1 //====- X86InstrSSE.td - Describe the X86 Instruction Set --*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the X86 SSE instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
17 //===----------------------------------------------------------------------===//
18 // SSE specific DAG Nodes.
19 //===----------------------------------------------------------------------===//
21 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
22 SDTCisFP<0>, SDTCisInt<2> ]>;
23 def SDTX86VFCMP : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>,
24 SDTCisFP<1>, SDTCisVT<3, i8>]>;
26 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
27 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
28 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
29 [SDNPCommutative, SDNPAssociative]>;
30 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
31 [SDNPCommutative, SDNPAssociative]>;
32 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
33 [SDNPCommutative, SDNPAssociative]>;
34 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
35 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
36 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
37 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>;
38 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>;
39 def X86pshufb : SDNode<"X86ISD::PSHUFB",
40 SDTypeProfile<1, 2, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
42 def X86pextrb : SDNode<"X86ISD::PEXTRB",
43 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
44 def X86pextrw : SDNode<"X86ISD::PEXTRW",
45 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
46 def X86pinsrb : SDNode<"X86ISD::PINSRB",
47 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
48 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
49 def X86pinsrw : SDNode<"X86ISD::PINSRW",
50 SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
51 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
52 def X86insrtps : SDNode<"X86ISD::INSERTPS",
53 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
54 SDTCisVT<2, v4f32>, SDTCisPtrTy<3>]>>;
55 def X86vzmovl : SDNode<"X86ISD::VZEXT_MOVL",
56 SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>;
57 def X86vzload : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
58 [SDNPHasChain, SDNPMayLoad]>;
59 def X86vshl : SDNode<"X86ISD::VSHL", SDTIntShiftOp>;
60 def X86vshr : SDNode<"X86ISD::VSRL", SDTIntShiftOp>;
61 def X86cmpps : SDNode<"X86ISD::CMPPS", SDTX86VFCMP>;
62 def X86cmppd : SDNode<"X86ISD::CMPPD", SDTX86VFCMP>;
63 def X86pcmpeqb : SDNode<"X86ISD::PCMPEQB", SDTIntBinOp, [SDNPCommutative]>;
64 def X86pcmpeqw : SDNode<"X86ISD::PCMPEQW", SDTIntBinOp, [SDNPCommutative]>;
65 def X86pcmpeqd : SDNode<"X86ISD::PCMPEQD", SDTIntBinOp, [SDNPCommutative]>;
66 def X86pcmpeqq : SDNode<"X86ISD::PCMPEQQ", SDTIntBinOp, [SDNPCommutative]>;
67 def X86pcmpgtb : SDNode<"X86ISD::PCMPGTB", SDTIntBinOp>;
68 def X86pcmpgtw : SDNode<"X86ISD::PCMPGTW", SDTIntBinOp>;
69 def X86pcmpgtd : SDNode<"X86ISD::PCMPGTD", SDTIntBinOp>;
70 def X86pcmpgtq : SDNode<"X86ISD::PCMPGTQ", SDTIntBinOp>;
72 def SDTX86CmpPTest : SDTypeProfile<0, 2, [SDTCisVT<0, v4f32>,
74 def X86ptest : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
76 //===----------------------------------------------------------------------===//
77 // SSE Complex Patterns
78 //===----------------------------------------------------------------------===//
80 // These are 'extloads' from a scalar to the low element of a vector, zeroing
81 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
83 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
84 [SDNPHasChain, SDNPMayLoad]>;
85 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
86 [SDNPHasChain, SDNPMayLoad]>;
88 def ssmem : Operand<v4f32> {
89 let PrintMethod = "printf32mem";
90 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
91 let ParserMatchClass = X86MemAsmOperand;
93 def sdmem : Operand<v2f64> {
94 let PrintMethod = "printf64mem";
95 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
96 let ParserMatchClass = X86MemAsmOperand;
99 //===----------------------------------------------------------------------===//
100 // SSE pattern fragments
101 //===----------------------------------------------------------------------===//
103 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
104 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
105 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
106 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
108 // Like 'store', but always requires vector alignment.
109 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
110 (store node:$val, node:$ptr), [{
111 return cast<StoreSDNode>(N)->getAlignment() >= 16;
114 // Like 'load', but always requires vector alignment.
115 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
116 return cast<LoadSDNode>(N)->getAlignment() >= 16;
119 def alignedloadfsf32 : PatFrag<(ops node:$ptr),
120 (f32 (alignedload node:$ptr))>;
121 def alignedloadfsf64 : PatFrag<(ops node:$ptr),
122 (f64 (alignedload node:$ptr))>;
123 def alignedloadv4f32 : PatFrag<(ops node:$ptr),
124 (v4f32 (alignedload node:$ptr))>;
125 def alignedloadv2f64 : PatFrag<(ops node:$ptr),
126 (v2f64 (alignedload node:$ptr))>;
127 def alignedloadv4i32 : PatFrag<(ops node:$ptr),
128 (v4i32 (alignedload node:$ptr))>;
129 def alignedloadv2i64 : PatFrag<(ops node:$ptr),
130 (v2i64 (alignedload node:$ptr))>;
132 // Like 'load', but uses special alignment checks suitable for use in
133 // memory operands in most SSE instructions, which are required to
134 // be naturally aligned on some targets but not on others. If the subtarget
135 // allows unaligned accesses, match any load, though this may require
136 // setting a feature bit in the processor (on startup, for example).
137 // Opteron 10h and later implement such a feature.
138 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
139 return Subtarget->hasVectorUAMem()
140 || cast<LoadSDNode>(N)->getAlignment() >= 16;
143 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
144 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
145 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
146 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
147 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
148 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
149 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
151 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
153 // FIXME: 8 byte alignment for mmx reads is not required
154 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
155 return cast<LoadSDNode>(N)->getAlignment() >= 8;
158 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
159 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
160 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
161 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
163 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
164 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
165 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
166 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
167 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
168 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
170 def vzmovl_v2i64 : PatFrag<(ops node:$src),
171 (bitconvert (v2i64 (X86vzmovl
172 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
173 def vzmovl_v4i32 : PatFrag<(ops node:$src),
174 (bitconvert (v4i32 (X86vzmovl
175 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
177 def vzload_v2i64 : PatFrag<(ops node:$src),
178 (bitconvert (v2i64 (X86vzload node:$src)))>;
181 def fp32imm0 : PatLeaf<(f32 fpimm), [{
182 return N->isExactlyValue(+0.0);
185 // BYTE_imm - Transform bit immediates into byte immediates.
186 def BYTE_imm : SDNodeXForm<imm, [{
187 // Transformation function: imm >> 3
188 return getI32Imm(N->getZExtValue() >> 3);
191 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
193 def SHUFFLE_get_shuf_imm : SDNodeXForm<vector_shuffle, [{
194 return getI8Imm(X86::getShuffleSHUFImmediate(N));
197 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
199 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<vector_shuffle, [{
200 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
203 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
205 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<vector_shuffle, [{
206 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
209 // SHUFFLE_get_palign_imm xform function: convert vector_shuffle mask to
211 def SHUFFLE_get_palign_imm : SDNodeXForm<vector_shuffle, [{
212 return getI8Imm(X86::getShufflePALIGNRImmediate(N));
215 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
216 (vector_shuffle node:$lhs, node:$rhs), [{
217 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
218 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
221 def movddup : PatFrag<(ops node:$lhs, node:$rhs),
222 (vector_shuffle node:$lhs, node:$rhs), [{
223 return X86::isMOVDDUPMask(cast<ShuffleVectorSDNode>(N));
226 def movhlps : PatFrag<(ops node:$lhs, node:$rhs),
227 (vector_shuffle node:$lhs, node:$rhs), [{
228 return X86::isMOVHLPSMask(cast<ShuffleVectorSDNode>(N));
231 def movhlps_undef : PatFrag<(ops node:$lhs, node:$rhs),
232 (vector_shuffle node:$lhs, node:$rhs), [{
233 return X86::isMOVHLPS_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
236 def movlhps : PatFrag<(ops node:$lhs, node:$rhs),
237 (vector_shuffle node:$lhs, node:$rhs), [{
238 return X86::isMOVLHPSMask(cast<ShuffleVectorSDNode>(N));
241 def movlp : PatFrag<(ops node:$lhs, node:$rhs),
242 (vector_shuffle node:$lhs, node:$rhs), [{
243 return X86::isMOVLPMask(cast<ShuffleVectorSDNode>(N));
246 def movl : PatFrag<(ops node:$lhs, node:$rhs),
247 (vector_shuffle node:$lhs, node:$rhs), [{
248 return X86::isMOVLMask(cast<ShuffleVectorSDNode>(N));
251 def movshdup : PatFrag<(ops node:$lhs, node:$rhs),
252 (vector_shuffle node:$lhs, node:$rhs), [{
253 return X86::isMOVSHDUPMask(cast<ShuffleVectorSDNode>(N));
256 def movsldup : PatFrag<(ops node:$lhs, node:$rhs),
257 (vector_shuffle node:$lhs, node:$rhs), [{
258 return X86::isMOVSLDUPMask(cast<ShuffleVectorSDNode>(N));
261 def unpckl : PatFrag<(ops node:$lhs, node:$rhs),
262 (vector_shuffle node:$lhs, node:$rhs), [{
263 return X86::isUNPCKLMask(cast<ShuffleVectorSDNode>(N));
266 def unpckh : PatFrag<(ops node:$lhs, node:$rhs),
267 (vector_shuffle node:$lhs, node:$rhs), [{
268 return X86::isUNPCKHMask(cast<ShuffleVectorSDNode>(N));
271 def unpckl_undef : PatFrag<(ops node:$lhs, node:$rhs),
272 (vector_shuffle node:$lhs, node:$rhs), [{
273 return X86::isUNPCKL_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
276 def unpckh_undef : PatFrag<(ops node:$lhs, node:$rhs),
277 (vector_shuffle node:$lhs, node:$rhs), [{
278 return X86::isUNPCKH_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
281 def pshufd : PatFrag<(ops node:$lhs, node:$rhs),
282 (vector_shuffle node:$lhs, node:$rhs), [{
283 return X86::isPSHUFDMask(cast<ShuffleVectorSDNode>(N));
284 }], SHUFFLE_get_shuf_imm>;
286 def shufp : PatFrag<(ops node:$lhs, node:$rhs),
287 (vector_shuffle node:$lhs, node:$rhs), [{
288 return X86::isSHUFPMask(cast<ShuffleVectorSDNode>(N));
289 }], SHUFFLE_get_shuf_imm>;
291 def pshufhw : PatFrag<(ops node:$lhs, node:$rhs),
292 (vector_shuffle node:$lhs, node:$rhs), [{
293 return X86::isPSHUFHWMask(cast<ShuffleVectorSDNode>(N));
294 }], SHUFFLE_get_pshufhw_imm>;
296 def pshuflw : PatFrag<(ops node:$lhs, node:$rhs),
297 (vector_shuffle node:$lhs, node:$rhs), [{
298 return X86::isPSHUFLWMask(cast<ShuffleVectorSDNode>(N));
299 }], SHUFFLE_get_pshuflw_imm>;
301 def palign : PatFrag<(ops node:$lhs, node:$rhs),
302 (vector_shuffle node:$lhs, node:$rhs), [{
303 return X86::isPALIGNRMask(cast<ShuffleVectorSDNode>(N));
304 }], SHUFFLE_get_palign_imm>;
306 //===----------------------------------------------------------------------===//
307 // SSE scalar FP Instructions
308 //===----------------------------------------------------------------------===//
310 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded after
311 // instruction selection into a branch sequence.
