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 X86pextrb : SDNode<"X86ISD::PEXTRB",
40 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
41 def X86pextrw : SDNode<"X86ISD::PEXTRW",
42 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
43 def X86pinsrb : SDNode<"X86ISD::PINSRB",
44 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
45 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
46 def X86pinsrw : SDNode<"X86ISD::PINSRW",
47 SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
48 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
49 def X86insrtps : SDNode<"X86ISD::INSERTPS",
50 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
51 SDTCisVT<2, f32>, SDTCisPtrTy<3>]>>;
52 def X86vzmovl : SDNode<"X86ISD::VZEXT_MOVL",
53 SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>;
54 def X86vzload : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
55 [SDNPHasChain, SDNPMayLoad]>;
56 def X86vshl : SDNode<"X86ISD::VSHL", SDTIntShiftOp>;
57 def X86vshr : SDNode<"X86ISD::VSRL", SDTIntShiftOp>;
58 def X86cmpps : SDNode<"X86ISD::CMPPS", SDTX86VFCMP>;
59 def X86cmppd : SDNode<"X86ISD::CMPPD", SDTX86VFCMP>;
60 def X86pcmpeqb : SDNode<"X86ISD::PCMPEQB", SDTIntBinOp, [SDNPCommutative]>;
61 def X86pcmpeqw : SDNode<"X86ISD::PCMPEQW", SDTIntBinOp, [SDNPCommutative]>;
62 def X86pcmpeqd : SDNode<"X86ISD::PCMPEQD", SDTIntBinOp, [SDNPCommutative]>;
63 def X86pcmpeqq : SDNode<"X86ISD::PCMPEQQ", SDTIntBinOp, [SDNPCommutative]>;
64 def X86pcmpgtb : SDNode<"X86ISD::PCMPGTB", SDTIntBinOp>;
65 def X86pcmpgtw : SDNode<"X86ISD::PCMPGTW", SDTIntBinOp>;
66 def X86pcmpgtd : SDNode<"X86ISD::PCMPGTD", SDTIntBinOp>;
67 def X86pcmpgtq : SDNode<"X86ISD::PCMPGTQ", SDTIntBinOp>;
69 //===----------------------------------------------------------------------===//
70 // SSE Complex Patterns
71 //===----------------------------------------------------------------------===//
73 // These are 'extloads' from a scalar to the low element of a vector, zeroing
74 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
76 def sse_load_f32 : ComplexPattern<v4f32, 4, "SelectScalarSSELoad", [],
77 [SDNPHasChain, SDNPMayLoad]>;
78 def sse_load_f64 : ComplexPattern<v2f64, 4, "SelectScalarSSELoad", [],
79 [SDNPHasChain, SDNPMayLoad]>;
81 def ssmem : Operand<v4f32> {
82 let PrintMethod = "printf32mem";
83 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
85 def sdmem : Operand<v2f64> {
86 let PrintMethod = "printf64mem";
87 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
90 //===----------------------------------------------------------------------===//
91 // SSE pattern fragments
92 //===----------------------------------------------------------------------===//
94 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
95 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
96 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
97 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
99 // Like 'store', but always requires vector alignment.
100 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
101 (st node:$val, node:$ptr), [{
102 StoreSDNode *ST = cast<StoreSDNode>(N);
103 return !ST->isTruncatingStore() &&
104 ST->getAddressingMode() == ISD::UNINDEXED &&
105 ST->getAlignment() >= 16;
108 // Like 'load', but always requires vector alignment.
109 def alignedload : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
110 LoadSDNode *LD = cast<LoadSDNode>(N);
111 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
112 LD->getAddressingMode() == ISD::UNINDEXED &&
113 LD->getAlignment() >= 16;
116 def alignedloadfsf32 : PatFrag<(ops node:$ptr), (f32 (alignedload node:$ptr))>;
117 def alignedloadfsf64 : PatFrag<(ops node:$ptr), (f64 (alignedload node:$ptr))>;
118 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
119 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
120 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
121 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
123 // Like 'load', but uses special alignment checks suitable for use in
124 // memory operands in most SSE instructions, which are required to
125 // be naturally aligned on some targets but not on others.
126 // FIXME: Actually implement support for targets that don't require the
127 // alignment. This probably wants a subtarget predicate.
128 def memop : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
129 LoadSDNode *LD = cast<LoadSDNode>(N);
130 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
131 LD->getAddressingMode() == ISD::UNINDEXED &&
132 LD->getAlignment() >= 16;
135 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
136 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
137 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
138 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
139 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
140 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
141 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
143 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
145 // FIXME: 8 byte alignment for mmx reads is not required
146 def memop64 : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
147 LoadSDNode *LD = cast<LoadSDNode>(N);
148 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
149 LD->getAddressingMode() == ISD::UNINDEXED &&
150 LD->getAlignment() >= 8;
153 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
154 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
155 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
156 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
158 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
159 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
160 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
161 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
162 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
163 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
165 def vzmovl_v2i64 : PatFrag<(ops node:$src),
166 (bitconvert (v2i64 (X86vzmovl
167 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
168 def vzmovl_v4i32 : PatFrag<(ops node:$src),
169 (bitconvert (v4i32 (X86vzmovl
170 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
172 def vzload_v2i64 : PatFrag<(ops node:$src),
173 (bitconvert (v2i64 (X86vzload node:$src)))>;
176 def fp32imm0 : PatLeaf<(f32 fpimm), [{
177 return N->isExactlyValue(+0.0);
180 def PSxLDQ_imm : SDNodeXForm<imm, [{
181 // Transformation function: imm >> 3
182 return getI32Imm(N->getZExtValue() >> 3);
185 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
187 def SHUFFLE_get_shuf_imm : SDNodeXForm<build_vector, [{
188 return getI8Imm(X86::getShuffleSHUFImmediate(N));
191 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
193 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<build_vector, [{
194 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
197 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
199 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<build_vector, [{
200 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
203 def SSE_splat_mask : PatLeaf<(build_vector), [{
204 return X86::isSplatMask(N);
205 }], SHUFFLE_get_shuf_imm>;
207 def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
208 return X86::isSplatLoMask(N);
211 def MOVDDUP_shuffle_mask : PatLeaf<(build_vector), [{
212 return X86::isMOVDDUPMask(N);
215 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
216 return X86::isMOVHLPSMask(N);
219 def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
220 return X86::isMOVHLPS_v_undef_Mask(N);
223 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
224 return X86::isMOVHPMask(N);
227 def MOVLP_shuffle_mask : PatLeaf<(build_vector), [{
228 return X86::isMOVLPMask(N);
231 def MOVL_shuffle_mask : PatLeaf<(build_vector), [{
232 return X86::isMOVLMask(N);
235 def MOVSHDUP_shuffle_mask : PatLeaf<(build_vector), [{
236 return X86::isMOVSHDUPMask(N);
239 def MOVSLDUP_shuffle_mask : PatLeaf<(build_vector), [{
240 return X86::isMOVSLDUPMask(N);
243 def UNPCKL_shuffle_mask : PatLeaf<(build_vector), [{
244 return X86::isUNPCKLMask(N);
247 def UNPCKH_shuffle_mask : PatLeaf<(build_vector), [{
248 return X86::isUNPCKHMask(N);
251 def UNPCKL_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
252 return X86::isUNPCKL_v_undef_Mask(N);
255 def UNPCKH_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
256 return X86::isUNPCKH_v_undef_Mask(N);
259 def PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
260 return X86::isPSHUFDMask(N);
261 }], SHUFFLE_get_shuf_imm>;
263 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
264 return X86::isPSHUFHWMask(N);
265 }], SHUFFLE_get_pshufhw_imm>;
267 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
268 return X86::isPSHUFLWMask(N);
269 }], SHUFFLE_get_pshuflw_imm>;
271 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
272 return X86::isPSHUFDMask(N);
273 }], SHUFFLE_get_shuf_imm>;
275 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
276 return X86::isSHUFPMask(N);
277 }], SHUFFLE_get_shuf_imm>;
279 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
280 return X86::isSHUFPMask(N);
281 }], SHUFFLE_get_shuf_imm>;
284 //===----------------------------------------------------------------------===//
285 // SSE scalar FP Instructions
286 //===----------------------------------------------------------------------===//
288 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
289 // scheduler into a branch sequence.
290 // These are expanded by the scheduler.
291 let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
292 def CMOV_FR32 : I<0, Pseudo,
293 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
294 "#CMOV_FR32 PSEUDO!",
295 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
297 def CMOV_FR64 : I<0, Pseudo,
298 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
299 "#CMOV_FR64 PSEUDO!",
300 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
302 def CMOV_V4F32 : I<0, Pseudo,
303 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
304 "#CMOV_V4F32 PSEUDO!",
306 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
308 def CMOV_V2F64 : I<0, Pseudo,
309 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
310 "#CMOV_V2F64 PSEUDO!",
312 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
314 def CMOV_V2I64 : I<0, Pseudo,
315 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
316 "#CMOV_V2I64 PSEUDO!",
318 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
322 //===----------------------------------------------------------------------===//
324 //===----------------------------------------------------------------------===//
327 let neverHasSideEffects = 1 in
328 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
329 "movss\t{$src, $dst|$dst, $src}", []>;
330 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
331 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
332 "movss\t{$src, $dst|$dst, $src}",
333 [(set FR32:$dst, (loadf32 addr:$src))]>;
334 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
335 "movss\t{$src, $dst|$dst, $src}",
336 [(store FR32:$src, addr:$dst)]>;
338 // Conversion instructions
339 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
340 "cvttss2si\t{$src, $dst|$dst, $src}",
341 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
342 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
343 "cvttss2si\t{$src, $dst|$dst, $src}",
344 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
345 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
346 "cvtsi2ss\t{$src, $dst|$dst, $src}",
347 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
348 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
349 "cvtsi2ss\t{$src, $dst|$dst, $src}",
350 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
352 // Match intrinsics which expect XMM operand(s).
353 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
354 "cvtss2si\t{$src, $dst|$dst, $src}",
355 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
356 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
357 "cvtss2si\t{$src, $dst|$dst, $src}",
358 [(set GR32:$dst, (int_x86_sse_cvtss2si
359 (load addr:$src)))]>;
361 // Match intrinisics which expect MM and XMM operand(s).
362 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
363 "cvtps2pi\t{$src, $dst|$dst, $src}",
364 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
365 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
366 "cvtps2pi\t{$src, $dst|$dst, $src}",
367 [(set VR64:$dst, (int_x86_sse_cvtps2pi
368 (load addr:$src)))]>;
369 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
370 "cvttps2pi\t{$src, $dst|$dst, $src}",
371 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
372 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
373 "cvttps2pi\t{$src, $dst|$dst, $src}",
374 [(set VR64:$dst, (int_x86_sse_cvttps2pi
375 (load addr:$src)))]>;
376 let Constraints = "$src1 = $dst" in {
377 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
378 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
379 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
380 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
382 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
383 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
384 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
385 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
386 (load addr:$src2)))]>;
389 // Aliases for intrinsics
390 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
391 "cvttss2si\t{$src, $dst|$dst, $src}",
393 (int_x86_sse_cvttss2si VR128:$src))]>;
394 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
395 "cvttss2si\t{$src, $dst|$dst, $src}",
397 (int_x86_sse_cvttss2si(load addr:$src)))]>;
399 let Constraints = "$src1 = $dst" in {
400 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
401 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
402 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
403 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
405 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
406 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
407 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
408 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
409 (loadi32 addr:$src2)))]>;
412 // Comparison instructions
413 let Constraints = "$src1 = $dst" in {
414 let neverHasSideEffects = 1 in
415 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
416 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
417 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
418 let neverHasSideEffects = 1, mayLoad = 1 in
419 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
420 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
421 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
424 let Defs = [EFLAGS] in {
425 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
426 "ucomiss\t{$src2, $src1|$src1, $src2}",
427 [(X86cmp FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
428 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
429 "ucomiss\t{$src2, $src1|$src1, $src2}",
430 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
434 // Aliases to match intrinsics which expect XMM operand(s).
