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))>;
164 // Like 'store', but requires the non-temporal bit to be set
165 def nontemporalstore : PatFrag<(ops node:$val, node:$ptr),
166 (st node:$val, node:$ptr), [{
167 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
168 return ST->isNonTemporal();
172 def alignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
173 (st node:$val, node:$ptr), [{
174 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
175 return ST->isNonTemporal() && !ST->isTruncatingStore() &&
176 ST->getAddressingMode() == ISD::UNINDEXED &&
177 ST->getAlignment() >= 16;
181 def unalignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
182 (st node:$val, node:$ptr), [{
183 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
184 return ST->isNonTemporal() &&
185 ST->getAlignment() < 16;
189 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
190 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
191 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
192 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
193 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
194 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
196 def vzmovl_v2i64 : PatFrag<(ops node:$src),
197 (bitconvert (v2i64 (X86vzmovl
198 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
199 def vzmovl_v4i32 : PatFrag<(ops node:$src),
200 (bitconvert (v4i32 (X86vzmovl
201 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
203 def vzload_v2i64 : PatFrag<(ops node:$src),
204 (bitconvert (v2i64 (X86vzload node:$src)))>;
207 def fp32imm0 : PatLeaf<(f32 fpimm), [{
208 return N->isExactlyValue(+0.0);
211 // BYTE_imm - Transform bit immediates into byte immediates.
212 def BYTE_imm : SDNodeXForm<imm, [{
213 // Transformation function: imm >> 3
214 return getI32Imm(N->getZExtValue() >> 3);
217 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
219 def SHUFFLE_get_shuf_imm : SDNodeXForm<vector_shuffle, [{
220 return getI8Imm(X86::getShuffleSHUFImmediate(N));
223 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
225 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<vector_shuffle, [{
226 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
229 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
231 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<vector_shuffle, [{
232 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
235 // SHUFFLE_get_palign_imm xform function: convert vector_shuffle mask to
237 def SHUFFLE_get_palign_imm : SDNodeXForm<vector_shuffle, [{
238 return getI8Imm(X86::getShufflePALIGNRImmediate(N));
241 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
242 (vector_shuffle node:$lhs, node:$rhs), [{
243 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
244 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
247 def movddup : PatFrag<(ops node:$lhs, node:$rhs),
248 (vector_shuffle node:$lhs, node:$rhs), [{
249 return X86::isMOVDDUPMask(cast<ShuffleVectorSDNode>(N));
252 def movhlps : PatFrag<(ops node:$lhs, node:$rhs),
253 (vector_shuffle node:$lhs, node:$rhs), [{
254 return X86::isMOVHLPSMask(cast<ShuffleVectorSDNode>(N));
257 def movhlps_undef : PatFrag<(ops node:$lhs, node:$rhs),
258 (vector_shuffle node:$lhs, node:$rhs), [{
259 return X86::isMOVHLPS_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
262 def movlhps : PatFrag<(ops node:$lhs, node:$rhs),
263 (vector_shuffle node:$lhs, node:$rhs), [{
264 return X86::isMOVLHPSMask(cast<ShuffleVectorSDNode>(N));
267 def movlp : PatFrag<(ops node:$lhs, node:$rhs),
268 (vector_shuffle node:$lhs, node:$rhs), [{
269 return X86::isMOVLPMask(cast<ShuffleVectorSDNode>(N));
272 def movl : PatFrag<(ops node:$lhs, node:$rhs),
273 (vector_shuffle node:$lhs, node:$rhs), [{
274 return X86::isMOVLMask(cast<ShuffleVectorSDNode>(N));
277 def movshdup : PatFrag<(ops node:$lhs, node:$rhs),
278 (vector_shuffle node:$lhs, node:$rhs), [{
279 return X86::isMOVSHDUPMask(cast<ShuffleVectorSDNode>(N));
282 def movsldup : PatFrag<(ops node:$lhs, node:$rhs),
283 (vector_shuffle node:$lhs, node:$rhs), [{
284 return X86::isMOVSLDUPMask(cast<ShuffleVectorSDNode>(N));
287 def unpckl : PatFrag<(ops node:$lhs, node:$rhs),
288 (vector_shuffle node:$lhs, node:$rhs), [{
289 return X86::isUNPCKLMask(cast<ShuffleVectorSDNode>(N));
292 def unpckh : PatFrag<(ops node:$lhs, node:$rhs),
293 (vector_shuffle node:$lhs, node:$rhs), [{
294 return X86::isUNPCKHMask(cast<ShuffleVectorSDNode>(N));
297 def unpckl_undef : PatFrag<(ops node:$lhs, node:$rhs),
298 (vector_shuffle node:$lhs, node:$rhs), [{
299 return X86::isUNPCKL_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
302 def unpckh_undef : PatFrag<(ops node:$lhs, node:$rhs),
303 (vector_shuffle node:$lhs, node:$rhs), [{
304 return X86::isUNPCKH_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
307 def pshufd : PatFrag<(ops node:$lhs, node:$rhs),
308 (vector_shuffle node:$lhs, node:$rhs), [{
309 return X86::isPSHUFDMask(cast<ShuffleVectorSDNode>(N));
310 }], SHUFFLE_get_shuf_imm>;
312 def shufp : PatFrag<(ops node:$lhs, node:$rhs),
313 (vector_shuffle node:$lhs, node:$rhs), [{
314 return X86::isSHUFPMask(cast<ShuffleVectorSDNode>(N));
315 }], SHUFFLE_get_shuf_imm>;
317 def pshufhw : PatFrag<(ops node:$lhs, node:$rhs),
318 (vector_shuffle node:$lhs, node:$rhs), [{
319 return X86::isPSHUFHWMask(cast<ShuffleVectorSDNode>(N));
320 }], SHUFFLE_get_pshufhw_imm>;
322 def pshuflw : PatFrag<(ops node:$lhs, node:$rhs),
323 (vector_shuffle node:$lhs, node:$rhs), [{
324 return X86::isPSHUFLWMask(cast<ShuffleVectorSDNode>(N));
325 }], SHUFFLE_get_pshuflw_imm>;
327 def palign : PatFrag<(ops node:$lhs, node:$rhs),
328 (vector_shuffle node:$lhs, node:$rhs), [{
329 return X86::isPALIGNRMask(cast<ShuffleVectorSDNode>(N));
330 }], SHUFFLE_get_palign_imm>;
332 //===----------------------------------------------------------------------===//
333 // SSE scalar FP Instructions
334 //===----------------------------------------------------------------------===//
336 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded after
337 // instruction selection into a branch sequence.
338 let Uses = [EFLAGS], usesCustomInserter = 1 in {
339 def CMOV_FR32 : I<0, Pseudo,
340 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
341 "#CMOV_FR32 PSEUDO!",
342 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
344 def CMOV_FR64 : I<0, Pseudo,
345 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
346 "#CMOV_FR64 PSEUDO!",
347 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
349 def CMOV_V4F32 : I<0, Pseudo,
350 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
351 "#CMOV_V4F32 PSEUDO!",
353 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
355 def CMOV_V2F64 : I<0, Pseudo,
356 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
357 "#CMOV_V2F64 PSEUDO!",
359 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
361 def CMOV_V2I64 : I<0, Pseudo,
362 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
363 "#CMOV_V2I64 PSEUDO!",
365 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
369 //===----------------------------------------------------------------------===//
371 //===----------------------------------------------------------------------===//
373 // Move Instructions. Register-to-register movss is not used for FR32
374 // register copies because it's a partial register update; FsMOVAPSrr is
375 // used instead. Register-to-register movss is not modeled as an INSERT_SUBREG
376 // because INSERT_SUBREG requires that the insert be implementable in terms of
377 // a copy, and just mentioned, we don't use movss for copies.
378 let Constraints = "$src1 = $dst" in
379 def MOVSSrr : SSI<0x10, MRMSrcReg,
380 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
381 "movss\t{$src2, $dst|$dst, $src2}",
382 [(set (v4f32 VR128:$dst),
383 (movl VR128:$src1, (scalar_to_vector FR32:$src2)))]>;
385 // Extract the low 32-bit value from one vector and insert it into another.
386 let AddedComplexity = 15 in
387 def : Pat<(v4f32 (movl VR128:$src1, VR128:$src2)),
388 (MOVSSrr (v4f32 VR128:$src1),
389 (EXTRACT_SUBREG (v4f32 VR128:$src2), x86_subreg_ss))>;
391 // Implicitly promote a 32-bit scalar to a vector.
392 def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
393 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), FR32:$src, x86_subreg_ss)>;
395 // Loading from memory automatically zeroing upper bits.
396 let canFoldAsLoad = 1, isReMaterializable = 1 in
397 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
398 "movss\t{$src, $dst|$dst, $src}",
399 [(set FR32:$dst, (loadf32 addr:$src))]>;
401 // MOVSSrm zeros the high parts of the register; represent this
402 // with SUBREG_TO_REG.
403 let AddedComplexity = 20 in {
404 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
405 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), x86_subreg_ss)>;
406 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
407 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), x86_subreg_ss)>;
408 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
409 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), x86_subreg_ss)>;
412 // Store scalar value to memory.
413 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
414 "movss\t{$src, $dst|$dst, $src}",
415 [(store FR32:$src, addr:$dst)]>;
417 // Extract and store.
418 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
421 (EXTRACT_SUBREG (v4f32 VR128:$src), x86_subreg_ss))>;
423 // Conversion instructions
424 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
425 "cvttss2si\t{$src, $dst|$dst, $src}",
426 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
427 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
428 "cvttss2si\t{$src, $dst|$dst, $src}",
429 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
430 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
431 "cvtsi2ss\t{$src, $dst|$dst, $src}",
432 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
433 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
434 "cvtsi2ss\t{$src, $dst|$dst, $src}",
435 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
437 // Match intrinsics which expect XMM operand(s).
438 def CVTSS2SIrr: SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
439 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
440 def CVTSS2SIrm: SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
441 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
443 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
444 "cvtss2si\t{$src, $dst|$dst, $src}",
445 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
446 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
447 "cvtss2si\t{$src, $dst|$dst, $src}",
448 [(set GR32:$dst, (int_x86_sse_cvtss2si
449 (load addr:$src)))]>;
451 // Match intrinisics which expect MM and XMM operand(s).
452 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
453 "cvtps2pi\t{$src, $dst|$dst, $src}",
454 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
455 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
456 "cvtps2pi\t{$src, $dst|$dst, $src}",
457 [(set VR64:$dst, (int_x86_sse_cvtps2pi
458 (load addr:$src)))]>;
459 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
460 "cvttps2pi\t{$src, $dst|$dst, $src}",
461 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
462 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
463 "cvttps2pi\t{$src, $dst|$dst, $src}",
464 [(set VR64:$dst, (int_x86_sse_cvttps2pi
465 (load addr:$src)))]>;
466 let Constraints = "$src1 = $dst" in {
467 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
468 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
469 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
470 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
472 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
473 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
474 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
475 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
476 (load addr:$src2)))]>;
479 // Aliases for intrinsics
480 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
481 "cvttss2si\t{$src, $dst|$dst, $src}",
483 (int_x86_sse_cvttss2si VR128:$src))]>;
484 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
485 "cvttss2si\t{$src, $dst|$dst, $src}",
487 (int_x86_sse_cvttss2si(load addr:$src)))]>;
489 let Constraints = "$src1 = $dst" in {
490 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
491 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
492 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
493 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
495 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
496 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
497 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
498 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
499 (loadi32 addr:$src2)))]>;
502 // Comparison instructions
503 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
504 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
505 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
506 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
508 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
509 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
510 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
513 let Defs = [EFLAGS] in {
514 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
515 "ucomiss\t{$src2, $src1|$src1, $src2}",
516 [(set EFLAGS, (X86cmp FR32:$src1, FR32:$src2))]>;
517 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
518 "ucomiss\t{$src2, $src1|$src1, $src2}",
519 [(set EFLAGS, (X86cmp FR32:$src1, (loadf32 addr:$src2)))]>;
521 def COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
522 "comiss\t{$src2, $src1|$src1, $src2}", []>;
523 def COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
524 "comiss\t{$src2, $src1|$src1, $src2}", []>;
528 // Aliases to match intrinsics which expect XMM operand(s).