312 let Uses = [EFLAGS], usesCustomInserter = 1 in {
313 def CMOV_FR32 : I<0, Pseudo,
314 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
315 "#CMOV_FR32 PSEUDO!",
316 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
318 def CMOV_FR64 : I<0, Pseudo,
319 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
320 "#CMOV_FR64 PSEUDO!",
321 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
323 def CMOV_V4F32 : I<0, Pseudo,
324 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
325 "#CMOV_V4F32 PSEUDO!",
327 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
329 def CMOV_V2F64 : I<0, Pseudo,
330 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
331 "#CMOV_V2F64 PSEUDO!",
333 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
335 def CMOV_V2I64 : I<0, Pseudo,
336 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
337 "#CMOV_V2I64 PSEUDO!",
339 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
343 //===----------------------------------------------------------------------===//
345 //===----------------------------------------------------------------------===//
348 let neverHasSideEffects = 1 in
349 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
350 "movss\t{$src, $dst|$dst, $src}", []>;
351 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
352 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
353 "movss\t{$src, $dst|$dst, $src}",
354 [(set FR32:$dst, (loadf32 addr:$src))]>;
355 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
356 "movss\t{$src, $dst|$dst, $src}",
357 [(store FR32:$src, addr:$dst)]>;
359 // Conversion instructions
360 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
361 "cvttss2si\t{$src, $dst|$dst, $src}",
362 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
363 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
364 "cvttss2si\t{$src, $dst|$dst, $src}",
365 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
366 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
367 "cvtsi2ss\t{$src, $dst|$dst, $src}",
368 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
369 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
370 "cvtsi2ss\t{$src, $dst|$dst, $src}",
371 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
373 // Match intrinsics which expect XMM operand(s).
374 def CVTSS2SIrr: SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
375 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
376 def CVTSS2SIrm: SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
377 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
379 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
380 "cvtss2si\t{$src, $dst|$dst, $src}",
381 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
382 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
383 "cvtss2si\t{$src, $dst|$dst, $src}",
384 [(set GR32:$dst, (int_x86_sse_cvtss2si
385 (load addr:$src)))]>;
387 // Match intrinisics which expect MM and XMM operand(s).
388 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
389 "cvtps2pi\t{$src, $dst|$dst, $src}",
390 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
391 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
392 "cvtps2pi\t{$src, $dst|$dst, $src}",
393 [(set VR64:$dst, (int_x86_sse_cvtps2pi
394 (load addr:$src)))]>;
395 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
396 "cvttps2pi\t{$src, $dst|$dst, $src}",
397 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
398 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
399 "cvttps2pi\t{$src, $dst|$dst, $src}",
400 [(set VR64:$dst, (int_x86_sse_cvttps2pi
401 (load addr:$src)))]>;
402 let Constraints = "$src1 = $dst" in {
403 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
404 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
405 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
406 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
408 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
409 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
410 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
411 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
412 (load addr:$src2)))]>;
415 // Aliases for intrinsics
416 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
417 "cvttss2si\t{$src, $dst|$dst, $src}",
419 (int_x86_sse_cvttss2si VR128:$src))]>;
420 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
421 "cvttss2si\t{$src, $dst|$dst, $src}",
423 (int_x86_sse_cvttss2si(load addr:$src)))]>;
425 let Constraints = "$src1 = $dst" in {
426 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
427 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
428 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
429 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
431 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
432 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
433 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
434 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
435 (loadi32 addr:$src2)))]>;
438 // Comparison instructions
439 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
440 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
441 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
442 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
444 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
445 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
446 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
449 let Defs = [EFLAGS] in {
450 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
451 "ucomiss\t{$src2, $src1|$src1, $src2}",
452 [(X86cmp FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
453 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
454 "ucomiss\t{$src2, $src1|$src1, $src2}",
455 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
458 def COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
459 "comiss\t{$src2, $src1|$src1, $src2}", []>;
460 def COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
461 "comiss\t{$src2, $src1|$src1, $src2}", []>;
465 // Aliases to match intrinsics which expect XMM operand(s).
466 let Constraints = "$src1 = $dst" in {
467 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
469 (ins VR128:$src1, VR128:$src, SSECC:$cc),
470 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
471 [(set VR128:$dst, (int_x86_sse_cmp_ss
473 VR128:$src, imm:$cc))]>;
474 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
476 (ins VR128:$src1, f32mem:$src, SSECC:$cc),
477 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
478 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
479 (load addr:$src), imm:$cc))]>;
482 let Defs = [EFLAGS] in {
483 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
484 "ucomiss\t{$src2, $src1|$src1, $src2}",
485 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
487 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
488 "ucomiss\t{$src2, $src1|$src1, $src2}",
489 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
492 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
493 "comiss\t{$src2, $src1|$src1, $src2}",
494 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
496 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
497 "comiss\t{$src2, $src1|$src1, $src2}",
498 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
502 // Aliases of packed SSE1 instructions for scalar use. These all have names
503 // that start with 'Fs'.
505 // Alias instructions that map fld0 to pxor for sse.
506 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
508 // FIXME: Set encoding to pseudo!
509 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
510 [(set FR32:$dst, fp32imm0)]>,
511 Requires<[HasSSE1]>, TB, OpSize;
513 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
515 let neverHasSideEffects = 1 in
516 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
517 "movaps\t{$src, $dst|$dst, $src}", []>;
519 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
521 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
522 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
523 "movaps\t{$src, $dst|$dst, $src}",
524 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
526 // Alias bitwise logical operations using SSE logical ops on packed FP values.
527 let Constraints = "$src1 = $dst" in {
528 let isCommutable = 1 in {
529 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst),
530 (ins FR32:$src1, FR32:$src2),
531 "andps\t{$src2, $dst|$dst, $src2}",
532 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
533 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst),
534 (ins FR32:$src1, FR32:$src2),
535 "orps\t{$src2, $dst|$dst, $src2}",
536 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
537 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst),
538 (ins FR32:$src1, FR32:$src2),
539 "xorps\t{$src2, $dst|$dst, $src2}",
540 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
543 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst),
544 (ins FR32:$src1, f128mem:$src2),
545 "andps\t{$src2, $dst|$dst, $src2}",
546 [(set FR32:$dst, (X86fand FR32:$src1,
547 (memopfsf32 addr:$src2)))]>;
548 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst),
549 (ins FR32:$src1, f128mem:$src2),
550 "orps\t{$src2, $dst|$dst, $src2}",
551 [(set FR32:$dst, (X86for FR32:$src1,
552 (memopfsf32 addr:$src2)))]>;
553 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst),
554 (ins FR32:$src1, f128mem:$src2),
555 "xorps\t{$src2, $dst|$dst, $src2}",
556 [(set FR32:$dst, (X86fxor FR32:$src1,
557 (memopfsf32 addr:$src2)))]>;
559 let neverHasSideEffects = 1 in {
560 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
561 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
562 "andnps\t{$src2, $dst|$dst, $src2}", []>;
564 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
565 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
566 "andnps\t{$src2, $dst|$dst, $src2}", []>;
570 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
572 /// In addition, we also have a special variant of the scalar form here to
573 /// represent the associated intrinsic operation. This form is unlike the
574 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
575 /// and leaves the top elements unmodified (therefore these cannot be commuted).
577 /// These three forms can each be reg+reg or reg+mem, so there are a total of
578 /// six "instructions".
580 let Constraints = "$src1 = $dst" in {
581 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
582 SDNode OpNode, Intrinsic F32Int,
583 bit Commutable = 0> {
584 // Scalar operation, reg+reg.
585 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
586 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
587 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
588 let isCommutable = Commutable;
591 // Scalar operation, reg+mem.
592 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
593 (ins FR32:$src1, f32mem:$src2),
594 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
595 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
597 // Vector operation, reg+reg.
598 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
599 (ins VR128:$src1, VR128:$src2),
600 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
601 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
602 let isCommutable = Commutable;
605 // Vector operation, reg+mem.
606 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
607 (ins VR128:$src1, f128mem:$src2),
608 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
609 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
611 // Intrinsic operation, reg+reg.
612 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
613 (ins VR128:$src1, VR128:$src2),
614 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
615 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]>;
617 // Intrinsic operation, reg+mem.
618 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
619 (ins VR128:$src1, ssmem:$src2),
620 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
621 [(set VR128:$dst, (F32Int VR128:$src1,
622 sse_load_f32:$src2))]>;
626 // Arithmetic instructions
627 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
628 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
629 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
630 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
632 /// sse1_fp_binop_rm - Other SSE1 binops
634 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
635 /// instructions for a full-vector intrinsic form. Operations that map
636 /// onto C operators don't use this form since they just use the plain
637 /// vector form instead of having a separate vector intrinsic form.
639 /// This provides a total of eight "instructions".
641 let Constraints = "$src1 = $dst" in {
642 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
646 bit Commutable = 0> {
648 // Scalar operation, reg+reg.
649 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
650 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
651 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
652 let isCommutable = Commutable;
655 // Scalar operation, reg+mem.
656 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
657 (ins FR32:$src1, f32mem:$src2),
658 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
659 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
661 // Vector operation, reg+reg.
662 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
663 (ins VR128:$src1, VR128:$src2),
664 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
665 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
666 let isCommutable = Commutable;
669 // Vector operation, reg+mem.
670 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
671 (ins VR128:$src1, f128mem:$src2),
672 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
673 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
675 // Intrinsic operation, reg+reg.
676 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
677 (ins VR128:$src1, VR128:$src2),
678 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
679 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
680 let isCommutable = Commutable;
683 // Intrinsic operation, reg+mem.
684 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
685 (ins VR128:$src1, ssmem:$src2),
686 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
687 [(set VR128:$dst, (F32Int VR128:$src1,
688 sse_load_f32:$src2))]>;
690 // Vector intrinsic operation, reg+reg.
691 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
692 (ins VR128:$src1, VR128:$src2),
693 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
694 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
695 let isCommutable = Commutable;
698 // Vector intrinsic operation, reg+mem.
699 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
700 (ins VR128:$src1, f128mem:$src2),
701 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
702 [(set VR128:$dst, (V4F32Int VR128:$src1, (memopv4f32 addr:$src2)))]>;
706 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
707 int_x86_sse_max_ss, int_x86_sse_max_ps>;
708 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
709 int_x86_sse_min_ss, int_x86_sse_min_ps>;
711 //===----------------------------------------------------------------------===//
712 // SSE packed FP Instructions
715 let neverHasSideEffects = 1 in
716 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
717 "movaps\t{$src, $dst|$dst, $src}", []>;
718 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
719 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
720 "movaps\t{$src, $dst|$dst, $src}",
721 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
723 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
724 "movaps\t{$src, $dst|$dst, $src}",
725 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
727 let neverHasSideEffects = 1 in
728 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
729 "movups\t{$src, $dst|$dst, $src}", []>;
730 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
731 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
732 "movups\t{$src, $dst|$dst, $src}",
733 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
734 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
735 "movups\t{$src, $dst|$dst, $src}",
736 [(store (v4f32 VR128:$src), addr:$dst)]>;
738 // Intrinsic forms of MOVUPS load and store
739 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
740 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
741 "movups\t{$src, $dst|$dst, $src}",
742 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
743 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
744 "movups\t{$src, $dst|$dst, $src}",
745 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
747 let Constraints = "$src1 = $dst" in {
748 let AddedComplexity = 20 in {
749 def MOVLPSrm : PSI<0x12, MRMSrcMem,
750 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
751 "movlps\t{$src2, $dst|$dst, $src2}",
754 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
755 def MOVHPSrm : PSI<0x16, MRMSrcMem,
756 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
757 "movhps\t{$src2, $dst|$dst, $src2}",
759 (movlhps VR128:$src1,
760 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
762 } // Constraints = "$src1 = $dst"
765 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
766 "movlps\t{$src, $dst|$dst, $src}",
767 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
768 (iPTR 0))), addr:$dst)]>;
770 // v2f64 extract element 1 is always custom lowered to unpack high to low
771 // and extract element 0 so the non-store version isn't too horrible.