435 let Constraints = "$src1 = $dst" in {
436 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
437 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
438 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
439 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
440 VR128:$src, imm:$cc))]>;
441 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
442 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src, SSECC:$cc),
443 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
444 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
445 (load addr:$src), imm:$cc))]>;
448 let Defs = [EFLAGS] in {
449 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs),
450 (ins VR128:$src1, VR128:$src2),
451 "ucomiss\t{$src2, $src1|$src1, $src2}",
452 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
454 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),
455 (ins VR128:$src1, f128mem:$src2),
456 "ucomiss\t{$src2, $src1|$src1, $src2}",
457 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
460 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs),
461 (ins VR128:$src1, VR128:$src2),
462 "comiss\t{$src2, $src1|$src1, $src2}",
463 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
465 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs),
466 (ins VR128:$src1, f128mem:$src2),
467 "comiss\t{$src2, $src1|$src1, $src2}",
468 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
472 // Aliases of packed SSE1 instructions for scalar use. These all have names that
475 // Alias instructions that map fld0 to pxor for sse.
476 let isReMaterializable = 1, isAsCheapAsAMove = 1 in
477 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
478 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
479 Requires<[HasSSE1]>, TB, OpSize;
481 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
483 let neverHasSideEffects = 1 in
484 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
485 "movaps\t{$src, $dst|$dst, $src}", []>;
487 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
489 let isSimpleLoad = 1 in
490 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
491 "movaps\t{$src, $dst|$dst, $src}",
492 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
494 // Alias bitwise logical operations using SSE logical ops on packed FP values.
495 let Constraints = "$src1 = $dst" in {
496 let isCommutable = 1 in {
497 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
498 "andps\t{$src2, $dst|$dst, $src2}",
499 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
500 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
501 "orps\t{$src2, $dst|$dst, $src2}",
502 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
503 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
504 "xorps\t{$src2, $dst|$dst, $src2}",
505 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
508 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
509 "andps\t{$src2, $dst|$dst, $src2}",
510 [(set FR32:$dst, (X86fand FR32:$src1,
511 (memopfsf32 addr:$src2)))]>;
512 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
513 "orps\t{$src2, $dst|$dst, $src2}",
514 [(set FR32:$dst, (X86for FR32:$src1,
515 (memopfsf32 addr:$src2)))]>;
516 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
517 "xorps\t{$src2, $dst|$dst, $src2}",
518 [(set FR32:$dst, (X86fxor FR32:$src1,
519 (memopfsf32 addr:$src2)))]>;
520 let neverHasSideEffects = 1 in {
521 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
522 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
523 "andnps\t{$src2, $dst|$dst, $src2}", []>;
526 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
527 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
528 "andnps\t{$src2, $dst|$dst, $src2}", []>;
532 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
534 /// In addition, we also have a special variant of the scalar form here to
535 /// represent the associated intrinsic operation. This form is unlike the
536 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
537 /// and leaves the top elements undefined.
539 /// These three forms can each be reg+reg or reg+mem, so there are a total of
540 /// six "instructions".
542 let Constraints = "$src1 = $dst" in {
543 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
544 SDNode OpNode, Intrinsic F32Int,
545 bit Commutable = 0> {
546 // Scalar operation, reg+reg.
547 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
548 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
549 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
550 let isCommutable = Commutable;
553 // Scalar operation, reg+mem.
554 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
555 (ins FR32:$src1, f32mem:$src2),
556 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
557 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
559 // Vector operation, reg+reg.
560 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
561 (ins VR128:$src1, VR128:$src2),
562 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
563 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
564 let isCommutable = Commutable;
567 // Vector operation, reg+mem.
568 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
569 (ins VR128:$src1, f128mem:$src2),
570 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
571 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
573 // Intrinsic operation, reg+reg.
574 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
575 (ins VR128:$src1, VR128:$src2),
576 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
577 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
578 let isCommutable = Commutable;
581 // Intrinsic operation, reg+mem.
582 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
583 (ins VR128:$src1, ssmem:$src2),
584 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
585 [(set VR128:$dst, (F32Int VR128:$src1,
586 sse_load_f32:$src2))]>;
590 // Arithmetic instructions
591 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
592 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
593 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
594 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
596 /// sse1_fp_binop_rm - Other SSE1 binops
598 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
599 /// instructions for a full-vector intrinsic form. Operations that map
600 /// onto C operators don't use this form since they just use the plain
601 /// vector form instead of having a separate vector intrinsic form.
603 /// This provides a total of eight "instructions".
605 let Constraints = "$src1 = $dst" in {
606 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
610 bit Commutable = 0> {
612 // Scalar operation, reg+reg.
613 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
614 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
615 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
616 let isCommutable = Commutable;
619 // Scalar operation, reg+mem.
620 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
621 (ins FR32:$src1, f32mem:$src2),
622 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
623 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
625 // Vector operation, reg+reg.
626 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
627 (ins VR128:$src1, VR128:$src2),
628 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
629 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
630 let isCommutable = Commutable;
633 // Vector operation, reg+mem.
634 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
635 (ins VR128:$src1, f128mem:$src2),
636 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
637 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
639 // Intrinsic operation, reg+reg.
640 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
641 (ins VR128:$src1, VR128:$src2),
642 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
643 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
644 let isCommutable = Commutable;
647 // Intrinsic operation, reg+mem.
648 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
649 (ins VR128:$src1, ssmem:$src2),
650 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
651 [(set VR128:$dst, (F32Int VR128:$src1,
652 sse_load_f32:$src2))]>;
654 // Vector intrinsic operation, reg+reg.
655 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
656 (ins VR128:$src1, VR128:$src2),
657 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
658 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
659 let isCommutable = Commutable;
662 // Vector intrinsic operation, reg+mem.
663 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
664 (ins VR128:$src1, f128mem:$src2),
665 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
666 [(set VR128:$dst, (V4F32Int VR128:$src1, (memopv4f32 addr:$src2)))]>;
670 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
671 int_x86_sse_max_ss, int_x86_sse_max_ps>;
672 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
673 int_x86_sse_min_ss, int_x86_sse_min_ps>;
675 //===----------------------------------------------------------------------===//
676 // SSE packed FP Instructions
679 let neverHasSideEffects = 1 in
680 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
681 "movaps\t{$src, $dst|$dst, $src}", []>;
682 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
683 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
684 "movaps\t{$src, $dst|$dst, $src}",
685 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
687 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
688 "movaps\t{$src, $dst|$dst, $src}",
689 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
691 let neverHasSideEffects = 1 in
692 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
693 "movups\t{$src, $dst|$dst, $src}", []>;
694 let isSimpleLoad = 1 in
695 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
696 "movups\t{$src, $dst|$dst, $src}",
697 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
698 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
699 "movups\t{$src, $dst|$dst, $src}",
700 [(store (v4f32 VR128:$src), addr:$dst)]>;
702 // Intrinsic forms of MOVUPS load and store
703 let isSimpleLoad = 1 in
704 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
705 "movups\t{$src, $dst|$dst, $src}",
706 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
707 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
708 "movups\t{$src, $dst|$dst, $src}",
709 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
711 let Constraints = "$src1 = $dst" in {
712 let AddedComplexity = 20 in {
713 def MOVLPSrm : PSI<0x12, MRMSrcMem,
714 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
715 "movlps\t{$src2, $dst|$dst, $src2}",
717 (v4f32 (vector_shuffle VR128:$src1,
718 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
719 MOVLP_shuffle_mask)))]>;
720 def MOVHPSrm : PSI<0x16, MRMSrcMem,
721 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
722 "movhps\t{$src2, $dst|$dst, $src2}",
724 (v4f32 (vector_shuffle VR128:$src1,
725 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
726 MOVHP_shuffle_mask)))]>;
728 } // Constraints = "$src1 = $dst"
731 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
732 "movlps\t{$src, $dst|$dst, $src}",
733 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
734 (iPTR 0))), addr:$dst)]>;
736 // v2f64 extract element 1 is always custom lowered to unpack high to low
737 // and extract element 0 so the non-store version isn't too horrible.
738 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
739 "movhps\t{$src, $dst|$dst, $src}",
740 [(store (f64 (vector_extract
741 (v2f64 (vector_shuffle
742 (bc_v2f64 (v4f32 VR128:$src)), (undef),
743 UNPCKH_shuffle_mask)), (iPTR 0))),
746 let Constraints = "$src1 = $dst" in {
747 let AddedComplexity = 20 in {
748 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
749 "movlhps\t{$src2, $dst|$dst, $src2}",
751 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
752 MOVHP_shuffle_mask)))]>;
754 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
755 "movhlps\t{$src2, $dst|$dst, $src2}",
757 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
758 MOVHLPS_shuffle_mask)))]>;
760 } // Constraints = "$src1 = $dst"
762 let AddedComplexity = 20 in
763 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef), MOVDDUP_shuffle_mask)),
764 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
771 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
773 /// In addition, we also have a special variant of the scalar form here to
774 /// represent the associated intrinsic operation. This form is unlike the
775 /// plain scalar form, in that it takes an entire vector (instead of a
776 /// scalar) and leaves the top elements undefined.
778 /// And, we have a special variant form for a full-vector intrinsic form.
780 /// These four forms can each have a reg or a mem operand, so there are a
781 /// total of eight "instructions".
783 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
787 bit Commutable = 0> {
788 // Scalar operation, reg.
789 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
790 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
791 [(set FR32:$dst, (OpNode FR32:$src))]> {
792 let isCommutable = Commutable;
795 // Scalar operation, mem.
796 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
797 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
798 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
800 // Vector operation, reg.
801 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
802 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
803 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
804 let isCommutable = Commutable;
807 // Vector operation, mem.
808 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
809 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
810 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
812 // Intrinsic operation, reg.
813 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
814 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
815 [(set VR128:$dst, (F32Int VR128:$src))]> {
816 let isCommutable = Commutable;
819 // Intrinsic operation, mem.
820 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
821 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
822 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
824 // Vector intrinsic operation, reg
825 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
826 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
827 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
828 let isCommutable = Commutable;
831 // Vector intrinsic operation, mem
832 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
833 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
834 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
838 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
839 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
841 // Reciprocal approximations. Note that these typically require refinement
842 // in order to obtain suitable precision.
