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:$src2)),
471 // Aliases of packed SSE1 instructions for scalar use. These all have names
472 // that start with 'Fs'.
474 // Alias instructions that map fld0 to pxor for sse.
475 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
477 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
478 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
479 Requires<[HasSSE1]>, TB, OpSize;
481 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
483 let neverHasSideEffects = 1 in
484 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
485 "movaps\t{$src, $dst|$dst, $src}", []>;
487 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
489 let canFoldAsLoad = 1 in
490 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
491 "movaps\t{$src, $dst|$dst, $src}",
492 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
494 // Alias bitwise logical operations using SSE logical ops on packed FP values.
495 let Constraints = "$src1 = $dst" in {
496 let isCommutable = 1 in {
497 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst),
498 (ins FR32:$src1, FR32:$src2),
499 "andps\t{$src2, $dst|$dst, $src2}",
500 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
501 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst),
502 (ins FR32:$src1, FR32:$src2),
503 "orps\t{$src2, $dst|$dst, $src2}",
504 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
505 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst),
506 (ins FR32:$src1, FR32:$src2),
507 "xorps\t{$src2, $dst|$dst, $src2}",
508 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
511 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst),
512 (ins FR32:$src1, f128mem:$src2),
513 "andps\t{$src2, $dst|$dst, $src2}",
514 [(set FR32:$dst, (X86fand FR32:$src1,
515 (memopfsf32 addr:$src2)))]>;
516 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst),
517 (ins FR32:$src1, f128mem:$src2),
518 "orps\t{$src2, $dst|$dst, $src2}",
519 [(set FR32:$dst, (X86for FR32:$src1,
520 (memopfsf32 addr:$src2)))]>;
521 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst),
522 (ins FR32:$src1, f128mem:$src2),
523 "xorps\t{$src2, $dst|$dst, $src2}",
524 [(set FR32:$dst, (X86fxor FR32:$src1,
525 (memopfsf32 addr:$src2)))]>;
527 let neverHasSideEffects = 1 in {
528 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
529 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
530 "andnps\t{$src2, $dst|$dst, $src2}", []>;
532 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
533 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
534 "andnps\t{$src2, $dst|$dst, $src2}", []>;
538 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
540 /// In addition, we also have a special variant of the scalar form here to
541 /// represent the associated intrinsic operation. This form is unlike the
542 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
543 /// and leaves the top elements unmodified (therefore these cannot be commuted).
545 /// These three forms can each be reg+reg or reg+mem, so there are a total of
546 /// six "instructions".
548 let Constraints = "$src1 = $dst" in {
549 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
550 SDNode OpNode, Intrinsic F32Int,
551 bit Commutable = 0> {
552 // Scalar operation, reg+reg.
553 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
554 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
555 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
556 let isCommutable = Commutable;
559 // Scalar operation, reg+mem.
560 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
561 (ins FR32:$src1, f32mem:$src2),
562 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
563 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
565 // Vector operation, reg+reg.
566 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
567 (ins VR128:$src1, VR128:$src2),
568 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
569 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
570 let isCommutable = Commutable;
573 // Vector operation, reg+mem.
574 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
575 (ins VR128:$src1, f128mem:$src2),
576 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
577 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
579 // Intrinsic operation, reg+reg.
580 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
581 (ins VR128:$src1, VR128:$src2),
582 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
583 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]>;
585 // Intrinsic operation, reg+mem.
586 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
587 (ins VR128:$src1, ssmem:$src2),
588 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
589 [(set VR128:$dst, (F32Int VR128:$src1,
590 sse_load_f32:$src2))]>;
594 // Arithmetic instructions
595 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
596 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
597 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
598 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
600 /// sse1_fp_binop_rm - Other SSE1 binops
602 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
603 /// instructions for a full-vector intrinsic form. Operations that map
604 /// onto C operators don't use this form since they just use the plain
605 /// vector form instead of having a separate vector intrinsic form.
607 /// This provides a total of eight "instructions".
609 let Constraints = "$src1 = $dst" in {
610 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
614 bit Commutable = 0> {
616 // Scalar operation, reg+reg.
617 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
618 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
619 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
620 let isCommutable = Commutable;
623 // Scalar operation, reg+mem.
624 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
625 (ins FR32:$src1, f32mem:$src2),
626 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
627 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
629 // Vector operation, reg+reg.
630 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
631 (ins VR128:$src1, VR128:$src2),
632 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
633 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
634 let isCommutable = Commutable;
637 // Vector operation, reg+mem.
638 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
639 (ins VR128:$src1, f128mem:$src2),
640 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
641 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
643 // Intrinsic operation, reg+reg.
644 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
645 (ins VR128:$src1, VR128:$src2),
646 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
647 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
648 let isCommutable = Commutable;
651 // Intrinsic operation, reg+mem.
652 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
653 (ins VR128:$src1, ssmem:$src2),
654 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
655 [(set VR128:$dst, (F32Int VR128:$src1,
656 sse_load_f32:$src2))]>;
658 // Vector intrinsic operation, reg+reg.
659 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
660 (ins VR128:$src1, VR128:$src2),
661 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
662 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
663 let isCommutable = Commutable;
666 // Vector intrinsic operation, reg+mem.
667 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
668 (ins VR128:$src1, f128mem:$src2),
669 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
670 [(set VR128:$dst, (V4F32Int VR128:$src1, (memopv4f32 addr:$src2)))]>;
674 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
675 int_x86_sse_max_ss, int_x86_sse_max_ps>;
676 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
677 int_x86_sse_min_ss, int_x86_sse_min_ps>;
679 //===----------------------------------------------------------------------===//
680 // SSE packed FP Instructions
683 let neverHasSideEffects = 1 in
684 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
685 "movaps\t{$src, $dst|$dst, $src}", []>;
686 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
687 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
688 "movaps\t{$src, $dst|$dst, $src}",
689 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
691 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
692 "movaps\t{$src, $dst|$dst, $src}",
693 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
695 let neverHasSideEffects = 1 in
696 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
697 "movups\t{$src, $dst|$dst, $src}", []>;
698 let canFoldAsLoad = 1 in
699 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
700 "movups\t{$src, $dst|$dst, $src}",
701 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
702 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
703 "movups\t{$src, $dst|$dst, $src}",
704 [(store (v4f32 VR128:$src), addr:$dst)]>;
706 // Intrinsic forms of MOVUPS load and store
707 let canFoldAsLoad = 1 in
708 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
709 "movups\t{$src, $dst|$dst, $src}",
710 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
711 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
712 "movups\t{$src, $dst|$dst, $src}",
713 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
715 let Constraints = "$src1 = $dst" in {
716 let AddedComplexity = 20 in {
717 def MOVLPSrm : PSI<0x12, MRMSrcMem,
718 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
719 "movlps\t{$src2, $dst|$dst, $src2}",
722 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
723 def MOVHPSrm : PSI<0x16, MRMSrcMem,
724 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
725 "movhps\t{$src2, $dst|$dst, $src2}",
728 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
730 } // Constraints = "$src1 = $dst"
733 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
734 "movlps\t{$src, $dst|$dst, $src}",
735 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
736 (iPTR 0))), addr:$dst)]>;
738 // v2f64 extract element 1 is always custom lowered to unpack high to low
739 // and extract element 0 so the non-store version isn't too horrible.
740 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
741 "movhps\t{$src, $dst|$dst, $src}",
742 [(store (f64 (vector_extract
743 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
744 (undef)), (iPTR 0))), addr:$dst)]>;
746 let Constraints = "$src1 = $dst" in {
747 let AddedComplexity = 20 in {
748 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
749 (ins VR128:$src1, VR128:$src2),
750 "movlhps\t{$src2, $dst|$dst, $src2}",
752 (v4f32 (movhp VR128:$src1, VR128:$src2)))]>;
754 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
755 (ins VR128:$src1, VR128:$src2),
756 "movhlps\t{$src2, $dst|$dst, $src2}",
758 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
760 } // Constraints = "$src1 = $dst"
762 let AddedComplexity = 20 in {
763 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
764 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
765 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
766 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
773 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
775 /// In addition, we also have a special variant of the scalar form here to
776 /// represent the associated intrinsic operation. This form is unlike the
777 /// plain scalar form, in that it takes an entire vector (instead of a
778 /// scalar) and leaves the top elements undefined.
780 /// And, we have a special variant form for a full-vector intrinsic form.
782 /// These four forms can each have a reg or a mem operand, so there are a
783 /// total of eight "instructions".
785 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
789 bit Commutable = 0> {
790 // Scalar operation, reg.
791 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
792 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
793 [(set FR32:$dst, (OpNode FR32:$src))]> {
794 let isCommutable = Commutable;
797 // Scalar operation, mem.
798 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
799 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
800 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
802 // Vector operation, reg.
803 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
804 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
805 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
806 let isCommutable = Commutable;
809 // Vector operation, mem.
810 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
811 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
812 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
814 // Intrinsic operation, reg.
815 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
816 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
817 [(set VR128:$dst, (F32Int VR128:$src))]> {
818 let isCommutable = Commutable;
821 // Intrinsic operation, mem.
822 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
823 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
824 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
826 // Vector intrinsic operation, reg
827 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
828 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
829 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
830 let isCommutable = Commutable;
833 // Vector intrinsic operation, mem
834 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
835 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
836 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
840 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
841 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
843 // Reciprocal approximations. Note that these typically require refinement
844 // in order to obtain suitable precision.
