1 //====- X86InstrSSE.td - Describe the X86 Instruction Set --*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the X86 SSE instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
17 //===----------------------------------------------------------------------===//
18 // SSE specific DAG Nodes.
19 //===----------------------------------------------------------------------===//
21 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
22 SDTCisFP<0>, SDTCisInt<2> ]>;
24 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
25 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
26 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
27 [SDNPCommutative, SDNPAssociative]>;
28 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
29 [SDNPCommutative, SDNPAssociative]>;
30 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
31 [SDNPCommutative, SDNPAssociative]>;
32 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
33 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
34 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
35 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>;
36 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>;
37 def X86s2vec : SDNode<"X86ISD::S2VEC", SDTypeProfile<1, 1, []>, []>;
38 def X86pextrw : SDNode<"X86ISD::PEXTRW", SDTypeProfile<1, 2, []>, []>;
39 def X86pinsrw : SDNode<"X86ISD::PINSRW", SDTypeProfile<1, 3, []>, []>;
41 //===----------------------------------------------------------------------===//
42 // SSE 'Special' Instructions
43 //===----------------------------------------------------------------------===//
45 let isImplicitDef = 1 in
46 def IMPLICIT_DEF_VR128 : I<0, Pseudo, (outs VR128:$dst), (ins),
48 [(set VR128:$dst, (v4f32 (undef)))]>,
50 def IMPLICIT_DEF_FR32 : I<0, Pseudo, (outs FR32:$dst), (ins),
52 [(set FR32:$dst, (undef))]>, Requires<[HasSSE1]>;
53 def IMPLICIT_DEF_FR64 : I<0, Pseudo, (outs FR64:$dst), (ins),
55 [(set FR64:$dst, (undef))]>, Requires<[HasSSE2]>;
57 //===----------------------------------------------------------------------===//
58 // SSE Complex Patterns
59 //===----------------------------------------------------------------------===//
61 // These are 'extloads' from a scalar to the low element of a vector, zeroing
62 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
64 def sse_load_f32 : ComplexPattern<v4f32, 4, "SelectScalarSSELoad", [],
66 def sse_load_f64 : ComplexPattern<v2f64, 4, "SelectScalarSSELoad", [],
69 def ssmem : Operand<v4f32> {
70 let PrintMethod = "printf32mem";
71 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
73 def sdmem : Operand<v2f64> {
74 let PrintMethod = "printf64mem";
75 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
78 //===----------------------------------------------------------------------===//
79 // SSE pattern fragments
80 //===----------------------------------------------------------------------===//
82 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
83 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
84 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
85 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
87 // Like 'store', but always requires vector alignment.
88 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
89 (st node:$val, node:$ptr), [{
90 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
91 return !ST->isTruncatingStore() &&
92 ST->getAddressingMode() == ISD::UNINDEXED &&
93 ST->getAlignment() >= 16;
97 // Like 'load', but always requires vector alignment.
98 def alignedload : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
99 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
100 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
101 LD->getAddressingMode() == ISD::UNINDEXED &&
102 LD->getAlignment() >= 16;
106 def alignedloadfsf32 : PatFrag<(ops node:$ptr), (f32 (alignedload node:$ptr))>;
107 def alignedloadfsf64 : PatFrag<(ops node:$ptr), (f64 (alignedload node:$ptr))>;
108 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
109 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
110 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
111 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
113 // Like 'load', but uses special alignment checks suitable for use in
114 // memory operands in most SSE instructions, which are required to
115 // be naturally aligned on some targets but not on others.
116 // FIXME: Actually implement support for targets that don't require the
117 // alignment. This probably wants a subtarget predicate.
118 def memop : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
119 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
120 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
121 LD->getAddressingMode() == ISD::UNINDEXED &&
122 LD->getAlignment() >= 16;
126 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
127 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
128 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
129 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
130 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
131 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
133 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
135 def memop64 : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
136 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
137 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
138 LD->getAddressingMode() == ISD::UNINDEXED &&
139 LD->getAlignment() >= 8;
143 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
144 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop64 node:$ptr))>;
145 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
146 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
147 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
149 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
150 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
151 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
152 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
153 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
154 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
156 def fp32imm0 : PatLeaf<(f32 fpimm), [{
157 return N->isExactlyValue(+0.0);
160 def PSxLDQ_imm : SDNodeXForm<imm, [{
161 // Transformation function: imm >> 3
162 return getI32Imm(N->getValue() >> 3);
165 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
167 def SHUFFLE_get_shuf_imm : SDNodeXForm<build_vector, [{
168 return getI8Imm(X86::getShuffleSHUFImmediate(N));
171 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
173 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<build_vector, [{
174 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
177 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
179 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<build_vector, [{
180 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
183 def SSE_splat_mask : PatLeaf<(build_vector), [{
184 return X86::isSplatMask(N);
185 }], SHUFFLE_get_shuf_imm>;
187 def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
188 return X86::isSplatLoMask(N);
191 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
192 return X86::isMOVHLPSMask(N);
195 def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
196 return X86::isMOVHLPS_v_undef_Mask(N);
199 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
200 return X86::isMOVHPMask(N);
203 def MOVLP_shuffle_mask : PatLeaf<(build_vector), [{
204 return X86::isMOVLPMask(N);
207 def MOVL_shuffle_mask : PatLeaf<(build_vector), [{
208 return X86::isMOVLMask(N);
211 def MOVSHDUP_shuffle_mask : PatLeaf<(build_vector), [{
212 return X86::isMOVSHDUPMask(N);
215 def MOVSLDUP_shuffle_mask : PatLeaf<(build_vector), [{
216 return X86::isMOVSLDUPMask(N);
219 def UNPCKL_shuffle_mask : PatLeaf<(build_vector), [{
220 return X86::isUNPCKLMask(N);
223 def UNPCKH_shuffle_mask : PatLeaf<(build_vector), [{
224 return X86::isUNPCKHMask(N);
227 def UNPCKL_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
228 return X86::isUNPCKL_v_undef_Mask(N);
231 def UNPCKH_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
232 return X86::isUNPCKH_v_undef_Mask(N);
235 def PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
236 return X86::isPSHUFDMask(N);
237 }], SHUFFLE_get_shuf_imm>;
239 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
240 return X86::isPSHUFHWMask(N);
241 }], SHUFFLE_get_pshufhw_imm>;
243 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
244 return X86::isPSHUFLWMask(N);
245 }], SHUFFLE_get_pshuflw_imm>;
247 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
248 return X86::isPSHUFDMask(N);
249 }], SHUFFLE_get_shuf_imm>;
251 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
252 return X86::isSHUFPMask(N);
253 }], SHUFFLE_get_shuf_imm>;
255 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
256 return X86::isSHUFPMask(N);
257 }], SHUFFLE_get_shuf_imm>;
259 //===----------------------------------------------------------------------===//
260 // SSE scalar FP Instructions
261 //===----------------------------------------------------------------------===//
263 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
264 // scheduler into a branch sequence.
265 // These are expanded by the scheduler.