843 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
844 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
845 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
846 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
849 let Constraints = "$src1 = $dst" in {
850 let isCommutable = 1 in {
851 def ANDPSrr : PSI<0x54, MRMSrcReg,
852 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
853 "andps\t{$src2, $dst|$dst, $src2}",
854 [(set VR128:$dst, (v2i64
855 (and VR128:$src1, VR128:$src2)))]>;
856 def ORPSrr : PSI<0x56, MRMSrcReg,
857 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
858 "orps\t{$src2, $dst|$dst, $src2}",
859 [(set VR128:$dst, (v2i64
860 (or VR128:$src1, VR128:$src2)))]>;
861 def XORPSrr : PSI<0x57, MRMSrcReg,
862 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
863 "xorps\t{$src2, $dst|$dst, $src2}",
864 [(set VR128:$dst, (v2i64
865 (xor VR128:$src1, VR128:$src2)))]>;
868 def ANDPSrm : PSI<0x54, MRMSrcMem,
869 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
870 "andps\t{$src2, $dst|$dst, $src2}",
871 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
872 (memopv2i64 addr:$src2)))]>;
873 def ORPSrm : PSI<0x56, MRMSrcMem,
874 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
875 "orps\t{$src2, $dst|$dst, $src2}",
876 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
877 (memopv2i64 addr:$src2)))]>;
878 def XORPSrm : PSI<0x57, MRMSrcMem,
879 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
880 "xorps\t{$src2, $dst|$dst, $src2}",
881 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
882 (memopv2i64 addr:$src2)))]>;
883 def ANDNPSrr : PSI<0x55, MRMSrcReg,
884 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
885 "andnps\t{$src2, $dst|$dst, $src2}",
887 (v2i64 (and (xor VR128:$src1,
888 (bc_v2i64 (v4i32 immAllOnesV))),
890 def ANDNPSrm : PSI<0x55, MRMSrcMem,
891 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
892 "andnps\t{$src2, $dst|$dst, $src2}",
894 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
895 (bc_v2i64 (v4i32 immAllOnesV))),
896 (memopv2i64 addr:$src2))))]>;
899 let Constraints = "$src1 = $dst" in {
900 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
901 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
902 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
903 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
904 VR128:$src, imm:$cc))]>;
905 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
906 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
907 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
908 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
909 (memop addr:$src), imm:$cc))]>;
911 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
912 (CMPPSrri VR128:$src1, VR128:$src2, imm:$cc)>;
913 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
914 (CMPPSrmi VR128:$src1, addr:$src2, imm:$cc)>;
916 // Shuffle and unpack instructions
917 let Constraints = "$src1 = $dst" in {
918 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
919 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
920 (outs VR128:$dst), (ins VR128:$src1,
921 VR128:$src2, i32i8imm:$src3),
922 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
924 (v4f32 (vector_shuffle
925 VR128:$src1, VR128:$src2,
926 SHUFP_shuffle_mask:$src3)))]>;
927 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
928 (outs VR128:$dst), (ins VR128:$src1,
929 f128mem:$src2, i32i8imm:$src3),
930 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
932 (v4f32 (vector_shuffle
933 VR128:$src1, (memopv4f32 addr:$src2),
934 SHUFP_shuffle_mask:$src3)))]>;
936 let AddedComplexity = 10 in {
937 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
938 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
939 "unpckhps\t{$src2, $dst|$dst, $src2}",
941 (v4f32 (vector_shuffle
942 VR128:$src1, VR128:$src2,
943 UNPCKH_shuffle_mask)))]>;
944 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
945 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
946 "unpckhps\t{$src2, $dst|$dst, $src2}",
948 (v4f32 (vector_shuffle
949 VR128:$src1, (memopv4f32 addr:$src2),
950 UNPCKH_shuffle_mask)))]>;
952 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
953 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
954 "unpcklps\t{$src2, $dst|$dst, $src2}",
956 (v4f32 (vector_shuffle
957 VR128:$src1, VR128:$src2,
958 UNPCKL_shuffle_mask)))]>;
959 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
960 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
961 "unpcklps\t{$src2, $dst|$dst, $src2}",
963 (v4f32 (vector_shuffle
964 VR128:$src1, (memopv4f32 addr:$src2),
965 UNPCKL_shuffle_mask)))]>;
967 } // Constraints = "$src1 = $dst"
970 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
971 "movmskps\t{$src, $dst|$dst, $src}",
972 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
973 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
974 "movmskpd\t{$src, $dst|$dst, $src}",
975 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
977 // Prefetch intrinsic.
978 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
979 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
980 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
981 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
982 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
983 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
984 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
985 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
987 // Non-temporal stores
988 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
989 "movntps\t{$src, $dst|$dst, $src}",
990 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
992 // Load, store, and memory fence
993 def SFENCE : PSI<0xAE, MRM7m, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
996 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
997 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
998 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
999 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1001 // Alias instructions that map zero vector to pxor / xorp* for sse.
1002 let isReMaterializable = 1 in
1003 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
1004 "xorps\t$dst, $dst",
1005 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1007 let Predicates = [HasSSE1] in {
1008 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
1009 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
1010 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
1011 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
1012 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
1015 // FR32 to 128-bit vector conversion.
1016 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
1017 "movss\t{$src, $dst|$dst, $src}",
1019 (v4f32 (scalar_to_vector FR32:$src)))]>;
1020 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1021 "movss\t{$src, $dst|$dst, $src}",
1023 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
1025 // FIXME: may not be able to eliminate this movss with coalescing the src and
1026 // dest register classes are different. We really want to write this pattern
1028 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1029 // (f32 FR32:$src)>;
1030 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
1031 "movss\t{$src, $dst|$dst, $src}",
1032 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
1034 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
1035 "movss\t{$src, $dst|$dst, $src}",
1036 [(store (f32 (vector_extract (v4f32 VR128:$src),
1037 (iPTR 0))), addr:$dst)]>;
1040 // Move to lower bits of a VR128, leaving upper bits alone.
1041 // Three operand (but two address) aliases.
1042 let Constraints = "$src1 = $dst" in {
1043 let neverHasSideEffects = 1 in
1044 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
1045 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
1046 "movss\t{$src2, $dst|$dst, $src2}", []>;
1048 let AddedComplexity = 15 in
1049 def MOVLPSrr : SSI<0x10, MRMSrcReg,
1050 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1051 "movss\t{$src2, $dst|$dst, $src2}",
1053 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
1054 MOVL_shuffle_mask)))]>;
1057 // Move to lower bits of a VR128 and zeroing upper bits.
1058 // Loading from memory automatically zeroing upper bits.
1059 let AddedComplexity = 20 in
1060 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1061 "movss\t{$src, $dst|$dst, $src}",
1062 [(set VR128:$dst, (v4f32 (X86vzmovl (v4f32 (scalar_to_vector
1063 (loadf32 addr:$src))))))]>;
1065 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
1066 (MOVZSS2PSrm addr:$src)>;
1068 //===----------------------------------------------------------------------===//
1069 // SSE2 Instructions
1070 //===----------------------------------------------------------------------===//
1072 // Move Instructions
1073 let neverHasSideEffects = 1 in
1074 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1075 "movsd\t{$src, $dst|$dst, $src}", []>;
1076 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1077 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1078 "movsd\t{$src, $dst|$dst, $src}",
1079 [(set FR64:$dst, (loadf64 addr:$src))]>;
1080 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1081 "movsd\t{$src, $dst|$dst, $src}",
1082 [(store FR64:$src, addr:$dst)]>;
1084 // Conversion instructions
1085 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1086 "cvttsd2si\t{$src, $dst|$dst, $src}",
1087 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1088 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1089 "cvttsd2si\t{$src, $dst|$dst, $src}",
1090 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1091 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1092 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1093 [(set FR32:$dst, (fround FR64:$src))]>;
1094 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1095 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1096 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1097 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1098 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1099 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1100 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1101 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1102 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1104 // SSE2 instructions with XS prefix
1105 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1106 "cvtss2sd\t{$src, $dst|$dst, $src}",
1107 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1108 Requires<[HasSSE2]>;
1109 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1110 "cvtss2sd\t{$src, $dst|$dst, $src}",
1111 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1112 Requires<[HasSSE2]>;
1114 // Match intrinsics which expect XMM operand(s).
1115 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1116 "cvtsd2si\t{$src, $dst|$dst, $src}",
1117 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1118 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1119 "cvtsd2si\t{$src, $dst|$dst, $src}",
1120 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1121 (load addr:$src)))]>;
1123 // Match intrinisics which expect MM and XMM operand(s).
1124 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1125 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1126 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1127 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1128 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1129 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1130 (memop addr:$src)))]>;
1131 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1132 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1133 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1134 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1135 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1136 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1137 (memop addr:$src)))]>;
1138 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1139 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1140 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1141 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1142 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1143 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1144 (load addr:$src)))]>;
1146 // Aliases for intrinsics
1147 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1148 "cvttsd2si\t{$src, $dst|$dst, $src}",
1150 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1151 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1152 "cvttsd2si\t{$src, $dst|$dst, $src}",
1153 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1154 (load addr:$src)))]>;
1156 // Comparison instructions
1157 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1158 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1159 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1160 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1162 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1163 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1164 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1167 let Defs = [EFLAGS] in {
1168 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1169 "ucomisd\t{$src2, $src1|$src1, $src2}",
1170 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1171 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1172 "ucomisd\t{$src2, $src1|$src1, $src2}",
1173 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1174 (implicit EFLAGS)]>;
1177 // Aliases to match intrinsics which expect XMM operand(s).
1178 let Constraints = "$src1 = $dst" in {
1179 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1180 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1181 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1182 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1183 VR128:$src, imm:$cc))]>;
1184 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1185 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1186 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1187 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1188 (load addr:$src), imm:$cc))]>;
1191 let Defs = [EFLAGS] in {
1192 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1193 "ucomisd\t{$src2, $src1|$src1, $src2}",
1194 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1195 (implicit EFLAGS)]>;
1196 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1197 "ucomisd\t{$src2, $src1|$src1, $src2}",
1198 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1199 (implicit EFLAGS)]>;
1201 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1202 "comisd\t{$src2, $src1|$src1, $src2}",
1203 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1204 (implicit EFLAGS)]>;
1205 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1206 "comisd\t{$src2, $src1|$src1, $src2}",
1207 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1208 (implicit EFLAGS)]>;
1211 // Aliases of packed SSE2 instructions for scalar use. These all have names that
1214 // Alias instructions that map fld0 to pxor for sse.
1215 let isReMaterializable = 1, isAsCheapAsAMove = 1 in
1216 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1217 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1218 Requires<[HasSSE2]>, TB, OpSize;
1220 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1222 let neverHasSideEffects = 1 in
1223 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1224 "movapd\t{$src, $dst|$dst, $src}", []>;
1226 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1228 let isSimpleLoad = 1 in
1229 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1230 "movapd\t{$src, $dst|$dst, $src}",
1231 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1233 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1234 let Constraints = "$src1 = $dst" in {
1235 let isCommutable = 1 in {
1236 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1237 (ins FR64:$src1, FR64:$src2),
1238 "andpd\t{$src2, $dst|$dst, $src2}",
1239 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1240 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1241 (ins FR64:$src1, FR64:$src2),
1242 "orpd\t{$src2, $dst|$dst, $src2}",
1243 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1244 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1245 (ins FR64:$src1, FR64:$src2),
1246 "xorpd\t{$src2, $dst|$dst, $src2}",
1247 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1250 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1251 (ins FR64:$src1, f128mem:$src2),
1252 "andpd\t{$src2, $dst|$dst, $src2}",
1253 [(set FR64:$dst, (X86fand FR64:$src1,
1254 (memopfsf64 addr:$src2)))]>;
1255 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1256 (ins FR64:$src1, f128mem:$src2),
1257 "orpd\t{$src2, $dst|$dst, $src2}",
1258 [(set FR64:$dst, (X86for FR64:$src1,
1259 (memopfsf64 addr:$src2)))]>;
1260 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1261 (ins FR64:$src1, f128mem:$src2),
1262 "xorpd\t{$src2, $dst|$dst, $src2}",
1263 [(set FR64:$dst, (X86fxor FR64:$src1,
1264 (memopfsf64 addr:$src2)))]>;
1266 let neverHasSideEffects = 1 in {
1267 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1268 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1269 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1271 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1272 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1273 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1277 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1279 /// In addition, we also have a special variant of the scalar form here to
1280 /// represent the associated intrinsic operation. This form is unlike the
1281 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1282 /// and leaves the top elements undefined.
