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:$src2)),
502 // Aliases of packed SSE1 instructions for scalar use. These all have names
503 // that start with 'Fs'.
505 // Alias instructions that map fld0 to pxor for sse.
506 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
508 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
509 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
510 Requires<[HasSSE1]>, TB, OpSize;
512 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
514 let neverHasSideEffects = 1 in
515 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
516 "movaps\t{$src, $dst|$dst, $src}", []>;
518 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
520 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
521 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
522 "movaps\t{$src, $dst|$dst, $src}",
523 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
525 // Alias bitwise logical operations using SSE logical ops on packed FP values.
526 let Constraints = "$src1 = $dst" in {
527 let isCommutable = 1 in {
528 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst),
529 (ins FR32:$src1, FR32:$src2),
530 "andps\t{$src2, $dst|$dst, $src2}",
531 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
532 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst),
533 (ins FR32:$src1, FR32:$src2),
534 "orps\t{$src2, $dst|$dst, $src2}",
535 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
536 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst),
537 (ins FR32:$src1, FR32:$src2),
538 "xorps\t{$src2, $dst|$dst, $src2}",
539 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
542 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst),
543 (ins FR32:$src1, f128mem:$src2),
544 "andps\t{$src2, $dst|$dst, $src2}",
545 [(set FR32:$dst, (X86fand FR32:$src1,
546 (memopfsf32 addr:$src2)))]>;
547 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst),
548 (ins FR32:$src1, f128mem:$src2),
549 "orps\t{$src2, $dst|$dst, $src2}",
550 [(set FR32:$dst, (X86for FR32:$src1,
551 (memopfsf32 addr:$src2)))]>;
552 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst),
553 (ins FR32:$src1, f128mem:$src2),
554 "xorps\t{$src2, $dst|$dst, $src2}",
555 [(set FR32:$dst, (X86fxor FR32:$src1,
556 (memopfsf32 addr:$src2)))]>;
558 let neverHasSideEffects = 1 in {
559 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
560 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
561 "andnps\t{$src2, $dst|$dst, $src2}", []>;
563 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
564 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
565 "andnps\t{$src2, $dst|$dst, $src2}", []>;
569 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
571 /// In addition, we also have a special variant of the scalar form here to
572 /// represent the associated intrinsic operation. This form is unlike the
573 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
574 /// and leaves the top elements unmodified (therefore these cannot be commuted).
576 /// These three forms can each be reg+reg or reg+mem, so there are a total of
577 /// six "instructions".
579 let Constraints = "$src1 = $dst" in {
580 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
581 SDNode OpNode, Intrinsic F32Int,
582 bit Commutable = 0> {
583 // Scalar operation, reg+reg.
584 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
585 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
586 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
587 let isCommutable = Commutable;
590 // Scalar operation, reg+mem.
591 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
592 (ins FR32:$src1, f32mem:$src2),
593 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
594 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
596 // Vector operation, reg+reg.
597 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
598 (ins VR128:$src1, VR128:$src2),
599 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
600 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
601 let isCommutable = Commutable;
604 // Vector operation, reg+mem.
605 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
606 (ins VR128:$src1, f128mem:$src2),
607 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
608 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
610 // Intrinsic operation, reg+reg.
611 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
612 (ins VR128:$src1, VR128:$src2),
613 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
614 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]>;
616 // Intrinsic operation, reg+mem.
617 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
618 (ins VR128:$src1, ssmem:$src2),
619 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
620 [(set VR128:$dst, (F32Int VR128:$src1,
621 sse_load_f32:$src2))]>;
625 // Arithmetic instructions
626 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
627 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
628 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
629 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
631 /// sse1_fp_binop_rm - Other SSE1 binops
633 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
634 /// instructions for a full-vector intrinsic form. Operations that map
635 /// onto C operators don't use this form since they just use the plain
636 /// vector form instead of having a separate vector intrinsic form.
638 /// This provides a total of eight "instructions".
640 let Constraints = "$src1 = $dst" in {
641 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
645 bit Commutable = 0> {
647 // Scalar operation, reg+reg.
648 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
649 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
650 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
651 let isCommutable = Commutable;
654 // Scalar operation, reg+mem.
655 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
656 (ins FR32:$src1, f32mem:$src2),
657 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
658 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
660 // Vector operation, reg+reg.
661 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
662 (ins VR128:$src1, VR128:$src2),
663 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
664 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
665 let isCommutable = Commutable;
668 // Vector operation, reg+mem.
669 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
670 (ins VR128:$src1, f128mem:$src2),
671 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
672 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
674 // Intrinsic operation, reg+reg.
675 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
676 (ins VR128:$src1, VR128:$src2),
677 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
678 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
679 let isCommutable = Commutable;
682 // Intrinsic operation, reg+mem.
683 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
684 (ins VR128:$src1, ssmem:$src2),
685 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
686 [(set VR128:$dst, (F32Int VR128:$src1,
687 sse_load_f32:$src2))]>;
689 // Vector intrinsic operation, reg+reg.
690 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
691 (ins VR128:$src1, VR128:$src2),
692 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
693 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
694 let isCommutable = Commutable;
697 // Vector intrinsic operation, reg+mem.
698 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
699 (ins VR128:$src1, f128mem:$src2),
700 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
701 [(set VR128:$dst, (V4F32Int VR128:$src1, (memopv4f32 addr:$src2)))]>;
705 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
706 int_x86_sse_max_ss, int_x86_sse_max_ps>;
707 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
708 int_x86_sse_min_ss, int_x86_sse_min_ps>;
710 //===----------------------------------------------------------------------===//
711 // SSE packed FP Instructions
714 let neverHasSideEffects = 1 in
715 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
716 "movaps\t{$src, $dst|$dst, $src}", []>;
717 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
718 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
719 "movaps\t{$src, $dst|$dst, $src}",
720 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
722 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
723 "movaps\t{$src, $dst|$dst, $src}",
724 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
726 let neverHasSideEffects = 1 in
727 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
728 "movups\t{$src, $dst|$dst, $src}", []>;
729 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
730 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
731 "movups\t{$src, $dst|$dst, $src}",
732 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
733 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
734 "movups\t{$src, $dst|$dst, $src}",
735 [(store (v4f32 VR128:$src), addr:$dst)]>;
737 // Intrinsic forms of MOVUPS load and store
738 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
739 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
740 "movups\t{$src, $dst|$dst, $src}",
741 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
742 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
743 "movups\t{$src, $dst|$dst, $src}",
744 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
746 let Constraints = "$src1 = $dst" in {
747 let AddedComplexity = 20 in {
748 def MOVLPSrm : PSI<0x12, MRMSrcMem,
749 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
750 "movlps\t{$src2, $dst|$dst, $src2}",
753 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
754 def MOVHPSrm : PSI<0x16, MRMSrcMem,
755 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
756 "movhps\t{$src2, $dst|$dst, $src2}",
758 (movlhps VR128:$src1,
759 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
761 } // Constraints = "$src1 = $dst"
764 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
765 "movlps\t{$src, $dst|$dst, $src}",
766 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
767 (iPTR 0))), addr:$dst)]>;
769 // v2f64 extract element 1 is always custom lowered to unpack high to low
770 // and extract element 0 so the non-store version isn't too horrible.
771 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
772 "movhps\t{$src, $dst|$dst, $src}",
773 [(store (f64 (vector_extract
774 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
775 (undef)), (iPTR 0))), addr:$dst)]>;
777 let Constraints = "$src1 = $dst" in {
778 let AddedComplexity = 20 in {
779 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
780 (ins VR128:$src1, VR128:$src2),
781 "movlhps\t{$src2, $dst|$dst, $src2}",
783 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
785 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
786 (ins VR128:$src1, VR128:$src2),
787 "movhlps\t{$src2, $dst|$dst, $src2}",
789 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
791 } // Constraints = "$src1 = $dst"
793 let AddedComplexity = 20 in {
794 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
795 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
796 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
797 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
804 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
806 /// In addition, we also have a special variant of the scalar form here to
807 /// represent the associated intrinsic operation. This form is unlike the
808 /// plain scalar form, in that it takes an entire vector (instead of a
809 /// scalar) and leaves the top elements undefined.
811 /// And, we have a special variant form for a full-vector intrinsic form.
813 /// These four forms can each have a reg or a mem operand, so there are a
814 /// total of eight "instructions".
816 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
820 bit Commutable = 0> {
821 // Scalar operation, reg.
822 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
823 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
824 [(set FR32:$dst, (OpNode FR32:$src))]> {
825 let isCommutable = Commutable;
828 // Scalar operation, mem.
829 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
830 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
831 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
832 Requires<[HasSSE1, OptForSize]>;
834 // Vector operation, reg.
835 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
836 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
837 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
838 let isCommutable = Commutable;
841 // Vector operation, mem.
842 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
843 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
844 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
846 // Intrinsic operation, reg.
847 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
848 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
849 [(set VR128:$dst, (F32Int VR128:$src))]> {
850 let isCommutable = Commutable;
853 // Intrinsic operation, mem.
854 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
855 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
856 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
858 // Vector intrinsic operation, reg
859 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
860 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
861 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
862 let isCommutable = Commutable;
865 // Vector intrinsic operation, mem
866 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
867 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
868 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
872 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
873 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
875 // Reciprocal approximations. Note that these typically require refinement
876 // in order to obtain suitable precision.
