1 //====- X86InstrSSE.td - Describe the X86 Instruction Set --*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the X86 SSE instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
17 //===----------------------------------------------------------------------===//
18 // SSE specific DAG Nodes.
19 //===----------------------------------------------------------------------===//
21 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
22 SDTCisFP<0>, SDTCisInt<2> ]>;
23 def SDTX86VFCMP : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>,
24 SDTCisFP<1>, SDTCisVT<3, i8>]>;
26 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
27 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
28 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
29 [SDNPCommutative, SDNPAssociative]>;
30 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
31 [SDNPCommutative, SDNPAssociative]>;
32 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
33 [SDNPCommutative, SDNPAssociative]>;
34 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
35 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
36 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
37 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>;
38 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>;
39 def X86pshufb : SDNode<"X86ISD::PSHUFB",
40 SDTypeProfile<1, 2, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
42 def X86pextrb : SDNode<"X86ISD::PEXTRB",
43 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
44 def X86pextrw : SDNode<"X86ISD::PEXTRW",
45 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
46 def X86pinsrb : SDNode<"X86ISD::PINSRB",
47 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
48 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
49 def X86pinsrw : SDNode<"X86ISD::PINSRW",
50 SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
51 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
52 def X86insrtps : SDNode<"X86ISD::INSERTPS",
53 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
54 SDTCisVT<2, f32>, SDTCisPtrTy<3>]>>;
55 def X86vzmovl : SDNode<"X86ISD::VZEXT_MOVL",
56 SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>;
57 def X86vzload : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
58 [SDNPHasChain, SDNPMayLoad]>;
59 def X86vshl : SDNode<"X86ISD::VSHL", SDTIntShiftOp>;
60 def X86vshr : SDNode<"X86ISD::VSRL", SDTIntShiftOp>;
61 def X86cmpps : SDNode<"X86ISD::CMPPS", SDTX86VFCMP>;
62 def X86cmppd : SDNode<"X86ISD::CMPPD", SDTX86VFCMP>;
63 def X86pcmpeqb : SDNode<"X86ISD::PCMPEQB", SDTIntBinOp, [SDNPCommutative]>;
64 def X86pcmpeqw : SDNode<"X86ISD::PCMPEQW", SDTIntBinOp, [SDNPCommutative]>;
65 def X86pcmpeqd : SDNode<"X86ISD::PCMPEQD", SDTIntBinOp, [SDNPCommutative]>;
66 def X86pcmpeqq : SDNode<"X86ISD::PCMPEQQ", SDTIntBinOp, [SDNPCommutative]>;
67 def X86pcmpgtb : SDNode<"X86ISD::PCMPGTB", SDTIntBinOp>;
68 def X86pcmpgtw : SDNode<"X86ISD::PCMPGTW", SDTIntBinOp>;
69 def X86pcmpgtd : SDNode<"X86ISD::PCMPGTD", SDTIntBinOp>;
70 def X86pcmpgtq : SDNode<"X86ISD::PCMPGTQ", SDTIntBinOp>;
72 //===----------------------------------------------------------------------===//
73 // SSE Complex Patterns
74 //===----------------------------------------------------------------------===//
76 // These are 'extloads' from a scalar to the low element of a vector, zeroing
77 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
79 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
80 [SDNPHasChain, SDNPMayLoad]>;
81 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
82 [SDNPHasChain, SDNPMayLoad]>;
84 def ssmem : Operand<v4f32> {
85 let PrintMethod = "printf32mem";
86 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm, i8imm);
88 def sdmem : Operand<v2f64> {
89 let PrintMethod = "printf64mem";
90 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm, i8imm);
93 //===----------------------------------------------------------------------===//
94 // SSE pattern fragments
95 //===----------------------------------------------------------------------===//
97 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
98 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
99 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
100 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
102 // Like 'store', but always requires vector alignment.
103 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
104 (store node:$val, node:$ptr), [{
105 return cast<StoreSDNode>(N)->getAlignment() >= 16;
108 // Like 'load', but always requires vector alignment.
109 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
110 return cast<LoadSDNode>(N)->getAlignment() >= 16;
113 def alignedloadfsf32 : PatFrag<(ops node:$ptr), (f32 (alignedload node:$ptr))>;
114 def alignedloadfsf64 : PatFrag<(ops node:$ptr), (f64 (alignedload node:$ptr))>;
115 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
116 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
117 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
118 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
120 // Like 'load', but uses special alignment checks suitable for use in
121 // memory operands in most SSE instructions, which are required to
122 // be naturally aligned on some targets but not on others.
123 // FIXME: Actually implement support for targets that don't require the
124 // alignment. This probably wants a subtarget predicate.
125 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
126 return cast<LoadSDNode>(N)->getAlignment() >= 16;
129 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
130 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
131 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
132 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
133 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
134 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
135 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
137 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
139 // FIXME: 8 byte alignment for mmx reads is not required
140 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
141 return cast<LoadSDNode>(N)->getAlignment() >= 8;
144 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
145 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
146 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
147 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
149 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
150 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
151 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
152 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
153 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
154 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
156 def vzmovl_v2i64 : PatFrag<(ops node:$src),
157 (bitconvert (v2i64 (X86vzmovl
158 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
159 def vzmovl_v4i32 : PatFrag<(ops node:$src),
160 (bitconvert (v4i32 (X86vzmovl
161 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
163 def vzload_v2i64 : PatFrag<(ops node:$src),
164 (bitconvert (v2i64 (X86vzload node:$src)))>;
167 def fp32imm0 : PatLeaf<(f32 fpimm), [{
168 return N->isExactlyValue(+0.0);
171 def PSxLDQ_imm : SDNodeXForm<imm, [{
172 // Transformation function: imm >> 3
173 return getI32Imm(N->getZExtValue() >> 3);
176 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
178 def SHUFFLE_get_shuf_imm : SDNodeXForm<build_vector, [{
179 return getI8Imm(X86::getShuffleSHUFImmediate(N));
182 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
184 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<build_vector, [{
185 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
188 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
190 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<build_vector, [{
191 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
194 def SSE_splat_mask : PatLeaf<(build_vector), [{
195 return X86::isSplatMask(N);
196 }], SHUFFLE_get_shuf_imm>;
198 def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
199 return X86::isSplatLoMask(N);
202 def MOVDDUP_shuffle_mask : PatLeaf<(build_vector), [{
203 return X86::isMOVDDUPMask(N);
206 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
207 return X86::isMOVHLPSMask(N);
210 def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
211 return X86::isMOVHLPS_v_undef_Mask(N);
214 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
215 return X86::isMOVHPMask(N);
218 def MOVLP_shuffle_mask : PatLeaf<(build_vector), [{
219 return X86::isMOVLPMask(N);
222 def MOVL_shuffle_mask : PatLeaf<(build_vector), [{
223 return X86::isMOVLMask(N);
226 def MOVSHDUP_shuffle_mask : PatLeaf<(build_vector), [{
227 return X86::isMOVSHDUPMask(N);
230 def MOVSLDUP_shuffle_mask : PatLeaf<(build_vector), [{
231 return X86::isMOVSLDUPMask(N);
234 def UNPCKL_shuffle_mask : PatLeaf<(build_vector), [{
235 return X86::isUNPCKLMask(N);
238 def UNPCKH_shuffle_mask : PatLeaf<(build_vector), [{
239 return X86::isUNPCKHMask(N);
242 def UNPCKL_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
243 return X86::isUNPCKL_v_undef_Mask(N);
246 def UNPCKH_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
247 return X86::isUNPCKH_v_undef_Mask(N);
250 def PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
251 return X86::isPSHUFDMask(N);
252 }], SHUFFLE_get_shuf_imm>;
254 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
255 return X86::isPSHUFHWMask(N);
256 }], SHUFFLE_get_pshufhw_imm>;
258 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
259 return X86::isPSHUFLWMask(N);
260 }], SHUFFLE_get_pshuflw_imm>;
262 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
263 return X86::isPSHUFDMask(N);
264 }], SHUFFLE_get_shuf_imm>;
266 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
267 return X86::isSHUFPMask(N);
268 }], SHUFFLE_get_shuf_imm>;
270 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
271 return X86::isSHUFPMask(N);
272 }], SHUFFLE_get_shuf_imm>;
275 //===----------------------------------------------------------------------===//
276 // SSE scalar FP Instructions
277 //===----------------------------------------------------------------------===//
279 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
280 // scheduler into a branch sequence.
281 // These are expanded by the scheduler.
282 let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
283 def CMOV_FR32 : I<0, Pseudo,
284 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
285 "#CMOV_FR32 PSEUDO!",
286 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
288 def CMOV_FR64 : I<0, Pseudo,
289 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
290 "#CMOV_FR64 PSEUDO!",
291 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
293 def CMOV_V4F32 : I<0, Pseudo,
294 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
295 "#CMOV_V4F32 PSEUDO!",
297 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
299 def CMOV_V2F64 : I<0, Pseudo,
300 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
301 "#CMOV_V2F64 PSEUDO!",
303 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
305 def CMOV_V2I64 : I<0, Pseudo,
306 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
307 "#CMOV_V2I64 PSEUDO!",
309 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
313 //===----------------------------------------------------------------------===//
315 //===----------------------------------------------------------------------===//
318 let neverHasSideEffects = 1 in
319 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
320 "movss\t{$src, $dst|$dst, $src}", []>;
321 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
322 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
323 "movss\t{$src, $dst|$dst, $src}",
324 [(set FR32:$dst, (loadf32 addr:$src))]>;
325 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
326 "movss\t{$src, $dst|$dst, $src}",
327 [(store FR32:$src, addr:$dst)]>;
329 // Conversion instructions
330 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
331 "cvttss2si\t{$src, $dst|$dst, $src}",
332 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
333 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
334 "cvttss2si\t{$src, $dst|$dst, $src}",
335 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
336 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
337 "cvtsi2ss\t{$src, $dst|$dst, $src}",
338 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
339 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
340 "cvtsi2ss\t{$src, $dst|$dst, $src}",
341 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
343 // Match intrinsics which expect XMM operand(s).
344 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
345 "cvtss2si\t{$src, $dst|$dst, $src}",
346 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
347 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
348 "cvtss2si\t{$src, $dst|$dst, $src}",
349 [(set GR32:$dst, (int_x86_sse_cvtss2si
350 (load addr:$src)))]>;
352 // Match intrinisics which expect MM and XMM operand(s).
353 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
354 "cvtps2pi\t{$src, $dst|$dst, $src}",
355 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
356 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
357 "cvtps2pi\t{$src, $dst|$dst, $src}",
358 [(set VR64:$dst, (int_x86_sse_cvtps2pi
359 (load addr:$src)))]>;
360 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
361 "cvttps2pi\t{$src, $dst|$dst, $src}",
362 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
363 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
364 "cvttps2pi\t{$src, $dst|$dst, $src}",
365 [(set VR64:$dst, (int_x86_sse_cvttps2pi
366 (load addr:$src)))]>;
367 let Constraints = "$src1 = $dst" in {
368 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
369 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
370 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
371 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
373 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
374 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
375 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
376 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
377 (load addr:$src2)))]>;
380 // Aliases for intrinsics
381 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
382 "cvttss2si\t{$src, $dst|$dst, $src}",
384 (int_x86_sse_cvttss2si VR128:$src))]>;
385 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
386 "cvttss2si\t{$src, $dst|$dst, $src}",
388 (int_x86_sse_cvttss2si(load addr:$src)))]>;
390 let Constraints = "$src1 = $dst" in {
391 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
392 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
393 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
394 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
396 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
397 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
398 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
399 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
400 (loadi32 addr:$src2)))]>;
403 // Comparison instructions
404 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
405 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
406 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
407 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
409 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
410 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
411 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
414 let Defs = [EFLAGS] in {
415 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
416 "ucomiss\t{$src2, $src1|$src1, $src2}",
417 [(X86cmp FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
418 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
419 "ucomiss\t{$src2, $src1|$src1, $src2}",
420 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
424 // Aliases to match intrinsics which expect XMM operand(s).