1284 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1285 /// six "instructions".
1287 let Constraints = "$src1 = $dst" in {
1288 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1289 SDNode OpNode, Intrinsic F64Int,
1290 bit Commutable = 0> {
1291 // Scalar operation, reg+reg.
1292 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1293 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1294 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1295 let isCommutable = Commutable;
1298 // Scalar operation, reg+mem.
1299 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1300 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1301 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1303 // Vector operation, reg+reg.
1304 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1305 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1306 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1307 let isCommutable = Commutable;
1310 // Vector operation, reg+mem.
1311 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1312 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1313 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1315 // Intrinsic operation, reg+reg.
1316 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1317 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1318 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1319 let isCommutable = Commutable;
1322 // Intrinsic operation, reg+mem.
1323 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1324 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1325 [(set VR128:$dst, (F64Int VR128:$src1,
1326 sse_load_f64:$src2))]>;
1330 // Arithmetic instructions
1331 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1332 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1333 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1334 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1336 /// sse2_fp_binop_rm - Other SSE2 binops
1338 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1339 /// instructions for a full-vector intrinsic form. Operations that map
1340 /// onto C operators don't use this form since they just use the plain
1341 /// vector form instead of having a separate vector intrinsic form.
1343 /// This provides a total of eight "instructions".
1345 let Constraints = "$src1 = $dst" in {
1346 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1350 bit Commutable = 0> {
1352 // Scalar operation, reg+reg.
1353 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1354 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1355 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1356 let isCommutable = Commutable;
1359 // Scalar operation, reg+mem.
1360 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1361 (ins FR64:$src1, f64mem:$src2),
1362 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1363 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1365 // Vector operation, reg+reg.
1366 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1367 (ins VR128:$src1, VR128:$src2),
1368 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1369 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1370 let isCommutable = Commutable;
1373 // Vector operation, reg+mem.
1374 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1375 (ins VR128:$src1, f128mem:$src2),
1376 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1377 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1379 // Intrinsic operation, reg+reg.
1380 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1381 (ins VR128:$src1, VR128:$src2),
1382 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1383 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1384 let isCommutable = Commutable;
1387 // Intrinsic operation, reg+mem.
1388 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1389 (ins VR128:$src1, sdmem:$src2),
1390 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1391 [(set VR128:$dst, (F64Int VR128:$src1,
1392 sse_load_f64:$src2))]>;
1394 // Vector intrinsic operation, reg+reg.
1395 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1396 (ins VR128:$src1, VR128:$src2),
1397 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1398 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1399 let isCommutable = Commutable;
1402 // Vector intrinsic operation, reg+mem.
1403 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1404 (ins VR128:$src1, f128mem:$src2),
1405 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1406 [(set VR128:$dst, (V2F64Int VR128:$src1,
1407 (memopv2f64 addr:$src2)))]>;
1411 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1412 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1413 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1414 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1416 //===----------------------------------------------------------------------===//
1417 // SSE packed FP Instructions
1419 // Move Instructions
1420 let neverHasSideEffects = 1 in
1421 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1422 "movapd\t{$src, $dst|$dst, $src}", []>;
1423 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1424 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1425 "movapd\t{$src, $dst|$dst, $src}",
1426 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1428 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1429 "movapd\t{$src, $dst|$dst, $src}",
1430 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1432 let neverHasSideEffects = 1 in
1433 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1434 "movupd\t{$src, $dst|$dst, $src}", []>;
1435 let isSimpleLoad = 1 in
1436 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1437 "movupd\t{$src, $dst|$dst, $src}",
1438 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1439 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1440 "movupd\t{$src, $dst|$dst, $src}",
1441 [(store (v2f64 VR128:$src), addr:$dst)]>;
1443 // Intrinsic forms of MOVUPD load and store
1444 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1445 "movupd\t{$src, $dst|$dst, $src}",
1446 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1447 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1448 "movupd\t{$src, $dst|$dst, $src}",
1449 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1451 let Constraints = "$src1 = $dst" in {
1452 let AddedComplexity = 20 in {
1453 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1454 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1455 "movlpd\t{$src2, $dst|$dst, $src2}",
1457 (v2f64 (vector_shuffle VR128:$src1,
1458 (scalar_to_vector (loadf64 addr:$src2)),
1459 MOVLP_shuffle_mask)))]>;
1460 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1461 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1462 "movhpd\t{$src2, $dst|$dst, $src2}",
1464 (v2f64 (vector_shuffle VR128:$src1,
1465 (scalar_to_vector (loadf64 addr:$src2)),
1466 MOVHP_shuffle_mask)))]>;
1467 } // AddedComplexity
1468 } // Constraints = "$src1 = $dst"
1470 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1471 "movlpd\t{$src, $dst|$dst, $src}",
1472 [(store (f64 (vector_extract (v2f64 VR128:$src),
1473 (iPTR 0))), addr:$dst)]>;
1475 // v2f64 extract element 1 is always custom lowered to unpack high to low
1476 // and extract element 0 so the non-store version isn't too horrible.
1477 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1478 "movhpd\t{$src, $dst|$dst, $src}",
1479 [(store (f64 (vector_extract
1480 (v2f64 (vector_shuffle VR128:$src, (undef),
1481 UNPCKH_shuffle_mask)), (iPTR 0))),
1484 // SSE2 instructions without OpSize prefix
1485 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1486 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1487 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1488 TB, Requires<[HasSSE2]>;
1489 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1490 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1491 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1492 (bitconvert (memopv2i64 addr:$src))))]>,
1493 TB, Requires<[HasSSE2]>;
1495 // SSE2 instructions with XS prefix
1496 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1497 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1498 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1499 XS, Requires<[HasSSE2]>;
1500 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1501 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1502 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1503 (bitconvert (memopv2i64 addr:$src))))]>,
1504 XS, Requires<[HasSSE2]>;
1506 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1507 "cvtps2dq\t{$src, $dst|$dst, $src}",
1508 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1509 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1510 "cvtps2dq\t{$src, $dst|$dst, $src}",
1511 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1512 (memop addr:$src)))]>;
1513 // SSE2 packed instructions with XS prefix
1514 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1515 "cvttps2dq\t{$src, $dst|$dst, $src}",
1516 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1517 XS, Requires<[HasSSE2]>;
1518 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1519 "cvttps2dq\t{$src, $dst|$dst, $src}",
1520 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1521 (memop addr:$src)))]>,
1522 XS, Requires<[HasSSE2]>;
1524 // SSE2 packed instructions with XD prefix
1525 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1526 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1527 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1528 XD, Requires<[HasSSE2]>;
1529 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1530 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1531 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1532 (memop addr:$src)))]>,
1533 XD, Requires<[HasSSE2]>;
1535 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1536 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1537 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1538 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1539 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1540 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1541 (memop addr:$src)))]>;
1543 // SSE2 instructions without OpSize prefix
1544 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1545 "cvtps2pd\t{$src, $dst|$dst, $src}",
1546 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1547 TB, Requires<[HasSSE2]>;
1548 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1549 "cvtps2pd\t{$src, $dst|$dst, $src}",
1550 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1551 (load addr:$src)))]>,
1552 TB, Requires<[HasSSE2]>;
1554 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1555 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1556 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1557 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1558 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1559 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1560 (memop addr:$src)))]>;
1562 // Match intrinsics which expect XMM operand(s).
1563 // Aliases for intrinsics
1564 let Constraints = "$src1 = $dst" in {
1565 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1566 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1567 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1568 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1570 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1571 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1572 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1573 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1574 (loadi32 addr:$src2)))]>;
1575 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1576 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1577 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1578 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1580 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1581 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1582 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1583 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1584 (load addr:$src2)))]>;
1585 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1586 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1587 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1588 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1589 VR128:$src2))]>, XS,
1590 Requires<[HasSSE2]>;
1591 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1592 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1593 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1594 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1595 (load addr:$src2)))]>, XS,
1596 Requires<[HasSSE2]>;
1601 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1603 /// In addition, we also have a special variant of the scalar form here to
1604 /// represent the associated intrinsic operation. This form is unlike the
1605 /// plain scalar form, in that it takes an entire vector (instead of a
1606 /// scalar) and leaves the top elements undefined.
1608 /// And, we have a special variant form for a full-vector intrinsic form.
1610 /// These four forms can each have a reg or a mem operand, so there are a
1611 /// total of eight "instructions".
1613 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1617 bit Commutable = 0> {
1618 // Scalar operation, reg.
1619 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1620 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1621 [(set FR64:$dst, (OpNode FR64:$src))]> {
1622 let isCommutable = Commutable;
1625 // Scalar operation, mem.
1626 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1627 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1628 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1630 // Vector operation, reg.
1631 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1632 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1633 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1634 let isCommutable = Commutable;
1637 // Vector operation, mem.
1638 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1639 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1640 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1642 // Intrinsic operation, reg.
1643 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1644 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1645 [(set VR128:$dst, (F64Int VR128:$src))]> {
1646 let isCommutable = Commutable;
1649 // Intrinsic operation, mem.
1650 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1651 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1652 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1654 // Vector intrinsic operation, reg
1655 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1656 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1657 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1658 let isCommutable = Commutable;
1661 // Vector intrinsic operation, mem
1662 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1663 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1664 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1668 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1669 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1671 // There is no f64 version of the reciprocal approximation instructions.
