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>,
74 def X86ptest : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
76 //===----------------------------------------------------------------------===//
77 // SSE Complex Patterns
78 //===----------------------------------------------------------------------===//
80 // These are 'extloads' from a scalar to the low element of a vector, zeroing
81 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
83 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
84 [SDNPHasChain, SDNPMayLoad]>;
85 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
86 [SDNPHasChain, SDNPMayLoad]>;
88 def ssmem : Operand<v4f32> {
89 let PrintMethod = "printf32mem";
90 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
91 let ParserMatchClass = X86MemAsmOperand;
93 def sdmem : Operand<v2f64> {
94 let PrintMethod = "printf64mem";
95 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
96 let ParserMatchClass = X86MemAsmOperand;
99 //===----------------------------------------------------------------------===//
100 // SSE pattern fragments
101 //===----------------------------------------------------------------------===//
103 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
104 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
105 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
106 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
108 // Like 'store', but always requires vector alignment.
109 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
110 (store node:$val, node:$ptr), [{
111 return cast<StoreSDNode>(N)->getAlignment() >= 16;
114 // Like 'load', but always requires vector alignment.
115 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
116 return cast<LoadSDNode>(N)->getAlignment() >= 16;
119 def alignedloadfsf32 : PatFrag<(ops node:$ptr),
120 (f32 (alignedload node:$ptr))>;
121 def alignedloadfsf64 : PatFrag<(ops node:$ptr),
122 (f64 (alignedload node:$ptr))>;
123 def alignedloadv4f32 : PatFrag<(ops node:$ptr),
124 (v4f32 (alignedload node:$ptr))>;
125 def alignedloadv2f64 : PatFrag<(ops node:$ptr),
126 (v2f64 (alignedload node:$ptr))>;
127 def alignedloadv4i32 : PatFrag<(ops node:$ptr),
128 (v4i32 (alignedload node:$ptr))>;
129 def alignedloadv2i64 : PatFrag<(ops node:$ptr),
130 (v2i64 (alignedload node:$ptr))>;
132 // Like 'load', but uses special alignment checks suitable for use in
133 // memory operands in most SSE instructions, which are required to
134 // be naturally aligned on some targets but not on others.
135 // FIXME: Actually implement support for targets that don't require the
136 // alignment. This probably wants a subtarget predicate.
137 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
138 return cast<LoadSDNode>(N)->getAlignment() >= 16;
141 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
142 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
143 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
144 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
145 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
146 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
147 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
149 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
151 // FIXME: 8 byte alignment for mmx reads is not required
152 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
153 return cast<LoadSDNode>(N)->getAlignment() >= 8;
156 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
157 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
158 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
159 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
161 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
162 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
163 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
164 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
165 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
166 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
168 def vzmovl_v2i64 : PatFrag<(ops node:$src),
169 (bitconvert (v2i64 (X86vzmovl
170 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
171 def vzmovl_v4i32 : PatFrag<(ops node:$src),
172 (bitconvert (v4i32 (X86vzmovl
173 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
175 def vzload_v2i64 : PatFrag<(ops node:$src),
176 (bitconvert (v2i64 (X86vzload node:$src)))>;
179 def fp32imm0 : PatLeaf<(f32 fpimm), [{
180 return N->isExactlyValue(+0.0);
183 // BYTE_imm - Transform bit immediates into byte immediates.
184 def BYTE_imm : SDNodeXForm<imm, [{
185 // Transformation function: imm >> 3
186 return getI32Imm(N->getZExtValue() >> 3);
189 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
191 def SHUFFLE_get_shuf_imm : SDNodeXForm<vector_shuffle, [{
192 return getI8Imm(X86::getShuffleSHUFImmediate(N));
195 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
197 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<vector_shuffle, [{
198 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
201 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
203 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<vector_shuffle, [{
204 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
207 // SHUFFLE_get_palign_imm xform function: convert vector_shuffle mask to
209 def SHUFFLE_get_palign_imm : SDNodeXForm<vector_shuffle, [{
210 return getI8Imm(X86::getShufflePALIGNRImmediate(N));
213 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
214 (vector_shuffle node:$lhs, node:$rhs), [{
215 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
216 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
219 def movddup : PatFrag<(ops node:$lhs, node:$rhs),
220 (vector_shuffle node:$lhs, node:$rhs), [{
221 return X86::isMOVDDUPMask(cast<ShuffleVectorSDNode>(N));
224 def movhlps : PatFrag<(ops node:$lhs, node:$rhs),
225 (vector_shuffle node:$lhs, node:$rhs), [{
226 return X86::isMOVHLPSMask(cast<ShuffleVectorSDNode>(N));
229 def movhlps_undef : PatFrag<(ops node:$lhs, node:$rhs),
230 (vector_shuffle node:$lhs, node:$rhs), [{
231 return X86::isMOVHLPS_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
234 def movlhps : PatFrag<(ops node:$lhs, node:$rhs),
235 (vector_shuffle node:$lhs, node:$rhs), [{
236 return X86::isMOVLHPSMask(cast<ShuffleVectorSDNode>(N));
239 def movlp : PatFrag<(ops node:$lhs, node:$rhs),
240 (vector_shuffle node:$lhs, node:$rhs), [{
241 return X86::isMOVLPMask(cast<ShuffleVectorSDNode>(N));
244 def movl : PatFrag<(ops node:$lhs, node:$rhs),
245 (vector_shuffle node:$lhs, node:$rhs), [{
246 return X86::isMOVLMask(cast<ShuffleVectorSDNode>(N));
249 def movshdup : PatFrag<(ops node:$lhs, node:$rhs),
250 (vector_shuffle node:$lhs, node:$rhs), [{
251 return X86::isMOVSHDUPMask(cast<ShuffleVectorSDNode>(N));
254 def movsldup : PatFrag<(ops node:$lhs, node:$rhs),
255 (vector_shuffle node:$lhs, node:$rhs), [{
256 return X86::isMOVSLDUPMask(cast<ShuffleVectorSDNode>(N));
259 def unpckl : PatFrag<(ops node:$lhs, node:$rhs),
260 (vector_shuffle node:$lhs, node:$rhs), [{
261 return X86::isUNPCKLMask(cast<ShuffleVectorSDNode>(N));
264 def unpckh : PatFrag<(ops node:$lhs, node:$rhs),
265 (vector_shuffle node:$lhs, node:$rhs), [{
266 return X86::isUNPCKHMask(cast<ShuffleVectorSDNode>(N));
269 def unpckl_undef : PatFrag<(ops node:$lhs, node:$rhs),
270 (vector_shuffle node:$lhs, node:$rhs), [{
271 return X86::isUNPCKL_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
274 def unpckh_undef : PatFrag<(ops node:$lhs, node:$rhs),
275 (vector_shuffle node:$lhs, node:$rhs), [{
276 return X86::isUNPCKH_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
279 def pshufd : PatFrag<(ops node:$lhs, node:$rhs),
280 (vector_shuffle node:$lhs, node:$rhs), [{
281 return X86::isPSHUFDMask(cast<ShuffleVectorSDNode>(N));
282 }], SHUFFLE_get_shuf_imm>;
284 def shufp : PatFrag<(ops node:$lhs, node:$rhs),
285 (vector_shuffle node:$lhs, node:$rhs), [{
286 return X86::isSHUFPMask(cast<ShuffleVectorSDNode>(N));
287 }], SHUFFLE_get_shuf_imm>;
289 def pshufhw : PatFrag<(ops node:$lhs, node:$rhs),
290 (vector_shuffle node:$lhs, node:$rhs), [{
291 return X86::isPSHUFHWMask(cast<ShuffleVectorSDNode>(N));
292 }], SHUFFLE_get_pshufhw_imm>;
294 def pshuflw : PatFrag<(ops node:$lhs, node:$rhs),
295 (vector_shuffle node:$lhs, node:$rhs), [{
296 return X86::isPSHUFLWMask(cast<ShuffleVectorSDNode>(N));
297 }], SHUFFLE_get_pshuflw_imm>;
299 def palign : PatFrag<(ops node:$lhs, node:$rhs),
300 (vector_shuffle node:$lhs, node:$rhs), [{
301 return X86::isPALIGNRMask(cast<ShuffleVectorSDNode>(N));
302 }], SHUFFLE_get_palign_imm>;
304 //===----------------------------------------------------------------------===//
305 // SSE scalar FP Instructions
306 //===----------------------------------------------------------------------===//
308 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded after
309 // instruction selection into a branch sequence.
310 let Uses = [EFLAGS], usesCustomInserter = 1 in {
311 def CMOV_FR32 : I<0, Pseudo,
312 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
313 "#CMOV_FR32 PSEUDO!",
314 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
316 def CMOV_FR64 : I<0, Pseudo,
317 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
318 "#CMOV_FR64 PSEUDO!",
319 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
321 def CMOV_V4F32 : I<0, Pseudo,
322 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
323 "#CMOV_V4F32 PSEUDO!",
325 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
327 def CMOV_V2F64 : I<0, Pseudo,
328 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
329 "#CMOV_V2F64 PSEUDO!",
331 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
333 def CMOV_V2I64 : I<0, Pseudo,
334 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
335 "#CMOV_V2I64 PSEUDO!",
337 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
341 //===----------------------------------------------------------------------===//
343 //===----------------------------------------------------------------------===//
346 let neverHasSideEffects = 1 in
347 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
348 "movss\t{$src, $dst|$dst, $src}", []>;
349 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
350 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
351 "movss\t{$src, $dst|$dst, $src}",
352 [(set FR32:$dst, (loadf32 addr:$src))]>;
353 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
354 "movss\t{$src, $dst|$dst, $src}",
355 [(store FR32:$src, addr:$dst)]>;
357 // Conversion instructions
358 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
359 "cvttss2si\t{$src, $dst|$dst, $src}",
360 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
361 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
362 "cvttss2si\t{$src, $dst|$dst, $src}",
363 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
364 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
365 "cvtsi2ss\t{$src, $dst|$dst, $src}",
366 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
367 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
368 "cvtsi2ss\t{$src, $dst|$dst, $src}",
369 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
371 // Match intrinsics which expect XMM operand(s).
372 def CVTSS2SIrr: SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
373 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
374 def CVTSS2SIrm: SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
375 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
377 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
378 "cvtss2si\t{$src, $dst|$dst, $src}",
379 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
380 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
381 "cvtss2si\t{$src, $dst|$dst, $src}",
382 [(set GR32:$dst, (int_x86_sse_cvtss2si
383 (load addr:$src)))]>;
385 // Match intrinisics which expect MM and XMM operand(s).