772 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
773 "movhps\t{$src, $dst|$dst, $src}",
774 [(store (f64 (vector_extract
775 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
776 (undef)), (iPTR 0))), addr:$dst)]>;
778 let Constraints = "$src1 = $dst" in {
779 let AddedComplexity = 20 in {
780 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
781 (ins VR128:$src1, VR128:$src2),
782 "movlhps\t{$src2, $dst|$dst, $src2}",
784 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
786 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
787 (ins VR128:$src1, VR128:$src2),
788 "movhlps\t{$src2, $dst|$dst, $src2}",
790 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
792 } // Constraints = "$src1 = $dst"
794 let AddedComplexity = 20 in {
795 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
796 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
797 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
798 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
805 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
807 /// In addition, we also have a special variant of the scalar form here to
808 /// represent the associated intrinsic operation. This form is unlike the
809 /// plain scalar form, in that it takes an entire vector (instead of a
810 /// scalar) and leaves the top elements undefined.
812 /// And, we have a special variant form for a full-vector intrinsic form.
814 /// These four forms can each have a reg or a mem operand, so there are a
815 /// total of eight "instructions".
817 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
821 bit Commutable = 0> {
822 // Scalar operation, reg.
823 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
824 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
825 [(set FR32:$dst, (OpNode FR32:$src))]> {
826 let isCommutable = Commutable;
829 // Scalar operation, mem.
830 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
831 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
832 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
833 Requires<[HasSSE1, OptForSize]>;
835 // Vector operation, reg.
836 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
837 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
838 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
839 let isCommutable = Commutable;
842 // Vector operation, mem.
843 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
844 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
845 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
847 // Intrinsic operation, reg.
848 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
849 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
850 [(set VR128:$dst, (F32Int VR128:$src))]> {
851 let isCommutable = Commutable;
854 // Intrinsic operation, mem.
855 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
856 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
857 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
859 // Vector intrinsic operation, reg
860 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
861 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
862 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
863 let isCommutable = Commutable;
866 // Vector intrinsic operation, mem
867 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
868 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
869 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
873 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
874 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
876 // Reciprocal approximations. Note that these typically require refinement
877 // in order to obtain suitable precision.
878 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
879 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
880 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
881 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
884 let Constraints = "$src1 = $dst" in {
885 let isCommutable = 1 in {
886 def ANDPSrr : PSI<0x54, MRMSrcReg,
887 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
888 "andps\t{$src2, $dst|$dst, $src2}",
889 [(set VR128:$dst, (v2i64
890 (and VR128:$src1, VR128:$src2)))]>;
891 def ORPSrr : PSI<0x56, MRMSrcReg,
892 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
893 "orps\t{$src2, $dst|$dst, $src2}",
894 [(set VR128:$dst, (v2i64
895 (or VR128:$src1, VR128:$src2)))]>;
896 def XORPSrr : PSI<0x57, MRMSrcReg,
897 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
898 "xorps\t{$src2, $dst|$dst, $src2}",
899 [(set VR128:$dst, (v2i64
900 (xor VR128:$src1, VR128:$src2)))]>;
903 def ANDPSrm : PSI<0x54, MRMSrcMem,
904 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
905 "andps\t{$src2, $dst|$dst, $src2}",
906 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
907 (memopv2i64 addr:$src2)))]>;
908 def ORPSrm : PSI<0x56, MRMSrcMem,
909 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
910 "orps\t{$src2, $dst|$dst, $src2}",
911 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
912 (memopv2i64 addr:$src2)))]>;
913 def XORPSrm : PSI<0x57, MRMSrcMem,
914 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
915 "xorps\t{$src2, $dst|$dst, $src2}",
916 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
917 (memopv2i64 addr:$src2)))]>;
918 def ANDNPSrr : PSI<0x55, MRMSrcReg,
919 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
920 "andnps\t{$src2, $dst|$dst, $src2}",
922 (v2i64 (and (xor VR128:$src1,
923 (bc_v2i64 (v4i32 immAllOnesV))),
925 def ANDNPSrm : PSI<0x55, MRMSrcMem,
926 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
927 "andnps\t{$src2, $dst|$dst, $src2}",
929 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
930 (bc_v2i64 (v4i32 immAllOnesV))),
931 (memopv2i64 addr:$src2))))]>;
934 let Constraints = "$src1 = $dst" in {
935 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
936 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
937 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
938 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
939 VR128:$src, imm:$cc))]>;
940 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
941 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
942 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
943 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
944 (memop addr:$src), imm:$cc))]>;
946 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
947 (CMPPSrri VR128:$src1, VR128:$src2, imm:$cc)>;
948 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
949 (CMPPSrmi VR128:$src1, addr:$src2, imm:$cc)>;
951 // Shuffle and unpack instructions
952 let Constraints = "$src1 = $dst" in {
953 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
954 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
955 (outs VR128:$dst), (ins VR128:$src1,
956 VR128:$src2, i8imm:$src3),
957 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
959 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
960 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
961 (outs VR128:$dst), (ins VR128:$src1,
962 f128mem:$src2, i8imm:$src3),
963 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
966 VR128:$src1, (memopv4f32 addr:$src2))))]>;
968 let AddedComplexity = 10 in {
969 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
970 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
971 "unpckhps\t{$src2, $dst|$dst, $src2}",
973 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
974 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
975 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
976 "unpckhps\t{$src2, $dst|$dst, $src2}",
978 (v4f32 (unpckh VR128:$src1,
979 (memopv4f32 addr:$src2))))]>;
981 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
982 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
983 "unpcklps\t{$src2, $dst|$dst, $src2}",
985 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
986 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
987 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
988 "unpcklps\t{$src2, $dst|$dst, $src2}",
990 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
992 } // Constraints = "$src1 = $dst"
995 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
996 "movmskps\t{$src, $dst|$dst, $src}",
997 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
998 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
999 "movmskpd\t{$src, $dst|$dst, $src}",
1000 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1002 // Prefetch intrinsic.
1003 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1004 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1005 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1006 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1007 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1008 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1009 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1010 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1012 // Non-temporal stores
1013 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1014 "movntps\t{$src, $dst|$dst, $src}",
1015 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1017 // Load, store, and memory fence
1018 def SFENCE : PSI<0xAE, MRM7r, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
1021 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1022 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1023 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1024 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1026 // Alias instructions that map zero vector to pxor / xorp* for sse.
1027 // We set canFoldAsLoad because this can be converted to a constant-pool
1028 // load of an all-zeros value if folding it would be beneficial.
1029 // FIXME: Change encoding to pseudo!
1030 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1031 isCodeGenOnly = 1 in
1032 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1033 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1035 let Predicates = [HasSSE1] in {
1036 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
1037 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
1038 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
1039 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
1040 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
1043 // FR32 to 128-bit vector conversion.
1044 let isAsCheapAsAMove = 1 in
1045 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
1046 "movss\t{$src, $dst|$dst, $src}",
1048 (v4f32 (scalar_to_vector FR32:$src)))]>;
1049 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1050 "movss\t{$src, $dst|$dst, $src}",
1052 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
1054 // FIXME: may not be able to eliminate this movss with coalescing the src and
1055 // dest register classes are different. We really want to write this pattern
1057 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1058 // (f32 FR32:$src)>;
1059 let isAsCheapAsAMove = 1 in
1060 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
1061 "movss\t{$src, $dst|$dst, $src}",
1062 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
1064 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
1065 "movss\t{$src, $dst|$dst, $src}",
1066 [(store (f32 (vector_extract (v4f32 VR128:$src),
1067 (iPTR 0))), addr:$dst)]>;
1070 // Move to lower bits of a VR128, leaving upper bits alone.
1071 // Three operand (but two address) aliases.
1072 let Constraints = "$src1 = $dst" in {
1073 let neverHasSideEffects = 1 in
1074 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
1075 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
1076 "movss\t{$src2, $dst|$dst, $src2}", []>;
1078 let AddedComplexity = 15 in
1079 def MOVLPSrr : SSI<0x10, MRMSrcReg,
1080 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1081 "movss\t{$src2, $dst|$dst, $src2}",
1083 (v4f32 (movl VR128:$src1, VR128:$src2)))]>;
1086 // Move to lower bits of a VR128 and zeroing upper bits.
1087 // Loading from memory automatically zeroing upper bits.
1088 let AddedComplexity = 20 in
1089 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1090 "movss\t{$src, $dst|$dst, $src}",
1091 [(set VR128:$dst, (v4f32 (X86vzmovl (v4f32 (scalar_to_vector
1092 (loadf32 addr:$src))))))]>;
1094 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
1095 (MOVZSS2PSrm addr:$src)>;
1097 //===---------------------------------------------------------------------===//
1098 // SSE2 Instructions
1099 //===---------------------------------------------------------------------===//
1101 // Move Instructions
1102 let neverHasSideEffects = 1 in
1103 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1104 "movsd\t{$src, $dst|$dst, $src}", []>;
1105 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1106 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1107 "movsd\t{$src, $dst|$dst, $src}",
1108 [(set FR64:$dst, (loadf64 addr:$src))]>;
1109 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1110 "movsd\t{$src, $dst|$dst, $src}",
1111 [(store FR64:$src, addr:$dst)]>;
1113 // Conversion instructions
1114 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1115 "cvttsd2si\t{$src, $dst|$dst, $src}",
1116 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1117 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1118 "cvttsd2si\t{$src, $dst|$dst, $src}",
1119 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1120 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1121 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1122 [(set FR32:$dst, (fround FR64:$src))]>;
1123 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1124 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1125 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
1126 Requires<[HasSSE2, OptForSize]>;
1127 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1128 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1129 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1130 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1131 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1132 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1134 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1135 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1136 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1137 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1138 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1139 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1140 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1141 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1142 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1143 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1144 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1145 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1146 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1147 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1148 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1149 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1150 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1151 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1152 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1153 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1155 // SSE2 instructions with XS prefix
1156 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1157 "cvtss2sd\t{$src, $dst|$dst, $src}",
1158 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1159 Requires<[HasSSE2]>;
1160 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1161 "cvtss2sd\t{$src, $dst|$dst, $src}",
1162 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1163 Requires<[HasSSE2, OptForSize]>;
1165 def : Pat<(extloadf32 addr:$src),
1166 (CVTSS2SDrr (MOVSSrm addr:$src))>, Requires<[HasSSE2, OptForSpeed]>;
1168 // Match intrinsics which expect XMM operand(s).
1169 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1170 "cvtsd2si\t{$src, $dst|$dst, $src}",
1171 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1172 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1173 "cvtsd2si\t{$src, $dst|$dst, $src}",
1174 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1175 (load addr:$src)))]>;
1177 // Match intrinisics which expect MM and XMM operand(s).
1178 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1179 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1180 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1181 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1182 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1183 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1184 (memop addr:$src)))]>;
1185 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1186 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1187 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1188 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1189 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1190 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1191 (memop addr:$src)))]>;
1192 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1193 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1194 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1195 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1196 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1197 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1198 (load addr:$src)))]>;
1200 // Aliases for intrinsics
1201 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1202 "cvttsd2si\t{$src, $dst|$dst, $src}",
1204 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1205 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1206 "cvttsd2si\t{$src, $dst|$dst, $src}",
1207 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1208 (load addr:$src)))]>;
1210 // Comparison instructions
1211 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1212 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1213 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1214 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1216 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1217 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1218 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1221 let Defs = [EFLAGS] in {
1222 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1223 "ucomisd\t{$src2, $src1|$src1, $src2}",
1224 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1225 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1226 "ucomisd\t{$src2, $src1|$src1, $src2}",
1227 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1228 (implicit EFLAGS)]>;
1229 } // Defs = [EFLAGS]
1231 // Aliases to match intrinsics which expect XMM operand(s).
