1 //====- X86InstrSSE.td - Describe the X86 Instruction Set -------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Evan Cheng and is distributed under the University
6 // of Illinois Open Source 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 [SDNPHasChain, SDNPOutFlag]>;
37 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest,
38 [SDNPHasChain, SDNPOutFlag]>;
39 def X86comi_new: SDNode<"X86ISD::COMI_NEW", SDTX86CmpTest,
41 def X86ucomi_new: SDNode<"X86ISD::UCOMI_NEW",SDTX86CmpTest>;
42 def X86s2vec : SDNode<"X86ISD::S2VEC", SDTypeProfile<1, 1, []>, []>;
43 def X86pextrw : SDNode<"X86ISD::PEXTRW", SDTypeProfile<1, 2, []>, []>;
44 def X86pinsrw : SDNode<"X86ISD::PINSRW", SDTypeProfile<1, 3, []>, []>;
46 //===----------------------------------------------------------------------===//
47 // SSE 'Special' Instructions
48 //===----------------------------------------------------------------------===//
50 def IMPLICIT_DEF_VR128 : I<0, Pseudo, (outs VR128:$dst), (ins),
52 [(set VR128:$dst, (v4f32 (undef)))]>,
54 def IMPLICIT_DEF_FR32 : I<0, Pseudo, (outs FR32:$dst), (ins),
56 [(set FR32:$dst, (undef))]>, Requires<[HasSSE1]>;
57 def IMPLICIT_DEF_FR64 : I<0, Pseudo, (outs FR64:$dst), (ins),
59 [(set FR64:$dst, (undef))]>, Requires<[HasSSE2]>;
61 //===----------------------------------------------------------------------===//
62 // SSE Complex Patterns
63 //===----------------------------------------------------------------------===//
65 // These are 'extloads' from a scalar to the low element of a vector, zeroing
66 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
68 def sse_load_f32 : ComplexPattern<v4f32, 4, "SelectScalarSSELoad", [],
70 def sse_load_f64 : ComplexPattern<v2f64, 4, "SelectScalarSSELoad", [],
73 def ssmem : Operand<v4f32> {
74 let PrintMethod = "printf32mem";
75 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
77 def sdmem : Operand<v2f64> {
78 let PrintMethod = "printf64mem";
79 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
82 //===----------------------------------------------------------------------===//
83 // SSE pattern fragments
84 //===----------------------------------------------------------------------===//
86 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
87 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
88 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
89 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
91 // Like 'store', but always requires vector alignment.
92 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
93 (st node:$val, node:$ptr), [{
94 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
95 return !ST->isTruncatingStore() &&
96 ST->getAddressingMode() == ISD::UNINDEXED &&
97 ST->getAlignment() >= 16;
101 // Like 'load', but always requires vector alignment.
102 def alignedload : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
103 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
104 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
105 LD->getAddressingMode() == ISD::UNINDEXED &&
106 LD->getAlignment() >= 16;
110 def alignedloadfsf32 : PatFrag<(ops node:$ptr), (f32 (alignedload node:$ptr))>;
111 def alignedloadfsf64 : PatFrag<(ops node:$ptr), (f64 (alignedload node:$ptr))>;
112 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
113 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
114 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
115 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
117 // Like 'load', but uses special alignment checks suitable for use in
118 // memory operands in most SSE instructions, which are required to
119 // be naturally aligned on some targets but not on others.
120 // FIXME: Actually implement support for targets that don't require the
121 // alignment. This probably wants a subtarget predicate.
122 def memop : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
123 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
124 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
125 LD->getAddressingMode() == ISD::UNINDEXED &&
126 LD->getAlignment() >= 16;
130 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
131 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
132 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
133 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
134 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
135 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
137 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
139 def memop64 : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
140 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
141 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
142 LD->getAddressingMode() == ISD::UNINDEXED &&
143 LD->getAlignment() >= 8;
147 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
148 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop64 node:$ptr))>;
149 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
150 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
151 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
153 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
154 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
155 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
156 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
157 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
158 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
160 def fp32imm0 : PatLeaf<(f32 fpimm), [{
161 return N->isExactlyValue(+0.0);
164 def PSxLDQ_imm : SDNodeXForm<imm, [{
165 // Transformation function: imm >> 3
166 return getI32Imm(N->getValue() >> 3);
169 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
171 def SHUFFLE_get_shuf_imm : SDNodeXForm<build_vector, [{
172 return getI8Imm(X86::getShuffleSHUFImmediate(N));
175 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
177 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<build_vector, [{
178 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
181 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
183 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<build_vector, [{
184 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
187 def SSE_splat_mask : PatLeaf<(build_vector), [{
188 return X86::isSplatMask(N);
189 }], SHUFFLE_get_shuf_imm>;
191 def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
192 return X86::isSplatLoMask(N);
195 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
196 return X86::isMOVHLPSMask(N);
199 def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
200 return X86::isMOVHLPS_v_undef_Mask(N);
203 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
204 return X86::isMOVHPMask(N);
207 def MOVLP_shuffle_mask : PatLeaf<(build_vector), [{
208 return X86::isMOVLPMask(N);
211 def MOVL_shuffle_mask : PatLeaf<(build_vector), [{
212 return X86::isMOVLMask(N);
215 def MOVSHDUP_shuffle_mask : PatLeaf<(build_vector), [{
216 return X86::isMOVSHDUPMask(N);
219 def MOVSLDUP_shuffle_mask : PatLeaf<(build_vector), [{
220 return X86::isMOVSLDUPMask(N);
223 def UNPCKL_shuffle_mask : PatLeaf<(build_vector), [{
224 return X86::isUNPCKLMask(N);
227 def UNPCKH_shuffle_mask : PatLeaf<(build_vector), [{
228 return X86::isUNPCKHMask(N);
231 def UNPCKL_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
232 return X86::isUNPCKL_v_undef_Mask(N);
235 def UNPCKH_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
236 return X86::isUNPCKH_v_undef_Mask(N);
239 def PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
240 return X86::isPSHUFDMask(N);
241 }], SHUFFLE_get_shuf_imm>;
243 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
244 return X86::isPSHUFHWMask(N);
245 }], SHUFFLE_get_pshufhw_imm>;
247 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
248 return X86::isPSHUFLWMask(N);
249 }], SHUFFLE_get_pshuflw_imm>;
251 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
252 return X86::isPSHUFDMask(N);
253 }], SHUFFLE_get_shuf_imm>;
255 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
256 return X86::isSHUFPMask(N);
257 }], SHUFFLE_get_shuf_imm>;
259 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
260 return X86::isSHUFPMask(N);
261 }], SHUFFLE_get_shuf_imm>;
263 //===----------------------------------------------------------------------===//
264 // SSE scalar FP Instructions
265 //===----------------------------------------------------------------------===//
267 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
268 // scheduler into a branch sequence.
269 // These are expanded by the scheduler.
270 let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
271 def CMOV_FR32 : I<0, Pseudo,
272 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
273 "#CMOV_FR32 PSEUDO!",
274 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond))]>;
275 def CMOV_FR64 : I<0, Pseudo,
276 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
277 "#CMOV_FR64 PSEUDO!",
278 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond))]>;
279 def CMOV_V4F32 : I<0, Pseudo,
280 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
281 "#CMOV_V4F32 PSEUDO!",
283 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
284 def CMOV_V2F64 : I<0, Pseudo,
285 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
286 "#CMOV_V2F64 PSEUDO!",
288 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
289 def CMOV_V2I64 : I<0, Pseudo,
290 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
291 "#CMOV_V2I64 PSEUDO!",
293 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
295 def NEW_CMOV_FR32 : I<0, Pseudo,
296 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
297 "#CMOV_FR32 PSEUDO!",
298 [(set FR32:$dst, (X86cmov_new FR32:$t, FR32:$f, imm:$cond,
300 def NEW_CMOV_FR64 : I<0, Pseudo,
301 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
302 "#CMOV_FR64 PSEUDO!",
303 [(set FR64:$dst, (X86cmov_new FR64:$t, FR64:$f, imm:$cond,
305 def NEW_CMOV_V4F32 : I<0, Pseudo,
306 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
307 "#CMOV_V4F32 PSEUDO!",
309 (v4f32 (X86cmov_new VR128:$t, VR128:$f, imm:$cond,
311 def NEW_CMOV_V2F64 : I<0, Pseudo,
312 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
313 "#CMOV_V2F64 PSEUDO!",
315 (v2f64 (X86cmov_new VR128:$t, VR128:$f, imm:$cond,
317 def NEW_CMOV_V2I64 : I<0, Pseudo,
318 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
319 "#CMOV_V2I64 PSEUDO!",
321 (v2i64 (X86cmov_new VR128:$t, VR128:$f, imm:$cond,
325 //===----------------------------------------------------------------------===//
327 //===----------------------------------------------------------------------===//
330 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
331 "movss\t{$src, $dst|$dst, $src}", []>;
332 let isLoad = 1, isReMaterializable = 1 in
333 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
334 "movss\t{$src, $dst|$dst, $src}",
335 [(set FR32:$dst, (loadf32 addr:$src))]>;
336 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
337 "movss\t{$src, $dst|$dst, $src}",
338 [(store FR32:$src, addr:$dst)]>;
340 // Conversion instructions
341 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
342 "cvttss2si\t{$src, $dst|$dst, $src}",
343 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
344 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
345 "cvttss2si\t{$src, $dst|$dst, $src}",
346 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
347 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
348 "cvtsi2ss\t{$src, $dst|$dst, $src}",
349 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
350 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
351 "cvtsi2ss\t{$src, $dst|$dst, $src}",
352 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
354 // Match intrinsics which expect XMM operand(s).
