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> ]>;
24 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
25 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
26 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
27 [SDNPCommutative, SDNPAssociative]>;
28 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
29 [SDNPCommutative, SDNPAssociative]>;
30 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
31 [SDNPCommutative, SDNPAssociative]>;
32 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
33 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
34 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
35 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>;
36 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>;
37 def X86pextrb : SDNode<"X86ISD::PEXTRB",
38 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
39 def X86pextrw : SDNode<"X86ISD::PEXTRW",
40 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
41 def X86pinsrb : SDNode<"X86ISD::PINSRB",
42 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
43 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
44 def X86pinsrw : SDNode<"X86ISD::PINSRW",
45 SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
46 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
47 def X86insrtps : SDNode<"X86ISD::INSERTPS",
48 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
49 SDTCisVT<2, f32>, SDTCisPtrTy<3>]>>;
51 //===----------------------------------------------------------------------===//
52 // SSE 'Special' Instructions
53 //===----------------------------------------------------------------------===//
55 let isImplicitDef = 1 in {
56 def IMPLICIT_DEF_VR128 : I<0, Pseudo, (outs VR128:$dst), (ins),
58 [(set VR128:$dst, (v4f32 (undef)))]>,
60 def IMPLICIT_DEF_FR32 : I<0, Pseudo, (outs FR32:$dst), (ins),
62 [(set FR32:$dst, (undef))]>, Requires<[HasSSE1]>;
63 def IMPLICIT_DEF_FR64 : I<0, Pseudo, (outs FR64:$dst), (ins),
65 [(set FR64:$dst, (undef))]>, Requires<[HasSSE2]>;
68 //===----------------------------------------------------------------------===//
69 // SSE Complex Patterns
70 //===----------------------------------------------------------------------===//
72 // These are 'extloads' from a scalar to the low element of a vector, zeroing
73 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
75 def sse_load_f32 : ComplexPattern<v4f32, 4, "SelectScalarSSELoad", [],
76 [SDNPHasChain, SDNPMayLoad]>;
77 def sse_load_f64 : ComplexPattern<v2f64, 4, "SelectScalarSSELoad", [],
78 [SDNPHasChain, SDNPMayLoad]>;
80 def ssmem : Operand<v4f32> {
81 let PrintMethod = "printf32mem";
82 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
84 def sdmem : Operand<v2f64> {
85 let PrintMethod = "printf64mem";
86 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
89 //===----------------------------------------------------------------------===//
90 // SSE pattern fragments
91 //===----------------------------------------------------------------------===//
93 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
94 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
95 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
96 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
98 // Like 'store', but always requires vector alignment.
99 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
100 (st node:$val, node:$ptr), [{
101 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
102 return !ST->isTruncatingStore() &&
103 ST->getAddressingMode() == ISD::UNINDEXED &&
104 ST->getAlignment() >= 16;
108 // Like 'load', but always requires vector alignment.
109 def alignedload : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
110 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
111 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
112 LD->getAddressingMode() == ISD::UNINDEXED &&
113 LD->getAlignment() >= 16;
117 def alignedloadfsf32 : PatFrag<(ops node:$ptr), (f32 (alignedload node:$ptr))>;
118 def alignedloadfsf64 : PatFrag<(ops node:$ptr), (f64 (alignedload node:$ptr))>;
119 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
120 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
121 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
122 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
124 // Like 'load', but uses special alignment checks suitable for use in
125 // memory operands in most SSE instructions, which are required to
126 // be naturally aligned on some targets but not on others.
127 // FIXME: Actually implement support for targets that don't require the
128 // alignment. This probably wants a subtarget predicate.
129 def memop : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
130 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
131 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
132 LD->getAddressingMode() == ISD::UNINDEXED &&
133 LD->getAlignment() >= 16;
137 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
138 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
139 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
140 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
141 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
142 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
143 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
145 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
147 // FIXME: 8 byte alignment for mmx reads is not required
148 def memop64 : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
149 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
150 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
151 LD->getAddressingMode() == ISD::UNINDEXED &&
152 LD->getAlignment() >= 8;
156 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
157 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
158 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
159 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
161 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
162 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
163 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
164 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
165 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
166 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
168 def fp32imm0 : PatLeaf<(f32 fpimm), [{
169 return N->isExactlyValue(+0.0);
172 def PSxLDQ_imm : SDNodeXForm<imm, [{
173 // Transformation function: imm >> 3
174 return getI32Imm(N->getValue() >> 3);
177 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
179 def SHUFFLE_get_shuf_imm : SDNodeXForm<build_vector, [{
180 return getI8Imm(X86::getShuffleSHUFImmediate(N));
183 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
185 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<build_vector, [{
186 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
189 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
191 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<build_vector, [{
192 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
195 def SSE_splat_mask : PatLeaf<(build_vector), [{
196 return X86::isSplatMask(N);
197 }], SHUFFLE_get_shuf_imm>;
199 def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
200 return X86::isSplatLoMask(N);
203 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
204 return X86::isMOVHLPSMask(N);
207 def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
208 return X86::isMOVHLPS_v_undef_Mask(N);
211 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
212 return X86::isMOVHPMask(N);
215 def MOVLP_shuffle_mask : PatLeaf<(build_vector), [{
216 return X86::isMOVLPMask(N);
219 def MOVL_shuffle_mask : PatLeaf<(build_vector), [{
220 return X86::isMOVLMask(N);
223 def MOVSHDUP_shuffle_mask : PatLeaf<(build_vector), [{
224 return X86::isMOVSHDUPMask(N);
227 def MOVSLDUP_shuffle_mask : PatLeaf<(build_vector), [{
228 return X86::isMOVSLDUPMask(N);
231 def UNPCKL_shuffle_mask : PatLeaf<(build_vector), [{
232 return X86::isUNPCKLMask(N);
235 def UNPCKH_shuffle_mask : PatLeaf<(build_vector), [{
236 return X86::isUNPCKHMask(N);
239 def UNPCKL_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
240 return X86::isUNPCKL_v_undef_Mask(N);
243 def UNPCKH_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
244 return X86::isUNPCKH_v_undef_Mask(N);
247 def PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
248 return X86::isPSHUFDMask(N);
249 }], SHUFFLE_get_shuf_imm>;
251 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
252 return X86::isPSHUFHWMask(N);
253 }], SHUFFLE_get_pshufhw_imm>;
255 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
256 return X86::isPSHUFLWMask(N);
257 }], SHUFFLE_get_pshuflw_imm>;
259 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
260 return X86::isPSHUFDMask(N);
261 }], SHUFFLE_get_shuf_imm>;
263 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
264 return X86::isSHUFPMask(N);
265 }], SHUFFLE_get_shuf_imm>;
267 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
268 return X86::isSHUFPMask(N);
269 }], SHUFFLE_get_shuf_imm>;
271 //===----------------------------------------------------------------------===//
272 // SSE scalar FP Instructions
273 //===----------------------------------------------------------------------===//
275 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
276 // scheduler into a branch sequence.
277 // These are expanded by the scheduler.
278 let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
279 def CMOV_FR32 : I<0, Pseudo,
280 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
281 "#CMOV_FR32 PSEUDO!",
282 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
284 def CMOV_FR64 : I<0, Pseudo,
285 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
286 "#CMOV_FR64 PSEUDO!",
287 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
289 def CMOV_V4F32 : I<0, Pseudo,
290 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
291 "#CMOV_V4F32 PSEUDO!",
293 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
295 def CMOV_V2F64 : I<0, Pseudo,
296 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
297 "#CMOV_V2F64 PSEUDO!",
299 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
301 def CMOV_V2I64 : I<0, Pseudo,
302 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
303 "#CMOV_V2I64 PSEUDO!",
305 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
309 //===----------------------------------------------------------------------===//
311 //===----------------------------------------------------------------------===//
314 let neverHasSideEffects = 1 in
315 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
316 "movss\t{$src, $dst|$dst, $src}", []>;
317 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
318 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
319 "movss\t{$src, $dst|$dst, $src}",
320 [(set FR32:$dst, (loadf32 addr:$src))]>;
321 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
322 "movss\t{$src, $dst|$dst, $src}",
323 [(store FR32:$src, addr:$dst)]>;
325 // Conversion instructions
326 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
327 "cvttss2si\t{$src, $dst|$dst, $src}",
328 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
329 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
330 "cvttss2si\t{$src, $dst|$dst, $src}",
331 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
332 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
333 "cvtsi2ss\t{$src, $dst|$dst, $src}",
334 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
335 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
336 "cvtsi2ss\t{$src, $dst|$dst, $src}",
337 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
339 // Match intrinsics which expect XMM operand(s).
340 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
341 "cvtss2si\t{$src, $dst|$dst, $src}",
342 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
343 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
344 "cvtss2si\t{$src, $dst|$dst, $src}",
345 [(set GR32:$dst, (int_x86_sse_cvtss2si
346 (load addr:$src)))]>;
348 // Match intrinisics which expect MM and XMM operand(s).
349 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
350 "cvtps2pi\t{$src, $dst|$dst, $src}",
351 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
352 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
353 "cvtps2pi\t{$src, $dst|$dst, $src}",
354 [(set VR64:$dst, (int_x86_sse_cvtps2pi
355 (load addr:$src)))]>;
356 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
357 "cvttps2pi\t{$src, $dst|$dst, $src}",
358 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
359 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
360 "cvttps2pi\t{$src, $dst|$dst, $src}",
361 [(set VR64:$dst, (int_x86_sse_cvttps2pi
362 (load addr:$src)))]>;
363 let Constraints = "$src1 = $dst" in {
364 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
365 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
366 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
367 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
369 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
370 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
371 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
372 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
373 (load addr:$src2)))]>;
376 // Aliases for intrinsics
377 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
378 "cvttss2si\t{$src, $dst|$dst, $src}",
380 (int_x86_sse_cvttss2si VR128:$src))]>;
381 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
382 "cvttss2si\t{$src, $dst|$dst, $src}",
384 (int_x86_sse_cvttss2si(load addr:$src)))]>;
386 let Constraints = "$src1 = $dst" in {
387 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
388 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
389 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
390 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
392 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
393 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
394 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
395 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
396 (loadi32 addr:$src2)))]>;
399 // Comparison instructions
400 let Constraints = "$src1 = $dst" in {
401 let neverHasSideEffects = 1 in
402 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
403 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
404 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
405 let neverHasSideEffects = 1, mayLoad = 1 in
406 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
407 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
408 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
411 let Defs = [EFLAGS] in {
412 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
413 "ucomiss\t{$src2, $src1|$src1, $src2}",
414 [(X86cmp FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
415 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
416 "ucomiss\t{$src2, $src1|$src1, $src2}",
417 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
421 // Aliases to match intrinsics which expect XMM operand(s).
