src}",
2201 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2202 TB, Requires<[HasSSE2]>;
2205 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2206 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2207 TB, Requires<[HasSSE2]>;
2209 // Load, store, and memory fence
2210 def LFENCE : I<0xAE, MRM5r, (outs), (ins),
2211 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2212 def MFENCE : I<0xAE, MRM6r, (outs), (ins),
2213 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2215 //TODO: custom lower this so as to never even generate the noop
2216 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2218 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2219 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2220 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2223 // Alias instructions that map zero vector to pxor / xorp* for sse.
2224 // We set canFoldAsLoad because this can be converted to a constant-pool
2225 // load of an all-ones value if folding it would be beneficial.
2226 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1 in
2227 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2228 "pcmpeqd\t$dst, $dst",
2229 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2231 // FR64 to 128-bit vector conversion.
2232 let isAsCheapAsAMove = 1 in
2233 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2234 "movsd\t{$src, $dst|$dst, $src}",
2236 (v2f64 (scalar_to_vector FR64:$src)))]>;
2237 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2238 "movsd\t{$src, $dst|$dst, $src}",
2240 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2242 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2243 "movd\t{$src, $dst|$dst, $src}",
2245 (v4i32 (scalar_to_vector GR32:$src)))]>;
2246 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2247 "movd\t{$src, $dst|$dst, $src}",
2249 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2251 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2252 "movd\t{$src, $dst|$dst, $src}",
2253 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2255 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2256 "movd\t{$src, $dst|$dst, $src}",
2257 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2259 // SSE2 instructions with XS prefix
2260 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2261 "movq\t{$src, $dst|$dst, $src}",
2263 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2264 Requires<[HasSSE2]>;
2265 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2266 "movq\t{$src, $dst|$dst, $src}",
2267 [(store (i64 (vector_extract (v2i64 VR128:$src),
2268 (iPTR 0))), addr:$dst)]>;
2270 // FIXME: may not be able to eliminate this movss with coalescing the src and
2271 // dest register classes are different. We really want to write this pattern
2273 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2274 // (f32 FR32:$src)>;
2275 let isAsCheapAsAMove = 1 in
2276 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2277 "movsd\t{$src, $dst|$dst, $src}",
2278 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2280 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2281 "movsd\t{$src, $dst|$dst, $src}",
2282 [(store (f64 (vector_extract (v2f64 VR128:$src),
2283 (iPTR 0))), addr:$dst)]>;
2284 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2285 "movd\t{$src, $dst|$dst, $src}",
2286 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2288 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2289 "movd\t{$src, $dst|$dst, $src}",
2290 [(store (i32 (vector_extract (v4i32 VR128:$src),
2291 (iPTR 0))), addr:$dst)]>;
2293 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2294 "movd\t{$src, $dst|$dst, $src}",
2295 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2296 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2297 "movd\t{$src, $dst|$dst, $src}",
2298 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2301 // Move to lower bits of a VR128, leaving upper bits alone.
2302 // Three operand (but two address) aliases.
2303 let Constraints = "$src1 = $dst" in {
2304 let neverHasSideEffects = 1 in
2305 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2306 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2307 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2309 let AddedComplexity = 15 in
2310 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2311 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2312 "movsd\t{$src2, $dst|$dst, $src2}",
2314 (v2f64 (movl VR128:$src1, VR128:$src2)))]>;
2317 // Store / copy lower 64-bits of a XMM register.
2318 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2319 "movq\t{$src, $dst|$dst, $src}",
2320 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2322 // Move to lower bits of a VR128 and zeroing upper bits.
2323 // Loading from memory automatically zeroing upper bits.
2324 let AddedComplexity = 20 in {
2325 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2326 "movsd\t{$src, $dst|$dst, $src}",
2328 (v2f64 (X86vzmovl (v2f64 (scalar_to_vector
2329 (loadf64 addr:$src))))))]>;
2331 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
2332 (MOVZSD2PDrm addr:$src)>;
2333 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
2334 (MOVZSD2PDrm addr:$src)>;
2335 def : Pat<(v2f64 (X86vzload addr:$src)), (MOVZSD2PDrm addr:$src)>;
2338 // movd / movq to XMM register zero-extends
2339 let AddedComplexity = 15 in {
2340 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2341 "movd\t{$src, $dst|$dst, $src}",
2342 [(set VR128:$dst, (v4i32 (X86vzmovl
2343 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2344 // This is X86-64 only.
