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<1, 2, [SDTCisVT<0, i32>,
75 def X86ptest : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
77 //===----------------------------------------------------------------------===//
78 // SSE Complex Patterns
79 //===----------------------------------------------------------------------===//
81 // These are 'extloads' from a scalar to the low element of a vector, zeroing
82 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
84 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
85 [SDNPHasChain, SDNPMayLoad]>;
86 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
87 [SDNPHasChain, SDNPMayLoad]>;
89 def ssmem : Operand<v4f32> {
90 let PrintMethod = "printf32mem";
91 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
92 let ParserMatchClass = X86MemAsmOperand;
94 def sdmem : Operand<v2f64> {
95 let PrintMethod = "printf64mem";
96 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
97 let ParserMatchClass = X86MemAsmOperand;
100 //===----------------------------------------------------------------------===//
101 // SSE pattern fragments
102 //===----------------------------------------------------------------------===//
104 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
105 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
106 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
107 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
109 // Like 'store', but always requires vector alignment.
110 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
111 (store node:$val, node:$ptr), [{
112 return cast<StoreSDNode>(N)->getAlignment() >= 16;
115 // Like 'load', but always requires vector alignment.
116 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
117 return cast<LoadSDNode>(N)->getAlignment() >= 16;
120 def alignedloadfsf32 : PatFrag<(ops node:$ptr),
121 (f32 (alignedload node:$ptr))>;
122 def alignedloadfsf64 : PatFrag<(ops node:$ptr),
123 (f64 (alignedload node:$ptr))>;
124 def alignedloadv4f32 : PatFrag<(ops node:$ptr),
125 (v4f32 (alignedload node:$ptr))>;
126 def alignedloadv2f64 : PatFrag<(ops node:$ptr),
127 (v2f64 (alignedload node:$ptr))>;
128 def alignedloadv4i32 : PatFrag<(ops node:$ptr),
129 (v4i32 (alignedload node:$ptr))>;
130 def alignedloadv2i64 : PatFrag<(ops node:$ptr),
131 (v2i64 (alignedload node:$ptr))>;
133 // Like 'load', but uses special alignment checks suitable for use in
134 // memory operands in most SSE instructions, which are required to
135 // be naturally aligned on some targets but not on others. If the subtarget
136 // allows unaligned accesses, match any load, though this may require
137 // setting a feature bit in the processor (on startup, for example).
138 // Opteron 10h and later implement such a feature.
139 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
140 return Subtarget->hasVectorUAMem()
141 || cast<LoadSDNode>(N)->getAlignment() >= 16;
144 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
145 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
146 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
147 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
148 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
149 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
150 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
152 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
154 // FIXME: 8 byte alignment for mmx reads is not required
155 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
156 return cast<LoadSDNode>(N)->getAlignment() >= 8;
159 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
160 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
161 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
162 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
165 // Like 'store', but requires the non-temporal bit to be set
166 def nontemporalstore : PatFrag<(ops node:$val, node:$ptr),
167 (st node:$val, node:$ptr), [{
168 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
169 return ST->isNonTemporal();
173 def alignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
174 (st node:$val, node:$ptr), [{
175 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
176 return ST->isNonTemporal() && !ST->isTruncatingStore() &&
177 ST->getAddressingMode() == ISD::UNINDEXED &&
178 ST->getAlignment() >= 16;
182 def unalignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
183 (st node:$val, node:$ptr), [{
184 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
185 return ST->isNonTemporal() &&
186 ST->getAlignment() < 16;
190 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
191 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
192 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
193 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
194 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
195 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
197 def vzmovl_v2i64 : PatFrag<(ops node:$src),
198 (bitconvert (v2i64 (X86vzmovl
199 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
200 def vzmovl_v4i32 : PatFrag<(ops node:$src),
201 (bitconvert (v4i32 (X86vzmovl
202 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
204 def vzload_v2i64 : PatFrag<(ops node:$src),
205 (bitconvert (v2i64 (X86vzload node:$src)))>;
208 def fp32imm0 : PatLeaf<(f32 fpimm), [{
209 return N->isExactlyValue(+0.0);
212 // BYTE_imm - Transform bit immediates into byte immediates.
213 def BYTE_imm : SDNodeXForm<imm, [{
214 // Transformation function: imm >> 3
215 return getI32Imm(N->getZExtValue() >> 3);
218 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
220 def SHUFFLE_get_shuf_imm : SDNodeXForm<vector_shuffle, [{
221 return getI8Imm(X86::getShuffleSHUFImmediate(N));
224 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
226 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<vector_shuffle, [{
227 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
230 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
232 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<vector_shuffle, [{
233 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
236 // SHUFFLE_get_palign_imm xform function: convert vector_shuffle mask to
238 def SHUFFLE_get_palign_imm : SDNodeXForm<vector_shuffle, [{
239 return getI8Imm(X86::getShufflePALIGNRImmediate(N));
242 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
243 (vector_shuffle node:$lhs, node:$rhs), [{
244 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
245 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
248 def movddup : PatFrag<(ops node:$lhs, node:$rhs),
249 (vector_shuffle node:$lhs, node:$rhs), [{
250 return X86::isMOVDDUPMask(cast<ShuffleVectorSDNode>(N));
253 def movhlps : PatFrag<(ops node:$lhs, node:$rhs),
254 (vector_shuffle node:$lhs, node:$rhs), [{
255 return X86::isMOVHLPSMask(cast<ShuffleVectorSDNode>(N));
258 def movhlps_undef : PatFrag<(ops node:$lhs, node:$rhs),
259 (vector_shuffle node:$lhs, node:$rhs), [{
260 return X86::isMOVHLPS_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
263 def movlhps : PatFrag<(ops node:$lhs, node:$rhs),
264 (vector_shuffle node:$lhs, node:$rhs), [{
265 return X86::isMOVLHPSMask(cast<ShuffleVectorSDNode>(N));
268 def movlp : PatFrag<(ops node:$lhs, node:$rhs),
269 (vector_shuffle node:$lhs, node:$rhs), [{
270 return X86::isMOVLPMask(cast<ShuffleVectorSDNode>(N));
273 def movl : PatFrag<(ops node:$lhs, node:$rhs),
274 (vector_shuffle node:$lhs, node:$rhs), [{
275 return X86::isMOVLMask(cast<ShuffleVectorSDNode>(N));
278 def movshdup : PatFrag<(ops node:$lhs, node:$rhs),
279 (vector_shuffle node:$lhs, node:$rhs), [{
280 return X86::isMOVSHDUPMask(cast<ShuffleVectorSDNode>(N));
283 def movsldup : PatFrag<(ops node:$lhs, node:$rhs),
284 (vector_shuffle node:$lhs, node:$rhs), [{
285 return X86::isMOVSLDUPMask(cast<ShuffleVectorSDNode>(N));
288 def unpckl : PatFrag<(ops node:$lhs, node:$rhs),
289 (vector_shuffle node:$lhs, node:$rhs), [{
290 return X86::isUNPCKLMask(cast<ShuffleVectorSDNode>(N));
293 def unpckh : PatFrag<(ops node:$lhs, node:$rhs),
294 (vector_shuffle node:$lhs, node:$rhs), [{
295 return X86::isUNPCKHMask(cast<ShuffleVectorSDNode>(N));
298 def unpckl_undef : PatFrag<(ops node:$lhs, node:$rhs),
299 (vector_shuffle node:$lhs, node:$rhs), [{
300 return X86::isUNPCKL_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
303 def unpckh_undef : PatFrag<(ops node:$lhs, node:$rhs),
304 (vector_shuffle node:$lhs, node:$rhs), [{
305 return X86::isUNPCKH_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
308 def pshufd : PatFrag<(ops node:$lhs, node:$rhs),
309 (vector_shuffle node:$lhs, node:$rhs), [{
310 return X86::isPSHUFDMask(cast<ShuffleVectorSDNode>(N));
311 }], SHUFFLE_get_shuf_imm>;
313 def shufp : PatFrag<(ops node:$lhs, node:$rhs),
314 (vector_shuffle node:$lhs, node:$rhs), [{
315 return X86::isSHUFPMask(cast<ShuffleVectorSDNode>(N));
316 }], SHUFFLE_get_shuf_imm>;
318 def pshufhw : PatFrag<(ops node:$lhs, node:$rhs),
319 (vector_shuffle node:$lhs, node:$rhs), [{
320 return X86::isPSHUFHWMask(cast<ShuffleVectorSDNode>(N));
321 }], SHUFFLE_get_pshufhw_imm>;
323 def pshuflw : PatFrag<(ops node:$lhs, node:$rhs),
324 (vector_shuffle node:$lhs, node:$rhs), [{
325 return X86::isPSHUFLWMask(cast<ShuffleVectorSDNode>(N));
326 }], SHUFFLE_get_pshuflw_imm>;
328 def palign : PatFrag<(ops node:$lhs, node:$rhs),
329 (vector_shuffle node:$lhs, node:$rhs), [{
330 return X86::isPALIGNRMask(cast<ShuffleVectorSDNode>(N));
331 }], SHUFFLE_get_palign_imm>;
333 //===----------------------------------------------------------------------===//
334 // SSE scalar FP Instructions
335 //===----------------------------------------------------------------------===//
337 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded after
338 // instruction selection into a branch sequence.
339 let Uses = [EFLAGS], usesCustomInserter = 1 in {
340 def CMOV_FR32 : I<0, Pseudo,
341 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
342 "#CMOV_FR32 PSEUDO!",
343 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
345 def CMOV_FR64 : I<0, Pseudo,
346 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
347 "#CMOV_FR64 PSEUDO!",
348 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
350 def CMOV_V4F32 : I<0, Pseudo,
351 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
352 "#CMOV_V4F32 PSEUDO!",
354 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
356 def CMOV_V2F64 : I<0, Pseudo,
357 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
358 "#CMOV_V2F64 PSEUDO!",
360 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
362 def CMOV_V2I64 : I<0, Pseudo,
363 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
364 "#CMOV_V2I64 PSEUDO!",
366 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
370 //===----------------------------------------------------------------------===//
372 //===----------------------------------------------------------------------===//
374 // Move Instructions. Register-to-register movss is not used for FR32
375 // register copies because it's a partial register update; FsMOVAPSrr is
376 // used instead. Register-to-register movss is not modeled as an INSERT_SUBREG
377 // because INSERT_SUBREG requires that the insert be implementable in terms of
378 // a copy, and just mentioned, we don't use movss for copies.
379 let Constraints = "$src1 = $dst" in
380 def MOVSSrr : SSI<0x10, MRMSrcReg,
381 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
382 "movss\t{$src2, $dst|$dst, $src2}",
383 [(set (v4f32 VR128:$dst),
384 (movl VR128:$src1, (scalar_to_vector FR32:$src2)))]>;
386 // Extract the low 32-bit value from one vector and insert it into another.
387 let AddedComplexity = 15 in
388 def : Pat<(v4f32 (movl VR128:$src1, VR128:$src2)),
389 (MOVSSrr (v4f32 VR128:$src1),
390 (EXTRACT_SUBREG (v4f32 VR128:$src2), sub_ss))>;
392 // Implicitly promote a 32-bit scalar to a vector.
393 def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
394 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), FR32:$src, sub_ss)>;
396 // Loading from memory automatically zeroing upper bits.
397 let canFoldAsLoad = 1, isReMaterializable = 1 in
398 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
399 "movss\t{$src, $dst|$dst, $src}",
400 [(set FR32:$dst, (loadf32 addr:$src))]>;
402 // MOVSSrm zeros the high parts of the register; represent this
403 // with SUBREG_TO_REG.