877 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
878 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
879 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
880 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
883 let Constraints = "$src1 = $dst" in {
884 let isCommutable = 1 in {
885 def ANDPSrr : PSI<0x54, MRMSrcReg,
886 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
887 "andps\t{$src2, $dst|$dst, $src2}",
888 [(set VR128:$dst, (v2i64
889 (and VR128:$src1, VR128:$src2)))]>;
890 def ORPSrr : PSI<0x56, MRMSrcReg,
891 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
892 "orps\t{$src2, $dst|$dst, $src2}",
893 [(set VR128:$dst, (v2i64
894 (or VR128:$src1, VR128:$src2)))]>;
895 def XORPSrr : PSI<0x57, MRMSrcReg,
896 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
897 "xorps\t{$src2, $dst|$dst, $src2}",
898 [(set VR128:$dst, (v2i64
899 (xor VR128:$src1, VR128:$src2)))]>;
902 def ANDPSrm : PSI<0x54, MRMSrcMem,
903 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
904 "andps\t{$src2, $dst|$dst, $src2}",
905 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
906 (memopv2i64 addr:$src2)))]>;
907 def ORPSrm : PSI<0x56, MRMSrcMem,
908 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
909 "orps\t{$src2, $dst|$dst, $src2}",
910 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
911 (memopv2i64 addr:$src2)))]>;
912 def XORPSrm : PSI<0x57, MRMSrcMem,
913 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
914 "xorps\t{$src2, $dst|$dst, $src2}",
915 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
916 (memopv2i64 addr:$src2)))]>;
917 def ANDNPSrr : PSI<0x55, MRMSrcReg,
918 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
919 "andnps\t{$src2, $dst|$dst, $src2}",
921 (v2i64 (and (xor VR128:$src1,
922 (bc_v2i64 (v4i32 immAllOnesV))),
924 def ANDNPSrm : PSI<0x55, MRMSrcMem,
925 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
926 "andnps\t{$src2, $dst|$dst, $src2}",
928 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
929 (bc_v2i64 (v4i32 immAllOnesV))),
930 (memopv2i64 addr:$src2))))]>;
933 let Constraints = "$src1 = $dst" in {
934 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
935 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
936 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
937 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
938 VR128:$src, imm:$cc))]>;
939 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
940 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
941 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
942 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
943 (memop addr:$src), imm:$cc))]>;
945 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
946 (CMPPSrri VR128:$src1, VR128:$src2, imm:$cc)>;
947 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
948 (CMPPSrmi VR128:$src1, addr:$src2, imm:$cc)>;
950 // Shuffle and unpack instructions
951 let Constraints = "$src1 = $dst" in {
952 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
953 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
954 (outs VR128:$dst), (ins VR128:$src1,
955 VR128:$src2, i8imm:$src3),
956 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
958 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
959 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
960 (outs VR128:$dst), (ins VR128:$src1,
961 f128mem:$src2, i8imm:$src3),
962 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
965 VR128:$src1, (memopv4f32 addr:$src2))))]>;
967 let AddedComplexity = 10 in {
968 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
969 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
970 "unpckhps\t{$src2, $dst|$dst, $src2}",
972 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
973 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
974 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
975 "unpckhps\t{$src2, $dst|$dst, $src2}",
977 (v4f32 (unpckh VR128:$src1,
978 (memopv4f32 addr:$src2))))]>;
980 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
981 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
982 "unpcklps\t{$src2, $dst|$dst, $src2}",
984 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
985 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
986 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
987 "unpcklps\t{$src2, $dst|$dst, $src2}",
989 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
991 } // Constraints = "$src1 = $dst"
994 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
995 "movmskps\t{$src, $dst|$dst, $src}",
996 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
997 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
998 "movmskpd\t{$src, $dst|$dst, $src}",
999 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1001 // Prefetch intrinsic.
1002 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1003 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1004 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1005 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1006 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1007 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1008 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1009 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1011 // Non-temporal stores
1012 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1013 "movntps\t{$src, $dst|$dst, $src}",
1014 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1016 // Load, store, and memory fence
1017 def SFENCE : PSI<0xAE, MRM7r, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
1020 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1021 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1022 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1023 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1025 // Alias instructions that map zero vector to pxor / xorp* for sse.
1026 // We set canFoldAsLoad because this can be converted to a constant-pool
1027 // load of an all-zeros value if folding it would be beneficial.
1028 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1029 isCodeGenOnly = 1 in
1030 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
1031 "xorps\t$dst, $dst",
1032 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1034 let Predicates = [HasSSE1] in {
1035 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
1036 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
1037 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
1038 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
1039 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
1042 // FR32 to 128-bit vector conversion.
1043 let isAsCheapAsAMove = 1 in
1044 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
1045 "movss\t{$src, $dst|$dst, $src}",
1047 (v4f32 (scalar_to_vector FR32:$src)))]>;
1048 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1049 "movss\t{$src, $dst|$dst, $src}",
1051 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
1053 // FIXME: may not be able to eliminate this movss with coalescing the src and
1054 // dest register classes are different. We really want to write this pattern
1056 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1057 // (f32 FR32:$src)>;
1058 let isAsCheapAsAMove = 1 in
1059 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
1060 "movss\t{$src, $dst|$dst, $src}",
1061 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
1063 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
1064 "movss\t{$src, $dst|$dst, $src}",
1065 [(store (f32 (vector_extract (v4f32 VR128:$src),
1066 (iPTR 0))), addr:$dst)]>;
1069 // Move to lower bits of a VR128, leaving upper bits alone.
1070 // Three operand (but two address) aliases.
1071 let Constraints = "$src1 = $dst" in {
1072 let neverHasSideEffects = 1 in
1073 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
1074 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
1075 "movss\t{$src2, $dst|$dst, $src2}", []>;
1077 let AddedComplexity = 15 in
1078 def MOVLPSrr : SSI<0x10, MRMSrcReg,
1079 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1080 "movss\t{$src2, $dst|$dst, $src2}",
1082 (v4f32 (movl VR128:$src1, VR128:$src2)))]>;
1085 // Move to lower bits of a VR128 and zeroing upper bits.
1086 // Loading from memory automatically zeroing upper bits.
1087 let AddedComplexity = 20 in
1088 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1089 "movss\t{$src, $dst|$dst, $src}",
1090 [(set VR128:$dst, (v4f32 (X86vzmovl (v4f32 (scalar_to_vector
1091 (loadf32 addr:$src))))))]>;
1093 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
1094 (MOVZSS2PSrm addr:$src)>;
1096 //===---------------------------------------------------------------------===//
1097 // SSE2 Instructions
1098 //===---------------------------------------------------------------------===//
1100 // Move Instructions
1101 let neverHasSideEffects = 1 in
1102 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1103 "movsd\t{$src, $dst|$dst, $src}", []>;
1104 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1105 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1106 "movsd\t{$src, $dst|$dst, $src}",
1107 [(set FR64:$dst, (loadf64 addr:$src))]>;
1108 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1109 "movsd\t{$src, $dst|$dst, $src}",
1110 [(store FR64:$src, addr:$dst)]>;
1112 // Conversion instructions
1113 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1114 "cvttsd2si\t{$src, $dst|$dst, $src}",
1115 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1116 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1117 "cvttsd2si\t{$src, $dst|$dst, $src}",
1118 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1119 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1120 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1121 [(set FR32:$dst, (fround FR64:$src))]>;
1122 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1123 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1124 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
1125 Requires<[HasSSE2, OptForSize]>;
1126 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1127 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1128 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1129 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1130 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1131 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1133 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1134 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1135 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1136 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1137 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1138 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1139 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1140 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1141 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1142 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1143 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1144 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1145 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1146 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1147 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1148 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1149 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1150 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1151 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1152 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1154 // SSE2 instructions with XS prefix
1155 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1156 "cvtss2sd\t{$src, $dst|$dst, $src}",
1157 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1158 Requires<[HasSSE2]>;
1159 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1160 "cvtss2sd\t{$src, $dst|$dst, $src}",
1161 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1162 Requires<[HasSSE2, OptForSize]>;
1164 def : Pat<(extloadf32 addr:$src),
1165 (CVTSS2SDrr (MOVSSrm addr:$src))>, Requires<[HasSSE2, OptForSpeed]>;
1167 // Match intrinsics which expect XMM operand(s).
1168 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1169 "cvtsd2si\t{$src, $dst|$dst, $src}",
1170 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1171 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1172 "cvtsd2si\t{$src, $dst|$dst, $src}",
1173 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1174 (load addr:$src)))]>;
1176 // Match intrinisics which expect MM and XMM operand(s).
1177 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1178 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1179 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1180 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1181 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1182 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1183 (memop addr:$src)))]>;
1184 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1185 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1186 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1187 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1188 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1189 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1190 (memop addr:$src)))]>;
1191 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1192 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1193 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1194 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1195 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1196 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1197 (load addr:$src)))]>;
1199 // Aliases for intrinsics
1200 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1201 "cvttsd2si\t{$src, $dst|$dst, $src}",
1203 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1204 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1205 "cvttsd2si\t{$src, $dst|$dst, $src}",
1206 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1207 (load addr:$src)))]>;
1209 // Comparison instructions
1210 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1211 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1212 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1213 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1215 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1216 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1217 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1220 let Defs = [EFLAGS] in {
1221 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1222 "ucomisd\t{$src2, $src1|$src1, $src2}",
1223 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1224 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1225 "ucomisd\t{$src2, $src1|$src1, $src2}",
1226 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1227 (implicit EFLAGS)]>;
1228 } // Defs = [EFLAGS]
1230 // Aliases to match intrinsics which expect XMM operand(s).