355 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
356 "cvtss2si\t{$src, $dst|$dst, $src}",
357 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
358 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
359 "cvtss2si\t{$src, $dst|$dst, $src}",
360 [(set GR32:$dst, (int_x86_sse_cvtss2si
361 (load addr:$src)))]>;
363 // Aliases for intrinsics
364 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
365 "cvttss2si\t{$src, $dst|$dst, $src}",
367 (int_x86_sse_cvttss2si VR128:$src))]>;
368 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
369 "cvttss2si\t{$src, $dst|$dst, $src}",
371 (int_x86_sse_cvttss2si(load addr:$src)))]>;
373 let isTwoAddress = 1 in {
374 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
375 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
376 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
377 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
379 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
380 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
381 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
382 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
383 (loadi32 addr:$src2)))]>;
386 // Comparison instructions
387 let isTwoAddress = 1 in {
388 def CMPSSrr : SSI<0xC2, MRMSrcReg,
389 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
390 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
391 def CMPSSrm : SSI<0xC2, MRMSrcMem,
392 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
393 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
396 let Defs = [EFLAGS] in {
397 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
398 "ucomiss\t{$src2, $src1|$src1, $src2}",
399 [(X86cmp FR32:$src1, FR32:$src2)]>;
400 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
401 "ucomiss\t{$src2, $src1|$src1, $src2}",
402 [(X86cmp FR32:$src1, (loadf32 addr:$src2))]>;
404 def NEW_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
405 "ucomiss\t{$src2, $src1|$src1, $src2}",
406 [(X86cmp_new FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
407 def NEW_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
408 "ucomiss\t{$src2, $src1|$src1, $src2}",
409 [(X86cmp_new FR32:$src1, (loadf32 addr:$src2)),
413 // Aliases to match intrinsics which expect XMM operand(s).
414 let isTwoAddress = 1 in {
415 def Int_CMPSSrr : SSI<0xC2, MRMSrcReg,
416 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
417 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
418 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
419 VR128:$src, imm:$cc))]>;
420 def Int_CMPSSrm : SSI<0xC2, MRMSrcMem,
421 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src, SSECC:$cc),
422 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
423 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
424 (load addr:$src), imm:$cc))]>;
427 let Defs = [EFLAGS] in {
428 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
429 "ucomiss\t{$src2, $src1|$src1, $src2}",
430 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2)]>;
431 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
432 "ucomiss\t{$src2, $src1|$src1, $src2}",
433 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2))]>;
435 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
436 "comiss\t{$src2, $src1|$src1, $src2}",
437 [(X86comi (v4f32 VR128:$src1), VR128:$src2)]>;
438 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
439 "comiss\t{$src2, $src1|$src1, $src2}",
440 [(X86comi (v4f32 VR128:$src1), (load addr:$src2))]>;
442 def NEW_Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs),
443 (ins VR128:$src1, VR128:$src2),
444 "ucomiss\t{$src2, $src1|$src1, $src2}",
445 [(X86ucomi_new (v4f32 VR128:$src1), VR128:$src2),
447 def NEW_Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),
448 (ins VR128:$src1, f128mem:$src2),
449 "ucomiss\t{$src2, $src1|$src1, $src2}",
450 [(X86ucomi_new (v4f32 VR128:$src1), (load addr:$src2)),
453 def NEW_Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs),
454 (ins VR128:$src1, VR128:$src2),
455 "comiss\t{$src2, $src1|$src1, $src2}",
456 [(X86comi_new (v4f32 VR128:$src1), VR128:$src2),
458 def NEW_Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs),
459 (ins VR128:$src1, f128mem:$src2),
460 "comiss\t{$src2, $src1|$src1, $src2}",
461 [(X86comi_new (v4f32 VR128:$src1), (load addr:$src2)),
465 // Aliases of packed SSE1 instructions for scalar use. These all have names that
468 // Alias instructions that map fld0 to pxor for sse.
469 let isReMaterializable = 1 in
470 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
471 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
472 Requires<[HasSSE1]>, TB, OpSize;
474 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
476 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
477 "movaps\t{$src, $dst|$dst, $src}", []>;
479 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
482 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
483 "movaps\t{$src, $dst|$dst, $src}",
484 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
486 // Alias bitwise logical operations using SSE logical ops on packed FP values.
487 let isTwoAddress = 1 in {
488 let isCommutable = 1 in {
489 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
490 "andps\t{$src2, $dst|$dst, $src2}",
491 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
492 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
493 "orps\t{$src2, $dst|$dst, $src2}",
494 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
495 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
496 "xorps\t{$src2, $dst|$dst, $src2}",
497 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
500 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
501 "andps\t{$src2, $dst|$dst, $src2}",
502 [(set FR32:$dst, (X86fand FR32:$src1,
503 (memopfsf32 addr:$src2)))]>;
504 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
505 "orps\t{$src2, $dst|$dst, $src2}",
506 [(set FR32:$dst, (X86for FR32:$src1,
507 (memopfsf32 addr:$src2)))]>;
508 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
509 "xorps\t{$src2, $dst|$dst, $src2}",
510 [(set FR32:$dst, (X86fxor FR32:$src1,
511 (memopfsf32 addr:$src2)))]>;
513 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
514 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
515 "andnps\t{$src2, $dst|$dst, $src2}", []>;
516 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
517 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
518 "andnps\t{$src2, $dst|$dst, $src2}", []>;
521 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
523 /// In addition, we also have a special variant of the scalar form here to
524 /// represent the associated intrinsic operation. This form is unlike the
525 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
526 /// and leaves the top elements undefined.
528 /// These three forms can each be reg+reg or reg+mem, so there are a total of
529 /// six "instructions".
531 let isTwoAddress = 1 in {
532 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
533 SDNode OpNode, Intrinsic F32Int,
534 bit Commutable = 0> {
535 // Scalar operation, reg+reg.
536 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
537 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
538 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
539 let isCommutable = Commutable;
542 // Scalar operation, reg+mem.
543 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
544 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
545 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
547 // Vector operation, reg+reg.
548 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
549 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
550 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
551 let isCommutable = Commutable;
554 // Vector operation, reg+mem.
555 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
556 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
557 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
559 // Intrinsic operation, reg+reg.
560 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
561 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
562 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
563 let isCommutable = Commutable;
566 // Intrinsic operation, reg+mem.
567 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
568 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
569 [(set VR128:$dst, (F32Int VR128:$src1,
570 sse_load_f32:$src2))]>;
574 // Arithmetic instructions
575 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
576 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
577 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
578 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
580 /// sse1_fp_binop_rm - Other SSE1 binops
582 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
583 /// instructions for a full-vector intrinsic form. Operations that map
584 /// onto C operators don't use this form since they just use the plain
585 /// vector form instead of having a separate vector intrinsic form.
587 /// This provides a total of eight "instructions".
589 let isTwoAddress = 1 in {
590 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
594 bit Commutable = 0> {
596 // Scalar operation, reg+reg.
597 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
598 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
599 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
600 let isCommutable = Commutable;
603 // Scalar operation, reg+mem.
604 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
605 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
606 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
608 // Vector operation, reg+reg.
609 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
610 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
611 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
612 let isCommutable = Commutable;
615 // Vector operation, reg+mem.
616 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
617 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
618 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
620 // Intrinsic operation, reg+reg.
621 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
622 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
623 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
624 let isCommutable = Commutable;
627 // Intrinsic operation, reg+mem.
628 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
629 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
630 [(set VR128:$dst, (F32Int VR128:$src1,
631 sse_load_f32:$src2))]>;
633 // Vector intrinsic operation, reg+reg.
634 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
635 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
636 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
637 let isCommutable = Commutable;
640 // Vector intrinsic operation, reg+mem.
641 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
642 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
643 [(set VR128:$dst, (V4F32Int VR128:$src1, (load addr:$src2)))]>;
647 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
648 int_x86_sse_max_ss, int_x86_sse_max_ps>;
649 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
650 int_x86_sse_min_ss, int_x86_sse_min_ps>;
652 //===----------------------------------------------------------------------===//
653 // SSE packed FP Instructions
656 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
657 "movaps\t{$src, $dst|$dst, $src}", []>;
658 let isLoad = 1, isReMaterializable = 1 in
659 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
660 "movaps\t{$src, $dst|$dst, $src}",
661 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
663 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
664 "movaps\t{$src, $dst|$dst, $src}",
665 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
667 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
668 "movups\t{$src, $dst|$dst, $src}", []>;
670 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$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:$dst)]>;
677 // Intrinsic forms of MOVUPS load and store
679 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$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:$src)]>;
686 let isTwoAddress = 1 in {
687 let AddedComplexity = 20 in {
688 def MOVLPSrm : PSI<0x12, MRMSrcMem,
689 (outs VR128:$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:$src2)))),
701 MOVHP_shuffle_mask)))]>;
705 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$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 isTwoAddress = 1 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:$src2,
732 MOVHLPS_shuffle_mask)))]>;
740 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
742 /// In addition, we also have a special variant of the scalar form here to
743 /// represent the associated intrinsic operation. This form is unlike the
744 /// plain scalar form, in that it takes an entire vector (instead of a
745 /// scalar) and leaves the top elements undefined.