386 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
387 "cvtps2pi\t{$src, $dst|$dst, $src}",
388 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
389 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
390 "cvtps2pi\t{$src, $dst|$dst, $src}",
391 [(set VR64:$dst, (int_x86_sse_cvtps2pi
392 (load addr:$src)))]>;
393 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
394 "cvttps2pi\t{$src, $dst|$dst, $src}",
395 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
396 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
397 "cvttps2pi\t{$src, $dst|$dst, $src}",
398 [(set VR64:$dst, (int_x86_sse_cvttps2pi
399 (load addr:$src)))]>;
400 let Constraints = "$src1 = $dst" in {
401 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
402 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
403 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
404 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
406 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
407 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
408 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
409 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
410 (load addr:$src2)))]>;
413 // Aliases for intrinsics
414 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
415 "cvttss2si\t{$src, $dst|$dst, $src}",
417 (int_x86_sse_cvttss2si VR128:$src))]>;
418 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
419 "cvttss2si\t{$src, $dst|$dst, $src}",
421 (int_x86_sse_cvttss2si(load addr:$src)))]>;
423 let Constraints = "$src1 = $dst" in {
424 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
425 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
426 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
427 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
429 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
430 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
431 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
432 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
433 (loadi32 addr:$src2)))]>;
436 // Comparison instructions
437 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
438 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
439 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
440 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
442 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
443 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
444 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
447 let Defs = [EFLAGS] in {
448 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
449 "ucomiss\t{$src2, $src1|$src1, $src2}",
450 [(X86cmp FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
451 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
452 "ucomiss\t{$src2, $src1|$src1, $src2}",
453 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
456 def COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
457 "comiss\t{$src2, $src1|$src1, $src2}", []>;
458 def COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
459 "comiss\t{$src2, $src1|$src1, $src2}", []>;
463 // Aliases to match intrinsics which expect XMM operand(s).
464 let Constraints = "$src1 = $dst" in {
465 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
467 (ins VR128:$src1, VR128:$src, SSECC:$cc),
468 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
469 [(set VR128:$dst, (int_x86_sse_cmp_ss
471 VR128:$src, imm:$cc))]>;
472 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
474 (ins VR128:$src1, f32mem:$src, SSECC:$cc),
475 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
476 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
477 (load addr:$src), imm:$cc))]>;
480 let Defs = [EFLAGS] in {
481 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
482 "ucomiss\t{$src2, $src1|$src1, $src2}",
483 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
485 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
486 "ucomiss\t{$src2, $src1|$src1, $src2}",
487 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
490 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
491 "comiss\t{$src2, $src1|$src1, $src2}",
492 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
494 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
495 "comiss\t{$src2, $src1|$src1, $src2}",
496 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
500 // Aliases of packed SSE1 instructions for scalar use. These all have names
501 // that start with 'Fs'.
503 // Alias instructions that map fld0 to pxor for sse.
504 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
506 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
507 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
508 Requires<[HasSSE1]>, TB, OpSize;
510 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
512 let neverHasSideEffects = 1 in
513 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
514 "movaps\t{$src, $dst|$dst, $src}", []>;
516 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
518 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
519 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
520 "movaps\t{$src, $dst|$dst, $src}",
521 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
523 // Alias bitwise logical operations using SSE logical ops on packed FP values.
524 let Constraints = "$src1 = $dst" in {
525 let isCommutable = 1 in {
526 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst),
527 (ins FR32:$src1, FR32:$src2),
528 "andps\t{$src2, $dst|$dst, $src2}",
529 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
530 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst),
531 (ins FR32:$src1, FR32:$src2),
532 "orps\t{$src2, $dst|$dst, $src2}",
533 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
534 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst),
535 (ins FR32:$src1, FR32:$src2),
536 "xorps\t{$src2, $dst|$dst, $src2}",
537 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
540 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst),
541 (ins FR32:$src1, f128mem:$src2),
542 "andps\t{$src2, $dst|$dst, $src2}",
543 [(set FR32:$dst, (X86fand FR32:$src1,
544 (memopfsf32 addr:$src2)))]>;
545 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst),
546 (ins FR32:$src1, f128mem:$src2),
547 "orps\t{$src2, $dst|$dst, $src2}",
548 [(set FR32:$dst, (X86for FR32:$src1,
549 (memopfsf32 addr:$src2)))]>;
550 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst),
551 (ins FR32:$src1, f128mem:$src2),
552 "xorps\t{$src2, $dst|$dst, $src2}",
553 [(set FR32:$dst, (X86fxor FR32:$src1,
554 (memopfsf32 addr:$src2)))]>;
556 let neverHasSideEffects = 1 in {
557 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
558 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
559 "andnps\t{$src2, $dst|$dst, $src2}", []>;
561 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
562 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
563 "andnps\t{$src2, $dst|$dst, $src2}", []>;
567 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
569 /// In addition, we also have a special variant of the scalar form here to
570 /// represent the associated intrinsic operation. This form is unlike the
571 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
572 /// and leaves the top elements unmodified (therefore these cannot be commuted).
574 /// These three forms can each be reg+reg or reg+mem, so there are a total of
575 /// six "instructions".
577 let Constraints = "$src1 = $dst" in {
578 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
579 SDNode OpNode, Intrinsic F32Int,
580 bit Commutable = 0> {
581 // Scalar operation, reg+reg.
582 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
583 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
584 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
585 let isCommutable = Commutable;
588 // Scalar operation, reg+mem.
589 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
590 (ins FR32:$src1, f32mem:$src2),
591 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
592 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
594 // Vector operation, reg+reg.
595 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
596 (ins VR128:$src1, VR128:$src2),
597 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
598 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
599 let isCommutable = Commutable;
602 // Vector operation, reg+mem.
603 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
604 (ins VR128:$src1, f128mem:$src2),
605 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
606 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
608 // Intrinsic operation, reg+reg.
609 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
610 (ins VR128:$src1, VR128:$src2),
611 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
612 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]>;
614 // Intrinsic operation, reg+mem.
615 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
616 (ins VR128:$src1, ssmem:$src2),
617 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
618 [(set VR128:$dst, (F32Int VR128:$src1,
619 sse_load_f32:$src2))]>;
623 // Arithmetic instructions
624 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
625 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
626 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
627 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
629 /// sse1_fp_binop_rm - Other SSE1 binops
631 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
632 /// instructions for a full-vector intrinsic form. Operations that map
633 /// onto C operators don't use this form since they just use the plain
634 /// vector form instead of having a separate vector intrinsic form.
636 /// This provides a total of eight "instructions".
638 let Constraints = "$src1 = $dst" in {
639 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
643 bit Commutable = 0> {
645 // Scalar operation, reg+reg.
646 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
647 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
648 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
649 let isCommutable = Commutable;
652 // Scalar operation, reg+mem.
653 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
654 (ins FR32:$src1, f32mem:$src2),
655 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
656 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
658 // Vector operation, reg+reg.
659 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
660 (ins VR128:$src1, VR128:$src2),
661 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
662 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
663 let isCommutable = Commutable;
666 // Vector operation, reg+mem.
667 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
668 (ins VR128:$src1, f128mem:$src2),
669 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
670 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
672 // Intrinsic operation, reg+reg.
673 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
674 (ins VR128:$src1, VR128:$src2),
675 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
676 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
677 let isCommutable = Commutable;
680 // Intrinsic operation, reg+mem.
681 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
682 (ins VR128:$src1, ssmem:$src2),
683 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
684 [(set VR128:$dst, (F32Int VR128:$src1,
685 sse_load_f32:$src2))]>;
687 // Vector intrinsic operation, reg+reg.
688 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
689 (ins VR128:$src1, VR128:$src2),
690 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
691 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
692 let isCommutable = Commutable;
695 // Vector intrinsic operation, reg+mem.
696 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
697 (ins VR128:$src1, f128mem:$src2),
698 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
699 [(set VR128:$dst, (V4F32Int VR128:$src1, (memopv4f32 addr:$src2)))]>;
703 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
704 int_x86_sse_max_ss, int_x86_sse_max_ps>;
705 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
706 int_x86_sse_min_ss, int_x86_sse_min_ps>;
708 //===----------------------------------------------------------------------===//
709 // SSE packed FP Instructions
712 let neverHasSideEffects = 1 in
713 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
714 "movaps\t{$src, $dst|$dst, $src}", []>;
715 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
716 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
717 "movaps\t{$src, $dst|$dst, $src}",
718 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
720 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
721 "movaps\t{$src, $dst|$dst, $src}",
722 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
724 let neverHasSideEffects = 1 in
725 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
726 "movups\t{$src, $dst|$dst, $src}", []>;
727 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
728 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
729 "movups\t{$src, $dst|$dst, $src}",
730 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
731 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
732 "movups\t{$src, $dst|$dst, $src}",
733 [(store (v4f32 VR128:$src), addr:$dst)]>;
735 // Intrinsic forms of MOVUPS load and store
736 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
737 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
738 "movups\t{$src, $dst|$dst, $src}",
739 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
740 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
741 "movups\t{$src, $dst|$dst, $src}",
742 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
744 let Constraints = "$src1 = $dst" in {
745 let AddedComplexity = 20 in {
746 def MOVLPSrm : PSI<0x12, MRMSrcMem,
747 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
748 "movlps\t{$src2, $dst|$dst, $src2}",
751 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
752 def MOVHPSrm : PSI<0x16, MRMSrcMem,
753 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
754 "movhps\t{$src2, $dst|$dst, $src2}",
756 (movlhps VR128:$src1,
757 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
759 } // Constraints = "$src1 = $dst"
762 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
763 "movlps\t{$src, $dst|$dst, $src}",
764 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
765 (iPTR 0))), addr:$dst)]>;
767 // v2f64 extract element 1 is always custom lowered to unpack high to low
768 // and extract element 0 so the non-store version isn't too horrible.
769 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
770 "movhps\t{$src, $dst|$dst, $src}",
771 [(store (f64 (vector_extract
772 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
773 (undef)), (iPTR 0))), addr:$dst)]>;
775 let Constraints = "$src1 = $dst" in {
776 let AddedComplexity = 20 in {
777 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
778 (ins VR128:$src1, VR128:$src2),
779 "movlhps\t{$src2, $dst|$dst, $src2}",
781 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
783 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
784 (ins VR128:$src1, VR128:$src2),
785 "movhlps\t{$src2, $dst|$dst, $src2}",
787 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
789 } // Constraints = "$src1 = $dst"
791 let AddedComplexity = 20 in {
792 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
793 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
794 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
795 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
802 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
804 /// In addition, we also have a special variant of the scalar form here to
805 /// represent the associated intrinsic operation. This form is unlike the
806 /// plain scalar form, in that it takes an entire vector (instead of a
807 /// scalar) and leaves the top elements undefined.
809 /// And, we have a special variant form for a full-vector intrinsic form.
811 /// These four forms can each have a reg or a mem operand, so there are a
812 /// total of eight "instructions".
814 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
818 bit Commutable = 0> {
819 // Scalar operation, reg.
820 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
821 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
822 [(set FR32:$dst, (OpNode FR32:$src))]> {
823 let isCommutable = Commutable;
826 // Scalar operation, mem.
827 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
828 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
829 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
831 // Vector operation, reg.
832 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
833 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
834 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
835 let isCommutable = Commutable;
838 // Vector operation, mem.
839 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
840 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
841 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
843 // Intrinsic operation, reg.
844 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
845 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
846 [(set VR128:$dst, (F32Int VR128:$src))]> {
847 let isCommutable = Commutable;
850 // Intrinsic operation, mem.
851 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
852 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
853 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
855 // Vector intrinsic operation, reg
856 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
857 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
858 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
859 let isCommutable = Commutable;
862 // Vector intrinsic operation, mem
863 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
864 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
865 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
869 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
870 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
872 // Reciprocal approximations. Note that these typically require refinement
873 // in order to obtain suitable precision.