266 let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
267 def CMOV_FR32 : I<0, Pseudo,
268 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
269 "#CMOV_FR32 PSEUDO!",
270 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
272 def CMOV_FR64 : I<0, Pseudo,
273 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
274 "#CMOV_FR64 PSEUDO!",
275 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
277 def CMOV_V4F32 : I<0, Pseudo,
278 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
279 "#CMOV_V4F32 PSEUDO!",
281 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
283 def CMOV_V2F64 : I<0, Pseudo,
284 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
285 "#CMOV_V2F64 PSEUDO!",
287 (v2f64 (X86cmov VR128:$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:$cond,
297 //===----------------------------------------------------------------------===//
299 //===----------------------------------------------------------------------===//
302 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
303 "movss\t{$src, $dst|$dst, $src}", []>;
304 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
305 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
306 "movss\t{$src, $dst|$dst, $src}",
307 [(set FR32:$dst, (loadf32 addr:$src))]>;
308 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
309 "movss\t{$src, $dst|$dst, $src}",
310 [(store FR32:$src, addr:$dst)]>;
312 // Conversion instructions
313 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
314 "cvttss2si\t{$src, $dst|$dst, $src}",
315 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
316 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
317 "cvttss2si\t{$src, $dst|$dst, $src}",
318 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
319 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
320 "cvtsi2ss\t{$src, $dst|$dst, $src}",
321 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
322 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
323 "cvtsi2ss\t{$src, $dst|$dst, $src}",
324 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
326 // Match intrinsics which expect XMM operand(s).
327 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
328 "cvtss2si\t{$src, $dst|$dst, $src}",
329 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
330 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
331 "cvtss2si\t{$src, $dst|$dst, $src}",
332 [(set GR32:$dst, (int_x86_sse_cvtss2si
333 (load addr:$src)))]>;
335 // Match intrinisics which expect MM and XMM operand(s).
336 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
337 "cvtps2pi\t{$src, $dst|$dst, $src}",
338 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
339 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
340 "cvtps2pi\t{$src, $dst|$dst, $src}",
341 [(set VR64:$dst, (int_x86_sse_cvtps2pi
342 (load addr:$src)))]>;
343 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
344 "cvttps2pi\t{$src, $dst|$dst, $src}",
345 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
346 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
347 "cvttps2pi\t{$src, $dst|$dst, $src}",
348 [(set VR64:$dst, (int_x86_sse_cvttps2pi
349 (load addr:$src)))]>;
350 let isTwoAddress = 1 in {
351 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
352 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
353 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
354 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
356 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
357 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
358 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
359 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
360 (load addr:$src2)))]>;
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 : SSIi8<0xC2, MRMSrcReg,
389 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
390 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
391 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
392 (outs FR32:$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:$src2), (implicit EFLAGS)]>;
400 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
401 "ucomiss\t{$src2, $src1|$src1, $src2}",
402 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
406 // Aliases to match intrinsics which expect XMM operand(s).
407 let isTwoAddress = 1 in {
408 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
409 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
410 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
411 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
412 VR128:$src, imm:$cc))]>;
413 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
414 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src, SSECC:$cc),
415 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
416 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
417 (load addr:$src), imm:$cc))]>;
420 let Defs = [EFLAGS] in {
421 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs),
422 (ins VR128:$src1, VR128:$src2),
423 "ucomiss\t{$src2, $src1|$src1, $src2}",
424 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
426 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),
427 (ins VR128:$src1, f128mem:$src2),
428 "ucomiss\t{$src2, $src1|$src1, $src2}",
429 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
432 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs),
433 (ins VR128:$src1, VR128:$src2),
434 "comiss\t{$src2, $src1|$src1, $src2}",
435 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
437 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs),
438 (ins VR128:$src1, f128mem:$src2),
439 "comiss\t{$src2, $src1|$src1, $src2}",
440 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
444 // Aliases of packed SSE1 instructions for scalar use. These all have names that
447 // Alias instructions that map fld0 to pxor for sse.
448 let isReMaterializable = 1, neverHasSideEffects = 1 in
449 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
450 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
451 Requires<[HasSSE1]>, TB, OpSize;
453 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
455 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
456 "movaps\t{$src, $dst|$dst, $src}", []>;
458 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
460 let isSimpleLoad = 1 in
461 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
462 "movaps\t{$src, $dst|$dst, $src}",
463 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
465 // Alias bitwise logical operations using SSE logical ops on packed FP values.
466 let isTwoAddress = 1 in {
467 let isCommutable = 1 in {
468 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
469 "andps\t{$src2, $dst|$dst, $src2}",
470 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
471 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
472 "orps\t{$src2, $dst|$dst, $src2}",
473 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
474 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
475 "xorps\t{$src2, $dst|$dst, $src2}",
476 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
479 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
480 "andps\t{$src2, $dst|$dst, $src2}",
481 [(set FR32:$dst, (X86fand FR32:$src1,
482 (memopfsf32 addr:$src2)))]>;
483 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
484 "orps\t{$src2, $dst|$dst, $src2}",
485 [(set FR32:$dst, (X86for FR32:$src1,
486 (memopfsf32 addr:$src2)))]>;
487 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
488 "xorps\t{$src2, $dst|$dst, $src2}",
489 [(set FR32:$dst, (X86fxor FR32:$src1,
490 (memopfsf32 addr:$src2)))]>;
492 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
493 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
494 "andnps\t{$src2, $dst|$dst, $src2}", []>;
495 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
496 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
497 "andnps\t{$src2, $dst|$dst, $src2}", []>;
500 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
502 /// In addition, we also have a special variant of the scalar form here to
503 /// represent the associated intrinsic operation. This form is unlike the
504 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
505 /// and leaves the top elements undefined.
507 /// These three forms can each be reg+reg or reg+mem, so there are a total of
508 /// six "instructions".
510 let isTwoAddress = 1 in {
511 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
512 SDNode OpNode, Intrinsic F32Int,
513 bit Commutable = 0> {
514 // Scalar operation, reg+reg.
515 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
516 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
517 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
518 let isCommutable = Commutable;
521 // Scalar operation, reg+mem.
522 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
523 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
524 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
526 // Vector operation, reg+reg.
527 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
528 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
529 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
530 let isCommutable = Commutable;
533 // Vector operation, reg+mem.
534 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
535 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
536 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
538 // Intrinsic operation, reg+reg.
539 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
540 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
541 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
542 let isCommutable = Commutable;
545 // Intrinsic operation, reg+mem.
546 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
547 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
548 [(set VR128:$dst, (F32Int VR128:$src1,
549 sse_load_f32:$src2))]>;
553 // Arithmetic instructions
554 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
555 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
556 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
557 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
559 /// sse1_fp_binop_rm - Other SSE1 binops
561 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
562 /// instructions for a full-vector intrinsic form. Operations that map
563 /// onto C operators don't use this form since they just use the plain
564 /// vector form instead of having a separate vector intrinsic form.
566 /// This provides a total of eight "instructions".
568 let isTwoAddress = 1 in {
569 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
573 bit Commutable = 0> {
575 // Scalar operation, reg+reg.
576 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
577 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
578 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
579 let isCommutable = Commutable;
582 // Scalar operation, reg+mem.
583 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
584 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
585 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
587 // Vector operation, reg+reg.
588 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
589 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
590 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
591 let isCommutable = Commutable;
594 // Vector operation, reg+mem.
595 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
596 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
597 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
599 // Intrinsic operation, reg+reg.
600 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
601 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
602 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
603 let isCommutable = Commutable;
606 // Intrinsic operation, reg+mem.
607 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
608 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
609 [(set VR128:$dst, (F32Int VR128:$src1,
610 sse_load_f32:$src2))]>;
612 // Vector intrinsic operation, reg+reg.
613 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
614 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
615 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
616 let isCommutable = Commutable;
619 // Vector intrinsic operation, reg+mem.
