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, $src2}",
383 (movl VR128:$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:$src2)),
388 (MOVSSrr VR128:$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, $src2}",
516 [(X86cmp FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
517 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
518 "ucomiss\t{$src2, $src1|$src1, $src2}",
519 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
522 def COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
523 "comiss\t{$src2, $src1|$src1, $src2}", []>;
524 def COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
525 "comiss\t{$src2, $src1|$src1, $src2}", []>;
529 // Aliases to match intrinsics which expect XMM operand(s).
530 let Constraints = "$src1 = $dst" in {
531 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
533 (ins VR128:$src1, VR128:$src, SSECC:$cc),
534 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
535 [(set VR128:$dst, (int_x86_sse_cmp_ss
537 VR128:$src, imm:$cc))]>;
538 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
540 (ins VR128:$src1, f32mem:$src, SSECC:$cc),
541 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
542 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
543 (load addr:$src), imm:$cc))]>;
546 let Defs = [EFLAGS] in {
547 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
548 "ucomiss\t{$src2, $src1|$src1, $src2}",
549 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
551 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
552 "ucomiss\t{$src2, $src1|$src1, $src2}",
553 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
556 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
557 "comiss\t{$src2, $src1|$src1, $src2}",
558 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
560 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
561 "comiss\t{$src2, $src1|$src1, $src2}",
562 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
566 // Aliases of packed SSE1 instructions for scalar use. These all have names
567 // that start with 'Fs'.
569 // Alias instructions that map fld0 to pxor for sse.
570 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
572 // FIXME: Set encoding to pseudo!
573 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
574 [(set FR32:$dst, fp32imm0)]>,
575 Requires<[HasSSE1]>, TB, OpSize;
577 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
579 let neverHasSideEffects = 1 in
580 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
581 "movaps\t{$src, $dst|$dst, $src}", []>;
583 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
585 let canFoldAsLoad = 1, isReMaterializable = 1 in
586 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
587 "movaps\t{$src, $dst|$dst, $src}",
588 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
590 // Alias bitwise logical operations using SSE logical ops on packed FP values.
591 let Constraints = "$src1 = $dst" in {
592 let isCommutable = 1 in {
593 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst),
594 (ins FR32:$src1, FR32:$src2),
595 "andps\t{$src2, $dst|$dst, $src2}",
596 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
597 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst),
598 (ins FR32:$src1, FR32:$src2),
599 "orps\t{$src2, $dst|$dst, $src2}",
600 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
601 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst),
602 (ins FR32:$src1, FR32:$src2),
603 "xorps\t{$src2, $dst|$dst, $src2}",
604 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
607 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst),
608 (ins FR32:$src1, f128mem:$src2),
609 "andps\t{$src2, $dst|$dst, $src2}",
610 [(set FR32:$dst, (X86fand FR32:$src1,
611 (memopfsf32 addr:$src2)))]>;
612 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst),
613 (ins FR32:$src1, f128mem:$src2),
614 "orps\t{$src2, $dst|$dst, $src2}",
615 [(set FR32:$dst, (X86for FR32:$src1,
616 (memopfsf32 addr:$src2)))]>;
617 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst),
618 (ins FR32:$src1, f128mem:$src2),
619 "xorps\t{$src2, $dst|$dst, $src2}",
620 [(set FR32:$dst, (X86fxor FR32:$src1,
621 (memopfsf32 addr:$src2)))]>;
623 let neverHasSideEffects = 1 in {
624 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
625 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
626 "andnps\t{$src2, $dst|$dst, $src2}", []>;
628 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
629 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
630 "andnps\t{$src2, $dst|$dst, $src2}", []>;
634 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
636 /// In addition, we also have a special variant of the scalar form here to
637 /// represent the associated intrinsic operation. This form is unlike the
638 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
639 /// and leaves the top elements unmodified (therefore these cannot be commuted).
641 /// These three forms can each be reg+reg or reg+mem, so there are a total of
642 /// six "instructions".
644 let Constraints = "$src1 = $dst" in {
645 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
646 SDNode OpNode, Intrinsic F32Int,
647 bit Commutable = 0> {
648 // Scalar operation, reg+reg.
649 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
650 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
651 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
652 let isCommutable = Commutable;
655 // Scalar operation, reg+mem.
656 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
657 (ins FR32:$src1, f32mem:$src2),
658 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
659 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
661 // Vector operation, reg+reg.
662 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
663 (ins VR128:$src1, VR128:$src2),
664 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
665 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
666 let isCommutable = Commutable;
669 // Vector operation, reg+mem.
670 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
671 (ins VR128:$src1, f128mem:$src2),
672 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
673 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
675 // Intrinsic operation, reg+reg.
676 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
677 (ins VR128:$src1, VR128:$src2),
678 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
679 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]>;
681 // Intrinsic operation, reg+mem.
682 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
683 (ins VR128:$src1, ssmem:$src2),
684 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
685 [(set VR128:$dst, (F32Int VR128:$src1,
686 sse_load_f32:$src2))]>;
690 // Arithmetic instructions
691 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
692 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
693 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
694 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
696 /// sse1_fp_binop_rm - Other SSE1 binops
698 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
699 /// instructions for a full-vector intrinsic form. Operations that map
700 /// onto C operators don't use this form since they just use the plain
701 /// vector form instead of having a separate vector intrinsic form.
703 /// This provides a total of eight "instructions".
705 let Constraints = "$src1 = $dst" in {
706 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
710 bit Commutable = 0> {
712 // Scalar operation, reg+reg.
713 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
714 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
715 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
716 let isCommutable = Commutable;
719 // Scalar operation, reg+mem.
720 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
721 (ins FR32:$src1, f32mem:$src2),
722 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
723 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
725 // Vector operation, reg+reg.
726 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
727 (ins VR128:$src1, VR128:$src2),
728 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
729 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
730 let isCommutable = Commutable;
733 // Vector operation, reg+mem.
734 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
735 (ins VR128:$src1, f128mem:$src2),
736 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
737 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
739 // Intrinsic operation, reg+reg.
740 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
741 (ins VR128:$src1, VR128:$src2),
742 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
743 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
744 let isCommutable = Commutable;
747 // Intrinsic operation, reg+mem.
748 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
749 (ins VR128:$src1, ssmem:$src2),
750 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
751 [(set VR128:$dst, (F32Int VR128:$src1,
752 sse_load_f32:$src2))]>;
754 // Vector intrinsic operation, reg+reg.
755 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
756 (ins VR128:$src1, VR128:$src2),
757 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
758 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
759 let isCommutable = Commutable;
762 // Vector intrinsic operation, reg+mem.
763 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
764 (ins VR128:$src1, f128mem:$src2),
765 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
766 [(set VR128:$dst, (V4F32Int VR128:$src1, (memopv4f32 addr:$src2)))]>;
770 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
771 int_x86_sse_max_ss, int_x86_sse_max_ps>;
772 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
773 int_x86_sse_min_ss, int_x86_sse_min_ps>;
775 //===----------------------------------------------------------------------===//
776 // SSE packed FP Instructions
779 let neverHasSideEffects = 1 in
780 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
781 "movaps\t{$src, $dst|$dst, $src}", []>;
782 let canFoldAsLoad = 1, isReMaterializable = 1 in
783 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
784 "movaps\t{$src, $dst|$dst, $src}",
785 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
787 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
788 "movaps\t{$src, $dst|$dst, $src}",
789 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
791 let neverHasSideEffects = 1 in
792 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
793 "movups\t{$src, $dst|$dst, $src}", []>;
794 let canFoldAsLoad = 1, isReMaterializable = 1 in
795 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
796 "movups\t{$src, $dst|$dst, $src}",
797 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
798 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
799 "movups\t{$src, $dst|$dst, $src}",
800 [(store (v4f32 VR128:$src), addr:$dst)]>;
802 // Intrinsic forms of MOVUPS load and store
803 let canFoldAsLoad = 1, isReMaterializable = 1 in
804 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
805 "movups\t{$src, $dst|$dst, $src}",
806 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
807 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
808 "movups\t{$src, $dst|$dst, $src}",
809 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
811 let Constraints = "$src1 = $dst" in {
812 let AddedComplexity = 20 in {
813 def MOVLPSrm : PSI<0x12, MRMSrcMem,
814 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
815 "movlps\t{$src2, $dst|$dst, $src2}",
818 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
819 def MOVHPSrm : PSI<0x16, MRMSrcMem,
820 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
821 "movhps\t{$src2, $dst|$dst, $src2}",
823 (movlhps VR128:$src1,
824 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
826 } // Constraints = "$src1 = $dst"
829 def : Pat<(movlhps VR128:$src1, (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
830 (MOVHPSrm VR128:$src1, addr:$src2)>;
832 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
833 "movlps\t{$src, $dst|$dst, $src}",
834 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
835 (iPTR 0))), addr:$dst)]>;
837 // v2f64 extract element 1 is always custom lowered to unpack high to low
838 // and extract element 0 so the non-store version isn't too horrible.
839 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
840 "movhps\t{$src, $dst|$dst, $src}",
841 [(store (f64 (vector_extract
842 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
843 (undef)), (iPTR 0))), addr:$dst)]>;
845 let Constraints = "$src1 = $dst" in {
846 let AddedComplexity = 20 in {
847 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
848 (ins VR128:$src1, VR128:$src2),
849 "movlhps\t{$src2, $dst|$dst, $src2}",
851 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
853 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
854 (ins VR128:$src1, VR128:$src2),
855 "movhlps\t{$src2, $dst|$dst, $src2}",
857 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
859 } // Constraints = "$src1 = $dst"
861 let AddedComplexity = 20 in {
862 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
863 (MOVLHPSrr VR128:$src, VR128:$src)>;
864 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
865 (MOVLHPSrr VR128:$src, VR128:$src)>;
872 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
874 /// In addition, we also have a special variant of the scalar form here to
875 /// represent the associated intrinsic operation. This form is unlike the
876 /// plain scalar form, in that it takes an entire vector (instead of a
877 /// scalar) and leaves the top elements undefined.