422 let Constraints = "$src1 = $dst" in {
423 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
424 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
425 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
426 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
427 VR128:$src, imm:$cc))]>;
428 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
429 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src, SSECC:$cc),
430 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
431 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
432 (load addr:$src), imm:$cc))]>;
435 let Defs = [EFLAGS] in {
436 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs),
437 (ins VR128:$src1, VR128:$src2),
438 "ucomiss\t{$src2, $src1|$src1, $src2}",
439 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
441 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),
442 (ins VR128:$src1, f128mem:$src2),
443 "ucomiss\t{$src2, $src1|$src1, $src2}",
444 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
447 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs),
448 (ins VR128:$src1, VR128:$src2),
449 "comiss\t{$src2, $src1|$src1, $src2}",
450 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
452 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs),
453 (ins VR128:$src1, f128mem:$src2),
454 "comiss\t{$src2, $src1|$src1, $src2}",
455 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
459 // Aliases of packed SSE1 instructions for scalar use. These all have names that
462 // Alias instructions that map fld0 to pxor for sse.
463 let isReMaterializable = 1 in
464 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
465 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
466 Requires<[HasSSE1]>, TB, OpSize;
468 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
470 let neverHasSideEffects = 1 in
471 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
472 "movaps\t{$src, $dst|$dst, $src}", []>;
474 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
476 let isSimpleLoad = 1 in
477 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
478 "movaps\t{$src, $dst|$dst, $src}",
479 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
481 // Alias bitwise logical operations using SSE logical ops on packed FP values.
482 let Constraints = "$src1 = $dst" in {
483 let isCommutable = 1 in {
484 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst), (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:$dst), (ins FR32:$src1, FR32:$src2),
488 "orps\t{$src2, $dst|$dst, $src2}",
489 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
490 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
491 "xorps\t{$src2, $dst|$dst, $src2}",
492 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
495 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
496 "andps\t{$src2, $dst|$dst, $src2}",
497 [(set FR32:$dst, (X86fand FR32:$src1,
498 (memopfsf32 addr:$src2)))]>;
499 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
500 "orps\t{$src2, $dst|$dst, $src2}",
501 [(set FR32:$dst, (X86for FR32:$src1,
502 (memopfsf32 addr:$src2)))]>;
503 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
504 "xorps\t{$src2, $dst|$dst, $src2}",
505 [(set FR32:$dst, (X86fxor FR32:$src1,
506 (memopfsf32 addr:$src2)))]>;
507 let neverHasSideEffects = 1 in {
508 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
509 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
510 "andnps\t{$src2, $dst|$dst, $src2}", []>;
513 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
514 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
515 "andnps\t{$src2, $dst|$dst, $src2}", []>;
519 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
521 /// In addition, we also have a special variant of the scalar form here to
522 /// represent the associated intrinsic operation. This form is unlike the
523 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
524 /// and leaves the top elements undefined.
526 /// These three forms can each be reg+reg or reg+mem, so there are a total of
527 /// six "instructions".
529 let Constraints = "$src1 = $dst" in {
530 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
531 SDNode OpNode, Intrinsic F32Int,
532 bit Commutable = 0> {
533 // Scalar operation, reg+reg.
534 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
535 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
536 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
537 let isCommutable = Commutable;
540 // Scalar operation, reg+mem.
541 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
542 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
543 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
545 // Vector operation, reg+reg.
546 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
547 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
548 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
549 let isCommutable = Commutable;
552 // Vector operation, reg+mem.
553 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
554 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
555 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
557 // Intrinsic operation, reg+reg.
558 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
559 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
560 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
561 let isCommutable = Commutable;
564 // Intrinsic operation, reg+mem.
565 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
566 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
567 [(set VR128:$dst, (F32Int VR128:$src1,
568 sse_load_f32:$src2))]>;
572 // Arithmetic instructions
573 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
574 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
575 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
576 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
578 /// sse1_fp_binop_rm - Other SSE1 binops
580 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
581 /// instructions for a full-vector intrinsic form. Operations that map
582 /// onto C operators don't use this form since they just use the plain
583 /// vector form instead of having a separate vector intrinsic form.
585 /// This provides a total of eight "instructions".
587 let Constraints = "$src1 = $dst" in {
588 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
592 bit Commutable = 0> {
594 // Scalar operation, reg+reg.
595 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
596 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
597 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
598 let isCommutable = Commutable;
601 // Scalar operation, reg+mem.
602 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
603 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
604 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
606 // Vector operation, reg+reg.
607 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
608 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
609 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
610 let isCommutable = Commutable;
613 // Vector operation, reg+mem.
614 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
615 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
616 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
618 // Intrinsic operation, reg+reg.
619 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
620 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
621 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
622 let isCommutable = Commutable;
625 // Intrinsic operation, reg+mem.
626 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
627 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
628 [(set VR128:$dst, (F32Int VR128:$src1,
629 sse_load_f32:$src2))]>;
631 // Vector intrinsic operation, reg+reg.
632 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
633 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
634 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
635 let isCommutable = Commutable;
638 // Vector intrinsic operation, reg+mem.
639 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
640 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
641 [(set VR128:$dst, (V4F32Int VR128:$src1, (load addr:$src2)))]>;
645 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
646 int_x86_sse_max_ss, int_x86_sse_max_ps>;
647 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
648 int_x86_sse_min_ss, int_x86_sse_min_ps>;
650 //===----------------------------------------------------------------------===//
651 // SSE packed FP Instructions
654 let neverHasSideEffects = 1 in
655 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
656 "movaps\t{$src, $dst|$dst, $src}", []>;
657 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
658 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
659 "movaps\t{$src, $dst|$dst, $src}",
660 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
662 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
663 "movaps\t{$src, $dst|$dst, $src}",
664 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
666 let neverHasSideEffects = 1 in
667 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
668 "movups\t{$src, $dst|$dst, $src}", []>;
669 let isSimpleLoad = 1 in
670 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
671 "movups\t{$src, $dst|$dst, $src}",
672 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
673 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
674 "movups\t{$src, $dst|$dst, $src}",
675 [(store (v4f32 VR128:$src), addr:$dst)]>;
677 // Intrinsic forms of MOVUPS load and store
678 let isSimpleLoad = 1 in
679 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
680 "movups\t{$src, $dst|$dst, $src}",
681 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
682 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
683 "movups\t{$src, $dst|$dst, $src}",
684 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
686 let Constraints = "$src1 = $dst" in {
687 let AddedComplexity = 20 in {
688 def MOVLPSrm : PSI<0x12, MRMSrcMem,
689 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
690 "movlps\t{$src2, $dst|$dst, $src2}",
692 (v4f32 (vector_shuffle VR128:$src1,
693 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
694 MOVLP_shuffle_mask)))]>;
695 def MOVHPSrm : PSI<0x16, MRMSrcMem,
696 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
697 "movhps\t{$src2, $dst|$dst, $src2}",
699 (v4f32 (vector_shuffle VR128:$src1,
700 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
701 MOVHP_shuffle_mask)))]>;
703 } // Constraints = "$src1 = $dst"
705 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
706 "movlps\t{$src, $dst|$dst, $src}",
707 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
708 (iPTR 0))), addr:$dst)]>;
710 // v2f64 extract element 1 is always custom lowered to unpack high to low
711 // and extract element 0 so the non-store version isn't too horrible.
712 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
713 "movhps\t{$src, $dst|$dst, $src}",
714 [(store (f64 (vector_extract
715 (v2f64 (vector_shuffle
716 (bc_v2f64 (v4f32 VR128:$src)), (undef),
717 UNPCKH_shuffle_mask)), (iPTR 0))),
720 let Constraints = "$src1 = $dst" in {
721 let AddedComplexity = 15 in {
722 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
723 "movlhps\t{$src2, $dst|$dst, $src2}",
725 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
726 MOVHP_shuffle_mask)))]>;
728 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
729 "movhlps\t{$src2, $dst|$dst, $src2}",
731 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
732 MOVHLPS_shuffle_mask)))]>;
734 } // Constraints = "$src1 = $dst"
740 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
742 /// In addition, we also have a special variant of the scalar form here to
743 /// represent the associated intrinsic operation. This form is unlike the
744 /// plain scalar form, in that it takes an entire vector (instead of a
745 /// scalar) and leaves the top elements undefined.
747 /// And, we have a special variant form for a full-vector intrinsic form.
749 /// These four forms can each have a reg or a mem operand, so there are a
750 /// total of eight "instructions".
752 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
756 bit Commutable = 0> {
757 // Scalar operation, reg.
758 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
759 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
760 [(set FR32:$dst, (OpNode FR32:$src))]> {
761 let isCommutable = Commutable;
764 // Scalar operation, mem.
765 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
766 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
767 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
769 // Vector operation, reg.
770 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
771 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
772 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
773 let isCommutable = Commutable;
776 // Vector operation, mem.
777 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
778 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
779 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
781 // Intrinsic operation, reg.
782 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
783 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
784 [(set VR128:$dst, (F32Int VR128:$src))]> {
785 let isCommutable = Commutable;
788 // Intrinsic operation, mem.
789 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
790 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
791 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
793 // Vector intrinsic operation, reg
794 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
795 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
796 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
797 let isCommutable = Commutable;
800 // Vector intrinsic operation, mem
801 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
802 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
803 [(set VR128:$dst, (V4F32Int (load addr:$src)))]>;
807 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
808 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
810 // Reciprocal approximations. Note that these typically require refinement
811 // in order to obtain suitable precision.