1231 let Constraints = "$src1 = $dst" in {
1232 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1234 (ins VR128:$src1, VR128:$src, SSECC:$cc),
1235 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1236 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1237 VR128:$src, imm:$cc))]>;
1238 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1240 (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1241 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1242 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1243 (load addr:$src), imm:$cc))]>;
1246 let Defs = [EFLAGS] in {
1247 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1248 "ucomisd\t{$src2, $src1|$src1, $src2}",
1249 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1250 (implicit EFLAGS)]>;
1251 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1252 "ucomisd\t{$src2, $src1|$src1, $src2}",
1253 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1254 (implicit EFLAGS)]>;
1256 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1257 "comisd\t{$src2, $src1|$src1, $src2}",
1258 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1259 (implicit EFLAGS)]>;
1260 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1261 "comisd\t{$src2, $src1|$src1, $src2}",
1262 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1263 (implicit EFLAGS)]>;
1264 } // Defs = [EFLAGS]
1266 // Aliases of packed SSE2 instructions for scalar use. These all have names
1267 // that start with 'Fs'.
1269 // Alias instructions that map fld0 to pxor for sse.
1270 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1271 canFoldAsLoad = 1 in
1272 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1273 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1274 Requires<[HasSSE2]>, TB, OpSize;
1276 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1278 let neverHasSideEffects = 1 in
1279 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1280 "movapd\t{$src, $dst|$dst, $src}", []>;
1282 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1284 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1285 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1286 "movapd\t{$src, $dst|$dst, $src}",
1287 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1289 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1290 let Constraints = "$src1 = $dst" in {
1291 let isCommutable = 1 in {
1292 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1293 (ins FR64:$src1, FR64:$src2),
1294 "andpd\t{$src2, $dst|$dst, $src2}",
1295 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1296 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1297 (ins FR64:$src1, FR64:$src2),
1298 "orpd\t{$src2, $dst|$dst, $src2}",
1299 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1300 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1301 (ins FR64:$src1, FR64:$src2),
1302 "xorpd\t{$src2, $dst|$dst, $src2}",
1303 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1306 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1307 (ins FR64:$src1, f128mem:$src2),
1308 "andpd\t{$src2, $dst|$dst, $src2}",
1309 [(set FR64:$dst, (X86fand FR64:$src1,
1310 (memopfsf64 addr:$src2)))]>;
1311 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1312 (ins FR64:$src1, f128mem:$src2),
1313 "orpd\t{$src2, $dst|$dst, $src2}",
1314 [(set FR64:$dst, (X86for FR64:$src1,
1315 (memopfsf64 addr:$src2)))]>;
1316 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1317 (ins FR64:$src1, f128mem:$src2),
1318 "xorpd\t{$src2, $dst|$dst, $src2}",
1319 [(set FR64:$dst, (X86fxor FR64:$src1,
1320 (memopfsf64 addr:$src2)))]>;
1322 let neverHasSideEffects = 1 in {
1323 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1324 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1325 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1327 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1328 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1329 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1333 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1335 /// In addition, we also have a special variant of the scalar form here to
1336 /// represent the associated intrinsic operation. This form is unlike the
1337 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1338 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1340 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1341 /// six "instructions".
1343 let Constraints = "$src1 = $dst" in {
1344 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1345 SDNode OpNode, Intrinsic F64Int,
1346 bit Commutable = 0> {
1347 // Scalar operation, reg+reg.
1348 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1349 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1350 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1351 let isCommutable = Commutable;
1354 // Scalar operation, reg+mem.
1355 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1356 (ins FR64:$src1, f64mem:$src2),
1357 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1358 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1360 // Vector operation, reg+reg.
1361 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1362 (ins VR128:$src1, VR128:$src2),
1363 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1364 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1365 let isCommutable = Commutable;
1368 // Vector operation, reg+mem.
1369 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1370 (ins VR128:$src1, f128mem:$src2),
1371 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1372 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1374 // Intrinsic operation, reg+reg.
1375 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1376 (ins VR128:$src1, VR128:$src2),
1377 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1378 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]>;
1380 // Intrinsic operation, reg+mem.
1381 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1382 (ins VR128:$src1, sdmem:$src2),
1383 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1384 [(set VR128:$dst, (F64Int VR128:$src1,
1385 sse_load_f64:$src2))]>;
1389 // Arithmetic instructions
1390 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1391 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1392 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1393 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1395 /// sse2_fp_binop_rm - Other SSE2 binops
1397 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1398 /// instructions for a full-vector intrinsic form. Operations that map
1399 /// onto C operators don't use this form since they just use the plain
1400 /// vector form instead of having a separate vector intrinsic form.
1402 /// This provides a total of eight "instructions".
1404 let Constraints = "$src1 = $dst" in {
1405 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1409 bit Commutable = 0> {
1411 // Scalar operation, reg+reg.
1412 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1413 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1414 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1415 let isCommutable = Commutable;
1418 // Scalar operation, reg+mem.
1419 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1420 (ins FR64:$src1, f64mem:$src2),
1421 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1422 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1424 // Vector operation, reg+reg.
1425 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1426 (ins VR128:$src1, VR128:$src2),
1427 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1428 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1429 let isCommutable = Commutable;
1432 // Vector operation, reg+mem.
1433 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1434 (ins VR128:$src1, f128mem:$src2),
1435 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1436 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1438 // Intrinsic operation, reg+reg.
1439 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1440 (ins VR128:$src1, VR128:$src2),
1441 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1442 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1443 let isCommutable = Commutable;
1446 // Intrinsic operation, reg+mem.
1447 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1448 (ins VR128:$src1, sdmem:$src2),
1449 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1450 [(set VR128:$dst, (F64Int VR128:$src1,
1451 sse_load_f64:$src2))]>;
1453 // Vector intrinsic operation, reg+reg.
1454 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1455 (ins VR128:$src1, VR128:$src2),
1456 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1457 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1458 let isCommutable = Commutable;
1461 // Vector intrinsic operation, reg+mem.
1462 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1463 (ins VR128:$src1, f128mem:$src2),
1464 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1465 [(set VR128:$dst, (V2F64Int VR128:$src1,
1466 (memopv2f64 addr:$src2)))]>;
1470 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1471 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1472 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1473 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1475 //===---------------------------------------------------------------------===//
1476 // SSE packed FP Instructions
1478 // Move Instructions
1479 let neverHasSideEffects = 1 in
1480 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1481 "movapd\t{$src, $dst|$dst, $src}", []>;
1482 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1483 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1484 "movapd\t{$src, $dst|$dst, $src}",
1485 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1487 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1488 "movapd\t{$src, $dst|$dst, $src}",
1489 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1491 let neverHasSideEffects = 1 in
1492 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1493 "movupd\t{$src, $dst|$dst, $src}", []>;
1494 let canFoldAsLoad = 1 in
1495 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1496 "movupd\t{$src, $dst|$dst, $src}",
1497 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1498 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1499 "movupd\t{$src, $dst|$dst, $src}",
1500 [(store (v2f64 VR128:$src), addr:$dst)]>;
1502 // Intrinsic forms of MOVUPD load and store
1503 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1504 "movupd\t{$src, $dst|$dst, $src}",
1505 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1506 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1507 "movupd\t{$src, $dst|$dst, $src}",
1508 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1510 let Constraints = "$src1 = $dst" in {
1511 let AddedComplexity = 20 in {
1512 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1513 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1514 "movlpd\t{$src2, $dst|$dst, $src2}",
1516 (v2f64 (movlp VR128:$src1,
1517 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1518 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1519 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1520 "movhpd\t{$src2, $dst|$dst, $src2}",
1522 (v2f64 (movlhps VR128:$src1,
1523 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1524 } // AddedComplexity
1525 } // Constraints = "$src1 = $dst"
1527 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1528 "movlpd\t{$src, $dst|$dst, $src}",
1529 [(store (f64 (vector_extract (v2f64 VR128:$src),
1530 (iPTR 0))), addr:$dst)]>;
1532 // v2f64 extract element 1 is always custom lowered to unpack high to low
1533 // and extract element 0 so the non-store version isn't too horrible.
1534 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1535 "movhpd\t{$src, $dst|$dst, $src}",
1536 [(store (f64 (vector_extract
1537 (v2f64 (unpckh VR128:$src, (undef))),
1538 (iPTR 0))), addr:$dst)]>;
1540 // SSE2 instructions without OpSize prefix
1541 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1542 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1543 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1544 TB, Requires<[HasSSE2]>;
1545 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1546 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1547 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1548 (bitconvert (memopv2i64 addr:$src))))]>,
1549 TB, Requires<[HasSSE2]>;
1551 // SSE2 instructions with XS prefix
1552 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1553 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1554 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1555 XS, Requires<[HasSSE2]>;
1556 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1557 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1558 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1559 (bitconvert (memopv2i64 addr:$src))))]>,
1560 XS, Requires<[HasSSE2]>;
1562 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1563 "cvtps2dq\t{$src, $dst|$dst, $src}",
1564 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1565 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1566 "cvtps2dq\t{$src, $dst|$dst, $src}",
1567 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1568 (memop addr:$src)))]>;
1569 // SSE2 packed instructions with XS prefix
1570 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1571 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1572 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1573 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1575 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1576 "cvttps2dq\t{$src, $dst|$dst, $src}",
1578 (int_x86_sse2_cvttps2dq VR128:$src))]>,
1579 XS, Requires<[HasSSE2]>;
1580 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1581 "cvttps2dq\t{$src, $dst|$dst, $src}",
1582 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1583 (memop addr:$src)))]>,
1584 XS, Requires<[HasSSE2]>;
1586 // SSE2 packed instructions with XD prefix
1587 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1588 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1589 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1590 XD, Requires<[HasSSE2]>;
1591 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1592 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1593 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1594 (memop addr:$src)))]>,
1595 XD, Requires<[HasSSE2]>;
1597 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1598 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1599 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1600 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1601 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1602 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1603 (memop addr:$src)))]>;
1605 // SSE2 instructions without OpSize prefix
1606 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1607 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1608 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1609 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1611 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1612 "cvtps2pd\t{$src, $dst|$dst, $src}",
1613 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1614 TB, Requires<[HasSSE2]>;
1615 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1616 "cvtps2pd\t{$src, $dst|$dst, $src}",
1617 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1618 (load addr:$src)))]>,
1619 TB, Requires<[HasSSE2]>;
1621 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1622 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1623 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1624 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1627 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1628 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1629 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1630 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1631 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1632 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1633 (memop addr:$src)))]>;
1635 // Match intrinsics which expect XMM operand(s).