879 /// And, we have a special variant form for a full-vector intrinsic form.
881 /// These four forms can each have a reg or a mem operand, so there are a
882 /// total of eight "instructions".
884 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
888 bit Commutable = 0> {
889 // Scalar operation, reg.
890 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
891 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
892 [(set FR32:$dst, (OpNode FR32:$src))]> {
893 let isCommutable = Commutable;
896 // Scalar operation, mem.
897 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
898 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
899 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
900 Requires<[HasSSE1, OptForSize]>;
902 // Vector operation, reg.
903 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
904 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
905 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
906 let isCommutable = Commutable;
909 // Vector operation, mem.
910 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
911 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
912 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
914 // Intrinsic operation, reg.
915 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
916 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
917 [(set VR128:$dst, (F32Int VR128:$src))]> {
918 let isCommutable = Commutable;
921 // Intrinsic operation, mem.
922 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
923 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
924 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
926 // Vector intrinsic operation, reg
927 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
928 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
929 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
930 let isCommutable = Commutable;
933 // Vector intrinsic operation, mem
934 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
935 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
936 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
940 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
941 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
943 // Reciprocal approximations. Note that these typically require refinement
944 // in order to obtain suitable precision.
945 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
946 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
947 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
948 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
951 let Constraints = "$src1 = $dst" in {
952 let isCommutable = 1 in {
953 def ANDPSrr : PSI<0x54, MRMSrcReg,
954 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
955 "andps\t{$src2, $dst|$dst, $src2}",
956 [(set VR128:$dst, (v2i64
957 (and VR128:$src1, VR128:$src2)))]>;
958 def ORPSrr : PSI<0x56, MRMSrcReg,
959 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
960 "orps\t{$src2, $dst|$dst, $src2}",
961 [(set VR128:$dst, (v2i64
962 (or VR128:$src1, VR128:$src2)))]>;
963 def XORPSrr : PSI<0x57, MRMSrcReg,
964 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
965 "xorps\t{$src2, $dst|$dst, $src2}",
966 [(set VR128:$dst, (v2i64
967 (xor VR128:$src1, VR128:$src2)))]>;
970 def ANDPSrm : PSI<0x54, MRMSrcMem,
971 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
972 "andps\t{$src2, $dst|$dst, $src2}",
973 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
974 (memopv2i64 addr:$src2)))]>;
975 def ORPSrm : PSI<0x56, MRMSrcMem,
976 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
977 "orps\t{$src2, $dst|$dst, $src2}",
978 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
979 (memopv2i64 addr:$src2)))]>;
980 def XORPSrm : PSI<0x57, MRMSrcMem,
981 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
982 "xorps\t{$src2, $dst|$dst, $src2}",
983 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
984 (memopv2i64 addr:$src2)))]>;
985 def ANDNPSrr : PSI<0x55, MRMSrcReg,
986 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
987 "andnps\t{$src2, $dst|$dst, $src2}",
989 (v2i64 (and (xor VR128:$src1,
990 (bc_v2i64 (v4i32 immAllOnesV))),
992 def ANDNPSrm : PSI<0x55, MRMSrcMem,
993 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
994 "andnps\t{$src2, $dst|$dst, $src2}",
996 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
997 (bc_v2i64 (v4i32 immAllOnesV))),
998 (memopv2i64 addr:$src2))))]>;
1001 let Constraints = "$src1 = $dst" in {
1002 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
1003 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1004 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1005 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1006 VR128:$src, imm:$cc))]>;
1007 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
1008 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1009 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1010 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1011 (memop addr:$src), imm:$cc))]>;
1013 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1014 (CMPPSrri VR128:$src1, VR128:$src2, imm:$cc)>;
1015 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1016 (CMPPSrmi VR128:$src1, addr:$src2, imm:$cc)>;
1018 // Shuffle and unpack instructions
1019 let Constraints = "$src1 = $dst" in {
1020 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
1021 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
1022 (outs VR128:$dst), (ins VR128:$src1,
1023 VR128:$src2, i8imm:$src3),
1024 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1026 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1027 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
1028 (outs VR128:$dst), (ins VR128:$src1,
1029 f128mem:$src2, i8imm:$src3),
1030 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1033 VR128:$src1, (memopv4f32 addr:$src2))))]>;
1035 let AddedComplexity = 10 in {
1036 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
1037 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1038 "unpckhps\t{$src2, $dst|$dst, $src2}",
1040 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
1041 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
1042 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1043 "unpckhps\t{$src2, $dst|$dst, $src2}",
1045 (v4f32 (unpckh VR128:$src1,
1046 (memopv4f32 addr:$src2))))]>;
1048 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
1049 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1050 "unpcklps\t{$src2, $dst|$dst, $src2}",
1052 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
1053 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
1054 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1055 "unpcklps\t{$src2, $dst|$dst, $src2}",
1057 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
1058 } // AddedComplexity
1059 } // Constraints = "$src1 = $dst"
1062 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1063 "movmskps\t{$src, $dst|$dst, $src}",
1064 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
1065 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1066 "movmskpd\t{$src, $dst|$dst, $src}",
1067 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1069 // Prefetch intrinsic.
1070 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1071 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1072 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1073 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1074 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1075 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1076 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1077 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1079 // Non-temporal stores
1080 def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1081 "movntps\t{$src, $dst|$dst, $src}",
1082 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1084 let AddedComplexity = 400 in { // Prefer non-temporal versions
1085 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1086 "movntps\t{$src, $dst|$dst, $src}",
1087 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
1089 def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1090 "movntdq\t{$src, $dst|$dst, $src}",
1091 [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
1093 def : Pat<(alignednontemporalstore (v2i64 VR128:$src), addr:$dst),
1094 (MOVNTDQ_64mr VR128:$src, addr:$dst)>;
1096 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
1097 "movnti\t{$src, $dst|$dst, $src}",
1098 [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
1099 TB, Requires<[HasSSE2]>;
1101 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
1102 "movnti\t{$src, $dst|$dst, $src}",
1103 [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
1104 TB, Requires<[HasSSE2]>;
1107 // Load, store, and memory fence
1108 def SFENCE : PSI<0xAE, MRM7r, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
1111 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1112 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1113 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1114 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1116 // Alias instructions that map zero vector to pxor / xorp* for sse.
1117 // We set canFoldAsLoad because this can be converted to a constant-pool
1118 // load of an all-zeros value if folding it would be beneficial.
1119 // FIXME: Change encoding to pseudo!
1120 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1121 isCodeGenOnly = 1 in
1122 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1123 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1125 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
1126 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
1127 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
1128 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
1129 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
1131 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1132 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), x86_subreg_ss))>;
1134 //===---------------------------------------------------------------------===//
1135 // SSE2 Instructions
1136 //===---------------------------------------------------------------------===//
1138 // Move Instructions. Register-to-register movsd is not used for FR64
1139 // register copies because it's a partial register update; FsMOVAPDrr is
1140 // used instead. Register-to-register movsd is not modeled as an INSERT_SUBREG
1141 // because INSERT_SUBREG requires that the insert be implementable in terms of
1142 // a copy, and just mentioned, we don't use movsd for copies.
1143 let Constraints = "$src1 = $dst" in
1144 def MOVSDrr : SDI<0x10, MRMSrcReg,
1145 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
1146 "movsd\t{$src2, $dst|$dst, $src2}",
1148 (movl VR128:$src1, (scalar_to_vector FR64:$src2)))]>;
1150 // Extract the low 64-bit value from one vector and insert it into another.
1151 let AddedComplexity = 15 in
1152 def : Pat<(v2f64 (movl VR128:$src1, VR128:$src2)),
1153 (MOVSDrr VR128:$src1,
1154 (EXTRACT_SUBREG (v2f64 VR128:$src2), x86_subreg_sd))>;
1156 // Implicitly promote a 64-bit scalar to a vector.
1157 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
1158 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), FR64:$src, x86_subreg_sd)>;
1160 // Loading from memory automatically zeroing upper bits.
1161 let canFoldAsLoad = 1, isReMaterializable = 1, AddedComplexity = 20 in
1162 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1163 "movsd\t{$src, $dst|$dst, $src}",
1164 [(set FR64:$dst, (loadf64 addr:$src))]>;
1166 // MOVSDrm zeros the high parts of the register; represent this
1167 // with SUBREG_TO_REG.
1168 let AddedComplexity = 20 in {
1169 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
1170 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>;
1171 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
1172 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>;
1173 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
1174 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>;
1175 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
1176 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>;
1177 def : Pat<(v2f64 (X86vzload addr:$src)),
1178 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), x86_subreg_sd)>;
1181 // Store scalar value to memory.
1182 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1183 "movsd\t{$src, $dst|$dst, $src}",
1184 [(store FR64:$src, addr:$dst)]>;
1186 // Extract and store.
1187 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
1190 (EXTRACT_SUBREG (v2f64 VR128:$src), x86_subreg_sd))>;
1192 // Conversion instructions
1193 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1194 "cvttsd2si\t{$src, $dst|$dst, $src}",
1195 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1196 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1197 "cvttsd2si\t{$src, $dst|$dst, $src}",
1198 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1199 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1200 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1201 [(set FR32:$dst, (fround FR64:$src))]>;
1202 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1203 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1204 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
1205 Requires<[HasSSE2, OptForSize]>;
1206 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1207 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1208 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1209 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1210 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1211 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1213 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1214 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1215 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1216 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1217 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1218 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1219 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1220 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1221 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1222 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1223 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1224 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1225 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1226 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1227 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1228 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1229 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1230 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1231 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1232 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1234 // SSE2 instructions with XS prefix
1235 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1236 "cvtss2sd\t{$src, $dst|$dst, $src}",
1237 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1238 Requires<[HasSSE2]>;
1239 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1240 "cvtss2sd\t{$src, $dst|$dst, $src}",
1241 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1242 Requires<[HasSSE2, OptForSize]>;
1244 def : Pat<(extloadf32 addr:$src),
1245 (CVTSS2SDrr (MOVSSrm addr:$src))>,
1246 Requires<[HasSSE2, OptForSpeed]>;
1248 // Match intrinsics which expect XMM operand(s).