845 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
846 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
847 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
848 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
851 let Constraints = "$src1 = $dst" in {
852 let isCommutable = 1 in {
853 def ANDPSrr : PSI<0x54, MRMSrcReg,
854 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
855 "andps\t{$src2, $dst|$dst, $src2}",
856 [(set VR128:$dst, (v2i64
857 (and VR128:$src1, VR128:$src2)))]>;
858 def ORPSrr : PSI<0x56, MRMSrcReg,
859 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
860 "orps\t{$src2, $dst|$dst, $src2}",
861 [(set VR128:$dst, (v2i64
862 (or VR128:$src1, VR128:$src2)))]>;
863 def XORPSrr : PSI<0x57, MRMSrcReg,
864 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
865 "xorps\t{$src2, $dst|$dst, $src2}",
866 [(set VR128:$dst, (v2i64
867 (xor VR128:$src1, VR128:$src2)))]>;
870 def ANDPSrm : PSI<0x54, MRMSrcMem,
871 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
872 "andps\t{$src2, $dst|$dst, $src2}",
873 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
874 (memopv2i64 addr:$src2)))]>;
875 def ORPSrm : PSI<0x56, MRMSrcMem,
876 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
877 "orps\t{$src2, $dst|$dst, $src2}",
878 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
879 (memopv2i64 addr:$src2)))]>;
880 def XORPSrm : PSI<0x57, MRMSrcMem,
881 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
882 "xorps\t{$src2, $dst|$dst, $src2}",
883 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
884 (memopv2i64 addr:$src2)))]>;
885 def ANDNPSrr : PSI<0x55, MRMSrcReg,
886 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
887 "andnps\t{$src2, $dst|$dst, $src2}",
889 (v2i64 (and (xor VR128:$src1,
890 (bc_v2i64 (v4i32 immAllOnesV))),
892 def ANDNPSrm : PSI<0x55, MRMSrcMem,
893 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
894 "andnps\t{$src2, $dst|$dst, $src2}",
896 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
897 (bc_v2i64 (v4i32 immAllOnesV))),
898 (memopv2i64 addr:$src2))))]>;
901 let Constraints = "$src1 = $dst" in {
902 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
903 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
904 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
905 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
906 VR128:$src, imm:$cc))]>;
907 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
908 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
909 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
910 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
911 (memop addr:$src), imm:$cc))]>;
913 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
914 (CMPPSrri VR128:$src1, VR128:$src2, imm:$cc)>;
915 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
916 (CMPPSrmi VR128:$src1, addr:$src2, imm:$cc)>;
918 // Shuffle and unpack instructions
919 let Constraints = "$src1 = $dst" in {
920 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
921 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
922 (outs VR128:$dst), (ins VR128:$src1,
923 VR128:$src2, i8imm:$src3),
924 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
926 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
927 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
928 (outs VR128:$dst), (ins VR128:$src1,
929 f128mem:$src2, i8imm:$src3),
930 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
933 VR128:$src1, (memopv4f32 addr:$src2))))]>;
935 let AddedComplexity = 10 in {
936 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
937 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
938 "unpckhps\t{$src2, $dst|$dst, $src2}",
940 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
941 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
942 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
943 "unpckhps\t{$src2, $dst|$dst, $src2}",
945 (v4f32 (unpckh VR128:$src1,
946 (memopv4f32 addr:$src2))))]>;
948 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
949 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
950 "unpcklps\t{$src2, $dst|$dst, $src2}",
952 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
953 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
954 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
955 "unpcklps\t{$src2, $dst|$dst, $src2}",
957 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
959 } // Constraints = "$src1 = $dst"
962 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
963 "movmskps\t{$src, $dst|$dst, $src}",
964 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
965 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
966 "movmskpd\t{$src, $dst|$dst, $src}",
967 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
969 // Prefetch intrinsic.
970 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
971 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
972 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
973 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
974 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
975 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
976 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
977 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
979 // Non-temporal stores
980 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
981 "movntps\t{$src, $dst|$dst, $src}",
982 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
984 // Load, store, and memory fence
985 def SFENCE : PSI<0xAE, MRM7r, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
988 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
989 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
990 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
991 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
993 // Alias instructions that map zero vector to pxor / xorp* for sse.
994 // We set canFoldAsLoad because this can be converted to a constant-pool
995 // load of an all-zeros value if folding it would be beneficial.
996 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
998 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
1000 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1002 let Predicates = [HasSSE1] in {
1003 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
1004 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
1005 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
1006 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
1007 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
1010 // FR32 to 128-bit vector conversion.
1011 let isAsCheapAsAMove = 1 in
1012 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
1013 "movss\t{$src, $dst|$dst, $src}",
1015 (v4f32 (scalar_to_vector FR32:$src)))]>;
1016 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1017 "movss\t{$src, $dst|$dst, $src}",
1019 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
1021 // FIXME: may not be able to eliminate this movss with coalescing the src and
1022 // dest register classes are different. We really want to write this pattern
1024 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1025 // (f32 FR32:$src)>;
1026 let isAsCheapAsAMove = 1 in
1027 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
1028 "movss\t{$src, $dst|$dst, $src}",
1029 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
1031 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
1032 "movss\t{$src, $dst|$dst, $src}",
1033 [(store (f32 (vector_extract (v4f32 VR128:$src),
1034 (iPTR 0))), addr:$dst)]>;
1037 // Move to lower bits of a VR128, leaving upper bits alone.
1038 // Three operand (but two address) aliases.
1039 let Constraints = "$src1 = $dst" in {
1040 let neverHasSideEffects = 1 in
1041 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
1042 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
1043 "movss\t{$src2, $dst|$dst, $src2}", []>;
1045 let AddedComplexity = 15 in
1046 def MOVLPSrr : SSI<0x10, MRMSrcReg,
1047 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1048 "movss\t{$src2, $dst|$dst, $src2}",
1050 (v4f32 (movl VR128:$src1, VR128:$src2)))]>;
1053 // Move to lower bits of a VR128 and zeroing upper bits.
1054 // Loading from memory automatically zeroing upper bits.
1055 let AddedComplexity = 20 in
1056 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1057 "movss\t{$src, $dst|$dst, $src}",
1058 [(set VR128:$dst, (v4f32 (X86vzmovl (v4f32 (scalar_to_vector
1059 (loadf32 addr:$src))))))]>;
1061 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
1062 (MOVZSS2PSrm addr:$src)>;
1064 //===---------------------------------------------------------------------===//
1065 // SSE2 Instructions
1066 //===---------------------------------------------------------------------===//
1068 // Move Instructions
1069 let neverHasSideEffects = 1 in
1070 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1071 "movsd\t{$src, $dst|$dst, $src}", []>;
1072 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1073 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1074 "movsd\t{$src, $dst|$dst, $src}",
1075 [(set FR64:$dst, (loadf64 addr:$src))]>;
1076 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1077 "movsd\t{$src, $dst|$dst, $src}",
1078 [(store FR64:$src, addr:$dst)]>;
1080 // Conversion instructions
1081 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1082 "cvttsd2si\t{$src, $dst|$dst, $src}",
1083 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1084 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1085 "cvttsd2si\t{$src, $dst|$dst, $src}",
1086 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1087 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1088 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1089 [(set FR32:$dst, (fround FR64:$src))]>;
1090 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1091 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1092 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1093 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1094 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1095 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1096 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1097 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1098 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1100 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1101 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1102 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1103 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1104 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1105 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1106 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1107 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1108 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1109 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1110 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1111 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1112 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1113 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1114 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1115 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1116 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1117 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1118 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1119 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1121 // SSE2 instructions with XS prefix
1122 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1123 "cvtss2sd\t{$src, $dst|$dst, $src}",
1124 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1125 Requires<[HasSSE2]>;
1126 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1127 "cvtss2sd\t{$src, $dst|$dst, $src}",
1128 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1129 Requires<[HasSSE2]>;
1131 // Match intrinsics which expect XMM operand(s).
1132 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1133 "cvtsd2si\t{$src, $dst|$dst, $src}",
1134 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1135 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1136 "cvtsd2si\t{$src, $dst|$dst, $src}",
1137 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1138 (load addr:$src)))]>;
1140 // Match intrinisics which expect MM and XMM operand(s).
1141 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1142 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1143 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1144 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1145 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1146 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1147 (memop addr:$src)))]>;
1148 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1149 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1150 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1151 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1152 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1153 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1154 (memop addr:$src)))]>;
1155 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1156 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1157 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1158 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1159 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1160 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1161 (load addr:$src)))]>;
1163 // Aliases for intrinsics
1164 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1165 "cvttsd2si\t{$src, $dst|$dst, $src}",
1167 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1168 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1169 "cvttsd2si\t{$src, $dst|$dst, $src}",
1170 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1171 (load addr:$src)))]>;
1173 // Comparison instructions
1174 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1175 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1176 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1177 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1179 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1180 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1181 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1184 let Defs = [EFLAGS] in {
1185 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1186 "ucomisd\t{$src2, $src1|$src1, $src2}",
1187 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1188 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1189 "ucomisd\t{$src2, $src1|$src1, $src2}",
1190 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1191 (implicit EFLAGS)]>;
1192 } // Defs = [EFLAGS]
1194 // Aliases to match intrinsics which expect XMM operand(s).