2345 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2346 "mov{d|q}\t{$src, $dst|$dst, $src}",
2347 [(set VR128:$dst, (v2i64 (X86vzmovl
2348 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2351 let AddedComplexity = 20 in {
2352 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2353 "movd\t{$src, $dst|$dst, $src}",
2355 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2356 (loadi32 addr:$src))))))]>;
2358 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2359 (MOVZDI2PDIrm addr:$src)>;
2360 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2361 (MOVZDI2PDIrm addr:$src)>;
2362 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2363 (MOVZDI2PDIrm addr:$src)>;
2365 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2366 "movq\t{$src, $dst|$dst, $src}",
2368 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2369 (loadi64 addr:$src))))))]>, XS,
2370 Requires<[HasSSE2]>;
2372 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2373 (MOVZQI2PQIrm addr:$src)>;
2374 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2375 (MOVZQI2PQIrm addr:$src)>;
2376 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2379 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2380 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2381 let AddedComplexity = 15 in
2382 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2383 "movq\t{$src, $dst|$dst, $src}",
2384 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2385 XS, Requires<[HasSSE2]>;
2387 let AddedComplexity = 20 in {
2388 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2389 "movq\t{$src, $dst|$dst, $src}",
2390 [(set VR128:$dst, (v2i64 (X86vzmovl
2391 (loadv2i64 addr:$src))))]>,
2392 XS, Requires<[HasSSE2]>;
2394 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2395 (MOVZPQILo2PQIrm addr:$src)>;
2398 //===----------------------------------------------------------------------===//
2399 // SSE3 Instructions
2400 //===----------------------------------------------------------------------===//
2402 // Move Instructions
2403 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2404 "movshdup\t{$src, $dst|$dst, $src}",
2405 [(set VR128:$dst, (v4f32 (movshdup
2406 VR128:$src, (undef))))]>;
2407 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2408 "movshdup\t{$src, $dst|$dst, $src}",
2409 [(set VR128:$dst, (movshdup
2410 (memopv4f32 addr:$src), (undef)))]>;
2412 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2413 "movsldup\t{$src, $dst|$dst, $src}",
2414 [(set VR128:$dst, (v4f32 (movsldup
2415 VR128:$src, (undef))))]>;
2416 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2417 "movsldup\t{$src, $dst|$dst, $src}",
2418 [(set VR128:$dst, (movsldup
2419 (memopv4f32 addr:$src), (undef)))]>;
2421 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2422 "movddup\t{$src, $dst|$dst, $src}",
2423 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2424 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2425 "movddup\t{$src, $dst|$dst, $src}",
2427 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2430 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2432 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2433 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2434 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2438 let Constraints = "$src1 = $dst" in {
2439 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2440 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2441 "addsubps\t{$src2, $dst|$dst, $src2}",
2442 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2444 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2445 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2446 "addsubps\t{$src2, $dst|$dst, $src2}",
2447 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2448 (memop addr:$src2)))]>;
2449 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2450 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2451 "addsubpd\t{$src2, $dst|$dst, $src2}",
2452 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2454 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2455 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2456 "addsubpd\t{$src2, $dst|$dst, $src2}",
2457 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2458 (memop addr:$src2)))]>;
2461 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2462 "lddqu\t{$src, $dst|$dst, $src}",
2463 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2466 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2467 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2468 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2469 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2470 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2471 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2472 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2473 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2474 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2475 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2476 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2477 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2478 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2479 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2480 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2481 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2483 let Constraints = "$src1 = $dst" in {
2484 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2485 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2486 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2487 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2488 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2489 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2490 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2491 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2494 // Thread synchronization
2495 def MONITOR : I<0xC8, RawFrm, (outs), (ins), "monitor",
2496 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2497 def MWAIT : I<0xC9, RawFrm, (outs), (ins), "mwait",
2498 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2500 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2501 let AddedComplexity = 15 in
2502 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2503 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2504 let AddedComplexity = 20 in
2505 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2506 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2508 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2509 let AddedComplexity = 15 in
2510 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2511 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2512 let AddedComplexity = 20 in
2513 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2514 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2516 //===----------------------------------------------------------------------===//
2517 // SSSE3 Instructions
2518 //===----------------------------------------------------------------------===//