1249 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1250 "cvtsd2si\t{$src, $dst|$dst, $src}",
1251 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1252 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1253 "cvtsd2si\t{$src, $dst|$dst, $src}",
1254 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1255 (load addr:$src)))]>;
1257 // Match intrinisics which expect MM and XMM operand(s).
1258 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1259 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1260 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1261 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1262 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1263 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1264 (memop addr:$src)))]>;
1265 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1266 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1267 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1268 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1269 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1270 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1271 (memop addr:$src)))]>;
1272 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1273 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1274 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1275 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1276 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1277 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1278 (load addr:$src)))]>;
1280 // Aliases for intrinsics
1281 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1282 "cvttsd2si\t{$src, $dst|$dst, $src}",
1284 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1285 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1286 "cvttsd2si\t{$src, $dst|$dst, $src}",
1287 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1288 (load addr:$src)))]>;
1290 // Comparison instructions
1291 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1292 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1293 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1294 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1296 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1297 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1298 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1301 let Defs = [EFLAGS] in {
1302 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1303 "ucomisd\t{$src2, $src1|$src1, $src2}",
1304 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1305 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1306 "ucomisd\t{$src2, $src1|$src1, $src2}",
1307 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1308 (implicit EFLAGS)]>;
1309 } // Defs = [EFLAGS]
1311 // Aliases to match intrinsics which expect XMM operand(s).
1312 let Constraints = "$src1 = $dst" in {
1313 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1315 (ins VR128:$src1, VR128:$src, SSECC:$cc),
1316 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1317 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1318 VR128:$src, imm:$cc))]>;
1319 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1321 (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1322 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1323 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1324 (load addr:$src), imm:$cc))]>;
1327 let Defs = [EFLAGS] in {
1328 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1329 "ucomisd\t{$src2, $src1|$src1, $src2}",
1330 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1331 (implicit EFLAGS)]>;
1332 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1333 "ucomisd\t{$src2, $src1|$src1, $src2}",
1334 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1335 (implicit EFLAGS)]>;
1337 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1338 "comisd\t{$src2, $src1|$src1, $src2}",
1339 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1340 (implicit EFLAGS)]>;
1341 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1342 "comisd\t{$src2, $src1|$src1, $src2}",
1343 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1344 (implicit EFLAGS)]>;
1345 } // Defs = [EFLAGS]
1347 // Aliases of packed SSE2 instructions for scalar use. These all have names
1348 // that start with 'Fs'.
1350 // Alias instructions that map fld0 to pxor for sse.
1351 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1352 canFoldAsLoad = 1 in
1353 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1354 [(set FR64:$dst, fpimm0)]>,
1355 Requires<[HasSSE2]>, TB, OpSize;
1357 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1359 let neverHasSideEffects = 1 in
1360 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1361 "movapd\t{$src, $dst|$dst, $src}", []>;
1363 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1365 let canFoldAsLoad = 1, isReMaterializable = 1 in
1366 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1367 "movapd\t{$src, $dst|$dst, $src}",
1368 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1370 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1371 let Constraints = "$src1 = $dst" in {
1372 let isCommutable = 1 in {
1373 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1374 (ins FR64:$src1, FR64:$src2),
1375 "andpd\t{$src2, $dst|$dst, $src2}",
1376 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1377 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1378 (ins FR64:$src1, FR64:$src2),
1379 "orpd\t{$src2, $dst|$dst, $src2}",
1380 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1381 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1382 (ins FR64:$src1, FR64:$src2),
1383 "xorpd\t{$src2, $dst|$dst, $src2}",
1384 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1387 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1388 (ins FR64:$src1, f128mem:$src2),
1389 "andpd\t{$src2, $dst|$dst, $src2}",
1390 [(set FR64:$dst, (X86fand FR64:$src1,
1391 (memopfsf64 addr:$src2)))]>;
1392 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1393 (ins FR64:$src1, f128mem:$src2),
1394 "orpd\t{$src2, $dst|$dst, $src2}",
1395 [(set FR64:$dst, (X86for FR64:$src1,
1396 (memopfsf64 addr:$src2)))]>;
1397 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1398 (ins FR64:$src1, f128mem:$src2),
1399 "xorpd\t{$src2, $dst|$dst, $src2}",
1400 [(set FR64:$dst, (X86fxor FR64:$src1,
1401 (memopfsf64 addr:$src2)))]>;
1403 let neverHasSideEffects = 1 in {
1404 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1405 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1406 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1408 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1409 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1410 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1414 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1416 /// In addition, we also have a special variant of the scalar form here to
1417 /// represent the associated intrinsic operation. This form is unlike the
1418 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1419 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1421 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1422 /// six "instructions".
1424 let Constraints = "$src1 = $dst" in {
1425 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1426 SDNode OpNode, Intrinsic F64Int,
1427 bit Commutable = 0> {
1428 // Scalar operation, reg+reg.
1429 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1430 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1431 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1432 let isCommutable = Commutable;
1435 // Scalar operation, reg+mem.
1436 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1437 (ins FR64:$src1, f64mem:$src2),
1438 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1439 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1441 // Vector operation, reg+reg.
1442 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1443 (ins VR128:$src1, VR128:$src2),
1444 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1445 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1446 let isCommutable = Commutable;
1449 // Vector operation, reg+mem.
1450 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1451 (ins VR128:$src1, f128mem:$src2),
1452 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1453 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1455 // Intrinsic operation, reg+reg.
1456 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1457 (ins VR128:$src1, VR128:$src2),
1458 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1459 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]>;
1461 // Intrinsic operation, reg+mem.
1462 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1463 (ins VR128:$src1, sdmem:$src2),
1464 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1465 [(set VR128:$dst, (F64Int VR128:$src1,
1466 sse_load_f64:$src2))]>;
1470 // Arithmetic instructions
1471 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1472 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1473 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1474 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1476 /// sse2_fp_binop_rm - Other SSE2 binops
1478 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1479 /// instructions for a full-vector intrinsic form. Operations that map
1480 /// onto C operators don't use this form since they just use the plain
1481 /// vector form instead of having a separate vector intrinsic form.
1483 /// This provides a total of eight "instructions".
1485 let Constraints = "$src1 = $dst" in {
1486 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1490 bit Commutable = 0> {
1492 // Scalar operation, reg+reg.
1493 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1494 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1495 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1496 let isCommutable = Commutable;
1499 // Scalar operation, reg+mem.
1500 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1501 (ins FR64:$src1, f64mem:$src2),
1502 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1503 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1505 // Vector operation, reg+reg.
1506 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1507 (ins VR128:$src1, VR128:$src2),
1508 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1509 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1510 let isCommutable = Commutable;
1513 // Vector operation, reg+mem.
1514 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1515 (ins VR128:$src1, f128mem:$src2),
1516 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1517 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1519 // Intrinsic operation, reg+reg.
1520 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1521 (ins VR128:$src1, VR128:$src2),
1522 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1523 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1524 let isCommutable = Commutable;
1527 // Intrinsic operation, reg+mem.
1528 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1529 (ins VR128:$src1, sdmem:$src2),
1530 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1531 [(set VR128:$dst, (F64Int VR128:$src1,
1532 sse_load_f64:$src2))]>;
1534 // Vector intrinsic operation, reg+reg.
1535 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1536 (ins VR128:$src1, VR128:$src2),
1537 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1538 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1539 let isCommutable = Commutable;
1542 // Vector intrinsic operation, reg+mem.
1543 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1544 (ins VR128:$src1, f128mem:$src2),
1545 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1546 [(set VR128:$dst, (V2F64Int VR128:$src1,
1547 (memopv2f64 addr:$src2)))]>;
1551 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1552 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1553 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1554 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1556 //===---------------------------------------------------------------------===//
1557 // SSE packed FP Instructions
1559 // Move Instructions
1560 let neverHasSideEffects = 1 in
1561 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1562 "movapd\t{$src, $dst|$dst, $src}", []>;
1563 let canFoldAsLoad = 1, isReMaterializable = 1 in
1564 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1565 "movapd\t{$src, $dst|$dst, $src}",
1566 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1568 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1569 "movapd\t{$src, $dst|$dst, $src}",
1570 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1572 let neverHasSideEffects = 1 in
1573 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1574 "movupd\t{$src, $dst|$dst, $src}", []>;
1575 let canFoldAsLoad = 1 in
1576 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1577 "movupd\t{$src, $dst|$dst, $src}",
1578 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1579 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1580 "movupd\t{$src, $dst|$dst, $src}",
1581 [(store (v2f64 VR128:$src), addr:$dst)]>;
1583 // Intrinsic forms of MOVUPD load and store
1584 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1585 "movupd\t{$src, $dst|$dst, $src}",
1586 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1587 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1588 "movupd\t{$src, $dst|$dst, $src}",
1589 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1591 let Constraints = "$src1 = $dst" in {
1592 let AddedComplexity = 20 in {
1593 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1594 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1595 "movlpd\t{$src2, $dst|$dst, $src2}",
1597 (v2f64 (movlp VR128:$src1,
1598 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1599 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1600 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1601 "movhpd\t{$src2, $dst|$dst, $src2}",
1603 (v2f64 (movlhps VR128:$src1,
1604 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1605 } // AddedComplexity
1606 } // Constraints = "$src1 = $dst"
1608 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1609 "movlpd\t{$src, $dst|$dst, $src}",
1610 [(store (f64 (vector_extract (v2f64 VR128:$src),
1611 (iPTR 0))), addr:$dst)]>;
1613 // v2f64 extract element 1 is always custom lowered to unpack high to low
1614 // and extract element 0 so the non-store version isn't too horrible.
