1 //====- X86InstrSSE.td - Describe the X86 Instruction Set --*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the X86 SSE instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
17 //===----------------------------------------------------------------------===//
18 // SSE specific DAG Nodes.
19 //===----------------------------------------------------------------------===//
21 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
22 SDTCisFP<0>, SDTCisInt<2> ]>;
23 def SDTX86VFCMP : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>,
24 SDTCisFP<1>, SDTCisVT<3, i8>]>;
26 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
27 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
28 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
29 [SDNPCommutative, SDNPAssociative]>;
30 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
31 [SDNPCommutative, SDNPAssociative]>;
32 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
33 [SDNPCommutative, SDNPAssociative]>;
34 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
35 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
36 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
37 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>;
38 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>;
39 def X86pshufb : SDNode<"X86ISD::PSHUFB",
40 SDTypeProfile<1, 2, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
42 def X86pextrb : SDNode<"X86ISD::PEXTRB",
43 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
44 def X86pextrw : SDNode<"X86ISD::PEXTRW",
45 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
46 def X86pinsrb : SDNode<"X86ISD::PINSRB",
47 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
48 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
49 def X86pinsrw : SDNode<"X86ISD::PINSRW",
50 SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
51 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
52 def X86insrtps : SDNode<"X86ISD::INSERTPS",
53 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
54 SDTCisVT<2, v4f32>, SDTCisPtrTy<3>]>>;
55 def X86vzmovl : SDNode<"X86ISD::VZEXT_MOVL",
56 SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>;
57 def X86vzload : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
58 [SDNPHasChain, SDNPMayLoad]>;
59 def X86vshl : SDNode<"X86ISD::VSHL", SDTIntShiftOp>;
60 def X86vshr : SDNode<"X86ISD::VSRL", SDTIntShiftOp>;
61 def X86cmpps : SDNode<"X86ISD::CMPPS", SDTX86VFCMP>;
62 def X86cmppd : SDNode<"X86ISD::CMPPD", SDTX86VFCMP>;
63 def X86pcmpeqb : SDNode<"X86ISD::PCMPEQB", SDTIntBinOp, [SDNPCommutative]>;
64 def X86pcmpeqw : SDNode<"X86ISD::PCMPEQW", SDTIntBinOp, [SDNPCommutative]>;
65 def X86pcmpeqd : SDNode<"X86ISD::PCMPEQD", SDTIntBinOp, [SDNPCommutative]>;
66 def X86pcmpeqq : SDNode<"X86ISD::PCMPEQQ", SDTIntBinOp, [SDNPCommutative]>;
67 def X86pcmpgtb : SDNode<"X86ISD::PCMPGTB", SDTIntBinOp>;
68 def X86pcmpgtw : SDNode<"X86ISD::PCMPGTW", SDTIntBinOp>;
69 def X86pcmpgtd : SDNode<"X86ISD::PCMPGTD", SDTIntBinOp>;
70 def X86pcmpgtq : SDNode<"X86ISD::PCMPGTQ", SDTIntBinOp>;
72 def SDTX86CmpPTest : SDTypeProfile<1, 2, [SDTCisVT<0, i32>,
75 def X86ptest : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
77 //===----------------------------------------------------------------------===//
78 // SSE Complex Patterns
79 //===----------------------------------------------------------------------===//
81 // These are 'extloads' from a scalar to the low element of a vector, zeroing
82 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
84 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
85 [SDNPHasChain, SDNPMayLoad]>;
86 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
87 [SDNPHasChain, SDNPMayLoad]>;
89 def ssmem : Operand<v4f32> {
90 let PrintMethod = "printf32mem";
91 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
92 let ParserMatchClass = X86MemAsmOperand;
94 def sdmem : Operand<v2f64> {
95 let PrintMethod = "printf64mem";
96 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
97 let ParserMatchClass = X86MemAsmOperand;
100 //===----------------------------------------------------------------------===//
101 // SSE pattern fragments
102 //===----------------------------------------------------------------------===//
104 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
105 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
106 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
107 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
109 // Like 'store', but always requires vector alignment.
110 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
111 (store node:$val, node:$ptr), [{
112 return cast<StoreSDNode>(N)->getAlignment() >= 16;
115 // Like 'load', but always requires vector alignment.
116 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
117 return cast<LoadSDNode>(N)->getAlignment() >= 16;
120 def alignedloadfsf32 : PatFrag<(ops node:$ptr),
121 (f32 (alignedload node:$ptr))>;
122 def alignedloadfsf64 : PatFrag<(ops node:$ptr),
123 (f64 (alignedload node:$ptr))>;
124 def alignedloadv4f32 : PatFrag<(ops node:$ptr),
125 (v4f32 (alignedload node:$ptr))>;
126 def alignedloadv2f64 : PatFrag<(ops node:$ptr),
127 (v2f64 (alignedload node:$ptr))>;
128 def alignedloadv4i32 : PatFrag<(ops node:$ptr),
129 (v4i32 (alignedload node:$ptr))>;
130 def alignedloadv2i64 : PatFrag<(ops node:$ptr),
131 (v2i64 (alignedload node:$ptr))>;
133 // Like 'load', but uses special alignment checks suitable for use in
134 // memory operands in most SSE instructions, which are required to
135 // be naturally aligned on some targets but not on others. If the subtarget
136 // allows unaligned accesses, match any load, though this may require
137 // setting a feature bit in the processor (on startup, for example).
138 // Opteron 10h and later implement such a feature.
139 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
140 return Subtarget->hasVectorUAMem()
141 || cast<LoadSDNode>(N)->getAlignment() >= 16;
144 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
145 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
146 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
147 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
148 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
149 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
150 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
152 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
154 // FIXME: 8 byte alignment for mmx reads is not required
155 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
156 return cast<LoadSDNode>(N)->getAlignment() >= 8;
159 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
160 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
161 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
162 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
165 // Like 'store', but requires the non-temporal bit to be set
166 def nontemporalstore : PatFrag<(ops node:$val, node:$ptr),
167 (st node:$val, node:$ptr), [{
168 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
169 return ST->isNonTemporal();
173 def alignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
174 (st node:$val, node:$ptr), [{
175 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
176 return ST->isNonTemporal() && !ST->isTruncatingStore() &&
177 ST->getAddressingMode() == ISD::UNINDEXED &&
178 ST->getAlignment() >= 16;
182 def unalignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
183 (st node:$val, node:$ptr), [{
184 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
185 return ST->isNonTemporal() &&
186 ST->getAlignment() < 16;
190 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
191 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
192 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
193 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
194 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
195 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
197 def vzmovl_v2i64 : PatFrag<(ops node:$src),
198 (bitconvert (v2i64 (X86vzmovl
199 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
200 def vzmovl_v4i32 : PatFrag<(ops node:$src),
201 (bitconvert (v4i32 (X86vzmovl
202 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
204 def vzload_v2i64 : PatFrag<(ops node:$src),
205 (bitconvert (v2i64 (X86vzload node:$src)))>;
208 def fp32imm0 : PatLeaf<(f32 fpimm), [{
209 return N->isExactlyValue(+0.0);
212 // BYTE_imm - Transform bit immediates into byte immediates.
213 def BYTE_imm : SDNodeXForm<imm, [{
214 // Transformation function: imm >> 3
215 return getI32Imm(N->getZExtValue() >> 3);
218 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
220 def SHUFFLE_get_shuf_imm : SDNodeXForm<vector_shuffle, [{
221 return getI8Imm(X86::getShuffleSHUFImmediate(N));
224 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
226 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<vector_shuffle, [{
227 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
230 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
232 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<vector_shuffle, [{
233 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
236 // SHUFFLE_get_palign_imm xform function: convert vector_shuffle mask to
238 def SHUFFLE_get_palign_imm : SDNodeXForm<vector_shuffle, [{
239 return getI8Imm(X86::getShufflePALIGNRImmediate(N));
242 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
243 (vector_shuffle node:$lhs, node:$rhs), [{
244 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
245 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
248 def movddup : PatFrag<(ops node:$lhs, node:$rhs),
249 (vector_shuffle node:$lhs, node:$rhs), [{
250 return X86::isMOVDDUPMask(cast<ShuffleVectorSDNode>(N));
253 def movhlps : PatFrag<(ops node:$lhs, node:$rhs),
254 (vector_shuffle node:$lhs, node:$rhs), [{
255 return X86::isMOVHLPSMask(cast<ShuffleVectorSDNode>(N));
258 def movhlps_undef : PatFrag<(ops node:$lhs, node:$rhs),
259 (vector_shuffle node:$lhs, node:$rhs), [{
260 return X86::isMOVHLPS_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
263 def movlhps : PatFrag<(ops node:$lhs, node:$rhs),
264 (vector_shuffle node:$lhs, node:$rhs), [{
265 return X86::isMOVLHPSMask(cast<ShuffleVectorSDNode>(N));
268 def movlp : PatFrag<(ops node:$lhs, node:$rhs),
269 (vector_shuffle node:$lhs, node:$rhs), [{
270 return X86::isMOVLPMask(cast<ShuffleVectorSDNode>(N));
273 def movl : PatFrag<(ops node:$lhs, node:$rhs),
274 (vector_shuffle node:$lhs, node:$rhs), [{
275 return X86::isMOVLMask(cast<ShuffleVectorSDNode>(N));
278 def movshdup : PatFrag<(ops node:$lhs, node:$rhs),
279 (vector_shuffle node:$lhs, node:$rhs), [{
280 return X86::isMOVSHDUPMask(cast<ShuffleVectorSDNode>(N));
283 def movsldup : PatFrag<(ops node:$lhs, node:$rhs),
284 (vector_shuffle node:$lhs, node:$rhs), [{
285 return X86::isMOVSLDUPMask(cast<ShuffleVectorSDNode>(N));
288 def unpckl : PatFrag<(ops node:$lhs, node:$rhs),
289 (vector_shuffle node:$lhs, node:$rhs), [{
290 return X86::isUNPCKLMask(cast<ShuffleVectorSDNode>(N));
293 def unpckh : PatFrag<(ops node:$lhs, node:$rhs),
294 (vector_shuffle node:$lhs, node:$rhs), [{
295 return X86::isUNPCKHMask(cast<ShuffleVectorSDNode>(N));
298 def unpckl_undef : PatFrag<(ops node:$lhs, node:$rhs),
299 (vector_shuffle node:$lhs, node:$rhs), [{
300 return X86::isUNPCKL_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
303 def unpckh_undef : PatFrag<(ops node:$lhs, node:$rhs),
304 (vector_shuffle node:$lhs, node:$rhs), [{
305 return X86::isUNPCKH_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
308 def pshufd : PatFrag<(ops node:$lhs, node:$rhs),
309 (vector_shuffle node:$lhs, node:$rhs), [{
310 return X86::isPSHUFDMask(cast<ShuffleVectorSDNode>(N));
311 }], SHUFFLE_get_shuf_imm>;
313 def shufp : PatFrag<(ops node:$lhs, node:$rhs),
314 (vector_shuffle node:$lhs, node:$rhs), [{
315 return X86::isSHUFPMask(cast<ShuffleVectorSDNode>(N));
316 }], SHUFFLE_get_shuf_imm>;
318 def pshufhw : PatFrag<(ops node:$lhs, node:$rhs),
319 (vector_shuffle node:$lhs, node:$rhs), [{
320 return X86::isPSHUFHWMask(cast<ShuffleVectorSDNode>(N));
321 }], SHUFFLE_get_pshufhw_imm>;
323 def pshuflw : PatFrag<(ops node:$lhs, node:$rhs),
324 (vector_shuffle node:$lhs, node:$rhs), [{
325 return X86::isPSHUFLWMask(cast<ShuffleVectorSDNode>(N));
326 }], SHUFFLE_get_pshuflw_imm>;
328 def palign : PatFrag<(ops node:$lhs, node:$rhs),
329 (vector_shuffle node:$lhs, node:$rhs), [{
330 return X86::isPALIGNRMask(cast<ShuffleVectorSDNode>(N));
331 }], SHUFFLE_get_palign_imm>;
333 //===----------------------------------------------------------------------===//
334 // SSE scalar FP Instructions
335 //===----------------------------------------------------------------------===//
337 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded after
338 // instruction selection into a branch sequence.