1636 // Aliases for intrinsics
1637 let Constraints = "$src1 = $dst" in {
1638 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1639 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1640 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1641 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1643 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1644 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1645 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1646 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1647 (loadi32 addr:$src2)))]>;
1648 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1649 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1650 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1651 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1653 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1654 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1655 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1656 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1657 (load addr:$src2)))]>;
1658 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1659 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1660 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1661 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1662 VR128:$src2))]>, XS,
1663 Requires<[HasSSE2]>;
1664 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1665 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1666 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1667 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1668 (load addr:$src2)))]>, XS,
1669 Requires<[HasSSE2]>;
1674 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1676 /// In addition, we also have a special variant of the scalar form here to
1677 /// represent the associated intrinsic operation. This form is unlike the
1678 /// plain scalar form, in that it takes an entire vector (instead of a
1679 /// scalar) and leaves the top elements undefined.
1681 /// And, we have a special variant form for a full-vector intrinsic form.
1683 /// These four forms can each have a reg or a mem operand, so there are a
1684 /// total of eight "instructions".
1686 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1690 bit Commutable = 0> {
1691 // Scalar operation, reg.
1692 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1693 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1694 [(set FR64:$dst, (OpNode FR64:$src))]> {
1695 let isCommutable = Commutable;
1698 // Scalar operation, mem.
1699 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1700 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1701 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1703 // Vector operation, reg.
1704 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1705 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1706 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1707 let isCommutable = Commutable;
1710 // Vector operation, mem.
1711 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1712 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1713 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1715 // Intrinsic operation, reg.
1716 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1717 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1718 [(set VR128:$dst, (F64Int VR128:$src))]> {
1719 let isCommutable = Commutable;
1722 // Intrinsic operation, mem.
1723 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1724 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1725 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1727 // Vector intrinsic operation, reg
1728 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1729 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1730 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1731 let isCommutable = Commutable;
1734 // Vector intrinsic operation, mem
1735 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1736 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1737 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1741 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1742 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1744 // There is no f64 version of the reciprocal approximation instructions.
1747 let Constraints = "$src1 = $dst" in {
1748 let isCommutable = 1 in {
1749 def ANDPDrr : PDI<0x54, MRMSrcReg,
1750 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1751 "andpd\t{$src2, $dst|$dst, $src2}",
1753 (and (bc_v2i64 (v2f64 VR128:$src1)),
1754 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1755 def ORPDrr : PDI<0x56, MRMSrcReg,
1756 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1757 "orpd\t{$src2, $dst|$dst, $src2}",
1759 (or (bc_v2i64 (v2f64 VR128:$src1)),
1760 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1761 def XORPDrr : PDI<0x57, MRMSrcReg,
1762 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1763 "xorpd\t{$src2, $dst|$dst, $src2}",
1765 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1766 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1769 def ANDPDrm : PDI<0x54, MRMSrcMem,
1770 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1771 "andpd\t{$src2, $dst|$dst, $src2}",
1773 (and (bc_v2i64 (v2f64 VR128:$src1)),
1774 (memopv2i64 addr:$src2)))]>;
1775 def ORPDrm : PDI<0x56, MRMSrcMem,
1776 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1777 "orpd\t{$src2, $dst|$dst, $src2}",
1779 (or (bc_v2i64 (v2f64 VR128:$src1)),
1780 (memopv2i64 addr:$src2)))]>;
1781 def XORPDrm : PDI<0x57, MRMSrcMem,
1782 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1783 "xorpd\t{$src2, $dst|$dst, $src2}",
1785 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1786 (memopv2i64 addr:$src2)))]>;
1787 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1788 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1789 "andnpd\t{$src2, $dst|$dst, $src2}",
1791 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1792 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1793 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1794 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1795 "andnpd\t{$src2, $dst|$dst, $src2}",
1797 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1798 (memopv2i64 addr:$src2)))]>;
1801 let Constraints = "$src1 = $dst" in {
1802 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1803 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1804 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1805 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1806 VR128:$src, imm:$cc))]>;
1807 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1808 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1809 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1810 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1811 (memop addr:$src), imm:$cc))]>;
1813 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1814 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1815 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1816 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1818 // Shuffle and unpack instructions
1819 let Constraints = "$src1 = $dst" in {
1820 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1821 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1822 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1824 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1825 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1826 (outs VR128:$dst), (ins VR128:$src1,
1827 f128mem:$src2, i8imm:$src3),
1828 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1831 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1833 let AddedComplexity = 10 in {
1834 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1835 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1836 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1838 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1839 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1840 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1841 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1843 (v2f64 (unpckh VR128:$src1,
1844 (memopv2f64 addr:$src2))))]>;
1846 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1847 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1848 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1850 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1851 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1852 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1853 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1855 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1856 } // AddedComplexity
1857 } // Constraints = "$src1 = $dst"
1860 //===---------------------------------------------------------------------===//
1861 // SSE integer instructions
1863 // Move Instructions
1864 let neverHasSideEffects = 1 in
1865 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1866 "movdqa\t{$src, $dst|$dst, $src}", []>;
1867 let canFoldAsLoad = 1, mayLoad = 1 in
1868 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1869 "movdqa\t{$src, $dst|$dst, $src}",
1870 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1872 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1873 "movdqa\t{$src, $dst|$dst, $src}",
1874 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1875 let canFoldAsLoad = 1, mayLoad = 1 in
1876 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1877 "movdqu\t{$src, $dst|$dst, $src}",
1878 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1879 XS, Requires<[HasSSE2]>;
1881 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1882 "movdqu\t{$src, $dst|$dst, $src}",
1883 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1884 XS, Requires<[HasSSE2]>;
1886 // Intrinsic forms of MOVDQU load and store
1887 let canFoldAsLoad = 1 in
1888 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1889 "movdqu\t{$src, $dst|$dst, $src}",
1890 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1891 XS, Requires<[HasSSE2]>;
1892 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1893 "movdqu\t{$src, $dst|$dst, $src}",
1894 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1895 XS, Requires<[HasSSE2]>;
1897 let Constraints = "$src1 = $dst" in {
1899 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1900 bit Commutable = 0> {
1901 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1902 (ins VR128:$src1, VR128:$src2),
1903 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1904 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1905 let isCommutable = Commutable;
1907 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1908 (ins VR128:$src1, i128mem:$src2),
1909 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1910 [(set VR128:$dst, (IntId VR128:$src1,
1911 (bitconvert (memopv2i64
1915 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1917 Intrinsic IntId, Intrinsic IntId2> {
1918 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1919 (ins VR128:$src1, VR128:$src2),
1920 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1921 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1922 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1923 (ins VR128:$src1, i128mem:$src2),
1924 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1925 [(set VR128:$dst, (IntId VR128:$src1,
1926 (bitconvert (memopv2i64 addr:$src2))))]>;
1927 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
1928 (ins VR128:$src1, i32i8imm:$src2),
1929 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1930 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1933 /// PDI_binop_rm - Simple SSE2 binary operator.
1934 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1935 ValueType OpVT, bit Commutable = 0> {
1936 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1937 (ins VR128:$src1, VR128:$src2),
1938 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1939 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1940 let isCommutable = Commutable;
1942 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1943 (ins VR128:$src1, i128mem:$src2),
1944 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1945 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1946 (bitconvert (memopv2i64 addr:$src2)))))]>;
1949 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1951 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1952 /// to collapse (bitconvert VT to VT) into its operand.
1954 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1955 bit Commutable = 0> {
1956 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1957 (ins VR128:$src1, VR128:$src2),
1958 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1959 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1960 let isCommutable = Commutable;
1962 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1963 (ins VR128:$src1, i128mem:$src2),
1964 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1965 [(set VR128:$dst, (OpNode VR128:$src1,
1966 (memopv2i64 addr:$src2)))]>;
1969 } // Constraints = "$src1 = $dst"
1971 // 128-bit Integer Arithmetic
1973 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1974 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1975 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1976 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1978 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1979 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1980 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1981 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1983 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1984 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1985 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1986 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1988 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1989 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1990 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1991 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1993 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1995 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1996 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1997 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1999 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
2001 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
2002 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
2005 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
2006 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
2007 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
2008 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
2009 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
2012 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
2013 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
2014 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
2015 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
2016 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
2017 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
2019 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2020 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2021 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2022 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2023 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2024 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2026 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2027 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2028 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2029 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2031 // 128-bit logical shifts.
2032 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
2033 def PSLLDQri : PDIi8<0x73, MRM7r,
2034 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2035 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2036 def PSRLDQri : PDIi8<0x73, MRM3r,
2037 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2038 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2039 // PSRADQri doesn't exist in SSE[1-3].
2042 let Predicates = [HasSSE2] in {
2043 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2044 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2045 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2046 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2047 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2048 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2049 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2050 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2051 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2052 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2054 // Shift up / down and insert zero's.