1195 let Constraints = "$src1 = $dst" in {
1196 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1197 (outs VR128:$dst), (ins VR128:$src1, VR128:$src,
1199 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1200 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1201 VR128:$src, imm:$cc))]>;
1202 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1203 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src,
1205 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1206 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1207 (load addr:$src), imm:$cc))]>;
1210 let Defs = [EFLAGS] in {
1211 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1212 "ucomisd\t{$src2, $src1|$src1, $src2}",
1213 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1214 (implicit EFLAGS)]>;
1215 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1216 "ucomisd\t{$src2, $src1|$src1, $src2}",
1217 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1218 (implicit EFLAGS)]>;
1220 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1221 "comisd\t{$src2, $src1|$src1, $src2}",
1222 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1223 (implicit EFLAGS)]>;
1224 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1225 "comisd\t{$src2, $src1|$src1, $src2}",
1226 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1227 (implicit EFLAGS)]>;
1228 } // Defs = [EFLAGS]
1230 // Aliases of packed SSE2 instructions for scalar use. These all have names
1231 // that start with 'Fs'.
1233 // Alias instructions that map fld0 to pxor for sse.
1234 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1235 canFoldAsLoad = 1 in
1236 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1237 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1238 Requires<[HasSSE2]>, TB, OpSize;
1240 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1242 let neverHasSideEffects = 1 in
1243 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1244 "movapd\t{$src, $dst|$dst, $src}", []>;
1246 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1248 let canFoldAsLoad = 1 in
1249 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1250 "movapd\t{$src, $dst|$dst, $src}",
1251 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1253 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1254 let Constraints = "$src1 = $dst" in {
1255 let isCommutable = 1 in {
1256 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1257 (ins FR64:$src1, FR64:$src2),
1258 "andpd\t{$src2, $dst|$dst, $src2}",
1259 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1260 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1261 (ins FR64:$src1, FR64:$src2),
1262 "orpd\t{$src2, $dst|$dst, $src2}",
1263 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1264 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1265 (ins FR64:$src1, FR64:$src2),
1266 "xorpd\t{$src2, $dst|$dst, $src2}",
1267 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1270 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1271 (ins FR64:$src1, f128mem:$src2),
1272 "andpd\t{$src2, $dst|$dst, $src2}",
1273 [(set FR64:$dst, (X86fand FR64:$src1,
1274 (memopfsf64 addr:$src2)))]>;
1275 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1276 (ins FR64:$src1, f128mem:$src2),
1277 "orpd\t{$src2, $dst|$dst, $src2}",
1278 [(set FR64:$dst, (X86for FR64:$src1,
1279 (memopfsf64 addr:$src2)))]>;
1280 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1281 (ins FR64:$src1, f128mem:$src2),
1282 "xorpd\t{$src2, $dst|$dst, $src2}",
1283 [(set FR64:$dst, (X86fxor FR64:$src1,
1284 (memopfsf64 addr:$src2)))]>;
1286 let neverHasSideEffects = 1 in {
1287 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1288 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1289 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1291 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1292 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1293 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1297 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1299 /// In addition, we also have a special variant of the scalar form here to
1300 /// represent the associated intrinsic operation. This form is unlike the
1301 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1302 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1304 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1305 /// six "instructions".
1307 let Constraints = "$src1 = $dst" in {
1308 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1309 SDNode OpNode, Intrinsic F64Int,
1310 bit Commutable = 0> {
1311 // Scalar operation, reg+reg.
1312 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1313 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1314 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1315 let isCommutable = Commutable;
1318 // Scalar operation, reg+mem.
1319 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1320 (ins FR64:$src1, f64mem:$src2),
1321 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1322 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1324 // Vector operation, reg+reg.
1325 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1326 (ins VR128:$src1, VR128:$src2),
1327 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1328 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1329 let isCommutable = Commutable;
1332 // Vector operation, reg+mem.
1333 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1334 (ins VR128:$src1, f128mem:$src2),
1335 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1336 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1338 // Intrinsic operation, reg+reg.
1339 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1340 (ins VR128:$src1, VR128:$src2),
1341 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1342 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]>;
1344 // Intrinsic operation, reg+mem.
1345 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1346 (ins VR128:$src1, sdmem:$src2),
1347 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1348 [(set VR128:$dst, (F64Int VR128:$src1,
1349 sse_load_f64:$src2))]>;
1353 // Arithmetic instructions
1354 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1355 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1356 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1357 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1359 /// sse2_fp_binop_rm - Other SSE2 binops
1361 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1362 /// instructions for a full-vector intrinsic form. Operations that map
1363 /// onto C operators don't use this form since they just use the plain
1364 /// vector form instead of having a separate vector intrinsic form.
1366 /// This provides a total of eight "instructions".
1368 let Constraints = "$src1 = $dst" in {
1369 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1373 bit Commutable = 0> {
1375 // Scalar operation, reg+reg.
1376 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1377 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1378 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1379 let isCommutable = Commutable;
1382 // Scalar operation, reg+mem.
1383 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1384 (ins FR64:$src1, f64mem:$src2),
1385 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1386 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1388 // Vector operation, reg+reg.
1389 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1390 (ins VR128:$src1, VR128:$src2),
1391 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1392 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1393 let isCommutable = Commutable;
1396 // Vector operation, reg+mem.
1397 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1398 (ins VR128:$src1, f128mem:$src2),
1399 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1400 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1402 // Intrinsic operation, reg+reg.
1403 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1404 (ins VR128:$src1, VR128:$src2),
1405 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1406 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1407 let isCommutable = Commutable;
1410 // Intrinsic operation, reg+mem.
1411 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1412 (ins VR128:$src1, sdmem:$src2),
1413 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1414 [(set VR128:$dst, (F64Int VR128:$src1,
1415 sse_load_f64:$src2))]>;
1417 // Vector intrinsic operation, reg+reg.
1418 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1419 (ins VR128:$src1, VR128:$src2),
1420 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1421 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1422 let isCommutable = Commutable;
1425 // Vector intrinsic operation, reg+mem.
1426 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1427 (ins VR128:$src1, f128mem:$src2),
1428 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1429 [(set VR128:$dst, (V2F64Int VR128:$src1,
1430 (memopv2f64 addr:$src2)))]>;
1434 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1435 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1436 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1437 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1439 //===---------------------------------------------------------------------===//
1440 // SSE packed FP Instructions
1442 // Move Instructions
1443 let neverHasSideEffects = 1 in
1444 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1445 "movapd\t{$src, $dst|$dst, $src}", []>;
1446 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1447 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1448 "movapd\t{$src, $dst|$dst, $src}",
1449 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1451 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1452 "movapd\t{$src, $dst|$dst, $src}",
1453 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1455 let neverHasSideEffects = 1 in
1456 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1457 "movupd\t{$src, $dst|$dst, $src}", []>;
1458 let canFoldAsLoad = 1 in
1459 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1460 "movupd\t{$src, $dst|$dst, $src}",
1461 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1462 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1463 "movupd\t{$src, $dst|$dst, $src}",
1464 [(store (v2f64 VR128:$src), addr:$dst)]>;
1466 // Intrinsic forms of MOVUPD load and store
1467 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1468 "movupd\t{$src, $dst|$dst, $src}",
1469 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1470 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1471 "movupd\t{$src, $dst|$dst, $src}",
1472 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1474 let Constraints = "$src1 = $dst" in {
1475 let AddedComplexity = 20 in {
1476 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1477 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1478 "movlpd\t{$src2, $dst|$dst, $src2}",
1480 (v2f64 (movlp VR128:$src1,
1481 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1482 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1483 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1484 "movhpd\t{$src2, $dst|$dst, $src2}",
1486 (v2f64 (movhp VR128:$src1,
1487 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1488 } // AddedComplexity
1489 } // Constraints = "$src1 = $dst"
1491 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1492 "movlpd\t{$src, $dst|$dst, $src}",
1493 [(store (f64 (vector_extract (v2f64 VR128:$src),
1494 (iPTR 0))), addr:$dst)]>;
1496 // v2f64 extract element 1 is always custom lowered to unpack high to low
1497 // and extract element 0 so the non-store version isn't too horrible.
1498 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1499 "movhpd\t{$src, $dst|$dst, $src}",
1500 [(store (f64 (vector_extract
1501 (v2f64 (unpckh VR128:$src, (undef))),
1502 (iPTR 0))), addr:$dst)]>;
1504 // SSE2 instructions without OpSize prefix
1505 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1506 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1507 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1508 TB, Requires<[HasSSE2]>;
1509 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1510 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1511 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1512 (bitconvert (memopv2i64 addr:$src))))]>,
1513 TB, Requires<[HasSSE2]>;
1515 // SSE2 instructions with XS prefix
1516 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1517 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1518 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1519 XS, Requires<[HasSSE2]>;
1520 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1521 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1522 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1523 (bitconvert (memopv2i64 addr:$src))))]>,
1524 XS, Requires<[HasSSE2]>;
1526 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1527 "cvtps2dq\t{$src, $dst|$dst, $src}",
1528 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1529 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1530 "cvtps2dq\t{$src, $dst|$dst, $src}",
1531 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1532 (memop addr:$src)))]>;
1533 // SSE2 packed instructions with XS prefix
1534 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1535 "cvttps2dq\t{$src, $dst|$dst, $src}",
1536 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1537 XS, Requires<[HasSSE2]>;
1538 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1539 "cvttps2dq\t{$src, $dst|$dst, $src}",
1540 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1541 (memop addr:$src)))]>,
1542 XS, Requires<[HasSSE2]>;
1544 // SSE2 packed instructions with XD prefix
1545 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1546 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1547 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1548 XD, Requires<[HasSSE2]>;
1549 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1550 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1551 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1552 (memop addr:$src)))]>,
1553 XD, Requires<[HasSSE2]>;
1555 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1556 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1557 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1558 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1559 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1560 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1561 (memop addr:$src)))]>;
1563 // SSE2 instructions without OpSize prefix
1564 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1565 "cvtps2pd\t{$src, $dst|$dst, $src}",
1566 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1567 TB, Requires<[HasSSE2]>;
1568 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1569 "cvtps2pd\t{$src, $dst|$dst, $src}",
1570 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1571 (load addr:$src)))]>,
1572 TB, Requires<[HasSSE2]>;
1574 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1575 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1576 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1577 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1578 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1579 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1580 (memop addr:$src)))]>;
1582 // Match intrinsics which expect XMM operand(s).