874 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
875 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
876 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
877 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
880 let Constraints = "$src1 = $dst" in {
881 let isCommutable = 1 in {
882 def ANDPSrr : PSI<0x54, MRMSrcReg,
883 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
884 "andps\t{$src2, $dst|$dst, $src2}",
885 [(set VR128:$dst, (v2i64
886 (and VR128:$src1, VR128:$src2)))]>;
887 def ORPSrr : PSI<0x56, MRMSrcReg,
888 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
889 "orps\t{$src2, $dst|$dst, $src2}",
890 [(set VR128:$dst, (v2i64
891 (or VR128:$src1, VR128:$src2)))]>;
892 def XORPSrr : PSI<0x57, MRMSrcReg,
893 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
894 "xorps\t{$src2, $dst|$dst, $src2}",
895 [(set VR128:$dst, (v2i64
896 (xor VR128:$src1, VR128:$src2)))]>;
899 def ANDPSrm : PSI<0x54, MRMSrcMem,
900 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
901 "andps\t{$src2, $dst|$dst, $src2}",
902 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
903 (memopv2i64 addr:$src2)))]>;
904 def ORPSrm : PSI<0x56, MRMSrcMem,
905 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
906 "orps\t{$src2, $dst|$dst, $src2}",
907 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
908 (memopv2i64 addr:$src2)))]>;
909 def XORPSrm : PSI<0x57, MRMSrcMem,
910 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
911 "xorps\t{$src2, $dst|$dst, $src2}",
912 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
913 (memopv2i64 addr:$src2)))]>;
914 def ANDNPSrr : PSI<0x55, MRMSrcReg,
915 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
916 "andnps\t{$src2, $dst|$dst, $src2}",
918 (v2i64 (and (xor VR128:$src1,
919 (bc_v2i64 (v4i32 immAllOnesV))),
921 def ANDNPSrm : PSI<0x55, MRMSrcMem,
922 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
923 "andnps\t{$src2, $dst|$dst, $src2}",
925 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
926 (bc_v2i64 (v4i32 immAllOnesV))),
927 (memopv2i64 addr:$src2))))]>;
930 let Constraints = "$src1 = $dst" in {
931 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
932 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
933 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
934 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
935 VR128:$src, imm:$cc))]>;
936 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
937 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
938 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
939 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
940 (memop addr:$src), imm:$cc))]>;
942 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
943 (CMPPSrri VR128:$src1, VR128:$src2, imm:$cc)>;
944 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
945 (CMPPSrmi VR128:$src1, addr:$src2, imm:$cc)>;
947 // Shuffle and unpack instructions
948 let Constraints = "$src1 = $dst" in {
949 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
950 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
951 (outs VR128:$dst), (ins VR128:$src1,
952 VR128:$src2, i8imm:$src3),
953 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
955 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
956 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
957 (outs VR128:$dst), (ins VR128:$src1,
958 f128mem:$src2, i8imm:$src3),
959 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
962 VR128:$src1, (memopv4f32 addr:$src2))))]>;
964 let AddedComplexity = 10 in {
965 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
966 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
967 "unpckhps\t{$src2, $dst|$dst, $src2}",
969 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
970 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
971 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
972 "unpckhps\t{$src2, $dst|$dst, $src2}",
974 (v4f32 (unpckh VR128:$src1,
975 (memopv4f32 addr:$src2))))]>;
977 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
978 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
979 "unpcklps\t{$src2, $dst|$dst, $src2}",
981 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
982 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
983 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
984 "unpcklps\t{$src2, $dst|$dst, $src2}",
986 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
988 } // Constraints = "$src1 = $dst"
991 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
992 "movmskps\t{$src, $dst|$dst, $src}",
993 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
994 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
995 "movmskpd\t{$src, $dst|$dst, $src}",
996 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
998 // Prefetch intrinsic.
999 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1000 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1001 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1002 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1003 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1004 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1005 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1006 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1008 // Non-temporal stores
1009 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1010 "movntps\t{$src, $dst|$dst, $src}",
1011 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1013 // Load, store, and memory fence
1014 def SFENCE : PSI<0xAE, MRM7r, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
1017 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1018 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1019 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1020 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1022 // Alias instructions that map zero vector to pxor / xorp* for sse.
1023 // We set canFoldAsLoad because this can be converted to a constant-pool
1024 // load of an all-zeros value if folding it would be beneficial.
1025 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1026 isCodeGenOnly = 1 in
1027 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
1028 "xorps\t$dst, $dst",
1029 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1031 let Predicates = [HasSSE1] in {
1032 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
1033 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
1034 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
1035 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
1036 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
1039 // FR32 to 128-bit vector conversion.
1040 let isAsCheapAsAMove = 1 in
1041 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
1042 "movss\t{$src, $dst|$dst, $src}",
1044 (v4f32 (scalar_to_vector FR32:$src)))]>;
1045 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1046 "movss\t{$src, $dst|$dst, $src}",
1048 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
1050 // FIXME: may not be able to eliminate this movss with coalescing the src and
1051 // dest register classes are different. We really want to write this pattern
1053 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1054 // (f32 FR32:$src)>;
1055 let isAsCheapAsAMove = 1 in
1056 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
1057 "movss\t{$src, $dst|$dst, $src}",
1058 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
1060 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
1061 "movss\t{$src, $dst|$dst, $src}",
1062 [(store (f32 (vector_extract (v4f32 VR128:$src),
1063 (iPTR 0))), addr:$dst)]>;
1066 // Move to lower bits of a VR128, leaving upper bits alone.
1067 // Three operand (but two address) aliases.
1068 let Constraints = "$src1 = $dst" in {
1069 let neverHasSideEffects = 1 in
1070 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
1071 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
1072 "movss\t{$src2, $dst|$dst, $src2}", []>;
1074 let AddedComplexity = 15 in
1075 def MOVLPSrr : SSI<0x10, MRMSrcReg,
1076 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1077 "movss\t{$src2, $dst|$dst, $src2}",
1079 (v4f32 (movl VR128:$src1, VR128:$src2)))]>;
1082 // Move to lower bits of a VR128 and zeroing upper bits.
1083 // Loading from memory automatically zeroing upper bits.
1084 let AddedComplexity = 20 in
1085 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1086 "movss\t{$src, $dst|$dst, $src}",
1087 [(set VR128:$dst, (v4f32 (X86vzmovl (v4f32 (scalar_to_vector
1088 (loadf32 addr:$src))))))]>;
1090 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
1091 (MOVZSS2PSrm addr:$src)>;
1093 //===---------------------------------------------------------------------===//
1094 // SSE2 Instructions
1095 //===---------------------------------------------------------------------===//
1097 // Move Instructions
1098 let neverHasSideEffects = 1 in
1099 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1100 "movsd\t{$src, $dst|$dst, $src}", []>;
1101 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1102 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1103 "movsd\t{$src, $dst|$dst, $src}",
1104 [(set FR64:$dst, (loadf64 addr:$src))]>;
1105 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1106 "movsd\t{$src, $dst|$dst, $src}",
1107 [(store FR64:$src, addr:$dst)]>;
1109 // Conversion instructions
1110 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1111 "cvttsd2si\t{$src, $dst|$dst, $src}",
1112 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1113 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1114 "cvttsd2si\t{$src, $dst|$dst, $src}",
1115 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1116 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1117 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1118 [(set FR32:$dst, (fround FR64:$src))]>;
1119 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1120 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1121 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1122 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1123 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1124 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1125 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1126 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1127 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1129 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1130 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1131 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1132 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1133 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1134 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1135 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1136 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1137 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1138 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1139 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1140 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1141 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1142 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1143 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1144 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1145 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1146 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1147 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1148 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1150 // SSE2 instructions with XS prefix
1151 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1152 "cvtss2sd\t{$src, $dst|$dst, $src}",
1153 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1154 Requires<[HasSSE2]>;
1155 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1156 "cvtss2sd\t{$src, $dst|$dst, $src}",
1157 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1158 Requires<[HasSSE2]>;
1160 // Match intrinsics which expect XMM operand(s).
1161 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1162 "cvtsd2si\t{$src, $dst|$dst, $src}",
1163 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1164 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1165 "cvtsd2si\t{$src, $dst|$dst, $src}",
1166 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1167 (load addr:$src)))]>;
1169 // Match intrinisics which expect MM and XMM operand(s).
1170 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1171 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1172 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1173 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1174 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1175 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1176 (memop addr:$src)))]>;
1177 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1178 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1179 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1180 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1181 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1182 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1183 (memop addr:$src)))]>;
1184 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1185 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1186 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1187 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1188 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1189 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1190 (load addr:$src)))]>;
1192 // Aliases for intrinsics
1193 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1194 "cvttsd2si\t{$src, $dst|$dst, $src}",
1196 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1197 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1198 "cvttsd2si\t{$src, $dst|$dst, $src}",
1199 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1200 (load addr:$src)))]>;
1202 // Comparison instructions
1203 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1204 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1205 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1206 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1208 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1209 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1210 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1213 let Defs = [EFLAGS] in {
1214 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1215 "ucomisd\t{$src2, $src1|$src1, $src2}",
1216 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1217 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1218 "ucomisd\t{$src2, $src1|$src1, $src2}",
1219 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1220 (implicit EFLAGS)]>;
1221 } // Defs = [EFLAGS]
1223 // Aliases to match intrinsics which expect XMM operand(s).
1224 let Constraints = "$src1 = $dst" in {
1225 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1227 (ins VR128:$src1, VR128:$src, SSECC:$cc),
1228 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1229 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1230 VR128:$src, imm:$cc))]>;
1231 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1233 (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1234 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1235 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1236 (load addr:$src), imm:$cc))]>;
1239 let Defs = [EFLAGS] in {
1240 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1241 "ucomisd\t{$src2, $src1|$src1, $src2}",
1242 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1243 (implicit EFLAGS)]>;
1244 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1245 "ucomisd\t{$src2, $src1|$src1, $src2}",
1246 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1247 (implicit EFLAGS)]>;
1249 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1250 "comisd\t{$src2, $src1|$src1, $src2}",
1251 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1252 (implicit EFLAGS)]>;
1253 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1254 "comisd\t{$src2, $src1|$src1, $src2}",
1255 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1256 (implicit EFLAGS)]>;
1257 } // Defs = [EFLAGS]
1259 // Aliases of packed SSE2 instructions for scalar use. These all have names
1260 // that start with 'Fs'.
1262 // Alias instructions that map fld0 to pxor for sse.
