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:$src)))]>;
452 // Match intrinisics which expect MM and XMM operand(s).
453 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$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, $src2}", []>;
645 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
647 /// In addition, we also have a special variant of the scalar form here to
648 /// represent the associated intrinsic operation. This form is unlike the
649 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
650 /// and leaves the top elements unmodified (therefore these cannot be commuted).
652 /// These three forms can each be reg+reg or reg+mem, so there are a total of
653 /// six "instructions".
655 let Constraints = "$src1 = $dst" in {
656 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
657 SDNode OpNode, Intrinsic F32Int,
658 bit Commutable = 0> {
659 // Scalar operation, reg+reg.
660 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$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 // Scalar operation, reg+mem.
667 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
668 (ins FR32:$src1, f32mem:$src2),
669 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
670 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
672 // Vector operation, reg+reg.
673 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
674 (ins VR128:$src1, VR128:$src2),
675 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
676 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
677 let isCommutable = Commutable;
680 // Vector operation, reg+mem.
681 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
682 (ins VR128:$src1, f128mem:$src2),
683 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
684 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
686 // Intrinsic operation, reg+reg.
687 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
688 (ins VR128:$src1, VR128:$src2),
689 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
690 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]>;
692 // Intrinsic operation, reg+mem.
693 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
694 (ins VR128:$src1, ssmem:$src2),
695 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
696 [(set VR128:$dst, (F32Int VR128:$src1,
697 sse_load_f32:$src2))]>;
701 // Arithmetic instructions
702 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
703 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
704 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
705 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
707 /// sse1_fp_binop_rm - Other SSE1 binops
709 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
710 /// instructions for a full-vector intrinsic form. Operations that map
711 /// onto C operators don't use this form since they just use the plain
712 /// vector form instead of having a separate vector intrinsic form.
714 /// This provides a total of eight "instructions".
716 let Constraints = "$src1 = $dst" in {
717 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
721 bit Commutable = 0> {
723 // Scalar operation, reg+reg.
724 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
725 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
726 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
727 let isCommutable = Commutable;
730 // Scalar operation, reg+mem.
731 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
732 (ins FR32:$src1, f32mem:$src2),
733 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
734 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
736 // Vector operation, reg+reg.
737 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
738 (ins VR128:$src1, VR128:$src2),
739 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
740 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
741 let isCommutable = Commutable;
744 // Vector operation, reg+mem.
745 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
746 (ins VR128:$src1, f128mem:$src2),
747 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
748 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
750 // Intrinsic operation, reg+reg.
751 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
752 (ins VR128:$src1, VR128:$src2),
753 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
754 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
755 let isCommutable = Commutable;
758 // Intrinsic operation, reg+mem.
759 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
760 (ins VR128:$src1, ssmem:$src2),
761 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
762 [(set VR128:$dst, (F32Int VR128:$src1,
763 sse_load_f32:$src2))]>;
765 // Vector intrinsic operation, reg+reg.
766 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
767 (ins VR128:$src1, VR128:$src2),
768 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
769 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
770 let isCommutable = Commutable;
773 // Vector intrinsic operation, reg+mem.
774 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
775 (ins VR128:$src1, f128mem:$src2),
776 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
777 [(set VR128:$dst, (V4F32Int VR128:$src1, (memopv4f32 addr:$src2)))]>;
781 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
782 int_x86_sse_max_ss, int_x86_sse_max_ps>;
783 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
784 int_x86_sse_min_ss, int_x86_sse_min_ps>;
786 //===----------------------------------------------------------------------===//
787 // SSE packed FP Instructions
790 let neverHasSideEffects = 1 in
791 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
792 "movaps\t{$src, $dst|$dst, $src}", []>;
793 let canFoldAsLoad = 1, isReMaterializable = 1 in
794 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
795 "movaps\t{$src, $dst|$dst, $src}",
796 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
798 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
799 "movaps\t{$src, $dst|$dst, $src}",
800 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
802 let neverHasSideEffects = 1 in
803 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
804 "movups\t{$src, $dst|$dst, $src}", []>;
805 let canFoldAsLoad = 1, isReMaterializable = 1 in
806 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
807 "movups\t{$src, $dst|$dst, $src}",
808 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
809 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
810 "movups\t{$src, $dst|$dst, $src}",
811 [(store (v4f32 VR128:$src), addr:$dst)]>;
813 // Intrinsic forms of MOVUPS load and store
814 let canFoldAsLoad = 1, isReMaterializable = 1 in
815 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
816 "movups\t{$src, $dst|$dst, $src}",
817 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
818 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
819 "movups\t{$src, $dst|$dst, $src}",
820 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
822 let Constraints = "$src1 = $dst" in {
823 let AddedComplexity = 20 in {
824 def MOVLPSrm : PSI<0x12, MRMSrcMem,
825 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
826 "movlps\t{$src2, $dst|$dst, $src2}",
829 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
830 def MOVHPSrm : PSI<0x16, MRMSrcMem,
831 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
832 "movhps\t{$src2, $dst|$dst, $src2}",
834 (movlhps VR128:$src1,
835 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
837 } // Constraints = "$src1 = $dst"
840 def : Pat<(movlhps VR128:$src1, (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
841 (MOVHPSrm (v4i32 VR128:$src1), addr:$src2)>;
843 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
844 "movlps\t{$src, $dst|$dst, $src}",
845 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
846 (iPTR 0))), addr:$dst)]>;
848 // v2f64 extract element 1 is always custom lowered to unpack high to low
849 // and extract element 0 so the non-store version isn't too horrible.
850 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
851 "movhps\t{$src, $dst|$dst, $src}",
852 [(store (f64 (vector_extract
853 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
854 (undef)), (iPTR 0))), addr:$dst)]>;
856 let Constraints = "$src1 = $dst" in {
857 let AddedComplexity = 20 in {
858 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
859 (ins VR128:$src1, VR128:$src2),
860 "movlhps\t{$src2, $dst|$dst, $src2}",
862 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
864 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
865 (ins VR128:$src1, VR128:$src2),
866 "movhlps\t{$src2, $dst|$dst, $src2}",
868 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
870 } // Constraints = "$src1 = $dst"
872 let AddedComplexity = 20 in {
873 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
874 (MOVLHPSrr (v4f32 VR128:$src), (v4f32 VR128:$src))>;
875 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
876 (MOVLHPSrr (v2i64 VR128:$src), (v2i64 VR128:$src))>;
883 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
885 /// In addition, we also have a special variant of the scalar form here to
886 /// represent the associated intrinsic operation. This form is unlike the
887 /// plain scalar form, in that it takes an entire vector (instead of a
888 /// scalar) and leaves the top elements undefined.
890 /// And, we have a special variant form for a full-vector intrinsic form.
892 /// These four forms can each have a reg or a mem operand, so there are a
893 /// total of eight "instructions".
895 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
899 bit Commutable = 0> {
900 // Scalar operation, reg.
901 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
902 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
903 [(set FR32:$dst, (OpNode FR32:$src))]> {
904 let isCommutable = Commutable;
907 // Scalar operation, mem.
908 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
909 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
910 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
911 Requires<[HasSSE1, OptForSize]>;
913 // Vector operation, reg.
914 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
915 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
916 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
917 let isCommutable = Commutable;
920 // Vector operation, mem.
921 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
922 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
923 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
925 // Intrinsic operation, reg.
926 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
927 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
928 [(set VR128:$dst, (F32Int VR128:$src))]> {
929 let isCommutable = Commutable;
932 // Intrinsic operation, mem.
933 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
934 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
935 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
937 // Vector intrinsic operation, reg
938 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
939 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
940 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
941 let isCommutable = Commutable;
944 // Vector intrinsic operation, mem
945 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
946 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
947 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
951 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
952 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
954 // Reciprocal approximations. Note that these typically require refinement
955 // in order to obtain suitable precision.
956 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
957 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
958 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
959 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
962 let Constraints = "$src1 = $dst" in {
963 let isCommutable = 1 in {
964 def ANDPSrr : PSI<0x54, MRMSrcReg,
965 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
966 "andps\t{$src2, $dst|$dst, $src2}",
967 [(set VR128:$dst, (v2i64
968 (and VR128:$src1, VR128:$src2)))]>;
969 def ORPSrr : PSI<0x56, MRMSrcReg,
970 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
971 "orps\t{$src2, $dst|$dst, $src2}",
972 [(set VR128:$dst, (v2i64
973 (or VR128:$src1, VR128:$src2)))]>;
974 def XORPSrr : PSI<0x57, MRMSrcReg,
975 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
976 "xorps\t{$src2, $dst|$dst, $src2}",
977 [(set VR128:$dst, (v2i64
978 (xor VR128:$src1, VR128:$src2)))]>;
981 def ANDPSrm : PSI<0x54, MRMSrcMem,
982 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
983 "andps\t{$src2, $dst|$dst, $src2}",
984 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
985 (memopv2i64 addr:$src2)))]>;
986 def ORPSrm : PSI<0x56, MRMSrcMem,
987 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
988 "orps\t{$src2, $dst|$dst, $src2}",
989 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
990 (memopv2i64 addr:$src2)))]>;
991 def XORPSrm : PSI<0x57, MRMSrcMem,
992 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
993 "xorps\t{$src2, $dst|$dst, $src2}",
994 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
995 (memopv2i64 addr:$src2)))]>;
996 def ANDNPSrr : PSI<0x55, MRMSrcReg,
997 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
998 "andnps\t{$src2, $dst|$dst, $src2}",
1000 (v2i64 (and (xor VR128:$src1,
1001 (bc_v2i64 (v4i32 immAllOnesV))),
1003 def ANDNPSrm : PSI<0x55, MRMSrcMem,
1004 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1005 "andnps\t{$src2, $dst|$dst, $src2}",
1007 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
1008 (bc_v2i64 (v4i32 immAllOnesV))),
1009 (memopv2i64 addr:$src2))))]>;
1012 let Constraints = "$src1 = $dst" in {
1013 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
1014 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1015 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1016 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1017 VR128:$src, imm:$cc))]>;
1018 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
1019 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1020 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1021 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1022 (memop addr:$src), imm:$cc))]>;
1024 // Accept explicit immediate argument form instead of comparison code.