2520 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2521 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2522 Intrinsic IntId64, Intrinsic IntId128> {
2523 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2524 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2525 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2527 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2528 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2530 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2532 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2534 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2535 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2538 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2540 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2543 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2546 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2547 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2548 Intrinsic IntId64, Intrinsic IntId128> {
2549 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2551 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2552 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2554 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2556 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2559 (bitconvert (memopv4i16 addr:$src))))]>;
2561 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2563 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2564 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2567 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2569 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2572 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2575 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2576 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2577 Intrinsic IntId64, Intrinsic IntId128> {
2578 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2580 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2581 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2583 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2585 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2588 (bitconvert (memopv2i32 addr:$src))))]>;
2590 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2592 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2593 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2596 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2598 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2601 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2604 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2605 int_x86_ssse3_pabs_b,
2606 int_x86_ssse3_pabs_b_128>;
2607 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2608 int_x86_ssse3_pabs_w,
2609 int_x86_ssse3_pabs_w_128>;
2610 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2611 int_x86_ssse3_pabs_d,
2612 int_x86_ssse3_pabs_d_128>;
2614 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2615 let Constraints = "$src1 = $dst" in {
2616 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2617 Intrinsic IntId64, Intrinsic IntId128,
2618 bit Commutable = 0> {
2619 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2620 (ins VR64:$src1, VR64:$src2),
2621 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2622 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2623 let isCommutable = Commutable;
2625 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2626 (ins VR64:$src1, i64mem:$src2),
2627 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2629 (IntId64 VR64:$src1,
2630 (bitconvert (memopv8i8 addr:$src2))))]>;
2632 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2633 (ins VR128:$src1, VR128:$src2),
2634 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2635 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2637 let isCommutable = Commutable;
2639 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2640 (ins VR128:$src1, i128mem:$src2),
2641 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2643 (IntId128 VR128:$src1,
2644 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2648 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2649 let Constraints = "$src1 = $dst" in {
2650 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2651 Intrinsic IntId64, Intrinsic IntId128,
2652 bit Commutable = 0> {
2653 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2654 (ins VR64:$src1, VR64:$src2),
2655 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2656 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2657 let isCommutable = Commutable;
2659 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2660 (ins VR64:$src1, i64mem:$src2),
2661 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2663 (IntId64 VR64:$src1,
2664 (bitconvert (memopv4i16 addr:$src2))))]>;
2666 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2667 (ins VR128:$src1, VR128:$src2),
2668 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2669 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2671 let isCommutable = Commutable;
2673 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2674 (ins VR128:$src1, i128mem:$src2),
2675 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2677 (IntId128 VR128:$src1,
2678 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2682 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2683 let Constraints = "$src1 = $dst" in {
2684 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2685 Intrinsic IntId64, Intrinsic IntId128,
2686 bit Commutable = 0> {
2687 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2688 (ins VR64:$src1, VR64:$src2),
2689 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2690 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2691 let isCommutable = Commutable;
2693 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2694 (ins VR64:$src1, i64mem:$src2),
2695 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2697 (IntId64 VR64:$src1,
2698 (bitconvert (memopv2i32 addr:$src2))))]>;
2700 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2701 (ins VR128:$src1, VR128:$src2),
2702 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2703 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2705 let isCommutable = Commutable;
2707 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2708 (ins VR128:$src1, i128mem:$src2),
2709 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2711 (IntId128 VR128:$src1,
2712 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2716 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2717 int_x86_ssse3_phadd_w,
2718 int_x86_ssse3_phadd_w_128>;
2719 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2720 int_x86_ssse3_phadd_d,
2721 int_x86_ssse3_phadd_d_128>;
2722 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2723 int_x86_ssse3_phadd_sw,
2724 int_x86_ssse3_phadd_sw_128>;
2725 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2726 int_x86_ssse3_phsub_w,
2727 int_x86_ssse3_phsub_w_128>;
2728 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2729 int_x86_ssse3_phsub_d,
2730 int_x86_ssse3_phsub_d_128>;
2731 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2732 int_x86_ssse3_phsub_sw,