1263 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1264 canFoldAsLoad = 1 in
1265 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1266 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1267 Requires<[HasSSE2]>, TB, OpSize;
1269 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1271 let neverHasSideEffects = 1 in
1272 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1273 "movapd\t{$src, $dst|$dst, $src}", []>;
1275 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1277 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1278 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1279 "movapd\t{$src, $dst|$dst, $src}",
1280 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1282 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1283 let Constraints = "$src1 = $dst" in {
1284 let isCommutable = 1 in {
1285 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1286 (ins FR64:$src1, FR64:$src2),
1287 "andpd\t{$src2, $dst|$dst, $src2}",
1288 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1289 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1290 (ins FR64:$src1, FR64:$src2),
1291 "orpd\t{$src2, $dst|$dst, $src2}",
1292 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1293 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1294 (ins FR64:$src1, FR64:$src2),
1295 "xorpd\t{$src2, $dst|$dst, $src2}",
1296 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1299 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1300 (ins FR64:$src1, f128mem:$src2),
1301 "andpd\t{$src2, $dst|$dst, $src2}",
1302 [(set FR64:$dst, (X86fand FR64:$src1,
1303 (memopfsf64 addr:$src2)))]>;
1304 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1305 (ins FR64:$src1, f128mem:$src2),
1306 "orpd\t{$src2, $dst|$dst, $src2}",
1307 [(set FR64:$dst, (X86for FR64:$src1,
1308 (memopfsf64 addr:$src2)))]>;
1309 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1310 (ins FR64:$src1, f128mem:$src2),
1311 "xorpd\t{$src2, $dst|$dst, $src2}",
1312 [(set FR64:$dst, (X86fxor FR64:$src1,
1313 (memopfsf64 addr:$src2)))]>;
1315 let neverHasSideEffects = 1 in {
1316 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1317 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1318 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1320 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1321 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1322 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1326 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1328 /// In addition, we also have a special variant of the scalar form here to
1329 /// represent the associated intrinsic operation. This form is unlike the
1330 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1331 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1333 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1334 /// six "instructions".
1336 let Constraints = "$src1 = $dst" in {
1337 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1338 SDNode OpNode, Intrinsic F64Int,
1339 bit Commutable = 0> {
1340 // Scalar operation, reg+reg.
1341 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1342 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1343 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1344 let isCommutable = Commutable;
1347 // Scalar operation, reg+mem.
1348 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1349 (ins FR64:$src1, f64mem:$src2),
1350 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1351 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1353 // Vector operation, reg+reg.
1354 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1355 (ins VR128:$src1, VR128:$src2),
1356 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1357 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1358 let isCommutable = Commutable;
1361 // Vector operation, reg+mem.
1362 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1363 (ins VR128:$src1, f128mem:$src2),
1364 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1365 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1367 // Intrinsic operation, reg+reg.
1368 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1369 (ins VR128:$src1, VR128:$src2),
1370 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1371 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]>;
1373 // Intrinsic operation, reg+mem.
1374 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1375 (ins VR128:$src1, sdmem:$src2),
1376 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1377 [(set VR128:$dst, (F64Int VR128:$src1,
1378 sse_load_f64:$src2))]>;
1382 // Arithmetic instructions
1383 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1384 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1385 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1386 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1388 /// sse2_fp_binop_rm - Other SSE2 binops
1390 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1391 /// instructions for a full-vector intrinsic form. Operations that map
1392 /// onto C operators don't use this form since they just use the plain
1393 /// vector form instead of having a separate vector intrinsic form.
1395 /// This provides a total of eight "instructions".
1397 let Constraints = "$src1 = $dst" in {
1398 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1402 bit Commutable = 0> {
1404 // Scalar operation, reg+reg.
1405 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1406 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1407 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1408 let isCommutable = Commutable;
1411 // Scalar operation, reg+mem.
1412 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1413 (ins FR64:$src1, f64mem:$src2),
1414 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1415 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1417 // Vector operation, reg+reg.
1418 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1419 (ins VR128:$src1, VR128:$src2),
1420 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1421 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1422 let isCommutable = Commutable;
1425 // Vector operation, reg+mem.
1426 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1427 (ins VR128:$src1, f128mem:$src2),
1428 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1429 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1431 // Intrinsic operation, reg+reg.
1432 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1433 (ins VR128:$src1, VR128:$src2),
1434 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1435 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1436 let isCommutable = Commutable;
1439 // Intrinsic operation, reg+mem.
1440 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1441 (ins VR128:$src1, sdmem:$src2),
1442 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1443 [(set VR128:$dst, (F64Int VR128:$src1,
1444 sse_load_f64:$src2))]>;
1446 // Vector intrinsic operation, reg+reg.
1447 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1448 (ins VR128:$src1, VR128:$src2),
1449 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1450 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1451 let isCommutable = Commutable;
1454 // Vector intrinsic operation, reg+mem.
1455 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1456 (ins VR128:$src1, f128mem:$src2),
1457 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1458 [(set VR128:$dst, (V2F64Int VR128:$src1,
1459 (memopv2f64 addr:$src2)))]>;
1463 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1464 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1465 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1466 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1468 //===---------------------------------------------------------------------===//
1469 // SSE packed FP Instructions
1471 // Move Instructions
1472 let neverHasSideEffects = 1 in
1473 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1474 "movapd\t{$src, $dst|$dst, $src}", []>;
1475 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1476 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1477 "movapd\t{$src, $dst|$dst, $src}",
1478 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1480 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1481 "movapd\t{$src, $dst|$dst, $src}",
1482 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1484 let neverHasSideEffects = 1 in
1485 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1486 "movupd\t{$src, $dst|$dst, $src}", []>;
1487 let canFoldAsLoad = 1 in
1488 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1489 "movupd\t{$src, $dst|$dst, $src}",
1490 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1491 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1492 "movupd\t{$src, $dst|$dst, $src}",
1493 [(store (v2f64 VR128:$src), addr:$dst)]>;
1495 // Intrinsic forms of MOVUPD load and store
1496 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1497 "movupd\t{$src, $dst|$dst, $src}",
1498 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1499 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1500 "movupd\t{$src, $dst|$dst, $src}",
1501 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1503 let Constraints = "$src1 = $dst" in {
1504 let AddedComplexity = 20 in {
1505 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1506 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1507 "movlpd\t{$src2, $dst|$dst, $src2}",
1509 (v2f64 (movlp VR128:$src1,
1510 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1511 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1512 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1513 "movhpd\t{$src2, $dst|$dst, $src2}",
1515 (v2f64 (movlhps VR128:$src1,
1516 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1517 } // AddedComplexity
1518 } // Constraints = "$src1 = $dst"
1520 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1521 "movlpd\t{$src, $dst|$dst, $src}",
1522 [(store (f64 (vector_extract (v2f64 VR128:$src),
1523 (iPTR 0))), addr:$dst)]>;
1525 // v2f64 extract element 1 is always custom lowered to unpack high to low
1526 // and extract element 0 so the non-store version isn't too horrible.
1527 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1528 "movhpd\t{$src, $dst|$dst, $src}",
1529 [(store (f64 (vector_extract
1530 (v2f64 (unpckh VR128:$src, (undef))),
1531 (iPTR 0))), addr:$dst)]>;
1533 // SSE2 instructions without OpSize prefix
1534 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1535 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1536 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1537 TB, Requires<[HasSSE2]>;
1538 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1539 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1540 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1541 (bitconvert (memopv2i64 addr:$src))))]>,
1542 TB, Requires<[HasSSE2]>;
1544 // SSE2 instructions with XS prefix
1545 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1546 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1547 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1548 XS, Requires<[HasSSE2]>;
1549 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1550 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1551 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1552 (bitconvert (memopv2i64 addr:$src))))]>,
1553 XS, Requires<[HasSSE2]>;
1555 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1556 "cvtps2dq\t{$src, $dst|$dst, $src}",
1557 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1558 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1559 "cvtps2dq\t{$src, $dst|$dst, $src}",
1560 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1561 (memop addr:$src)))]>;
1562 // SSE2 packed instructions with XS prefix
1563 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1564 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1565 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1566 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1568 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1569 "cvttps2dq\t{$src, $dst|$dst, $src}",
1571 (int_x86_sse2_cvttps2dq VR128:$src))]>,
1572 XS, Requires<[HasSSE2]>;
1573 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1574 "cvttps2dq\t{$src, $dst|$dst, $src}",
1575 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1576 (memop addr:$src)))]>,
1577 XS, Requires<[HasSSE2]>;
1579 // SSE2 packed instructions with XD prefix
1580 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1581 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1582 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1583 XD, Requires<[HasSSE2]>;
1584 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1585 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1586 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1587 (memop addr:$src)))]>,
1588 XD, Requires<[HasSSE2]>;
1590 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1591 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1592 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1593 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1594 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1595 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1596 (memop addr:$src)))]>;
1598 // SSE2 instructions without OpSize prefix
1599 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1600 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1601 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1602 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1604 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1605 "cvtps2pd\t{$src, $dst|$dst, $src}",
1606 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1607 TB, Requires<[HasSSE2]>;
1608 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1609 "cvtps2pd\t{$src, $dst|$dst, $src}",
1610 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1611 (load addr:$src)))]>,
1612 TB, Requires<[HasSSE2]>;
1614 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1615 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1616 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1617 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1620 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1621 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1622 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1623 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1624 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1625 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1626 (memop addr:$src)))]>;
1628 // Match intrinsics which expect XMM operand(s).
1629 // Aliases for intrinsics
1630 let Constraints = "$src1 = $dst" in {
1631 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1632 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1633 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1634 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1636 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1637 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1638 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1639 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1640 (loadi32 addr:$src2)))]>;
1641 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1642 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1643 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1644 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1646 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1647 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1648 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1649 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1650 (load addr:$src2)))]>;
1651 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1652 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1653 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1654 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1655 VR128:$src2))]>, XS,
1656 Requires<[HasSSE2]>;
1657 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1658 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1659 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1660 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1661 (load addr:$src2)))]>, XS,
1662 Requires<[HasSSE2]>;
1667 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1669 /// In addition, we also have a special variant of the scalar form here to
1670 /// represent the associated intrinsic operation. This form is unlike the
1671 /// plain scalar form, in that it takes an entire vector (instead of a
1672 /// scalar) and leaves the top elements undefined.
1674 /// And, we have a special variant form for a full-vector intrinsic form.
1676 /// These four forms can each have a reg or a mem operand, so there are a
1677 /// total of eight "instructions".
1679 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1683 bit Commutable = 0> {
1684 // Scalar operation, reg.
1685 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1686 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1687 [(set FR64:$dst, (OpNode FR64:$src))]> {
1688 let isCommutable = Commutable;
1691 // Scalar operation, mem.
1692 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1693 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1694 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1696 // Vector operation, reg.
1697 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1698 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1699 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1700 let isCommutable = Commutable;
1703 // Vector operation, mem.
1704 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1705 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1706 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1708 // Intrinsic operation, reg.
1709 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1710 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1711 [(set VR128:$dst, (F64Int VR128:$src))]> {
1712 let isCommutable = Commutable;
1715 // Intrinsic operation, mem.
1716 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1717 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1718 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1720 // Vector intrinsic operation, reg
1721 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1722 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1723 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1724 let isCommutable = Commutable;
1727 // Vector intrinsic operation, mem
1728 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1729 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1730 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1734 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1735 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1737 // There is no f64 version of the reciprocal approximation instructions.