2055 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2056 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2057 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2058 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2062 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2063 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2064 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2066 let Constraints = "$src1 = $dst" in {
2067 def PANDNrr : PDI<0xDF, MRMSrcReg,
2068 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2069 "pandn\t{$src2, $dst|$dst, $src2}",
2070 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2073 def PANDNrm : PDI<0xDF, MRMSrcMem,
2074 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2075 "pandn\t{$src2, $dst|$dst, $src2}",
2076 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2077 (memopv2i64 addr:$src2))))]>;
2080 // SSE2 Integer comparison
2081 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2082 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2083 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2084 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2085 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2086 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2088 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2089 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2090 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2091 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2092 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2093 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2094 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2095 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2096 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2097 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2098 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2099 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2101 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2102 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2103 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2104 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2105 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2106 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2107 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2108 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2109 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2110 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2111 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2112 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2115 // Pack instructions
2116 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2117 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2118 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2120 // Shuffle and unpack instructions
2121 let AddedComplexity = 5 in {
2122 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2123 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2124 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2125 [(set VR128:$dst, (v4i32 (pshufd:$src2
2126 VR128:$src1, (undef))))]>;
2127 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2128 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2129 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2130 [(set VR128:$dst, (v4i32 (pshufd:$src2
2131 (bc_v4i32 (memopv2i64 addr:$src1)),
2135 // SSE2 with ImmT == Imm8 and XS prefix.
2136 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2137 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2138 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2139 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2141 XS, Requires<[HasSSE2]>;
2142 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2143 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2144 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2145 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2146 (bc_v8i16 (memopv2i64 addr:$src1)),
2148 XS, Requires<[HasSSE2]>;
2150 // SSE2 with ImmT == Imm8 and XD prefix.
2151 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2152 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2153 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2154 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2156 XD, Requires<[HasSSE2]>;
2157 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2158 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2159 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2160 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2161 (bc_v8i16 (memopv2i64 addr:$src1)),
2163 XD, Requires<[HasSSE2]>;
2166 let Constraints = "$src1 = $dst" in {
2167 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2168 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2169 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2171 (v16i8 (unpckl VR128:$src1, VR128:$src2)))]>;
2172 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2173 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2174 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2176 (unpckl VR128:$src1,
2177 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2178 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2179 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2180 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2182 (v8i16 (unpckl VR128:$src1, VR128:$src2)))]>;
2183 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2184 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2185 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2187 (unpckl VR128:$src1,
2188 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2189 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2190 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2191 "punpckldq\t{$src2, $dst|$dst, $src2}",
2193 (v4i32 (unpckl VR128:$src1, VR128:$src2)))]>;
2194 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2195 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2196 "punpckldq\t{$src2, $dst|$dst, $src2}",
2198 (unpckl VR128:$src1,
2199 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2200 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2201 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2202 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2204 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2205 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2206 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2207 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2209 (v2i64 (unpckl VR128:$src1,
2210 (memopv2i64 addr:$src2))))]>;
2212 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2213 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2214 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2216 (v16i8 (unpckh VR128:$src1, VR128:$src2)))]>;
2217 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2218 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2219 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2221 (unpckh VR128:$src1,
2222 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2223 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2224 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2225 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2227 (v8i16 (unpckh VR128:$src1, VR128:$src2)))]>;
2228 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2229 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2230 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2232 (unpckh VR128:$src1,
2233 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2234 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2235 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2236 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2238 (v4i32 (unpckh VR128:$src1, VR128:$src2)))]>;
2239 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2240 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2241 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2243 (unpckh VR128:$src1,
2244 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2245 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2246 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2247 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2249 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2250 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2251 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2252 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2254 (v2i64 (unpckh VR128:$src1,
2255 (memopv2i64 addr:$src2))))]>;
2259 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2260 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2261 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2262 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2264 let Constraints = "$src1 = $dst" in {
2265 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2266 (outs VR128:$dst), (ins VR128:$src1,
2267 GR32:$src2, i32i8imm:$src3),
2268 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2270 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2271 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2272 (outs VR128:$dst), (ins VR128:$src1,
2273 i16mem:$src2, i32i8imm:$src3),
2274 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2276 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2281 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2282 "pmovmskb\t{$src, $dst|$dst, $src}",
2283 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2285 // Conditional store
2287 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2288 "maskmovdqu\t{$mask, $src|$src, $mask}",
2289 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2292 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2293 "maskmovdqu\t{$mask, $src|$src, $mask}",
2294 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2296 // Non-temporal stores
2297 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2298 "movntpd\t{$src, $dst|$dst, $src}",
2299 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2300 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2301 "movntdq\t{$src, $dst|$dst, $src}",
2302 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2303 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2304 "movnti\t{$src, $dst|$dst, $src}",
2305 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2306 TB, Requires<[HasSSE2]>;
2309 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2310 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2311 TB, Requires<[HasSSE2]>;
2313 // Load, store, and memory fence
2314 def LFENCE : I<0xAE, MRM5r, (outs), (ins),
2315 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2316 def MFENCE : I<0xAE, MRM6r, (outs), (ins),
2317 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2319 //TODO: custom lower this so as to never even generate the noop
2320 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2322 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2323 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2324 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2327 // Alias instructions that map zero vector to pxor / xorp* for sse.
2328 // We set canFoldAsLoad because this can be converted to a constant-pool
2329 // load of an all-ones value if folding it would be beneficial.
2330 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2331 isCodeGenOnly = 1 in
2332 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2333 "pcmpeqd\t$dst, $dst",
2334 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2336 // FR64 to 128-bit vector conversion.
2337 let isAsCheapAsAMove = 1 in
2338 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2339 "movsd\t{$src, $dst|$dst, $src}",
2341 (v2f64 (scalar_to_vector FR64:$src)))]>;
2342 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2343 "movsd\t{$src, $dst|$dst, $src}",
2345 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2347 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2348 "movd\t{$src, $dst|$dst, $src}",
2350 (v4i32 (scalar_to_vector GR32:$src)))]>;
2351 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2352 "movd\t{$src, $dst|$dst, $src}",
2354 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2356 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2357 "movd\t{$src, $dst|$dst, $src}",
2358 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2360 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2361 "movd\t{$src, $dst|$dst, $src}",
2362 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2364 // SSE2 instructions with XS prefix
2365 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2366 "movq\t{$src, $dst|$dst, $src}",
2368 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2369 Requires<[HasSSE2]>;
2370 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2371 "movq\t{$src, $dst|$dst, $src}",
2372 [(store (i64 (vector_extract (v2i64 VR128:$src),
2373 (iPTR 0))), addr:$dst)]>;
2375 // FIXME: may not be able to eliminate this movss with coalescing the src and
2376 // dest register classes are different. We really want to write this pattern
2378 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2379 // (f32 FR32:$src)>;
2380 let isAsCheapAsAMove = 1 in
2381 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2382 "movsd\t{$src, $dst|$dst, $src}",
2383 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2385 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2386 "movsd\t{$src, $dst|$dst, $src}",
2387 [(store (f64 (vector_extract (v2f64 VR128:$src),
2388 (iPTR 0))), addr:$dst)]>;
2389 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2390 "movd\t{$src, $dst|$dst, $src}",
2391 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2393 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2394 "movd\t{$src, $dst|$dst, $src}",
2395 [(store (i32 (vector_extract (v4i32 VR128:$src),
2396 (iPTR 0))), addr:$dst)]>;
2398 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2399 "movd\t{$src, $dst|$dst, $src}",
2400 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2401 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2402 "movd\t{$src, $dst|$dst, $src}",
2403 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2406 // Move to lower bits of a VR128, leaving upper bits alone.
2407 // Three operand (but two address) aliases.
2408 let Constraints = "$src1 = $dst" in {
2409 let neverHasSideEffects = 1 in
2410 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2411 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2412 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2414 let AddedComplexity = 15 in
2415 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2416 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2417 "movsd\t{$src2, $dst|$dst, $src2}",
2419 (v2f64 (movl VR128:$src1, VR128:$src2)))]>;
2422 // Store / copy lower 64-bits of a XMM register.
2423 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2424 "movq\t{$src, $dst|$dst, $src}",
2425 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2427 // Move to lower bits of a VR128 and zeroing upper bits.
2428 // Loading from memory automatically zeroing upper bits.
2429 let AddedComplexity = 20 in {
2430 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2431 "movsd\t{$src, $dst|$dst, $src}",
2433 (v2f64 (X86vzmovl (v2f64 (scalar_to_vector
2434 (loadf64 addr:$src))))))]>;
2436 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
2437 (MOVZSD2PDrm addr:$src)>;
2438 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
2439 (MOVZSD2PDrm addr:$src)>;
2440 def : Pat<(v2f64 (X86vzload addr:$src)), (MOVZSD2PDrm addr:$src)>;
2443 // movd / movq to XMM register zero-extends
2444 let AddedComplexity = 15 in {
2445 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2446 "movd\t{$src, $dst|$dst, $src}",
2447 [(set VR128:$dst, (v4i32 (X86vzmovl
2448 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2449 // This is X86-64 only.