339 let Uses = [EFLAGS], usesCustomInserter = 1 in {
340 def CMOV_FR32 : I<0, Pseudo,
341 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
342 "#CMOV_FR32 PSEUDO!",
343 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
345 def CMOV_FR64 : I<0, Pseudo,
346 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
347 "#CMOV_FR64 PSEUDO!",
348 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
350 def CMOV_V4F32 : I<0, Pseudo,
351 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
352 "#CMOV_V4F32 PSEUDO!",
354 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
356 def CMOV_V2F64 : I<0, Pseudo,
357 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
358 "#CMOV_V2F64 PSEUDO!",
360 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
362 def CMOV_V2I64 : I<0, Pseudo,
363 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
364 "#CMOV_V2I64 PSEUDO!",
366 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
370 //===----------------------------------------------------------------------===//
372 //===----------------------------------------------------------------------===//
374 // Move Instructions. Register-to-register movss is not used for FR32
375 // register copies because it's a partial register update; FsMOVAPSrr is
376 // used instead. Register-to-register movss is not modeled as an INSERT_SUBREG
377 // because INSERT_SUBREG requires that the insert be implementable in terms of
378 // a copy, and just mentioned, we don't use movss for copies.
379 let Constraints = "$src1 = $dst" in
380 def MOVSSrr : SSI<0x10, MRMSrcReg,
381 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
382 "movss\t{$src2, $dst|$dst, $src2}",
383 [(set (v4f32 VR128:$dst),
384 (movl VR128:$src1, (scalar_to_vector FR32:$src2)))]>;
386 // Extract the low 32-bit value from one vector and insert it into another.
387 let AddedComplexity = 15 in
388 def : Pat<(v4f32 (movl VR128:$src1, VR128:$src2)),
389 (MOVSSrr (v4f32 VR128:$src1),
390 (EXTRACT_SUBREG (v4f32 VR128:$src2), sub_ss))>;
392 // Implicitly promote a 32-bit scalar to a vector.
393 def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
394 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), FR32:$src, sub_ss)>;
396 // Loading from memory automatically zeroing upper bits.
397 let canFoldAsLoad = 1, isReMaterializable = 1 in
398 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
399 "movss\t{$src, $dst|$dst, $src}",
400 [(set FR32:$dst, (loadf32 addr:$src))]>;
402 // MOVSSrm zeros the high parts of the register; represent this
403 // with SUBREG_TO_REG.
404 let AddedComplexity = 20 in {
405 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
406 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
407 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
408 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
409 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
410 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
413 // Store scalar value to memory.
414 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
415 "movss\t{$src, $dst|$dst, $src}",
416 [(store FR32:$src, addr:$dst)]>;
418 // Extract and store.
419 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
422 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
424 // Conversion instructions
425 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
426 "cvttss2si\t{$src, $dst|$dst, $src}",
427 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
428 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
429 "cvttss2si\t{$src, $dst|$dst, $src}",
430 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
431 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
432 "cvtsi2ss\t{$src, $dst|$dst, $src}",
433 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
434 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
435 "cvtsi2ss\t{$src, $dst|$dst, $src}",
436 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
438 // Match intrinsics which expect XMM operand(s).
439 def CVTSS2SIrr: SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
440 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
441 def CVTSS2SIrm: SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
442 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
444 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
445 "cvtss2si\t{$src, $dst|$dst, $src}",
446 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
447 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
448 "cvtss2si\t{$src, $dst|$dst, $src}",
449 [(set GR32:$dst, (int_x86_sse_cvtss2si
450 (load addr:$src)))]>;
452 // Match intrinsics which expect MM and XMM operand(s).
453 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
454 "cvtps2pi\t{$src, $dst|$dst, $src}",
455 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
456 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
457 "cvtps2pi\t{$src, $dst|$dst, $src}",
458 [(set VR64:$dst, (int_x86_sse_cvtps2pi
459 (load addr:$src)))]>;
460 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
461 "cvttps2pi\t{$src, $dst|$dst, $src}",
462 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
463 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
464 "cvttps2pi\t{$src, $dst|$dst, $src}",
465 [(set VR64:$dst, (int_x86_sse_cvttps2pi
466 (load addr:$src)))]>;
467 let Constraints = "$src1 = $dst" in {
468 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
469 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
470 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
471 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
473 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
474 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
475 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
476 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
477 (load addr:$src2)))]>;
480 // Aliases for intrinsics
481 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
482 "cvttss2si\t{$src, $dst|$dst, $src}",
484 (int_x86_sse_cvttss2si VR128:$src))]>;
485 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
486 "cvttss2si\t{$src, $dst|$dst, $src}",
488 (int_x86_sse_cvttss2si(load addr:$src)))]>;
490 let Constraints = "$src1 = $dst" in {
491 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
492 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
493 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
494 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
496 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
497 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
498 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
499 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
500 (loadi32 addr:$src2)))]>;
503 // Comparison instructions
504 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
505 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
506 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
507 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
509 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
510 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
511 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
513 // Accept explicit immediate argument form instead of comparison code.
514 let isAsmParserOnly = 1 in {
515 def CMPSSrr_alt : SSIi8<0xC2, MRMSrcReg,
516 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, i8imm:$src2),
517 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
519 def CMPSSrm_alt : SSIi8<0xC2, MRMSrcMem,
520 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, i8imm:$src2),
521 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
525 let Defs = [EFLAGS] in {
526 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
527 "ucomiss\t{$src2, $src1|$src1, $src2}",
528 [(set EFLAGS, (X86cmp FR32:$src1, FR32:$src2))]>;
529 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
530 "ucomiss\t{$src2, $src1|$src1, $src2}",
531 [(set EFLAGS, (X86cmp FR32:$src1, (loadf32 addr:$src2)))]>;
533 def COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
534 "comiss\t{$src2, $src1|$src1, $src2}", []>;
535 def COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
536 "comiss\t{$src2, $src1|$src1, $src2}", []>;
540 // Aliases to match intrinsics which expect XMM operand(s).
541 let Constraints = "$src1 = $dst" in {
542 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
544 (ins VR128:$src1, VR128:$src, SSECC:$cc),
545 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
546 [(set VR128:$dst, (int_x86_sse_cmp_ss
548 VR128:$src, imm:$cc))]>;
549 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
551 (ins VR128:$src1, f32mem:$src, SSECC:$cc),
552 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
553 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
554 (load addr:$src), imm:$cc))]>;
557 let Defs = [EFLAGS] in {
558 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
559 "ucomiss\t{$src2, $src1|$src1, $src2}",
560 [(set EFLAGS, (X86ucomi (v4f32 VR128:$src1),
562 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
563 "ucomiss\t{$src2, $src1|$src1, $src2}",
564 [(set EFLAGS, (X86ucomi (v4f32 VR128:$src1),
565 (load addr:$src2)))]>;
567 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
568 "comiss\t{$src2, $src1|$src1, $src2}",
569 [(set EFLAGS, (X86comi (v4f32 VR128:$src1),
571 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
572 "comiss\t{$src2, $src1|$src1, $src2}",
573 [(set EFLAGS, (X86comi (v4f32 VR128:$src1),
574 (load addr:$src2)))]>;
577 // Aliases of packed SSE1 instructions for scalar use. These all have names
578 // that start with 'Fs'.
580 // Alias instructions that map fld0 to pxor for sse.
581 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
583 // FIXME: Set encoding to pseudo!
584 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
585 [(set FR32:$dst, fp32imm0)]>,
586 Requires<[HasSSE1]>, TB, OpSize;
588 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
590 let neverHasSideEffects = 1 in
591 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
592 "movaps\t{$src, $dst|$dst, $src}", []>;
594 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
596 let canFoldAsLoad = 1, isReMaterializable = 1 in
597 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
598 "movaps\t{$src, $dst|$dst, $src}",
599 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
601 /// sse12_fp_alias_pack_logical - SSE 1 & 2 aliased packed FP logical ops
603 multiclass sse12_fp_alias_pack_logical<bits<8> opc, string OpcodeStr,
604 SDNode OpNode, int NoPat = 0,
605 bit MayLoad = 0, bit Commutable = 1> {
606 def PSrr : PSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
607 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
609 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))])> {
610 let isCommutable = Commutable;
613 def PDrr : PDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
614 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
616 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))])> {
617 let isCommutable = Commutable;
620 def PSrm : PSI<opc, MRMSrcMem, (outs FR32:$dst),
621 (ins FR32:$src1, f128mem:$src2),
622 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
624 [(set FR32:$dst, (OpNode FR32:$src1,
625 (memopfsf32 addr:$src2)))])> {
626 let mayLoad = MayLoad;
629 def PDrm : PDI<opc, MRMSrcMem, (outs FR64:$dst),
630 (ins FR64:$src1, f128mem:$src2),
631 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
633 [(set FR64:$dst, (OpNode FR64:$src1,
634 (memopfsf64 addr:$src2)))])> {
635 let mayLoad = MayLoad;
639 // Alias bitwise logical operations using SSE logical ops on packed FP values.
640 let Constraints = "$src1 = $dst" in {
641 defm FsAND : sse12_fp_alias_pack_logical<0x54, "and", X86fand>;
642 defm FsOR : sse12_fp_alias_pack_logical<0x56, "or", X86for>;
643 defm FsXOR : sse12_fp_alias_pack_logical<0x57, "xor", X86fxor>;
645 let neverHasSideEffects = 1 in
646 defm FsANDN : sse12_fp_alias_pack_logical<0x55, "andn", undef, 1, 1, 0>;
649 /// sse12_fp_scalar - SSE 1 & 2 scalar instructions class
650 multiclass sse12_fp_scalar<bits<8> opc, string OpcodeStr, SDNode OpNode,
651 RegisterClass RC, X86MemOperand memop> {
652 let isCommutable = 1 in {
653 def rr : SI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
654 OpcodeStr, [(set RC:$dst, (OpNode RC:$src1, RC:$src2))]>;
656 def rm : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memop:$src2),
657 OpcodeStr, [(set RC:$dst, (OpNode RC:$src1, (load addr:$src2)))]>;
660 /// sse12_fp_scalar_int - SSE 1 & 2 scalar instructions intrinsics class
661 multiclass sse12_fp_scalar_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
662 string asm, string SSEVer, string FPSizeStr,
663 Operand memop, ComplexPattern mem_cpat> {
664 def rr_Int : SI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
665 asm, [(set RC:$dst, (
666 !nameconcat<Intrinsic>("int_x86_sse",
667 !strconcat(SSEVer, !strconcat("_",
668 !strconcat(OpcodeStr, FPSizeStr))))
669 RC:$src1, RC:$src2))]>;
670 def rm_Int : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memop:$src2),
671 asm, [(set RC:$dst, (
672 !nameconcat<Intrinsic>("int_x86_sse",
673 !strconcat(SSEVer, !strconcat("_",
674 !strconcat(OpcodeStr, FPSizeStr))))
675 RC:$src1, mem_cpat:$src2))]>;
678 /// sse12_fp_packed - SSE 1 & 2 packed instructions class
679 multiclass sse12_fp_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
680 RegisterClass RC, ValueType vt,
681 X86MemOperand x86memop, PatFrag mem_frag,
683 let isCommutable = 1 in
684 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
685 OpcodeStr, [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))],d>;
686 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
687 OpcodeStr, [(set RC:$dst, (OpNode RC:$src1,
688 (mem_frag addr:$src2)))],d>;
691 /// basic_sse12_fp_binop_rm - SSE 1 & 2 binops come in both scalar and
694 /// In addition, we also have a special variant of the scalar form here to
695 /// represent the associated intrinsic operation. This form is unlike the
696 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
697 /// and leaves the top elements unmodified (therefore these cannot be commuted).
699 /// These three forms can each be reg+reg or reg+mem, so there are a total of
700 /// six "instructions".