529 let Constraints = "$src1 = $dst" in {
530 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
532 (ins VR128:$src1, VR128:$src, SSECC:$cc),
533 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
534 [(set VR128:$dst, (int_x86_sse_cmp_ss
536 VR128:$src, imm:$cc))]>;
537 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
539 (ins VR128:$src1, f32mem:$src, SSECC:$cc),
540 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
541 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
542 (load addr:$src), imm:$cc))]>;
545 let Defs = [EFLAGS] in {
546 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
547 "ucomiss\t{$src2, $src1|$src1, $src2}",
548 [(set EFLAGS, (X86ucomi (v4f32 VR128:$src1),
550 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
551 "ucomiss\t{$src2, $src1|$src1, $src2}",
552 [(set EFLAGS, (X86ucomi (v4f32 VR128:$src1),
553 (load addr:$src2)))]>;
555 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
556 "comiss\t{$src2, $src1|$src1, $src2}",
557 [(set EFLAGS, (X86comi (v4f32 VR128:$src1),
559 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
560 "comiss\t{$src2, $src1|$src1, $src2}",
561 [(set EFLAGS, (X86comi (v4f32 VR128:$src1),
562 (load addr:$src2)))]>;
565 // Aliases of packed SSE1 instructions for scalar use. These all have names
566 // that start with 'Fs'.
568 // Alias instructions that map fld0 to pxor for sse.
569 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
571 // FIXME: Set encoding to pseudo!
572 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
573 [(set FR32:$dst, fp32imm0)]>,
574 Requires<[HasSSE1]>, TB, OpSize;
576 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
578 let neverHasSideEffects = 1 in
579 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
580 "movaps\t{$src, $dst|$dst, $src}", []>;
582 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
584 let canFoldAsLoad = 1, isReMaterializable = 1 in
585 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
586 "movaps\t{$src, $dst|$dst, $src}",
587 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
589 // Alias bitwise logical operations using SSE logical ops on packed FP values.
590 let Constraints = "$src1 = $dst" in {
591 let isCommutable = 1 in {
592 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst),
593 (ins FR32:$src1, FR32:$src2),
594 "andps\t{$src2, $dst|$dst, $src2}",
595 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
596 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst),
597 (ins FR32:$src1, FR32:$src2),
598 "orps\t{$src2, $dst|$dst, $src2}",
599 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
600 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst),
601 (ins FR32:$src1, FR32:$src2),
602 "xorps\t{$src2, $dst|$dst, $src2}",
603 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
606 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst),
607 (ins FR32:$src1, f128mem:$src2),
608 "andps\t{$src2, $dst|$dst, $src2}",
609 [(set FR32:$dst, (X86fand FR32:$src1,
610 (memopfsf32 addr:$src2)))]>;
611 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst),
612 (ins FR32:$src1, f128mem:$src2),
613 "orps\t{$src2, $dst|$dst, $src2}",
614 [(set FR32:$dst, (X86for FR32:$src1,
615 (memopfsf32 addr:$src2)))]>;
616 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst),
617 (ins FR32:$src1, f128mem:$src2),
618 "xorps\t{$src2, $dst|$dst, $src2}",
619 [(set FR32:$dst, (X86fxor FR32:$src1,
620 (memopfsf32 addr:$src2)))]>;
622 let neverHasSideEffects = 1 in {
623 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
624 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
625 "andnps\t{$src2, $dst|$dst, $src2}", []>;
627 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
628 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
629 "andnps\t{$src2, $dst|$dst, $src2}", []>;
633 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
635 /// In addition, we also have a special variant of the scalar form here to
636 /// represent the associated intrinsic operation. This form is unlike the
637 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
638 /// and leaves the top elements unmodified (therefore these cannot be commuted).
640 /// These three forms can each be reg+reg or reg+mem, so there are a total of
641 /// six "instructions".
643 let Constraints = "$src1 = $dst" in {
644 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
645 SDNode OpNode, Intrinsic F32Int,
646 bit Commutable = 0> {
647 // Scalar operation, reg+reg.
648 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$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))]>;
680 // Intrinsic operation, reg+mem.
681 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
682 (ins VR128:$src1, ssmem:$src2),
683 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
684 [(set VR128:$dst, (F32Int VR128:$src1,
685 sse_load_f32:$src2))]>;
689 // Arithmetic instructions
690 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
691 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
692 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
693 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
695 /// sse1_fp_binop_rm - Other SSE1 binops
697 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
698 /// instructions for a full-vector intrinsic form. Operations that map
699 /// onto C operators don't use this form since they just use the plain
700 /// vector form instead of having a separate vector intrinsic form.
702 /// This provides a total of eight "instructions".
704 let Constraints = "$src1 = $dst" in {
705 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
709 bit Commutable = 0> {
711 // Scalar operation, reg+reg.
712 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
713 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
714 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
715 let isCommutable = Commutable;
718 // Scalar operation, reg+mem.
719 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
720 (ins FR32:$src1, f32mem:$src2),
721 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
722 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
724 // Vector operation, reg+reg.
725 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
726 (ins VR128:$src1, VR128:$src2),
727 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
728 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
729 let isCommutable = Commutable;
732 // Vector operation, reg+mem.
733 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
734 (ins VR128:$src1, f128mem:$src2),
735 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
736 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
738 // Intrinsic operation, reg+reg.
739 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
740 (ins VR128:$src1, VR128:$src2),
741 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
742 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
743 let isCommutable = Commutable;
746 // Intrinsic operation, reg+mem.
747 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
748 (ins VR128:$src1, ssmem:$src2),
749 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
750 [(set VR128:$dst, (F32Int VR128:$src1,
751 sse_load_f32:$src2))]>;
753 // Vector intrinsic operation, reg+reg.
754 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
755 (ins VR128:$src1, VR128:$src2),
756 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
757 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
758 let isCommutable = Commutable;
761 // Vector intrinsic operation, reg+mem.
762 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
763 (ins VR128:$src1, f128mem:$src2),
764 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
765 [(set VR128:$dst, (V4F32Int VR128:$src1, (memopv4f32 addr:$src2)))]>;
769 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
770 int_x86_sse_max_ss, int_x86_sse_max_ps>;
771 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
772 int_x86_sse_min_ss, int_x86_sse_min_ps>;
774 //===----------------------------------------------------------------------===//
775 // SSE packed FP Instructions
778 let neverHasSideEffects = 1 in
779 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
780 "movaps\t{$src, $dst|$dst, $src}", []>;
781 let canFoldAsLoad = 1, isReMaterializable = 1 in
782 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
783 "movaps\t{$src, $dst|$dst, $src}",
784 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
786 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
787 "movaps\t{$src, $dst|$dst, $src}",
788 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
790 let neverHasSideEffects = 1 in
791 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
792 "movups\t{$src, $dst|$dst, $src}", []>;
793 let canFoldAsLoad = 1, isReMaterializable = 1 in
794 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
795 "movups\t{$src, $dst|$dst, $src}",
796 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
797 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
798 "movups\t{$src, $dst|$dst, $src}",
799 [(store (v4f32 VR128:$src), addr:$dst)]>;
801 // Intrinsic forms of MOVUPS load and store
802 let canFoldAsLoad = 1, isReMaterializable = 1 in
803 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
804 "movups\t{$src, $dst|$dst, $src}",
805 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
806 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
807 "movups\t{$src, $dst|$dst, $src}",
808 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
810 let Constraints = "$src1 = $dst" in {
811 let AddedComplexity = 20 in {
812 def MOVLPSrm : PSI<0x12, MRMSrcMem,
813 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
814 "movlps\t{$src2, $dst|$dst, $src2}",
817 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
818 def MOVHPSrm : PSI<0x16, MRMSrcMem,
819 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
820 "movhps\t{$src2, $dst|$dst, $src2}",
822 (movlhps VR128:$src1,
823 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
825 } // Constraints = "$src1 = $dst"
828 def : Pat<(movlhps VR128:$src1, (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
829 (MOVHPSrm (v4i32 VR128:$src1), addr:$src2)>;
831 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
832 "movlps\t{$src, $dst|$dst, $src}",
833 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
834 (iPTR 0))), addr:$dst)]>;
836 // v2f64 extract element 1 is always custom lowered to unpack high to low
837 // and extract element 0 so the non-store version isn't too horrible.
838 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
839 "movhps\t{$src, $dst|$dst, $src}",
840 [(store (f64 (vector_extract
841 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
842 (undef)), (iPTR 0))), addr:$dst)]>;
844 let Constraints = "$src1 = $dst" in {
845 let AddedComplexity = 20 in {
846 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
847 (ins VR128:$src1, VR128:$src2),
848 "movlhps\t{$src2, $dst|$dst, $src2}",
850 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
852 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
853 (ins VR128:$src1, VR128:$src2),
854 "movhlps\t{$src2, $dst|$dst, $src2}",
856 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
858 } // Constraints = "$src1 = $dst"
860 let AddedComplexity = 20 in {
861 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
862 (MOVLHPSrr (v4f32 VR128:$src), (v4f32 VR128:$src))>;
863 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
864 (MOVLHPSrr (v2i64 VR128:$src), (v2i64 VR128:$src))>;
871 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
873 /// In addition, we also have a special variant of the scalar form here to
874 /// represent the associated intrinsic operation. This form is unlike the
875 /// plain scalar form, in that it takes an entire vector (instead of a
876 /// scalar) and leaves the top elements undefined.
878 /// And, we have a special variant form for a full-vector intrinsic form.
880 /// These four forms can each have a reg or a mem operand, so there are a
881 /// total of eight "instructions".
883 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
887 bit Commutable = 0> {
888 // Scalar operation, reg.
889 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
890 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
891 [(set FR32:$dst, (OpNode FR32:$src))]> {
892 let isCommutable = Commutable;
895 // Scalar operation, mem.
896 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
897 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
898 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
899 Requires<[HasSSE1, OptForSize]>;
901 // Vector operation, reg.
902 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
903 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
904 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
905 let isCommutable = Commutable;
908 // Vector operation, mem.
909 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
910 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
911 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
913 // Intrinsic operation, reg.
914 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
915 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
916 [(set VR128:$dst, (F32Int VR128:$src))]> {
917 let isCommutable = Commutable;
920 // Intrinsic operation, mem.
921 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
922 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
923 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
925 // Vector intrinsic operation, reg
926 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
927 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
928 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
929 let isCommutable = Commutable;
932 // Vector intrinsic operation, mem
933 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
934 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
935 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
939 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
940 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
942 // Reciprocal approximations. Note that these typically require refinement
943 // in order to obtain suitable precision.