620 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
621 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
622 [(set VR128:$dst, (V4F32Int VR128:$src1, (load addr:$src2)))]>;
626 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
627 int_x86_sse_max_ss, int_x86_sse_max_ps>;
628 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
629 int_x86_sse_min_ss, int_x86_sse_min_ps>;
631 //===----------------------------------------------------------------------===//
632 // SSE packed FP Instructions
635 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
636 "movaps\t{$src, $dst|$dst, $src}", []>;
637 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
638 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
639 "movaps\t{$src, $dst|$dst, $src}",
640 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
642 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
643 "movaps\t{$src, $dst|$dst, $src}",
644 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
646 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
647 "movups\t{$src, $dst|$dst, $src}", []>;
648 let isSimpleLoad = 1 in
649 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
650 "movups\t{$src, $dst|$dst, $src}",
651 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
652 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
653 "movups\t{$src, $dst|$dst, $src}",
654 [(store (v4f32 VR128:$src), addr:$dst)]>;
656 // Intrinsic forms of MOVUPS load and store
657 let isSimpleLoad = 1 in
658 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
659 "movups\t{$src, $dst|$dst, $src}",
660 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
661 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
662 "movups\t{$src, $dst|$dst, $src}",
663 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
665 let isTwoAddress = 1 in {
666 let AddedComplexity = 20 in {
667 def MOVLPSrm : PSI<0x12, MRMSrcMem,
668 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
669 "movlps\t{$src2, $dst|$dst, $src2}",
671 (v4f32 (vector_shuffle VR128:$src1,
672 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
673 MOVLP_shuffle_mask)))]>;
674 def MOVHPSrm : PSI<0x16, MRMSrcMem,
675 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
676 "movhps\t{$src2, $dst|$dst, $src2}",
678 (v4f32 (vector_shuffle VR128:$src1,
679 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
680 MOVHP_shuffle_mask)))]>;
684 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
685 "movlps\t{$src, $dst|$dst, $src}",
686 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
687 (iPTR 0))), addr:$dst)]>;
689 // v2f64 extract element 1 is always custom lowered to unpack high to low
690 // and extract element 0 so the non-store version isn't too horrible.
691 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
692 "movhps\t{$src, $dst|$dst, $src}",
693 [(store (f64 (vector_extract
694 (v2f64 (vector_shuffle
695 (bc_v2f64 (v4f32 VR128:$src)), (undef),
696 UNPCKH_shuffle_mask)), (iPTR 0))),
699 let isTwoAddress = 1 in {
700 let AddedComplexity = 15 in {
701 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
702 "movlhps\t{$src2, $dst|$dst, $src2}",
704 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
705 MOVHP_shuffle_mask)))]>;
707 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
708 "movhlps\t{$src2, $dst|$dst, $src2}",
710 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
711 MOVHLPS_shuffle_mask)))]>;
719 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
721 /// In addition, we also have a special variant of the scalar form here to
722 /// represent the associated intrinsic operation. This form is unlike the
723 /// plain scalar form, in that it takes an entire vector (instead of a
724 /// scalar) and leaves the top elements undefined.
726 /// And, we have a special variant form for a full-vector intrinsic form.
728 /// These four forms can each have a reg or a mem operand, so there are a
729 /// total of eight "instructions".
731 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
735 bit Commutable = 0> {
736 // Scalar operation, reg.
737 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
738 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
739 [(set FR32:$dst, (OpNode FR32:$src))]> {
740 let isCommutable = Commutable;
743 // Scalar operation, mem.
744 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
745 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
746 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
748 // Vector operation, reg.
749 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
750 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
751 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
752 let isCommutable = Commutable;
755 // Vector operation, mem.
756 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
757 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
758 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
760 // Intrinsic operation, reg.
761 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
762 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
763 [(set VR128:$dst, (F32Int VR128:$src))]> {
764 let isCommutable = Commutable;
767 // Intrinsic operation, mem.
768 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
769 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
770 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
772 // Vector intrinsic operation, reg
773 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
774 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
775 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
776 let isCommutable = Commutable;
779 // Vector intrinsic operation, mem
780 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
781 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
782 [(set VR128:$dst, (V4F32Int (load addr:$src)))]>;
786 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
787 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
789 // Reciprocal approximations. Note that these typically require refinement
790 // in order to obtain suitable precision.
791 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
792 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
793 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
794 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
797 let isTwoAddress = 1 in {
798 let isCommutable = 1 in {
799 def ANDPSrr : PSI<0x54, MRMSrcReg,
800 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
801 "andps\t{$src2, $dst|$dst, $src2}",
802 [(set VR128:$dst, (v2i64
803 (and VR128:$src1, VR128:$src2)))]>;
804 def ORPSrr : PSI<0x56, MRMSrcReg,
805 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
806 "orps\t{$src2, $dst|$dst, $src2}",
807 [(set VR128:$dst, (v2i64
808 (or VR128:$src1, VR128:$src2)))]>;
809 def XORPSrr : PSI<0x57, MRMSrcReg,
810 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
811 "xorps\t{$src2, $dst|$dst, $src2}",
812 [(set VR128:$dst, (v2i64
813 (xor VR128:$src1, VR128:$src2)))]>;
816 def ANDPSrm : PSI<0x54, MRMSrcMem,
817 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
818 "andps\t{$src2, $dst|$dst, $src2}",
819 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
820 (memopv2i64 addr:$src2)))]>;
821 def ORPSrm : PSI<0x56, MRMSrcMem,
822 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
823 "orps\t{$src2, $dst|$dst, $src2}",
824 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
825 (memopv2i64 addr:$src2)))]>;
826 def XORPSrm : PSI<0x57, MRMSrcMem,
827 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
828 "xorps\t{$src2, $dst|$dst, $src2}",
829 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
830 (memopv2i64 addr:$src2)))]>;
831 def ANDNPSrr : PSI<0x55, MRMSrcReg,
832 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
833 "andnps\t{$src2, $dst|$dst, $src2}",
835 (v2i64 (and (xor VR128:$src1,
836 (bc_v2i64 (v4i32 immAllOnesV))),
838 def ANDNPSrm : PSI<0x55, MRMSrcMem,
839 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
840 "andnps\t{$src2, $dst|$dst, $src2}",
842 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
843 (bc_v2i64 (v4i32 immAllOnesV))),
844 (memopv2i64 addr:$src2))))]>;
847 let isTwoAddress = 1 in {
848 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
849 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
850 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
851 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
852 VR128:$src, imm:$cc))]>;
853 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
854 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
855 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
856 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
857 (load addr:$src), imm:$cc))]>;
860 // Shuffle and unpack instructions
861 let isTwoAddress = 1 in {
862 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
863 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
864 (outs VR128:$dst), (ins VR128:$src1,
865 VR128:$src2, i32i8imm:$src3),
866 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
868 (v4f32 (vector_shuffle
869 VR128:$src1, VR128:$src2,
870 SHUFP_shuffle_mask:$src3)))]>;
871 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
872 (outs VR128:$dst), (ins VR128:$src1,
873 f128mem:$src2, i32i8imm:$src3),
874 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
876 (v4f32 (vector_shuffle
877 VR128:$src1, (memopv4f32 addr:$src2),
878 SHUFP_shuffle_mask:$src3)))]>;
880 let AddedComplexity = 10 in {
881 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
882 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
883 "unpckhps\t{$src2, $dst|$dst, $src2}",
885 (v4f32 (vector_shuffle
886 VR128:$src1, VR128:$src2,
887 UNPCKH_shuffle_mask)))]>;
888 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
889 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
890 "unpckhps\t{$src2, $dst|$dst, $src2}",
892 (v4f32 (vector_shuffle
893 VR128:$src1, (memopv4f32 addr:$src2),
894 UNPCKH_shuffle_mask)))]>;
896 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
897 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
898 "unpcklps\t{$src2, $dst|$dst, $src2}",
900 (v4f32 (vector_shuffle
901 VR128:$src1, VR128:$src2,
902 UNPCKL_shuffle_mask)))]>;
903 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
904 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
905 "unpcklps\t{$src2, $dst|$dst, $src2}",
907 (v4f32 (vector_shuffle
908 VR128:$src1, (memopv4f32 addr:$src2),
909 UNPCKL_shuffle_mask)))]>;
914 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
915 "movmskps\t{$src, $dst|$dst, $src}",
916 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
917 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
918 "movmskpd\t{$src, $dst|$dst, $src}",
919 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
921 // Prefetching loads.