425 let Constraints = "$src1 = $dst" in {
426 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
427 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
428 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
429 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
430 VR128:$src, imm:$cc))]>;
431 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
432 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src, SSECC:$cc),
433 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
434 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
435 (load addr:$src), imm:$cc))]>;
438 let Defs = [EFLAGS] in {
439 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
440 "ucomiss\t{$src2, $src1|$src1, $src2}",
441 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
443 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
444 "ucomiss\t{$src2, $src1|$src1, $src2}",
445 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
448 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
449 "comiss\t{$src2, $src1|$src1, $src2}",
450 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
452 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
453 "comiss\t{$src2, $src1|$src1, $src2}",
454 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
458 // Aliases of packed SSE1 instructions for scalar use. These all have names that
461 // Alias instructions that map fld0 to pxor for sse.
462 let isReMaterializable = 1, isAsCheapAsAMove = 1 in
463 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
464 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
465 Requires<[HasSSE1]>, TB, OpSize;
467 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
469 let neverHasSideEffects = 1 in
470 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
471 "movaps\t{$src, $dst|$dst, $src}", []>;
473 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
475 let canFoldAsLoad = 1 in
476 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
477 "movaps\t{$src, $dst|$dst, $src}",
478 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
480 // Alias bitwise logical operations using SSE logical ops on packed FP values.
481 let Constraints = "$src1 = $dst" in {
482 let isCommutable = 1 in {
483 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst),
484 (ins FR32:$src1, FR32:$src2),
485 "andps\t{$src2, $dst|$dst, $src2}",
486 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
487 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst),
488 (ins FR32:$src1, FR32:$src2),
489 "orps\t{$src2, $dst|$dst, $src2}",
490 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
491 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst),
492 (ins FR32:$src1, FR32:$src2),
493 "xorps\t{$src2, $dst|$dst, $src2}",
494 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
497 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst),
498 (ins FR32:$src1, f128mem:$src2),
499 "andps\t{$src2, $dst|$dst, $src2}",
500 [(set FR32:$dst, (X86fand FR32:$src1,
501 (memopfsf32 addr:$src2)))]>;
502 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst),
503 (ins FR32:$src1, f128mem:$src2),
504 "orps\t{$src2, $dst|$dst, $src2}",
505 [(set FR32:$dst, (X86for FR32:$src1,
506 (memopfsf32 addr:$src2)))]>;
507 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst),
508 (ins FR32:$src1, f128mem:$src2),
509 "xorps\t{$src2, $dst|$dst, $src2}",
510 [(set FR32:$dst, (X86fxor FR32:$src1,
511 (memopfsf32 addr:$src2)))]>;
513 let neverHasSideEffects = 1 in {
514 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
515 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
516 "andnps\t{$src2, $dst|$dst, $src2}", []>;
518 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
519 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
520 "andnps\t{$src2, $dst|$dst, $src2}", []>;
524 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
526 /// In addition, we also have a special variant of the scalar form here to
527 /// represent the associated intrinsic operation. This form is unlike the
528 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
529 /// and leaves the top elements unmodified (therefore these cannot be commuted).
531 /// These three forms can each be reg+reg or reg+mem, so there are a total of
532 /// six "instructions".
534 let Constraints = "$src1 = $dst" in {
535 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
536 SDNode OpNode, Intrinsic F32Int,
537 bit Commutable = 0> {
538 // Scalar operation, reg+reg.
539 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
540 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
541 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
542 let isCommutable = Commutable;
545 // Scalar operation, reg+mem.
546 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
547 (ins FR32:$src1, f32mem:$src2),
548 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
549 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
551 // Vector operation, reg+reg.
552 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
553 (ins VR128:$src1, VR128:$src2),
554 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
555 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
556 let isCommutable = Commutable;
559 // Vector operation, reg+mem.
560 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
561 (ins VR128:$src1, f128mem:$src2),
562 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
563 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
565 // Intrinsic operation, reg+reg.
566 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
567 (ins VR128:$src1, VR128:$src2),
568 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
569 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]>;
571 // Intrinsic operation, reg+mem.
572 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
573 (ins VR128:$src1, ssmem:$src2),
574 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
575 [(set VR128:$dst, (F32Int VR128:$src1,
576 sse_load_f32:$src2))]>;
580 // Arithmetic instructions
581 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
582 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
583 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
584 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
586 /// sse1_fp_binop_rm - Other SSE1 binops
588 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
589 /// instructions for a full-vector intrinsic form. Operations that map
590 /// onto C operators don't use this form since they just use the plain
591 /// vector form instead of having a separate vector intrinsic form.
593 /// This provides a total of eight "instructions".
595 let Constraints = "$src1 = $dst" in {
596 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
600 bit Commutable = 0> {
602 // Scalar operation, reg+reg.
603 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
604 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
605 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
606 let isCommutable = Commutable;
609 // Scalar operation, reg+mem.
610 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
611 (ins FR32:$src1, f32mem:$src2),
612 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
613 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
615 // Vector operation, reg+reg.
616 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
617 (ins VR128:$src1, VR128:$src2),
618 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
619 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
620 let isCommutable = Commutable;
623 // Vector operation, reg+mem.
624 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
625 (ins VR128:$src1, f128mem:$src2),
626 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
627 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
629 // Intrinsic operation, reg+reg.
630 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
631 (ins VR128:$src1, VR128:$src2),
632 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
633 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
634 let isCommutable = Commutable;
637 // Intrinsic operation, reg+mem.
638 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
639 (ins VR128:$src1, ssmem:$src2),
640 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
641 [(set VR128:$dst, (F32Int VR128:$src1,
642 sse_load_f32:$src2))]>;
644 // Vector intrinsic operation, reg+reg.
645 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
646 (ins VR128:$src1, VR128:$src2),
647 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
648 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
649 let isCommutable = Commutable;
652 // Vector intrinsic operation, reg+mem.
653 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
654 (ins VR128:$src1, f128mem:$src2),
655 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
656 [(set VR128:$dst, (V4F32Int VR128:$src1, (memopv4f32 addr:$src2)))]>;
660 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
661 int_x86_sse_max_ss, int_x86_sse_max_ps>;
662 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
663 int_x86_sse_min_ss, int_x86_sse_min_ps>;
665 //===----------------------------------------------------------------------===//
666 // SSE packed FP Instructions
669 let neverHasSideEffects = 1 in
670 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
671 "movaps\t{$src, $dst|$dst, $src}", []>;
672 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
673 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
674 "movaps\t{$src, $dst|$dst, $src}",
675 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
677 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
678 "movaps\t{$src, $dst|$dst, $src}",
679 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
681 let neverHasSideEffects = 1 in
682 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
683 "movups\t{$src, $dst|$dst, $src}", []>;
684 let canFoldAsLoad = 1 in
685 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
686 "movups\t{$src, $dst|$dst, $src}",
687 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
688 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
689 "movups\t{$src, $dst|$dst, $src}",
690 [(store (v4f32 VR128:$src), addr:$dst)]>;
692 // Intrinsic forms of MOVUPS load and store
693 let canFoldAsLoad = 1 in
694 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
695 "movups\t{$src, $dst|$dst, $src}",
696 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
697 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
698 "movups\t{$src, $dst|$dst, $src}",
699 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
701 let Constraints = "$src1 = $dst" in {
702 let AddedComplexity = 20 in {
703 def MOVLPSrm : PSI<0x12, MRMSrcMem,
704 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
705 "movlps\t{$src2, $dst|$dst, $src2}",
707 (v4f32 (vector_shuffle VR128:$src1,
708 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
709 MOVLP_shuffle_mask)))]>;
710 def MOVHPSrm : PSI<0x16, MRMSrcMem,
711 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
712 "movhps\t{$src2, $dst|$dst, $src2}",
714 (v4f32 (vector_shuffle VR128:$src1,
715 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
716 MOVHP_shuffle_mask)))]>;
718 } // Constraints = "$src1 = $dst"
721 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
722 "movlps\t{$src, $dst|$dst, $src}",
723 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
724 (iPTR 0))), addr:$dst)]>;
726 // v2f64 extract element 1 is always custom lowered to unpack high to low
727 // and extract element 0 so the non-store version isn't too horrible.
728 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
729 "movhps\t{$src, $dst|$dst, $src}",
730 [(store (f64 (vector_extract
731 (v2f64 (vector_shuffle
732 (bc_v2f64 (v4f32 VR128:$src)), (undef),
733 UNPCKH_shuffle_mask)), (iPTR 0))),
736 let Constraints = "$src1 = $dst" in {
737 let AddedComplexity = 20 in {
738 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
739 "movlhps\t{$src2, $dst|$dst, $src2}",
741 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
742 MOVHP_shuffle_mask)))]>;
744 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
745 "movhlps\t{$src2, $dst|$dst, $src2}",
747 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
748 MOVHLPS_shuffle_mask)))]>;
750 } // Constraints = "$src1 = $dst"
752 let AddedComplexity = 20 in
753 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef), MOVDDUP_shuffle_mask)),
754 (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
761 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
763 /// In addition, we also have a special variant of the scalar form here to
764 /// represent the associated intrinsic operation. This form is unlike the
765 /// plain scalar form, in that it takes an entire vector (instead of a
766 /// scalar) and leaves the top elements undefined.
768 /// And, we have a special variant form for a full-vector intrinsic form.
770 /// These four forms can each have a reg or a mem operand, so there are a
771 /// total of eight "instructions".
773 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
777 bit Commutable = 0> {
778 // Scalar operation, reg.
779 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
780 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
781 [(set FR32:$dst, (OpNode FR32:$src))]> {
782 let isCommutable = Commutable;
785 // Scalar operation, mem.
786 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
787 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
788 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
790 // Vector operation, reg.
791 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
792 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
793 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
794 let isCommutable = Commutable;
797 // Vector operation, mem.
798 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
799 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
800 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
802 // Intrinsic operation, reg.
803 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
804 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
805 [(set VR128:$dst, (F32Int VR128:$src))]> {
806 let isCommutable = Commutable;
809 // Intrinsic operation, mem.
810 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
811 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
812 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
814 // Vector intrinsic operation, reg
815 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
816 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
817 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
818 let isCommutable = Commutable;
821 // Vector intrinsic operation, mem
822 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
823 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
824 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
828 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
829 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
831 // Reciprocal approximations. Note that these typically require refinement
832 // in order to obtain suitable precision.
