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:$src2)))],d>;
691 /// sse12_fp_packed_int - SSE 1 & 2 packed instructions intrinsics class
692 multiclass sse12_fp_packed_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
693 string asm, string SSEVer, string FPSizeStr,
694 X86MemOperand memop, PatFrag mem_frag,
696 def rr_Int : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
697 asm, [(set RC:$dst, (
698 !nameconcat<Intrinsic>("int_x86_sse",
699 !strconcat(SSEVer, !strconcat("_",
700 !strconcat(OpcodeStr, FPSizeStr))))
701 RC:$src1, RC:$src2))], d>;
702 def rm_Int : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memop:$src2),
703 asm, [(set RC:$dst, (
704 !nameconcat<Intrinsic>("int_x86_sse",
705 !strconcat(SSEVer, !strconcat("_",
706 !strconcat(OpcodeStr, FPSizeStr))))
707 RC:$src1, (mem_frag addr:$src2)))], d>;
710 /// basic_sse12_fp_binop_rm - SSE 1 & 2 binops come in both scalar and
713 /// In addition, we also have a special variant of the scalar form here to
714 /// represent the associated intrinsic operation. This form is unlike the
715 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
716 /// and leaves the top elements unmodified (therefore these cannot be commuted).
718 /// These three forms can each be reg+reg or reg+mem, so there are a total of
719 /// six "instructions".
721 multiclass basic_sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
724 let isAsmParserOnly = 1 in {
725 defm V#NAME#SS : sse12_fp_scalar<opc,
726 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
727 OpNode, FR32, f32mem>, XS, VEX_4V;
729 defm V#NAME#SD : sse12_fp_scalar<opc,
730 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
731 OpNode, FR64, f64mem>, XD, VEX_4V;
733 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
734 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
735 VR128, v4f32, f128mem, memopv4f32, SSEPackedSingle>,
738 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
739 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
740 VR128, v2f64, f128mem, memopv2f64, SSEPackedDouble>,
743 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
744 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
745 "", "_ss", ssmem, sse_load_f32>, XS, VEX_4V;
747 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
748 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
749 "2", "_sd", sdmem, sse_load_f64>, XD, VEX_4V;
752 let Constraints = "$src1 = $dst" in {
753 defm SS : sse12_fp_scalar<opc,
754 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
755 OpNode, FR32, f32mem>, XS;
757 defm SD : sse12_fp_scalar<opc,
758 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
759 OpNode, FR64, f64mem>, XD;
761 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
762 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v4f32,
763 f128mem, memopv4f32, SSEPackedSingle>, TB;
765 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
766 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v2f64,
767 f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
769 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
770 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
771 "", "_ss", ssmem, sse_load_f32>, XS;
773 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
774 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
775 "2", "_sd", sdmem, sse_load_f64>, XD;
779 // Arithmetic instructions
780 defm ADD : basic_sse12_fp_binop_rm<0x58, "add", fadd>;
781 defm MUL : basic_sse12_fp_binop_rm<0x59, "mul", fmul>;
783 let isCommutable = 0 in {
784 defm SUB : basic_sse12_fp_binop_rm<0x5C, "sub", fsub>;
785 defm DIV : basic_sse12_fp_binop_rm<0x5E, "div", fdiv>;
788 /// sse12_fp_binop_rm - Other SSE 1 & 2 binops
790 /// This multiclass is like basic_sse12_fp_binop_rm, with the addition of
791 /// instructions for a full-vector intrinsic form. Operations that map
792 /// onto C operators don't use this form since they just use the plain
793 /// vector form instead of having a separate vector intrinsic form.
795 /// This provides a total of eight "instructions".
797 multiclass sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
800 let isAsmParserOnly = 1 in {
801 // Scalar operation, reg+reg.
802 defm V#NAME#SS : sse12_fp_scalar<opc,
803 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
804 OpNode, FR32, f32mem>, XS, VEX_4V;
806 defm V#NAME#SD : sse12_fp_scalar<opc,
807 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
808 OpNode, FR64, f64mem>, XD, VEX_4V;
810 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
811 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
812 VR128, v4f32, f128mem, memopv4f32, SSEPackedSingle>,
815 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
816 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
817 VR128, v2f64, f128mem, memopv2f64, SSEPackedDouble>,
820 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
821 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
822 "", "_ss", ssmem, sse_load_f32>, XS, VEX_4V;
824 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
825 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
826 "2", "_sd", sdmem, sse_load_f64>, XD, VEX_4V;
828 defm V#NAME#PS : sse12_fp_packed_int<opc, OpcodeStr, VR128,
829 !strconcat(OpcodeStr, "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
830 "", "_ps", f128mem, memopv4f32, SSEPackedSingle>, VEX_4V;
832 defm V#NAME#PD : sse12_fp_packed_int<opc, OpcodeStr, VR128,
833 !strconcat(OpcodeStr, "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
834 "2", "_pd", f128mem, memopv2f64, SSEPackedDouble>, OpSize,
838 let Constraints = "$src1 = $dst" in {
839 // Scalar operation, reg+reg.
840 defm SS : sse12_fp_scalar<opc,
841 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
842 OpNode, FR32, f32mem>, XS;
843 defm SD : sse12_fp_scalar<opc,
844 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
845 OpNode, FR64, f64mem>, XD;
846 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
847 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v4f32,
848 f128mem, memopv4f32, SSEPackedSingle>, TB;
850 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
851 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v2f64,
852 f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
854 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
855 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
856 "", "_ss", ssmem, sse_load_f32>, XS;
858 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
859 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
860 "2", "_sd", sdmem, sse_load_f64>, XD;
862 defm PS : sse12_fp_packed_int<opc, OpcodeStr, VR128,
863 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
864 "", "_ps", f128mem, memopv4f32, SSEPackedSingle>, TB;
866 defm PD : sse12_fp_packed_int<opc, OpcodeStr, VR128,
867 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
868 "2", "_pd", f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
872 let isCommutable = 0 in {
873 defm MAX : sse12_fp_binop_rm<0x5F, "max", X86fmax>;
874 defm MIN : sse12_fp_binop_rm<0x5D, "min", X86fmin>;
877 //===----------------------------------------------------------------------===//
878 // SSE packed FP Instructions
881 let neverHasSideEffects = 1 in
882 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
883 "movaps\t{$src, $dst|$dst, $src}", []>;
884 let canFoldAsLoad = 1, isReMaterializable = 1 in
885 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
886 "movaps\t{$src, $dst|$dst, $src}",
887 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
889 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
890 "movaps\t{$src, $dst|$dst, $src}",
891 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
893 let neverHasSideEffects = 1 in
894 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
895 "movups\t{$src, $dst|$dst, $src}", []>;
896 let canFoldAsLoad = 1, isReMaterializable = 1 in
897 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
898 "movups\t{$src, $dst|$dst, $src}",
899 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
900 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
901 "movups\t{$src, $dst|$dst, $src}",
902 [(store (v4f32 VR128:$src), addr:$dst)]>;
904 // Intrinsic forms of MOVUPS load and store
905 let canFoldAsLoad = 1, isReMaterializable = 1 in
906 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
907 "movups\t{$src, $dst|$dst, $src}",
908 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
909 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
910 "movups\t{$src, $dst|$dst, $src}",
911 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
913 let Constraints = "$src1 = $dst" in {
914 let AddedComplexity = 20 in {
915 def MOVLPSrm : PSI<0x12, MRMSrcMem,
916 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
917 "movlps\t{$src2, $dst|$dst, $src2}",
920 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
921 def MOVHPSrm : PSI<0x16, MRMSrcMem,
922 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
923 "movhps\t{$src2, $dst|$dst, $src2}",
925 (movlhps VR128:$src1,
926 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
928 } // Constraints = "$src1 = $dst"
931 def : Pat<(movlhps VR128:$src1, (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
932 (MOVHPSrm (v4i32 VR128:$src1), addr:$src2)>;
934 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
935 "movlps\t{$src, $dst|$dst, $src}",
936 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
937 (iPTR 0))), addr:$dst)]>;
939 // v2f64 extract element 1 is always custom lowered to unpack high to low
940 // and extract element 0 so the non-store version isn't too horrible.
941 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
942 "movhps\t{$src, $dst|$dst, $src}",
943 [(store (f64 (vector_extract
944 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
945 (undef)), (iPTR 0))), addr:$dst)]>;
947 let Constraints = "$src1 = $dst" in {
948 let AddedComplexity = 20 in {
949 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
950 (ins VR128:$src1, VR128:$src2),
951 "movlhps\t{$src2, $dst|$dst, $src2}",
953 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
955 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
956 (ins VR128:$src1, VR128:$src2),
957 "movhlps\t{$src2, $dst|$dst, $src2}",
959 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
961 } // Constraints = "$src1 = $dst"
963 let AddedComplexity = 20 in {
964 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
965 (MOVLHPSrr (v4f32 VR128:$src), (v4f32 VR128:$src))>;
966 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
967 (MOVLHPSrr (v2i64 VR128:$src), (v2i64 VR128:$src))>;
974 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
976 /// In addition, we also have a special variant of the scalar form here to
977 /// represent the associated intrinsic operation. This form is unlike the
978 /// plain scalar form, in that it takes an entire vector (instead of a
979 /// scalar) and leaves the top elements undefined.
981 /// And, we have a special variant form for a full-vector intrinsic form.
983 /// These four forms can each have a reg or a mem operand, so there are a
984 /// total of eight "instructions".
986 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
990 bit Commutable = 0> {
991 // Scalar operation, reg.
992 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
993 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
994 [(set FR32:$dst, (OpNode FR32:$src))]> {
995 let isCommutable = Commutable;
998 // Scalar operation, mem.