922 // TODO: no intrinsics for these?
923 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src), "prefetcht0\t$src", []>;
924 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src), "prefetcht1\t$src", []>;
925 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src), "prefetcht2\t$src", []>;
926 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src), "prefetchnta\t$src", []>;
928 // Non-temporal stores
929 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
930 "movntps\t{$src, $dst|$dst, $src}",
931 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
933 // Load, store, and memory fence
934 def SFENCE : PSI<0xAE, MRM7m, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
937 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
938 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
939 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
940 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
942 // Alias instructions that map zero vector to pxor / xorp* for sse.
943 let isReMaterializable = 1, neverHasSideEffects = 1 in
944 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
946 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
948 // FR32 to 128-bit vector conversion.
949 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
950 "movss\t{$src, $dst|$dst, $src}",
952 (v4f32 (scalar_to_vector FR32:$src)))]>;
953 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
954 "movss\t{$src, $dst|$dst, $src}",
956 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
958 // FIXME: may not be able to eliminate this movss with coalescing the src and
959 // dest register classes are different. We really want to write this pattern
961 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
963 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
964 "movss\t{$src, $dst|$dst, $src}",
965 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
967 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
968 "movss\t{$src, $dst|$dst, $src}",
969 [(store (f32 (vector_extract (v4f32 VR128:$src),
970 (iPTR 0))), addr:$dst)]>;
973 // Move to lower bits of a VR128, leaving upper bits alone.
974 // Three operand (but two address) aliases.
975 let isTwoAddress = 1 in {
976 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
977 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
978 "movss\t{$src2, $dst|$dst, $src2}", []>;
980 let AddedComplexity = 15 in
981 def MOVLPSrr : SSI<0x10, MRMSrcReg,
982 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
983 "movss\t{$src2, $dst|$dst, $src2}",
985 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
986 MOVL_shuffle_mask)))]>;
989 // Move to lower bits of a VR128 and zeroing upper bits.
990 // Loading from memory automatically zeroing upper bits.
991 let AddedComplexity = 20 in
992 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
993 "movss\t{$src, $dst|$dst, $src}",
994 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV_bc,
995 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
996 MOVL_shuffle_mask)))]>;
999 //===----------------------------------------------------------------------===//
1000 // SSE2 Instructions
1001 //===----------------------------------------------------------------------===//
1003 // Move Instructions
1004 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1005 "movsd\t{$src, $dst|$dst, $src}", []>;
1006 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1007 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1008 "movsd\t{$src, $dst|$dst, $src}",
1009 [(set FR64:$dst, (loadf64 addr:$src))]>;
1010 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1011 "movsd\t{$src, $dst|$dst, $src}",
1012 [(store FR64:$src, addr:$dst)]>;
1014 // Conversion instructions
1015 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1016 "cvttsd2si\t{$src, $dst|$dst, $src}",
1017 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1018 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1019 "cvttsd2si\t{$src, $dst|$dst, $src}",
1020 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1021 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1022 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1023 [(set FR32:$dst, (fround FR64:$src))]>;
1024 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1025 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1026 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1027 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1028 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1029 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1030 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1031 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1032 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1034 // SSE2 instructions with XS prefix
1035 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1036 "cvtss2sd\t{$src, $dst|$dst, $src}",
1037 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1038 Requires<[HasSSE2]>;
1039 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1040 "cvtss2sd\t{$src, $dst|$dst, $src}",
1041 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1042 Requires<[HasSSE2]>;
1044 // Match intrinsics which expect XMM operand(s).
1045 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1046 "cvtsd2si\t{$src, $dst|$dst, $src}",
1047 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1048 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1049 "cvtsd2si\t{$src, $dst|$dst, $src}",
1050 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1051 (load addr:$src)))]>;
1053 // Match intrinisics which expect MM and XMM operand(s).
1054 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1055 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1056 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1057 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1058 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1059 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1060 (load addr:$src)))]>;
1061 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1062 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1063 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1064 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1065 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1066 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1067 (load addr:$src)))]>;
1068 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1069 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1070 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1071 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1072 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1073 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1074 (load addr:$src)))]>;
1076 // Aliases for intrinsics
1077 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1078 "cvttsd2si\t{$src, $dst|$dst, $src}",
1080 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1081 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1082 "cvttsd2si\t{$src, $dst|$dst, $src}",
1083 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1084 (load addr:$src)))]>;
1086 // Comparison instructions
1087 let isTwoAddress = 1 in {
1088 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1089 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1090 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1091 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1092 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1093 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1096 let Defs = [EFLAGS] in {
1097 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1098 "ucomisd\t{$src2, $src1|$src1, $src2}",
1099 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1100 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1101 "ucomisd\t{$src2, $src1|$src1, $src2}",
1102 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1103 (implicit EFLAGS)]>;
1106 // Aliases to match intrinsics which expect XMM operand(s).
1107 let isTwoAddress = 1 in {
1108 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1109 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1110 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1111 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1112 VR128:$src, imm:$cc))]>;
1113 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1114 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1115 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1116 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1117 (load addr:$src), imm:$cc))]>;
1120 let Defs = [EFLAGS] in {
1121 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1122 "ucomisd\t{$src2, $src1|$src1, $src2}",
1123 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1124 (implicit EFLAGS)]>;
1125 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1126 "ucomisd\t{$src2, $src1|$src1, $src2}",
1127 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1128 (implicit EFLAGS)]>;
1130 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1131 "comisd\t{$src2, $src1|$src1, $src2}",
1132 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1133 (implicit EFLAGS)]>;
1134 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1135 "comisd\t{$src2, $src1|$src1, $src2}",
1136 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1137 (implicit EFLAGS)]>;
1140 // Aliases of packed SSE2 instructions for scalar use. These all have names that
1143 // Alias instructions that map fld0 to pxor for sse.