944 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
945 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
946 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
947 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
950 let Constraints = "$src1 = $dst" in {
951 let isCommutable = 1 in {
952 def ANDPSrr : PSI<0x54, MRMSrcReg,
953 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
954 "andps\t{$src2, $dst|$dst, $src2}",
955 [(set VR128:$dst, (v2i64
956 (and VR128:$src1, VR128:$src2)))]>;
957 def ORPSrr : PSI<0x56, MRMSrcReg,
958 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
959 "orps\t{$src2, $dst|$dst, $src2}",
960 [(set VR128:$dst, (v2i64
961 (or VR128:$src1, VR128:$src2)))]>;
962 def XORPSrr : PSI<0x57, MRMSrcReg,
963 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
964 "xorps\t{$src2, $dst|$dst, $src2}",
965 [(set VR128:$dst, (v2i64
966 (xor VR128:$src1, VR128:$src2)))]>;
969 def ANDPSrm : PSI<0x54, MRMSrcMem,
970 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
971 "andps\t{$src2, $dst|$dst, $src2}",
972 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
973 (memopv2i64 addr:$src2)))]>;
974 def ORPSrm : PSI<0x56, MRMSrcMem,
975 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
976 "orps\t{$src2, $dst|$dst, $src2}",
977 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
978 (memopv2i64 addr:$src2)))]>;
979 def XORPSrm : PSI<0x57, MRMSrcMem,
980 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
981 "xorps\t{$src2, $dst|$dst, $src2}",
982 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
983 (memopv2i64 addr:$src2)))]>;
984 def ANDNPSrr : PSI<0x55, MRMSrcReg,
985 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
986 "andnps\t{$src2, $dst|$dst, $src2}",
988 (v2i64 (and (xor VR128:$src1,
989 (bc_v2i64 (v4i32 immAllOnesV))),
991 def ANDNPSrm : PSI<0x55, MRMSrcMem,
992 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
993 "andnps\t{$src2, $dst|$dst, $src2}",
995 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
996 (bc_v2i64 (v4i32 immAllOnesV))),
997 (memopv2i64 addr:$src2))))]>;
1000 let Constraints = "$src1 = $dst" in {
1001 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
1002 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1003 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1004 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1005 VR128:$src, imm:$cc))]>;
1006 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
1007 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1008 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1009 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1010 (memop addr:$src), imm:$cc))]>;
1012 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1013 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
1014 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1015 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
1017 // Shuffle and unpack instructions
1018 let Constraints = "$src1 = $dst" in {
1019 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
1020 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
1021 (outs VR128:$dst), (ins VR128:$src1,
1022 VR128:$src2, i8imm:$src3),
1023 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1025 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1026 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
1027 (outs VR128:$dst), (ins VR128:$src1,
1028 f128mem:$src2, i8imm:$src3),
1029 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1032 VR128:$src1, (memopv4f32 addr:$src2))))]>;
1034 let AddedComplexity = 10 in {
1035 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
1036 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1037 "unpckhps\t{$src2, $dst|$dst, $src2}",
1039 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
1040 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
1041 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1042 "unpckhps\t{$src2, $dst|$dst, $src2}",
1044 (v4f32 (unpckh VR128:$src1,
1045 (memopv4f32 addr:$src2))))]>;
1047 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
1048 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1049 "unpcklps\t{$src2, $dst|$dst, $src2}",
1051 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
1052 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
1053 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1054 "unpcklps\t{$src2, $dst|$dst, $src2}",
1056 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
1057 } // AddedComplexity
1058 } // Constraints = "$src1 = $dst"
1061 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1062 "movmskps\t{$src, $dst|$dst, $src}",
1063 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
1064 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1065 "movmskpd\t{$src, $dst|$dst, $src}",
1066 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1068 // Prefetch intrinsic.
1069 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1070 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1071 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1072 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1073 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1074 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1075 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1076 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1078 // Non-temporal stores
1079 def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1080 "movntps\t{$src, $dst|$dst, $src}",
1081 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1083 let AddedComplexity = 400 in { // Prefer non-temporal versions
1084 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1085 "movntps\t{$src, $dst|$dst, $src}",
1086 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
1088 def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1089 "movntdq\t{$src, $dst|$dst, $src}",
1090 [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
1092 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
1093 "movnti\t{$src, $dst|$dst, $src}",
1094 [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
1095 TB, Requires<[HasSSE2]>;
1097 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
1098 "movnti\t{$src, $dst|$dst, $src}",
1099 [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
1100 TB, Requires<[HasSSE2]>;
1103 // Load, store, and memory fence
1104 def SFENCE : PSI<0xAE, MRM7r, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
1107 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1108 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1109 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1110 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1112 // Alias instructions that map zero vector to pxor / xorp* for sse.
1113 // We set canFoldAsLoad because this can be converted to a constant-pool
1114 // load of an all-zeros value if folding it would be beneficial.
1115 // FIXME: Change encoding to pseudo!
1116 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1117 isCodeGenOnly = 1 in
1118 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1119 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1121 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
1122 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
1123 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
1124 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
1125 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
1127 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1128 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), x86_subreg_ss))>;
1130 //===---------------------------------------------------------------------===//
1131 // SSE2 Instructions
1132 //===---------------------------------------------------------------------===//
1134 // Move Instructions. Register-to-register movsd is not used for FR64
1135 // register copies because it's a partial register update; FsMOVAPDrr is
1136 // used instead. Register-to-register movsd is not modeled as an INSERT_SUBREG
1137 // because INSERT_SUBREG requires that the insert be implementable in terms of
1138 // a copy, and just mentioned, we don't use movsd for copies.
1139 let Constraints = "$src1 = $dst" in
1140 def MOVSDrr : SDI<0x10, MRMSrcReg,
1141 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
1142 "movsd\t{$src2, $dst|$dst, $src2}",
1143 [(set (v2f64 VR128:$dst),
1144 (movl VR128:$src1, (scalar_to_vector FR64:$src2)))]>;
1146 // Extract the low 64-bit value from one vector and insert it into another.
1147 let AddedComplexity = 15 in
1148 def : Pat<(v2f64 (movl VR128:$src1, VR128:$src2)),
1149 (MOVSDrr (v2f64 VR128:$src1),
1150 (EXTRACT_SUBREG (v2f64 VR128:$src2), x86_subreg_sd))>;
1152 // Implicitly promote a 64-bit scalar to a vector.
1153 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
1154 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), FR64:$src, x86_subreg_sd)>;
1156 // Loading from memory automatically zeroing upper bits.
1157 let canFoldAsLoad = 1, isReMaterializable = 1, AddedComplexity = 20 in
1158 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1159 "movsd\t{$src, $dst|$dst, $src}",
1160 [(set FR64:$dst, (loadf64 addr:$src))]>;
1162 // MOVSDrm zeros the high parts of the register; represent this
1163 // with SUBREG_TO_REG.
1164 let AddedComplexity = 20 in {
1165 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
1166 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>;
1167 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
1168 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>;
1169 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
1170 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>;
1171 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
1172 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>;
1173 def : Pat<(v2f64 (X86vzload addr:$src)),
1174 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>;
1177 // Store scalar value to memory.
1178 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1179 "movsd\t{$src, $dst|$dst, $src}",
1180 [(store FR64:$src, addr:$dst)]>;
1182 // Extract and store.
1183 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
1186 (EXTRACT_SUBREG (v2f64 VR128:$src), x86_subreg_sd))>;
1188 // Conversion instructions
1189 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1190 "cvttsd2si\t{$src, $dst|$dst, $src}",
1191 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1192 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1193 "cvttsd2si\t{$src, $dst|$dst, $src}",
1194 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1195 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1196 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1197 [(set FR32:$dst, (fround FR64:$src))]>;
1198 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1199 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1200 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
1201 Requires<[HasSSE2, OptForSize]>;
1202 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1203 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1204 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1205 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1206 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1207 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1209 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1210 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1211 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1212 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1213 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1214 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1215 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1216 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1217 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1218 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1219 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1220 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1221 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1222 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1223 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1224 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1225 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1226 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1227 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1228 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1230 // SSE2 instructions with XS prefix
1231 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1232 "cvtss2sd\t{$src, $dst|$dst, $src}",
1233 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1234 Requires<[HasSSE2]>;
1235 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1236 "cvtss2sd\t{$src, $dst|$dst, $src}",
1237 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1238 Requires<[HasSSE2, OptForSize]>;
1240 def : Pat<(extloadf32 addr:$src),
1241 (CVTSS2SDrr (MOVSSrm addr:$src))>,
1242 Requires<[HasSSE2, OptForSpeed]>;
1244 // Match intrinsics which expect XMM operand(s).
1245 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1246 "cvtsd2si\t{$src, $dst|$dst, $src}",
1247 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1248 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1249 "cvtsd2si\t{$src, $dst|$dst, $src}",
1250 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1251 (load addr:$src)))]>;
1253 // Match intrinisics which expect MM and XMM operand(s).
1254 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1255 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1256 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1257 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1258 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1259 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1260 (memop addr:$src)))]>;
1261 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1262 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1263 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1264 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1265 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1266 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1267 (memop addr:$src)))]>;
1268 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1269 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1270 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1271 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1272 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1273 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1274 (load addr:$src)))]>;
1276 // Aliases for intrinsics
1277 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1278 "cvttsd2si\t{$src, $dst|$dst, $src}",
1280 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1281 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1282 "cvttsd2si\t{$src, $dst|$dst, $src}",
1283 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1284 (load addr:$src)))]>;
1286 // Comparison instructions
1287 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1288 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1289 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1290 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1292 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1293 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1294 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1297 let Defs = [EFLAGS] in {
1298 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1299 "ucomisd\t{$src2, $src1|$src1, $src2}",
1300 [(set EFLAGS, (X86cmp FR64:$src1, FR64:$src2))]>;
1301 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1302 "ucomisd\t{$src2, $src1|$src1, $src2}",
1303 [(set EFLAGS, (X86cmp FR64:$src1, (loadf64 addr:$src2)))]>;
1304 } // Defs = [EFLAGS]
1306 // Aliases to match intrinsics which expect XMM operand(s).
1307 let Constraints = "$src1 = $dst" in {
1308 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1310 (ins VR128:$src1, VR128:$src, SSECC:$cc),
1311 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1312 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1313 VR128:$src, imm:$cc))]>;
1314 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1316 (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1317 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1318 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1319 (load addr:$src), imm:$cc))]>;
1322 let Defs = [EFLAGS] in {
1323 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1324 "ucomisd\t{$src2, $src1|$src1, $src2}",
1325 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1327 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1328 "ucomisd\t{$src2, $src1|$src1, $src2}",
1329 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1330 (load addr:$src2)))]>;
1332 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1333 "comisd\t{$src2, $src1|$src1, $src2}",
1334 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1336 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1337 "comisd\t{$src2, $src1|$src1, $src2}",
1338 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1339 (load addr:$src2)))]>;
1340 } // Defs = [EFLAGS]
1342 // Aliases of packed SSE2 instructions for scalar use. These all have names
1343 // that start with 'Fs'.
1345 // Alias instructions that map fld0 to pxor for sse.