2733 int_x86_ssse3_phsub_sw_128>;
2734 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2735 int_x86_ssse3_pmadd_ub_sw,
2736 int_x86_ssse3_pmadd_ub_sw_128>;
2737 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2738 int_x86_ssse3_pmul_hr_sw,
2739 int_x86_ssse3_pmul_hr_sw_128, 1>;
2740 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2741 int_x86_ssse3_pshuf_b,
2742 int_x86_ssse3_pshuf_b_128>;
2743 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2744 int_x86_ssse3_psign_b,
2745 int_x86_ssse3_psign_b_128>;
2746 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2747 int_x86_ssse3_psign_w,
2748 int_x86_ssse3_psign_w_128>;
2749 defm PSIGND : SS3I_binop_rm_int_32<0x09, "psignd",
2750 int_x86_ssse3_psign_d,
2751 int_x86_ssse3_psign_d_128>;
2753 let Constraints = "$src1 = $dst" in {
2754 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2755 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2756 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2758 (int_x86_ssse3_palign_r
2759 VR64:$src1, VR64:$src2,
2761 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2762 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2763 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2765 (int_x86_ssse3_palign_r
2767 (bitconvert (memopv2i32 addr:$src2)),
2770 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2771 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2772 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2774 (int_x86_ssse3_palign_r_128
2775 VR128:$src1, VR128:$src2,
2776 imm:$src3))]>, OpSize;
2777 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2778 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2779 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2781 (int_x86_ssse3_palign_r_128
2783 (bitconvert (memopv4i32 addr:$src2)),
2784 imm:$src3))]>, OpSize;
2787 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2788 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2789 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2790 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2792 //===----------------------------------------------------------------------===//
2793 // Non-Instruction Patterns
2794 //===----------------------------------------------------------------------===//
2796 // extload f32 -> f64. This matches load+fextend because we have a hack in
2797 // the isel (PreprocessForFPConvert) that can introduce loads after dag combine.
2798 // Since these loads aren't folded into the fextend, we have to match it
2800 let Predicates = [HasSSE2] in
2801 def : Pat<(fextend (loadf32 addr:$src)),
2802 (CVTSS2SDrm addr:$src)>;
2805 let Predicates = [HasSSE2] in {
2806 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2807 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2808 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2809 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2810 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2811 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2812 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2813 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2814 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2815 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2816 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2817 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2818 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2819 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2820 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2821 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2822 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2823 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2824 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2825 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2826 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2827 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2828 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2829 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2830 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2831 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2832 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2833 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2834 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2835 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2838 // Move scalar to XMM zero-extended
2839 // movd to XMM register zero-extends
2840 let AddedComplexity = 15 in {
2841 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2842 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2843 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2844 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2845 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE1]>;
2846 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2847 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2848 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2849 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2852 // Splat v2f64 / v2i64
2853 let AddedComplexity = 10 in {
2854 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
2855 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2856 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
2857 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2858 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
2859 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2860 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
2861 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2864 // Special unary SHUFPSrri case.
2865 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
2866 (SHUFPSrri VR128:$src1, VR128:$src1,
2867 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2868 Requires<[HasSSE1]>;
2869 let AddedComplexity = 5 in
2870 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
2871 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2872 Requires<[HasSSE2]>;
2873 // Special unary SHUFPDrri case.
2874 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
2875 (SHUFPDrri VR128:$src1, VR128:$src1,
2876 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2877 Requires<[HasSSE2]>;
2878 // Special unary SHUFPDrri case.
2879 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
2880 (SHUFPDrri VR128:$src1, VR128:$src1,
2881 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2882 Requires<[HasSSE2]>;
2883 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2884 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
2885 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2886 Requires<[HasSSE2]>;
2888 // Special binary v4i32 shuffle cases with SHUFPS.
2889 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
2890 (SHUFPSrri VR128:$src1, VR128:$src2,
2891 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2892 Requires<[HasSSE2]>;
2893 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
2894 (SHUFPSrmi VR128:$src1, addr:$src2,
2895 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2896 Requires<[HasSSE2]>;
2897 // Special binary v2i64 shuffle cases using SHUFPDrri.