1144 let isReMaterializable = 1, neverHasSideEffects = 1 in
1145 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1146 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1147 Requires<[HasSSE2]>, TB, OpSize;
1149 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1151 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1152 "movapd\t{$src, $dst|$dst, $src}", []>;
1154 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1156 let isSimpleLoad = 1 in
1157 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1158 "movapd\t{$src, $dst|$dst, $src}",
1159 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1161 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1162 let isTwoAddress = 1 in {
1163 let isCommutable = 1 in {
1164 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1165 "andpd\t{$src2, $dst|$dst, $src2}",
1166 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1167 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1168 "orpd\t{$src2, $dst|$dst, $src2}",
1169 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1170 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1171 "xorpd\t{$src2, $dst|$dst, $src2}",
1172 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1175 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1176 "andpd\t{$src2, $dst|$dst, $src2}",
1177 [(set FR64:$dst, (X86fand FR64:$src1,
1178 (memopfsf64 addr:$src2)))]>;
1179 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1180 "orpd\t{$src2, $dst|$dst, $src2}",
1181 [(set FR64:$dst, (X86for FR64:$src1,
1182 (memopfsf64 addr:$src2)))]>;
1183 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1184 "xorpd\t{$src2, $dst|$dst, $src2}",
1185 [(set FR64:$dst, (X86fxor FR64:$src1,
1186 (memopfsf64 addr:$src2)))]>;
1188 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1189 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1190 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1191 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1192 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1193 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1196 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1198 /// In addition, we also have a special variant of the scalar form here to
1199 /// represent the associated intrinsic operation. This form is unlike the
1200 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1201 /// and leaves the top elements undefined.
1203 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1204 /// six "instructions".
1206 let isTwoAddress = 1 in {
1207 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1208 SDNode OpNode, Intrinsic F64Int,
1209 bit Commutable = 0> {
1210 // Scalar operation, reg+reg.
1211 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1212 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1213 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1214 let isCommutable = Commutable;
1217 // Scalar operation, reg+mem.
1218 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1219 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1220 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1222 // Vector operation, reg+reg.
1223 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1224 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1225 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1226 let isCommutable = Commutable;
1229 // Vector operation, reg+mem.
1230 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1231 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1232 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1234 // Intrinsic operation, reg+reg.
1235 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1236 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1237 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1238 let isCommutable = Commutable;
1241 // Intrinsic operation, reg+mem.
1242 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1243 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1244 [(set VR128:$dst, (F64Int VR128:$src1,
1245 sse_load_f64:$src2))]>;
1249 // Arithmetic instructions
1250 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1251 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1252 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1253 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1255 /// sse2_fp_binop_rm - Other SSE2 binops
1257 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1258 /// instructions for a full-vector intrinsic form. Operations that map
1259 /// onto C operators don't use this form since they just use the plain
1260 /// vector form instead of having a separate vector intrinsic form.
1262 /// This provides a total of eight "instructions".
1264 let isTwoAddress = 1 in {
1265 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1269 bit Commutable = 0> {
1271 // Scalar operation, reg+reg.
1272 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1273 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1274 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1275 let isCommutable = Commutable;
1278 // Scalar operation, reg+mem.
1279 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1280 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1281 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1283 // Vector operation, reg+reg.
1284 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1285 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1286 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1287 let isCommutable = Commutable;
1290 // Vector operation, reg+mem.
1291 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1292 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1293 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1295 // Intrinsic operation, reg+reg.
1296 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1297 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1298 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1299 let isCommutable = Commutable;
1302 // Intrinsic operation, reg+mem.
1303 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1304 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1305 [(set VR128:$dst, (F64Int VR128:$src1,
1306 sse_load_f64:$src2))]>;
1308 // Vector intrinsic operation, reg+reg.
1309 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1310 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1311 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1312 let isCommutable = Commutable;
1315 // Vector intrinsic operation, reg+mem.
1316 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1317 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1318 [(set VR128:$dst, (V2F64Int VR128:$src1, (load addr:$src2)))]>;
1322 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1323 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1324 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1325 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1327 //===----------------------------------------------------------------------===//
1328 // SSE packed FP Instructions
1330 // Move Instructions
1331 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1332 "movapd\t{$src, $dst|$dst, $src}", []>;
1333 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1334 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1335 "movapd\t{$src, $dst|$dst, $src}",
1336 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1338 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1339 "movapd\t{$src, $dst|$dst, $src}",
1340 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1342 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1343 "movupd\t{$src, $dst|$dst, $src}", []>;
1344 let isSimpleLoad = 1 in
1345 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1346 "movupd\t{$src, $dst|$dst, $src}",
1347 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1348 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1349 "movupd\t{$src, $dst|$dst, $src}",
1350 [(store (v2f64 VR128:$src), addr:$dst)]>;
1352 // Intrinsic forms of MOVUPD load and store
1353 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1354 "movupd\t{$src, $dst|$dst, $src}",
1355 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1356 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1357 "movupd\t{$src, $dst|$dst, $src}",
1358 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1360 let isTwoAddress = 1 in {
1361 let AddedComplexity = 20 in {
1362 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1363 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1364 "movlpd\t{$src2, $dst|$dst, $src2}",
1366 (v2f64 (vector_shuffle VR128:$src1,
1367 (scalar_to_vector (loadf64 addr:$src2)),
1368 MOVLP_shuffle_mask)))]>;
1369 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1370 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1371 "movhpd\t{$src2, $dst|$dst, $src2}",
1373 (v2f64 (vector_shuffle VR128:$src1,
1374 (scalar_to_vector (loadf64 addr:$src2)),
1375 MOVHP_shuffle_mask)))]>;
1376 } // AddedComplexity
1379 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1380 "movlpd\t{$src, $dst|$dst, $src}",
1381 [(store (f64 (vector_extract (v2f64 VR128:$src),
1382 (iPTR 0))), addr:$dst)]>;
1384 // v2f64 extract element 1 is always custom lowered to unpack high to low
1385 // and extract element 0 so the non-store version isn't too horrible.
1386 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1387 "movhpd\t{$src, $dst|$dst, $src}",
1388 [(store (f64 (vector_extract
1389 (v2f64 (vector_shuffle VR128:$src, (undef),
1390 UNPCKH_shuffle_mask)), (iPTR 0))),
1393 // SSE2 instructions without OpSize prefix
1394 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1395 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1396 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1397 TB, Requires<[HasSSE2]>;
1398 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1399 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1400 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1401 (bitconvert (memopv2i64 addr:$src))))]>,
1402 TB, Requires<[HasSSE2]>;
1404 // SSE2 instructions with XS prefix
1405 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1406 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1407 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1408 XS, Requires<[HasSSE2]>;
1409 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1410 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1411 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1412 (bitconvert (memopv2i64 addr:$src))))]>,
1413 XS, Requires<[HasSSE2]>;
1415 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1416 "cvtps2dq\t{$src, $dst|$dst, $src}",
1417 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1418 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1419 "cvtps2dq\t{$src, $dst|$dst, $src}",
1420 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1421 (load addr:$src)))]>;
1422 // SSE2 packed instructions with XS prefix
1423 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1424 "cvttps2dq\t{$src, $dst|$dst, $src}",
1425 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1426 XS, Requires<[HasSSE2]>;
1427 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1428 "cvttps2dq\t{$src, $dst|$dst, $src}",
1429 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1430 (load addr:$src)))]>,
1431 XS, Requires<[HasSSE2]>;
1433 // SSE2 packed instructions with XD prefix
1434 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1435 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1436 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1437 XD, Requires<[HasSSE2]>;
1438 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1439 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1440 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1441 (load addr:$src)))]>,
1442 XD, Requires<[HasSSE2]>;
1444 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1445 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1446 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1447 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1448 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1449 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1450 (load addr:$src)))]>;
1452 // SSE2 instructions without OpSize prefix
1453 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1454 "cvtps2pd\t{$src, $dst|$dst, $src}",
1455 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1456 TB, Requires<[HasSSE2]>;
1457 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f64mem:$src),
1458 "cvtps2pd\t{$src, $dst|$dst, $src}",
1459 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1460 (load addr:$src)))]>,
1461 TB, Requires<[HasSSE2]>;
1463 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1464 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1465 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1466 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f128mem:$src),
1467 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1468 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1469 (load addr:$src)))]>;
1471 // Match intrinsics which expect XMM operand(s).
