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, v4f32>, 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 def SDTX86CmpPTest : SDTypeProfile<0, 2, [SDTCisVT<0, v4f32>, SDTCisVT<1, v4f32>]>;
73 def X86ptest : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
75 //===----------------------------------------------------------------------===//
76 // SSE Complex Patterns
77 //===----------------------------------------------------------------------===//
79 // These are 'extloads' from a scalar to the low element of a vector, zeroing
80 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
82 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
83 [SDNPHasChain, SDNPMayLoad]>;
84 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
85 [SDNPHasChain, SDNPMayLoad]>;
87 def ssmem : Operand<v4f32> {
88 let PrintMethod = "printf32mem";
89 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
91 def sdmem : Operand<v2f64> {
92 let PrintMethod = "printf64mem";
93 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
96 //===----------------------------------------------------------------------===//
97 // SSE pattern fragments
98 //===----------------------------------------------------------------------===//
100 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
101 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
102 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
103 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
105 // Like 'store', but always requires vector alignment.
106 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
107 (store node:$val, node:$ptr), [{
108 return cast<StoreSDNode>(N)->getAlignment() >= 16;
111 // Like 'load', but always requires vector alignment.
112 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
113 return cast<LoadSDNode>(N)->getAlignment() >= 16;
116 def alignedloadfsf32 : PatFrag<(ops node:$ptr), (f32 (alignedload node:$ptr))>;
117 def alignedloadfsf64 : PatFrag<(ops node:$ptr), (f64 (alignedload node:$ptr))>;
118 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
119 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
120 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
121 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
123 // Like 'load', but uses special alignment checks suitable for use in
124 // memory operands in most SSE instructions, which are required to
125 // be naturally aligned on some targets but not on others.
126 // FIXME: Actually implement support for targets that don't require the
127 // alignment. This probably wants a subtarget predicate.
128 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
129 return cast<LoadSDNode>(N)->getAlignment() >= 16;
132 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
133 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
134 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
135 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
136 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
137 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
138 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
140 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
142 // FIXME: 8 byte alignment for mmx reads is not required
143 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
144 return cast<LoadSDNode>(N)->getAlignment() >= 8;
147 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
148 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
149 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
150 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
152 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
153 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
154 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
155 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
156 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
157 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
159 def vzmovl_v2i64 : PatFrag<(ops node:$src),
160 (bitconvert (v2i64 (X86vzmovl
161 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
162 def vzmovl_v4i32 : PatFrag<(ops node:$src),
163 (bitconvert (v4i32 (X86vzmovl
164 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
166 def vzload_v2i64 : PatFrag<(ops node:$src),
167 (bitconvert (v2i64 (X86vzload node:$src)))>;
170 def fp32imm0 : PatLeaf<(f32 fpimm), [{
171 return N->isExactlyValue(+0.0);
174 def PSxLDQ_imm : SDNodeXForm<imm, [{
175 // Transformation function: imm >> 3
176 return getI32Imm(N->getZExtValue() >> 3);
179 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
181 def SHUFFLE_get_shuf_imm : SDNodeXForm<vector_shuffle, [{
182 return getI8Imm(X86::getShuffleSHUFImmediate(N));
185 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
187 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<vector_shuffle, [{
188 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
191 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
193 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<vector_shuffle, [{
194 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
197 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
198 (vector_shuffle node:$lhs, node:$rhs), [{
199 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
200 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
203 def movddup : PatFrag<(ops node:$lhs, node:$rhs),
204 (vector_shuffle node:$lhs, node:$rhs), [{
205 return X86::isMOVDDUPMask(cast<ShuffleVectorSDNode>(N));
208 def movhlps : PatFrag<(ops node:$lhs, node:$rhs),
209 (vector_shuffle node:$lhs, node:$rhs), [{
210 return X86::isMOVHLPSMask(cast<ShuffleVectorSDNode>(N));
213 def movhlps_undef : PatFrag<(ops node:$lhs, node:$rhs),
214 (vector_shuffle node:$lhs, node:$rhs), [{
215 return X86::isMOVHLPS_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
218 def movhp : PatFrag<(ops node:$lhs, node:$rhs),
219 (vector_shuffle node:$lhs, node:$rhs), [{
220 return X86::isMOVHPMask(cast<ShuffleVectorSDNode>(N));
223 def movlp : PatFrag<(ops node:$lhs, node:$rhs),
224 (vector_shuffle node:$lhs, node:$rhs), [{
225 return X86::isMOVLPMask(cast<ShuffleVectorSDNode>(N));
228 def movl : PatFrag<(ops node:$lhs, node:$rhs),
229 (vector_shuffle node:$lhs, node:$rhs), [{
230 return X86::isMOVLMask(cast<ShuffleVectorSDNode>(N));
233 def movshdup : PatFrag<(ops node:$lhs, node:$rhs),
234 (vector_shuffle node:$lhs, node:$rhs), [{
235 return X86::isMOVSHDUPMask(cast<ShuffleVectorSDNode>(N));
238 def movsldup : PatFrag<(ops node:$lhs, node:$rhs),
239 (vector_shuffle node:$lhs, node:$rhs), [{
240 return X86::isMOVSLDUPMask(cast<ShuffleVectorSDNode>(N));
243 def unpckl : PatFrag<(ops node:$lhs, node:$rhs),
244 (vector_shuffle node:$lhs, node:$rhs), [{
245 return X86::isUNPCKLMask(cast<ShuffleVectorSDNode>(N));
248 def unpckh : PatFrag<(ops node:$lhs, node:$rhs),
249 (vector_shuffle node:$lhs, node:$rhs), [{
250 return X86::isUNPCKHMask(cast<ShuffleVectorSDNode>(N));
253 def unpckl_undef : PatFrag<(ops node:$lhs, node:$rhs),
254 (vector_shuffle node:$lhs, node:$rhs), [{
255 return X86::isUNPCKL_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
258 def unpckh_undef : PatFrag<(ops node:$lhs, node:$rhs),
259 (vector_shuffle node:$lhs, node:$rhs), [{
260 return X86::isUNPCKH_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
263 def pshufd : PatFrag<(ops node:$lhs, node:$rhs),
264 (vector_shuffle node:$lhs, node:$rhs), [{
265 return X86::isPSHUFDMask(cast<ShuffleVectorSDNode>(N));
266 }], SHUFFLE_get_shuf_imm>;
268 def shufp : PatFrag<(ops node:$lhs, node:$rhs),
269 (vector_shuffle node:$lhs, node:$rhs), [{
270 return X86::isSHUFPMask(cast<ShuffleVectorSDNode>(N));
271 }], SHUFFLE_get_shuf_imm>;
273 def pshufhw : PatFrag<(ops node:$lhs, node:$rhs),
274 (vector_shuffle node:$lhs, node:$rhs), [{
275 return X86::isPSHUFHWMask(cast<ShuffleVectorSDNode>(N));
276 }], SHUFFLE_get_pshufhw_imm>;
278 def pshuflw : PatFrag<(ops node:$lhs, node:$rhs),
279 (vector_shuffle node:$lhs, node:$rhs), [{
280 return X86::isPSHUFLWMask(cast<ShuffleVectorSDNode>(N));
281 }], SHUFFLE_get_pshuflw_imm>;
283 //===----------------------------------------------------------------------===//
284 // SSE scalar FP Instructions
285 //===----------------------------------------------------------------------===//
287 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
288 // scheduler into a branch sequence.
289 // These are expanded by the scheduler.
290 let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
291 def CMOV_FR32 : I<0, Pseudo,
292 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
293 "#CMOV_FR32 PSEUDO!",
294 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
296 def CMOV_FR64 : I<0, Pseudo,
297 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
298 "#CMOV_FR64 PSEUDO!",
299 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
301 def CMOV_V4F32 : I<0, Pseudo,
302 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
303 "#CMOV_V4F32 PSEUDO!",
305 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
307 def CMOV_V2F64 : I<0, Pseudo,
308 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
309 "#CMOV_V2F64 PSEUDO!",
311 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
313 def CMOV_V2I64 : I<0, Pseudo,
314 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
315 "#CMOV_V2I64 PSEUDO!",
317 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
321 //===----------------------------------------------------------------------===//
323 //===----------------------------------------------------------------------===//
326 let neverHasSideEffects = 1 in
327 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
328 "movss\t{$src, $dst|$dst, $src}", []>;
329 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
330 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
331 "movss\t{$src, $dst|$dst, $src}",
332 [(set FR32:$dst, (loadf32 addr:$src))]>;
333 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
334 "movss\t{$src, $dst|$dst, $src}",
335 [(store FR32:$src, addr:$dst)]>;
337 // Conversion instructions
338 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
339 "cvttss2si\t{$src, $dst|$dst, $src}",
340 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
341 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
342 "cvttss2si\t{$src, $dst|$dst, $src}",
343 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
344 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
345 "cvtsi2ss\t{$src, $dst|$dst, $src}",
346 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
347 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
348 "cvtsi2ss\t{$src, $dst|$dst, $src}",
349 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
351 // Match intrinsics which expect XMM operand(s).
352 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
353 "cvtss2si\t{$src, $dst|$dst, $src}",
354 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
355 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
356 "cvtss2si\t{$src, $dst|$dst, $src}",
357 [(set GR32:$dst, (int_x86_sse_cvtss2si
358 (load addr:$src)))]>;
360 // Match intrinisics which expect MM and XMM operand(s).
361 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
362 "cvtps2pi\t{$src, $dst|$dst, $src}",
363 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
364 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
365 "cvtps2pi\t{$src, $dst|$dst, $src}",
366 [(set VR64:$dst, (int_x86_sse_cvtps2pi
367 (load addr:$src)))]>;
368 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
369 "cvttps2pi\t{$src, $dst|$dst, $src}",
370 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
371 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
372 "cvttps2pi\t{$src, $dst|$dst, $src}",
373 [(set VR64:$dst, (int_x86_sse_cvttps2pi
374 (load addr:$src)))]>;
375 let Constraints = "$src1 = $dst" in {
376 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
377 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
378 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
379 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
381 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
382 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
383 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
384 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
385 (load addr:$src2)))]>;
388 // Aliases for intrinsics
389 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
390 "cvttss2si\t{$src, $dst|$dst, $src}",
392 (int_x86_sse_cvttss2si VR128:$src))]>;
393 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
394 "cvttss2si\t{$src, $dst|$dst, $src}",
396 (int_x86_sse_cvttss2si(load addr:$src)))]>;
398 let Constraints = "$src1 = $dst" in {
399 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
400 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
401 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
402 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
404 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
405 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
406 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
407 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
408 (loadi32 addr:$src2)))]>;
411 // Comparison instructions
412 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
413 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
414 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
415 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
417 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
418 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
419 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
422 let Defs = [EFLAGS] in {
423 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
424 "ucomiss\t{$src2, $src1|$src1, $src2}",
425 [(X86cmp FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
426 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
427 "ucomiss\t{$src2, $src1|$src1, $src2}",
428 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
432 // Aliases to match intrinsics which expect XMM operand(s).