747 /// And, we have a special variant form for a full-vector intrinsic form.
749 /// These four forms can each have a reg or a mem operand, so there are a
750 /// total of eight "instructions".
752 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
756 bit Commutable = 0> {
757 // Scalar operation, reg.
758 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$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:$src)))]>;
807 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
808 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
810 // Reciprocal approximations. Note that these typically require refinement
811 // in order to obtain suitable precision.
812 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
813 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
814 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
815 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
818 let isTwoAddress = 1 in {
819 let isCommutable = 1 in {
820 def ANDPSrr : PSI<0x54, MRMSrcReg,
821 (outs VR128:$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:$src2))))]>;
868 let isTwoAddress = 1 in {
869 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
870 (outs VR128:$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:$cc))]>;
881 // Shuffle and unpack instructions
882 let isTwoAddress = 1 in {
883 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
884 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
885 (outs VR128:$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:$src2),
930 UNPCKL_shuffle_mask)))]>;
935 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$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:$src))]>;
942 // Prefetching loads.
943 // TODO: no intrinsics for these?
944 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src), "prefetcht0\t$src", []>;
945 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src), "prefetcht1\t$src", []>;
946 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src), "prefetcht2\t$src", []>;
947 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src), "prefetchnta\t$src", []>;
949 // Non-temporal stores
950 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
951 "movntps\t{$src, $dst|$dst, $src}",
952 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
954 // Load, store, and memory fence
955 def SFENCE : PSI<0xAE, MRM7m, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
958 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
959 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
960 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
961 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
963 // Alias instructions that map zero vector to pxor / xorp* for sse.
964 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
965 let isReMaterializable = 1 in
966 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
968 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
970 // FR32 to 128-bit vector conversion.
971 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
972 "movss\t{$src, $dst|$dst, $src}",
974 (v4f32 (scalar_to_vector FR32:$src)))]>;
975 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
976 "movss\t{$src, $dst|$dst, $src}",
978 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
980 // FIXME: may not be able to eliminate this movss with coalescing the src and
981 // dest register classes are different. We really want to write this pattern
983 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
985 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
986 "movss\t{$src, $dst|$dst, $src}",
987 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
989 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
990 "movss\t{$src, $dst|$dst, $src}",
991 [(store (f32 (vector_extract (v4f32 VR128:$src),
992 (iPTR 0))), addr:$dst)]>;
995 // Move to lower bits of a VR128, leaving upper bits alone.
996 // Three operand (but two address) aliases.
997 let isTwoAddress = 1 in {
998 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
999 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
1000 "movss\t{$src2, $dst|$dst, $src2}", []>;
1002 let AddedComplexity = 15 in
1003 def MOVLPSrr : SSI<0x10, MRMSrcReg,
1004 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1005 "movss\t{$src2, $dst|$dst, $src2}",
1007 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
1008 MOVL_shuffle_mask)))]>;
1011 // Move to lower bits of a VR128 and zeroing upper bits.
1012 // Loading from memory automatically zeroing upper bits.
1013 let AddedComplexity = 20 in
1014 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1015 "movss\t{$src, $dst|$dst, $src}",
1016 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV,
1017 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
1018 MOVL_shuffle_mask)))]>;
1021 //===----------------------------------------------------------------------===//
1022 // SSE2 Instructions
1023 //===----------------------------------------------------------------------===//
1025 // Move Instructions
1026 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1027 "movsd\t{$src, $dst|$dst, $src}", []>;
1028 let isLoad = 1, isReMaterializable = 1 in
1029 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1030 "movsd\t{$src, $dst|$dst, $src}",
1031 [(set FR64:$dst, (loadf64 addr:$src))]>;
1032 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1033 "movsd\t{$src, $dst|$dst, $src}",
1034 [(store FR64:$src, addr:$dst)]>;
1036 // Conversion instructions
1037 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1038 "cvttsd2si\t{$src, $dst|$dst, $src}",
1039 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1040 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1041 "cvttsd2si\t{$src, $dst|$dst, $src}",
1042 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1043 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1044 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1045 [(set FR32:$dst, (fround FR64:$src))]>;
1046 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1047 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1048 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1049 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1050 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1051 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1052 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1053 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1054 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1056 // SSE2 instructions with XS prefix
1057 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1058 "cvtss2sd\t{$src, $dst|$dst, $src}",
1059 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1060 Requires<[HasSSE2]>;
1061 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1062 "cvtss2sd\t{$src, $dst|$dst, $src}",
1063 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1064 Requires<[HasSSE2]>;
1066 // Match intrinsics which expect XMM operand(s).
1067 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1068 "cvtsd2si\t{$src, $dst|$dst, $src}",
1069 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1070 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1071 "cvtsd2si\t{$src, $dst|$dst, $src}",
1072 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1073 (load addr:$src)))]>;
1075 // Aliases for intrinsics
1076 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1077 "cvttsd2si\t{$src, $dst|$dst, $src}",
1079 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1080 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1081 "cvttsd2si\t{$src, $dst|$dst, $src}",
1082 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1083 (load addr:$src)))]>;
1085 // Comparison instructions
1086 let isTwoAddress = 1 in {
1087 def CMPSDrr : SDI<0xC2, MRMSrcReg,
1088 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1089 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1090 def CMPSDrm : SDI<0xC2, MRMSrcMem,
1091 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1092 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1095 let Defs = [EFLAGS] in {
1096 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1097 "ucomisd\t{$src2, $src1|$src1, $src2}",
1098 [(X86cmp FR64:$src1, FR64:$src2)]>;
1099 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1100 "ucomisd\t{$src2, $src1|$src1, $src2}",
1101 [(X86cmp FR64:$src1, (loadf64 addr:$src2))]>;
1103 def NEW_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1104 "ucomisd\t{$src2, $src1|$src1, $src2}",
1105 [(X86cmp_new FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1106 def NEW_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1107 "ucomisd\t{$src2, $src1|$src1, $src2}",
1108 [(X86cmp_new FR64:$src1, (loadf64 addr:$src2)),
1109 (implicit EFLAGS)]>;
1112 // Aliases to match intrinsics which expect XMM operand(s).
1113 let isTwoAddress = 1 in {
1114 def Int_CMPSDrr : SDI<0xC2, MRMSrcReg,
1115 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1116 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1117 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1118 VR128:$src, imm:$cc))]>;
1119 def Int_CMPSDrm : SDI<0xC2, MRMSrcMem,
1120 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1121 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1122 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1123 (load addr:$src), imm:$cc))]>;
1126 let Defs = [EFLAGS] in {
1127 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1128 "ucomisd\t{$src2, $src1|$src1, $src2}",
1129 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
1130 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1131 "ucomisd\t{$src2, $src1|$src1, $src2}",
1132 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2))]>;
1134 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1135 "comisd\t{$src2, $src1|$src1, $src2}",
1136 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
1137 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1138 "comisd\t{$src2, $src1|$src1, $src2}",
1139 [(X86comi (v2f64 VR128:$src1), (load addr:$src2))]>;
1141 def NEW_Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs),
1142 (ins VR128:$src1, VR128:$src2),
1143 "ucomisd\t{$src2, $src1|$src1, $src2}",
1144 [(X86ucomi_new (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1145 (implicit EFLAGS)]>;
1146 def NEW_Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),
1147 (ins VR128:$src1, f128mem:$src2),
1148 "ucomisd\t{$src2, $src1|$src1, $src2}",
1149 [(X86ucomi_new (v2f64 VR128:$src1), (load addr:$src2)),
1150 (implicit EFLAGS)]>;
1152 def NEW_Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs),
1153 (ins VR128:$src1, VR128:$src2),
1154 "comisd\t{$src2, $src1|$src1, $src2}",
1155 [(X86comi_new (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1156 (implicit EFLAGS)]>;
1157 def NEW_Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs),
1158 (ins VR128:$src1, f128mem:$src2),
1159 "comisd\t{$src2, $src1|$src1, $src2}",
1160 [(X86comi_new (v2f64 VR128:$src1), (load addr:$src2)),
1161 (implicit EFLAGS)]>;
1164 // Aliases of packed SSE2 instructions for scalar use. These all have names that
1167 // Alias instructions that map fld0 to pxor for sse.
1168 let isReMaterializable = 1 in
1169 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1170 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1171 Requires<[HasSSE2]>, TB, OpSize;
1173 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1175 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1176 "movapd\t{$src, $dst|$dst, $src}", []>;
1178 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1181 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1182 "movapd\t{$src, $dst|$dst, $src}",
1183 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1185 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1186 let isTwoAddress = 1 in {
1187 let isCommutable = 1 in {
1188 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1189 "andpd\t{$src2, $dst|$dst, $src2}",
1190 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1191 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1192 "orpd\t{$src2, $dst|$dst, $src2}",
1193 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1194 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1195 "xorpd\t{$src2, $dst|$dst, $src2}",
1196 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1199 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1200 "andpd\t{$src2, $dst|$dst, $src2}",
1201 [(set FR64:$dst, (X86fand FR64:$src1,
1202 (memopfsf64 addr:$src2)))]>;
1203 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1204 "orpd\t{$src2, $dst|$dst, $src2}",
1205 [(set FR64:$dst, (X86for FR64:$src1,
1206 (memopfsf64 addr:$src2)))]>;
1207 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1208 "xorpd\t{$src2, $dst|$dst, $src2}",
1209 [(set FR64:$dst, (X86fxor FR64:$src1,
1210 (memopfsf64 addr:$src2)))]>;
1212 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1213 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1214 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1215 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1216 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1217 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1220 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1222 /// In addition, we also have a special variant of the scalar form here to
1223 /// represent the associated intrinsic operation. This form is unlike the
1224 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1225 /// and leaves the top elements undefined.