812 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
813 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
814 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
815 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
818 let Constraints = "$src1 = $dst" in {
819 let isCommutable = 1 in {
820 def ANDPSrr : PSI<0x54, MRMSrcReg,
821 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
822 "andps\t{$src2, $dst|$dst, $src2}",
823 [(set VR128:$dst, (v2i64
824 (and VR128:$src1, VR128:$src2)))]>;
825 def ORPSrr : PSI<0x56, MRMSrcReg,
826 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
827 "orps\t{$src2, $dst|$dst, $src2}",
828 [(set VR128:$dst, (v2i64
829 (or VR128:$src1, VR128:$src2)))]>;
830 def XORPSrr : PSI<0x57, MRMSrcReg,
831 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
832 "xorps\t{$src2, $dst|$dst, $src2}",
833 [(set VR128:$dst, (v2i64
834 (xor VR128:$src1, VR128:$src2)))]>;
837 def ANDPSrm : PSI<0x54, MRMSrcMem,
838 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
839 "andps\t{$src2, $dst|$dst, $src2}",
840 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
841 (memopv2i64 addr:$src2)))]>;
842 def ORPSrm : PSI<0x56, MRMSrcMem,
843 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
844 "orps\t{$src2, $dst|$dst, $src2}",
845 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
846 (memopv2i64 addr:$src2)))]>;
847 def XORPSrm : PSI<0x57, MRMSrcMem,
848 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
849 "xorps\t{$src2, $dst|$dst, $src2}",
850 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
851 (memopv2i64 addr:$src2)))]>;
852 def ANDNPSrr : PSI<0x55, MRMSrcReg,
853 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
854 "andnps\t{$src2, $dst|$dst, $src2}",
856 (v2i64 (and (xor VR128:$src1,
857 (bc_v2i64 (v4i32 immAllOnesV))),
859 def ANDNPSrm : PSI<0x55, MRMSrcMem,
860 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
861 "andnps\t{$src2, $dst|$dst, $src2}",
863 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
864 (bc_v2i64 (v4i32 immAllOnesV))),
865 (memopv2i64 addr:$src2))))]>;
868 let Constraints = "$src1 = $dst" in {
869 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
870 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
871 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
872 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
873 VR128:$src, imm:$cc))]>;
874 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
875 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
876 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
877 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
878 (load addr:$src), imm:$cc))]>;
881 // Shuffle and unpack instructions
882 let Constraints = "$src1 = $dst" in {
883 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
884 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
885 (outs VR128:$dst), (ins VR128:$src1,
886 VR128:$src2, i32i8imm:$src3),
887 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
889 (v4f32 (vector_shuffle
890 VR128:$src1, VR128:$src2,
891 SHUFP_shuffle_mask:$src3)))]>;
892 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
893 (outs VR128:$dst), (ins VR128:$src1,
894 f128mem:$src2, i32i8imm:$src3),
895 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
897 (v4f32 (vector_shuffle
898 VR128:$src1, (memopv4f32 addr:$src2),
899 SHUFP_shuffle_mask:$src3)))]>;
901 let AddedComplexity = 10 in {
902 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
903 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
904 "unpckhps\t{$src2, $dst|$dst, $src2}",
906 (v4f32 (vector_shuffle
907 VR128:$src1, VR128:$src2,
908 UNPCKH_shuffle_mask)))]>;
909 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
910 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
911 "unpckhps\t{$src2, $dst|$dst, $src2}",
913 (v4f32 (vector_shuffle
914 VR128:$src1, (memopv4f32 addr:$src2),
915 UNPCKH_shuffle_mask)))]>;
917 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
918 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
919 "unpcklps\t{$src2, $dst|$dst, $src2}",
921 (v4f32 (vector_shuffle
922 VR128:$src1, VR128:$src2,
923 UNPCKL_shuffle_mask)))]>;
924 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
925 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
926 "unpcklps\t{$src2, $dst|$dst, $src2}",
928 (v4f32 (vector_shuffle
929 VR128:$src1, (memopv4f32 addr:$src2),
930 UNPCKL_shuffle_mask)))]>;
932 } // Constraints = "$src1 = $dst"
935 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
936 "movmskps\t{$src, $dst|$dst, $src}",
937 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
938 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
939 "movmskpd\t{$src, $dst|$dst, $src}",
940 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
942 // Prefetch intrinsic.
943 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
944 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
945 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
946 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
947 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
948 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
949 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
950 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
952 // Non-temporal stores
953 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
954 "movntps\t{$src, $dst|$dst, $src}",
955 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
957 // Load, store, and memory fence
958 def SFENCE : PSI<0xAE, MRM7m, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
961 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
962 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
963 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
964 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
966 // Alias instructions that map zero vector to pxor / xorp* for sse.
967 let isReMaterializable = 1 in
968 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
970 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
972 let Predicates = [HasSSE1] in {
973 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
974 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
975 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
976 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
977 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
980 // FR32 to 128-bit vector conversion.
981 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
982 "movss\t{$src, $dst|$dst, $src}",
984 (v4f32 (scalar_to_vector FR32:$src)))]>;
985 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
986 "movss\t{$src, $dst|$dst, $src}",
988 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
990 // FIXME: may not be able to eliminate this movss with coalescing the src and
991 // dest register classes are different. We really want to write this pattern
993 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
995 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
996 "movss\t{$src, $dst|$dst, $src}",
997 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
999 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
1000 "movss\t{$src, $dst|$dst, $src}",
1001 [(store (f32 (vector_extract (v4f32 VR128:$src),
1002 (iPTR 0))), addr:$dst)]>;
1005 // Move to lower bits of a VR128, leaving upper bits alone.
1006 // Three operand (but two address) aliases.
1007 let Constraints = "$src1 = $dst" in {
1008 let neverHasSideEffects = 1 in
1009 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
1010 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
1011 "movss\t{$src2, $dst|$dst, $src2}", []>;
1013 let AddedComplexity = 15 in
1014 def MOVLPSrr : SSI<0x10, MRMSrcReg,
1015 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1016 "movss\t{$src2, $dst|$dst, $src2}",
1018 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
1019 MOVL_shuffle_mask)))]>;
1022 // Move to lower bits of a VR128 and zeroing upper bits.
1023 // Loading from memory automatically zeroing upper bits.
1024 let AddedComplexity = 20 in
1025 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1026 "movss\t{$src, $dst|$dst, $src}",
1027 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV_bc,
1028 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
1029 MOVL_shuffle_mask)))]>;
1032 //===----------------------------------------------------------------------===//
1033 // SSE2 Instructions
1034 //===----------------------------------------------------------------------===//
1036 // Move Instructions
1037 let neverHasSideEffects = 1 in
1038 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1039 "movsd\t{$src, $dst|$dst, $src}", []>;
1040 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1041 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1042 "movsd\t{$src, $dst|$dst, $src}",
1043 [(set FR64:$dst, (loadf64 addr:$src))]>;
1044 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1045 "movsd\t{$src, $dst|$dst, $src}",
1046 [(store FR64:$src, addr:$dst)]>;
1048 // Conversion instructions
1049 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1050 "cvttsd2si\t{$src, $dst|$dst, $src}",
1051 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1052 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1053 "cvttsd2si\t{$src, $dst|$dst, $src}",
1054 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1055 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1056 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1057 [(set FR32:$dst, (fround FR64:$src))]>;
1058 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1059 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1060 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1061 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1062 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1063 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1064 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1065 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1066 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1068 // SSE2 instructions with XS prefix
1069 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1070 "cvtss2sd\t{$src, $dst|$dst, $src}",
1071 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1072 Requires<[HasSSE2]>;
1073 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1074 "cvtss2sd\t{$src, $dst|$dst, $src}",
1075 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1076 Requires<[HasSSE2]>;
1078 // Match intrinsics which expect XMM operand(s).
1079 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1080 "cvtsd2si\t{$src, $dst|$dst, $src}",
1081 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1082 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1083 "cvtsd2si\t{$src, $dst|$dst, $src}",
1084 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1085 (load addr:$src)))]>;
1087 // Match intrinisics which expect MM and XMM operand(s).
1088 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1089 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1090 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1091 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1092 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1093 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1094 (load addr:$src)))]>;
1095 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1096 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1097 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1098 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1099 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1100 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1101 (load addr:$src)))]>;
1102 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1103 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1104 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1105 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1106 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1107 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1108 (load addr:$src)))]>;
1110 // Aliases for intrinsics
1111 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1112 "cvttsd2si\t{$src, $dst|$dst, $src}",
1114 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1115 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1116 "cvttsd2si\t{$src, $dst|$dst, $src}",
1117 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1118 (load addr:$src)))]>;
1120 // Comparison instructions
1121 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1122 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1123 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1124 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1126 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1127 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1128 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1131 let Defs = [EFLAGS] in {
1132 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1133 "ucomisd\t{$src2, $src1|$src1, $src2}",
1134 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1135 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1136 "ucomisd\t{$src2, $src1|$src1, $src2}",
1137 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1138 (implicit EFLAGS)]>;
1141 // Aliases to match intrinsics which expect XMM operand(s).
1142 let Constraints = "$src1 = $dst" in {
1143 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1144 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1145 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1146 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1147 VR128:$src, imm:$cc))]>;
1148 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1149 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1150 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1151 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1152 (load addr:$src), imm:$cc))]>;
1155 let Defs = [EFLAGS] in {
1156 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1157 "ucomisd\t{$src2, $src1|$src1, $src2}",
1158 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1159 (implicit EFLAGS)]>;
1160 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1161 "ucomisd\t{$src2, $src1|$src1, $src2}",
1162 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1163 (implicit EFLAGS)]>;
1165 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1166 "comisd\t{$src2, $src1|$src1, $src2}",
1167 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1168 (implicit EFLAGS)]>;
1169 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1170 "comisd\t{$src2, $src1|$src1, $src2}",
1171 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1172 (implicit EFLAGS)]>;
1175 // Aliases of packed SSE2 instructions for scalar use. These all have names that
1178 // Alias instructions that map fld0 to pxor for sse.
1179 let isReMaterializable = 1 in
1180 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1181 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1182 Requires<[HasSSE2]>, TB, OpSize;
1184 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1186 let neverHasSideEffects = 1 in
1187 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1188 "movapd\t{$src, $dst|$dst, $src}", []>;
1190 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1192 let isSimpleLoad = 1 in
1193 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1194 "movapd\t{$src, $dst|$dst, $src}",
1195 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1197 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1198 let Constraints = "$src1 = $dst" in {
1199 let isCommutable = 1 in {
1200 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1201 "andpd\t{$src2, $dst|$dst, $src2}",
1202 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1203 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1204 "orpd\t{$src2, $dst|$dst, $src2}",
1205 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1206 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1207 "xorpd\t{$src2, $dst|$dst, $src2}",
1208 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1211 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1212 "andpd\t{$src2, $dst|$dst, $src2}",
1213 [(set FR64:$dst, (X86fand FR64:$src1,
1214 (memopfsf64 addr:$src2)))]>;
1215 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1216 "orpd\t{$src2, $dst|$dst, $src2}",
1217 [(set FR64:$dst, (X86for FR64:$src1,
1218 (memopfsf64 addr:$src2)))]>;
1219 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1220 "xorpd\t{$src2, $dst|$dst, $src2}",
1221 [(set FR64:$dst, (X86fxor FR64:$src1,
1222 (memopfsf64 addr:$src2)))]>;
1224 let neverHasSideEffects = 1 in {
1225 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1226 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1227 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1229 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1230 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1231 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1235 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1237 /// In addition, we also have a special variant of the scalar form here to
1238 /// represent the associated intrinsic operation. This form is unlike the
1239 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1240 /// and leaves the top elements undefined.
1242 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1243 /// six "instructions".
1245 let Constraints = "$src1 = $dst" in {
1246 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1247 SDNode OpNode, Intrinsic F64Int,
1248 bit Commutable = 0> {
1249 // Scalar operation, reg+reg.
1250 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1251 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1252 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1253 let isCommutable = Commutable;
1256 // Scalar operation, reg+mem.
1257 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1258 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1259 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1261 // Vector operation, reg+reg.
1262 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1263 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1264 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1265 let isCommutable = Commutable;
1268 // Vector operation, reg+mem.
1269 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1270 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1271 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1273 // Intrinsic operation, reg+reg.
1274 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1275 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1276 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1277 let isCommutable = Commutable;
1280 // Intrinsic operation, reg+mem.
1281 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1282 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1283 [(set VR128:$dst, (F64Int VR128:$src1,
1284 sse_load_f64:$src2))]>;
1288 // Arithmetic instructions
1289 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1290 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1291 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1292 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1294 /// sse2_fp_binop_rm - Other SSE2 binops
1296 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1297 /// instructions for a full-vector intrinsic form. Operations that map
1298 /// onto C operators don't use this form since they just use the plain
1299 /// vector form instead of having a separate vector intrinsic form.