433 let Constraints = "$src1 = $dst" in {
434 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
435 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
436 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
437 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
438 VR128:$src, imm:$cc))]>;
439 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
440 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src, SSECC:$cc),
441 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
442 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
443 (load addr:$src), imm:$cc))]>;
446 let Defs = [EFLAGS] in {
447 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
448 "ucomiss\t{$src2, $src1|$src1, $src2}",
449 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
451 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
452 "ucomiss\t{$src2, $src1|$src1, $src2}",
453 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
456 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
457 "comiss\t{$src2, $src1|$src1, $src2}",
458 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
460 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
461 "comiss\t{$src2, $src1|$src1, $src2}",
462 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
466 // Aliases of packed SSE1 instructions for scalar use. These all have names that
469 // Alias instructions that map fld0 to pxor for sse.
470 let isReMaterializable = 1, isAsCheapAsAMove = 1 in
471 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
472 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
473 Requires<[HasSSE1]>, TB, OpSize;
475 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
477 let neverHasSideEffects = 1 in
478 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
479 "movaps\t{$src, $dst|$dst, $src}", []>;
481 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
483 let canFoldAsLoad = 1 in
484 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
485 "movaps\t{$src, $dst|$dst, $src}",
486 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
488 // Alias bitwise logical operations using SSE logical ops on packed FP values.
489 let Constraints = "$src1 = $dst" in {
490 let isCommutable = 1 in {
491 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst),
492 (ins FR32:$src1, FR32:$src2),
493 "andps\t{$src2, $dst|$dst, $src2}",
494 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
495 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst),
496 (ins FR32:$src1, FR32:$src2),
497 "orps\t{$src2, $dst|$dst, $src2}",
498 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
499 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst),
500 (ins FR32:$src1, FR32:$src2),
501 "xorps\t{$src2, $dst|$dst, $src2}",
502 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
505 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst),
506 (ins FR32:$src1, f128mem:$src2),
507 "andps\t{$src2, $dst|$dst, $src2}",
508 [(set FR32:$dst, (X86fand FR32:$src1,
509 (memopfsf32 addr:$src2)))]>;
510 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst),
511 (ins FR32:$src1, f128mem:$src2),
512 "orps\t{$src2, $dst|$dst, $src2}",
513 [(set FR32:$dst, (X86for FR32:$src1,
514 (memopfsf32 addr:$src2)))]>;
515 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst),
516 (ins FR32:$src1, f128mem:$src2),
517 "xorps\t{$src2, $dst|$dst, $src2}",
518 [(set FR32:$dst, (X86fxor FR32:$src1,
519 (memopfsf32 addr:$src2)))]>;
521 let neverHasSideEffects = 1 in {
522 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
523 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
524 "andnps\t{$src2, $dst|$dst, $src2}", []>;
526 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
527 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
528 "andnps\t{$src2, $dst|$dst, $src2}", []>;
532 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
534 /// In addition, we also have a special variant of the scalar form here to
535 /// represent the associated intrinsic operation. This form is unlike the
536 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
537 /// and leaves the top elements unmodified (therefore these cannot be commuted).
539 /// These three forms can each be reg+reg or reg+mem, so there are a total of
540 /// six "instructions".
542 let Constraints = "$src1 = $dst" in {
543 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
544 SDNode OpNode, Intrinsic F32Int,
545 bit Commutable = 0> {
546 // Scalar operation, reg+reg.
547 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
548 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
549 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
550 let isCommutable = Commutable;
553 // Scalar operation, reg+mem.
554 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
555 (ins FR32:$src1, f32mem:$src2),
556 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
557 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
559 // Vector operation, reg+reg.
560 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
561 (ins VR128:$src1, VR128:$src2),
562 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
563 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
564 let isCommutable = Commutable;
567 // Vector operation, reg+mem.
568 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
569 (ins VR128:$src1, f128mem:$src2),
570 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
571 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
573 // Intrinsic operation, reg+reg.
574 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
575 (ins VR128:$src1, VR128:$src2),
576 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
577 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]>;
579 // Intrinsic operation, reg+mem.
580 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
581 (ins VR128:$src1, ssmem:$src2),
582 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
583 [(set VR128:$dst, (F32Int VR128:$src1,
584 sse_load_f32:$src2))]>;
588 // Arithmetic instructions
589 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
590 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
591 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
592 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
594 /// sse1_fp_binop_rm - Other SSE1 binops
596 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
597 /// instructions for a full-vector intrinsic form. Operations that map
598 /// onto C operators don't use this form since they just use the plain
599 /// vector form instead of having a separate vector intrinsic form.
601 /// This provides a total of eight "instructions".
603 let Constraints = "$src1 = $dst" in {
604 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
608 bit Commutable = 0> {
610 // Scalar operation, reg+reg.
611 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
612 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
613 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
614 let isCommutable = Commutable;
617 // Scalar operation, reg+mem.
618 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
619 (ins FR32:$src1, f32mem:$src2),
620 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
621 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
623 // Vector operation, reg+reg.
624 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
625 (ins VR128:$src1, VR128:$src2),
626 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
627 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
628 let isCommutable = Commutable;
631 // Vector operation, reg+mem.
632 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
633 (ins VR128:$src1, f128mem:$src2),
634 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
635 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
637 // Intrinsic operation, reg+reg.
638 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
639 (ins VR128:$src1, VR128:$src2),
640 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
641 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
642 let isCommutable = Commutable;
645 // Intrinsic operation, reg+mem.
646 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
647 (ins VR128:$src1, ssmem:$src2),
648 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
649 [(set VR128:$dst, (F32Int VR128:$src1,
650 sse_load_f32:$src2))]>;
652 // Vector intrinsic operation, reg+reg.
653 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
654 (ins VR128:$src1, VR128:$src2),
655 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
656 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
657 let isCommutable = Commutable;
660 // Vector intrinsic operation, reg+mem.
661 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
662 (ins VR128:$src1, f128mem:$src2),
663 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
664 [(set VR128:$dst, (V4F32Int VR128:$src1, (memopv4f32 addr:$src2)))]>;
668 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
669 int_x86_sse_max_ss, int_x86_sse_max_ps>;
670 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
671 int_x86_sse_min_ss, int_x86_sse_min_ps>;
673 //===----------------------------------------------------------------------===//
674 // SSE packed FP Instructions
677 let neverHasSideEffects = 1 in
678 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
679 "movaps\t{$src, $dst|$dst, $src}", []>;
680 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
681 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
682 "movaps\t{$src, $dst|$dst, $src}",
683 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
685 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
686 "movaps\t{$src, $dst|$dst, $src}",
687 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
689 let neverHasSideEffects = 1 in
690 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
691 "movups\t{$src, $dst|$dst, $src}", []>;
692 let canFoldAsLoad = 1 in
693 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
694 "movups\t{$src, $dst|$dst, $src}",
695 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
696 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
697 "movups\t{$src, $dst|$dst, $src}",
698 [(store (v4f32 VR128:$src), addr:$dst)]>;
700 // Intrinsic forms of MOVUPS load and store
701 let canFoldAsLoad = 1 in
702 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
703 "movups\t{$src, $dst|$dst, $src}",
704 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
705 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
706 "movups\t{$src, $dst|$dst, $src}",
707 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
709 let Constraints = "$src1 = $dst" in {
710 let AddedComplexity = 20 in {
711 def MOVLPSrm : PSI<0x12, MRMSrcMem,
712 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
713 "movlps\t{$src2, $dst|$dst, $src2}",
716 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
717 def MOVHPSrm : PSI<0x16, MRMSrcMem,
718 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
719 "movhps\t{$src2, $dst|$dst, $src2}",
722 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
724 } // Constraints = "$src1 = $dst"
727 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
728 "movlps\t{$src, $dst|$dst, $src}",
729 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
730 (iPTR 0))), addr:$dst)]>;
732 // v2f64 extract element 1 is always custom lowered to unpack high to low
733 // and extract element 0 so the non-store version isn't too horrible.
734 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
735 "movhps\t{$src, $dst|$dst, $src}",
736 [(store (f64 (vector_extract
737 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
738 (undef)), (iPTR 0))), addr:$dst)]>;
740 let Constraints = "$src1 = $dst" in {
741 let AddedComplexity = 20 in {
742 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
743 (ins VR128:$src1, VR128:$src2),
744 "movlhps\t{$src2, $dst|$dst, $src2}",
746 (v4f32 (movhp VR128:$src1, VR128:$src2)))]>;
748 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
749 (ins VR128:$src1, VR128:$src2),
750 "movhlps\t{$src2, $dst|$dst, $src2}",
752 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
754 } // Constraints = "$src1 = $dst"
756 let AddedComplexity = 20 in {
757 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
758 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
759 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
760 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
767 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
769 /// In addition, we also have a special variant of the scalar form here to
770 /// represent the associated intrinsic operation. This form is unlike the
771 /// plain scalar form, in that it takes an entire vector (instead of a
772 /// scalar) and leaves the top elements undefined.
774 /// And, we have a special variant form for a full-vector intrinsic form.
776 /// These four forms can each have a reg or a mem operand, so there are a
777 /// total of eight "instructions".
779 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
783 bit Commutable = 0> {
784 // Scalar operation, reg.
785 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
786 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
787 [(set FR32:$dst, (OpNode FR32:$src))]> {
788 let isCommutable = Commutable;
791 // Scalar operation, mem.
792 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
793 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
794 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
796 // Vector operation, reg.
797 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
798 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
799 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
800 let isCommutable = Commutable;
803 // Vector operation, mem.
804 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
805 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
806 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
808 // Intrinsic operation, reg.
809 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
810 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
811 [(set VR128:$dst, (F32Int VR128:$src))]> {
812 let isCommutable = Commutable;
815 // Intrinsic operation, mem.