1227 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1228 /// six "instructions".
1230 let isTwoAddress = 1 in {
1231 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1232 SDNode OpNode, Intrinsic F64Int,
1233 bit Commutable = 0> {
1234 // Scalar operation, reg+reg.
1235 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1236 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1237 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1238 let isCommutable = Commutable;
1241 // Scalar operation, reg+mem.
1242 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1243 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1244 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1246 // Vector operation, reg+reg.
1247 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1248 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1249 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1250 let isCommutable = Commutable;
1253 // Vector operation, reg+mem.
1254 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1255 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1256 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1258 // Intrinsic operation, reg+reg.
1259 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1260 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1261 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1262 let isCommutable = Commutable;
1265 // Intrinsic operation, reg+mem.
1266 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1267 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1268 [(set VR128:$dst, (F64Int VR128:$src1,
1269 sse_load_f64:$src2))]>;
1273 // Arithmetic instructions
1274 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1275 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1276 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1277 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1279 /// sse2_fp_binop_rm - Other SSE2 binops
1281 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1282 /// instructions for a full-vector intrinsic form. Operations that map
1283 /// onto C operators don't use this form since they just use the plain
1284 /// vector form instead of having a separate vector intrinsic form.
1286 /// This provides a total of eight "instructions".
1288 let isTwoAddress = 1 in {
1289 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1293 bit Commutable = 0> {
1295 // Scalar operation, reg+reg.
1296 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1297 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1298 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1299 let isCommutable = Commutable;
1302 // Scalar operation, reg+mem.
1303 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1304 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1305 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1307 // Vector operation, reg+reg.
1308 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1309 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1310 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1311 let isCommutable = Commutable;
1314 // Vector operation, reg+mem.
1315 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1316 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1317 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1319 // Intrinsic operation, reg+reg.
1320 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1321 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1322 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1323 let isCommutable = Commutable;
1326 // Intrinsic operation, reg+mem.
1327 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1328 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1329 [(set VR128:$dst, (F64Int VR128:$src1,
1330 sse_load_f64:$src2))]>;
1332 // Vector intrinsic operation, reg+reg.
1333 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1334 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1335 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1336 let isCommutable = Commutable;
1339 // Vector intrinsic operation, reg+mem.
1340 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1341 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1342 [(set VR128:$dst, (V2F64Int VR128:$src1, (load addr:$src2)))]>;
1346 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1347 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1348 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1349 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1351 //===----------------------------------------------------------------------===//
1352 // SSE packed FP Instructions
1354 // Move Instructions
1355 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1356 "movapd\t{$src, $dst|$dst, $src}", []>;
1357 let isLoad = 1, isReMaterializable = 1 in
1358 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1359 "movapd\t{$src, $dst|$dst, $src}",
1360 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1362 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1363 "movapd\t{$src, $dst|$dst, $src}",
1364 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1366 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1367 "movupd\t{$src, $dst|$dst, $src}", []>;
1369 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1370 "movupd\t{$src, $dst|$dst, $src}",
1371 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1372 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1373 "movupd\t{$src, $dst|$dst, $src}",
1374 [(store (v2f64 VR128:$src), addr:$dst)]>;
1376 // Intrinsic forms of MOVUPD load and store
1377 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1378 "movupd\t{$src, $dst|$dst, $src}",
1379 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1380 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1381 "movupd\t{$src, $dst|$dst, $src}",
1382 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1384 let isTwoAddress = 1 in {
1385 let AddedComplexity = 20 in {
1386 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1387 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1388 "movlpd\t{$src2, $dst|$dst, $src2}",
1390 (v2f64 (vector_shuffle VR128:$src1,
1391 (scalar_to_vector (loadf64 addr:$src2)),
1392 MOVLP_shuffle_mask)))]>;
1393 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1394 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1395 "movhpd\t{$src2, $dst|$dst, $src2}",
1397 (v2f64 (vector_shuffle VR128:$src1,
1398 (scalar_to_vector (loadf64 addr:$src2)),
1399 MOVHP_shuffle_mask)))]>;
1400 } // AddedComplexity
1403 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1404 "movlpd\t{$src, $dst|$dst, $src}",
1405 [(store (f64 (vector_extract (v2f64 VR128:$src),
1406 (iPTR 0))), addr:$dst)]>;
1408 // v2f64 extract element 1 is always custom lowered to unpack high to low
1409 // and extract element 0 so the non-store version isn't too horrible.
1410 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1411 "movhpd\t{$src, $dst|$dst, $src}",
1412 [(store (f64 (vector_extract
1413 (v2f64 (vector_shuffle VR128:$src, (undef),
1414 UNPCKH_shuffle_mask)), (iPTR 0))),
1417 // SSE2 instructions without OpSize prefix
1418 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1419 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1420 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1421 TB, Requires<[HasSSE2]>;
1422 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1423 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1424 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1425 (bitconvert (memopv2i64 addr:$src))))]>,
1426 TB, Requires<[HasSSE2]>;
1428 // SSE2 instructions with XS prefix
1429 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1430 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1431 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1432 XS, Requires<[HasSSE2]>;
1433 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1434 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1435 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1436 (bitconvert (memopv2i64 addr:$src))))]>,
1437 XS, Requires<[HasSSE2]>;
1439 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1440 "cvtps2dq\t{$src, $dst|$dst, $src}",
1441 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1442 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1443 "cvtps2dq\t{$src, $dst|$dst, $src}",
1444 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1445 (load addr:$src)))]>;
1446 // SSE2 packed instructions with XS prefix
1447 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1448 "cvttps2dq\t{$src, $dst|$dst, $src}",
1449 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1450 XS, Requires<[HasSSE2]>;
1451 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1452 "cvttps2dq\t{$src, $dst|$dst, $src}",
1453 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1454 (load addr:$src)))]>,
1455 XS, Requires<[HasSSE2]>;
1457 // SSE2 packed instructions with XD prefix
1458 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1459 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1460 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1461 XD, Requires<[HasSSE2]>;
1462 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1463 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1464 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1465 (load addr:$src)))]>,
1466 XD, Requires<[HasSSE2]>;
1468 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1469 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1470 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1471 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1472 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1473 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1474 (load addr:$src)))]>;
1476 // SSE2 instructions without OpSize prefix
1477 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1478 "cvtps2pd\t{$src, $dst|$dst, $src}",
1479 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1480 TB, Requires<[HasSSE2]>;
1481 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f64mem:$src),
1482 "cvtps2pd\t{$src, $dst|$dst, $src}",
1483 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1484 (load addr:$src)))]>,
1485 TB, Requires<[HasSSE2]>;
1487 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1488 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1489 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1490 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f128mem:$src),
1491 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1492 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1493 (load addr:$src)))]>;
1495 // Match intrinsics which expect XMM operand(s).
1496 // Aliases for intrinsics
1497 let isTwoAddress = 1 in {
1498 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1499 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1500 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1501 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1503 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1504 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1505 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1506 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1507 (loadi32 addr:$src2)))]>;
1508 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1509 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1510 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1511 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1513 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1514 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1515 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1516 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1517 (load addr:$src2)))]>;
1518 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1519 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1520 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1521 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1522 VR128:$src2))]>, XS,
1523 Requires<[HasSSE2]>;
1524 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1525 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1526 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1527 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1528 (load addr:$src2)))]>, XS,
1529 Requires<[HasSSE2]>;
1534 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1536 /// In addition, we also have a special variant of the scalar form here to
1537 /// represent the associated intrinsic operation. This form is unlike the
1538 /// plain scalar form, in that it takes an entire vector (instead of a
1539 /// scalar) and leaves the top elements undefined.
1541 /// And, we have a special variant form for a full-vector intrinsic form.
1543 /// These four forms can each have a reg or a mem operand, so there are a
1544 /// total of eight "instructions".
1546 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1550 bit Commutable = 0> {
1551 // Scalar operation, reg.
1552 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1553 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1554 [(set FR64:$dst, (OpNode FR64:$src))]> {
1555 let isCommutable = Commutable;
1558 // Scalar operation, mem.
1559 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1560 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1561 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1563 // Vector operation, reg.
1564 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1565 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1566 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1567 let isCommutable = Commutable;
1570 // Vector operation, mem.
1571 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1572 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1573 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1575 // Intrinsic operation, reg.
1576 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1577 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1578 [(set VR128:$dst, (F64Int VR128:$src))]> {
1579 let isCommutable = Commutable;
1582 // Intrinsic operation, mem.
1583 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1584 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1585 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1587 // Vector intrinsic operation, reg
1588 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1589 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1590 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1591 let isCommutable = Commutable;
1594 // Vector intrinsic operation, mem
1595 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1596 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1597 [(set VR128:$dst, (V2F64Int (load addr:$src)))]>;
1601 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1602 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1604 // There is no f64 version of the reciprocal approximation instructions.