1615 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1616 "movhpd\t{$src, $dst|$dst, $src}",
1617 [(store (f64 (vector_extract
1618 (v2f64 (unpckh VR128:$src, (undef))),
1619 (iPTR 0))), addr:$dst)]>;
1621 // SSE2 instructions without OpSize prefix
1622 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1623 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1624 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1625 TB, Requires<[HasSSE2]>;
1626 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1627 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1628 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1629 (bitconvert (memopv2i64 addr:$src))))]>,
1630 TB, Requires<[HasSSE2]>;
1632 // SSE2 instructions with XS prefix
1633 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1634 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1635 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1636 XS, Requires<[HasSSE2]>;
1637 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1638 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1639 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1640 (bitconvert (memopv2i64 addr:$src))))]>,
1641 XS, Requires<[HasSSE2]>;
1643 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1644 "cvtps2dq\t{$src, $dst|$dst, $src}",
1645 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1646 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1647 "cvtps2dq\t{$src, $dst|$dst, $src}",
1648 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1649 (memop addr:$src)))]>;
1650 // SSE2 packed instructions with XS prefix
1651 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1652 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1653 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1654 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1656 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1657 "cvttps2dq\t{$src, $dst|$dst, $src}",
1659 (int_x86_sse2_cvttps2dq VR128:$src))]>,
1660 XS, Requires<[HasSSE2]>;
1661 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1662 "cvttps2dq\t{$src, $dst|$dst, $src}",
1663 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1664 (memop addr:$src)))]>,
1665 XS, Requires<[HasSSE2]>;
1667 // SSE2 packed instructions with XD prefix
1668 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1669 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1670 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1671 XD, Requires<[HasSSE2]>;
1672 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1673 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1674 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1675 (memop addr:$src)))]>,
1676 XD, Requires<[HasSSE2]>;
1678 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1679 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1680 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1681 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1682 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1683 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1684 (memop addr:$src)))]>;
1686 // SSE2 instructions without OpSize prefix
1687 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1688 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1689 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1690 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1692 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1693 "cvtps2pd\t{$src, $dst|$dst, $src}",
1694 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1695 TB, Requires<[HasSSE2]>;
1696 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1697 "cvtps2pd\t{$src, $dst|$dst, $src}",
1698 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1699 (load addr:$src)))]>,
1700 TB, Requires<[HasSSE2]>;
1702 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1703 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1704 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1705 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1708 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1709 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1710 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1711 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1712 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1713 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1714 (memop addr:$src)))]>;
1716 // Match intrinsics which expect XMM operand(s).
1717 // Aliases for intrinsics
1718 let Constraints = "$src1 = $dst" in {
1719 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1720 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1721 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1722 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1724 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1725 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1726 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1727 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1728 (loadi32 addr:$src2)))]>;
1729 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1730 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1731 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1732 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1734 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1735 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1736 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1737 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1738 (load addr:$src2)))]>;
1739 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1740 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1741 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1742 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1743 VR128:$src2))]>, XS,
1744 Requires<[HasSSE2]>;
1745 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1746 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1747 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1748 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1749 (load addr:$src2)))]>, XS,
1750 Requires<[HasSSE2]>;
1755 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1757 /// In addition, we also have a special variant of the scalar form here to
1758 /// represent the associated intrinsic operation. This form is unlike the
1759 /// plain scalar form, in that it takes an entire vector (instead of a
1760 /// scalar) and leaves the top elements undefined.
1762 /// And, we have a special variant form for a full-vector intrinsic form.
1764 /// These four forms can each have a reg or a mem operand, so there are a
1765 /// total of eight "instructions".
1767 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1771 bit Commutable = 0> {
1772 // Scalar operation, reg.
1773 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1774 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1775 [(set FR64:$dst, (OpNode FR64:$src))]> {
1776 let isCommutable = Commutable;
1779 // Scalar operation, mem.
1780 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1781 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1782 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1784 // Vector operation, reg.
1785 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1786 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1787 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1788 let isCommutable = Commutable;
1791 // Vector operation, mem.
1792 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1793 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1794 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1796 // Intrinsic operation, reg.
1797 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1798 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1799 [(set VR128:$dst, (F64Int VR128:$src))]> {
1800 let isCommutable = Commutable;
1803 // Intrinsic operation, mem.
1804 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1805 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1806 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1808 // Vector intrinsic operation, reg
1809 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1810 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1811 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1812 let isCommutable = Commutable;
1815 // Vector intrinsic operation, mem
1816 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1817 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1818 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1822 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1823 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1825 // There is no f64 version of the reciprocal approximation instructions.
1828 let Constraints = "$src1 = $dst" in {
1829 let isCommutable = 1 in {
1830 def ANDPDrr : PDI<0x54, MRMSrcReg,
1831 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1832 "andpd\t{$src2, $dst|$dst, $src2}",
1834 (and (bc_v2i64 (v2f64 VR128:$src1)),
1835 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1836 def ORPDrr : PDI<0x56, MRMSrcReg,
1837 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1838 "orpd\t{$src2, $dst|$dst, $src2}",
1840 (or (bc_v2i64 (v2f64 VR128:$src1)),
1841 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1842 def XORPDrr : PDI<0x57, MRMSrcReg,
1843 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1844 "xorpd\t{$src2, $dst|$dst, $src2}",
1846 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1847 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1850 def ANDPDrm : PDI<0x54, MRMSrcMem,
1851 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1852 "andpd\t{$src2, $dst|$dst, $src2}",
1854 (and (bc_v2i64 (v2f64 VR128:$src1)),
1855 (memopv2i64 addr:$src2)))]>;
1856 def ORPDrm : PDI<0x56, MRMSrcMem,
1857 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1858 "orpd\t{$src2, $dst|$dst, $src2}",
1860 (or (bc_v2i64 (v2f64 VR128:$src1)),
1861 (memopv2i64 addr:$src2)))]>;
1862 def XORPDrm : PDI<0x57, MRMSrcMem,
1863 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1864 "xorpd\t{$src2, $dst|$dst, $src2}",
1866 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1867 (memopv2i64 addr:$src2)))]>;
1868 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1869 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1870 "andnpd\t{$src2, $dst|$dst, $src2}",
1872 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1873 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1874 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1875 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1876 "andnpd\t{$src2, $dst|$dst, $src2}",
1878 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1879 (memopv2i64 addr:$src2)))]>;
1882 let Constraints = "$src1 = $dst" in {
1883 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1884 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1885 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1886 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1887 VR128:$src, imm:$cc))]>;
1888 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1889 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1890 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1891 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1892 (memop addr:$src), imm:$cc))]>;
1894 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1895 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1896 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1897 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1899 // Shuffle and unpack instructions
1900 let Constraints = "$src1 = $dst" in {
1901 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1902 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1903 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1905 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1906 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1907 (outs VR128:$dst), (ins VR128:$src1,
1908 f128mem:$src2, i8imm:$src3),
1909 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1912 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1914 let AddedComplexity = 10 in {
1915 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1916 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1917 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1919 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1920 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1921 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1922 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1924 (v2f64 (unpckh VR128:$src1,
1925 (memopv2f64 addr:$src2))))]>;
1927 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1928 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1929 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1931 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1932 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1933 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1934 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1936 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1937 } // AddedComplexity
1938 } // Constraints = "$src1 = $dst"
1941 //===---------------------------------------------------------------------===//
1942 // SSE integer instructions
1944 // Move Instructions
1945 let neverHasSideEffects = 1 in
1946 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1947 "movdqa\t{$src, $dst|$dst, $src}", []>;
1948 let canFoldAsLoad = 1, mayLoad = 1 in
1949 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1950 "movdqa\t{$src, $dst|$dst, $src}",
1951 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1953 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1954 "movdqa\t{$src, $dst|$dst, $src}",
1955 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1956 let canFoldAsLoad = 1, mayLoad = 1 in
1957 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1958 "movdqu\t{$src, $dst|$dst, $src}",
1959 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1960 XS, Requires<[HasSSE2]>;
1962 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1963 "movdqu\t{$src, $dst|$dst, $src}",
1964 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1965 XS, Requires<[HasSSE2]>;
1967 // Intrinsic forms of MOVDQU load and store
1968 let canFoldAsLoad = 1 in
1969 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1970 "movdqu\t{$src, $dst|$dst, $src}",
1971 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1972 XS, Requires<[HasSSE2]>;
1973 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1974 "movdqu\t{$src, $dst|$dst, $src}",
1975 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1976 XS, Requires<[HasSSE2]>;
1978 let Constraints = "$src1 = $dst" in {
1980 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1981 bit Commutable = 0> {
1982 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1983 (ins VR128:$src1, VR128:$src2),
1984 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1985 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1986 let isCommutable = Commutable;
1988 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1989 (ins VR128:$src1, i128mem:$src2),
1990 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1991 [(set VR128:$dst, (IntId VR128:$src1,
1992 (bitconvert (memopv2i64
1996 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1998 Intrinsic IntId, Intrinsic IntId2> {
1999 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2000 (ins VR128:$src1, VR128:$src2),
2001 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2002 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
2003 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2004 (ins VR128:$src1, i128mem:$src2),
2005 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2006 [(set VR128:$dst, (IntId VR128:$src1,
2007 (bitconvert (memopv2i64 addr:$src2))))]>;
2008 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
2009 (ins VR128:$src1, i32i8imm:$src2),
2010 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2011 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
2014 /// PDI_binop_rm - Simple SSE2 binary operator.
2015 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
2016 ValueType OpVT, bit Commutable = 0> {
2017 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2018 (ins VR128:$src1, VR128:$src2),
2019 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2020 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
2021 let isCommutable = Commutable;
2023 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2024 (ins VR128:$src1, i128mem:$src2),
2025 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2026 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
2027 (bitconvert (memopv2i64 addr:$src2)))))]>;
2030 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
2032 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
2033 /// to collapse (bitconvert VT to VT) into its operand.
2035 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
2036 bit Commutable = 0> {
2037 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2038 (ins VR128:$src1, VR128:$src2),
2039 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2040 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
2041 let isCommutable = Commutable;
2043 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2044 (ins VR128:$src1, i128mem:$src2),
2045 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2046 [(set VR128:$dst, (OpNode VR128:$src1,
2047 (memopv2i64 addr:$src2)))]>;
2050 } // Constraints = "$src1 = $dst"
2052 // 128-bit Integer Arithmetic
2054 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
2055 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
2056 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
2057 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
2059 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
2060 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
2061 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
2062 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
2064 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
2065 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
2066 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
2067 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
2069 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
2070 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
2071 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
2072 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
2074 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
2076 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
2077 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
2078 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
2080 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
2082 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
2083 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
2086 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
2087 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
2088 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
2089 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
2090 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
2093 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
2094 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
2095 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
2096 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
2097 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
2098 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
2100 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2101 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2102 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2103 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2104 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2105 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2107 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2108 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2109 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2110 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2112 // 128-bit logical shifts.