816 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
817 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
818 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
820 // Vector intrinsic operation, reg
821 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
822 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
823 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
824 let isCommutable = Commutable;
827 // Vector intrinsic operation, mem
828 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
829 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
830 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
834 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
835 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
837 // Reciprocal approximations. Note that these typically require refinement
838 // in order to obtain suitable precision.
839 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
840 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
841 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
842 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
845 let Constraints = "$src1 = $dst" in {
846 let isCommutable = 1 in {
847 def ANDPSrr : PSI<0x54, MRMSrcReg,
848 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
849 "andps\t{$src2, $dst|$dst, $src2}",
850 [(set VR128:$dst, (v2i64
851 (and VR128:$src1, VR128:$src2)))]>;
852 def ORPSrr : PSI<0x56, MRMSrcReg,
853 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
854 "orps\t{$src2, $dst|$dst, $src2}",
855 [(set VR128:$dst, (v2i64
856 (or VR128:$src1, VR128:$src2)))]>;
857 def XORPSrr : PSI<0x57, MRMSrcReg,
858 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
859 "xorps\t{$src2, $dst|$dst, $src2}",
860 [(set VR128:$dst, (v2i64
861 (xor VR128:$src1, VR128:$src2)))]>;
864 def ANDPSrm : PSI<0x54, MRMSrcMem,
865 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
866 "andps\t{$src2, $dst|$dst, $src2}",
867 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
868 (memopv2i64 addr:$src2)))]>;
869 def ORPSrm : PSI<0x56, MRMSrcMem,
870 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
871 "orps\t{$src2, $dst|$dst, $src2}",
872 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
873 (memopv2i64 addr:$src2)))]>;
874 def XORPSrm : PSI<0x57, MRMSrcMem,
875 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
876 "xorps\t{$src2, $dst|$dst, $src2}",
877 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
878 (memopv2i64 addr:$src2)))]>;
879 def ANDNPSrr : PSI<0x55, MRMSrcReg,
880 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
881 "andnps\t{$src2, $dst|$dst, $src2}",
883 (v2i64 (and (xor VR128:$src1,
884 (bc_v2i64 (v4i32 immAllOnesV))),
886 def ANDNPSrm : PSI<0x55, MRMSrcMem,
887 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
888 "andnps\t{$src2, $dst|$dst, $src2}",
890 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
891 (bc_v2i64 (v4i32 immAllOnesV))),
892 (memopv2i64 addr:$src2))))]>;
895 let Constraints = "$src1 = $dst" in {
896 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
897 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
898 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
899 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
900 VR128:$src, imm:$cc))]>;
901 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
902 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
903 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
904 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
905 (memop addr:$src), imm:$cc))]>;
907 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
908 (CMPPSrri VR128:$src1, VR128:$src2, imm:$cc)>;
909 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
910 (CMPPSrmi VR128:$src1, addr:$src2, imm:$cc)>;
912 // Shuffle and unpack instructions
913 let Constraints = "$src1 = $dst" in {
914 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
915 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
916 (outs VR128:$dst), (ins VR128:$src1,
917 VR128:$src2, i8imm:$src3),
918 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
920 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
921 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
922 (outs VR128:$dst), (ins VR128:$src1,
923 f128mem:$src2, i8imm:$src3),
924 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
927 VR128:$src1, (memopv4f32 addr:$src2))))]>;
929 let AddedComplexity = 10 in {
930 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
931 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
932 "unpckhps\t{$src2, $dst|$dst, $src2}",
934 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
935 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
936 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
937 "unpckhps\t{$src2, $dst|$dst, $src2}",
939 (v4f32 (unpckh VR128:$src1,
940 (memopv4f32 addr:$src2))))]>;
942 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
943 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
944 "unpcklps\t{$src2, $dst|$dst, $src2}",
946 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
947 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
948 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
949 "unpcklps\t{$src2, $dst|$dst, $src2}",
951 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
953 } // Constraints = "$src1 = $dst"
956 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
957 "movmskps\t{$src, $dst|$dst, $src}",
958 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
959 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
960 "movmskpd\t{$src, $dst|$dst, $src}",
961 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
963 // Prefetch intrinsic.
964 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
965 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
966 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
967 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
968 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
969 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
970 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
971 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
973 // Non-temporal stores
974 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
975 "movntps\t{$src, $dst|$dst, $src}",
976 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
978 // Load, store, and memory fence
979 def SFENCE : PSI<0xAE, MRM7r, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
982 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
983 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
984 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
985 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
987 // Alias instructions that map zero vector to pxor / xorp* for sse.
988 // We set canFoldAsLoad because this can be converted to a constant-pool
989 // load of an all-zeros value if folding it would be beneficial.
990 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1 in
991 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
993 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
995 let Predicates = [HasSSE1] in {
996 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
997 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
998 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
999 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
1000 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
1003 // FR32 to 128-bit vector conversion.
1004 let isAsCheapAsAMove = 1 in
1005 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
1006 "movss\t{$src, $dst|$dst, $src}",
1008 (v4f32 (scalar_to_vector FR32:$src)))]>;
1009 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1010 "movss\t{$src, $dst|$dst, $src}",
1012 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
1014 // FIXME: may not be able to eliminate this movss with coalescing the src and
1015 // dest register classes are different. We really want to write this pattern
1017 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1018 // (f32 FR32:$src)>;
1019 let isAsCheapAsAMove = 1 in
1020 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
1021 "movss\t{$src, $dst|$dst, $src}",
1022 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
1024 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
1025 "movss\t{$src, $dst|$dst, $src}",
1026 [(store (f32 (vector_extract (v4f32 VR128:$src),
1027 (iPTR 0))), addr:$dst)]>;
1030 // Move to lower bits of a VR128, leaving upper bits alone.
1031 // Three operand (but two address) aliases.
1032 let Constraints = "$src1 = $dst" in {
1033 let neverHasSideEffects = 1 in
1034 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
1035 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
1036 "movss\t{$src2, $dst|$dst, $src2}", []>;
1038 let AddedComplexity = 15 in
1039 def MOVLPSrr : SSI<0x10, MRMSrcReg,
1040 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1041 "movss\t{$src2, $dst|$dst, $src2}",
1043 (v4f32 (movl VR128:$src1, VR128:$src2)))]>;
1046 // Move to lower bits of a VR128 and zeroing upper bits.
1047 // Loading from memory automatically zeroing upper bits.
1048 let AddedComplexity = 20 in
1049 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1050 "movss\t{$src, $dst|$dst, $src}",
1051 [(set VR128:$dst, (v4f32 (X86vzmovl (v4f32 (scalar_to_vector
1052 (loadf32 addr:$src))))))]>;
1054 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
1055 (MOVZSS2PSrm addr:$src)>;
1057 //===----------------------------------------------------------------------===//
1058 // SSE2 Instructions
1059 //===----------------------------------------------------------------------===//
1061 // Move Instructions
1062 let neverHasSideEffects = 1 in
1063 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1064 "movsd\t{$src, $dst|$dst, $src}", []>;
1065 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1066 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1067 "movsd\t{$src, $dst|$dst, $src}",
1068 [(set FR64:$dst, (loadf64 addr:$src))]>;
1069 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1070 "movsd\t{$src, $dst|$dst, $src}",
1071 [(store FR64:$src, addr:$dst)]>;
1073 // Conversion instructions
1074 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1075 "cvttsd2si\t{$src, $dst|$dst, $src}",
1076 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1077 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1078 "cvttsd2si\t{$src, $dst|$dst, $src}",
1079 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1080 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1081 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1082 [(set FR32:$dst, (fround FR64:$src))]>;
1083 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1084 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1085 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1086 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1087 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1088 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1089 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1090 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1091 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1093 // SSE2 instructions with XS prefix
1094 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1095 "cvtss2sd\t{$src, $dst|$dst, $src}",
1096 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1097 Requires<[HasSSE2]>;
1098 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1099 "cvtss2sd\t{$src, $dst|$dst, $src}",
1100 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1101 Requires<[HasSSE2]>;
1103 // Match intrinsics which expect XMM operand(s).
1104 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1105 "cvtsd2si\t{$src, $dst|$dst, $src}",
1106 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1107 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1108 "cvtsd2si\t{$src, $dst|$dst, $src}",
1109 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1110 (load addr:$src)))]>;
1112 // Match intrinisics which expect MM and XMM operand(s).
1113 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1114 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1115 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1116 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1117 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1118 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1119 (memop addr:$src)))]>;
1120 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1121 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1122 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1123 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1124 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1125 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1126 (memop addr:$src)))]>;
1127 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1128 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1129 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1130 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1131 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1132 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1133 (load addr:$src)))]>;
1135 // Aliases for intrinsics
1136 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1137 "cvttsd2si\t{$src, $dst|$dst, $src}",
1139 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1140 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1141 "cvttsd2si\t{$src, $dst|$dst, $src}",
1142 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1143 (load addr:$src)))]>;
1145 // Comparison instructions
1146 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1147 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1148 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1149 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1151 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1152 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1153 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1156 let Defs = [EFLAGS] in {
1157 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1158 "ucomisd\t{$src2, $src1|$src1, $src2}",
1159 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1160 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1161 "ucomisd\t{$src2, $src1|$src1, $src2}",
1162 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1163 (implicit EFLAGS)]>;
1164 } // Defs = [EFLAGS]
1166 // Aliases to match intrinsics which expect XMM operand(s).
1167 let Constraints = "$src1 = $dst" in {
1168 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1169 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1170 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1171 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1172 VR128:$src, imm:$cc))]>;
1173 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1174 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1175 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1176 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1177 (load addr:$src), imm:$cc))]>;
1180 let Defs = [EFLAGS] in {
1181 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1182 "ucomisd\t{$src2, $src1|$src1, $src2}",
1183 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1184 (implicit EFLAGS)]>;
1185 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1186 "ucomisd\t{$src2, $src1|$src1, $src2}",
1187 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1188 (implicit EFLAGS)]>;
1190 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1191 "comisd\t{$src2, $src1|$src1, $src2}",
1192 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1193 (implicit EFLAGS)]>;
1194 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1195 "comisd\t{$src2, $src1|$src1, $src2}",
1196 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1197 (implicit EFLAGS)]>;
1198 } // Defs = [EFLAGS]
1200 // Aliases of packed SSE2 instructions for scalar use. These all have names that
1203 // Alias instructions that map fld0 to pxor for sse.