999 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
1000 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1001 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
1002 Requires<[HasSSE1, OptForSize]>;
1004 // Vector operation, reg.
1005 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1006 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1007 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
1008 let isCommutable = Commutable;
1011 // Vector operation, mem.
1012 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1013 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1014 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
1016 // Intrinsic operation, reg.
1017 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1018 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1019 [(set VR128:$dst, (F32Int VR128:$src))]> {
1020 let isCommutable = Commutable;
1023 // Intrinsic operation, mem.
1024 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
1025 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1026 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
1028 // Vector intrinsic operation, reg
1029 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1030 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1031 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
1032 let isCommutable = Commutable;
1035 // Vector intrinsic operation, mem
1036 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1037 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1038 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
1042 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
1043 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
1045 // Reciprocal approximations. Note that these typically require refinement
1046 // in order to obtain suitable precision.
1047 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
1048 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
1049 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
1050 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
1052 /// sse12_fp_pack_logical - SSE 1 & 2 packed FP logical ops
1054 multiclass sse12_fp_pack_logical<bits<8> opc, string OpcodeStr,
1055 SDNode OpNode, int HasPat = 0,
1057 list<list<dag>> Pattern = []> {
1058 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
1059 (ins VR128:$src1, VR128:$src2),
1060 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
1061 !if(HasPat, Pattern[0],
1062 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1,
1064 { let isCommutable = Commutable; }
1066 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1067 (ins VR128:$src1, VR128:$src2),
1068 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1069 !if(HasPat, Pattern[1],
1070 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1071 (bc_v2i64 (v2f64 VR128:$src2))))])>
1072 { let isCommutable = Commutable; }
1074 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
1075 (ins VR128:$src1, f128mem:$src2),
1076 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
1077 !if(HasPat, Pattern[2],
1078 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
1079 (memopv2i64 addr:$src2)))])>;
1081 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1082 (ins VR128:$src1, f128mem:$src2),
1083 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1084 !if(HasPat, Pattern[3],
1085 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1086 (memopv2i64 addr:$src2)))])>;
1090 let Constraints = "$src1 = $dst" in {
1091 defm AND : sse12_fp_pack_logical<0x54, "and", and>;
1092 defm OR : sse12_fp_pack_logical<0x56, "or", or>;
1093 defm XOR : sse12_fp_pack_logical<0x57, "xor", xor>;
1094 defm ANDN : sse12_fp_pack_logical<0x55, "andn", undef /* dummy */, 1, 0, [
1096 [(set VR128:$dst, (v2i64 (and (xor VR128:$src1,
1097 (bc_v2i64 (v4i32 immAllOnesV))),
1100 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1101 (bc_v2i64 (v2f64 VR128:$src2))))],
1103 [(set VR128:$dst, (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
1104 (bc_v2i64 (v4i32 immAllOnesV))),
1105 (memopv2i64 addr:$src2))))],
1107 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1108 (memopv2i64 addr:$src2)))]]>;
1111 let Constraints = "$src1 = $dst" in {
1112 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
1113 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1114 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1115 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1116 VR128:$src, imm:$cc))]>;
1117 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
1118 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1119 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1120 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1121 (memop addr:$src), imm:$cc))]>;
1123 // Accept explicit immediate argument form instead of comparison code.
1124 let isAsmParserOnly = 1 in {
1125 def CMPPSrri_alt : PSIi8<0xC2, MRMSrcReg,
1126 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1127 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1128 def CMPPSrmi_alt : PSIi8<0xC2, MRMSrcMem,
1129 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1130 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1133 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1134 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
1135 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1136 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
1138 // Shuffle and unpack instructions
1139 let Constraints = "$src1 = $dst" in {
1140 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
1141 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
1142 (outs VR128:$dst), (ins VR128:$src1,
1143 VR128:$src2, i8imm:$src3),
1144 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1146 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1147 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
1148 (outs VR128:$dst), (ins VR128:$src1,
1149 f128mem:$src2, i8imm:$src3),
1150 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1153 VR128:$src1, (memopv4f32 addr:$src2))))]>;
1155 let AddedComplexity = 10 in {
1156 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
1157 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1158 "unpckhps\t{$src2, $dst|$dst, $src2}",
1160 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
1161 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
1162 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1163 "unpckhps\t{$src2, $dst|$dst, $src2}",
1165 (v4f32 (unpckh VR128:$src1,
1166 (memopv4f32 addr:$src2))))]>;
1168 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
1169 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1170 "unpcklps\t{$src2, $dst|$dst, $src2}",
1172 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
1173 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
1174 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1175 "unpcklps\t{$src2, $dst|$dst, $src2}",
1177 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
1178 } // AddedComplexity
1179 } // Constraints = "$src1 = $dst"
1182 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1183 "movmskps\t{$src, $dst|$dst, $src}",
1184 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
1185 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1186 "movmskpd\t{$src, $dst|$dst, $src}",
1187 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1189 // Prefetch intrinsic.
1190 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1191 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1192 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1193 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1194 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1195 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1196 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1197 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1199 // Non-temporal stores
1200 def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1201 "movntps\t{$src, $dst|$dst, $src}",
1202 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1204 let AddedComplexity = 400 in { // Prefer non-temporal versions
1205 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1206 "movntps\t{$src, $dst|$dst, $src}",
1207 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
1209 def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1210 "movntdq\t{$src, $dst|$dst, $src}",
1211 [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
1213 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
1214 "movnti\t{$src, $dst|$dst, $src}",
1215 [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
1216 TB, Requires<[HasSSE2]>;
1218 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
1219 "movnti\t{$src, $dst|$dst, $src}",
1220 [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
1221 TB, Requires<[HasSSE2]>;
1224 // Load, store, and memory fence
1225 def SFENCE : I<0xAE, MRM_F8, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>,
1226 TB, Requires<[HasSSE1]>;
1229 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1230 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1231 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1232 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1234 // Alias instructions that map zero vector to pxor / xorp* for sse.
1235 // We set canFoldAsLoad because this can be converted to a constant-pool
1236 // load of an all-zeros value if folding it would be beneficial.
1237 // FIXME: Change encoding to pseudo!
1238 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1239 isCodeGenOnly = 1 in {
1240 def V_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1241 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
1242 def V_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1243 [(set VR128:$dst, (v2f64 immAllZerosV))]>;
1244 let ExeDomain = SSEPackedInt in
1245 def V_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
1246 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1249 def : Pat<(v2i64 immAllZerosV), (V_SET0PI)>;
1250 def : Pat<(v8i16 immAllZerosV), (V_SET0PI)>;
1251 def : Pat<(v16i8 immAllZerosV), (V_SET0PI)>;
1253 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1254 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
1256 //===---------------------------------------------------------------------===//
1257 // SSE2 Instructions
1258 //===---------------------------------------------------------------------===//
1260 // Move Instructions. Register-to-register movsd is not used for FR64
1261 // register copies because it's a partial register update; FsMOVAPDrr is
1262 // used instead. Register-to-register movsd is not modeled as an INSERT_SUBREG
1263 // because INSERT_SUBREG requires that the insert be implementable in terms of
1264 // a copy, and just mentioned, we don't use movsd for copies.
1265 let Constraints = "$src1 = $dst" in
1266 def MOVSDrr : SDI<0x10, MRMSrcReg,
1267 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
1268 "movsd\t{$src2, $dst|$dst, $src2}",
1269 [(set (v2f64 VR128:$dst),
1270 (movl VR128:$src1, (scalar_to_vector FR64:$src2)))]>;
1272 // Extract the low 64-bit value from one vector and insert it into another.
1273 let AddedComplexity = 15 in
1274 def : Pat<(v2f64 (movl VR128:$src1, VR128:$src2)),
1275 (MOVSDrr (v2f64 VR128:$src1),
1276 (EXTRACT_SUBREG (v2f64 VR128:$src2), sub_sd))>;
1278 // Implicitly promote a 64-bit scalar to a vector.
1279 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
1280 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), FR64:$src, sub_sd)>;
1282 // Loading from memory automatically zeroing upper bits.
1283 let canFoldAsLoad = 1, isReMaterializable = 1, AddedComplexity = 20 in
1284 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1285 "movsd\t{$src, $dst|$dst, $src}",
1286 [(set FR64:$dst, (loadf64 addr:$src))]>;
1288 // MOVSDrm zeros the high parts of the register; represent this
1289 // with SUBREG_TO_REG.
1290 let AddedComplexity = 20 in {
1291 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
1292 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1293 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
1294 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1295 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
1296 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1297 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
1298 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1299 def : Pat<(v2f64 (X86vzload addr:$src)),
1300 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1303 // Store scalar value to memory.
1304 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1305 "movsd\t{$src, $dst|$dst, $src}",
1306 [(store FR64:$src, addr:$dst)]>;
1308 // Extract and store.
1309 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
1312 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
1314 // Conversion instructions
1315 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1316 "cvttsd2si\t{$src, $dst|$dst, $src}",
1317 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1318 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1319 "cvttsd2si\t{$src, $dst|$dst, $src}",
1320 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1321 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1322 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1323 [(set FR32:$dst, (fround FR64:$src))]>;
1324 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1325 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1326 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
1327 Requires<[HasSSE2, OptForSize]>;
1328 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1329 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1330 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1331 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1332 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1333 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1335 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1336 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1337 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1338 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1339 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1340 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1341 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1342 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1343 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1344 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1345 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1346 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1347 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1348 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1349 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1350 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1351 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1352 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1353 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1354 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1356 // SSE2 instructions with XS prefix
1357 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1358 "cvtss2sd\t{$src, $dst|$dst, $src}",
1359 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1360 Requires<[HasSSE2]>;
1361 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1362 "cvtss2sd\t{$src, $dst|$dst, $src}",
1363 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1364 Requires<[HasSSE2, OptForSize]>;
1366 def : Pat<(extloadf32 addr:$src),
1367 (CVTSS2SDrr (MOVSSrm addr:$src))>,
1368 Requires<[HasSSE2, OptForSpeed]>;
1370 // Match intrinsics which expect XMM operand(s).