1232 let Constraints = "$src1 = $dst" in {
1233 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1235 (ins VR128:$src1, VR128:$src, SSECC:$cc),
1236 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1237 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1238 VR128:$src, imm:$cc))]>;
1239 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1241 (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1242 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1243 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1244 (load addr:$src), imm:$cc))]>;
1247 let Defs = [EFLAGS] in {
1248 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1249 "ucomisd\t{$src2, $src1|$src1, $src2}",
1250 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1251 (implicit EFLAGS)]>;
1252 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1253 "ucomisd\t{$src2, $src1|$src1, $src2}",
1254 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1255 (implicit EFLAGS)]>;
1257 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1258 "comisd\t{$src2, $src1|$src1, $src2}",
1259 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1260 (implicit EFLAGS)]>;
1261 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1262 "comisd\t{$src2, $src1|$src1, $src2}",
1263 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1264 (implicit EFLAGS)]>;
1265 } // Defs = [EFLAGS]
1267 // Aliases of packed SSE2 instructions for scalar use. These all have names
1268 // that start with 'Fs'.
1270 // Alias instructions that map fld0 to pxor for sse.
1271 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1272 canFoldAsLoad = 1 in
1273 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1274 [(set FR64:$dst, fpimm0)]>,
1275 Requires<[HasSSE2]>, TB, OpSize;
1277 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1279 let neverHasSideEffects = 1 in
1280 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1281 "movapd\t{$src, $dst|$dst, $src}", []>;
1283 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1285 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1286 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1287 "movapd\t{$src, $dst|$dst, $src}",
1288 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1290 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1291 let Constraints = "$src1 = $dst" in {
1292 let isCommutable = 1 in {
1293 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1294 (ins FR64:$src1, FR64:$src2),
1295 "andpd\t{$src2, $dst|$dst, $src2}",
1296 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1297 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1298 (ins FR64:$src1, FR64:$src2),
1299 "orpd\t{$src2, $dst|$dst, $src2}",
1300 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1301 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1302 (ins FR64:$src1, FR64:$src2),
1303 "xorpd\t{$src2, $dst|$dst, $src2}",
1304 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1307 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1308 (ins FR64:$src1, f128mem:$src2),
1309 "andpd\t{$src2, $dst|$dst, $src2}",
1310 [(set FR64:$dst, (X86fand FR64:$src1,
1311 (memopfsf64 addr:$src2)))]>;
1312 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1313 (ins FR64:$src1, f128mem:$src2),
1314 "orpd\t{$src2, $dst|$dst, $src2}",
1315 [(set FR64:$dst, (X86for FR64:$src1,
1316 (memopfsf64 addr:$src2)))]>;
1317 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1318 (ins FR64:$src1, f128mem:$src2),
1319 "xorpd\t{$src2, $dst|$dst, $src2}",
1320 [(set FR64:$dst, (X86fxor FR64:$src1,
1321 (memopfsf64 addr:$src2)))]>;
1323 let neverHasSideEffects = 1 in {
1324 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1325 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1326 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1328 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1329 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1330 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1334 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1336 /// In addition, we also have a special variant of the scalar form here to
1337 /// represent the associated intrinsic operation. This form is unlike the
1338 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1339 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1341 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1342 /// six "instructions".
1344 let Constraints = "$src1 = $dst" in {
1345 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1346 SDNode OpNode, Intrinsic F64Int,
1347 bit Commutable = 0> {
1348 // Scalar operation, reg+reg.
1349 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1350 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1351 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1352 let isCommutable = Commutable;
1355 // Scalar operation, reg+mem.
1356 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1357 (ins FR64:$src1, f64mem:$src2),
1358 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1359 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1361 // Vector operation, reg+reg.
1362 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1363 (ins VR128:$src1, VR128:$src2),
1364 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1365 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1366 let isCommutable = Commutable;
1369 // Vector operation, reg+mem.
1370 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1371 (ins VR128:$src1, f128mem:$src2),
1372 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1373 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1375 // Intrinsic operation, reg+reg.
1376 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1377 (ins VR128:$src1, VR128:$src2),
1378 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1379 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]>;
1381 // Intrinsic operation, reg+mem.
1382 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1383 (ins VR128:$src1, sdmem:$src2),
1384 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1385 [(set VR128:$dst, (F64Int VR128:$src1,
1386 sse_load_f64:$src2))]>;
1390 // Arithmetic instructions
1391 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1392 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1393 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1394 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1396 /// sse2_fp_binop_rm - Other SSE2 binops
1398 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1399 /// instructions for a full-vector intrinsic form. Operations that map
1400 /// onto C operators don't use this form since they just use the plain
1401 /// vector form instead of having a separate vector intrinsic form.
1403 /// This provides a total of eight "instructions".
1405 let Constraints = "$src1 = $dst" in {
1406 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1410 bit Commutable = 0> {
1412 // Scalar operation, reg+reg.
1413 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1414 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1415 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1416 let isCommutable = Commutable;
1419 // Scalar operation, reg+mem.
1420 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1421 (ins FR64:$src1, f64mem:$src2),
1422 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1423 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1425 // Vector operation, reg+reg.
1426 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1427 (ins VR128:$src1, VR128:$src2),
1428 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1429 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1430 let isCommutable = Commutable;
1433 // Vector operation, reg+mem.
1434 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1435 (ins VR128:$src1, f128mem:$src2),
1436 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1437 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1439 // Intrinsic operation, reg+reg.
1440 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1441 (ins VR128:$src1, VR128:$src2),
1442 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1443 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1444 let isCommutable = Commutable;
1447 // Intrinsic operation, reg+mem.
1448 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1449 (ins VR128:$src1, sdmem:$src2),
1450 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1451 [(set VR128:$dst, (F64Int VR128:$src1,
1452 sse_load_f64:$src2))]>;
1454 // Vector intrinsic operation, reg+reg.
1455 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1456 (ins VR128:$src1, VR128:$src2),
1457 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1458 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1459 let isCommutable = Commutable;
1462 // Vector intrinsic operation, reg+mem.
1463 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1464 (ins VR128:$src1, f128mem:$src2),
1465 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1466 [(set VR128:$dst, (V2F64Int VR128:$src1,
1467 (memopv2f64 addr:$src2)))]>;
1471 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1472 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1473 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1474 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1476 //===---------------------------------------------------------------------===//
1477 // SSE packed FP Instructions
1479 // Move Instructions
1480 let neverHasSideEffects = 1 in
1481 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1482 "movapd\t{$src, $dst|$dst, $src}", []>;
1483 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1484 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1485 "movapd\t{$src, $dst|$dst, $src}",
1486 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1488 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1489 "movapd\t{$src, $dst|$dst, $src}",
1490 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1492 let neverHasSideEffects = 1 in
1493 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1494 "movupd\t{$src, $dst|$dst, $src}", []>;
1495 let canFoldAsLoad = 1 in
1496 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1497 "movupd\t{$src, $dst|$dst, $src}",
1498 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1499 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1500 "movupd\t{$src, $dst|$dst, $src}",
1501 [(store (v2f64 VR128:$src), addr:$dst)]>;
1503 // Intrinsic forms of MOVUPD load and store
1504 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1505 "movupd\t{$src, $dst|$dst, $src}",
1506 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1507 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1508 "movupd\t{$src, $dst|$dst, $src}",
1509 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1511 let Constraints = "$src1 = $dst" in {
1512 let AddedComplexity = 20 in {
1513 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1514 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1515 "movlpd\t{$src2, $dst|$dst, $src2}",
1517 (v2f64 (movlp VR128:$src1,
1518 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1519 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1520 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1521 "movhpd\t{$src2, $dst|$dst, $src2}",
1523 (v2f64 (movlhps VR128:$src1,
1524 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1525 } // AddedComplexity
1526 } // Constraints = "$src1 = $dst"
1528 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1529 "movlpd\t{$src, $dst|$dst, $src}",
1530 [(store (f64 (vector_extract (v2f64 VR128:$src),
1531 (iPTR 0))), addr:$dst)]>;
1533 // v2f64 extract element 1 is always custom lowered to unpack high to low
1534 // and extract element 0 so the non-store version isn't too horrible.
1535 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1536 "movhpd\t{$src, $dst|$dst, $src}",
1537 [(store (f64 (vector_extract
1538 (v2f64 (unpckh VR128:$src, (undef))),
1539 (iPTR 0))), addr:$dst)]>;
1541 // SSE2 instructions without OpSize prefix
1542 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1543 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1544 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1545 TB, Requires<[HasSSE2]>;
1546 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1547 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1548 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1549 (bitconvert (memopv2i64 addr:$src))))]>,
1550 TB, Requires<[HasSSE2]>;
1552 // SSE2 instructions with XS prefix
1553 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1554 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1555 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1556 XS, Requires<[HasSSE2]>;
1557 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1558 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1559 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1560 (bitconvert (memopv2i64 addr:$src))))]>,
1561 XS, Requires<[HasSSE2]>;
1563 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1564 "cvtps2dq\t{$src, $dst|$dst, $src}",
1565 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1566 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1567 "cvtps2dq\t{$src, $dst|$dst, $src}",
1568 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1569 (memop addr:$src)))]>;
1570 // SSE2 packed instructions with XS prefix
1571 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1572 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1573 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1574 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1576 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1577 "cvttps2dq\t{$src, $dst|$dst, $src}",
1579 (int_x86_sse2_cvttps2dq VR128:$src))]>,
1580 XS, Requires<[HasSSE2]>;
1581 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1582 "cvttps2dq\t{$src, $dst|$dst, $src}",
1583 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1584 (memop addr:$src)))]>,
1585 XS, Requires<[HasSSE2]>;
1587 // SSE2 packed instructions with XD prefix
1588 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1589 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1590 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1591 XD, Requires<[HasSSE2]>;
1592 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1593 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1594 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1595 (memop addr:$src)))]>,
1596 XD, Requires<[HasSSE2]>;
1598 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1599 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1600 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1601 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1602 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1603 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1604 (memop addr:$src)))]>;
1606 // SSE2 instructions without OpSize prefix
1607 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1608 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1609 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1610 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1612 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1613 "cvtps2pd\t{$src, $dst|$dst, $src}",
1614 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1615 TB, Requires<[HasSSE2]>;
1616 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1617 "cvtps2pd\t{$src, $dst|$dst, $src}",
1618 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1619 (load addr:$src)))]>,
1620 TB, Requires<[HasSSE2]>;
1622 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1623 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1624 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1625 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1628 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1629 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1630 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1631 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1632 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1633 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1634 (memop addr:$src)))]>;
1636 // Match intrinsics which expect XMM operand(s).
1637 // Aliases for intrinsics
1638 let Constraints = "$src1 = $dst" in {
1639 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1640 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1641 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1642 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1644 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1645 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1646 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1647 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1648 (loadi32 addr:$src2)))]>;
1649 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1650 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1651 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1652 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1654 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1655 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1656 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1657 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1658 (load addr:$src2)))]>;
1659 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1660 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1661 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1662 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1663 VR128:$src2))]>, XS,
1664 Requires<[HasSSE2]>;
1665 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1666 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1667 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1668 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1669 (load addr:$src2)))]>, XS,
1670 Requires<[HasSSE2]>;
1675 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1677 /// In addition, we also have a special variant of the scalar form here to
1678 /// represent the associated intrinsic operation. This form is unlike the
1679 /// plain scalar form, in that it takes an entire vector (instead of a
1680 /// scalar) and leaves the top elements undefined.