1346 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1347 canFoldAsLoad = 1 in
1348 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1349 [(set FR64:$dst, fpimm0)]>,
1350 Requires<[HasSSE2]>, TB, OpSize;
1352 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1354 let neverHasSideEffects = 1 in
1355 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1356 "movapd\t{$src, $dst|$dst, $src}", []>;
1358 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1360 let canFoldAsLoad = 1, isReMaterializable = 1 in
1361 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1362 "movapd\t{$src, $dst|$dst, $src}",
1363 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1365 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1366 let Constraints = "$src1 = $dst" in {
1367 let isCommutable = 1 in {
1368 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1369 (ins FR64:$src1, FR64:$src2),
1370 "andpd\t{$src2, $dst|$dst, $src2}",
1371 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1372 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1373 (ins FR64:$src1, FR64:$src2),
1374 "orpd\t{$src2, $dst|$dst, $src2}",
1375 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1376 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1377 (ins FR64:$src1, FR64:$src2),
1378 "xorpd\t{$src2, $dst|$dst, $src2}",
1379 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1382 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1383 (ins FR64:$src1, f128mem:$src2),
1384 "andpd\t{$src2, $dst|$dst, $src2}",
1385 [(set FR64:$dst, (X86fand FR64:$src1,
1386 (memopfsf64 addr:$src2)))]>;
1387 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1388 (ins FR64:$src1, f128mem:$src2),
1389 "orpd\t{$src2, $dst|$dst, $src2}",
1390 [(set FR64:$dst, (X86for FR64:$src1,
1391 (memopfsf64 addr:$src2)))]>;
1392 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1393 (ins FR64:$src1, f128mem:$src2),
1394 "xorpd\t{$src2, $dst|$dst, $src2}",
1395 [(set FR64:$dst, (X86fxor FR64:$src1,
1396 (memopfsf64 addr:$src2)))]>;
1398 let neverHasSideEffects = 1 in {
1399 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1400 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1401 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1403 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1404 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1405 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1409 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1411 /// In addition, we also have a special variant of the scalar form here to
1412 /// represent the associated intrinsic operation. This form is unlike the
1413 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1414 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1416 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1417 /// six "instructions".
1419 let Constraints = "$src1 = $dst" in {
1420 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1421 SDNode OpNode, Intrinsic F64Int,
1422 bit Commutable = 0> {
1423 // Scalar operation, reg+reg.
1424 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1425 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1426 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1427 let isCommutable = Commutable;
1430 // Scalar operation, reg+mem.
1431 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1432 (ins FR64:$src1, f64mem:$src2),
1433 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1434 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1436 // Vector operation, reg+reg.
1437 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1438 (ins VR128:$src1, VR128:$src2),
1439 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1440 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1441 let isCommutable = Commutable;
1444 // Vector operation, reg+mem.
1445 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1446 (ins VR128:$src1, f128mem:$src2),
1447 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1448 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1450 // Intrinsic operation, reg+reg.
1451 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1452 (ins VR128:$src1, VR128:$src2),
1453 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1454 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]>;
1456 // Intrinsic operation, reg+mem.
1457 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1458 (ins VR128:$src1, sdmem:$src2),
1459 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1460 [(set VR128:$dst, (F64Int VR128:$src1,
1461 sse_load_f64:$src2))]>;
1465 // Arithmetic instructions
1466 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1467 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1468 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1469 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1471 /// sse2_fp_binop_rm - Other SSE2 binops
1473 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1474 /// instructions for a full-vector intrinsic form. Operations that map
1475 /// onto C operators don't use this form since they just use the plain
1476 /// vector form instead of having a separate vector intrinsic form.
1478 /// This provides a total of eight "instructions".
1480 let Constraints = "$src1 = $dst" in {
1481 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1485 bit Commutable = 0> {
1487 // Scalar operation, reg+reg.
1488 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1489 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1490 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1491 let isCommutable = Commutable;
1494 // Scalar operation, reg+mem.
1495 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1496 (ins FR64:$src1, f64mem:$src2),
1497 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1498 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1500 // Vector operation, reg+reg.
1501 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1502 (ins VR128:$src1, VR128:$src2),
1503 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1504 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1505 let isCommutable = Commutable;
1508 // Vector operation, reg+mem.
1509 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1510 (ins VR128:$src1, f128mem:$src2),
1511 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1512 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1514 // Intrinsic operation, reg+reg.
1515 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1516 (ins VR128:$src1, VR128:$src2),
1517 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1518 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1519 let isCommutable = Commutable;
1522 // Intrinsic operation, reg+mem.
1523 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1524 (ins VR128:$src1, sdmem:$src2),
1525 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1526 [(set VR128:$dst, (F64Int VR128:$src1,
1527 sse_load_f64:$src2))]>;
1529 // Vector intrinsic operation, reg+reg.
1530 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1531 (ins VR128:$src1, VR128:$src2),
1532 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1533 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1534 let isCommutable = Commutable;
1537 // Vector intrinsic operation, reg+mem.
1538 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1539 (ins VR128:$src1, f128mem:$src2),
1540 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1541 [(set VR128:$dst, (V2F64Int VR128:$src1,
1542 (memopv2f64 addr:$src2)))]>;
1546 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1547 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1548 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1549 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1551 //===---------------------------------------------------------------------===//
1552 // SSE packed FP Instructions
1554 // Move Instructions
1555 let neverHasSideEffects = 1 in
1556 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1557 "movapd\t{$src, $dst|$dst, $src}", []>;
1558 let canFoldAsLoad = 1, isReMaterializable = 1 in
1559 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1560 "movapd\t{$src, $dst|$dst, $src}",
1561 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1563 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1564 "movapd\t{$src, $dst|$dst, $src}",
1565 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1567 let neverHasSideEffects = 1 in
1568 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1569 "movupd\t{$src, $dst|$dst, $src}", []>;
1570 let canFoldAsLoad = 1 in
1571 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1572 "movupd\t{$src, $dst|$dst, $src}",
1573 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1574 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1575 "movupd\t{$src, $dst|$dst, $src}",
1576 [(store (v2f64 VR128:$src), addr:$dst)]>;
1578 // Intrinsic forms of MOVUPD load and store
1579 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1580 "movupd\t{$src, $dst|$dst, $src}",
1581 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1582 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1583 "movupd\t{$src, $dst|$dst, $src}",
1584 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1586 let Constraints = "$src1 = $dst" in {
1587 let AddedComplexity = 20 in {
1588 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1589 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1590 "movlpd\t{$src2, $dst|$dst, $src2}",
1592 (v2f64 (movlp VR128:$src1,
1593 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1594 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1595 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1596 "movhpd\t{$src2, $dst|$dst, $src2}",
1598 (v2f64 (movlhps VR128:$src1,
1599 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1600 } // AddedComplexity
1601 } // Constraints = "$src1 = $dst"
1603 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1604 "movlpd\t{$src, $dst|$dst, $src}",
1605 [(store (f64 (vector_extract (v2f64 VR128:$src),
1606 (iPTR 0))), addr:$dst)]>;
1608 // v2f64 extract element 1 is always custom lowered to unpack high to low
1609 // and extract element 0 so the non-store version isn't too horrible.
1610 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1611 "movhpd\t{$src, $dst|$dst, $src}",
1612 [(store (f64 (vector_extract
1613 (v2f64 (unpckh VR128:$src, (undef))),
1614 (iPTR 0))), addr:$dst)]>;
1616 // SSE2 instructions without OpSize prefix
1617 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1618 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1619 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1620 TB, Requires<[HasSSE2]>;
1621 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1622 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1623 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1624 (bitconvert (memopv2i64 addr:$src))))]>,
1625 TB, Requires<[HasSSE2]>;
1627 // SSE2 instructions with XS prefix
1628 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1629 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1630 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1631 XS, Requires<[HasSSE2]>;
1632 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1633 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1634 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1635 (bitconvert (memopv2i64 addr:$src))))]>,
1636 XS, Requires<[HasSSE2]>;
1638 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1639 "cvtps2dq\t{$src, $dst|$dst, $src}",
1640 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1641 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1642 "cvtps2dq\t{$src, $dst|$dst, $src}",
1643 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1644 (memop addr:$src)))]>;
1645 // SSE2 packed instructions with XS prefix
1646 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1647 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1648 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1649 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1651 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1652 "cvttps2dq\t{$src, $dst|$dst, $src}",
1654 (int_x86_sse2_cvttps2dq VR128:$src))]>,
1655 XS, Requires<[HasSSE2]>;
1656 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1657 "cvttps2dq\t{$src, $dst|$dst, $src}",
1658 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1659 (memop addr:$src)))]>,
1660 XS, Requires<[HasSSE2]>;
1662 // SSE2 packed instructions with XD prefix
1663 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1664 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1665 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1666 XD, Requires<[HasSSE2]>;
1667 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1668 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1669 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1670 (memop addr:$src)))]>,
1671 XD, Requires<[HasSSE2]>;
1673 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1674 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1675 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1676 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1677 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1678 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1679 (memop addr:$src)))]>;
1681 // SSE2 instructions without OpSize prefix
1682 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1683 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1684 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1685 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1687 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1688 "cvtps2pd\t{$src, $dst|$dst, $src}",
1689 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1690 TB, Requires<[HasSSE2]>;
1691 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1692 "cvtps2pd\t{$src, $dst|$dst, $src}",
1693 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1694 (load addr:$src)))]>,
1695 TB, Requires<[HasSSE2]>;
1697 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1698 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1699 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1700 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1703 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1704 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1705 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1706 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1707 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1708 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1709 (memop addr:$src)))]>;
1711 // Match intrinsics which expect XMM operand(s).
1712 // Aliases for intrinsics
1713 let Constraints = "$src1 = $dst" in {
1714 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1715 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1716 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1717 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1719 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1720 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1721 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1722 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1723 (loadi32 addr:$src2)))]>;
1724 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1725 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1726 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1727 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1729 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1730 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1731 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1732 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1733 (load addr:$src2)))]>;
1734 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1735 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1736 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1737 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1738 VR128:$src2))]>, XS,
1739 Requires<[HasSSE2]>;
1740 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1741 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1742 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1743 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1744 (load addr:$src2)))]>, XS,
1745 Requires<[HasSSE2]>;
1750 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1752 /// In addition, we also have a special variant of the scalar form here to
1753 /// represent the associated intrinsic operation. This form is unlike the
1754 /// plain scalar form, in that it takes an entire vector (instead of a
1755 /// scalar) and leaves the top elements undefined.
1757 /// And, we have a special variant form for a full-vector intrinsic form.
1759 /// These four forms can each have a reg or a mem operand, so there are a
1760 /// total of eight "instructions".
1762 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1766 bit Commutable = 0> {
1767 // Scalar operation, reg.
1768 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1769 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1770 [(set FR64:$dst, (OpNode FR64:$src))]> {
1771 let isCommutable = Commutable;
1774 // Scalar operation, mem.
1775 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1776 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1777 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1779 // Vector operation, reg.
1780 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1781 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1782 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1783 let isCommutable = Commutable;
1786 // Vector operation, mem.
1787 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1788 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1789 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1791 // Intrinsic operation, reg.
1792 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1793 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1794 [(set VR128:$dst, (F64Int VR128:$src))]> {
1795 let isCommutable = Commutable;
1798 // Intrinsic operation, mem.
1799 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1800 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1801 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1803 // Vector intrinsic operation, reg
1804 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1805 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1806 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1807 let isCommutable = Commutable;
1810 // Vector intrinsic operation, mem
1811 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1812 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1813 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1817 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1818 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1820 // There is no f64 version of the reciprocal approximation instructions.