1371 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1372 "cvtsd2si\t{$src, $dst|$dst, $src}",
1373 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1374 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1375 "cvtsd2si\t{$src, $dst|$dst, $src}",
1376 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1377 (load addr:$src)))]>;
1379 // Match intrinsics which expect MM and XMM operand(s).
1380 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1381 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1382 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1383 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1384 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1385 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1386 (memop addr:$src)))]>;
1387 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1388 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1389 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1390 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1391 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1392 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1393 (memop addr:$src)))]>;
1394 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1395 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1396 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1397 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1398 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1399 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1400 (load addr:$src)))]>;
1402 // Aliases for intrinsics
1403 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1404 "cvttsd2si\t{$src, $dst|$dst, $src}",
1406 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1407 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1408 "cvttsd2si\t{$src, $dst|$dst, $src}",
1409 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1410 (load addr:$src)))]>;
1412 // Comparison instructions
1413 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1414 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1415 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1416 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1418 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1419 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1420 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1422 // Accept explicit immediate argument form instead of comparison code.
1423 let isAsmParserOnly = 1 in {
1424 def CMPSDrr_alt : SDIi8<0xC2, MRMSrcReg,
1425 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, i8imm:$src2),
1426 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1428 def CMPSDrm_alt : SDIi8<0xC2, MRMSrcMem,
1429 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, i8imm:$src2),
1430 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1434 let Defs = [EFLAGS] in {
1435 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1436 "ucomisd\t{$src2, $src1|$src1, $src2}",
1437 [(set EFLAGS, (X86cmp FR64:$src1, FR64:$src2))]>;
1438 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1439 "ucomisd\t{$src2, $src1|$src1, $src2}",
1440 [(set EFLAGS, (X86cmp FR64:$src1, (loadf64 addr:$src2)))]>;
1441 } // Defs = [EFLAGS]
1443 // Aliases to match intrinsics which expect XMM operand(s).
1444 let Constraints = "$src1 = $dst" in {
1445 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1447 (ins VR128:$src1, VR128:$src, SSECC:$cc),
1448 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1449 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1450 VR128:$src, imm:$cc))]>;
1451 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1453 (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1454 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1455 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1456 (load addr:$src), imm:$cc))]>;
1459 let Defs = [EFLAGS] in {
1460 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1461 "ucomisd\t{$src2, $src1|$src1, $src2}",
1462 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1464 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1465 "ucomisd\t{$src2, $src1|$src1, $src2}",
1466 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1467 (load addr:$src2)))]>;
1469 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1470 "comisd\t{$src2, $src1|$src1, $src2}",
1471 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1473 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1474 "comisd\t{$src2, $src1|$src1, $src2}",
1475 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1476 (load addr:$src2)))]>;
1477 } // Defs = [EFLAGS]
1479 // Aliases of packed SSE2 instructions for scalar use. These all have names
1480 // that start with 'Fs'.
1482 // Alias instructions that map fld0 to pxor for sse.
1483 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1484 canFoldAsLoad = 1 in
1485 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1486 [(set FR64:$dst, fpimm0)]>,
1487 Requires<[HasSSE2]>, TB, OpSize;
1489 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1491 let neverHasSideEffects = 1 in
1492 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1493 "movapd\t{$src, $dst|$dst, $src}", []>;
1495 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1497 let canFoldAsLoad = 1, isReMaterializable = 1 in
1498 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1499 "movapd\t{$src, $dst|$dst, $src}",
1500 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1502 //===---------------------------------------------------------------------===//
1503 // SSE packed FP Instructions
1505 // Move Instructions
1506 let neverHasSideEffects = 1 in
1507 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1508 "movapd\t{$src, $dst|$dst, $src}", []>;
1509 let canFoldAsLoad = 1, isReMaterializable = 1 in
1510 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1511 "movapd\t{$src, $dst|$dst, $src}",
1512 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1514 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1515 "movapd\t{$src, $dst|$dst, $src}",
1516 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1518 let neverHasSideEffects = 1 in
1519 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1520 "movupd\t{$src, $dst|$dst, $src}", []>;
1521 let canFoldAsLoad = 1 in
1522 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1523 "movupd\t{$src, $dst|$dst, $src}",
1524 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1525 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1526 "movupd\t{$src, $dst|$dst, $src}",
1527 [(store (v2f64 VR128:$src), addr:$dst)]>;
1529 // Intrinsic forms of MOVUPD load and store
1530 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1531 "movupd\t{$src, $dst|$dst, $src}",
1532 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1533 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1534 "movupd\t{$src, $dst|$dst, $src}",
1535 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1537 let Constraints = "$src1 = $dst" in {
1538 let AddedComplexity = 20 in {
1539 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1540 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1541 "movlpd\t{$src2, $dst|$dst, $src2}",
1543 (v2f64 (movlp VR128:$src1,
1544 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1545 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1546 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1547 "movhpd\t{$src2, $dst|$dst, $src2}",
1549 (v2f64 (movlhps VR128:$src1,
1550 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1551 } // AddedComplexity
1552 } // Constraints = "$src1 = $dst"
1554 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1555 "movlpd\t{$src, $dst|$dst, $src}",
1556 [(store (f64 (vector_extract (v2f64 VR128:$src),
1557 (iPTR 0))), addr:$dst)]>;
1559 // v2f64 extract element 1 is always custom lowered to unpack high to low
1560 // and extract element 0 so the non-store version isn't too horrible.
1561 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1562 "movhpd\t{$src, $dst|$dst, $src}",
1563 [(store (f64 (vector_extract
1564 (v2f64 (unpckh VR128:$src, (undef))),
1565 (iPTR 0))), addr:$dst)]>;
1567 // SSE2 instructions without OpSize prefix
1568 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1569 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1570 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1571 TB, Requires<[HasSSE2]>;
1572 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1573 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1574 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1575 (bitconvert (memopv2i64 addr:$src))))]>,
1576 TB, Requires<[HasSSE2]>;
1578 // SSE2 instructions with XS prefix
1579 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1580 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1581 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1582 XS, Requires<[HasSSE2]>;
1583 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1584 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1585 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1586 (bitconvert (memopv2i64 addr:$src))))]>,
1587 XS, Requires<[HasSSE2]>;
1589 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1590 "cvtps2dq\t{$src, $dst|$dst, $src}",
1591 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1592 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1593 "cvtps2dq\t{$src, $dst|$dst, $src}",
1594 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1595 (memop addr:$src)))]>;
1596 // SSE2 packed instructions with XS prefix
1597 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1598 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1599 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1600 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1602 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1603 "cvttps2dq\t{$src, $dst|$dst, $src}",
1605 (int_x86_sse2_cvttps2dq VR128:$src))]>,
1606 XS, Requires<[HasSSE2]>;
1607 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1608 "cvttps2dq\t{$src, $dst|$dst, $src}",
1609 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1610 (memop addr:$src)))]>,
1611 XS, Requires<[HasSSE2]>;
1613 // SSE2 packed instructions with XD prefix
1614 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1615 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1616 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1617 XD, Requires<[HasSSE2]>;
1618 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1619 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1620 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1621 (memop addr:$src)))]>,
1622 XD, Requires<[HasSSE2]>;
1624 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1625 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1626 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1627 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1628 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1629 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1630 (memop addr:$src)))]>;
1632 // SSE2 instructions without OpSize prefix
1633 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1634 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1635 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1636 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1638 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1639 "cvtps2pd\t{$src, $dst|$dst, $src}",
1640 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1641 TB, Requires<[HasSSE2]>;
1642 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1643 "cvtps2pd\t{$src, $dst|$dst, $src}",
1644 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1645 (load addr:$src)))]>,
1646 TB, Requires<[HasSSE2]>;
1648 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1649 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1650 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1651 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1654 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1655 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1656 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1657 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1658 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1659 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1660 (memop addr:$src)))]>;
1662 // Match intrinsics which expect XMM operand(s).
1663 // Aliases for intrinsics
1664 let Constraints = "$src1 = $dst" in {
1665 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1666 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1667 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1668 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1670 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1671 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1672 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1673 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1674 (loadi32 addr:$src2)))]>;
1675 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1676 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1677 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1678 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1680 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1681 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1682 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1683 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1684 (load addr:$src2)))]>;
1685 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1686 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1687 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1688 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1689 VR128:$src2))]>, XS,
1690 Requires<[HasSSE2]>;
1691 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1692 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1693 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1694 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1695 (load addr:$src2)))]>, XS,
1696 Requires<[HasSSE2]>;
1701 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1703 /// In addition, we also have a special variant of the scalar form here to
1704 /// represent the associated intrinsic operation. This form is unlike the
1705 /// plain scalar form, in that it takes an entire vector (instead of a
1706 /// scalar) and leaves the top elements undefined.
1708 /// And, we have a special variant form for a full-vector intrinsic form.
1710 /// These four forms can each have a reg or a mem operand, so there are a
1711 /// total of eight "instructions".
1713 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1717 bit Commutable = 0> {
1718 // Scalar operation, reg.
1719 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1720 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1721 [(set FR64:$dst, (OpNode FR64:$src))]> {
1722 let isCommutable = Commutable;
1725 // Scalar operation, mem.
1726 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1727 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1728 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1730 // Vector operation, reg.
1731 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1732 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1733 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1734 let isCommutable = Commutable;
1737 // Vector operation, mem.