1682 /// And, we have a special variant form for a full-vector intrinsic form.
1684 /// These four forms can each have a reg or a mem operand, so there are a
1685 /// total of eight "instructions".
1687 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1691 bit Commutable = 0> {
1692 // Scalar operation, reg.
1693 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1694 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1695 [(set FR64:$dst, (OpNode FR64:$src))]> {
1696 let isCommutable = Commutable;
1699 // Scalar operation, mem.
1700 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1701 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1702 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1704 // Vector operation, reg.
1705 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1706 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1707 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1708 let isCommutable = Commutable;
1711 // Vector operation, mem.
1712 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1713 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1714 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1716 // Intrinsic operation, reg.
1717 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1718 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1719 [(set VR128:$dst, (F64Int VR128:$src))]> {
1720 let isCommutable = Commutable;
1723 // Intrinsic operation, mem.
1724 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1725 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1726 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1728 // Vector intrinsic operation, reg
1729 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1730 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1731 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1732 let isCommutable = Commutable;
1735 // Vector intrinsic operation, mem
1736 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1737 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1738 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1742 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1743 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1745 // There is no f64 version of the reciprocal approximation instructions.
1748 let Constraints = "$src1 = $dst" in {
1749 let isCommutable = 1 in {
1750 def ANDPDrr : PDI<0x54, MRMSrcReg,
1751 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1752 "andpd\t{$src2, $dst|$dst, $src2}",
1754 (and (bc_v2i64 (v2f64 VR128:$src1)),
1755 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1756 def ORPDrr : PDI<0x56, MRMSrcReg,
1757 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1758 "orpd\t{$src2, $dst|$dst, $src2}",
1760 (or (bc_v2i64 (v2f64 VR128:$src1)),
1761 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1762 def XORPDrr : PDI<0x57, MRMSrcReg,
1763 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1764 "xorpd\t{$src2, $dst|$dst, $src2}",
1766 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1767 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1770 def ANDPDrm : PDI<0x54, MRMSrcMem,
1771 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1772 "andpd\t{$src2, $dst|$dst, $src2}",
1774 (and (bc_v2i64 (v2f64 VR128:$src1)),
1775 (memopv2i64 addr:$src2)))]>;
1776 def ORPDrm : PDI<0x56, MRMSrcMem,
1777 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1778 "orpd\t{$src2, $dst|$dst, $src2}",
1780 (or (bc_v2i64 (v2f64 VR128:$src1)),
1781 (memopv2i64 addr:$src2)))]>;
1782 def XORPDrm : PDI<0x57, MRMSrcMem,
1783 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1784 "xorpd\t{$src2, $dst|$dst, $src2}",
1786 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1787 (memopv2i64 addr:$src2)))]>;
1788 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1789 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1790 "andnpd\t{$src2, $dst|$dst, $src2}",
1792 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1793 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1794 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1795 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1796 "andnpd\t{$src2, $dst|$dst, $src2}",
1798 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1799 (memopv2i64 addr:$src2)))]>;
1802 let Constraints = "$src1 = $dst" in {
1803 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1804 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1805 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1806 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1807 VR128:$src, imm:$cc))]>;
1808 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1809 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1810 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1811 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1812 (memop addr:$src), imm:$cc))]>;
1814 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1815 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1816 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1817 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1819 // Shuffle and unpack instructions
1820 let Constraints = "$src1 = $dst" in {
1821 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1822 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1823 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1825 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1826 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1827 (outs VR128:$dst), (ins VR128:$src1,
1828 f128mem:$src2, i8imm:$src3),
1829 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1832 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1834 let AddedComplexity = 10 in {
1835 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1836 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1837 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1839 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1840 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1841 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1842 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1844 (v2f64 (unpckh VR128:$src1,
1845 (memopv2f64 addr:$src2))))]>;
1847 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1848 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1849 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1851 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1852 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1853 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1854 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1856 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1857 } // AddedComplexity
1858 } // Constraints = "$src1 = $dst"
1861 //===---------------------------------------------------------------------===//
1862 // SSE integer instructions
1864 // Move Instructions
1865 let neverHasSideEffects = 1 in
1866 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1867 "movdqa\t{$src, $dst|$dst, $src}", []>;
1868 let canFoldAsLoad = 1, mayLoad = 1 in
1869 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1870 "movdqa\t{$src, $dst|$dst, $src}",
1871 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1873 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1874 "movdqa\t{$src, $dst|$dst, $src}",
1875 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1876 let canFoldAsLoad = 1, mayLoad = 1 in
1877 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1878 "movdqu\t{$src, $dst|$dst, $src}",
1879 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1880 XS, Requires<[HasSSE2]>;
1882 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1883 "movdqu\t{$src, $dst|$dst, $src}",
1884 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1885 XS, Requires<[HasSSE2]>;
1887 // Intrinsic forms of MOVDQU load and store
1888 let canFoldAsLoad = 1 in
1889 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1890 "movdqu\t{$src, $dst|$dst, $src}",
1891 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1892 XS, Requires<[HasSSE2]>;
1893 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1894 "movdqu\t{$src, $dst|$dst, $src}",
1895 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1896 XS, Requires<[HasSSE2]>;
1898 let Constraints = "$src1 = $dst" in {
1900 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1901 bit Commutable = 0> {
1902 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1903 (ins VR128:$src1, VR128:$src2),
1904 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1905 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1906 let isCommutable = Commutable;
1908 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1909 (ins VR128:$src1, i128mem:$src2),
1910 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1911 [(set VR128:$dst, (IntId VR128:$src1,
1912 (bitconvert (memopv2i64
1916 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1918 Intrinsic IntId, Intrinsic IntId2> {
1919 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1920 (ins VR128:$src1, VR128:$src2),
1921 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1922 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1923 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1924 (ins VR128:$src1, i128mem:$src2),
1925 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1926 [(set VR128:$dst, (IntId VR128:$src1,
1927 (bitconvert (memopv2i64 addr:$src2))))]>;
1928 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
1929 (ins VR128:$src1, i32i8imm:$src2),
1930 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1931 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1934 /// PDI_binop_rm - Simple SSE2 binary operator.
1935 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1936 ValueType OpVT, bit Commutable = 0> {
1937 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1938 (ins VR128:$src1, VR128:$src2),
1939 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1940 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1941 let isCommutable = Commutable;
1943 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1944 (ins VR128:$src1, i128mem:$src2),
1945 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1946 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1947 (bitconvert (memopv2i64 addr:$src2)))))]>;
1950 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1952 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1953 /// to collapse (bitconvert VT to VT) into its operand.
1955 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1956 bit Commutable = 0> {
1957 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1958 (ins VR128:$src1, VR128:$src2),
1959 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1960 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1961 let isCommutable = Commutable;
1963 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1964 (ins VR128:$src1, i128mem:$src2),
1965 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1966 [(set VR128:$dst, (OpNode VR128:$src1,
1967 (memopv2i64 addr:$src2)))]>;
1970 } // Constraints = "$src1 = $dst"
1972 // 128-bit Integer Arithmetic
1974 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1975 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1976 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1977 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1979 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1980 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1981 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1982 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1984 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1985 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1986 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1987 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1989 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1990 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1991 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1992 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1994 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1996 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1997 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1998 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
2000 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
2002 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
2003 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
2006 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
2007 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
2008 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
2009 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
2010 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
2013 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
2014 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
2015 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
2016 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
2017 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
2018 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
2020 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2021 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2022 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2023 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2024 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2025 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2027 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2028 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2029 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2030 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2032 // 128-bit logical shifts.
2033 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
2034 def PSLLDQri : PDIi8<0x73, MRM7r,
2035 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2036 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2037 def PSRLDQri : PDIi8<0x73, MRM3r,
2038 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2039 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2040 // PSRADQri doesn't exist in SSE[1-3].
2043 let Predicates = [HasSSE2] in {
2044 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2045 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2046 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2047 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2048 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2049 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2050 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2051 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2052 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2053 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2055 // Shift up / down and insert zero's.
2056 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2057 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2058 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2059 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2063 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2064 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2065 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2067 let Constraints = "$src1 = $dst" in {
2068 def PANDNrr : PDI<0xDF, MRMSrcReg,
2069 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2070 "pandn\t{$src2, $dst|$dst, $src2}",
2071 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2074 def PANDNrm : PDI<0xDF, MRMSrcMem,
2075 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2076 "pandn\t{$src2, $dst|$dst, $src2}",
2077 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2078 (memopv2i64 addr:$src2))))]>;
2081 // SSE2 Integer comparison
2082 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2083 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2084 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2085 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2086 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2087 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2089 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2090 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2091 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2092 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2093 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2094 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2095 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2096 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2097 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2098 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2099 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2100 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2102 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2103 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2104 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2105 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2106 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2107 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2108 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2109 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2110 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2111 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2112 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2113 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2116 // Pack instructions
2117 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2118 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2119 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2121 // Shuffle and unpack instructions
2122 let AddedComplexity = 5 in {
2123 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2124 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2125 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2126 [(set VR128:$dst, (v4i32 (pshufd:$src2
2127 VR128:$src1, (undef))))]>;
2128 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2129 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2130 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2131 [(set VR128:$dst, (v4i32 (pshufd:$src2
2132 (bc_v4i32 (memopv2i64 addr:$src1)),
2136 // SSE2 with ImmT == Imm8 and XS prefix.
2137 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2138 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2139 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2140 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2142 XS, Requires<[HasSSE2]>;
2143 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2144 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2145 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2146 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2147 (bc_v8i16 (memopv2i64 addr:$src1)),
2149 XS, Requires<[HasSSE2]>;
2151 // SSE2 with ImmT == Imm8 and XD prefix.
2152 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2153 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2154 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2155 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2157 XD, Requires<[HasSSE2]>;
2158 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2159 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2160 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2161 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2162 (bc_v8i16 (memopv2i64 addr:$src1)),
2164 XD, Requires<[HasSSE2]>;
2167 let Constraints = "$src1 = $dst" in {
2168 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2169 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2170 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2172 (v16i8 (unpckl VR128:$src1, VR128:$src2)))]>;
2173 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2174 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2175 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2177 (unpckl VR128:$src1,
2178 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2179 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2180 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2181 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2183 (v8i16 (unpckl VR128:$src1, VR128:$src2)))]>;
2184 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2185 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2186 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2188 (unpckl VR128:$src1,
2189 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2190 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2191 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2192 "punpckldq\t{$src2, $dst|$dst, $src2}",
2194 (v4i32 (unpckl VR128:$src1, VR128:$src2)))]>;
2195 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2196 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2197 "punpckldq\t{$src2, $dst|$dst, $src2}",
2199 (unpckl VR128:$src1,
2200 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2201 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2202 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2203 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2205 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2206 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2207 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2208 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2210 (v2i64 (unpckl VR128:$src1,
2211 (memopv2i64 addr:$src2))))]>;
2213 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2214 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2215 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2217 (v16i8 (unpckh VR128:$src1, VR128:$src2)))]>;
2218 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2219 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2220 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2222 (unpckh VR128:$src1,
2223 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2224 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2225 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2226 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2228 (v8i16 (unpckh VR128:$src1, VR128:$src2)))]>;
2229 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2230 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2231 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2233 (unpckh VR128:$src1,
2234 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2235 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2236 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2237 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2239 (v4i32 (unpckh VR128:$src1, VR128:$src2)))]>;
2240 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2241 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2242 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2244 (unpckh VR128:$src1,
2245 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2246 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2247 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2248 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2250 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2251 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2252 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2253 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2255 (v2i64 (unpckh VR128:$src1,
2256 (memopv2i64 addr:$src2))))]>;
2260 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2261 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2262 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2263 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2265 let Constraints = "$src1 = $dst" in {
2266 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2267 (outs VR128:$dst), (ins VR128:$src1,
2268 GR32:$src2, i32i8imm:$src3),
2269 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2271 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2272 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2273 (outs VR128:$dst), (ins VR128:$src1,
2274 i16mem:$src2, i32i8imm:$src3),
2275 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2277 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2282 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2283 "pmovmskb\t{$src, $dst|$dst, $src}",
2284 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2286 // Conditional store
2288 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2289 "maskmovdqu\t{$mask, $src|$src, $mask}",
2290 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2293 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2294 "maskmovdqu\t{$mask, $src|$src, $mask}",
2295 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2297 // Non-temporal stores
2298 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2299 "movntpd\t{$src, $dst|$dst, $src}",
2300 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2301 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2302 "movntdq\t{$src, $dst|$dst, $src}",
2303 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2304 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2305 "movnti\t{$src, $dst|$dst, $src}",
2306 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2307 TB, Requires<[HasSSE2]>;
2310 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2311 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2312 TB, Requires<[HasSSE2]>;
2314 // Load, store, and memory fence
2315 def LFENCE : I<0xAE, MRM5r, (outs), (ins),
2316 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2317 def MFENCE : I<0xAE, MRM6r, (outs), (ins),
2318 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2320 //TODO: custom lower this so as to never even generate the noop
2321 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2323 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2324 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2325 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2328 // Alias instructions that map zero vector to pxor / xorp* for sse.