2113 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
2114 def PSLLDQri : PDIi8<0x73, MRM7r,
2115 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2116 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2117 def PSRLDQri : PDIi8<0x73, MRM3r,
2118 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2119 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2120 // PSRADQri doesn't exist in SSE[1-3].
2123 let Predicates = [HasSSE2] in {
2124 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2125 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2126 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2127 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2128 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2129 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2130 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2131 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2132 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2133 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2135 // Shift up / down and insert zero's.
2136 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2137 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2138 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2139 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2143 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2144 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2145 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2147 let Constraints = "$src1 = $dst" in {
2148 def PANDNrr : PDI<0xDF, MRMSrcReg,
2149 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2150 "pandn\t{$src2, $dst|$dst, $src2}",
2151 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2154 def PANDNrm : PDI<0xDF, MRMSrcMem,
2155 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2156 "pandn\t{$src2, $dst|$dst, $src2}",
2157 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2158 (memopv2i64 addr:$src2))))]>;
2161 // SSE2 Integer comparison
2162 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2163 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2164 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2165 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2166 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2167 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2169 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2170 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2171 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2172 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2173 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2174 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2175 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2176 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2177 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2178 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2179 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2180 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2182 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2183 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2184 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2185 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2186 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2187 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2188 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2189 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2190 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2191 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2192 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2193 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2196 // Pack instructions
2197 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2198 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2199 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2201 // Shuffle and unpack instructions
2202 let AddedComplexity = 5 in {
2203 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2204 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2205 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2206 [(set VR128:$dst, (v4i32 (pshufd:$src2
2207 VR128:$src1, (undef))))]>;
2208 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2209 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2210 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2211 [(set VR128:$dst, (v4i32 (pshufd:$src2
2212 (bc_v4i32 (memopv2i64 addr:$src1)),
2216 // SSE2 with ImmT == Imm8 and XS prefix.
2217 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2218 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2219 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2220 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2222 XS, Requires<[HasSSE2]>;
2223 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2224 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2225 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2226 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2227 (bc_v8i16 (memopv2i64 addr:$src1)),
2229 XS, Requires<[HasSSE2]>;
2231 // SSE2 with ImmT == Imm8 and XD prefix.
2232 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2233 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2234 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2235 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2237 XD, Requires<[HasSSE2]>;
2238 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2239 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2240 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2241 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2242 (bc_v8i16 (memopv2i64 addr:$src1)),
2244 XD, Requires<[HasSSE2]>;
2247 let Constraints = "$src1 = $dst" in {
2248 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2249 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2250 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2252 (v16i8 (unpckl VR128:$src1, VR128:$src2)))]>;
2253 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2254 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2255 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2257 (unpckl VR128:$src1,
2258 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2259 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2260 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2261 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2263 (v8i16 (unpckl VR128:$src1, VR128:$src2)))]>;
2264 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2265 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2266 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2268 (unpckl VR128:$src1,
2269 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2270 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2271 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2272 "punpckldq\t{$src2, $dst|$dst, $src2}",
2274 (v4i32 (unpckl VR128:$src1, VR128:$src2)))]>;
2275 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2276 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2277 "punpckldq\t{$src2, $dst|$dst, $src2}",
2279 (unpckl VR128:$src1,
2280 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2281 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2282 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2283 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2285 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2286 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2287 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2288 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2290 (v2i64 (unpckl VR128:$src1,
2291 (memopv2i64 addr:$src2))))]>;
2293 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2294 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2295 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2297 (v16i8 (unpckh VR128:$src1, VR128:$src2)))]>;
2298 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2299 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2300 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2302 (unpckh VR128:$src1,
2303 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2304 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2305 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2306 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2308 (v8i16 (unpckh VR128:$src1, VR128:$src2)))]>;
2309 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2310 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2311 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2313 (unpckh VR128:$src1,
2314 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2315 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2316 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2317 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2319 (v4i32 (unpckh VR128:$src1, VR128:$src2)))]>;
2320 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2321 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2322 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2324 (unpckh VR128:$src1,
2325 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2326 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2327 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2328 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2330 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2331 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2332 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2333 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2335 (v2i64 (unpckh VR128:$src1,
2336 (memopv2i64 addr:$src2))))]>;
2340 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2341 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2342 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2343 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2345 let Constraints = "$src1 = $dst" in {
2346 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2347 (outs VR128:$dst), (ins VR128:$src1,
2348 GR32:$src2, i32i8imm:$src3),
2349 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2351 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2352 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2353 (outs VR128:$dst), (ins VR128:$src1,
2354 i16mem:$src2, i32i8imm:$src3),
2355 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2357 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2362 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2363 "pmovmskb\t{$src, $dst|$dst, $src}",
2364 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2366 // Conditional store
2368 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2369 "maskmovdqu\t{$mask, $src|$src, $mask}",
2370 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2373 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2374 "maskmovdqu\t{$mask, $src|$src, $mask}",
2375 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2377 // Non-temporal stores
2378 def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2379 "movntpd\t{$src, $dst|$dst, $src}",
2380 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2381 def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2382 "movntdq\t{$src, $dst|$dst, $src}",
2383 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2384 def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2385 "movnti\t{$src, $dst|$dst, $src}",
2386 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2387 TB, Requires<[HasSSE2]>;
2389 let AddedComplexity = 400 in { // Prefer non-temporal versions
2390 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2391 "movntpd\t{$src, $dst|$dst, $src}",
2392 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
2394 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2395 "movntdq\t{$src, $dst|$dst, $src}",
2396 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2398 def : Pat<(alignednontemporalstore (v4i32 VR128:$src), addr:$dst),
2399 (MOVNTDQmr VR128:$src, addr:$dst)>;
2403 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2404 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2405 TB, Requires<[HasSSE2]>;
2407 // Load, store, and memory fence
2408 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
2409 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2410 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
2411 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2413 //TODO: custom lower this so as to never even generate the noop
2414 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2416 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2417 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2418 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2421 // Alias instructions that map zero vector to pxor / xorp* for sse.
2422 // We set canFoldAsLoad because this can be converted to a constant-pool
2423 // load of an all-ones value if folding it would be beneficial.
2424 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2425 isCodeGenOnly = 1 in
2426 // FIXME: Change encoding to pseudo.
2427 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
2428 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2430 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2431 "movd\t{$src, $dst|$dst, $src}",
2433 (v4i32 (scalar_to_vector GR32:$src)))]>;
2434 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2435 "movd\t{$src, $dst|$dst, $src}",
2437 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2439 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2440 "movd\t{$src, $dst|$dst, $src}",
2441 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2443 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2444 "movd\t{$src, $dst|$dst, $src}",
2445 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2447 // SSE2 instructions with XS prefix
2448 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2449 "movq\t{$src, $dst|$dst, $src}",
2451 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2452 Requires<[HasSSE2]>;
2453 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2454 "movq\t{$src, $dst|$dst, $src}",
2455 [(store (i64 (vector_extract (v2i64 VR128:$src),
2456 (iPTR 0))), addr:$dst)]>;
2458 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
2459 (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), x86_subreg_sd))>;
2461 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2462 "movd\t{$src, $dst|$dst, $src}",
2463 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2465 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2466 "movd\t{$src, $dst|$dst, $src}",
2467 [(store (i32 (vector_extract (v4i32 VR128:$src),
2468 (iPTR 0))), addr:$dst)]>;
2470 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2471 "movd\t{$src, $dst|$dst, $src}",
2472 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2473 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2474 "movd\t{$src, $dst|$dst, $src}",
2475 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2477 // Store / copy lower 64-bits of a XMM register.
2478 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2479 "movq\t{$src, $dst|$dst, $src}",
2480 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2482 // movd / movq to XMM register zero-extends
2483 let AddedComplexity = 15 in {
2484 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2485 "movd\t{$src, $dst|$dst, $src}",
2486 [(set VR128:$dst, (v4i32 (X86vzmovl
2487 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2488 // This is X86-64 only.