702 multiclass basic_sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
703 SDNode OpNode, bit Commutable = 0> {
705 let isAsmParserOnly = 1 in {
706 defm V#NAME#SS : sse12_fp_scalar<opc,
707 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
708 OpNode, FR32, f32mem>, XS, VEX_4V;
710 defm V#NAME#SD : sse12_fp_scalar<opc,
711 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
712 OpNode, FR64, f64mem>, XD, VEX_4V;
714 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
715 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
716 VR128, v4f32, f128mem, memopv4f32, SSEPackedSingle>,
719 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
720 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
721 VR128, v2f64, f128mem, memopv2f64, SSEPackedDouble>,
724 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
725 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
726 "", "_ss", ssmem, sse_load_f32>, XS, VEX_4V;
728 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
729 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
730 "2", "_sd", sdmem, sse_load_f64>, XD, VEX_4V;
733 let Constraints = "$src1 = $dst" in {
734 defm SS : sse12_fp_scalar<opc,
735 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
736 OpNode, FR32, f32mem>, XS;
738 defm SD : sse12_fp_scalar<opc,
739 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
740 OpNode, FR64, f64mem>, XD;
742 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
743 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v4f32,
744 f128mem, memopv4f32, SSEPackedSingle>, TB;
746 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
747 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v2f64,
748 f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
750 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
751 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
752 "", "_ss", ssmem, sse_load_f32>, XS;
754 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
755 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
756 "2", "_sd", sdmem, sse_load_f64>, XD;
760 // Arithmetic instructions
761 defm ADD : basic_sse12_fp_binop_rm<0x58, "add", fadd, 1>;
762 defm MUL : basic_sse12_fp_binop_rm<0x59, "mul", fmul, 1>;
764 let isCommutable = 0 in {
765 defm SUB : basic_sse12_fp_binop_rm<0x5C, "sub", fsub>;
766 defm DIV : basic_sse12_fp_binop_rm<0x5E, "div", fdiv>;
769 /// sse12_fp_binop_rm - Other SSE 1 & 2 binops
771 /// This multiclass is like basic_sse12_fp_binop_rm, with the addition of
772 /// instructions for a full-vector intrinsic form. Operations that map
773 /// onto C operators don't use this form since they just use the plain
774 /// vector form instead of having a separate vector intrinsic form.
776 /// This provides a total of eight "instructions".
778 let Constraints = "$src1 = $dst" in {
779 multiclass sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
780 SDNode OpNode, bit Commutable = 0> {
782 let Constraints = "", isAsmParserOnly = 1 in {
783 // Scalar operation, reg+reg.
784 defm V#NAME#SS : sse12_fp_scalar<opc,
785 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
786 OpNode, FR32, f32mem>, XS, VEX_4V;
788 defm V#NAME#SD : sse12_fp_scalar<opc,
789 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
790 OpNode, FR64, f64mem>, XD, VEX_4V;
792 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
793 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
794 VR128, v4f32, f128mem, memopv4f32, SSEPackedSingle>,
797 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
798 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
799 VR128, v2f64, f128mem, memopv2f64, SSEPackedDouble>,
802 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
803 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
804 "", "_ss", ssmem, sse_load_f32>, XS, VEX_4V;
806 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
807 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
808 "2", "_sd", sdmem, sse_load_f64>, XD, VEX_4V;
811 let Constraints = "$src1 = $dst" in {
812 // Scalar operation, reg+reg.
813 defm SS : sse12_fp_scalar<opc,
814 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
815 OpNode, FR32, f32mem>, XS;
816 defm SD : sse12_fp_scalar<opc,
817 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
818 OpNode, FR64, f64mem>, XD;
819 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
820 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v4f32,
821 f128mem, memopv4f32, SSEPackedSingle>, TB;
823 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
824 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v2f64,
825 f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
827 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
828 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
829 "", "_ss", ssmem, sse_load_f32>, XS;
831 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
832 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
833 "2", "_sd", sdmem, sse_load_f64>, XD;
836 // Vector intrinsic operation, reg+reg.
837 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
838 (ins VR128:$src1, VR128:$src2),
839 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
840 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
841 !strconcat(OpcodeStr, "_ps")) VR128:$src1,
843 // int_x86_sse_xxx_ps
844 let isCommutable = Commutable;
847 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
848 (ins VR128:$src1, VR128:$src2),
849 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
850 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
851 !strconcat(OpcodeStr, "_pd")) VR128:$src1,
853 // int_x86_sse2_xxx_pd
854 let isCommutable = Commutable;
857 // Vector intrinsic operation, reg+mem.
858 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
859 (ins VR128:$src1, f128mem:$src2),
860 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
861 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
862 !strconcat(OpcodeStr, "_ps")) VR128:$src1,
863 (memopv4f32 addr:$src2)))]>;
864 // int_x86_sse_xxx_ps
866 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
867 (ins VR128:$src1, f128mem:$src2),
868 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
869 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
870 !strconcat(OpcodeStr, "_pd")) VR128:$src1,
871 (memopv2f64 addr:$src2)))]>;
872 // int_x86_sse2_xxx_pd
876 let isCommutable = 0 in {
877 defm MAX : sse12_fp_binop_rm<0x5F, "max", X86fmax>;
878 defm MIN : sse12_fp_binop_rm<0x5D, "min", X86fmin>;
881 //===----------------------------------------------------------------------===//
882 // SSE packed FP Instructions
885 let neverHasSideEffects = 1 in
886 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
887 "movaps\t{$src, $dst|$dst, $src}", []>;
888 let canFoldAsLoad = 1, isReMaterializable = 1 in
889 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
890 "movaps\t{$src, $dst|$dst, $src}",
891 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
893 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
894 "movaps\t{$src, $dst|$dst, $src}",
895 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
897 let neverHasSideEffects = 1 in
898 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
899 "movups\t{$src, $dst|$dst, $src}", []>;
900 let canFoldAsLoad = 1, isReMaterializable = 1 in
901 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
902 "movups\t{$src, $dst|$dst, $src}",
903 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
904 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
905 "movups\t{$src, $dst|$dst, $src}",
906 [(store (v4f32 VR128:$src), addr:$dst)]>;
908 // Intrinsic forms of MOVUPS load and store
909 let canFoldAsLoad = 1, isReMaterializable = 1 in
910 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
911 "movups\t{$src, $dst|$dst, $src}",
912 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
913 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
914 "movups\t{$src, $dst|$dst, $src}",
915 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
917 let Constraints = "$src1 = $dst" in {
918 let AddedComplexity = 20 in {
919 def MOVLPSrm : PSI<0x12, MRMSrcMem,
920 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
921 "movlps\t{$src2, $dst|$dst, $src2}",
924 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
925 def MOVHPSrm : PSI<0x16, MRMSrcMem,
926 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
927 "movhps\t{$src2, $dst|$dst, $src2}",
929 (movlhps VR128:$src1,
930 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
932 } // Constraints = "$src1 = $dst"
935 def : Pat<(movlhps VR128:$src1, (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
936 (MOVHPSrm (v4i32 VR128:$src1), addr:$src2)>;
938 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
939 "movlps\t{$src, $dst|$dst, $src}",
940 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
941 (iPTR 0))), addr:$dst)]>;
943 // v2f64 extract element 1 is always custom lowered to unpack high to low
944 // and extract element 0 so the non-store version isn't too horrible.
945 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
946 "movhps\t{$src, $dst|$dst, $src}",
947 [(store (f64 (vector_extract
948 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
949 (undef)), (iPTR 0))), addr:$dst)]>;
951 let Constraints = "$src1 = $dst" in {
952 let AddedComplexity = 20 in {
953 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
954 (ins VR128:$src1, VR128:$src2),
955 "movlhps\t{$src2, $dst|$dst, $src2}",
957 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
959 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
960 (ins VR128:$src1, VR128:$src2),
961 "movhlps\t{$src2, $dst|$dst, $src2}",
963 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
965 } // Constraints = "$src1 = $dst"
967 let AddedComplexity = 20 in {
968 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
969 (MOVLHPSrr (v4f32 VR128:$src), (v4f32 VR128:$src))>;
970 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
971 (MOVLHPSrr (v2i64 VR128:$src), (v2i64 VR128:$src))>;
978 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
980 /// In addition, we also have a special variant of the scalar form here to
981 /// represent the associated intrinsic operation. This form is unlike the
982 /// plain scalar form, in that it takes an entire vector (instead of a
983 /// scalar) and leaves the top elements undefined.
985 /// And, we have a special variant form for a full-vector intrinsic form.
987 /// These four forms can each have a reg or a mem operand, so there are a
988 /// total of eight "instructions".
990 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
994 bit Commutable = 0> {
995 // Scalar operation, reg.
996 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
997 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
998 [(set FR32:$dst, (OpNode FR32:$src))]> {
999 let isCommutable = Commutable;
1002 // Scalar operation, mem.
1003 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
1004 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1005 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
1006 Requires<[HasSSE1, OptForSize]>;
1008 // Vector operation, reg.
1009 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1010 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1011 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
1012 let isCommutable = Commutable;
1015 // Vector operation, mem.
1016 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1017 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1018 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
1020 // Intrinsic operation, reg.
1021 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1022 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1023 [(set VR128:$dst, (F32Int VR128:$src))]> {
1024 let isCommutable = Commutable;
1027 // Intrinsic operation, mem.
1028 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
1029 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1030 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
1032 // Vector intrinsic operation, reg
1033 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1034 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1035 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
1036 let isCommutable = Commutable;
1039 // Vector intrinsic operation, mem
1040 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1041 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1042 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
1046 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
1047 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
1049 // Reciprocal approximations. Note that these typically require refinement
1050 // in order to obtain suitable precision.
1051 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
1052 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
1053 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
1054 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
1056 /// sse12_fp_pack_logical - SSE 1 & 2 packed FP logical ops
1058 multiclass sse12_fp_pack_logical<bits<8> opc, string OpcodeStr,
1059 SDNode OpNode, int HasPat = 0,
1061 list<list<dag>> Pattern = []> {
1062 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
1063 (ins VR128:$src1, VR128:$src2),
1064 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
1065 !if(HasPat, Pattern[0],
1066 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1,
1068 { let isCommutable = Commutable; }
1070 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1071 (ins VR128:$src1, VR128:$src2),
1072 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1073 !if(HasPat, Pattern[1],
1074 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1075 (bc_v2i64 (v2f64 VR128:$src2))))])>
1076 { let isCommutable = Commutable; }
1078 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
1079 (ins VR128:$src1, f128mem:$src2),
1080 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
1081 !if(HasPat, Pattern[2],
1082 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
1083 (memopv2i64 addr:$src2)))])>;
1085 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1086 (ins VR128:$src1, f128mem:$src2),
1087 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1088 !if(HasPat, Pattern[3],
1089 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1090 (memopv2i64 addr:$src2)))])>;
1094 let Constraints = "$src1 = $dst" in {
1095 defm AND : sse12_fp_pack_logical<0x54, "and", and>;
1096 defm OR : sse12_fp_pack_logical<0x56, "or", or>;
1097 defm XOR : sse12_fp_pack_logical<0x57, "xor", xor>;
1098 defm ANDN : sse12_fp_pack_logical<0x55, "andn", undef /* dummy */, 1, 0, [
1100 [(set VR128:$dst, (v2i64 (and (xor VR128:$src1,
1101 (bc_v2i64 (v4i32 immAllOnesV))),
1104 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1105 (bc_v2i64 (v2f64 VR128:$src2))))],
1107 [(set VR128:$dst, (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
1108 (bc_v2i64 (v4i32 immAllOnesV))),
1109 (memopv2i64 addr:$src2))))],
1111 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1112 (memopv2i64 addr:$src2)))]]>;
1115 let Constraints = "$src1 = $dst" in {
1116 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
1117 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1118 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1119 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1120 VR128:$src, imm:$cc))]>;
1121 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
1122 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1123 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1124 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1125 (memop addr:$src), imm:$cc))]>;
1127 // Accept explicit immediate argument form instead of comparison code.
1128 let isAsmParserOnly = 1 in {
1129 def CMPPSrri_alt : PSIi8<0xC2, MRMSrcReg,
1130 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1131 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1132 def CMPPSrmi_alt : PSIi8<0xC2, MRMSrcMem,
1133 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1134 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1137 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1138 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
1139 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1140 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
1142 // Shuffle and unpack instructions
1143 let Constraints = "$src1 = $dst" in {
1144 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
1145 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
1146 (outs VR128:$dst), (ins VR128:$src1,
1147 VR128:$src2, i8imm:$src3),
1148 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1150 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1151 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
1152 (outs VR128:$dst), (ins VR128:$src1,
1153 f128mem:$src2, i8imm:$src3),
1154 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1157 VR128:$src1, (memopv4f32 addr:$src2))))]>;
1159 let AddedComplexity = 10 in {
1160 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
1161 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1162 "unpckhps\t{$src2, $dst|$dst, $src2}",
1164 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
1165 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
1166 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1167 "unpckhps\t{$src2, $dst|$dst, $src2}",
1169 (v4f32 (unpckh VR128:$src1,
1170 (memopv4f32 addr:$src2))))]>;
1172 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
1173 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1174 "unpcklps\t{$src2, $dst|$dst, $src2}",
1176 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
1177 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
1178 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1179 "unpcklps\t{$src2, $dst|$dst, $src2}",
1181 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
1182 } // AddedComplexity
1183 } // Constraints = "$src1 = $dst"
1186 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1187 "movmskps\t{$src, $dst|$dst, $src}",
1188 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
1189 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1190 "movmskpd\t{$src, $dst|$dst, $src}",
1191 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1193 // Prefetch intrinsic.