404 let AddedComplexity = 20 in {
405 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
406 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
407 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
408 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
409 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
410 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
413 // Store scalar value to memory.
414 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
415 "movss\t{$src, $dst|$dst, $src}",
416 [(store FR32:$src, addr:$dst)]>;
418 // Extract and store.
419 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
422 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
424 // Conversion instructions
425 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
426 "cvttss2si\t{$src, $dst|$dst, $src}",
427 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
428 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
429 "cvttss2si\t{$src, $dst|$dst, $src}",
430 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
431 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
432 "cvtsi2ss\t{$src, $dst|$dst, $src}",
433 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
434 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
435 "cvtsi2ss\t{$src, $dst|$dst, $src}",
436 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
438 // Match intrinsics which expect XMM operand(s).
439 def CVTSS2SIrr: SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
440 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
441 def CVTSS2SIrm: SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
442 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
444 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
445 "cvtss2si\t{$src, $dst|$dst, $src}",
446 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
447 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
448 "cvtss2si\t{$src, $dst|$dst, $src}",
449 [(set GR32:$dst, (int_x86_sse_cvtss2si
450 (load addr:$src)))]>;
452 // Match intrinsics which expect MM and XMM operand(s).
453 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
454 "cvtps2pi\t{$src, $dst|$dst, $src}",
455 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
456 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
457 "cvtps2pi\t{$src, $dst|$dst, $src}",
458 [(set VR64:$dst, (int_x86_sse_cvtps2pi
459 (load addr:$src)))]>;
460 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
461 "cvttps2pi\t{$src, $dst|$dst, $src}",
462 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
463 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
464 "cvttps2pi\t{$src, $dst|$dst, $src}",
465 [(set VR64:$dst, (int_x86_sse_cvttps2pi
466 (load addr:$src)))]>;
467 let Constraints = "$src1 = $dst" in {
468 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
469 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
470 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
471 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
473 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
474 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
475 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
476 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
477 (load addr:$src2)))]>;
480 // Aliases for intrinsics
481 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
482 "cvttss2si\t{$src, $dst|$dst, $src}",
484 (int_x86_sse_cvttss2si VR128:$src))]>;
485 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
486 "cvttss2si\t{$src, $dst|$dst, $src}",
488 (int_x86_sse_cvttss2si(load addr:$src)))]>;
490 let Constraints = "$src1 = $dst" in {
491 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
492 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
493 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
494 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
496 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
497 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
498 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
499 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
500 (loadi32 addr:$src2)))]>;
503 // Comparison instructions
504 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
505 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
506 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
507 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
509 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
510 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
511 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
513 // Accept explicit immediate argument form instead of comparison code.
514 let isAsmParserOnly = 1 in {
515 def CMPSSrr_alt : SSIi8<0xC2, MRMSrcReg,
516 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, i8imm:$src2),
517 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
519 def CMPSSrm_alt : SSIi8<0xC2, MRMSrcMem,
520 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, i8imm:$src2),
521 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
525 let Defs = [EFLAGS] in {
526 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
527 "ucomiss\t{$src2, $src1|$src1, $src2}",
528 [(set EFLAGS, (X86cmp FR32:$src1, FR32:$src2))]>;
529 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
530 "ucomiss\t{$src2, $src1|$src1, $src2}",
531 [(set EFLAGS, (X86cmp FR32:$src1, (loadf32 addr:$src2)))]>;
533 def COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
534 "comiss\t{$src2, $src1|$src1, $src2}", []>;
535 def COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
536 "comiss\t{$src2, $src1|$src1, $src2}", []>;
540 // Aliases to match intrinsics which expect XMM operand(s).
541 let Constraints = "$src1 = $dst" in {
542 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
544 (ins VR128:$src1, VR128:$src, SSECC:$cc),
545 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
546 [(set VR128:$dst, (int_x86_sse_cmp_ss
548 VR128:$src, imm:$cc))]>;
549 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
551 (ins VR128:$src1, f32mem:$src, SSECC:$cc),
552 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
553 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
554 (load addr:$src), imm:$cc))]>;
557 let Defs = [EFLAGS] in {
558 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
559 "ucomiss\t{$src2, $src1|$src1, $src2}",
560 [(set EFLAGS, (X86ucomi (v4f32 VR128:$src1),
562 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
563 "ucomiss\t{$src2, $src1|$src1, $src2}",
564 [(set EFLAGS, (X86ucomi (v4f32 VR128:$src1),
565 (load addr:$src2)))]>;
567 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
568 "comiss\t{$src2, $src1|$src1, $src2}",
569 [(set EFLAGS, (X86comi (v4f32 VR128:$src1),
571 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
572 "comiss\t{$src2, $src1|$src1, $src2}",
573 [(set EFLAGS, (X86comi (v4f32 VR128:$src1),
574 (load addr:$src2)))]>;
577 // Aliases of packed SSE1 instructions for scalar use. These all have names
578 // that start with 'Fs'.
580 // Alias instructions that map fld0 to pxor for sse.
581 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
583 // FIXME: Set encoding to pseudo!
584 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
585 [(set FR32:$dst, fp32imm0)]>,
586 Requires<[HasSSE1]>, TB, OpSize;
588 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
590 let neverHasSideEffects = 1 in
591 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
592 "movaps\t{$src, $dst|$dst, $src}", []>;
594 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
596 let canFoldAsLoad = 1, isReMaterializable = 1 in
597 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
598 "movaps\t{$src, $dst|$dst, $src}",
599 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
601 // Alias bitwise logical operations using SSE logical ops on packed FP values.
602 let Constraints = "$src1 = $dst" in {
603 let isCommutable = 1 in {
604 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst),
605 (ins FR32:$src1, FR32:$src2),
606 "andps\t{$src2, $dst|$dst, $src2}",
607 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
608 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst),
609 (ins FR32:$src1, FR32:$src2),
610 "orps\t{$src2, $dst|$dst, $src2}",
611 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
612 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst),
613 (ins FR32:$src1, FR32:$src2),
614 "xorps\t{$src2, $dst|$dst, $src2}",
615 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
618 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst),
619 (ins FR32:$src1, f128mem:$src2),
620 "andps\t{$src2, $dst|$dst, $src2}",
621 [(set FR32:$dst, (X86fand FR32:$src1,
622 (memopfsf32 addr:$src2)))]>;
623 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst),
624 (ins FR32:$src1, f128mem:$src2),
625 "orps\t{$src2, $dst|$dst, $src2}",
626 [(set FR32:$dst, (X86for FR32:$src1,
627 (memopfsf32 addr:$src2)))]>;
628 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst),
629 (ins FR32:$src1, f128mem:$src2),
630 "xorps\t{$src2, $dst|$dst, $src2}",
631 [(set FR32:$dst, (X86fxor FR32:$src1,
632 (memopfsf32 addr:$src2)))]>;
634 let neverHasSideEffects = 1 in {
635 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
636 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
637 "andnps\t{$src2, $dst|$dst, $src2}", []>;
639 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
640 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
641 "andnps\t{$src2, $dst|$dst, $src2}", []>;
645 /// basic_sse12_fp_binop_rm - SSE 1 & 2 binops come in both scalar and
648 /// In addition, we also have a special variant of the scalar form here to
649 /// represent the associated intrinsic operation. This form is unlike the
650 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
651 /// and leaves the top elements unmodified (therefore these cannot be commuted).
653 /// These three forms can each be reg+reg or reg+mem, so there are a total of
654 /// six "instructions".
656 let Constraints = "$src1 = $dst" in {
657 multiclass basic_sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
658 SDNode OpNode, bit Commutable = 0> {
659 // Scalar operation, reg+reg.
660 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
661 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
662 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
663 let isCommutable = Commutable;
666 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
667 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
668 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
669 let isCommutable = Commutable;
672 // Scalar operation, reg+mem.
673 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
674 (ins FR32:$src1, f32mem:$src2),
675 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
676 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
678 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
679 (ins FR64:$src1, f64mem:$src2),
680 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
681 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
683 // Vector operation, reg+reg.
684 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
685 (ins VR128:$src1, VR128:$src2),
686 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
687 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
688 let isCommutable = Commutable;
691 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
692 (ins VR128:$src1, VR128:$src2),
693 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
694 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
695 let isCommutable = Commutable;
698 // Vector operation, reg+mem.
699 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
700 (ins VR128:$src1, f128mem:$src2),
701 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
702 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
704 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
705 (ins VR128:$src1, f128mem:$src2),
706 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
707 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
709 // Intrinsic operation, reg+reg.
710 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
711 (ins VR128:$src1, VR128:$src2),
712 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
713 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
714 !strconcat(OpcodeStr, "_ss")) VR128:$src1,
716 // int_x86_sse_xxx_ss
718 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
719 (ins VR128:$src1, VR128:$src2),
720 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
721 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
722 !strconcat(OpcodeStr, "_sd")) VR128:$src1,
724 // int_x86_sse2_xxx_sd
726 // Intrinsic operation, reg+mem.
727 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
728 (ins VR128:$src1, ssmem:$src2),
729 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
730 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
731 !strconcat(OpcodeStr, "_ss")) VR128:$src1,
732 sse_load_f32:$src2))]>;
733 // int_x86_sse_xxx_ss
735 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
736 (ins VR128:$src1, sdmem:$src2),
737 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
738 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
739 !strconcat(OpcodeStr, "_sd")) VR128:$src1,
740 sse_load_f64:$src2))]>;
741 // int_x86_sse2_xxx_sd
745 // Arithmetic instructions
746 defm ADD : basic_sse12_fp_binop_rm<0x58, "add", fadd, 1>;
747 defm MUL : basic_sse12_fp_binop_rm<0x59, "mul", fmul, 1>;
748 defm SUB : basic_sse12_fp_binop_rm<0x5C, "sub", fsub>;
749 defm DIV : basic_sse12_fp_binop_rm<0x5E, "div", fdiv>;
751 /// sse12_fp_binop_rm - Other SSE 1 & 2 binops
753 /// This multiclass is like basic_sse12_fp_binop_rm, with the addition of
754 /// instructions for a full-vector intrinsic form. Operations that map
755 /// onto C operators don't use this form since they just use the plain
756 /// vector form instead of having a separate vector intrinsic form.
758 /// This provides a total of eight "instructions".
760 let Constraints = "$src1 = $dst" in {
761 multiclass sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
762 SDNode OpNode, bit Commutable = 0> {
764 // Scalar operation, reg+reg.
765 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
766 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
767 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
768 let isCommutable = Commutable;
771 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
772 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
773 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
774 let isCommutable = Commutable;
777 // Scalar operation, reg+mem.
778 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
779 (ins FR32:$src1, f32mem:$src2),
780 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
781 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
783 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
784 (ins FR64:$src1, f64mem:$src2),
785 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
786 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
788 // Vector operation, reg+reg.
789 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
790 (ins VR128:$src1, VR128:$src2),
791 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
792 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
793 let isCommutable = Commutable;
796 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
797 (ins VR128:$src1, VR128:$src2),
798 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
799 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
800 let isCommutable = Commutable;
803 // Vector operation, reg+mem.
804 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
805 (ins VR128:$src1, f128mem:$src2),
806 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
807 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
809 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
810 (ins VR128:$src1, f128mem:$src2),
811 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
812 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
814 // Intrinsic operation, reg+reg.