1301 /// This provides a total of eight "instructions".
1303 let Constraints = "$src1 = $dst" in {
1304 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1308 bit Commutable = 0> {
1310 // Scalar operation, reg+reg.
1311 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1312 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1313 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1314 let isCommutable = Commutable;
1317 // Scalar operation, reg+mem.
1318 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1319 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1320 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1322 // Vector operation, reg+reg.
1323 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1324 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1325 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1326 let isCommutable = Commutable;
1329 // Vector operation, reg+mem.
1330 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1331 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1332 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1334 // Intrinsic operation, reg+reg.
1335 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1336 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1337 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1338 let isCommutable = Commutable;
1341 // Intrinsic operation, reg+mem.
1342 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1343 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1344 [(set VR128:$dst, (F64Int VR128:$src1,
1345 sse_load_f64:$src2))]>;
1347 // Vector intrinsic operation, reg+reg.
1348 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1349 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1350 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1351 let isCommutable = Commutable;
1354 // Vector intrinsic operation, reg+mem.
1355 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1356 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1357 [(set VR128:$dst, (V2F64Int VR128:$src1, (load addr:$src2)))]>;
1361 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1362 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1363 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1364 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1366 //===----------------------------------------------------------------------===//
1367 // SSE packed FP Instructions
1369 // Move Instructions
1370 let neverHasSideEffects = 1 in
1371 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1372 "movapd\t{$src, $dst|$dst, $src}", []>;
1373 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1374 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1375 "movapd\t{$src, $dst|$dst, $src}",
1376 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1378 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1379 "movapd\t{$src, $dst|$dst, $src}",
1380 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1382 let neverHasSideEffects = 1 in
1383 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1384 "movupd\t{$src, $dst|$dst, $src}", []>;
1385 let isSimpleLoad = 1 in
1386 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1387 "movupd\t{$src, $dst|$dst, $src}",
1388 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1389 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1390 "movupd\t{$src, $dst|$dst, $src}",
1391 [(store (v2f64 VR128:$src), addr:$dst)]>;
1393 // Intrinsic forms of MOVUPD load and store
1394 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1395 "movupd\t{$src, $dst|$dst, $src}",
1396 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1397 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1398 "movupd\t{$src, $dst|$dst, $src}",
1399 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1401 let Constraints = "$src1 = $dst" in {
1402 let AddedComplexity = 20 in {
1403 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1404 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1405 "movlpd\t{$src2, $dst|$dst, $src2}",
1407 (v2f64 (vector_shuffle VR128:$src1,
1408 (scalar_to_vector (loadf64 addr:$src2)),
1409 MOVLP_shuffle_mask)))]>;
1410 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1411 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1412 "movhpd\t{$src2, $dst|$dst, $src2}",
1414 (v2f64 (vector_shuffle VR128:$src1,
1415 (scalar_to_vector (loadf64 addr:$src2)),
1416 MOVHP_shuffle_mask)))]>;
1417 } // AddedComplexity
1418 } // Constraints = "$src1 = $dst"
1420 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1421 "movlpd\t{$src, $dst|$dst, $src}",
1422 [(store (f64 (vector_extract (v2f64 VR128:$src),
1423 (iPTR 0))), addr:$dst)]>;
1425 // v2f64 extract element 1 is always custom lowered to unpack high to low
1426 // and extract element 0 so the non-store version isn't too horrible.
1427 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1428 "movhpd\t{$src, $dst|$dst, $src}",
1429 [(store (f64 (vector_extract
1430 (v2f64 (vector_shuffle VR128:$src, (undef),
1431 UNPCKH_shuffle_mask)), (iPTR 0))),
1434 // SSE2 instructions without OpSize prefix
1435 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1436 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1437 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1438 TB, Requires<[HasSSE2]>;
1439 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1440 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1441 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1442 (bitconvert (memopv2i64 addr:$src))))]>,
1443 TB, Requires<[HasSSE2]>;
1445 // SSE2 instructions with XS prefix
1446 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1447 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1448 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1449 XS, Requires<[HasSSE2]>;
1450 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1451 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1452 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1453 (bitconvert (memopv2i64 addr:$src))))]>,
1454 XS, Requires<[HasSSE2]>;
1456 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1457 "cvtps2dq\t{$src, $dst|$dst, $src}",
1458 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1459 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1460 "cvtps2dq\t{$src, $dst|$dst, $src}",
1461 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1462 (load addr:$src)))]>;
1463 // SSE2 packed instructions with XS prefix
1464 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1465 "cvttps2dq\t{$src, $dst|$dst, $src}",
1466 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1467 XS, Requires<[HasSSE2]>;
1468 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1469 "cvttps2dq\t{$src, $dst|$dst, $src}",
1470 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1471 (load addr:$src)))]>,
1472 XS, Requires<[HasSSE2]>;
1474 // SSE2 packed instructions with XD prefix
1475 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1476 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1477 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1478 XD, Requires<[HasSSE2]>;
1479 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1480 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1481 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1482 (load addr:$src)))]>,
1483 XD, Requires<[HasSSE2]>;
1485 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1486 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1487 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1488 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1489 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1490 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1491 (load addr:$src)))]>;
1493 // SSE2 instructions without OpSize prefix
1494 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1495 "cvtps2pd\t{$src, $dst|$dst, $src}",
1496 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1497 TB, Requires<[HasSSE2]>;
1498 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f64mem:$src),
1499 "cvtps2pd\t{$src, $dst|$dst, $src}",
1500 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1501 (load addr:$src)))]>,
1502 TB, Requires<[HasSSE2]>;
1504 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1505 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1506 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1507 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f128mem:$src),
1508 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1509 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1510 (load addr:$src)))]>;
1512 // Match intrinsics which expect XMM operand(s).
1513 // Aliases for intrinsics
1514 let Constraints = "$src1 = $dst" in {
1515 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1516 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1517 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1518 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1520 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1521 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1522 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1523 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1524 (loadi32 addr:$src2)))]>;
1525 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1526 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1527 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1528 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1530 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1531 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1532 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1533 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1534 (load addr:$src2)))]>;
1535 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1536 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1537 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1538 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1539 VR128:$src2))]>, XS,
1540 Requires<[HasSSE2]>;
1541 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1542 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1543 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1544 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1545 (load addr:$src2)))]>, XS,
1546 Requires<[HasSSE2]>;
1551 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1553 /// In addition, we also have a special variant of the scalar form here to
1554 /// represent the associated intrinsic operation. This form is unlike the
1555 /// plain scalar form, in that it takes an entire vector (instead of a
1556 /// scalar) and leaves the top elements undefined.
1558 /// And, we have a special variant form for a full-vector intrinsic form.
1560 /// These four forms can each have a reg or a mem operand, so there are a
1561 /// total of eight "instructions".
1563 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1567 bit Commutable = 0> {
1568 // Scalar operation, reg.
1569 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1570 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1571 [(set FR64:$dst, (OpNode FR64:$src))]> {
1572 let isCommutable = Commutable;
1575 // Scalar operation, mem.
1576 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1577 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1578 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1580 // Vector operation, reg.
1581 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1582 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1583 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1584 let isCommutable = Commutable;
1587 // Vector operation, mem.
1588 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1589 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1590 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1592 // Intrinsic operation, reg.
1593 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1594 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1595 [(set VR128:$dst, (F64Int VR128:$src))]> {
1596 let isCommutable = Commutable;
1599 // Intrinsic operation, mem.
1600 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1601 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1602 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1604 // Vector intrinsic operation, reg
1605 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1606 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1607 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1608 let isCommutable = Commutable;
1611 // Vector intrinsic operation, mem
1612 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1613 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1614 [(set VR128:$dst, (V2F64Int (load addr:$src)))]>;
1618 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1619 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1621 // There is no f64 version of the reciprocal approximation instructions.
1624 let Constraints = "$src1 = $dst" in {
1625 let isCommutable = 1 in {
1626 def ANDPDrr : PDI<0x54, MRMSrcReg,
1627 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1628 "andpd\t{$src2, $dst|$dst, $src2}",
1630 (and (bc_v2i64 (v2f64 VR128:$src1)),
1631 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1632 def ORPDrr : PDI<0x56, MRMSrcReg,
1633 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1634 "orpd\t{$src2, $dst|$dst, $src2}",
1636 (or (bc_v2i64 (v2f64 VR128:$src1)),
1637 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1638 def XORPDrr : PDI<0x57, MRMSrcReg,
1639 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1640 "xorpd\t{$src2, $dst|$dst, $src2}",
1642 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1643 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1646 def ANDPDrm : PDI<0x54, MRMSrcMem,
1647 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1648 "andpd\t{$src2, $dst|$dst, $src2}",
1650 (and (bc_v2i64 (v2f64 VR128:$src1)),
1651 (memopv2i64 addr:$src2)))]>;
1652 def ORPDrm : PDI<0x56, MRMSrcMem,
1653 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1654 "orpd\t{$src2, $dst|$dst, $src2}",
1656 (or (bc_v2i64 (v2f64 VR128:$src1)),
1657 (memopv2i64 addr:$src2)))]>;
1658 def XORPDrm : PDI<0x57, MRMSrcMem,
1659 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1660 "xorpd\t{$src2, $dst|$dst, $src2}",
1662 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1663 (memopv2i64 addr:$src2)))]>;
1664 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1665 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1666 "andnpd\t{$src2, $dst|$dst, $src2}",
1668 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1669 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1670 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1671 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1672 "andnpd\t{$src2, $dst|$dst, $src2}",
1674 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1675 (memopv2i64 addr:$src2)))]>;
1678 let Constraints = "$src1 = $dst" in {
1679 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1680 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1681 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1682 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1683 VR128:$src, imm:$cc))]>;
1684 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1685 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1686 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1687 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1688 (load addr:$src), imm:$cc))]>;
1691 // Shuffle and unpack instructions
1692 let Constraints = "$src1 = $dst" in {
1693 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1694 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1695 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1696 [(set VR128:$dst, (v2f64 (vector_shuffle
1697 VR128:$src1, VR128:$src2,
1698 SHUFP_shuffle_mask:$src3)))]>;
1699 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1700 (outs VR128:$dst), (ins VR128:$src1,
1701 f128mem:$src2, i8imm:$src3),
1702 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1704 (v2f64 (vector_shuffle
1705 VR128:$src1, (memopv2f64 addr:$src2),
1706 SHUFP_shuffle_mask:$src3)))]>;
1708 let AddedComplexity = 10 in {
1709 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1710 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1711 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1713 (v2f64 (vector_shuffle
1714 VR128:$src1, VR128:$src2,
1715 UNPCKH_shuffle_mask)))]>;
1716 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1717 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1718 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1720 (v2f64 (vector_shuffle
1721 VR128:$src1, (memopv2f64 addr:$src2),
1722 UNPCKH_shuffle_mask)))]>;
1724 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1725 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1726 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1728 (v2f64 (vector_shuffle
1729 VR128:$src1, VR128:$src2,
1730 UNPCKL_shuffle_mask)))]>;
1731 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1732 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1733 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1735 (v2f64 (vector_shuffle
1736 VR128:$src1, (memopv2f64 addr:$src2),
1737 UNPCKL_shuffle_mask)))]>;
1738 } // AddedComplexity
1739 } // Constraints = "$src1 = $dst"
1742 //===----------------------------------------------------------------------===//
1743 // SSE integer instructions
1745 // Move Instructions
1746 let neverHasSideEffects = 1 in
1747 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1748 "movdqa\t{$src, $dst|$dst, $src}", []>;
1749 let isSimpleLoad = 1, mayLoad = 1 in
1750 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1751 "movdqa\t{$src, $dst|$dst, $src}",
1752 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1754 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1755 "movdqa\t{$src, $dst|$dst, $src}",
1756 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1757 let isSimpleLoad = 1, mayLoad = 1 in
1758 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1759 "movdqu\t{$src, $dst|$dst, $src}",
1760 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1761 XS, Requires<[HasSSE2]>;
1763 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1764 "movdqu\t{$src, $dst|$dst, $src}",
1765 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1766 XS, Requires<[HasSSE2]>;
1768 // Intrinsic forms of MOVDQU load and store
1769 let isSimpleLoad = 1 in
1770 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1771 "movdqu\t{$src, $dst|$dst, $src}",
1772 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1773 XS, Requires<[HasSSE2]>;
1774 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1775 "movdqu\t{$src, $dst|$dst, $src}",
1776 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1777 XS, Requires<[HasSSE2]>;
1779 let Constraints = "$src1 = $dst" in {
1781 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1782 bit Commutable = 0> {
1783 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1784 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1785 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1786 let isCommutable = Commutable;
1788 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1789 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1790 [(set VR128:$dst, (IntId VR128:$src1,
1791 (bitconvert (memopv2i64 addr:$src2))))]>;
1794 /// PDI_binop_rm - Simple SSE2 binary operator.