1194 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1195 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1196 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1197 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1198 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1199 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1200 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1201 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1203 // Non-temporal stores
1204 def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1205 "movntps\t{$src, $dst|$dst, $src}",
1206 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1208 let AddedComplexity = 400 in { // Prefer non-temporal versions
1209 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1210 "movntps\t{$src, $dst|$dst, $src}",
1211 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
1213 def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1214 "movntdq\t{$src, $dst|$dst, $src}",
1215 [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
1217 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
1218 "movnti\t{$src, $dst|$dst, $src}",
1219 [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
1220 TB, Requires<[HasSSE2]>;
1222 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
1223 "movnti\t{$src, $dst|$dst, $src}",
1224 [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
1225 TB, Requires<[HasSSE2]>;
1228 // Load, store, and memory fence
1229 def SFENCE : I<0xAE, MRM_F8, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>,
1230 TB, Requires<[HasSSE1]>;
1233 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1234 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1235 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1236 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1238 // Alias instructions that map zero vector to pxor / xorp* for sse.
1239 // We set canFoldAsLoad because this can be converted to a constant-pool
1240 // load of an all-zeros value if folding it would be beneficial.
1241 // FIXME: Change encoding to pseudo!
1242 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1243 isCodeGenOnly = 1 in {
1244 def V_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1245 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
1246 def V_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1247 [(set VR128:$dst, (v2f64 immAllZerosV))]>;
1248 let ExeDomain = SSEPackedInt in
1249 def V_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
1250 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1253 def : Pat<(v2i64 immAllZerosV), (V_SET0PI)>;
1254 def : Pat<(v8i16 immAllZerosV), (V_SET0PI)>;
1255 def : Pat<(v16i8 immAllZerosV), (V_SET0PI)>;
1257 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1258 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
1260 //===---------------------------------------------------------------------===//
1261 // SSE2 Instructions
1262 //===---------------------------------------------------------------------===//
1264 // Move Instructions. Register-to-register movsd is not used for FR64
1265 // register copies because it's a partial register update; FsMOVAPDrr is
1266 // used instead. Register-to-register movsd is not modeled as an INSERT_SUBREG
1267 // because INSERT_SUBREG requires that the insert be implementable in terms of
1268 // a copy, and just mentioned, we don't use movsd for copies.
1269 let Constraints = "$src1 = $dst" in
1270 def MOVSDrr : SDI<0x10, MRMSrcReg,
1271 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
1272 "movsd\t{$src2, $dst|$dst, $src2}",
1273 [(set (v2f64 VR128:$dst),
1274 (movl VR128:$src1, (scalar_to_vector FR64:$src2)))]>;
1276 // Extract the low 64-bit value from one vector and insert it into another.
1277 let AddedComplexity = 15 in
1278 def : Pat<(v2f64 (movl VR128:$src1, VR128:$src2)),
1279 (MOVSDrr (v2f64 VR128:$src1),
1280 (EXTRACT_SUBREG (v2f64 VR128:$src2), sub_sd))>;
1282 // Implicitly promote a 64-bit scalar to a vector.
1283 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
1284 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), FR64:$src, sub_sd)>;
1286 // Loading from memory automatically zeroing upper bits.
1287 let canFoldAsLoad = 1, isReMaterializable = 1, AddedComplexity = 20 in
1288 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1289 "movsd\t{$src, $dst|$dst, $src}",
1290 [(set FR64:$dst, (loadf64 addr:$src))]>;
1292 // MOVSDrm zeros the high parts of the register; represent this
1293 // with SUBREG_TO_REG.
1294 let AddedComplexity = 20 in {
1295 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
1296 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1297 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
1298 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1299 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
1300 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1301 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
1302 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1303 def : Pat<(v2f64 (X86vzload addr:$src)),
1304 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1307 // Store scalar value to memory.
1308 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1309 "movsd\t{$src, $dst|$dst, $src}",
1310 [(store FR64:$src, addr:$dst)]>;
1312 // Extract and store.
1313 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
1316 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
1318 // Conversion instructions
1319 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1320 "cvttsd2si\t{$src, $dst|$dst, $src}",
1321 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1322 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1323 "cvttsd2si\t{$src, $dst|$dst, $src}",
1324 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1325 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1326 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1327 [(set FR32:$dst, (fround FR64:$src))]>;
1328 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1329 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1330 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
1331 Requires<[HasSSE2, OptForSize]>;
1332 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1333 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1334 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1335 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1336 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1337 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1339 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1340 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1341 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1342 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1343 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1344 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1345 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1346 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1347 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1348 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1349 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1350 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1351 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1352 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1353 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1354 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1355 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1356 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1357 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1358 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1360 // SSE2 instructions with XS prefix
1361 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1362 "cvtss2sd\t{$src, $dst|$dst, $src}",
1363 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1364 Requires<[HasSSE2]>;
1365 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1366 "cvtss2sd\t{$src, $dst|$dst, $src}",
1367 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1368 Requires<[HasSSE2, OptForSize]>;
1370 def : Pat<(extloadf32 addr:$src),
1371 (CVTSS2SDrr (MOVSSrm addr:$src))>,
1372 Requires<[HasSSE2, OptForSpeed]>;
1374 // Match intrinsics which expect XMM operand(s).
1375 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1376 "cvtsd2si\t{$src, $dst|$dst, $src}",
1377 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1378 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1379 "cvtsd2si\t{$src, $dst|$dst, $src}",
1380 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1381 (load addr:$src)))]>;
1383 // Match intrinsics which expect MM and XMM operand(s).
1384 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1385 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1386 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1387 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1388 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1389 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1390 (memop addr:$src)))]>;
1391 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1392 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1393 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1394 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1395 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1396 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1397 (memop addr:$src)))]>;
1398 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1399 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1400 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1401 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1402 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1403 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1404 (load addr:$src)))]>;
1406 // Aliases for intrinsics
1407 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1408 "cvttsd2si\t{$src, $dst|$dst, $src}",
1410 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1411 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1412 "cvttsd2si\t{$src, $dst|$dst, $src}",
1413 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1414 (load addr:$src)))]>;
1416 // Comparison instructions
1417 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1418 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1419 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1420 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1422 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1423 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1424 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1426 // Accept explicit immediate argument form instead of comparison code.
1427 let isAsmParserOnly = 1 in {
1428 def CMPSDrr_alt : SDIi8<0xC2, MRMSrcReg,
1429 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, i8imm:$src2),
1430 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1432 def CMPSDrm_alt : SDIi8<0xC2, MRMSrcMem,
1433 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, i8imm:$src2),
1434 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1438 let Defs = [EFLAGS] in {
1439 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1440 "ucomisd\t{$src2, $src1|$src1, $src2}",
1441 [(set EFLAGS, (X86cmp FR64:$src1, FR64:$src2))]>;
1442 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1443 "ucomisd\t{$src2, $src1|$src1, $src2}",
1444 [(set EFLAGS, (X86cmp FR64:$src1, (loadf64 addr:$src2)))]>;
1445 } // Defs = [EFLAGS]
1447 // Aliases to match intrinsics which expect XMM operand(s).
1448 let Constraints = "$src1 = $dst" in {
1449 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1451 (ins VR128:$src1, VR128:$src, SSECC:$cc),
1452 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1453 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1454 VR128:$src, imm:$cc))]>;
1455 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1457 (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1458 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1459 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1460 (load addr:$src), imm:$cc))]>;
1463 let Defs = [EFLAGS] in {
1464 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1465 "ucomisd\t{$src2, $src1|$src1, $src2}",
1466 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1468 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1469 "ucomisd\t{$src2, $src1|$src1, $src2}",
1470 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1471 (load addr:$src2)))]>;
1473 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1474 "comisd\t{$src2, $src1|$src1, $src2}",
1475 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1477 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1478 "comisd\t{$src2, $src1|$src1, $src2}",
1479 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1480 (load addr:$src2)))]>;
1481 } // Defs = [EFLAGS]
1483 // Aliases of packed SSE2 instructions for scalar use. These all have names
1484 // that start with 'Fs'.
1486 // Alias instructions that map fld0 to pxor for sse.
1487 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1488 canFoldAsLoad = 1 in
1489 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1490 [(set FR64:$dst, fpimm0)]>,
1491 Requires<[HasSSE2]>, TB, OpSize;
1493 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1495 let neverHasSideEffects = 1 in
1496 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1497 "movapd\t{$src, $dst|$dst, $src}", []>;
1499 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1501 let canFoldAsLoad = 1, isReMaterializable = 1 in
1502 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1503 "movapd\t{$src, $dst|$dst, $src}",
1504 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1506 //===---------------------------------------------------------------------===//
1507 // SSE packed FP Instructions
1509 // Move Instructions
1510 let neverHasSideEffects = 1 in
1511 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1512 "movapd\t{$src, $dst|$dst, $src}", []>;
1513 let canFoldAsLoad = 1, isReMaterializable = 1 in
1514 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1515 "movapd\t{$src, $dst|$dst, $src}",
1516 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1518 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1519 "movapd\t{$src, $dst|$dst, $src}",
1520 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1522 let neverHasSideEffects = 1 in
1523 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1524 "movupd\t{$src, $dst|$dst, $src}", []>;
1525 let canFoldAsLoad = 1 in
1526 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1527 "movupd\t{$src, $dst|$dst, $src}",
1528 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1529 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1530 "movupd\t{$src, $dst|$dst, $src}",
1531 [(store (v2f64 VR128:$src), addr:$dst)]>;
1533 // Intrinsic forms of MOVUPD load and store
1534 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1535 "movupd\t{$src, $dst|$dst, $src}",
1536 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1537 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1538 "movupd\t{$src, $dst|$dst, $src}",
1539 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1541 let Constraints = "$src1 = $dst" in {
1542 let AddedComplexity = 20 in {
1543 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1544 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1545 "movlpd\t{$src2, $dst|$dst, $src2}",
1547 (v2f64 (movlp VR128:$src1,
1548 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1549 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1550 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1551 "movhpd\t{$src2, $dst|$dst, $src2}",
1553 (v2f64 (movlhps VR128:$src1,
1554 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1555 } // AddedComplexity
1556 } // Constraints = "$src1 = $dst"
1558 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1559 "movlpd\t{$src, $dst|$dst, $src}",
1560 [(store (f64 (vector_extract (v2f64 VR128:$src),
1561 (iPTR 0))), addr:$dst)]>;
1563 // v2f64 extract element 1 is always custom lowered to unpack high to low
1564 // and extract element 0 so the non-store version isn't too horrible.
1565 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1566 "movhpd\t{$src, $dst|$dst, $src}",
1567 [(store (f64 (vector_extract
1568 (v2f64 (unpckh VR128:$src, (undef))),
1569 (iPTR 0))), addr:$dst)]>;
1571 // SSE2 instructions without OpSize prefix
1572 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1573 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1574 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1575 TB, Requires<[HasSSE2]>;
1576 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1577 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1578 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1579 (bitconvert (memopv2i64 addr:$src))))]>,
1580 TB, Requires<[HasSSE2]>;
1582 // SSE2 instructions with XS prefix
1583 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1584 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1585 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1586 XS, Requires<[HasSSE2]>;
1587 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1588 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1589 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1590 (bitconvert (memopv2i64 addr:$src))))]>,
1591 XS, Requires<[HasSSE2]>;
1593 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1594 "cvtps2dq\t{$src, $dst|$dst, $src}",
1595 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1596 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1597 "cvtps2dq\t{$src, $dst|$dst, $src}",
1598 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1599 (memop addr:$src)))]>;
1600 // SSE2 packed instructions with XS prefix
1601 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1602 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1603 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1604 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1606 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1607 "cvttps2dq\t{$src, $dst|$dst, $src}",
1609 (int_x86_sse2_cvttps2dq VR128:$src))]>,
1610 XS, Requires<[HasSSE2]>;
1611 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1612 "cvttps2dq\t{$src, $dst|$dst, $src}",
1613 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1614 (memop addr:$src)))]>,
1615 XS, Requires<[HasSSE2]>;
1617 // SSE2 packed instructions with XD prefix
1618 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1619 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1620 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1621 XD, Requires<[HasSSE2]>;
1622 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1623 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1624 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1625 (memop addr:$src)))]>,
1626 XD, Requires<[HasSSE2]>;
1628 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1629 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1630 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1631 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1632 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1633 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1634 (memop addr:$src)))]>;
1636 // SSE2 instructions without OpSize prefix
1637 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1638 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1639 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1640 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1642 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1643 "cvtps2pd\t{$src, $dst|$dst, $src}",
1644 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1645 TB, Requires<[HasSSE2]>;
1646 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1647 "cvtps2pd\t{$src, $dst|$dst, $src}",
1648 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1649 (load addr:$src)))]>,
1650 TB, Requires<[HasSSE2]>;
1652 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1653 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1654 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1655 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1658 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1659 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1660 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1661 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1662 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1663 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1664 (memop addr:$src)))]>;
1666 // Match intrinsics which expect XMM operand(s).