1607 let isTwoAddress = 1 in {
1608 let isCommutable = 1 in {
1609 def ANDPDrr : PDI<0x54, MRMSrcReg,
1610 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1611 "andpd\t{$src2, $dst|$dst, $src2}",
1613 (and (bc_v2i64 (v2f64 VR128:$src1)),
1614 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1615 def ORPDrr : PDI<0x56, MRMSrcReg,
1616 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1617 "orpd\t{$src2, $dst|$dst, $src2}",
1619 (or (bc_v2i64 (v2f64 VR128:$src1)),
1620 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1621 def XORPDrr : PDI<0x57, MRMSrcReg,
1622 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1623 "xorpd\t{$src2, $dst|$dst, $src2}",
1625 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1626 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1629 def ANDPDrm : PDI<0x54, MRMSrcMem,
1630 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1631 "andpd\t{$src2, $dst|$dst, $src2}",
1633 (and (bc_v2i64 (v2f64 VR128:$src1)),
1634 (memopv2i64 addr:$src2)))]>;
1635 def ORPDrm : PDI<0x56, MRMSrcMem,
1636 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1637 "orpd\t{$src2, $dst|$dst, $src2}",
1639 (or (bc_v2i64 (v2f64 VR128:$src1)),
1640 (memopv2i64 addr:$src2)))]>;
1641 def XORPDrm : PDI<0x57, MRMSrcMem,
1642 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1643 "xorpd\t{$src2, $dst|$dst, $src2}",
1645 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1646 (memopv2i64 addr:$src2)))]>;
1647 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1648 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1649 "andnpd\t{$src2, $dst|$dst, $src2}",
1651 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1652 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1653 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1654 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1655 "andnpd\t{$src2, $dst|$dst, $src2}",
1657 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1658 (memopv2i64 addr:$src2)))]>;
1661 let isTwoAddress = 1 in {
1662 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1663 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1664 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1665 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1666 VR128:$src, imm:$cc))]>;
1667 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1668 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1669 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1670 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1671 (load addr:$src), imm:$cc))]>;
1674 // Shuffle and unpack instructions
1675 let isTwoAddress = 1 in {
1676 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1677 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1678 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1679 [(set VR128:$dst, (v2f64 (vector_shuffle
1680 VR128:$src1, VR128:$src2,
1681 SHUFP_shuffle_mask:$src3)))]>;
1682 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1683 (outs VR128:$dst), (ins VR128:$src1,
1684 f128mem:$src2, i8imm:$src3),
1685 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1687 (v2f64 (vector_shuffle
1688 VR128:$src1, (memopv2f64 addr:$src2),
1689 SHUFP_shuffle_mask:$src3)))]>;
1691 let AddedComplexity = 10 in {
1692 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1693 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1694 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1696 (v2f64 (vector_shuffle
1697 VR128:$src1, VR128:$src2,
1698 UNPCKH_shuffle_mask)))]>;
1699 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1700 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1701 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1703 (v2f64 (vector_shuffle
1704 VR128:$src1, (memopv2f64 addr:$src2),
1705 UNPCKH_shuffle_mask)))]>;
1707 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1708 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1709 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1711 (v2f64 (vector_shuffle
1712 VR128:$src1, VR128:$src2,
1713 UNPCKL_shuffle_mask)))]>;
1714 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1715 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1716 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1718 (v2f64 (vector_shuffle
1719 VR128:$src1, (memopv2f64 addr:$src2),
1720 UNPCKL_shuffle_mask)))]>;
1721 } // AddedComplexity
1725 //===----------------------------------------------------------------------===//
1726 // SSE integer instructions
1728 // Move Instructions
1729 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1730 "movdqa\t{$src, $dst|$dst, $src}", []>;
1732 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1733 "movdqa\t{$src, $dst|$dst, $src}",
1734 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1735 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1736 "movdqa\t{$src, $dst|$dst, $src}",
1737 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1739 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1740 "movdqu\t{$src, $dst|$dst, $src}",
1741 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1742 XS, Requires<[HasSSE2]>;
1743 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1744 "movdqu\t{$src, $dst|$dst, $src}",
1745 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1746 XS, Requires<[HasSSE2]>;
1748 // Intrinsic forms of MOVDQU load and store
1750 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1751 "movdqu\t{$src, $dst|$dst, $src}",
1752 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1753 XS, Requires<[HasSSE2]>;
1754 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1755 "movdqu\t{$src, $dst|$dst, $src}",
1756 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1757 XS, Requires<[HasSSE2]>;
1759 let isTwoAddress = 1 in {
1761 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1762 bit Commutable = 0> {
1763 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1764 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1765 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1766 let isCommutable = Commutable;
1768 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1769 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1770 [(set VR128:$dst, (IntId VR128:$src1,
1771 (bitconvert (memopv2i64 addr:$src2))))]>;
1774 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1775 string OpcodeStr, Intrinsic IntId> {
1776 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1777 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1778 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1779 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1780 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1781 [(set VR128:$dst, (IntId VR128:$src1,
1782 (bitconvert (memopv2i64 addr:$src2))))]>;
1783 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1784 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1785 [(set VR128:$dst, (IntId VR128:$src1,
1786 (scalar_to_vector (i32 imm:$src2))))]>;
1790 /// PDI_binop_rm - Simple SSE2 binary operator.
1791 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1792 ValueType OpVT, bit Commutable = 0> {
1793 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1794 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1795 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1796 let isCommutable = Commutable;
1798 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1799 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1800 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1801 (bitconvert (memopv2i64 addr:$src2)))))]>;
1804 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1806 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1807 /// to collapse (bitconvert VT to VT) into its operand.
1809 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1810 bit Commutable = 0> {
1811 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1812 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1813 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1814 let isCommutable = Commutable;
1816 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1817 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1818 [(set VR128:$dst, (OpNode VR128:$src1,(memopv2i64 addr:$src2)))]>;
1823 // 128-bit Integer Arithmetic
1825 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1826 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1827 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1828 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1830 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1831 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1832 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1833 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1835 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1836 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1837 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1838 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1840 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1841 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1842 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1843 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1845 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1847 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1848 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1849 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1851 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1853 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1854 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1857 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1858 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1859 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1860 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1861 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1864 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw", int_x86_sse2_psll_w>;
1865 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld", int_x86_sse2_psll_d>;
1866 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq", int_x86_sse2_psll_q>;
1868 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw", int_x86_sse2_psrl_w>;
1869 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld", int_x86_sse2_psrl_d>;
1870 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq", int_x86_sse2_psrl_q>;
1872 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw", int_x86_sse2_psra_w>;
1873 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad", int_x86_sse2_psra_d>;
1874 // PSRAQ doesn't exist in SSE[1-3].
1876 // 128-bit logical shifts.
1877 let isTwoAddress = 1 in {
1878 def PSLLDQri : PDIi8<0x73, MRM7r,
1879 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1880 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1881 def PSRLDQri : PDIi8<0x73, MRM3r,
1882 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1883 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1884 // PSRADQri doesn't exist in SSE[1-3].
1887 let Predicates = [HasSSE2] in {
1888 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1889 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1890 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1891 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1892 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1893 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1897 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1898 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1899 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1901 let isTwoAddress = 1 in {
1902 def PANDNrr : PDI<0xDF, MRMSrcReg,
1903 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1904 "pandn\t{$src2, $dst|$dst, $src2}",
1905 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1908 def PANDNrm : PDI<0xDF, MRMSrcMem,
1909 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1910 "pandn\t{$src2, $dst|$dst, $src2}",
1911 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1912 (memopv2i64 addr:$src2))))]>;
1915 // SSE2 Integer comparison
1916 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1917 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1918 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1919 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1920 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1921 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1923 // Pack instructions
1924 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1925 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1926 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1928 // Shuffle and unpack instructions
1929 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1930 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1931 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1932 [(set VR128:$dst, (v4i32 (vector_shuffle
1933 VR128:$src1, (undef),
1934 PSHUFD_shuffle_mask:$src2)))]>;
1935 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1936 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1937 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1938 [(set VR128:$dst, (v4i32 (vector_shuffle
1939 (bc_v4i32(memopv2i64 addr:$src1)),
1941 PSHUFD_shuffle_mask:$src2)))]>;
1943 // SSE2 with ImmT == Imm8 and XS prefix.
1944 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1945 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1946 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1947 [(set VR128:$dst, (v8i16 (vector_shuffle
1948 VR128:$src1, (undef),
1949 PSHUFHW_shuffle_mask:$src2)))]>,
1950 XS, Requires<[HasSSE2]>;
1951 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
1952 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1953 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1954 [(set VR128:$dst, (v8i16 (vector_shuffle
1955 (bc_v8i16 (memopv2i64 addr:$src1)),
1957 PSHUFHW_shuffle_mask:$src2)))]>,
1958 XS, Requires<[HasSSE2]>;
1960 // SSE2 with ImmT == Imm8 and XD prefix.