1674 let Constraints = "$src1 = $dst" in {
1675 let isCommutable = 1 in {
1676 def ANDPDrr : PDI<0x54, MRMSrcReg,
1677 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1678 "andpd\t{$src2, $dst|$dst, $src2}",
1680 (and (bc_v2i64 (v2f64 VR128:$src1)),
1681 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1682 def ORPDrr : PDI<0x56, MRMSrcReg,
1683 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1684 "orpd\t{$src2, $dst|$dst, $src2}",
1686 (or (bc_v2i64 (v2f64 VR128:$src1)),
1687 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1688 def XORPDrr : PDI<0x57, MRMSrcReg,
1689 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1690 "xorpd\t{$src2, $dst|$dst, $src2}",
1692 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1693 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1696 def ANDPDrm : PDI<0x54, MRMSrcMem,
1697 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1698 "andpd\t{$src2, $dst|$dst, $src2}",
1700 (and (bc_v2i64 (v2f64 VR128:$src1)),
1701 (memopv2i64 addr:$src2)))]>;
1702 def ORPDrm : PDI<0x56, MRMSrcMem,
1703 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1704 "orpd\t{$src2, $dst|$dst, $src2}",
1706 (or (bc_v2i64 (v2f64 VR128:$src1)),
1707 (memopv2i64 addr:$src2)))]>;
1708 def XORPDrm : PDI<0x57, MRMSrcMem,
1709 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1710 "xorpd\t{$src2, $dst|$dst, $src2}",
1712 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1713 (memopv2i64 addr:$src2)))]>;
1714 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1715 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1716 "andnpd\t{$src2, $dst|$dst, $src2}",
1718 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1719 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1720 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1721 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1722 "andnpd\t{$src2, $dst|$dst, $src2}",
1724 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1725 (memopv2i64 addr:$src2)))]>;
1728 let Constraints = "$src1 = $dst" in {
1729 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1730 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1731 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1732 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1733 VR128:$src, imm:$cc))]>;
1734 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1735 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1736 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1737 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1738 (memop addr:$src), imm:$cc))]>;
1740 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1741 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1742 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1743 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1745 // Shuffle and unpack instructions
1746 let Constraints = "$src1 = $dst" in {
1747 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1748 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1749 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1750 [(set VR128:$dst, (v2f64 (vector_shuffle
1751 VR128:$src1, VR128:$src2,
1752 SHUFP_shuffle_mask:$src3)))]>;
1753 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1754 (outs VR128:$dst), (ins VR128:$src1,
1755 f128mem:$src2, i8imm:$src3),
1756 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1758 (v2f64 (vector_shuffle
1759 VR128:$src1, (memopv2f64 addr:$src2),
1760 SHUFP_shuffle_mask:$src3)))]>;
1762 let AddedComplexity = 10 in {
1763 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1764 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1765 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1767 (v2f64 (vector_shuffle
1768 VR128:$src1, VR128:$src2,
1769 UNPCKH_shuffle_mask)))]>;
1770 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1771 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1772 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1774 (v2f64 (vector_shuffle
1775 VR128:$src1, (memopv2f64 addr:$src2),
1776 UNPCKH_shuffle_mask)))]>;
1778 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1779 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1780 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1782 (v2f64 (vector_shuffle
1783 VR128:$src1, VR128:$src2,
1784 UNPCKL_shuffle_mask)))]>;
1785 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1786 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1787 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1789 (v2f64 (vector_shuffle
1790 VR128:$src1, (memopv2f64 addr:$src2),
1791 UNPCKL_shuffle_mask)))]>;
1792 } // AddedComplexity
1793 } // Constraints = "$src1 = $dst"
1796 //===----------------------------------------------------------------------===//
1797 // SSE integer instructions
1799 // Move Instructions
1800 let neverHasSideEffects = 1 in
1801 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1802 "movdqa\t{$src, $dst|$dst, $src}", []>;
1803 let isSimpleLoad = 1, mayLoad = 1 in
1804 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1805 "movdqa\t{$src, $dst|$dst, $src}",
1806 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1808 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1809 "movdqa\t{$src, $dst|$dst, $src}",
1810 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1811 let isSimpleLoad = 1, mayLoad = 1 in
1812 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1813 "movdqu\t{$src, $dst|$dst, $src}",
1814 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1815 XS, Requires<[HasSSE2]>;
1817 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1818 "movdqu\t{$src, $dst|$dst, $src}",
1819 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1820 XS, Requires<[HasSSE2]>;
1822 // Intrinsic forms of MOVDQU load and store
1823 let isSimpleLoad = 1 in
1824 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1825 "movdqu\t{$src, $dst|$dst, $src}",
1826 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1827 XS, Requires<[HasSSE2]>;
1828 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1829 "movdqu\t{$src, $dst|$dst, $src}",
1830 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1831 XS, Requires<[HasSSE2]>;
1833 let Constraints = "$src1 = $dst" in {
1835 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1836 bit Commutable = 0> {
1837 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1838 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1839 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1840 let isCommutable = Commutable;
1842 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1843 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1844 [(set VR128:$dst, (IntId VR128:$src1,
1845 (bitconvert (memopv2i64 addr:$src2))))]>;
1848 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1850 Intrinsic IntId, Intrinsic IntId2> {
1851 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1852 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1853 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1854 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1855 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1856 [(set VR128:$dst, (IntId VR128:$src1,
1857 (bitconvert (memopv2i64 addr:$src2))))]>;
1858 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1859 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1860 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1863 /// PDI_binop_rm - Simple SSE2 binary operator.
1864 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1865 ValueType OpVT, bit Commutable = 0> {
1866 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1867 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1868 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1869 let isCommutable = Commutable;
1871 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1872 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1873 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1874 (bitconvert (memopv2i64 addr:$src2)))))]>;
1877 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1879 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1880 /// to collapse (bitconvert VT to VT) into its operand.
1882 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1883 bit Commutable = 0> {
1884 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1885 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1886 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1887 let isCommutable = Commutable;
1889 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1890 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1891 [(set VR128:$dst, (OpNode VR128:$src1,(memopv2i64 addr:$src2)))]>;
1894 } // Constraints = "$src1 = $dst"
1896 // 128-bit Integer Arithmetic
1898 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1899 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1900 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1901 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1903 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1904 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1905 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1906 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1908 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1909 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1910 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1911 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1913 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1914 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1915 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1916 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1918 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1920 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1921 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1922 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1924 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1926 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1927 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1930 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1931 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1932 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1933 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1934 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1937 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
1938 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
1939 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
1940 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
1941 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
1942 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
1944 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
1945 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
1946 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
1947 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
1948 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
1949 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
1951 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
1952 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
1953 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
1954 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
1956 // 128-bit logical shifts.
1957 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1958 def PSLLDQri : PDIi8<0x73, MRM7r,
1959 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1960 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1961 def PSRLDQri : PDIi8<0x73, MRM3r,
1962 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1963 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1964 // PSRADQri doesn't exist in SSE[1-3].
1967 let Predicates = [HasSSE2] in {
1968 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1969 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1970 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1971 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1972 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
1973 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
1974 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
1975 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
1976 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1977 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1979 // Shift up / down and insert zero's.
1980 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
1981 (v2i64 (PSLLDQri VR128:$src, (PSxLDQ_imm imm:$amt)))>;
1982 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
1983 (v2i64 (PSRLDQri VR128:$src, (PSxLDQ_imm imm:$amt)))>;
1987 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1988 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1989 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1991 let Constraints = "$src1 = $dst" in {
1992 def PANDNrr : PDI<0xDF, MRMSrcReg,
1993 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1994 "pandn\t{$src2, $dst|$dst, $src2}",
1995 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1998 def PANDNrm : PDI<0xDF, MRMSrcMem,
1999 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2000 "pandn\t{$src2, $dst|$dst, $src2}",
2001 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2002 (memopv2i64 addr:$src2))))]>;
2005 // SSE2 Integer comparison
2006 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2007 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2008 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2009 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2010 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2011 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2013 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2014 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2015 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2016 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2017 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2018 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2019 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2020 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2021 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2022 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2023 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2024 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2026 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2027 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2028 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2029 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2030 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2031 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2032 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2033 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2034 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2035 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2036 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2037 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2040 // Pack instructions
2041 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2042 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2043 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2045 // Shuffle and unpack instructions
2046 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2047 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2048 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2049 [(set VR128:$dst, (v4i32 (vector_shuffle
2050 VR128:$src1, (undef),
2051 PSHUFD_shuffle_mask:$src2)))]>;
2052 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2053 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2054 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2055 [(set VR128:$dst, (v4i32 (vector_shuffle
2056 (bc_v4i32(memopv2i64 addr:$src1)),
2058 PSHUFD_shuffle_mask:$src2)))]>;
2060 // SSE2 with ImmT == Imm8 and XS prefix.
2061 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2062 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2063 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2064 [(set VR128:$dst, (v8i16 (vector_shuffle
2065 VR128:$src1, (undef),
2066 PSHUFHW_shuffle_mask:$src2)))]>,
2067 XS, Requires<[HasSSE2]>;
2068 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2069 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2070 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2071 [(set VR128:$dst, (v8i16 (vector_shuffle
2072 (bc_v8i16 (memopv2i64 addr:$src1)),
2074 PSHUFHW_shuffle_mask:$src2)))]>,
2075 XS, Requires<[HasSSE2]>;
2077 // SSE2 with ImmT == Imm8 and XD prefix.
2078 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2079 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2080 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2081 [(set VR128:$dst, (v8i16 (vector_shuffle
2082 VR128:$src1, (undef),
2083 PSHUFLW_shuffle_mask:$src2)))]>,
2084 XD, Requires<[HasSSE2]>;
2085 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2086 (outs VR128:$dst), (ins i128mem:$src1, i32i8imm:$src2),
2087 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2088 [(set VR128:$dst, (v8i16 (vector_shuffle
2089 (bc_v8i16 (memopv2i64 addr:$src1)),
2091 PSHUFLW_shuffle_mask:$src2)))]>,
2092 XD, Requires<[HasSSE2]>;
2095 let Constraints = "$src1 = $dst" in {
2096 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2097 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2098 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2100 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
2101 UNPCKL_shuffle_mask)))]>;
2102 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2103 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2104 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2106 (v16i8 (vector_shuffle VR128:$src1,
2107 (bc_v16i8 (memopv2i64 addr:$src2)),
2108 UNPCKL_shuffle_mask)))]>;
2109 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2110 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2111 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2113 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2114 UNPCKL_shuffle_mask)))]>;
2115 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2116 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2117 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2119 (v8i16 (vector_shuffle VR128:$src1,
2120 (bc_v8i16 (memopv2i64 addr:$src2)),
2121 UNPCKL_shuffle_mask)))]>;
2122 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2123 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2124 "punpckldq\t{$src2, $dst|$dst, $src2}",
2126 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2127 UNPCKL_shuffle_mask)))]>;
2128 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2129 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2130 "punpckldq\t{$src2, $dst|$dst, $src2}",
2132 (v4i32 (vector_shuffle VR128:$src1,
2133 (bc_v4i32 (memopv2i64 addr:$src2)),
2134 UNPCKL_shuffle_mask)))]>;
2135 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2136 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2137 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2139 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2140 UNPCKL_shuffle_mask)))]>;
2141 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2142 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2143 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2145 (v2i64 (vector_shuffle VR128:$src1,
2146 (memopv2i64 addr:$src2),
2147 UNPCKL_shuffle_mask)))]>;
2149 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2150 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2151 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2153 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
2154 UNPCKH_shuffle_mask)))]>;
2155 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2156 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2157 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2159 (v16i8 (vector_shuffle VR128:$src1,
2160 (bc_v16i8 (memopv2i64 addr:$src2)),
2161 UNPCKH_shuffle_mask)))]>;
2162 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2163 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2164 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2166 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2167 UNPCKH_shuffle_mask)))]>;
2168 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2169 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2170 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2172 (v8i16 (vector_shuffle VR128:$src1,
2173 (bc_v8i16 (memopv2i64 addr:$src2)),
2174 UNPCKH_shuffle_mask)))]>;
2175 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2176 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2177 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2179 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2180 UNPCKH_shuffle_mask)))]>;
2181 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2182 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2183 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2185 (v4i32 (vector_shuffle VR128:$src1,
2186 (bc_v4i32 (memopv2i64 addr:$src2)),
2187 UNPCKH_shuffle_mask)))]>;
2188 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2189 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2190 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2192 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2193 UNPCKH_shuffle_mask)))]>;
2194 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2195 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2196 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2198 (v2i64 (vector_shuffle VR128:$src1,
2199 (memopv2i64 addr:$src2),
2200 UNPCKH_shuffle_mask)))]>;
2204 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2205 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2206 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2207 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2209 let Constraints = "$src1 = $dst" in {
2210 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2211 (outs VR128:$dst), (ins VR128:$src1,
2212 GR32:$src2, i32i8imm:$src3),
2213 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2215 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2216 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2217 (outs VR128:$dst), (ins VR128:$src1,
2218 i16mem:$src2, i32i8imm:$src3),
2219 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2221 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2226 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2227 "pmovmskb\t{$src, $dst|$dst, $src}",
2228 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2230 // Conditional store
2232 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2233 "maskmovdqu\t{$mask, $src|$src, $mask}",
2234 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2236 // Non-temporal stores
2237 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2238 "movntpd\t{$src, $dst|$dst, $src}",
2239 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2240 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2241 "movntdq\t{$src, $dst|$dst, $src}",
2242 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2243 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2244 "movnti\t{$src, $dst|$dst, $src}",
2245 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2246 TB, Requires<[HasSSE2]>;
2249 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2250 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2251 TB, Requires<[HasSSE2]>;
2253 // Load, store, and memory fence
2254 def LFENCE : I<0xAE, MRM5m, (outs), (ins),
2255 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2256 def MFENCE : I<0xAE, MRM6m, (outs), (ins),
2257 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2259 //TODO: custom lower this so as to never even generate the noop
2260 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2262 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2263 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2264 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2267 // Alias instructions that map zero vector to pxor / xorp* for sse.