1738 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1739 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1740 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1742 // Intrinsic operation, reg.
1743 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1744 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1745 [(set VR128:$dst, (F64Int VR128:$src))]> {
1746 let isCommutable = Commutable;
1749 // Intrinsic operation, mem.
1750 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1751 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1752 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1754 // Vector intrinsic operation, reg
1755 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1756 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1757 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1758 let isCommutable = Commutable;
1761 // Vector intrinsic operation, mem
1762 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1763 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1764 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1768 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1769 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1771 // There is no f64 version of the reciprocal approximation instructions.
1773 let Constraints = "$src1 = $dst" in {
1774 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1775 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1776 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1777 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1778 VR128:$src, imm:$cc))]>;
1779 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1780 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1781 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1782 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1783 (memop addr:$src), imm:$cc))]>;
1785 // Accept explicit immediate argument form instead of comparison code.
1786 let isAsmParserOnly = 1 in {
1787 def CMPPDrri_alt : PDIi8<0xC2, MRMSrcReg,
1788 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1789 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1790 def CMPPDrmi_alt : PDIi8<0xC2, MRMSrcMem,
1791 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1792 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1795 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1796 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1797 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1798 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1800 // Shuffle and unpack instructions
1801 let Constraints = "$src1 = $dst" in {
1802 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1803 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1804 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1806 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1807 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1808 (outs VR128:$dst), (ins VR128:$src1,
1809 f128mem:$src2, i8imm:$src3),
1810 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1813 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1815 let AddedComplexity = 10 in {
1816 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1817 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1818 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1820 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1821 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1822 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1823 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1825 (v2f64 (unpckh VR128:$src1,
1826 (memopv2f64 addr:$src2))))]>;
1828 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1829 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1830 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1832 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1833 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1834 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1835 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1837 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1838 } // AddedComplexity
1839 } // Constraints = "$src1 = $dst"
1842 //===---------------------------------------------------------------------===//
1843 // SSE integer instructions
1844 let ExeDomain = SSEPackedInt in {
1846 // Move Instructions
1847 let neverHasSideEffects = 1 in
1848 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1849 "movdqa\t{$src, $dst|$dst, $src}", []>;
1850 let canFoldAsLoad = 1, mayLoad = 1 in
1851 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1852 "movdqa\t{$src, $dst|$dst, $src}",
1853 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1855 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1856 "movdqa\t{$src, $dst|$dst, $src}",
1857 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1858 let canFoldAsLoad = 1, mayLoad = 1 in
1859 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1860 "movdqu\t{$src, $dst|$dst, $src}",
1861 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1862 XS, Requires<[HasSSE2]>;
1864 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1865 "movdqu\t{$src, $dst|$dst, $src}",
1866 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1867 XS, Requires<[HasSSE2]>;
1869 // Intrinsic forms of MOVDQU load and store
1870 let canFoldAsLoad = 1 in
1871 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1872 "movdqu\t{$src, $dst|$dst, $src}",
1873 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1874 XS, Requires<[HasSSE2]>;
1875 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1876 "movdqu\t{$src, $dst|$dst, $src}",
1877 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1878 XS, Requires<[HasSSE2]>;
1880 let Constraints = "$src1 = $dst" in {
1882 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1883 bit Commutable = 0> {
1884 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1885 (ins VR128:$src1, VR128:$src2),
1886 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1887 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1888 let isCommutable = Commutable;
1890 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1891 (ins VR128:$src1, i128mem:$src2),
1892 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1893 [(set VR128:$dst, (IntId VR128:$src1,
1894 (bitconvert (memopv2i64
1898 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1900 Intrinsic IntId, Intrinsic IntId2> {
1901 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1902 (ins VR128:$src1, VR128:$src2),
1903 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1904 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1905 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1906 (ins VR128:$src1, i128mem:$src2),
1907 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1908 [(set VR128:$dst, (IntId VR128:$src1,
1909 (bitconvert (memopv2i64 addr:$src2))))]>;
1910 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
1911 (ins VR128:$src1, i32i8imm:$src2),
1912 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1913 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1916 /// PDI_binop_rm - Simple SSE2 binary operator.
1917 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1918 ValueType OpVT, bit Commutable = 0> {
1919 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1920 (ins VR128:$src1, VR128:$src2),
1921 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1922 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1923 let isCommutable = Commutable;
1925 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1926 (ins VR128:$src1, i128mem:$src2),
1927 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1928 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1929 (bitconvert (memopv2i64 addr:$src2)))))]>;
1932 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1934 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1935 /// to collapse (bitconvert VT to VT) into its operand.
1937 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1938 bit Commutable = 0> {
1939 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1940 (ins VR128:$src1, VR128:$src2),
1941 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1942 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1943 let isCommutable = Commutable;
1945 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1946 (ins VR128:$src1, i128mem:$src2),
1947 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1948 [(set VR128:$dst, (OpNode VR128:$src1,
1949 (memopv2i64 addr:$src2)))]>;
1952 } // Constraints = "$src1 = $dst"
1953 } // ExeDomain = SSEPackedInt
1955 // 128-bit Integer Arithmetic
1957 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1958 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1959 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1960 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1962 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1963 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1964 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1965 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1967 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1968 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1969 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1970 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1972 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1973 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1974 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1975 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1977 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1979 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1980 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1981 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1983 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1985 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1986 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1989 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1990 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1991 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1992 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1993 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
1996 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
1997 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
1998 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
1999 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
2000 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
2001 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
2003 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2004 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2005 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2006 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2007 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2008 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2010 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2011 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2012 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2013 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2015 // 128-bit logical shifts.
2016 let Constraints = "$src1 = $dst", neverHasSideEffects = 1,
2017 ExeDomain = SSEPackedInt in {
2018 def PSLLDQri : PDIi8<0x73, MRM7r,
2019 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2020 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2021 def PSRLDQri : PDIi8<0x73, MRM3r,
2022 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2023 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2024 // PSRADQri doesn't exist in SSE[1-3].
2027 let Predicates = [HasSSE2] in {
2028 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2029 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2030 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2031 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2032 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2033 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2034 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2035 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2036 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2037 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2039 // Shift up / down and insert zero's.
2040 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2041 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2042 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2043 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2047 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2048 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2049 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2051 let Constraints = "$src1 = $dst", ExeDomain = SSEPackedInt in {
2052 def PANDNrr : PDI<0xDF, MRMSrcReg,
2053 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2054 "pandn\t{$src2, $dst|$dst, $src2}",
2055 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2058 def PANDNrm : PDI<0xDF, MRMSrcMem,
2059 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2060 "pandn\t{$src2, $dst|$dst, $src2}",
2061 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2062 (memopv2i64 addr:$src2))))]>;
2065 // SSE2 Integer comparison
2066 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2067 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2068 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2069 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2070 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2071 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2073 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2074 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2075 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2076 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2077 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2078 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2079 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2080 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2081 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2082 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2083 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2084 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2086 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2087 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2088 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2089 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2090 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2091 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2092 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2093 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2094 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2095 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2096 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2097 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2100 // Pack instructions
2101 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2102 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2103 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2105 let ExeDomain = SSEPackedInt in {
2107 // Shuffle and unpack instructions
2108 let AddedComplexity = 5 in {
2109 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2110 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2111 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2112 [(set VR128:$dst, (v4i32 (pshufd:$src2
2113 VR128:$src1, (undef))))]>;
2114 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2115 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2116 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2117 [(set VR128:$dst, (v4i32 (pshufd:$src2
2118 (bc_v4i32 (memopv2i64 addr:$src1)),
2122 // SSE2 with ImmT == Imm8 and XS prefix.
2123 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2124 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2125 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2126 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2128 XS, Requires<[HasSSE2]>;
2129 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2130 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2131 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2132 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2133 (bc_v8i16 (memopv2i64 addr:$src1)),
2135 XS, Requires<[HasSSE2]>;
2137 // SSE2 with ImmT == Imm8 and XD prefix.
2138 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2139 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2140 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2141 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2143 XD, Requires<[HasSSE2]>;
2144 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2145 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2146 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2147 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2148 (bc_v8i16 (memopv2i64 addr:$src1)),
2150 XD, Requires<[HasSSE2]>;
2152 // Unpack instructions
2153 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
2154 PatFrag unp_frag, PatFrag bc_frag> {
2155 def rr : PDI<opc, MRMSrcReg,
2156 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2157 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2158 [(set VR128:$dst, (vt (unp_frag VR128:$src1, VR128:$src2)))]>;
2159 def rm : PDI<opc, MRMSrcMem,
2160 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2161 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2162 [(set VR128:$dst, (unp_frag VR128:$src1,
2163 (bc_frag (memopv2i64
2167 let Constraints = "$src1 = $dst" in {
2168 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, unpckl, bc_v16i8>;
2169 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, unpckl, bc_v8i16>;
2170 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, unpckl, bc_v4i32>;
2172 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2173 /// knew to collapse (bitconvert VT to VT) into its operand.
2174 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2175 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2176 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2178 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2179 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2180 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2181 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2183 (v2i64 (unpckl VR128:$src1,
2184 (memopv2i64 addr:$src2))))]>;
2186 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, unpckh, bc_v16i8>;
2187 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, unpckh, bc_v8i16>;
2188 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, unpckh, bc_v4i32>;
2190 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2191 /// knew to collapse (bitconvert VT to VT) into its operand.