833 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
834 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
835 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
836 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
839 let Constraints = "$src1 = $dst" in {
840 let isCommutable = 1 in {
841 def ANDPSrr : PSI<0x54, MRMSrcReg,
842 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
843 "andps\t{$src2, $dst|$dst, $src2}",
844 [(set VR128:$dst, (v2i64
845 (and VR128:$src1, VR128:$src2)))]>;
846 def ORPSrr : PSI<0x56, MRMSrcReg,
847 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
848 "orps\t{$src2, $dst|$dst, $src2}",
849 [(set VR128:$dst, (v2i64
850 (or VR128:$src1, VR128:$src2)))]>;
851 def XORPSrr : PSI<0x57, MRMSrcReg,
852 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
853 "xorps\t{$src2, $dst|$dst, $src2}",
854 [(set VR128:$dst, (v2i64
855 (xor VR128:$src1, VR128:$src2)))]>;
858 def ANDPSrm : PSI<0x54, MRMSrcMem,
859 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
860 "andps\t{$src2, $dst|$dst, $src2}",
861 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
862 (memopv2i64 addr:$src2)))]>;
863 def ORPSrm : PSI<0x56, MRMSrcMem,
864 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
865 "orps\t{$src2, $dst|$dst, $src2}",
866 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
867 (memopv2i64 addr:$src2)))]>;
868 def XORPSrm : PSI<0x57, MRMSrcMem,
869 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
870 "xorps\t{$src2, $dst|$dst, $src2}",
871 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
872 (memopv2i64 addr:$src2)))]>;
873 def ANDNPSrr : PSI<0x55, MRMSrcReg,
874 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
875 "andnps\t{$src2, $dst|$dst, $src2}",
877 (v2i64 (and (xor VR128:$src1,
878 (bc_v2i64 (v4i32 immAllOnesV))),
880 def ANDNPSrm : PSI<0x55, MRMSrcMem,
881 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
882 "andnps\t{$src2, $dst|$dst, $src2}",
884 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
885 (bc_v2i64 (v4i32 immAllOnesV))),
886 (memopv2i64 addr:$src2))))]>;
889 let Constraints = "$src1 = $dst" in {
890 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
891 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
892 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
893 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
894 VR128:$src, imm:$cc))]>;
895 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
896 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
897 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
898 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
899 (memop addr:$src), imm:$cc))]>;
901 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
902 (CMPPSrri VR128:$src1, VR128:$src2, imm:$cc)>;
903 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
904 (CMPPSrmi VR128:$src1, addr:$src2, imm:$cc)>;
906 // Shuffle and unpack instructions
907 let Constraints = "$src1 = $dst" in {
908 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
909 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
910 (outs VR128:$dst), (ins VR128:$src1,
911 VR128:$src2, i32i8imm:$src3),
912 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
914 (v4f32 (vector_shuffle
915 VR128:$src1, VR128:$src2,
916 SHUFP_shuffle_mask:$src3)))]>;
917 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
918 (outs VR128:$dst), (ins VR128:$src1,
919 f128mem:$src2, i32i8imm:$src3),
920 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
922 (v4f32 (vector_shuffle
923 VR128:$src1, (memopv4f32 addr:$src2),
924 SHUFP_shuffle_mask:$src3)))]>;
926 let AddedComplexity = 10 in {
927 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
928 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
929 "unpckhps\t{$src2, $dst|$dst, $src2}",
931 (v4f32 (vector_shuffle
932 VR128:$src1, VR128:$src2,
933 UNPCKH_shuffle_mask)))]>;
934 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
935 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
936 "unpckhps\t{$src2, $dst|$dst, $src2}",
938 (v4f32 (vector_shuffle
939 VR128:$src1, (memopv4f32 addr:$src2),
940 UNPCKH_shuffle_mask)))]>;
942 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
943 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
944 "unpcklps\t{$src2, $dst|$dst, $src2}",
946 (v4f32 (vector_shuffle
947 VR128:$src1, VR128:$src2,
948 UNPCKL_shuffle_mask)))]>;
949 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
950 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
951 "unpcklps\t{$src2, $dst|$dst, $src2}",
953 (v4f32 (vector_shuffle
954 VR128:$src1, (memopv4f32 addr:$src2),
955 UNPCKL_shuffle_mask)))]>;
957 } // Constraints = "$src1 = $dst"
960 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
961 "movmskps\t{$src, $dst|$dst, $src}",
962 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
963 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
964 "movmskpd\t{$src, $dst|$dst, $src}",
965 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
967 // Prefetch intrinsic.
968 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
969 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
970 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
971 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
972 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
973 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
974 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
975 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
977 // Non-temporal stores
978 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
979 "movntps\t{$src, $dst|$dst, $src}",
980 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
982 // Load, store, and memory fence
983 def SFENCE : PSI<0xAE, MRM7m, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
986 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
987 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
988 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
989 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
991 // Alias instructions that map zero vector to pxor / xorp* for sse.
992 // We set canFoldAsLoad because this can be converted to a constant-pool
993 // load of an all-zeros value if folding it would be beneficial.
994 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1 in
995 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
997 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
999 let Predicates = [HasSSE1] in {
1000 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
1001 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
1002 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
1003 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
1004 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
1007 // FR32 to 128-bit vector conversion.
1008 let isAsCheapAsAMove = 1 in
1009 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
1010 "movss\t{$src, $dst|$dst, $src}",
1012 (v4f32 (scalar_to_vector FR32:$src)))]>;
1013 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1014 "movss\t{$src, $dst|$dst, $src}",
1016 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
1018 // FIXME: may not be able to eliminate this movss with coalescing the src and
1019 // dest register classes are different. We really want to write this pattern
1021 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1022 // (f32 FR32:$src)>;
1023 let isAsCheapAsAMove = 1 in
1024 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
1025 "movss\t{$src, $dst|$dst, $src}",
1026 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
1028 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
1029 "movss\t{$src, $dst|$dst, $src}",
1030 [(store (f32 (vector_extract (v4f32 VR128:$src),
1031 (iPTR 0))), addr:$dst)]>;
1034 // Move to lower bits of a VR128, leaving upper bits alone.
1035 // Three operand (but two address) aliases.
1036 let Constraints = "$src1 = $dst" in {
1037 let neverHasSideEffects = 1 in
1038 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
1039 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
1040 "movss\t{$src2, $dst|$dst, $src2}", []>;
1042 let AddedComplexity = 15 in
1043 def MOVLPSrr : SSI<0x10, MRMSrcReg,
1044 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1045 "movss\t{$src2, $dst|$dst, $src2}",
1047 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
1048 MOVL_shuffle_mask)))]>;
1051 // Move to lower bits of a VR128 and zeroing upper bits.
1052 // Loading from memory automatically zeroing upper bits.
1053 let AddedComplexity = 20 in
1054 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1055 "movss\t{$src, $dst|$dst, $src}",
1056 [(set VR128:$dst, (v4f32 (X86vzmovl (v4f32 (scalar_to_vector
1057 (loadf32 addr:$src))))))]>;
1059 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
1060 (MOVZSS2PSrm addr:$src)>;
1062 //===----------------------------------------------------------------------===//
1063 // SSE2 Instructions
1064 //===----------------------------------------------------------------------===//
1066 // Move Instructions
1067 let neverHasSideEffects = 1 in
1068 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1069 "movsd\t{$src, $dst|$dst, $src}", []>;
1070 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1071 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1072 "movsd\t{$src, $dst|$dst, $src}",
1073 [(set FR64:$dst, (loadf64 addr:$src))]>;
1074 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1075 "movsd\t{$src, $dst|$dst, $src}",
1076 [(store FR64:$src, addr:$dst)]>;
1078 // Conversion instructions
1079 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1080 "cvttsd2si\t{$src, $dst|$dst, $src}",
1081 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1082 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1083 "cvttsd2si\t{$src, $dst|$dst, $src}",
1084 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1085 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1086 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1087 [(set FR32:$dst, (fround FR64:$src))]>;
1088 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1089 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1090 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1091 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1092 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1093 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1094 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1095 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1096 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1098 // SSE2 instructions with XS prefix
1099 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1100 "cvtss2sd\t{$src, $dst|$dst, $src}",
1101 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1102 Requires<[HasSSE2]>;
1103 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1104 "cvtss2sd\t{$src, $dst|$dst, $src}",
1105 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1106 Requires<[HasSSE2]>;
1108 // Match intrinsics which expect XMM operand(s).
1109 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1110 "cvtsd2si\t{$src, $dst|$dst, $src}",
1111 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1112 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1113 "cvtsd2si\t{$src, $dst|$dst, $src}",
1114 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1115 (load addr:$src)))]>;
1117 // Match intrinisics which expect MM and XMM operand(s).
1118 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1119 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1120 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1121 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1122 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1123 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1124 (memop addr:$src)))]>;
1125 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1126 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1127 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1128 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1129 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1130 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1131 (memop addr:$src)))]>;
1132 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1133 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1134 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1135 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1136 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1137 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1138 (load addr:$src)))]>;
1140 // Aliases for intrinsics
1141 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1142 "cvttsd2si\t{$src, $dst|$dst, $src}",
1144 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1145 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1146 "cvttsd2si\t{$src, $dst|$dst, $src}",
1147 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1148 (load addr:$src)))]>;
1150 // Comparison instructions
1151 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1152 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1153 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1154 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1156 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1157 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1158 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1161 let Defs = [EFLAGS] in {
1162 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1163 "ucomisd\t{$src2, $src1|$src1, $src2}",
1164 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1165 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1166 "ucomisd\t{$src2, $src1|$src1, $src2}",
1167 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1168 (implicit EFLAGS)]>;
1169 } // Defs = [EFLAGS]
1171 // Aliases to match intrinsics which expect XMM operand(s).
1172 let Constraints = "$src1 = $dst" in {
1173 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1174 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1175 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1176 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1177 VR128:$src, imm:$cc))]>;
1178 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1179 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1180 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1181 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1182 (load addr:$src), imm:$cc))]>;
1185 let Defs = [EFLAGS] in {
1186 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1187 "ucomisd\t{$src2, $src1|$src1, $src2}",
1188 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1189 (implicit EFLAGS)]>;
1190 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1191 "ucomisd\t{$src2, $src1|$src1, $src2}",
1192 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1193 (implicit EFLAGS)]>;
1195 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1196 "comisd\t{$src2, $src1|$src1, $src2}",
1197 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1198 (implicit EFLAGS)]>;
1199 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1200 "comisd\t{$src2, $src1|$src1, $src2}",
1201 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1202 (implicit EFLAGS)]>;
1203 } // Defs = [EFLAGS]
1205 // Aliases of packed SSE2 instructions for scalar use. These all have names that
1208 // Alias instructions that map fld0 to pxor for sse.
1209 let isReMaterializable = 1, isAsCheapAsAMove = 1 in
1210 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1211 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1212 Requires<[HasSSE2]>, TB, OpSize;
1214 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1216 let neverHasSideEffects = 1 in
1217 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1218 "movapd\t{$src, $dst|$dst, $src}", []>;
1220 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1222 let canFoldAsLoad = 1 in
1223 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1224 "movapd\t{$src, $dst|$dst, $src}",
1225 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1227 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1228 let Constraints = "$src1 = $dst" in {
1229 let isCommutable = 1 in {
1230 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1231 (ins FR64:$src1, FR64:$src2),
1232 "andpd\t{$src2, $dst|$dst, $src2}",
1233 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1234 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1235 (ins FR64:$src1, FR64:$src2),
1236 "orpd\t{$src2, $dst|$dst, $src2}",
1237 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1238 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1239 (ins FR64:$src1, FR64:$src2),
1240 "xorpd\t{$src2, $dst|$dst, $src2}",
1241 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1244 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1245 (ins FR64:$src1, f128mem:$src2),
1246 "andpd\t{$src2, $dst|$dst, $src2}",
1247 [(set FR64:$dst, (X86fand FR64:$src1,
1248 (memopfsf64 addr:$src2)))]>;
1249 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1250 (ins FR64:$src1, f128mem:$src2),
1251 "orpd\t{$src2, $dst|$dst, $src2}",
1252 [(set FR64:$dst, (X86for FR64:$src1,
1253 (memopfsf64 addr:$src2)))]>;
1254 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1255 (ins FR64:$src1, f128mem:$src2),
1256 "xorpd\t{$src2, $dst|$dst, $src2}",
1257 [(set FR64:$dst, (X86fxor FR64:$src1,
1258 (memopfsf64 addr:$src2)))]>;
1260 let neverHasSideEffects = 1 in {
1261 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1262 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1263 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1265 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1266 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1267 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1271 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1273 /// In addition, we also have a special variant of the scalar form here to
1274 /// represent the associated intrinsic operation. This form is unlike the
1275 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1276 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1278 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1279 /// six "instructions".