1823 let Constraints = "$src1 = $dst" in {
1824 let isCommutable = 1 in {
1825 def ANDPDrr : PDI<0x54, MRMSrcReg,
1826 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1827 "andpd\t{$src2, $dst|$dst, $src2}",
1829 (and (bc_v2i64 (v2f64 VR128:$src1)),
1830 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1831 def ORPDrr : PDI<0x56, MRMSrcReg,
1832 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1833 "orpd\t{$src2, $dst|$dst, $src2}",
1835 (or (bc_v2i64 (v2f64 VR128:$src1)),
1836 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1837 def XORPDrr : PDI<0x57, MRMSrcReg,
1838 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1839 "xorpd\t{$src2, $dst|$dst, $src2}",
1841 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1842 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1845 def ANDPDrm : PDI<0x54, MRMSrcMem,
1846 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1847 "andpd\t{$src2, $dst|$dst, $src2}",
1849 (and (bc_v2i64 (v2f64 VR128:$src1)),
1850 (memopv2i64 addr:$src2)))]>;
1851 def ORPDrm : PDI<0x56, MRMSrcMem,
1852 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1853 "orpd\t{$src2, $dst|$dst, $src2}",
1855 (or (bc_v2i64 (v2f64 VR128:$src1)),
1856 (memopv2i64 addr:$src2)))]>;
1857 def XORPDrm : PDI<0x57, MRMSrcMem,
1858 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1859 "xorpd\t{$src2, $dst|$dst, $src2}",
1861 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1862 (memopv2i64 addr:$src2)))]>;
1863 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1864 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1865 "andnpd\t{$src2, $dst|$dst, $src2}",
1867 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1868 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1869 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1870 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1871 "andnpd\t{$src2, $dst|$dst, $src2}",
1873 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1874 (memopv2i64 addr:$src2)))]>;
1877 let Constraints = "$src1 = $dst" in {
1878 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1879 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1880 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1881 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1882 VR128:$src, imm:$cc))]>;
1883 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1884 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1885 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1886 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1887 (memop addr:$src), imm:$cc))]>;
1889 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1890 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1891 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1892 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1894 // Shuffle and unpack instructions
1895 let Constraints = "$src1 = $dst" in {
1896 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1897 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1898 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1900 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1901 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1902 (outs VR128:$dst), (ins VR128:$src1,
1903 f128mem:$src2, i8imm:$src3),
1904 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1907 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1909 let AddedComplexity = 10 in {
1910 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1911 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1912 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1914 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1915 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1916 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1917 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1919 (v2f64 (unpckh VR128:$src1,
1920 (memopv2f64 addr:$src2))))]>;
1922 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1923 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1924 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1926 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1927 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1928 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1929 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1931 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1932 } // AddedComplexity
1933 } // Constraints = "$src1 = $dst"
1936 //===---------------------------------------------------------------------===//
1937 // SSE integer instructions
1939 // Move Instructions
1940 let neverHasSideEffects = 1 in
1941 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1942 "movdqa\t{$src, $dst|$dst, $src}", []>;
1943 let canFoldAsLoad = 1, mayLoad = 1 in
1944 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1945 "movdqa\t{$src, $dst|$dst, $src}",
1946 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1948 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1949 "movdqa\t{$src, $dst|$dst, $src}",
1950 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1951 let canFoldAsLoad = 1, mayLoad = 1 in
1952 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1953 "movdqu\t{$src, $dst|$dst, $src}",
1954 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1955 XS, Requires<[HasSSE2]>;
1957 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1958 "movdqu\t{$src, $dst|$dst, $src}",
1959 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1960 XS, Requires<[HasSSE2]>;
1962 // Intrinsic forms of MOVDQU load and store
1963 let canFoldAsLoad = 1 in
1964 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1965 "movdqu\t{$src, $dst|$dst, $src}",
1966 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1967 XS, Requires<[HasSSE2]>;
1968 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1969 "movdqu\t{$src, $dst|$dst, $src}",
1970 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1971 XS, Requires<[HasSSE2]>;
1973 let Constraints = "$src1 = $dst" in {
1975 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1976 bit Commutable = 0> {
1977 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1978 (ins VR128:$src1, VR128:$src2),
1979 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1980 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1981 let isCommutable = Commutable;
1983 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1984 (ins VR128:$src1, i128mem:$src2),
1985 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1986 [(set VR128:$dst, (IntId VR128:$src1,
1987 (bitconvert (memopv2i64
1991 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1993 Intrinsic IntId, Intrinsic IntId2> {
1994 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1995 (ins VR128:$src1, VR128:$src2),
1996 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1997 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1998 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1999 (ins VR128:$src1, i128mem:$src2),
2000 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2001 [(set VR128:$dst, (IntId VR128:$src1,
2002 (bitconvert (memopv2i64 addr:$src2))))]>;
2003 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
2004 (ins VR128:$src1, i32i8imm:$src2),
2005 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2006 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
2009 /// PDI_binop_rm - Simple SSE2 binary operator.
2010 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
2011 ValueType OpVT, bit Commutable = 0> {
2012 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2013 (ins VR128:$src1, VR128:$src2),
2014 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2015 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
2016 let isCommutable = Commutable;
2018 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2019 (ins VR128:$src1, i128mem:$src2),
2020 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2021 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
2022 (bitconvert (memopv2i64 addr:$src2)))))]>;
2025 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
2027 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
2028 /// to collapse (bitconvert VT to VT) into its operand.
2030 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
2031 bit Commutable = 0> {
2032 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2033 (ins VR128:$src1, VR128:$src2),
2034 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2035 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
2036 let isCommutable = Commutable;
2038 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2039 (ins VR128:$src1, i128mem:$src2),
2040 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2041 [(set VR128:$dst, (OpNode VR128:$src1,
2042 (memopv2i64 addr:$src2)))]>;
2045 } // Constraints = "$src1 = $dst"
2047 // 128-bit Integer Arithmetic
2049 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
2050 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
2051 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
2052 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
2054 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
2055 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
2056 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
2057 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
2059 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
2060 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
2061 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
2062 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
2064 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
2065 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
2066 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
2067 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
2069 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
2071 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
2072 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
2073 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
2075 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
2077 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
2078 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
2081 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
2082 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
2083 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
2084 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
2085 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
2088 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
2089 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
2090 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
2091 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
2092 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
2093 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
2095 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2096 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2097 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2098 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2099 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2100 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2102 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2103 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2104 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2105 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2107 // 128-bit logical shifts.
2108 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
2109 def PSLLDQri : PDIi8<0x73, MRM7r,
2110 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2111 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2112 def PSRLDQri : PDIi8<0x73, MRM3r,
2113 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2114 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2115 // PSRADQri doesn't exist in SSE[1-3].
2118 let Predicates = [HasSSE2] in {
2119 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2120 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2121 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2122 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2123 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2124 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2125 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2126 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2127 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2128 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2130 // Shift up / down and insert zero's.
2131 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2132 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2133 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2134 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2138 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2139 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2140 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2142 let Constraints = "$src1 = $dst" in {
2143 def PANDNrr : PDI<0xDF, MRMSrcReg,
2144 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2145 "pandn\t{$src2, $dst|$dst, $src2}",
2146 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2149 def PANDNrm : PDI<0xDF, MRMSrcMem,
2150 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2151 "pandn\t{$src2, $dst|$dst, $src2}",
2152 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2153 (memopv2i64 addr:$src2))))]>;
2156 // SSE2 Integer comparison
2157 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2158 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2159 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2160 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2161 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2162 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2164 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2165 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2166 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2167 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2168 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2169 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2170 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2171 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2172 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2173 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2174 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2175 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2177 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2178 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2179 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2180 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2181 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2182 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2183 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2184 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2185 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2186 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2187 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2188 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2191 // Pack instructions
2192 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2193 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2194 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2196 // Shuffle and unpack instructions
2197 let AddedComplexity = 5 in {
2198 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2199 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2200 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2201 [(set VR128:$dst, (v4i32 (pshufd:$src2
2202 VR128:$src1, (undef))))]>;
2203 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2204 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2205 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2206 [(set VR128:$dst, (v4i32 (pshufd:$src2
2207 (bc_v4i32 (memopv2i64 addr:$src1)),
2211 // SSE2 with ImmT == Imm8 and XS prefix.
2212 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2213 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2214 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2215 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2217 XS, Requires<[HasSSE2]>;
2218 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2219 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2220 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2221 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2222 (bc_v8i16 (memopv2i64 addr:$src1)),
2224 XS, Requires<[HasSSE2]>;
2226 // SSE2 with ImmT == Imm8 and XD prefix.
2227 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2228 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2229 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2230 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2232 XD, Requires<[HasSSE2]>;
2233 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2234 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2235 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2236 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2237 (bc_v8i16 (memopv2i64 addr:$src1)),
2239 XD, Requires<[HasSSE2]>;
2242 let Constraints = "$src1 = $dst" in {
2243 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2244 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2245 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2247 (v16i8 (unpckl VR128:$src1, VR128:$src2)))]>;
2248 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2249 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2250 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2252 (unpckl VR128:$src1,
2253 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2254 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2255 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2256 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2258 (v8i16 (unpckl VR128:$src1, VR128:$src2)))]>;
2259 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2260 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2261 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2263 (unpckl VR128:$src1,
2264 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2265 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2266 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2267 "punpckldq\t{$src2, $dst|$dst, $src2}",
2269 (v4i32 (unpckl VR128:$src1, VR128:$src2)))]>;
2270 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2271 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2272 "punpckldq\t{$src2, $dst|$dst, $src2}",
2274 (unpckl VR128:$src1,
2275 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2276 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2277 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2278 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2280 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2281 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2282 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2283 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2285 (v2i64 (unpckl VR128:$src1,
2286 (memopv2i64 addr:$src2))))]>;
2288 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2289 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2290 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2292 (v16i8 (unpckh VR128:$src1, VR128:$src2)))]>;
2293 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2294 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2295 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2297 (unpckh VR128:$src1,
2298 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2299 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2300 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2301 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2303 (v8i16 (unpckh VR128:$src1, VR128:$src2)))]>;
2304 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2305 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2306 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2308 (unpckh VR128:$src1,
2309 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2310 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2311 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2312 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2314 (v4i32 (unpckh VR128:$src1, VR128:$src2)))]>;
2315 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2316 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2317 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2319 (unpckh VR128:$src1,
2320 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2321 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2322 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2323 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2325 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2326 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2327 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2328 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2330 (v2i64 (unpckh VR128:$src1,
2331 (memopv2i64 addr:$src2))))]>;
2335 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2336 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2337 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2338 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2340 let Constraints = "$src1 = $dst" in {
2341 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2342 (outs VR128:$dst), (ins VR128:$src1,
2343 GR32:$src2, i32i8imm:$src3),
2344 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2346 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2347 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2348 (outs VR128:$dst), (ins VR128:$src1,
2349 i16mem:$src2, i32i8imm:$src3),
2350 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2352 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2357 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2358 "pmovmskb\t{$src, $dst|$dst, $src}",
2359 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2361 // Conditional store
2363 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2364 "maskmovdqu\t{$mask, $src|$src, $mask}",
2365 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2368 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2369 "maskmovdqu\t{$mask, $src|$src, $mask}",
2370 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2372 // Non-temporal stores
2373 def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2374 "movntpd\t{$src, $dst|$dst, $src}",
2375 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2376 def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2377 "movntdq\t{$src, $dst|$dst, $src}",
2378 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2379 def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2380 "movnti\t{$src, $dst|$dst, $src}",
2381 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2382 TB, Requires<[HasSSE2]>;
2384 let AddedComplexity = 400 in { // Prefer non-temporal versions
2385 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2386 "movntpd\t{$src, $dst|$dst, $src}",
2387 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
2389 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2390 "movntdq\t{$src, $dst|$dst, $src}",
2391 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2395 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2396 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2397 TB, Requires<[HasSSE2]>;
2399 // Load, store, and memory fence
2400 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
2401 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2402 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
2403 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2405 //TODO: custom lower this so as to never even generate the noop
2406 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2408 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2409 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2410 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2413 // Alias instructions that map zero vector to pxor / xorp* for sse.