1961 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
1962 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1963 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1964 [(set VR128:$dst, (v8i16 (vector_shuffle
1965 VR128:$src1, (undef),
1966 PSHUFLW_shuffle_mask:$src2)))]>,
1967 XD, Requires<[HasSSE2]>;
1968 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
1969 (outs VR128:$dst), (ins i128mem:$src1, i32i8imm:$src2),
1970 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1971 [(set VR128:$dst, (v8i16 (vector_shuffle
1972 (bc_v8i16 (memopv2i64 addr:$src1)),
1974 PSHUFLW_shuffle_mask:$src2)))]>,
1975 XD, Requires<[HasSSE2]>;
1978 let isTwoAddress = 1 in {
1979 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
1980 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1981 "punpcklbw\t{$src2, $dst|$dst, $src2}",
1983 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1984 UNPCKL_shuffle_mask)))]>;
1985 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
1986 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1987 "punpcklbw\t{$src2, $dst|$dst, $src2}",
1989 (v16i8 (vector_shuffle VR128:$src1,
1990 (bc_v16i8 (memopv2i64 addr:$src2)),
1991 UNPCKL_shuffle_mask)))]>;
1992 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
1993 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1994 "punpcklwd\t{$src2, $dst|$dst, $src2}",
1996 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1997 UNPCKL_shuffle_mask)))]>;
1998 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
1999 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2000 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2002 (v8i16 (vector_shuffle VR128:$src1,
2003 (bc_v8i16 (memopv2i64 addr:$src2)),
2004 UNPCKL_shuffle_mask)))]>;
2005 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2006 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2007 "punpckldq\t{$src2, $dst|$dst, $src2}",
2009 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2010 UNPCKL_shuffle_mask)))]>;
2011 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2012 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2013 "punpckldq\t{$src2, $dst|$dst, $src2}",
2015 (v4i32 (vector_shuffle VR128:$src1,
2016 (bc_v4i32 (memopv2i64 addr:$src2)),
2017 UNPCKL_shuffle_mask)))]>;
2018 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2019 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2020 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2022 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2023 UNPCKL_shuffle_mask)))]>;
2024 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2025 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2026 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2028 (v2i64 (vector_shuffle VR128:$src1,
2029 (memopv2i64 addr:$src2),
2030 UNPCKL_shuffle_mask)))]>;
2032 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2033 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2034 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2036 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
2037 UNPCKH_shuffle_mask)))]>;
2038 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2039 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2040 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2042 (v16i8 (vector_shuffle VR128:$src1,
2043 (bc_v16i8 (memopv2i64 addr:$src2)),
2044 UNPCKH_shuffle_mask)))]>;
2045 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2046 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2047 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2049 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2050 UNPCKH_shuffle_mask)))]>;
2051 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2052 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2053 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2055 (v8i16 (vector_shuffle VR128:$src1,
2056 (bc_v8i16 (memopv2i64 addr:$src2)),
2057 UNPCKH_shuffle_mask)))]>;
2058 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2059 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2060 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2062 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2063 UNPCKH_shuffle_mask)))]>;
2064 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2065 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2066 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2068 (v4i32 (vector_shuffle VR128:$src1,
2069 (bc_v4i32 (memopv2i64 addr:$src2)),
2070 UNPCKH_shuffle_mask)))]>;
2071 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2072 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2073 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2075 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2076 UNPCKH_shuffle_mask)))]>;
2077 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2078 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2079 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2081 (v2i64 (vector_shuffle VR128:$src1,
2082 (memopv2i64 addr:$src2),
2083 UNPCKH_shuffle_mask)))]>;
2087 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2088 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2089 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2090 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2091 (iPTR imm:$src2)))]>;
2092 let isTwoAddress = 1 in {
2093 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2094 (outs VR128:$dst), (ins VR128:$src1,
2095 GR32:$src2, i32i8imm:$src3),
2096 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2098 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
2099 GR32:$src2, (iPTR imm:$src3))))]>;
2100 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2101 (outs VR128:$dst), (ins VR128:$src1,
2102 i16mem:$src2, i32i8imm:$src3),
2103 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2105 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
2106 (i32 (anyext (loadi16 addr:$src2))),
2107 (iPTR imm:$src3))))]>;
2111 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2112 "pmovmskb\t{$src, $dst|$dst, $src}",
2113 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2115 // Conditional store
2117 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2118 "maskmovdqu\t{$mask, $src|$src, $mask}",
2119 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2121 // Non-temporal stores
2122 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2123 "movntpd\t{$src, $dst|$dst, $src}",
2124 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2125 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2126 "movntdq\t{$src, $dst|$dst, $src}",
2127 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2128 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2129 "movnti\t{$src, $dst|$dst, $src}",
2130 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2131 TB, Requires<[HasSSE2]>;
2134 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2135 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2136 TB, Requires<[HasSSE2]>;
2138 // Load, store, and memory fence
2139 def LFENCE : I<0xAE, MRM5m, (outs), (ins),
2140 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2141 def MFENCE : I<0xAE, MRM6m, (outs), (ins),
2142 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2145 // Alias instructions that map zero vector to pxor / xorp* for sse.
2146 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
2147 let isReMaterializable = 1 in
2148 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2149 "pcmpeqd\t$dst, $dst",
2150 [(set VR128:$dst, (v2f64 immAllOnesV))]>;
2152 // FR64 to 128-bit vector conversion.
2153 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2154 "movsd\t{$src, $dst|$dst, $src}",
2156 (v2f64 (scalar_to_vector FR64:$src)))]>;
2157 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2158 "movsd\t{$src, $dst|$dst, $src}",
2160 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2162 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2163 "movd\t{$src, $dst|$dst, $src}",
2165 (v4i32 (scalar_to_vector GR32:$src)))]>;
2166 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2167 "movd\t{$src, $dst|$dst, $src}",
2169 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2171 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2172 "movd\t{$src, $dst|$dst, $src}",
2173 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2175 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2176 "movd\t{$src, $dst|$dst, $src}",
2177 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2179 // SSE2 instructions with XS prefix
2180 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2181 "movq\t{$src, $dst|$dst, $src}",
2183 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2184 Requires<[HasSSE2]>;
2185 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2186 "movq\t{$src, $dst|$dst, $src}",
2187 [(store (i64 (vector_extract (v2i64 VR128:$src),
2188 (iPTR 0))), addr:$dst)]>;
2190 // FIXME: may not be able to eliminate this movss with coalescing the src and
2191 // dest register classes are different. We really want to write this pattern
2193 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2194 // (f32 FR32:$src)>;
2195 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2196 "movsd\t{$src, $dst|$dst, $src}",
2197 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2199 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2200 "movsd\t{$src, $dst|$dst, $src}",
2201 [(store (f64 (vector_extract (v2f64 VR128:$src),
2202 (iPTR 0))), addr:$dst)]>;
2203 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2204 "movd\t{$src, $dst|$dst, $src}",
2205 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2207 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2208 "movd\t{$src, $dst|$dst, $src}",
2209 [(store (i32 (vector_extract (v4i32 VR128:$src),
2210 (iPTR 0))), addr:$dst)]>;
2212 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2213 "movd\t{$src, $dst|$dst, $src}",
2214 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2215 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2216 "movd\t{$src, $dst|$dst, $src}",
2217 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2220 // Move to lower bits of a VR128, leaving upper bits alone.
2221 // Three operand (but two address) aliases.
2222 let isTwoAddress = 1 in {
2223 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2224 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2225 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2227 let AddedComplexity = 15 in
2228 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2229 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2230 "movsd\t{$src2, $dst|$dst, $src2}",
2232 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
2233 MOVL_shuffle_mask)))]>;
2236 // Store / copy lower 64-bits of a XMM register.
2237 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2238 "movq\t{$src, $dst|$dst, $src}",
2239 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2241 // Move to lower bits of a VR128 and zeroing upper bits.
2242 // Loading from memory automatically zeroing upper bits.
2243 let AddedComplexity = 20 in
2244 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2245 "movsd\t{$src, $dst|$dst, $src}",
2247 (v2f64 (vector_shuffle immAllZerosV,
2248 (v2f64 (scalar_to_vector
2249 (loadf64 addr:$src))),
2250 MOVL_shuffle_mask)))]>;
2252 let AddedComplexity = 15 in
2253 // movd / movq to XMM register zero-extends
2254 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2255 "movd\t{$src, $dst|$dst, $src}",
2257 (v4i32 (vector_shuffle immAllZerosV,
2258 (v4i32 (scalar_to_vector GR32:$src)),
2259 MOVL_shuffle_mask)))]>;
2260 let AddedComplexity = 20 in
2261 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2262 "movd\t{$src, $dst|$dst, $src}",
2264 (v4i32 (vector_shuffle immAllZerosV,
2265 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
2266 MOVL_shuffle_mask)))]>;
2268 // Moving from XMM to XMM but still clear upper 64 bits.