1281 let Constraints = "$src1 = $dst" in {
1282 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1283 SDNode OpNode, Intrinsic F64Int,
1284 bit Commutable = 0> {
1285 // Scalar operation, reg+reg.
1286 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1287 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1288 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1289 let isCommutable = Commutable;
1292 // Scalar operation, reg+mem.
1293 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1294 (ins FR64:$src1, f64mem:$src2),
1295 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1296 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1298 // Vector operation, reg+reg.
1299 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1300 (ins VR128:$src1, VR128:$src2),
1301 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1302 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1303 let isCommutable = Commutable;
1306 // Vector operation, reg+mem.
1307 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1308 (ins VR128:$src1, f128mem:$src2),
1309 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1310 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1312 // Intrinsic operation, reg+reg.
1313 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1314 (ins VR128:$src1, VR128:$src2),
1315 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1316 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]>;
1318 // Intrinsic operation, reg+mem.
1319 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1320 (ins VR128:$src1, sdmem:$src2),
1321 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1322 [(set VR128:$dst, (F64Int VR128:$src1,
1323 sse_load_f64:$src2))]>;
1327 // Arithmetic instructions
1328 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1329 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1330 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1331 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1333 /// sse2_fp_binop_rm - Other SSE2 binops
1335 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1336 /// instructions for a full-vector intrinsic form. Operations that map
1337 /// onto C operators don't use this form since they just use the plain
1338 /// vector form instead of having a separate vector intrinsic form.
1340 /// This provides a total of eight "instructions".
1342 let Constraints = "$src1 = $dst" in {
1343 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1347 bit Commutable = 0> {
1349 // Scalar operation, reg+reg.
1350 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1351 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1352 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1353 let isCommutable = Commutable;
1356 // Scalar operation, reg+mem.
1357 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1358 (ins FR64:$src1, f64mem:$src2),
1359 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1360 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1362 // Vector operation, reg+reg.
1363 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1364 (ins VR128:$src1, VR128:$src2),
1365 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1366 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1367 let isCommutable = Commutable;
1370 // Vector operation, reg+mem.
1371 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1372 (ins VR128:$src1, f128mem:$src2),
1373 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1374 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1376 // Intrinsic operation, reg+reg.
1377 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
1378 (ins VR128:$src1, VR128:$src2),
1379 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1380 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1381 let isCommutable = Commutable;
1384 // Intrinsic operation, reg+mem.
1385 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
1386 (ins VR128:$src1, sdmem:$src2),
1387 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1388 [(set VR128:$dst, (F64Int VR128:$src1,
1389 sse_load_f64:$src2))]>;
1391 // Vector intrinsic operation, reg+reg.
1392 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1393 (ins VR128:$src1, VR128:$src2),
1394 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1395 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1396 let isCommutable = Commutable;
1399 // Vector intrinsic operation, reg+mem.
1400 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1401 (ins VR128:$src1, f128mem:$src2),
1402 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1403 [(set VR128:$dst, (V2F64Int VR128:$src1,
1404 (memopv2f64 addr:$src2)))]>;
1408 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1409 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1410 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1411 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1413 //===----------------------------------------------------------------------===//
1414 // SSE packed FP Instructions
1416 // Move Instructions
1417 let neverHasSideEffects = 1 in
1418 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1419 "movapd\t{$src, $dst|$dst, $src}", []>;
1420 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1421 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1422 "movapd\t{$src, $dst|$dst, $src}",
1423 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1425 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1426 "movapd\t{$src, $dst|$dst, $src}",
1427 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1429 let neverHasSideEffects = 1 in
1430 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1431 "movupd\t{$src, $dst|$dst, $src}", []>;
1432 let canFoldAsLoad = 1 in
1433 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1434 "movupd\t{$src, $dst|$dst, $src}",
1435 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1436 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1437 "movupd\t{$src, $dst|$dst, $src}",
1438 [(store (v2f64 VR128:$src), addr:$dst)]>;
1440 // Intrinsic forms of MOVUPD load and store
1441 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1442 "movupd\t{$src, $dst|$dst, $src}",
1443 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1444 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1445 "movupd\t{$src, $dst|$dst, $src}",
1446 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1448 let Constraints = "$src1 = $dst" in {
1449 let AddedComplexity = 20 in {
1450 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1451 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1452 "movlpd\t{$src2, $dst|$dst, $src2}",
1454 (v2f64 (vector_shuffle VR128:$src1,
1455 (scalar_to_vector (loadf64 addr:$src2)),
1456 MOVLP_shuffle_mask)))]>;
1457 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1458 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1459 "movhpd\t{$src2, $dst|$dst, $src2}",
1461 (v2f64 (vector_shuffle VR128:$src1,
1462 (scalar_to_vector (loadf64 addr:$src2)),
1463 MOVHP_shuffle_mask)))]>;
1464 } // AddedComplexity
1465 } // Constraints = "$src1 = $dst"
1467 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1468 "movlpd\t{$src, $dst|$dst, $src}",
1469 [(store (f64 (vector_extract (v2f64 VR128:$src),
1470 (iPTR 0))), addr:$dst)]>;
1472 // v2f64 extract element 1 is always custom lowered to unpack high to low
1473 // and extract element 0 so the non-store version isn't too horrible.
1474 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1475 "movhpd\t{$src, $dst|$dst, $src}",
1476 [(store (f64 (vector_extract
1477 (v2f64 (vector_shuffle VR128:$src, (undef),
1478 UNPCKH_shuffle_mask)), (iPTR 0))),
1481 // SSE2 instructions without OpSize prefix
1482 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1483 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1484 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1485 TB, Requires<[HasSSE2]>;
1486 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1487 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1488 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1489 (bitconvert (memopv2i64 addr:$src))))]>,
1490 TB, Requires<[HasSSE2]>;
1492 // SSE2 instructions with XS prefix
1493 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1494 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1495 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1496 XS, Requires<[HasSSE2]>;
1497 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1498 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1499 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1500 (bitconvert (memopv2i64 addr:$src))))]>,
1501 XS, Requires<[HasSSE2]>;
1503 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1504 "cvtps2dq\t{$src, $dst|$dst, $src}",
1505 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1506 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1507 "cvtps2dq\t{$src, $dst|$dst, $src}",
1508 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1509 (memop addr:$src)))]>;
1510 // SSE2 packed instructions with XS prefix
1511 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1512 "cvttps2dq\t{$src, $dst|$dst, $src}",
1513 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1514 XS, Requires<[HasSSE2]>;
1515 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1516 "cvttps2dq\t{$src, $dst|$dst, $src}",
1517 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1518 (memop addr:$src)))]>,
1519 XS, Requires<[HasSSE2]>;
1521 // SSE2 packed instructions with XD prefix
1522 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1523 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1524 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1525 XD, Requires<[HasSSE2]>;
1526 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1527 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1528 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1529 (memop addr:$src)))]>,
1530 XD, Requires<[HasSSE2]>;
1532 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1533 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1534 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1535 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1536 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1537 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1538 (memop addr:$src)))]>;
1540 // SSE2 instructions without OpSize prefix
1541 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1542 "cvtps2pd\t{$src, $dst|$dst, $src}",
1543 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1544 TB, Requires<[HasSSE2]>;
1545 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1546 "cvtps2pd\t{$src, $dst|$dst, $src}",
1547 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1548 (load addr:$src)))]>,
1549 TB, Requires<[HasSSE2]>;
1551 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1552 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1553 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1554 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1555 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1556 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1557 (memop addr:$src)))]>;
1559 // Match intrinsics which expect XMM operand(s).
1560 // Aliases for intrinsics
1561 let Constraints = "$src1 = $dst" in {
1562 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1563 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1564 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1565 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1567 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1568 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1569 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1570 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1571 (loadi32 addr:$src2)))]>;
1572 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1573 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1574 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1575 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1577 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1578 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1579 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1580 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1581 (load addr:$src2)))]>;
1582 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1583 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1584 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1585 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1586 VR128:$src2))]>, XS,
1587 Requires<[HasSSE2]>;
1588 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1589 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1590 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1591 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1592 (load addr:$src2)))]>, XS,
1593 Requires<[HasSSE2]>;
1598 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1600 /// In addition, we also have a special variant of the scalar form here to
1601 /// represent the associated intrinsic operation. This form is unlike the
1602 /// plain scalar form, in that it takes an entire vector (instead of a
1603 /// scalar) and leaves the top elements undefined.
1605 /// And, we have a special variant form for a full-vector intrinsic form.
1607 /// These four forms can each have a reg or a mem operand, so there are a
1608 /// total of eight "instructions".
1610 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1614 bit Commutable = 0> {
1615 // Scalar operation, reg.
1616 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1617 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1618 [(set FR64:$dst, (OpNode FR64:$src))]> {
1619 let isCommutable = Commutable;
1622 // Scalar operation, mem.
1623 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1624 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1625 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1627 // Vector operation, reg.
1628 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1629 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1630 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1631 let isCommutable = Commutable;
1634 // Vector operation, mem.
1635 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1636 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1637 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1639 // Intrinsic operation, reg.
1640 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1641 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1642 [(set VR128:$dst, (F64Int VR128:$src))]> {
1643 let isCommutable = Commutable;
1646 // Intrinsic operation, mem.
1647 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1648 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1649 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1651 // Vector intrinsic operation, reg
1652 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1653 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1654 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1655 let isCommutable = Commutable;
1658 // Vector intrinsic operation, mem
1659 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1660 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1661 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1665 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1666 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1668 // There is no f64 version of the reciprocal approximation instructions.