1025 let isAsmParserOnly = 1 in {
1026 def CMPPSrri_alt : PSIi8<0xC2, MRMSrcReg,
1027 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1028 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1029 def CMPPSrmi_alt : PSIi8<0xC2, MRMSrcMem,
1030 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1031 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1034 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1035 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
1036 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1037 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
1039 // Shuffle and unpack instructions
1040 let Constraints = "$src1 = $dst" in {
1041 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
1042 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
1043 (outs VR128:$dst), (ins VR128:$src1,
1044 VR128:$src2, i8imm:$src3),
1045 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1047 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1048 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
1049 (outs VR128:$dst), (ins VR128:$src1,
1050 f128mem:$src2, i8imm:$src3),
1051 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1054 VR128:$src1, (memopv4f32 addr:$src2))))]>;
1056 let AddedComplexity = 10 in {
1057 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
1058 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1059 "unpckhps\t{$src2, $dst|$dst, $src2}",
1061 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
1062 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
1063 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1064 "unpckhps\t{$src2, $dst|$dst, $src2}",
1066 (v4f32 (unpckh VR128:$src1,
1067 (memopv4f32 addr:$src2))))]>;
1069 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
1070 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1071 "unpcklps\t{$src2, $dst|$dst, $src2}",
1073 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
1074 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
1075 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1076 "unpcklps\t{$src2, $dst|$dst, $src2}",
1078 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
1079 } // AddedComplexity
1080 } // Constraints = "$src1 = $dst"
1083 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1084 "movmskps\t{$src, $dst|$dst, $src}",
1085 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
1086 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1087 "movmskpd\t{$src, $dst|$dst, $src}",
1088 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1090 // Prefetch intrinsic.
1091 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1092 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1093 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1094 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1095 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1096 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1097 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1098 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1100 // Non-temporal stores
1101 def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1102 "movntps\t{$src, $dst|$dst, $src}",
1103 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1105 let AddedComplexity = 400 in { // Prefer non-temporal versions
1106 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1107 "movntps\t{$src, $dst|$dst, $src}",
1108 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
1110 def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1111 "movntdq\t{$src, $dst|$dst, $src}",
1112 [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
1114 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
1115 "movnti\t{$src, $dst|$dst, $src}",
1116 [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
1117 TB, Requires<[HasSSE2]>;
1119 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
1120 "movnti\t{$src, $dst|$dst, $src}",
1121 [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
1122 TB, Requires<[HasSSE2]>;
1125 // Load, store, and memory fence
1126 def SFENCE : I<0xAE, MRM_F8, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>,
1127 TB, Requires<[HasSSE1]>;
1130 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1131 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1132 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1133 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1135 // Alias instructions that map zero vector to pxor / xorp* for sse.
1136 // We set canFoldAsLoad because this can be converted to a constant-pool
1137 // load of an all-zeros value if folding it would be beneficial.
1138 // FIXME: Change encoding to pseudo!
1139 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1140 isCodeGenOnly = 1 in {
1141 def V_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1142 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
1143 def V_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1144 [(set VR128:$dst, (v2f64 immAllZerosV))]>;
1145 let ExeDomain = SSEPackedInt in
1146 def V_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
1147 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1150 def : Pat<(v2i64 immAllZerosV), (V_SET0PI)>;
1151 def : Pat<(v8i16 immAllZerosV), (V_SET0PI)>;
1152 def : Pat<(v16i8 immAllZerosV), (V_SET0PI)>;
1154 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1155 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
1157 //===---------------------------------------------------------------------===//
1158 // SSE2 Instructions
1159 //===---------------------------------------------------------------------===//
1161 // Move Instructions. Register-to-register movsd is not used for FR64
1162 // register copies because it's a partial register update; FsMOVAPDrr is
1163 // used instead. Register-to-register movsd is not modeled as an INSERT_SUBREG
1164 // because INSERT_SUBREG requires that the insert be implementable in terms of
1165 // a copy, and just mentioned, we don't use movsd for copies.
1166 let Constraints = "$src1 = $dst" in
1167 def MOVSDrr : SDI<0x10, MRMSrcReg,
1168 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
1169 "movsd\t{$src2, $dst|$dst, $src2}",
1170 [(set (v2f64 VR128:$dst),
1171 (movl VR128:$src1, (scalar_to_vector FR64:$src2)))]>;
1173 // Extract the low 64-bit value from one vector and insert it into another.
1174 let AddedComplexity = 15 in
1175 def : Pat<(v2f64 (movl VR128:$src1, VR128:$src2)),
1176 (MOVSDrr (v2f64 VR128:$src1),
1177 (EXTRACT_SUBREG (v2f64 VR128:$src2), sub_sd))>;
1179 // Implicitly promote a 64-bit scalar to a vector.
1180 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
1181 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), FR64:$src, sub_sd)>;
1183 // Loading from memory automatically zeroing upper bits.
1184 let canFoldAsLoad = 1, isReMaterializable = 1, AddedComplexity = 20 in
1185 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1186 "movsd\t{$src, $dst|$dst, $src}",
1187 [(set FR64:$dst, (loadf64 addr:$src))]>;
1189 // MOVSDrm zeros the high parts of the register; represent this
1190 // with SUBREG_TO_REG.
1191 let AddedComplexity = 20 in {
1192 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
1193 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1194 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
1195 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1196 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
1197 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1198 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
1199 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1200 def : Pat<(v2f64 (X86vzload addr:$src)),
1201 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1204 // Store scalar value to memory.
1205 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1206 "movsd\t{$src, $dst|$dst, $src}",
1207 [(store FR64:$src, addr:$dst)]>;
1209 // Extract and store.
1210 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
1213 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
1215 // Conversion instructions
1216 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1217 "cvttsd2si\t{$src, $dst|$dst, $src}",
1218 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1219 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1220 "cvttsd2si\t{$src, $dst|$dst, $src}",
1221 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1222 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1223 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1224 [(set FR32:$dst, (fround FR64:$src))]>;
1225 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1226 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1227 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
1228 Requires<[HasSSE2, OptForSize]>;
1229 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1230 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1231 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1232 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1233 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1234 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1236 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1237 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1238 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1239 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1240 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1241 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1242 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1243 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1244 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1245 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1246 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1247 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1248 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1249 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1250 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1251 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1252 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1253 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1254 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1255 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1257 // SSE2 instructions with XS prefix
1258 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1259 "cvtss2sd\t{$src, $dst|$dst, $src}",
1260 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1261 Requires<[HasSSE2]>;
1262 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1263 "cvtss2sd\t{$src, $dst|$dst, $src}",
1264 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1265 Requires<[HasSSE2, OptForSize]>;
1267 def : Pat<(extloadf32 addr:$src),
1268 (CVTSS2SDrr (MOVSSrm addr:$src))>,
1269 Requires<[HasSSE2, OptForSpeed]>;
1271 // Match intrinsics which expect XMM operand(s).
1272 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1273 "cvtsd2si\t{$src, $dst|$dst, $src}",
1274 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1275 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1276 "cvtsd2si\t{$src, $dst|$dst, $src}",
1277 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1278 (load addr:$src)))]>;
1280 // Match intrinisics which expect MM and XMM operand(s).
1281 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1282 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1283 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1284 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1285 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1286 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1287 (memop addr:$src)))]>;
1288 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1289 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1290 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1291 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1292 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1293 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1294 (memop addr:$src)))]>;
1295 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1296 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1297 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1298 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1299 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1300 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1301 (load addr:$src)))]>;
1303 // Aliases for intrinsics
1304 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1305 "cvttsd2si\t{$src, $dst|$dst, $src}",
1307 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1308 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1309 "cvttsd2si\t{$src, $dst|$dst, $src}",
1310 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1311 (load addr:$src)))]>;
1313 // Comparison instructions
1314 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1315 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1316 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1317 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1319 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1320 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1321 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1323 // Accept explicit immediate argument form instead of comparison code.
1324 let isAsmParserOnly = 1 in {
1325 def CMPSDrr_alt : SDIi8<0xC2, MRMSrcReg,
1326 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, i8imm:$src2),
1327 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1329 def CMPSDrm_alt : SDIi8<0xC2, MRMSrcMem,
1330 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, i8imm:$src2),
1331 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1335 let Defs = [EFLAGS] in {
1336 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1337 "ucomisd\t{$src2, $src1|$src1, $src2}",
1338 [(set EFLAGS, (X86cmp FR64:$src1, FR64:$src2))]>;
1339 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1340 "ucomisd\t{$src2, $src1|$src1, $src2}",
1341 [(set EFLAGS, (X86cmp FR64:$src1, (loadf64 addr:$src2)))]>;
1342 } // Defs = [EFLAGS]
1344 // Aliases to match intrinsics which expect XMM operand(s).
1345 let Constraints = "$src1 = $dst" in {
1346 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1348 (ins VR128:$src1, VR128:$src, SSECC:$cc),
1349 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1350 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1351 VR128:$src, imm:$cc))]>;
1352 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1354 (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1355 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1356 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1357 (load addr:$src), imm:$cc))]>;
1360 let Defs = [EFLAGS] in {
1361 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1362 "ucomisd\t{$src2, $src1|$src1, $src2}",
1363 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1365 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1366 "ucomisd\t{$src2, $src1|$src1, $src2}",
1367 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1368 (load addr:$src2)))]>;
1370 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1371 "comisd\t{$src2, $src1|$src1, $src2}",
1372 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1374 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1375 "comisd\t{$src2, $src1|$src1, $src2}",
1376 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1377 (load addr:$src2)))]>;
1378 } // Defs = [EFLAGS]
1380 // Aliases of packed SSE2 instructions for scalar use. These all have names
1381 // that start with 'Fs'.
1383 // Alias instructions that map fld0 to pxor for sse.