2269 let AddedComplexity = 15 in
2270 def MOVZQI2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2271 "movq\t{$src, $dst|$dst, $src}",
2272 [(set VR128:$dst, (int_x86_sse2_movl_dq VR128:$src))]>,
2273 XS, Requires<[HasSSE2]>;
2274 let AddedComplexity = 20 in
2275 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2276 "movq\t{$src, $dst|$dst, $src}",
2277 [(set VR128:$dst, (int_x86_sse2_movl_dq
2278 (bitconvert (memopv2i64 addr:$src))))]>,
2279 XS, Requires<[HasSSE2]>;
2282 //===----------------------------------------------------------------------===//
2283 // SSE3 Instructions
2284 //===----------------------------------------------------------------------===//
2286 // Move Instructions
2287 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2288 "movshdup\t{$src, $dst|$dst, $src}",
2289 [(set VR128:$dst, (v4f32 (vector_shuffle
2290 VR128:$src, (undef),
2291 MOVSHDUP_shuffle_mask)))]>;
2292 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2293 "movshdup\t{$src, $dst|$dst, $src}",
2294 [(set VR128:$dst, (v4f32 (vector_shuffle
2295 (memopv4f32 addr:$src), (undef),
2296 MOVSHDUP_shuffle_mask)))]>;
2298 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2299 "movsldup\t{$src, $dst|$dst, $src}",
2300 [(set VR128:$dst, (v4f32 (vector_shuffle
2301 VR128:$src, (undef),
2302 MOVSLDUP_shuffle_mask)))]>;
2303 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2304 "movsldup\t{$src, $dst|$dst, $src}",
2305 [(set VR128:$dst, (v4f32 (vector_shuffle
2306 (memopv4f32 addr:$src), (undef),
2307 MOVSLDUP_shuffle_mask)))]>;
2309 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2310 "movddup\t{$src, $dst|$dst, $src}",
2311 [(set VR128:$dst, (v2f64 (vector_shuffle
2312 VR128:$src, (undef),
2313 SSE_splat_lo_mask)))]>;
2314 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2315 "movddup\t{$src, $dst|$dst, $src}",
2317 (v2f64 (vector_shuffle
2318 (scalar_to_vector (loadf64 addr:$src)),
2320 SSE_splat_lo_mask)))]>;
2323 let isTwoAddress = 1 in {
2324 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2325 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2326 "addsubps\t{$src2, $dst|$dst, $src2}",
2327 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2329 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2330 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2331 "addsubps\t{$src2, $dst|$dst, $src2}",
2332 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2333 (load addr:$src2)))]>;
2334 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2335 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2336 "addsubpd\t{$src2, $dst|$dst, $src2}",
2337 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2339 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2340 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2341 "addsubpd\t{$src2, $dst|$dst, $src2}",
2342 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2343 (load addr:$src2)))]>;
2346 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2347 "lddqu\t{$src, $dst|$dst, $src}",
2348 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2351 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2352 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2353 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2354 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2355 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2356 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2357 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2358 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
2359 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2360 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2361 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2362 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2363 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2364 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2365 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2366 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
2368 let isTwoAddress = 1 in {
2369 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2370 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2371 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2372 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2373 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2374 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2375 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2376 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2379 // Thread synchronization
2380 def MONITOR : I<0xC8, RawFrm, (outs), (ins), "monitor",
2381 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2382 def MWAIT : I<0xC9, RawFrm, (outs), (ins), "mwait",
2383 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2385 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2386 let AddedComplexity = 15 in
2387 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2388 MOVSHDUP_shuffle_mask)),
2389 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2390 let AddedComplexity = 20 in
2391 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2392 MOVSHDUP_shuffle_mask)),
2393 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2395 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2396 let AddedComplexity = 15 in
2397 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2398 MOVSLDUP_shuffle_mask)),
2399 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2400 let AddedComplexity = 20 in
2401 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2402 MOVSLDUP_shuffle_mask)),
2403 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2405 //===----------------------------------------------------------------------===//
2406 // SSSE3 Instructions
2407 //===----------------------------------------------------------------------===//
2409 // SSSE3 Instruction Templates:
2411 // SS38I - SSSE3 instructions with T8 prefix.
2412 // SS3AI - SSSE3 instructions with TA prefix.
2414 // Note: SSSE3 instructions have 64-bit and 128-bit versions. The 64-bit version
2415 // uses the MMX registers. We put those instructions here because they better
2416 // fit into the SSSE3 instruction category rather than the MMX category.
2418 class SS38I<bits<8> o, Format F, dag outs, dag ins, string asm,
2420 : I<o, F, outs, ins, asm, pattern>, T8, Requires<[HasSSSE3]>;
2421 class SS3AI<bits<8> o, Format F, dag outs, dag ins, string asm,
2423 : I<o, F, outs, ins, asm, pattern>, TA, Requires<[HasSSSE3]>;
2425 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2426 let isTwoAddress = 1 in {
2427 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2428 Intrinsic IntId64, Intrinsic IntId128,
2429 bit Commutable = 0> {
2430 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2431 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2432 [(set VR64:$dst, (IntId64 VR64:$src))]> {
2433 let isCommutable = Commutable;
2435 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2436 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2438 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2440 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2442 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2443 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2445 let isCommutable = Commutable;
2447 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2449 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2452 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2456 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2457 let isTwoAddress = 1 in {
2458 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2459 Intrinsic IntId64, Intrinsic IntId128,
2460 bit Commutable = 0> {
2461 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2463 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2464 [(set VR64:$dst, (IntId64 VR64:$src))]> {
2465 let isCommutable = Commutable;
2467 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2469 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2472 (bitconvert (memopv4i16 addr:$src))))]>;
2474 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2476 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2477 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2479 let isCommutable = Commutable;
2481 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2483 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2486 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2490 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2491 let isTwoAddress = 1 in {
2492 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2493 Intrinsic IntId64, Intrinsic IntId128,
2494 bit Commutable = 0> {
2495 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2497 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2498 [(set VR64:$dst, (IntId64 VR64:$src))]> {
2499 let isCommutable = Commutable;
2501 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2503 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2506 (bitconvert (memopv2i32 addr:$src))))]>;
2508 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2510 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2511 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2513 let isCommutable = Commutable;
2515 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2517 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2520 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2524 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2525 int_x86_ssse3_pabs_b,
2526 int_x86_ssse3_pabs_b_128>;
2527 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2528 int_x86_ssse3_pabs_w,
2529 int_x86_ssse3_pabs_w_128>;
2530 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2531 int_x86_ssse3_pabs_d,
2532 int_x86_ssse3_pabs_d_128>;
2534 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2535 let isTwoAddress = 1 in {
2536 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2537 Intrinsic IntId64, Intrinsic IntId128,
2538 bit Commutable = 0> {
2539 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2540 (ins VR64:$src1, VR64:$src2),
2541 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2542 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2543 let isCommutable = Commutable;
2545 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2546 (ins VR64:$src1, i64mem:$src2),
2547 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2549 (IntId64 VR64:$src1,
2550 (bitconvert (memopv8i8 addr:$src2))))]>;
2552 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2553 (ins VR128:$src1, VR128:$src2),
2554 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2555 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2557 let isCommutable = Commutable;
2559 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2560 (ins VR128:$src1, i128mem:$src2),
2561 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2563 (IntId128 VR128:$src1,
2564 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2568 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2569 let isTwoAddress = 1 in {
2570 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2571 Intrinsic IntId64, Intrinsic IntId128,
2572 bit Commutable = 0> {
2573 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2574 (ins VR64:$src1, VR64:$src2),
2575 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2576 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2577 let isCommutable = Commutable;
2579 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2580 (ins VR64:$src1, i64mem:$src2),
2581 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2583 (IntId64 VR64:$src1,
2584 (bitconvert (memopv4i16 addr:$src2))))]>;
2586 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2587 (ins VR128:$src1, VR128:$src2),
2588 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2589 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2591 let isCommutable = Commutable;
2593 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2594 (ins VR128:$src1, i128mem:$src2),
2595 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2597 (IntId128 VR128:$src1,
2598 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2602 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2603 let isTwoAddress = 1 in {
2604 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2605 Intrinsic IntId64, Intrinsic IntId128,
2606 bit Commutable = 0> {
2607 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2608 (ins VR64:$src1, VR64:$src2),
2609 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2610 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2611 let isCommutable = Commutable;
2613 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2614 (ins VR64:$src1, i64mem:$src2),
2615 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2617 (IntId64 VR64:$src1,
2618 (bitconvert (memopv2i32 addr:$src2))))]>;
2620 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2621 (ins VR128:$src1, VR128:$src2),
2622 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2623 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2625 let isCommutable = Commutable;
2627 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2628 (ins VR128:$src1, i128mem:$src2),
2629 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2631 (IntId128 VR128:$src1,
2632 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2636 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2637 int_x86_ssse3_phadd_w,
2638 int_x86_ssse3_phadd_w_128, 1>;
2639 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2640 int_x86_ssse3_phadd_d,
2641 int_x86_ssse3_phadd_d_128, 1>;
2642 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2643 int_x86_ssse3_phadd_sw,
2644 int_x86_ssse3_phadd_sw_128, 1>;
2645 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2646 int_x86_ssse3_phsub_w,
2647 int_x86_ssse3_phsub_w_128>;
2648 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2649 int_x86_ssse3_phsub_d,
2650 int_x86_ssse3_phsub_d_128>;
2651 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2652 int_x86_ssse3_phsub_sw,
2653 int_x86_ssse3_phsub_sw_128>;
2654 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2655 int_x86_ssse3_pmadd_ub_sw,
2656 int_x86_ssse3_pmadd_ub_sw_128, 1>;
2657 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2658 int_x86_ssse3_pmul_hr_sw,
2659 int_x86_ssse3_pmul_hr_sw_128, 1>;
2660 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2661 int_x86_ssse3_pshuf_b,
2662 int_x86_ssse3_pshuf_b_128>;
2663 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2664 int_x86_ssse3_psign_b,
2665 int_x86_ssse3_psign_b_128>;
2666 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2667 int_x86_ssse3_psign_w,
2668 int_x86_ssse3_psign_w_128>;
2669 defm PSIGND : SS3I_binop_rm_int_32<0x09, "psignd",
2670 int_x86_ssse3_psign_d,
2671 int_x86_ssse3_psign_d_128>;
2673 let isTwoAddress = 1 in {
2674 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2675 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2676 "palignr\t{$src2, $dst|$dst, $src2}",
2678 (int_x86_ssse3_palign_r
2679 VR64:$src1, VR64:$src2,
2681 def PALIGNR64rm : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2682 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2683 "palignr\t{$src2, $dst|$dst, $src2}",
2685 (int_x86_ssse3_palign_r
2687 (bitconvert (memopv2i32 addr:$src2)),
2690 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2691 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2692 "palignr\t{$src2, $dst|$dst, $src2}",
2694 (int_x86_ssse3_palign_r_128
2695 VR128:$src1, VR128:$src2,
2696 imm:$src3))]>, OpSize;
2697 def PALIGNR128rm : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2698 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2699 "palignr\t{$src2, $dst|$dst, $src2}",
2701 (int_x86_ssse3_palign_r_128
2703 (bitconvert (memopv4i32 addr:$src2)),
2704 imm:$src3))]>, OpSize;
2707 //===----------------------------------------------------------------------===//
2708 // Non-Instruction Patterns
2709 //===----------------------------------------------------------------------===//
2711 // 128-bit vector undef's.