1671 let Constraints = "$src1 = $dst" in {
1672 let isCommutable = 1 in {
1673 def ANDPDrr : PDI<0x54, MRMSrcReg,
1674 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1675 "andpd\t{$src2, $dst|$dst, $src2}",
1677 (and (bc_v2i64 (v2f64 VR128:$src1)),
1678 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1679 def ORPDrr : PDI<0x56, MRMSrcReg,
1680 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1681 "orpd\t{$src2, $dst|$dst, $src2}",
1683 (or (bc_v2i64 (v2f64 VR128:$src1)),
1684 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1685 def XORPDrr : PDI<0x57, MRMSrcReg,
1686 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1687 "xorpd\t{$src2, $dst|$dst, $src2}",
1689 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1690 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1693 def ANDPDrm : PDI<0x54, MRMSrcMem,
1694 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1695 "andpd\t{$src2, $dst|$dst, $src2}",
1697 (and (bc_v2i64 (v2f64 VR128:$src1)),
1698 (memopv2i64 addr:$src2)))]>;
1699 def ORPDrm : PDI<0x56, MRMSrcMem,
1700 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1701 "orpd\t{$src2, $dst|$dst, $src2}",
1703 (or (bc_v2i64 (v2f64 VR128:$src1)),
1704 (memopv2i64 addr:$src2)))]>;
1705 def XORPDrm : PDI<0x57, MRMSrcMem,
1706 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1707 "xorpd\t{$src2, $dst|$dst, $src2}",
1709 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1710 (memopv2i64 addr:$src2)))]>;
1711 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1712 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1713 "andnpd\t{$src2, $dst|$dst, $src2}",
1715 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1716 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1717 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1718 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1719 "andnpd\t{$src2, $dst|$dst, $src2}",
1721 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1722 (memopv2i64 addr:$src2)))]>;
1725 let Constraints = "$src1 = $dst" in {
1726 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1727 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1728 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1729 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1730 VR128:$src, imm:$cc))]>;
1731 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1732 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1733 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1734 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1735 (memop addr:$src), imm:$cc))]>;
1737 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1738 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1739 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1740 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1742 // Shuffle and unpack instructions
1743 let Constraints = "$src1 = $dst" in {
1744 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1745 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1746 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1747 [(set VR128:$dst, (v2f64 (vector_shuffle
1748 VR128:$src1, VR128:$src2,
1749 SHUFP_shuffle_mask:$src3)))]>;
1750 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1751 (outs VR128:$dst), (ins VR128:$src1,
1752 f128mem:$src2, i8imm:$src3),
1753 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1755 (v2f64 (vector_shuffle
1756 VR128:$src1, (memopv2f64 addr:$src2),
1757 SHUFP_shuffle_mask:$src3)))]>;
1759 let AddedComplexity = 10 in {
1760 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1761 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1762 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1764 (v2f64 (vector_shuffle
1765 VR128:$src1, VR128:$src2,
1766 UNPCKH_shuffle_mask)))]>;
1767 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1768 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1769 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1771 (v2f64 (vector_shuffle
1772 VR128:$src1, (memopv2f64 addr:$src2),
1773 UNPCKH_shuffle_mask)))]>;
1775 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1776 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1777 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1779 (v2f64 (vector_shuffle
1780 VR128:$src1, VR128:$src2,
1781 UNPCKL_shuffle_mask)))]>;
1782 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1783 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1784 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1786 (v2f64 (vector_shuffle
1787 VR128:$src1, (memopv2f64 addr:$src2),
1788 UNPCKL_shuffle_mask)))]>;
1789 } // AddedComplexity
1790 } // Constraints = "$src1 = $dst"
1793 //===----------------------------------------------------------------------===//
1794 // SSE integer instructions
1796 // Move Instructions
1797 let neverHasSideEffects = 1 in
1798 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1799 "movdqa\t{$src, $dst|$dst, $src}", []>;
1800 let canFoldAsLoad = 1, mayLoad = 1 in
1801 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1802 "movdqa\t{$src, $dst|$dst, $src}",
1803 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1805 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1806 "movdqa\t{$src, $dst|$dst, $src}",
1807 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1808 let canFoldAsLoad = 1, mayLoad = 1 in
1809 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1810 "movdqu\t{$src, $dst|$dst, $src}",
1811 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1812 XS, Requires<[HasSSE2]>;
1814 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1815 "movdqu\t{$src, $dst|$dst, $src}",
1816 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1817 XS, Requires<[HasSSE2]>;
1819 // Intrinsic forms of MOVDQU load and store
1820 let canFoldAsLoad = 1 in
1821 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1822 "movdqu\t{$src, $dst|$dst, $src}",
1823 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1824 XS, Requires<[HasSSE2]>;
1825 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1826 "movdqu\t{$src, $dst|$dst, $src}",
1827 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1828 XS, Requires<[HasSSE2]>;
1830 let Constraints = "$src1 = $dst" in {
1832 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1833 bit Commutable = 0> {
1834 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1835 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1836 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1837 let isCommutable = Commutable;
1839 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1840 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1841 [(set VR128:$dst, (IntId VR128:$src1,
1842 (bitconvert (memopv2i64 addr:$src2))))]>;
1845 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1847 Intrinsic IntId, Intrinsic IntId2> {
1848 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1849 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1850 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1851 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1852 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1853 [(set VR128:$dst, (IntId VR128:$src1,
1854 (bitconvert (memopv2i64 addr:$src2))))]>;
1855 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1856 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1857 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1860 /// PDI_binop_rm - Simple SSE2 binary operator.
1861 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1862 ValueType OpVT, bit Commutable = 0> {
1863 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1864 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1865 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1866 let isCommutable = Commutable;
1868 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1869 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1870 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1871 (bitconvert (memopv2i64 addr:$src2)))))]>;
1874 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1876 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1877 /// to collapse (bitconvert VT to VT) into its operand.
1879 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1880 bit Commutable = 0> {
1881 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1882 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1883 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1884 let isCommutable = Commutable;
1886 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1887 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1888 [(set VR128:$dst, (OpNode VR128:$src1,(memopv2i64 addr:$src2)))]>;
1891 } // Constraints = "$src1 = $dst"
1893 // 128-bit Integer Arithmetic
1895 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1896 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1897 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1898 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1900 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1901 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1902 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1903 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1905 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1906 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1907 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1908 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1910 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1911 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1912 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1913 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1915 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1917 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1918 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1919 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1921 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1923 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1924 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1927 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1928 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1929 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1930 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1931 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1934 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
1935 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
1936 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
1937 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
1938 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
1939 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
1941 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
1942 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
1943 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
1944 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
1945 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
1946 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
1948 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
1949 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
1950 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
1951 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
1953 // 128-bit logical shifts.
1954 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1955 def PSLLDQri : PDIi8<0x73, MRM7r,
1956 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1957 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1958 def PSRLDQri : PDIi8<0x73, MRM3r,
1959 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1960 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1961 // PSRADQri doesn't exist in SSE[1-3].
1964 let Predicates = [HasSSE2] in {
1965 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1966 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1967 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1968 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1969 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
1970 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
1971 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
1972 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
1973 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1974 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1976 // Shift up / down and insert zero's.
1977 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
1978 (v2i64 (PSLLDQri VR128:$src, (PSxLDQ_imm imm:$amt)))>;
1979 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
1980 (v2i64 (PSRLDQri VR128:$src, (PSxLDQ_imm imm:$amt)))>;
1984 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1985 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1986 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1988 let Constraints = "$src1 = $dst" in {
1989 def PANDNrr : PDI<0xDF, MRMSrcReg,
1990 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1991 "pandn\t{$src2, $dst|$dst, $src2}",
1992 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1995 def PANDNrm : PDI<0xDF, MRMSrcMem,
1996 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1997 "pandn\t{$src2, $dst|$dst, $src2}",
1998 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1999 (memopv2i64 addr:$src2))))]>;
2002 // SSE2 Integer comparison
2003 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2004 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2005 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2006 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2007 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2008 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2010 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2011 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2012 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2013 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2014 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2015 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2016 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2017 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2018 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2019 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2020 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2021 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2023 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2024 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2025 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2026 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2027 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2028 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2029 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2030 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2031 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2032 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2033 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2034 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2037 // Pack instructions
2038 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2039 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2040 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2042 // Shuffle and unpack instructions
2043 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2044 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2045 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2046 [(set VR128:$dst, (v4i32 (vector_shuffle
2047 VR128:$src1, (undef),
2048 PSHUFD_shuffle_mask:$src2)))]>;
2049 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2050 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2051 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2052 [(set VR128:$dst, (v4i32 (vector_shuffle
2053 (bc_v4i32(memopv2i64 addr:$src1)),
2055 PSHUFD_shuffle_mask:$src2)))]>;
2057 // SSE2 with ImmT == Imm8 and XS prefix.
2058 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2059 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2060 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2061 [(set VR128:$dst, (v8i16 (vector_shuffle
2062 VR128:$src1, (undef),
2063 PSHUFHW_shuffle_mask:$src2)))]>,
2064 XS, Requires<[HasSSE2]>;
2065 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2066 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2067 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2068 [(set VR128:$dst, (v8i16 (vector_shuffle
2069 (bc_v8i16 (memopv2i64 addr:$src1)),
2071 PSHUFHW_shuffle_mask:$src2)))]>,
2072 XS, Requires<[HasSSE2]>;
2074 // SSE2 with ImmT == Imm8 and XD prefix.
2075 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2076 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2077 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2078 [(set VR128:$dst, (v8i16 (vector_shuffle
2079 VR128:$src1, (undef),
2080 PSHUFLW_shuffle_mask:$src2)))]>,
2081 XD, Requires<[HasSSE2]>;
2082 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2083 (outs VR128:$dst), (ins i128mem:$src1, i32i8imm:$src2),
2084 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2085 [(set VR128:$dst, (v8i16 (vector_shuffle
2086 (bc_v8i16 (memopv2i64 addr:$src1)),
2088 PSHUFLW_shuffle_mask:$src2)))]>,
2089 XD, Requires<[HasSSE2]>;
2092 let Constraints = "$src1 = $dst" in {
2093 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2094 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2095 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2097 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
2098 UNPCKL_shuffle_mask)))]>;
2099 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2100 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2101 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2103 (v16i8 (vector_shuffle VR128:$src1,
2104 (bc_v16i8 (memopv2i64 addr:$src2)),
2105 UNPCKL_shuffle_mask)))]>;
2106 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2107 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2108 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2110 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2111 UNPCKL_shuffle_mask)))]>;
2112 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2113 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2114 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2116 (v8i16 (vector_shuffle VR128:$src1,
2117 (bc_v8i16 (memopv2i64 addr:$src2)),
2118 UNPCKL_shuffle_mask)))]>;
2119 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2120 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2121 "punpckldq\t{$src2, $dst|$dst, $src2}",
2123 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2124 UNPCKL_shuffle_mask)))]>;
2125 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2126 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2127 "punpckldq\t{$src2, $dst|$dst, $src2}",
2129 (v4i32 (vector_shuffle VR128:$src1,
2130 (bc_v4i32 (memopv2i64 addr:$src2)),
2131 UNPCKL_shuffle_mask)))]>;
2132 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2133 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2134 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2136 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2137 UNPCKL_shuffle_mask)))]>;
2138 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2139 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2140 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2142 (v2i64 (vector_shuffle VR128:$src1,
2143 (memopv2i64 addr:$src2),
2144 UNPCKL_shuffle_mask)))]>;
2146 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2147 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2148 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2150 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
2151 UNPCKH_shuffle_mask)))]>;
2152 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2153 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2154 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2156 (v16i8 (vector_shuffle VR128:$src1,
2157 (bc_v16i8 (memopv2i64 addr:$src2)),
2158 UNPCKH_shuffle_mask)))]>;
2159 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2160 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2161 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2163 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2164 UNPCKH_shuffle_mask)))]>;
2165 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2166 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2167 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2169 (v8i16 (vector_shuffle VR128:$src1,
2170 (bc_v8i16 (memopv2i64 addr:$src2)),
2171 UNPCKH_shuffle_mask)))]>;
2172 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2173 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2174 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2176 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2177 UNPCKH_shuffle_mask)))]>;
2178 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2179 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2180 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2182 (v4i32 (vector_shuffle VR128:$src1,
2183 (bc_v4i32 (memopv2i64 addr:$src2)),
2184 UNPCKH_shuffle_mask)))]>;
2185 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2186 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2187 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2189 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2190 UNPCKH_shuffle_mask)))]>;
2191 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2192 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2193 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2195 (v2i64 (vector_shuffle VR128:$src1,
2196 (memopv2i64 addr:$src2),
2197 UNPCKH_shuffle_mask)))]>;
2201 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2202 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2203 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2204 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2206 let Constraints = "$src1 = $dst" in {
2207 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2208 (outs VR128:$dst), (ins VR128:$src1,
2209 GR32:$src2, i32i8imm:$src3),
2210 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2212 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2213 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2214 (outs VR128:$dst), (ins VR128:$src1,
2215 i16mem:$src2, i32i8imm:$src3),
2216 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2218 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2223 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2224 "pmovmskb\t{$src, $dst|$dst, $src}",
2225 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2227 // Conditional store
2229 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2230 "maskmovdqu\t{$mask, $src|$src, $mask}",
2231 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2234 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2235 "maskmovdqu\t{$mask, $src|$src, $mask}",
2236 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2238 // Non-temporal stores
2239 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2240 "movntpd\t{$src, $dst|$dst, $src}",
2241 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2242 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2243 "movntdq\t{$src, $dst|$dst, $src}",
2244 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2245 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2246 "movnti\t{$src, $dst|$dst, $src}",
2247 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2248 TB, Requires<[HasSSE2]>;
2251 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2252 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2253 TB, Requires<[HasSSE2]>;
2255 // Load, store, and memory fence
2256 def LFENCE : I<0xAE, MRM5r, (outs), (ins),
2257 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2258 def MFENCE : I<0xAE, MRM6r, (outs), (ins),
2259 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2261 //TODO: custom lower this so as to never even generate the noop
2262 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2264 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2265 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2266 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2269 // Alias instructions that map zero vector to pxor / xorp* for sse.