1204 let isReMaterializable = 1, isAsCheapAsAMove = 1 in
1205 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1206 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1207 Requires<[HasSSE2]>, TB, OpSize;
1209 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1211 let neverHasSideEffects = 1 in
1212 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1213 "movapd\t{$src, $dst|$dst, $src}", []>;
1215 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1217 let canFoldAsLoad = 1 in
1218 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1219 "movapd\t{$src, $dst|$dst, $src}",
1220 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1222 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1223 let Constraints = "$src1 = $dst" in {
1224 let isCommutable = 1 in {
1225 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1226 (ins FR64:$src1, FR64:$src2),
1227 "andpd\t{$src2, $dst|$dst, $src2}",
1228 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1229 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1230 (ins FR64:$src1, FR64:$src2),
1231 "orpd\t{$src2, $dst|$dst, $src2}",
1232 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1233 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1234 (ins FR64:$src1, FR64:$src2),
1235 "xorpd\t{$src2, $dst|$dst, $src2}",
1236 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1239 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1240 (ins FR64:$src1, f128mem:$src2),
1241 "andpd\t{$src2, $dst|$dst, $src2}",
1242 [(set FR64:$dst, (X86fand FR64:$src1,
1243 (memopfsf64 addr:$src2)))]>;
1244 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1245 (ins FR64:$src1, f128mem:$src2),
1246 "orpd\t{$src2, $dst|$dst, $src2}",
1247 [(set FR64:$dst, (X86for FR64:$src1,
1248 (memopfsf64 addr:$src2)))]>;
1249 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1250 (ins FR64:$src1, f128mem:$src2),
1251 "xorpd\t{$src2, $dst|$dst, $src2}",
1252 [(set FR64:$dst, (X86fxor FR64:$src1,
1253 (memopfsf64 addr:$src2)))]>;
1255 let neverHasSideEffects = 1 in {
1256 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1257 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1258 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1260 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1261 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1262 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1266 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1268 /// In addition, we also have a special variant of the scalar form here to
1269 /// represent the associated intrinsic operation. This form is unlike the
1270 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1271 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1273 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1274 /// six "instructions".
1276 let Constraints = "$src1 = $dst" in {
1277 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1278 SDNode OpNode, Intrinsic F64Int,
1279 bit Commutable = 0> {
1280 // Scalar operation, reg+reg.
1281 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1282 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1283 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1284 let isCommutable = Commutable;
1287 // Scalar operation, reg+mem.
1288 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1289 (ins FR64:$src1, f64mem:$src2),
1290 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1291 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1293 // Vector operation, reg+reg.
1294 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1295 (ins VR128:$src1, VR128:$src2),
1296 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1297 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1298 let isCommutable = Commutable;
1301 // Vector operation, reg+mem.
1302 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1303 (ins VR128:$src1, f128mem:$src2),
1304 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1305 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1307 // Intrinsic operation, reg+reg.
1308 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1309 (ins VR128:$src1, VR128:$src2),
1310 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1311 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]>;
1313 // Intrinsic operation, reg+mem.
1314 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1315 (ins VR128:$src1, sdmem:$src2),
1316 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1317 [(set VR128:$dst, (F64Int VR128:$src1,
1318 sse_load_f64:$src2))]>;
1322 // Arithmetic instructions
1323 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1324 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1325 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1326 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1328 /// sse2_fp_binop_rm - Other SSE2 binops
1330 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1331 /// instructions for a full-vector intrinsic form. Operations that map
1332 /// onto C operators don't use this form since they just use the plain
1333 /// vector form instead of having a separate vector intrinsic form.
1335 /// This provides a total of eight "instructions".
1337 let Constraints = "$src1 = $dst" in {
1338 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1342 bit Commutable = 0> {
1344 // Scalar operation, reg+reg.
1345 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1346 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1347 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1348 let isCommutable = Commutable;
1351 // Scalar operation, reg+mem.
1352 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1353 (ins FR64:$src1, f64mem:$src2),
1354 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1355 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1357 // Vector operation, reg+reg.
1358 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1359 (ins VR128:$src1, VR128:$src2),
1360 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1361 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1362 let isCommutable = Commutable;
1365 // Vector operation, reg+mem.
1366 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1367 (ins VR128:$src1, f128mem:$src2),
1368 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1369 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1371 // Intrinsic operation, reg+reg.
1372 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1373 (ins VR128:$src1, VR128:$src2),
1374 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1375 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1376 let isCommutable = Commutable;
1379 // Intrinsic operation, reg+mem.
1380 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1381 (ins VR128:$src1, sdmem:$src2),
1382 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1383 [(set VR128:$dst, (F64Int VR128:$src1,
1384 sse_load_f64:$src2))]>;
1386 // Vector intrinsic operation, reg+reg.
1387 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1388 (ins VR128:$src1, VR128:$src2),
1389 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1390 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1391 let isCommutable = Commutable;
1394 // Vector intrinsic operation, reg+mem.
1395 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1396 (ins VR128:$src1, f128mem:$src2),
1397 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1398 [(set VR128:$dst, (V2F64Int VR128:$src1,
1399 (memopv2f64 addr:$src2)))]>;
1403 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1404 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1405 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1406 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1408 //===----------------------------------------------------------------------===//
1409 // SSE packed FP Instructions
1411 // Move Instructions
1412 let neverHasSideEffects = 1 in
1413 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1414 "movapd\t{$src, $dst|$dst, $src}", []>;
1415 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1416 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1417 "movapd\t{$src, $dst|$dst, $src}",
1418 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1420 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1421 "movapd\t{$src, $dst|$dst, $src}",
1422 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1424 let neverHasSideEffects = 1 in
1425 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1426 "movupd\t{$src, $dst|$dst, $src}", []>;
1427 let canFoldAsLoad = 1 in
1428 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1429 "movupd\t{$src, $dst|$dst, $src}",
1430 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1431 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1432 "movupd\t{$src, $dst|$dst, $src}",
1433 [(store (v2f64 VR128:$src), addr:$dst)]>;
1435 // Intrinsic forms of MOVUPD load and store
1436 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1437 "movupd\t{$src, $dst|$dst, $src}",
1438 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1439 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1440 "movupd\t{$src, $dst|$dst, $src}",
1441 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1443 let Constraints = "$src1 = $dst" in {
1444 let AddedComplexity = 20 in {
1445 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1446 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1447 "movlpd\t{$src2, $dst|$dst, $src2}",
1449 (v2f64 (movlp VR128:$src1,
1450 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1451 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1452 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1453 "movhpd\t{$src2, $dst|$dst, $src2}",
1455 (v2f64 (movhp VR128:$src1,
1456 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1457 } // AddedComplexity
1458 } // Constraints = "$src1 = $dst"
1460 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1461 "movlpd\t{$src, $dst|$dst, $src}",
1462 [(store (f64 (vector_extract (v2f64 VR128:$src),
1463 (iPTR 0))), addr:$dst)]>;
1465 // v2f64 extract element 1 is always custom lowered to unpack high to low
1466 // and extract element 0 so the non-store version isn't too horrible.
1467 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1468 "movhpd\t{$src, $dst|$dst, $src}",
1469 [(store (f64 (vector_extract
1470 (v2f64 (unpckh VR128:$src, (undef))),
1471 (iPTR 0))), addr:$dst)]>;
1473 // SSE2 instructions without OpSize prefix
1474 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1475 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1476 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1477 TB, Requires<[HasSSE2]>;
1478 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1479 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1480 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1481 (bitconvert (memopv2i64 addr:$src))))]>,
1482 TB, Requires<[HasSSE2]>;
1484 // SSE2 instructions with XS prefix
1485 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1486 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1487 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1488 XS, Requires<[HasSSE2]>;
1489 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1490 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1491 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1492 (bitconvert (memopv2i64 addr:$src))))]>,
1493 XS, Requires<[HasSSE2]>;
1495 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1496 "cvtps2dq\t{$src, $dst|$dst, $src}",
1497 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1498 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1499 "cvtps2dq\t{$src, $dst|$dst, $src}",
1500 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1501 (memop addr:$src)))]>;
1502 // SSE2 packed instructions with XS prefix
1503 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1504 "cvttps2dq\t{$src, $dst|$dst, $src}",
1505 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1506 XS, Requires<[HasSSE2]>;
1507 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1508 "cvttps2dq\t{$src, $dst|$dst, $src}",
1509 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1510 (memop addr:$src)))]>,
1511 XS, Requires<[HasSSE2]>;
1513 // SSE2 packed instructions with XD prefix
1514 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1515 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1516 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1517 XD, Requires<[HasSSE2]>;
1518 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1519 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1520 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1521 (memop addr:$src)))]>,
1522 XD, Requires<[HasSSE2]>;
1524 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1525 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1526 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1527 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1528 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1529 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1530 (memop addr:$src)))]>;
1532 // SSE2 instructions without OpSize prefix
1533 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1534 "cvtps2pd\t{$src, $dst|$dst, $src}",
1535 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1536 TB, Requires<[HasSSE2]>;
1537 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1538 "cvtps2pd\t{$src, $dst|$dst, $src}",
1539 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1540 (load addr:$src)))]>,
1541 TB, Requires<[HasSSE2]>;
1543 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1544 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1545 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1546 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1547 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1548 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1549 (memop addr:$src)))]>;
1551 // Match intrinsics which expect XMM operand(s).
1552 // Aliases for intrinsics
1553 let Constraints = "$src1 = $dst" in {
1554 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1555 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1556 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1557 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1559 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1560 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1561 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1562 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1563 (loadi32 addr:$src2)))]>;
1564 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1565 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1566 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1567 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1569 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1570 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1571 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1572 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1573 (load addr:$src2)))]>;
1574 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1575 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1576 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1577 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1578 VR128:$src2))]>, XS,
1579 Requires<[HasSSE2]>;
1580 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1581 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1582 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1583 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1584 (load addr:$src2)))]>, XS,
1585 Requires<[HasSSE2]>;
1590 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1592 /// In addition, we also have a special variant of the scalar form here to
1593 /// represent the associated intrinsic operation. This form is unlike the
1594 /// plain scalar form, in that it takes an entire vector (instead of a
1595 /// scalar) and leaves the top elements undefined.
1597 /// And, we have a special variant form for a full-vector intrinsic form.
1599 /// These four forms can each have a reg or a mem operand, so there are a
1600 /// total of eight "instructions".
1602 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1606 bit Commutable = 0> {
1607 // Scalar operation, reg.
1608 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1609 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1610 [(set FR64:$dst, (OpNode FR64:$src))]> {
1611 let isCommutable = Commutable;
1614 // Scalar operation, mem.