1667 // Aliases for intrinsics
1668 let Constraints = "$src1 = $dst" in {
1669 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1670 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1671 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1672 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1674 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1675 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1676 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1677 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1678 (loadi32 addr:$src2)))]>;
1679 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1680 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1681 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1682 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1684 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1685 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1686 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1687 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1688 (load addr:$src2)))]>;
1689 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1690 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1691 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1692 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1693 VR128:$src2))]>, XS,
1694 Requires<[HasSSE2]>;
1695 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1696 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1697 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1698 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1699 (load addr:$src2)))]>, XS,
1700 Requires<[HasSSE2]>;
1705 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1707 /// In addition, we also have a special variant of the scalar form here to
1708 /// represent the associated intrinsic operation. This form is unlike the
1709 /// plain scalar form, in that it takes an entire vector (instead of a
1710 /// scalar) and leaves the top elements undefined.
1712 /// And, we have a special variant form for a full-vector intrinsic form.
1714 /// These four forms can each have a reg or a mem operand, so there are a
1715 /// total of eight "instructions".
1717 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1721 bit Commutable = 0> {
1722 // Scalar operation, reg.
1723 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1724 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1725 [(set FR64:$dst, (OpNode FR64:$src))]> {
1726 let isCommutable = Commutable;
1729 // Scalar operation, mem.
1730 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1731 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1732 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1734 // Vector operation, reg.
1735 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1736 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1737 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1738 let isCommutable = Commutable;
1741 // Vector operation, mem.
1742 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1743 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1744 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1746 // Intrinsic operation, reg.
1747 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1748 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1749 [(set VR128:$dst, (F64Int VR128:$src))]> {
1750 let isCommutable = Commutable;
1753 // Intrinsic operation, mem.
1754 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1755 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1756 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1758 // Vector intrinsic operation, reg
1759 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1760 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1761 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1762 let isCommutable = Commutable;
1765 // Vector intrinsic operation, mem
1766 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1767 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1768 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1772 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1773 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1775 // There is no f64 version of the reciprocal approximation instructions.
1777 let Constraints = "$src1 = $dst" in {
1778 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1779 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1780 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1781 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1782 VR128:$src, imm:$cc))]>;
1783 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1784 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1785 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1786 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1787 (memop addr:$src), imm:$cc))]>;
1789 // Accept explicit immediate argument form instead of comparison code.
1790 let isAsmParserOnly = 1 in {
1791 def CMPPDrri_alt : PDIi8<0xC2, MRMSrcReg,
1792 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1793 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1794 def CMPPDrmi_alt : PDIi8<0xC2, MRMSrcMem,
1795 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1796 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1799 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1800 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1801 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1802 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1804 // Shuffle and unpack instructions
1805 let Constraints = "$src1 = $dst" in {
1806 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1807 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1808 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1810 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1811 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1812 (outs VR128:$dst), (ins VR128:$src1,
1813 f128mem:$src2, i8imm:$src3),
1814 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1817 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1819 let AddedComplexity = 10 in {
1820 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1821 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1822 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1824 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1825 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1826 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1827 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1829 (v2f64 (unpckh VR128:$src1,
1830 (memopv2f64 addr:$src2))))]>;
1832 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1833 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1834 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1836 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1837 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1838 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1839 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1841 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1842 } // AddedComplexity
1843 } // Constraints = "$src1 = $dst"
1846 //===---------------------------------------------------------------------===//
1847 // SSE integer instructions
1848 let ExeDomain = SSEPackedInt in {
1850 // Move Instructions
1851 let neverHasSideEffects = 1 in
1852 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1853 "movdqa\t{$src, $dst|$dst, $src}", []>;
1854 let canFoldAsLoad = 1, mayLoad = 1 in
1855 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1856 "movdqa\t{$src, $dst|$dst, $src}",
1857 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1859 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1860 "movdqa\t{$src, $dst|$dst, $src}",
1861 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1862 let canFoldAsLoad = 1, mayLoad = 1 in
1863 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1864 "movdqu\t{$src, $dst|$dst, $src}",
1865 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1866 XS, Requires<[HasSSE2]>;
1868 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1869 "movdqu\t{$src, $dst|$dst, $src}",
1870 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1871 XS, Requires<[HasSSE2]>;
1873 // Intrinsic forms of MOVDQU load and store
1874 let canFoldAsLoad = 1 in
1875 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1876 "movdqu\t{$src, $dst|$dst, $src}",
1877 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1878 XS, Requires<[HasSSE2]>;
1879 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1880 "movdqu\t{$src, $dst|$dst, $src}",
1881 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1882 XS, Requires<[HasSSE2]>;
1884 let Constraints = "$src1 = $dst" in {
1886 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1887 bit Commutable = 0> {
1888 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1889 (ins VR128:$src1, VR128:$src2),
1890 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1891 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1892 let isCommutable = Commutable;
1894 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1895 (ins VR128:$src1, i128mem:$src2),
1896 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1897 [(set VR128:$dst, (IntId VR128:$src1,
1898 (bitconvert (memopv2i64
1902 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1904 Intrinsic IntId, Intrinsic IntId2> {
1905 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1906 (ins VR128:$src1, VR128:$src2),
1907 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1908 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1909 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1910 (ins VR128:$src1, i128mem:$src2),
1911 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1912 [(set VR128:$dst, (IntId VR128:$src1,
1913 (bitconvert (memopv2i64 addr:$src2))))]>;
1914 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
1915 (ins VR128:$src1, i32i8imm:$src2),
1916 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1917 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1920 /// PDI_binop_rm - Simple SSE2 binary operator.
1921 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1922 ValueType OpVT, bit Commutable = 0> {
1923 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1924 (ins VR128:$src1, VR128:$src2),
1925 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1926 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1927 let isCommutable = Commutable;
1929 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1930 (ins VR128:$src1, i128mem:$src2),
1931 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1932 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1933 (bitconvert (memopv2i64 addr:$src2)))))]>;
1936 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1938 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1939 /// to collapse (bitconvert VT to VT) into its operand.
1941 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1942 bit Commutable = 0> {
1943 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1944 (ins VR128:$src1, VR128:$src2),
1945 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1946 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1947 let isCommutable = Commutable;
1949 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1950 (ins VR128:$src1, i128mem:$src2),
1951 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1952 [(set VR128:$dst, (OpNode VR128:$src1,
1953 (memopv2i64 addr:$src2)))]>;
1956 } // Constraints = "$src1 = $dst"
1957 } // ExeDomain = SSEPackedInt
1959 // 128-bit Integer Arithmetic
1961 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1962 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1963 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1964 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1966 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1967 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1968 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1969 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1971 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1972 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1973 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1974 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1976 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1977 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1978 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1979 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1981 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1983 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1984 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1985 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1987 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1989 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1990 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1993 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1994 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1995 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1996 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1997 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
2000 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
2001 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
2002 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
2003 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
2004 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
2005 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
2007 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2008 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2009 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2010 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2011 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2012 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2014 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2015 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2016 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2017 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2019 // 128-bit logical shifts.
2020 let Constraints = "$src1 = $dst", neverHasSideEffects = 1,
2021 ExeDomain = SSEPackedInt in {
2022 def PSLLDQri : PDIi8<0x73, MRM7r,
2023 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2024 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2025 def PSRLDQri : PDIi8<0x73, MRM3r,
2026 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2027 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2028 // PSRADQri doesn't exist in SSE[1-3].
2031 let Predicates = [HasSSE2] in {
2032 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2033 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2034 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2035 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2036 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2037 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2038 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2039 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2040 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2041 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2043 // Shift up / down and insert zero's.
2044 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2045 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2046 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2047 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2051 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2052 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2053 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2055 let Constraints = "$src1 = $dst", ExeDomain = SSEPackedInt in {
2056 def PANDNrr : PDI<0xDF, MRMSrcReg,
2057 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2058 "pandn\t{$src2, $dst|$dst, $src2}",
2059 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2062 def PANDNrm : PDI<0xDF, MRMSrcMem,
2063 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2064 "pandn\t{$src2, $dst|$dst, $src2}",
2065 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2066 (memopv2i64 addr:$src2))))]>;
2069 // SSE2 Integer comparison
2070 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2071 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2072 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2073 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2074 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2075 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2077 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2078 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2079 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2080 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2081 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2082 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2083 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2084 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2085 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2086 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2087 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2088 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2090 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2091 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2092 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2093 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2094 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2095 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2096 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2097 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2098 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2099 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2100 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2101 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2104 // Pack instructions
2105 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2106 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2107 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2109 let ExeDomain = SSEPackedInt in {
2111 // Shuffle and unpack instructions
2112 let AddedComplexity = 5 in {
2113 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2114 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2115 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2116 [(set VR128:$dst, (v4i32 (pshufd:$src2
2117 VR128:$src1, (undef))))]>;
2118 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2119 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2120 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2121 [(set VR128:$dst, (v4i32 (pshufd:$src2
2122 (bc_v4i32 (memopv2i64 addr:$src1)),
2126 // SSE2 with ImmT == Imm8 and XS prefix.
2127 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2128 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2129 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2130 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2132 XS, Requires<[HasSSE2]>;
2133 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2134 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2135 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2136 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2137 (bc_v8i16 (memopv2i64 addr:$src1)),
2139 XS, Requires<[HasSSE2]>;
2141 // SSE2 with ImmT == Imm8 and XD prefix.
2142 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2143 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2144 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2145 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2147 XD, Requires<[HasSSE2]>;
2148 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2149 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2150 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2151 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2152 (bc_v8i16 (memopv2i64 addr:$src1)),
2154 XD, Requires<[HasSSE2]>;
2156 // Unpack instructions
2157 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
2158 PatFrag unp_frag, PatFrag bc_frag> {
2159 def rr : PDI<opc, MRMSrcReg,
2160 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2161 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2162 [(set VR128:$dst, (vt (unp_frag VR128:$src1, VR128:$src2)))]>;
2163 def rm : PDI<opc, MRMSrcMem,
2164 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2165 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2166 [(set VR128:$dst, (unp_frag VR128:$src1,
2167 (bc_frag (memopv2i64
2171 let Constraints = "$src1 = $dst" in {
2172 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, unpckl, bc_v16i8>;
2173 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, unpckl, bc_v8i16>;
2174 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, unpckl, bc_v4i32>;
2176 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2177 /// knew to collapse (bitconvert VT to VT) into its operand.
2178 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2179 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2180 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2182 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2183 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2184 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2185 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2187 (v2i64 (unpckl VR128:$src1,
2188 (memopv2i64 addr:$src2))))]>;
2190 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, unpckh, bc_v16i8>;
2191 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, unpckh, bc_v8i16>;
2192 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, unpckh, bc_v4i32>;
2194 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2195 /// knew to collapse (bitconvert VT to VT) into its operand.
2196 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2197 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2198 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2200 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2201 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2202 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2203 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2205 (v2i64 (unpckh VR128:$src1,
2206 (memopv2i64 addr:$src2))))]>;
2210 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2211 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2212 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2213 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2215 let Constraints = "$src1 = $dst" in {
2216 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2217 (outs VR128:$dst), (ins VR128:$src1,
2218 GR32:$src2, i32i8imm:$src3),
2219 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2221 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2222 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2223 (outs VR128:$dst), (ins VR128:$src1,
2224 i16mem:$src2, i32i8imm:$src3),
2225 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2227 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2232 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2233 "pmovmskb\t{$src, $dst|$dst, $src}",
2234 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2236 // Conditional store
2238 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2239 "maskmovdqu\t{$mask, $src|$src, $mask}",
2240 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2243 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2244 "maskmovdqu\t{$mask, $src|$src, $mask}",
2245 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2247 } // ExeDomain = SSEPackedInt
2249 // Non-temporal stores
2250 def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2251 "movntpd\t{$src, $dst|$dst, $src}",
2252 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2253 let ExeDomain = SSEPackedInt in
2254 def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2255 "movntdq\t{$src, $dst|$dst, $src}",
2256 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2257 def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2258 "movnti\t{$src, $dst|$dst, $src}",
2259 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2260 TB, Requires<[HasSSE2]>;
2262 let AddedComplexity = 400 in { // Prefer non-temporal versions
2263 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2264 "movntpd\t{$src, $dst|$dst, $src}",
2265 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
2267 let ExeDomain = SSEPackedInt in
2268 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2269 "movntdq\t{$src, $dst|$dst, $src}",
2270 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2274 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2275 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2276 TB, Requires<[HasSSE2]>;
2278 // Load, store, and memory fence
2279 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
2280 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2281 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
2282 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2284 // Pause. This "instruction" is encoded as "rep; nop", so even though it
2285 // was introduced with SSE2, it's backward compatible.