815 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
816 (ins VR128:$src1, VR128:$src2),
817 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
818 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
819 !strconcat(OpcodeStr, "_ss")) VR128:$src1,
821 // int_x86_sse_xxx_ss
822 let isCommutable = Commutable;
825 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
826 (ins VR128:$src1, VR128:$src2),
827 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
828 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
829 !strconcat(OpcodeStr, "_sd")) VR128:$src1,
831 // int_x86_sse2_xxx_sd
832 let isCommutable = Commutable;
835 // Intrinsic operation, reg+mem.
836 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
837 (ins VR128:$src1, ssmem:$src2),
838 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
839 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
840 !strconcat(OpcodeStr, "_ss")) VR128:$src1,
841 sse_load_f32:$src2))]>;
842 // int_x86_sse_xxx_ss
844 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
845 (ins VR128:$src1, sdmem:$src2),
846 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
847 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
848 !strconcat(OpcodeStr, "_sd")) VR128:$src1,
849 sse_load_f64:$src2))]>;
850 // int_x86_sse2_xxx_sd
852 // Vector intrinsic operation, reg+reg.
853 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
854 (ins VR128:$src1, VR128:$src2),
855 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
856 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
857 !strconcat(OpcodeStr, "_ps")) VR128:$src1,
859 // int_x86_sse_xxx_ps
860 let isCommutable = Commutable;
863 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
864 (ins VR128:$src1, VR128:$src2),
865 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
866 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
867 !strconcat(OpcodeStr, "_pd")) VR128:$src1,
869 // int_x86_sse2_xxx_pd
870 let isCommutable = Commutable;
873 // Vector intrinsic operation, reg+mem.
874 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
875 (ins VR128:$src1, f128mem:$src2),
876 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
877 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
878 !strconcat(OpcodeStr, "_ps")) VR128:$src1,
879 (memopv4f32 addr:$src2)))]>;
880 // int_x86_sse_xxx_ps
882 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
883 (ins VR128:$src1, f128mem:$src2),
884 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
885 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
886 !strconcat(OpcodeStr, "_pd")) VR128:$src1,
887 (memopv2f64 addr:$src2)))]>;
888 // int_x86_sse2_xxx_pd
892 defm MAX : sse12_fp_binop_rm<0x5F, "max", X86fmax>;
893 defm MIN : sse12_fp_binop_rm<0x5D, "min", X86fmin>;
895 //===----------------------------------------------------------------------===//
896 // SSE packed FP Instructions
899 let neverHasSideEffects = 1 in
900 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
901 "movaps\t{$src, $dst|$dst, $src}", []>;
902 let canFoldAsLoad = 1, isReMaterializable = 1 in
903 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
904 "movaps\t{$src, $dst|$dst, $src}",
905 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
907 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
908 "movaps\t{$src, $dst|$dst, $src}",
909 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
911 let neverHasSideEffects = 1 in
912 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
913 "movups\t{$src, $dst|$dst, $src}", []>;
914 let canFoldAsLoad = 1, isReMaterializable = 1 in
915 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
916 "movups\t{$src, $dst|$dst, $src}",
917 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
918 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
919 "movups\t{$src, $dst|$dst, $src}",
920 [(store (v4f32 VR128:$src), addr:$dst)]>;
922 // Intrinsic forms of MOVUPS load and store
923 let canFoldAsLoad = 1, isReMaterializable = 1 in
924 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
925 "movups\t{$src, $dst|$dst, $src}",
926 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
927 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
928 "movups\t{$src, $dst|$dst, $src}",
929 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
931 let Constraints = "$src1 = $dst" in {
932 let AddedComplexity = 20 in {
933 def MOVLPSrm : PSI<0x12, MRMSrcMem,
934 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
935 "movlps\t{$src2, $dst|$dst, $src2}",
938 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
939 def MOVHPSrm : PSI<0x16, MRMSrcMem,
940 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
941 "movhps\t{$src2, $dst|$dst, $src2}",
943 (movlhps VR128:$src1,
944 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
946 } // Constraints = "$src1 = $dst"
949 def : Pat<(movlhps VR128:$src1, (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
950 (MOVHPSrm (v4i32 VR128:$src1), addr:$src2)>;
952 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
953 "movlps\t{$src, $dst|$dst, $src}",
954 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
955 (iPTR 0))), addr:$dst)]>;
957 // v2f64 extract element 1 is always custom lowered to unpack high to low
958 // and extract element 0 so the non-store version isn't too horrible.
959 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
960 "movhps\t{$src, $dst|$dst, $src}",
961 [(store (f64 (vector_extract
962 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
963 (undef)), (iPTR 0))), addr:$dst)]>;
965 let Constraints = "$src1 = $dst" in {
966 let AddedComplexity = 20 in {
967 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
968 (ins VR128:$src1, VR128:$src2),
969 "movlhps\t{$src2, $dst|$dst, $src2}",
971 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
973 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
974 (ins VR128:$src1, VR128:$src2),
975 "movhlps\t{$src2, $dst|$dst, $src2}",
977 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
979 } // Constraints = "$src1 = $dst"
981 let AddedComplexity = 20 in {
982 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
983 (MOVLHPSrr (v4f32 VR128:$src), (v4f32 VR128:$src))>;
984 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
985 (MOVLHPSrr (v2i64 VR128:$src), (v2i64 VR128:$src))>;
992 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
994 /// In addition, we also have a special variant of the scalar form here to
995 /// represent the associated intrinsic operation. This form is unlike the
996 /// plain scalar form, in that it takes an entire vector (instead of a
997 /// scalar) and leaves the top elements undefined.
999 /// And, we have a special variant form for a full-vector intrinsic form.
1001 /// These four forms can each have a reg or a mem operand, so there are a
1002 /// total of eight "instructions".
1004 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
1008 bit Commutable = 0> {
1009 // Scalar operation, reg.
1010 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
1011 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1012 [(set FR32:$dst, (OpNode FR32:$src))]> {
1013 let isCommutable = Commutable;
1016 // Scalar operation, mem.
1017 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
1018 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1019 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
1020 Requires<[HasSSE1, OptForSize]>;
1022 // Vector operation, reg.
1023 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1024 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1025 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
1026 let isCommutable = Commutable;
1029 // Vector operation, mem.
1030 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1031 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1032 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
1034 // Intrinsic operation, reg.
1035 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1036 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1037 [(set VR128:$dst, (F32Int VR128:$src))]> {
1038 let isCommutable = Commutable;
1041 // Intrinsic operation, mem.
1042 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
1043 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1044 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
1046 // Vector intrinsic operation, reg
1047 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1048 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1049 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
1050 let isCommutable = Commutable;
1053 // Vector intrinsic operation, mem
1054 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1055 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1056 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
1060 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
1061 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
1063 // Reciprocal approximations. Note that these typically require refinement
1064 // in order to obtain suitable precision.
1065 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
1066 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
1067 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
1068 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
1071 let Constraints = "$src1 = $dst" in {
1072 let isCommutable = 1 in {
1073 def ANDPSrr : PSI<0x54, MRMSrcReg,
1074 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1075 "andps\t{$src2, $dst|$dst, $src2}",
1076 [(set VR128:$dst, (v2i64
1077 (and VR128:$src1, VR128:$src2)))]>;
1078 def ORPSrr : PSI<0x56, MRMSrcReg,
1079 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1080 "orps\t{$src2, $dst|$dst, $src2}",
1081 [(set VR128:$dst, (v2i64
1082 (or VR128:$src1, VR128:$src2)))]>;
1083 def XORPSrr : PSI<0x57, MRMSrcReg,
1084 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1085 "xorps\t{$src2, $dst|$dst, $src2}",
1086 [(set VR128:$dst, (v2i64
1087 (xor VR128:$src1, VR128:$src2)))]>;
1090 def ANDPSrm : PSI<0x54, MRMSrcMem,
1091 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1092 "andps\t{$src2, $dst|$dst, $src2}",
1093 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
1094 (memopv2i64 addr:$src2)))]>;
1095 def ORPSrm : PSI<0x56, MRMSrcMem,
1096 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1097 "orps\t{$src2, $dst|$dst, $src2}",
1098 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
1099 (memopv2i64 addr:$src2)))]>;
1100 def XORPSrm : PSI<0x57, MRMSrcMem,
1101 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1102 "xorps\t{$src2, $dst|$dst, $src2}",
1103 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
1104 (memopv2i64 addr:$src2)))]>;
1105 def ANDNPSrr : PSI<0x55, MRMSrcReg,
1106 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1107 "andnps\t{$src2, $dst|$dst, $src2}",
1109 (v2i64 (and (xor VR128:$src1,
1110 (bc_v2i64 (v4i32 immAllOnesV))),
1112 def ANDNPSrm : PSI<0x55, MRMSrcMem,
1113 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1114 "andnps\t{$src2, $dst|$dst, $src2}",
1116 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
1117 (bc_v2i64 (v4i32 immAllOnesV))),
1118 (memopv2i64 addr:$src2))))]>;
1121 let Constraints = "$src1 = $dst" in {
1122 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
1123 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1124 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1125 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1126 VR128:$src, imm:$cc))]>;
1127 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
1128 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1129 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1130 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1131 (memop addr:$src), imm:$cc))]>;
1133 // Accept explicit immediate argument form instead of comparison code.
1134 let isAsmParserOnly = 1 in {
1135 def CMPPSrri_alt : PSIi8<0xC2, MRMSrcReg,
1136 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1137 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1138 def CMPPSrmi_alt : PSIi8<0xC2, MRMSrcMem,
1139 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1140 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1143 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1144 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
1145 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1146 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
1148 // Shuffle and unpack instructions
1149 let Constraints = "$src1 = $dst" in {
1150 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
1151 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
1152 (outs VR128:$dst), (ins VR128:$src1,
1153 VR128:$src2, i8imm:$src3),
1154 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1156 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1157 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
1158 (outs VR128:$dst), (ins VR128:$src1,
1159 f128mem:$src2, i8imm:$src3),
1160 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1163 VR128:$src1, (memopv4f32 addr:$src2))))]>;
1165 let AddedComplexity = 10 in {
1166 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
1167 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1168 "unpckhps\t{$src2, $dst|$dst, $src2}",
1170 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
1171 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
1172 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1173 "unpckhps\t{$src2, $dst|$dst, $src2}",
1175 (v4f32 (unpckh VR128:$src1,
1176 (memopv4f32 addr:$src2))))]>;
1178 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
1179 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1180 "unpcklps\t{$src2, $dst|$dst, $src2}",
1182 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
1183 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
1184 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1185 "unpcklps\t{$src2, $dst|$dst, $src2}",
1187 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
1188 } // AddedComplexity
1189 } // Constraints = "$src1 = $dst"
1192 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1193 "movmskps\t{$src, $dst|$dst, $src}",
1194 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
1195 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1196 "movmskpd\t{$src, $dst|$dst, $src}",
1197 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1199 // Prefetch intrinsic.
1200 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1201 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1202 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1203 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1204 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1205 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1206 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1207 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1209 // Non-temporal stores
1210 def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1211 "movntps\t{$src, $dst|$dst, $src}",
1212 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1214 let AddedComplexity = 400 in { // Prefer non-temporal versions
1215 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1216 "movntps\t{$src, $dst|$dst, $src}",
1217 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
1219 def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1220 "movntdq\t{$src, $dst|$dst, $src}",
1221 [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
1223 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
1224 "movnti\t{$src, $dst|$dst, $src}",
1225 [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
1226 TB, Requires<[HasSSE2]>;
1228 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
1229 "movnti\t{$src, $dst|$dst, $src}",
1230 [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
1231 TB, Requires<[HasSSE2]>;
1234 // Load, store, and memory fence
1235 def SFENCE : I<0xAE, MRM_F8, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>,
1236 TB, Requires<[HasSSE1]>;
1239 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1240 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1241 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1242 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1244 // Alias instructions that map zero vector to pxor / xorp* for sse.