2898 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
2899 (SHUFPDrri VR128:$src1, VR128:$src2,
2900 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2901 Requires<[HasSSE2]>;
2903 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2904 let AddedComplexity = 15 in {
2905 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
2906 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2907 Requires<[OptForSpeed, HasSSE2]>;
2908 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
2909 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2910 Requires<[OptForSpeed, HasSSE2]>;
2912 let AddedComplexity = 10 in {
2913 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
2914 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2915 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
2916 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2917 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
2918 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2919 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
2920 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2923 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2924 let AddedComplexity = 15 in {
2925 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
2926 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2927 Requires<[OptForSpeed, HasSSE2]>;
2928 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
2929 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2930 Requires<[OptForSpeed, HasSSE2]>;
2932 let AddedComplexity = 10 in {
2933 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
2934 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2935 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
2936 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2937 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
2938 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2939 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
2940 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2943 let AddedComplexity = 20 in {
2944 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2945 def : Pat<(v4i32 (movhp VR128:$src1, VR128:$src2)),
2946 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2948 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2949 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
2950 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2952 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2953 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
2954 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2955 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
2956 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2959 let AddedComplexity = 20 in {
2960 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2961 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2962 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
2963 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2964 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
2965 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2966 def : Pat<(v4f32 (movhp VR128:$src1, (load addr:$src2))),
2967 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2968 def : Pat<(v2f64 (movhp VR128:$src1, (load addr:$src2))),
2969 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2971 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
2972 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2973 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
2974 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2975 def : Pat<(v4i32 (movhp VR128:$src1, (load addr:$src2))),
2976 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2977 def : Pat<(v2i64 (movhp VR128:$src1, (load addr:$src2))),
2978 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2981 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
2982 // (store (vector_shuffle (load addr), v2, <0, 1, 4, 5>), addr) using MOVHPS
2983 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
2984 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
2985 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
2986 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2987 def : Pat<(store (v4f32 (movhp (load addr:$src1), VR128:$src2)), addr:$src1),
2988 (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
2989 def : Pat<(store (v2f64 (movhp (load addr:$src1), VR128:$src2)), addr:$src1),
2990 (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2992 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
2994 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
2995 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
2996 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2997 def : Pat<(store (v4i32 (movhp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
2999 (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3000 def : Pat<(store (v2i64 (movhp (load addr:$src1), VR128:$src2)), addr:$src1),
3001 (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3004 let AddedComplexity = 15 in {
3005 // Setting the lowest element in the vector.
3006 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3007 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3008 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3009 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3011 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
3012 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3013 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3014 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3015 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3018 // Set lowest element and zero upper elements.
3019 let AddedComplexity = 15 in
3020 def : Pat<(v2f64 (movl immAllZerosV_bc, VR128:$src)),
3021 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3022 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3023 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3025 // Some special case pandn patterns.
3026 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3028 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3029 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3031 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3032 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3034 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3036 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3037 (memop addr:$src2))),
3038 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3039 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3040 (memop addr:$src2))),
3041 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3042 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3043 (memop addr:$src2))),
3044 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3046 // vector -> vector casts
3047 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3048 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3049 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3050 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3051 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3052 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3053 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3054 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3056 // Use movaps / movups for SSE integer load / store (one byte shorter).
3057 def : Pat<(alignedloadv4i32 addr:$src),
3058 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
3059 def : Pat<(loadv4i32 addr:$src),
3060 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3061 def : Pat<(alignedloadv2i64 addr:$src),
3062 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3063 def : Pat<(loadv2i64 addr:$src),
3064 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3066 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3067 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3068 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3069 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3070 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3071 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3072 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3073 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3074 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3075 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3076 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3077 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3078 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3079 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3080 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3081 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3083 //===----------------------------------------------------------------------===//
3084 // SSE4.1 Instructions
3085 //===----------------------------------------------------------------------===//
3087 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3090 Intrinsic V2F64Int> {
3091 // Intrinsic operation, reg.
3092 // Vector intrinsic operation, reg
3093 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3094 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3095 !strconcat(OpcodeStr,
3096 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3097 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3100 // Vector intrinsic operation, mem
3101 def PSm_Int : SS4AIi8<opcps, MRMSrcMem,
3102 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3103 !strconcat(OpcodeStr,
3104 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3106 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3109 // Vector intrinsic operation, reg
3110 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3111 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3112 !strconcat(OpcodeStr,
3113 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3114 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3117 // Vector intrinsic operation, mem
3118 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3119 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3120 !strconcat(OpcodeStr,
3121 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3123 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3127 let Constraints = "$src1 = $dst" in {
3128 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3132 // Intrinsic operation, reg.