2270 // We set canFoldAsLoad because this can be converted to a constant-pool
2271 // load of an all-ones value if folding it would be beneficial.
2272 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1 in
2273 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2274 "pcmpeqd\t$dst, $dst",
2275 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2277 // FR64 to 128-bit vector conversion.
2278 let isAsCheapAsAMove = 1 in
2279 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2280 "movsd\t{$src, $dst|$dst, $src}",
2282 (v2f64 (scalar_to_vector FR64:$src)))]>;
2283 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2284 "movsd\t{$src, $dst|$dst, $src}",
2286 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2288 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2289 "movd\t{$src, $dst|$dst, $src}",
2291 (v4i32 (scalar_to_vector GR32:$src)))]>;
2292 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2293 "movd\t{$src, $dst|$dst, $src}",
2295 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2297 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2298 "movd\t{$src, $dst|$dst, $src}",
2299 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2301 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2302 "movd\t{$src, $dst|$dst, $src}",
2303 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2305 // SSE2 instructions with XS prefix
2306 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2307 "movq\t{$src, $dst|$dst, $src}",
2309 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2310 Requires<[HasSSE2]>;
2311 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2312 "movq\t{$src, $dst|$dst, $src}",
2313 [(store (i64 (vector_extract (v2i64 VR128:$src),
2314 (iPTR 0))), addr:$dst)]>;
2316 // FIXME: may not be able to eliminate this movss with coalescing the src and
2317 // dest register classes are different. We really want to write this pattern
2319 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2320 // (f32 FR32:$src)>;
2321 let isAsCheapAsAMove = 1 in
2322 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2323 "movsd\t{$src, $dst|$dst, $src}",
2324 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2326 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2327 "movsd\t{$src, $dst|$dst, $src}",
2328 [(store (f64 (vector_extract (v2f64 VR128:$src),
2329 (iPTR 0))), addr:$dst)]>;
2330 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2331 "movd\t{$src, $dst|$dst, $src}",
2332 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2334 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2335 "movd\t{$src, $dst|$dst, $src}",
2336 [(store (i32 (vector_extract (v4i32 VR128:$src),
2337 (iPTR 0))), addr:$dst)]>;
2339 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2340 "movd\t{$src, $dst|$dst, $src}",
2341 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2342 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2343 "movd\t{$src, $dst|$dst, $src}",
2344 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2347 // Move to lower bits of a VR128, leaving upper bits alone.
2348 // Three operand (but two address) aliases.
2349 let Constraints = "$src1 = $dst" in {
2350 let neverHasSideEffects = 1 in
2351 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2352 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2353 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2355 let AddedComplexity = 15 in
2356 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2357 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2358 "movsd\t{$src2, $dst|$dst, $src2}",
2360 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
2361 MOVL_shuffle_mask)))]>;
2364 // Store / copy lower 64-bits of a XMM register.
2365 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2366 "movq\t{$src, $dst|$dst, $src}",
2367 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2369 // Move to lower bits of a VR128 and zeroing upper bits.
2370 // Loading from memory automatically zeroing upper bits.
2371 let AddedComplexity = 20 in {
2372 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2373 "movsd\t{$src, $dst|$dst, $src}",
2375 (v2f64 (X86vzmovl (v2f64 (scalar_to_vector
2376 (loadf64 addr:$src))))))]>;
2378 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
2379 (MOVZSD2PDrm addr:$src)>;
2380 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
2381 (MOVZSD2PDrm addr:$src)>;
2382 def : Pat<(v2f64 (X86vzload addr:$src)), (MOVZSD2PDrm addr:$src)>;
2385 // movd / movq to XMM register zero-extends
2386 let AddedComplexity = 15 in {
2387 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2388 "movd\t{$src, $dst|$dst, $src}",
2389 [(set VR128:$dst, (v4i32 (X86vzmovl
2390 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2391 // This is X86-64 only.
2392 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2393 "mov{d|q}\t{$src, $dst|$dst, $src}",
2394 [(set VR128:$dst, (v2i64 (X86vzmovl
2395 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2398 let AddedComplexity = 20 in {
2399 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2400 "movd\t{$src, $dst|$dst, $src}",
2402 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2403 (loadi32 addr:$src))))))]>;
2405 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2406 (MOVZDI2PDIrm addr:$src)>;
2407 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2408 (MOVZDI2PDIrm addr:$src)>;
2409 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2410 (MOVZDI2PDIrm addr:$src)>;
2412 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2413 "movq\t{$src, $dst|$dst, $src}",
2415 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2416 (loadi64 addr:$src))))))]>, XS,
2417 Requires<[HasSSE2]>;
2419 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2420 (MOVZQI2PQIrm addr:$src)>;
2421 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2422 (MOVZQI2PQIrm addr:$src)>;
2423 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2426 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2427 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2428 let AddedComplexity = 15 in
2429 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2430 "movq\t{$src, $dst|$dst, $src}",
2431 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2432 XS, Requires<[HasSSE2]>;
2434 let AddedComplexity = 20 in {
2435 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2436 "movq\t{$src, $dst|$dst, $src}",
2437 [(set VR128:$dst, (v2i64 (X86vzmovl
2438 (loadv2i64 addr:$src))))]>,
2439 XS, Requires<[HasSSE2]>;
2441 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2442 (MOVZPQILo2PQIrm addr:$src)>;
2445 //===----------------------------------------------------------------------===//
2446 // SSE3 Instructions
2447 //===----------------------------------------------------------------------===//
2449 // Move Instructions
2450 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2451 "movshdup\t{$src, $dst|$dst, $src}",
2452 [(set VR128:$dst, (v4f32 (vector_shuffle
2453 VR128:$src, (undef),
2454 MOVSHDUP_shuffle_mask)))]>;
2455 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2456 "movshdup\t{$src, $dst|$dst, $src}",
2457 [(set VR128:$dst, (v4f32 (vector_shuffle
2458 (memopv4f32 addr:$src), (undef),
2459 MOVSHDUP_shuffle_mask)))]>;
2461 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2462 "movsldup\t{$src, $dst|$dst, $src}",
2463 [(set VR128:$dst, (v4f32 (vector_shuffle
2464 VR128:$src, (undef),
2465 MOVSLDUP_shuffle_mask)))]>;
2466 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2467 "movsldup\t{$src, $dst|$dst, $src}",
2468 [(set VR128:$dst, (v4f32 (vector_shuffle
2469 (memopv4f32 addr:$src), (undef),
2470 MOVSLDUP_shuffle_mask)))]>;
2472 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2473 "movddup\t{$src, $dst|$dst, $src}",
2475 (v2f64 (vector_shuffle VR128:$src, (undef),
2476 MOVDDUP_shuffle_mask)))]>;
2477 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2478 "movddup\t{$src, $dst|$dst, $src}",
2480 (v2f64 (vector_shuffle
2481 (scalar_to_vector (loadf64 addr:$src)),
2482 (undef), MOVDDUP_shuffle_mask)))]>;
2484 def : Pat<(vector_shuffle
2485 (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2486 (undef), MOVDDUP_shuffle_mask),
2487 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2488 def : Pat<(vector_shuffle
2489 (memopv2f64 addr:$src), (undef), MOVDDUP_shuffle_mask),
2490 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2494 let Constraints = "$src1 = $dst" in {
2495 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2496 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2497 "addsubps\t{$src2, $dst|$dst, $src2}",
2498 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2500 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2501 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2502 "addsubps\t{$src2, $dst|$dst, $src2}",
2503 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2504 (memop addr:$src2)))]>;
2505 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2506 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2507 "addsubpd\t{$src2, $dst|$dst, $src2}",
2508 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2510 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2511 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2512 "addsubpd\t{$src2, $dst|$dst, $src2}",
2513 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2514 (memop addr:$src2)))]>;
2517 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2518 "lddqu\t{$src, $dst|$dst, $src}",
2519 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2522 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2523 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2524 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2525 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2526 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2527 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2528 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2529 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2530 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2531 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2532 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2533 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2534 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2535 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2536 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2537 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2539 let Constraints = "$src1 = $dst" in {
2540 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2541 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2542 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2543 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2544 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2545 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2546 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2547 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2550 // Thread synchronization
2551 def MONITOR : I<0xC8, RawFrm, (outs), (ins), "monitor",
2552 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2553 def MWAIT : I<0xC9, RawFrm, (outs), (ins), "mwait",
2554 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2556 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2557 let AddedComplexity = 15 in
2558 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2559 MOVSHDUP_shuffle_mask)),
2560 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2561 let AddedComplexity = 20 in
2562 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2563 MOVSHDUP_shuffle_mask)),
2564 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2566 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2567 let AddedComplexity = 15 in
2568 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2569 MOVSLDUP_shuffle_mask)),
2570 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2571 let AddedComplexity = 20 in
2572 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2573 MOVSLDUP_shuffle_mask)),
2574 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2576 //===----------------------------------------------------------------------===//
2577 // SSSE3 Instructions
2578 //===----------------------------------------------------------------------===//
2580 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2581 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2582 Intrinsic IntId64, Intrinsic IntId128> {
2583 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2584 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2585 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2587 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2588 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2590 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2592 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2594 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2595 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2598 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2600 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2603 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2606 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2607 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2608 Intrinsic IntId64, Intrinsic IntId128> {
2609 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2611 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2612 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2614 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2616 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2619 (bitconvert (memopv4i16 addr:$src))))]>;
2621 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2623 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2624 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2627 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2629 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2632 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2635 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2636 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2637 Intrinsic IntId64, Intrinsic IntId128> {
2638 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2640 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2641 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2643 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2645 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2648 (bitconvert (memopv2i32 addr:$src))))]>;
2650 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2652 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2653 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2656 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2658 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2661 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2664 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2665 int_x86_ssse3_pabs_b,
2666 int_x86_ssse3_pabs_b_128>;
2667 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2668 int_x86_ssse3_pabs_w,
2669 int_x86_ssse3_pabs_w_128>;
2670 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2671 int_x86_ssse3_pabs_d,
2672 int_x86_ssse3_pabs_d_128>;
2674 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2675 let Constraints = "$src1 = $dst" in {
2676 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2677 Intrinsic IntId64, Intrinsic IntId128,
2678 bit Commutable = 0> {
2679 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2680 (ins VR64:$src1, VR64:$src2),
2681 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2682 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2683 let isCommutable = Commutable;
2685 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2686 (ins VR64:$src1, i64mem:$src2),
2687 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2689 (IntId64 VR64:$src1,
2690 (bitconvert (memopv8i8 addr:$src2))))]>;
2692 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2693 (ins VR128:$src1, VR128:$src2),
2694 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2695 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2697 let isCommutable = Commutable;
2699 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2700 (ins VR128:$src1, i128mem:$src2),
2701 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2703 (IntId128 VR128:$src1,
2704 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2708 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2709 let Constraints = "$src1 = $dst" in {