1245 // We set canFoldAsLoad because this can be converted to a constant-pool
1246 // load of an all-zeros value if folding it would be beneficial.
1247 // FIXME: Change encoding to pseudo!
1248 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1249 isCodeGenOnly = 1 in {
1250 def V_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1251 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
1252 def V_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1253 [(set VR128:$dst, (v2f64 immAllZerosV))]>;
1254 let ExeDomain = SSEPackedInt in
1255 def V_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
1256 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1259 def : Pat<(v2i64 immAllZerosV), (V_SET0PI)>;
1260 def : Pat<(v8i16 immAllZerosV), (V_SET0PI)>;
1261 def : Pat<(v16i8 immAllZerosV), (V_SET0PI)>;
1263 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1264 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
1266 //===---------------------------------------------------------------------===//
1267 // SSE2 Instructions
1268 //===---------------------------------------------------------------------===//
1270 // Move Instructions. Register-to-register movsd is not used for FR64
1271 // register copies because it's a partial register update; FsMOVAPDrr is
1272 // used instead. Register-to-register movsd is not modeled as an INSERT_SUBREG
1273 // because INSERT_SUBREG requires that the insert be implementable in terms of
1274 // a copy, and just mentioned, we don't use movsd for copies.
1275 let Constraints = "$src1 = $dst" in
1276 def MOVSDrr : SDI<0x10, MRMSrcReg,
1277 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
1278 "movsd\t{$src2, $dst|$dst, $src2}",
1279 [(set (v2f64 VR128:$dst),
1280 (movl VR128:$src1, (scalar_to_vector FR64:$src2)))]>;
1282 // Extract the low 64-bit value from one vector and insert it into another.
1283 let AddedComplexity = 15 in
1284 def : Pat<(v2f64 (movl VR128:$src1, VR128:$src2)),
1285 (MOVSDrr (v2f64 VR128:$src1),
1286 (EXTRACT_SUBREG (v2f64 VR128:$src2), sub_sd))>;
1288 // Implicitly promote a 64-bit scalar to a vector.
1289 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
1290 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), FR64:$src, sub_sd)>;
1292 // Loading from memory automatically zeroing upper bits.
1293 let canFoldAsLoad = 1, isReMaterializable = 1, AddedComplexity = 20 in
1294 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1295 "movsd\t{$src, $dst|$dst, $src}",
1296 [(set FR64:$dst, (loadf64 addr:$src))]>;
1298 // MOVSDrm zeros the high parts of the register; represent this
1299 // with SUBREG_TO_REG.
1300 let AddedComplexity = 20 in {
1301 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
1302 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1303 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
1304 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1305 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
1306 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1307 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
1308 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1309 def : Pat<(v2f64 (X86vzload addr:$src)),
1310 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1313 // Store scalar value to memory.
1314 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1315 "movsd\t{$src, $dst|$dst, $src}",
1316 [(store FR64:$src, addr:$dst)]>;
1318 // Extract and store.
1319 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
1322 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
1324 // Conversion instructions
1325 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1326 "cvttsd2si\t{$src, $dst|$dst, $src}",
1327 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1328 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1329 "cvttsd2si\t{$src, $dst|$dst, $src}",
1330 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1331 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1332 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1333 [(set FR32:$dst, (fround FR64:$src))]>;
1334 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1335 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1336 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
1337 Requires<[HasSSE2, OptForSize]>;
1338 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1339 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1340 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1341 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1342 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1343 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1345 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1346 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1347 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1348 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1349 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1350 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1351 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1352 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1353 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1354 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1355 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1356 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1357 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1358 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1359 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1360 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1361 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1362 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1363 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1364 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1366 // SSE2 instructions with XS prefix
1367 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1368 "cvtss2sd\t{$src, $dst|$dst, $src}",
1369 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1370 Requires<[HasSSE2]>;
1371 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1372 "cvtss2sd\t{$src, $dst|$dst, $src}",
1373 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1374 Requires<[HasSSE2, OptForSize]>;
1376 def : Pat<(extloadf32 addr:$src),
1377 (CVTSS2SDrr (MOVSSrm addr:$src))>,
1378 Requires<[HasSSE2, OptForSpeed]>;
1380 // Match intrinsics which expect XMM operand(s).
1381 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1382 "cvtsd2si\t{$src, $dst|$dst, $src}",
1383 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1384 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1385 "cvtsd2si\t{$src, $dst|$dst, $src}",
1386 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1387 (load addr:$src)))]>;
1389 // Match intrinsics which expect MM and XMM operand(s).
1390 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1391 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1392 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1393 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1394 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1395 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1396 (memop addr:$src)))]>;
1397 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1398 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1399 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1400 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1401 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1402 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1403 (memop addr:$src)))]>;
1404 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1405 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1406 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1407 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1408 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1409 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1410 (load addr:$src)))]>;
1412 // Aliases for intrinsics
1413 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1414 "cvttsd2si\t{$src, $dst|$dst, $src}",
1416 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1417 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1418 "cvttsd2si\t{$src, $dst|$dst, $src}",
1419 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1420 (load addr:$src)))]>;
1422 // Comparison instructions
1423 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1424 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1425 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1426 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1428 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1429 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1430 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1432 // Accept explicit immediate argument form instead of comparison code.
1433 let isAsmParserOnly = 1 in {
1434 def CMPSDrr_alt : SDIi8<0xC2, MRMSrcReg,
1435 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, i8imm:$src2),
1436 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1438 def CMPSDrm_alt : SDIi8<0xC2, MRMSrcMem,
1439 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, i8imm:$src2),
1440 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1444 let Defs = [EFLAGS] in {
1445 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1446 "ucomisd\t{$src2, $src1|$src1, $src2}",
1447 [(set EFLAGS, (X86cmp FR64:$src1, FR64:$src2))]>;
1448 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1449 "ucomisd\t{$src2, $src1|$src1, $src2}",
1450 [(set EFLAGS, (X86cmp FR64:$src1, (loadf64 addr:$src2)))]>;
1451 } // Defs = [EFLAGS]
1453 // Aliases to match intrinsics which expect XMM operand(s).
1454 let Constraints = "$src1 = $dst" in {
1455 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1457 (ins VR128:$src1, VR128:$src, SSECC:$cc),
1458 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1459 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1460 VR128:$src, imm:$cc))]>;
1461 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1463 (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1464 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1465 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1466 (load addr:$src), imm:$cc))]>;
1469 let Defs = [EFLAGS] in {
1470 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1471 "ucomisd\t{$src2, $src1|$src1, $src2}",
1472 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1474 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1475 "ucomisd\t{$src2, $src1|$src1, $src2}",
1476 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1477 (load addr:$src2)))]>;
1479 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1480 "comisd\t{$src2, $src1|$src1, $src2}",
1481 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1483 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1484 "comisd\t{$src2, $src1|$src1, $src2}",
1485 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1486 (load addr:$src2)))]>;
1487 } // Defs = [EFLAGS]
1489 // Aliases of packed SSE2 instructions for scalar use. These all have names
1490 // that start with 'Fs'.
1492 // Alias instructions that map fld0 to pxor for sse.
1493 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1494 canFoldAsLoad = 1 in
1495 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1496 [(set FR64:$dst, fpimm0)]>,
1497 Requires<[HasSSE2]>, TB, OpSize;
1499 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1501 let neverHasSideEffects = 1 in
1502 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1503 "movapd\t{$src, $dst|$dst, $src}", []>;
1505 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1507 let canFoldAsLoad = 1, isReMaterializable = 1 in
1508 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1509 "movapd\t{$src, $dst|$dst, $src}",
1510 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1512 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1513 let Constraints = "$src1 = $dst" in {
1514 let isCommutable = 1 in {
1515 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1516 (ins FR64:$src1, FR64:$src2),
1517 "andpd\t{$src2, $dst|$dst, $src2}",
1518 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1519 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1520 (ins FR64:$src1, FR64:$src2),
1521 "orpd\t{$src2, $dst|$dst, $src2}",
1522 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1523 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1524 (ins FR64:$src1, FR64:$src2),
1525 "xorpd\t{$src2, $dst|$dst, $src2}",
1526 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1529 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1530 (ins FR64:$src1, f128mem:$src2),
1531 "andpd\t{$src2, $dst|$dst, $src2}",
1532 [(set FR64:$dst, (X86fand FR64:$src1,
1533 (memopfsf64 addr:$src2)))]>;
1534 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1535 (ins FR64:$src1, f128mem:$src2),
1536 "orpd\t{$src2, $dst|$dst, $src2}",
1537 [(set FR64:$dst, (X86for FR64:$src1,
1538 (memopfsf64 addr:$src2)))]>;
1539 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1540 (ins FR64:$src1, f128mem:$src2),
1541 "xorpd\t{$src2, $dst|$dst, $src2}",
1542 [(set FR64:$dst, (X86fxor FR64:$src1,
1543 (memopfsf64 addr:$src2)))]>;
1545 let neverHasSideEffects = 1 in {
1546 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1547 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1548 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1550 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1551 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1552 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1556 //===---------------------------------------------------------------------===//
1557 // SSE packed FP Instructions
1559 // Move Instructions
1560 let neverHasSideEffects = 1 in
1561 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$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:$cc))]>;
1894 // Accept explicit immediate argument form instead of comparison code.
1895 let isAsmParserOnly = 1 in {
1896 def CMPPDrri_alt : PDIi8<0xC2, MRMSrcReg,
1897 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1898 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1899 def CMPPDrmi_alt : PDIi8<0xC2, MRMSrcMem,
1900 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1901 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1904 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1905 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1906 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1907 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1909 // Shuffle and unpack instructions
1910 let Constraints = "$src1 = $dst" in {
1911 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1912 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1913 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1915 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1916 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1917 (outs VR128:$dst), (ins VR128:$src1,
1918 f128mem:$src2, i8imm:$src3),
1919 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1922 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1924 let AddedComplexity = 10 in {
1925 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1926 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1927 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1929 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1930 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1931 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1932 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1934 (v2f64 (unpckh VR128:$src1,
1935 (memopv2f64 addr:$src2))))]>;
1937 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1938 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1939 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1941 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1942 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1943 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1944 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1946 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1947 } // AddedComplexity
1948 } // Constraints = "$src1 = $dst"
1951 //===---------------------------------------------------------------------===//
1952 // SSE integer instructions
1953 let ExeDomain = SSEPackedInt in {
1955 // Move Instructions
1956 let neverHasSideEffects = 1 in
1957 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1958 "movdqa\t{$src, $dst|$dst, $src}", []>;
1959 let canFoldAsLoad = 1, mayLoad = 1 in
1960 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1961 "movdqa\t{$src, $dst|$dst, $src}",
1962 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1964 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1965 "movdqa\t{$src, $dst|$dst, $src}",
1966 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1967 let canFoldAsLoad = 1, mayLoad = 1 in
1968 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1969 "movdqu\t{$src, $dst|$dst, $src}",
1970 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1971 XS, Requires<[HasSSE2]>;
1973 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1974 "movdqu\t{$src, $dst|$dst, $src}",
1975 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1976 XS, Requires<[HasSSE2]>;
1978 // Intrinsic forms of MOVDQU load and store
1979 let canFoldAsLoad = 1 in
1980 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1981 "movdqu\t{$src, $dst|$dst, $src}",
1982 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1983 XS, Requires<[HasSSE2]>;
1984 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1985 "movdqu\t{$src, $dst|$dst, $src}",
1986 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1987 XS, Requires<[HasSSE2]>;
1989 let Constraints = "$src1 = $dst" in {
1991 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1992 bit Commutable = 0> {
1993 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1994 (ins VR128:$src1, VR128:$src2),
1995 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1996 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1997 let isCommutable = Commutable;
1999 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2000 (ins VR128:$src1, i128mem:$src2),
2001 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2002 [(set VR128:$dst, (IntId VR128:$src1,
2003 (bitconvert (memopv2i64
2007 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
2009 Intrinsic IntId, Intrinsic IntId2> {
2010 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2011 (ins VR128:$src1, VR128:$src2),
2012 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2013 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
2014 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2015 (ins VR128:$src1, i128mem:$src2),
2016 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2017 [(set VR128:$dst, (IntId VR128:$src1,
2018 (bitconvert (memopv2i64 addr:$src2))))]>;
2019 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
2020 (ins VR128:$src1, i32i8imm:$src2),
2021 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2022 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
2025 /// PDI_binop_rm - Simple SSE2 binary operator.
