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:$ptr), (f32 (alignedload node:$ptr))>;
120 def alignedloadfsf64 : PatFrag<(ops node:$ptr), (f64 (alignedload node:$ptr))>;
121 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
122 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
123 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
124 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
126 // Like 'load', but uses special alignment checks suitable for use in
127 // memory operands in most SSE instructions, which are required to
128 // be naturally aligned on some targets but not on others.
129 // FIXME: Actually implement support for targets that don't require the
130 // alignment. This probably wants a subtarget predicate.
131 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
132 return cast<LoadSDNode>(N)->getAlignment() >= 16;
135 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
136 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
137 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
138 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
139 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
140 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
141 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
143 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
145 // FIXME: 8 byte alignment for mmx reads is not required
146 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
147 return cast<LoadSDNode>(N)->getAlignment() >= 8;
150 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
151 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
152 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
153 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
155 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
156 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
157 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
158 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
159 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
160 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
162 def vzmovl_v2i64 : PatFrag<(ops node:$src),
163 (bitconvert (v2i64 (X86vzmovl
164 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
165 def vzmovl_v4i32 : PatFrag<(ops node:$src),
166 (bitconvert (v4i32 (X86vzmovl
167 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
169 def vzload_v2i64 : PatFrag<(ops node:$src),
170 (bitconvert (v2i64 (X86vzload node:$src)))>;
173 def fp32imm0 : PatLeaf<(f32 fpimm), [{
174 return N->isExactlyValue(+0.0);
177 def PSxLDQ_imm : SDNodeXForm<imm, [{
178 // Transformation function: imm >> 3
179 return getI32Imm(N->getZExtValue() >> 3);
182 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
184 def SHUFFLE_get_shuf_imm : SDNodeXForm<vector_shuffle, [{
185 return getI8Imm(X86::getShuffleSHUFImmediate(N));
188 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
190 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<vector_shuffle, [{
191 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
194 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
196 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<vector_shuffle, [{
197 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
200 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
201 (vector_shuffle node:$lhs, node:$rhs), [{
202 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
203 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
206 def movddup : PatFrag<(ops node:$lhs, node:$rhs),
207 (vector_shuffle node:$lhs, node:$rhs), [{
208 return X86::isMOVDDUPMask(cast<ShuffleVectorSDNode>(N));
211 def movhlps : PatFrag<(ops node:$lhs, node:$rhs),
212 (vector_shuffle node:$lhs, node:$rhs), [{
213 return X86::isMOVHLPSMask(cast<ShuffleVectorSDNode>(N));
216 def movhlps_undef : PatFrag<(ops node:$lhs, node:$rhs),
217 (vector_shuffle node:$lhs, node:$rhs), [{
218 return X86::isMOVHLPS_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
221 def movhp : PatFrag<(ops node:$lhs, node:$rhs),
222 (vector_shuffle node:$lhs, node:$rhs), [{
223 return X86::isMOVHPMask(cast<ShuffleVectorSDNode>(N));
226 def movlp : PatFrag<(ops node:$lhs, node:$rhs),
227 (vector_shuffle node:$lhs, node:$rhs), [{
228 return X86::isMOVLPMask(cast<ShuffleVectorSDNode>(N));
231 def movl : PatFrag<(ops node:$lhs, node:$rhs),
232 (vector_shuffle node:$lhs, node:$rhs), [{
233 return X86::isMOVLMask(cast<ShuffleVectorSDNode>(N));
236 def movshdup : PatFrag<(ops node:$lhs, node:$rhs),
237 (vector_shuffle node:$lhs, node:$rhs), [{
238 return X86::isMOVSHDUPMask(cast<ShuffleVectorSDNode>(N));
241 def movsldup : PatFrag<(ops node:$lhs, node:$rhs),
242 (vector_shuffle node:$lhs, node:$rhs), [{
243 return X86::isMOVSLDUPMask(cast<ShuffleVectorSDNode>(N));
246 def unpckl : PatFrag<(ops node:$lhs, node:$rhs),
247 (vector_shuffle node:$lhs, node:$rhs), [{
248 return X86::isUNPCKLMask(cast<ShuffleVectorSDNode>(N));
251 def unpckh : PatFrag<(ops node:$lhs, node:$rhs),
252 (vector_shuffle node:$lhs, node:$rhs), [{
253 return X86::isUNPCKHMask(cast<ShuffleVectorSDNode>(N));
256 def unpckl_undef : PatFrag<(ops node:$lhs, node:$rhs),
257 (vector_shuffle node:$lhs, node:$rhs), [{
258 return X86::isUNPCKL_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
261 def unpckh_undef : PatFrag<(ops node:$lhs, node:$rhs),
262 (vector_shuffle node:$lhs, node:$rhs), [{
263 return X86::isUNPCKH_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
266 def pshufd : PatFrag<(ops node:$lhs, node:$rhs),
267 (vector_shuffle node:$lhs, node:$rhs), [{
268 return X86::isPSHUFDMask(cast<ShuffleVectorSDNode>(N));
269 }], SHUFFLE_get_shuf_imm>;
271 def shufp : PatFrag<(ops node:$lhs, node:$rhs),
272 (vector_shuffle node:$lhs, node:$rhs), [{
273 return X86::isSHUFPMask(cast<ShuffleVectorSDNode>(N));
274 }], SHUFFLE_get_shuf_imm>;
276 def pshufhw : PatFrag<(ops node:$lhs, node:$rhs),
277 (vector_shuffle node:$lhs, node:$rhs), [{
278 return X86::isPSHUFHWMask(cast<ShuffleVectorSDNode>(N));
279 }], SHUFFLE_get_pshufhw_imm>;
281 def pshuflw : PatFrag<(ops node:$lhs, node:$rhs),
282 (vector_shuffle node:$lhs, node:$rhs), [{
283 return X86::isPSHUFLWMask(cast<ShuffleVectorSDNode>(N));
284 }], SHUFFLE_get_pshuflw_imm>;
286 //===----------------------------------------------------------------------===//
287 // SSE scalar FP Instructions
288 //===----------------------------------------------------------------------===//
290 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
291 // scheduler into a branch sequence.
292 // These are expanded by the scheduler.
293 let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
294 def CMOV_FR32 : I<0, Pseudo,
295 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
296 "#CMOV_FR32 PSEUDO!",
297 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
299 def CMOV_FR64 : I<0, Pseudo,
300 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
301 "#CMOV_FR64 PSEUDO!",
302 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
304 def CMOV_V4F32 : I<0, Pseudo,
305 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
306 "#CMOV_V4F32 PSEUDO!",
308 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
310 def CMOV_V2F64 : I<0, Pseudo,
311 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
312 "#CMOV_V2F64 PSEUDO!",
314 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
316 def CMOV_V2I64 : I<0, Pseudo,
317 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
318 "#CMOV_V2I64 PSEUDO!",
320 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
324 //===----------------------------------------------------------------------===//
326 //===----------------------------------------------------------------------===//
329 let neverHasSideEffects = 1 in
330 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
331 "movss\t{$src, $dst|$dst, $src}", []>;
332 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
333 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
334 "movss\t{$src, $dst|$dst, $src}",
335 [(set FR32:$dst, (loadf32 addr:$src))]>;
336 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
337 "movss\t{$src, $dst|$dst, $src}",
338 [(store FR32:$src, addr:$dst)]>;
340 // Conversion instructions
341 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
342 "cvttss2si\t{$src, $dst|$dst, $src}",
343 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
344 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
345 "cvttss2si\t{$src, $dst|$dst, $src}",
346 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
347 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
348 "cvtsi2ss\t{$src, $dst|$dst, $src}",
349 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
350 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
351 "cvtsi2ss\t{$src, $dst|$dst, $src}",
352 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
354 // Match intrinsics which expect XMM operand(s).
355 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
356 "cvtss2si\t{$src, $dst|$dst, $src}",
357 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
358 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
359 "cvtss2si\t{$src, $dst|$dst, $src}",
360 [(set GR32:$dst, (int_x86_sse_cvtss2si
361 (load addr:$src)))]>;
363 // Match intrinisics which expect MM and XMM operand(s).
364 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
365 "cvtps2pi\t{$src, $dst|$dst, $src}",
366 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
367 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
368 "cvtps2pi\t{$src, $dst|$dst, $src}",
369 [(set VR64:$dst, (int_x86_sse_cvtps2pi
370 (load addr:$src)))]>;
371 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
372 "cvttps2pi\t{$src, $dst|$dst, $src}",
373 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
374 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
375 "cvttps2pi\t{$src, $dst|$dst, $src}",
376 [(set VR64:$dst, (int_x86_sse_cvttps2pi
377 (load addr:$src)))]>;
378 let Constraints = "$src1 = $dst" in {
379 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
380 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
381 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
382 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
384 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
385 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
386 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
387 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
388 (load addr:$src2)))]>;
391 // Aliases for intrinsics
392 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
393 "cvttss2si\t{$src, $dst|$dst, $src}",
395 (int_x86_sse_cvttss2si VR128:$src))]>;
396 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
397 "cvttss2si\t{$src, $dst|$dst, $src}",
399 (int_x86_sse_cvttss2si(load addr:$src)))]>;
401 let Constraints = "$src1 = $dst" in {
402 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
403 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
404 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
405 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
407 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
408 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
409 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
410 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
411 (loadi32 addr:$src2)))]>;
414 // Comparison instructions
415 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
416 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
417 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
418 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
420 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
421 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
422 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
425 let Defs = [EFLAGS] in {
426 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
427 "ucomiss\t{$src2, $src1|$src1, $src2}",
428 [(X86cmp FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
429 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
430 "ucomiss\t{$src2, $src1|$src1, $src2}",
431 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
435 // Aliases to match intrinsics which expect XMM operand(s).
436 let Constraints = "$src1 = $dst" in {
437 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
438 (outs VR128:$dst), (ins VR128:$src1, VR128:$src,
440 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
441 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
442 VR128:$src, imm:$cc))]>;
443 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
444 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src,
446 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
447 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
448 (load addr:$src), imm:$cc))]>;
451 let Defs = [EFLAGS] in {
452 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
453 "ucomiss\t{$src2, $src1|$src1, $src2}",
454 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
456 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
457 "ucomiss\t{$src2, $src1|$src1, $src2}",
458 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
461 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
462 "comiss\t{$src2, $src1|$src1, $src2}",
463 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
465 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
466 "comiss\t{$src2, $src1|$src1, $src2}",
467 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
471 // Aliases of packed SSE1 instructions for scalar use. These all have names
472 // that start with 'Fs'.
474 // Alias instructions that map fld0 to pxor for sse.