2414 // We set canFoldAsLoad because this can be converted to a constant-pool
2415 // load of an all-ones value if folding it would be beneficial.
2416 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2417 isCodeGenOnly = 1 in
2418 // FIXME: Change encoding to pseudo.
2419 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
2420 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2422 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2423 "movd\t{$src, $dst|$dst, $src}",
2425 (v4i32 (scalar_to_vector GR32:$src)))]>;
2426 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2427 "movd\t{$src, $dst|$dst, $src}",
2429 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2431 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2432 "movd\t{$src, $dst|$dst, $src}",
2433 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2435 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2436 "movd\t{$src, $dst|$dst, $src}",
2437 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2439 // SSE2 instructions with XS prefix
2440 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2441 "movq\t{$src, $dst|$dst, $src}",
2443 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2444 Requires<[HasSSE2]>;
2445 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2446 "movq\t{$src, $dst|$dst, $src}",
2447 [(store (i64 (vector_extract (v2i64 VR128:$src),
2448 (iPTR 0))), addr:$dst)]>;
2450 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
2451 (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), x86_subreg_sd))>;
2453 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2454 "movd\t{$src, $dst|$dst, $src}",
2455 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2457 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2458 "movd\t{$src, $dst|$dst, $src}",
2459 [(store (i32 (vector_extract (v4i32 VR128:$src),
2460 (iPTR 0))), addr:$dst)]>;
2462 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2463 "movd\t{$src, $dst|$dst, $src}",
2464 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2465 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2466 "movd\t{$src, $dst|$dst, $src}",
2467 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2469 // Store / copy lower 64-bits of a XMM register.
2470 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2471 "movq\t{$src, $dst|$dst, $src}",
2472 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2474 // movd / movq to XMM register zero-extends
2475 let AddedComplexity = 15 in {
2476 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2477 "movd\t{$src, $dst|$dst, $src}",
2478 [(set VR128:$dst, (v4i32 (X86vzmovl
2479 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2480 // This is X86-64 only.
2481 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2482 "mov{d|q}\t{$src, $dst|$dst, $src}",
2483 [(set VR128:$dst, (v2i64 (X86vzmovl
2484 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2487 let AddedComplexity = 20 in {
2488 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2489 "movd\t{$src, $dst|$dst, $src}",
2491 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2492 (loadi32 addr:$src))))))]>;
2494 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2495 (MOVZDI2PDIrm addr:$src)>;
2496 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2497 (MOVZDI2PDIrm addr:$src)>;
2498 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2499 (MOVZDI2PDIrm addr:$src)>;
2501 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2502 "movq\t{$src, $dst|$dst, $src}",
2504 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2505 (loadi64 addr:$src))))))]>, XS,
2506 Requires<[HasSSE2]>;
2508 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2509 (MOVZQI2PQIrm addr:$src)>;
2510 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2511 (MOVZQI2PQIrm addr:$src)>;
2512 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2515 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2516 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2517 let AddedComplexity = 15 in
2518 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2519 "movq\t{$src, $dst|$dst, $src}",
2520 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2521 XS, Requires<[HasSSE2]>;
2523 let AddedComplexity = 20 in {
2524 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2525 "movq\t{$src, $dst|$dst, $src}",
2526 [(set VR128:$dst, (v2i64 (X86vzmovl
2527 (loadv2i64 addr:$src))))]>,
2528 XS, Requires<[HasSSE2]>;
2530 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2531 (MOVZPQILo2PQIrm addr:$src)>;
2534 // Instructions for the disassembler
2535 // xr = XMM register
2538 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2539 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2541 //===---------------------------------------------------------------------===//
2542 // SSE3 Instructions
2543 //===---------------------------------------------------------------------===//
2545 // Move Instructions
2546 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2547 "movshdup\t{$src, $dst|$dst, $src}",
2548 [(set VR128:$dst, (v4f32 (movshdup
2549 VR128:$src, (undef))))]>;
2550 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2551 "movshdup\t{$src, $dst|$dst, $src}",
2552 [(set VR128:$dst, (movshdup
2553 (memopv4f32 addr:$src), (undef)))]>;
2555 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2556 "movsldup\t{$src, $dst|$dst, $src}",
2557 [(set VR128:$dst, (v4f32 (movsldup
2558 VR128:$src, (undef))))]>;
2559 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2560 "movsldup\t{$src, $dst|$dst, $src}",
2561 [(set VR128:$dst, (movsldup
2562 (memopv4f32 addr:$src), (undef)))]>;
2564 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2565 "movddup\t{$src, $dst|$dst, $src}",
2566 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2567 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2568 "movddup\t{$src, $dst|$dst, $src}",
2570 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2573 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2575 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2577 let AddedComplexity = 5 in {
2578 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2579 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2580 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2581 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2582 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2583 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2584 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2585 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2589 let Constraints = "$src1 = $dst" in {
2590 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2591 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2592 "addsubps\t{$src2, $dst|$dst, $src2}",
2593 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2595 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2596 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2597 "addsubps\t{$src2, $dst|$dst, $src2}",
2598 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2599 (memop addr:$src2)))]>;
2600 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2601 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2602 "addsubpd\t{$src2, $dst|$dst, $src2}",
2603 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2605 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2606 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2607 "addsubpd\t{$src2, $dst|$dst, $src2}",
2608 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2609 (memop addr:$src2)))]>;
2612 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2613 "lddqu\t{$src, $dst|$dst, $src}",
2614 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2617 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2618 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2619 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2620 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2621 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2622 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2623 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2624 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2625 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2626 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2627 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2628 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2629 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2630 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2631 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2632 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2634 let Constraints = "$src1 = $dst" in {
2635 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2636 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2637 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2638 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2639 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2640 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2641 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2642 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2645 // Thread synchronization
2646 def MONITOR : I<0x01, MRM_C8, (outs), (ins), "monitor",
2647 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2648 def MWAIT : I<0x01, MRM_C9, (outs), (ins), "mwait",
2649 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2651 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2652 let AddedComplexity = 15 in
2653 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2654 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2655 let AddedComplexity = 20 in
2656 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2657 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2659 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2660 let AddedComplexity = 15 in
2661 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2662 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2663 let AddedComplexity = 20 in
2664 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2665 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2667 //===---------------------------------------------------------------------===//
2668 // SSSE3 Instructions
2669 //===---------------------------------------------------------------------===//
2671 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2672 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2673 Intrinsic IntId64, Intrinsic IntId128> {
2674 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2675 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2676 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2678 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2679 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2681 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2683 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2685 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2686 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2689 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2691 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2694 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2697 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2698 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2699 Intrinsic IntId64, Intrinsic IntId128> {
2700 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2702 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2703 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2705 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2707 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2710 (bitconvert (memopv4i16 addr:$src))))]>;
2712 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2714 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2715 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2718 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2720 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2723 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2726 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2727 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2728 Intrinsic IntId64, Intrinsic IntId128> {
2729 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2731 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2732 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2734 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2736 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2739 (bitconvert (memopv2i32 addr:$src))))]>;
2741 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2743 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2744 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2747 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2749 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2752 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2755 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2756 int_x86_ssse3_pabs_b,
2757 int_x86_ssse3_pabs_b_128>;
2758 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2759 int_x86_ssse3_pabs_w,
2760 int_x86_ssse3_pabs_w_128>;
2761 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2762 int_x86_ssse3_pabs_d,
2763 int_x86_ssse3_pabs_d_128>;
2765 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2766 let Constraints = "$src1 = $dst" in {
2767 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2768 Intrinsic IntId64, Intrinsic IntId128,
2769 bit Commutable = 0> {
2770 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2771 (ins VR64:$src1, VR64:$src2),
2772 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2773 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2774 let isCommutable = Commutable;
2776 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2777 (ins VR64:$src1, i64mem:$src2),
2778 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2780 (IntId64 VR64:$src1,
2781 (bitconvert (memopv8i8 addr:$src2))))]>;
2783 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2784 (ins VR128:$src1, VR128:$src2),
2785 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2786 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2788 let isCommutable = Commutable;
2790 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2791 (ins VR128:$src1, i128mem:$src2),
2792 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2794 (IntId128 VR128:$src1,
2795 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2799 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2800 let Constraints = "$src1 = $dst" in {
2801 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2802 Intrinsic IntId64, Intrinsic IntId128,
2803 bit Commutable = 0> {
2804 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2805 (ins VR64:$src1, VR64:$src2),
2806 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2807 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2808 let isCommutable = Commutable;
2810 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2811 (ins VR64:$src1, i64mem:$src2),
2812 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2814 (IntId64 VR64:$src1,
2815 (bitconvert (memopv4i16 addr:$src2))))]>;
2817 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2818 (ins VR128:$src1, VR128:$src2),
2819 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2820 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2822 let isCommutable = Commutable;
2824 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2825 (ins VR128:$src1, i128mem:$src2),
2826 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2828 (IntId128 VR128:$src1,
2829 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2833 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2834 let Constraints = "$src1 = $dst" in {
2835 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2836 Intrinsic IntId64, Intrinsic IntId128,
2837 bit Commutable = 0> {
2838 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2839 (ins VR64:$src1, VR64:$src2),
2840 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2841 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2842 let isCommutable = Commutable;
2844 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2845 (ins VR64:$src1, i64mem:$src2),
2846 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2848 (IntId64 VR64:$src1,
2849 (bitconvert (memopv2i32 addr:$src2))))]>;
2851 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2852 (ins VR128:$src1, VR128:$src2),
2853 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2854 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2856 let isCommutable = Commutable;
2858 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2859 (ins VR128:$src1, i128mem:$src2),
2860 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2862 (IntId128 VR128:$src1,
2863 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2867 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2868 int_x86_ssse3_phadd_w,
2869 int_x86_ssse3_phadd_w_128>;
2870 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2871 int_x86_ssse3_phadd_d,
2872 int_x86_ssse3_phadd_d_128>;
2873 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2874 int_x86_ssse3_phadd_sw,
2875 int_x86_ssse3_phadd_sw_128>;
2876 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2877 int_x86_ssse3_phsub_w,
2878 int_x86_ssse3_phsub_w_128>;
2879 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2880 int_x86_ssse3_phsub_d,
2881 int_x86_ssse3_phsub_d_128>;
2882 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2883 int_x86_ssse3_phsub_sw,
2884 int_x86_ssse3_phsub_sw_128>;
2885 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2886 int_x86_ssse3_pmadd_ub_sw,
2887 int_x86_ssse3_pmadd_ub_sw_128>;
2888 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2889 int_x86_ssse3_pmul_hr_sw,
2890 int_x86_ssse3_pmul_hr_sw_128, 1>;
2891 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2892 int_x86_ssse3_pshuf_b,
2893 int_x86_ssse3_pshuf_b_128>;
2894 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2895 int_x86_ssse3_psign_b,
2896 int_x86_ssse3_psign_b_128>;
2897 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2898 int_x86_ssse3_psign_w,
2899 int_x86_ssse3_psign_w_128>;
2900 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2901 int_x86_ssse3_psign_d,
2902 int_x86_ssse3_psign_d_128>;
2904 let Constraints = "$src1 = $dst" in {
2905 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2906 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2907 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2909 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2910 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2911 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2914 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2915 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2916 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2918 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2919 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2920 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2924 // palignr patterns.