1795 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1796 ValueType OpVT, bit Commutable = 0> {
1797 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1798 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1799 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1800 let isCommutable = Commutable;
1802 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1803 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1804 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1805 (bitconvert (memopv2i64 addr:$src2)))))]>;
1808 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1810 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1811 /// to collapse (bitconvert VT to VT) into its operand.
1813 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1814 bit Commutable = 0> {
1815 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1816 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1817 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1818 let isCommutable = Commutable;
1820 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1821 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1822 [(set VR128:$dst, (OpNode VR128:$src1,(memopv2i64 addr:$src2)))]>;
1825 } // Constraints = "$src1 = $dst"
1827 // 128-bit Integer Arithmetic
1829 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1830 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1831 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1832 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1834 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1835 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1836 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1837 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1839 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1840 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1841 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1842 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1844 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1845 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1846 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1847 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1849 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1851 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1852 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1853 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1855 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1857 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1858 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1861 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1862 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1863 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1864 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1865 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1868 defm PSLLW : PDI_binop_rm_int<0xF1, "psllw", int_x86_sse2_psll_w>;
1869 defm PSLLD : PDI_binop_rm_int<0xF2, "pslld", int_x86_sse2_psll_d>;
1870 defm PSLLQ : PDI_binop_rm_int<0xF3, "psllq", int_x86_sse2_psll_q>;
1872 defm PSRLW : PDI_binop_rm_int<0xD1, "psrlw", int_x86_sse2_psrl_w>;
1873 defm PSRLD : PDI_binop_rm_int<0xD2, "psrld", int_x86_sse2_psrl_d>;
1874 defm PSRLQ : PDI_binop_rm_int<0xD3, "psrlq", int_x86_sse2_psrl_q>;
1876 defm PSRAW : PDI_binop_rm_int<0xE1, "psraw", int_x86_sse2_psra_w>;
1877 defm PSRAD : PDI_binop_rm_int<0xE2, "psrad", int_x86_sse2_psra_d>;
1879 // Some immediate variants need to match a bit_convert.
1880 let Constraints = "$src1 = $dst" in {
1881 def PSLLWri : PDIi8<0x71, MRM6r, (outs VR128:$dst),
1882 (ins VR128:$src1, i32i8imm:$src2),
1883 "psllw\t{$src2, $dst|$dst, $src2}",
1884 [(set VR128:$dst, (int_x86_sse2_psll_w VR128:$src1,
1885 (bc_v8i16 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
1886 def PSLLDri : PDIi8<0x72, MRM6r, (outs VR128:$dst),
1887 (ins VR128:$src1, i32i8imm:$src2),
1888 "pslld\t{$src2, $dst|$dst, $src2}",
1889 [(set VR128:$dst, (int_x86_sse2_psll_d VR128:$src1,
1890 (scalar_to_vector (i32 imm:$src2))))]>;
1891 def PSLLQri : PDIi8<0x73, MRM6r, (outs VR128:$dst),
1892 (ins VR128:$src1, i32i8imm:$src2),
1893 "psllq\t{$src2, $dst|$dst, $src2}",
1894 [(set VR128:$dst, (int_x86_sse2_psll_q VR128:$src1,
1895 (bc_v2i64 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
1897 def PSRLWri : PDIi8<0x71, MRM2r, (outs VR128:$dst),
1898 (ins VR128:$src1, i32i8imm:$src2),
1899 "psrlw\t{$src2, $dst|$dst, $src2}",
1900 [(set VR128:$dst, (int_x86_sse2_psrl_w VR128:$src1,
1901 (bc_v8i16 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
1902 def PSRLDri : PDIi8<0x72, MRM2r, (outs VR128:$dst),
1903 (ins VR128:$src1, i32i8imm:$src2),
1904 "psrld\t{$src2, $dst|$dst, $src2}",
1905 [(set VR128:$dst, (int_x86_sse2_psrl_d VR128:$src1,
1906 (scalar_to_vector (i32 imm:$src2))))]>;
1907 def PSRLQri : PDIi8<0x73, MRM2r, (outs VR128:$dst),
1908 (ins VR128:$src1, i32i8imm:$src2),
1909 "psrlq\t{$src2, $dst|$dst, $src2}",
1910 [(set VR128:$dst, (int_x86_sse2_psrl_q VR128:$src1,
1911 (bc_v2i64 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
1913 def PSRAWri : PDIi8<0x71, MRM4r, (outs VR128:$dst),
1914 (ins VR128:$src1, i32i8imm:$src2),
1915 "psraw\t{$src2, $dst|$dst, $src2}",
1916 [(set VR128:$dst, (int_x86_sse2_psra_w VR128:$src1,
1917 (bc_v8i16 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
1918 def PSRADri : PDIi8<0x72, MRM4r, (outs VR128:$dst),
1919 (ins VR128:$src1, i32i8imm:$src2),
1920 "psrad\t{$src2, $dst|$dst, $src2}",
1921 [(set VR128:$dst, (int_x86_sse2_psra_d VR128:$src1,
1922 (scalar_to_vector (i32 imm:$src2))))]>;
1925 // PSRAQ doesn't exist in SSE[1-3].
1927 // 128-bit logical shifts.
1928 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1929 def PSLLDQri : PDIi8<0x73, MRM7r,
1930 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1931 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1932 def PSRLDQri : PDIi8<0x73, MRM3r,
1933 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1934 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1935 // PSRADQri doesn't exist in SSE[1-3].
1938 let Predicates = [HasSSE2] in {
1939 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1940 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1941 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1942 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1943 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1944 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1948 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1949 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1950 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1952 let Constraints = "$src1 = $dst" in {
1953 def PANDNrr : PDI<0xDF, MRMSrcReg,
1954 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1955 "pandn\t{$src2, $dst|$dst, $src2}",
1956 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1959 def PANDNrm : PDI<0xDF, MRMSrcMem,
1960 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1961 "pandn\t{$src2, $dst|$dst, $src2}",
1962 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1963 (memopv2i64 addr:$src2))))]>;
1966 // SSE2 Integer comparison
1967 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1968 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1969 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1970 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1971 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1972 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1974 // Pack instructions
1975 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1976 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1977 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1979 // Shuffle and unpack instructions
1980 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1981 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1982 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1983 [(set VR128:$dst, (v4i32 (vector_shuffle
1984 VR128:$src1, (undef),
1985 PSHUFD_shuffle_mask:$src2)))]>;
1986 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1987 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1988 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1989 [(set VR128:$dst, (v4i32 (vector_shuffle
1990 (bc_v4i32(memopv2i64 addr:$src1)),
1992 PSHUFD_shuffle_mask:$src2)))]>;
1994 // SSE2 with ImmT == Imm8 and XS prefix.
1995 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1996 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1997 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1998 [(set VR128:$dst, (v8i16 (vector_shuffle
1999 VR128:$src1, (undef),
2000 PSHUFHW_shuffle_mask:$src2)))]>,
2001 XS, Requires<[HasSSE2]>;
2002 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2003 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2004 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2005 [(set VR128:$dst, (v8i16 (vector_shuffle
2006 (bc_v8i16 (memopv2i64 addr:$src1)),
2008 PSHUFHW_shuffle_mask:$src2)))]>,
2009 XS, Requires<[HasSSE2]>;
2011 // SSE2 with ImmT == Imm8 and XD prefix.