475 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1 in
476 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
477 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
478 Requires<[HasSSE1]>, TB, OpSize;
480 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
482 let neverHasSideEffects = 1 in
483 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
484 "movaps\t{$src, $dst|$dst, $src}", []>;
486 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
488 let canFoldAsLoad = 1 in
489 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
490 "movaps\t{$src, $dst|$dst, $src}",
491 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
493 // Alias bitwise logical operations using SSE logical ops on packed FP values.
494 let Constraints = "$src1 = $dst" in {
495 let isCommutable = 1 in {
496 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst),
497 (ins FR32:$src1, FR32:$src2),
498 "andps\t{$src2, $dst|$dst, $src2}",
499 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
500 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst),
501 (ins FR32:$src1, FR32:$src2),
502 "orps\t{$src2, $dst|$dst, $src2}",
503 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
504 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst),
505 (ins FR32:$src1, FR32:$src2),
506 "xorps\t{$src2, $dst|$dst, $src2}",
507 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
510 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst),
511 (ins FR32:$src1, f128mem:$src2),
512 "andps\t{$src2, $dst|$dst, $src2}",
513 [(set FR32:$dst, (X86fand FR32:$src1,
514 (memopfsf32 addr:$src2)))]>;
515 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst),
516 (ins FR32:$src1, f128mem:$src2),
517 "orps\t{$src2, $dst|$dst, $src2}",
518 [(set FR32:$dst, (X86for FR32:$src1,
519 (memopfsf32 addr:$src2)))]>;
520 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst),
521 (ins FR32:$src1, f128mem:$src2),
522 "xorps\t{$src2, $dst|$dst, $src2}",
523 [(set FR32:$dst, (X86fxor FR32:$src1,
524 (memopfsf32 addr:$src2)))]>;
526 let neverHasSideEffects = 1 in {
527 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
528 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
529 "andnps\t{$src2, $dst|$dst, $src2}", []>;
531 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
532 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
533 "andnps\t{$src2, $dst|$dst, $src2}", []>;
537 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
539 /// In addition, we also have a special variant of the scalar form here to
540 /// represent the associated intrinsic operation. This form is unlike the
541 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
542 /// and leaves the top elements unmodified (therefore these cannot be commuted).
544 /// These three forms can each be reg+reg or reg+mem, so there are a total of
545 /// six "instructions".
547 let Constraints = "$src1 = $dst" in {
548 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
549 SDNode OpNode, Intrinsic F32Int,
550 bit Commutable = 0> {
551 // Scalar operation, reg+reg.
552 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
553 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
554 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
555 let isCommutable = Commutable;
558 // Scalar operation, reg+mem.
559 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
560 (ins FR32:$src1, f32mem:$src2),
561 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
562 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
564 // Vector operation, reg+reg.
565 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
566 (ins VR128:$src1, VR128:$src2),
567 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
568 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
569 let isCommutable = Commutable;
572 // Vector operation, reg+mem.
573 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
574 (ins VR128:$src1, f128mem:$src2),
575 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
576 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
578 // Intrinsic operation, reg+reg.
579 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
580 (ins VR128:$src1, VR128:$src2),
581 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
582 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]>;
584 // Intrinsic operation, reg+mem.
585 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
586 (ins VR128:$src1, ssmem:$src2),
587 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
588 [(set VR128:$dst, (F32Int VR128:$src1,
589 sse_load_f32:$src2))]>;
593 // Arithmetic instructions
594 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
595 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
596 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
597 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
599 /// sse1_fp_binop_rm - Other SSE1 binops
601 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
602 /// instructions for a full-vector intrinsic form. Operations that map
603 /// onto C operators don't use this form since they just use the plain
604 /// vector form instead of having a separate vector intrinsic form.
606 /// This provides a total of eight "instructions".
608 let Constraints = "$src1 = $dst" in {
609 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
613 bit Commutable = 0> {
615 // Scalar operation, reg+reg.
616 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
617 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
618 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
619 let isCommutable = Commutable;
622 // Scalar operation, reg+mem.
623 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
624 (ins FR32:$src1, f32mem:$src2),
625 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
626 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
628 // Vector operation, reg+reg.
629 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
630 (ins VR128:$src1, VR128:$src2),
631 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
632 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
633 let isCommutable = Commutable;
636 // Vector operation, reg+mem.
637 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
638 (ins VR128:$src1, f128mem:$src2),
639 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
640 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
642 // Intrinsic operation, reg+reg.
643 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
644 (ins VR128:$src1, VR128:$src2),
645 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
646 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
647 let isCommutable = Commutable;
650 // Intrinsic operation, reg+mem.
651 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
652 (ins VR128:$src1, ssmem:$src2),
653 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
654 [(set VR128:$dst, (F32Int VR128:$src1,
655 sse_load_f32:$src2))]>;
657 // Vector intrinsic operation, reg+reg.
658 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
659 (ins VR128:$src1, VR128:$src2),
660 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
661 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
662 let isCommutable = Commutable;
665 // Vector intrinsic operation, reg+mem.
666 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
667 (ins VR128:$src1, f128mem:$src2),
668 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
669 [(set VR128:$dst, (V4F32Int VR128:$src1, (memopv4f32 addr:$src2)))]>;
673 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
674 int_x86_sse_max_ss, int_x86_sse_max_ps>;
675 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
676 int_x86_sse_min_ss, int_x86_sse_min_ps>;
678 //===----------------------------------------------------------------------===//
679 // SSE packed FP Instructions
682 let neverHasSideEffects = 1 in
683 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
684 "movaps\t{$src, $dst|$dst, $src}", []>;
685 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
686 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
687 "movaps\t{$src, $dst|$dst, $src}",
688 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
690 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
691 "movaps\t{$src, $dst|$dst, $src}",
692 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
694 let neverHasSideEffects = 1 in
695 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
696 "movups\t{$src, $dst|$dst, $src}", []>;
697 let canFoldAsLoad = 1 in
698 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
699 "movups\t{$src, $dst|$dst, $src}",
700 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
701 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
702 "movups\t{$src, $dst|$dst, $src}",
703 [(store (v4f32 VR128:$src), addr:$dst)]>;
705 // Intrinsic forms of MOVUPS load and store
706 let canFoldAsLoad = 1 in
707 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
708 "movups\t{$src, $dst|$dst, $src}",
709 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
710 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
711 "movups\t{$src, $dst|$dst, $src}",
712 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
714 let Constraints = "$src1 = $dst" in {
715 let AddedComplexity = 20 in {
716 def MOVLPSrm : PSI<0x12, MRMSrcMem,
717 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
718 "movlps\t{$src2, $dst|$dst, $src2}",
721 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
722 def MOVHPSrm : PSI<0x16, MRMSrcMem,
723 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
724 "movhps\t{$src2, $dst|$dst, $src2}",
727 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
729 } // Constraints = "$src1 = $dst"
732 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
733 "movlps\t{$src, $dst|$dst, $src}",
734 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
735 (iPTR 0))), addr:$dst)]>;
737 // v2f64 extract element 1 is always custom lowered to unpack high to low
738 // and extract element 0 so the non-store version isn't too horrible.
739 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
740 "movhps\t{$src, $dst|$dst, $src}",
741 [(store (f64 (vector_extract
742 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
743 (undef)), (iPTR 0))), addr:$dst)]>;
745 let Constraints = "$src1 = $dst" in {
746 let AddedComplexity = 20 in {
747 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
748 (ins VR128:$src1, VR128:$src2),
749 "movlhps\t{$src2, $dst|$dst, $src2}",
751 (v4f32 (movhp VR128:$src1, VR128:$src2)))]>;
753 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
754 (ins VR128:$src1, VR128:$src2),
755 "movhlps\t{$src2, $dst|$dst, $src2}",
757 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
759 } // Constraints = "$src1 = $dst"
761 let AddedComplexity = 20 in {
762 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
763 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
764 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
765 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
772 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
774 /// In addition, we also have a special variant of the scalar form here to
775 /// represent the associated intrinsic operation. This form is unlike the
776 /// plain scalar form, in that it takes an entire vector (instead of a
777 /// scalar) and leaves the top elements undefined.
779 /// And, we have a special variant form for a full-vector intrinsic form.
781 /// These four forms can each have a reg or a mem operand, so there are a
782 /// total of eight "instructions".
784 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
788 bit Commutable = 0> {
789 // Scalar operation, reg.
790 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
791 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
792 [(set FR32:$dst, (OpNode FR32:$src))]> {
793 let isCommutable = Commutable;
796 // Scalar operation, mem.
797 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
798 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
799 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
801 // Vector operation, reg.
802 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
803 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
804 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
805 let isCommutable = Commutable;
808 // Vector operation, mem.
809 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
810 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
811 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
813 // Intrinsic operation, reg.
814 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
815 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
816 [(set VR128:$dst, (F32Int VR128:$src))]> {
817 let isCommutable = Commutable;
820 // Intrinsic operation, mem.
821 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
822 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
823 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
825 // Vector intrinsic operation, reg
826 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
827 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
828 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
829 let isCommutable = Commutable;
832 // Vector intrinsic operation, mem
833 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
834 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
835 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
839 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
840 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
842 // Reciprocal approximations. Note that these typically require refinement
843 // in order to obtain suitable precision.