1384 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1385 canFoldAsLoad = 1 in
1386 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1387 [(set FR64:$dst, fpimm0)]>,
1388 Requires<[HasSSE2]>, TB, OpSize;
1390 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1392 let neverHasSideEffects = 1 in
1393 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1394 "movapd\t{$src, $dst|$dst, $src}", []>;
1396 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1398 let canFoldAsLoad = 1, isReMaterializable = 1 in
1399 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1400 "movapd\t{$src, $dst|$dst, $src}",
1401 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1403 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1404 let Constraints = "$src1 = $dst" in {
1405 let isCommutable = 1 in {
1406 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1407 (ins FR64:$src1, FR64:$src2),
1408 "andpd\t{$src2, $dst|$dst, $src2}",
1409 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1410 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1411 (ins FR64:$src1, FR64:$src2),
1412 "orpd\t{$src2, $dst|$dst, $src2}",
1413 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1414 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1415 (ins FR64:$src1, FR64:$src2),
1416 "xorpd\t{$src2, $dst|$dst, $src2}",
1417 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1420 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1421 (ins FR64:$src1, f128mem:$src2),
1422 "andpd\t{$src2, $dst|$dst, $src2}",
1423 [(set FR64:$dst, (X86fand FR64:$src1,
1424 (memopfsf64 addr:$src2)))]>;
1425 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1426 (ins FR64:$src1, f128mem:$src2),
1427 "orpd\t{$src2, $dst|$dst, $src2}",
1428 [(set FR64:$dst, (X86for FR64:$src1,
1429 (memopfsf64 addr:$src2)))]>;
1430 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1431 (ins FR64:$src1, f128mem:$src2),
1432 "xorpd\t{$src2, $dst|$dst, $src2}",
1433 [(set FR64:$dst, (X86fxor FR64:$src1,
1434 (memopfsf64 addr:$src2)))]>;
1436 let neverHasSideEffects = 1 in {
1437 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1438 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1439 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1441 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1442 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1443 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1447 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1449 /// In addition, we also have a special variant of the scalar form here to
1450 /// represent the associated intrinsic operation. This form is unlike the
1451 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1452 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1454 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1455 /// six "instructions".
1457 let Constraints = "$src1 = $dst" in {
1458 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1459 SDNode OpNode, Intrinsic F64Int,
1460 bit Commutable = 0> {
1461 // Scalar operation, reg+reg.
1462 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1463 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1464 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1465 let isCommutable = Commutable;
1468 // Scalar operation, reg+mem.
1469 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1470 (ins FR64:$src1, f64mem:$src2),
1471 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1472 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1474 // Vector operation, reg+reg.
1475 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1476 (ins VR128:$src1, VR128:$src2),
1477 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1478 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1479 let isCommutable = Commutable;
1482 // Vector operation, reg+mem.
1483 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1484 (ins VR128:$src1, f128mem:$src2),
1485 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1486 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1488 // Intrinsic operation, reg+reg.
1489 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1490 (ins VR128:$src1, VR128:$src2),
1491 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1492 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]>;
1494 // Intrinsic operation, reg+mem.
1495 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1496 (ins VR128:$src1, sdmem:$src2),
1497 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1498 [(set VR128:$dst, (F64Int VR128:$src1,
1499 sse_load_f64:$src2))]>;
1503 // Arithmetic instructions
1504 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1505 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1506 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1507 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1509 /// sse2_fp_binop_rm - Other SSE2 binops
1511 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1512 /// instructions for a full-vector intrinsic form. Operations that map
1513 /// onto C operators don't use this form since they just use the plain
1514 /// vector form instead of having a separate vector intrinsic form.
1516 /// This provides a total of eight "instructions".
1518 let Constraints = "$src1 = $dst" in {
1519 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1523 bit Commutable = 0> {
1525 // Scalar operation, reg+reg.
1526 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1527 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1528 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1529 let isCommutable = Commutable;
1532 // Scalar operation, reg+mem.
1533 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1534 (ins FR64:$src1, f64mem:$src2),
1535 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1536 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1538 // Vector operation, reg+reg.
1539 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1540 (ins VR128:$src1, VR128:$src2),
1541 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1542 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1543 let isCommutable = Commutable;
1546 // Vector operation, reg+mem.
1547 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1548 (ins VR128:$src1, f128mem:$src2),
1549 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1550 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1552 // Intrinsic operation, reg+reg.
1553 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1554 (ins VR128:$src1, VR128:$src2),
1555 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1556 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1557 let isCommutable = Commutable;
1560 // Intrinsic operation, reg+mem.
1561 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1562 (ins VR128:$src1, sdmem:$src2),
1563 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1564 [(set VR128:$dst, (F64Int VR128:$src1,
1565 sse_load_f64:$src2))]>;
1567 // Vector intrinsic operation, reg+reg.
1568 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1569 (ins VR128:$src1, VR128:$src2),
1570 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1571 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1572 let isCommutable = Commutable;
1575 // Vector intrinsic operation, reg+mem.
1576 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1577 (ins VR128:$src1, f128mem:$src2),
1578 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1579 [(set VR128:$dst, (V2F64Int VR128:$src1,
1580 (memopv2f64 addr:$src2)))]>;
1584 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1585 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1586 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1587 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1589 //===---------------------------------------------------------------------===//
1590 // SSE packed FP Instructions
1592 // Move Instructions
1593 let neverHasSideEffects = 1 in
1594 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1595 "movapd\t{$src, $dst|$dst, $src}", []>;
1596 let canFoldAsLoad = 1, isReMaterializable = 1 in
1597 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1598 "movapd\t{$src, $dst|$dst, $src}",
1599 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1601 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1602 "movapd\t{$src, $dst|$dst, $src}",
1603 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1605 let neverHasSideEffects = 1 in
1606 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1607 "movupd\t{$src, $dst|$dst, $src}", []>;
1608 let canFoldAsLoad = 1 in
1609 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1610 "movupd\t{$src, $dst|$dst, $src}",
1611 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1612 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1613 "movupd\t{$src, $dst|$dst, $src}",
1614 [(store (v2f64 VR128:$src), addr:$dst)]>;
1616 // Intrinsic forms of MOVUPD load and store
1617 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1618 "movupd\t{$src, $dst|$dst, $src}",
1619 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1620 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1621 "movupd\t{$src, $dst|$dst, $src}",
1622 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1624 let Constraints = "$src1 = $dst" in {
1625 let AddedComplexity = 20 in {
1626 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1627 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1628 "movlpd\t{$src2, $dst|$dst, $src2}",
1630 (v2f64 (movlp VR128:$src1,
1631 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1632 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1633 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1634 "movhpd\t{$src2, $dst|$dst, $src2}",
1636 (v2f64 (movlhps VR128:$src1,
1637 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1638 } // AddedComplexity
1639 } // Constraints = "$src1 = $dst"
1641 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1642 "movlpd\t{$src, $dst|$dst, $src}",
1643 [(store (f64 (vector_extract (v2f64 VR128:$src),
1644 (iPTR 0))), addr:$dst)]>;
1646 // v2f64 extract element 1 is always custom lowered to unpack high to low
1647 // and extract element 0 so the non-store version isn't too horrible.
1648 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1649 "movhpd\t{$src, $dst|$dst, $src}",
1650 [(store (f64 (vector_extract
1651 (v2f64 (unpckh VR128:$src, (undef))),
1652 (iPTR 0))), addr:$dst)]>;
1654 // SSE2 instructions without OpSize prefix
1655 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1656 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1657 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1658 TB, Requires<[HasSSE2]>;
1659 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1660 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1661 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1662 (bitconvert (memopv2i64 addr:$src))))]>,
1663 TB, Requires<[HasSSE2]>;
1665 // SSE2 instructions with XS prefix
1666 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1667 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1668 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1669 XS, Requires<[HasSSE2]>;
1670 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1671 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1672 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1673 (bitconvert (memopv2i64 addr:$src))))]>,
1674 XS, Requires<[HasSSE2]>;
1676 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1677 "cvtps2dq\t{$src, $dst|$dst, $src}",
1678 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1679 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1680 "cvtps2dq\t{$src, $dst|$dst, $src}",
1681 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1682 (memop addr:$src)))]>;
1683 // SSE2 packed instructions with XS prefix
1684 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1685 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1686 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1687 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1689 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1690 "cvttps2dq\t{$src, $dst|$dst, $src}",
1692 (int_x86_sse2_cvttps2dq VR128:$src))]>,
1693 XS, Requires<[HasSSE2]>;
1694 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1695 "cvttps2dq\t{$src, $dst|$dst, $src}",
1696 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1697 (memop addr:$src)))]>,
1698 XS, Requires<[HasSSE2]>;
1700 // SSE2 packed instructions with XD prefix
1701 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1702 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1703 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1704 XD, Requires<[HasSSE2]>;
1705 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1706 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1707 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1708 (memop addr:$src)))]>,
1709 XD, Requires<[HasSSE2]>;
1711 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1712 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1713 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1714 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1715 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1716 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1717 (memop addr:$src)))]>;
1719 // SSE2 instructions without OpSize prefix
1720 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1721 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1722 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1723 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1725 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1726 "cvtps2pd\t{$src, $dst|$dst, $src}",
1727 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1728 TB, Requires<[HasSSE2]>;
1729 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1730 "cvtps2pd\t{$src, $dst|$dst, $src}",
1731 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1732 (load addr:$src)))]>,
1733 TB, Requires<[HasSSE2]>;
1735 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1736 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1737 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1738 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1741 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1742 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1743 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1744 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1745 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1746 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1747 (memop addr:$src)))]>;
1749 // Match intrinsics which expect XMM operand(s).
1750 // Aliases for intrinsics
1751 let Constraints = "$src1 = $dst" in {
1752 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1753 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1754 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1755 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1757 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1758 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1759 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1760 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1761 (loadi32 addr:$src2)))]>;
1762 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1763 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1764 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1765 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1767 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1768 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1769 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1770 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1771 (load addr:$src2)))]>;
1772 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1773 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1774 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1775 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1776 VR128:$src2))]>, XS,
1777 Requires<[HasSSE2]>;
1778 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1779 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1780 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1781 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1782 (load addr:$src2)))]>, XS,
1783 Requires<[HasSSE2]>;
1788 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1790 /// In addition, we also have a special variant of the scalar form here to
1791 /// represent the associated intrinsic operation. This form is unlike the
1792 /// plain scalar form, in that it takes an entire vector (instead of a
1793 /// scalar) and leaves the top elements undefined.
1795 /// And, we have a special variant form for a full-vector intrinsic form.
1797 /// These four forms can each have a reg or a mem operand, so there are a
1798 /// total of eight "instructions".
1800 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1804 bit Commutable = 0> {
1805 // Scalar operation, reg.
1806 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1807 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1808 [(set FR64:$dst, (OpNode FR64:$src))]> {
1809 let isCommutable = Commutable;
1812 // Scalar operation, mem.
1813 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1814 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1815 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1817 // Vector operation, reg.