2268 let isReMaterializable = 1, isAsCheapAsAMove = 1 in
2269 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2270 "pcmpeqd\t$dst, $dst",
2271 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2273 // FR64 to 128-bit vector conversion.
2274 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2275 "movsd\t{$src, $dst|$dst, $src}",
2277 (v2f64 (scalar_to_vector FR64:$src)))]>;
2278 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2279 "movsd\t{$src, $dst|$dst, $src}",
2281 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2283 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2284 "movd\t{$src, $dst|$dst, $src}",
2286 (v4i32 (scalar_to_vector GR32:$src)))]>;
2287 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2288 "movd\t{$src, $dst|$dst, $src}",
2290 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2292 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2293 "movd\t{$src, $dst|$dst, $src}",
2294 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2296 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2297 "movd\t{$src, $dst|$dst, $src}",
2298 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2300 // SSE2 instructions with XS prefix
2301 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2302 "movq\t{$src, $dst|$dst, $src}",
2304 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2305 Requires<[HasSSE2]>;
2306 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2307 "movq\t{$src, $dst|$dst, $src}",
2308 [(store (i64 (vector_extract (v2i64 VR128:$src),
2309 (iPTR 0))), addr:$dst)]>;
2311 // FIXME: may not be able to eliminate this movss with coalescing the src and
2312 // dest register classes are different. We really want to write this pattern
2314 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2315 // (f32 FR32:$src)>;
2316 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2317 "movsd\t{$src, $dst|$dst, $src}",
2318 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2320 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2321 "movsd\t{$src, $dst|$dst, $src}",
2322 [(store (f64 (vector_extract (v2f64 VR128:$src),
2323 (iPTR 0))), addr:$dst)]>;
2324 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2325 "movd\t{$src, $dst|$dst, $src}",
2326 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2328 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2329 "movd\t{$src, $dst|$dst, $src}",
2330 [(store (i32 (vector_extract (v4i32 VR128:$src),
2331 (iPTR 0))), addr:$dst)]>;
2333 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2334 "movd\t{$src, $dst|$dst, $src}",
2335 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2336 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2337 "movd\t{$src, $dst|$dst, $src}",
2338 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2341 // Move to lower bits of a VR128, leaving upper bits alone.
2342 // Three operand (but two address) aliases.
2343 let Constraints = "$src1 = $dst" in {
2344 let neverHasSideEffects = 1 in
2345 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2346 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2347 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2349 let AddedComplexity = 15 in
2350 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2351 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2352 "movsd\t{$src2, $dst|$dst, $src2}",
2354 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
2355 MOVL_shuffle_mask)))]>;
2358 // Store / copy lower 64-bits of a XMM register.
2359 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2360 "movq\t{$src, $dst|$dst, $src}",
2361 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2363 // Move to lower bits of a VR128 and zeroing upper bits.
2364 // Loading from memory automatically zeroing upper bits.
2365 let AddedComplexity = 20 in {
2366 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2367 "movsd\t{$src, $dst|$dst, $src}",
2369 (v2f64 (X86vzmovl (v2f64 (scalar_to_vector
2370 (loadf64 addr:$src))))))]>;
2372 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
2373 (MOVZSD2PDrm addr:$src)>;
2374 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
2375 (MOVZSD2PDrm addr:$src)>;
2376 def : Pat<(v2f64 (X86vzload addr:$src)), (MOVZSD2PDrm addr:$src)>;
2379 // movd / movq to XMM register zero-extends
2380 let AddedComplexity = 15 in {
2381 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2382 "movd\t{$src, $dst|$dst, $src}",
2383 [(set VR128:$dst, (v4i32 (X86vzmovl
2384 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2385 // This is X86-64 only.
2386 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2387 "mov{d|q}\t{$src, $dst|$dst, $src}",
2388 [(set VR128:$dst, (v2i64 (X86vzmovl
2389 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2392 let AddedComplexity = 20 in {
2393 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2394 "movd\t{$src, $dst|$dst, $src}",
2396 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2397 (loadi32 addr:$src))))))]>;
2399 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2400 (MOVZDI2PDIrm addr:$src)>;
2401 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2402 (MOVZDI2PDIrm addr:$src)>;
2403 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2404 (MOVZDI2PDIrm addr:$src)>;
2406 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2407 "movq\t{$src, $dst|$dst, $src}",
2409 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2410 (loadi64 addr:$src))))))]>, XS,
2411 Requires<[HasSSE2]>;
2413 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2414 (MOVZQI2PQIrm addr:$src)>;
2415 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2416 (MOVZQI2PQIrm addr:$src)>;
2417 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2420 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2421 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2422 let AddedComplexity = 15 in
2423 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2424 "movq\t{$src, $dst|$dst, $src}",
2425 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2426 XS, Requires<[HasSSE2]>;
2428 let AddedComplexity = 20 in {
2429 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2430 "movq\t{$src, $dst|$dst, $src}",
2431 [(set VR128:$dst, (v2i64 (X86vzmovl
2432 (loadv2i64 addr:$src))))]>,
2433 XS, Requires<[HasSSE2]>;
2435 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2436 (MOVZPQILo2PQIrm addr:$src)>;
2439 //===----------------------------------------------------------------------===//
2440 // SSE3 Instructions
2441 //===----------------------------------------------------------------------===//
2443 // Move Instructions
2444 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2445 "movshdup\t{$src, $dst|$dst, $src}",
2446 [(set VR128:$dst, (v4f32 (vector_shuffle
2447 VR128:$src, (undef),
2448 MOVSHDUP_shuffle_mask)))]>;
2449 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2450 "movshdup\t{$src, $dst|$dst, $src}",
2451 [(set VR128:$dst, (v4f32 (vector_shuffle
2452 (memopv4f32 addr:$src), (undef),
2453 MOVSHDUP_shuffle_mask)))]>;
2455 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2456 "movsldup\t{$src, $dst|$dst, $src}",
2457 [(set VR128:$dst, (v4f32 (vector_shuffle
2458 VR128:$src, (undef),
2459 MOVSLDUP_shuffle_mask)))]>;
2460 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2461 "movsldup\t{$src, $dst|$dst, $src}",
2462 [(set VR128:$dst, (v4f32 (vector_shuffle
2463 (memopv4f32 addr:$src), (undef),
2464 MOVSLDUP_shuffle_mask)))]>;
2466 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2467 "movddup\t{$src, $dst|$dst, $src}",
2469 (v2f64 (vector_shuffle VR128:$src, (undef),
2470 MOVDDUP_shuffle_mask)))]>;
2471 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2472 "movddup\t{$src, $dst|$dst, $src}",
2474 (v2f64 (vector_shuffle
2475 (scalar_to_vector (loadf64 addr:$src)),
2476 (undef), MOVDDUP_shuffle_mask)))]>;
2478 def : Pat<(vector_shuffle
2479 (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2480 (undef), MOVDDUP_shuffle_mask),
2481 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2482 def : Pat<(vector_shuffle
2483 (memopv2f64 addr:$src), (undef), MOVDDUP_shuffle_mask),
2484 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2488 let Constraints = "$src1 = $dst" in {
2489 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2490 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2491 "addsubps\t{$src2, $dst|$dst, $src2}",
2492 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2494 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2495 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2496 "addsubps\t{$src2, $dst|$dst, $src2}",
2497 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2498 (memop addr:$src2)))]>;
2499 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2500 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2501 "addsubpd\t{$src2, $dst|$dst, $src2}",
2502 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2504 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2505 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2506 "addsubpd\t{$src2, $dst|$dst, $src2}",
2507 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2508 (memop addr:$src2)))]>;
2511 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2512 "lddqu\t{$src, $dst|$dst, $src}",
2513 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2516 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2517 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2518 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2519 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2520 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2521 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2522 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2523 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2524 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2525 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2526 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2527 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2528 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2529 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2530 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2531 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2533 let Constraints = "$src1 = $dst" in {
2534 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2535 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2536 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2537 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2538 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2539 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2540 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2541 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2544 // Thread synchronization
2545 def MONITOR : I<0xC8, RawFrm, (outs), (ins), "monitor",
2546 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2547 def MWAIT : I<0xC9, RawFrm, (outs), (ins), "mwait",
2548 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2550 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2551 let AddedComplexity = 15 in
2552 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2553 MOVSHDUP_shuffle_mask)),
2554 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2555 let AddedComplexity = 20 in
2556 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2557 MOVSHDUP_shuffle_mask)),
2558 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2560 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2561 let AddedComplexity = 15 in
2562 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2563 MOVSLDUP_shuffle_mask)),
2564 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2565 let AddedComplexity = 20 in
2566 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2567 MOVSLDUP_shuffle_mask)),
2568 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2570 //===----------------------------------------------------------------------===//
2571 // SSSE3 Instructions
2572 //===----------------------------------------------------------------------===//
2574 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2575 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2576 Intrinsic IntId64, Intrinsic IntId128> {
2577 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2578 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2579 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2581 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2582 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2584 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2586 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2588 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2589 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2592 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2594 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2597 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2600 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2601 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2602 Intrinsic IntId64, Intrinsic IntId128> {
2603 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2605 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2606 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2608 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2610 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2613 (bitconvert (memopv4i16 addr:$src))))]>;
2615 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2617 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2618 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2621 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2623 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2626 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2629 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2630 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2631 Intrinsic IntId64, Intrinsic IntId128> {
2632 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2634 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2635 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2637 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2639 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2642 (bitconvert (memopv2i32 addr:$src))))]>;
2644 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2646 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2647 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2650 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2652 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2655 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2658 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2659 int_x86_ssse3_pabs_b,
2660 int_x86_ssse3_pabs_b_128>;
2661 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2662 int_x86_ssse3_pabs_w,
2663 int_x86_ssse3_pabs_w_128>;
2664 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2665 int_x86_ssse3_pabs_d,
2666 int_x86_ssse3_pabs_d_128>;
2668 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2669 let Constraints = "$src1 = $dst" in {
2670 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2671 Intrinsic IntId64, Intrinsic IntId128,
2672 bit Commutable = 0> {
2673 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2674 (ins VR64:$src1, VR64:$src2),
2675 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2676 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2677 let isCommutable = Commutable;
2679 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2680 (ins VR64:$src1, i64mem:$src2),
2681 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2683 (IntId64 VR64:$src1,
2684 (bitconvert (memopv8i8 addr:$src2))))]>;
2686 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2687 (ins VR128:$src1, VR128:$src2),
2688 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2689 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2691 let isCommutable = Commutable;
2693 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2694 (ins VR128:$src1, i128mem:$src2),
2695 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2697 (IntId128 VR128:$src1,
2698 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2702 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2703 let Constraints = "$src1 = $dst" in {