2329 // We set canFoldAsLoad because this can be converted to a constant-pool
2330 // load of an all-ones value if folding it would be beneficial.
2331 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2332 isCodeGenOnly = 1 in
2333 // FIXME: Change encoding to pseudo.
2334 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
2335 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2337 // FR64 to 128-bit vector conversion.
2338 let isAsCheapAsAMove = 1 in
2339 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2340 "movsd\t{$src, $dst|$dst, $src}",
2342 (v2f64 (scalar_to_vector FR64:$src)))]>;
2343 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2344 "movsd\t{$src, $dst|$dst, $src}",
2346 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2348 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2349 "movd\t{$src, $dst|$dst, $src}",
2351 (v4i32 (scalar_to_vector GR32:$src)))]>;
2352 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2353 "movd\t{$src, $dst|$dst, $src}",
2355 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2357 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2358 "movd\t{$src, $dst|$dst, $src}",
2359 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2361 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2362 "movd\t{$src, $dst|$dst, $src}",
2363 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2365 // SSE2 instructions with XS prefix
2366 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2367 "movq\t{$src, $dst|$dst, $src}",
2369 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2370 Requires<[HasSSE2]>;
2371 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2372 "movq\t{$src, $dst|$dst, $src}",
2373 [(store (i64 (vector_extract (v2i64 VR128:$src),
2374 (iPTR 0))), addr:$dst)]>;
2376 // FIXME: may not be able to eliminate this movss with coalescing the src and
2377 // dest register classes are different. We really want to write this pattern
2379 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2380 // (f32 FR32:$src)>;
2381 let isAsCheapAsAMove = 1 in
2382 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2383 "movsd\t{$src, $dst|$dst, $src}",
2384 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2386 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2387 "movsd\t{$src, $dst|$dst, $src}",
2388 [(store (f64 (vector_extract (v2f64 VR128:$src),
2389 (iPTR 0))), addr:$dst)]>;
2390 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2391 "movd\t{$src, $dst|$dst, $src}",
2392 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2394 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2395 "movd\t{$src, $dst|$dst, $src}",
2396 [(store (i32 (vector_extract (v4i32 VR128:$src),
2397 (iPTR 0))), addr:$dst)]>;
2399 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2400 "movd\t{$src, $dst|$dst, $src}",
2401 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2402 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2403 "movd\t{$src, $dst|$dst, $src}",
2404 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2407 // Move to lower bits of a VR128, leaving upper bits alone.
2408 // Three operand (but two address) aliases.
2409 let Constraints = "$src1 = $dst" in {
2410 let neverHasSideEffects = 1 in
2411 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2412 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2413 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2415 let AddedComplexity = 15 in
2416 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2417 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2418 "movsd\t{$src2, $dst|$dst, $src2}",
2420 (v2f64 (movl VR128:$src1, VR128:$src2)))]>;
2423 // Store / copy lower 64-bits of a XMM register.
2424 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2425 "movq\t{$src, $dst|$dst, $src}",
2426 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2428 // Move to lower bits of a VR128 and zeroing upper bits.
2429 // Loading from memory automatically zeroing upper bits.
2430 let AddedComplexity = 20 in {
2431 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2432 "movsd\t{$src, $dst|$dst, $src}",
2434 (v2f64 (X86vzmovl (v2f64 (scalar_to_vector
2435 (loadf64 addr:$src))))))]>;
2437 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
2438 (MOVZSD2PDrm addr:$src)>;
2439 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
2440 (MOVZSD2PDrm addr:$src)>;
2441 def : Pat<(v2f64 (X86vzload addr:$src)), (MOVZSD2PDrm addr:$src)>;
2444 // movd / movq to XMM register zero-extends
2445 let AddedComplexity = 15 in {
2446 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2447 "movd\t{$src, $dst|$dst, $src}",
2448 [(set VR128:$dst, (v4i32 (X86vzmovl
2449 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2450 // This is X86-64 only.
2451 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2452 "mov{d|q}\t{$src, $dst|$dst, $src}",
2453 [(set VR128:$dst, (v2i64 (X86vzmovl
2454 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2457 let AddedComplexity = 20 in {
2458 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2459 "movd\t{$src, $dst|$dst, $src}",
2461 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2462 (loadi32 addr:$src))))))]>;
2464 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2465 (MOVZDI2PDIrm addr:$src)>;
2466 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2467 (MOVZDI2PDIrm addr:$src)>;
2468 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2469 (MOVZDI2PDIrm addr:$src)>;
2471 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2472 "movq\t{$src, $dst|$dst, $src}",
2474 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2475 (loadi64 addr:$src))))))]>, XS,
2476 Requires<[HasSSE2]>;
2478 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2479 (MOVZQI2PQIrm addr:$src)>;
2480 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2481 (MOVZQI2PQIrm addr:$src)>;
2482 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2485 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2486 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2487 let AddedComplexity = 15 in
2488 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2489 "movq\t{$src, $dst|$dst, $src}",
2490 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2491 XS, Requires<[HasSSE2]>;
2493 let AddedComplexity = 20 in {
2494 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2495 "movq\t{$src, $dst|$dst, $src}",
2496 [(set VR128:$dst, (v2i64 (X86vzmovl
2497 (loadv2i64 addr:$src))))]>,
2498 XS, Requires<[HasSSE2]>;
2500 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2501 (MOVZPQILo2PQIrm addr:$src)>;
2504 // Instructions for the disassembler
2505 // xr = XMM register
2508 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2509 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2511 //===---------------------------------------------------------------------===//
2512 // SSE3 Instructions
2513 //===---------------------------------------------------------------------===//
2515 // Move Instructions
2516 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2517 "movshdup\t{$src, $dst|$dst, $src}",
2518 [(set VR128:$dst, (v4f32 (movshdup
2519 VR128:$src, (undef))))]>;
2520 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2521 "movshdup\t{$src, $dst|$dst, $src}",
2522 [(set VR128:$dst, (movshdup
2523 (memopv4f32 addr:$src), (undef)))]>;
2525 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2526 "movsldup\t{$src, $dst|$dst, $src}",
2527 [(set VR128:$dst, (v4f32 (movsldup
2528 VR128:$src, (undef))))]>;
2529 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2530 "movsldup\t{$src, $dst|$dst, $src}",
2531 [(set VR128:$dst, (movsldup
2532 (memopv4f32 addr:$src), (undef)))]>;
2534 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2535 "movddup\t{$src, $dst|$dst, $src}",
2536 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2537 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2538 "movddup\t{$src, $dst|$dst, $src}",
2540 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2543 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2545 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2547 let AddedComplexity = 5 in {
2548 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2549 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2550 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2551 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2552 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2553 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2554 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2555 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2559 let Constraints = "$src1 = $dst" in {
2560 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2561 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2562 "addsubps\t{$src2, $dst|$dst, $src2}",
2563 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2565 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2566 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2567 "addsubps\t{$src2, $dst|$dst, $src2}",
2568 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2569 (memop addr:$src2)))]>;
2570 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2571 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2572 "addsubpd\t{$src2, $dst|$dst, $src2}",
2573 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2575 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2576 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2577 "addsubpd\t{$src2, $dst|$dst, $src2}",
2578 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2579 (memop addr:$src2)))]>;
2582 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2583 "lddqu\t{$src, $dst|$dst, $src}",
2584 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2587 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2588 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2589 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2590 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2591 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2592 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2593 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2594 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2595 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2596 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2597 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2598 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2599 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2600 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2601 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2602 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2604 let Constraints = "$src1 = $dst" in {
2605 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2606 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2607 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2608 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2609 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2610 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2611 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2612 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2615 // Thread synchronization
2616 def MONITOR : I<0x01, MRM1r, (outs), (ins), "monitor",
2617 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2618 def MWAIT : I<0x01, MRM1r, (outs), (ins), "mwait",
2619 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2621 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2622 let AddedComplexity = 15 in
2623 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2624 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2625 let AddedComplexity = 20 in
2626 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2627 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2629 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2630 let AddedComplexity = 15 in
2631 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2632 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2633 let AddedComplexity = 20 in
2634 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2635 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2637 //===---------------------------------------------------------------------===//
2638 // SSSE3 Instructions
2639 //===---------------------------------------------------------------------===//
2641 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2642 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2643 Intrinsic IntId64, Intrinsic IntId128> {
2644 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2645 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2646 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2648 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2649 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2651 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2653 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2655 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2656 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2659 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2661 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2664 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2667 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2668 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2669 Intrinsic IntId64, Intrinsic IntId128> {
2670 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2672 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2673 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2675 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2677 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2680 (bitconvert (memopv4i16 addr:$src))))]>;
2682 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2684 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2685 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2688 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2690 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2693 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2696 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2697 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2698 Intrinsic IntId64, Intrinsic IntId128> {
2699 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2701 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2702 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2704 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2706 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2709 (bitconvert (memopv2i32 addr:$src))))]>;
2711 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2713 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2714 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2717 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2719 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2722 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2725 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2726 int_x86_ssse3_pabs_b,
2727 int_x86_ssse3_pabs_b_128>;
2728 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2729 int_x86_ssse3_pabs_w,
2730 int_x86_ssse3_pabs_w_128>;
2731 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2732 int_x86_ssse3_pabs_d,
2733 int_x86_ssse3_pabs_d_128>;
2735 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2736 let Constraints = "$src1 = $dst" in {
2737 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2738 Intrinsic IntId64, Intrinsic IntId128,
2739 bit Commutable = 0> {
2740 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2741 (ins VR64:$src1, VR64:$src2),
2742 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2743 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2744 let isCommutable = Commutable;
2746 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2747 (ins VR64:$src1, i64mem:$src2),
2748 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2750 (IntId64 VR64:$src1,
2751 (bitconvert (memopv8i8 addr:$src2))))]>;
2753 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2754 (ins VR128:$src1, VR128:$src2),
2755 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2756 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2758 let isCommutable = Commutable;
2760 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2761 (ins VR128:$src1, i128mem:$src2),
2762 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2764 (IntId128 VR128:$src1,
2765 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2769 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2770 let Constraints = "$src1 = $dst" in {
2771 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2772 Intrinsic IntId64, Intrinsic IntId128,
2773 bit Commutable = 0> {
2774 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2775 (ins VR64:$src1, VR64:$src2),
2776 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2777 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2778 let isCommutable = Commutable;
2780 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2781 (ins VR64:$src1, i64mem:$src2),
2782 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2784 (IntId64 VR64:$src1,
2785 (bitconvert (memopv4i16 addr:$src2))))]>;
2787 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2788 (ins VR128:$src1, VR128:$src2),
2789 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2790 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2792 let isCommutable = Commutable;
2794 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2795 (ins VR128:$src1, i128mem:$src2),
2796 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2798 (IntId128 VR128:$src1,
2799 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2803 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2804 let Constraints = "$src1 = $dst" in {
2805 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2806 Intrinsic IntId64, Intrinsic IntId128,
2807 bit Commutable = 0> {
2808 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2809 (ins VR64:$src1, VR64:$src2),
2810 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2811 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2812 let isCommutable = Commutable;
2814 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2815 (ins VR64:$src1, i64mem:$src2),
2816 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2818 (IntId64 VR64:$src1,
2819 (bitconvert (memopv2i32 addr:$src2))))]>;
2821 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2822 (ins VR128:$src1, VR128:$src2),
2823 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2824 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2826 let isCommutable = Commutable;
2828 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2829 (ins VR128:$src1, i128mem:$src2),
2830 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2832 (IntId128 VR128:$src1,
2833 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2837 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2838 int_x86_ssse3_phadd_w,
2839 int_x86_ssse3_phadd_w_128>;
2840 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2841 int_x86_ssse3_phadd_d,
2842 int_x86_ssse3_phadd_d_128>;
2843 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2844 int_x86_ssse3_phadd_sw,
2845 int_x86_ssse3_phadd_sw_128>;
2846 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2847 int_x86_ssse3_phsub_w,
2848 int_x86_ssse3_phsub_w_128>;
2849 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2850 int_x86_ssse3_phsub_d,
2851 int_x86_ssse3_phsub_d_128>;
2852 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2853 int_x86_ssse3_phsub_sw,
2854 int_x86_ssse3_phsub_sw_128>;
2855 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2856 int_x86_ssse3_pmadd_ub_sw,
2857 int_x86_ssse3_pmadd_ub_sw_128>;
2858 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2859 int_x86_ssse3_pmul_hr_sw,
2860 int_x86_ssse3_pmul_hr_sw_128, 1>;
2861 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2862 int_x86_ssse3_pshuf_b,
2863 int_x86_ssse3_pshuf_b_128>;
2864 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2865 int_x86_ssse3_psign_b,
2866 int_x86_ssse3_psign_b_128>;
2867 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2868 int_x86_ssse3_psign_w,
2869 int_x86_ssse3_psign_w_128>;
2870 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2871 int_x86_ssse3_psign_d,
2872 int_x86_ssse3_psign_d_128>;
2874 let Constraints = "$src1 = $dst" in {
2875 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2876 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2877 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2879 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2880 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2881 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2884 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2885 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2886 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2888 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2889 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2890 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2894 // palignr patterns.