2286 def PAUSE : I<0x90, RawFrm, (outs), (ins), "pause", []>, REP;
2288 //TODO: custom lower this so as to never even generate the noop
2289 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2291 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2292 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2293 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2296 // Alias instructions that map zero vector to pxor / xorp* for sse.
2297 // We set canFoldAsLoad because this can be converted to a constant-pool
2298 // load of an all-ones value if folding it would be beneficial.
2299 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2300 isCodeGenOnly = 1, ExeDomain = SSEPackedInt in
2301 // FIXME: Change encoding to pseudo.
2302 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
2303 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2305 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2306 "movd\t{$src, $dst|$dst, $src}",
2308 (v4i32 (scalar_to_vector GR32:$src)))]>;
2309 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2310 "movd\t{$src, $dst|$dst, $src}",
2312 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2314 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2315 "movd\t{$src, $dst|$dst, $src}",
2316 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2318 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2319 "movd\t{$src, $dst|$dst, $src}",
2320 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2322 // SSE2 instructions with XS prefix
2323 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2324 "movq\t{$src, $dst|$dst, $src}",
2326 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2327 Requires<[HasSSE2]>;
2328 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2329 "movq\t{$src, $dst|$dst, $src}",
2330 [(store (i64 (vector_extract (v2i64 VR128:$src),
2331 (iPTR 0))), addr:$dst)]>;
2333 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
2334 (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
2336 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2337 "movd\t{$src, $dst|$dst, $src}",
2338 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2340 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2341 "movd\t{$src, $dst|$dst, $src}",
2342 [(store (i32 (vector_extract (v4i32 VR128:$src),
2343 (iPTR 0))), addr:$dst)]>;
2345 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2346 "movd\t{$src, $dst|$dst, $src}",
2347 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2348 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2349 "movd\t{$src, $dst|$dst, $src}",
2350 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2352 // Store / copy lower 64-bits of a XMM register.
2353 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2354 "movq\t{$src, $dst|$dst, $src}",
2355 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2357 // movd / movq to XMM register zero-extends
2358 let AddedComplexity = 15 in {
2359 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2360 "movd\t{$src, $dst|$dst, $src}",
2361 [(set VR128:$dst, (v4i32 (X86vzmovl
2362 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2363 // This is X86-64 only.
2364 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2365 "mov{d|q}\t{$src, $dst|$dst, $src}",
2366 [(set VR128:$dst, (v2i64 (X86vzmovl
2367 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2370 let AddedComplexity = 20 in {
2371 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2372 "movd\t{$src, $dst|$dst, $src}",
2374 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2375 (loadi32 addr:$src))))))]>;
2377 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2378 (MOVZDI2PDIrm addr:$src)>;
2379 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2380 (MOVZDI2PDIrm addr:$src)>;
2381 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2382 (MOVZDI2PDIrm addr:$src)>;
2384 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2385 "movq\t{$src, $dst|$dst, $src}",
2387 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2388 (loadi64 addr:$src))))))]>, XS,
2389 Requires<[HasSSE2]>;
2391 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2392 (MOVZQI2PQIrm addr:$src)>;
2393 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2394 (MOVZQI2PQIrm addr:$src)>;
2395 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2398 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2399 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2400 let AddedComplexity = 15 in
2401 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2402 "movq\t{$src, $dst|$dst, $src}",
2403 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2404 XS, Requires<[HasSSE2]>;
2406 let AddedComplexity = 20 in {
2407 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2408 "movq\t{$src, $dst|$dst, $src}",
2409 [(set VR128:$dst, (v2i64 (X86vzmovl
2410 (loadv2i64 addr:$src))))]>,
2411 XS, Requires<[HasSSE2]>;
2413 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2414 (MOVZPQILo2PQIrm addr:$src)>;
2417 // Instructions for the disassembler
2418 // xr = XMM register
2421 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2422 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2424 //===---------------------------------------------------------------------===//
2425 // SSE3 Instructions
2426 //===---------------------------------------------------------------------===//
2428 // Move Instructions
2429 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2430 "movshdup\t{$src, $dst|$dst, $src}",
2431 [(set VR128:$dst, (v4f32 (movshdup
2432 VR128:$src, (undef))))]>;
2433 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2434 "movshdup\t{$src, $dst|$dst, $src}",
2435 [(set VR128:$dst, (movshdup
2436 (memopv4f32 addr:$src), (undef)))]>;
2438 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2439 "movsldup\t{$src, $dst|$dst, $src}",
2440 [(set VR128:$dst, (v4f32 (movsldup
2441 VR128:$src, (undef))))]>;
2442 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2443 "movsldup\t{$src, $dst|$dst, $src}",
2444 [(set VR128:$dst, (movsldup
2445 (memopv4f32 addr:$src), (undef)))]>;
2447 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2448 "movddup\t{$src, $dst|$dst, $src}",
2449 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2450 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2451 "movddup\t{$src, $dst|$dst, $src}",
2453 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2456 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2458 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2460 let AddedComplexity = 5 in {
2461 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2462 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2463 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2464 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2465 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2466 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2467 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2468 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2472 let Constraints = "$src1 = $dst" in {
2473 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2474 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2475 "addsubps\t{$src2, $dst|$dst, $src2}",
2476 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2478 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2479 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2480 "addsubps\t{$src2, $dst|$dst, $src2}",
2481 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2482 (memop addr:$src2)))]>;
2483 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2484 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2485 "addsubpd\t{$src2, $dst|$dst, $src2}",
2486 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2488 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2489 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2490 "addsubpd\t{$src2, $dst|$dst, $src2}",
2491 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2492 (memop addr:$src2)))]>;
2495 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2496 "lddqu\t{$src, $dst|$dst, $src}",
2497 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2500 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2501 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2502 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2503 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2504 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2505 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2506 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2507 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2508 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2509 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2510 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2511 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2512 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2513 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2514 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2515 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2517 let Constraints = "$src1 = $dst" in {
2518 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2519 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2520 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2521 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2522 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2523 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2524 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2525 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2528 // Thread synchronization
2529 def MONITOR : I<0x01, MRM_C8, (outs), (ins), "monitor",
2530 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2531 def MWAIT : I<0x01, MRM_C9, (outs), (ins), "mwait",
2532 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2534 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2535 let AddedComplexity = 15 in
2536 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2537 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2538 let AddedComplexity = 20 in
2539 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2540 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2542 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2543 let AddedComplexity = 15 in
2544 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2545 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2546 let AddedComplexity = 20 in
2547 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2548 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2550 //===---------------------------------------------------------------------===//
2551 // SSSE3 Instructions
2552 //===---------------------------------------------------------------------===//
2554 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2555 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2556 Intrinsic IntId64, Intrinsic IntId128> {
2557 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2558 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2559 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2561 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2562 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2564 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2566 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2568 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2569 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2572 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2574 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2577 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2580 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2581 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2582 Intrinsic IntId64, Intrinsic IntId128> {
2583 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2585 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2586 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2588 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2590 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2593 (bitconvert (memopv4i16 addr:$src))))]>;
2595 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2597 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2598 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2601 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2603 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2606 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2609 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2610 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2611 Intrinsic IntId64, Intrinsic IntId128> {
2612 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2614 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2615 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2617 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2619 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2622 (bitconvert (memopv2i32 addr:$src))))]>;
2624 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2626 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2627 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2630 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2632 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2635 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2638 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2639 int_x86_ssse3_pabs_b,
2640 int_x86_ssse3_pabs_b_128>;
2641 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2642 int_x86_ssse3_pabs_w,
2643 int_x86_ssse3_pabs_w_128>;
2644 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2645 int_x86_ssse3_pabs_d,
2646 int_x86_ssse3_pabs_d_128>;
2648 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2649 let Constraints = "$src1 = $dst" in {
2650 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2651 Intrinsic IntId64, Intrinsic IntId128,
2652 bit Commutable = 0> {
2653 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2654 (ins VR64:$src1, VR64:$src2),
2655 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2656 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2657 let isCommutable = Commutable;
2659 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2660 (ins VR64:$src1, i64mem:$src2),
2661 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2663 (IntId64 VR64:$src1,
2664 (bitconvert (memopv8i8 addr:$src2))))]>;
2666 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2667 (ins VR128:$src1, VR128:$src2),
2668 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2669 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2671 let isCommutable = Commutable;
2673 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2674 (ins VR128:$src1, i128mem:$src2),
2675 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2677 (IntId128 VR128:$src1,
2678 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2682 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2683 let Constraints = "$src1 = $dst" in {
2684 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2685 Intrinsic IntId64, Intrinsic IntId128,
2686 bit Commutable = 0> {
2687 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2688 (ins VR64:$src1, VR64:$src2),
2689 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2690 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2691 let isCommutable = Commutable;
2693 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2694 (ins VR64:$src1, i64mem:$src2),
2695 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2697 (IntId64 VR64:$src1,
2698 (bitconvert (memopv4i16 addr:$src2))))]>;
2700 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2701 (ins VR128:$src1, VR128:$src2),
2702 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2703 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2705 let isCommutable = Commutable;
2707 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2708 (ins VR128:$src1, i128mem:$src2),
2709 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2711 (IntId128 VR128:$src1,
2712 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2716 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2717 let Constraints = "$src1 = $dst" in {
2718 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2719 Intrinsic IntId64, Intrinsic IntId128,
2720 bit Commutable = 0> {
2721 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2722 (ins VR64:$src1, VR64:$src2),
2723 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2724 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2725 let isCommutable = Commutable;
2727 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2728 (ins VR64:$src1, i64mem:$src2),
2729 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2731 (IntId64 VR64:$src1,
2732 (bitconvert (memopv2i32 addr:$src2))))]>;
2734 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2735 (ins VR128:$src1, VR128:$src2),
2736 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2737 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2739 let isCommutable = Commutable;
2741 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2742 (ins VR128:$src1, i128mem:$src2),
2743 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2745 (IntId128 VR128:$src1,
2746 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2750 let ImmT = NoImm in { // None of these have i8 immediate fields.
2751 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2752 int_x86_ssse3_phadd_w,
2753 int_x86_ssse3_phadd_w_128>;
2754 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2755 int_x86_ssse3_phadd_d,
2756 int_x86_ssse3_phadd_d_128>;
2757 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2758 int_x86_ssse3_phadd_sw,
2759 int_x86_ssse3_phadd_sw_128>;
2760 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2761 int_x86_ssse3_phsub_w,
2762 int_x86_ssse3_phsub_w_128>;
2763 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2764 int_x86_ssse3_phsub_d,
2765 int_x86_ssse3_phsub_d_128>;
2766 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2767 int_x86_ssse3_phsub_sw,
2768 int_x86_ssse3_phsub_sw_128>;
2769 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2770 int_x86_ssse3_pmadd_ub_sw,
2771 int_x86_ssse3_pmadd_ub_sw_128>;
2772 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2773 int_x86_ssse3_pmul_hr_sw,
2774 int_x86_ssse3_pmul_hr_sw_128, 1>;
2776 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2777 int_x86_ssse3_pshuf_b,
2778 int_x86_ssse3_pshuf_b_128>;
2779 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2780 int_x86_ssse3_psign_b,
2781 int_x86_ssse3_psign_b_128>;
2782 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2783 int_x86_ssse3_psign_w,
2784 int_x86_ssse3_psign_w_128>;
2785 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2786 int_x86_ssse3_psign_d,
2787 int_x86_ssse3_psign_d_128>;
2790 // palignr patterns.