1740 let Constraints = "$src1 = $dst" in {
1741 let isCommutable = 1 in {
1742 def ANDPDrr : PDI<0x54, MRMSrcReg,
1743 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1744 "andpd\t{$src2, $dst|$dst, $src2}",
1746 (and (bc_v2i64 (v2f64 VR128:$src1)),
1747 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1748 def ORPDrr : PDI<0x56, MRMSrcReg,
1749 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1750 "orpd\t{$src2, $dst|$dst, $src2}",
1752 (or (bc_v2i64 (v2f64 VR128:$src1)),
1753 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1754 def XORPDrr : PDI<0x57, MRMSrcReg,
1755 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1756 "xorpd\t{$src2, $dst|$dst, $src2}",
1758 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1759 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1762 def ANDPDrm : PDI<0x54, MRMSrcMem,
1763 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1764 "andpd\t{$src2, $dst|$dst, $src2}",
1766 (and (bc_v2i64 (v2f64 VR128:$src1)),
1767 (memopv2i64 addr:$src2)))]>;
1768 def ORPDrm : PDI<0x56, MRMSrcMem,
1769 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1770 "orpd\t{$src2, $dst|$dst, $src2}",
1772 (or (bc_v2i64 (v2f64 VR128:$src1)),
1773 (memopv2i64 addr:$src2)))]>;
1774 def XORPDrm : PDI<0x57, MRMSrcMem,
1775 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1776 "xorpd\t{$src2, $dst|$dst, $src2}",
1778 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1779 (memopv2i64 addr:$src2)))]>;
1780 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1781 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1782 "andnpd\t{$src2, $dst|$dst, $src2}",
1784 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1785 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1786 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1787 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1788 "andnpd\t{$src2, $dst|$dst, $src2}",
1790 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1791 (memopv2i64 addr:$src2)))]>;
1794 let Constraints = "$src1 = $dst" in {
1795 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1796 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1797 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1798 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1799 VR128:$src, imm:$cc))]>;
1800 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1801 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1802 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1803 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1804 (memop addr:$src), imm:$cc))]>;
1806 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1807 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1808 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1809 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1811 // Shuffle and unpack instructions
1812 let Constraints = "$src1 = $dst" in {
1813 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1814 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1815 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1817 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1818 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1819 (outs VR128:$dst), (ins VR128:$src1,
1820 f128mem:$src2, i8imm:$src3),
1821 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1824 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1826 let AddedComplexity = 10 in {
1827 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1828 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1829 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1831 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1832 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1833 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1834 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1836 (v2f64 (unpckh VR128:$src1,
1837 (memopv2f64 addr:$src2))))]>;
1839 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1840 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1841 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1843 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1844 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1845 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1846 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1848 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1849 } // AddedComplexity
1850 } // Constraints = "$src1 = $dst"
1853 //===---------------------------------------------------------------------===//
1854 // SSE integer instructions
1856 // Move Instructions
1857 let neverHasSideEffects = 1 in
1858 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1859 "movdqa\t{$src, $dst|$dst, $src}", []>;
1860 let canFoldAsLoad = 1, mayLoad = 1 in
1861 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1862 "movdqa\t{$src, $dst|$dst, $src}",
1863 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1865 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1866 "movdqa\t{$src, $dst|$dst, $src}",
1867 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1868 let canFoldAsLoad = 1, mayLoad = 1 in
1869 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1870 "movdqu\t{$src, $dst|$dst, $src}",
1871 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1872 XS, Requires<[HasSSE2]>;
1874 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1875 "movdqu\t{$src, $dst|$dst, $src}",
1876 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1877 XS, Requires<[HasSSE2]>;
1879 // Intrinsic forms of MOVDQU load and store
1880 let canFoldAsLoad = 1 in
1881 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1882 "movdqu\t{$src, $dst|$dst, $src}",
1883 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1884 XS, Requires<[HasSSE2]>;
1885 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1886 "movdqu\t{$src, $dst|$dst, $src}",
1887 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1888 XS, Requires<[HasSSE2]>;
1890 let Constraints = "$src1 = $dst" in {
1892 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1893 bit Commutable = 0> {
1894 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1895 (ins VR128:$src1, VR128:$src2),
1896 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1897 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1898 let isCommutable = Commutable;
1900 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1901 (ins VR128:$src1, i128mem:$src2),
1902 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1903 [(set VR128:$dst, (IntId VR128:$src1,
1904 (bitconvert (memopv2i64
1908 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1910 Intrinsic IntId, Intrinsic IntId2> {
1911 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1912 (ins VR128:$src1, VR128:$src2),
1913 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1914 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1915 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1916 (ins VR128:$src1, i128mem:$src2),
1917 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1918 [(set VR128:$dst, (IntId VR128:$src1,
1919 (bitconvert (memopv2i64 addr:$src2))))]>;
1920 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
1921 (ins VR128:$src1, i32i8imm:$src2),
1922 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1923 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1926 /// PDI_binop_rm - Simple SSE2 binary operator.
1927 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1928 ValueType OpVT, bit Commutable = 0> {
1929 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1930 (ins VR128:$src1, VR128:$src2),
1931 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1932 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1933 let isCommutable = Commutable;
1935 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1936 (ins VR128:$src1, i128mem:$src2),
1937 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1938 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1939 (bitconvert (memopv2i64 addr:$src2)))))]>;
1942 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1944 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1945 /// to collapse (bitconvert VT to VT) into its operand.
1947 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1948 bit Commutable = 0> {
1949 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1950 (ins VR128:$src1, VR128:$src2),
1951 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1952 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1953 let isCommutable = Commutable;
1955 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1956 (ins VR128:$src1, i128mem:$src2),
1957 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1958 [(set VR128:$dst, (OpNode VR128:$src1,
1959 (memopv2i64 addr:$src2)))]>;
1962 } // Constraints = "$src1 = $dst"
1964 // 128-bit Integer Arithmetic
1966 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1967 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1968 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1969 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1971 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1972 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1973 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1974 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1976 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1977 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1978 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1979 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1981 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1982 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1983 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1984 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1986 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1988 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1989 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1990 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1992 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1994 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1995 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1998 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1999 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
2000 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
2001 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
2002 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
2005 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
2006 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
2007 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
2008 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
2009 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
2010 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
2012 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2013 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2014 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2015 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2016 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2017 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2019 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2020 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2021 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2022 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2024 // 128-bit logical shifts.
2025 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
2026 def PSLLDQri : PDIi8<0x73, MRM7r,
2027 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2028 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2029 def PSRLDQri : PDIi8<0x73, MRM3r,
2030 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2031 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2032 // PSRADQri doesn't exist in SSE[1-3].
2035 let Predicates = [HasSSE2] in {
2036 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2037 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2038 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2039 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2040 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2041 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2042 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2043 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2044 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2045 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2047 // Shift up / down and insert zero's.
2048 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2049 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2050 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2051 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2055 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2056 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2057 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2059 let Constraints = "$src1 = $dst" in {
2060 def PANDNrr : PDI<0xDF, MRMSrcReg,
2061 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2062 "pandn\t{$src2, $dst|$dst, $src2}",
2063 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2066 def PANDNrm : PDI<0xDF, MRMSrcMem,
2067 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2068 "pandn\t{$src2, $dst|$dst, $src2}",
2069 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2070 (memopv2i64 addr:$src2))))]>;
2073 // SSE2 Integer comparison
2074 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2075 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2076 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2077 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2078 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2079 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2081 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2082 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2083 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2084 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2085 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2086 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2087 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2088 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2089 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2090 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2091 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2092 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2094 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2095 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2096 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2097 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2098 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2099 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2100 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2101 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2102 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2103 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2104 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2105 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2108 // Pack instructions
2109 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2110 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2111 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2113 // Shuffle and unpack instructions
2114 let AddedComplexity = 5 in {
2115 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2116 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2117 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2118 [(set VR128:$dst, (v4i32 (pshufd:$src2
2119 VR128:$src1, (undef))))]>;
2120 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2121 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2122 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2123 [(set VR128:$dst, (v4i32 (pshufd:$src2
2124 (bc_v4i32 (memopv2i64 addr:$src1)),
2128 // SSE2 with ImmT == Imm8 and XS prefix.
2129 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2130 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2131 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2132 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2134 XS, Requires<[HasSSE2]>;
2135 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2136 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2137 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2138 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2139 (bc_v8i16 (memopv2i64 addr:$src1)),
2141 XS, Requires<[HasSSE2]>;
2143 // SSE2 with ImmT == Imm8 and XD prefix.
2144 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2145 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2146 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2147 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2149 XD, Requires<[HasSSE2]>;
2150 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2151 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2152 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2153 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2154 (bc_v8i16 (memopv2i64 addr:$src1)),
2156 XD, Requires<[HasSSE2]>;
2159 let Constraints = "$src1 = $dst" in {
2160 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2161 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2162 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2164 (v16i8 (unpckl VR128:$src1, VR128:$src2)))]>;
2165 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2166 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2167 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2169 (unpckl VR128:$src1,
2170 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2171 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2172 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2173 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2175 (v8i16 (unpckl VR128:$src1, VR128:$src2)))]>;
2176 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2177 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2178 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2180 (unpckl VR128:$src1,
2181 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2182 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2183 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2184 "punpckldq\t{$src2, $dst|$dst, $src2}",
2186 (v4i32 (unpckl VR128:$src1, VR128:$src2)))]>;
2187 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2188 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2189 "punpckldq\t{$src2, $dst|$dst, $src2}",
2191 (unpckl VR128:$src1,
2192 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2193 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2194 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2195 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2197 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2198 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2199 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2200 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2202 (v2i64 (unpckl VR128:$src1,
2203 (memopv2i64 addr:$src2))))]>;
2205 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2206 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2207 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2209 (v16i8 (unpckh VR128:$src1, VR128:$src2)))]>;
2210 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2211 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2212 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2214 (unpckh VR128:$src1,
2215 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2216 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2217 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2218 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2220 (v8i16 (unpckh VR128:$src1, VR128:$src2)))]>;
2221 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2222 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2223 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2225 (unpckh VR128:$src1,
2226 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2227 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2228 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2229 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2231 (v4i32 (unpckh VR128:$src1, VR128:$src2)))]>;
2232 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2233 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2234 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2236 (unpckh VR128:$src1,
2237 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2238 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2239 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2240 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2242 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2243 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2244 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2245 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2247 (v2i64 (unpckh VR128:$src1,
2248 (memopv2i64 addr:$src2))))]>;
2252 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2253 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2254 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2255 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2257 let Constraints = "$src1 = $dst" in {
2258 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2259 (outs VR128:$dst), (ins VR128:$src1,
2260 GR32:$src2, i32i8imm:$src3),
2261 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2263 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2264 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2265 (outs VR128:$dst), (ins VR128:$src1,
2266 i16mem:$src2, i32i8imm:$src3),
2267 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2269 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2274 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2275 "pmovmskb\t{$src, $dst|$dst, $src}",
2276 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2278 // Conditional store
2280 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2281 "maskmovdqu\t{$mask, $src|$src, $mask}",
2282 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2285 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2286 "maskmovdqu\t{$mask, $src|$src, $mask}",
2287 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2289 // Non-temporal stores
2290 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2291 "movntpd\t{$src, $dst|$dst, $src}",
2292 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2293 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2294 "movntdq\t{$src, $dst|$dst, $src}",
2295 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2296 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2297 "movnti\t{$src, $dst|$dst, $src}",
2298 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2299 TB, Requires<[HasSSE2]>;
2302 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2303 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2304 TB, Requires<[HasSSE2]>;
2306 // Load, store, and memory fence
2307 def LFENCE : I<0xAE, MRM5r, (outs), (ins),
2308 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2309 def MFENCE : I<0xAE, MRM6r, (outs), (ins),
2310 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2312 //TODO: custom lower this so as to never even generate the noop
2313 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2315 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2316 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2317 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2320 // Alias instructions that map zero vector to pxor / xorp* for sse.