1615 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1616 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1617 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1619 // Vector operation, reg.
1620 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1621 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1622 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1623 let isCommutable = Commutable;
1626 // Vector operation, mem.
1627 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1628 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1629 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1631 // Intrinsic operation, reg.
1632 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1633 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1634 [(set VR128:$dst, (F64Int VR128:$src))]> {
1635 let isCommutable = Commutable;
1638 // Intrinsic operation, mem.
1639 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1640 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1641 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1643 // Vector intrinsic operation, reg
1644 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1645 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1646 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1647 let isCommutable = Commutable;
1650 // Vector intrinsic operation, mem
1651 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1652 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1653 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1657 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1658 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1660 // There is no f64 version of the reciprocal approximation instructions.
1663 let Constraints = "$src1 = $dst" in {
1664 let isCommutable = 1 in {
1665 def ANDPDrr : PDI<0x54, MRMSrcReg,
1666 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1667 "andpd\t{$src2, $dst|$dst, $src2}",
1669 (and (bc_v2i64 (v2f64 VR128:$src1)),
1670 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1671 def ORPDrr : PDI<0x56, MRMSrcReg,
1672 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1673 "orpd\t{$src2, $dst|$dst, $src2}",
1675 (or (bc_v2i64 (v2f64 VR128:$src1)),
1676 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1677 def XORPDrr : PDI<0x57, MRMSrcReg,
1678 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1679 "xorpd\t{$src2, $dst|$dst, $src2}",
1681 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1682 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1685 def ANDPDrm : PDI<0x54, MRMSrcMem,
1686 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1687 "andpd\t{$src2, $dst|$dst, $src2}",
1689 (and (bc_v2i64 (v2f64 VR128:$src1)),
1690 (memopv2i64 addr:$src2)))]>;
1691 def ORPDrm : PDI<0x56, MRMSrcMem,
1692 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1693 "orpd\t{$src2, $dst|$dst, $src2}",
1695 (or (bc_v2i64 (v2f64 VR128:$src1)),
1696 (memopv2i64 addr:$src2)))]>;
1697 def XORPDrm : PDI<0x57, MRMSrcMem,
1698 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1699 "xorpd\t{$src2, $dst|$dst, $src2}",
1701 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1702 (memopv2i64 addr:$src2)))]>;
1703 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1704 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1705 "andnpd\t{$src2, $dst|$dst, $src2}",
1707 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1708 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1709 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1710 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1711 "andnpd\t{$src2, $dst|$dst, $src2}",
1713 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1714 (memopv2i64 addr:$src2)))]>;
1717 let Constraints = "$src1 = $dst" in {
1718 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1719 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1720 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1721 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1722 VR128:$src, imm:$cc))]>;
1723 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1724 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1725 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1726 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1727 (memop addr:$src), imm:$cc))]>;
1729 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1730 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1731 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1732 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1734 // Shuffle and unpack instructions
1735 let Constraints = "$src1 = $dst" in {
1736 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1737 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1738 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1740 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1741 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1742 (outs VR128:$dst), (ins VR128:$src1,
1743 f128mem:$src2, i8imm:$src3),
1744 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1747 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1749 let AddedComplexity = 10 in {
1750 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1751 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1752 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1754 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1755 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1756 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1757 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1759 (v2f64 (unpckh VR128:$src1,
1760 (memopv2f64 addr:$src2))))]>;
1762 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1763 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1764 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1766 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1767 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1768 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1769 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1771 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1772 } // AddedComplexity
1773 } // Constraints = "$src1 = $dst"
1776 //===----------------------------------------------------------------------===//
1777 // SSE integer instructions
1779 // Move Instructions
1780 let neverHasSideEffects = 1 in
1781 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1782 "movdqa\t{$src, $dst|$dst, $src}", []>;
1783 let canFoldAsLoad = 1, mayLoad = 1 in
1784 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1785 "movdqa\t{$src, $dst|$dst, $src}",
1786 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1788 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1789 "movdqa\t{$src, $dst|$dst, $src}",
1790 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1791 let canFoldAsLoad = 1, mayLoad = 1 in
1792 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1793 "movdqu\t{$src, $dst|$dst, $src}",
1794 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1795 XS, Requires<[HasSSE2]>;
1797 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1798 "movdqu\t{$src, $dst|$dst, $src}",
1799 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1800 XS, Requires<[HasSSE2]>;
1802 // Intrinsic forms of MOVDQU load and store
1803 let canFoldAsLoad = 1 in
1804 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1805 "movdqu\t{$src, $dst|$dst, $src}",
1806 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1807 XS, Requires<[HasSSE2]>;
1808 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1809 "movdqu\t{$src, $dst|$dst, $src}",
1810 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1811 XS, Requires<[HasSSE2]>;
1813 let Constraints = "$src1 = $dst" in {
1815 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1816 bit Commutable = 0> {
1817 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1818 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1819 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1820 let isCommutable = Commutable;
1822 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1823 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1824 [(set VR128:$dst, (IntId VR128:$src1,
1825 (bitconvert (memopv2i64 addr:$src2))))]>;
1828 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1830 Intrinsic IntId, Intrinsic IntId2> {
1831 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1832 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1833 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1834 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1835 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1836 [(set VR128:$dst, (IntId VR128:$src1,
1837 (bitconvert (memopv2i64 addr:$src2))))]>;
1838 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1839 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1840 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1843 /// PDI_binop_rm - Simple SSE2 binary operator.
1844 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1845 ValueType OpVT, bit Commutable = 0> {
1846 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1847 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1848 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1849 let isCommutable = Commutable;
1851 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1852 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1853 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1854 (bitconvert (memopv2i64 addr:$src2)))))]>;
1857 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1859 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1860 /// to collapse (bitconvert VT to VT) into its operand.
1862 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1863 bit Commutable = 0> {
1864 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1865 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1866 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1867 let isCommutable = Commutable;
1869 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1870 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1871 [(set VR128:$dst, (OpNode VR128:$src1,(memopv2i64 addr:$src2)))]>;
1874 } // Constraints = "$src1 = $dst"
1876 // 128-bit Integer Arithmetic
1878 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1879 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1880 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1881 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1883 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1884 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1885 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1886 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1888 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1889 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1890 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1891 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1893 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1894 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1895 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1896 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1898 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1900 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1901 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1902 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1904 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1906 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1907 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1910 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1911 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1912 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1913 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1914 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
1917 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
1918 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
1919 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
1920 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
1921 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
1922 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
1924 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
1925 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
1926 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
1927 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
1928 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
1929 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
1931 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
1932 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
1933 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
1934 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
1936 // 128-bit logical shifts.
1937 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1938 def PSLLDQri : PDIi8<0x73, MRM7r,
1939 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1940 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1941 def PSRLDQri : PDIi8<0x73, MRM3r,
1942 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1943 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1944 // PSRADQri doesn't exist in SSE[1-3].
1947 let Predicates = [HasSSE2] in {
1948 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1949 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1950 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1951 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1952 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
1953 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
1954 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
1955 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
1956 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1957 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1959 // Shift up / down and insert zero's.
1960 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
1961 (v2i64 (PSLLDQri VR128:$src, (PSxLDQ_imm imm:$amt)))>;
1962 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
1963 (v2i64 (PSRLDQri VR128:$src, (PSxLDQ_imm imm:$amt)))>;
1967 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1968 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1969 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1971 let Constraints = "$src1 = $dst" in {
1972 def PANDNrr : PDI<0xDF, MRMSrcReg,
1973 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1974 "pandn\t{$src2, $dst|$dst, $src2}",
1975 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1978 def PANDNrm : PDI<0xDF, MRMSrcMem,
1979 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1980 "pandn\t{$src2, $dst|$dst, $src2}",
1981 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1982 (memopv2i64 addr:$src2))))]>;
1985 // SSE2 Integer comparison
1986 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1987 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1988 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1989 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1990 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1991 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1993 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
1994 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
1995 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
1996 (PCMPEQBrm VR128:$src1, addr:$src2)>;
1997 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
1998 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
1999 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2000 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2001 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2002 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2003 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2004 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2006 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2007 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2008 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2009 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2010 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2011 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2012 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2013 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2014 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2015 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2016 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2017 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2020 // Pack instructions
2021 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2022 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2023 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2025 // Shuffle and unpack instructions
2026 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2027 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2028 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2029 [(set VR128:$dst, (v4i32 (pshufd:$src2
2030 VR128:$src1, (undef))))]>;
2031 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2032 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2033 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2034 [(set VR128:$dst, (v4i32 (pshufd:$src2
2035 (bc_v4i32(memopv2i64 addr:$src1)),
2038 // SSE2 with ImmT == Imm8 and XS prefix.
2039 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2040 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2041 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2042 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2044 XS, Requires<[HasSSE2]>;
2045 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2046 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2047 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2048 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2049 (bc_v8i16 (memopv2i64 addr:$src1)),
2051 XS, Requires<[HasSSE2]>;
2053 // SSE2 with ImmT == Imm8 and XD prefix.