844 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
845 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
846 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
847 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
850 let Constraints = "$src1 = $dst" in {
851 let isCommutable = 1 in {
852 def ANDPSrr : PSI<0x54, MRMSrcReg,
853 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
854 "andps\t{$src2, $dst|$dst, $src2}",
855 [(set VR128:$dst, (v2i64
856 (and VR128:$src1, VR128:$src2)))]>;
857 def ORPSrr : PSI<0x56, MRMSrcReg,
858 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
859 "orps\t{$src2, $dst|$dst, $src2}",
860 [(set VR128:$dst, (v2i64
861 (or VR128:$src1, VR128:$src2)))]>;
862 def XORPSrr : PSI<0x57, MRMSrcReg,
863 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
864 "xorps\t{$src2, $dst|$dst, $src2}",
865 [(set VR128:$dst, (v2i64
866 (xor VR128:$src1, VR128:$src2)))]>;
869 def ANDPSrm : PSI<0x54, MRMSrcMem,
870 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
871 "andps\t{$src2, $dst|$dst, $src2}",
872 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
873 (memopv2i64 addr:$src2)))]>;
874 def ORPSrm : PSI<0x56, MRMSrcMem,
875 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
876 "orps\t{$src2, $dst|$dst, $src2}",
877 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
878 (memopv2i64 addr:$src2)))]>;
879 def XORPSrm : PSI<0x57, MRMSrcMem,
880 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
881 "xorps\t{$src2, $dst|$dst, $src2}",
882 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
883 (memopv2i64 addr:$src2)))]>;
884 def ANDNPSrr : PSI<0x55, MRMSrcReg,
885 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
886 "andnps\t{$src2, $dst|$dst, $src2}",
888 (v2i64 (and (xor VR128:$src1,
889 (bc_v2i64 (v4i32 immAllOnesV))),
891 def ANDNPSrm : PSI<0x55, MRMSrcMem,
892 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
893 "andnps\t{$src2, $dst|$dst, $src2}",
895 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
896 (bc_v2i64 (v4i32 immAllOnesV))),
897 (memopv2i64 addr:$src2))))]>;
900 let Constraints = "$src1 = $dst" in {
901 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
902 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
903 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
904 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
905 VR128:$src, imm:$cc))]>;
906 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
907 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
908 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
909 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
910 (memop addr:$src), imm:$cc))]>;
912 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
913 (CMPPSrri VR128:$src1, VR128:$src2, imm:$cc)>;
914 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
915 (CMPPSrmi VR128:$src1, addr:$src2, imm:$cc)>;
917 // Shuffle and unpack instructions
918 let Constraints = "$src1 = $dst" in {
919 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
920 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
921 (outs VR128:$dst), (ins VR128:$src1,
922 VR128:$src2, i8imm:$src3),
923 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
925 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
926 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
927 (outs VR128:$dst), (ins VR128:$src1,
928 f128mem:$src2, i8imm:$src3),
929 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
932 VR128:$src1, (memopv4f32 addr:$src2))))]>;
934 let AddedComplexity = 10 in {
935 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
936 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
937 "unpckhps\t{$src2, $dst|$dst, $src2}",
939 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
940 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
941 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
942 "unpckhps\t{$src2, $dst|$dst, $src2}",
944 (v4f32 (unpckh VR128:$src1,
945 (memopv4f32 addr:$src2))))]>;
947 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
948 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
949 "unpcklps\t{$src2, $dst|$dst, $src2}",
951 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
952 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
953 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
954 "unpcklps\t{$src2, $dst|$dst, $src2}",
956 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
958 } // Constraints = "$src1 = $dst"
961 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
962 "movmskps\t{$src, $dst|$dst, $src}",
963 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
964 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
965 "movmskpd\t{$src, $dst|$dst, $src}",
966 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
968 // Prefetch intrinsic.
969 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
970 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
971 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
972 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
973 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
974 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
975 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
976 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
978 // Non-temporal stores
979 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
980 "movntps\t{$src, $dst|$dst, $src}",
981 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
983 // Load, store, and memory fence
984 def SFENCE : PSI<0xAE, MRM7r, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
987 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
988 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
989 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
990 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
992 // Alias instructions that map zero vector to pxor / xorp* for sse.
993 // We set canFoldAsLoad because this can be converted to a constant-pool
994 // load of an all-zeros value if folding it would be beneficial.
995 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
997 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
999 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1001 let Predicates = [HasSSE1] in {
1002 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
1003 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
1004 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
1005 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
1006 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
1009 // FR32 to 128-bit vector conversion.
1010 let isAsCheapAsAMove = 1 in
1011 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
1012 "movss\t{$src, $dst|$dst, $src}",
1014 (v4f32 (scalar_to_vector FR32:$src)))]>;
1015 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1016 "movss\t{$src, $dst|$dst, $src}",
1018 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
1020 // FIXME: may not be able to eliminate this movss with coalescing the src and
1021 // dest register classes are different. We really want to write this pattern
1023 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1024 // (f32 FR32:$src)>;
1025 let isAsCheapAsAMove = 1 in
1026 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
1027 "movss\t{$src, $dst|$dst, $src}",
1028 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
1030 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
1031 "movss\t{$src, $dst|$dst, $src}",
1032 [(store (f32 (vector_extract (v4f32 VR128:$src),
1033 (iPTR 0))), addr:$dst)]>;
1036 // Move to lower bits of a VR128, leaving upper bits alone.
1037 // Three operand (but two address) aliases.
1038 let Constraints = "$src1 = $dst" in {
1039 let neverHasSideEffects = 1 in
1040 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
1041 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
1042 "movss\t{$src2, $dst|$dst, $src2}", []>;
1044 let AddedComplexity = 15 in
1045 def MOVLPSrr : SSI<0x10, MRMSrcReg,
1046 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1047 "movss\t{$src2, $dst|$dst, $src2}",
1049 (v4f32 (movl VR128:$src1, VR128:$src2)))]>;
1052 // Move to lower bits of a VR128 and zeroing upper bits.
1053 // Loading from memory automatically zeroing upper bits.
1054 let AddedComplexity = 20 in
1055 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1056 "movss\t{$src, $dst|$dst, $src}",
1057 [(set VR128:$dst, (v4f32 (X86vzmovl (v4f32 (scalar_to_vector
1058 (loadf32 addr:$src))))))]>;
1060 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
1061 (MOVZSS2PSrm addr:$src)>;
1063 //===---------------------------------------------------------------------===//
1064 // SSE2 Instructions
1065 //===---------------------------------------------------------------------===//
1067 // Move Instructions
1068 let neverHasSideEffects = 1 in
1069 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1070 "movsd\t{$src, $dst|$dst, $src}", []>;
1071 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1072 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1073 "movsd\t{$src, $dst|$dst, $src}",
1074 [(set FR64:$dst, (loadf64 addr:$src))]>;
1075 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1076 "movsd\t{$src, $dst|$dst, $src}",
1077 [(store FR64:$src, addr:$dst)]>;
1079 // Conversion instructions
1080 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1081 "cvttsd2si\t{$src, $dst|$dst, $src}",
1082 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1083 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1084 "cvttsd2si\t{$src, $dst|$dst, $src}",
1085 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1086 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1087 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1088 [(set FR32:$dst, (fround FR64:$src))]>;
1089 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1090 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1091 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1092 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1093 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1094 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1095 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1096 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1097 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1099 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1100 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1101 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1102 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1103 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1104 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1105 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1106 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1107 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1108 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1109 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1110 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1111 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1112 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1113 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1114 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1115 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1116 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1117 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1118 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1120 // SSE2 instructions with XS prefix
1121 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1122 "cvtss2sd\t{$src, $dst|$dst, $src}",
1123 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1124 Requires<[HasSSE2]>;
1125 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1126 "cvtss2sd\t{$src, $dst|$dst, $src}",
1127 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1128 Requires<[HasSSE2]>;
1130 // Match intrinsics which expect XMM operand(s).
1131 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1132 "cvtsd2si\t{$src, $dst|$dst, $src}",
1133 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1134 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1135 "cvtsd2si\t{$src, $dst|$dst, $src}",
1136 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1137 (load addr:$src)))]>;
1139 // Match intrinisics which expect MM and XMM operand(s).
1140 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1141 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1142 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1143 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1144 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1145 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1146 (memop addr:$src)))]>;
1147 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1148 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1149 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1150 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1151 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1152 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1153 (memop addr:$src)))]>;
1154 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1155 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1156 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1157 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1158 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1159 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1160 (load addr:$src)))]>;
1162 // Aliases for intrinsics
1163 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1164 "cvttsd2si\t{$src, $dst|$dst, $src}",
1166 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1167 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1168 "cvttsd2si\t{$src, $dst|$dst, $src}",
1169 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1170 (load addr:$src)))]>;
1172 // Comparison instructions
1173 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1174 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1175 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1176 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1178 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1179 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1180 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1183 let Defs = [EFLAGS] in {
1184 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1185 "ucomisd\t{$src2, $src1|$src1, $src2}",
1186 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1187 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1188 "ucomisd\t{$src2, $src1|$src1, $src2}",
1189 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1190 (implicit EFLAGS)]>;
1191 } // Defs = [EFLAGS]
1193 // Aliases to match intrinsics which expect XMM operand(s).
1194 let Constraints = "$src1 = $dst" in {
1195 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1196 (outs VR128:$dst), (ins VR128:$src1, VR128:$src,
1198 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1199 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1200 VR128:$src, imm:$cc))]>;
1201 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1202 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src,
1204 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1205 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1206 (load addr:$src), imm:$cc))]>;
1209 let Defs = [EFLAGS] in {
1210 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1211 "ucomisd\t{$src2, $src1|$src1, $src2}",
1212 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1213 (implicit EFLAGS)]>;
1214 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1215 "ucomisd\t{$src2, $src1|$src1, $src2}",
1216 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1217 (implicit EFLAGS)]>;
1219 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1220 "comisd\t{$src2, $src1|$src1, $src2}",
1221 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1222 (implicit EFLAGS)]>;
1223 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1224 "comisd\t{$src2, $src1|$src1, $src2}",
1225 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1226 (implicit EFLAGS)]>;
1227 } // Defs = [EFLAGS]
1229 // Aliases of packed SSE2 instructions for scalar use. These all have names
1230 // that start with 'Fs'.
1232 // Alias instructions that map fld0 to pxor for sse.
1233 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1 in
1234 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1235 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1236 Requires<[HasSSE2]>, TB, OpSize;
1238 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1240 let neverHasSideEffects = 1 in
1241 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1242 "movapd\t{$src, $dst|$dst, $src}", []>;
1244 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1246 let canFoldAsLoad = 1 in
1247 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1248 "movapd\t{$src, $dst|$dst, $src}",
1249 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1251 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1252 let Constraints = "$src1 = $dst" in {
1253 let isCommutable = 1 in {
1254 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1255 (ins FR64:$src1, FR64:$src2),
1256 "andpd\t{$src2, $dst|$dst, $src2}",
1257 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1258 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1259 (ins FR64:$src1, FR64:$src2),
1260 "orpd\t{$src2, $dst|$dst, $src2}",
1261 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1262 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1263 (ins FR64:$src1, FR64:$src2),
1264 "xorpd\t{$src2, $dst|$dst, $src2}",
1265 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1268 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1269 (ins FR64:$src1, f128mem:$src2),
1270 "andpd\t{$src2, $dst|$dst, $src2}",
1271 [(set FR64:$dst, (X86fand FR64:$src1,
1272 (memopfsf64 addr:$src2)))]>;
1273 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1274 (ins FR64:$src1, f128mem:$src2),
1275 "orpd\t{$src2, $dst|$dst, $src2}",
1276 [(set FR64:$dst, (X86for FR64:$src1,
1277 (memopfsf64 addr:$src2)))]>;
1278 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1279 (ins FR64:$src1, f128mem:$src2),
1280 "xorpd\t{$src2, $dst|$dst, $src2}",
1281 [(set FR64:$dst, (X86fxor FR64:$src1,
1282 (memopfsf64 addr:$src2)))]>;
1284 let neverHasSideEffects = 1 in {
1285 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1286 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1287 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1289 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1290 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1291 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1295 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1297 /// In addition, we also have a special variant of the scalar form here to
1298 /// represent the associated intrinsic operation. This form is unlike the
1299 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1300 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1302 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1303 /// six "instructions".
1305 let Constraints = "$src1 = $dst" in {
1306 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1307 SDNode OpNode, Intrinsic F64Int,
1308 bit Commutable = 0> {
1309 // Scalar operation, reg+reg.