2012 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2013 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2014 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2015 [(set VR128:$dst, (v8i16 (vector_shuffle
2016 VR128:$src1, (undef),
2017 PSHUFLW_shuffle_mask:$src2)))]>,
2018 XD, Requires<[HasSSE2]>;
2019 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2020 (outs VR128:$dst), (ins i128mem:$src1, i32i8imm:$src2),
2021 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2022 [(set VR128:$dst, (v8i16 (vector_shuffle
2023 (bc_v8i16 (memopv2i64 addr:$src1)),
2025 PSHUFLW_shuffle_mask:$src2)))]>,
2026 XD, Requires<[HasSSE2]>;
2029 let Constraints = "$src1 = $dst" in {
2030 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2031 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2032 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2034 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
2035 UNPCKL_shuffle_mask)))]>;
2036 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2037 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2038 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2040 (v16i8 (vector_shuffle VR128:$src1,
2041 (bc_v16i8 (memopv2i64 addr:$src2)),
2042 UNPCKL_shuffle_mask)))]>;
2043 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2044 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2045 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2047 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2048 UNPCKL_shuffle_mask)))]>;
2049 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2050 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2051 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2053 (v8i16 (vector_shuffle VR128:$src1,
2054 (bc_v8i16 (memopv2i64 addr:$src2)),
2055 UNPCKL_shuffle_mask)))]>;
2056 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2057 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2058 "punpckldq\t{$src2, $dst|$dst, $src2}",
2060 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2061 UNPCKL_shuffle_mask)))]>;
2062 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2063 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2064 "punpckldq\t{$src2, $dst|$dst, $src2}",
2066 (v4i32 (vector_shuffle VR128:$src1,
2067 (bc_v4i32 (memopv2i64 addr:$src2)),
2068 UNPCKL_shuffle_mask)))]>;
2069 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2070 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2071 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2073 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2074 UNPCKL_shuffle_mask)))]>;
2075 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2076 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2077 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2079 (v2i64 (vector_shuffle VR128:$src1,
2080 (memopv2i64 addr:$src2),
2081 UNPCKL_shuffle_mask)))]>;
2083 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2084 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2085 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2087 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
2088 UNPCKH_shuffle_mask)))]>;
2089 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2090 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2091 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2093 (v16i8 (vector_shuffle VR128:$src1,
2094 (bc_v16i8 (memopv2i64 addr:$src2)),
2095 UNPCKH_shuffle_mask)))]>;
2096 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2097 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2098 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2100 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2101 UNPCKH_shuffle_mask)))]>;
2102 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2103 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2104 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2106 (v8i16 (vector_shuffle VR128:$src1,
2107 (bc_v8i16 (memopv2i64 addr:$src2)),
2108 UNPCKH_shuffle_mask)))]>;
2109 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2110 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2111 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2113 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2114 UNPCKH_shuffle_mask)))]>;
2115 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2116 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2117 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2119 (v4i32 (vector_shuffle VR128:$src1,
2120 (bc_v4i32 (memopv2i64 addr:$src2)),
2121 UNPCKH_shuffle_mask)))]>;
2122 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2123 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2124 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2126 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2127 UNPCKH_shuffle_mask)))]>;
2128 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2129 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2130 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2132 (v2i64 (vector_shuffle VR128:$src1,
2133 (memopv2i64 addr:$src2),
2134 UNPCKH_shuffle_mask)))]>;
2138 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2139 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2140 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2141 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2143 let Constraints = "$src1 = $dst" in {
2144 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2145 (outs VR128:$dst), (ins VR128:$src1,
2146 GR32:$src2, i32i8imm:$src3),
2147 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2149 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2150 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2151 (outs VR128:$dst), (ins VR128:$src1,
2152 i16mem:$src2, i32i8imm:$src3),
2153 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2155 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2160 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2161 "pmovmskb\t{$src, $dst|$dst, $src}",
2162 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2164 // Conditional store
2166 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2167 "maskmovdqu\t{$mask, $src|$src, $mask}",
2168 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2170 // Non-temporal stores
2171 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2172 "movntpd\t{$src, $dst|$dst, $src}",
2173 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2174 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2175 "movntdq\t{$src, $dst|$dst, $src}",
2176 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2177 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2178 "movnti\t{$src, $dst|$dst, $src}",
2179 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2180 TB, Requires<[HasSSE2]>;
2183 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2184 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2185 TB, Requires<[HasSSE2]>;
2187 // Load, store, and memory fence
2188 def LFENCE : I<0xAE, MRM5m, (outs), (ins),
2189 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2190 def MFENCE : I<0xAE, MRM6m, (outs), (ins),
2191 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2193 //TODO: custom lower this so as to never even generate the noop
2194 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2196 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2197 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2198 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2201 // Alias instructions that map zero vector to pxor / xorp* for sse.
2202 let isReMaterializable = 1 in
2203 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2204 "pcmpeqd\t$dst, $dst",
2205 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2207 // FR64 to 128-bit vector conversion.
2208 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2209 "movsd\t{$src, $dst|$dst, $src}",
2211 (v2f64 (scalar_to_vector FR64:$src)))]>;
2212 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2213 "movsd\t{$src, $dst|$dst, $src}",
2215 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2217 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2218 "movd\t{$src, $dst|$dst, $src}",
2220 (v4i32 (scalar_to_vector GR32:$src)))]>;
2221 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2222 "movd\t{$src, $dst|$dst, $src}",
2224 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2226 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2227 "movd\t{$src, $dst|$dst, $src}",
2228 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2230 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2231 "movd\t{$src, $dst|$dst, $src}",
2232 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2234 // SSE2 instructions with XS prefix
2235 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2236 "movq\t{$src, $dst|$dst, $src}",
2238 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2239 Requires<[HasSSE2]>;
2240 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2241 "movq\t{$src, $dst|$dst, $src}",
2242 [(store (i64 (vector_extract (v2i64 VR128:$src),
2243 (iPTR 0))), addr:$dst)]>;
2245 // FIXME: may not be able to eliminate this movss with coalescing the src and
2246 // dest register classes are different. We really want to write this pattern
2248 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2249 // (f32 FR32:$src)>;
2250 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2251 "movsd\t{$src, $dst|$dst, $src}",
2252 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2254 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2255 "movsd\t{$src, $dst|$dst, $src}",
2256 [(store (f64 (vector_extract (v2f64 VR128:$src),
2257 (iPTR 0))), addr:$dst)]>;
2258 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2259 "movd\t{$src, $dst|$dst, $src}",
2260 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2262 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2263 "movd\t{$src, $dst|$dst, $src}",
2264 [(store (i32 (vector_extract (v4i32 VR128:$src),
2265 (iPTR 0))), addr:$dst)]>;
2267 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2268 "movd\t{$src, $dst|$dst, $src}",
2269 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2270 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2271 "movd\t{$src, $dst|$dst, $src}",
2272 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2275 // Move to lower bits of a VR128, leaving upper bits alone.
2276 // Three operand (but two address) aliases.
2277 let Constraints = "$src1 = $dst" in {
2278 let neverHasSideEffects = 1 in
2279 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2280 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2281 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2283 let AddedComplexity = 15 in
2284 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2285 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2286 "movsd\t{$src2, $dst|$dst, $src2}",
2288 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
2289 MOVL_shuffle_mask)))]>;
2292 // Store / copy lower 64-bits of a XMM register.
2293 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2294 "movq\t{$src, $dst|$dst, $src}",
2295 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2297 // Move to lower bits of a VR128 and zeroing upper bits.
2298 // Loading from memory automatically zeroing upper bits.
2299 let AddedComplexity = 20 in
2300 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2301 "movsd\t{$src, $dst|$dst, $src}",
2303 (v2f64 (vector_shuffle immAllZerosV_bc,
2304 (v2f64 (scalar_to_vector
2305 (loadf64 addr:$src))),
2306 MOVL_shuffle_mask)))]>;
2308 // movd / movq to XMM register zero-extends
2309 let AddedComplexity = 15 in {
2310 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2311 "movd\t{$src, $dst|$dst, $src}",
2313 (v4i32 (vector_shuffle immAllZerosV,
2314 (v4i32 (scalar_to_vector GR32:$src)),
2315 MOVL_shuffle_mask)))]>;
2316 // This is X86-64 only.
2317 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2318 "mov{d|q}\t{$src, $dst|$dst, $src}",
2320 (v2i64 (vector_shuffle immAllZerosV_bc,
2321 (v2i64 (scalar_to_vector GR64:$src)),
2322 MOVL_shuffle_mask)))]>;
2325 let AddedComplexity = 20 in {
2326 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2327 "movd\t{$src, $dst|$dst, $src}",
2329 (v4i32 (vector_shuffle immAllZerosV,
2330 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
2331 MOVL_shuffle_mask)))]>;
2332 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2333 "movq\t{$src, $dst|$dst, $src}",
2335 (v2i64 (vector_shuffle immAllZerosV_bc,
2336 (v2i64 (scalar_to_vector (loadi64 addr:$src))),
2337 MOVL_shuffle_mask)))]>, XS,
2338 Requires<[HasSSE2]>;
2341 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2342 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2343 let AddedComplexity = 15 in
2344 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2345 "movq\t{$src, $dst|$dst, $src}",
2346 [(set VR128:$dst, (v2i64 (vector_shuffle immAllZerosV_bc,
2348 MOVL_shuffle_mask)))]>,
2349 XS, Requires<[HasSSE2]>;
2351 let AddedComplexity = 20 in
2352 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2353 "movq\t{$src, $dst|$dst, $src}",
2354 [(set VR128:$dst, (v2i64 (vector_shuffle immAllZerosV_bc,
2355 (memopv2i64 addr:$src),
2356 MOVL_shuffle_mask)))]>,
2357 XS, Requires<[HasSSE2]>;
2359 //===----------------------------------------------------------------------===//
2360 // SSE3 Instructions
2361 //===----------------------------------------------------------------------===//
2363 // Move Instructions
2364 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2365 "movshdup\t{$src, $dst|$dst, $src}",
2366 [(set VR128:$dst, (v4f32 (vector_shuffle
2367 VR128:$src, (undef),
2368 MOVSHDUP_shuffle_mask)))]>;
2369 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2370 "movshdup\t{$src, $dst|$dst, $src}",
2371 [(set VR128:$dst, (v4f32 (vector_shuffle
2372 (memopv4f32 addr:$src), (undef),
2373 MOVSHDUP_shuffle_mask)))]>;
2375 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2376 "movsldup\t{$src, $dst|$dst, $src}",
2377 [(set VR128:$dst, (v4f32 (vector_shuffle
2378 VR128:$src, (undef),
2379 MOVSLDUP_shuffle_mask)))]>;
2380 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2381 "movsldup\t{$src, $dst|$dst, $src}",
2382 [(set VR128:$dst, (v4f32 (vector_shuffle
2383 (memopv4f32 addr:$src), (undef),
2384 MOVSLDUP_shuffle_mask)))]>;
2386 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2387 "movddup\t{$src, $dst|$dst, $src}",
2388 [(set VR128:$dst, (v2f64 (vector_shuffle
2389 VR128:$src, (undef),
2390 SSE_splat_lo_mask)))]>;
2391 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2392 "movddup\t{$src, $dst|$dst, $src}",
2394 (v2f64 (vector_shuffle
2395 (scalar_to_vector (loadf64 addr:$src)),
2397 SSE_splat_lo_mask)))]>;
2400 let Constraints = "$src1 = $dst" in {
2401 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2402 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2403 "addsubps\t{$src2, $dst|$dst, $src2}",
2404 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2406 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2407 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2408 "addsubps\t{$src2, $dst|$dst, $src2}",
2409 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2410 (load addr:$src2)))]>;
2411 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2412 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2413 "addsubpd\t{$src2, $dst|$dst, $src2}",
2414 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2416 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2417 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2418 "addsubpd\t{$src2, $dst|$dst, $src2}",
2419 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2420 (load addr:$src2)))]>;
2423 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2424 "lddqu\t{$src, $dst|$dst, $src}",
2425 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2428 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2429 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2430 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2431 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2432 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2433 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2434 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2435 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
2436 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2437 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2438 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2439 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2440 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2441 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2442 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2443 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
2445 let Constraints = "$src1 = $dst" in {
2446 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2447 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2448 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2449 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2450 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2451 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2452 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2453 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2456 // Thread synchronization
2457 def MONITOR : I<0xC8, RawFrm, (outs), (ins), "monitor",
2458 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2459 def MWAIT : I<0xC9, RawFrm, (outs), (ins), "mwait",
2460 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2462 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2463 let AddedComplexity = 15 in
2464 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2465 MOVSHDUP_shuffle_mask)),
2466 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2467 let AddedComplexity = 20 in
2468 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2469 MOVSHDUP_shuffle_mask)),
2470 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2472 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2473 let AddedComplexity = 15 in
2474 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2475 MOVSLDUP_shuffle_mask)),
2476 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2477 let AddedComplexity = 20 in
2478 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2479 MOVSLDUP_shuffle_mask)),
2480 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2482 //===----------------------------------------------------------------------===//
2483 // SSSE3 Instructions