2054 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2055 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2056 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2057 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2059 XD, Requires<[HasSSE2]>;
2060 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2061 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2062 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2063 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2064 (bc_v8i16 (memopv2i64 addr:$src1)),
2066 XD, Requires<[HasSSE2]>;
2069 let Constraints = "$src1 = $dst" in {
2070 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2071 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2072 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2074 (v16i8 (unpckl VR128:$src1, VR128:$src2)))]>;
2075 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2076 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2077 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2079 (unpckl VR128:$src1,
2080 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2081 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2082 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2083 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2085 (v8i16 (unpckl VR128:$src1, VR128:$src2)))]>;
2086 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2087 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2088 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2090 (unpckl VR128:$src1,
2091 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2092 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2093 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2094 "punpckldq\t{$src2, $dst|$dst, $src2}",
2096 (v4i32 (unpckl VR128:$src1, VR128:$src2)))]>;
2097 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2098 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2099 "punpckldq\t{$src2, $dst|$dst, $src2}",
2101 (unpckl VR128:$src1,
2102 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2103 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2104 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2105 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2107 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2108 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2109 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2110 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2112 (v2i64 (unpckl VR128:$src1,
2113 (memopv2i64 addr:$src2))))]>;
2115 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2116 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2117 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2119 (v16i8 (unpckh VR128:$src1, VR128:$src2)))]>;
2120 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2121 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2122 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2124 (unpckh VR128:$src1,
2125 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2126 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2127 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2128 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2130 (v8i16 (unpckh VR128:$src1, VR128:$src2)))]>;
2131 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2132 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2133 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2135 (unpckh VR128:$src1,
2136 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2137 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2138 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2139 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2141 (v4i32 (unpckh VR128:$src1, VR128:$src2)))]>;
2142 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2143 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2144 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2146 (unpckh VR128:$src1,
2147 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2148 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2149 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2150 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2152 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2153 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2154 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2155 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2157 (v2i64 (unpckh VR128:$src1,
2158 (memopv2i64 addr:$src2))))]>;
2162 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2163 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2164 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2165 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2167 let Constraints = "$src1 = $dst" in {
2168 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2169 (outs VR128:$dst), (ins VR128:$src1,
2170 GR32:$src2, i32i8imm:$src3),
2171 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2173 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2174 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2175 (outs VR128:$dst), (ins VR128:$src1,
2176 i16mem:$src2, i32i8imm:$src3),
2177 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2179 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2184 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2185 "pmovmskb\t{$src, $dst|$dst, $src}",
2186 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2188 // Conditional store
2190 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2191 "maskmovdqu\t{$mask, $src|$src, $mask}",
2192 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2195 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2196 "maskmovdqu\t{$mask, $src|$src, $mask}",
2197 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2199 // Non-temporal stores
2200 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2201 "movntpd\t{$src, $dst|$dst, $src}",
2202 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2203 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2204 "movntdq\t{$src, $dst|$dst, $src}",
2205 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2206 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2207 "movnti\t{$src, $dst|$dst, $src}",
2208 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2209 TB, Requires<[HasSSE2]>;
2212 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2213 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2214 TB, Requires<[HasSSE2]>;
2216 // Load, store, and memory fence
2217 def LFENCE : I<0xAE, MRM5r, (outs), (ins),
2218 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2219 def MFENCE : I<0xAE, MRM6r, (outs), (ins),
2220 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2222 //TODO: custom lower this so as to never even generate the noop
2223 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2225 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2226 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2227 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2230 // Alias instructions that map zero vector to pxor / xorp* for sse.
2231 // We set canFoldAsLoad because this can be converted to a constant-pool
2232 // load of an all-ones value if folding it would be beneficial.
2233 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1 in
2234 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2235 "pcmpeqd\t$dst, $dst",
2236 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2238 // FR64 to 128-bit vector conversion.
2239 let isAsCheapAsAMove = 1 in
2240 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2241 "movsd\t{$src, $dst|$dst, $src}",
2243 (v2f64 (scalar_to_vector FR64:$src)))]>;
2244 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2245 "movsd\t{$src, $dst|$dst, $src}",
2247 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2249 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2250 "movd\t{$src, $dst|$dst, $src}",
2252 (v4i32 (scalar_to_vector GR32:$src)))]>;
2253 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2254 "movd\t{$src, $dst|$dst, $src}",
2256 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2258 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2259 "movd\t{$src, $dst|$dst, $src}",
2260 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2262 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2263 "movd\t{$src, $dst|$dst, $src}",
2264 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2266 // SSE2 instructions with XS prefix
2267 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2268 "movq\t{$src, $dst|$dst, $src}",
2270 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2271 Requires<[HasSSE2]>;
2272 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2273 "movq\t{$src, $dst|$dst, $src}",
2274 [(store (i64 (vector_extract (v2i64 VR128:$src),
2275 (iPTR 0))), addr:$dst)]>;
2277 // FIXME: may not be able to eliminate this movss with coalescing the src and
2278 // dest register classes are different. We really want to write this pattern
2280 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2281 // (f32 FR32:$src)>;
2282 let isAsCheapAsAMove = 1 in
2283 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2284 "movsd\t{$src, $dst|$dst, $src}",
2285 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2287 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2288 "movsd\t{$src, $dst|$dst, $src}",
2289 [(store (f64 (vector_extract (v2f64 VR128:$src),
2290 (iPTR 0))), addr:$dst)]>;
2291 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2292 "movd\t{$src, $dst|$dst, $src}",
2293 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2295 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2296 "movd\t{$src, $dst|$dst, $src}",
2297 [(store (i32 (vector_extract (v4i32 VR128:$src),
2298 (iPTR 0))), addr:$dst)]>;
2300 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2301 "movd\t{$src, $dst|$dst, $src}",
2302 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2303 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2304 "movd\t{$src, $dst|$dst, $src}",
2305 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2308 // Move to lower bits of a VR128, leaving upper bits alone.
2309 // Three operand (but two address) aliases.
2310 let Constraints = "$src1 = $dst" in {
2311 let neverHasSideEffects = 1 in
2312 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2313 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2314 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2316 let AddedComplexity = 15 in
2317 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2318 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2319 "movsd\t{$src2, $dst|$dst, $src2}",
2321 (v2f64 (movl VR128:$src1, VR128:$src2)))]>;
2324 // Store / copy lower 64-bits of a XMM register.
2325 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2326 "movq\t{$src, $dst|$dst, $src}",
2327 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2329 // Move to lower bits of a VR128 and zeroing upper bits.
2330 // Loading from memory automatically zeroing upper bits.
2331 let AddedComplexity = 20 in {
2332 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2333 "movsd\t{$src, $dst|$dst, $src}",
2335 (v2f64 (X86vzmovl (v2f64 (scalar_to_vector
2336 (loadf64 addr:$src))))))]>;
2338 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
2339 (MOVZSD2PDrm addr:$src)>;
2340 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
2341 (MOVZSD2PDrm addr:$src)>;
2342 def : Pat<(v2f64 (X86vzload addr:$src)), (MOVZSD2PDrm addr:$src)>;
2345 // movd / movq to XMM register zero-extends
2346 let AddedComplexity = 15 in {
2347 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2348 "movd\t{$src, $dst|$dst, $src}",
2349 [(set VR128:$dst, (v4i32 (X86vzmovl
2350 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2351 // This is X86-64 only.
2352 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2353 "mov{d|q}\t{$src, $dst|$dst, $src}",
2354 [(set VR128:$dst, (v2i64 (X86vzmovl
2355 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2358 let AddedComplexity = 20 in {
2359 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2360 "movd\t{$src, $dst|$dst, $src}",
2362 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2363 (loadi32 addr:$src))))))]>;
2365 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2366 (MOVZDI2PDIrm addr:$src)>;
2367 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2368 (MOVZDI2PDIrm addr:$src)>;
2369 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2370 (MOVZDI2PDIrm addr:$src)>;
2372 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2373 "movq\t{$src, $dst|$dst, $src}",
2375 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2376 (loadi64 addr:$src))))))]>, XS,
2377 Requires<[HasSSE2]>;
2379 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2380 (MOVZQI2PQIrm addr:$src)>;
2381 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2382 (MOVZQI2PQIrm addr:$src)>;
2383 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2386 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2387 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2388 let AddedComplexity = 15 in
2389 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2390 "movq\t{$src, $dst|$dst, $src}",
2391 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2392 XS, Requires<[HasSSE2]>;
2394 let AddedComplexity = 20 in {
2395 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2396 "movq\t{$src, $dst|$dst, $src}",
2397 [(set VR128:$dst, (v2i64 (X86vzmovl
2398 (loadv2i64 addr:$src))))]>,
2399 XS, Requires<[HasSSE2]>;
2401 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2402 (MOVZPQILo2PQIrm addr:$src)>;
2405 //===----------------------------------------------------------------------===//
2406 // SSE3 Instructions
2407 //===----------------------------------------------------------------------===//
2409 // Move Instructions
2410 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2411 "movshdup\t{$src, $dst|$dst, $src}",
2412 [(set VR128:$dst, (v4f32 (movshdup
2413 VR128:$src, (undef))))]>;
2414 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2415 "movshdup\t{$src, $dst|$dst, $src}",
2416 [(set VR128:$dst, (movshdup
2417 (memopv4f32 addr:$src), (undef)))]>;
2419 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2420 "movsldup\t{$src, $dst|$dst, $src}",
2421 [(set VR128:$dst, (v4f32 (movsldup
2422 VR128:$src, (undef))))]>;
2423 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2424 "movsldup\t{$src, $dst|$dst, $src}",
2425 [(set VR128:$dst, (movsldup
2426 (memopv4f32 addr:$src), (undef)))]>;
2428 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2429 "movddup\t{$src, $dst|$dst, $src}",
2430 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2431 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2432 "movddup\t{$src, $dst|$dst, $src}",
2434 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2437 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2439 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2441 let AddedComplexity = 5 in {
2442 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2443 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2444 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2445 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2446 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2447 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2448 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2449 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2453 let Constraints = "$src1 = $dst" in {
2454 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2455 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2456 "addsubps\t{$src2, $dst|$dst, $src2}",
2457 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2459 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2460 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2461 "addsubps\t{$src2, $dst|$dst, $src2}",
2462 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2463 (memop addr:$src2)))]>;
2464 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2465 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2466 "addsubpd\t{$src2, $dst|$dst, $src2}",
2467 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2469 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2470 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2471 "addsubpd\t{$src2, $dst|$dst, $src2}",
2472 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2473 (memop addr:$src2)))]>;
2476 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2477 "lddqu\t{$src, $dst|$dst, $src}",
2478 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2481 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2482 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2483 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2484 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2485 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2486 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2487 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2488 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2489 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2490 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2491 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2492 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2493 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2494 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2495 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2496 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2498 let Constraints = "$src1 = $dst" in {
2499 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2500 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2501 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2502 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2503 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2504 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2505 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2506 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2509 // Thread synchronization
2510 def MONITOR : I<0x01, MRM1r, (outs), (ins), "monitor",
2511 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2512 def MWAIT : I<0x01, MRM1r, (outs), (ins), "mwait",
2513 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2515 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2516 let AddedComplexity = 15 in
2517 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2518 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2519 let AddedComplexity = 20 in
2520 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2521 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2523 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2524 let AddedComplexity = 15 in
2525 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2526 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2527 let AddedComplexity = 20 in
2528 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2529 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2531 //===----------------------------------------------------------------------===//
2532 // SSSE3 Instructions
2533 //===----------------------------------------------------------------------===//
2535 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2536 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2537 Intrinsic IntId64, Intrinsic IntId128> {