2895 def : Pat<(int_x86_ssse3_palign_r VR64:$src1, VR64:$src2, (i8 imm:$src3)),
2896 (PALIGNR64rr VR64:$src1, VR64:$src2, (BYTE_imm imm:$src3))>,
2897 Requires<[HasSSSE3]>;
2898 def : Pat<(int_x86_ssse3_palign_r VR64:$src1,
2899 (memop64 addr:$src2),
2901 (PALIGNR64rm VR64:$src1, addr:$src2, (BYTE_imm imm:$src3))>,
2902 Requires<[HasSSSE3]>;
2904 def : Pat<(int_x86_ssse3_palign_r_128 VR128:$src1, VR128:$src2, (i8 imm:$src3)),
2905 (PALIGNR128rr VR128:$src1, VR128:$src2, (BYTE_imm imm:$src3))>,
2906 Requires<[HasSSSE3]>;
2907 def : Pat<(int_x86_ssse3_palign_r_128 VR128:$src1,
2908 (memopv2i64 addr:$src2),
2910 (PALIGNR128rm VR128:$src1, addr:$src2, (BYTE_imm imm:$src3))>,
2911 Requires<[HasSSSE3]>;
2913 let AddedComplexity = 5 in {
2914 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2915 (PALIGNR128rr VR128:$src2, VR128:$src1,
2916 (SHUFFLE_get_palign_imm VR128:$src3))>,
2917 Requires<[HasSSSE3]>;
2918 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2919 (PALIGNR128rr VR128:$src2, VR128:$src1,
2920 (SHUFFLE_get_palign_imm VR128:$src3))>,
2921 Requires<[HasSSSE3]>;
2922 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2923 (PALIGNR128rr VR128:$src2, VR128:$src1,
2924 (SHUFFLE_get_palign_imm VR128:$src3))>,
2925 Requires<[HasSSSE3]>;
2926 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2927 (PALIGNR128rr VR128:$src2, VR128:$src1,
2928 (SHUFFLE_get_palign_imm VR128:$src3))>,
2929 Requires<[HasSSSE3]>;
2932 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2933 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2934 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2935 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2937 //===---------------------------------------------------------------------===//
2938 // Non-Instruction Patterns
2939 //===---------------------------------------------------------------------===//
2941 // extload f32 -> f64. This matches load+fextend because we have a hack in
2942 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2944 // Since these loads aren't folded into the fextend, we have to match it
2946 let Predicates = [HasSSE2] in
2947 def : Pat<(fextend (loadf32 addr:$src)),
2948 (CVTSS2SDrm addr:$src)>;
2951 let Predicates = [HasSSE2] in {
2952 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2953 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2954 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2955 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2956 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2957 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2958 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2959 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2960 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2961 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2962 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2963 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2964 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2965 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2966 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2967 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2968 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2969 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2970 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2971 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2972 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2973 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2974 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2975 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2976 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2977 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2978 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2979 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2980 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2981 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2984 // Move scalar to XMM zero-extended
2985 // movd to XMM register zero-extends
2986 let AddedComplexity = 15 in {
2987 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2988 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2989 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2990 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2991 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE1]>;
2992 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2993 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2994 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2995 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2998 // Splat v2f64 / v2i64
2999 let AddedComplexity = 10 in {
3000 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
3001 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3002 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
3003 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3004 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
3005 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3006 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
3007 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3010 // Special unary SHUFPSrri case.
3011 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
3012 (SHUFPSrri VR128:$src1, VR128:$src1,
3013 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3014 Requires<[HasSSE1]>;
3015 let AddedComplexity = 5 in
3016 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
3017 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3018 Requires<[HasSSE2]>;
3019 // Special unary SHUFPDrri case.
3020 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
3021 (SHUFPDrri VR128:$src1, VR128:$src1,
3022 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3023 Requires<[HasSSE2]>;
3024 // Special unary SHUFPDrri case.
3025 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
3026 (SHUFPDrri VR128:$src1, VR128:$src1,
3027 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3028 Requires<[HasSSE2]>;
3029 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
3030 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
3031 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3032 Requires<[HasSSE2]>;
3034 // Special binary v4i32 shuffle cases with SHUFPS.
3035 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
3036 (SHUFPSrri VR128:$src1, VR128:$src2,
3037 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3038 Requires<[HasSSE2]>;
3039 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
3040 (SHUFPSrmi VR128:$src1, addr:$src2,
3041 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3042 Requires<[HasSSE2]>;
3043 // Special binary v2i64 shuffle cases using SHUFPDrri.
3044 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
3045 (SHUFPDrri VR128:$src1, VR128:$src2,
3046 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3047 Requires<[HasSSE2]>;
3049 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
3050 let AddedComplexity = 15 in {
3051 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
3052 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3053 Requires<[OptForSpeed, HasSSE2]>;
3054 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
3055 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3056 Requires<[OptForSpeed, HasSSE2]>;
3058 let AddedComplexity = 10 in {
3059 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
3060 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
3061 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
3062 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3063 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
3064 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3065 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
3066 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3069 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
3070 let AddedComplexity = 15 in {
3071 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
3072 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3073 Requires<[OptForSpeed, HasSSE2]>;
3074 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
3075 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3076 Requires<[OptForSpeed, HasSSE2]>;
3078 let AddedComplexity = 10 in {
3079 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
3080 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
3081 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
3082 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3083 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
3084 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3085 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
3086 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3089 let AddedComplexity = 20 in {
3090 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3091 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3092 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3094 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3095 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3096 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3098 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3099 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3100 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3101 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3102 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3105 let AddedComplexity = 20 in {
3106 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3107 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3108 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3109 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3110 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3111 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3112 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3113 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3114 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3117 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3118 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3119 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3120 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3121 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3122 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3124 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3125 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3126 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3128 let AddedComplexity = 15 in {
3129 // Setting the lowest element in the vector.
3130 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3131 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3132 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3133 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3135 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
3136 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3137 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3138 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3139 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3142 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3143 // fall back to this for SSE1)
3144 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3145 (SHUFPSrri VR128:$src2, VR128:$src1,
3146 (SHUFFLE_get_shuf_imm VR128:$src3))>, Requires<[HasSSE1]>;
3148 // Set lowest element and zero upper elements.
3149 let AddedComplexity = 15 in
3150 def : Pat<(v2f64 (movl immAllZerosV_bc, VR128:$src)),
3151 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3152 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3153 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3155 // Some special case pandn patterns.
3156 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3158 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3159 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3161 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3162 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3164 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3166 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3167 (memop addr:$src2))),
3168 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3169 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3170 (memop addr:$src2))),
3171 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3172 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3173 (memop addr:$src2))),
3174 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3176 // vector -> vector casts
3177 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3178 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3179 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3180 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3181 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3182 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3183 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3184 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3186 // Use movaps / movups for SSE integer load / store (one byte shorter).
3187 def : Pat<(alignedloadv4i32 addr:$src),
3188 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
3189 def : Pat<(loadv4i32 addr:$src),
3190 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3191 def : Pat<(alignedloadv2i64 addr:$src),
3192 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3193 def : Pat<(loadv2i64 addr:$src),
3194 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3196 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3197 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3198 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3199 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3200 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3201 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3202 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3203 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3204 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3205 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3206 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3207 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3208 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3209 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3210 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3211 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3213 //===----------------------------------------------------------------------===//
3214 // SSE4.1 Instructions
3215 //===----------------------------------------------------------------------===//
3217 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3220 Intrinsic V2F64Int> {
3221 // Intrinsic operation, reg.
3222 // Vector intrinsic operation, reg
3223 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3224 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3225 !strconcat(OpcodeStr,
3226 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3227 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3230 // Vector intrinsic operation, mem
3231 def PSm_Int : Ii8<opcps, MRMSrcMem,
3232 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3233 !strconcat(OpcodeStr,
3234 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3236 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3238 Requires<[HasSSE41]>;
3240 // Vector intrinsic operation, reg
3241 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3242 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3243 !strconcat(OpcodeStr,
3244 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3245 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3248 // Vector intrinsic operation, mem
3249 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3250 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3251 !strconcat(OpcodeStr,
3252 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3254 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3258 let Constraints = "$src1 = $dst" in {
3259 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3263 // Intrinsic operation, reg.
3264 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3266 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3267 !strconcat(OpcodeStr,
3268 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3270 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3273 // Intrinsic operation, mem.
3274 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3276 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3277 !strconcat(OpcodeStr,
3278 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3280 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3283 // Intrinsic operation, reg.