2484 //===----------------------------------------------------------------------===//
2486 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2487 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2488 Intrinsic IntId64, Intrinsic IntId128> {
2489 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2490 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2491 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2493 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2494 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2496 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2498 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2500 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2501 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2504 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2506 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2509 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2512 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2513 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2514 Intrinsic IntId64, Intrinsic IntId128> {
2515 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2517 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2518 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2520 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2522 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2525 (bitconvert (memopv4i16 addr:$src))))]>;
2527 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2529 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2530 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2533 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2535 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2538 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2541 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2542 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2543 Intrinsic IntId64, Intrinsic IntId128> {
2544 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2546 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2547 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2549 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2551 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2554 (bitconvert (memopv2i32 addr:$src))))]>;
2556 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2558 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2559 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2562 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2564 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2567 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2570 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2571 int_x86_ssse3_pabs_b,
2572 int_x86_ssse3_pabs_b_128>;
2573 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2574 int_x86_ssse3_pabs_w,
2575 int_x86_ssse3_pabs_w_128>;
2576 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2577 int_x86_ssse3_pabs_d,
2578 int_x86_ssse3_pabs_d_128>;
2580 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2581 let Constraints = "$src1 = $dst" in {
2582 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2583 Intrinsic IntId64, Intrinsic IntId128,
2584 bit Commutable = 0> {
2585 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2586 (ins VR64:$src1, VR64:$src2),
2587 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2588 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2589 let isCommutable = Commutable;
2591 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2592 (ins VR64:$src1, i64mem:$src2),
2593 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2595 (IntId64 VR64:$src1,
2596 (bitconvert (memopv8i8 addr:$src2))))]>;
2598 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2599 (ins VR128:$src1, VR128:$src2),
2600 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2601 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2603 let isCommutable = Commutable;
2605 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2606 (ins VR128:$src1, i128mem:$src2),
2607 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2609 (IntId128 VR128:$src1,
2610 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2614 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2615 let Constraints = "$src1 = $dst" in {
2616 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2617 Intrinsic IntId64, Intrinsic IntId128,
2618 bit Commutable = 0> {
2619 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2620 (ins VR64:$src1, VR64:$src2),
2621 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2622 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2623 let isCommutable = Commutable;
2625 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2626 (ins VR64:$src1, i64mem:$src2),
2627 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2629 (IntId64 VR64:$src1,
2630 (bitconvert (memopv4i16 addr:$src2))))]>;
2632 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2633 (ins VR128:$src1, VR128:$src2),
2634 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2635 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2637 let isCommutable = Commutable;
2639 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2640 (ins VR128:$src1, i128mem:$src2),
2641 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2643 (IntId128 VR128:$src1,
2644 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2648 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2649 let Constraints = "$src1 = $dst" in {
2650 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2651 Intrinsic IntId64, Intrinsic IntId128,
2652 bit Commutable = 0> {
2653 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2654 (ins VR64:$src1, VR64:$src2),
2655 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2656 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2657 let isCommutable = Commutable;
2659 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2660 (ins VR64:$src1, i64mem:$src2),
2661 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2663 (IntId64 VR64:$src1,
2664 (bitconvert (memopv2i32 addr:$src2))))]>;
2666 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2667 (ins VR128:$src1, VR128:$src2),
2668 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2669 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2671 let isCommutable = Commutable;
2673 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2674 (ins VR128:$src1, i128mem:$src2),
2675 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2677 (IntId128 VR128:$src1,
2678 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2682 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2683 int_x86_ssse3_phadd_w,
2684 int_x86_ssse3_phadd_w_128, 1>;
2685 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2686 int_x86_ssse3_phadd_d,
2687 int_x86_ssse3_phadd_d_128, 1>;
2688 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2689 int_x86_ssse3_phadd_sw,
2690 int_x86_ssse3_phadd_sw_128, 1>;
2691 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2692 int_x86_ssse3_phsub_w,
2693 int_x86_ssse3_phsub_w_128>;
2694 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2695 int_x86_ssse3_phsub_d,
2696 int_x86_ssse3_phsub_d_128>;
2697 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2698 int_x86_ssse3_phsub_sw,
2699 int_x86_ssse3_phsub_sw_128>;
2700 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2701 int_x86_ssse3_pmadd_ub_sw,
2702 int_x86_ssse3_pmadd_ub_sw_128, 1>;
2703 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2704 int_x86_ssse3_pmul_hr_sw,
2705 int_x86_ssse3_pmul_hr_sw_128, 1>;
2706 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2707 int_x86_ssse3_pshuf_b,
2708 int_x86_ssse3_pshuf_b_128>;
2709 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2710 int_x86_ssse3_psign_b,
2711 int_x86_ssse3_psign_b_128>;
2712 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2713 int_x86_ssse3_psign_w,
2714 int_x86_ssse3_psign_w_128>;
2715 defm PSIGND : SS3I_binop_rm_int_32<0x09, "psignd",
2716 int_x86_ssse3_psign_d,
2717 int_x86_ssse3_psign_d_128>;
2719 let Constraints = "$src1 = $dst" in {
2720 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2721 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2722 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2724 (int_x86_ssse3_palign_r
2725 VR64:$src1, VR64:$src2,
2727 def PALIGNR64rm : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2728 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2729 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2731 (int_x86_ssse3_palign_r
2733 (bitconvert (memopv2i32 addr:$src2)),
2736 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2737 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2738 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2740 (int_x86_ssse3_palign_r_128
2741 VR128:$src1, VR128:$src2,
2742 imm:$src3))]>, OpSize;
2743 def PALIGNR128rm : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2744 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2745 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2747 (int_x86_ssse3_palign_r_128
2749 (bitconvert (memopv4i32 addr:$src2)),
2750 imm:$src3))]>, OpSize;
2753 //===----------------------------------------------------------------------===//
2754 // Non-Instruction Patterns
2755 //===----------------------------------------------------------------------===//
2757 // 128-bit vector undef's.
2758 def : Pat<(v4f32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2759 def : Pat<(v2f64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2760 def : Pat<(v16i8 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2761 def : Pat<(v8i16 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2762 def : Pat<(v4i32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2763 def : Pat<(v2i64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2765 // extload f32 -> f64. This matches load+fextend because we have a hack in
2766 // the isel (PreprocessForFPConvert) that can introduce loads after dag combine.
2767 // Since these loads aren't folded into the fextend, we have to match it
2769 let Predicates = [HasSSE2] in
2770 def : Pat<(fextend (loadf32 addr:$src)),
2771 (CVTSS2SDrm addr:$src)>;
2774 let Predicates = [HasSSE2] in {
2775 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2776 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2777 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2778 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2779 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2780 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2781 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2782 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2783 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2784 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2785 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2786 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2787 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2788 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2789 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2790 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2791 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2792 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2793 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2794 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2795 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2796 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2797 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2798 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2799 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2800 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2801 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2802 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2803 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2804 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2807 // Move scalar to XMM zero-extended
2808 // movd to XMM register zero-extends
2809 let AddedComplexity = 15 in {
2810 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2811 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc,
2812 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
2813 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2814 def : Pat<(v4f32 (vector_shuffle immAllZerosV_bc,
2815 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
2816 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
2819 // Splat v2f64 / v2i64
2820 let AddedComplexity = 10 in {
2821 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2822 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2823 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2824 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2825 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2826 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2827 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2828 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2832 def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SSE_splat_mask:$sm),
2833 (SHUFPSrri VR128:$src, VR128:$src, SSE_splat_mask:$sm)>,
2834 Requires<[HasSSE1]>;
2836 // Special unary SHUFPSrri case.
2837 // FIXME: when we want non two-address code, then we should use PSHUFD?
2838 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2839 SHUFP_unary_shuffle_mask:$sm)),
2840 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2841 Requires<[HasSSE1]>;
2842 // Special unary SHUFPDrri case.
2843 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (undef),
2844 SHUFP_unary_shuffle_mask:$sm)),
2845 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2846 Requires<[HasSSE2]>;
2847 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2848 def : Pat<(vector_shuffle (memopv4f32 addr:$src1), (undef),
2849 SHUFP_unary_shuffle_mask:$sm),
2850 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2851 Requires<[HasSSE2]>;
2852 // Special binary v4i32 shuffle cases with SHUFPS.
2853 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (v4i32 VR128:$src2),
2854 PSHUFD_binary_shuffle_mask:$sm)),
2855 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2856 Requires<[HasSSE2]>;
2857 def : Pat<(v4i32 (vector_shuffle VR128:$src1,
2858 (bc_v4i32 (memopv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm)),
2859 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2860 Requires<[HasSSE2]>;
2861 // Special binary v2i64 shuffle cases using SHUFPDrri.
2862 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2863 SHUFP_shuffle_mask:$sm)),
2864 (SHUFPDrri VR128:$src1, VR128:$src2, SHUFP_shuffle_mask:$sm)>,
2865 Requires<[HasSSE2]>;
2866 // Special unary SHUFPDrri case.
2867 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (undef),
2868 SHUFP_unary_shuffle_mask:$sm)),
2869 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2870 Requires<[HasSSE2]>;
2872 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2873 let AddedComplexity = 10 in {
2874 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2875 UNPCKL_v_undef_shuffle_mask)),
2876 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2877 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2878 UNPCKL_v_undef_shuffle_mask)),
2879 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2880 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2881 UNPCKL_v_undef_shuffle_mask)),
2882 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2883 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2884 UNPCKL_v_undef_shuffle_mask)),
2885 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2888 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2889 let AddedComplexity = 10 in {
2890 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2891 UNPCKH_v_undef_shuffle_mask)),
2892 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2893 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2894 UNPCKH_v_undef_shuffle_mask)),
2895 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2896 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2897 UNPCKH_v_undef_shuffle_mask)),
2898 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2899 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2900 UNPCKH_v_undef_shuffle_mask)),
2901 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2904 let AddedComplexity = 15 in {
2905 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2906 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2907 MOVHP_shuffle_mask)),
2908 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2910 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2911 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2912 MOVHLPS_shuffle_mask)),
2913 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2915 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2916 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2917 MOVHLPS_v_undef_shuffle_mask)),
2918 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2919 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2920 MOVHLPS_v_undef_shuffle_mask)),
2921 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2924 let AddedComplexity = 20 in {
2925 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2926 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2927 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2928 MOVLP_shuffle_mask)),
2929 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2930 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2931 MOVLP_shuffle_mask)),
2932 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2933 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2934 MOVHP_shuffle_mask)),
2935 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2936 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2937 MOVHP_shuffle_mask)),
2938 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2940 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2941 MOVLP_shuffle_mask)),
2942 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2943 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2944 MOVLP_shuffle_mask)),
2945 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2946 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2947 MOVHP_shuffle_mask)),
2948 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2949 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2950 MOVLP_shuffle_mask)),
2951 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2954 let AddedComplexity = 15 in {
2955 // Setting the lowest element in the vector.