2026 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
2027 ValueType OpVT, bit Commutable = 0> {
2028 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2029 (ins VR128:$src1, VR128:$src2),
2030 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2031 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
2032 let isCommutable = Commutable;
2034 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2035 (ins VR128:$src1, i128mem:$src2),
2036 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2037 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
2038 (bitconvert (memopv2i64 addr:$src2)))))]>;
2041 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
2043 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
2044 /// to collapse (bitconvert VT to VT) into its operand.
2046 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
2047 bit Commutable = 0> {
2048 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2049 (ins VR128:$src1, VR128:$src2),
2050 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2051 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
2052 let isCommutable = Commutable;
2054 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2055 (ins VR128:$src1, i128mem:$src2),
2056 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2057 [(set VR128:$dst, (OpNode VR128:$src1,
2058 (memopv2i64 addr:$src2)))]>;
2061 } // Constraints = "$src1 = $dst"
2062 } // ExeDomain = SSEPackedInt
2064 // 128-bit Integer Arithmetic
2066 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
2067 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
2068 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
2069 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
2071 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
2072 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
2073 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
2074 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
2076 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
2077 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
2078 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
2079 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
2081 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
2082 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
2083 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
2084 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
2086 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
2088 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
2089 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
2090 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
2092 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
2094 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
2095 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
2098 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
2099 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
2100 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
2101 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
2102 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
2105 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
2106 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
2107 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
2108 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
2109 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
2110 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
2112 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2113 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2114 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2115 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2116 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2117 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2119 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2120 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2121 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2122 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2124 // 128-bit logical shifts.
2125 let Constraints = "$src1 = $dst", neverHasSideEffects = 1,
2126 ExeDomain = SSEPackedInt in {
2127 def PSLLDQri : PDIi8<0x73, MRM7r,
2128 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2129 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2130 def PSRLDQri : PDIi8<0x73, MRM3r,
2131 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2132 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2133 // PSRADQri doesn't exist in SSE[1-3].
2136 let Predicates = [HasSSE2] in {
2137 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2138 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2139 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2140 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2141 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2142 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2143 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2144 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2145 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2146 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2148 // Shift up / down and insert zero's.
2149 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2150 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2151 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2152 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2156 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2157 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2158 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2160 let Constraints = "$src1 = $dst", ExeDomain = SSEPackedInt in {
2161 def PANDNrr : PDI<0xDF, MRMSrcReg,
2162 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2163 "pandn\t{$src2, $dst|$dst, $src2}",
2164 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2167 def PANDNrm : PDI<0xDF, MRMSrcMem,
2168 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2169 "pandn\t{$src2, $dst|$dst, $src2}",
2170 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2171 (memopv2i64 addr:$src2))))]>;
2174 // SSE2 Integer comparison
2175 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2176 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2177 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2178 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2179 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2180 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2182 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2183 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2184 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2185 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2186 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2187 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2188 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2189 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2190 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2191 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2192 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2193 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2195 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2196 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2197 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2198 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2199 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2200 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2201 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2202 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2203 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2204 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2205 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2206 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2209 // Pack instructions
2210 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2211 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2212 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2214 let ExeDomain = SSEPackedInt in {
2216 // Shuffle and unpack instructions
2217 let AddedComplexity = 5 in {
2218 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2219 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2220 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2221 [(set VR128:$dst, (v4i32 (pshufd:$src2
2222 VR128:$src1, (undef))))]>;
2223 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2224 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2225 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2226 [(set VR128:$dst, (v4i32 (pshufd:$src2
2227 (bc_v4i32 (memopv2i64 addr:$src1)),
2231 // SSE2 with ImmT == Imm8 and XS prefix.
2232 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2233 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2234 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2235 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2237 XS, Requires<[HasSSE2]>;
2238 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2239 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2240 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2241 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2242 (bc_v8i16 (memopv2i64 addr:$src1)),
2244 XS, Requires<[HasSSE2]>;
2246 // SSE2 with ImmT == Imm8 and XD prefix.
2247 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2248 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2249 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2250 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2252 XD, Requires<[HasSSE2]>;
2253 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2254 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2255 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2256 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2257 (bc_v8i16 (memopv2i64 addr:$src1)),
2259 XD, Requires<[HasSSE2]>;
2262 let Constraints = "$src1 = $dst" in {
2263 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2264 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2265 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2267 (v16i8 (unpckl VR128:$src1, VR128:$src2)))]>;
2268 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2269 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2270 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2272 (unpckl VR128:$src1,
2273 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2274 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2275 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2276 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2278 (v8i16 (unpckl VR128:$src1, VR128:$src2)))]>;
2279 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2280 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2281 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2283 (unpckl VR128:$src1,
2284 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2285 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2286 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2287 "punpckldq\t{$src2, $dst|$dst, $src2}",
2289 (v4i32 (unpckl VR128:$src1, VR128:$src2)))]>;
2290 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2291 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2292 "punpckldq\t{$src2, $dst|$dst, $src2}",
2294 (unpckl VR128:$src1,
2295 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2296 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2297 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2298 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2300 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2301 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2302 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2303 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2305 (v2i64 (unpckl VR128:$src1,
2306 (memopv2i64 addr:$src2))))]>;
2308 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2309 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2310 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2312 (v16i8 (unpckh VR128:$src1, VR128:$src2)))]>;
2313 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2314 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2315 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2317 (unpckh VR128:$src1,
2318 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2319 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2320 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2321 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2323 (v8i16 (unpckh VR128:$src1, VR128:$src2)))]>;
2324 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2325 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2326 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2328 (unpckh VR128:$src1,
2329 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2330 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2331 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2332 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2334 (v4i32 (unpckh VR128:$src1, VR128:$src2)))]>;
2335 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2336 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2337 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2339 (unpckh VR128:$src1,
2340 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2341 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2342 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2343 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2345 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2346 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2347 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2348 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2350 (v2i64 (unpckh VR128:$src1,
2351 (memopv2i64 addr:$src2))))]>;
2355 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2356 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2357 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2358 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2360 let Constraints = "$src1 = $dst" in {
2361 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2362 (outs VR128:$dst), (ins VR128:$src1,
2363 GR32:$src2, i32i8imm:$src3),
2364 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2366 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2367 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2368 (outs VR128:$dst), (ins VR128:$src1,
2369 i16mem:$src2, i32i8imm:$src3),
2370 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2372 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2377 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2378 "pmovmskb\t{$src, $dst|$dst, $src}",
2379 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2381 // Conditional store
2383 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2384 "maskmovdqu\t{$mask, $src|$src, $mask}",
2385 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2388 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2389 "maskmovdqu\t{$mask, $src|$src, $mask}",
2390 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2392 } // ExeDomain = SSEPackedInt
2394 // Non-temporal stores
2395 def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2396 "movntpd\t{$src, $dst|$dst, $src}",
2397 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2398 let ExeDomain = SSEPackedInt in
2399 def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2400 "movntdq\t{$src, $dst|$dst, $src}",
2401 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2402 def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2403 "movnti\t{$src, $dst|$dst, $src}",
2404 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2405 TB, Requires<[HasSSE2]>;
2407 let AddedComplexity = 400 in { // Prefer non-temporal versions
2408 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2409 "movntpd\t{$src, $dst|$dst, $src}",
2410 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
2412 let ExeDomain = SSEPackedInt in
2413 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2414 "movntdq\t{$src, $dst|$dst, $src}",
2415 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2419 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2420 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2421 TB, Requires<[HasSSE2]>;
2423 // Load, store, and memory fence
2424 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
2425 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2426 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
2427 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2429 // Pause. This "instruction" is encoded as "rep; nop", so even though it
2430 // was introduced with SSE2, it's backward compatible.
2431 def PAUSE : I<0x90, RawFrm, (outs), (ins), "pause", []>, REP;
2433 //TODO: custom lower this so as to never even generate the noop
2434 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2436 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2437 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2438 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2441 // Alias instructions that map zero vector to pxor / xorp* for sse.
2442 // We set canFoldAsLoad because this can be converted to a constant-pool
2443 // load of an all-ones value if folding it would be beneficial.
2444 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2445 isCodeGenOnly = 1, ExeDomain = SSEPackedInt in
2446 // FIXME: Change encoding to pseudo.
2447 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
2448 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2450 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2451 "movd\t{$src, $dst|$dst, $src}",
2453 (v4i32 (scalar_to_vector GR32:$src)))]>;
2454 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2455 "movd\t{$src, $dst|$dst, $src}",
2457 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2459 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2460 "movd\t{$src, $dst|$dst, $src}",
2461 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2463 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2464 "movd\t{$src, $dst|$dst, $src}",
2465 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2467 // SSE2 instructions with XS prefix
2468 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2469 "movq\t{$src, $dst|$dst, $src}",
2471 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2472 Requires<[HasSSE2]>;
2473 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2474 "movq\t{$src, $dst|$dst, $src}",
2475 [(store (i64 (vector_extract (v2i64 VR128:$src),
2476 (iPTR 0))), addr:$dst)]>;
2478 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
2479 (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
2481 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2482 "movd\t{$src, $dst|$dst, $src}",
2483 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2485 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2486 "movd\t{$src, $dst|$dst, $src}",
2487 [(store (i32 (vector_extract (v4i32 VR128:$src),
2488 (iPTR 0))), addr:$dst)]>;
2490 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2491 "movd\t{$src, $dst|$dst, $src}",
2492 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2493 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2494 "movd\t{$src, $dst|$dst, $src}",
2495 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2497 // Store / copy lower 64-bits of a XMM register.
2498 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2499 "movq\t{$src, $dst|$dst, $src}",
2500 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2502 // movd / movq to XMM register zero-extends
2503 let AddedComplexity = 15 in {
2504 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2505 "movd\t{$src, $dst|$dst, $src}",
2506 [(set VR128:$dst, (v4i32 (X86vzmovl
2507 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2508 // This is X86-64 only.