1310 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1311 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1312 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1313 let isCommutable = Commutable;
1316 // Scalar operation, reg+mem.
1317 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1318 (ins FR64:$src1, f64mem:$src2),
1319 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1320 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1322 // Vector operation, reg+reg.
1323 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1324 (ins VR128:$src1, VR128:$src2),
1325 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1326 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1327 let isCommutable = Commutable;
1330 // Vector operation, reg+mem.
1331 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1332 (ins VR128:$src1, f128mem:$src2),
1333 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1334 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1336 // Intrinsic operation, reg+reg.
1337 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1338 (ins VR128:$src1, VR128:$src2),
1339 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1340 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]>;
1342 // Intrinsic operation, reg+mem.
1343 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1344 (ins VR128:$src1, sdmem:$src2),
1345 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1346 [(set VR128:$dst, (F64Int VR128:$src1,
1347 sse_load_f64:$src2))]>;
1351 // Arithmetic instructions
1352 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1353 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1354 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1355 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1357 /// sse2_fp_binop_rm - Other SSE2 binops
1359 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1360 /// instructions for a full-vector intrinsic form. Operations that map
1361 /// onto C operators don't use this form since they just use the plain
1362 /// vector form instead of having a separate vector intrinsic form.
1364 /// This provides a total of eight "instructions".
1366 let Constraints = "$src1 = $dst" in {
1367 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1371 bit Commutable = 0> {
1373 // Scalar operation, reg+reg.
1374 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1375 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1376 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1377 let isCommutable = Commutable;
1380 // Scalar operation, reg+mem.
1381 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1382 (ins FR64:$src1, f64mem:$src2),
1383 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1384 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1386 // Vector operation, reg+reg.
1387 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1388 (ins VR128:$src1, VR128:$src2),
1389 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1390 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1391 let isCommutable = Commutable;
1394 // Vector operation, reg+mem.
1395 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1396 (ins VR128:$src1, f128mem:$src2),
1397 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1398 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1400 // Intrinsic operation, reg+reg.
1401 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1402 (ins VR128:$src1, VR128:$src2),
1403 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1404 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1405 let isCommutable = Commutable;
1408 // Intrinsic operation, reg+mem.
1409 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1410 (ins VR128:$src1, sdmem:$src2),
1411 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1412 [(set VR128:$dst, (F64Int VR128:$src1,
1413 sse_load_f64:$src2))]>;
1415 // Vector intrinsic operation, reg+reg.
1416 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1417 (ins VR128:$src1, VR128:$src2),
1418 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1419 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1420 let isCommutable = Commutable;
1423 // Vector intrinsic operation, reg+mem.
1424 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1425 (ins VR128:$src1, f128mem:$src2),
1426 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1427 [(set VR128:$dst, (V2F64Int VR128:$src1,
1428 (memopv2f64 addr:$src2)))]>;
1432 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1433 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1434 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1435 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1437 //===---------------------------------------------------------------------===//
1438 // SSE packed FP Instructions
1440 // Move Instructions
1441 let neverHasSideEffects = 1 in
1442 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1443 "movapd\t{$src, $dst|$dst, $src}", []>;
1444 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1445 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1446 "movapd\t{$src, $dst|$dst, $src}",
1447 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1449 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1450 "movapd\t{$src, $dst|$dst, $src}",
1451 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1453 let neverHasSideEffects = 1 in
1454 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1455 "movupd\t{$src, $dst|$dst, $src}", []>;
1456 let canFoldAsLoad = 1 in
1457 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1458 "movupd\t{$src, $dst|$dst, $src}",
1459 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1460 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1461 "movupd\t{$src, $dst|$dst, $src}",
1462 [(store (v2f64 VR128:$src), addr:$dst)]>;
1464 // Intrinsic forms of MOVUPD load and store
1465 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1466 "movupd\t{$src, $dst|$dst, $src}",
1467 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1468 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1469 "movupd\t{$src, $dst|$dst, $src}",
1470 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1472 let Constraints = "$src1 = $dst" in {
1473 let AddedComplexity = 20 in {
1474 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1475 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1476 "movlpd\t{$src2, $dst|$dst, $src2}",
1478 (v2f64 (movlp VR128:$src1,
1479 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1480 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1481 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1482 "movhpd\t{$src2, $dst|$dst, $src2}",
1484 (v2f64 (movhp VR128:$src1,
1485 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1486 } // AddedComplexity
1487 } // Constraints = "$src1 = $dst"
1489 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1490 "movlpd\t{$src, $dst|$dst, $src}",
1491 [(store (f64 (vector_extract (v2f64 VR128:$src),
1492 (iPTR 0))), addr:$dst)]>;
1494 // v2f64 extract element 1 is always custom lowered to unpack high to low
1495 // and extract element 0 so the non-store version isn't too horrible.
1496 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1497 "movhpd\t{$src, $dst|$dst, $src}",
1498 [(store (f64 (vector_extract
1499 (v2f64 (unpckh VR128:$src, (undef))),
1500 (iPTR 0))), addr:$dst)]>;
1502 // SSE2 instructions without OpSize prefix
1503 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1504 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1505 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1506 TB, Requires<[HasSSE2]>;
1507 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1508 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1509 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1510 (bitconvert (memopv2i64 addr:$src))))]>,
1511 TB, Requires<[HasSSE2]>;
1513 // SSE2 instructions with XS prefix
1514 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1515 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1516 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1517 XS, Requires<[HasSSE2]>;
1518 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1519 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1520 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1521 (bitconvert (memopv2i64 addr:$src))))]>,
1522 XS, Requires<[HasSSE2]>;
1524 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1525 "cvtps2dq\t{$src, $dst|$dst, $src}",
1526 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1527 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1528 "cvtps2dq\t{$src, $dst|$dst, $src}",
1529 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1530 (memop addr:$src)))]>;
1531 // SSE2 packed instructions with XS prefix
1532 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1533 "cvttps2dq\t{$src, $dst|$dst, $src}",
1534 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1535 XS, Requires<[HasSSE2]>;
1536 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1537 "cvttps2dq\t{$src, $dst|$dst, $src}",
1538 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1539 (memop addr:$src)))]>,
1540 XS, Requires<[HasSSE2]>;
1542 // SSE2 packed instructions with XD prefix
1543 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1544 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1545 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1546 XD, Requires<[HasSSE2]>;
1547 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1548 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1549 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1550 (memop addr:$src)))]>,
1551 XD, Requires<[HasSSE2]>;
1553 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1554 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1555 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1556 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1557 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1558 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1559 (memop addr:$src)))]>;
1561 // SSE2 instructions without OpSize prefix
1562 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1563 "cvtps2pd\t{$src, $dst|$dst, $src}",
1564 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1565 TB, Requires<[HasSSE2]>;
1566 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1567 "cvtps2pd\t{$src, $dst|$dst, $src}",
1568 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1569 (load addr:$src)))]>,
1570 TB, Requires<[HasSSE2]>;
1572 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1573 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1574 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1575 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1576 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1577 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1578 (memop addr:$src)))]>;
1580 // Match intrinsics which expect XMM operand(s).
1581 // Aliases for intrinsics
1582 let Constraints = "$src1 = $dst" in {
1583 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1584 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1585 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1586 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1588 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1589 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1590 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1591 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1592 (loadi32 addr:$src2)))]>;
1593 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1594 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1595 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1596 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1598 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1599 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1600 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1601 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1602 (load addr:$src2)))]>;
1603 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1604 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1605 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1606 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1607 VR128:$src2))]>, XS,
1608 Requires<[HasSSE2]>;
1609 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1610 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1611 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1612 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1613 (load addr:$src2)))]>, XS,
1614 Requires<[HasSSE2]>;
1619 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1621 /// In addition, we also have a special variant of the scalar form here to
1622 /// represent the associated intrinsic operation. This form is unlike the
1623 /// plain scalar form, in that it takes an entire vector (instead of a
1624 /// scalar) and leaves the top elements undefined.
1626 /// And, we have a special variant form for a full-vector intrinsic form.
1628 /// These four forms can each have a reg or a mem operand, so there are a
1629 /// total of eight "instructions".
1631 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1635 bit Commutable = 0> {
1636 // Scalar operation, reg.
1637 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1638 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1639 [(set FR64:$dst, (OpNode FR64:$src))]> {
1640 let isCommutable = Commutable;
1643 // Scalar operation, mem.
1644 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1645 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1646 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1648 // Vector operation, reg.
1649 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1650 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1651 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1652 let isCommutable = Commutable;
1655 // Vector operation, mem.
1656 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1657 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1658 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1660 // Intrinsic operation, reg.
1661 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1662 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1663 [(set VR128:$dst, (F64Int VR128:$src))]> {
1664 let isCommutable = Commutable;
1667 // Intrinsic operation, mem.
1668 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1669 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1670 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1672 // Vector intrinsic operation, reg
1673 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1674 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1675 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1676 let isCommutable = Commutable;
1679 // Vector intrinsic operation, mem
1680 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1681 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1682 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1686 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1687 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1689 // There is no f64 version of the reciprocal approximation instructions.