2704 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2705 Intrinsic IntId64, Intrinsic IntId128,
2706 bit Commutable = 0> {
2707 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2708 (ins VR64:$src1, VR64:$src2),
2709 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2710 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2711 let isCommutable = Commutable;
2713 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2714 (ins VR64:$src1, i64mem:$src2),
2715 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2717 (IntId64 VR64:$src1,
2718 (bitconvert (memopv4i16 addr:$src2))))]>;
2720 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2721 (ins VR128:$src1, VR128:$src2),
2722 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2723 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2725 let isCommutable = Commutable;
2727 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2728 (ins VR128:$src1, i128mem:$src2),
2729 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2731 (IntId128 VR128:$src1,
2732 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2736 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2737 let Constraints = "$src1 = $dst" in {
2738 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2739 Intrinsic IntId64, Intrinsic IntId128,
2740 bit Commutable = 0> {
2741 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2742 (ins VR64:$src1, VR64:$src2),
2743 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2744 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2745 let isCommutable = Commutable;
2747 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2748 (ins VR64:$src1, i64mem:$src2),
2749 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2751 (IntId64 VR64:$src1,
2752 (bitconvert (memopv2i32 addr:$src2))))]>;
2754 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2755 (ins VR128:$src1, VR128:$src2),
2756 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2757 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2759 let isCommutable = Commutable;
2761 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2762 (ins VR128:$src1, i128mem:$src2),
2763 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2765 (IntId128 VR128:$src1,
2766 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2770 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2771 int_x86_ssse3_phadd_w,
2772 int_x86_ssse3_phadd_w_128>;
2773 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2774 int_x86_ssse3_phadd_d,
2775 int_x86_ssse3_phadd_d_128>;
2776 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2777 int_x86_ssse3_phadd_sw,
2778 int_x86_ssse3_phadd_sw_128>;
2779 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2780 int_x86_ssse3_phsub_w,
2781 int_x86_ssse3_phsub_w_128>;
2782 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2783 int_x86_ssse3_phsub_d,
2784 int_x86_ssse3_phsub_d_128>;
2785 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2786 int_x86_ssse3_phsub_sw,
2787 int_x86_ssse3_phsub_sw_128>;
2788 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2789 int_x86_ssse3_pmadd_ub_sw,
2790 int_x86_ssse3_pmadd_ub_sw_128>;
2791 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2792 int_x86_ssse3_pmul_hr_sw,
2793 int_x86_ssse3_pmul_hr_sw_128, 1>;
2794 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2795 int_x86_ssse3_pshuf_b,
2796 int_x86_ssse3_pshuf_b_128>;
2797 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2798 int_x86_ssse3_psign_b,
2799 int_x86_ssse3_psign_b_128>;
2800 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2801 int_x86_ssse3_psign_w,
2802 int_x86_ssse3_psign_w_128>;
2803 defm PSIGND : SS3I_binop_rm_int_32<0x09, "psignd",
2804 int_x86_ssse3_psign_d,
2805 int_x86_ssse3_psign_d_128>;
2807 let Constraints = "$src1 = $dst" in {
2808 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2809 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2810 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2812 (int_x86_ssse3_palign_r
2813 VR64:$src1, VR64:$src2,
2815 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2816 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2817 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2819 (int_x86_ssse3_palign_r
2821 (bitconvert (memopv2i32 addr:$src2)),
2824 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2825 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2826 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2828 (int_x86_ssse3_palign_r_128
2829 VR128:$src1, VR128:$src2,
2830 imm:$src3))]>, OpSize;
2831 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2832 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2833 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2835 (int_x86_ssse3_palign_r_128
2837 (bitconvert (memopv4i32 addr:$src2)),
2838 imm:$src3))]>, OpSize;
2841 //===----------------------------------------------------------------------===//
2842 // Non-Instruction Patterns
2843 //===----------------------------------------------------------------------===//
2845 // extload f32 -> f64. This matches load+fextend because we have a hack in
2846 // the isel (PreprocessForFPConvert) that can introduce loads after dag combine.
2847 // Since these loads aren't folded into the fextend, we have to match it
2849 let Predicates = [HasSSE2] in
2850 def : Pat<(fextend (loadf32 addr:$src)),
2851 (CVTSS2SDrm addr:$src)>;
2854 let Predicates = [HasSSE2] in {
2855 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2856 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2857 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2858 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2859 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2860 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2861 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2862 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2863 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2864 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2865 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2866 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2867 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2868 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2869 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2870 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2871 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2872 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2873 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2874 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2875 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2876 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2877 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2878 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2879 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2880 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2881 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2882 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2883 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2884 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2887 // Move scalar to XMM zero-extended
2888 // movd to XMM register zero-extends
2889 let AddedComplexity = 15 in {
2890 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2891 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2892 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2893 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2894 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
2895 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2896 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE2]>;
2897 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2898 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE2]>;
2901 // Splat v2f64 / v2i64
2902 let AddedComplexity = 10 in {
2903 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2904 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2905 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2906 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2907 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2908 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2909 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2910 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2913 // Special unary SHUFPSrri case.
2914 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2915 SHUFP_unary_shuffle_mask:$sm)),
2916 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2917 Requires<[HasSSE1]>;
2918 // Special unary SHUFPDrri case.
2919 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (undef),
2920 SHUFP_unary_shuffle_mask:$sm)),
2921 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2922 Requires<[HasSSE2]>;
2923 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2924 def : Pat<(vector_shuffle (bc_v4i32 (memopv4f32 addr:$src1)), (undef),
2925 SHUFP_unary_shuffle_mask:$sm),
2926 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2927 Requires<[HasSSE2]>;
2929 // Special binary v4i32 shuffle cases with SHUFPS.
2930 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (v4i32 VR128:$src2),
2931 PSHUFD_binary_shuffle_mask:$sm)),
2932 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2933 Requires<[HasSSE2]>;
2934 def : Pat<(v4i32 (vector_shuffle VR128:$src1,
2935 (bc_v4i32 (memopv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm)),
2936 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2937 Requires<[HasSSE2]>;
2938 // Special binary v2i64 shuffle cases using SHUFPDrri.
2939 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2940 SHUFP_shuffle_mask:$sm)),
2941 (SHUFPDrri VR128:$src1, VR128:$src2, SHUFP_shuffle_mask:$sm)>,
2942 Requires<[HasSSE2]>;
2943 // Special unary SHUFPDrri case.
2944 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (undef),
2945 SHUFP_unary_shuffle_mask:$sm)),
2946 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2947 Requires<[HasSSE2]>;
2949 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2950 let AddedComplexity = 15 in {
2951 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2952 UNPCKL_v_undef_shuffle_mask:$sm)),
2953 (PSHUFDri VR128:$src, PSHUFD_shuffle_mask:$sm)>,
2954 Requires<[OptForSpeed, HasSSE2]>;
2955 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2956 UNPCKL_v_undef_shuffle_mask:$sm)),
2957 (PSHUFDri VR128:$src, PSHUFD_shuffle_mask:$sm)>,
2958 Requires<[OptForSpeed, HasSSE2]>;
2960 let AddedComplexity = 10 in {
2961 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2962 UNPCKL_v_undef_shuffle_mask)),
2963 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2964 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2965 UNPCKL_v_undef_shuffle_mask)),
2966 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2967 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2968 UNPCKL_v_undef_shuffle_mask)),
2969 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2970 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2971 UNPCKL_v_undef_shuffle_mask)),
2972 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2975 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2976 let AddedComplexity = 15 in {
2977 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2978 UNPCKH_v_undef_shuffle_mask:$sm)),
2979 (PSHUFDri VR128:$src, PSHUFD_shuffle_mask:$sm)>,
2980 Requires<[OptForSpeed, HasSSE2]>;
2981 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2982 UNPCKH_v_undef_shuffle_mask:$sm)),
2983 (PSHUFDri VR128:$src, PSHUFD_shuffle_mask:$sm)>,
2984 Requires<[OptForSpeed, HasSSE2]>;
2986 let AddedComplexity = 10 in {
2987 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2988 UNPCKH_v_undef_shuffle_mask)),
2989 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2990 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2991 UNPCKH_v_undef_shuffle_mask)),
2992 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2993 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2994 UNPCKH_v_undef_shuffle_mask)),
2995 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2996 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2997 UNPCKH_v_undef_shuffle_mask)),
2998 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3001 let AddedComplexity = 20 in {
3002 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3003 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
3004 MOVHP_shuffle_mask)),
3005 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3007 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3008 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
3009 MOVHLPS_shuffle_mask)),
3010 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3012 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3013 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
3014 MOVHLPS_v_undef_shuffle_mask)),
3015 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3016 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
3017 MOVHLPS_v_undef_shuffle_mask)),
3018 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3021 let AddedComplexity = 20 in {
3022 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3023 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
3024 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memop addr:$src2),
3025 MOVLP_shuffle_mask)),
3026 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3027 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memop addr:$src2),
3028 MOVLP_shuffle_mask)),
3029 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3030 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memop addr:$src2),
3031 MOVHP_shuffle_mask)),
3032 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3033 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memop addr:$src2),
3034 MOVHP_shuffle_mask)),
3035 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3037 def : Pat<(v4i32 (vector_shuffle VR128:$src1,
3038 (bc_v4i32 (memopv2i64 addr:$src2)),
3039 MOVLP_shuffle_mask)),
3040 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3041 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memop addr:$src2),
3042 MOVLP_shuffle_mask)),
3043 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3044 def : Pat<(v4i32 (vector_shuffle VR128:$src1,
3045 (bc_v4i32 (memopv2i64 addr:$src2)),
3046 MOVHP_shuffle_mask)),
3047 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3048 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memop addr:$src2),
3049 MOVHP_shuffle_mask)),
3050 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3053 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3054 // (store (vector_shuffle (load addr), v2, <0, 1, 4, 5>), addr) using MOVHPS
3055 def : Pat<(store (v4f32 (vector_shuffle (memop addr:$src1), VR128:$src2,
3056 MOVLP_shuffle_mask)), addr:$src1),
3057 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3058 def : Pat<(store (v2f64 (vector_shuffle (memop addr:$src1), VR128:$src2,
3059 MOVLP_shuffle_mask)), addr:$src1),
3060 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3061 def : Pat<(store (v4f32 (vector_shuffle (memop addr:$src1), VR128:$src2,
3062 MOVHP_shuffle_mask)), addr:$src1),
3063 (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3064 def : Pat<(store (v2f64 (vector_shuffle (memop addr:$src1), VR128:$src2,
3065 MOVHP_shuffle_mask)), addr:$src1),
3066 (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3068 def : Pat<(store (v4i32 (vector_shuffle
3069 (bc_v4i32 (memopv2i64 addr:$src1)), VR128:$src2,
3070 MOVLP_shuffle_mask)), addr:$src1),
3071 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3072 def : Pat<(store (v2i64 (vector_shuffle (memop addr:$src1), VR128:$src2,
3073 MOVLP_shuffle_mask)), addr:$src1),
3074 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3075 def : Pat<(store (v4i32 (vector_shuffle
3076 (bc_v4i32 (memopv2i64 addr:$src1)), VR128:$src2,
3077 MOVHP_shuffle_mask)), addr:$src1),
3078 (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3079 def : Pat<(store (v2i64 (vector_shuffle (memop addr:$src1), VR128:$src2,
3080 MOVHP_shuffle_mask)), addr:$src1),
3081 (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3084 let AddedComplexity = 15 in {
3085 // Setting the lowest element in the vector.