1583 // Aliases for intrinsics
1584 let Constraints = "$src1 = $dst" in {
1585 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1586 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1587 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1588 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1590 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1591 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1592 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1593 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1594 (loadi32 addr:$src2)))]>;
1595 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1596 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1597 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1598 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1600 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1601 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1602 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1603 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1604 (load addr:$src2)))]>;
1605 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1606 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1607 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1608 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1609 VR128:$src2))]>, XS,
1610 Requires<[HasSSE2]>;
1611 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1612 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1613 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1614 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1615 (load addr:$src2)))]>, XS,
1616 Requires<[HasSSE2]>;
1621 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1623 /// In addition, we also have a special variant of the scalar form here to
1624 /// represent the associated intrinsic operation. This form is unlike the
1625 /// plain scalar form, in that it takes an entire vector (instead of a
1626 /// scalar) and leaves the top elements undefined.
1628 /// And, we have a special variant form for a full-vector intrinsic form.
1630 /// These four forms can each have a reg or a mem operand, so there are a
1631 /// total of eight "instructions".
1633 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1637 bit Commutable = 0> {
1638 // Scalar operation, reg.
1639 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1640 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1641 [(set FR64:$dst, (OpNode FR64:$src))]> {
1642 let isCommutable = Commutable;
1645 // Scalar operation, mem.
1646 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1647 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1648 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1650 // Vector operation, reg.
1651 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1652 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1653 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1654 let isCommutable = Commutable;
1657 // Vector operation, mem.
1658 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1659 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1660 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1662 // Intrinsic operation, reg.
1663 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1664 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1665 [(set VR128:$dst, (F64Int VR128:$src))]> {
1666 let isCommutable = Commutable;
1669 // Intrinsic operation, mem.
1670 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1671 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1672 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1674 // Vector intrinsic operation, reg
1675 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1676 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1677 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1678 let isCommutable = Commutable;
1681 // Vector intrinsic operation, mem
1682 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1683 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1684 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1688 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1689 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1691 // There is no f64 version of the reciprocal approximation instructions.
1694 let Constraints = "$src1 = $dst" in {
1695 let isCommutable = 1 in {
1696 def ANDPDrr : PDI<0x54, MRMSrcReg,
1697 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1698 "andpd\t{$src2, $dst|$dst, $src2}",
1700 (and (bc_v2i64 (v2f64 VR128:$src1)),
1701 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1702 def ORPDrr : PDI<0x56, MRMSrcReg,
1703 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1704 "orpd\t{$src2, $dst|$dst, $src2}",
1706 (or (bc_v2i64 (v2f64 VR128:$src1)),
1707 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1708 def XORPDrr : PDI<0x57, MRMSrcReg,
1709 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1710 "xorpd\t{$src2, $dst|$dst, $src2}",
1712 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1713 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1716 def ANDPDrm : PDI<0x54, MRMSrcMem,
1717 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1718 "andpd\t{$src2, $dst|$dst, $src2}",
1720 (and (bc_v2i64 (v2f64 VR128:$src1)),
1721 (memopv2i64 addr:$src2)))]>;
1722 def ORPDrm : PDI<0x56, MRMSrcMem,
1723 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1724 "orpd\t{$src2, $dst|$dst, $src2}",
1726 (or (bc_v2i64 (v2f64 VR128:$src1)),
1727 (memopv2i64 addr:$src2)))]>;
1728 def XORPDrm : PDI<0x57, MRMSrcMem,
1729 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1730 "xorpd\t{$src2, $dst|$dst, $src2}",
1732 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1733 (memopv2i64 addr:$src2)))]>;
1734 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1735 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1736 "andnpd\t{$src2, $dst|$dst, $src2}",
1738 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1739 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1740 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1741 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1742 "andnpd\t{$src2, $dst|$dst, $src2}",
1744 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1745 (memopv2i64 addr:$src2)))]>;
1748 let Constraints = "$src1 = $dst" in {
1749 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1750 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1751 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1752 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1753 VR128:$src, imm:$cc))]>;
1754 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1755 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1756 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1757 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1758 (memop addr:$src), imm:$cc))]>;
1760 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1761 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1762 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1763 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1765 // Shuffle and unpack instructions
1766 let Constraints = "$src1 = $dst" in {
1767 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1768 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1769 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1771 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1772 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1773 (outs VR128:$dst), (ins VR128:$src1,
1774 f128mem:$src2, i8imm:$src3),
1775 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1778 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1780 let AddedComplexity = 10 in {
1781 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1782 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1783 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1785 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1786 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1787 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1788 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1790 (v2f64 (unpckh VR128:$src1,
1791 (memopv2f64 addr:$src2))))]>;
1793 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1794 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1795 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1797 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1798 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1799 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1800 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1802 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1803 } // AddedComplexity
1804 } // Constraints = "$src1 = $dst"
1807 //===---------------------------------------------------------------------===//
1808 // SSE integer instructions
1810 // Move Instructions
1811 let neverHasSideEffects = 1 in
1812 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1813 "movdqa\t{$src, $dst|$dst, $src}", []>;
1814 let canFoldAsLoad = 1, mayLoad = 1 in
1815 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1816 "movdqa\t{$src, $dst|$dst, $src}",
1817 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1819 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1820 "movdqa\t{$src, $dst|$dst, $src}",
1821 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1822 let canFoldAsLoad = 1, mayLoad = 1 in
1823 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1824 "movdqu\t{$src, $dst|$dst, $src}",
1825 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1826 XS, Requires<[HasSSE2]>;
1828 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1829 "movdqu\t{$src, $dst|$dst, $src}",
1830 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1831 XS, Requires<[HasSSE2]>;
1833 // Intrinsic forms of MOVDQU load and store
1834 let canFoldAsLoad = 1 in
1835 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1836 "movdqu\t{$src, $dst|$dst, $src}",
1837 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1838 XS, Requires<[HasSSE2]>;
1839 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1840 "movdqu\t{$src, $dst|$dst, $src}",
1841 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1842 XS, Requires<[HasSSE2]>;
1844 let Constraints = "$src1 = $dst" in {
1846 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1847 bit Commutable = 0> {
1848 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1849 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1850 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1851 let isCommutable = Commutable;
1853 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1854 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1855 [(set VR128:$dst, (IntId VR128:$src1,
1856 (bitconvert (memopv2i64 addr:$src2))))]>;
1859 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1861 Intrinsic IntId, Intrinsic IntId2> {
1862 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1,
1864 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1865 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1866 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1,
1868 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1869 [(set VR128:$dst, (IntId VR128:$src1,
1870 (bitconvert (memopv2i64 addr:$src2))))]>;
1871 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1,
1873 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1874 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1877 /// PDI_binop_rm - Simple SSE2 binary operator.
1878 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1879 ValueType OpVT, bit Commutable = 0> {
1880 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1,
1882 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1883 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1884 let isCommutable = Commutable;
1886 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1,
1888 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1889 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1890 (bitconvert (memopv2i64 addr:$src2)))))]>;
1893 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1895 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1896 /// to collapse (bitconvert VT to VT) into its operand.
1898 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1899 bit Commutable = 0> {
1900 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1901 (ins VR128:$src1, VR128:$src2),
1902 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1903 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1904 let isCommutable = Commutable;
1906 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1907 (ins VR128:$src1, i128mem:$src2),
1908 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1909 [(set VR128:$dst, (OpNode VR128:$src1,
1910 (memopv2i64 addr:$src2)))]>;
1913 } // Constraints = "$src1 = $dst"
1915 // 128-bit Integer Arithmetic
1917 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1918 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1919 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1920 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1922 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1923 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1924 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1925 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1927 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1928 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1929 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1930 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1932 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1933 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1934 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1935 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1937 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1939 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1940 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1941 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1943 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1945 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1946 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1949 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1950 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1951 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1952 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1953 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
1956 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
1957 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
1958 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
1959 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
1960 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
1961 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
1963 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
1964 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
1965 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
1966 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
1967 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
1968 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
1970 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
1971 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
1972 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
1973 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
1975 // 128-bit logical shifts.
1976 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1977 def PSLLDQri : PDIi8<0x73, MRM7r,
1978 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1979 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1980 def PSRLDQri : PDIi8<0x73, MRM3r,
1981 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1982 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1983 // PSRADQri doesn't exist in SSE[1-3].
1986 let Predicates = [HasSSE2] in {
1987 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1988 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1989 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1990 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1991 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
1992 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
1993 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
1994 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
1995 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1996 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1998 // Shift up / down and insert zero's.