2538 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2539 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2540 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2542 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2543 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2545 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2547 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2549 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2550 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2553 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2555 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2558 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2561 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2562 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2563 Intrinsic IntId64, Intrinsic IntId128> {
2564 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2566 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2567 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2569 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2571 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2574 (bitconvert (memopv4i16 addr:$src))))]>;
2576 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2578 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2579 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2582 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2584 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2587 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2590 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2591 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2592 Intrinsic IntId64, Intrinsic IntId128> {
2593 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2595 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2596 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2598 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2600 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2603 (bitconvert (memopv2i32 addr:$src))))]>;
2605 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2607 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2608 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2611 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2613 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2616 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2619 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2620 int_x86_ssse3_pabs_b,
2621 int_x86_ssse3_pabs_b_128>;
2622 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2623 int_x86_ssse3_pabs_w,
2624 int_x86_ssse3_pabs_w_128>;
2625 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2626 int_x86_ssse3_pabs_d,
2627 int_x86_ssse3_pabs_d_128>;
2629 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2630 let Constraints = "$src1 = $dst" in {
2631 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2632 Intrinsic IntId64, Intrinsic IntId128,
2633 bit Commutable = 0> {
2634 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2635 (ins VR64:$src1, VR64:$src2),
2636 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2637 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2638 let isCommutable = Commutable;
2640 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2641 (ins VR64:$src1, i64mem:$src2),
2642 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2644 (IntId64 VR64:$src1,
2645 (bitconvert (memopv8i8 addr:$src2))))]>;
2647 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2648 (ins VR128:$src1, VR128:$src2),
2649 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2650 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2652 let isCommutable = Commutable;
2654 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2655 (ins VR128:$src1, i128mem:$src2),
2656 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2658 (IntId128 VR128:$src1,
2659 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2663 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2664 let Constraints = "$src1 = $dst" in {
2665 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2666 Intrinsic IntId64, Intrinsic IntId128,
2667 bit Commutable = 0> {
2668 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2669 (ins VR64:$src1, VR64:$src2),
2670 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2671 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2672 let isCommutable = Commutable;
2674 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2675 (ins VR64:$src1, i64mem:$src2),
2676 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2678 (IntId64 VR64:$src1,
2679 (bitconvert (memopv4i16 addr:$src2))))]>;
2681 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2682 (ins VR128:$src1, VR128:$src2),
2683 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2684 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2686 let isCommutable = Commutable;
2688 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2689 (ins VR128:$src1, i128mem:$src2),
2690 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2692 (IntId128 VR128:$src1,
2693 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2697 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2698 let Constraints = "$src1 = $dst" in {
2699 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2700 Intrinsic IntId64, Intrinsic IntId128,
2701 bit Commutable = 0> {
2702 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2703 (ins VR64:$src1, VR64:$src2),
2704 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2705 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2706 let isCommutable = Commutable;
2708 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2709 (ins VR64:$src1, i64mem:$src2),
2710 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2712 (IntId64 VR64:$src1,
2713 (bitconvert (memopv2i32 addr:$src2))))]>;
2715 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2716 (ins VR128:$src1, VR128:$src2),
2717 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2718 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2720 let isCommutable = Commutable;
2722 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2723 (ins VR128:$src1, i128mem:$src2),
2724 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2726 (IntId128 VR128:$src1,
2727 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2731 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2732 int_x86_ssse3_phadd_w,
2733 int_x86_ssse3_phadd_w_128>;
2734 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2735 int_x86_ssse3_phadd_d,
2736 int_x86_ssse3_phadd_d_128>;
2737 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2738 int_x86_ssse3_phadd_sw,
2739 int_x86_ssse3_phadd_sw_128>;
2740 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2741 int_x86_ssse3_phsub_w,
2742 int_x86_ssse3_phsub_w_128>;
2743 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2744 int_x86_ssse3_phsub_d,
2745 int_x86_ssse3_phsub_d_128>;
2746 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2747 int_x86_ssse3_phsub_sw,
2748 int_x86_ssse3_phsub_sw_128>;
2749 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2750 int_x86_ssse3_pmadd_ub_sw,
2751 int_x86_ssse3_pmadd_ub_sw_128>;
2752 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2753 int_x86_ssse3_pmul_hr_sw,
2754 int_x86_ssse3_pmul_hr_sw_128, 1>;
2755 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2756 int_x86_ssse3_pshuf_b,
2757 int_x86_ssse3_pshuf_b_128>;
2758 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2759 int_x86_ssse3_psign_b,
2760 int_x86_ssse3_psign_b_128>;
2761 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2762 int_x86_ssse3_psign_w,
2763 int_x86_ssse3_psign_w_128>;
2764 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2765 int_x86_ssse3_psign_d,
2766 int_x86_ssse3_psign_d_128>;
2768 let Constraints = "$src1 = $dst" in {
2769 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2770 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2771 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2773 (int_x86_ssse3_palign_r
2774 VR64:$src1, VR64:$src2,
2776 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2777 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2778 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2780 (int_x86_ssse3_palign_r
2782 (bitconvert (memopv2i32 addr:$src2)),
2785 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2786 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2787 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2789 (int_x86_ssse3_palign_r_128
2790 VR128:$src1, VR128:$src2,
2791 imm:$src3))]>, OpSize;
2792 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2793 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2794 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2796 (int_x86_ssse3_palign_r_128
2798 (bitconvert (memopv4i32 addr:$src2)),
2799 imm:$src3))]>, OpSize;
2802 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2803 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2804 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2805 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2807 //===----------------------------------------------------------------------===//
2808 // Non-Instruction Patterns
2809 //===----------------------------------------------------------------------===//
2811 // extload f32 -> f64. This matches load+fextend because we have a hack in
2812 // the isel (PreprocessForFPConvert) that can introduce loads after dag combine.
2813 // Since these loads aren't folded into the fextend, we have to match it
2815 let Predicates = [HasSSE2] in
2816 def : Pat<(fextend (loadf32 addr:$src)),
2817 (CVTSS2SDrm addr:$src)>;
2820 let Predicates = [HasSSE2] in {
2821 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2822 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2823 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2824 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2825 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2826 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2827 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2828 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2829 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2830 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2831 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2832 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2833 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2834 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2835 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2836 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2837 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2838 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2839 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2840 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2841 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2842 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2843 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2844 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2845 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2846 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2847 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2848 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2849 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2850 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2853 // Move scalar to XMM zero-extended
2854 // movd to XMM register zero-extends
2855 let AddedComplexity = 15 in {
2856 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2857 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2858 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2859 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2860 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE1]>;
2861 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2862 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2863 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2864 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2867 // Splat v2f64 / v2i64
2868 let AddedComplexity = 10 in {
2869 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
2870 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2871 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
2872 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2873 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
2874 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2875 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
2876 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2879 // Special unary SHUFPSrri case.
2880 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
2881 (SHUFPSrri VR128:$src1, VR128:$src1,
2882 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2883 Requires<[HasSSE1]>;
2884 let AddedComplexity = 5 in
2885 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
2886 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2887 Requires<[HasSSE2]>;
2888 // Special unary SHUFPDrri case.
2889 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
2890 (SHUFPDrri VR128:$src1, VR128:$src1,
2891 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2892 Requires<[HasSSE2]>;
2893 // Special unary SHUFPDrri case.
2894 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
2895 (SHUFPDrri VR128:$src1, VR128:$src1,
2896 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2897 Requires<[HasSSE2]>;
2898 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2899 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
2900 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2901 Requires<[HasSSE2]>;
2903 // Special binary v4i32 shuffle cases with SHUFPS.
2904 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
2905 (SHUFPSrri VR128:$src1, VR128:$src2,
2906 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2907 Requires<[HasSSE2]>;
2908 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
2909 (SHUFPSrmi VR128:$src1, addr:$src2,
2910 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2911 Requires<[HasSSE2]>;
2912 // Special binary v2i64 shuffle cases using SHUFPDrri.
2913 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
2914 (SHUFPDrri VR128:$src1, VR128:$src2,
2915 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2916 Requires<[HasSSE2]>;
2918 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2919 let AddedComplexity = 15 in {
2920 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
2921 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2922 Requires<[OptForSpeed, HasSSE2]>;
2923 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
2924 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2925 Requires<[OptForSpeed, HasSSE2]>;
2927 let AddedComplexity = 10 in {
2928 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
2929 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2930 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
2931 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2932 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
2933 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2934 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
2935 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2938 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2939 let AddedComplexity = 15 in {
2940 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
2941 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2942 Requires<[OptForSpeed, HasSSE2]>;
2943 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
2944 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2945 Requires<[OptForSpeed, HasSSE2]>;
2947 let AddedComplexity = 10 in {
2948 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
2949 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2950 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
2951 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2952 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
2953 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2954 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
2955 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2958 let AddedComplexity = 20 in {
2959 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2960 def : Pat<(v4i32 (movhp VR128:$src1, VR128:$src2)),
2961 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2963 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2964 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
2965 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2967 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2968 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
2969 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2970 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
2971 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2974 let AddedComplexity = 20 in {
2975 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2976 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2977 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
2978 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2979 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
2980 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2981 def : Pat<(v4f32 (movhp VR128:$src1, (load addr:$src2))),
2982 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2983 def : Pat<(v2f64 (movhp VR128:$src1, (load addr:$src2))),
2984 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2986 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
2987 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2988 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
2989 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2990 def : Pat<(v4i32 (movhp VR128:$src1, (load addr:$src2))),
2991 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2992 def : Pat<(v2i64 (movhp VR128:$src1, (load addr:$src2))),
2993 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2996 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
2997 // (store (vector_shuffle (load addr), v2, <0, 1, 4, 5>), addr) using MOVHPS
2998 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
2999 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3000 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3001 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3002 def : Pat<(store (v4f32 (movhp (load addr:$src1), VR128:$src2)), addr:$src1),
3003 (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3004 def : Pat<(store (v2f64 (movhp (load addr:$src1), VR128:$src2)), addr:$src1),
3005 (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3007 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3009 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3010 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3011 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3012 def : Pat<(store (v4i32 (movhp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3014 (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3015 def : Pat<(store (v2i64 (movhp (load addr:$src1), VR128:$src2)), addr:$src1),
3016 (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3019 let AddedComplexity = 15 in {
3020 // Setting the lowest element in the vector.
3021 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3022 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3023 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3024 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3026 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
3027 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3028 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3029 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3030 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3033 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3034 // fall back to this for SSE1)
3035 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3036 (SHUFPSrri VR128:$src2, VR128:$src1,
3037 (SHUFFLE_get_shuf_imm VR128:$src3))>, Requires<[HasSSE1]>;
3039 // Set lowest element and zero upper elements.