3284 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3286 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3287 !strconcat(OpcodeStr,
3288 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3290 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3293 // Intrinsic operation, mem.
3294 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3296 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3297 !strconcat(OpcodeStr,
3298 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3300 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3305 // FP round - roundss, roundps, roundsd, roundpd
3306 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3307 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3308 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3309 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3311 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3312 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3313 Intrinsic IntId128> {
3314 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3316 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3317 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3318 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3320 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3323 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3326 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3327 int_x86_sse41_phminposuw>;
3329 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3330 let Constraints = "$src1 = $dst" in {
3331 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3332 Intrinsic IntId128, bit Commutable = 0> {
3333 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3334 (ins VR128:$src1, VR128:$src2),
3335 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3336 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3338 let isCommutable = Commutable;
3340 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3341 (ins VR128:$src1, i128mem:$src2),
3342 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3344 (IntId128 VR128:$src1,
3345 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3349 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3350 int_x86_sse41_pcmpeqq, 1>;
3351 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3352 int_x86_sse41_packusdw, 0>;
3353 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3354 int_x86_sse41_pminsb, 1>;
3355 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3356 int_x86_sse41_pminsd, 1>;
3357 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3358 int_x86_sse41_pminud, 1>;
3359 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3360 int_x86_sse41_pminuw, 1>;
3361 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3362 int_x86_sse41_pmaxsb, 1>;
3363 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3364 int_x86_sse41_pmaxsd, 1>;
3365 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3366 int_x86_sse41_pmaxud, 1>;
3367 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3368 int_x86_sse41_pmaxuw, 1>;
3370 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3372 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3373 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3374 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3375 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3377 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3378 let Constraints = "$src1 = $dst" in {
3379 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3380 SDNode OpNode, Intrinsic IntId128,
3381 bit Commutable = 0> {
3382 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3383 (ins VR128:$src1, VR128:$src2),
3384 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3385 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3386 VR128:$src2))]>, OpSize {
3387 let isCommutable = Commutable;
3389 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3390 (ins VR128:$src1, VR128:$src2),
3391 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3392 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3394 let isCommutable = Commutable;
3396 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3397 (ins VR128:$src1, i128mem:$src2),
3398 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3400 (OpNode VR128:$src1, (memop addr:$src2)))]>, OpSize;
3401 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3402 (ins VR128:$src1, i128mem:$src2),
3403 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3405 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3409 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", v4i32, mul,
3410 int_x86_sse41_pmulld, 1>;
3412 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3413 let Constraints = "$src1 = $dst" in {
3414 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3415 Intrinsic IntId128, bit Commutable = 0> {
3416 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3417 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3418 !strconcat(OpcodeStr,
3419 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3421 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3423 let isCommutable = Commutable;
3425 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3426 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3427 !strconcat(OpcodeStr,
3428 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3430 (IntId128 VR128:$src1,
3431 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3436 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3437 int_x86_sse41_blendps, 0>;
3438 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3439 int_x86_sse41_blendpd, 0>;
3440 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3441 int_x86_sse41_pblendw, 0>;
3442 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3443 int_x86_sse41_dpps, 1>;
3444 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3445 int_x86_sse41_dppd, 1>;
3446 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3447 int_x86_sse41_mpsadbw, 1>;
3450 /// SS41I_ternary_int - SSE 4.1 ternary operator
3451 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3452 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3453 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3454 (ins VR128:$src1, VR128:$src2),
3455 !strconcat(OpcodeStr,
3456 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3457 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3460 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3461 (ins VR128:$src1, i128mem:$src2),
3462 !strconcat(OpcodeStr,
3463 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3466 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3470 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3471 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3472 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3475 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3476 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3477 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3478 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3480 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3481 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3483 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3487 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3488 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3489 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3490 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3491 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3492 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3494 // Common patterns involving scalar load.
3495 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3496 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3497 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3498 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3500 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3501 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3502 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3503 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3505 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3506 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3507 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3508 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3510 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3511 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3512 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3513 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3515 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3516 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3517 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3518 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3520 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3521 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3522 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3523 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3526 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3527 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3528 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3529 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3531 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3532 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3534 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3538 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3539 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3540 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3541 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3543 // Common patterns involving scalar load
3544 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3545 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3546 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3547 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3549 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3550 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3551 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3552 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3555 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3556 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3557 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3558 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3560 // Expecting a i16 load any extended to i32 value.
3561 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3562 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3563 [(set VR128:$dst, (IntId (bitconvert
3564 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3568 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3569 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3571 // Common patterns involving scalar load
3572 def : Pat<(int_x86_sse41_pmovsxbq
3573 (bitconvert (v4i32 (X86vzmovl
3574 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3575 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3577 def : Pat<(int_x86_sse41_pmovzxbq
3578 (bitconvert (v4i32 (X86vzmovl
3579 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3580 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3583 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3584 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3585 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3586 (ins VR128:$src1, i32i8imm:$src2),
3587 !strconcat(OpcodeStr,
3588 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3589 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3591 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3592 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3593 !strconcat(OpcodeStr,
3594 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3597 // There's an AssertZext in the way of writing the store pattern
3598 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3601 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3604 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3605 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3606 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3607 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3608 !strconcat(OpcodeStr,
3609 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3612 // There's an AssertZext in the way of writing the store pattern
3613 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3616 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3619 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3620 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3621 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3622 (ins VR128:$src1, i32i8imm:$src2),
3623 !strconcat(OpcodeStr,
3624 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3626 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3627 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3628 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3629 !strconcat(OpcodeStr,
3630 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3631 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3632 addr:$dst)]>, OpSize;
3635 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3638 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3640 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3641 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3642 (ins VR128:$src1, i32i8imm:$src2),
3643 !strconcat(OpcodeStr,
3644 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3646 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3648 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3649 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3650 !strconcat(OpcodeStr,
3651 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3652 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3653 addr:$dst)]>, OpSize;
3656 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3658 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3659 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3662 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3663 Requires<[HasSSE41]>;
3665 let Constraints = "$src1 = $dst" in {
3666 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3667 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3668 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3669 !strconcat(OpcodeStr,
3670 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3672 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3673 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3674 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3675 !strconcat(OpcodeStr,
3676 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3678 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3679 imm:$src3))]>, OpSize;
3683 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3685 let Constraints = "$src1 = $dst" in {
3686 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3687 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3688 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3689 !strconcat(OpcodeStr,
3690 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3692 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3694 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3695 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3696 !strconcat(OpcodeStr,
3697 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3699 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3700 imm:$src3)))]>, OpSize;
3704 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3706 // insertps has a few different modes, there's the first two here below which
3707 // are optimized inserts that won't zero arbitrary elements in the destination
3708 // vector. The next one matches the intrinsic and could zero arbitrary elements
3709 // in the target vector.
3710 let Constraints = "$src1 = $dst" in {
3711 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3712 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3713 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3714 !strconcat(OpcodeStr,
3715 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3717 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3719 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3720 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3721 !strconcat(OpcodeStr,
3722 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3724 (X86insrtps VR128:$src1,
3725 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3726 imm:$src3))]>, OpSize;
3730 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3732 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3733 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3735 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3736 // the intel intrinsic that corresponds to this.
3737 let Defs = [EFLAGS] in {
3738 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3739 "ptest \t{$src2, $src1|$src1, $src2}",
3740 [(X86ptest VR128:$src1, VR128:$src2),
3741 (implicit EFLAGS)]>, OpSize;
3742 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3743 "ptest \t{$src2, $src1|$src1, $src2}",
3744 [(X86ptest VR128:$src1, (load addr:$src2)),
3745 (implicit EFLAGS)]>, OpSize;
3748 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3749 "movntdqa\t{$src, $dst|$dst, $src}",
3750 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;
3753 //===----------------------------------------------------------------------===//
3754 // SSE4.2 Instructions
3755 //===----------------------------------------------------------------------===//
3757 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3758 let Constraints = "$src1 = $dst" in {
3759 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3760 Intrinsic IntId128, bit Commutable = 0> {
3761 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3762 (ins VR128:$src1, VR128:$src2),
3763 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3764 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3766 let isCommutable = Commutable;
3768 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3769 (ins VR128:$src1, i128mem:$src2),
3770 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3772 (IntId128 VR128:$src1,
3773 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3777 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3779 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3780 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3781 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3782 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3784 // crc intrinsic instruction
3785 // This set of instructions are only rm, the only difference is the size
3787 let Constraints = "$src1 = $dst" in {
3788 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3789 (ins GR32:$src1, i8mem:$src2),
3790 "crc32 \t{$src2, $src1|$src1, $src2}",
3792 (int_x86_sse42_crc32_8 GR32:$src1,
3793 (load addr:$src2)))]>, OpSize;
3794 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3795 (ins GR32:$src1, GR8:$src2),
3796 "crc32 \t{$src2, $src1|$src1, $src2}",
3798 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>,
3800 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3801 (ins GR32:$src1, i16mem:$src2),
3802 "crc32 \t{$src2, $src1|$src1, $src2}",
3804 (int_x86_sse42_crc32_16 GR32:$src1,
3805 (load addr:$src2)))]>,
3807 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3808 (ins GR32:$src1, GR16:$src2),
3809 "crc32 \t{$src2, $src1|$src1, $src2}",
3811 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3813 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3814 (ins GR32:$src1, i32mem:$src2),
3815 "crc32 \t{$src2, $src1|$src1, $src2}",
3817 (int_x86_sse42_crc32_32 GR32:$src1,
3818 (load addr:$src2)))]>, OpSize;
3819 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3820 (ins GR32:$src1, GR32:$src2),
3821 "crc32 \t{$src2, $src1|$src1, $src2}",
3823 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>,
3825 def CRC64m64 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3826 (ins GR64:$src1, i64mem:$src2),
3827 "crc32 \t{$src2, $src1|$src1, $src2}",
3829 (int_x86_sse42_crc32_64 GR64:$src1,
3830 (load addr:$src2)))]>,
3832 def CRC64r64 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3833 (ins GR64:$src1, GR64:$src2),
3834 "crc32 \t{$src2, $src1|$src1, $src2}",
3836 (int_x86_sse42_crc32_64 GR64:$src1, GR64:$src2))]>,
3840 // String/text processing instructions.
3841 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3842 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3843 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3844 "#PCMPISTRM128rr PSEUDO!",
3845 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3846 imm:$src3))]>, OpSize;
3847 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3848 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3849 "#PCMPISTRM128rm PSEUDO!",
3850 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3851 imm:$src3))]>, OpSize;
3854 let Defs = [XMM0, EFLAGS] in {
3855 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3856 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3857 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3858 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3859 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3860 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3863 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3864 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3865 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3866 "#PCMPESTRM128rr PSEUDO!",
3868 (int_x86_sse42_pcmpestrm128
3869 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3871 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3872 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3873 "#PCMPESTRM128rm PSEUDO!",
3874 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3875 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3879 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3880 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3881 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3882 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3883 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3884 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3885 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3888 let Defs = [ECX, EFLAGS] in {
3889 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3890 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3891 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3892 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3893 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3894 (implicit EFLAGS)]>, OpSize;
3895 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3896 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3897 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3898 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3899 (implicit EFLAGS)]>, OpSize;
3903 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3904 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3905 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3906 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3907 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3908 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3910 let Defs = [ECX, EFLAGS] in {
3911 let Uses = [EAX, EDX] in {
3912 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3913 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3914 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3915 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3916 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3917 (implicit EFLAGS)]>, OpSize;
3918 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3919 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3920 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3922 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
3923 (implicit EFLAGS)]>, OpSize;
3928 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3929 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3930 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3931 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3932 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3933 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;