2450 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2451 "mov{d|q}\t{$src, $dst|$dst, $src}",
2452 [(set VR128:$dst, (v2i64 (X86vzmovl
2453 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2456 let AddedComplexity = 20 in {
2457 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2458 "movd\t{$src, $dst|$dst, $src}",
2460 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2461 (loadi32 addr:$src))))))]>;
2463 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2464 (MOVZDI2PDIrm addr:$src)>;
2465 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2466 (MOVZDI2PDIrm addr:$src)>;
2467 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2468 (MOVZDI2PDIrm addr:$src)>;
2470 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2471 "movq\t{$src, $dst|$dst, $src}",
2473 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2474 (loadi64 addr:$src))))))]>, XS,
2475 Requires<[HasSSE2]>;
2477 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2478 (MOVZQI2PQIrm addr:$src)>;
2479 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2480 (MOVZQI2PQIrm addr:$src)>;
2481 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2484 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2485 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2486 let AddedComplexity = 15 in
2487 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2488 "movq\t{$src, $dst|$dst, $src}",
2489 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2490 XS, Requires<[HasSSE2]>;
2492 let AddedComplexity = 20 in {
2493 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2494 "movq\t{$src, $dst|$dst, $src}",
2495 [(set VR128:$dst, (v2i64 (X86vzmovl
2496 (loadv2i64 addr:$src))))]>,
2497 XS, Requires<[HasSSE2]>;
2499 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2500 (MOVZPQILo2PQIrm addr:$src)>;
2503 // Instructions for the disassembler
2504 // xr = XMM register
2507 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2508 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2510 //===---------------------------------------------------------------------===//
2511 // SSE3 Instructions
2512 //===---------------------------------------------------------------------===//
2514 // Move Instructions
2515 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2516 "movshdup\t{$src, $dst|$dst, $src}",
2517 [(set VR128:$dst, (v4f32 (movshdup
2518 VR128:$src, (undef))))]>;
2519 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2520 "movshdup\t{$src, $dst|$dst, $src}",
2521 [(set VR128:$dst, (movshdup
2522 (memopv4f32 addr:$src), (undef)))]>;
2524 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2525 "movsldup\t{$src, $dst|$dst, $src}",
2526 [(set VR128:$dst, (v4f32 (movsldup
2527 VR128:$src, (undef))))]>;
2528 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2529 "movsldup\t{$src, $dst|$dst, $src}",
2530 [(set VR128:$dst, (movsldup
2531 (memopv4f32 addr:$src), (undef)))]>;
2533 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2534 "movddup\t{$src, $dst|$dst, $src}",
2535 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2536 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2537 "movddup\t{$src, $dst|$dst, $src}",
2539 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2542 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2544 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2546 let AddedComplexity = 5 in {
2547 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2548 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2549 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2550 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2551 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2552 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2553 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2554 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2558 let Constraints = "$src1 = $dst" in {
2559 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2560 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2561 "addsubps\t{$src2, $dst|$dst, $src2}",
2562 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2564 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2565 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2566 "addsubps\t{$src2, $dst|$dst, $src2}",
2567 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2568 (memop addr:$src2)))]>;
2569 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2570 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2571 "addsubpd\t{$src2, $dst|$dst, $src2}",
2572 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2574 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2575 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2576 "addsubpd\t{$src2, $dst|$dst, $src2}",
2577 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2578 (memop addr:$src2)))]>;
2581 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2582 "lddqu\t{$src, $dst|$dst, $src}",
2583 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2586 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2587 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2588 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2589 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2590 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2591 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2592 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2593 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2594 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2595 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2596 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2597 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2598 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2599 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2600 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2601 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2603 let Constraints = "$src1 = $dst" in {
2604 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2605 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2606 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2607 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2608 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2609 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2610 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2611 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2614 // Thread synchronization
2615 def MONITOR : I<0x01, MRM1r, (outs), (ins), "monitor",
2616 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2617 def MWAIT : I<0x01, MRM1r, (outs), (ins), "mwait",
2618 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2620 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2621 let AddedComplexity = 15 in
2622 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2623 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2624 let AddedComplexity = 20 in
2625 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2626 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2628 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2629 let AddedComplexity = 15 in
2630 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2631 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2632 let AddedComplexity = 20 in
2633 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2634 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2636 //===---------------------------------------------------------------------===//
2637 // SSSE3 Instructions
2638 //===---------------------------------------------------------------------===//
2640 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2641 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2642 Intrinsic IntId64, Intrinsic IntId128> {
2643 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2644 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2645 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2647 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2648 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2650 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2652 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2654 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2655 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2658 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2660 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2663 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2666 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2667 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2668 Intrinsic IntId64, Intrinsic IntId128> {
2669 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2671 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2672 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2674 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2676 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2679 (bitconvert (memopv4i16 addr:$src))))]>;
2681 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2683 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2684 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2687 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2689 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2692 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2695 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2696 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2697 Intrinsic IntId64, Intrinsic IntId128> {
2698 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2700 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2701 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2703 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2705 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2708 (bitconvert (memopv2i32 addr:$src))))]>;
2710 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2712 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2713 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2716 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2718 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2721 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2724 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2725 int_x86_ssse3_pabs_b,
2726 int_x86_ssse3_pabs_b_128>;
2727 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2728 int_x86_ssse3_pabs_w,
2729 int_x86_ssse3_pabs_w_128>;
2730 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2731 int_x86_ssse3_pabs_d,
2732 int_x86_ssse3_pabs_d_128>;
2734 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2735 let Constraints = "$src1 = $dst" in {
2736 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2737 Intrinsic IntId64, Intrinsic IntId128,
2738 bit Commutable = 0> {
2739 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2740 (ins VR64:$src1, VR64:$src2),
2741 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2742 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2743 let isCommutable = Commutable;
2745 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2746 (ins VR64:$src1, i64mem:$src2),
2747 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2749 (IntId64 VR64:$src1,
2750 (bitconvert (memopv8i8 addr:$src2))))]>;
2752 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2753 (ins VR128:$src1, VR128:$src2),
2754 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2755 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2757 let isCommutable = Commutable;
2759 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2760 (ins VR128:$src1, i128mem:$src2),
2761 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2763 (IntId128 VR128:$src1,
2764 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2768 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2769 let Constraints = "$src1 = $dst" in {
2770 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2771 Intrinsic IntId64, Intrinsic IntId128,
2772 bit Commutable = 0> {
2773 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2774 (ins VR64:$src1, VR64:$src2),
2775 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2776 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2777 let isCommutable = Commutable;
2779 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2780 (ins VR64:$src1, i64mem:$src2),
2781 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2783 (IntId64 VR64:$src1,
2784 (bitconvert (memopv4i16 addr:$src2))))]>;
2786 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2787 (ins VR128:$src1, VR128:$src2),
2788 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2789 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2791 let isCommutable = Commutable;
2793 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2794 (ins VR128:$src1, i128mem:$src2),
2795 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2797 (IntId128 VR128:$src1,
2798 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2802 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2803 let Constraints = "$src1 = $dst" in {
2804 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2805 Intrinsic IntId64, Intrinsic IntId128,
2806 bit Commutable = 0> {
2807 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2808 (ins VR64:$src1, VR64:$src2),
2809 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2810 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2811 let isCommutable = Commutable;
2813 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2814 (ins VR64:$src1, i64mem:$src2),
2815 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2817 (IntId64 VR64:$src1,
2818 (bitconvert (memopv2i32 addr:$src2))))]>;
2820 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2821 (ins VR128:$src1, VR128:$src2),
2822 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2823 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2825 let isCommutable = Commutable;
2827 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2828 (ins VR128:$src1, i128mem:$src2),
2829 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2831 (IntId128 VR128:$src1,
2832 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2836 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2837 int_x86_ssse3_phadd_w,
2838 int_x86_ssse3_phadd_w_128>;
2839 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2840 int_x86_ssse3_phadd_d,
2841 int_x86_ssse3_phadd_d_128>;
2842 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2843 int_x86_ssse3_phadd_sw,
2844 int_x86_ssse3_phadd_sw_128>;
2845 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2846 int_x86_ssse3_phsub_w,
2847 int_x86_ssse3_phsub_w_128>;
2848 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2849 int_x86_ssse3_phsub_d,
2850 int_x86_ssse3_phsub_d_128>;
2851 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2852 int_x86_ssse3_phsub_sw,
2853 int_x86_ssse3_phsub_sw_128>;
2854 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2855 int_x86_ssse3_pmadd_ub_sw,
2856 int_x86_ssse3_pmadd_ub_sw_128>;
2857 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2858 int_x86_ssse3_pmul_hr_sw,
2859 int_x86_ssse3_pmul_hr_sw_128, 1>;
2860 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2861 int_x86_ssse3_pshuf_b,
2862 int_x86_ssse3_pshuf_b_128>;
2863 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2864 int_x86_ssse3_psign_b,
2865 int_x86_ssse3_psign_b_128>;
2866 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2867 int_x86_ssse3_psign_w,
2868 int_x86_ssse3_psign_w_128>;
2869 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2870 int_x86_ssse3_psign_d,
2871 int_x86_ssse3_psign_d_128>;
2873 let Constraints = "$src1 = $dst" in {
2874 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2875 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2876 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2878 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2879 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2880 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2883 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2884 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2885 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2887 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2888 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2889 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2893 // palignr patterns.