3133 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3135 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3136 !strconcat(OpcodeStr,
3137 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3139 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3142 // Intrinsic operation, mem.
3143 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3145 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3146 !strconcat(OpcodeStr,
3147 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3149 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3152 // Intrinsic operation, reg.
3153 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3155 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3156 !strconcat(OpcodeStr,
3157 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3159 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3162 // Intrinsic operation, mem.
3163 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3165 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3166 !strconcat(OpcodeStr,
3167 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3169 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3174 // FP round - roundss, roundps, roundsd, roundpd
3175 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3176 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3177 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3178 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3180 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3181 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3182 Intrinsic IntId128> {
3183 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3185 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3186 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3187 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3189 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3192 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3195 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3196 int_x86_sse41_phminposuw>;
3198 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3199 let Constraints = "$src1 = $dst" in {
3200 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3201 Intrinsic IntId128, bit Commutable = 0> {
3202 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3203 (ins VR128:$src1, VR128:$src2),
3204 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3205 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3207 let isCommutable = Commutable;
3209 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3210 (ins VR128:$src1, i128mem:$src2),
3211 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3213 (IntId128 VR128:$src1,
3214 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3218 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3219 int_x86_sse41_pcmpeqq, 1>;
3220 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3221 int_x86_sse41_packusdw, 0>;
3222 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3223 int_x86_sse41_pminsb, 1>;
3224 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3225 int_x86_sse41_pminsd, 1>;
3226 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3227 int_x86_sse41_pminud, 1>;
3228 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3229 int_x86_sse41_pminuw, 1>;
3230 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3231 int_x86_sse41_pmaxsb, 1>;
3232 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3233 int_x86_sse41_pmaxsd, 1>;
3234 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3235 int_x86_sse41_pmaxud, 1>;
3236 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3237 int_x86_sse41_pmaxuw, 1>;
3239 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3241 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3242 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3243 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3244 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3246 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3247 let Constraints = "$src1 = $dst" in {
3248 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3249 SDNode OpNode, Intrinsic IntId128,
3250 bit Commutable = 0> {
3251 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3252 (ins VR128:$src1, VR128:$src2),
3253 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3254 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3255 VR128:$src2))]>, OpSize {
3256 let isCommutable = Commutable;
3258 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3259 (ins VR128:$src1, VR128:$src2),
3260 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3261 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3263 let isCommutable = Commutable;
3265 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3266 (ins VR128:$src1, i128mem:$src2),
3267 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3269 (OpNode VR128:$src1, (memop addr:$src2)))]>, OpSize;
3270 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3271 (ins VR128:$src1, i128mem:$src2),
3272 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3274 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3278 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", v4i32, mul,
3279 int_x86_sse41_pmulld, 1>;
3281 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3282 let Constraints = "$src1 = $dst" in {
3283 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3284 Intrinsic IntId128, bit Commutable = 0> {
3285 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3286 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3287 !strconcat(OpcodeStr,
3288 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3290 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3292 let isCommutable = Commutable;
3294 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3295 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3296 !strconcat(OpcodeStr,
3297 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3299 (IntId128 VR128:$src1,
3300 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3305 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3306 int_x86_sse41_blendps, 0>;
3307 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3308 int_x86_sse41_blendpd, 0>;
3309 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3310 int_x86_sse41_pblendw, 0>;
3311 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3312 int_x86_sse41_dpps, 1>;
3313 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3314 int_x86_sse41_dppd, 1>;
3315 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3316 int_x86_sse41_mpsadbw, 1>;
3319 /// SS41I_ternary_int - SSE 4.1 ternary operator
3320 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3321 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3322 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3323 (ins VR128:$src1, VR128:$src2),
3324 !strconcat(OpcodeStr,
3325 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3326 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3329 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3330 (ins VR128:$src1, i128mem:$src2),
3331 !strconcat(OpcodeStr,
3332 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3335 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3339 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3340 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3341 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3344 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3345 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3346 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3347 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3349 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3350 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3352 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3356 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3357 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3358 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3359 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3360 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3361 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3363 // Common patterns involving scalar load.