3086 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
3087 MOVL_shuffle_mask)),
3088 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3089 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
3090 MOVL_shuffle_mask)),
3091 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3093 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
3094 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
3095 MOVLP_shuffle_mask)),
3096 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3097 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
3098 MOVLP_shuffle_mask)),
3099 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3102 // Set lowest element and zero upper elements.
3103 let AddedComplexity = 15 in
3104 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc, VR128:$src,
3105 MOVL_shuffle_mask)),
3106 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3107 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3108 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3110 // Some special case pandn patterns.
3111 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3113 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3114 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3116 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3117 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3119 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3121 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3122 (memop addr:$src2))),
3123 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3124 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3125 (memop addr:$src2))),
3126 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3127 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3128 (memop addr:$src2))),
3129 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3131 // vector -> vector casts
3132 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3133 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3134 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3135 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3136 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3137 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3138 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3139 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3141 // Use movaps / movups for SSE integer load / store (one byte shorter).
3142 def : Pat<(alignedloadv4i32 addr:$src),
3143 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
3144 def : Pat<(loadv4i32 addr:$src),
3145 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3146 def : Pat<(alignedloadv2i64 addr:$src),
3147 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3148 def : Pat<(loadv2i64 addr:$src),
3149 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3151 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3152 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3153 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3154 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3155 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3156 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3157 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3158 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3159 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3160 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3161 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3162 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3163 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3164 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3165 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3166 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3168 //===----------------------------------------------------------------------===//
3169 // SSE4.1 Instructions
3170 //===----------------------------------------------------------------------===//
3172 multiclass sse41_fp_unop_rm<bits<8> opcss, bits<8> opcps,
3173 bits<8> opcsd, bits<8> opcpd,
3178 Intrinsic V2F64Int> {
3179 // Intrinsic operation, reg.
3180 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3181 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3182 !strconcat(OpcodeStr,
3183 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3184 [(set VR128:$dst, (F32Int VR128:$src1, imm:$src2))]>,
3187 // Intrinsic operation, mem.
3188 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3189 (outs VR128:$dst), (ins ssmem:$src1, i32i8imm:$src2),
3190 !strconcat(OpcodeStr,
3191 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3192 [(set VR128:$dst, (F32Int sse_load_f32:$src1, imm:$src2))]>,
3195 // Vector intrinsic operation, reg
3196 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3197 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3198 !strconcat(OpcodeStr,
3199 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3200 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3203 // Vector intrinsic operation, mem
3204 def PSm_Int : SS4AIi8<opcps, MRMSrcMem,
3205 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3206 !strconcat(OpcodeStr,
3207 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3209 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3212 // Intrinsic operation, reg.
3213 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3214 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3215 !strconcat(OpcodeStr,
3216 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3217 [(set VR128:$dst, (F64Int VR128:$src1, imm:$src2))]>,
3220 // Intrinsic operation, mem.
3221 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3222 (outs VR128:$dst), (ins sdmem:$src1, i32i8imm:$src2),
3223 !strconcat(OpcodeStr,
3224 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3225 [(set VR128:$dst, (F64Int sse_load_f64:$src1, imm:$src2))]>,
3228 // Vector intrinsic operation, reg
3229 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3230 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3231 !strconcat(OpcodeStr,
3232 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3233 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3236 // Vector intrinsic operation, mem
3237 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3238 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3239 !strconcat(OpcodeStr,
3240 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3242 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3246 // FP round - roundss, roundps, roundsd, roundpd
3247 defm ROUND : sse41_fp_unop_rm<0x0A, 0x08, 0x0B, 0x09, "round",
3248 int_x86_sse41_round_ss, int_x86_sse41_round_ps,
3249 int_x86_sse41_round_sd, int_x86_sse41_round_pd>;
3251 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3252 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3253 Intrinsic IntId128> {
3254 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3256 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3257 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3258 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3260 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3263 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3266 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3267 int_x86_sse41_phminposuw>;
3269 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3270 let Constraints = "$src1 = $dst" in {
3271 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3272 Intrinsic IntId128, bit Commutable = 0> {
3273 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3274 (ins VR128:$src1, VR128:$src2),
3275 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3276 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3278 let isCommutable = Commutable;
3280 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3281 (ins VR128:$src1, i128mem:$src2),
3282 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3284 (IntId128 VR128:$src1,
3285 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3289 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3290 int_x86_sse41_pcmpeqq, 1>;
3291 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3292 int_x86_sse41_packusdw, 0>;
3293 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3294 int_x86_sse41_pminsb, 1>;
3295 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3296 int_x86_sse41_pminsd, 1>;
3297 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3298 int_x86_sse41_pminud, 1>;
3299 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3300 int_x86_sse41_pminuw, 1>;
3301 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3302 int_x86_sse41_pmaxsb, 1>;
3303 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3304 int_x86_sse41_pmaxsd, 1>;
3305 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3306 int_x86_sse41_pmaxud, 1>;
3307 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3308 int_x86_sse41_pmaxuw, 1>;
3310 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3311 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3312 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3313 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3316 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3317 let Constraints = "$src1 = $dst" in {
3318 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3319 SDNode OpNode, Intrinsic IntId128,
3320 bit Commutable = 0> {
3321 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3322 (ins VR128:$src1, VR128:$src2),
3323 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3324 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3325 VR128:$src2))]>, OpSize {
3326 let isCommutable = Commutable;
3328 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3329 (ins VR128:$src1, VR128:$src2),
3330 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3331 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3333 let isCommutable = Commutable;
3335 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3336 (ins VR128:$src1, i128mem:$src2),
3337 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3339 (OpNode VR128:$src1, (memop addr:$src2)))]>, OpSize;
3340 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3341 (ins VR128:$src1, i128mem:$src2),
3342 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3344 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3348 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", v4i32, mul,
3349 int_x86_sse41_pmulld, 1>;
3350 defm PMULDQ : SS41I_binop_patint<0x28, "pmuldq", v2i64, mul,
3351 int_x86_sse41_pmuldq, 1>;
3354 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3355 let Constraints = "$src1 = $dst" in {
3356 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3357 Intrinsic IntId128, bit Commutable = 0> {
3358 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3359 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3360 !strconcat(OpcodeStr,
3361 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3363 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3365 let isCommutable = Commutable;
3367 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3368 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3369 !strconcat(OpcodeStr,
3370 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3372 (IntId128 VR128:$src1,
3373 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3378 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3379 int_x86_sse41_blendps, 0>;
3380 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3381 int_x86_sse41_blendpd, 0>;
3382 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3383 int_x86_sse41_pblendw, 0>;
3384 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3385 int_x86_sse41_dpps, 1>;
3386 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3387 int_x86_sse41_dppd, 1>;
3388 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3389 int_x86_sse41_mpsadbw, 1>;
3392 /// SS41I_ternary_int - SSE 4.1 ternary operator
3393 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3394 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3395 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3396 (ins VR128:$src1, VR128:$src2),
3397 !strconcat(OpcodeStr,
3398 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3399 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3402 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3403 (ins VR128:$src1, i128mem:$src2),
3404 !strconcat(OpcodeStr,
3405 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3408 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3412 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3413 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3414 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3417 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3418 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3419 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3420 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3422 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3423 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3425 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3429 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3430 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3431 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3432 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3433 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3434 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3436 // Common patterns involving scalar load.
3437 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3438 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3439 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3440 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3442 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3443 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3444 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3445 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3447 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3448 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3449 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3450 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3452 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3453 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3454 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3455 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3457 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3458 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3459 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3460 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3462 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3463 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3464 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3465 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3468 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3469 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3470 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3471 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3473 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3474 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3476 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3480 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3481 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3482 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3483 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3485 // Common patterns involving scalar load
3486 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3487 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3488 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3489 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3491 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3492 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3493 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3494 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3497 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3498 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3499 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3500 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3502 // Expecting a i16 load any extended to i32 value.
3503 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3504 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3505 [(set VR128:$dst, (IntId (bitconvert
3506 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3510 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3511 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovsxbq", int_x86_sse41_pmovzxbq>;
3513 // Common patterns involving scalar load
3514 def : Pat<(int_x86_sse41_pmovsxbq
3515 (bitconvert (v4i32 (X86vzmovl
3516 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3517 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3519 def : Pat<(int_x86_sse41_pmovzxbq
3520 (bitconvert (v4i32 (X86vzmovl
3521 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3522 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3525 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3526 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3527 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3528 (ins VR128:$src1, i32i8imm:$src2),
3529 !strconcat(OpcodeStr,
3530 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3531 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3533 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3534 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3535 !strconcat(OpcodeStr,
3536 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3539 // There's an AssertZext in the way of writing the store pattern
3540 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3543 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3546 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3547 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3548 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3549 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3550 !strconcat(OpcodeStr,
3551 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3554 // There's an AssertZext in the way of writing the store pattern
3555 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3558 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3561 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3562 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3563 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3564 (ins VR128:$src1, i32i8imm:$src2),
3565 !strconcat(OpcodeStr,
3566 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3568 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3569 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3570 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3571 !strconcat(OpcodeStr,
3572 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3573 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3574 addr:$dst)]>, OpSize;
3577 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3580 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3582 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3583 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3584 (ins VR128:$src1, i32i8imm:$src2),
3585 !strconcat(OpcodeStr,
3586 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3588 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3590 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3591 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3592 !strconcat(OpcodeStr,
3593 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3594 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3595 addr:$dst)]>, OpSize;
3598 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3600 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3601 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3604 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3605 Requires<[HasSSE41]>;
3607 let Constraints = "$src1 = $dst" in {
3608 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3609 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3610 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3611 !strconcat(OpcodeStr,
3612 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3614 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3615 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3616 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3617 !strconcat(OpcodeStr,
3618 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3620 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3621 imm:$src3))]>, OpSize;
3625 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3627 let Constraints = "$src1 = $dst" in {
3628 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3629 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3630 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3631 !strconcat(OpcodeStr,
3632 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3634 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3636 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3637 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3638 !strconcat(OpcodeStr,
3639 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3641 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3642 imm:$src3)))]>, OpSize;
3646 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3648 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:$src1, FR32:$src2, i32i8imm:$src3),
3652 !strconcat(OpcodeStr,
3653 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3655 (X86insrtps VR128:$src1, FR32:$src2, imm:$src3))]>, OpSize;
3656 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3657 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3658 !strconcat(OpcodeStr,
3659 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3661 (X86insrtps VR128:$src1, (loadf32 addr:$src2),
3662 imm:$src3))]>, OpSize;
3666 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3668 let Defs = [EFLAGS] in {
3669 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3670 "ptest \t{$src2, $src1|$src1, $src2}", []>, OpSize;
3671 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3672 "ptest \t{$src2, $src1|$src1, $src2}", []>, OpSize;
3675 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3676 "movntdqa\t{$src, $dst|$dst, $src}",
3677 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;
3679 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3680 let Constraints = "$src1 = $dst" in {
3681 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3682 Intrinsic IntId128, bit Commutable = 0> {
3683 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3684 (ins VR128:$src1, VR128:$src2),
3685 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3686 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3688 let isCommutable = Commutable;
3690 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3691 (ins VR128:$src1, i128mem:$src2),
3692 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3694 (IntId128 VR128:$src1,
3695 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3699 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3701 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3702 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3703 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3704 (PCMPGTQrm VR128:$src1, addr:$src2)>;