2509 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2510 "mov{d|q}\t{$src, $dst|$dst, $src}",
2511 [(set VR128:$dst, (v2i64 (X86vzmovl
2512 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2515 let AddedComplexity = 20 in {
2516 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2517 "movd\t{$src, $dst|$dst, $src}",
2519 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2520 (loadi32 addr:$src))))))]>;
2522 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2523 (MOVZDI2PDIrm addr:$src)>;
2524 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2525 (MOVZDI2PDIrm addr:$src)>;
2526 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2527 (MOVZDI2PDIrm addr:$src)>;
2529 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2530 "movq\t{$src, $dst|$dst, $src}",
2532 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2533 (loadi64 addr:$src))))))]>, XS,
2534 Requires<[HasSSE2]>;
2536 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2537 (MOVZQI2PQIrm addr:$src)>;
2538 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2539 (MOVZQI2PQIrm addr:$src)>;
2540 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2543 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2544 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2545 let AddedComplexity = 15 in
2546 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2547 "movq\t{$src, $dst|$dst, $src}",
2548 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2549 XS, Requires<[HasSSE2]>;
2551 let AddedComplexity = 20 in {
2552 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2553 "movq\t{$src, $dst|$dst, $src}",
2554 [(set VR128:$dst, (v2i64 (X86vzmovl
2555 (loadv2i64 addr:$src))))]>,
2556 XS, Requires<[HasSSE2]>;
2558 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2559 (MOVZPQILo2PQIrm addr:$src)>;
2562 // Instructions for the disassembler
2563 // xr = XMM register
2566 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2567 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2569 //===---------------------------------------------------------------------===//
2570 // SSE3 Instructions
2571 //===---------------------------------------------------------------------===//
2573 // Move Instructions
2574 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2575 "movshdup\t{$src, $dst|$dst, $src}",
2576 [(set VR128:$dst, (v4f32 (movshdup
2577 VR128:$src, (undef))))]>;
2578 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2579 "movshdup\t{$src, $dst|$dst, $src}",
2580 [(set VR128:$dst, (movshdup
2581 (memopv4f32 addr:$src), (undef)))]>;
2583 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2584 "movsldup\t{$src, $dst|$dst, $src}",
2585 [(set VR128:$dst, (v4f32 (movsldup
2586 VR128:$src, (undef))))]>;
2587 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2588 "movsldup\t{$src, $dst|$dst, $src}",
2589 [(set VR128:$dst, (movsldup
2590 (memopv4f32 addr:$src), (undef)))]>;
2592 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2593 "movddup\t{$src, $dst|$dst, $src}",
2594 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2595 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2596 "movddup\t{$src, $dst|$dst, $src}",
2598 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2601 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2603 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2605 let AddedComplexity = 5 in {
2606 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2607 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2608 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2609 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2610 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2611 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2612 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2613 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2617 let Constraints = "$src1 = $dst" in {
2618 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2619 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2620 "addsubps\t{$src2, $dst|$dst, $src2}",
2621 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2623 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2624 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2625 "addsubps\t{$src2, $dst|$dst, $src2}",
2626 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2627 (memop addr:$src2)))]>;
2628 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2629 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2630 "addsubpd\t{$src2, $dst|$dst, $src2}",
2631 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2633 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2634 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2635 "addsubpd\t{$src2, $dst|$dst, $src2}",
2636 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2637 (memop addr:$src2)))]>;
2640 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2641 "lddqu\t{$src, $dst|$dst, $src}",
2642 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2645 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2646 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2647 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2648 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2649 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2650 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2651 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2652 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2653 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2654 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2655 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2656 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2657 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2658 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2659 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2660 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2662 let Constraints = "$src1 = $dst" in {
2663 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2664 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2665 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2666 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2667 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2668 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2669 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2670 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2673 // Thread synchronization
2674 def MONITOR : I<0x01, MRM_C8, (outs), (ins), "monitor",
2675 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2676 def MWAIT : I<0x01, MRM_C9, (outs), (ins), "mwait",
2677 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2679 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2680 let AddedComplexity = 15 in
2681 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2682 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2683 let AddedComplexity = 20 in
2684 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2685 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2687 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2688 let AddedComplexity = 15 in
2689 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2690 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2691 let AddedComplexity = 20 in
2692 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2693 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2695 //===---------------------------------------------------------------------===//
2696 // SSSE3 Instructions
2697 //===---------------------------------------------------------------------===//
2699 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2700 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2701 Intrinsic IntId64, Intrinsic IntId128> {
2702 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2703 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2704 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2706 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2707 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2709 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2711 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2713 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2714 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2717 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2719 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2722 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2725 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2726 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2727 Intrinsic IntId64, Intrinsic IntId128> {
2728 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2730 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2731 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2733 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2735 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2738 (bitconvert (memopv4i16 addr:$src))))]>;
2740 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2742 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2743 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2746 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2748 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2751 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2754 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2755 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2756 Intrinsic IntId64, Intrinsic IntId128> {
2757 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2759 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2760 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2762 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2764 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2767 (bitconvert (memopv2i32 addr:$src))))]>;
2769 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2771 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2772 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2775 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2777 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2780 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2783 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2784 int_x86_ssse3_pabs_b,
2785 int_x86_ssse3_pabs_b_128>;
2786 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2787 int_x86_ssse3_pabs_w,
2788 int_x86_ssse3_pabs_w_128>;
2789 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2790 int_x86_ssse3_pabs_d,
2791 int_x86_ssse3_pabs_d_128>;
2793 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2794 let Constraints = "$src1 = $dst" in {
2795 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2796 Intrinsic IntId64, Intrinsic IntId128,
2797 bit Commutable = 0> {
2798 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2799 (ins VR64:$src1, VR64:$src2),
2800 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2801 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2802 let isCommutable = Commutable;
2804 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2805 (ins VR64:$src1, i64mem:$src2),
2806 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2808 (IntId64 VR64:$src1,
2809 (bitconvert (memopv8i8 addr:$src2))))]>;
2811 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2812 (ins VR128:$src1, VR128:$src2),
2813 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2814 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2816 let isCommutable = Commutable;
2818 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2819 (ins VR128:$src1, i128mem:$src2),
2820 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2822 (IntId128 VR128:$src1,
2823 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2827 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2828 let Constraints = "$src1 = $dst" in {
2829 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2830 Intrinsic IntId64, Intrinsic IntId128,
2831 bit Commutable = 0> {
2832 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2833 (ins VR64:$src1, VR64:$src2),
2834 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2835 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2836 let isCommutable = Commutable;
2838 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2839 (ins VR64:$src1, i64mem:$src2),
2840 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2842 (IntId64 VR64:$src1,
2843 (bitconvert (memopv4i16 addr:$src2))))]>;
2845 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2846 (ins VR128:$src1, VR128:$src2),
2847 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2848 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2850 let isCommutable = Commutable;
2852 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2853 (ins VR128:$src1, i128mem:$src2),
2854 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2856 (IntId128 VR128:$src1,
2857 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2861 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2862 let Constraints = "$src1 = $dst" in {
2863 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2864 Intrinsic IntId64, Intrinsic IntId128,
2865 bit Commutable = 0> {
2866 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2867 (ins VR64:$src1, VR64:$src2),
2868 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2869 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2870 let isCommutable = Commutable;
2872 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2873 (ins VR64:$src1, i64mem:$src2),
2874 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2876 (IntId64 VR64:$src1,
2877 (bitconvert (memopv2i32 addr:$src2))))]>;
2879 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2880 (ins VR128:$src1, VR128:$src2),
2881 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2882 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2884 let isCommutable = Commutable;
2886 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2887 (ins VR128:$src1, i128mem:$src2),
2888 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2890 (IntId128 VR128:$src1,
2891 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2895 let ImmT = NoImm in { // None of these have i8 immediate fields.
2896 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2897 int_x86_ssse3_phadd_w,
2898 int_x86_ssse3_phadd_w_128>;
2899 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2900 int_x86_ssse3_phadd_d,
2901 int_x86_ssse3_phadd_d_128>;
2902 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2903 int_x86_ssse3_phadd_sw,
2904 int_x86_ssse3_phadd_sw_128>;
2905 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2906 int_x86_ssse3_phsub_w,
2907 int_x86_ssse3_phsub_w_128>;
2908 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2909 int_x86_ssse3_phsub_d,
2910 int_x86_ssse3_phsub_d_128>;
2911 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2912 int_x86_ssse3_phsub_sw,
2913 int_x86_ssse3_phsub_sw_128>;
2914 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2915 int_x86_ssse3_pmadd_ub_sw,
2916 int_x86_ssse3_pmadd_ub_sw_128>;
2917 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2918 int_x86_ssse3_pmul_hr_sw,
2919 int_x86_ssse3_pmul_hr_sw_128, 1>;
2921 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2922 int_x86_ssse3_pshuf_b,
2923 int_x86_ssse3_pshuf_b_128>;
2924 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2925 int_x86_ssse3_psign_b,
2926 int_x86_ssse3_psign_b_128>;
2927 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2928 int_x86_ssse3_psign_w,
2929 int_x86_ssse3_psign_w_128>;
2930 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2931 int_x86_ssse3_psign_d,
2932 int_x86_ssse3_psign_d_128>;
2935 // palignr patterns.
2936 let Constraints = "$src1 = $dst" in {
2937 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2938 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2939 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2941 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2942 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2943 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2946 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2947 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2948 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2950 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2951 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2952 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2956 let AddedComplexity = 5 in {
2958 def : Pat<(v1i64 (palign:$src3 VR64:$src1, VR64:$src2)),
2959 (PALIGNR64rr VR64:$src2, VR64:$src1,
2960 (SHUFFLE_get_palign_imm VR64:$src3))>,
2961 Requires<[HasSSSE3]>;
2962 def : Pat<(v2i32 (palign:$src3 VR64:$src1, VR64:$src2)),
2963 (PALIGNR64rr VR64:$src2, VR64:$src1,
2964 (SHUFFLE_get_palign_imm VR64:$src3))>,
2965 Requires<[HasSSSE3]>;
2966 def : Pat<(v2f32 (palign:$src3 VR64:$src1, VR64:$src2)),
2967 (PALIGNR64rr VR64:$src2, VR64:$src1,
2968 (SHUFFLE_get_palign_imm VR64:$src3))>,
2969 Requires<[HasSSSE3]>;
2970 def : Pat<(v4i16 (palign:$src3 VR64:$src1, VR64:$src2)),
2971 (PALIGNR64rr VR64:$src2, VR64:$src1,
2972 (SHUFFLE_get_palign_imm VR64:$src3))>,
2973 Requires<[HasSSSE3]>;
2974 def : Pat<(v8i8 (palign:$src3 VR64:$src1, VR64:$src2)),
2975 (PALIGNR64rr VR64:$src2, VR64:$src1,
2976 (SHUFFLE_get_palign_imm VR64:$src3))>,
2977 Requires<[HasSSSE3]>;
2979 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2980 (PALIGNR128rr VR128:$src2, VR128:$src1,
2981 (SHUFFLE_get_palign_imm VR128:$src3))>,
2982 Requires<[HasSSSE3]>;
2983 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2984 (PALIGNR128rr VR128:$src2, VR128:$src1,
2985 (SHUFFLE_get_palign_imm VR128:$src3))>,
2986 Requires<[HasSSSE3]>;
2987 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2988 (PALIGNR128rr VR128:$src2, VR128:$src1,
2989 (SHUFFLE_get_palign_imm VR128:$src3))>,
2990 Requires<[HasSSSE3]>;
2991 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2992 (PALIGNR128rr VR128:$src2, VR128:$src1,
2993 (SHUFFLE_get_palign_imm VR128:$src3))>,
2994 Requires<[HasSSSE3]>;
2997 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2998 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2999 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
3000 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
3002 //===---------------------------------------------------------------------===//
3003 // Non-Instruction Patterns
3004 //===---------------------------------------------------------------------===//
3006 // extload f32 -> f64. This matches load+fextend because we have a hack in
3007 // the isel (PreprocessForFPConvert) that can introduce loads after dag
3009 // Since these loads aren't folded into the fextend, we have to match it
3011 let Predicates = [HasSSE2] in
3012 def : Pat<(fextend (loadf32 addr:$src)),
3013 (CVTSS2SDrm addr:$src)>;
3016 let Predicates = [HasSSE2] in {
3017 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
3018 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
3019 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
3020 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
3021 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
3022 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
3023 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
3024 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
3025 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
3026 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
3027 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
3028 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
3029 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
3030 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
3031 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
3032 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
3033 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
3034 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
3035 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
3036 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
3037 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
3038 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
3039 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
3040 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
3041 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
3042 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
3043 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
3044 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
3045 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
3046 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
3049 // Move scalar to XMM zero-extended
3050 // movd to XMM register zero-extends
3051 let AddedComplexity = 15 in {
3052 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
3053 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
3054 (MOVSDrr (v2f64 (V_SET0PS)), FR64:$src)>;
3055 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
3056 (MOVSSrr (v4f32 (V_SET0PS)), FR32:$src)>;
3057 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
3058 (MOVSSrr (v4f32 (V_SET0PS)),
3059 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss)))>;
3060 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
3061 (MOVSSrr (v4i32 (V_SET0PI)),
3062 (EXTRACT_SUBREG (v4i32 VR128:$src), sub_ss))>;
3065 // Splat v2f64 / v2i64
3066 let AddedComplexity = 10 in {
3067 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
3068 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3069 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
3070 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3071 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
3072 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3073 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
3074 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3077 // Special unary SHUFPSrri case.