2192 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2193 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2194 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2196 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2197 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2198 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2199 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2201 (v2i64 (unpckh VR128:$src1,
2202 (memopv2i64 addr:$src2))))]>;
2206 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2207 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2208 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2209 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2211 let Constraints = "$src1 = $dst" in {
2212 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2213 (outs VR128:$dst), (ins VR128:$src1,
2214 GR32:$src2, i32i8imm:$src3),
2215 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2217 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2218 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2219 (outs VR128:$dst), (ins VR128:$src1,
2220 i16mem:$src2, i32i8imm:$src3),
2221 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2223 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2228 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2229 "pmovmskb\t{$src, $dst|$dst, $src}",
2230 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2232 // Conditional store
2234 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2235 "maskmovdqu\t{$mask, $src|$src, $mask}",
2236 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2239 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2240 "maskmovdqu\t{$mask, $src|$src, $mask}",
2241 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2243 } // ExeDomain = SSEPackedInt
2245 // Non-temporal stores
2246 def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2247 "movntpd\t{$src, $dst|$dst, $src}",
2248 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2249 let ExeDomain = SSEPackedInt in
2250 def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2251 "movntdq\t{$src, $dst|$dst, $src}",
2252 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2253 def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2254 "movnti\t{$src, $dst|$dst, $src}",
2255 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2256 TB, Requires<[HasSSE2]>;
2258 let AddedComplexity = 400 in { // Prefer non-temporal versions
2259 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2260 "movntpd\t{$src, $dst|$dst, $src}",
2261 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
2263 let ExeDomain = SSEPackedInt in
2264 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2265 "movntdq\t{$src, $dst|$dst, $src}",
2266 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2270 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2271 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2272 TB, Requires<[HasSSE2]>;
2274 // Load, store, and memory fence
2275 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
2276 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2277 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
2278 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2280 // Pause. This "instruction" is encoded as "rep; nop", so even though it
2281 // was introduced with SSE2, it's backward compatible.
2282 def PAUSE : I<0x90, RawFrm, (outs), (ins), "pause", []>, REP;
2284 //TODO: custom lower this so as to never even generate the noop
2285 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2287 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2288 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2289 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2292 // Alias instructions that map zero vector to pxor / xorp* for sse.
2293 // We set canFoldAsLoad because this can be converted to a constant-pool
2294 // load of an all-ones value if folding it would be beneficial.
2295 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2296 isCodeGenOnly = 1, ExeDomain = SSEPackedInt in
2297 // FIXME: Change encoding to pseudo.
2298 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
2299 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2301 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2302 "movd\t{$src, $dst|$dst, $src}",
2304 (v4i32 (scalar_to_vector GR32:$src)))]>;
2305 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2306 "movd\t{$src, $dst|$dst, $src}",
2308 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2310 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2311 "movd\t{$src, $dst|$dst, $src}",
2312 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2314 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2315 "movd\t{$src, $dst|$dst, $src}",
2316 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2318 // SSE2 instructions with XS prefix
2319 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2320 "movq\t{$src, $dst|$dst, $src}",
2322 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2323 Requires<[HasSSE2]>;
2324 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2325 "movq\t{$src, $dst|$dst, $src}",
2326 [(store (i64 (vector_extract (v2i64 VR128:$src),
2327 (iPTR 0))), addr:$dst)]>;
2329 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
2330 (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
2332 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2333 "movd\t{$src, $dst|$dst, $src}",
2334 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2336 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2337 "movd\t{$src, $dst|$dst, $src}",
2338 [(store (i32 (vector_extract (v4i32 VR128:$src),
2339 (iPTR 0))), addr:$dst)]>;
2341 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2342 "movd\t{$src, $dst|$dst, $src}",
2343 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2344 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2345 "movd\t{$src, $dst|$dst, $src}",
2346 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2348 // Store / copy lower 64-bits of a XMM register.
2349 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2350 "movq\t{$src, $dst|$dst, $src}",
2351 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2353 // movd / movq to XMM register zero-extends
2354 let AddedComplexity = 15 in {
2355 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2356 "movd\t{$src, $dst|$dst, $src}",
2357 [(set VR128:$dst, (v4i32 (X86vzmovl
2358 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2359 // This is X86-64 only.
2360 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2361 "mov{d|q}\t{$src, $dst|$dst, $src}",
2362 [(set VR128:$dst, (v2i64 (X86vzmovl
2363 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2366 let AddedComplexity = 20 in {
2367 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2368 "movd\t{$src, $dst|$dst, $src}",
2370 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2371 (loadi32 addr:$src))))))]>;
2373 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2374 (MOVZDI2PDIrm addr:$src)>;
2375 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2376 (MOVZDI2PDIrm addr:$src)>;
2377 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2378 (MOVZDI2PDIrm addr:$src)>;
2380 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2381 "movq\t{$src, $dst|$dst, $src}",
2383 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2384 (loadi64 addr:$src))))))]>, XS,
2385 Requires<[HasSSE2]>;
2387 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2388 (MOVZQI2PQIrm addr:$src)>;
2389 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2390 (MOVZQI2PQIrm addr:$src)>;
2391 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2394 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2395 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2396 let AddedComplexity = 15 in
2397 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2398 "movq\t{$src, $dst|$dst, $src}",
2399 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2400 XS, Requires<[HasSSE2]>;
2402 let AddedComplexity = 20 in {
2403 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2404 "movq\t{$src, $dst|$dst, $src}",
2405 [(set VR128:$dst, (v2i64 (X86vzmovl
2406 (loadv2i64 addr:$src))))]>,
2407 XS, Requires<[HasSSE2]>;
2409 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2410 (MOVZPQILo2PQIrm addr:$src)>;
2413 // Instructions for the disassembler
2414 // xr = XMM register
2417 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2418 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2420 //===---------------------------------------------------------------------===//
2421 // SSE3 Instructions
2422 //===---------------------------------------------------------------------===//
2424 // Move Instructions
2425 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2426 "movshdup\t{$src, $dst|$dst, $src}",
2427 [(set VR128:$dst, (v4f32 (movshdup
2428 VR128:$src, (undef))))]>;
2429 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2430 "movshdup\t{$src, $dst|$dst, $src}",
2431 [(set VR128:$dst, (movshdup
2432 (memopv4f32 addr:$src), (undef)))]>;
2434 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2435 "movsldup\t{$src, $dst|$dst, $src}",
2436 [(set VR128:$dst, (v4f32 (movsldup
2437 VR128:$src, (undef))))]>;
2438 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2439 "movsldup\t{$src, $dst|$dst, $src}",
2440 [(set VR128:$dst, (movsldup
2441 (memopv4f32 addr:$src), (undef)))]>;
2443 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2444 "movddup\t{$src, $dst|$dst, $src}",
2445 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2446 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2447 "movddup\t{$src, $dst|$dst, $src}",
2449 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2452 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2454 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2456 let AddedComplexity = 5 in {
2457 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2458 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2459 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2460 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2461 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2462 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2463 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2464 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2468 let Constraints = "$src1 = $dst" in {
2469 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2470 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2471 "addsubps\t{$src2, $dst|$dst, $src2}",
2472 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2474 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2475 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2476 "addsubps\t{$src2, $dst|$dst, $src2}",
2477 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2478 (memop addr:$src2)))]>;
2479 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2480 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2481 "addsubpd\t{$src2, $dst|$dst, $src2}",
2482 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2484 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2485 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2486 "addsubpd\t{$src2, $dst|$dst, $src2}",
2487 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2488 (memop addr:$src2)))]>;
2491 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2492 "lddqu\t{$src, $dst|$dst, $src}",
2493 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2496 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2497 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2498 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2499 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2500 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2501 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2502 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2503 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2504 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2505 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2506 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2507 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2508 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2509 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2510 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2511 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2513 let Constraints = "$src1 = $dst" in {
2514 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2515 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2516 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2517 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2518 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2519 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2520 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2521 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2524 // Thread synchronization
2525 def MONITOR : I<0x01, MRM_C8, (outs), (ins), "monitor",
2526 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2527 def MWAIT : I<0x01, MRM_C9, (outs), (ins), "mwait",
2528 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2530 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2531 let AddedComplexity = 15 in
2532 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2533 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2534 let AddedComplexity = 20 in
2535 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2536 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2538 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2539 let AddedComplexity = 15 in
2540 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2541 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2542 let AddedComplexity = 20 in
2543 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2544 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2546 //===---------------------------------------------------------------------===//
2547 // SSSE3 Instructions
2548 //===---------------------------------------------------------------------===//
2550 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2551 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2552 Intrinsic IntId64, Intrinsic IntId128> {
2553 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2554 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2555 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2557 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2558 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2560 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2562 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2564 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2565 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2568 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2570 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2573 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2576 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2577 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2578 Intrinsic IntId64, Intrinsic IntId128> {
2579 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2581 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2582 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2584 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2586 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2589 (bitconvert (memopv4i16 addr:$src))))]>;
2591 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2593 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2594 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2597 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2599 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2602 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2605 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2606 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2607 Intrinsic IntId64, Intrinsic IntId128> {
2608 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2610 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2611 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2613 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2615 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2618 (bitconvert (memopv2i32 addr:$src))))]>;
2620 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2622 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2623 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2626 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2628 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2631 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2634 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2635 int_x86_ssse3_pabs_b,
2636 int_x86_ssse3_pabs_b_128>;
2637 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2638 int_x86_ssse3_pabs_w,
2639 int_x86_ssse3_pabs_w_128>;
2640 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2641 int_x86_ssse3_pabs_d,
2642 int_x86_ssse3_pabs_d_128>;
2644 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2645 let Constraints = "$src1 = $dst" in {
2646 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2647 Intrinsic IntId64, Intrinsic IntId128,
2648 bit Commutable = 0> {
2649 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2650 (ins VR64:$src1, VR64:$src2),
2651 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2652 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2653 let isCommutable = Commutable;
2655 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2656 (ins VR64:$src1, i64mem:$src2),
2657 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2659 (IntId64 VR64:$src1,
2660 (bitconvert (memopv8i8 addr:$src2))))]>;
2662 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2663 (ins VR128:$src1, VR128:$src2),
2664 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2665 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2667 let isCommutable = Commutable;
2669 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2670 (ins VR128:$src1, i128mem:$src2),
2671 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2673 (IntId128 VR128:$src1,
2674 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2678 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2679 let Constraints = "$src1 = $dst" in {
2680 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2681 Intrinsic IntId64, Intrinsic IntId128,
2682 bit Commutable = 0> {
2683 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2684 (ins VR64:$src1, VR64:$src2),
2685 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2686 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2687 let isCommutable = Commutable;
2689 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2690 (ins VR64:$src1, i64mem:$src2),
2691 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2693 (IntId64 VR64:$src1,
2694 (bitconvert (memopv4i16 addr:$src2))))]>;
2696 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2697 (ins VR128:$src1, VR128:$src2),
2698 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2699 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2701 let isCommutable = Commutable;
2703 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2704 (ins VR128:$src1, i128mem:$src2),
2705 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2707 (IntId128 VR128:$src1,
2708 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2712 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2713 let Constraints = "$src1 = $dst" in {
2714 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2715 Intrinsic IntId64, Intrinsic IntId128,
2716 bit Commutable = 0> {
2717 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2718 (ins VR64:$src1, VR64:$src2),
2719 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2720 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2721 let isCommutable = Commutable;
2723 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2724 (ins VR64:$src1, i64mem:$src2),
2725 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2727 (IntId64 VR64:$src1,
2728 (bitconvert (memopv2i32 addr:$src2))))]>;
2730 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2731 (ins VR128:$src1, VR128:$src2),
2732 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2733 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2735 let isCommutable = Commutable;
2737 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2738 (ins VR128:$src1, i128mem:$src2),
2739 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2741 (IntId128 VR128:$src1,
2742 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2746 let ImmT = NoImm in { // None of these have i8 immediate fields.