2710 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2711 Intrinsic IntId64, Intrinsic IntId128,
2712 bit Commutable = 0> {
2713 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2714 (ins VR64:$src1, VR64:$src2),
2715 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2716 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2717 let isCommutable = Commutable;
2719 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2720 (ins VR64:$src1, i64mem:$src2),
2721 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2723 (IntId64 VR64:$src1,
2724 (bitconvert (memopv4i16 addr:$src2))))]>;
2726 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2727 (ins VR128:$src1, VR128:$src2),
2728 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2729 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2731 let isCommutable = Commutable;
2733 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2734 (ins VR128:$src1, i128mem:$src2),
2735 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2737 (IntId128 VR128:$src1,
2738 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2742 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2743 let Constraints = "$src1 = $dst" in {
2744 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2745 Intrinsic IntId64, Intrinsic IntId128,
2746 bit Commutable = 0> {
2747 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2748 (ins VR64:$src1, VR64:$src2),
2749 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2750 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2751 let isCommutable = Commutable;
2753 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2754 (ins VR64:$src1, i64mem:$src2),
2755 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2757 (IntId64 VR64:$src1,
2758 (bitconvert (memopv2i32 addr:$src2))))]>;
2760 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2761 (ins VR128:$src1, VR128:$src2),
2762 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2763 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2765 let isCommutable = Commutable;
2767 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2768 (ins VR128:$src1, i128mem:$src2),
2769 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2771 (IntId128 VR128:$src1,
2772 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2776 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2777 int_x86_ssse3_phadd_w,
2778 int_x86_ssse3_phadd_w_128>;
2779 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2780 int_x86_ssse3_phadd_d,
2781 int_x86_ssse3_phadd_d_128>;
2782 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2783 int_x86_ssse3_phadd_sw,
2784 int_x86_ssse3_phadd_sw_128>;
2785 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2786 int_x86_ssse3_phsub_w,
2787 int_x86_ssse3_phsub_w_128>;
2788 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2789 int_x86_ssse3_phsub_d,
2790 int_x86_ssse3_phsub_d_128>;
2791 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2792 int_x86_ssse3_phsub_sw,
2793 int_x86_ssse3_phsub_sw_128>;
2794 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2795 int_x86_ssse3_pmadd_ub_sw,
2796 int_x86_ssse3_pmadd_ub_sw_128>;
2797 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2798 int_x86_ssse3_pmul_hr_sw,
2799 int_x86_ssse3_pmul_hr_sw_128, 1>;
2800 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2801 int_x86_ssse3_pshuf_b,
2802 int_x86_ssse3_pshuf_b_128>;
2803 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2804 int_x86_ssse3_psign_b,
2805 int_x86_ssse3_psign_b_128>;
2806 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2807 int_x86_ssse3_psign_w,
2808 int_x86_ssse3_psign_w_128>;
2809 defm PSIGND : SS3I_binop_rm_int_32<0x09, "psignd",
2810 int_x86_ssse3_psign_d,
2811 int_x86_ssse3_psign_d_128>;
2813 let Constraints = "$src1 = $dst" in {
2814 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2815 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2816 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2818 (int_x86_ssse3_palign_r
2819 VR64:$src1, VR64:$src2,
2821 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2822 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2823 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2825 (int_x86_ssse3_palign_r
2827 (bitconvert (memopv2i32 addr:$src2)),
2830 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2831 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2832 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2834 (int_x86_ssse3_palign_r_128
2835 VR128:$src1, VR128:$src2,
2836 imm:$src3))]>, OpSize;
2837 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2838 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2839 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2841 (int_x86_ssse3_palign_r_128
2843 (bitconvert (memopv4i32 addr:$src2)),
2844 imm:$src3))]>, OpSize;
2847 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2848 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2849 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2850 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2852 //===----------------------------------------------------------------------===//
2853 // Non-Instruction Patterns
2854 //===----------------------------------------------------------------------===//
2856 // extload f32 -> f64. This matches load+fextend because we have a hack in
2857 // the isel (PreprocessForFPConvert) that can introduce loads after dag combine.
2858 // Since these loads aren't folded into the fextend, we have to match it
2860 let Predicates = [HasSSE2] in
2861 def : Pat<(fextend (loadf32 addr:$src)),
2862 (CVTSS2SDrm addr:$src)>;
2865 let Predicates = [HasSSE2] in {
2866 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2867 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2868 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2869 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2870 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2871 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2872 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2873 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2874 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2875 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2876 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2877 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2878 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2879 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2880 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2881 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2882 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2883 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2884 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2885 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2886 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2887 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2888 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2889 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2890 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2891 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2892 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2893 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2894 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2895 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2898 // Move scalar to XMM zero-extended
2899 // movd to XMM register zero-extends
2900 let AddedComplexity = 15 in {
2901 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2902 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2903 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2904 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2905 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE1]>;
2906 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2907 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2908 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2909 (MOVLPSrr (V_SET0), VR128:$src)>, Requires<[HasSSE1]>;
2912 // Splat v2f64 / v2i64
2913 let AddedComplexity = 10 in {
2914 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2915 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2916 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2917 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2918 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2919 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2920 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2921 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2924 // Special unary SHUFPSrri case.
2925 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2926 SHUFP_unary_shuffle_mask:$sm)),
2927 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2928 Requires<[HasSSE1]>;
2929 // Special unary SHUFPDrri case.
2930 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (undef),
2931 SHUFP_unary_shuffle_mask:$sm)),
2932 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2933 Requires<[HasSSE2]>;
2934 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2935 def : Pat<(vector_shuffle (bc_v4i32 (memopv4f32 addr:$src1)), (undef),
2936 SHUFP_unary_shuffle_mask:$sm),
2937 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2938 Requires<[HasSSE2]>;
2940 // Special binary v4i32 shuffle cases with SHUFPS.
2941 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (v4i32 VR128:$src2),
2942 PSHUFD_binary_shuffle_mask:$sm)),
2943 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2944 Requires<[HasSSE2]>;
2945 def : Pat<(v4i32 (vector_shuffle VR128:$src1,
2946 (bc_v4i32 (memopv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm)),
2947 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2948 Requires<[HasSSE2]>;
2949 // Special binary v2i64 shuffle cases using SHUFPDrri.
2950 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2951 SHUFP_shuffle_mask:$sm)),
2952 (SHUFPDrri VR128:$src1, VR128:$src2, SHUFP_shuffle_mask:$sm)>,
2953 Requires<[HasSSE2]>;
2954 // Special unary SHUFPDrri case.
2955 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (undef),
2956 SHUFP_unary_shuffle_mask:$sm)),
2957 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2958 Requires<[HasSSE2]>;
2960 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2961 let AddedComplexity = 15 in {
2962 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2963 UNPCKL_v_undef_shuffle_mask:$sm)),
2964 (PSHUFDri VR128:$src, PSHUFD_shuffle_mask:$sm)>,
2965 Requires<[OptForSpeed, HasSSE2]>;
2966 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2967 UNPCKL_v_undef_shuffle_mask:$sm)),
2968 (PSHUFDri VR128:$src, PSHUFD_shuffle_mask:$sm)>,
2969 Requires<[OptForSpeed, HasSSE2]>;
2971 let AddedComplexity = 10 in {
2972 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2973 UNPCKL_v_undef_shuffle_mask)),
2974 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2975 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2976 UNPCKL_v_undef_shuffle_mask)),
2977 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2978 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2979 UNPCKL_v_undef_shuffle_mask)),
2980 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2981 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2982 UNPCKL_v_undef_shuffle_mask)),
2983 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2986 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2987 let AddedComplexity = 15 in {
2988 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2989 UNPCKH_v_undef_shuffle_mask:$sm)),
2990 (PSHUFDri VR128:$src, PSHUFD_shuffle_mask:$sm)>,
2991 Requires<[OptForSpeed, HasSSE2]>;
2992 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2993 UNPCKH_v_undef_shuffle_mask:$sm)),
2994 (PSHUFDri VR128:$src, PSHUFD_shuffle_mask:$sm)>,
2995 Requires<[OptForSpeed, HasSSE2]>;
2997 let AddedComplexity = 10 in {
2998 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2999 UNPCKH_v_undef_shuffle_mask)),
3000 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
3001 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
3002 UNPCKH_v_undef_shuffle_mask)),
3003 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3004 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
3005 UNPCKH_v_undef_shuffle_mask)),
3006 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3007 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
3008 UNPCKH_v_undef_shuffle_mask)),
3009 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3012 let AddedComplexity = 20 in {
3013 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3014 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
3015 MOVHP_shuffle_mask)),
3016 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3018 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3019 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
3020 MOVHLPS_shuffle_mask)),
3021 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3023 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3024 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
3025 MOVHLPS_v_undef_shuffle_mask)),
3026 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3027 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
3028 MOVHLPS_v_undef_shuffle_mask)),
3029 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3032 let AddedComplexity = 20 in {
3033 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3034 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
3035 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (load addr:$src2),
3036 MOVLP_shuffle_mask)),
3037 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3038 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (load addr:$src2),
3039 MOVLP_shuffle_mask)),
3040 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3041 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (load addr:$src2),
3042 MOVHP_shuffle_mask)),
3043 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3044 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (load addr:$src2),
3045 MOVHP_shuffle_mask)),
3046 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3048 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (load addr:$src2),
3049 MOVLP_shuffle_mask)),
3050 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3051 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (load addr:$src2),
3052 MOVLP_shuffle_mask)),
3053 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3054 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (load addr:$src2),
3055 MOVHP_shuffle_mask)),
3056 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
3057 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (load addr:$src2),
3058 MOVHP_shuffle_mask)),
3059 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3062 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3063 // (store (vector_shuffle (load addr), v2, <0, 1, 4, 5>), addr) using MOVHPS
3064 def : Pat<(store (v4f32 (vector_shuffle (load addr:$src1), VR128:$src2,
3065 MOVLP_shuffle_mask)), addr:$src1),
3066 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3067 def : Pat<(store (v2f64 (vector_shuffle (load addr:$src1), VR128:$src2,
3068 MOVLP_shuffle_mask)), addr:$src1),
3069 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3070 def : Pat<(store (v4f32 (vector_shuffle (load addr:$src1), VR128:$src2,
3071 MOVHP_shuffle_mask)), addr:$src1),
3072 (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3073 def : Pat<(store (v2f64 (vector_shuffle (load addr:$src1), VR128:$src2,
3074 MOVHP_shuffle_mask)), addr:$src1),
3075 (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3077 def : Pat<(store (v4i32 (vector_shuffle
3078 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2,
3079 MOVLP_shuffle_mask)), addr:$src1),
3080 (MOVLPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3081 def : Pat<(store (v2i64 (vector_shuffle (load addr:$src1), VR128:$src2,
3082 MOVLP_shuffle_mask)), addr:$src1),
3083 (MOVLPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3084 def : Pat<(store (v4i32 (vector_shuffle
3085 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2,
3086 MOVHP_shuffle_mask)), addr:$src1),
3087 (MOVHPSmr addr:$src1, VR128:$src2)>, Requires<[HasSSE1]>;
3088 def : Pat<(store (v2i64 (vector_shuffle (load addr:$src1), VR128:$src2,
3089 MOVHP_shuffle_mask)), addr:$src1),
3090 (MOVHPDmr addr:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3093 let AddedComplexity = 15 in {
3094 // Setting the lowest element in the vector.