2956 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2957 MOVL_shuffle_mask)),
2958 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2959 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2960 MOVL_shuffle_mask)),
2961 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2963 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2964 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2965 MOVLP_shuffle_mask)),
2966 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2967 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2968 MOVLP_shuffle_mask)),
2969 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2972 // Set lowest element and zero upper elements.
2973 let AddedComplexity = 15 in
2974 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc, VR128:$src,
2975 MOVL_shuffle_mask)),
2976 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
2979 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2980 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2981 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2982 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2983 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2984 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2985 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2986 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2987 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2988 Requires<[HasSSE2]>;
2989 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2990 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2991 Requires<[HasSSE2]>;
2992 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2993 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2994 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2995 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2996 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2997 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2998 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2999 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
3000 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
3001 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
3002 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
3003 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
3004 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
3005 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
3006 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
3007 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3009 // Some special case pandn patterns.
3010 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3012 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3013 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3015 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3016 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3018 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3020 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3021 (memopv2i64 addr:$src2))),
3022 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3023 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3024 (memopv2i64 addr:$src2))),
3025 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3026 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3027 (memopv2i64 addr:$src2))),
3028 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3030 // vector -> vector casts
3031 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3032 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3033 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3034 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3036 // Use movaps / movups for SSE integer load / store (one byte shorter).
3037 def : Pat<(alignedloadv4i32 addr:$src),
3038 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
3039 def : Pat<(loadv4i32 addr:$src),
3040 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3041 def : Pat<(alignedloadv2i64 addr:$src),
3042 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3043 def : Pat<(loadv2i64 addr:$src),
3044 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3046 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3047 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3048 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3049 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3050 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3051 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3052 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3053 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3054 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3055 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3056 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3057 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3058 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3059 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3060 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3061 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3063 //===----------------------------------------------------------------------===//
3064 // SSE4.1 Instructions
3065 //===----------------------------------------------------------------------===//
3067 multiclass sse41_fp_unop_rm<bits<8> opcss, bits<8> opcps,
3068 bits<8> opcsd, bits<8> opcpd,
3073 Intrinsic V2F64Int> {
3074 // Intrinsic operation, reg.
3075 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3076 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3077 !strconcat(OpcodeStr,
3078 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3079 [(set VR128:$dst, (F32Int VR128:$src1, imm:$src2))]>,
3082 // Intrinsic operation, mem.
3083 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3084 (outs VR128:$dst), (ins ssmem:$src1, i32i8imm:$src2),
3085 !strconcat(OpcodeStr,
3086 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3087 [(set VR128:$dst, (F32Int sse_load_f32:$src1, imm:$src2))]>,
3090 // Vector intrinsic operation, reg
3091 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3092 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3093 !strconcat(OpcodeStr,
3094 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3095 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3098 // Vector intrinsic operation, mem
3099 def PSm_Int : SS4AIi8<opcps, MRMSrcMem,
3100 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3101 !strconcat(OpcodeStr,
3102 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3103 [(set VR128:$dst, (V4F32Int (load addr:$src1),imm:$src2))]>,
3106 // Intrinsic operation, reg.
3107 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3108 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3109 !strconcat(OpcodeStr,
3110 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3111 [(set VR128:$dst, (F64Int VR128:$src1, imm:$src2))]>,
3114 // Intrinsic operation, mem.
3115 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3116 (outs VR128:$dst), (ins sdmem:$src1, i32i8imm:$src2),
3117 !strconcat(OpcodeStr,
3118 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3119 [(set VR128:$dst, (F64Int sse_load_f64:$src1, imm:$src2))]>,
3122 // Vector intrinsic operation, reg
3123 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3124 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3125 !strconcat(OpcodeStr,
3126 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3127 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3130 // Vector intrinsic operation, mem
3131 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3132 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3133 !strconcat(OpcodeStr,
3134 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3135 [(set VR128:$dst, (V2F64Int (load addr:$src1),imm:$src2))]>,
3139 // FP round - roundss, roundps, roundsd, roundpd
3140 defm ROUND : sse41_fp_unop_rm<0x0A, 0x08, 0x0B, 0x09, "round",
3141 int_x86_sse41_round_ss, int_x86_sse41_round_ps,
3142 int_x86_sse41_round_sd, int_x86_sse41_round_pd>;
3144 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3145 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3146 Intrinsic IntId128> {
3147 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3149 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3150 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3151 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3153 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3156 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3159 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3160 int_x86_sse41_phminposuw>;
3162 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3163 let Constraints = "$src1 = $dst" in {
3164 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3165 Intrinsic IntId128, bit Commutable = 0> {
3166 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3167 (ins VR128:$src1, VR128:$src2),
3168 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3169 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3171 let isCommutable = Commutable;
3173 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3174 (ins VR128:$src1, i128mem:$src2),
3175 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3177 (IntId128 VR128:$src1,
3178 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3182 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3183 int_x86_sse41_pcmpeqq, 1>;
3184 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3185 int_x86_sse41_packusdw, 0>;
3186 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3187 int_x86_sse41_pminsb, 1>;
3188 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3189 int_x86_sse41_pminsd, 1>;
3190 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3191 int_x86_sse41_pminud, 1>;
3192 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3193 int_x86_sse41_pminuw, 1>;
3194 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3195 int_x86_sse41_pmaxsb, 1>;
3196 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3197 int_x86_sse41_pmaxsd, 1>;
3198 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3199 int_x86_sse41_pmaxud, 1>;
3200 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3201 int_x86_sse41_pmaxuw, 1>;
3202 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq",
3203 int_x86_sse41_pmuldq, 1>;
3206 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3207 let Constraints = "$src1 = $dst" in {
3208 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, SDNode OpNode,
3209 Intrinsic IntId128, bit Commutable = 0> {
3210 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3211 (ins VR128:$src1, VR128:$src2),
3212 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3213 [(set VR128:$dst, (OpNode (v4i32 VR128:$src1),
3214 VR128:$src2))]>, OpSize {
3215 let isCommutable = Commutable;
3217 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3218 (ins VR128:$src1, VR128:$src2),
3219 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3220 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3222 let isCommutable = Commutable;
3224 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3225 (ins VR128:$src1, i128mem:$src2),
3226 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3228 (OpNode VR128:$src1, (memopv4i32 addr:$src2)))]>, OpSize;
3229 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3230 (ins VR128:$src1, i128mem:$src2),
3231 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3233 (IntId128 VR128:$src1, (memopv4i32 addr:$src2)))]>,
3237 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", mul,
3238 int_x86_sse41_pmulld, 1>;
3241 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3242 let Constraints = "$src1 = $dst" in {
3243 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3244 Intrinsic IntId128, bit Commutable = 0> {
3245 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3246 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3247 !strconcat(OpcodeStr,
3248 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3250 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3252 let isCommutable = Commutable;
3254 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3255 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3256 !strconcat(OpcodeStr,
3257 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3259 (IntId128 VR128:$src1,
3260 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3265 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3266 int_x86_sse41_blendps, 0>;
3267 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3268 int_x86_sse41_blendpd, 0>;
3269 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3270 int_x86_sse41_pblendw, 0>;
3271 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3272 int_x86_sse41_dpps, 1>;
3273 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3274 int_x86_sse41_dppd, 1>;
3275 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3276 int_x86_sse41_mpsadbw, 0>;
3279 /// SS41I_ternary_int - SSE 4.1 ternary operator
3280 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3281 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3282 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3283 (ins VR128:$src1, VR128:$src2),
3284 !strconcat(OpcodeStr,
3285 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3286 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3289 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3290 (ins VR128:$src1, i128mem:$src2),
3291 !strconcat(OpcodeStr,
3292 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3295 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3299 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3300 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3301 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3304 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3305 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3306 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3307 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3309 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3310 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3312 (IntId (bitconvert (v4i32 (load addr:$src)))))]>, OpSize;
3315 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3316 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3317 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3318 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3319 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3320 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3322 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3323 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3324 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3325 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3327 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3328 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3330 (IntId (bitconvert (v4i32 (load addr:$src)))))]>, OpSize;
3333 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3334 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3335 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3336 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3338 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3339 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3340 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3341 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3343 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3344 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3346 (IntId (bitconvert (v4i32 (load addr:$src)))))]>, OpSize;
3349 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3350 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovsxbq", int_x86_sse41_pmovzxbq>;
3353 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3354 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3355 def rr : SS4AIi8<opc, MRMSrcReg, (outs GR32:$dst),
3356 (ins VR128:$src1, i32i8imm:$src2),
3357 !strconcat(OpcodeStr,
3358 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3359 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3361 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3362 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3363 !strconcat(OpcodeStr,
3364 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3367 // There's an AssertZext in the way of writing the store pattern
3368 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3371 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3374 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3375 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3376 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3377 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3378 !strconcat(OpcodeStr,
3379 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3382 // There's an AssertZext in the way of writing the store pattern
3383 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3386 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3389 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3390 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3391 def rr : SS4AIi8<opc, MRMSrcReg, (outs GR32:$dst),
3392 (ins VR128:$src1, i32i8imm:$src2),
3393 !strconcat(OpcodeStr,
3394 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3396 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3397 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3398 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3399 !strconcat(OpcodeStr,
3400 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3401 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3402 addr:$dst)]>, OpSize;
3405 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3408 /// SS41I_extractf32 - SSE 4.1 extract 32 bits to fp reg or memory destination
3409 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3410 def rr : SS4AIi8<opc, MRMSrcReg, (outs FR32:$dst),
3411 (ins VR128:$src1, i32i8imm:$src2),
3412 !strconcat(OpcodeStr,
3413 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3415 (extractelt (v4f32 VR128:$src1), imm:$src2))]>, OpSize;
3416 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3417 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3418 !strconcat(OpcodeStr,
3419 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3420 [(store (extractelt (v4f32 VR128:$src1), imm:$src2),
3421 addr:$dst)]>, OpSize;
3424 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3426 let Constraints = "$src1 = $dst" in {
3427 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3428 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3429 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3430 !strconcat(OpcodeStr,
3431 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3433 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3434 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3435 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3436 !strconcat(OpcodeStr,
3437 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3439 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3440 imm:$src3))]>, OpSize;
3444 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3446 let Constraints = "$src1 = $dst" in {
3447 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3448 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3449 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3450 !strconcat(OpcodeStr,
3451 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3453 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3455 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3456 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3457 !strconcat(OpcodeStr,
3458 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3460 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3461 imm:$src3)))]>, OpSize;
3465 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3467 let Constraints = "$src1 = $dst" in {
3468 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3469 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3470 (ins VR128:$src1, FR32:$src2, i32i8imm:$src3),
3471 !strconcat(OpcodeStr,
3472 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3474 (X86insrtps VR128:$src1, FR32:$src2, imm:$src3))]>, OpSize;
3475 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3476 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3477 !strconcat(OpcodeStr,
3478 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3480 (X86insrtps VR128:$src1, (loadf32 addr:$src2),
3481 imm:$src3))]>, OpSize;
3485 defm INSERTPS : SS41I_insertf32<0x31, "insertps">;