2747 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2748 int_x86_ssse3_phadd_w,
2749 int_x86_ssse3_phadd_w_128>;
2750 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2751 int_x86_ssse3_phadd_d,
2752 int_x86_ssse3_phadd_d_128>;
2753 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2754 int_x86_ssse3_phadd_sw,
2755 int_x86_ssse3_phadd_sw_128>;
2756 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2757 int_x86_ssse3_phsub_w,
2758 int_x86_ssse3_phsub_w_128>;
2759 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2760 int_x86_ssse3_phsub_d,
2761 int_x86_ssse3_phsub_d_128>;
2762 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2763 int_x86_ssse3_phsub_sw,
2764 int_x86_ssse3_phsub_sw_128>;
2765 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2766 int_x86_ssse3_pmadd_ub_sw,
2767 int_x86_ssse3_pmadd_ub_sw_128>;
2768 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2769 int_x86_ssse3_pmul_hr_sw,
2770 int_x86_ssse3_pmul_hr_sw_128, 1>;
2772 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2773 int_x86_ssse3_pshuf_b,
2774 int_x86_ssse3_pshuf_b_128>;
2775 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2776 int_x86_ssse3_psign_b,
2777 int_x86_ssse3_psign_b_128>;
2778 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2779 int_x86_ssse3_psign_w,
2780 int_x86_ssse3_psign_w_128>;
2781 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2782 int_x86_ssse3_psign_d,
2783 int_x86_ssse3_psign_d_128>;
2786 // palignr patterns.
2787 let Constraints = "$src1 = $dst" in {
2788 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2789 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2790 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2792 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2793 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2794 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2797 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2798 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2799 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2801 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2802 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2803 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2807 let AddedComplexity = 5 in {
2809 def : Pat<(v1i64 (palign:$src3 VR64:$src1, VR64:$src2)),
2810 (PALIGNR64rr VR64:$src2, VR64:$src1,
2811 (SHUFFLE_get_palign_imm VR64:$src3))>,
2812 Requires<[HasSSSE3]>;
2813 def : Pat<(v2i32 (palign:$src3 VR64:$src1, VR64:$src2)),
2814 (PALIGNR64rr VR64:$src2, VR64:$src1,
2815 (SHUFFLE_get_palign_imm VR64:$src3))>,
2816 Requires<[HasSSSE3]>;
2817 def : Pat<(v2f32 (palign:$src3 VR64:$src1, VR64:$src2)),
2818 (PALIGNR64rr VR64:$src2, VR64:$src1,
2819 (SHUFFLE_get_palign_imm VR64:$src3))>,
2820 Requires<[HasSSSE3]>;
2821 def : Pat<(v4i16 (palign:$src3 VR64:$src1, VR64:$src2)),
2822 (PALIGNR64rr VR64:$src2, VR64:$src1,
2823 (SHUFFLE_get_palign_imm VR64:$src3))>,
2824 Requires<[HasSSSE3]>;
2825 def : Pat<(v8i8 (palign:$src3 VR64:$src1, VR64:$src2)),
2826 (PALIGNR64rr VR64:$src2, VR64:$src1,
2827 (SHUFFLE_get_palign_imm VR64:$src3))>,
2828 Requires<[HasSSSE3]>;
2830 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2831 (PALIGNR128rr VR128:$src2, VR128:$src1,
2832 (SHUFFLE_get_palign_imm VR128:$src3))>,
2833 Requires<[HasSSSE3]>;
2834 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2835 (PALIGNR128rr VR128:$src2, VR128:$src1,
2836 (SHUFFLE_get_palign_imm VR128:$src3))>,
2837 Requires<[HasSSSE3]>;
2838 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2839 (PALIGNR128rr VR128:$src2, VR128:$src1,
2840 (SHUFFLE_get_palign_imm VR128:$src3))>,
2841 Requires<[HasSSSE3]>;
2842 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2843 (PALIGNR128rr VR128:$src2, VR128:$src1,
2844 (SHUFFLE_get_palign_imm VR128:$src3))>,
2845 Requires<[HasSSSE3]>;
2848 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2849 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2850 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2851 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2853 //===---------------------------------------------------------------------===//
2854 // Non-Instruction Patterns
2855 //===---------------------------------------------------------------------===//
2857 // extload f32 -> f64. This matches load+fextend because we have a hack in
2858 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2860 // Since these loads aren't folded into the fextend, we have to match it
2862 let Predicates = [HasSSE2] in
2863 def : Pat<(fextend (loadf32 addr:$src)),
2864 (CVTSS2SDrm addr:$src)>;
2867 let Predicates = [HasSSE2] in {
2868 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2869 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2870 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2871 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2872 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2873 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2874 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2875 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2876 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2877 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2878 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2879 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2880 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2881 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2882 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2883 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2884 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2885 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2886 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2887 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2888 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2889 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2890 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2891 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2892 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2893 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2894 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2895 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2896 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2897 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2900 // Move scalar to XMM zero-extended
2901 // movd to XMM register zero-extends
2902 let AddedComplexity = 15 in {
2903 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2904 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2905 (MOVSDrr (v2f64 (V_SET0PS)), FR64:$src)>;
2906 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2907 (MOVSSrr (v4f32 (V_SET0PS)), FR32:$src)>;
2908 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2909 (MOVSSrr (v4f32 (V_SET0PS)),
2910 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss)))>;
2911 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2912 (MOVSSrr (v4i32 (V_SET0PI)),
2913 (EXTRACT_SUBREG (v4i32 VR128:$src), sub_ss))>;
2916 // Splat v2f64 / v2i64
2917 let AddedComplexity = 10 in {
2918 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
2919 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2920 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
2921 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2922 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
2923 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2924 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
2925 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2928 // Special unary SHUFPSrri case.
2929 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
2930 (SHUFPSrri VR128:$src1, VR128:$src1,
2931 (SHUFFLE_get_shuf_imm VR128:$src3))>;
2932 let AddedComplexity = 5 in
2933 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
2934 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2935 Requires<[HasSSE2]>;
2936 // Special unary SHUFPDrri case.
2937 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
2938 (SHUFPDrri VR128:$src1, VR128:$src1,
2939 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2940 Requires<[HasSSE2]>;
2941 // Special unary SHUFPDrri case.
2942 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
2943 (SHUFPDrri VR128:$src1, VR128:$src1,
2944 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2945 Requires<[HasSSE2]>;
2946 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2947 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
2948 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2949 Requires<[HasSSE2]>;
2951 // Special binary v4i32 shuffle cases with SHUFPS.
2952 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
2953 (SHUFPSrri VR128:$src1, VR128:$src2,
2954 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2955 Requires<[HasSSE2]>;
2956 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
2957 (SHUFPSrmi VR128:$src1, addr:$src2,
2958 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2959 Requires<[HasSSE2]>;
2960 // Special binary v2i64 shuffle cases using SHUFPDrri.