2894 def : Pat<(int_x86_ssse3_palign_r VR64:$src1, VR64:$src2, (i8 imm:$src3)),
2895 (PALIGNR64rr VR64:$src1, VR64:$src2, (BYTE_imm imm:$src3))>,
2896 Requires<[HasSSSE3]>;
2897 def : Pat<(int_x86_ssse3_palign_r VR64:$src1,
2898 (memop64 addr:$src2),
2900 (PALIGNR64rm VR64:$src1, addr:$src2, (BYTE_imm imm:$src3))>,
2901 Requires<[HasSSSE3]>;
2903 def : Pat<(int_x86_ssse3_palign_r_128 VR128:$src1, VR128:$src2, (i8 imm:$src3)),
2904 (PALIGNR128rr VR128:$src1, VR128:$src2, (BYTE_imm imm:$src3))>,
2905 Requires<[HasSSSE3]>;
2906 def : Pat<(int_x86_ssse3_palign_r_128 VR128:$src1,
2907 (memopv2i64 addr:$src2),
2909 (PALIGNR128rm VR128:$src1, addr:$src2, (BYTE_imm imm:$src3))>,
2910 Requires<[HasSSSE3]>;
2912 let AddedComplexity = 5 in {
2913 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2914 (PALIGNR128rr VR128:$src2, VR128:$src1,
2915 (SHUFFLE_get_palign_imm VR128:$src3))>,
2916 Requires<[HasSSSE3]>;
2917 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2918 (PALIGNR128rr VR128:$src2, VR128:$src1,
2919 (SHUFFLE_get_palign_imm VR128:$src3))>,
2920 Requires<[HasSSSE3]>;
2921 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2922 (PALIGNR128rr VR128:$src2, VR128:$src1,
2923 (SHUFFLE_get_palign_imm VR128:$src3))>,
2924 Requires<[HasSSSE3]>;
2925 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2926 (PALIGNR128rr VR128:$src2, VR128:$src1,
2927 (SHUFFLE_get_palign_imm VR128:$src3))>,
2928 Requires<[HasSSSE3]>;
2931 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2932 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2933 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2934 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2936 //===---------------------------------------------------------------------===//
2937 // Non-Instruction Patterns
2938 //===---------------------------------------------------------------------===//
2940 // extload f32 -> f64. This matches load+fextend because we have a hack in
2941 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2943 // Since these loads aren't folded into the fextend, we have to match it
2945 let Predicates = [HasSSE2] in
2946 def : Pat<(fextend (loadf32 addr:$src)),
2947 (CVTSS2SDrm addr:$src)>;
2950 let Predicates = [HasSSE2] in {
2951 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2952 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2953 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2954 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2955 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2956 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2957 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2958 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2959 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2960 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2961 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2962 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2963 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2964 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2965 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2966 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2967 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2968 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2969 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2970 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2971 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2972 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2973 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2974 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2975 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2976 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2977 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2978 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2979 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2980 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2983 // Move scalar to XMM zero-extended
2984 // movd to XMM register zero-extends
2985 let AddedComplexity = 15 in {
2986 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2987 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2988 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2989 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2990 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE1]>;
2991 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2992 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2993 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2994 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2997 // Splat v2f64 / v2i64
2998 let AddedComplexity = 10 in {
2999 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
3000 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3001 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
3002 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3003 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
3004 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3005 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
3006 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3009 // Special unary SHUFPSrri case.
3010 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
3011 (SHUFPSrri VR128:$src1, VR128:$src1,
3012 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3013 Requires<[HasSSE1]>;
3014 let AddedComplexity = 5 in
3015 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
3016 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3017 Requires<[HasSSE2]>;
3018 // Special unary SHUFPDrri case.
3019 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
3020 (SHUFPDrri VR128:$src1, VR128:$src1,
3021 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3022 Requires<[HasSSE2]>;
3023 // Special unary SHUFPDrri case.
3024 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
3025 (SHUFPDrri VR128:$src1, VR128:$src1,
3026 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3027 Requires<[HasSSE2]>;
3028 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
3029 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
3030 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3031 Requires<[HasSSE2]>;
3033 // Special binary v4i32 shuffle cases with SHUFPS.
3034 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
3035 (SHUFPSrri VR128:$src1, VR128:$src2,
3036 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3037 Requires<[HasSSE2]>;
3038 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
3039 (SHUFPSrmi VR128:$src1, addr:$src2,
3040 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3041 Requires<[HasSSE2]>;
3042 // Special binary v2i64 shuffle cases using SHUFPDrri.
3043 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
3044 (SHUFPDrri VR128:$src1, VR128:$src2,
3045 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3046 Requires<[HasSSE2]>;
3048 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
3049 let AddedComplexity = 15 in {
3050 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
3051 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3052 Requires<[OptForSpeed, HasSSE2]>;
3053 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
3054 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3055 Requires<[OptForSpeed, HasSSE2]>;
3057 let AddedComplexity = 10 in {
3058 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
3059 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
3060 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
3061 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3062 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
3063 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3064 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
3065 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3068 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
3069 let AddedComplexity = 15 in {
3070 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
3071 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3072 Requires<[OptForSpeed, HasSSE2]>;
3073 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
3074 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3075 Requires<[OptForSpeed, HasSSE2]>;
3077 let AddedComplexity = 10 in {
3078 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
3079 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
3080 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
3081 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3082 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
3083 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3084 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
3085 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3088 let AddedComplexity = 20 in {
3089 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3090 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3091 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3093 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3094 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3095 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3097 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3098 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3099 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3100 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3101 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3104 let AddedComplexity = 20 in {
3105 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3106 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3107 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3108 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3109 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3110 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3111 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3112 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3113 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3116 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3117 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3118 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3119 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3120 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3121 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3123 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3124 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3125 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3127 let AddedComplexity = 15 in {
3128 // Setting the lowest element in the vector.
3129 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3130 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3131 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3132 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3134 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
3135 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3136 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3137 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3138 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3141 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3142 // fall back to this for SSE1)
3143 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3144 (SHUFPSrri VR128:$src2, VR128:$src1,
3145 (SHUFFLE_get_shuf_imm VR128:$src3))>, Requires<[HasSSE1]>;
3147 // Set lowest element and zero upper elements.
3148 let AddedComplexity = 15 in
3149 def : Pat<(v2f64 (movl immAllZerosV_bc, VR128:$src)),
3150 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3151 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3152 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3154 // Some special case pandn patterns.
3155 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3157 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3158 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3160 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3161 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3163 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3165 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3166 (memop addr:$src2))),
3167 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3168 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3169 (memop addr:$src2))),
3170 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3171 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3172 (memop addr:$src2))),
3173 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3175 // vector -> vector casts
3176 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3177 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3178 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3179 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3180 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3181 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3182 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3183 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3185 // Use movaps / movups for SSE integer load / store (one byte shorter).
3186 def : Pat<(alignedloadv4i32 addr:$src),
3187 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
3188 def : Pat<(loadv4i32 addr:$src),
3189 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3190 def : Pat<(alignedloadv2i64 addr:$src),
3191 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3192 def : Pat<(loadv2i64 addr:$src),
3193 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3195 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3196 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3197 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3198 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3199 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3200 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3201 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3202 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3203 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3204 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3205 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3206 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3207 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3208 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3209 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3210 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3212 //===----------------------------------------------------------------------===//
3213 // SSE4.1 Instructions
3214 //===----------------------------------------------------------------------===//
3216 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3219 Intrinsic V2F64Int> {
3220 // Intrinsic operation, reg.
3221 // Vector intrinsic operation, reg
3222 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3223 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3224 !strconcat(OpcodeStr,
3225 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3226 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3229 // Vector intrinsic operation, mem
3230 def PSm_Int : Ii8<opcps, MRMSrcMem,
3231 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3232 !strconcat(OpcodeStr,
3233 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3235 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3237 Requires<[HasSSE41]>;
3239 // Vector intrinsic operation, reg
3240 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3241 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3242 !strconcat(OpcodeStr,
3243 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3244 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3247 // Vector intrinsic operation, mem
3248 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3249 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3250 !strconcat(OpcodeStr,
3251 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3253 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3257 let Constraints = "$src1 = $dst" in {
3258 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3262 // Intrinsic operation, reg.
3263 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3265 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3266 !strconcat(OpcodeStr,
3267 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3269 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3272 // Intrinsic operation, mem.
3273 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3275 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3276 !strconcat(OpcodeStr,
3277 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3279 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3282 // Intrinsic operation, reg.
3283 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3285 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3286 !strconcat(OpcodeStr,
3287 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3289 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3292 // Intrinsic operation, mem.
3293 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3295 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3296 !strconcat(OpcodeStr,
3297 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3299 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3304 // FP round - roundss, roundps, roundsd, roundpd
3305 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3306 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3307 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3308 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3310 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3311 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3312 Intrinsic IntId128> {
3313 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3315 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3316 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3317 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3319 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3322 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3325 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3326 int_x86_sse41_phminposuw>;
3328 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3329 let Constraints = "$src1 = $dst" in {
3330 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3331 Intrinsic IntId128, bit Commutable = 0> {
3332 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3333 (ins VR128:$src1, VR128:$src2),
3334 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3335 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3337 let isCommutable = Commutable;
3339 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3340 (ins VR128:$src1, i128mem:$src2),
3341 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3343 (IntId128 VR128:$src1,
3344 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3348 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3349 int_x86_sse41_pcmpeqq, 1>;
3350 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3351 int_x86_sse41_packusdw, 0>;
3352 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3353 int_x86_sse41_pminsb, 1>;
3354 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3355 int_x86_sse41_pminsd, 1>;
3356 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3357 int_x86_sse41_pminud, 1>;
3358 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3359 int_x86_sse41_pminuw, 1>;
3360 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3361 int_x86_sse41_pmaxsb, 1>;
3362 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3363 int_x86_sse41_pmaxsd, 1>;
3364 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3365 int_x86_sse41_pmaxud, 1>;
3366 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3367 int_x86_sse41_pmaxuw, 1>;
3369 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3371 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3372 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3373 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3374 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3376 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3377 let Constraints = "$src1 = $dst" in {
3378 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3379 SDNode OpNode, Intrinsic IntId128,
3380 bit Commutable = 0> {
3381 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3382 (ins VR128:$src1, VR128:$src2),
3383 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3384 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3385 VR128:$src2))]>, OpSize {
3386 let isCommutable = Commutable;
3388 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3389 (ins VR128:$src1, VR128:$src2),
3390 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3391 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3393 let isCommutable = Commutable;
3395 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3396 (ins VR128:$src1, i128mem:$src2),
3397 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3399 (OpNode VR128:$src1, (memop addr:$src2)))]>, OpSize;
3400 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3401 (ins VR128:$src1, i128mem:$src2),
3402 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3404 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3408 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", v4i32, mul,
3409 int_x86_sse41_pmulld, 1>;
3411 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3412 let Constraints = "$src1 = $dst" in {
3413 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3414 Intrinsic IntId128, bit Commutable = 0> {
3415 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3416 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3417 !strconcat(OpcodeStr,
3418 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3420 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3422 let isCommutable = Commutable;
3424 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3425 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3426 !strconcat(OpcodeStr,
3427 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3429 (IntId128 VR128:$src1,
3430 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3435 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3436 int_x86_sse41_blendps, 0>;
3437 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3438 int_x86_sse41_blendpd, 0>;
3439 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3440 int_x86_sse41_pblendw, 0>;
3441 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3442 int_x86_sse41_dpps, 1>;
3443 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3444 int_x86_sse41_dppd, 1>;
3445 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3446 int_x86_sse41_mpsadbw, 1>;
3449 /// SS41I_ternary_int - SSE 4.1 ternary operator
3450 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3451 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3452 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3453 (ins VR128:$src1, VR128:$src2),
3454 !strconcat(OpcodeStr,
3455 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3456 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3459 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3460 (ins VR128:$src1, i128mem:$src2),
3461 !strconcat(OpcodeStr,
3462 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3465 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3469 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3470 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3471 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3474 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3475 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3476 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3477 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3479 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3480 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3482 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3486 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3487 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3488 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3489 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3490 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3491 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3493 // Common patterns involving scalar load.