2925 def : Pat<(int_x86_ssse3_palign_r VR64:$src1, VR64:$src2, (i8 imm:$src3)),
2926 (PALIGNR64rr VR64:$src1, VR64:$src2, (BYTE_imm imm:$src3))>,
2927 Requires<[HasSSSE3]>;
2928 def : Pat<(int_x86_ssse3_palign_r VR64:$src1,
2929 (memop64 addr:$src2),
2931 (PALIGNR64rm VR64:$src1, addr:$src2, (BYTE_imm imm:$src3))>,
2932 Requires<[HasSSSE3]>;
2934 def : Pat<(int_x86_ssse3_palign_r_128 VR128:$src1, VR128:$src2, (i8 imm:$src3)),
2935 (PALIGNR128rr VR128:$src1, VR128:$src2, (BYTE_imm imm:$src3))>,
2936 Requires<[HasSSSE3]>;
2937 def : Pat<(int_x86_ssse3_palign_r_128 VR128:$src1,
2938 (memopv2i64 addr:$src2),
2940 (PALIGNR128rm VR128:$src1, addr:$src2, (BYTE_imm imm:$src3))>,
2941 Requires<[HasSSSE3]>;
2943 let AddedComplexity = 5 in {
2944 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2945 (PALIGNR128rr VR128:$src2, VR128:$src1,
2946 (SHUFFLE_get_palign_imm VR128:$src3))>,
2947 Requires<[HasSSSE3]>;
2948 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2949 (PALIGNR128rr VR128:$src2, VR128:$src1,
2950 (SHUFFLE_get_palign_imm VR128:$src3))>,
2951 Requires<[HasSSSE3]>;
2952 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2953 (PALIGNR128rr VR128:$src2, VR128:$src1,
2954 (SHUFFLE_get_palign_imm VR128:$src3))>,
2955 Requires<[HasSSSE3]>;
2956 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2957 (PALIGNR128rr VR128:$src2, VR128:$src1,
2958 (SHUFFLE_get_palign_imm VR128:$src3))>,
2959 Requires<[HasSSSE3]>;
2962 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2963 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2964 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2965 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2967 //===---------------------------------------------------------------------===//
2968 // Non-Instruction Patterns
2969 //===---------------------------------------------------------------------===//
2971 // extload f32 -> f64. This matches load+fextend because we have a hack in
2972 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2974 // Since these loads aren't folded into the fextend, we have to match it
2976 let Predicates = [HasSSE2] in
2977 def : Pat<(fextend (loadf32 addr:$src)),
2978 (CVTSS2SDrm addr:$src)>;
2981 let Predicates = [HasSSE2] in {
2982 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2983 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2984 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2985 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2986 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2987 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2988 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2989 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2990 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2991 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2992 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2993 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2994 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2995 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2996 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2997 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2998 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2999 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
3000 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
3001 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
3002 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
3003 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
3004 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
3005 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
3006 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
3007 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
3008 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
3009 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
3010 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
3011 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
3014 // Move scalar to XMM zero-extended
3015 // movd to XMM register zero-extends
3016 let AddedComplexity = 15 in {
3017 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
3018 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
3019 (MOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
3020 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
3021 (MOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
3022 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
3023 (MOVSSrr (v4f32 (V_SET0)),
3024 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), x86_subreg_ss)))>;
3025 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
3026 (MOVSSrr (v4i32 (V_SET0)),
3027 (EXTRACT_SUBREG (v4i32 VR128:$src), x86_subreg_ss))>;
3030 // Splat v2f64 / v2i64
3031 let AddedComplexity = 10 in {
3032 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
3033 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3034 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
3035 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3036 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
3037 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3038 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
3039 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3042 // Special unary SHUFPSrri case.
3043 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
3044 (SHUFPSrri VR128:$src1, VR128:$src1,
3045 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3046 let AddedComplexity = 5 in
3047 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
3048 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3049 Requires<[HasSSE2]>;
3050 // Special unary SHUFPDrri case.
3051 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
3052 (SHUFPDrri VR128:$src1, VR128:$src1,
3053 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3054 Requires<[HasSSE2]>;
3055 // Special unary SHUFPDrri case.
3056 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
3057 (SHUFPDrri VR128:$src1, VR128:$src1,
3058 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3059 Requires<[HasSSE2]>;
3060 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
3061 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
3062 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3063 Requires<[HasSSE2]>;
3065 // Special binary v4i32 shuffle cases with SHUFPS.
3066 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
3067 (SHUFPSrri VR128:$src1, VR128:$src2,
3068 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3069 Requires<[HasSSE2]>;
3070 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
3071 (SHUFPSrmi VR128:$src1, addr:$src2,
3072 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3073 Requires<[HasSSE2]>;
3074 // Special binary v2i64 shuffle cases using SHUFPDrri.
3075 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
3076 (SHUFPDrri VR128:$src1, VR128:$src2,
3077 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3078 Requires<[HasSSE2]>;
3080 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
3081 let AddedComplexity = 15 in {
3082 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
3083 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3084 Requires<[OptForSpeed, HasSSE2]>;
3085 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
3086 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3087 Requires<[OptForSpeed, HasSSE2]>;
3089 let AddedComplexity = 10 in {
3090 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
3091 (UNPCKLPSrr VR128:$src, VR128:$src)>;
3092 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
3093 (PUNPCKLBWrr VR128:$src, VR128:$src)>;
3094 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
3095 (PUNPCKLWDrr VR128:$src, VR128:$src)>;
3096 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
3097 (PUNPCKLDQrr VR128:$src, VR128:$src)>;
3100 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
3101 let AddedComplexity = 15 in {
3102 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
3103 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3104 Requires<[OptForSpeed, HasSSE2]>;
3105 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
3106 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3107 Requires<[OptForSpeed, HasSSE2]>;
3109 let AddedComplexity = 10 in {
3110 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
3111 (UNPCKHPSrr VR128:$src, VR128:$src)>;
3112 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
3113 (PUNPCKHBWrr VR128:$src, VR128:$src)>;
3114 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
3115 (PUNPCKHWDrr VR128:$src, VR128:$src)>;
3116 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
3117 (PUNPCKHDQrr VR128:$src, VR128:$src)>;
3120 let AddedComplexity = 20 in {
3121 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3122 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3123 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3125 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3126 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3127 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3129 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3130 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3131 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3132 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3133 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3136 let AddedComplexity = 20 in {
3137 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3138 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3139 (MOVLPSrm VR128:$src1, addr:$src2)>;
3140 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3141 (MOVLPDrm VR128:$src1, addr:$src2)>;
3142 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3143 (MOVLPSrm VR128:$src1, addr:$src2)>;
3144 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3145 (MOVLPDrm VR128:$src1, addr:$src2)>;
3148 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3149 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3150 (MOVLPSmr addr:$src1, VR128:$src2)>;
3151 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3152 (MOVLPDmr addr:$src1, VR128:$src2)>;
3153 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3155 (MOVLPSmr addr:$src1, VR128:$src2)>;
3156 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3157 (MOVLPDmr addr:$src1, VR128:$src2)>;
3159 let AddedComplexity = 15 in {
3160 // Setting the lowest element in the vector.
3161 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3162 (MOVSSrr (v4i32 VR128:$src1),
3163 (EXTRACT_SUBREG (v4i32 VR128:$src2), x86_subreg_ss))>;
3164 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3165 (MOVSDrr (v2i64 VR128:$src1),
3166 (EXTRACT_SUBREG (v2i64 VR128:$src2), x86_subreg_sd))>;
3168 // vector_shuffle v1, v2 <4, 5, 2, 3> using movsd
3169 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3170 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, x86_subreg_sd))>,
3171 Requires<[HasSSE2]>;
3172 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3173 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, x86_subreg_sd))>,
3174 Requires<[HasSSE2]>;
3177 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3178 // fall back to this for SSE1)
3179 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3180 (SHUFPSrri VR128:$src2, VR128:$src1,
3181 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3183 // Set lowest element and zero upper elements.
3184 let AddedComplexity = 15 in
3185 def : Pat<(v2f64 (movl immAllZerosV_bc, VR128:$src)),
3186 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3187 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3188 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3190 // Some special case pandn patterns.
3191 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3193 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3194 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3196 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3197 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3199 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3201 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3202 (memop addr:$src2))),
3203 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3204 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3205 (memop addr:$src2))),
3206 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3207 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3208 (memop addr:$src2))),
3209 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3211 // vector -> vector casts
3212 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3213 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3214 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3215 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3216 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3217 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3218 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3219 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3221 // Use movaps / movups for SSE integer load / store (one byte shorter).
3222 def : Pat<(alignedloadv4i32 addr:$src),
3223 (MOVAPSrm addr:$src)>;
3224 def : Pat<(loadv4i32 addr:$src),
3225 (MOVUPSrm addr:$src)>;
3226 def : Pat<(alignedloadv2i64 addr:$src),
3227 (MOVAPSrm addr:$src)>;
3228 def : Pat<(loadv2i64 addr:$src),
3229 (MOVUPSrm addr:$src)>;
3231 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3232 (MOVAPSmr addr:$dst, VR128:$src)>;
3233 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3234 (MOVAPSmr addr:$dst, VR128:$src)>;
3235 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3236 (MOVAPSmr addr:$dst, VR128:$src)>;
3237 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3238 (MOVAPSmr addr:$dst, VR128:$src)>;
3239 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3240 (MOVUPSmr addr:$dst, VR128:$src)>;
3241 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3242 (MOVUPSmr addr:$dst, VR128:$src)>;
3243 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3244 (MOVUPSmr addr:$dst, VR128:$src)>;
3245 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3246 (MOVUPSmr addr:$dst, VR128:$src)>;
3248 //===----------------------------------------------------------------------===//
3249 // SSE4.1 Instructions
3250 //===----------------------------------------------------------------------===//
3252 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3255 Intrinsic V2F64Int> {
3256 // Intrinsic operation, reg.
3257 // Vector intrinsic operation, reg
3258 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3259 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3260 !strconcat(OpcodeStr,
3261 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3262 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3265 // Vector intrinsic operation, mem
3266 def PSm_Int : Ii8<opcps, MRMSrcMem,
3267 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3268 !strconcat(OpcodeStr,
3269 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3271 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3273 Requires<[HasSSE41]>;
3275 // Vector intrinsic operation, reg
3276 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3277 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3278 !strconcat(OpcodeStr,
3279 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3280 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3283 // Vector intrinsic operation, mem
3284 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3285 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3286 !strconcat(OpcodeStr,
3287 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3289 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3293 let Constraints = "$src1 = $dst" in {
3294 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3298 // Intrinsic operation, reg.
3299 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3301 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3302 !strconcat(OpcodeStr,
3303 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3305 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3308 // Intrinsic operation, mem.
3309 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3311 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3312 !strconcat(OpcodeStr,
3313 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3315 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3318 // Intrinsic operation, reg.
3319 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3321 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3322 !strconcat(OpcodeStr,
3323 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3325 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3328 // Intrinsic operation, mem.