2961 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
2962 (SHUFPDrri VR128:$src1, VR128:$src2,
2963 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2964 Requires<[HasSSE2]>;
2966 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2967 let AddedComplexity = 15 in {
2968 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
2969 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2970 Requires<[OptForSpeed, HasSSE2]>;
2971 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
2972 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2973 Requires<[OptForSpeed, HasSSE2]>;
2975 let AddedComplexity = 10 in {
2976 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
2977 (UNPCKLPSrr VR128:$src, VR128:$src)>;
2978 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
2979 (PUNPCKLBWrr VR128:$src, VR128:$src)>;
2980 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
2981 (PUNPCKLWDrr VR128:$src, VR128:$src)>;
2982 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
2983 (PUNPCKLDQrr VR128:$src, VR128:$src)>;
2986 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2987 let AddedComplexity = 15 in {
2988 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
2989 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2990 Requires<[OptForSpeed, HasSSE2]>;
2991 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
2992 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2993 Requires<[OptForSpeed, HasSSE2]>;
2995 let AddedComplexity = 10 in {
2996 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
2997 (UNPCKHPSrr VR128:$src, VR128:$src)>;
2998 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
2999 (PUNPCKHBWrr VR128:$src, VR128:$src)>;
3000 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
3001 (PUNPCKHWDrr VR128:$src, VR128:$src)>;
3002 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
3003 (PUNPCKHDQrr VR128:$src, VR128:$src)>;
3006 let AddedComplexity = 20 in {
3007 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3008 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3009 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3011 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3012 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3013 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3015 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3016 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3017 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3018 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3019 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3022 let AddedComplexity = 20 in {
3023 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3024 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3025 (MOVLPSrm VR128:$src1, addr:$src2)>;
3026 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3027 (MOVLPDrm VR128:$src1, addr:$src2)>;
3028 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3029 (MOVLPSrm VR128:$src1, addr:$src2)>;
3030 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3031 (MOVLPDrm VR128:$src1, addr:$src2)>;
3034 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3035 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3036 (MOVLPSmr addr:$src1, VR128:$src2)>;
3037 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3038 (MOVLPDmr addr:$src1, VR128:$src2)>;
3039 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3041 (MOVLPSmr addr:$src1, VR128:$src2)>;
3042 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3043 (MOVLPDmr addr:$src1, VR128:$src2)>;
3045 let AddedComplexity = 15 in {
3046 // Setting the lowest element in the vector.
3047 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3048 (MOVSSrr (v4i32 VR128:$src1),
3049 (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_ss))>;
3050 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3051 (MOVSDrr (v2i64 VR128:$src1),
3052 (EXTRACT_SUBREG (v2i64 VR128:$src2), sub_sd))>;
3054 // vector_shuffle v1, v2 <4, 5, 2, 3> using movsd
3055 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3056 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3057 Requires<[HasSSE2]>;
3058 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3059 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3060 Requires<[HasSSE2]>;
3063 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3064 // fall back to this for SSE1)
3065 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3066 (SHUFPSrri VR128:$src2, VR128:$src1,
3067 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3069 // Set lowest element and zero upper elements.
3070 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3071 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3073 // Some special case pandn patterns.
3074 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3076 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3077 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3079 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3080 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3082 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3084 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3085 (memop addr:$src2))),
3086 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3087 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3088 (memop addr:$src2))),
3089 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3090 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3091 (memop addr:$src2))),
3092 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3094 // vector -> vector casts
3095 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3096 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3097 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3098 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3099 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3100 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3101 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3102 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3104 // Use movaps / movups for SSE integer load / store (one byte shorter).
3105 def : Pat<(alignedloadv4i32 addr:$src),
3106 (MOVAPSrm addr:$src)>;
3107 def : Pat<(loadv4i32 addr:$src),
3108 (MOVUPSrm addr:$src)>;
3109 def : Pat<(alignedloadv2i64 addr:$src),
3110 (MOVAPSrm addr:$src)>;
3111 def : Pat<(loadv2i64 addr:$src),
3112 (MOVUPSrm addr:$src)>;
3114 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3115 (MOVAPSmr addr:$dst, VR128:$src)>;
3116 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3117 (MOVAPSmr addr:$dst, VR128:$src)>;
3118 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3119 (MOVAPSmr addr:$dst, VR128:$src)>;
3120 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3121 (MOVAPSmr addr:$dst, VR128:$src)>;
3122 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3123 (MOVUPSmr addr:$dst, VR128:$src)>;
3124 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3125 (MOVUPSmr addr:$dst, VR128:$src)>;
3126 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3127 (MOVUPSmr addr:$dst, VR128:$src)>;
3128 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3129 (MOVUPSmr addr:$dst, VR128:$src)>;
3131 //===----------------------------------------------------------------------===//
3132 // SSE4.1 Instructions
3133 //===----------------------------------------------------------------------===//
3135 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3138 Intrinsic V2F64Int> {
3139 // Intrinsic operation, reg.
3140 // Vector intrinsic operation, reg
3141 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3142 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3143 !strconcat(OpcodeStr,
3144 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3145 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3148 // Vector intrinsic operation, mem
3149 def PSm_Int : Ii8<opcps, MRMSrcMem,
3150 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3151 !strconcat(OpcodeStr,
3152 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3154 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3156 Requires<[HasSSE41]>;
3158 // Vector intrinsic operation, reg
3159 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3160 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3161 !strconcat(OpcodeStr,
3162 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3163 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3166 // Vector intrinsic operation, mem
3167 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3168 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3169 !strconcat(OpcodeStr,
3170 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3172 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3176 let Constraints = "$src1 = $dst" in {
3177 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3181 // Intrinsic operation, reg.
3182 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3184 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3185 !strconcat(OpcodeStr,
3186 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3188 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3191 // Intrinsic operation, mem.
3192 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3194 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3195 !strconcat(OpcodeStr,
3196 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3198 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3201 // Intrinsic operation, reg.
3202 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3204 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3205 !strconcat(OpcodeStr,
3206 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3208 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3211 // Intrinsic operation, mem.
3212 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3214 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3215 !strconcat(OpcodeStr,
3216 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3218 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3223 // FP round - roundss, roundps, roundsd, roundpd
3224 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3225 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3226 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3227 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3229 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3230 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3231 Intrinsic IntId128> {
3232 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3234 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3235 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3236 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3238 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3241 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3244 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3245 int_x86_sse41_phminposuw>;
3247 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3248 let Constraints = "$src1 = $dst" in {
3249 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3250 Intrinsic IntId128, bit Commutable = 0> {
3251 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3252 (ins VR128:$src1, VR128:$src2),
3253 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3254 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3256 let isCommutable = Commutable;
3258 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3259 (ins VR128:$src1, i128mem:$src2),
3260 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3262 (IntId128 VR128:$src1,
3263 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3267 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3268 int_x86_sse41_pcmpeqq, 1>;
3269 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3270 int_x86_sse41_packusdw, 0>;
3271 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3272 int_x86_sse41_pminsb, 1>;
3273 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3274 int_x86_sse41_pminsd, 1>;
3275 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3276 int_x86_sse41_pminud, 1>;
3277 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3278 int_x86_sse41_pminuw, 1>;
3279 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3280 int_x86_sse41_pmaxsb, 1>;
3281 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3282 int_x86_sse41_pmaxsd, 1>;
3283 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3284 int_x86_sse41_pmaxud, 1>;
3285 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3286 int_x86_sse41_pmaxuw, 1>;
3288 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3290 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3291 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3292 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3293 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3295 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3296 let Constraints = "$src1 = $dst" in {
3297 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3298 SDNode OpNode, Intrinsic IntId128,
3299 bit Commutable = 0> {
3300 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3301 (ins VR128:$src1, VR128:$src2),
3302 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3303 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3304 VR128:$src2))]>, OpSize {
3305 let isCommutable = Commutable;
3307 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3308 (ins VR128:$src1, VR128:$src2),
3309 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3310 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3312 let isCommutable = Commutable;
3314 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3315 (ins VR128:$src1, i128mem:$src2),
3316 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3318 (OpVT (OpNode VR128:$src1, (memop addr:$src2))))]>, OpSize;
3319 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3320 (ins VR128:$src1, i128mem:$src2),
3321 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3323 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3328 /// SS48I_binop_rm - Simple SSE41 binary operator.
3329 let Constraints = "$src1 = $dst" in {
3330 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
3331 ValueType OpVT, bit Commutable = 0> {
3332 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3333 (ins VR128:$src1, VR128:$src2),
3334 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3335 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]>,
3337 let isCommutable = Commutable;
3339 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3340 (ins VR128:$src1, i128mem:$src2),
3341 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3342 [(set VR128:$dst, (OpNode VR128:$src1,
3343 (bc_v4i32 (memopv2i64 addr:$src2))))]>,
3348 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, 1>;
3350 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3351 let Constraints = "$src1 = $dst" in {
3352 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3353 Intrinsic IntId128, bit Commutable = 0> {
3354 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3355 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3356 !strconcat(OpcodeStr,
3357 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3359 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3361 let isCommutable = Commutable;
3363 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3364 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3365 !strconcat(OpcodeStr,
3366 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3368 (IntId128 VR128:$src1,
3369 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3374 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3375 int_x86_sse41_blendps, 0>;
3376 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3377 int_x86_sse41_blendpd, 0>;
3378 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3379 int_x86_sse41_pblendw, 0>;
3380 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3381 int_x86_sse41_dpps, 1>;
3382 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3383 int_x86_sse41_dppd, 1>;
3384 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3385 int_x86_sse41_mpsadbw, 0>;
3388 /// SS41I_ternary_int - SSE 4.1 ternary operator
3389 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3390 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3391 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3392 (ins VR128:$src1, VR128:$src2),
3393 !strconcat(OpcodeStr,
3394 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3395 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3398 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3399 (ins VR128:$src1, i128mem:$src2),
3400 !strconcat(OpcodeStr,
3401 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3404 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3408 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3409 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3410 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3413 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3414 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3415 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3416 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3418 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3419 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3421 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3425 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3426 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3427 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3428 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3429 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3430 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3432 // Common patterns involving scalar load.
3433 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3434 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3435 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3436 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3438 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3439 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3440 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3441 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3443 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3444 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3445 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3446 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3448 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3449 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3450 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3451 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3453 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3454 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3455 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3456 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3458 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3459 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3460 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3461 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3464 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3465 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3466 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3467 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3469 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3470 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3472 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3476 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3477 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3478 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3479 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3481 // Common patterns involving scalar load
3482 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3483 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3484 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3485 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3487 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3488 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3489 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3490 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3493 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3494 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3495 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3496 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3498 // Expecting a i16 load any extended to i32 value.