1692 let Constraints = "$src1 = $dst" in {
1693 let isCommutable = 1 in {
1694 def ANDPDrr : PDI<0x54, MRMSrcReg,
1695 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1696 "andpd\t{$src2, $dst|$dst, $src2}",
1698 (and (bc_v2i64 (v2f64 VR128:$src1)),
1699 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1700 def ORPDrr : PDI<0x56, MRMSrcReg,
1701 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1702 "orpd\t{$src2, $dst|$dst, $src2}",
1704 (or (bc_v2i64 (v2f64 VR128:$src1)),
1705 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1706 def XORPDrr : PDI<0x57, MRMSrcReg,
1707 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1708 "xorpd\t{$src2, $dst|$dst, $src2}",
1710 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1711 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1714 def ANDPDrm : PDI<0x54, MRMSrcMem,
1715 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1716 "andpd\t{$src2, $dst|$dst, $src2}",
1718 (and (bc_v2i64 (v2f64 VR128:$src1)),
1719 (memopv2i64 addr:$src2)))]>;
1720 def ORPDrm : PDI<0x56, MRMSrcMem,
1721 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1722 "orpd\t{$src2, $dst|$dst, $src2}",
1724 (or (bc_v2i64 (v2f64 VR128:$src1)),
1725 (memopv2i64 addr:$src2)))]>;
1726 def XORPDrm : PDI<0x57, MRMSrcMem,
1727 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1728 "xorpd\t{$src2, $dst|$dst, $src2}",
1730 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1731 (memopv2i64 addr:$src2)))]>;
1732 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1733 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1734 "andnpd\t{$src2, $dst|$dst, $src2}",
1736 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1737 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1738 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1739 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1740 "andnpd\t{$src2, $dst|$dst, $src2}",
1742 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1743 (memopv2i64 addr:$src2)))]>;
1746 let Constraints = "$src1 = $dst" in {
1747 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1748 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1749 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1750 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1751 VR128:$src, imm:$cc))]>;
1752 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1753 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1754 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1755 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1756 (memop addr:$src), imm:$cc))]>;
1758 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1759 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1760 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1761 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1763 // Shuffle and unpack instructions
1764 let Constraints = "$src1 = $dst" in {
1765 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1766 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1767 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1769 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1770 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1771 (outs VR128:$dst), (ins VR128:$src1,
1772 f128mem:$src2, i8imm:$src3),
1773 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1776 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1778 let AddedComplexity = 10 in {
1779 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1780 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1781 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1783 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1784 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1785 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1786 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1788 (v2f64 (unpckh VR128:$src1,
1789 (memopv2f64 addr:$src2))))]>;
1791 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1792 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1793 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1795 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1796 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1797 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1798 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1800 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1801 } // AddedComplexity
1802 } // Constraints = "$src1 = $dst"
1805 //===---------------------------------------------------------------------===//
1806 // SSE integer instructions
1808 // Move Instructions
1809 let neverHasSideEffects = 1 in
1810 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1811 "movdqa\t{$src, $dst|$dst, $src}", []>;
1812 let canFoldAsLoad = 1, mayLoad = 1 in
1813 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1814 "movdqa\t{$src, $dst|$dst, $src}",
1815 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1817 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1818 "movdqa\t{$src, $dst|$dst, $src}",
1819 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1820 let canFoldAsLoad = 1, mayLoad = 1 in
1821 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1822 "movdqu\t{$src, $dst|$dst, $src}",
1823 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1824 XS, Requires<[HasSSE2]>;
1826 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1827 "movdqu\t{$src, $dst|$dst, $src}",
1828 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1829 XS, Requires<[HasSSE2]>;
1831 // Intrinsic forms of MOVDQU load and store
1832 let canFoldAsLoad = 1 in
1833 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1834 "movdqu\t{$src, $dst|$dst, $src}",
1835 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1836 XS, Requires<[HasSSE2]>;
1837 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1838 "movdqu\t{$src, $dst|$dst, $src}",
1839 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1840 XS, Requires<[HasSSE2]>;
1842 let Constraints = "$src1 = $dst" in {
1844 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1845 bit Commutable = 0> {
1846 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1847 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1848 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1849 let isCommutable = Commutable;
1851 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1852 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1853 [(set VR128:$dst, (IntId VR128:$src1,
1854 (bitconvert (memopv2i64 addr:$src2))))]>;
1857 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1859 Intrinsic IntId, Intrinsic IntId2> {
1860 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1,
1862 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1863 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1864 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1,
1866 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1867 [(set VR128:$dst, (IntId VR128:$src1,
1868 (bitconvert (memopv2i64 addr:$src2))))]>;
1869 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1,
1871 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1872 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1875 /// PDI_binop_rm - Simple SSE2 binary operator.
1876 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1877 ValueType OpVT, bit Commutable = 0> {
1878 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1,
1880 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1881 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1882 let isCommutable = Commutable;
1884 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1,
1886 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1887 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1888 (bitconvert (memopv2i64 addr:$src2)))))]>;
1891 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1893 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1894 /// to collapse (bitconvert VT to VT) into its operand.
1896 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1897 bit Commutable = 0> {
1898 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1899 (ins VR128:$src1, VR128:$src2),
1900 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1901 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1902 let isCommutable = Commutable;
1904 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1905 (ins VR128:$src1, i128mem:$src2),
1906 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1907 [(set VR128:$dst, (OpNode VR128:$src1,
1908 (memopv2i64 addr:$src2)))]>;
1911 } // Constraints = "$src1 = $dst"
1913 // 128-bit Integer Arithmetic
1915 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1916 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1917 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1918 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1920 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1921 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1922 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1923 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1925 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1926 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1927 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1928 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1930 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1931 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1932 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1933 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1935 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1937 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1938 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1939 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1941 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1943 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1944 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1947 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1948 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1949 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1950 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1951 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
1954 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
1955 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
1956 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
1957 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
1958 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
1959 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
1961 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
1962 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
1963 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
1964 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
1965 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
1966 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
1968 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
1969 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
1970 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
1971 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
1973 // 128-bit logical shifts.
1974 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1975 def PSLLDQri : PDIi8<0x73, MRM7r,
1976 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1977 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1978 def PSRLDQri : PDIi8<0x73, MRM3r,
1979 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1980 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1981 // PSRADQri doesn't exist in SSE[1-3].
1984 let Predicates = [HasSSE2] in {
1985 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1986 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1987 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1988 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1989 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
1990 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
1991 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
1992 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
1993 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1994 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1996 // Shift up / down and insert zero's.
1997 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
1998 (v2i64 (PSLLDQri VR128:$src, (PSxLDQ_imm imm:$amt)))>;
1999 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2000 (v2i64 (PSRLDQri VR128:$src, (PSxLDQ_imm imm:$amt)))>;
2004 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2005 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2006 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2008 let Constraints = "$src1 = $dst" in {
2009 def PANDNrr : PDI<0xDF, MRMSrcReg,
2010 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2011 "pandn\t{$src2, $dst|$dst, $src2}",
2012 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2015 def PANDNrm : PDI<0xDF, MRMSrcMem,
2016 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2017 "pandn\t{$src2, $dst|$dst, $src2}",
2018 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2019 (memopv2i64 addr:$src2))))]>;
2022 // SSE2 Integer comparison
2023 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2024 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2025 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2026 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2027 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2028 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2030 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2031 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2032 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2033 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2034 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2035 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2036 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2037 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2038 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2039 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2040 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2041 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2043 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2044 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2045 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2046 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2047 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2048 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2049 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2050 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2051 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2052 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2053 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2054 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2057 // Pack instructions
2058 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2059 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2060 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2062 // Shuffle and unpack instructions
2063 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2064 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2065 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2066 [(set VR128:$dst, (v4i32 (pshufd:$src2
2067 VR128:$src1, (undef))))]>;
2068 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2069 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2070 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2071 [(set VR128:$dst, (v4i32 (pshufd:$src2
2072 (bc_v4i32(memopv2i64 addr:$src1)),
2075 // SSE2 with ImmT == Imm8 and XS prefix.
2076 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2077 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2078 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2079 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2081 XS, Requires<[HasSSE2]>;
2082 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2083 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2084 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2085 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2086 (bc_v8i16 (memopv2i64 addr:$src1)),
2088 XS, Requires<[HasSSE2]>;
2090 // SSE2 with ImmT == Imm8 and XD prefix.
2091 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2092 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2093 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2094 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2096 XD, Requires<[HasSSE2]>;
2097 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2098 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2099 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2100 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2101 (bc_v8i16 (memopv2i64 addr:$src1)),
2103 XD, Requires<[HasSSE2]>;
2106 let Constraints = "$src1 = $dst" in {
2107 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2108 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2109 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2111 (v16i8 (unpckl VR128:$src1, VR128:$src2)))]>;
2112 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2113 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2114 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2116 (unpckl VR128:$src1,
2117 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2118 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2119 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2120 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2122 (v8i16 (unpckl VR128:$src1, VR128:$src2)))]>;
2123 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2124 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2125 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2127 (unpckl VR128:$src1,
2128 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2129 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2130 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2131 "punpckldq\t{$src2, $dst|$dst, $src2}",
2133 (v4i32 (unpckl VR128:$src1, VR128:$src2)))]>;
2134 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2135 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2136 "punpckldq\t{$src2, $dst|$dst, $src2}",
2138 (unpckl VR128:$src1,
2139 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2140 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2141 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2142 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2144 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2145 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2146 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2147 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2149 (v2i64 (unpckl VR128:$src1,
2150 (memopv2i64 addr:$src2))))]>;
2152 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2153 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2154 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2156 (v16i8 (unpckh VR128:$src1, VR128:$src2)))]>;
2157 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2158 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2159 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2161 (unpckh VR128:$src1,
2162 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2163 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2164 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2165 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2167 (v8i16 (unpckh VR128:$src1, VR128:$src2)))]>;
2168 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2169 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2170 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2172 (unpckh VR128:$src1,
2173 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2174 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2175 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2176 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2178 (v4i32 (unpckh VR128:$src1, VR128:$src2)))]>;
2179 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2180 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2181 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2183 (unpckh VR128:$src1,
2184 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2185 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2186 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2187 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2189 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2190 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2191 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2192 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2194 (v2i64 (unpckh VR128:$src1,
2195 (memopv2i64 addr:$src2))))]>;
2199 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2200 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2201 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2202 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2204 let Constraints = "$src1 = $dst" in {
2205 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2206 (outs VR128:$dst), (ins VR128:$src1,
2207 GR32:$src2, i32i8imm:$src3),
2208 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2210 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2211 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2212 (outs VR128:$dst), (ins VR128:$src1,
2213 i16mem:$src2, i32i8imm:$src3),
2214 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2216 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2221 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2222 "pmovmskb\t{$src, $dst|$dst, $src}",
2223 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2225 // Conditional store
2227 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2228 "maskmovdqu\t{$mask, $src|$src, $mask}",
2229 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2232 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2233 "maskmovdqu\t{$mask, $src|$src, $mask}",
2234 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2236 // Non-temporal stores
2237 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2238 "movntpd\t{$src, $dst|$dst, $src}",
2239 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2240 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2241 "movntdq\t{$src, $dst|$dst, $src}",
2242 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2243 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2244 "movnti\t{$src, $dst|$dst, $src}",
2245 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2246 TB, Requires<[HasSSE2]>;
2249 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2250 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2251 TB, Requires<[HasSSE2]>;
2253 // Load, store, and memory fence
2254 def LFENCE : I<0xAE, MRM5r, (outs), (ins),
2255 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2256 def MFENCE : I<0xAE, MRM6r, (outs), (ins),
2257 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2259 //TODO: custom lower this so as to never even generate the noop
2260 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2262 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2263 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2264 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2267 // Alias instructions that map zero vector to pxor / xorp* for sse.
2268 // We set canFoldAsLoad because this can be converted to a constant-pool
2269 // load of an all-ones value if folding it would be beneficial.
2270 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2271 isCodeGenOnly = 1 in
2272 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2273 "pcmpeqd\t$dst, $dst",
2274 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2276 // FR64 to 128-bit vector conversion.