1818 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1819 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1820 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1821 let isCommutable = Commutable;
1824 // Vector operation, mem.
1825 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1826 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1827 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1829 // Intrinsic operation, reg.
1830 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1831 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1832 [(set VR128:$dst, (F64Int VR128:$src))]> {
1833 let isCommutable = Commutable;
1836 // Intrinsic operation, mem.
1837 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1838 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1839 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1841 // Vector intrinsic operation, reg
1842 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1843 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1844 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1845 let isCommutable = Commutable;
1848 // Vector intrinsic operation, mem
1849 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1850 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1851 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1855 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1856 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1858 // There is no f64 version of the reciprocal approximation instructions.
1861 let Constraints = "$src1 = $dst" in {
1862 let isCommutable = 1 in {
1863 def ANDPDrr : PDI<0x54, MRMSrcReg,
1864 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1865 "andpd\t{$src2, $dst|$dst, $src2}",
1867 (and (bc_v2i64 (v2f64 VR128:$src1)),
1868 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1869 def ORPDrr : PDI<0x56, MRMSrcReg,
1870 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1871 "orpd\t{$src2, $dst|$dst, $src2}",
1873 (or (bc_v2i64 (v2f64 VR128:$src1)),
1874 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1875 def XORPDrr : PDI<0x57, MRMSrcReg,
1876 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1877 "xorpd\t{$src2, $dst|$dst, $src2}",
1879 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1880 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1883 def ANDPDrm : PDI<0x54, MRMSrcMem,
1884 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1885 "andpd\t{$src2, $dst|$dst, $src2}",
1887 (and (bc_v2i64 (v2f64 VR128:$src1)),
1888 (memopv2i64 addr:$src2)))]>;
1889 def ORPDrm : PDI<0x56, MRMSrcMem,
1890 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1891 "orpd\t{$src2, $dst|$dst, $src2}",
1893 (or (bc_v2i64 (v2f64 VR128:$src1)),
1894 (memopv2i64 addr:$src2)))]>;
1895 def XORPDrm : PDI<0x57, MRMSrcMem,
1896 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1897 "xorpd\t{$src2, $dst|$dst, $src2}",
1899 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1900 (memopv2i64 addr:$src2)))]>;
1901 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1902 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1903 "andnpd\t{$src2, $dst|$dst, $src2}",
1905 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1906 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1907 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1908 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1909 "andnpd\t{$src2, $dst|$dst, $src2}",
1911 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1912 (memopv2i64 addr:$src2)))]>;
1915 let Constraints = "$src1 = $dst" in {
1916 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1917 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1918 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1919 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1920 VR128:$src, imm:$cc))]>;
1921 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1922 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1923 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1924 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1925 (memop addr:$src), imm:$cc))]>;
1927 // Accept explicit immediate argument form instead of comparison code.
1928 let isAsmParserOnly = 1 in {
1929 def CMPPDrri_alt : PDIi8<0xC2, MRMSrcReg,
1930 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1931 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1932 def CMPPDrmi_alt : PDIi8<0xC2, MRMSrcMem,
1933 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1934 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1937 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1938 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1939 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1940 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1942 // Shuffle and unpack instructions
1943 let Constraints = "$src1 = $dst" in {
1944 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1945 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1946 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1948 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1949 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1950 (outs VR128:$dst), (ins VR128:$src1,
1951 f128mem:$src2, i8imm:$src3),
1952 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1955 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1957 let AddedComplexity = 10 in {
1958 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1959 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1960 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1962 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1963 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1964 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1965 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1967 (v2f64 (unpckh VR128:$src1,
1968 (memopv2f64 addr:$src2))))]>;
1970 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1971 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1972 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1974 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1975 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1976 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1977 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1979 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1980 } // AddedComplexity
1981 } // Constraints = "$src1 = $dst"
1984 //===---------------------------------------------------------------------===//
1985 // SSE integer instructions
1986 let ExeDomain = SSEPackedInt in {
1988 // Move Instructions
1989 let neverHasSideEffects = 1 in
1990 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1991 "movdqa\t{$src, $dst|$dst, $src}", []>;
1992 let canFoldAsLoad = 1, mayLoad = 1 in
1993 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1994 "movdqa\t{$src, $dst|$dst, $src}",
1995 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1997 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1998 "movdqa\t{$src, $dst|$dst, $src}",
1999 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
2000 let canFoldAsLoad = 1, mayLoad = 1 in
2001 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2002 "movdqu\t{$src, $dst|$dst, $src}",
2003 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
2004 XS, Requires<[HasSSE2]>;
2006 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2007 "movdqu\t{$src, $dst|$dst, $src}",
2008 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
2009 XS, Requires<[HasSSE2]>;
2011 // Intrinsic forms of MOVDQU load and store
2012 let canFoldAsLoad = 1 in
2013 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2014 "movdqu\t{$src, $dst|$dst, $src}",
2015 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
2016 XS, Requires<[HasSSE2]>;
2017 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2018 "movdqu\t{$src, $dst|$dst, $src}",
2019 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
2020 XS, Requires<[HasSSE2]>;
2022 let Constraints = "$src1 = $dst" in {
2024 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
2025 bit Commutable = 0> {
2026 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2027 (ins VR128:$src1, VR128:$src2),
2028 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2029 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
2030 let isCommutable = Commutable;
2032 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2033 (ins VR128:$src1, i128mem:$src2),
2034 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2035 [(set VR128:$dst, (IntId VR128:$src1,
2036 (bitconvert (memopv2i64
2040 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
2042 Intrinsic IntId, Intrinsic IntId2> {
2043 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2044 (ins VR128:$src1, VR128:$src2),
2045 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2046 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
2047 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2048 (ins VR128:$src1, i128mem:$src2),
2049 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2050 [(set VR128:$dst, (IntId VR128:$src1,
2051 (bitconvert (memopv2i64 addr:$src2))))]>;
2052 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
2053 (ins VR128:$src1, i32i8imm:$src2),
2054 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2055 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
2058 /// PDI_binop_rm - Simple SSE2 binary operator.
2059 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
2060 ValueType OpVT, bit Commutable = 0> {
2061 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2062 (ins VR128:$src1, VR128:$src2),
2063 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2064 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
2065 let isCommutable = Commutable;
2067 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2068 (ins VR128:$src1, i128mem:$src2),
2069 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2070 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
2071 (bitconvert (memopv2i64 addr:$src2)))))]>;
2074 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
2076 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
2077 /// to collapse (bitconvert VT to VT) into its operand.
2079 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
2080 bit Commutable = 0> {
2081 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2082 (ins VR128:$src1, VR128:$src2),
2083 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2084 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
2085 let isCommutable = Commutable;
2087 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2088 (ins VR128:$src1, i128mem:$src2),
2089 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2090 [(set VR128:$dst, (OpNode VR128:$src1,
2091 (memopv2i64 addr:$src2)))]>;
2094 } // Constraints = "$src1 = $dst"
2095 } // ExeDomain = SSEPackedInt
2097 // 128-bit Integer Arithmetic
2099 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
2100 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
2101 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
2102 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
2104 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
2105 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
2106 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
2107 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
2109 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
2110 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
2111 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
2112 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
2114 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
2115 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
2116 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
2117 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
2119 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
2121 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
2122 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
2123 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
2125 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
2127 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
2128 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
2131 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
2132 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
2133 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
2134 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
2135 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
2138 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
2139 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
2140 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
2141 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
2142 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
2143 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
2145 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2146 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2147 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2148 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2149 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2150 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2152 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2153 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2154 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2155 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2157 // 128-bit logical shifts.
2158 let Constraints = "$src1 = $dst", neverHasSideEffects = 1,
2159 ExeDomain = SSEPackedInt in {
2160 def PSLLDQri : PDIi8<0x73, MRM7r,
2161 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2162 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2163 def PSRLDQri : PDIi8<0x73, MRM3r,
2164 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2165 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2166 // PSRADQri doesn't exist in SSE[1-3].
2169 let Predicates = [HasSSE2] in {
2170 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2171 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2172 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2173 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2174 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2175 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2176 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2177 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2178 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2179 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2181 // Shift up / down and insert zero's.
2182 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2183 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2184 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2185 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2189 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2190 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2191 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2193 let Constraints = "$src1 = $dst", ExeDomain = SSEPackedInt in {
2194 def PANDNrr : PDI<0xDF, MRMSrcReg,
2195 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2196 "pandn\t{$src2, $dst|$dst, $src2}",
2197 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2200 def PANDNrm : PDI<0xDF, MRMSrcMem,
2201 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2202 "pandn\t{$src2, $dst|$dst, $src2}",
2203 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2204 (memopv2i64 addr:$src2))))]>;
2207 // SSE2 Integer comparison
2208 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2209 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2210 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2211 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2212 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2213 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2215 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2216 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2217 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2218 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2219 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2220 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2221 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2222 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2223 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2224 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2225 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2226 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2228 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2229 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2230 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2231 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2232 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2233 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2234 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2235 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2236 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2237 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2238 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2239 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2242 // Pack instructions
2243 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2244 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2245 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2247 let ExeDomain = SSEPackedInt in {
2249 // Shuffle and unpack instructions
2250 let AddedComplexity = 5 in {
2251 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2252 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2253 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2254 [(set VR128:$dst, (v4i32 (pshufd:$src2
2255 VR128:$src1, (undef))))]>;
2256 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2257 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2258 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2259 [(set VR128:$dst, (v4i32 (pshufd:$src2
2260 (bc_v4i32 (memopv2i64 addr:$src1)),
2264 // SSE2 with ImmT == Imm8 and XS prefix.
2265 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2266 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2267 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2268 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2270 XS, Requires<[HasSSE2]>;
2271 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2272 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2273 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2274 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2275 (bc_v8i16 (memopv2i64 addr:$src1)),
2277 XS, Requires<[HasSSE2]>;
2279 // SSE2 with ImmT == Imm8 and XD prefix.