3494 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3495 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3496 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3497 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3499 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3500 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3501 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3502 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3504 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3505 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3506 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3507 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3509 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3510 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3511 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3512 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3514 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3515 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3516 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3517 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3519 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3520 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3521 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3522 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3525 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3526 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3527 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3528 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3530 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3531 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3533 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3537 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3538 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3539 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3540 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3542 // Common patterns involving scalar load
3543 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3544 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3545 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3546 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3548 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3549 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3550 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3551 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3554 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3555 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3556 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3557 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3559 // Expecting a i16 load any extended to i32 value.
3560 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3561 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3562 [(set VR128:$dst, (IntId (bitconvert
3563 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3567 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3568 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3570 // Common patterns involving scalar load
3571 def : Pat<(int_x86_sse41_pmovsxbq
3572 (bitconvert (v4i32 (X86vzmovl
3573 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3574 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3576 def : Pat<(int_x86_sse41_pmovzxbq
3577 (bitconvert (v4i32 (X86vzmovl
3578 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3579 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3582 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3583 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3584 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3585 (ins VR128:$src1, i32i8imm:$src2),
3586 !strconcat(OpcodeStr,
3587 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3588 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3590 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3591 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3592 !strconcat(OpcodeStr,
3593 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3596 // There's an AssertZext in the way of writing the store pattern
3597 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3600 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3603 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3604 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3605 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3606 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3607 !strconcat(OpcodeStr,
3608 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3611 // There's an AssertZext in the way of writing the store pattern
3612 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3615 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3618 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3619 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3620 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3621 (ins VR128:$src1, i32i8imm:$src2),
3622 !strconcat(OpcodeStr,
3623 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3625 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3626 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3627 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3628 !strconcat(OpcodeStr,
3629 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3630 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3631 addr:$dst)]>, OpSize;
3634 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3637 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3639 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3640 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3641 (ins VR128:$src1, i32i8imm:$src2),
3642 !strconcat(OpcodeStr,
3643 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3645 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3647 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3648 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3649 !strconcat(OpcodeStr,
3650 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3651 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3652 addr:$dst)]>, OpSize;
3655 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3657 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3658 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3661 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3662 Requires<[HasSSE41]>;
3664 let Constraints = "$src1 = $dst" in {
3665 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3666 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3667 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3668 !strconcat(OpcodeStr,
3669 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3671 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3672 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3673 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3674 !strconcat(OpcodeStr,
3675 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3677 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3678 imm:$src3))]>, OpSize;
3682 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3684 let Constraints = "$src1 = $dst" in {
3685 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3686 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3687 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3688 !strconcat(OpcodeStr,
3689 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3691 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3693 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3694 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3695 !strconcat(OpcodeStr,
3696 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3698 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3699 imm:$src3)))]>, OpSize;
3703 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3705 // insertps has a few different modes, there's the first two here below which
3706 // are optimized inserts that won't zero arbitrary elements in the destination
3707 // vector. The next one matches the intrinsic and could zero arbitrary elements
3708 // in the target vector.
3709 let Constraints = "$src1 = $dst" in {
3710 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3711 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3712 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3713 !strconcat(OpcodeStr,
3714 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3716 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3718 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3719 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3720 !strconcat(OpcodeStr,
3721 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3723 (X86insrtps VR128:$src1,
3724 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3725 imm:$src3))]>, OpSize;
3729 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3731 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3732 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3734 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3735 // the intel intrinsic that corresponds to this.
3736 let Defs = [EFLAGS] in {
3737 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3738 "ptest \t{$src2, $src1|$src1, $src2}",
3739 [(X86ptest VR128:$src1, VR128:$src2),
3740 (implicit EFLAGS)]>, OpSize;
3741 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3742 "ptest \t{$src2, $src1|$src1, $src2}",
3743 [(X86ptest VR128:$src1, (load addr:$src2)),
3744 (implicit EFLAGS)]>, OpSize;
3747 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3748 "movntdqa\t{$src, $dst|$dst, $src}",
3749 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;
3752 //===----------------------------------------------------------------------===//
3753 // SSE4.2 Instructions
3754 //===----------------------------------------------------------------------===//
3756 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3757 let Constraints = "$src1 = $dst" in {
3758 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3759 Intrinsic IntId128, bit Commutable = 0> {
3760 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3761 (ins VR128:$src1, VR128:$src2),
3762 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3763 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3765 let isCommutable = Commutable;
3767 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3768 (ins VR128:$src1, i128mem:$src2),
3769 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3771 (IntId128 VR128:$src1,
3772 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3776 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3778 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3779 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3780 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3781 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3783 // crc intrinsic instruction
3784 // This set of instructions are only rm, the only difference is the size
3786 let Constraints = "$src1 = $dst" in {
3787 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3788 (ins GR32:$src1, i8mem:$src2),
3789 "crc32 \t{$src2, $src1|$src1, $src2}",
3791 (int_x86_sse42_crc32_8 GR32:$src1,
3792 (load addr:$src2)))]>, OpSize;
3793 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3794 (ins GR32:$src1, GR8:$src2),
3795 "crc32 \t{$src2, $src1|$src1, $src2}",
3797 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>,
3799 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3800 (ins GR32:$src1, i16mem:$src2),
3801 "crc32 \t{$src2, $src1|$src1, $src2}",
3803 (int_x86_sse42_crc32_16 GR32:$src1,
3804 (load addr:$src2)))]>,
3806 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3807 (ins GR32:$src1, GR16:$src2),
3808 "crc32 \t{$src2, $src1|$src1, $src2}",
3810 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3812 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3813 (ins GR32:$src1, i32mem:$src2),
3814 "crc32 \t{$src2, $src1|$src1, $src2}",
3816 (int_x86_sse42_crc32_32 GR32:$src1,
3817 (load addr:$src2)))]>, OpSize;
3818 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3819 (ins GR32:$src1, GR32:$src2),
3820 "crc32 \t{$src2, $src1|$src1, $src2}",
3822 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>,
3824 def CRC64m64 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3825 (ins GR64:$src1, i64mem:$src2),
3826 "crc32 \t{$src2, $src1|$src1, $src2}",
3828 (int_x86_sse42_crc32_64 GR64:$src1,
3829 (load addr:$src2)))]>,
3831 def CRC64r64 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3832 (ins GR64:$src1, GR64:$src2),
3833 "crc32 \t{$src2, $src1|$src1, $src2}",
3835 (int_x86_sse42_crc32_64 GR64:$src1, GR64:$src2))]>,
3839 // String/text processing instructions.
3840 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3841 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3842 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3843 "#PCMPISTRM128rr PSEUDO!",
3844 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3845 imm:$src3))]>, OpSize;
3846 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3847 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3848 "#PCMPISTRM128rm PSEUDO!",
3849 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3850 imm:$src3))]>, OpSize;
3853 let Defs = [XMM0, EFLAGS] in {
3854 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3855 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3856 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3857 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3858 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3859 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3862 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3863 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3864 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3865 "#PCMPESTRM128rr PSEUDO!",
3867 (int_x86_sse42_pcmpestrm128
3868 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3870 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3871 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3872 "#PCMPESTRM128rm PSEUDO!",
3873 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3874 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3878 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3879 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3880 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3881 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3882 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3883 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3884 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3887 let Defs = [ECX, EFLAGS] in {
3888 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3889 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3890 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3891 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3892 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3893 (implicit EFLAGS)]>, OpSize;
3894 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3895 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3896 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3897 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3898 (implicit EFLAGS)]>, OpSize;
3902 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3903 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3904 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3905 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3906 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3907 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3909 let Defs = [ECX, EFLAGS] in {
3910 let Uses = [EAX, EDX] in {
3911 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3912 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3913 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3914 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3915 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3916 (implicit EFLAGS)]>, OpSize;
3917 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3918 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3919 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3921 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
3922 (implicit EFLAGS)]>, OpSize;
3927 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3928 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3929 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3930 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3931 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3932 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;