3078 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
3079 (SHUFPSrri VR128:$src1, VR128:$src1,
3080 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3081 let AddedComplexity = 5 in
3082 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
3083 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3084 Requires<[HasSSE2]>;
3085 // Special unary SHUFPDrri case.
3086 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
3087 (SHUFPDrri VR128:$src1, VR128:$src1,
3088 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3089 Requires<[HasSSE2]>;
3090 // Special unary SHUFPDrri case.
3091 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
3092 (SHUFPDrri VR128:$src1, VR128:$src1,
3093 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3094 Requires<[HasSSE2]>;
3095 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
3096 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
3097 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3098 Requires<[HasSSE2]>;
3100 // Special binary v4i32 shuffle cases with SHUFPS.
3101 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
3102 (SHUFPSrri VR128:$src1, VR128:$src2,
3103 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3104 Requires<[HasSSE2]>;
3105 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
3106 (SHUFPSrmi VR128:$src1, addr:$src2,
3107 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3108 Requires<[HasSSE2]>;
3109 // Special binary v2i64 shuffle cases using SHUFPDrri.
3110 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
3111 (SHUFPDrri VR128:$src1, VR128:$src2,
3112 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3113 Requires<[HasSSE2]>;
3115 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
3116 let AddedComplexity = 15 in {
3117 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
3118 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3119 Requires<[OptForSpeed, HasSSE2]>;
3120 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
3121 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3122 Requires<[OptForSpeed, HasSSE2]>;
3124 let AddedComplexity = 10 in {
3125 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
3126 (UNPCKLPSrr VR128:$src, VR128:$src)>;
3127 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
3128 (PUNPCKLBWrr VR128:$src, VR128:$src)>;
3129 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
3130 (PUNPCKLWDrr VR128:$src, VR128:$src)>;
3131 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
3132 (PUNPCKLDQrr VR128:$src, VR128:$src)>;
3135 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
3136 let AddedComplexity = 15 in {
3137 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
3138 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3139 Requires<[OptForSpeed, HasSSE2]>;
3140 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
3141 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3142 Requires<[OptForSpeed, HasSSE2]>;
3144 let AddedComplexity = 10 in {
3145 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
3146 (UNPCKHPSrr VR128:$src, VR128:$src)>;
3147 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
3148 (PUNPCKHBWrr VR128:$src, VR128:$src)>;
3149 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
3150 (PUNPCKHWDrr VR128:$src, VR128:$src)>;
3151 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
3152 (PUNPCKHDQrr VR128:$src, VR128:$src)>;
3155 let AddedComplexity = 20 in {
3156 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3157 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3158 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3160 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3161 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3162 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3164 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3165 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3166 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3167 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3168 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3171 let AddedComplexity = 20 in {
3172 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3173 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3174 (MOVLPSrm VR128:$src1, addr:$src2)>;
3175 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3176 (MOVLPDrm VR128:$src1, addr:$src2)>;
3177 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3178 (MOVLPSrm VR128:$src1, addr:$src2)>;
3179 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3180 (MOVLPDrm VR128:$src1, addr:$src2)>;
3183 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3184 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3185 (MOVLPSmr addr:$src1, VR128:$src2)>;
3186 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3187 (MOVLPDmr addr:$src1, VR128:$src2)>;
3188 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3190 (MOVLPSmr addr:$src1, VR128:$src2)>;
3191 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3192 (MOVLPDmr addr:$src1, VR128:$src2)>;
3194 let AddedComplexity = 15 in {
3195 // Setting the lowest element in the vector.
3196 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3197 (MOVSSrr (v4i32 VR128:$src1),
3198 (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_ss))>;
3199 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3200 (MOVSDrr (v2i64 VR128:$src1),
3201 (EXTRACT_SUBREG (v2i64 VR128:$src2), sub_sd))>;
3203 // vector_shuffle v1, v2 <4, 5, 2, 3> using movsd
3204 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3205 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3206 Requires<[HasSSE2]>;
3207 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3208 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3209 Requires<[HasSSE2]>;
3212 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3213 // fall back to this for SSE1)
3214 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3215 (SHUFPSrri VR128:$src2, VR128:$src1,
3216 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3218 // Set lowest element and zero upper elements.
3219 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3220 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3222 // Some special case pandn patterns.
3223 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3225 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3226 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3228 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3229 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3231 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3233 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3234 (memop addr:$src2))),
3235 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3236 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3237 (memop addr:$src2))),
3238 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3239 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3240 (memop addr:$src2))),
3241 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3243 // vector -> vector casts
3244 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3245 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3246 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3247 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3248 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3249 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3250 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3251 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3253 // Use movaps / movups for SSE integer load / store (one byte shorter).
3254 def : Pat<(alignedloadv4i32 addr:$src),
3255 (MOVAPSrm addr:$src)>;
3256 def : Pat<(loadv4i32 addr:$src),
3257 (MOVUPSrm addr:$src)>;
3258 def : Pat<(alignedloadv2i64 addr:$src),
3259 (MOVAPSrm addr:$src)>;
3260 def : Pat<(loadv2i64 addr:$src),
3261 (MOVUPSrm addr:$src)>;
3263 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3264 (MOVAPSmr addr:$dst, VR128:$src)>;
3265 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3266 (MOVAPSmr addr:$dst, VR128:$src)>;
3267 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3268 (MOVAPSmr addr:$dst, VR128:$src)>;
3269 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3270 (MOVAPSmr addr:$dst, VR128:$src)>;
3271 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3272 (MOVUPSmr addr:$dst, VR128:$src)>;
3273 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3274 (MOVUPSmr addr:$dst, VR128:$src)>;
3275 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3276 (MOVUPSmr addr:$dst, VR128:$src)>;
3277 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3278 (MOVUPSmr addr:$dst, VR128:$src)>;
3280 //===----------------------------------------------------------------------===//
3281 // SSE4.1 Instructions
3282 //===----------------------------------------------------------------------===//
3284 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3287 Intrinsic V2F64Int> {
3288 // Intrinsic operation, reg.
3289 // Vector intrinsic operation, reg
3290 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3291 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3292 !strconcat(OpcodeStr,
3293 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3294 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3297 // Vector intrinsic operation, mem
3298 def PSm_Int : Ii8<opcps, MRMSrcMem,
3299 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3300 !strconcat(OpcodeStr,
3301 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3303 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3305 Requires<[HasSSE41]>;
3307 // Vector intrinsic operation, reg
3308 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3309 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3310 !strconcat(OpcodeStr,
3311 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3312 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3315 // Vector intrinsic operation, mem
3316 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3317 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3318 !strconcat(OpcodeStr,
3319 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3321 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3325 let Constraints = "$src1 = $dst" in {
3326 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3330 // Intrinsic operation, reg.
3331 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3333 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3334 !strconcat(OpcodeStr,
3335 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3337 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3340 // Intrinsic operation, mem.
3341 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3343 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3344 !strconcat(OpcodeStr,
3345 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3347 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3350 // Intrinsic operation, reg.
3351 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3353 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3354 !strconcat(OpcodeStr,
3355 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3357 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3360 // Intrinsic operation, mem.
3361 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3363 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3364 !strconcat(OpcodeStr,
3365 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3367 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3372 // FP round - roundss, roundps, roundsd, roundpd
3373 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3374 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3375 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3376 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3378 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3379 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3380 Intrinsic IntId128> {
3381 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3383 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3384 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3385 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3387 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3390 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3393 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3394 int_x86_sse41_phminposuw>;
3396 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3397 let Constraints = "$src1 = $dst" in {
3398 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3399 Intrinsic IntId128, bit Commutable = 0> {
3400 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3401 (ins VR128:$src1, VR128:$src2),
3402 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3403 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3405 let isCommutable = Commutable;
3407 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3408 (ins VR128:$src1, i128mem:$src2),
3409 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3411 (IntId128 VR128:$src1,
3412 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3416 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3417 int_x86_sse41_pcmpeqq, 1>;
3418 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3419 int_x86_sse41_packusdw, 0>;
3420 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3421 int_x86_sse41_pminsb, 1>;
3422 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3423 int_x86_sse41_pminsd, 1>;
3424 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3425 int_x86_sse41_pminud, 1>;
3426 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3427 int_x86_sse41_pminuw, 1>;
3428 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3429 int_x86_sse41_pmaxsb, 1>;
3430 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3431 int_x86_sse41_pmaxsd, 1>;
3432 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3433 int_x86_sse41_pmaxud, 1>;
3434 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3435 int_x86_sse41_pmaxuw, 1>;
3437 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3439 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3440 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3441 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3442 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3444 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3445 let Constraints = "$src1 = $dst" in {
3446 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3447 SDNode OpNode, Intrinsic IntId128,
3448 bit Commutable = 0> {
3449 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3450 (ins VR128:$src1, VR128:$src2),
3451 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3452 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3453 VR128:$src2))]>, OpSize {
3454 let isCommutable = Commutable;
3456 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3457 (ins VR128:$src1, VR128:$src2),
3458 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3459 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3461 let isCommutable = Commutable;
3463 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3464 (ins VR128:$src1, i128mem:$src2),
3465 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3467 (OpVT (OpNode VR128:$src1, (memop addr:$src2))))]>, OpSize;
3468 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3469 (ins VR128:$src1, i128mem:$src2),
3470 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3472 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3477 /// SS48I_binop_rm - Simple SSE41 binary operator.