1472 // Aliases for intrinsics
1473 let isTwoAddress = 1 in {
1474 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1475 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1476 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1477 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1479 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1480 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1481 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1482 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1483 (loadi32 addr:$src2)))]>;
1484 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1485 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1486 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1487 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1489 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1490 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1491 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1492 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1493 (load addr:$src2)))]>;
1494 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1495 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1496 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1497 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1498 VR128:$src2))]>, XS,
1499 Requires<[HasSSE2]>;
1500 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1501 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1502 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1503 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1504 (load addr:$src2)))]>, XS,
1505 Requires<[HasSSE2]>;
1510 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1512 /// In addition, we also have a special variant of the scalar form here to
1513 /// represent the associated intrinsic operation. This form is unlike the
1514 /// plain scalar form, in that it takes an entire vector (instead of a
1515 /// scalar) and leaves the top elements undefined.
1517 /// And, we have a special variant form for a full-vector intrinsic form.
1519 /// These four forms can each have a reg or a mem operand, so there are a
1520 /// total of eight "instructions".
1522 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1526 bit Commutable = 0> {
1527 // Scalar operation, reg.
1528 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1529 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1530 [(set FR64:$dst, (OpNode FR64:$src))]> {
1531 let isCommutable = Commutable;
1534 // Scalar operation, mem.
1535 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1536 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1537 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1539 // Vector operation, reg.
1540 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1541 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1542 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1543 let isCommutable = Commutable;
1546 // Vector operation, mem.
1547 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1548 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1549 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1551 // Intrinsic operation, reg.
1552 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1553 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1554 [(set VR128:$dst, (F64Int VR128:$src))]> {
1555 let isCommutable = Commutable;
1558 // Intrinsic operation, mem.
1559 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1560 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1561 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1563 // Vector intrinsic operation, reg
1564 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1565 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1566 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1567 let isCommutable = Commutable;
1570 // Vector intrinsic operation, mem
1571 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1572 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1573 [(set VR128:$dst, (V2F64Int (load addr:$src)))]>;
1577 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1578 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1580 // There is no f64 version of the reciprocal approximation instructions.
1583 let isTwoAddress = 1 in {
1584 let isCommutable = 1 in {
1585 def ANDPDrr : PDI<0x54, MRMSrcReg,
1586 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1587 "andpd\t{$src2, $dst|$dst, $src2}",
1589 (and (bc_v2i64 (v2f64 VR128:$src1)),
1590 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1591 def ORPDrr : PDI<0x56, MRMSrcReg,
1592 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1593 "orpd\t{$src2, $dst|$dst, $src2}",
1595 (or (bc_v2i64 (v2f64 VR128:$src1)),
1596 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1597 def XORPDrr : PDI<0x57, MRMSrcReg,
1598 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1599 "xorpd\t{$src2, $dst|$dst, $src2}",
1601 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1602 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1605 def ANDPDrm : PDI<0x54, MRMSrcMem,
1606 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1607 "andpd\t{$src2, $dst|$dst, $src2}",
1609 (and (bc_v2i64 (v2f64 VR128:$src1)),
1610 (memopv2i64 addr:$src2)))]>;
1611 def ORPDrm : PDI<0x56, MRMSrcMem,
1612 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1613 "orpd\t{$src2, $dst|$dst, $src2}",
1615 (or (bc_v2i64 (v2f64 VR128:$src1)),
1616 (memopv2i64 addr:$src2)))]>;
1617 def XORPDrm : PDI<0x57, MRMSrcMem,
1618 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1619 "xorpd\t{$src2, $dst|$dst, $src2}",
1621 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1622 (memopv2i64 addr:$src2)))]>;
1623 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1624 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1625 "andnpd\t{$src2, $dst|$dst, $src2}",
1627 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1628 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1629 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1630 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1631 "andnpd\t{$src2, $dst|$dst, $src2}",
1633 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1634 (memopv2i64 addr:$src2)))]>;
1637 let isTwoAddress = 1 in {
1638 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1639 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1640 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1641 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1642 VR128:$src, imm:$cc))]>;
1643 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1644 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1645 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1646 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1647 (load addr:$src), imm:$cc))]>;
1650 // Shuffle and unpack instructions
1651 let isTwoAddress = 1 in {
1652 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1653 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1654 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1655 [(set VR128:$dst, (v2f64 (vector_shuffle
1656 VR128:$src1, VR128:$src2,
1657 SHUFP_shuffle_mask:$src3)))]>;
1658 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1659 (outs VR128:$dst), (ins VR128:$src1,
1660 f128mem:$src2, i8imm:$src3),
1661 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1663 (v2f64 (vector_shuffle
1664 VR128:$src1, (memopv2f64 addr:$src2),
1665 SHUFP_shuffle_mask:$src3)))]>;
1667 let AddedComplexity = 10 in {
1668 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1669 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1670 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1672 (v2f64 (vector_shuffle
1673 VR128:$src1, VR128:$src2,
1674 UNPCKH_shuffle_mask)))]>;
1675 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1676 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1677 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1679 (v2f64 (vector_shuffle
1680 VR128:$src1, (memopv2f64 addr:$src2),
1681 UNPCKH_shuffle_mask)))]>;
1683 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1684 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1685 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1687 (v2f64 (vector_shuffle
1688 VR128:$src1, VR128:$src2,
1689 UNPCKL_shuffle_mask)))]>;
1690 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1691 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1692 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1694 (v2f64 (vector_shuffle
1695 VR128:$src1, (memopv2f64 addr:$src2),
1696 UNPCKL_shuffle_mask)))]>;
1697 } // AddedComplexity
1701 //===----------------------------------------------------------------------===//
1702 // SSE integer instructions
1704 // Move Instructions
1705 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1706 "movdqa\t{$src, $dst|$dst, $src}", []>;
1707 let isSimpleLoad = 1 in
1708 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1709 "movdqa\t{$src, $dst|$dst, $src}",
1710 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1711 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1712 "movdqa\t{$src, $dst|$dst, $src}",
1713 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1714 let isSimpleLoad = 1 in
1715 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1716 "movdqu\t{$src, $dst|$dst, $src}",
1717 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1718 XS, Requires<[HasSSE2]>;
1719 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1720 "movdqu\t{$src, $dst|$dst, $src}",
1721 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1722 XS, Requires<[HasSSE2]>;
1724 // Intrinsic forms of MOVDQU load and store
1725 let isSimpleLoad = 1 in
1726 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1727 "movdqu\t{$src, $dst|$dst, $src}",
1728 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1729 XS, Requires<[HasSSE2]>;
1730 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1731 "movdqu\t{$src, $dst|$dst, $src}",
1732 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1733 XS, Requires<[HasSSE2]>;
1735 let isTwoAddress = 1 in {
1737 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1738 bit Commutable = 0> {
1739 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1740 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1741 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1742 let isCommutable = Commutable;
1744 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1745 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1746 [(set VR128:$dst, (IntId VR128:$src1,
1747 (bitconvert (memopv2i64 addr:$src2))))]>;
1750 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1751 string OpcodeStr, Intrinsic IntId> {
1752 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1753 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1754 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1755 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1756 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1757 [(set VR128:$dst, (IntId VR128:$src1,
1758 (bitconvert (memopv2i64 addr:$src2))))]>;
1759 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1760 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1761 [(set VR128:$dst, (IntId VR128:$src1,
1762 (scalar_to_vector (i32 imm:$src2))))]>;
1766 /// PDI_binop_rm - Simple SSE2 binary operator.