1999 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2000 (v2i64 (PSLLDQri VR128:$src, (PSxLDQ_imm imm:$amt)))>;
2001 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2002 (v2i64 (PSRLDQri VR128:$src, (PSxLDQ_imm imm:$amt)))>;
2006 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2007 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2008 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2010 let Constraints = "$src1 = $dst" in {
2011 def PANDNrr : PDI<0xDF, MRMSrcReg,
2012 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2013 "pandn\t{$src2, $dst|$dst, $src2}",
2014 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2017 def PANDNrm : PDI<0xDF, MRMSrcMem,
2018 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2019 "pandn\t{$src2, $dst|$dst, $src2}",
2020 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2021 (memopv2i64 addr:$src2))))]>;
2024 // SSE2 Integer comparison
2025 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2026 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2027 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2028 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2029 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2030 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2032 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2033 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2034 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2035 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2036 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2037 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2038 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2039 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2040 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2041 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2042 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2043 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2045 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2046 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2047 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2048 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2049 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2050 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2051 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2052 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2053 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2054 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2055 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2056 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2059 // Pack instructions
2060 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2061 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2062 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2064 // Shuffle and unpack instructions
2065 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2066 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2067 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2068 [(set VR128:$dst, (v4i32 (pshufd:$src2
2069 VR128:$src1, (undef))))]>;
2070 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2071 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2072 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2073 [(set VR128:$dst, (v4i32 (pshufd:$src2
2074 (bc_v4i32(memopv2i64 addr:$src1)),
2077 // SSE2 with ImmT == Imm8 and XS prefix.
2078 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2079 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2080 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2081 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2083 XS, Requires<[HasSSE2]>;
2084 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2085 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2086 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2087 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2088 (bc_v8i16 (memopv2i64 addr:$src1)),
2090 XS, Requires<[HasSSE2]>;
2092 // SSE2 with ImmT == Imm8 and XD prefix.
2093 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2094 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2095 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2096 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2098 XD, Requires<[HasSSE2]>;
2099 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2100 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2101 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2102 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2103 (bc_v8i16 (memopv2i64 addr:$src1)),
2105 XD, Requires<[HasSSE2]>;
2108 let Constraints = "$src1 = $dst" in {
2109 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2110 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2111 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2113 (v16i8 (unpckl VR128:$src1, VR128:$src2)))]>;
2114 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2115 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2116 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2118 (unpckl VR128:$src1,
2119 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2120 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2121 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2122 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2124 (v8i16 (unpckl VR128:$src1, VR128:$src2)))]>;
2125 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2126 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2127 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2129 (unpckl VR128:$src1,
2130 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2131 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2132 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2133 "punpckldq\t{$src2, $dst|$dst, $src2}",
2135 (v4i32 (unpckl VR128:$src1, VR128:$src2)))]>;
2136 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2137 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2138 "punpckldq\t{$src2, $dst|$dst, $src2}",
2140 (unpckl VR128:$src1,
2141 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2142 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2143 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2144 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2146 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2147 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2148 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2149 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2151 (v2i64 (unpckl VR128:$src1,
2152 (memopv2i64 addr:$src2))))]>;
2154 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2155 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2156 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2158 (v16i8 (unpckh VR128:$src1, VR128:$src2)))]>;
2159 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2160 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2161 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2163 (unpckh VR128:$src1,
2164 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2165 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2166 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2167 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2169 (v8i16 (unpckh VR128:$src1, VR128:$src2)))]>;
2170 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2171 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2172 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2174 (unpckh VR128:$src1,
2175 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2176 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2177 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2178 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2180 (v4i32 (unpckh VR128:$src1, VR128:$src2)))]>;
2181 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2182 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2183 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2185 (unpckh VR128:$src1,
2186 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2187 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2188 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2189 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2191 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2192 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2193 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2194 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2196 (v2i64 (unpckh VR128:$src1,
2197 (memopv2i64 addr:$src2))))]>;
2201 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2202 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2203 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2204 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2206 let Constraints = "$src1 = $dst" in {
2207 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2208 (outs VR128:$dst), (ins VR128:$src1,
2209 GR32:$src2, i32i8imm:$src3),
2210 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2212 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2213 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2214 (outs VR128:$dst), (ins VR128:$src1,
2215 i16mem:$src2, i32i8imm:$src3),
2216 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2218 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2223 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2224 "pmovmskb\t{$src, $dst|$dst, $src}",
2225 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2227 // Conditional store
2229 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2230 "maskmovdqu\t{$mask, $src|$src, $mask}",
2231 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2234 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2235 "maskmovdqu\t{$mask, $src|$src, $mask}",
2236 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2238 // Non-temporal stores
2239 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2240 "movntpd\t{$src, $dst|$dst, $src}",
2241 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2242 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2243 "movntdq\t{$src, $dst|$dst, $src}",
2244 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2245 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2246 "movnti\t{$src, $dst|$dst, $src}",
2247 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2248 TB, Requires<[HasSSE2]>;
2251 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2252 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2253 TB, Requires<[HasSSE2]>;
2255 // Load, store, and memory fence
2256 def LFENCE : I<0xAE, MRM5r, (outs), (ins),
2257 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2258 def MFENCE : I<0xAE, MRM6r, (outs), (ins),
2259 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2261 //TODO: custom lower this so as to never even generate the noop
2262 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2264 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2265 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2266 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2269 // Alias instructions that map zero vector to pxor / xorp* for sse.
2270 // We set canFoldAsLoad because this can be converted to a constant-pool
2271 // load of an all-ones value if folding it would be beneficial.
2272 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2273 isCodeGenOnly = 1 in
2274 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2275 "pcmpeqd\t$dst, $dst",
2276 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2278 // FR64 to 128-bit vector conversion.
2279 let isAsCheapAsAMove = 1 in
2280 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2281 "movsd\t{$src, $dst|$dst, $src}",
2283 (v2f64 (scalar_to_vector FR64:$src)))]>;
2284 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2285 "movsd\t{$src, $dst|$dst, $src}",
2287 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2289 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2290 "movd\t{$src, $dst|$dst, $src}",
2292 (v4i32 (scalar_to_vector GR32:$src)))]>;
2293 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2294 "movd\t{$src, $dst|$dst, $src}",
2296 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2298 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2299 "movd\t{$src, $dst|$dst, $src}",
2300 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2302 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2303 "movd\t{$src, $dst|$dst, $src}",
2304 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2306 // SSE2 instructions with XS prefix
2307 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2308 "movq\t{$src, $dst|$dst, $src}",
2310 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2311 Requires<[HasSSE2]>;
2312 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2313 "movq\t{$src, $dst|$dst, $src}",
2314 [(store (i64 (vector_extract (v2i64 VR128:$src),
2315 (iPTR 0))), addr:$dst)]>;
2317 // FIXME: may not be able to eliminate this movss with coalescing the src and
2318 // dest register classes are different. We really want to write this pattern
2320 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2321 // (f32 FR32:$src)>;
2322 let isAsCheapAsAMove = 1 in
2323 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2324 "movsd\t{$src, $dst|$dst, $src}",
2325 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2327 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2328 "movsd\t{$src, $dst|$dst, $src}",
2329 [(store (f64 (vector_extract (v2f64 VR128:$src),
2330 (iPTR 0))), addr:$dst)]>;
2331 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2332 "movd\t{$src, $dst|$dst, $src}",
2333 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2335 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2336 "movd\t{$src, $dst|$dst, $src}",
2337 [(store (i32 (vector_extract (v4i32 VR128:$src),
2338 (iPTR 0))), addr:$dst)]>;
2340 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2341 "movd\t{$src, $dst|$dst, $src}",
2342 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2343 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2344 "movd\t{$src, $dst|$dst, $src}",
2345 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2348 // Move to lower bits of a VR128, leaving upper bits alone.
2349 // Three operand (but two address) aliases.
2350 let Constraints = "$src1 = $dst" in {
2351 let neverHasSideEffects = 1 in
2352 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2353 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2354 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2356 let AddedComplexity = 15 in
2357 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2358 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2359 "movsd\t{$src2, $dst|$dst, $src2}",
2361 (v2f64 (movl VR128:$src1, VR128:$src2)))]>;
2364 // Store / copy lower 64-bits of a XMM register.
2365 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2366 "movq\t{$src, $dst|$dst, $src}",
2367 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2369 // Move to lower bits of a VR128 and zeroing upper bits.
2370 // Loading from memory automatically zeroing upper bits.
2371 let AddedComplexity = 20 in {
2372 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2373 "movsd\t{$src, $dst|$dst, $src}",
2375 (v2f64 (X86vzmovl (v2f64 (scalar_to_vector
2376 (loadf64 addr:$src))))))]>;
2378 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
2379 (MOVZSD2PDrm addr:$src)>;
2380 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
2381 (MOVZSD2PDrm addr:$src)>;
2382 def : Pat<(v2f64 (X86vzload addr:$src)), (MOVZSD2PDrm addr:$src)>;
2385 // movd / movq to XMM register zero-extends
2386 let AddedComplexity = 15 in {
2387 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2388 "movd\t{$src, $dst|$dst, $src}",
2389 [(set VR128:$dst, (v4i32 (X86vzmovl
2390 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2391 // This is X86-64 only.