3478 let Constraints = "$src1 = $dst" in {
3479 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
3480 ValueType OpVT, bit Commutable = 0> {
3481 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3482 (ins VR128:$src1, VR128:$src2),
3483 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3484 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]>,
3486 let isCommutable = Commutable;
3488 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3489 (ins VR128:$src1, i128mem:$src2),
3490 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3491 [(set VR128:$dst, (OpNode VR128:$src1,
3492 (bc_v4i32 (memopv2i64 addr:$src2))))]>,
3497 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, 1>;
3499 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3500 let Constraints = "$src1 = $dst" in {
3501 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3502 Intrinsic IntId128, bit Commutable = 0> {
3503 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3504 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3505 !strconcat(OpcodeStr,
3506 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3508 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3510 let isCommutable = Commutable;
3512 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3513 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3514 !strconcat(OpcodeStr,
3515 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3517 (IntId128 VR128:$src1,
3518 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3523 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3524 int_x86_sse41_blendps, 0>;
3525 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3526 int_x86_sse41_blendpd, 0>;
3527 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3528 int_x86_sse41_pblendw, 0>;
3529 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3530 int_x86_sse41_dpps, 1>;
3531 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3532 int_x86_sse41_dppd, 1>;
3533 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3534 int_x86_sse41_mpsadbw, 0>;
3537 /// SS41I_ternary_int - SSE 4.1 ternary operator
3538 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3539 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3540 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3541 (ins VR128:$src1, VR128:$src2),
3542 !strconcat(OpcodeStr,
3543 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3544 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3547 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3548 (ins VR128:$src1, i128mem:$src2),
3549 !strconcat(OpcodeStr,
3550 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3553 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3557 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3558 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3559 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3562 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3563 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3564 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3565 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3567 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3568 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3570 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3574 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3575 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3576 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3577 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3578 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3579 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3581 // Common patterns involving scalar load.
3582 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3583 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3584 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3585 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3587 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3588 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3589 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3590 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3592 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3593 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3594 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3595 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3597 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3598 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3599 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3600 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3602 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3603 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3604 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3605 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3607 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3608 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3609 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3610 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3613 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3614 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3615 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3616 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3618 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3619 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3621 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3625 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3626 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3627 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3628 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3630 // Common patterns involving scalar load
3631 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3632 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3633 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3634 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3636 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3637 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3638 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3639 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3642 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3643 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3644 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3645 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3647 // Expecting a i16 load any extended to i32 value.
3648 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3649 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3650 [(set VR128:$dst, (IntId (bitconvert
3651 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3655 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3656 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3658 // Common patterns involving scalar load
3659 def : Pat<(int_x86_sse41_pmovsxbq
3660 (bitconvert (v4i32 (X86vzmovl
3661 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3662 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3664 def : Pat<(int_x86_sse41_pmovzxbq
3665 (bitconvert (v4i32 (X86vzmovl
3666 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3667 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3670 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3671 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3672 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3673 (ins VR128:$src1, i32i8imm:$src2),
3674 !strconcat(OpcodeStr,
3675 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3676 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3678 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3679 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3680 !strconcat(OpcodeStr,
3681 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3684 // There's an AssertZext in the way of writing the store pattern
3685 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3688 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3691 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3692 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3693 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3694 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3695 !strconcat(OpcodeStr,
3696 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3699 // There's an AssertZext in the way of writing the store pattern
3700 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3703 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3706 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3707 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3708 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3709 (ins VR128:$src1, i32i8imm:$src2),
3710 !strconcat(OpcodeStr,
3711 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3713 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3714 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3715 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3716 !strconcat(OpcodeStr,
3717 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3718 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3719 addr:$dst)]>, OpSize;
3722 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3725 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3727 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3728 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3729 (ins VR128:$src1, i32i8imm:$src2),
3730 !strconcat(OpcodeStr,
3731 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3733 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3735 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3736 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3737 !strconcat(OpcodeStr,
3738 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3739 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3740 addr:$dst)]>, OpSize;
3743 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3745 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3746 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3749 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3750 Requires<[HasSSE41]>;
3752 let Constraints = "$src1 = $dst" in {
3753 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3754 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3755 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3756 !strconcat(OpcodeStr,
3757 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3759 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3760 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3761 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3762 !strconcat(OpcodeStr,
3763 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3765 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3766 imm:$src3))]>, OpSize;
3770 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3772 let Constraints = "$src1 = $dst" in {
3773 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3774 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3775 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3776 !strconcat(OpcodeStr,
3777 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3779 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3781 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3782 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3783 !strconcat(OpcodeStr,
3784 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3786 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3787 imm:$src3)))]>, OpSize;
3791 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3793 // insertps has a few different modes, there's the first two here below which
3794 // are optimized inserts that won't zero arbitrary elements in the destination
3795 // vector. The next one matches the intrinsic and could zero arbitrary elements
3796 // in the target vector.
3797 let Constraints = "$src1 = $dst" in {
3798 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3799 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3800 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3801 !strconcat(OpcodeStr,
3802 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3804 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3806 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3807 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3808 !strconcat(OpcodeStr,
3809 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3811 (X86insrtps VR128:$src1,
3812 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3813 imm:$src3))]>, OpSize;
3817 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3819 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3820 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3822 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3823 // the intel intrinsic that corresponds to this.
3824 let Defs = [EFLAGS] in {
3825 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3826 "ptest \t{$src2, $src1|$src1, $src2}",
3827 [(set EFLAGS, (X86ptest VR128:$src1, VR128:$src2))]>,
3829 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3830 "ptest \t{$src2, $src1|$src1, $src2}",
3831 [(set EFLAGS, (X86ptest VR128:$src1, (load addr:$src2)))]>,
3835 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3836 "movntdqa\t{$src, $dst|$dst, $src}",
3837 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
3841 //===----------------------------------------------------------------------===//
3842 // SSE4.2 Instructions
3843 //===----------------------------------------------------------------------===//
3845 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3846 let Constraints = "$src1 = $dst" in {
3847 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3848 Intrinsic IntId128, bit Commutable = 0> {
3849 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3850 (ins VR128:$src1, VR128:$src2),
3851 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3852 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3854 let isCommutable = Commutable;
3856 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3857 (ins VR128:$src1, i128mem:$src2),
3858 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3860 (IntId128 VR128:$src1,
3861 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3865 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3867 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3868 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3869 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3870 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3872 // crc intrinsic instruction
3873 // This set of instructions are only rm, the only difference is the size
3875 let Constraints = "$src1 = $dst" in {
3876 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3877 (ins GR32:$src1, i8mem:$src2),
3878 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3880 (int_x86_sse42_crc32_8 GR32:$src1,
3881 (load addr:$src2)))]>;
3882 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3883 (ins GR32:$src1, GR8:$src2),
3884 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3886 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>;
3887 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3888 (ins GR32:$src1, i16mem:$src2),
3889 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3891 (int_x86_sse42_crc32_16 GR32:$src1,
3892 (load addr:$src2)))]>,
3894 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3895 (ins GR32:$src1, GR16:$src2),
3896 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3898 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3900 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3901 (ins GR32:$src1, i32mem:$src2),
3902 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3904 (int_x86_sse42_crc32_32 GR32:$src1,
3905 (load addr:$src2)))]>;
3906 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3907 (ins GR32:$src1, GR32:$src2),
3908 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3910 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>;
3911 def CRC64m8 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3912 (ins GR64:$src1, i8mem:$src2),
3913 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3915 (int_x86_sse42_crc64_8 GR64:$src1,
3916 (load addr:$src2)))]>,
3918 def CRC64r8 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3919 (ins GR64:$src1, GR8:$src2),
3920 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3922 (int_x86_sse42_crc64_8 GR64:$src1, GR8:$src2))]>,
3924 def CRC64m64 : SS42FI<0xF1, MRMSrcMem, (outs GR64:$dst),
3925 (ins GR64:$src1, i64mem:$src2),
3926 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3928 (int_x86_sse42_crc64_64 GR64:$src1,
3929 (load addr:$src2)))]>,
3931 def CRC64r64 : SS42FI<0xF1, MRMSrcReg, (outs GR64:$dst),
3932 (ins GR64:$src1, GR64:$src2),
3933 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3935 (int_x86_sse42_crc64_64 GR64:$src1, GR64:$src2))]>,
3939 // String/text processing instructions.
3940 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3941 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3942 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3943 "#PCMPISTRM128rr PSEUDO!",
3944 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3945 imm:$src3))]>, OpSize;
3946 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3947 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3948 "#PCMPISTRM128rm PSEUDO!",
3949 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3950 imm:$src3))]>, OpSize;
3953 let Defs = [XMM0, EFLAGS] in {
3954 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3955 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3956 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3957 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3958 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3959 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3962 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3963 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3964 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3965 "#PCMPESTRM128rr PSEUDO!",
3967 (int_x86_sse42_pcmpestrm128
3968 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3970 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3971 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3972 "#PCMPESTRM128rm PSEUDO!",
3973 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3974 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3978 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3979 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3980 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3981 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3982 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3983 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3984 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3987 let Defs = [ECX, EFLAGS] in {
3988 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3989 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3990 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3991 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3992 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3993 (implicit EFLAGS)]>, OpSize;
3994 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3995 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3996 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3997 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3998 (implicit EFLAGS)]>, OpSize;
4002 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
4003 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
4004 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
4005 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
4006 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
4007 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
4009 let Defs = [ECX, EFLAGS] in {
4010 let Uses = [EAX, EDX] in {
4011 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
4012 def rr : SS42AI<0x61, MRMSrcReg, (outs),
4013 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
4014 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
4015 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
4016 (implicit EFLAGS)]>, OpSize;
4017 def rm : SS42AI<0x61, MRMSrcMem, (outs),
4018 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
4019 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
4021 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
4022 (implicit EFLAGS)]>, OpSize;
4027 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
4028 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
4029 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
4030 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
4031 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
4032 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;
4034 //===----------------------------------------------------------------------===//
4035 // AES-NI Instructions
4036 //===----------------------------------------------------------------------===//
4038 let Constraints = "$src1 = $dst" in {
4039 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr,
4040 Intrinsic IntId128, bit Commutable = 0> {
4041 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
4042 (ins VR128:$src1, VR128:$src2),
4043 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4044 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
4046 let isCommutable = Commutable;
4048 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
4049 (ins VR128:$src1, i128mem:$src2),
4050 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4052 (IntId128 VR128:$src1,
4053 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
4057 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
4058 int_x86_aesni_aesenc>;
4059 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
4060 int_x86_aesni_aesenclast>;
4061 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
4062 int_x86_aesni_aesdec>;
4063 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
4064 int_x86_aesni_aesdeclast>;
4066 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, VR128:$src2)),
4067 (AESENCrr VR128:$src1, VR128:$src2)>;
4068 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, (memop addr:$src2))),
4069 (AESENCrm VR128:$src1, addr:$src2)>;
4070 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, VR128:$src2)),
4071 (AESENCLASTrr VR128:$src1, VR128:$src2)>;
4072 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, (memop addr:$src2))),
4073 (AESENCLASTrm VR128:$src1, addr:$src2)>;
4074 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, VR128:$src2)),
4075 (AESDECrr VR128:$src1, VR128:$src2)>;
4076 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, (memop addr:$src2))),
4077 (AESDECrm VR128:$src1, addr:$src2)>;
4078 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, VR128:$src2)),
4079 (AESDECLASTrr VR128:$src1, VR128:$src2)>;
4080 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, (memop addr:$src2))),
4081 (AESDECLASTrm VR128:$src1, addr:$src2)>;
4083 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
4085 "aesimc\t{$src1, $dst|$dst, $src1}",
4087 (int_x86_aesni_aesimc VR128:$src1))]>,
4090 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
4091 (ins i128mem:$src1),
4092 "aesimc\t{$src1, $dst|$dst, $src1}",
4094 (int_x86_aesni_aesimc (bitconvert (memopv2i64 addr:$src1))))]>,
4097 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
4098 (ins VR128:$src1, i8imm:$src2),
4099 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4101 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
4103 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
4104 (ins i128mem:$src1, i8imm:$src2),
4105 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4107 (int_x86_aesni_aeskeygenassist (bitconvert (memopv2i64 addr:$src1)),