1767 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1768 ValueType OpVT, bit Commutable = 0> {
1769 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1770 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1771 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1772 let isCommutable = Commutable;
1774 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1775 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1776 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1777 (bitconvert (memopv2i64 addr:$src2)))))]>;
1780 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1782 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1783 /// to collapse (bitconvert VT to VT) into its operand.
1785 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1786 bit Commutable = 0> {
1787 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1788 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1789 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1790 let isCommutable = Commutable;
1792 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1793 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1794 [(set VR128:$dst, (OpNode VR128:$src1,(memopv2i64 addr:$src2)))]>;
1799 // 128-bit Integer Arithmetic
1801 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1802 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1803 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1804 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1806 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1807 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1808 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1809 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1811 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1812 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1813 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1814 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1816 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1817 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1818 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1819 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1821 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1823 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1824 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1825 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1827 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1829 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1830 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1833 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1834 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1835 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1836 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1837 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1840 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw", int_x86_sse2_psll_w>;
1841 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld", int_x86_sse2_psll_d>;
1842 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq", int_x86_sse2_psll_q>;
1844 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw", int_x86_sse2_psrl_w>;
1845 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld", int_x86_sse2_psrl_d>;
1846 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq", int_x86_sse2_psrl_q>;
1848 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw", int_x86_sse2_psra_w>;
1849 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad", int_x86_sse2_psra_d>;
1850 // PSRAQ doesn't exist in SSE[1-3].
1852 // 128-bit logical shifts.
1853 let isTwoAddress = 1 in {
1854 def PSLLDQri : PDIi8<0x73, MRM7r,
1855 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1856 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1857 def PSRLDQri : PDIi8<0x73, MRM3r,
1858 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1859 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1860 // PSRADQri doesn't exist in SSE[1-3].
1863 let Predicates = [HasSSE2] in {
1864 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1865 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1866 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1867 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1868 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1869 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1873 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1874 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1875 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1877 let isTwoAddress = 1 in {
1878 def PANDNrr : PDI<0xDF, MRMSrcReg,
1879 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1880 "pandn\t{$src2, $dst|$dst, $src2}",
1881 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1884 def PANDNrm : PDI<0xDF, MRMSrcMem,
1885 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1886 "pandn\t{$src2, $dst|$dst, $src2}",
1887 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1888 (memopv2i64 addr:$src2))))]>;
1891 // SSE2 Integer comparison
1892 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1893 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1894 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1895 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1896 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1897 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1899 // Pack instructions
1900 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1901 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1902 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1904 // Shuffle and unpack instructions
1905 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1906 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1907 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1908 [(set VR128:$dst, (v4i32 (vector_shuffle
1909 VR128:$src1, (undef),
1910 PSHUFD_shuffle_mask:$src2)))]>;
1911 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1912 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1913 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1914 [(set VR128:$dst, (v4i32 (vector_shuffle
1915 (bc_v4i32(memopv2i64 addr:$src1)),
1917 PSHUFD_shuffle_mask:$src2)))]>;
1919 // SSE2 with ImmT == Imm8 and XS prefix.
1920 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1921 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1922 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1923 [(set VR128:$dst, (v8i16 (vector_shuffle
1924 VR128:$src1, (undef),
1925 PSHUFHW_shuffle_mask:$src2)))]>,
1926 XS, Requires<[HasSSE2]>;
1927 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
1928 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1929 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1930 [(set VR128:$dst, (v8i16 (vector_shuffle
1931 (bc_v8i16 (memopv2i64 addr:$src1)),
1933 PSHUFHW_shuffle_mask:$src2)))]>,
1934 XS, Requires<[HasSSE2]>;
1936 // SSE2 with ImmT == Imm8 and XD prefix.
1937 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
1938 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1939 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1940 [(set VR128:$dst, (v8i16 (vector_shuffle
1941 VR128:$src1, (undef),
1942 PSHUFLW_shuffle_mask:$src2)))]>,
1943 XD, Requires<[HasSSE2]>;
1944 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
1945 (outs VR128:$dst), (ins i128mem:$src1, i32i8imm:$src2),
1946 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1947 [(set VR128:$dst, (v8i16 (vector_shuffle
1948 (bc_v8i16 (memopv2i64 addr:$src1)),
1950 PSHUFLW_shuffle_mask:$src2)))]>,
1951 XD, Requires<[HasSSE2]>;
1954 let isTwoAddress = 1 in {
1955 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
1956 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1957 "punpcklbw\t{$src2, $dst|$dst, $src2}",
1959 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1960 UNPCKL_shuffle_mask)))]>;
1961 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
1962 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1963 "punpcklbw\t{$src2, $dst|$dst, $src2}",
1965 (v16i8 (vector_shuffle VR128:$src1,
1966 (bc_v16i8 (memopv2i64 addr:$src2)),
1967 UNPCKL_shuffle_mask)))]>;
1968 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
1969 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1970 "punpcklwd\t{$src2, $dst|$dst, $src2}",
1972 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1973 UNPCKL_shuffle_mask)))]>;
1974 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
1975 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1976 "punpcklwd\t{$src2, $dst|$dst, $src2}",
1978 (v8i16 (vector_shuffle VR128:$src1,
1979 (bc_v8i16 (memopv2i64 addr:$src2)),
1980 UNPCKL_shuffle_mask)))]>;
1981 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
1982 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1983 "punpckldq\t{$src2, $dst|$dst, $src2}",
1985 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1986 UNPCKL_shuffle_mask)))]>;
1987 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
1988 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1989 "punpckldq\t{$src2, $dst|$dst, $src2}",
1991 (v4i32 (vector_shuffle VR128:$src1,
1992 (bc_v4i32 (memopv2i64 addr:$src2)),
1993 UNPCKL_shuffle_mask)))]>;
1994 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
1995 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1996 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
1998 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1999 UNPCKL_shuffle_mask)))]>;
2000 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2001 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2002 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2004 (v2i64 (vector_shuffle VR128:$src1,
2005 (memopv2i64 addr:$src2),
2006 UNPCKL_shuffle_mask)))]>;
2008 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2009 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2010 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2012 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
2013 UNPCKH_shuffle_mask)))]>;
2014 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2015 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2016 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2018 (v16i8 (vector_shuffle VR128:$src1,
2019 (bc_v16i8 (memopv2i64 addr:$src2)),
2020 UNPCKH_shuffle_mask)))]>;
2021 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2022 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2023 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2025 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2026 UNPCKH_shuffle_mask)))]>;
2027 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2028 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2029 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2031 (v8i16 (vector_shuffle VR128:$src1,
2032 (bc_v8i16 (memopv2i64 addr:$src2)),
2033 UNPCKH_shuffle_mask)))]>;
2034 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2035 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2036 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2038 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2039 UNPCKH_shuffle_mask)))]>;
2040 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2041 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2042 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2044 (v4i32 (vector_shuffle VR128:$src1,
2045 (bc_v4i32 (memopv2i64 addr:$src2)),
2046 UNPCKH_shuffle_mask)))]>;
2047 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2048 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2049 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2051 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2052 UNPCKH_shuffle_mask)))]>;
2053 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2054 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2055 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2057 (v2i64 (vector_shuffle VR128:$src1,
2058 (memopv2i64 addr:$src2),
2059 UNPCKH_shuffle_mask)))]>;
2063 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2064 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2065 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2066 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2067 (iPTR imm:$src2)))]>;
2068 let isTwoAddress = 1 in {
2069 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2070 (outs VR128:$dst), (ins VR128:$src1,
2071 GR32:$src2, i32i8imm:$src3),
2072 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2074 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
2075 GR32:$src2, (iPTR imm:$src3))))]>;
2076 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2077 (outs VR128:$dst), (ins VR128:$src1,
2078 i16mem:$src2, i32i8imm:$src3),
2079 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2081 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
2082 (i32 (anyext (loadi16 addr:$src2))),
2083 (iPTR imm:$src3))))]>;
2087 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2088 "pmovmskb\t{$src, $dst|$dst, $src}",
2089 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2091 // Conditional store
2093 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2094 "maskmovdqu\t{$mask, $src|$src, $mask}",
2095 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2097 // Non-temporal stores
2098 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2099 "movntpd\t{$src, $dst|$dst, $src}",
2100 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2101 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2102 "movntdq\t{$src, $dst|$dst, $src}",
2103 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2104 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2105 "movnti\t{$src, $dst|$dst, $src}",
2106 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2107 TB, Requires<[HasSSE2]>;
2110 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2111 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2112 TB, Requires<[HasSSE2]>;
2114 // Load, store, and memory fence
2115 def LFENCE : I<0xAE, MRM5m, (outs), (ins),
2116 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2117 def MFENCE : I<0xAE, MRM6m, (outs), (ins),
2118 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2121 // Alias instructions that map zero vector to pxor / xorp* for sse.