2321 // We set canFoldAsLoad because this can be converted to a constant-pool
2322 // load of an all-ones value if folding it would be beneficial.
2323 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2324 isCodeGenOnly = 1 in
2325 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2326 "pcmpeqd\t$dst, $dst",
2327 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2329 // FR64 to 128-bit vector conversion.
2330 let isAsCheapAsAMove = 1 in
2331 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2332 "movsd\t{$src, $dst|$dst, $src}",
2334 (v2f64 (scalar_to_vector FR64:$src)))]>;
2335 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2336 "movsd\t{$src, $dst|$dst, $src}",
2338 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2340 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2341 "movd\t{$src, $dst|$dst, $src}",
2343 (v4i32 (scalar_to_vector GR32:$src)))]>;
2344 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2345 "movd\t{$src, $dst|$dst, $src}",
2347 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2349 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2350 "movd\t{$src, $dst|$dst, $src}",
2351 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2353 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2354 "movd\t{$src, $dst|$dst, $src}",
2355 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2357 // SSE2 instructions with XS prefix
2358 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2359 "movq\t{$src, $dst|$dst, $src}",
2361 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2362 Requires<[HasSSE2]>;
2363 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2364 "movq\t{$src, $dst|$dst, $src}",
2365 [(store (i64 (vector_extract (v2i64 VR128:$src),
2366 (iPTR 0))), addr:$dst)]>;
2368 // FIXME: may not be able to eliminate this movss with coalescing the src and
2369 // dest register classes are different. We really want to write this pattern
2371 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2372 // (f32 FR32:$src)>;
2373 let isAsCheapAsAMove = 1 in
2374 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2375 "movsd\t{$src, $dst|$dst, $src}",
2376 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2378 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2379 "movsd\t{$src, $dst|$dst, $src}",
2380 [(store (f64 (vector_extract (v2f64 VR128:$src),
2381 (iPTR 0))), addr:$dst)]>;
2382 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2383 "movd\t{$src, $dst|$dst, $src}",
2384 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2386 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2387 "movd\t{$src, $dst|$dst, $src}",
2388 [(store (i32 (vector_extract (v4i32 VR128:$src),
2389 (iPTR 0))), addr:$dst)]>;
2391 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2392 "movd\t{$src, $dst|$dst, $src}",
2393 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2394 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2395 "movd\t{$src, $dst|$dst, $src}",
2396 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2399 // Move to lower bits of a VR128, leaving upper bits alone.
2400 // Three operand (but two address) aliases.
2401 let Constraints = "$src1 = $dst" in {
2402 let neverHasSideEffects = 1 in
2403 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2404 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2405 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2407 let AddedComplexity = 15 in
2408 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2409 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2410 "movsd\t{$src2, $dst|$dst, $src2}",
2412 (v2f64 (movl VR128:$src1, VR128:$src2)))]>;
2415 // Store / copy lower 64-bits of a XMM register.
2416 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2417 "movq\t{$src, $dst|$dst, $src}",
2418 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2420 // Move to lower bits of a VR128 and zeroing upper bits.
2421 // Loading from memory automatically zeroing upper bits.
2422 let AddedComplexity = 20 in {
2423 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2424 "movsd\t{$src, $dst|$dst, $src}",
2426 (v2f64 (X86vzmovl (v2f64 (scalar_to_vector
2427 (loadf64 addr:$src))))))]>;
2429 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
2430 (MOVZSD2PDrm addr:$src)>;
2431 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
2432 (MOVZSD2PDrm addr:$src)>;
2433 def : Pat<(v2f64 (X86vzload addr:$src)), (MOVZSD2PDrm addr:$src)>;
2436 // movd / movq to XMM register zero-extends
2437 let AddedComplexity = 15 in {
2438 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2439 "movd\t{$src, $dst|$dst, $src}",
2440 [(set VR128:$dst, (v4i32 (X86vzmovl
2441 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2442 // This is X86-64 only.
2443 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2444 "mov{d|q}\t{$src, $dst|$dst, $src}",
2445 [(set VR128:$dst, (v2i64 (X86vzmovl
2446 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2449 let AddedComplexity = 20 in {
2450 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2451 "movd\t{$src, $dst|$dst, $src}",
2453 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2454 (loadi32 addr:$src))))))]>;
2456 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2457 (MOVZDI2PDIrm addr:$src)>;
2458 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2459 (MOVZDI2PDIrm addr:$src)>;
2460 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2461 (MOVZDI2PDIrm addr:$src)>;
2463 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2464 "movq\t{$src, $dst|$dst, $src}",
2466 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2467 (loadi64 addr:$src))))))]>, XS,
2468 Requires<[HasSSE2]>;
2470 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2471 (MOVZQI2PQIrm addr:$src)>;
2472 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2473 (MOVZQI2PQIrm addr:$src)>;
2474 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2477 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2478 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2479 let AddedComplexity = 15 in
2480 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2481 "movq\t{$src, $dst|$dst, $src}",
2482 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2483 XS, Requires<[HasSSE2]>;
2485 let AddedComplexity = 20 in {
2486 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2487 "movq\t{$src, $dst|$dst, $src}",
2488 [(set VR128:$dst, (v2i64 (X86vzmovl
2489 (loadv2i64 addr:$src))))]>,
2490 XS, Requires<[HasSSE2]>;
2492 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2493 (MOVZPQILo2PQIrm addr:$src)>;
2496 // Instructions for the disassembler
2497 // xr = XMM register
2500 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2501 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2503 //===---------------------------------------------------------------------===//
2504 // SSE3 Instructions
2505 //===---------------------------------------------------------------------===//
2507 // Move Instructions
2508 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2509 "movshdup\t{$src, $dst|$dst, $src}",
2510 [(set VR128:$dst, (v4f32 (movshdup
2511 VR128:$src, (undef))))]>;
2512 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2513 "movshdup\t{$src, $dst|$dst, $src}",
2514 [(set VR128:$dst, (movshdup
2515 (memopv4f32 addr:$src), (undef)))]>;
2517 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2518 "movsldup\t{$src, $dst|$dst, $src}",
2519 [(set VR128:$dst, (v4f32 (movsldup
2520 VR128:$src, (undef))))]>;
2521 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2522 "movsldup\t{$src, $dst|$dst, $src}",
2523 [(set VR128:$dst, (movsldup
2524 (memopv4f32 addr:$src), (undef)))]>;
2526 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2527 "movddup\t{$src, $dst|$dst, $src}",
2528 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2529 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2530 "movddup\t{$src, $dst|$dst, $src}",
2532 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2535 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2537 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2539 let AddedComplexity = 5 in {
2540 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2541 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2542 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2543 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2544 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2545 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2546 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2547 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2551 let Constraints = "$src1 = $dst" in {
2552 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2553 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2554 "addsubps\t{$src2, $dst|$dst, $src2}",
2555 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2557 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2558 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2559 "addsubps\t{$src2, $dst|$dst, $src2}",
2560 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2561 (memop addr:$src2)))]>;
2562 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2563 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2564 "addsubpd\t{$src2, $dst|$dst, $src2}",
2565 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2567 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2568 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2569 "addsubpd\t{$src2, $dst|$dst, $src2}",
2570 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2571 (memop addr:$src2)))]>;
2574 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2575 "lddqu\t{$src, $dst|$dst, $src}",
2576 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2579 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2580 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2581 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2582 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2583 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2584 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2585 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2586 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2587 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2588 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2589 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2590 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2591 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2592 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2593 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2594 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2596 let Constraints = "$src1 = $dst" in {
2597 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2598 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2599 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2600 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2601 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2602 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2603 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2604 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2607 // Thread synchronization
2608 def MONITOR : I<0x01, MRM1r, (outs), (ins), "monitor",
2609 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2610 def MWAIT : I<0x01, MRM1r, (outs), (ins), "mwait",
2611 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2613 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2614 let AddedComplexity = 15 in
2615 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2616 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2617 let AddedComplexity = 20 in
2618 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2619 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2621 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2622 let AddedComplexity = 15 in
2623 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2624 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2625 let AddedComplexity = 20 in
2626 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2627 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2629 //===---------------------------------------------------------------------===//
2630 // SSSE3 Instructions
2631 //===---------------------------------------------------------------------===//
2633 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2634 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2635 Intrinsic IntId64, Intrinsic IntId128> {
2636 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2637 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2638 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2640 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2641 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2643 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2645 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2647 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2648 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2651 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2653 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2656 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2659 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2660 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2661 Intrinsic IntId64, Intrinsic IntId128> {
2662 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2664 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2665 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2667 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2669 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2672 (bitconvert (memopv4i16 addr:$src))))]>;
2674 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2676 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2677 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2680 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2682 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2685 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2688 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2689 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2690 Intrinsic IntId64, Intrinsic IntId128> {
2691 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2693 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2694 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2696 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2698 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2701 (bitconvert (memopv2i32 addr:$src))))]>;
2703 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2705 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2706 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2709 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2711 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2714 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2717 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2718 int_x86_ssse3_pabs_b,
2719 int_x86_ssse3_pabs_b_128>;
2720 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2721 int_x86_ssse3_pabs_w,
2722 int_x86_ssse3_pabs_w_128>;
2723 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2724 int_x86_ssse3_pabs_d,
2725 int_x86_ssse3_pabs_d_128>;
2727 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2728 let Constraints = "$src1 = $dst" in {
2729 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2730 Intrinsic IntId64, Intrinsic IntId128,
2731 bit Commutable = 0> {
2732 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2733 (ins VR64:$src1, VR64:$src2),
2734 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2735 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2736 let isCommutable = Commutable;
2738 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2739 (ins VR64:$src1, i64mem:$src2),
2740 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2742 (IntId64 VR64:$src1,
2743 (bitconvert (memopv8i8 addr:$src2))))]>;
2745 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2746 (ins VR128:$src1, VR128:$src2),
2747 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2748 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2750 let isCommutable = Commutable;
2752 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2753 (ins VR128:$src1, i128mem:$src2),
2754 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2756 (IntId128 VR128:$src1,
2757 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2761 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2762 let Constraints = "$src1 = $dst" in {
2763 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2764 Intrinsic IntId64, Intrinsic IntId128,
2765 bit Commutable = 0> {
2766 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2767 (ins VR64:$src1, VR64:$src2),
2768 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2769 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2770 let isCommutable = Commutable;
2772 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2773 (ins VR64:$src1, i64mem:$src2),
2774 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2776 (IntId64 VR64:$src1,
2777 (bitconvert (memopv4i16 addr:$src2))))]>;
2779 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2780 (ins VR128:$src1, VR128:$src2),
2781 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2782 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2784 let isCommutable = Commutable;
2786 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2787 (ins VR128:$src1, i128mem:$src2),
2788 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2790 (IntId128 VR128:$src1,
2791 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2795 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2796 let Constraints = "$src1 = $dst" in {
2797 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2798 Intrinsic IntId64, Intrinsic IntId128,
2799 bit Commutable = 0> {
2800 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2801 (ins VR64:$src1, VR64:$src2),
2802 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2803 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2804 let isCommutable = Commutable;
2806 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2807 (ins VR64:$src1, i64mem:$src2),
2808 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2810 (IntId64 VR64:$src1,
2811 (bitconvert (memopv2i32 addr:$src2))))]>;
2813 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2814 (ins VR128:$src1, VR128:$src2),
2815 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2816 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2818 let isCommutable = Commutable;
2820 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2821 (ins VR128:$src1, i128mem:$src2),
2822 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2824 (IntId128 VR128:$src1,
2825 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2829 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2830 int_x86_ssse3_phadd_w,
2831 int_x86_ssse3_phadd_w_128>;
2832 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2833 int_x86_ssse3_phadd_d,
2834 int_x86_ssse3_phadd_d_128>;
2835 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2836 int_x86_ssse3_phadd_sw,
2837 int_x86_ssse3_phadd_sw_128>;
2838 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2839 int_x86_ssse3_phsub_w,
2840 int_x86_ssse3_phsub_w_128>;
2841 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2842 int_x86_ssse3_phsub_d,
2843 int_x86_ssse3_phsub_d_128>;
2844 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2845 int_x86_ssse3_phsub_sw,
2846 int_x86_ssse3_phsub_sw_128>;
2847 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2848 int_x86_ssse3_pmadd_ub_sw,
2849 int_x86_ssse3_pmadd_ub_sw_128>;
2850 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2851 int_x86_ssse3_pmul_hr_sw,
2852 int_x86_ssse3_pmul_hr_sw_128, 1>;
2853 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2854 int_x86_ssse3_pshuf_b,
2855 int_x86_ssse3_pshuf_b_128>;
2856 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2857 int_x86_ssse3_psign_b,
2858 int_x86_ssse3_psign_b_128>;
2859 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2860 int_x86_ssse3_psign_w,
2861 int_x86_ssse3_psign_w_128>;
2862 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2863 int_x86_ssse3_psign_d,
2864 int_x86_ssse3_psign_d_128>;
2866 let Constraints = "$src1 = $dst" in {
2867 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2868 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2869 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2871 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2872 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2873 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2876 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2877 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2878 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2880 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2881 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2882 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2886 // palignr patterns.