2277 let isAsCheapAsAMove = 1 in
2278 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2279 "movsd\t{$src, $dst|$dst, $src}",
2281 (v2f64 (scalar_to_vector FR64:$src)))]>;
2282 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2283 "movsd\t{$src, $dst|$dst, $src}",
2285 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2287 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2288 "movd\t{$src, $dst|$dst, $src}",
2290 (v4i32 (scalar_to_vector GR32:$src)))]>;
2291 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2292 "movd\t{$src, $dst|$dst, $src}",
2294 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2296 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2297 "movd\t{$src, $dst|$dst, $src}",
2298 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2300 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2301 "movd\t{$src, $dst|$dst, $src}",
2302 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2304 // SSE2 instructions with XS prefix
2305 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2306 "movq\t{$src, $dst|$dst, $src}",
2308 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2309 Requires<[HasSSE2]>;
2310 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2311 "movq\t{$src, $dst|$dst, $src}",
2312 [(store (i64 (vector_extract (v2i64 VR128:$src),
2313 (iPTR 0))), addr:$dst)]>;
2315 // FIXME: may not be able to eliminate this movss with coalescing the src and
2316 // dest register classes are different. We really want to write this pattern
2318 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2319 // (f32 FR32:$src)>;
2320 let isAsCheapAsAMove = 1 in
2321 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2322 "movsd\t{$src, $dst|$dst, $src}",
2323 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2325 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2326 "movsd\t{$src, $dst|$dst, $src}",
2327 [(store (f64 (vector_extract (v2f64 VR128:$src),
2328 (iPTR 0))), addr:$dst)]>;
2329 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2330 "movd\t{$src, $dst|$dst, $src}",
2331 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2333 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2334 "movd\t{$src, $dst|$dst, $src}",
2335 [(store (i32 (vector_extract (v4i32 VR128:$src),
2336 (iPTR 0))), addr:$dst)]>;
2338 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2339 "movd\t{$src, $dst|$dst, $src}",
2340 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2341 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2342 "movd\t{$src, $dst|$dst, $src}",
2343 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2346 // Move to lower bits of a VR128, leaving upper bits alone.
2347 // Three operand (but two address) aliases.
2348 let Constraints = "$src1 = $dst" in {
2349 let neverHasSideEffects = 1 in
2350 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2351 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2352 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2354 let AddedComplexity = 15 in
2355 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2356 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2357 "movsd\t{$src2, $dst|$dst, $src2}",
2359 (v2f64 (movl VR128:$src1, VR128:$src2)))]>;
2362 // Store / copy lower 64-bits of a XMM register.
2363 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2364 "movq\t{$src, $dst|$dst, $src}",
2365 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2367 // Move to lower bits of a VR128 and zeroing upper bits.
2368 // Loading from memory automatically zeroing upper bits.
2369 let AddedComplexity = 20 in {
2370 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2371 "movsd\t{$src, $dst|$dst, $src}",
2373 (v2f64 (X86vzmovl (v2f64 (scalar_to_vector
2374 (loadf64 addr:$src))))))]>;
2376 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
2377 (MOVZSD2PDrm addr:$src)>;
2378 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
2379 (MOVZSD2PDrm addr:$src)>;
2380 def : Pat<(v2f64 (X86vzload addr:$src)), (MOVZSD2PDrm addr:$src)>;
2383 // movd / movq to XMM register zero-extends
2384 let AddedComplexity = 15 in {
2385 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2386 "movd\t{$src, $dst|$dst, $src}",
2387 [(set VR128:$dst, (v4i32 (X86vzmovl
2388 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2389 // This is X86-64 only.
2390 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2391 "mov{d|q}\t{$src, $dst|$dst, $src}",
2392 [(set VR128:$dst, (v2i64 (X86vzmovl
2393 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2396 let AddedComplexity = 20 in {
2397 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2398 "movd\t{$src, $dst|$dst, $src}",
2400 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2401 (loadi32 addr:$src))))))]>;
2403 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2404 (MOVZDI2PDIrm addr:$src)>;
2405 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2406 (MOVZDI2PDIrm addr:$src)>;
2407 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2408 (MOVZDI2PDIrm addr:$src)>;
2410 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2411 "movq\t{$src, $dst|$dst, $src}",
2413 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2414 (loadi64 addr:$src))))))]>, XS,
2415 Requires<[HasSSE2]>;
2417 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2418 (MOVZQI2PQIrm addr:$src)>;
2419 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2420 (MOVZQI2PQIrm addr:$src)>;
2421 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2424 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2425 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2426 let AddedComplexity = 15 in
2427 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2428 "movq\t{$src, $dst|$dst, $src}",
2429 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2430 XS, Requires<[HasSSE2]>;
2432 let AddedComplexity = 20 in {
2433 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2434 "movq\t{$src, $dst|$dst, $src}",
2435 [(set VR128:$dst, (v2i64 (X86vzmovl
2436 (loadv2i64 addr:$src))))]>,
2437 XS, Requires<[HasSSE2]>;
2439 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2440 (MOVZPQILo2PQIrm addr:$src)>;
2443 //===---------------------------------------------------------------------===//
2444 // SSE3 Instructions
2445 //===---------------------------------------------------------------------===//
2447 // Move Instructions
2448 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2449 "movshdup\t{$src, $dst|$dst, $src}",
2450 [(set VR128:$dst, (v4f32 (movshdup
2451 VR128:$src, (undef))))]>;
2452 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2453 "movshdup\t{$src, $dst|$dst, $src}",
2454 [(set VR128:$dst, (movshdup
2455 (memopv4f32 addr:$src), (undef)))]>;
2457 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2458 "movsldup\t{$src, $dst|$dst, $src}",
2459 [(set VR128:$dst, (v4f32 (movsldup
2460 VR128:$src, (undef))))]>;
2461 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2462 "movsldup\t{$src, $dst|$dst, $src}",
2463 [(set VR128:$dst, (movsldup
2464 (memopv4f32 addr:$src), (undef)))]>;
2466 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2467 "movddup\t{$src, $dst|$dst, $src}",
2468 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2469 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2470 "movddup\t{$src, $dst|$dst, $src}",
2472 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2475 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2477 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2479 let AddedComplexity = 5 in {
2480 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2481 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2482 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2483 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2484 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2485 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2486 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2487 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2491 let Constraints = "$src1 = $dst" in {
2492 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2493 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2494 "addsubps\t{$src2, $dst|$dst, $src2}",
2495 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2497 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2498 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2499 "addsubps\t{$src2, $dst|$dst, $src2}",
2500 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2501 (memop addr:$src2)))]>;
2502 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2503 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2504 "addsubpd\t{$src2, $dst|$dst, $src2}",
2505 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2507 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2508 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2509 "addsubpd\t{$src2, $dst|$dst, $src2}",
2510 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2511 (memop addr:$src2)))]>;
2514 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2515 "lddqu\t{$src, $dst|$dst, $src}",
2516 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2519 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2520 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2521 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2522 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2523 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2524 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2525 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2526 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2527 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2528 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2529 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2530 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2531 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2532 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2533 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2534 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2536 let Constraints = "$src1 = $dst" in {
2537 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2538 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2539 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2540 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2541 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2542 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2543 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2544 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2547 // Thread synchronization
2548 def MONITOR : I<0x01, MRM1r, (outs), (ins), "monitor",
2549 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2550 def MWAIT : I<0x01, MRM1r, (outs), (ins), "mwait",
2551 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2553 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2554 let AddedComplexity = 15 in
2555 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2556 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2557 let AddedComplexity = 20 in
2558 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2559 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2561 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2562 let AddedComplexity = 15 in
2563 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2564 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2565 let AddedComplexity = 20 in
2566 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2567 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2569 //===---------------------------------------------------------------------===//
2570 // SSSE3 Instructions
2571 //===---------------------------------------------------------------------===//
2573 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2574 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2575 Intrinsic IntId64, Intrinsic IntId128> {
2576 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2577 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2578 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2580 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2581 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2583 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2585 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2587 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2588 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2591 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2593 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2596 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2599 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2600 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2601 Intrinsic IntId64, Intrinsic IntId128> {
2602 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2604 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2605 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2607 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2609 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2612 (bitconvert (memopv4i16 addr:$src))))]>;
2614 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2616 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2617 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2620 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2622 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2625 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2628 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2629 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2630 Intrinsic IntId64, Intrinsic IntId128> {
2631 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2633 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2634 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2636 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2638 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2641 (bitconvert (memopv2i32 addr:$src))))]>;
2643 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2645 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2646 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2649 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2651 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2654 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2657 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2658 int_x86_ssse3_pabs_b,
2659 int_x86_ssse3_pabs_b_128>;
2660 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2661 int_x86_ssse3_pabs_w,
2662 int_x86_ssse3_pabs_w_128>;
2663 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2664 int_x86_ssse3_pabs_d,
2665 int_x86_ssse3_pabs_d_128>;
2667 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2668 let Constraints = "$src1 = $dst" in {
2669 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2670 Intrinsic IntId64, Intrinsic IntId128,
2671 bit Commutable = 0> {
2672 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2673 (ins VR64:$src1, VR64:$src2),
2674 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2675 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2676 let isCommutable = Commutable;
2678 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2679 (ins VR64:$src1, i64mem:$src2),
2680 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2682 (IntId64 VR64:$src1,
2683 (bitconvert (memopv8i8 addr:$src2))))]>;
2685 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2686 (ins VR128:$src1, VR128:$src2),
2687 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2688 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2690 let isCommutable = Commutable;
2692 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2693 (ins VR128:$src1, i128mem:$src2),
2694 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2696 (IntId128 VR128:$src1,
2697 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2701 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2702 let Constraints = "$src1 = $dst" in {
2703 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2704 Intrinsic IntId64, Intrinsic IntId128,
2705 bit Commutable = 0> {
2706 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2707 (ins VR64:$src1, VR64:$src2),
2708 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2709 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2710 let isCommutable = Commutable;
2712 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2713 (ins VR64:$src1, i64mem:$src2),
2714 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2716 (IntId64 VR64:$src1,
2717 (bitconvert (memopv4i16 addr:$src2))))]>;
2719 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2720 (ins VR128:$src1, VR128:$src2),
2721 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2722 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2724 let isCommutable = Commutable;
2726 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2727 (ins VR128:$src1, i128mem:$src2),
2728 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2730 (IntId128 VR128:$src1,
2731 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2735 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2736 let Constraints = "$src1 = $dst" in {
2737 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2738 Intrinsic IntId64, Intrinsic IntId128,
2739 bit Commutable = 0> {
2740 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2741 (ins VR64:$src1, VR64:$src2),
2742 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2743 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2744 let isCommutable = Commutable;
2746 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2747 (ins VR64:$src1, i64mem:$src2),
2748 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2750 (IntId64 VR64:$src1,
2751 (bitconvert (memopv2i32 addr:$src2))))]>;
2753 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2754 (ins VR128:$src1, VR128:$src2),
2755 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2756 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2758 let isCommutable = Commutable;
2760 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2761 (ins VR128:$src1, i128mem:$src2),
2762 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2764 (IntId128 VR128:$src1,
2765 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2769 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2770 int_x86_ssse3_phadd_w,
2771 int_x86_ssse3_phadd_w_128>;
2772 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2773 int_x86_ssse3_phadd_d,
2774 int_x86_ssse3_phadd_d_128>;
2775 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2776 int_x86_ssse3_phadd_sw,
2777 int_x86_ssse3_phadd_sw_128>;
2778 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2779 int_x86_ssse3_phsub_w,
2780 int_x86_ssse3_phsub_w_128>;