2280 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2281 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2282 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2283 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2285 XD, Requires<[HasSSE2]>;
2286 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2287 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2288 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2289 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2290 (bc_v8i16 (memopv2i64 addr:$src1)),
2292 XD, Requires<[HasSSE2]>;
2295 let Constraints = "$src1 = $dst" in {
2296 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2297 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2298 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2300 (v16i8 (unpckl VR128:$src1, VR128:$src2)))]>;
2301 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2302 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2303 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2305 (unpckl VR128:$src1,
2306 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2307 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2308 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2309 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2311 (v8i16 (unpckl VR128:$src1, VR128:$src2)))]>;
2312 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2313 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2314 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2316 (unpckl VR128:$src1,
2317 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2318 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2319 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2320 "punpckldq\t{$src2, $dst|$dst, $src2}",
2322 (v4i32 (unpckl VR128:$src1, VR128:$src2)))]>;
2323 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2324 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2325 "punpckldq\t{$src2, $dst|$dst, $src2}",
2327 (unpckl VR128:$src1,
2328 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2329 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2330 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2331 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2333 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2334 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2335 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2336 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2338 (v2i64 (unpckl VR128:$src1,
2339 (memopv2i64 addr:$src2))))]>;
2341 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2342 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2343 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2345 (v16i8 (unpckh VR128:$src1, VR128:$src2)))]>;
2346 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2347 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2348 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2350 (unpckh VR128:$src1,
2351 (bc_v16i8 (memopv2i64 addr:$src2))))]>;
2352 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2353 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2354 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2356 (v8i16 (unpckh VR128:$src1, VR128:$src2)))]>;
2357 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2358 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2359 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2361 (unpckh VR128:$src1,
2362 (bc_v8i16 (memopv2i64 addr:$src2))))]>;
2363 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2364 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2365 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2367 (v4i32 (unpckh VR128:$src1, VR128:$src2)))]>;
2368 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2369 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2370 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2372 (unpckh VR128:$src1,
2373 (bc_v4i32 (memopv2i64 addr:$src2))))]>;
2374 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2375 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2376 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2378 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2379 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2380 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2381 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2383 (v2i64 (unpckh VR128:$src1,
2384 (memopv2i64 addr:$src2))))]>;
2388 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2389 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2390 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2391 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2393 let Constraints = "$src1 = $dst" in {
2394 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2395 (outs VR128:$dst), (ins VR128:$src1,
2396 GR32:$src2, i32i8imm:$src3),
2397 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2399 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2400 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2401 (outs VR128:$dst), (ins VR128:$src1,
2402 i16mem:$src2, i32i8imm:$src3),
2403 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2405 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2410 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2411 "pmovmskb\t{$src, $dst|$dst, $src}",
2412 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2414 // Conditional store
2416 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2417 "maskmovdqu\t{$mask, $src|$src, $mask}",
2418 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2421 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2422 "maskmovdqu\t{$mask, $src|$src, $mask}",
2423 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2425 } // ExeDomain = SSEPackedInt
2427 // Non-temporal stores
2428 def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2429 "movntpd\t{$src, $dst|$dst, $src}",
2430 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2431 let ExeDomain = SSEPackedInt in
2432 def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2433 "movntdq\t{$src, $dst|$dst, $src}",
2434 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2435 def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2436 "movnti\t{$src, $dst|$dst, $src}",
2437 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2438 TB, Requires<[HasSSE2]>;
2440 let AddedComplexity = 400 in { // Prefer non-temporal versions
2441 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2442 "movntpd\t{$src, $dst|$dst, $src}",
2443 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
2445 let ExeDomain = SSEPackedInt in
2446 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2447 "movntdq\t{$src, $dst|$dst, $src}",
2448 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2452 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2453 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2454 TB, Requires<[HasSSE2]>;
2456 // Load, store, and memory fence
2457 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
2458 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2459 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
2460 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2462 // Pause. This "instruction" is encoded as "rep; nop", so even though it
2463 // was introduced with SSE2, it's backward compabitle.
2464 def PAUSE : I<0x90, RawFrm, (outs), (ins), "pause", []>, REP;
2466 //TODO: custom lower this so as to never even generate the noop
2467 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2469 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2470 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2471 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2474 // Alias instructions that map zero vector to pxor / xorp* for sse.
2475 // We set canFoldAsLoad because this can be converted to a constant-pool
2476 // load of an all-ones value if folding it would be beneficial.
2477 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2478 isCodeGenOnly = 1, ExeDomain = SSEPackedInt in
2479 // FIXME: Change encoding to pseudo.
2480 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
2481 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2483 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2484 "movd\t{$src, $dst|$dst, $src}",
2486 (v4i32 (scalar_to_vector GR32:$src)))]>;
2487 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2488 "movd\t{$src, $dst|$dst, $src}",
2490 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2492 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2493 "movd\t{$src, $dst|$dst, $src}",
2494 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2496 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2497 "movd\t{$src, $dst|$dst, $src}",
2498 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2500 // SSE2 instructions with XS prefix
2501 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2502 "movq\t{$src, $dst|$dst, $src}",
2504 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2505 Requires<[HasSSE2]>;
2506 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2507 "movq\t{$src, $dst|$dst, $src}",
2508 [(store (i64 (vector_extract (v2i64 VR128:$src),
2509 (iPTR 0))), addr:$dst)]>;
2511 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
2512 (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
2514 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2515 "movd\t{$src, $dst|$dst, $src}",
2516 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2518 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2519 "movd\t{$src, $dst|$dst, $src}",
2520 [(store (i32 (vector_extract (v4i32 VR128:$src),
2521 (iPTR 0))), addr:$dst)]>;
2523 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2524 "movd\t{$src, $dst|$dst, $src}",
2525 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2526 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2527 "movd\t{$src, $dst|$dst, $src}",
2528 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2530 // Store / copy lower 64-bits of a XMM register.
2531 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2532 "movq\t{$src, $dst|$dst, $src}",
2533 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2535 // movd / movq to XMM register zero-extends
2536 let AddedComplexity = 15 in {
2537 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2538 "movd\t{$src, $dst|$dst, $src}",
2539 [(set VR128:$dst, (v4i32 (X86vzmovl
2540 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2541 // This is X86-64 only.
2542 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2543 "mov{d|q}\t{$src, $dst|$dst, $src}",
2544 [(set VR128:$dst, (v2i64 (X86vzmovl
2545 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2548 let AddedComplexity = 20 in {
2549 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2550 "movd\t{$src, $dst|$dst, $src}",
2552 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2553 (loadi32 addr:$src))))))]>;
2555 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2556 (MOVZDI2PDIrm addr:$src)>;
2557 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2558 (MOVZDI2PDIrm addr:$src)>;
2559 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2560 (MOVZDI2PDIrm addr:$src)>;
2562 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2563 "movq\t{$src, $dst|$dst, $src}",
2565 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2566 (loadi64 addr:$src))))))]>, XS,
2567 Requires<[HasSSE2]>;
2569 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2570 (MOVZQI2PQIrm addr:$src)>;
2571 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2572 (MOVZQI2PQIrm addr:$src)>;
2573 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2576 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2577 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2578 let AddedComplexity = 15 in
2579 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2580 "movq\t{$src, $dst|$dst, $src}",
2581 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2582 XS, Requires<[HasSSE2]>;
2584 let AddedComplexity = 20 in {
2585 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2586 "movq\t{$src, $dst|$dst, $src}",
2587 [(set VR128:$dst, (v2i64 (X86vzmovl
2588 (loadv2i64 addr:$src))))]>,
2589 XS, Requires<[HasSSE2]>;
2591 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2592 (MOVZPQILo2PQIrm addr:$src)>;
2595 // Instructions for the disassembler
2596 // xr = XMM register
2599 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2600 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2602 //===---------------------------------------------------------------------===//
2603 // SSE3 Instructions
2604 //===---------------------------------------------------------------------===//
2606 // Move Instructions
2607 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2608 "movshdup\t{$src, $dst|$dst, $src}",
2609 [(set VR128:$dst, (v4f32 (movshdup
2610 VR128:$src, (undef))))]>;
2611 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2612 "movshdup\t{$src, $dst|$dst, $src}",
2613 [(set VR128:$dst, (movshdup
2614 (memopv4f32 addr:$src), (undef)))]>;
2616 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2617 "movsldup\t{$src, $dst|$dst, $src}",
2618 [(set VR128:$dst, (v4f32 (movsldup
2619 VR128:$src, (undef))))]>;
2620 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2621 "movsldup\t{$src, $dst|$dst, $src}",
2622 [(set VR128:$dst, (movsldup
2623 (memopv4f32 addr:$src), (undef)))]>;
2625 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2626 "movddup\t{$src, $dst|$dst, $src}",
2627 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2628 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2629 "movddup\t{$src, $dst|$dst, $src}",
2631 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2634 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2636 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2638 let AddedComplexity = 5 in {
2639 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2640 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2641 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2642 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2643 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2644 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2645 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2646 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2650 let Constraints = "$src1 = $dst" in {
2651 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2652 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2653 "addsubps\t{$src2, $dst|$dst, $src2}",
2654 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2656 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2657 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2658 "addsubps\t{$src2, $dst|$dst, $src2}",
2659 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2660 (memop addr:$src2)))]>;
2661 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2662 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2663 "addsubpd\t{$src2, $dst|$dst, $src2}",
2664 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2666 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2667 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2668 "addsubpd\t{$src2, $dst|$dst, $src2}",
2669 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2670 (memop addr:$src2)))]>;
2673 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2674 "lddqu\t{$src, $dst|$dst, $src}",
2675 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2678 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2679 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2680 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2681 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2682 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2683 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2684 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2685 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2686 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2687 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2688 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2689 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2690 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2691 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2692 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2693 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2695 let Constraints = "$src1 = $dst" in {
2696 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2697 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2698 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2699 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2700 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2701 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2702 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2703 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2706 // Thread synchronization