2489 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2490 "mov{d|q}\t{$src, $dst|$dst, $src}",
2491 [(set VR128:$dst, (v2i64 (X86vzmovl
2492 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2495 let AddedComplexity = 20 in {
2496 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2497 "movd\t{$src, $dst|$dst, $src}",
2499 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2500 (loadi32 addr:$src))))))]>;
2502 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2503 (MOVZDI2PDIrm addr:$src)>;
2504 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2505 (MOVZDI2PDIrm addr:$src)>;
2506 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2507 (MOVZDI2PDIrm addr:$src)>;
2509 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2510 "movq\t{$src, $dst|$dst, $src}",
2512 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2513 (loadi64 addr:$src))))))]>, XS,
2514 Requires<[HasSSE2]>;
2516 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2517 (MOVZQI2PQIrm addr:$src)>;
2518 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2519 (MOVZQI2PQIrm addr:$src)>;
2520 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2523 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2524 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2525 let AddedComplexity = 15 in
2526 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2527 "movq\t{$src, $dst|$dst, $src}",
2528 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2529 XS, Requires<[HasSSE2]>;
2531 let AddedComplexity = 20 in {
2532 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2533 "movq\t{$src, $dst|$dst, $src}",
2534 [(set VR128:$dst, (v2i64 (X86vzmovl
2535 (loadv2i64 addr:$src))))]>,
2536 XS, Requires<[HasSSE2]>;
2538 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2539 (MOVZPQILo2PQIrm addr:$src)>;
2542 // Instructions for the disassembler
2543 // xr = XMM register
2546 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2547 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2549 //===---------------------------------------------------------------------===//
2550 // SSE3 Instructions
2551 //===---------------------------------------------------------------------===//
2553 // Move Instructions
2554 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2555 "movshdup\t{$src, $dst|$dst, $src}",
2556 [(set VR128:$dst, (v4f32 (movshdup
2557 VR128:$src, (undef))))]>;
2558 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2559 "movshdup\t{$src, $dst|$dst, $src}",
2560 [(set VR128:$dst, (movshdup
2561 (memopv4f32 addr:$src), (undef)))]>;
2563 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2564 "movsldup\t{$src, $dst|$dst, $src}",
2565 [(set VR128:$dst, (v4f32 (movsldup
2566 VR128:$src, (undef))))]>;
2567 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2568 "movsldup\t{$src, $dst|$dst, $src}",
2569 [(set VR128:$dst, (movsldup
2570 (memopv4f32 addr:$src), (undef)))]>;
2572 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2573 "movddup\t{$src, $dst|$dst, $src}",
2574 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2575 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2576 "movddup\t{$src, $dst|$dst, $src}",
2578 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2581 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2583 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2585 let AddedComplexity = 5 in {
2586 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2587 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2588 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2589 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2590 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2591 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2592 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2593 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2597 let Constraints = "$src1 = $dst" in {
2598 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2599 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2600 "addsubps\t{$src2, $dst|$dst, $src2}",
2601 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2603 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2604 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2605 "addsubps\t{$src2, $dst|$dst, $src2}",
2606 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2607 (memop addr:$src2)))]>;
2608 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2609 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2610 "addsubpd\t{$src2, $dst|$dst, $src2}",
2611 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2613 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2614 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2615 "addsubpd\t{$src2, $dst|$dst, $src2}",
2616 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2617 (memop addr:$src2)))]>;
2620 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2621 "lddqu\t{$src, $dst|$dst, $src}",
2622 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2625 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2626 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2627 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2628 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2629 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2630 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2631 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2632 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2633 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2634 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2635 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2636 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2637 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2638 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2639 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2640 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2642 let Constraints = "$src1 = $dst" in {
2643 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2644 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2645 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2646 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2647 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2648 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2649 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2650 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2653 // Thread synchronization
2654 def MONITOR : I<0x01, MRM_C8, (outs), (ins), "monitor",
2655 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2656 def MWAIT : I<0x01, MRM_C9, (outs), (ins), "mwait",
2657 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2659 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2660 let AddedComplexity = 15 in
2661 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2662 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2663 let AddedComplexity = 20 in
2664 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2665 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2667 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2668 let AddedComplexity = 15 in
2669 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2670 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2671 let AddedComplexity = 20 in
2672 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2673 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2675 //===---------------------------------------------------------------------===//
2676 // SSSE3 Instructions
2677 //===---------------------------------------------------------------------===//
2679 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2680 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2681 Intrinsic IntId64, Intrinsic IntId128> {
2682 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2683 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2684 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2686 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2687 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2689 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2691 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2693 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2694 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2697 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2699 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2702 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2705 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2706 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2707 Intrinsic IntId64, Intrinsic IntId128> {
2708 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2710 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2711 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2713 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2715 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2718 (bitconvert (memopv4i16 addr:$src))))]>;
2720 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2722 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2723 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2726 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2728 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2731 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2734 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2735 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2736 Intrinsic IntId64, Intrinsic IntId128> {
2737 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2739 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2740 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2742 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2744 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2747 (bitconvert (memopv2i32 addr:$src))))]>;
2749 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2751 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2752 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2755 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2757 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2760 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2763 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2764 int_x86_ssse3_pabs_b,
2765 int_x86_ssse3_pabs_b_128>;
2766 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2767 int_x86_ssse3_pabs_w,
2768 int_x86_ssse3_pabs_w_128>;
2769 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2770 int_x86_ssse3_pabs_d,
2771 int_x86_ssse3_pabs_d_128>;
2773 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2774 let Constraints = "$src1 = $dst" in {
2775 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2776 Intrinsic IntId64, Intrinsic IntId128,
2777 bit Commutable = 0> {
2778 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2779 (ins VR64:$src1, VR64:$src2),
2780 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2781 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2782 let isCommutable = Commutable;
2784 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2785 (ins VR64:$src1, i64mem:$src2),
2786 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2788 (IntId64 VR64:$src1,
2789 (bitconvert (memopv8i8 addr:$src2))))]>;
2791 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2792 (ins VR128:$src1, VR128:$src2),
2793 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2794 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2796 let isCommutable = Commutable;
2798 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2799 (ins VR128:$src1, i128mem:$src2),
2800 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2802 (IntId128 VR128:$src1,
2803 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2807 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2808 let Constraints = "$src1 = $dst" in {
2809 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2810 Intrinsic IntId64, Intrinsic IntId128,
2811 bit Commutable = 0> {
2812 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2813 (ins VR64:$src1, VR64:$src2),
2814 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2815 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2816 let isCommutable = Commutable;
2818 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2819 (ins VR64:$src1, i64mem:$src2),
2820 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2822 (IntId64 VR64:$src1,
2823 (bitconvert (memopv4i16 addr:$src2))))]>;
2825 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2826 (ins VR128:$src1, VR128:$src2),
2827 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2828 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2830 let isCommutable = Commutable;
2832 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2833 (ins VR128:$src1, i128mem:$src2),
2834 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2836 (IntId128 VR128:$src1,
2837 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2841 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2842 let Constraints = "$src1 = $dst" in {
2843 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2844 Intrinsic IntId64, Intrinsic IntId128,
2845 bit Commutable = 0> {
2846 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2847 (ins VR64:$src1, VR64:$src2),
2848 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2849 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2850 let isCommutable = Commutable;
2852 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2853 (ins VR64:$src1, i64mem:$src2),
2854 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2856 (IntId64 VR64:$src1,
2857 (bitconvert (memopv2i32 addr:$src2))))]>;
2859 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2860 (ins VR128:$src1, VR128:$src2),
2861 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2862 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2864 let isCommutable = Commutable;
2866 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2867 (ins VR128:$src1, i128mem:$src2),
2868 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2870 (IntId128 VR128:$src1,
2871 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2875 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2876 int_x86_ssse3_phadd_w,
2877 int_x86_ssse3_phadd_w_128>;
2878 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2879 int_x86_ssse3_phadd_d,
2880 int_x86_ssse3_phadd_d_128>;
2881 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2882 int_x86_ssse3_phadd_sw,
2883 int_x86_ssse3_phadd_sw_128>;
2884 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2885 int_x86_ssse3_phsub_w,
2886 int_x86_ssse3_phsub_w_128>;
2887 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2888 int_x86_ssse3_phsub_d,
2889 int_x86_ssse3_phsub_d_128>;
2890 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2891 int_x86_ssse3_phsub_sw,
2892 int_x86_ssse3_phsub_sw_128>;
2893 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2894 int_x86_ssse3_pmadd_ub_sw,
2895 int_x86_ssse3_pmadd_ub_sw_128>;
2896 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2897 int_x86_ssse3_pmul_hr_sw,
2898 int_x86_ssse3_pmul_hr_sw_128, 1>;
2899 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2900 int_x86_ssse3_pshuf_b,
2901 int_x86_ssse3_pshuf_b_128>;
2902 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2903 int_x86_ssse3_psign_b,
2904 int_x86_ssse3_psign_b_128>;
2905 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2906 int_x86_ssse3_psign_w,
2907 int_x86_ssse3_psign_w_128>;
2908 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2909 int_x86_ssse3_psign_d,
2910 int_x86_ssse3_psign_d_128>;
2912 let Constraints = "$src1 = $dst" in {
2913 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2914 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2915 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2917 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2918 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2919 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2922 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2923 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2924 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2926 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2927 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2928 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2932 // palignr patterns.
2933 def : Pat<(int_x86_ssse3_palign_r VR64:$src1, VR64:$src2, (i8 imm:$src3)),
2934 (PALIGNR64rr VR64:$src1, VR64:$src2, (BYTE_imm imm:$src3))>,
2935 Requires<[HasSSSE3]>;
2936 def : Pat<(int_x86_ssse3_palign_r VR64:$src1,
2937 (memop64 addr:$src2),
2939 (PALIGNR64rm VR64:$src1, addr:$src2, (BYTE_imm imm:$src3))>,
2940 Requires<[HasSSSE3]>;
2942 def : Pat<(int_x86_ssse3_palign_r_128 VR128:$src1, VR128:$src2, (i8 imm:$src3)),
2943 (PALIGNR128rr VR128:$src1, VR128:$src2, (BYTE_imm imm:$src3))>,
2944 Requires<[HasSSSE3]>;
2945 def : Pat<(int_x86_ssse3_palign_r_128 VR128:$src1,
2946 (memopv2i64 addr:$src2),
2948 (PALIGNR128rm VR128:$src1, addr:$src2, (BYTE_imm imm:$src3))>,
2949 Requires<[HasSSSE3]>;
2951 let AddedComplexity = 5 in {
2952 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2953 (PALIGNR128rr VR128:$src2, VR128:$src1,
2954 (SHUFFLE_get_palign_imm VR128:$src3))>,
2955 Requires<[HasSSSE3]>;
2956 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2957 (PALIGNR128rr VR128:$src2, VR128:$src1,
2958 (SHUFFLE_get_palign_imm VR128:$src3))>,
2959 Requires<[HasSSSE3]>;
2960 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2961 (PALIGNR128rr VR128:$src2, VR128:$src1,
2962 (SHUFFLE_get_palign_imm VR128:$src3))>,
2963 Requires<[HasSSSE3]>;
2964 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2965 (PALIGNR128rr VR128:$src2, VR128:$src1,
2966 (SHUFFLE_get_palign_imm VR128:$src3))>,
2967 Requires<[HasSSSE3]>;
2970 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2971 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2972 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2973 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2975 //===---------------------------------------------------------------------===//
2976 // Non-Instruction Patterns
2977 //===---------------------------------------------------------------------===//
2979 // extload f32 -> f64. This matches load+fextend because we have a hack in
2980 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2982 // Since these loads aren't folded into the fextend, we have to match it
2984 let Predicates = [HasSSE2] in
2985 def : Pat<(fextend (loadf32 addr:$src)),
2986 (CVTSS2SDrm addr:$src)>;
2989 let Predicates = [HasSSE2] in {
2990 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2991 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2992 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2993 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2994 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2995 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2996 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2997 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2998 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2999 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
3000 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
3001 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
3002 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
3003 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
3004 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
3005 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
3006 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
3007 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
3008 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
3009 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
3010 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
3011 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
3012 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
3013 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
3014 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
3015 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
3016 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
3017 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
3018 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
3019 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
3022 // Move scalar to XMM zero-extended
3023 // movd to XMM register zero-extends
3024 let AddedComplexity = 15 in {
3025 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
3026 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
3027 (MOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
3028 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
3029 (MOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
3030 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
3031 (MOVSSrr (v4f32 (V_SET0)),
3032 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), x86_subreg_ss)))>;
3033 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
3034 (MOVSSrr (v4i32 (V_SET0)),
3035 (EXTRACT_SUBREG (v4i32 VR128:$src), x86_subreg_ss))>;
3038 // Splat v2f64 / v2i64
3039 let AddedComplexity = 10 in {
3040 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
3041 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3042 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
3043 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3044 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
3045 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3046 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
3047 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3050 // Special unary SHUFPSrri case.