2781 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2782 int_x86_ssse3_phsub_d,
2783 int_x86_ssse3_phsub_d_128>;
2784 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2785 int_x86_ssse3_phsub_sw,
2786 int_x86_ssse3_phsub_sw_128>;
2787 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2788 int_x86_ssse3_pmadd_ub_sw,
2789 int_x86_ssse3_pmadd_ub_sw_128>;
2790 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2791 int_x86_ssse3_pmul_hr_sw,
2792 int_x86_ssse3_pmul_hr_sw_128, 1>;
2793 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2794 int_x86_ssse3_pshuf_b,
2795 int_x86_ssse3_pshuf_b_128>;
2796 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2797 int_x86_ssse3_psign_b,
2798 int_x86_ssse3_psign_b_128>;
2799 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2800 int_x86_ssse3_psign_w,
2801 int_x86_ssse3_psign_w_128>;
2802 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2803 int_x86_ssse3_psign_d,
2804 int_x86_ssse3_psign_d_128>;
2806 let Constraints = "$src1 = $dst" in {
2807 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2808 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2809 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2811 (int_x86_ssse3_palign_r
2812 VR64:$src1, VR64:$src2,
2814 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2815 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2816 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2818 (int_x86_ssse3_palign_r
2820 (bitconvert (memopv2i32 addr:$src2)),
2823 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2824 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2825 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2827 (int_x86_ssse3_palign_r_128
2828 VR128:$src1, VR128:$src2,
2829 imm:$src3))]>, OpSize;
2830 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2831 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2832 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2834 (int_x86_ssse3_palign_r_128
2836 (bitconvert (memopv4i32 addr:$src2)),
2837 imm:$src3))]>, OpSize;
2840 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2841 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2842 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2843 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2845 //===---------------------------------------------------------------------===//
2846 // Non-Instruction Patterns
2847 //===---------------------------------------------------------------------===//
2849 // extload f32 -> f64. This matches load+fextend because we have a hack in
2850 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2852 // Since these loads aren't folded into the fextend, we have to match it
2854 let Predicates = [HasSSE2] in
2855 def : Pat<(fextend (loadf32 addr:$src)),
2856 (CVTSS2SDrm addr:$src)>;
2859 let Predicates = [HasSSE2] in {
2860 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2861 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2862 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2863 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2864 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2865 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2866 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2867 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2868 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2869 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2870 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2871 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2872 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2873 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2874 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2875 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2876 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2877 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2878 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2879 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2880 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2881 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2882 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2883 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2884 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2885 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2886 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2887 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2888 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2889 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2892 // Move scalar to XMM zero-extended
2893 // movd to XMM register zero-extends
2894 let AddedComplexity = 15 in {
2895 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2896 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2897 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2898 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2899 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE1]>;
2900 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2901 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2902 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2903 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2906 // Splat v2f64 / v2i64
2907 let AddedComplexity = 10 in {
2908 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
2909 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2910 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
2911 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2912 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
2913 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2914 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
2915 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2918 // Special unary SHUFPSrri case.
2919 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
2920 (SHUFPSrri VR128:$src1, VR128:$src1,
2921 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2922 Requires<[HasSSE1]>;
2923 let AddedComplexity = 5 in
2924 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
2925 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2926 Requires<[HasSSE2]>;
2927 // Special unary SHUFPDrri case.
2928 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
2929 (SHUFPDrri VR128:$src1, VR128:$src1,
2930 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2931 Requires<[HasSSE2]>;
2932 // Special unary SHUFPDrri case.
2933 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
2934 (SHUFPDrri VR128:$src1, VR128:$src1,
2935 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2936 Requires<[HasSSE2]>;
2937 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2938 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
2939 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2940 Requires<[HasSSE2]>;
2942 // Special binary v4i32 shuffle cases with SHUFPS.
2943 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
2944 (SHUFPSrri VR128:$src1, VR128:$src2,
2945 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2946 Requires<[HasSSE2]>;
2947 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
2948 (SHUFPSrmi VR128:$src1, addr:$src2,
2949 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2950 Requires<[HasSSE2]>;
2951 // Special binary v2i64 shuffle cases using SHUFPDrri.
2952 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
2953 (SHUFPDrri VR128:$src1, VR128:$src2,
2954 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2955 Requires<[HasSSE2]>;
2957 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2958 let AddedComplexity = 15 in {
2959 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
2960 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2961 Requires<[OptForSpeed, HasSSE2]>;
2962 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
2963 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2964 Requires<[OptForSpeed, HasSSE2]>;
2966 let AddedComplexity = 10 in {
2967 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
2968 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2969 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
2970 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2971 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
2972 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2973 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
2974 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2977 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2978 let AddedComplexity = 15 in {
2979 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
2980 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2981 Requires<[OptForSpeed, HasSSE2]>;
2982 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
2983 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2984 Requires<[OptForSpeed, HasSSE2]>;
2986 let AddedComplexity = 10 in {
2987 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
2988 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2989 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
2990 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2991 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
2992 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2993 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
2994 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2997 let AddedComplexity = 20 in {
2998 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2999 def : Pat<(v4i32 (movhp VR128:$src1, VR128:$src2)),
3000 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3002 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3003 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3004 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3006 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3007 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3008 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3009 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3010 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3013 let AddedComplexity = 20 in {
3014 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3015 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
3016 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3017 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3018 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3019 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3020 def : Pat<(v4f32 (movhp VR128:$src1, (load addr:$src2))),
3021 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3022 def : Pat<(v2f64 (movhp VR128:$src1, (load addr:$src2))),
3023 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3025 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3026 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3027 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3028 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3029 def : Pat<(v4i32 (movhp VR128:$src1, (load addr:$src2))),
3030 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3031 def : Pat<(v2i64 (movhp VR128:$src1, (load addr:$src2))),
3032 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3035 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3036 // (store (vector_shuffle (load addr), v2, <0, 1, 4, 5>), addr) using MOVHPS
3037 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3038 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3039 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3040 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3041 def : Pat<(store (v4f32 (movhp (load addr:$src1), VR128:$src2)), addr:$src1),
3042 (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3043 def : Pat<(store (v2f64 (movhp (load addr:$src1), VR128:$src2)), addr:$src1),
3044 (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3046 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3048 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3049 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3050 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3051 def : Pat<(store (v4i32 (movhp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3053 (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3054 def : Pat<(store (v2i64 (movhp (load addr:$src1), VR128:$src2)), addr:$src1),
3055 (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3058 let AddedComplexity = 15 in {
3059 // Setting the lowest element in the vector.
3060 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3061 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3062 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3063 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3065 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
3066 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3067 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3068 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3069 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3072 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3073 // fall back to this for SSE1)
3074 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3075 (SHUFPSrri VR128:$src2, VR128:$src1,
3076 (SHUFFLE_get_shuf_imm VR128:$src3))>, Requires<[HasSSE1]>;
3078 // Set lowest element and zero upper elements.
3079 let AddedComplexity = 15 in
3080 def : Pat<(v2f64 (movl immAllZerosV_bc, VR128:$src)),
3081 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3082 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3083 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3085 // Some special case pandn patterns.
3086 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3088 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3089 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3091 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3092 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3094 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3096 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3097 (memop addr:$src2))),
3098 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3099 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3100 (memop addr:$src2))),
3101 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3102 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3103 (memop addr:$src2))),
3104 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3106 // vector -> vector casts
3107 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3108 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3109 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3110 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3111 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3112 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3113 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3114 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3116 // Use movaps / movups for SSE integer load / store (one byte shorter).
3117 def : Pat<(alignedloadv4i32 addr:$src),
3118 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
3119 def : Pat<(loadv4i32 addr:$src),
3120 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3121 def : Pat<(alignedloadv2i64 addr:$src),
3122 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3123 def : Pat<(loadv2i64 addr:$src),
3124 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3126 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3127 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3128 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3129 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3130 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3131 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3132 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3133 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3134 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3135 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3136 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3137 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3138 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3139 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3140 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3141 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3143 //===----------------------------------------------------------------------===//
3144 // SSE4.1 Instructions
3145 //===----------------------------------------------------------------------===//
3147 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3150 Intrinsic V2F64Int> {
3151 // Intrinsic operation, reg.
3152 // Vector intrinsic operation, reg
3153 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3154 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3155 !strconcat(OpcodeStr,
3156 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3157 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3160 // Vector intrinsic operation, mem
3161 def PSm_Int : SS4AIi8<opcps, MRMSrcMem,
3162 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3163 !strconcat(OpcodeStr,
3164 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3166 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3169 // Vector intrinsic operation, reg
3170 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3171 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3172 !strconcat(OpcodeStr,
3173 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3174 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3177 // Vector intrinsic operation, mem
3178 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3179 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3180 !strconcat(OpcodeStr,
3181 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3183 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3187 let Constraints = "$src1 = $dst" in {
3188 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3192 // Intrinsic operation, reg.
3193 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3195 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3196 !strconcat(OpcodeStr,
3197 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3199 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3202 // Intrinsic operation, mem.
3203 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3205 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3206 !strconcat(OpcodeStr,
3207 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3209 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3212 // Intrinsic operation, reg.
3213 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3215 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3216 !strconcat(OpcodeStr,
3217 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3219 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3222 // Intrinsic operation, mem.