2707 def MONITOR : I<0x01, MRM_C8, (outs), (ins), "monitor",
2708 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2709 def MWAIT : I<0x01, MRM_C9, (outs), (ins), "mwait",
2710 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2712 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2713 let AddedComplexity = 15 in
2714 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2715 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2716 let AddedComplexity = 20 in
2717 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2718 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2720 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2721 let AddedComplexity = 15 in
2722 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2723 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2724 let AddedComplexity = 20 in
2725 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2726 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2728 //===---------------------------------------------------------------------===//
2729 // SSSE3 Instructions
2730 //===---------------------------------------------------------------------===//
2732 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2733 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2734 Intrinsic IntId64, Intrinsic IntId128> {
2735 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2736 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2737 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2739 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2740 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2742 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2744 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2746 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2747 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2750 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2752 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2755 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2758 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2759 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2760 Intrinsic IntId64, Intrinsic IntId128> {
2761 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2763 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2764 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2766 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2768 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2771 (bitconvert (memopv4i16 addr:$src))))]>;
2773 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2775 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2776 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2779 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2781 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2784 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2787 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2788 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2789 Intrinsic IntId64, Intrinsic IntId128> {
2790 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2792 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2793 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2795 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2797 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2800 (bitconvert (memopv2i32 addr:$src))))]>;
2802 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2804 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2805 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2808 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2810 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2813 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2816 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2817 int_x86_ssse3_pabs_b,
2818 int_x86_ssse3_pabs_b_128>;
2819 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2820 int_x86_ssse3_pabs_w,
2821 int_x86_ssse3_pabs_w_128>;
2822 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2823 int_x86_ssse3_pabs_d,
2824 int_x86_ssse3_pabs_d_128>;
2826 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2827 let Constraints = "$src1 = $dst" in {
2828 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2829 Intrinsic IntId64, Intrinsic IntId128,
2830 bit Commutable = 0> {
2831 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2832 (ins VR64:$src1, VR64:$src2),
2833 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2834 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2835 let isCommutable = Commutable;
2837 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2838 (ins VR64:$src1, i64mem:$src2),
2839 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2841 (IntId64 VR64:$src1,
2842 (bitconvert (memopv8i8 addr:$src2))))]>;
2844 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2845 (ins VR128:$src1, VR128:$src2),
2846 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2847 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2849 let isCommutable = Commutable;
2851 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2852 (ins VR128:$src1, i128mem:$src2),
2853 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2855 (IntId128 VR128:$src1,
2856 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2860 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2861 let Constraints = "$src1 = $dst" in {
2862 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2863 Intrinsic IntId64, Intrinsic IntId128,
2864 bit Commutable = 0> {
2865 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2866 (ins VR64:$src1, VR64:$src2),
2867 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2868 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2869 let isCommutable = Commutable;
2871 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2872 (ins VR64:$src1, i64mem:$src2),
2873 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2875 (IntId64 VR64:$src1,
2876 (bitconvert (memopv4i16 addr:$src2))))]>;
2878 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2879 (ins VR128:$src1, VR128:$src2),
2880 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2881 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2883 let isCommutable = Commutable;
2885 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2886 (ins VR128:$src1, i128mem:$src2),
2887 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2889 (IntId128 VR128:$src1,
2890 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2894 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2895 let Constraints = "$src1 = $dst" in {
2896 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2897 Intrinsic IntId64, Intrinsic IntId128,
2898 bit Commutable = 0> {
2899 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2900 (ins VR64:$src1, VR64:$src2),
2901 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2902 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2903 let isCommutable = Commutable;
2905 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2906 (ins VR64:$src1, i64mem:$src2),
2907 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2909 (IntId64 VR64:$src1,
2910 (bitconvert (memopv2i32 addr:$src2))))]>;
2912 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2913 (ins VR128:$src1, VR128:$src2),
2914 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2915 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2917 let isCommutable = Commutable;
2919 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2920 (ins VR128:$src1, i128mem:$src2),
2921 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2923 (IntId128 VR128:$src1,
2924 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2928 let ImmT = NoImm in { // None of these have i8 immediate fields.
2929 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2930 int_x86_ssse3_phadd_w,
2931 int_x86_ssse3_phadd_w_128>;
2932 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2933 int_x86_ssse3_phadd_d,
2934 int_x86_ssse3_phadd_d_128>;
2935 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2936 int_x86_ssse3_phadd_sw,
2937 int_x86_ssse3_phadd_sw_128>;
2938 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2939 int_x86_ssse3_phsub_w,
2940 int_x86_ssse3_phsub_w_128>;
2941 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2942 int_x86_ssse3_phsub_d,
2943 int_x86_ssse3_phsub_d_128>;
2944 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2945 int_x86_ssse3_phsub_sw,
2946 int_x86_ssse3_phsub_sw_128>;
2947 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2948 int_x86_ssse3_pmadd_ub_sw,
2949 int_x86_ssse3_pmadd_ub_sw_128>;
2950 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2951 int_x86_ssse3_pmul_hr_sw,
2952 int_x86_ssse3_pmul_hr_sw_128, 1>;
2954 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2955 int_x86_ssse3_pshuf_b,
2956 int_x86_ssse3_pshuf_b_128>;
2957 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2958 int_x86_ssse3_psign_b,
2959 int_x86_ssse3_psign_b_128>;
2960 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2961 int_x86_ssse3_psign_w,
2962 int_x86_ssse3_psign_w_128>;
2963 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2964 int_x86_ssse3_psign_d,
2965 int_x86_ssse3_psign_d_128>;
2968 // palignr patterns.
2969 let Constraints = "$src1 = $dst" in {
2970 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2971 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2972 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2974 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2975 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2976 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2979 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2980 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2981 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2983 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2984 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2985 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2989 let AddedComplexity = 5 in {
2991 def : Pat<(v1i64 (palign:$src3 VR64:$src1, VR64:$src2)),
2992 (PALIGNR64rr VR64:$src2, VR64:$src1,
2993 (SHUFFLE_get_palign_imm VR64:$src3))>,
2994 Requires<[HasSSSE3]>;
2995 def : Pat<(v2i32 (palign:$src3 VR64:$src1, VR64:$src2)),
2996 (PALIGNR64rr VR64:$src2, VR64:$src1,
2997 (SHUFFLE_get_palign_imm VR64:$src3))>,
2998 Requires<[HasSSSE3]>;
2999 def : Pat<(v2f32 (palign:$src3 VR64:$src1, VR64:$src2)),
3000 (PALIGNR64rr VR64:$src2, VR64:$src1,
3001 (SHUFFLE_get_palign_imm VR64:$src3))>,
3002 Requires<[HasSSSE3]>;
3003 def : Pat<(v4i16 (palign:$src3 VR64:$src1, VR64:$src2)),
3004 (PALIGNR64rr VR64:$src2, VR64:$src1,
3005 (SHUFFLE_get_palign_imm VR64:$src3))>,
3006 Requires<[HasSSSE3]>;
3007 def : Pat<(v8i8 (palign:$src3 VR64:$src1, VR64:$src2)),
3008 (PALIGNR64rr VR64:$src2, VR64:$src1,
3009 (SHUFFLE_get_palign_imm VR64:$src3))>,
3010 Requires<[HasSSSE3]>;
3012 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
3013 (PALIGNR128rr VR128:$src2, VR128:$src1,
3014 (SHUFFLE_get_palign_imm VR128:$src3))>,
3015 Requires<[HasSSSE3]>;
3016 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
3017 (PALIGNR128rr VR128:$src2, VR128:$src1,
3018 (SHUFFLE_get_palign_imm VR128:$src3))>,
3019 Requires<[HasSSSE3]>;
3020 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
3021 (PALIGNR128rr VR128:$src2, VR128:$src1,
3022 (SHUFFLE_get_palign_imm VR128:$src3))>,
3023 Requires<[HasSSSE3]>;
3024 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
3025 (PALIGNR128rr VR128:$src2, VR128:$src1,
3026 (SHUFFLE_get_palign_imm VR128:$src3))>,
3027 Requires<[HasSSSE3]>;
3030 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
3031 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
3032 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
3033 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
3035 //===---------------------------------------------------------------------===//
3036 // Non-Instruction Patterns
3037 //===---------------------------------------------------------------------===//
3039 // extload f32 -> f64. This matches load+fextend because we have a hack in
3040 // the isel (PreprocessForFPConvert) that can introduce loads after dag
3042 // Since these loads aren't folded into the fextend, we have to match it
3044 let Predicates = [HasSSE2] in
3045 def : Pat<(fextend (loadf32 addr:$src)),
3046 (CVTSS2SDrm addr:$src)>;
3049 let Predicates = [HasSSE2] in {
3050 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
3051 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
3052 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
3053 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
3054 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
3055 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
3056 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
3057 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
3058 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
3059 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
3060 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
3061 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
3062 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
3063 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
3064 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
3065 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
3066 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
3067 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
3068 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
3069 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
3070 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
3071 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
3072 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
3073 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
3074 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
3075 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
3076 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
3077 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
3078 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
3079 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
3082 // Move scalar to XMM zero-extended
3083 // movd to XMM register zero-extends
3084 let AddedComplexity = 15 in {
3085 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
3086 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
3087 (MOVSDrr (v2f64 (V_SET0PS)), FR64:$src)>;
3088 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
3089 (MOVSSrr (v4f32 (V_SET0PS)), FR32:$src)>;
3090 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
3091 (MOVSSrr (v4f32 (V_SET0PS)),
3092 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss)))>;
3093 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
3094 (MOVSSrr (v4i32 (V_SET0PI)),
3095 (EXTRACT_SUBREG (v4i32 VR128:$src), sub_ss))>;
3098 // Splat v2f64 / v2i64
3099 let AddedComplexity = 10 in {
3100 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
3101 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3102 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
3103 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3104 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
3105 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3106 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
3107 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3110 // Special unary SHUFPSrri case.
3111 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
3112 (SHUFPSrri VR128:$src1, VR128:$src1,
3113 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3114 let AddedComplexity = 5 in
3115 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
3116 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3117 Requires<[HasSSE2]>;
3118 // Special unary SHUFPDrri case.
3119 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
3120 (SHUFPDrri VR128:$src1, VR128:$src1,
3121 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3122 Requires<[HasSSE2]>;
3123 // Special unary SHUFPDrri case.
3124 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
3125 (SHUFPDrri VR128:$src1, VR128:$src1,
3126 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3127 Requires<[HasSSE2]>;
3128 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
3129 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
3130 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3131 Requires<[HasSSE2]>;
3133 // Special binary v4i32 shuffle cases with SHUFPS.
3134 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
3135 (SHUFPSrri VR128:$src1, VR128:$src2,
3136 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3137 Requires<[HasSSE2]>;
3138 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
3139 (SHUFPSrmi VR128:$src1, addr:$src2,
3140 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3141 Requires<[HasSSE2]>;
3142 // Special binary v2i64 shuffle cases using SHUFPDrri.