2887 def : Pat<(int_x86_ssse3_palign_r VR64:$src1, VR64:$src2, (i8 imm:$src3)),
2888 (PALIGNR64rr VR64:$src1, VR64:$src2, (BYTE_imm imm:$src3))>,
2889 Requires<[HasSSSE3]>;
2890 def : Pat<(int_x86_ssse3_palign_r VR64:$src1,
2891 (memop64 addr:$src2),
2893 (PALIGNR64rm VR64:$src1, addr:$src2, (BYTE_imm imm:$src3))>,
2894 Requires<[HasSSSE3]>;
2896 def : Pat<(int_x86_ssse3_palign_r_128 VR128:$src1, VR128:$src2, (i8 imm:$src3)),
2897 (PALIGNR128rr VR128:$src1, VR128:$src2, (BYTE_imm imm:$src3))>,
2898 Requires<[HasSSSE3]>;
2899 def : Pat<(int_x86_ssse3_palign_r_128 VR128:$src1,
2900 (memopv2i64 addr:$src2),
2902 (PALIGNR128rm VR128:$src1, addr:$src2, (BYTE_imm imm:$src3))>,
2903 Requires<[HasSSSE3]>;
2905 let AddedComplexity = 5 in {
2906 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2907 (PALIGNR128rr VR128:$src2, VR128:$src1,
2908 (SHUFFLE_get_palign_imm VR128:$src3))>,
2909 Requires<[HasSSSE3]>;
2910 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2911 (PALIGNR128rr VR128:$src2, VR128:$src1,
2912 (SHUFFLE_get_palign_imm VR128:$src3))>,
2913 Requires<[HasSSSE3]>;
2914 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2915 (PALIGNR128rr VR128:$src2, VR128:$src1,
2916 (SHUFFLE_get_palign_imm VR128:$src3))>,
2917 Requires<[HasSSSE3]>;
2918 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2919 (PALIGNR128rr VR128:$src2, VR128:$src1,
2920 (SHUFFLE_get_palign_imm VR128:$src3))>,
2921 Requires<[HasSSSE3]>;
2924 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2925 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2926 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2927 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2929 //===---------------------------------------------------------------------===//
2930 // Non-Instruction Patterns
2931 //===---------------------------------------------------------------------===//
2933 // extload f32 -> f64. This matches load+fextend because we have a hack in
2934 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2936 // Since these loads aren't folded into the fextend, we have to match it
2938 let Predicates = [HasSSE2] in
2939 def : Pat<(fextend (loadf32 addr:$src)),
2940 (CVTSS2SDrm addr:$src)>;
2943 let Predicates = [HasSSE2] in {
2944 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2945 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2946 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2947 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2948 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2949 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2950 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2951 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2952 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2953 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2954 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2955 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2956 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2957 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2958 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2959 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2960 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2961 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2962 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2963 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2964 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2965 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2966 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2967 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2968 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2969 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2970 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2971 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2972 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2973 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2976 // Move scalar to XMM zero-extended
2977 // movd to XMM register zero-extends
2978 let AddedComplexity = 15 in {
2979 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2980 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2981 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2982 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2983 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE1]>;
2984 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2985 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2986 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2987 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2990 // Splat v2f64 / v2i64
2991 let AddedComplexity = 10 in {
2992 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
2993 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2994 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
2995 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2996 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
2997 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2998 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
2999 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3002 // Special unary SHUFPSrri case.
3003 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
3004 (SHUFPSrri VR128:$src1, VR128:$src1,
3005 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3006 Requires<[HasSSE1]>;
3007 let AddedComplexity = 5 in
3008 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
3009 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3010 Requires<[HasSSE2]>;
3011 // Special unary SHUFPDrri case.
3012 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
3013 (SHUFPDrri VR128:$src1, VR128:$src1,
3014 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3015 Requires<[HasSSE2]>;
3016 // Special unary SHUFPDrri case.
3017 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
3018 (SHUFPDrri VR128:$src1, VR128:$src1,
3019 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3020 Requires<[HasSSE2]>;
3021 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
3022 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
3023 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3024 Requires<[HasSSE2]>;
3026 // Special binary v4i32 shuffle cases with SHUFPS.
3027 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
3028 (SHUFPSrri VR128:$src1, VR128:$src2,
3029 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3030 Requires<[HasSSE2]>;
3031 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
3032 (SHUFPSrmi VR128:$src1, addr:$src2,
3033 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3034 Requires<[HasSSE2]>;
3035 // Special binary v2i64 shuffle cases using SHUFPDrri.
3036 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
3037 (SHUFPDrri VR128:$src1, VR128:$src2,
3038 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3039 Requires<[HasSSE2]>;
3041 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
3042 let AddedComplexity = 15 in {
3043 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
3044 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3045 Requires<[OptForSpeed, HasSSE2]>;
3046 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
3047 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3048 Requires<[OptForSpeed, HasSSE2]>;
3050 let AddedComplexity = 10 in {
3051 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
3052 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
3053 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
3054 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3055 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
3056 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3057 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
3058 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3061 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
3062 let AddedComplexity = 15 in {
3063 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
3064 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3065 Requires<[OptForSpeed, HasSSE2]>;
3066 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
3067 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3068 Requires<[OptForSpeed, HasSSE2]>;
3070 let AddedComplexity = 10 in {
3071 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
3072 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
3073 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
3074 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3075 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
3076 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3077 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
3078 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3081 let AddedComplexity = 20 in {
3082 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3083 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3084 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3086 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3087 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3088 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3090 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3091 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3092 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3093 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3094 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3097 let AddedComplexity = 20 in {
3098 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3099 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3100 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3101 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3102 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3103 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3104 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3105 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3106 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3109 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3110 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3111 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3112 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3113 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3114 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3116 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3117 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3118 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3120 let AddedComplexity = 15 in {
3121 // Setting the lowest element in the vector.
3122 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3123 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3124 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3125 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3127 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
3128 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3129 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3130 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3131 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3134 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3135 // fall back to this for SSE1)
3136 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3137 (SHUFPSrri VR128:$src2, VR128:$src1,
3138 (SHUFFLE_get_shuf_imm VR128:$src3))>, Requires<[HasSSE1]>;
3140 // Set lowest element and zero upper elements.
3141 let AddedComplexity = 15 in
3142 def : Pat<(v2f64 (movl immAllZerosV_bc, VR128:$src)),
3143 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3144 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3145 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3147 // Some special case pandn patterns.
3148 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3150 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3151 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3153 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3154 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3156 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3158 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3159 (memop addr:$src2))),
3160 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3161 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3162 (memop addr:$src2))),
3163 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3164 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3165 (memop addr:$src2))),
3166 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3168 // vector -> vector casts
3169 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3170 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3171 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3172 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3173 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3174 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3175 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3176 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3178 // Use movaps / movups for SSE integer load / store (one byte shorter).
3179 def : Pat<(alignedloadv4i32 addr:$src),
3180 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
3181 def : Pat<(loadv4i32 addr:$src),
3182 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3183 def : Pat<(alignedloadv2i64 addr:$src),
3184 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3185 def : Pat<(loadv2i64 addr:$src),
3186 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3188 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3189 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3190 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3191 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3192 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3193 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3194 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3195 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3196 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3197 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3198 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3199 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3200 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3201 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3202 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3203 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3205 //===----------------------------------------------------------------------===//
3206 // SSE4.1 Instructions
3207 //===----------------------------------------------------------------------===//
3209 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3212 Intrinsic V2F64Int> {
3213 // Intrinsic operation, reg.
3214 // Vector intrinsic operation, reg
3215 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3216 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3217 !strconcat(OpcodeStr,
3218 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3219 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3222 // Vector intrinsic operation, mem
3223 def PSm_Int : SS4AIi8<opcps, MRMSrcMem,
3224 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3225 !strconcat(OpcodeStr,
3226 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3228 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3231 // Vector intrinsic operation, reg
3232 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3233 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3234 !strconcat(OpcodeStr,
3235 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3236 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3239 // Vector intrinsic operation, mem
3240 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3241 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3242 !strconcat(OpcodeStr,
3243 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3245 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3249 let Constraints = "$src1 = $dst" in {
3250 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3254 // Intrinsic operation, reg.
3255 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3257 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3258 !strconcat(OpcodeStr,
3259 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3261 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3264 // Intrinsic operation, mem.
3265 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3267 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3268 !strconcat(OpcodeStr,
3269 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3271 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3274 // Intrinsic operation, reg.
3275 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3277 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3278 !strconcat(OpcodeStr,
3279 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3281 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3284 // Intrinsic operation, mem.