3040 let AddedComplexity = 15 in
3041 def : Pat<(v2f64 (movl immAllZerosV_bc, VR128:$src)),
3042 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3043 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3044 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3046 // Some special case pandn patterns.
3047 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3049 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3050 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3052 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3053 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3055 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3057 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3058 (memop addr:$src2))),
3059 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3060 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3061 (memop addr:$src2))),
3062 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3063 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3064 (memop addr:$src2))),
3065 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3067 // vector -> vector casts
3068 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3069 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3070 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3071 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3072 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3073 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3074 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3075 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3077 // Use movaps / movups for SSE integer load / store (one byte shorter).
3078 def : Pat<(alignedloadv4i32 addr:$src),
3079 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
3080 def : Pat<(loadv4i32 addr:$src),
3081 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3082 def : Pat<(alignedloadv2i64 addr:$src),
3083 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3084 def : Pat<(loadv2i64 addr:$src),
3085 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3087 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3088 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3089 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3090 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3091 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3092 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3093 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3094 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3095 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3096 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3097 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3098 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3099 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3100 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3101 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3102 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3104 //===----------------------------------------------------------------------===//
3105 // SSE4.1 Instructions
3106 //===----------------------------------------------------------------------===//
3108 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3111 Intrinsic V2F64Int> {
3112 // Intrinsic operation, reg.
3113 // Vector intrinsic operation, reg
3114 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3115 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3116 !strconcat(OpcodeStr,
3117 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3118 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3121 // Vector intrinsic operation, mem
3122 def PSm_Int : SS4AIi8<opcps, MRMSrcMem,
3123 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3124 !strconcat(OpcodeStr,
3125 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3127 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3130 // Vector intrinsic operation, reg
3131 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3132 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3133 !strconcat(OpcodeStr,
3134 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3135 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3138 // Vector intrinsic operation, mem
3139 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3140 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3141 !strconcat(OpcodeStr,
3142 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3144 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3148 let Constraints = "$src1 = $dst" in {
3149 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3153 // Intrinsic operation, reg.
3154 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3156 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3157 !strconcat(OpcodeStr,
3158 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3160 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3163 // Intrinsic operation, mem.
3164 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3166 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3167 !strconcat(OpcodeStr,
3168 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3170 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3173 // Intrinsic operation, reg.
3174 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3176 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3177 !strconcat(OpcodeStr,
3178 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3180 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3183 // Intrinsic operation, mem.
3184 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3186 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3187 !strconcat(OpcodeStr,
3188 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3190 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3195 // FP round - roundss, roundps, roundsd, roundpd
3196 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3197 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3198 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3199 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3201 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3202 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3203 Intrinsic IntId128> {
3204 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3206 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3207 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3208 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3210 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3213 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3216 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3217 int_x86_sse41_phminposuw>;
3219 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3220 let Constraints = "$src1 = $dst" in {
3221 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3222 Intrinsic IntId128, bit Commutable = 0> {
3223 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3224 (ins VR128:$src1, VR128:$src2),
3225 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3226 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3228 let isCommutable = Commutable;
3230 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3231 (ins VR128:$src1, i128mem:$src2),
3232 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3234 (IntId128 VR128:$src1,
3235 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3239 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3240 int_x86_sse41_pcmpeqq, 1>;
3241 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3242 int_x86_sse41_packusdw, 0>;
3243 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3244 int_x86_sse41_pminsb, 1>;
3245 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3246 int_x86_sse41_pminsd, 1>;
3247 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3248 int_x86_sse41_pminud, 1>;
3249 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3250 int_x86_sse41_pminuw, 1>;
3251 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3252 int_x86_sse41_pmaxsb, 1>;
3253 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3254 int_x86_sse41_pmaxsd, 1>;
3255 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3256 int_x86_sse41_pmaxud, 1>;
3257 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3258 int_x86_sse41_pmaxuw, 1>;
3260 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3262 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3263 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3264 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3265 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3267 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3268 let Constraints = "$src1 = $dst" in {
3269 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3270 SDNode OpNode, Intrinsic IntId128,
3271 bit Commutable = 0> {
3272 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3273 (ins VR128:$src1, VR128:$src2),
3274 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3275 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3276 VR128:$src2))]>, OpSize {
3277 let isCommutable = Commutable;
3279 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3280 (ins VR128:$src1, VR128:$src2),
3281 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3282 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3284 let isCommutable = Commutable;
3286 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3287 (ins VR128:$src1, i128mem:$src2),
3288 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3290 (OpNode VR128:$src1, (memop addr:$src2)))]>, OpSize;
3291 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3292 (ins VR128:$src1, i128mem:$src2),
3293 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3295 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3299 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", v4i32, mul,
3300 int_x86_sse41_pmulld, 1>;
3302 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3303 let Constraints = "$src1 = $dst" in {
3304 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3305 Intrinsic IntId128, bit Commutable = 0> {
3306 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3307 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3308 !strconcat(OpcodeStr,
3309 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3311 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3313 let isCommutable = Commutable;
3315 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3316 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3317 !strconcat(OpcodeStr,
3318 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3320 (IntId128 VR128:$src1,
3321 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3326 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3327 int_x86_sse41_blendps, 0>;
3328 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3329 int_x86_sse41_blendpd, 0>;
3330 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3331 int_x86_sse41_pblendw, 0>;
3332 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3333 int_x86_sse41_dpps, 1>;
3334 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3335 int_x86_sse41_dppd, 1>;
3336 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3337 int_x86_sse41_mpsadbw, 1>;
3340 /// SS41I_ternary_int - SSE 4.1 ternary operator
3341 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3342 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3343 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3344 (ins VR128:$src1, VR128:$src2),
3345 !strconcat(OpcodeStr,
3346 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3347 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3350 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3351 (ins VR128:$src1, i128mem:$src2),
3352 !strconcat(OpcodeStr,
3353 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3356 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3360 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3361 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3362 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3365 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3366 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3367 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3368 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3370 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3371 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3373 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3377 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3378 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3379 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3380 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3381 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3382 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3384 // Common patterns involving scalar load.
3385 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3386 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3387 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3388 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3390 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3391 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3392 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3393 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3395 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3396 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3397 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3398 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3400 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3401 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3402 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3403 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3405 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3406 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3407 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3408 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3410 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3411 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3412 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3413 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3416 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3417 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3418 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3419 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3421 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3422 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3424 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3428 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3429 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3430 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3431 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3433 // Common patterns involving scalar load
3434 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3435 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3436 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3437 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3439 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3440 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3441 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3442 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3445 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3446 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3447 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3448 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3450 // Expecting a i16 load any extended to i32 value.
3451 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3452 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3453 [(set VR128:$dst, (IntId (bitconvert
3454 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3458 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3459 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3461 // Common patterns involving scalar load
3462 def : Pat<(int_x86_sse41_pmovsxbq
3463 (bitconvert (v4i32 (X86vzmovl
3464 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3465 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3467 def : Pat<(int_x86_sse41_pmovzxbq
3468 (bitconvert (v4i32 (X86vzmovl
3469 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3470 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3473 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3474 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3475 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3476 (ins VR128:$src1, i32i8imm:$src2),
3477 !strconcat(OpcodeStr,
3478 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3479 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3481 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3482 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3483 !strconcat(OpcodeStr,
3484 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3487 // There's an AssertZext in the way of writing the store pattern
3488 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3491 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3494 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3495 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3496 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3497 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3498 !strconcat(OpcodeStr,
3499 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3502 // There's an AssertZext in the way of writing the store pattern
3503 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3506 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3509 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3510 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3511 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3512 (ins VR128:$src1, i32i8imm:$src2),
3513 !strconcat(OpcodeStr,
3514 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3516 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3517 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3518 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3519 !strconcat(OpcodeStr,
3520 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3521 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3522 addr:$dst)]>, OpSize;
3525 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3528 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3530 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3531 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3532 (ins VR128:$src1, i32i8imm:$src2),
3533 !strconcat(OpcodeStr,
3534 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3536 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3538 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3539 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3540 !strconcat(OpcodeStr,
3541 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3542 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3543 addr:$dst)]>, OpSize;
3546 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3548 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3549 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3552 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3553 Requires<[HasSSE41]>;
3555 let Constraints = "$src1 = $dst" in {
3556 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3557 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3558 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3559 !strconcat(OpcodeStr,
3560 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3562 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3563 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3564 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3565 !strconcat(OpcodeStr,
3566 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3568 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3569 imm:$src3))]>, OpSize;
3573 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3575 let Constraints = "$src1 = $dst" in {
3576 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3577 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3578 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3579 !strconcat(OpcodeStr,
3580 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3582 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3584 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3585 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3586 !strconcat(OpcodeStr,
3587 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3589 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3590 imm:$src3)))]>, OpSize;
3594 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3596 // insertps has a few different modes, there's the first two here below which
3597 // are optimized inserts that won't zero arbitrary elements in the destination
3598 // vector. The next one matches the intrinsic and could zero arbitrary elements
3599 // in the target vector.
3600 let Constraints = "$src1 = $dst" in {
3601 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3602 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3603 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3604 !strconcat(OpcodeStr,
3605 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3607 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>, OpSize;
3608 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3609 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3610 !strconcat(OpcodeStr,
3611 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3613 (X86insrtps VR128:$src1,
3614 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3615 imm:$src3))]>, OpSize;
3619 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3621 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3622 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3624 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3625 // the intel intrinsic that corresponds to this.
3626 let Defs = [EFLAGS] in {
3627 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3628 "ptest \t{$src2, $src1|$src1, $src2}",
3629 [(X86ptest VR128:$src1, VR128:$src2),
3630 (implicit EFLAGS)]>, OpSize;
3631 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3632 "ptest \t{$src2, $src1|$src1, $src2}",
3633 [(X86ptest VR128:$src1, (load addr:$src2)),
3634 (implicit EFLAGS)]>, OpSize;
3637 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3638 "movntdqa\t{$src, $dst|$dst, $src}",
3639 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;
3641 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3642 let Constraints = "$src1 = $dst" in {
3643 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3644 Intrinsic IntId128, bit Commutable = 0> {
3645 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3646 (ins VR128:$src1, VR128:$src2),
3647 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3648 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3650 let isCommutable = Commutable;
3652 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3653 (ins VR128:$src1, i128mem:$src2),
3654 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3656 (IntId128 VR128:$src1,
3657 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3661 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3663 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3664 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3665 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3666 (PCMPGTQrm VR128:$src1, addr:$src2)>;