2791 let Constraints = "$src1 = $dst" in {
2792 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2793 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2794 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2796 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2797 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2798 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2801 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2802 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2803 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2805 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2806 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2807 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2811 let AddedComplexity = 5 in {
2813 def : Pat<(v1i64 (palign:$src3 VR64:$src1, VR64:$src2)),
2814 (PALIGNR64rr VR64:$src2, VR64:$src1,
2815 (SHUFFLE_get_palign_imm VR64:$src3))>,
2816 Requires<[HasSSSE3]>;
2817 def : Pat<(v2i32 (palign:$src3 VR64:$src1, VR64:$src2)),
2818 (PALIGNR64rr VR64:$src2, VR64:$src1,
2819 (SHUFFLE_get_palign_imm VR64:$src3))>,
2820 Requires<[HasSSSE3]>;
2821 def : Pat<(v2f32 (palign:$src3 VR64:$src1, VR64:$src2)),
2822 (PALIGNR64rr VR64:$src2, VR64:$src1,
2823 (SHUFFLE_get_palign_imm VR64:$src3))>,
2824 Requires<[HasSSSE3]>;
2825 def : Pat<(v4i16 (palign:$src3 VR64:$src1, VR64:$src2)),
2826 (PALIGNR64rr VR64:$src2, VR64:$src1,
2827 (SHUFFLE_get_palign_imm VR64:$src3))>,
2828 Requires<[HasSSSE3]>;
2829 def : Pat<(v8i8 (palign:$src3 VR64:$src1, VR64:$src2)),
2830 (PALIGNR64rr VR64:$src2, VR64:$src1,
2831 (SHUFFLE_get_palign_imm VR64:$src3))>,
2832 Requires<[HasSSSE3]>;
2834 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2835 (PALIGNR128rr VR128:$src2, VR128:$src1,
2836 (SHUFFLE_get_palign_imm VR128:$src3))>,
2837 Requires<[HasSSSE3]>;
2838 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2839 (PALIGNR128rr VR128:$src2, VR128:$src1,
2840 (SHUFFLE_get_palign_imm VR128:$src3))>,
2841 Requires<[HasSSSE3]>;
2842 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2843 (PALIGNR128rr VR128:$src2, VR128:$src1,
2844 (SHUFFLE_get_palign_imm VR128:$src3))>,
2845 Requires<[HasSSSE3]>;
2846 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2847 (PALIGNR128rr VR128:$src2, VR128:$src1,
2848 (SHUFFLE_get_palign_imm VR128:$src3))>,
2849 Requires<[HasSSSE3]>;
2852 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2853 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2854 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2855 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2857 //===---------------------------------------------------------------------===//
2858 // Non-Instruction Patterns
2859 //===---------------------------------------------------------------------===//
2861 // extload f32 -> f64. This matches load+fextend because we have a hack in
2862 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2864 // Since these loads aren't folded into the fextend, we have to match it
2866 let Predicates = [HasSSE2] in
2867 def : Pat<(fextend (loadf32 addr:$src)),
2868 (CVTSS2SDrm addr:$src)>;
2871 let Predicates = [HasSSE2] in {
2872 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2873 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2874 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2875 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2876 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2877 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2878 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2879 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2880 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2881 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2882 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2883 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2884 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2885 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2886 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2887 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2888 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2889 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2890 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2891 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2892 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2893 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2894 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2895 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2896 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2897 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2898 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2899 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2900 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2901 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2904 // Move scalar to XMM zero-extended
2905 // movd to XMM register zero-extends
2906 let AddedComplexity = 15 in {
2907 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2908 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2909 (MOVSDrr (v2f64 (V_SET0PS)), FR64:$src)>;
2910 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2911 (MOVSSrr (v4f32 (V_SET0PS)), FR32:$src)>;
2912 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2913 (MOVSSrr (v4f32 (V_SET0PS)),
2914 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss)))>;
2915 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2916 (MOVSSrr (v4i32 (V_SET0PI)),
2917 (EXTRACT_SUBREG (v4i32 VR128:$src), sub_ss))>;
2920 // Splat v2f64 / v2i64
2921 let AddedComplexity = 10 in {
2922 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
2923 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2924 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
2925 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2926 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
2927 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2928 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
2929 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2932 // Special unary SHUFPSrri case.
2933 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
2934 (SHUFPSrri VR128:$src1, VR128:$src1,
2935 (SHUFFLE_get_shuf_imm VR128:$src3))>;
2936 let AddedComplexity = 5 in
2937 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
2938 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2939 Requires<[HasSSE2]>;
2940 // Special unary SHUFPDrri case.
2941 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
2942 (SHUFPDrri VR128:$src1, VR128:$src1,
2943 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2944 Requires<[HasSSE2]>;
2945 // Special unary SHUFPDrri case.
2946 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
2947 (SHUFPDrri VR128:$src1, VR128:$src1,
2948 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2949 Requires<[HasSSE2]>;
2950 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2951 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
2952 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2953 Requires<[HasSSE2]>;
2955 // Special binary v4i32 shuffle cases with SHUFPS.
2956 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
2957 (SHUFPSrri VR128:$src1, VR128:$src2,
2958 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2959 Requires<[HasSSE2]>;
2960 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
2961 (SHUFPSrmi VR128:$src1, addr:$src2,
2962 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2963 Requires<[HasSSE2]>;
2964 // Special binary v2i64 shuffle cases using SHUFPDrri.
2965 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
2966 (SHUFPDrri VR128:$src1, VR128:$src2,
2967 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2968 Requires<[HasSSE2]>;
2970 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2971 let AddedComplexity = 15 in {
2972 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
2973 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2974 Requires<[OptForSpeed, HasSSE2]>;
2975 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
2976 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2977 Requires<[OptForSpeed, HasSSE2]>;
2979 let AddedComplexity = 10 in {
2980 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
2981 (UNPCKLPSrr VR128:$src, VR128:$src)>;
2982 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
2983 (PUNPCKLBWrr VR128:$src, VR128:$src)>;
2984 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
2985 (PUNPCKLWDrr VR128:$src, VR128:$src)>;
2986 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
2987 (PUNPCKLDQrr VR128:$src, VR128:$src)>;
2990 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2991 let AddedComplexity = 15 in {
2992 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
2993 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2994 Requires<[OptForSpeed, HasSSE2]>;
2995 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
2996 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2997 Requires<[OptForSpeed, HasSSE2]>;
2999 let AddedComplexity = 10 in {
3000 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
3001 (UNPCKHPSrr VR128:$src, VR128:$src)>;
3002 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
3003 (PUNPCKHBWrr VR128:$src, VR128:$src)>;
3004 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
3005 (PUNPCKHWDrr VR128:$src, VR128:$src)>;
3006 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
3007 (PUNPCKHDQrr VR128:$src, VR128:$src)>;
3010 let AddedComplexity = 20 in {
3011 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3012 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3013 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3015 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3016 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3017 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3019 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3020 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3021 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3022 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3023 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3026 let AddedComplexity = 20 in {
3027 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3028 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3029 (MOVLPSrm VR128:$src1, addr:$src2)>;
3030 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3031 (MOVLPDrm VR128:$src1, addr:$src2)>;
3032 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3033 (MOVLPSrm VR128:$src1, addr:$src2)>;
3034 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3035 (MOVLPDrm VR128:$src1, addr:$src2)>;
3038 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3039 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3040 (MOVLPSmr addr:$src1, VR128:$src2)>;
3041 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3042 (MOVLPDmr addr:$src1, VR128:$src2)>;
3043 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3045 (MOVLPSmr addr:$src1, VR128:$src2)>;
3046 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3047 (MOVLPDmr addr:$src1, VR128:$src2)>;
3049 let AddedComplexity = 15 in {
3050 // Setting the lowest element in the vector.
3051 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3052 (MOVSSrr (v4i32 VR128:$src1),
3053 (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_ss))>;
3054 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3055 (MOVSDrr (v2i64 VR128:$src1),
3056 (EXTRACT_SUBREG (v2i64 VR128:$src2), sub_sd))>;
3058 // vector_shuffle v1, v2 <4, 5, 2, 3> using movsd
3059 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3060 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3061 Requires<[HasSSE2]>;
3062 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3063 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3064 Requires<[HasSSE2]>;
3067 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3068 // fall back to this for SSE1)
3069 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3070 (SHUFPSrri VR128:$src2, VR128:$src1,
3071 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3073 // Set lowest element and zero upper elements.
3074 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3075 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3077 // Some special case pandn patterns.
3078 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3080 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3081 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3083 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3084 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3086 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3088 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3089 (memop addr:$src2))),
3090 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3091 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3092 (memop addr:$src2))),
3093 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3094 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3095 (memop addr:$src2))),
3096 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3098 // vector -> vector casts
3099 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3100 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3101 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3102 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3103 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3104 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3105 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3106 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3108 // Use movaps / movups for SSE integer load / store (one byte shorter).
3109 def : Pat<(alignedloadv4i32 addr:$src),
3110 (MOVAPSrm addr:$src)>;
3111 def : Pat<(loadv4i32 addr:$src),
3112 (MOVUPSrm addr:$src)>;
3113 def : Pat<(alignedloadv2i64 addr:$src),
3114 (MOVAPSrm addr:$src)>;
3115 def : Pat<(loadv2i64 addr:$src),
3116 (MOVUPSrm addr:$src)>;
3118 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3119 (MOVAPSmr addr:$dst, VR128:$src)>;
3120 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3121 (MOVAPSmr addr:$dst, VR128:$src)>;
3122 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3123 (MOVAPSmr addr:$dst, VR128:$src)>;
3124 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3125 (MOVAPSmr addr:$dst, VR128:$src)>;
3126 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3127 (MOVUPSmr addr:$dst, VR128:$src)>;
3128 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3129 (MOVUPSmr addr:$dst, VR128:$src)>;
3130 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3131 (MOVUPSmr addr:$dst, VR128:$src)>;
3132 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3133 (MOVUPSmr addr:$dst, VR128:$src)>;
3135 //===----------------------------------------------------------------------===//
3136 // SSE4.1 Instructions
3137 //===----------------------------------------------------------------------===//
3139 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3142 Intrinsic V2F64Int> {
3143 // Intrinsic operation, reg.
3144 // Vector intrinsic operation, reg
3145 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3146 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3147 !strconcat(OpcodeStr,
3148 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3149 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3152 // Vector intrinsic operation, mem
3153 def PSm_Int : Ii8<opcps, MRMSrcMem,
3154 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3155 !strconcat(OpcodeStr,
3156 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3158 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3160 Requires<[HasSSE41]>;
3162 // Vector intrinsic operation, reg
3163 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3164 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3165 !strconcat(OpcodeStr,
3166 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3167 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3170 // Vector intrinsic operation, mem
3171 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3172 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3173 !strconcat(OpcodeStr,
3174 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3176 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3180 let Constraints = "$src1 = $dst" in {
3181 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3185 // Intrinsic operation, reg.
3186 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3188 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3189 !strconcat(OpcodeStr,
3190 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3192 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3195 // Intrinsic operation, mem.
3196 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3198 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3199 !strconcat(OpcodeStr,
3200 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3202 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3205 // Intrinsic operation, reg.
3206 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3208 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3209 !strconcat(OpcodeStr,
3210 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3212 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3215 // Intrinsic operation, mem.
3216 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3218 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3219 !strconcat(OpcodeStr,
3220 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3222 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3227 // FP round - roundss, roundps, roundsd, roundpd
3228 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3229 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3230 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3231 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3233 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3234 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3235 Intrinsic IntId128> {
3236 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3238 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3239 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3240 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3242 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3245 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3248 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3249 int_x86_sse41_phminposuw>;
3251 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3252 let Constraints = "$src1 = $dst" in {
3253 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3254 Intrinsic IntId128, bit Commutable = 0> {
3255 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3256 (ins VR128:$src1, VR128:$src2),
3257 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3258 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3260 let isCommutable = Commutable;
3262 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3263 (ins VR128:$src1, i128mem:$src2),
3264 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3266 (IntId128 VR128:$src1,
3267 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3271 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3272 int_x86_sse41_pcmpeqq, 1>;
3273 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3274 int_x86_sse41_packusdw, 0>;
3275 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3276 int_x86_sse41_pminsb, 1>;
3277 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3278 int_x86_sse41_pminsd, 1>;
3279 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3280 int_x86_sse41_pminud, 1>;
3281 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3282 int_x86_sse41_pminuw, 1>;
3283 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3284 int_x86_sse41_pmaxsb, 1>;
3285 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3286 int_x86_sse41_pmaxsd, 1>;
3287 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3288 int_x86_sse41_pmaxud, 1>;
3289 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3290 int_x86_sse41_pmaxuw, 1>;
3292 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3294 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3295 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3296 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3297 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3299 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3300 let Constraints = "$src1 = $dst" in {
3301 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3302 SDNode OpNode, Intrinsic IntId128,
3303 bit Commutable = 0> {
3304 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3305 (ins VR128:$src1, VR128:$src2),
3306 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3307 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3308 VR128:$src2))]>, OpSize {
3309 let isCommutable = Commutable;
3311 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3312 (ins VR128:$src1, VR128:$src2),
3313 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3314 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3316 let isCommutable = Commutable;
3318 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3319 (ins VR128:$src1, i128mem:$src2),
3320 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3322 (OpVT (OpNode VR128:$src1, (memop addr:$src2))))]>, OpSize;
3323 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3324 (ins VR128:$src1, i128mem:$src2),
3325 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3327 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3332 /// SS48I_binop_rm - Simple SSE41 binary operator.