3499 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3500 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3501 [(set VR128:$dst, (IntId (bitconvert
3502 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3506 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3507 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3509 // Common patterns involving scalar load
3510 def : Pat<(int_x86_sse41_pmovsxbq
3511 (bitconvert (v4i32 (X86vzmovl
3512 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3513 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3515 def : Pat<(int_x86_sse41_pmovzxbq
3516 (bitconvert (v4i32 (X86vzmovl
3517 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3518 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3521 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3522 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3523 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3524 (ins VR128:$src1, i32i8imm:$src2),
3525 !strconcat(OpcodeStr,
3526 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3527 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3529 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3530 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3531 !strconcat(OpcodeStr,
3532 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3535 // There's an AssertZext in the way of writing the store pattern
3536 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3539 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3542 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3543 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3544 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3545 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3546 !strconcat(OpcodeStr,
3547 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3550 // There's an AssertZext in the way of writing the store pattern
3551 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3554 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3557 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3558 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3559 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3560 (ins VR128:$src1, i32i8imm:$src2),
3561 !strconcat(OpcodeStr,
3562 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3564 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3565 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3566 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3567 !strconcat(OpcodeStr,
3568 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3569 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3570 addr:$dst)]>, OpSize;
3573 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3576 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3578 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3579 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3580 (ins VR128:$src1, i32i8imm:$src2),
3581 !strconcat(OpcodeStr,
3582 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3584 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3586 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3587 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3588 !strconcat(OpcodeStr,
3589 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3590 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3591 addr:$dst)]>, OpSize;
3594 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3596 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3597 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3600 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3601 Requires<[HasSSE41]>;
3603 let Constraints = "$src1 = $dst" in {
3604 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3605 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3606 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3607 !strconcat(OpcodeStr,
3608 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3610 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3611 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3612 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3613 !strconcat(OpcodeStr,
3614 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3616 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3617 imm:$src3))]>, OpSize;
3621 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3623 let Constraints = "$src1 = $dst" in {
3624 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3625 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3626 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3627 !strconcat(OpcodeStr,
3628 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3630 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3632 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3633 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3634 !strconcat(OpcodeStr,
3635 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3637 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3638 imm:$src3)))]>, OpSize;
3642 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3644 // insertps has a few different modes, there's the first two here below which
3645 // are optimized inserts that won't zero arbitrary elements in the destination
3646 // vector. The next one matches the intrinsic and could zero arbitrary elements
3647 // in the target vector.
3648 let Constraints = "$src1 = $dst" in {
3649 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3650 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3651 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3652 !strconcat(OpcodeStr,
3653 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3655 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3657 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3658 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3659 !strconcat(OpcodeStr,
3660 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3662 (X86insrtps VR128:$src1,
3663 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3664 imm:$src3))]>, OpSize;
3668 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3670 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3671 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3673 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3674 // the intel intrinsic that corresponds to this.
3675 let Defs = [EFLAGS] in {
3676 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3677 "ptest \t{$src2, $src1|$src1, $src2}",
3678 [(set EFLAGS, (X86ptest VR128:$src1, VR128:$src2))]>,
3680 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3681 "ptest \t{$src2, $src1|$src1, $src2}",
3682 [(set EFLAGS, (X86ptest VR128:$src1, (load addr:$src2)))]>,
3686 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3687 "movntdqa\t{$src, $dst|$dst, $src}",
3688 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
3692 //===----------------------------------------------------------------------===//
3693 // SSE4.2 Instructions
3694 //===----------------------------------------------------------------------===//
3696 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3697 let Constraints = "$src1 = $dst" in {
3698 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3699 Intrinsic IntId128, bit Commutable = 0> {
3700 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3701 (ins VR128:$src1, VR128:$src2),
3702 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3703 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3705 let isCommutable = Commutable;
3707 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3708 (ins VR128:$src1, i128mem:$src2),
3709 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3711 (IntId128 VR128:$src1,
3712 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3716 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3718 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3719 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3720 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3721 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3723 // crc intrinsic instruction
3724 // This set of instructions are only rm, the only difference is the size
3726 let Constraints = "$src1 = $dst" in {
3727 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3728 (ins GR32:$src1, i8mem:$src2),
3729 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3731 (int_x86_sse42_crc32_8 GR32:$src1,
3732 (load addr:$src2)))]>;
3733 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3734 (ins GR32:$src1, GR8:$src2),
3735 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3737 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>;
3738 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3739 (ins GR32:$src1, i16mem:$src2),
3740 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3742 (int_x86_sse42_crc32_16 GR32:$src1,
3743 (load addr:$src2)))]>,
3745 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3746 (ins GR32:$src1, GR16:$src2),
3747 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3749 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3751 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3752 (ins GR32:$src1, i32mem:$src2),
3753 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3755 (int_x86_sse42_crc32_32 GR32:$src1,
3756 (load addr:$src2)))]>;
3757 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3758 (ins GR32:$src1, GR32:$src2),
3759 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3761 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>;
3762 def CRC64m8 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3763 (ins GR64:$src1, i8mem:$src2),
3764 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3766 (int_x86_sse42_crc64_8 GR64:$src1,
3767 (load addr:$src2)))]>,
3769 def CRC64r8 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3770 (ins GR64:$src1, GR8:$src2),
3771 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3773 (int_x86_sse42_crc64_8 GR64:$src1, GR8:$src2))]>,
3775 def CRC64m64 : SS42FI<0xF1, MRMSrcMem, (outs GR64:$dst),
3776 (ins GR64:$src1, i64mem:$src2),
3777 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3779 (int_x86_sse42_crc64_64 GR64:$src1,
3780 (load addr:$src2)))]>,
3782 def CRC64r64 : SS42FI<0xF1, MRMSrcReg, (outs GR64:$dst),
3783 (ins GR64:$src1, GR64:$src2),
3784 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3786 (int_x86_sse42_crc64_64 GR64:$src1, GR64:$src2))]>,
3790 // String/text processing instructions.
3791 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3792 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3793 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3794 "#PCMPISTRM128rr PSEUDO!",
3795 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3796 imm:$src3))]>, OpSize;
3797 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3798 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3799 "#PCMPISTRM128rm PSEUDO!",
3800 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3801 imm:$src3))]>, OpSize;
3804 let Defs = [XMM0, EFLAGS] in {
3805 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3806 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3807 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3808 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3809 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3810 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3813 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3814 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3815 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3816 "#PCMPESTRM128rr PSEUDO!",
3818 (int_x86_sse42_pcmpestrm128
3819 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3821 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3822 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3823 "#PCMPESTRM128rm PSEUDO!",
3824 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3825 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3829 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3830 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3831 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3832 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3833 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3834 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3835 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3838 let Defs = [ECX, EFLAGS] in {
3839 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3840 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3841 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3842 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3843 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3844 (implicit EFLAGS)]>, OpSize;
3845 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3846 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3847 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3848 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3849 (implicit EFLAGS)]>, OpSize;
3853 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3854 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3855 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3856 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3857 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3858 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3860 let Defs = [ECX, EFLAGS] in {
3861 let Uses = [EAX, EDX] in {
3862 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3863 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3864 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3865 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3866 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3867 (implicit EFLAGS)]>, OpSize;
3868 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3869 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3870 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3872 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
3873 (implicit EFLAGS)]>, OpSize;
3878 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3879 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3880 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3881 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3882 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3883 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;
3885 //===----------------------------------------------------------------------===//
3886 // AES-NI Instructions
3887 //===----------------------------------------------------------------------===//
3889 let Constraints = "$src1 = $dst" in {
3890 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr,
3891 Intrinsic IntId128, bit Commutable = 0> {
3892 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
3893 (ins VR128:$src1, VR128:$src2),
3894 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3895 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3897 let isCommutable = Commutable;
3899 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
3900 (ins VR128:$src1, i128mem:$src2),
3901 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3903 (IntId128 VR128:$src1,
3904 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3908 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
3909 int_x86_aesni_aesenc>;
3910 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
3911 int_x86_aesni_aesenclast>;
3912 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
3913 int_x86_aesni_aesdec>;
3914 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
3915 int_x86_aesni_aesdeclast>;
3917 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, VR128:$src2)),
3918 (AESENCrr VR128:$src1, VR128:$src2)>;
3919 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, (memop addr:$src2))),
3920 (AESENCrm VR128:$src1, addr:$src2)>;
3921 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, VR128:$src2)),
3922 (AESENCLASTrr VR128:$src1, VR128:$src2)>;
3923 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, (memop addr:$src2))),
3924 (AESENCLASTrm VR128:$src1, addr:$src2)>;
3925 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, VR128:$src2)),
3926 (AESDECrr VR128:$src1, VR128:$src2)>;
3927 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, (memop addr:$src2))),
3928 (AESDECrm VR128:$src1, addr:$src2)>;
3929 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, VR128:$src2)),
3930 (AESDECLASTrr VR128:$src1, VR128:$src2)>;
3931 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, (memop addr:$src2))),
3932 (AESDECLASTrm VR128:$src1, addr:$src2)>;
3934 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
3936 "aesimc\t{$src1, $dst|$dst, $src1}",
3938 (int_x86_aesni_aesimc VR128:$src1))]>,
3941 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
3942 (ins i128mem:$src1),
3943 "aesimc\t{$src1, $dst|$dst, $src1}",
3945 (int_x86_aesni_aesimc (bitconvert (memopv2i64 addr:$src1))))]>,
3948 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
3949 (ins VR128:$src1, i8imm:$src2),
3950 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3952 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
3954 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
3955 (ins i128mem:$src1, i8imm:$src2),
3956 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3958 (int_x86_aesni_aeskeygenassist (bitconvert (memopv2i64 addr:$src1)),