3051 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
3052 (SHUFPSrri VR128:$src1, VR128:$src1,
3053 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3054 let AddedComplexity = 5 in
3055 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
3056 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3057 Requires<[HasSSE2]>;
3058 // Special unary SHUFPDrri case.
3059 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
3060 (SHUFPDrri VR128:$src1, VR128:$src1,
3061 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3062 Requires<[HasSSE2]>;
3063 // Special unary SHUFPDrri case.
3064 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
3065 (SHUFPDrri VR128:$src1, VR128:$src1,
3066 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3067 Requires<[HasSSE2]>;
3068 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
3069 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
3070 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3071 Requires<[HasSSE2]>;
3073 // Special binary v4i32 shuffle cases with SHUFPS.
3074 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
3075 (SHUFPSrri VR128:$src1, VR128:$src2,
3076 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3077 Requires<[HasSSE2]>;
3078 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
3079 (SHUFPSrmi VR128:$src1, addr:$src2,
3080 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3081 Requires<[HasSSE2]>;
3082 // Special binary v2i64 shuffle cases using SHUFPDrri.
3083 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
3084 (SHUFPDrri VR128:$src1, VR128:$src2,
3085 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3086 Requires<[HasSSE2]>;
3088 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
3089 let AddedComplexity = 15 in {
3090 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
3091 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3092 Requires<[OptForSpeed, HasSSE2]>;
3093 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
3094 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3095 Requires<[OptForSpeed, HasSSE2]>;
3097 let AddedComplexity = 10 in {
3098 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
3099 (UNPCKLPSrr VR128:$src, VR128:$src)>;
3100 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
3101 (PUNPCKLBWrr VR128:$src, VR128:$src)>;
3102 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
3103 (PUNPCKLWDrr VR128:$src, VR128:$src)>;
3104 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
3105 (PUNPCKLDQrr VR128:$src, VR128:$src)>;
3108 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
3109 let AddedComplexity = 15 in {
3110 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
3111 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3112 Requires<[OptForSpeed, HasSSE2]>;
3113 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
3114 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3115 Requires<[OptForSpeed, HasSSE2]>;
3117 let AddedComplexity = 10 in {
3118 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
3119 (UNPCKHPSrr VR128:$src, VR128:$src)>;
3120 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
3121 (PUNPCKHBWrr VR128:$src, VR128:$src)>;
3122 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
3123 (PUNPCKHWDrr VR128:$src, VR128:$src)>;
3124 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
3125 (PUNPCKHDQrr VR128:$src, VR128:$src)>;
3128 let AddedComplexity = 20 in {
3129 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3130 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3131 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3133 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3134 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3135 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3137 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3138 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3139 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3140 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3141 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3144 let AddedComplexity = 20 in {
3145 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3146 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3147 (MOVLPSrm VR128:$src1, addr:$src2)>;
3148 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3149 (MOVLPDrm VR128:$src1, addr:$src2)>;
3150 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3151 (MOVLPSrm VR128:$src1, addr:$src2)>;
3152 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3153 (MOVLPDrm VR128:$src1, addr:$src2)>;
3156 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3157 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3158 (MOVLPSmr addr:$src1, VR128:$src2)>;
3159 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3160 (MOVLPDmr addr:$src1, VR128:$src2)>;
3161 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3163 (MOVLPSmr addr:$src1, VR128:$src2)>;
3164 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3165 (MOVLPDmr addr:$src1, VR128:$src2)>;
3167 let AddedComplexity = 15 in {
3168 // Setting the lowest element in the vector.
3169 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3170 (MOVSSrr (v4i32 VR128:$src1),
3171 (EXTRACT_SUBREG (v4i32 VR128:$src2), x86_subreg_ss))>;
3172 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3173 (MOVSDrr (v2i64 VR128:$src1),
3174 (EXTRACT_SUBREG (v2i64 VR128:$src2), x86_subreg_sd))>;
3176 // vector_shuffle v1, v2 <4, 5, 2, 3> using movsd
3177 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3178 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, x86_subreg_sd))>,
3179 Requires<[HasSSE2]>;
3180 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3181 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, x86_subreg_sd))>,
3182 Requires<[HasSSE2]>;
3185 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3186 // fall back to this for SSE1)
3187 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3188 (SHUFPSrri VR128:$src2, VR128:$src1,
3189 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3191 // Set lowest element and zero upper elements.
3192 let AddedComplexity = 15 in
3193 def : Pat<(v2f64 (movl immAllZerosV_bc, VR128:$src)),
3194 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3195 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3196 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3198 // Some special case pandn patterns.
3199 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3201 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3202 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3204 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3205 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3207 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3209 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3210 (memop addr:$src2))),
3211 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3212 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3213 (memop addr:$src2))),
3214 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3215 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3216 (memop addr:$src2))),
3217 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3219 // vector -> vector casts
3220 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3221 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3222 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3223 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3224 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3225 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3226 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3227 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3229 // Use movaps / movups for SSE integer load / store (one byte shorter).
3230 def : Pat<(alignedloadv4i32 addr:$src),
3231 (MOVAPSrm addr:$src)>;
3232 def : Pat<(loadv4i32 addr:$src),
3233 (MOVUPSrm addr:$src)>;
3234 def : Pat<(alignedloadv2i64 addr:$src),
3235 (MOVAPSrm addr:$src)>;
3236 def : Pat<(loadv2i64 addr:$src),
3237 (MOVUPSrm addr:$src)>;
3239 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3240 (MOVAPSmr addr:$dst, VR128:$src)>;
3241 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3242 (MOVAPSmr addr:$dst, VR128:$src)>;
3243 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3244 (MOVAPSmr addr:$dst, VR128:$src)>;
3245 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3246 (MOVAPSmr addr:$dst, VR128:$src)>;
3247 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3248 (MOVUPSmr addr:$dst, VR128:$src)>;
3249 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3250 (MOVUPSmr addr:$dst, VR128:$src)>;
3251 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3252 (MOVUPSmr addr:$dst, VR128:$src)>;
3253 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3254 (MOVUPSmr addr:$dst, VR128:$src)>;
3256 //===----------------------------------------------------------------------===//
3257 // SSE4.1 Instructions
3258 //===----------------------------------------------------------------------===//
3260 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3263 Intrinsic V2F64Int> {
3264 // Intrinsic operation, reg.
3265 // Vector intrinsic operation, reg
3266 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3267 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3268 !strconcat(OpcodeStr,
3269 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3270 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3273 // Vector intrinsic operation, mem
3274 def PSm_Int : Ii8<opcps, MRMSrcMem,
3275 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3276 !strconcat(OpcodeStr,
3277 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3279 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3281 Requires<[HasSSE41]>;
3283 // Vector intrinsic operation, reg
3284 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3285 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3286 !strconcat(OpcodeStr,
3287 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3288 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3291 // Vector intrinsic operation, mem
3292 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3293 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3294 !strconcat(OpcodeStr,
3295 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3297 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3301 let Constraints = "$src1 = $dst" in {
3302 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3306 // Intrinsic operation, reg.
3307 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3309 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3310 !strconcat(OpcodeStr,
3311 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3313 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3316 // Intrinsic operation, mem.
3317 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3319 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3320 !strconcat(OpcodeStr,
3321 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3323 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3326 // Intrinsic operation, reg.
3327 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3329 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3330 !strconcat(OpcodeStr,
3331 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3333 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3336 // Intrinsic operation, mem.