2712 def : Pat<(v4f32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2713 def : Pat<(v2f64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2714 def : Pat<(v16i8 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2715 def : Pat<(v8i16 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2716 def : Pat<(v4i32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2717 def : Pat<(v2i64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2719 // 128-bit vector all zero's.
2720 def : Pat<(v16i8 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2721 def : Pat<(v8i16 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2722 def : Pat<(v4i32 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2723 def : Pat<(v2i64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2724 def : Pat<(v2f64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2726 // 128-bit vector all one's.
2727 def : Pat<(v16i8 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2728 def : Pat<(v8i16 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2729 def : Pat<(v4i32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2730 def : Pat<(v2i64 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2731 def : Pat<(v4f32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE1]>;
2734 // Scalar to v8i16 / v16i8. The source may be a GR32, but only the lower 8 or
2736 def : Pat<(v8i16 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2737 Requires<[HasSSE2]>;
2738 def : Pat<(v16i8 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2739 Requires<[HasSSE2]>;
2742 let Predicates = [HasSSE2] in {
2743 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2744 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2745 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2746 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2747 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2748 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2749 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2750 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2751 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2752 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2753 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2754 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2755 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2756 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2757 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2758 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2759 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2760 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2761 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2762 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2763 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2764 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2765 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2766 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2767 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2768 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2769 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2770 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2771 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2772 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2775 // Move scalar to XMM zero-extended
2776 // movd to XMM register zero-extends
2777 let AddedComplexity = 15 in {
2778 def : Pat<(v8i16 (vector_shuffle immAllZerosV,
2779 (v8i16 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
2780 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
2781 def : Pat<(v16i8 (vector_shuffle immAllZerosV,
2782 (v16i8 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
2783 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
2784 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2785 def : Pat<(v2f64 (vector_shuffle immAllZerosV,
2786 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
2787 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2788 def : Pat<(v4f32 (vector_shuffle immAllZerosV,
2789 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
2790 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
2793 // Splat v2f64 / v2i64
2794 let AddedComplexity = 10 in {
2795 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2796 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2797 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2798 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2799 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2800 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2801 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2802 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2806 def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SSE_splat_mask:$sm),
2807 (SHUFPSrri VR128:$src, VR128:$src, SSE_splat_mask:$sm)>,
2808 Requires<[HasSSE1]>;
2810 // Special unary SHUFPSrri case.
2811 // FIXME: when we want non two-address code, then we should use PSHUFD?
2812 def : Pat<(vector_shuffle (v4f32 VR128:$src1), (undef),
2813 SHUFP_unary_shuffle_mask:$sm),
2814 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2815 Requires<[HasSSE1]>;
2816 // Special unary SHUFPDrri case.
2817 def : Pat<(vector_shuffle (v2f64 VR128:$src1), (undef),
2818 SHUFP_unary_shuffle_mask:$sm),
2819 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2820 Requires<[HasSSE2]>;
2821 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2822 def : Pat<(vector_shuffle (memopv4f32 addr:$src1), (undef),
2823 SHUFP_unary_shuffle_mask:$sm),
2824 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2825 Requires<[HasSSE2]>;
2826 // Special binary v4i32 shuffle cases with SHUFPS.
2827 def : Pat<(vector_shuffle (v4i32 VR128:$src1), (v4i32 VR128:$src2),
2828 PSHUFD_binary_shuffle_mask:$sm),
2829 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2830 Requires<[HasSSE2]>;
2831 def : Pat<(vector_shuffle (v4i32 VR128:$src1),
2832 (bc_v4i32 (memopv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm),
2833 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2834 Requires<[HasSSE2]>;
2836 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2837 let AddedComplexity = 10 in {
2838 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2839 UNPCKL_v_undef_shuffle_mask)),
2840 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2841 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2842 UNPCKL_v_undef_shuffle_mask)),
2843 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2844 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2845 UNPCKL_v_undef_shuffle_mask)),
2846 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2847 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2848 UNPCKL_v_undef_shuffle_mask)),
2849 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2852 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2853 let AddedComplexity = 10 in {
2854 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2855 UNPCKH_v_undef_shuffle_mask)),
2856 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2857 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2858 UNPCKH_v_undef_shuffle_mask)),
2859 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2860 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2861 UNPCKH_v_undef_shuffle_mask)),
2862 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2863 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2864 UNPCKH_v_undef_shuffle_mask)),
2865 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2868 let AddedComplexity = 15 in {
2869 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2870 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2871 MOVHP_shuffle_mask)),
2872 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2874 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2875 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2876 MOVHLPS_shuffle_mask)),
2877 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2879 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2880 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2881 MOVHLPS_v_undef_shuffle_mask)),
2882 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2883 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2884 MOVHLPS_v_undef_shuffle_mask)),
2885 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2888 let AddedComplexity = 20 in {
2889 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2890 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2891 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2892 MOVLP_shuffle_mask)),
2893 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2894 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2895 MOVLP_shuffle_mask)),
2896 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2897 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2898 MOVHP_shuffle_mask)),
2899 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2900 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2901 MOVHP_shuffle_mask)),
2902 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2904 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2905 MOVLP_shuffle_mask)),
2906 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2907 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2908 MOVLP_shuffle_mask)),
2909 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2910 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2911 MOVHP_shuffle_mask)),
2912 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2913 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2914 MOVLP_shuffle_mask)),
2915 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2918 let AddedComplexity = 15 in {
2919 // Setting the lowest element in the vector.
2920 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2921 MOVL_shuffle_mask)),
2922 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2923 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2924 MOVL_shuffle_mask)),
2925 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2927 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2928 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2929 MOVLP_shuffle_mask)),
2930 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2931 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2932 MOVLP_shuffle_mask)),
2933 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2936 // Set lowest element and zero upper elements.
2937 let AddedComplexity = 20 in
2938 def : Pat<(bc_v2i64 (vector_shuffle immAllZerosV,
2939 (v2f64 (scalar_to_vector (loadf64 addr:$src))),
2940 MOVL_shuffle_mask)),
2941 (MOVZQI2PQIrm addr:$src)>, Requires<[HasSSE2]>;
2943 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2944 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2945 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2946 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2947 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2948 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2949 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2950 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2951 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2952 Requires<[HasSSE2]>;
2953 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2954 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2955 Requires<[HasSSE2]>;
2956 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2957 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2958 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2959 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2960 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2961 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2962 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2963 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2964 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
2965 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2966 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
2967 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2968 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
2969 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2970 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
2971 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2973 // Some special case pandn patterns.
2974 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2976 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2977 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2979 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2980 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2982 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2984 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2985 (memopv2i64 addr:$src2))),
2986 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2987 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2988 (memopv2i64 addr:$src2))),
2989 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2990 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2991 (memopv2i64 addr:$src2))),
2992 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2994 // Use movaps / movups for SSE integer load / store (one byte shorter).
2995 def : Pat<(alignedloadv4i32 addr:$src),
2996 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
2997 def : Pat<(loadv4i32 addr:$src),
2998 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
2999 def : Pat<(alignedloadv2i64 addr:$src),
3000 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3001 def : Pat<(loadv2i64 addr:$src),
3002 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3004 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3005 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3006 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3007 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3008 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3009 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3010 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3011 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3012 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3013 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3014 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3015 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3016 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3017 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3018 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3019 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3021 // (vextract (v4i32 bc (v4f32 s2v (f32 load $addr))), 0) -> (i32 load $addr)
3022 def : Pat<(vector_extract
3023 (bc_v4i32 (v4f32 (scalar_to_vector (loadf32 addr:$src)))), (iPTR 0)),
3024 (MOV32rm addr:$src)>, Requires<[HasSSE2]>;
3025 def : Pat<(vector_extract
3026 (bc_v2i64 (v2f64 (scalar_to_vector (loadf64 addr:$src)))), (iPTR 0)),
3027 (MOV64rm addr:$src)>, Requires<[HasSSE2, In64BitMode]>;