2392 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2393 "mov{d|q}\t{$src, $dst|$dst, $src}",
2394 [(set VR128:$dst, (v2i64 (X86vzmovl
2395 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2398 let AddedComplexity = 20 in {
2399 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2400 "movd\t{$src, $dst|$dst, $src}",
2402 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2403 (loadi32 addr:$src))))))]>;
2405 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2406 (MOVZDI2PDIrm addr:$src)>;
2407 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2408 (MOVZDI2PDIrm addr:$src)>;
2409 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2410 (MOVZDI2PDIrm addr:$src)>;
2412 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2413 "movq\t{$src, $dst|$dst, $src}",
2415 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2416 (loadi64 addr:$src))))))]>, XS,
2417 Requires<[HasSSE2]>;
2419 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2420 (MOVZQI2PQIrm addr:$src)>;
2421 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2422 (MOVZQI2PQIrm addr:$src)>;
2423 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2426 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2427 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2428 let AddedComplexity = 15 in
2429 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2430 "movq\t{$src, $dst|$dst, $src}",
2431 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2432 XS, Requires<[HasSSE2]>;
2434 let AddedComplexity = 20 in {
2435 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2436 "movq\t{$src, $dst|$dst, $src}",
2437 [(set VR128:$dst, (v2i64 (X86vzmovl
2438 (loadv2i64 addr:$src))))]>,
2439 XS, Requires<[HasSSE2]>;
2441 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2442 (MOVZPQILo2PQIrm addr:$src)>;
2445 //===---------------------------------------------------------------------===//
2446 // SSE3 Instructions
2447 //===---------------------------------------------------------------------===//
2449 // Move Instructions
2450 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2451 "movshdup\t{$src, $dst|$dst, $src}",
2452 [(set VR128:$dst, (v4f32 (movshdup
2453 VR128:$src, (undef))))]>;
2454 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2455 "movshdup\t{$src, $dst|$dst, $src}",
2456 [(set VR128:$dst, (movshdup
2457 (memopv4f32 addr:$src), (undef)))]>;
2459 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2460 "movsldup\t{$src, $dst|$dst, $src}",
2461 [(set VR128:$dst, (v4f32 (movsldup
2462 VR128:$src, (undef))))]>;
2463 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2464 "movsldup\t{$src, $dst|$dst, $src}",
2465 [(set VR128:$dst, (movsldup
2466 (memopv4f32 addr:$src), (undef)))]>;
2468 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2469 "movddup\t{$src, $dst|$dst, $src}",
2470 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2471 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2472 "movddup\t{$src, $dst|$dst, $src}",
2474 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2477 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2479 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2481 let AddedComplexity = 5 in {
2482 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2483 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2484 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2485 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2486 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2487 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2488 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2489 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2493 let Constraints = "$src1 = $dst" in {
2494 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2495 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2496 "addsubps\t{$src2, $dst|$dst, $src2}",
2497 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2499 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2500 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2501 "addsubps\t{$src2, $dst|$dst, $src2}",
2502 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2503 (memop addr:$src2)))]>;
2504 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2505 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2506 "addsubpd\t{$src2, $dst|$dst, $src2}",
2507 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2509 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2510 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2511 "addsubpd\t{$src2, $dst|$dst, $src2}",
2512 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2513 (memop addr:$src2)))]>;
2516 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2517 "lddqu\t{$src, $dst|$dst, $src}",
2518 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2521 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2522 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2523 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2524 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2525 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2526 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2527 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2528 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2529 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2530 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2531 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2532 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2533 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2534 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2535 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2536 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2538 let Constraints = "$src1 = $dst" in {
2539 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2540 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2541 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2542 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2543 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2544 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2545 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2546 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2549 // Thread synchronization
2550 def MONITOR : I<0x01, MRM1r, (outs), (ins), "monitor",
2551 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2552 def MWAIT : I<0x01, MRM1r, (outs), (ins), "mwait",
2553 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2555 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2556 let AddedComplexity = 15 in
2557 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2558 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2559 let AddedComplexity = 20 in
2560 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2561 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2563 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2564 let AddedComplexity = 15 in
2565 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2566 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2567 let AddedComplexity = 20 in
2568 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2569 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2571 //===---------------------------------------------------------------------===//
2572 // SSSE3 Instructions
2573 //===---------------------------------------------------------------------===//
2575 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2576 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2577 Intrinsic IntId64, Intrinsic IntId128> {
2578 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2579 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2580 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2582 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2583 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2585 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2587 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2589 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2590 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2593 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2595 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2598 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2601 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2602 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2603 Intrinsic IntId64, Intrinsic IntId128> {
2604 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2606 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2607 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2609 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2611 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2614 (bitconvert (memopv4i16 addr:$src))))]>;
2616 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2618 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2619 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2622 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2624 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2627 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2630 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2631 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2632 Intrinsic IntId64, Intrinsic IntId128> {
2633 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2635 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2636 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2638 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2640 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2643 (bitconvert (memopv2i32 addr:$src))))]>;
2645 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2647 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2648 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2651 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2653 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2656 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2659 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2660 int_x86_ssse3_pabs_b,
2661 int_x86_ssse3_pabs_b_128>;
2662 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2663 int_x86_ssse3_pabs_w,
2664 int_x86_ssse3_pabs_w_128>;
2665 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2666 int_x86_ssse3_pabs_d,
2667 int_x86_ssse3_pabs_d_128>;
2669 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2670 let Constraints = "$src1 = $dst" in {
2671 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2672 Intrinsic IntId64, Intrinsic IntId128,
2673 bit Commutable = 0> {
2674 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2675 (ins VR64:$src1, VR64:$src2),
2676 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2677 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2678 let isCommutable = Commutable;
2680 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2681 (ins VR64:$src1, i64mem:$src2),
2682 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2684 (IntId64 VR64:$src1,
2685 (bitconvert (memopv8i8 addr:$src2))))]>;
2687 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2688 (ins VR128:$src1, VR128:$src2),
2689 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2690 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2692 let isCommutable = Commutable;
2694 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2695 (ins VR128:$src1, i128mem:$src2),
2696 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2698 (IntId128 VR128:$src1,
2699 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2703 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2704 let Constraints = "$src1 = $dst" in {
2705 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2706 Intrinsic IntId64, Intrinsic IntId128,
2707 bit Commutable = 0> {
2708 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2709 (ins VR64:$src1, VR64:$src2),
2710 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2711 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2712 let isCommutable = Commutable;
2714 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2715 (ins VR64:$src1, i64mem:$src2),
2716 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2718 (IntId64 VR64:$src1,
2719 (bitconvert (memopv4i16 addr:$src2))))]>;
2721 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2722 (ins VR128:$src1, VR128:$src2),
2723 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2724 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2726 let isCommutable = Commutable;
2728 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2729 (ins VR128:$src1, i128mem:$src2),
2730 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2732 (IntId128 VR128:$src1,
2733 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2737 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2738 let Constraints = "$src1 = $dst" in {
2739 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2740 Intrinsic IntId64, Intrinsic IntId128,
2741 bit Commutable = 0> {
2742 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2743 (ins VR64:$src1, VR64:$src2),
2744 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2745 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2746 let isCommutable = Commutable;
2748 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2749 (ins VR64:$src1, i64mem:$src2),
2750 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2752 (IntId64 VR64:$src1,
2753 (bitconvert (memopv2i32 addr:$src2))))]>;
2755 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2756 (ins VR128:$src1, VR128:$src2),
2757 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2758 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2760 let isCommutable = Commutable;
2762 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2763 (ins VR128:$src1, i128mem:$src2),
2764 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2766 (IntId128 VR128:$src1,
2767 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2771 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2772 int_x86_ssse3_phadd_w,
2773 int_x86_ssse3_phadd_w_128>;
2774 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2775 int_x86_ssse3_phadd_d,
2776 int_x86_ssse3_phadd_d_128>;
2777 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2778 int_x86_ssse3_phadd_sw,
2779 int_x86_ssse3_phadd_sw_128>;
2780 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2781 int_x86_ssse3_phsub_w,
2782 int_x86_ssse3_phsub_w_128>;
2783 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2784 int_x86_ssse3_phsub_d,
2785 int_x86_ssse3_phsub_d_128>;
2786 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2787 int_x86_ssse3_phsub_sw,
2788 int_x86_ssse3_phsub_sw_128>;
2789 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2790 int_x86_ssse3_pmadd_ub_sw,
2791 int_x86_ssse3_pmadd_ub_sw_128>;
2792 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2793 int_x86_ssse3_pmul_hr_sw,
2794 int_x86_ssse3_pmul_hr_sw_128, 1>;
2795 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2796 int_x86_ssse3_pshuf_b,
2797 int_x86_ssse3_pshuf_b_128>;
2798 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2799 int_x86_ssse3_psign_b,
2800 int_x86_ssse3_psign_b_128>;
2801 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2802 int_x86_ssse3_psign_w,
2803 int_x86_ssse3_psign_w_128>;
2804 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2805 int_x86_ssse3_psign_d,
2806 int_x86_ssse3_psign_d_128>;
2808 let Constraints = "$src1 = $dst" in {
2809 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2810 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2811 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2813 (int_x86_ssse3_palign_r
2814 VR64:$src1, VR64:$src2,
2816 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2817 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2818 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2820 (int_x86_ssse3_palign_r
2822 (bitconvert (memopv2i32 addr:$src2)),
2825 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2826 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2827 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2829 (int_x86_ssse3_palign_r_128
2830 VR128:$src1, VR128:$src2,
2831 imm:$src3))]>, OpSize;
2832 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2833 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2834 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2836 (int_x86_ssse3_palign_r_128
2838 (bitconvert (memopv4i32 addr:$src2)),
2839 imm:$src3))]>, OpSize;
2842 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2843 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2844 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2845 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2847 //===---------------------------------------------------------------------===//
2848 // Non-Instruction Patterns
2849 //===---------------------------------------------------------------------===//
2851 // extload f32 -> f64. This matches load+fextend because we have a hack in
2852 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2854 // Since these loads aren't folded into the fextend, we have to match it
2856 let Predicates = [HasSSE2] in
2857 def : Pat<(fextend (loadf32 addr:$src)),
2858 (CVTSS2SDrm addr:$src)>;
2861 let Predicates = [HasSSE2] in {
2862 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2863 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2864 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2865 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2866 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2867 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2868 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2869 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2870 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2871 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2872 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2873 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2874 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2875 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2876 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2877 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2878 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2879 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2880 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2881 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2882 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2883 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2884 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2885 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2886 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2887 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2888 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2889 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2890 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2891 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2894 // Move scalar to XMM zero-extended
2895 // movd to XMM register zero-extends
2896 let AddedComplexity = 15 in {
2897 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2898 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2899 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2900 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2901 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE1]>;
2902 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2903 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2904 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2905 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2908 // Splat v2f64 / v2i64
2909 let AddedComplexity = 10 in {
2910 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
2911 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2912 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
2913 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2914 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
2915 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2916 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
2917 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2920 // Special unary SHUFPSrri case.