3337 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3339 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3340 !strconcat(OpcodeStr,
3341 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3343 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3348 // FP round - roundss, roundps, roundsd, roundpd
3349 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3350 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3351 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3352 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3354 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3355 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3356 Intrinsic IntId128> {
3357 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3359 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3360 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3361 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3363 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3366 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3369 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3370 int_x86_sse41_phminposuw>;
3372 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3373 let Constraints = "$src1 = $dst" in {
3374 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3375 Intrinsic IntId128, bit Commutable = 0> {
3376 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3377 (ins VR128:$src1, VR128:$src2),
3378 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3379 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3381 let isCommutable = Commutable;
3383 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3384 (ins VR128:$src1, i128mem:$src2),
3385 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3387 (IntId128 VR128:$src1,
3388 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3392 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3393 int_x86_sse41_pcmpeqq, 1>;
3394 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3395 int_x86_sse41_packusdw, 0>;
3396 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3397 int_x86_sse41_pminsb, 1>;
3398 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3399 int_x86_sse41_pminsd, 1>;
3400 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3401 int_x86_sse41_pminud, 1>;
3402 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3403 int_x86_sse41_pminuw, 1>;
3404 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3405 int_x86_sse41_pmaxsb, 1>;
3406 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3407 int_x86_sse41_pmaxsd, 1>;
3408 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3409 int_x86_sse41_pmaxud, 1>;
3410 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3411 int_x86_sse41_pmaxuw, 1>;
3413 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3415 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3416 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3417 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3418 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3420 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3421 let Constraints = "$src1 = $dst" in {
3422 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3423 SDNode OpNode, Intrinsic IntId128,
3424 bit Commutable = 0> {
3425 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3426 (ins VR128:$src1, VR128:$src2),
3427 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3428 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3429 VR128:$src2))]>, OpSize {
3430 let isCommutable = Commutable;
3432 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3433 (ins VR128:$src1, VR128:$src2),
3434 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3435 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3437 let isCommutable = Commutable;
3439 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3440 (ins VR128:$src1, i128mem:$src2),
3441 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3443 (OpVT (OpNode VR128:$src1, (memop addr:$src2))))]>, OpSize;
3444 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3445 (ins VR128:$src1, i128mem:$src2),
3446 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3448 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3452 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", v4i32, mul,
3453 int_x86_sse41_pmulld, 1>;
3455 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3456 let Constraints = "$src1 = $dst" in {
3457 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3458 Intrinsic IntId128, bit Commutable = 0> {
3459 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3460 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3461 !strconcat(OpcodeStr,
3462 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3464 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3466 let isCommutable = Commutable;
3468 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3469 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3470 !strconcat(OpcodeStr,
3471 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3473 (IntId128 VR128:$src1,
3474 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3479 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3480 int_x86_sse41_blendps, 0>;
3481 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3482 int_x86_sse41_blendpd, 0>;
3483 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3484 int_x86_sse41_pblendw, 0>;
3485 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3486 int_x86_sse41_dpps, 1>;
3487 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3488 int_x86_sse41_dppd, 1>;
3489 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3490 int_x86_sse41_mpsadbw, 1>;
3493 /// SS41I_ternary_int - SSE 4.1 ternary operator
3494 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3495 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3496 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3497 (ins VR128:$src1, VR128:$src2),
3498 !strconcat(OpcodeStr,
3499 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3500 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3503 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3504 (ins VR128:$src1, i128mem:$src2),
3505 !strconcat(OpcodeStr,
3506 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3509 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3513 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3514 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3515 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3518 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3519 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3520 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3521 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3523 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3524 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3526 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3530 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3531 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3532 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3533 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3534 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3535 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3537 // Common patterns involving scalar load.
3538 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3539 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3540 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3541 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3543 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3544 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3545 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3546 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3548 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3549 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3550 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3551 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3553 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3554 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3555 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3556 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3558 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3559 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3560 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3561 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3563 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3564 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3565 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3566 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3569 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3570 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3571 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3572 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3574 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3575 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3577 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3581 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3582 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3583 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3584 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3586 // Common patterns involving scalar load
3587 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3588 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3589 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3590 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3592 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3593 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3594 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3595 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3598 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3599 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3600 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3601 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3603 // Expecting a i16 load any extended to i32 value.
3604 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3605 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3606 [(set VR128:$dst, (IntId (bitconvert
3607 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3611 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3612 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3614 // Common patterns involving scalar load
3615 def : Pat<(int_x86_sse41_pmovsxbq
3616 (bitconvert (v4i32 (X86vzmovl
3617 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3618 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3620 def : Pat<(int_x86_sse41_pmovzxbq
3621 (bitconvert (v4i32 (X86vzmovl
3622 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3623 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3626 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3627 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3628 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3629 (ins VR128:$src1, i32i8imm:$src2),
3630 !strconcat(OpcodeStr,
3631 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3632 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3634 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3635 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3636 !strconcat(OpcodeStr,
3637 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3640 // There's an AssertZext in the way of writing the store pattern
3641 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3644 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3647 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3648 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3649 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3650 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3651 !strconcat(OpcodeStr,
3652 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3655 // There's an AssertZext in the way of writing the store pattern
3656 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3659 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3662 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3663 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3664 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3665 (ins VR128:$src1, i32i8imm:$src2),
3666 !strconcat(OpcodeStr,
3667 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3669 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3670 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3671 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3672 !strconcat(OpcodeStr,
3673 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3674 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3675 addr:$dst)]>, OpSize;
3678 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3681 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3683 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3684 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3685 (ins VR128:$src1, i32i8imm:$src2),
3686 !strconcat(OpcodeStr,
3687 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3689 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3691 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3692 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3693 !strconcat(OpcodeStr,
3694 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3695 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3696 addr:$dst)]>, OpSize;
3699 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3701 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3702 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3705 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3706 Requires<[HasSSE41]>;
3708 let Constraints = "$src1 = $dst" in {
3709 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3710 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3711 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3712 !strconcat(OpcodeStr,
3713 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3715 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3716 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3717 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3718 !strconcat(OpcodeStr,
3719 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3721 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3722 imm:$src3))]>, OpSize;
3726 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3728 let Constraints = "$src1 = $dst" in {
3729 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3730 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3731 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3732 !strconcat(OpcodeStr,
3733 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3735 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3737 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3738 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3739 !strconcat(OpcodeStr,
3740 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3742 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3743 imm:$src3)))]>, OpSize;
3747 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3749 // insertps has a few different modes, there's the first two here below which
3750 // are optimized inserts that won't zero arbitrary elements in the destination
3751 // vector. The next one matches the intrinsic and could zero arbitrary elements
3752 // in the target vector.
3753 let Constraints = "$src1 = $dst" in {
3754 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3755 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3756 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3757 !strconcat(OpcodeStr,
3758 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3760 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3762 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3763 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3764 !strconcat(OpcodeStr,
3765 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3767 (X86insrtps VR128:$src1,
3768 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3769 imm:$src3))]>, OpSize;
3773 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3775 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3776 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3778 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3779 // the intel intrinsic that corresponds to this.
3780 let Defs = [EFLAGS] in {
3781 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3782 "ptest \t{$src2, $src1|$src1, $src2}",
3783 [(X86ptest VR128:$src1, VR128:$src2),
3784 (implicit EFLAGS)]>, OpSize;
3785 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3786 "ptest \t{$src2, $src1|$src1, $src2}",
3787 [(X86ptest VR128:$src1, (load addr:$src2)),
3788 (implicit EFLAGS)]>, OpSize;
3791 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3792 "movntdqa\t{$src, $dst|$dst, $src}",
3793 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
3797 //===----------------------------------------------------------------------===//
3798 // SSE4.2 Instructions
3799 //===----------------------------------------------------------------------===//
3801 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3802 let Constraints = "$src1 = $dst" in {
3803 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3804 Intrinsic IntId128, bit Commutable = 0> {
3805 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3806 (ins VR128:$src1, VR128:$src2),
3807 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3808 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3810 let isCommutable = Commutable;
3812 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3813 (ins VR128:$src1, i128mem:$src2),
3814 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3816 (IntId128 VR128:$src1,
3817 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3821 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3823 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3824 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3825 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3826 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3828 // crc intrinsic instruction
3829 // This set of instructions are only rm, the only difference is the size
3831 let Constraints = "$src1 = $dst" in {
3832 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3833 (ins GR32:$src1, i8mem:$src2),
3834 "crc32 \t{$src2, $src1|$src1, $src2}",
3836 (int_x86_sse42_crc32_8 GR32:$src1,
3837 (load addr:$src2)))]>, OpSize;
3838 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3839 (ins GR32:$src1, GR8:$src2),
3840 "crc32 \t{$src2, $src1|$src1, $src2}",
3842 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>,
3844 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3845 (ins GR32:$src1, i16mem:$src2),
3846 "crc32 \t{$src2, $src1|$src1, $src2}",
3848 (int_x86_sse42_crc32_16 GR32:$src1,
3849 (load addr:$src2)))]>,
3851 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3852 (ins GR32:$src1, GR16:$src2),
3853 "crc32 \t{$src2, $src1|$src1, $src2}",
3855 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3857 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3858 (ins GR32:$src1, i32mem:$src2),
3859 "crc32 \t{$src2, $src1|$src1, $src2}",
3861 (int_x86_sse42_crc32_32 GR32:$src1,
3862 (load addr:$src2)))]>, OpSize;
3863 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3864 (ins GR32:$src1, GR32:$src2),
3865 "crc32 \t{$src2, $src1|$src1, $src2}",
3867 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>,
3869 def CRC64m64 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3870 (ins GR64:$src1, i64mem:$src2),
3871 "crc32 \t{$src2, $src1|$src1, $src2}",
3873 (int_x86_sse42_crc32_64 GR64:$src1,
3874 (load addr:$src2)))]>,
3876 def CRC64r64 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3877 (ins GR64:$src1, GR64:$src2),
3878 "crc32 \t{$src2, $src1|$src1, $src2}",
3880 (int_x86_sse42_crc32_64 GR64:$src1, GR64:$src2))]>,
3884 // String/text processing instructions.
3885 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3886 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3887 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3888 "#PCMPISTRM128rr PSEUDO!",
3889 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3890 imm:$src3))]>, OpSize;
3891 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3892 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3893 "#PCMPISTRM128rm PSEUDO!",
3894 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3895 imm:$src3))]>, OpSize;
3898 let Defs = [XMM0, EFLAGS] in {
3899 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3900 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3901 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3902 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3903 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3904 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3907 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3908 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3909 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3910 "#PCMPESTRM128rr PSEUDO!",
3912 (int_x86_sse42_pcmpestrm128
3913 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3915 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3916 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3917 "#PCMPESTRM128rm PSEUDO!",
3918 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3919 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3923 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3924 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3925 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3926 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3927 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3928 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3929 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3932 let Defs = [ECX, EFLAGS] in {
3933 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3934 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3935 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3936 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3937 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3938 (implicit EFLAGS)]>, OpSize;
3939 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3940 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3941 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3942 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3943 (implicit EFLAGS)]>, OpSize;
3947 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3948 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3949 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3950 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3951 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3952 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3954 let Defs = [ECX, EFLAGS] in {
3955 let Uses = [EAX, EDX] in {
3956 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3957 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3958 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3959 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3960 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3961 (implicit EFLAGS)]>, OpSize;
3962 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3963 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3964 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3966 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
3967 (implicit EFLAGS)]>, OpSize;
3972 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3973 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3974 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3975 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3976 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3977 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;