3143 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
3144 (SHUFPDrri VR128:$src1, VR128:$src2,
3145 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3146 Requires<[HasSSE2]>;
3148 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
3149 let AddedComplexity = 15 in {
3150 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
3151 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3152 Requires<[OptForSpeed, HasSSE2]>;
3153 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
3154 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3155 Requires<[OptForSpeed, HasSSE2]>;
3157 let AddedComplexity = 10 in {
3158 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
3159 (UNPCKLPSrr VR128:$src, VR128:$src)>;
3160 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
3161 (PUNPCKLBWrr VR128:$src, VR128:$src)>;
3162 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
3163 (PUNPCKLWDrr VR128:$src, VR128:$src)>;
3164 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
3165 (PUNPCKLDQrr VR128:$src, VR128:$src)>;
3168 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
3169 let AddedComplexity = 15 in {
3170 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
3171 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3172 Requires<[OptForSpeed, HasSSE2]>;
3173 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
3174 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3175 Requires<[OptForSpeed, HasSSE2]>;
3177 let AddedComplexity = 10 in {
3178 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
3179 (UNPCKHPSrr VR128:$src, VR128:$src)>;
3180 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
3181 (PUNPCKHBWrr VR128:$src, VR128:$src)>;
3182 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
3183 (PUNPCKHWDrr VR128:$src, VR128:$src)>;
3184 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
3185 (PUNPCKHDQrr VR128:$src, VR128:$src)>;
3188 let AddedComplexity = 20 in {
3189 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3190 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3191 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3193 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3194 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3195 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3197 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3198 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3199 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3200 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3201 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3204 let AddedComplexity = 20 in {
3205 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3206 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3207 (MOVLPSrm VR128:$src1, addr:$src2)>;
3208 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3209 (MOVLPDrm VR128:$src1, addr:$src2)>;
3210 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3211 (MOVLPSrm VR128:$src1, addr:$src2)>;
3212 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3213 (MOVLPDrm VR128:$src1, addr:$src2)>;
3216 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3217 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3218 (MOVLPSmr addr:$src1, VR128:$src2)>;
3219 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3220 (MOVLPDmr addr:$src1, VR128:$src2)>;
3221 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3223 (MOVLPSmr addr:$src1, VR128:$src2)>;
3224 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3225 (MOVLPDmr addr:$src1, VR128:$src2)>;
3227 let AddedComplexity = 15 in {
3228 // Setting the lowest element in the vector.
3229 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3230 (MOVSSrr (v4i32 VR128:$src1),
3231 (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_ss))>;
3232 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3233 (MOVSDrr (v2i64 VR128:$src1),
3234 (EXTRACT_SUBREG (v2i64 VR128:$src2), sub_sd))>;
3236 // vector_shuffle v1, v2 <4, 5, 2, 3> using movsd
3237 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3238 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3239 Requires<[HasSSE2]>;
3240 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3241 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3242 Requires<[HasSSE2]>;
3245 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3246 // fall back to this for SSE1)
3247 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3248 (SHUFPSrri VR128:$src2, VR128:$src1,
3249 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3251 // Set lowest element and zero upper elements.
3252 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3253 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3255 // Some special case pandn patterns.
3256 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3258 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3259 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3261 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3262 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3264 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3266 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3267 (memop addr:$src2))),
3268 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3269 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3270 (memop addr:$src2))),
3271 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3272 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3273 (memop addr:$src2))),
3274 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3276 // vector -> vector casts
3277 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3278 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3279 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3280 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3281 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3282 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3283 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3284 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3286 // Use movaps / movups for SSE integer load / store (one byte shorter).
3287 def : Pat<(alignedloadv4i32 addr:$src),
3288 (MOVAPSrm addr:$src)>;
3289 def : Pat<(loadv4i32 addr:$src),
3290 (MOVUPSrm addr:$src)>;
3291 def : Pat<(alignedloadv2i64 addr:$src),
3292 (MOVAPSrm addr:$src)>;
3293 def : Pat<(loadv2i64 addr:$src),
3294 (MOVUPSrm addr:$src)>;
3296 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3297 (MOVAPSmr addr:$dst, VR128:$src)>;
3298 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3299 (MOVAPSmr addr:$dst, VR128:$src)>;
3300 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3301 (MOVAPSmr addr:$dst, VR128:$src)>;
3302 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3303 (MOVAPSmr addr:$dst, VR128:$src)>;
3304 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3305 (MOVUPSmr addr:$dst, VR128:$src)>;
3306 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3307 (MOVUPSmr addr:$dst, VR128:$src)>;
3308 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3309 (MOVUPSmr addr:$dst, VR128:$src)>;
3310 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3311 (MOVUPSmr addr:$dst, VR128:$src)>;
3313 //===----------------------------------------------------------------------===//
3314 // SSE4.1 Instructions
3315 //===----------------------------------------------------------------------===//
3317 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3320 Intrinsic V2F64Int> {
3321 // Intrinsic operation, reg.
3322 // Vector intrinsic operation, reg
3323 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3324 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3325 !strconcat(OpcodeStr,
3326 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3327 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3330 // Vector intrinsic operation, mem
3331 def PSm_Int : Ii8<opcps, MRMSrcMem,
3332 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3333 !strconcat(OpcodeStr,
3334 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3336 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3338 Requires<[HasSSE41]>;
3340 // Vector intrinsic operation, reg
3341 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3342 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3343 !strconcat(OpcodeStr,
3344 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3345 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3348 // Vector intrinsic operation, mem
3349 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3350 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3351 !strconcat(OpcodeStr,
3352 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3354 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3358 let Constraints = "$src1 = $dst" in {
3359 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3363 // Intrinsic operation, reg.
3364 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3366 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3367 !strconcat(OpcodeStr,
3368 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3370 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3373 // Intrinsic operation, mem.
3374 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3376 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3377 !strconcat(OpcodeStr,
3378 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3380 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3383 // Intrinsic operation, reg.
3384 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3386 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3387 !strconcat(OpcodeStr,
3388 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3390 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3393 // Intrinsic operation, mem.
3394 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3396 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3397 !strconcat(OpcodeStr,
3398 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3400 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3405 // FP round - roundss, roundps, roundsd, roundpd
3406 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3407 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3408 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3409 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3411 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3412 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3413 Intrinsic IntId128> {
3414 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3416 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3417 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3418 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3420 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3423 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3426 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3427 int_x86_sse41_phminposuw>;
3429 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3430 let Constraints = "$src1 = $dst" in {
3431 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3432 Intrinsic IntId128, bit Commutable = 0> {
3433 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3434 (ins VR128:$src1, VR128:$src2),
3435 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3436 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3438 let isCommutable = Commutable;
3440 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3441 (ins VR128:$src1, i128mem:$src2),
3442 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3444 (IntId128 VR128:$src1,
3445 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3449 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3450 int_x86_sse41_pcmpeqq, 1>;
3451 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3452 int_x86_sse41_packusdw, 0>;
3453 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3454 int_x86_sse41_pminsb, 1>;
3455 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3456 int_x86_sse41_pminsd, 1>;
3457 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3458 int_x86_sse41_pminud, 1>;
3459 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3460 int_x86_sse41_pminuw, 1>;
3461 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3462 int_x86_sse41_pmaxsb, 1>;
3463 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3464 int_x86_sse41_pmaxsd, 1>;
3465 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3466 int_x86_sse41_pmaxud, 1>;
3467 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3468 int_x86_sse41_pmaxuw, 1>;
3470 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3472 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3473 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3474 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3475 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3477 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3478 let Constraints = "$src1 = $dst" in {
3479 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3480 SDNode OpNode, Intrinsic IntId128,
3481 bit Commutable = 0> {
3482 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3483 (ins VR128:$src1, VR128:$src2),
3484 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3485 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3486 VR128:$src2))]>, OpSize {
3487 let isCommutable = Commutable;
3489 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3490 (ins VR128:$src1, VR128:$src2),
3491 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3492 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3494 let isCommutable = Commutable;
3496 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3497 (ins VR128:$src1, i128mem:$src2),
3498 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3500 (OpVT (OpNode VR128:$src1, (memop addr:$src2))))]>, OpSize;
3501 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3502 (ins VR128:$src1, i128mem:$src2),
3503 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3505 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3510 /// SS48I_binop_rm - Simple SSE41 binary operator.
3511 let Constraints = "$src1 = $dst" in {
3512 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
3513 ValueType OpVT, bit Commutable = 0> {
3514 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3515 (ins VR128:$src1, VR128:$src2),
3516 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3517 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]>,
3519 let isCommutable = Commutable;
3521 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3522 (ins VR128:$src1, i128mem:$src2),
3523 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3524 [(set VR128:$dst, (OpNode VR128:$src1,
3525 (bc_v4i32 (memopv2i64 addr:$src2))))]>,
3530 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, 1>;
3532 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3533 let Constraints = "$src1 = $dst" in {
3534 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3535 Intrinsic IntId128, bit Commutable = 0> {
3536 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3537 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3538 !strconcat(OpcodeStr,
3539 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3541 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3543 let isCommutable = Commutable;
3545 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3546 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3547 !strconcat(OpcodeStr,
3548 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3550 (IntId128 VR128:$src1,
3551 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3556 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3557 int_x86_sse41_blendps, 0>;
3558 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3559 int_x86_sse41_blendpd, 0>;
3560 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3561 int_x86_sse41_pblendw, 0>;
3562 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3563 int_x86_sse41_dpps, 1>;
3564 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3565 int_x86_sse41_dppd, 1>;
3566 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3567 int_x86_sse41_mpsadbw, 0>;
3570 /// SS41I_ternary_int - SSE 4.1 ternary operator
3571 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3572 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3573 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3574 (ins VR128:$src1, VR128:$src2),
3575 !strconcat(OpcodeStr,
3576 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3577 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3580 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3581 (ins VR128:$src1, i128mem:$src2),
3582 !strconcat(OpcodeStr,
3583 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3586 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3590 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3591 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3592 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3595 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3596 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3597 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3598 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3600 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3601 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3603 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3607 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3608 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3609 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3610 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3611 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3612 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3614 // Common patterns involving scalar load.
3615 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3616 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3617 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3618 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3620 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3621 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3622 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3623 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3625 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3626 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3627 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3628 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3630 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3631 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3632 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3633 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3635 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3636 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3637 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3638 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3640 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3641 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3642 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3643 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3646 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3647 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3648 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3649 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3651 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3652 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3654 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3658 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3659 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3660 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3661 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3663 // Common patterns involving scalar load
3664 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3665 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3666 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3667 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3669 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3670 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3671 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3672 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3675 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3676 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3677 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3678 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3680 // Expecting a i16 load any extended to i32 value.