2921 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
2922 (SHUFPSrri VR128:$src1, VR128:$src1,
2923 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2924 Requires<[HasSSE1]>;
2925 let AddedComplexity = 5 in
2926 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
2927 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2928 Requires<[HasSSE2]>;
2929 // Special unary SHUFPDrri case.
2930 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
2931 (SHUFPDrri VR128:$src1, VR128:$src1,
2932 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2933 Requires<[HasSSE2]>;
2934 // Special unary SHUFPDrri case.
2935 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
2936 (SHUFPDrri VR128:$src1, VR128:$src1,
2937 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2938 Requires<[HasSSE2]>;
2939 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2940 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
2941 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2942 Requires<[HasSSE2]>;
2944 // Special binary v4i32 shuffle cases with SHUFPS.
2945 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
2946 (SHUFPSrri VR128:$src1, VR128:$src2,
2947 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2948 Requires<[HasSSE2]>;
2949 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
2950 (SHUFPSrmi VR128:$src1, addr:$src2,
2951 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2952 Requires<[HasSSE2]>;
2953 // Special binary v2i64 shuffle cases using SHUFPDrri.
2954 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
2955 (SHUFPDrri VR128:$src1, VR128:$src2,
2956 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2957 Requires<[HasSSE2]>;
2959 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2960 let AddedComplexity = 15 in {
2961 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
2962 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2963 Requires<[OptForSpeed, HasSSE2]>;
2964 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
2965 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2966 Requires<[OptForSpeed, HasSSE2]>;
2968 let AddedComplexity = 10 in {
2969 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
2970 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2971 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
2972 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2973 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
2974 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2975 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
2976 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2979 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2980 let AddedComplexity = 15 in {
2981 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
2982 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2983 Requires<[OptForSpeed, HasSSE2]>;
2984 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
2985 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2986 Requires<[OptForSpeed, HasSSE2]>;
2988 let AddedComplexity = 10 in {
2989 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
2990 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2991 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
2992 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2993 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
2994 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2995 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
2996 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2999 let AddedComplexity = 20 in {
3000 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3001 def : Pat<(v4i32 (movhp VR128:$src1, VR128:$src2)),
3002 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3004 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3005 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3006 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3008 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3009 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3010 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3011 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3012 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3015 let AddedComplexity = 20 in {
3016 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3017 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
3018 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3019 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3020 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3021 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3022 def : Pat<(v4f32 (movhp VR128:$src1, (load addr:$src2))),
3023 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3024 def : Pat<(v2f64 (movhp VR128:$src1, (load addr:$src2))),
3025 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3027 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3028 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3029 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3030 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3031 def : Pat<(v4i32 (movhp VR128:$src1, (load addr:$src2))),
3032 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3033 def : Pat<(v2i64 (movhp VR128:$src1, (load addr:$src2))),
3034 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3037 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3038 // (store (vector_shuffle (load addr), v2, <0, 1, 4, 5>), addr) using MOVHPS
3039 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3040 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3041 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3042 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3043 def : Pat<(store (v4f32 (movhp (load addr:$src1), VR128:$src2)), addr:$src1),
3044 (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3045 def : Pat<(store (v2f64 (movhp (load addr:$src1), VR128:$src2)), addr:$src1),
3046 (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3048 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3050 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3051 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3052 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3053 def : Pat<(store (v4i32 (movhp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3055 (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3056 def : Pat<(store (v2i64 (movhp (load addr:$src1), VR128:$src2)), addr:$src1),
3057 (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3060 let AddedComplexity = 15 in {
3061 // Setting the lowest element in the vector.
3062 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3063 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3064 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3065 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3067 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
3068 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3069 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3070 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3071 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3074 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3075 // fall back to this for SSE1)
3076 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3077 (SHUFPSrri VR128:$src2, VR128:$src1,
3078 (SHUFFLE_get_shuf_imm VR128:$src3))>, Requires<[HasSSE1]>;
3080 // Set lowest element and zero upper elements.
3081 let AddedComplexity = 15 in
3082 def : Pat<(v2f64 (movl immAllZerosV_bc, VR128:$src)),
3083 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3084 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3085 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3087 // Some special case pandn patterns.
3088 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3090 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3091 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3093 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3094 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3096 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3098 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3099 (memop addr:$src2))),
3100 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3101 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3102 (memop addr:$src2))),
3103 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3104 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3105 (memop addr:$src2))),
3106 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3108 // vector -> vector casts
3109 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3110 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3111 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3112 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3113 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3114 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3115 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3116 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3118 // Use movaps / movups for SSE integer load / store (one byte shorter).
3119 def : Pat<(alignedloadv4i32 addr:$src),
3120 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
3121 def : Pat<(loadv4i32 addr:$src),
3122 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3123 def : Pat<(alignedloadv2i64 addr:$src),
3124 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3125 def : Pat<(loadv2i64 addr:$src),
3126 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3128 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3129 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3130 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3131 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3132 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3133 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3134 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3135 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3136 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3137 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3138 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3139 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3140 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3141 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3142 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3143 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3145 //===----------------------------------------------------------------------===//
3146 // SSE4.1 Instructions
3147 //===----------------------------------------------------------------------===//
3149 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3152 Intrinsic V2F64Int> {
3153 // Intrinsic operation, reg.
3154 // Vector intrinsic operation, reg
3155 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3156 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3157 !strconcat(OpcodeStr,
3158 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3159 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3162 // Vector intrinsic operation, mem
3163 def PSm_Int : SS4AIi8<opcps, MRMSrcMem,
3164 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3165 !strconcat(OpcodeStr,
3166 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3168 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3171 // Vector intrinsic operation, reg
3172 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3173 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3174 !strconcat(OpcodeStr,
3175 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3176 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3179 // Vector intrinsic operation, mem
3180 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3181 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3182 !strconcat(OpcodeStr,
3183 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3185 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3189 let Constraints = "$src1 = $dst" in {
3190 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3194 // Intrinsic operation, reg.
3195 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3197 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3198 !strconcat(OpcodeStr,
3199 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3201 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3204 // Intrinsic operation, mem.
3205 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3207 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3208 !strconcat(OpcodeStr,
3209 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3211 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3214 // Intrinsic operation, reg.
3215 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3217 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3218 !strconcat(OpcodeStr,
3219 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3221 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3224 // Intrinsic operation, mem.
3225 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3227 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3228 !strconcat(OpcodeStr,
3229 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3231 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3236 // FP round - roundss, roundps, roundsd, roundpd
3237 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3238 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3239 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3240 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3242 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3243 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3244 Intrinsic IntId128> {
3245 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3247 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3248 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3249 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3251 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3254 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3257 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3258 int_x86_sse41_phminposuw>;
3260 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3261 let Constraints = "$src1 = $dst" in {
3262 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3263 Intrinsic IntId128, bit Commutable = 0> {
3264 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3265 (ins VR128:$src1, VR128:$src2),
3266 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3267 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3269 let isCommutable = Commutable;
3271 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3272 (ins VR128:$src1, i128mem:$src2),
3273 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3275 (IntId128 VR128:$src1,
3276 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3280 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3281 int_x86_sse41_pcmpeqq, 1>;
3282 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3283 int_x86_sse41_packusdw, 0>;
3284 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3285 int_x86_sse41_pminsb, 1>;
3286 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3287 int_x86_sse41_pminsd, 1>;
3288 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3289 int_x86_sse41_pminud, 1>;
3290 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3291 int_x86_sse41_pminuw, 1>;
3292 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3293 int_x86_sse41_pmaxsb, 1>;
3294 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3295 int_x86_sse41_pmaxsd, 1>;
3296 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3297 int_x86_sse41_pmaxud, 1>;
3298 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3299 int_x86_sse41_pmaxuw, 1>;
3301 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3303 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3304 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3305 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3306 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3308 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3309 let Constraints = "$src1 = $dst" in {
3310 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3311 SDNode OpNode, Intrinsic IntId128,
3312 bit Commutable = 0> {
3313 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3314 (ins VR128:$src1, VR128:$src2),
3315 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3316 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3317 VR128:$src2))]>, OpSize {
3318 let isCommutable = Commutable;
3320 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3321 (ins VR128:$src1, VR128:$src2),
3322 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3323 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3325 let isCommutable = Commutable;
3327 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3328 (ins VR128:$src1, i128mem:$src2),
3329 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3331 (OpNode VR128:$src1, (memop addr:$src2)))]>, OpSize;
3332 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3333 (ins VR128:$src1, i128mem:$src2),
3334 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3336 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3340 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", v4i32, mul,
3341 int_x86_sse41_pmulld, 1>;
3343 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3344 let Constraints = "$src1 = $dst" in {
3345 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3346 Intrinsic IntId128, bit Commutable = 0> {
3347 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3348 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3349 !strconcat(OpcodeStr,
3350 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3352 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3354 let isCommutable = Commutable;
3356 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3357 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3358 !strconcat(OpcodeStr,
3359 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3361 (IntId128 VR128:$src1,
3362 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3367 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3368 int_x86_sse41_blendps, 0>;
3369 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3370 int_x86_sse41_blendpd, 0>;
3371 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3372 int_x86_sse41_pblendw, 0>;
3373 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3374 int_x86_sse41_dpps, 1>;
3375 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3376 int_x86_sse41_dppd, 1>;
3377 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3378 int_x86_sse41_mpsadbw, 1>;
3381 /// SS41I_ternary_int - SSE 4.1 ternary operator
3382 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3383 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3384 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3385 (ins VR128:$src1, VR128:$src2),
3386 !strconcat(OpcodeStr,
3387 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3388 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3391 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3392 (ins VR128:$src1, i128mem:$src2),
3393 !strconcat(OpcodeStr,
3394 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3397 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3401 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3402 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3403 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3406 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3407 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3408 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3409 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3411 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3412 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3414 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3418 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3419 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3420 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3421 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3422 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3423 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3425 // Common patterns involving scalar load.