3285 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3287 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3288 !strconcat(OpcodeStr,
3289 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3291 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3296 // FP round - roundss, roundps, roundsd, roundpd
3297 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3298 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3299 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3300 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3302 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3303 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3304 Intrinsic IntId128> {
3305 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3307 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3308 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3309 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3311 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3314 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3317 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3318 int_x86_sse41_phminposuw>;
3320 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3321 let Constraints = "$src1 = $dst" in {
3322 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3323 Intrinsic IntId128, bit Commutable = 0> {
3324 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3325 (ins VR128:$src1, VR128:$src2),
3326 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3327 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3329 let isCommutable = Commutable;
3331 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3332 (ins VR128:$src1, i128mem:$src2),
3333 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3335 (IntId128 VR128:$src1,
3336 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3340 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3341 int_x86_sse41_pcmpeqq, 1>;
3342 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3343 int_x86_sse41_packusdw, 0>;
3344 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3345 int_x86_sse41_pminsb, 1>;
3346 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3347 int_x86_sse41_pminsd, 1>;
3348 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3349 int_x86_sse41_pminud, 1>;
3350 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3351 int_x86_sse41_pminuw, 1>;
3352 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3353 int_x86_sse41_pmaxsb, 1>;
3354 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3355 int_x86_sse41_pmaxsd, 1>;
3356 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3357 int_x86_sse41_pmaxud, 1>;
3358 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3359 int_x86_sse41_pmaxuw, 1>;
3361 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3363 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3364 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3365 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3366 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3368 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3369 let Constraints = "$src1 = $dst" in {
3370 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3371 SDNode OpNode, Intrinsic IntId128,
3372 bit Commutable = 0> {
3373 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3374 (ins VR128:$src1, VR128:$src2),
3375 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3376 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3377 VR128:$src2))]>, OpSize {
3378 let isCommutable = Commutable;
3380 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3381 (ins VR128:$src1, VR128:$src2),
3382 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3383 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3385 let isCommutable = Commutable;
3387 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3388 (ins VR128:$src1, i128mem:$src2),
3389 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3391 (OpNode VR128:$src1, (memop addr:$src2)))]>, OpSize;
3392 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3393 (ins VR128:$src1, i128mem:$src2),
3394 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3396 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3400 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", v4i32, mul,
3401 int_x86_sse41_pmulld, 1>;
3403 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3404 let Constraints = "$src1 = $dst" in {
3405 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3406 Intrinsic IntId128, bit Commutable = 0> {
3407 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3408 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3409 !strconcat(OpcodeStr,
3410 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3412 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3414 let isCommutable = Commutable;
3416 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3417 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3418 !strconcat(OpcodeStr,
3419 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3421 (IntId128 VR128:$src1,
3422 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3427 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3428 int_x86_sse41_blendps, 0>;
3429 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3430 int_x86_sse41_blendpd, 0>;
3431 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3432 int_x86_sse41_pblendw, 0>;
3433 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3434 int_x86_sse41_dpps, 1>;
3435 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3436 int_x86_sse41_dppd, 1>;
3437 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3438 int_x86_sse41_mpsadbw, 1>;
3441 /// SS41I_ternary_int - SSE 4.1 ternary operator
3442 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3443 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3444 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3445 (ins VR128:$src1, VR128:$src2),
3446 !strconcat(OpcodeStr,
3447 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3448 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3451 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3452 (ins VR128:$src1, i128mem:$src2),
3453 !strconcat(OpcodeStr,
3454 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3457 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3461 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3462 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3463 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3466 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3467 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3468 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3469 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3471 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3472 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3474 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3478 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3479 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3480 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3481 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3482 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3483 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3485 // Common patterns involving scalar load.
3486 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3487 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3488 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3489 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3491 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3492 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3493 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3494 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3496 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3497 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3498 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3499 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3501 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3502 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3503 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3504 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3506 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3507 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3508 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3509 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3511 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3512 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3513 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3514 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3517 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3518 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3519 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3520 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3522 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3523 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3525 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3529 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3530 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3531 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3532 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3534 // Common patterns involving scalar load
3535 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3536 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3537 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3538 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3540 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3541 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3542 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3543 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3546 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3547 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3548 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3549 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3551 // Expecting a i16 load any extended to i32 value.
3552 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3553 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3554 [(set VR128:$dst, (IntId (bitconvert
3555 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3559 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3560 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3562 // Common patterns involving scalar load
3563 def : Pat<(int_x86_sse41_pmovsxbq
3564 (bitconvert (v4i32 (X86vzmovl
3565 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3566 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3568 def : Pat<(int_x86_sse41_pmovzxbq
3569 (bitconvert (v4i32 (X86vzmovl
3570 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3571 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3574 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3575 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3576 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3577 (ins VR128:$src1, i32i8imm:$src2),
3578 !strconcat(OpcodeStr,
3579 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3580 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3582 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3583 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3584 !strconcat(OpcodeStr,
3585 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3588 // There's an AssertZext in the way of writing the store pattern
3589 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3592 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3595 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3596 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3597 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3598 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3599 !strconcat(OpcodeStr,
3600 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3603 // There's an AssertZext in the way of writing the store pattern
3604 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3607 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3610 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3611 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3612 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3613 (ins VR128:$src1, i32i8imm:$src2),
3614 !strconcat(OpcodeStr,
3615 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3617 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3618 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3619 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3620 !strconcat(OpcodeStr,
3621 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3622 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3623 addr:$dst)]>, OpSize;
3626 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3629 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3631 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3632 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3633 (ins VR128:$src1, i32i8imm:$src2),
3634 !strconcat(OpcodeStr,
3635 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3637 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3639 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3640 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3641 !strconcat(OpcodeStr,
3642 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3643 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3644 addr:$dst)]>, OpSize;
3647 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3649 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3650 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3653 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3654 Requires<[HasSSE41]>;
3656 let Constraints = "$src1 = $dst" in {
3657 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3658 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3659 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3660 !strconcat(OpcodeStr,
3661 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3663 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3664 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3665 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3666 !strconcat(OpcodeStr,
3667 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3669 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3670 imm:$src3))]>, OpSize;
3674 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3676 let Constraints = "$src1 = $dst" in {
3677 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3678 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3679 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3680 !strconcat(OpcodeStr,
3681 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3683 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3685 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3686 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3687 !strconcat(OpcodeStr,
3688 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3690 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3691 imm:$src3)))]>, OpSize;
3695 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3697 // insertps has a few different modes, there's the first two here below which
3698 // are optimized inserts that won't zero arbitrary elements in the destination
3699 // vector. The next one matches the intrinsic and could zero arbitrary elements
3700 // in the target vector.
3701 let Constraints = "$src1 = $dst" in {
3702 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3703 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3704 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3705 !strconcat(OpcodeStr,
3706 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3708 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3710 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3711 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3712 !strconcat(OpcodeStr,
3713 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3715 (X86insrtps VR128:$src1,
3716 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3717 imm:$src3))]>, OpSize;
3721 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3723 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3724 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3726 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3727 // the intel intrinsic that corresponds to this.
3728 let Defs = [EFLAGS] in {
3729 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3730 "ptest \t{$src2, $src1|$src1, $src2}",
3731 [(X86ptest VR128:$src1, VR128:$src2),
3732 (implicit EFLAGS)]>, OpSize;
3733 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3734 "ptest \t{$src2, $src1|$src1, $src2}",
3735 [(X86ptest VR128:$src1, (load addr:$src2)),
3736 (implicit EFLAGS)]>, OpSize;
3739 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3740 "movntdqa\t{$src, $dst|$dst, $src}",
3741 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;
3744 //===----------------------------------------------------------------------===//
3745 // SSE4.2 Instructions
3746 //===----------------------------------------------------------------------===//
3748 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3749 let Constraints = "$src1 = $dst" in {
3750 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3751 Intrinsic IntId128, bit Commutable = 0> {
3752 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3753 (ins VR128:$src1, VR128:$src2),
3754 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3755 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3757 let isCommutable = Commutable;
3759 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3760 (ins VR128:$src1, i128mem:$src2),
3761 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3763 (IntId128 VR128:$src1,
3764 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3768 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3770 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3771 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3772 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3773 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3775 // crc intrinsic instruction
3776 // This set of instructions are only rm, the only difference is the size
3778 let Constraints = "$src1 = $dst" in {
3779 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3780 (ins GR32:$src1, i8mem:$src2),
3781 "crc32 \t{$src2, $src1|$src1, $src2}",
3783 (int_x86_sse42_crc32_8 GR32:$src1,
3784 (load addr:$src2)))]>, OpSize;
3785 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3786 (ins GR32:$src1, GR8:$src2),
3787 "crc32 \t{$src2, $src1|$src1, $src2}",
3789 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>,
3791 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3792 (ins GR32:$src1, i16mem:$src2),
3793 "crc32 \t{$src2, $src1|$src1, $src2}",
3795 (int_x86_sse42_crc32_16 GR32:$src1,
3796 (load addr:$src2)))]>,
3798 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3799 (ins GR32:$src1, GR16:$src2),
3800 "crc32 \t{$src2, $src1|$src1, $src2}",
3802 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3804 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3805 (ins GR32:$src1, i32mem:$src2),
3806 "crc32 \t{$src2, $src1|$src1, $src2}",
3808 (int_x86_sse42_crc32_32 GR32:$src1,
3809 (load addr:$src2)))]>, OpSize;
3810 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3811 (ins GR32:$src1, GR32:$src2),
3812 "crc32 \t{$src2, $src1|$src1, $src2}",
3814 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>,
3816 def CRC64m64 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3817 (ins GR64:$src1, i64mem:$src2),
3818 "crc32 \t{$src2, $src1|$src1, $src2}",
3820 (int_x86_sse42_crc32_64 GR64:$src1,
3821 (load addr:$src2)))]>,
3823 def CRC64r64 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3824 (ins GR64:$src1, GR64:$src2),
3825 "crc32 \t{$src2, $src1|$src1, $src2}",
3827 (int_x86_sse42_crc32_64 GR64:$src1, GR64:$src2))]>,
3831 // String/text processing instructions.
3832 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3833 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3834 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3835 "#PCMPISTRM128rr PSEUDO!",
3836 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3837 imm:$src3))]>, OpSize;
3838 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3839 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3840 "#PCMPISTRM128rm PSEUDO!",
3841 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3842 imm:$src3))]>, OpSize;
3845 let Defs = [XMM0, EFLAGS] in {
3846 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3847 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3848 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3849 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3850 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3851 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3854 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3855 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3856 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3857 "#PCMPESTRM128rr PSEUDO!",
3859 (int_x86_sse42_pcmpestrm128
3860 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3862 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3863 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3864 "#PCMPESTRM128rm PSEUDO!",
3865 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3866 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3870 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3871 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3872 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3873 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3874 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3875 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3876 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3879 let Defs = [ECX, EFLAGS] in {
3880 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3881 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3882 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3883 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3884 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3885 (implicit EFLAGS)]>, OpSize;
3886 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3887 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3888 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3889 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3890 (implicit EFLAGS)]>, OpSize;
3894 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3895 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3896 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3897 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3898 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3899 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3901 let Defs = [ECX, EFLAGS] in {
3902 let Uses = [EAX, EDX] in {
3903 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3904 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3905 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3906 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3907 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3908 (implicit EFLAGS)]>, OpSize;
3909 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3910 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3911 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3913 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
3914 (implicit EFLAGS)]>, OpSize;
3919 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3920 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3921 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3922 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3923 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3924 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;