3681 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3682 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3683 [(set VR128:$dst, (IntId (bitconvert
3684 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3688 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3689 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3691 // Common patterns involving scalar load
3692 def : Pat<(int_x86_sse41_pmovsxbq
3693 (bitconvert (v4i32 (X86vzmovl
3694 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3695 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3697 def : Pat<(int_x86_sse41_pmovzxbq
3698 (bitconvert (v4i32 (X86vzmovl
3699 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3700 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3703 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3704 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3705 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3706 (ins VR128:$src1, i32i8imm:$src2),
3707 !strconcat(OpcodeStr,
3708 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3709 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3711 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3712 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3713 !strconcat(OpcodeStr,
3714 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3717 // There's an AssertZext in the way of writing the store pattern
3718 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3721 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3724 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3725 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3726 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3727 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3728 !strconcat(OpcodeStr,
3729 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3732 // There's an AssertZext in the way of writing the store pattern
3733 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3736 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3739 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3740 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3741 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3742 (ins VR128:$src1, i32i8imm:$src2),
3743 !strconcat(OpcodeStr,
3744 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3746 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3747 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3748 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3749 !strconcat(OpcodeStr,
3750 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3751 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3752 addr:$dst)]>, OpSize;
3755 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3758 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3760 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3761 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3762 (ins VR128:$src1, i32i8imm:$src2),
3763 !strconcat(OpcodeStr,
3764 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3766 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3768 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3769 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3770 !strconcat(OpcodeStr,
3771 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3772 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3773 addr:$dst)]>, OpSize;
3776 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3778 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3779 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3782 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3783 Requires<[HasSSE41]>;
3785 let Constraints = "$src1 = $dst" in {
3786 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3787 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3788 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3789 !strconcat(OpcodeStr,
3790 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3792 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3793 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3794 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3795 !strconcat(OpcodeStr,
3796 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3798 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3799 imm:$src3))]>, OpSize;
3803 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3805 let Constraints = "$src1 = $dst" in {
3806 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3807 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3808 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3809 !strconcat(OpcodeStr,
3810 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3812 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3814 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3815 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3816 !strconcat(OpcodeStr,
3817 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3819 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3820 imm:$src3)))]>, OpSize;
3824 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3826 // insertps has a few different modes, there's the first two here below which
3827 // are optimized inserts that won't zero arbitrary elements in the destination
3828 // vector. The next one matches the intrinsic and could zero arbitrary elements
3829 // in the target vector.
3830 let Constraints = "$src1 = $dst" in {
3831 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3832 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3833 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3834 !strconcat(OpcodeStr,
3835 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3837 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3839 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3840 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3841 !strconcat(OpcodeStr,
3842 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3844 (X86insrtps VR128:$src1,
3845 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3846 imm:$src3))]>, OpSize;
3850 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3852 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3853 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3855 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3856 // the intel intrinsic that corresponds to this.
3857 let Defs = [EFLAGS] in {
3858 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3859 "ptest \t{$src2, $src1|$src1, $src2}",
3860 [(set EFLAGS, (X86ptest VR128:$src1, VR128:$src2))]>,
3862 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3863 "ptest \t{$src2, $src1|$src1, $src2}",
3864 [(set EFLAGS, (X86ptest VR128:$src1, (load addr:$src2)))]>,
3868 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3869 "movntdqa\t{$src, $dst|$dst, $src}",
3870 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
3874 //===----------------------------------------------------------------------===//
3875 // SSE4.2 Instructions
3876 //===----------------------------------------------------------------------===//
3878 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3879 let Constraints = "$src1 = $dst" in {
3880 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3881 Intrinsic IntId128, bit Commutable = 0> {
3882 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3883 (ins VR128:$src1, VR128:$src2),
3884 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3885 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3887 let isCommutable = Commutable;
3889 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3890 (ins VR128:$src1, i128mem:$src2),
3891 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3893 (IntId128 VR128:$src1,
3894 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3898 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3900 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3901 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3902 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3903 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3905 // crc intrinsic instruction
3906 // This set of instructions are only rm, the only difference is the size
3908 let Constraints = "$src1 = $dst" in {
3909 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3910 (ins GR32:$src1, i8mem:$src2),
3911 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3913 (int_x86_sse42_crc32_8 GR32:$src1,
3914 (load addr:$src2)))]>;
3915 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3916 (ins GR32:$src1, GR8:$src2),
3917 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3919 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>;
3920 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3921 (ins GR32:$src1, i16mem:$src2),
3922 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3924 (int_x86_sse42_crc32_16 GR32:$src1,
3925 (load addr:$src2)))]>,
3927 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3928 (ins GR32:$src1, GR16:$src2),
3929 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3931 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3933 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3934 (ins GR32:$src1, i32mem:$src2),
3935 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3937 (int_x86_sse42_crc32_32 GR32:$src1,
3938 (load addr:$src2)))]>;
3939 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3940 (ins GR32:$src1, GR32:$src2),
3941 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3943 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>;
3944 def CRC64m8 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3945 (ins GR64:$src1, i8mem:$src2),
3946 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3948 (int_x86_sse42_crc64_8 GR64:$src1,
3949 (load addr:$src2)))]>,
3951 def CRC64r8 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3952 (ins GR64:$src1, GR8:$src2),
3953 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3955 (int_x86_sse42_crc64_8 GR64:$src1, GR8:$src2))]>,
3957 def CRC64m64 : SS42FI<0xF1, MRMSrcMem, (outs GR64:$dst),
3958 (ins GR64:$src1, i64mem:$src2),
3959 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3961 (int_x86_sse42_crc64_64 GR64:$src1,
3962 (load addr:$src2)))]>,
3964 def CRC64r64 : SS42FI<0xF1, MRMSrcReg, (outs GR64:$dst),
3965 (ins GR64:$src1, GR64:$src2),
3966 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3968 (int_x86_sse42_crc64_64 GR64:$src1, GR64:$src2))]>,
3972 // String/text processing instructions.
3973 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3974 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3975 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3976 "#PCMPISTRM128rr PSEUDO!",
3977 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3978 imm:$src3))]>, OpSize;
3979 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3980 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3981 "#PCMPISTRM128rm PSEUDO!",
3982 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3983 imm:$src3))]>, OpSize;
3986 let Defs = [XMM0, EFLAGS] in {
3987 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3988 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3989 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3990 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3991 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3992 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3995 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3996 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3997 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3998 "#PCMPESTRM128rr PSEUDO!",
4000 (int_x86_sse42_pcmpestrm128
4001 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
4003 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
4004 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
4005 "#PCMPESTRM128rm PSEUDO!",
4006 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
4007 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
4011 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
4012 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
4013 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
4014 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
4015 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
4016 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
4017 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
4020 let Defs = [ECX, EFLAGS] in {
4021 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
4022 def rr : SS42AI<0x63, MRMSrcReg, (outs),
4023 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
4024 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
4025 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
4026 (implicit EFLAGS)]>, OpSize;
4027 def rm : SS42AI<0x63, MRMSrcMem, (outs),
4028 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
4029 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
4030 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
4031 (implicit EFLAGS)]>, OpSize;
4035 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
4036 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
4037 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
4038 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
4039 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
4040 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
4042 let Defs = [ECX, EFLAGS] in {
4043 let Uses = [EAX, EDX] in {
4044 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
4045 def rr : SS42AI<0x61, MRMSrcReg, (outs),
4046 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
4047 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
4048 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
4049 (implicit EFLAGS)]>, OpSize;
4050 def rm : SS42AI<0x61, MRMSrcMem, (outs),
4051 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
4052 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
4054 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
4055 (implicit EFLAGS)]>, OpSize;
4060 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
4061 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
4062 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
4063 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
4064 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
4065 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;
4067 //===----------------------------------------------------------------------===//
4068 // AES-NI Instructions
4069 //===----------------------------------------------------------------------===//
4071 let Constraints = "$src1 = $dst" in {
4072 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr,
4073 Intrinsic IntId128, bit Commutable = 0> {
4074 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
4075 (ins VR128:$src1, VR128:$src2),
4076 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4077 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
4079 let isCommutable = Commutable;
4081 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
4082 (ins VR128:$src1, i128mem:$src2),
4083 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4085 (IntId128 VR128:$src1,
4086 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
4090 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
4091 int_x86_aesni_aesenc>;
4092 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
4093 int_x86_aesni_aesenclast>;
4094 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
4095 int_x86_aesni_aesdec>;
4096 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
4097 int_x86_aesni_aesdeclast>;
4099 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, VR128:$src2)),
4100 (AESENCrr VR128:$src1, VR128:$src2)>;
4101 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, (memop addr:$src2))),
4102 (AESENCrm VR128:$src1, addr:$src2)>;
4103 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, VR128:$src2)),
4104 (AESENCLASTrr VR128:$src1, VR128:$src2)>;
4105 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, (memop addr:$src2))),
4106 (AESENCLASTrm VR128:$src1, addr:$src2)>;
4107 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, VR128:$src2)),
4108 (AESDECrr VR128:$src1, VR128:$src2)>;
4109 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, (memop addr:$src2))),
4110 (AESDECrm VR128:$src1, addr:$src2)>;
4111 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, VR128:$src2)),
4112 (AESDECLASTrr VR128:$src1, VR128:$src2)>;
4113 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, (memop addr:$src2))),
4114 (AESDECLASTrm VR128:$src1, addr:$src2)>;
4116 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
4118 "aesimc\t{$src1, $dst|$dst, $src1}",
4120 (int_x86_aesni_aesimc VR128:$src1))]>,
4123 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
4124 (ins i128mem:$src1),
4125 "aesimc\t{$src1, $dst|$dst, $src1}",
4127 (int_x86_aesni_aesimc (bitconvert (memopv2i64 addr:$src1))))]>,
4130 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
4131 (ins VR128:$src1, i8imm:$src2),
4132 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4134 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
4136 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
4137 (ins i128mem:$src1, i8imm:$src2),
4138 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4140 (int_x86_aesni_aeskeygenassist (bitconvert (memopv2i64 addr:$src1)),