3329 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3331 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3332 !strconcat(OpcodeStr,
3333 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3335 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3340 // FP round - roundss, roundps, roundsd, roundpd
3341 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3342 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3343 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3344 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3346 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3347 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3348 Intrinsic IntId128> {
3349 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3351 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3352 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3353 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3355 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3358 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3361 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3362 int_x86_sse41_phminposuw>;
3364 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3365 let Constraints = "$src1 = $dst" in {
3366 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3367 Intrinsic IntId128, bit Commutable = 0> {
3368 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3369 (ins VR128:$src1, VR128:$src2),
3370 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3371 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3373 let isCommutable = Commutable;
3375 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3376 (ins VR128:$src1, i128mem:$src2),
3377 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3379 (IntId128 VR128:$src1,
3380 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3384 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3385 int_x86_sse41_pcmpeqq, 1>;
3386 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3387 int_x86_sse41_packusdw, 0>;
3388 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3389 int_x86_sse41_pminsb, 1>;
3390 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3391 int_x86_sse41_pminsd, 1>;
3392 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3393 int_x86_sse41_pminud, 1>;
3394 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3395 int_x86_sse41_pminuw, 1>;
3396 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3397 int_x86_sse41_pmaxsb, 1>;
3398 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3399 int_x86_sse41_pmaxsd, 1>;
3400 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3401 int_x86_sse41_pmaxud, 1>;
3402 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3403 int_x86_sse41_pmaxuw, 1>;
3405 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3407 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3408 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3409 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3410 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3412 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3413 let Constraints = "$src1 = $dst" in {
3414 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3415 SDNode OpNode, Intrinsic IntId128,
3416 bit Commutable = 0> {
3417 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3418 (ins VR128:$src1, VR128:$src2),
3419 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3420 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3421 VR128:$src2))]>, OpSize {
3422 let isCommutable = Commutable;
3424 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3425 (ins VR128:$src1, VR128:$src2),
3426 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3427 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3429 let isCommutable = Commutable;
3431 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3432 (ins VR128:$src1, i128mem:$src2),
3433 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3435 (OpVT (OpNode VR128:$src1, (memop addr:$src2))))]>, OpSize;
3436 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3437 (ins VR128:$src1, i128mem:$src2),
3438 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3440 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3444 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", v4i32, mul,
3445 int_x86_sse41_pmulld, 1>;
3447 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3448 let Constraints = "$src1 = $dst" in {
3449 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3450 Intrinsic IntId128, bit Commutable = 0> {
3451 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3452 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3453 !strconcat(OpcodeStr,
3454 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3456 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3458 let isCommutable = Commutable;
3460 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3461 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3462 !strconcat(OpcodeStr,
3463 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3465 (IntId128 VR128:$src1,
3466 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3471 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3472 int_x86_sse41_blendps, 0>;
3473 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3474 int_x86_sse41_blendpd, 0>;
3475 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3476 int_x86_sse41_pblendw, 0>;
3477 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3478 int_x86_sse41_dpps, 1>;
3479 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3480 int_x86_sse41_dppd, 1>;
3481 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3482 int_x86_sse41_mpsadbw, 1>;
3485 /// SS41I_ternary_int - SSE 4.1 ternary operator
3486 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3487 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3488 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3489 (ins VR128:$src1, VR128:$src2),
3490 !strconcat(OpcodeStr,
3491 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3492 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3495 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3496 (ins VR128:$src1, i128mem:$src2),
3497 !strconcat(OpcodeStr,
3498 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3501 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3505 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3506 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3507 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3510 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3511 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3512 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3513 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3515 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3516 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3518 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3522 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3523 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3524 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3525 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3526 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3527 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3529 // Common patterns involving scalar load.
3530 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3531 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3532 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3533 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3535 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3536 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3537 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3538 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3540 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3541 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3542 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3543 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3545 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3546 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3547 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3548 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3550 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3551 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3552 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3553 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3555 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3556 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3557 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3558 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3561 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3562 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3563 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3564 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3566 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3567 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3569 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3573 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3574 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3575 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3576 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3578 // Common patterns involving scalar load
3579 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3580 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3581 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3582 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3584 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3585 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3586 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3587 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3590 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3591 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3592 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3593 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3595 // Expecting a i16 load any extended to i32 value.
3596 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3597 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3598 [(set VR128:$dst, (IntId (bitconvert
3599 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3603 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3604 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3606 // Common patterns involving scalar load
3607 def : Pat<(int_x86_sse41_pmovsxbq
3608 (bitconvert (v4i32 (X86vzmovl
3609 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3610 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3612 def : Pat<(int_x86_sse41_pmovzxbq
3613 (bitconvert (v4i32 (X86vzmovl
3614 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3615 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3618 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3619 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3620 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3621 (ins VR128:$src1, i32i8imm:$src2),
3622 !strconcat(OpcodeStr,
3623 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3624 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3626 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3627 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3628 !strconcat(OpcodeStr,
3629 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3632 // There's an AssertZext in the way of writing the store pattern
3633 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3636 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3639 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3640 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3641 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3642 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3643 !strconcat(OpcodeStr,
3644 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3647 // There's an AssertZext in the way of writing the store pattern
3648 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3651 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3654 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3655 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3656 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3657 (ins VR128:$src1, i32i8imm:$src2),
3658 !strconcat(OpcodeStr,
3659 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3661 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3662 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3663 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3664 !strconcat(OpcodeStr,
3665 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3666 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3667 addr:$dst)]>, OpSize;
3670 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3673 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3675 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3676 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3677 (ins VR128:$src1, i32i8imm:$src2),
3678 !strconcat(OpcodeStr,
3679 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3681 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3683 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3684 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3685 !strconcat(OpcodeStr,
3686 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3687 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3688 addr:$dst)]>, OpSize;
3691 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3693 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3694 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3697 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3698 Requires<[HasSSE41]>;
3700 let Constraints = "$src1 = $dst" in {
3701 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3702 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3703 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3704 !strconcat(OpcodeStr,
3705 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3707 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3708 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3709 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3710 !strconcat(OpcodeStr,
3711 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3713 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3714 imm:$src3))]>, OpSize;
3718 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3720 let Constraints = "$src1 = $dst" in {
3721 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3722 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3723 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3724 !strconcat(OpcodeStr,
3725 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3727 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3729 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3730 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3731 !strconcat(OpcodeStr,
3732 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3734 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3735 imm:$src3)))]>, OpSize;
3739 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3741 // insertps has a few different modes, there's the first two here below which
3742 // are optimized inserts that won't zero arbitrary elements in the destination
3743 // vector. The next one matches the intrinsic and could zero arbitrary elements
3744 // in the target vector.
3745 let Constraints = "$src1 = $dst" in {
3746 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3747 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3748 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3749 !strconcat(OpcodeStr,
3750 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3752 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3754 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3755 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3756 !strconcat(OpcodeStr,
3757 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3759 (X86insrtps VR128:$src1,
3760 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3761 imm:$src3))]>, OpSize;
3765 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3767 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3768 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3770 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3771 // the intel intrinsic that corresponds to this.
3772 let Defs = [EFLAGS] in {
3773 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3774 "ptest \t{$src2, $src1|$src1, $src2}",
3775 [(X86ptest VR128:$src1, VR128:$src2),
3776 (implicit EFLAGS)]>, OpSize;
3777 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3778 "ptest \t{$src2, $src1|$src1, $src2}",
3779 [(X86ptest VR128:$src1, (load addr:$src2)),
3780 (implicit EFLAGS)]>, OpSize;
3783 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3784 "movntdqa\t{$src, $dst|$dst, $src}",
3785 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
3789 //===----------------------------------------------------------------------===//
3790 // SSE4.2 Instructions
3791 //===----------------------------------------------------------------------===//
3793 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3794 let Constraints = "$src1 = $dst" in {
3795 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3796 Intrinsic IntId128, bit Commutable = 0> {
3797 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3798 (ins VR128:$src1, VR128:$src2),
3799 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3800 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3802 let isCommutable = Commutable;
3804 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3805 (ins VR128:$src1, i128mem:$src2),
3806 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3808 (IntId128 VR128:$src1,
3809 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3813 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3815 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3816 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3817 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3818 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3820 // crc intrinsic instruction
3821 // This set of instructions are only rm, the only difference is the size
3823 let Constraints = "$src1 = $dst" in {
3824 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3825 (ins GR32:$src1, i8mem:$src2),
3826 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3828 (int_x86_sse42_crc32_8 GR32:$src1,
3829 (load addr:$src2)))]>;
3830 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3831 (ins GR32:$src1, GR8:$src2),
3832 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3834 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>;
3835 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3836 (ins GR32:$src1, i16mem:$src2),
3837 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3839 (int_x86_sse42_crc32_16 GR32:$src1,
3840 (load addr:$src2)))]>,
3842 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3843 (ins GR32:$src1, GR16:$src2),
3844 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3846 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3848 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3849 (ins GR32:$src1, i32mem:$src2),
3850 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3852 (int_x86_sse42_crc32_32 GR32:$src1,
3853 (load addr:$src2)))]>;
3854 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3855 (ins GR32:$src1, GR32:$src2),
3856 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3858 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>;
3859 def CRC64m8 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3860 (ins GR64:$src1, i8mem:$src2),
3861 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3863 (int_x86_sse42_crc64_8 GR64:$src1,
3864 (load addr:$src2)))]>,
3866 def CRC64r8 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3867 (ins GR64:$src1, GR8:$src2),
3868 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3870 (int_x86_sse42_crc64_8 GR64:$src1, GR8:$src2))]>,
3872 def CRC64m64 : SS42FI<0xF1, MRMSrcMem, (outs GR64:$dst),
3873 (ins GR64:$src1, i64mem:$src2),
3874 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3876 (int_x86_sse42_crc64_64 GR64:$src1,
3877 (load addr:$src2)))]>,
3879 def CRC64r64 : SS42FI<0xF1, MRMSrcReg, (outs GR64:$dst),
3880 (ins GR64:$src1, GR64:$src2),
3881 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3883 (int_x86_sse42_crc64_64 GR64:$src1, GR64:$src2))]>,
3887 // String/text processing instructions.
3888 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3889 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3890 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3891 "#PCMPISTRM128rr PSEUDO!",
3892 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3893 imm:$src3))]>, OpSize;
3894 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3895 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3896 "#PCMPISTRM128rm PSEUDO!",
3897 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3898 imm:$src3))]>, OpSize;
3901 let Defs = [XMM0, EFLAGS] in {
3902 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3903 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3904 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3905 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3906 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3907 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3910 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3911 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3912 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3913 "#PCMPESTRM128rr PSEUDO!",
3915 (int_x86_sse42_pcmpestrm128
3916 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3918 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3919 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3920 "#PCMPESTRM128rm PSEUDO!",
3921 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3922 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3926 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3927 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3928 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3929 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3930 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3931 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3932 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3935 let Defs = [ECX, EFLAGS] in {
3936 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3937 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3938 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3939 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3940 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3941 (implicit EFLAGS)]>, OpSize;
3942 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3943 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3944 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3945 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3946 (implicit EFLAGS)]>, OpSize;
3950 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3951 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3952 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3953 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3954 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3955 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3957 let Defs = [ECX, EFLAGS] in {
3958 let Uses = [EAX, EDX] in {
3959 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3960 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3961 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3962 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3963 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3964 (implicit EFLAGS)]>, OpSize;
3965 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3966 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3967 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3969 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
3970 (implicit EFLAGS)]>, OpSize;
3975 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3976 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3977 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3978 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3979 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3980 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;