3426 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3427 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3428 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3429 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3431 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3432 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3433 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3434 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3436 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3437 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3438 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3439 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3441 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3442 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3443 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3444 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3446 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3447 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3448 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3449 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3451 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3452 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3453 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3454 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3457 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3458 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3459 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3460 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3462 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3463 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3465 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3469 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3470 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3471 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3472 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3474 // Common patterns involving scalar load
3475 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3476 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3477 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3478 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3480 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3481 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3482 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3483 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3486 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3487 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3488 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3489 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3491 // Expecting a i16 load any extended to i32 value.
3492 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3493 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3494 [(set VR128:$dst, (IntId (bitconvert
3495 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3499 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3500 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3502 // Common patterns involving scalar load
3503 def : Pat<(int_x86_sse41_pmovsxbq
3504 (bitconvert (v4i32 (X86vzmovl
3505 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3506 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3508 def : Pat<(int_x86_sse41_pmovzxbq
3509 (bitconvert (v4i32 (X86vzmovl
3510 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3511 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3514 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3515 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3516 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3517 (ins VR128:$src1, i32i8imm:$src2),
3518 !strconcat(OpcodeStr,
3519 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3520 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3522 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3523 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3524 !strconcat(OpcodeStr,
3525 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3528 // There's an AssertZext in the way of writing the store pattern
3529 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3532 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3535 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3536 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3537 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3538 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3539 !strconcat(OpcodeStr,
3540 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3543 // There's an AssertZext in the way of writing the store pattern
3544 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3547 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3550 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3551 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3552 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3553 (ins VR128:$src1, i32i8imm:$src2),
3554 !strconcat(OpcodeStr,
3555 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3557 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3558 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3559 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3560 !strconcat(OpcodeStr,
3561 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3562 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3563 addr:$dst)]>, OpSize;
3566 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3569 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3571 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3572 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3573 (ins VR128:$src1, i32i8imm:$src2),
3574 !strconcat(OpcodeStr,
3575 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3577 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3579 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3580 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3581 !strconcat(OpcodeStr,
3582 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3583 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3584 addr:$dst)]>, OpSize;
3587 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3589 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3590 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3593 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3594 Requires<[HasSSE41]>;
3596 let Constraints = "$src1 = $dst" in {
3597 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3598 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3599 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3600 !strconcat(OpcodeStr,
3601 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3603 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3604 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3605 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3606 !strconcat(OpcodeStr,
3607 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3609 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3610 imm:$src3))]>, OpSize;
3614 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3616 let Constraints = "$src1 = $dst" in {
3617 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3618 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3619 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3620 !strconcat(OpcodeStr,
3621 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3623 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3625 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3626 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3627 !strconcat(OpcodeStr,
3628 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3630 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3631 imm:$src3)))]>, OpSize;
3635 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3637 // insertps has a few different modes, there's the first two here below which
3638 // are optimized inserts that won't zero arbitrary elements in the destination
3639 // vector. The next one matches the intrinsic and could zero arbitrary elements
3640 // in the target vector.
3641 let Constraints = "$src1 = $dst" in {
3642 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3643 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3644 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3645 !strconcat(OpcodeStr,
3646 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3648 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3650 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3651 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3652 !strconcat(OpcodeStr,
3653 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3655 (X86insrtps VR128:$src1,
3656 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3657 imm:$src3))]>, OpSize;
3661 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3663 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3664 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3666 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3667 // the intel intrinsic that corresponds to this.
3668 let Defs = [EFLAGS] in {
3669 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3670 "ptest \t{$src2, $src1|$src1, $src2}",
3671 [(X86ptest VR128:$src1, VR128:$src2),
3672 (implicit EFLAGS)]>, OpSize;
3673 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3674 "ptest \t{$src2, $src1|$src1, $src2}",
3675 [(X86ptest VR128:$src1, (load addr:$src2)),
3676 (implicit EFLAGS)]>, OpSize;
3679 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3680 "movntdqa\t{$src, $dst|$dst, $src}",
3681 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;
3684 //===----------------------------------------------------------------------===//
3685 // SSE4.2 Instructions
3686 //===----------------------------------------------------------------------===//
3688 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3689 let Constraints = "$src1 = $dst" in {
3690 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3691 Intrinsic IntId128, bit Commutable = 0> {
3692 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3693 (ins VR128:$src1, VR128:$src2),
3694 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3695 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3697 let isCommutable = Commutable;
3699 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3700 (ins VR128:$src1, i128mem:$src2),
3701 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3703 (IntId128 VR128:$src1,
3704 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3708 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3710 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3711 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3712 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3713 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3715 // crc intrinsic instruction
3716 // This set of instructions are only rm, the only difference is the size
3718 let Constraints = "$src1 = $dst" in {
3719 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3720 (ins GR32:$src1, i8mem:$src2),
3721 "crc32 \t{$src2, $src1|$src1, $src2}",
3723 (int_x86_sse42_crc32_8 GR32:$src1,
3724 (load addr:$src2)))]>, OpSize;
3725 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3726 (ins GR32:$src1, GR8:$src2),
3727 "crc32 \t{$src2, $src1|$src1, $src2}",
3729 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>,
3731 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3732 (ins GR32:$src1, i16mem:$src2),
3733 "crc32 \t{$src2, $src1|$src1, $src2}",
3735 (int_x86_sse42_crc32_16 GR32:$src1,
3736 (load addr:$src2)))]>,
3738 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3739 (ins GR32:$src1, GR16:$src2),
3740 "crc32 \t{$src2, $src1|$src1, $src2}",
3742 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3744 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3745 (ins GR32:$src1, i32mem:$src2),
3746 "crc32 \t{$src2, $src1|$src1, $src2}",
3748 (int_x86_sse42_crc32_32 GR32:$src1,
3749 (load addr:$src2)))]>, OpSize;
3750 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3751 (ins GR32:$src1, GR32:$src2),
3752 "crc32 \t{$src2, $src1|$src1, $src2}",
3754 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>,
3756 def CRC64m64 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3757 (ins GR64:$src1, i64mem:$src2),
3758 "crc32 \t{$src2, $src1|$src1, $src2}",
3760 (int_x86_sse42_crc32_64 GR64:$src1,
3761 (load addr:$src2)))]>,
3763 def CRC64r64 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3764 (ins GR64:$src1, GR64:$src2),
3765 "crc32 \t{$src2, $src1|$src1, $src2}",
3767 (int_x86_sse42_crc32_64 GR64:$src1, GR64:$src2))]>,
3771 // String/text processing instructions.
3772 let Defs = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
3773 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3774 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3775 "#PCMPISTRM128rr PSEUDO!",
3777 (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3778 imm:$src3))]>, OpSize;
3779 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3780 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3781 "#PCMPISTRM128rm PSEUDO!",
3783 (int_x86_sse42_pcmpistrm128 VR128:$src1,
3785 imm:$src3))]>, OpSize;
3788 let Defs = [XMM0, EFLAGS] in {
3789 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3790 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3791 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3793 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3794 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3795 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3799 let Defs = [EFLAGS], Uses = [EAX, EDX],
3800 usesCustomDAGSchedInserter = 1 in {
3801 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3802 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3803 "#PCMPESTRM128rr PSEUDO!",
3805 (int_x86_sse42_pcmpestrm128 VR128:$src1, EAX,
3807 EDX, imm:$src5))]>, OpSize;
3808 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3809 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3810 "#PCMPESTRM128rm PSEUDO!",
3812 (int_x86_sse42_pcmpestrm128 VR128:$src1, EAX,
3814 EDX, imm:$src5))]>, OpSize;
3817 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3818 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3819 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3820 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3822 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3823 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3824 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3828 let Defs = [ECX, EFLAGS] in {
3829 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3830 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3831 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3832 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3834 (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3835 (implicit EFLAGS)]>,
3837 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3838 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3839 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3841 (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3842 (implicit EFLAGS)]>,
3847 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3848 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3849 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3850 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3851 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3852 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3854 let Defs = [ECX, EFLAGS] in {
3855 let Uses = [EAX, EDX] in {
3856 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3857 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3858 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3859 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3861 (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3862 (implicit EFLAGS)]>,
3864 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3865 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3866 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3868 (IntId128 VR128:$src1, EAX, (load addr:$src3),
3870 (implicit EFLAGS)]>,
3876 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3877 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3878 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3879 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3880 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3881 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;