3333 let Constraints = "$src1 = $dst" in {
3334 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
3335 ValueType OpVT, bit Commutable = 0> {
3336 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3337 (ins VR128:$src1, VR128:$src2),
3338 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3339 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]>,
3341 let isCommutable = Commutable;
3343 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3344 (ins VR128:$src1, i128mem:$src2),
3345 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3346 [(set VR128:$dst, (OpNode VR128:$src1,
3347 (bc_v4i32 (memopv2i64 addr:$src2))))]>,
3352 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, 1>;
3354 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3355 let Constraints = "$src1 = $dst" in {
3356 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3357 Intrinsic IntId128, bit Commutable = 0> {
3358 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3359 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3360 !strconcat(OpcodeStr,
3361 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3363 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3365 let isCommutable = Commutable;
3367 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3368 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3369 !strconcat(OpcodeStr,
3370 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3372 (IntId128 VR128:$src1,
3373 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3378 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3379 int_x86_sse41_blendps, 0>;
3380 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3381 int_x86_sse41_blendpd, 0>;
3382 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3383 int_x86_sse41_pblendw, 0>;
3384 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3385 int_x86_sse41_dpps, 1>;
3386 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3387 int_x86_sse41_dppd, 1>;
3388 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3389 int_x86_sse41_mpsadbw, 0>;
3392 /// SS41I_ternary_int - SSE 4.1 ternary operator
3393 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3394 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3395 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3396 (ins VR128:$src1, VR128:$src2),
3397 !strconcat(OpcodeStr,
3398 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3399 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3402 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3403 (ins VR128:$src1, i128mem:$src2),
3404 !strconcat(OpcodeStr,
3405 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3408 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3412 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3413 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3414 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3417 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3418 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3419 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3420 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3422 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3423 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3425 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3429 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3430 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3431 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3432 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3433 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3434 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3436 // Common patterns involving scalar load.
3437 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3438 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3439 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3440 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3442 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3443 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3444 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3445 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3447 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3448 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3449 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3450 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3452 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3453 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3454 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3455 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3457 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3458 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3459 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3460 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3462 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3463 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3464 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3465 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3468 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3469 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3470 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3471 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3473 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3474 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3476 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3480 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3481 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3482 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3483 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3485 // Common patterns involving scalar load
3486 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3487 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3488 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3489 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3491 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3492 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3493 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3494 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3497 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3498 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3499 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3500 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3502 // Expecting a i16 load any extended to i32 value.
3503 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3504 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3505 [(set VR128:$dst, (IntId (bitconvert
3506 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3510 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3511 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3513 // Common patterns involving scalar load
3514 def : Pat<(int_x86_sse41_pmovsxbq
3515 (bitconvert (v4i32 (X86vzmovl
3516 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3517 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3519 def : Pat<(int_x86_sse41_pmovzxbq
3520 (bitconvert (v4i32 (X86vzmovl
3521 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3522 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3525 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3526 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3527 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3528 (ins VR128:$src1, i32i8imm:$src2),
3529 !strconcat(OpcodeStr,
3530 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3531 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3533 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3534 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3535 !strconcat(OpcodeStr,
3536 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3539 // There's an AssertZext in the way of writing the store pattern
3540 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3543 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3546 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3547 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3548 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3549 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3550 !strconcat(OpcodeStr,
3551 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3554 // There's an AssertZext in the way of writing the store pattern
3555 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3558 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3561 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3562 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3563 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3564 (ins VR128:$src1, i32i8imm:$src2),
3565 !strconcat(OpcodeStr,
3566 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3568 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3569 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3570 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3571 !strconcat(OpcodeStr,
3572 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3573 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3574 addr:$dst)]>, OpSize;
3577 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3580 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3582 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3583 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3584 (ins VR128:$src1, i32i8imm:$src2),
3585 !strconcat(OpcodeStr,
3586 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3588 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3590 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3591 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3592 !strconcat(OpcodeStr,
3593 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3594 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3595 addr:$dst)]>, OpSize;
3598 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3600 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3601 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3604 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3605 Requires<[HasSSE41]>;
3607 let Constraints = "$src1 = $dst" in {
3608 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3609 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3610 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3611 !strconcat(OpcodeStr,
3612 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3614 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3615 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3616 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3617 !strconcat(OpcodeStr,
3618 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3620 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3621 imm:$src3))]>, OpSize;
3625 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3627 let Constraints = "$src1 = $dst" in {
3628 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3629 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3630 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3631 !strconcat(OpcodeStr,
3632 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3634 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3636 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3637 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3638 !strconcat(OpcodeStr,
3639 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3641 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3642 imm:$src3)))]>, OpSize;
3646 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3648 // insertps has a few different modes, there's the first two here below which
3649 // are optimized inserts that won't zero arbitrary elements in the destination
3650 // vector. The next one matches the intrinsic and could zero arbitrary elements
3651 // in the target vector.
3652 let Constraints = "$src1 = $dst" in {
3653 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3654 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3655 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3656 !strconcat(OpcodeStr,
3657 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3659 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3661 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3662 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3663 !strconcat(OpcodeStr,
3664 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3666 (X86insrtps VR128:$src1,
3667 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3668 imm:$src3))]>, OpSize;
3672 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3674 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3675 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3677 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3678 // the intel intrinsic that corresponds to this.
3679 let Defs = [EFLAGS] in {
3680 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3681 "ptest \t{$src2, $src1|$src1, $src2}",
3682 [(set EFLAGS, (X86ptest VR128:$src1, VR128:$src2))]>,
3684 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3685 "ptest \t{$src2, $src1|$src1, $src2}",
3686 [(set EFLAGS, (X86ptest VR128:$src1, (load addr:$src2)))]>,
3690 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3691 "movntdqa\t{$src, $dst|$dst, $src}",
3692 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
3696 //===----------------------------------------------------------------------===//
3697 // SSE4.2 Instructions
3698 //===----------------------------------------------------------------------===//
3700 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3701 let Constraints = "$src1 = $dst" in {
3702 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3703 Intrinsic IntId128, bit Commutable = 0> {
3704 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3705 (ins VR128:$src1, VR128:$src2),
3706 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3707 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3709 let isCommutable = Commutable;
3711 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3712 (ins VR128:$src1, i128mem:$src2),
3713 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3715 (IntId128 VR128:$src1,
3716 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3720 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3722 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3723 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3724 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3725 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3727 // crc intrinsic instruction
3728 // This set of instructions are only rm, the only difference is the size
3730 let Constraints = "$src1 = $dst" in {
3731 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3732 (ins GR32:$src1, i8mem:$src2),
3733 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3735 (int_x86_sse42_crc32_8 GR32:$src1,
3736 (load addr:$src2)))]>;
3737 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3738 (ins GR32:$src1, GR8:$src2),
3739 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3741 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>;
3742 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3743 (ins GR32:$src1, i16mem:$src2),
3744 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3746 (int_x86_sse42_crc32_16 GR32:$src1,
3747 (load addr:$src2)))]>,
3749 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3750 (ins GR32:$src1, GR16:$src2),
3751 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3753 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3755 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3756 (ins GR32:$src1, i32mem:$src2),
3757 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3759 (int_x86_sse42_crc32_32 GR32:$src1,
3760 (load addr:$src2)))]>;
3761 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3762 (ins GR32:$src1, GR32:$src2),
3763 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3765 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>;
3766 def CRC64m8 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3767 (ins GR64:$src1, i8mem:$src2),
3768 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3770 (int_x86_sse42_crc64_8 GR64:$src1,
3771 (load addr:$src2)))]>,
3773 def CRC64r8 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3774 (ins GR64:$src1, GR8:$src2),
3775 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3777 (int_x86_sse42_crc64_8 GR64:$src1, GR8:$src2))]>,
3779 def CRC64m64 : SS42FI<0xF1, MRMSrcMem, (outs GR64:$dst),
3780 (ins GR64:$src1, i64mem:$src2),
3781 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3783 (int_x86_sse42_crc64_64 GR64:$src1,
3784 (load addr:$src2)))]>,
3786 def CRC64r64 : SS42FI<0xF1, MRMSrcReg, (outs GR64:$dst),
3787 (ins GR64:$src1, GR64:$src2),
3788 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3790 (int_x86_sse42_crc64_64 GR64:$src1, GR64:$src2))]>,
3794 // String/text processing instructions.
3795 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3796 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3797 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3798 "#PCMPISTRM128rr PSEUDO!",
3799 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3800 imm:$src3))]>, OpSize;
3801 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3802 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3803 "#PCMPISTRM128rm PSEUDO!",
3804 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3805 imm:$src3))]>, OpSize;
3808 let Defs = [XMM0, EFLAGS] in {
3809 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3810 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3811 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3812 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3813 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3814 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3817 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3818 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3819 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3820 "#PCMPESTRM128rr PSEUDO!",
3822 (int_x86_sse42_pcmpestrm128
3823 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3825 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3826 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3827 "#PCMPESTRM128rm PSEUDO!",
3828 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3829 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3833 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3834 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3835 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3836 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3837 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3838 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3839 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3842 let Defs = [ECX, EFLAGS] in {
3843 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3844 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3845 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3846 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3847 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3848 (implicit EFLAGS)]>, OpSize;
3849 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3850 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3851 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3852 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3853 (implicit EFLAGS)]>, OpSize;
3857 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3858 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3859 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3860 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3861 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3862 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3864 let Defs = [ECX, EFLAGS] in {
3865 let Uses = [EAX, EDX] in {
3866 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3867 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3868 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3869 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3870 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3871 (implicit EFLAGS)]>, OpSize;
3872 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3873 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3874 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3876 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
3877 (implicit EFLAGS)]>, OpSize;
3882 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3883 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3884 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3885 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3886 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3887 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;
3889 //===----------------------------------------------------------------------===//
3890 // AES-NI Instructions
3891 //===----------------------------------------------------------------------===//
3893 let Constraints = "$src1 = $dst" in {
3894 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr,
3895 Intrinsic IntId128, bit Commutable = 0> {
3896 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
3897 (ins VR128:$src1, VR128:$src2),
3898 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3899 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3901 let isCommutable = Commutable;
3903 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
3904 (ins VR128:$src1, i128mem:$src2),
3905 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3907 (IntId128 VR128:$src1,
3908 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3912 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
3913 int_x86_aesni_aesenc>;
3914 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
3915 int_x86_aesni_aesenclast>;
3916 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
3917 int_x86_aesni_aesdec>;
3918 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
3919 int_x86_aesni_aesdeclast>;
3921 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, VR128:$src2)),
3922 (AESENCrr VR128:$src1, VR128:$src2)>;
3923 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, (memop addr:$src2))),
3924 (AESENCrm VR128:$src1, addr:$src2)>;
3925 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, VR128:$src2)),
3926 (AESENCLASTrr VR128:$src1, VR128:$src2)>;
3927 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, (memop addr:$src2))),
3928 (AESENCLASTrm VR128:$src1, addr:$src2)>;
3929 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, VR128:$src2)),
3930 (AESDECrr VR128:$src1, VR128:$src2)>;
3931 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, (memop addr:$src2))),
3932 (AESDECrm VR128:$src1, addr:$src2)>;
3933 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, VR128:$src2)),
3934 (AESDECLASTrr VR128:$src1, VR128:$src2)>;
3935 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, (memop addr:$src2))),
3936 (AESDECLASTrm VR128:$src1, addr:$src2)>;
3938 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
3940 "aesimc\t{$src1, $dst|$dst, $src1}",
3942 (int_x86_aesni_aesimc VR128:$src1))]>,
3945 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
3946 (ins i128mem:$src1),
3947 "aesimc\t{$src1, $dst|$dst, $src1}",
3949 (int_x86_aesni_aesimc (bitconvert (memopv2i64 addr:$src1))))]>,
3952 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
3953 (ins VR128:$src1, i8imm:$src2),
3954 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3956 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
3958 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
3959 (ins i128mem:$src1, i8imm:$src2),
3960 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3962 (int_x86_aesni_aeskeygenassist (bitconvert (memopv2i64 addr:$src1)),