3364 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3365 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3366 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3367 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3369 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3370 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3371 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3372 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3374 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3375 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3376 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3377 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3379 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3380 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3381 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3382 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3384 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3385 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3386 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3387 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3389 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3390 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3391 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3392 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3395 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3396 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3397 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3398 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3400 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3401 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3403 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3407 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3408 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3409 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3410 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3412 // Common patterns involving scalar load
3413 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3414 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3415 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3416 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3418 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3419 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3420 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3421 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3424 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3425 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3426 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3427 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3429 // Expecting a i16 load any extended to i32 value.
3430 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3431 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3432 [(set VR128:$dst, (IntId (bitconvert
3433 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3437 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3438 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovsxbq", int_x86_sse41_pmovzxbq>;
3440 // Common patterns involving scalar load
3441 def : Pat<(int_x86_sse41_pmovsxbq
3442 (bitconvert (v4i32 (X86vzmovl
3443 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3444 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3446 def : Pat<(int_x86_sse41_pmovzxbq
3447 (bitconvert (v4i32 (X86vzmovl
3448 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3449 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3452 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3453 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3454 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3455 (ins VR128:$src1, i32i8imm:$src2),
3456 !strconcat(OpcodeStr,
3457 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3458 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3460 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3461 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3462 !strconcat(OpcodeStr,
3463 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3466 // There's an AssertZext in the way of writing the store pattern
3467 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3470 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3473 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3474 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3475 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3476 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3477 !strconcat(OpcodeStr,
3478 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3481 // There's an AssertZext in the way of writing the store pattern
3482 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3485 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3488 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3489 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3490 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3491 (ins VR128:$src1, i32i8imm:$src2),
3492 !strconcat(OpcodeStr,
3493 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3495 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3496 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3497 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3498 !strconcat(OpcodeStr,
3499 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3500 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3501 addr:$dst)]>, OpSize;
3504 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3507 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3509 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3510 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3511 (ins VR128:$src1, i32i8imm:$src2),
3512 !strconcat(OpcodeStr,
3513 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3515 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3517 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3518 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3519 !strconcat(OpcodeStr,
3520 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3521 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3522 addr:$dst)]>, OpSize;
3525 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3527 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3528 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3531 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3532 Requires<[HasSSE41]>;
3534 let Constraints = "$src1 = $dst" in {
3535 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3536 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3537 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3538 !strconcat(OpcodeStr,
3539 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3541 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3542 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3543 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3544 !strconcat(OpcodeStr,
3545 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3547 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3548 imm:$src3))]>, OpSize;
3552 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3554 let Constraints = "$src1 = $dst" in {
3555 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3556 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3557 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3558 !strconcat(OpcodeStr,
3559 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3561 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3563 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3564 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3565 !strconcat(OpcodeStr,
3566 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3568 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3569 imm:$src3)))]>, OpSize;
3573 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3575 let Constraints = "$src1 = $dst" in {
3576 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3577 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3578 (ins VR128:$src1, FR32:$src2, i32i8imm:$src3),
3579 !strconcat(OpcodeStr,
3580 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3582 (X86insrtps VR128:$src1, FR32:$src2, imm:$src3))]>, OpSize;
3583 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3584 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3585 !strconcat(OpcodeStr,
3586 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3588 (X86insrtps VR128:$src1, (loadf32 addr:$src2),
3589 imm:$src3))]>, OpSize;
3593 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3595 let Defs = [EFLAGS] in {
3596 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3597 "ptest \t{$src2, $src1|$src1, $src2}", []>, OpSize;
3598 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3599 "ptest \t{$src2, $src1|$src1, $src2}", []>, OpSize;
3602 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3603 "movntdqa\t{$src, $dst|$dst, $src}",
3604 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;
3606 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3607 let Constraints = "$src1 = $dst" in {
3608 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3609 Intrinsic IntId128, bit Commutable = 0> {
3610 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3611 (ins VR128:$src1, VR128:$src2),
3612 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3613 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3615 let isCommutable = Commutable;
3617 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3618 (ins VR128:$src1, i128mem:$src2),
3619 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3621 (IntId128 VR128:$src1,
3622 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3626 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3628 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3629 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3630 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3631 (PCMPGTQrm VR128:$src1, addr:$src2)>;