3095 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
3096 MOVL_shuffle_mask)),
3097 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3098 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
3099 MOVL_shuffle_mask)),
3100 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3102 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
3103 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
3104 MOVLP_shuffle_mask)),
3105 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3106 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
3107 MOVLP_shuffle_mask)),
3108 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3111 // Set lowest element and zero upper elements.
3112 let AddedComplexity = 15 in
3113 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc, VR128:$src,
3114 MOVL_shuffle_mask)),
3115 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3116 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3117 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3119 // Some special case pandn patterns.
3120 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3122 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3123 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3125 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3126 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3128 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3130 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3131 (memop addr:$src2))),
3132 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3133 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3134 (memop addr:$src2))),
3135 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3136 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3137 (memop addr:$src2))),
3138 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3140 // vector -> vector casts
3141 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3142 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3143 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3144 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3145 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3146 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3147 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3148 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3150 // Use movaps / movups for SSE integer load / store (one byte shorter).
3151 def : Pat<(alignedloadv4i32 addr:$src),
3152 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
3153 def : Pat<(loadv4i32 addr:$src),
3154 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3155 def : Pat<(alignedloadv2i64 addr:$src),
3156 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3157 def : Pat<(loadv2i64 addr:$src),
3158 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3160 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3161 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3162 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3163 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3164 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3165 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3166 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3167 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3168 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3169 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3170 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3171 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3172 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3173 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3174 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3175 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3177 //===----------------------------------------------------------------------===//
3178 // SSE4.1 Instructions
3179 //===----------------------------------------------------------------------===//
3181 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3184 Intrinsic V2F64Int> {
3185 // Intrinsic operation, reg.
3186 // Vector intrinsic operation, reg
3187 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3188 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3189 !strconcat(OpcodeStr,
3190 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3191 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3194 // Vector intrinsic operation, mem
3195 def PSm_Int : SS4AIi8<opcps, MRMSrcMem,
3196 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3197 !strconcat(OpcodeStr,
3198 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3200 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3203 // Vector intrinsic operation, reg
3204 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3205 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3206 !strconcat(OpcodeStr,
3207 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3208 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3211 // Vector intrinsic operation, mem
3212 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3213 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3214 !strconcat(OpcodeStr,
3215 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3217 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3221 let Constraints = "$src1 = $dst" in {
3222 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3226 // Intrinsic operation, reg.
3227 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3229 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3230 !strconcat(OpcodeStr,
3231 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3233 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3236 // Intrinsic operation, mem.
3237 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3239 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3240 !strconcat(OpcodeStr,
3241 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3243 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3246 // Intrinsic operation, reg.
3247 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3249 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3250 !strconcat(OpcodeStr,
3251 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3253 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3256 // Intrinsic operation, mem.
3257 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3259 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3260 !strconcat(OpcodeStr,
3261 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3263 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3268 // FP round - roundss, roundps, roundsd, roundpd
3269 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3270 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3271 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3272 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3274 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3275 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3276 Intrinsic IntId128> {
3277 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3279 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3280 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3281 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3283 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3286 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3289 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3290 int_x86_sse41_phminposuw>;
3292 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3293 let Constraints = "$src1 = $dst" in {
3294 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3295 Intrinsic IntId128, bit Commutable = 0> {
3296 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3297 (ins VR128:$src1, VR128:$src2),
3298 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3299 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3301 let isCommutable = Commutable;
3303 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3304 (ins VR128:$src1, i128mem:$src2),
3305 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3307 (IntId128 VR128:$src1,
3308 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3312 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3313 int_x86_sse41_pcmpeqq, 1>;
3314 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3315 int_x86_sse41_packusdw, 0>;
3316 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3317 int_x86_sse41_pminsb, 1>;
3318 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3319 int_x86_sse41_pminsd, 1>;
3320 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3321 int_x86_sse41_pminud, 1>;
3322 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3323 int_x86_sse41_pminuw, 1>;
3324 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3325 int_x86_sse41_pmaxsb, 1>;
3326 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3327 int_x86_sse41_pmaxsd, 1>;
3328 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3329 int_x86_sse41_pmaxud, 1>;
3330 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3331 int_x86_sse41_pmaxuw, 1>;
3333 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3335 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3336 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3337 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3338 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3340 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3341 let Constraints = "$src1 = $dst" in {
3342 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3343 SDNode OpNode, Intrinsic IntId128,
3344 bit Commutable = 0> {
3345 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3346 (ins VR128:$src1, VR128:$src2),
3347 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3348 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3349 VR128:$src2))]>, OpSize {
3350 let isCommutable = Commutable;
3352 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3353 (ins VR128:$src1, VR128:$src2),
3354 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3355 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3357 let isCommutable = Commutable;
3359 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3360 (ins VR128:$src1, i128mem:$src2),
3361 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3363 (OpNode VR128:$src1, (memop addr:$src2)))]>, OpSize;
3364 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3365 (ins VR128:$src1, i128mem:$src2),
3366 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3368 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3372 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", v4i32, mul,
3373 int_x86_sse41_pmulld, 1>;
3375 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3376 let Constraints = "$src1 = $dst" in {
3377 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3378 Intrinsic IntId128, bit Commutable = 0> {
3379 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3380 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3381 !strconcat(OpcodeStr,
3382 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3384 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3386 let isCommutable = Commutable;
3388 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3389 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3390 !strconcat(OpcodeStr,
3391 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3393 (IntId128 VR128:$src1,
3394 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3399 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3400 int_x86_sse41_blendps, 0>;
3401 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3402 int_x86_sse41_blendpd, 0>;
3403 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3404 int_x86_sse41_pblendw, 0>;
3405 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3406 int_x86_sse41_dpps, 1>;
3407 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3408 int_x86_sse41_dppd, 1>;
3409 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3410 int_x86_sse41_mpsadbw, 1>;
3413 /// SS41I_ternary_int - SSE 4.1 ternary operator
3414 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3415 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3416 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3417 (ins VR128:$src1, VR128:$src2),
3418 !strconcat(OpcodeStr,
3419 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3420 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3423 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3424 (ins VR128:$src1, i128mem:$src2),
3425 !strconcat(OpcodeStr,
3426 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3429 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3433 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3434 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3435 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3438 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3439 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3440 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3441 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3443 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3444 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3446 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3450 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3451 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3452 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3453 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3454 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3455 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3457 // Common patterns involving scalar load.
3458 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3459 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3460 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3461 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3463 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3464 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3465 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3466 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3468 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3469 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3470 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3471 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3473 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3474 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3475 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3476 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3478 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3479 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3480 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3481 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3483 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3484 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3485 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3486 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3489 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3490 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3491 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3492 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3494 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3495 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3497 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3501 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3502 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3503 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3504 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3506 // Common patterns involving scalar load
3507 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3508 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3509 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3510 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3512 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3513 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3514 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3515 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3518 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3519 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3520 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3521 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3523 // Expecting a i16 load any extended to i32 value.
3524 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3525 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3526 [(set VR128:$dst, (IntId (bitconvert
3527 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3531 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3532 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovsxbq", int_x86_sse41_pmovzxbq>;
3534 // Common patterns involving scalar load
3535 def : Pat<(int_x86_sse41_pmovsxbq
3536 (bitconvert (v4i32 (X86vzmovl
3537 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3538 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3540 def : Pat<(int_x86_sse41_pmovzxbq
3541 (bitconvert (v4i32 (X86vzmovl
3542 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3543 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3546 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3547 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3548 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3549 (ins VR128:$src1, i32i8imm:$src2),
3550 !strconcat(OpcodeStr,
3551 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3552 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3554 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3555 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3556 !strconcat(OpcodeStr,
3557 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3560 // There's an AssertZext in the way of writing the store pattern
3561 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3564 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3567 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3568 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3569 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3570 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3571 !strconcat(OpcodeStr,
3572 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3575 // There's an AssertZext in the way of writing the store pattern
3576 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3579 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3582 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3583 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3584 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3585 (ins VR128:$src1, i32i8imm:$src2),
3586 !strconcat(OpcodeStr,
3587 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3589 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3590 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3591 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3592 !strconcat(OpcodeStr,
3593 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3594 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3595 addr:$dst)]>, OpSize;
3598 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3601 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3603 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3604 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3605 (ins VR128:$src1, i32i8imm:$src2),
3606 !strconcat(OpcodeStr,
3607 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3609 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3611 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3612 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3613 !strconcat(OpcodeStr,
3614 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3615 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3616 addr:$dst)]>, OpSize;
3619 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3621 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3622 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3625 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3626 Requires<[HasSSE41]>;
3628 let Constraints = "$src1 = $dst" in {
3629 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3630 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3631 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3632 !strconcat(OpcodeStr,
3633 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3635 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3636 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3637 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3638 !strconcat(OpcodeStr,
3639 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3641 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3642 imm:$src3))]>, OpSize;
3646 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3648 let Constraints = "$src1 = $dst" in {
3649 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3650 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3651 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3652 !strconcat(OpcodeStr,
3653 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3655 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3657 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3658 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3659 !strconcat(OpcodeStr,
3660 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3662 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3663 imm:$src3)))]>, OpSize;
3667 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3669 let Constraints = "$src1 = $dst" in {
3670 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3671 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3672 (ins VR128:$src1, FR32:$src2, i32i8imm:$src3),
3673 !strconcat(OpcodeStr,
3674 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3676 (X86insrtps VR128:$src1, FR32:$src2, imm:$src3))]>, OpSize;
3677 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3678 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3679 !strconcat(OpcodeStr,
3680 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3682 (X86insrtps VR128:$src1, (loadf32 addr:$src2),
3683 imm:$src3))]>, OpSize;
3687 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3689 let Defs = [EFLAGS] in {
3690 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3691 "ptest \t{$src2, $src1|$src1, $src2}", []>, OpSize;
3692 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3693 "ptest \t{$src2, $src1|$src1, $src2}", []>, OpSize;
3696 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3697 "movntdqa\t{$src, $dst|$dst, $src}",
3698 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;
3700 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3701 let Constraints = "$src1 = $dst" in {
3702 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3703 Intrinsic IntId128, bit Commutable = 0> {
3704 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3705 (ins VR128:$src1, VR128:$src2),
3706 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3707 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3709 let isCommutable = Commutable;
3711 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3712 (ins VR128:$src1, i128mem:$src2),
3713 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3715 (IntId128 VR128:$src1,
3716 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3720 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3722 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3723 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3724 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3725 (PCMPGTQrm VR128:$src1, addr:$src2)>;