3223 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3225 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3226 !strconcat(OpcodeStr,
3227 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3229 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3234 // FP round - roundss, roundps, roundsd, roundpd
3235 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3236 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3237 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3238 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3240 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3241 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3242 Intrinsic IntId128> {
3243 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3245 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3246 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3247 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3249 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3252 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3255 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3256 int_x86_sse41_phminposuw>;
3258 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3259 let Constraints = "$src1 = $dst" in {
3260 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3261 Intrinsic IntId128, bit Commutable = 0> {
3262 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3263 (ins VR128:$src1, VR128:$src2),
3264 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3265 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3267 let isCommutable = Commutable;
3269 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3270 (ins VR128:$src1, i128mem:$src2),
3271 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3273 (IntId128 VR128:$src1,
3274 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3278 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3279 int_x86_sse41_pcmpeqq, 1>;
3280 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3281 int_x86_sse41_packusdw, 0>;
3282 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3283 int_x86_sse41_pminsb, 1>;
3284 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3285 int_x86_sse41_pminsd, 1>;
3286 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3287 int_x86_sse41_pminud, 1>;
3288 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3289 int_x86_sse41_pminuw, 1>;
3290 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3291 int_x86_sse41_pmaxsb, 1>;
3292 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3293 int_x86_sse41_pmaxsd, 1>;
3294 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3295 int_x86_sse41_pmaxud, 1>;
3296 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3297 int_x86_sse41_pmaxuw, 1>;
3299 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3301 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3302 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3303 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3304 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3306 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3307 let Constraints = "$src1 = $dst" in {
3308 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3309 SDNode OpNode, Intrinsic IntId128,
3310 bit Commutable = 0> {
3311 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3312 (ins VR128:$src1, VR128:$src2),
3313 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3314 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3315 VR128:$src2))]>, OpSize {
3316 let isCommutable = Commutable;
3318 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3319 (ins VR128:$src1, VR128:$src2),
3320 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3321 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3323 let isCommutable = Commutable;
3325 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3326 (ins VR128:$src1, i128mem:$src2),
3327 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3329 (OpNode VR128:$src1, (memop addr:$src2)))]>, OpSize;
3330 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3331 (ins VR128:$src1, i128mem:$src2),
3332 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3334 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3338 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", v4i32, mul,
3339 int_x86_sse41_pmulld, 1>;
3341 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3342 let Constraints = "$src1 = $dst" in {
3343 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3344 Intrinsic IntId128, bit Commutable = 0> {
3345 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3346 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3347 !strconcat(OpcodeStr,
3348 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3350 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3352 let isCommutable = Commutable;
3354 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3355 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3356 !strconcat(OpcodeStr,
3357 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3359 (IntId128 VR128:$src1,
3360 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3365 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3366 int_x86_sse41_blendps, 0>;
3367 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3368 int_x86_sse41_blendpd, 0>;
3369 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3370 int_x86_sse41_pblendw, 0>;
3371 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3372 int_x86_sse41_dpps, 1>;
3373 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3374 int_x86_sse41_dppd, 1>;
3375 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3376 int_x86_sse41_mpsadbw, 1>;
3379 /// SS41I_ternary_int - SSE 4.1 ternary operator
3380 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3381 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3382 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3383 (ins VR128:$src1, VR128:$src2),
3384 !strconcat(OpcodeStr,
3385 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3386 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3389 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3390 (ins VR128:$src1, i128mem:$src2),
3391 !strconcat(OpcodeStr,
3392 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3395 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3399 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3400 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3401 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3404 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3405 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3406 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3407 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3409 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3410 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3412 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3416 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3417 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3418 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3419 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3420 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3421 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3423 // Common patterns involving scalar load.
3424 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3425 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3426 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3427 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3429 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3430 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3431 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3432 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3434 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3435 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3436 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3437 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3439 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3440 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3441 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3442 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3444 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3445 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3446 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3447 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3449 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3450 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3451 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3452 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3455 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3456 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3457 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3458 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3460 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3461 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3463 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3467 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3468 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3469 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3470 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3472 // Common patterns involving scalar load
3473 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3474 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3475 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3476 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3478 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3479 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3480 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3481 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3484 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3485 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3486 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3487 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3489 // Expecting a i16 load any extended to i32 value.
3490 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3491 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3492 [(set VR128:$dst, (IntId (bitconvert
3493 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3497 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3498 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3500 // Common patterns involving scalar load
3501 def : Pat<(int_x86_sse41_pmovsxbq
3502 (bitconvert (v4i32 (X86vzmovl
3503 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3504 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3506 def : Pat<(int_x86_sse41_pmovzxbq
3507 (bitconvert (v4i32 (X86vzmovl
3508 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3509 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3512 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3513 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3514 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3515 (ins VR128:$src1, i32i8imm:$src2),
3516 !strconcat(OpcodeStr,
3517 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3518 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3520 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3521 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3522 !strconcat(OpcodeStr,
3523 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3526 // There's an AssertZext in the way of writing the store pattern
3527 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3530 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3533 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3534 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3535 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3536 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3537 !strconcat(OpcodeStr,
3538 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3541 // There's an AssertZext in the way of writing the store pattern
3542 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3545 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3548 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3549 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3550 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3551 (ins VR128:$src1, i32i8imm:$src2),
3552 !strconcat(OpcodeStr,
3553 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3555 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3556 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3557 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3558 !strconcat(OpcodeStr,
3559 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3560 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3561 addr:$dst)]>, OpSize;
3564 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3567 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3569 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3570 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3571 (ins VR128:$src1, i32i8imm:$src2),
3572 !strconcat(OpcodeStr,
3573 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3575 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3577 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3578 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3579 !strconcat(OpcodeStr,
3580 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3581 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3582 addr:$dst)]>, OpSize;
3585 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3587 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3588 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3591 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3592 Requires<[HasSSE41]>;
3594 let Constraints = "$src1 = $dst" in {
3595 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3596 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3597 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3598 !strconcat(OpcodeStr,
3599 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3601 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3602 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3603 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3604 !strconcat(OpcodeStr,
3605 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3607 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3608 imm:$src3))]>, OpSize;
3612 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3614 let Constraints = "$src1 = $dst" in {
3615 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3616 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3617 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3618 !strconcat(OpcodeStr,
3619 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3621 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3623 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3624 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3625 !strconcat(OpcodeStr,
3626 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3628 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3629 imm:$src3)))]>, OpSize;
3633 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3635 // insertps has a few different modes, there's the first two here below which
3636 // are optimized inserts that won't zero arbitrary elements in the destination
3637 // vector. The next one matches the intrinsic and could zero arbitrary elements
3638 // in the target vector.
3639 let Constraints = "$src1 = $dst" in {
3640 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3641 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3642 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3643 !strconcat(OpcodeStr,
3644 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3646 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3648 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3649 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3650 !strconcat(OpcodeStr,
3651 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3653 (X86insrtps VR128:$src1,
3654 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3655 imm:$src3))]>, OpSize;
3659 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3661 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3662 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3664 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3665 // the intel intrinsic that corresponds to this.
3666 let Defs = [EFLAGS] in {
3667 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3668 "ptest \t{$src2, $src1|$src1, $src2}",
3669 [(X86ptest VR128:$src1, VR128:$src2),
3670 (implicit EFLAGS)]>, OpSize;
3671 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3672 "ptest \t{$src2, $src1|$src1, $src2}",
3673 [(X86ptest VR128:$src1, (load addr:$src2)),
3674 (implicit EFLAGS)]>, OpSize;
3677 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3678 "movntdqa\t{$src, $dst|$dst, $src}",
3679 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;
3682 //===----------------------------------------------------------------------===//
3683 // SSE4.2 Instructions
3684 //===----------------------------------------------------------------------===//
3686 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3687 let Constraints = "$src1 = $dst" in {
3688 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3689 Intrinsic IntId128, bit Commutable = 0> {
3690 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3691 (ins VR128:$src1, VR128:$src2),
3692 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3693 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3695 let isCommutable = Commutable;
3697 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3698 (ins VR128:$src1, i128mem:$src2),
3699 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3701 (IntId128 VR128:$src1,
3702 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3706 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3708 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3709 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3710 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3711 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3713 // crc intrinsic instruction
3714 // This set of instructions are only rm, the only difference is the size
3716 let Constraints = "$src1 = $dst" in {
3717 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3718 (ins GR32:$src1, i8mem:$src2),
3719 "crc32 \t{$src2, $src1|$src1, $src2}",
3721 (int_x86_sse42_crc32_8 GR32:$src1,
3722 (load addr:$src2)))]>, OpSize;
3723 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3724 (ins GR32:$src1, GR8:$src2),
3725 "crc32 \t{$src2, $src1|$src1, $src2}",
3727 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>,
3729 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3730 (ins GR32:$src1, i16mem:$src2),
3731 "crc32 \t{$src2, $src1|$src1, $src2}",
3733 (int_x86_sse42_crc32_16 GR32:$src1,
3734 (load addr:$src2)))]>,
3736 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3737 (ins GR32:$src1, GR16:$src2),
3738 "crc32 \t{$src2, $src1|$src1, $src2}",
3740 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3742 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3743 (ins GR32:$src1, i32mem:$src2),
3744 "crc32 \t{$src2, $src1|$src1, $src2}",
3746 (int_x86_sse42_crc32_32 GR32:$src1,
3747 (load addr:$src2)))]>, OpSize;
3748 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3749 (ins GR32:$src1, GR32:$src2),
3750 "crc32 \t{$src2, $src1|$src1, $src2}",
3752 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>,
3754 def CRC64m64 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3755 (ins GR64:$src1, i64mem:$src2),
3756 "crc32 \t{$src2, $src1|$src1, $src2}",
3758 (int_x86_sse42_crc32_64 GR64:$src1,
3759 (load addr:$src2)))]>,
3761 def CRC64r64 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3762 (ins GR64:$src1, GR64:$src2),
3763 "crc32 \t{$src2, $src1|$src1, $src2}",
3765 (int_x86_sse42_crc32_64 GR64:$src1, GR64:$src2))]>,
3769 // String/text processing instructions.
3770 let Defs = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
3771 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3772 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3773 "#PCMPISTRM128rr PSEUDO!",
3775 (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3776 imm:$src3))]>, OpSize;
3777 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3778 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3779 "#PCMPISTRM128rm PSEUDO!",
3781 (int_x86_sse42_pcmpistrm128 VR128:$src1,
3783 imm:$src3))]>, OpSize;
3786 let Defs = [XMM0, EFLAGS] in {
3787 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3788 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3789 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3791 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3792 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3793 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3797 let Defs = [EFLAGS], Uses = [EAX, EDX],
3798 usesCustomDAGSchedInserter = 1 in {
3799 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3800 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3801 "#PCMPESTRM128rr PSEUDO!",
3803 (int_x86_sse42_pcmpestrm128 VR128:$src1, EAX,
3805 EDX, imm:$src5))]>, OpSize;
3806 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3807 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3808 "#PCMPESTRM128rm PSEUDO!",
3810 (int_x86_sse42_pcmpestrm128 VR128:$src1, EAX,
3812 EDX, imm:$src5))]>, OpSize;
3815 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3816 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3817 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3818 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3820 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3821 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3822 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3826 let Defs = [ECX, EFLAGS] in {
3827 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3828 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3829 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3830 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3832 (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3833 (implicit EFLAGS)]>,
3835 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3836 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3837 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3839 (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3840 (implicit EFLAGS)]>,
3845 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3846 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3847 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3848 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3849 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3850 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3852 let Defs = [ECX, EFLAGS] in {
3853 let Uses = [EAX, EDX] in {
3854 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3855 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3856 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3857 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3859 (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3860 (implicit EFLAGS)]>,
3862 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3863 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3864 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3866 (IntId128 VR128:$src1, EAX, (load addr:$src3),
3868 (implicit EFLAGS)]>,
3874 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3875 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3876 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3877 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3878 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3879 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;