2122 let isReMaterializable = 1, neverHasSideEffects = 1 in
2123 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2124 "pcmpeqd\t$dst, $dst",
2125 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2127 // FR64 to 128-bit vector conversion.
2128 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2129 "movsd\t{$src, $dst|$dst, $src}",
2131 (v2f64 (scalar_to_vector FR64:$src)))]>;
2132 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2133 "movsd\t{$src, $dst|$dst, $src}",
2135 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2137 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2138 "movd\t{$src, $dst|$dst, $src}",
2140 (v4i32 (scalar_to_vector GR32:$src)))]>;
2141 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2142 "movd\t{$src, $dst|$dst, $src}",
2144 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2146 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2147 "movd\t{$src, $dst|$dst, $src}",
2148 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2150 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2151 "movd\t{$src, $dst|$dst, $src}",
2152 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2154 // SSE2 instructions with XS prefix
2155 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2156 "movq\t{$src, $dst|$dst, $src}",
2158 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2159 Requires<[HasSSE2]>;
2160 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2161 "movq\t{$src, $dst|$dst, $src}",
2162 [(store (i64 (vector_extract (v2i64 VR128:$src),
2163 (iPTR 0))), addr:$dst)]>;
2165 // FIXME: may not be able to eliminate this movss with coalescing the src and
2166 // dest register classes are different. We really want to write this pattern
2168 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2169 // (f32 FR32:$src)>;
2170 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2171 "movsd\t{$src, $dst|$dst, $src}",
2172 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2174 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2175 "movsd\t{$src, $dst|$dst, $src}",
2176 [(store (f64 (vector_extract (v2f64 VR128:$src),
2177 (iPTR 0))), addr:$dst)]>;
2178 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2179 "movd\t{$src, $dst|$dst, $src}",
2180 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2182 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2183 "movd\t{$src, $dst|$dst, $src}",
2184 [(store (i32 (vector_extract (v4i32 VR128:$src),
2185 (iPTR 0))), addr:$dst)]>;
2187 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2188 "movd\t{$src, $dst|$dst, $src}",
2189 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2190 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2191 "movd\t{$src, $dst|$dst, $src}",
2192 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2195 // Move to lower bits of a VR128, leaving upper bits alone.
2196 // Three operand (but two address) aliases.
2197 let isTwoAddress = 1 in {
2198 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2199 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2200 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2202 let AddedComplexity = 15 in
2203 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2204 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2205 "movsd\t{$src2, $dst|$dst, $src2}",
2207 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
2208 MOVL_shuffle_mask)))]>;
2211 // Store / copy lower 64-bits of a XMM register.
2212 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2213 "movq\t{$src, $dst|$dst, $src}",
2214 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2216 // Move to lower bits of a VR128 and zeroing upper bits.
2217 // Loading from memory automatically zeroing upper bits.
2218 let AddedComplexity = 20 in
2219 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2220 "movsd\t{$src, $dst|$dst, $src}",
2222 (v2f64 (vector_shuffle immAllZerosV_bc,
2223 (v2f64 (scalar_to_vector
2224 (loadf64 addr:$src))),
2225 MOVL_shuffle_mask)))]>;
2227 // movd / movq to XMM register zero-extends
2228 let AddedComplexity = 15 in {
2229 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2230 "movd\t{$src, $dst|$dst, $src}",
2232 (v4i32 (vector_shuffle immAllZerosV,
2233 (v4i32 (scalar_to_vector GR32:$src)),
2234 MOVL_shuffle_mask)))]>;
2235 // This is X86-64 only.
2236 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2237 "mov{d|q}\t{$src, $dst|$dst, $src}",
2239 (v2i64 (vector_shuffle immAllZerosV_bc,
2240 (v2i64 (scalar_to_vector GR64:$src)),
2241 MOVL_shuffle_mask)))]>;
2244 let AddedComplexity = 20 in {
2245 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2246 "movd\t{$src, $dst|$dst, $src}",
2248 (v4i32 (vector_shuffle immAllZerosV,
2249 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
2250 MOVL_shuffle_mask)))]>;
2251 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2252 "movq\t{$src, $dst|$dst, $src}",
2254 (v2i64 (vector_shuffle immAllZerosV_bc,
2255 (v2i64 (scalar_to_vector (loadi64 addr:$src))),
2256 MOVL_shuffle_mask)))]>, XS,
2257 Requires<[HasSSE2]>;
2260 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2261 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2262 let AddedComplexity = 15 in
2263 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2264 "movq\t{$src, $dst|$dst, $src}",
2265 [(set VR128:$dst, (v2i64 (vector_shuffle immAllZerosV_bc,
2267 MOVL_shuffle_mask)))]>,
2268 XS, Requires<[HasSSE2]>;
2270 let AddedComplexity = 20 in
2271 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2272 "movq\t{$src, $dst|$dst, $src}",
2273 [(set VR128:$dst, (v2i64 (vector_shuffle immAllZerosV_bc,
2274 (memopv2i64 addr:$src),
2275 MOVL_shuffle_mask)))]>,
2276 XS, Requires<[HasSSE2]>;
2278 //===----------------------------------------------------------------------===//
2279 // SSE3 Instructions
2280 //===----------------------------------------------------------------------===//
2282 // Move Instructions
2283 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2284 "movshdup\t{$src, $dst|$dst, $src}",
2285 [(set VR128:$dst, (v4f32 (vector_shuffle
2286 VR128:$src, (undef),
2287 MOVSHDUP_shuffle_mask)))]>;
2288 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2289 "movshdup\t{$src, $dst|$dst, $src}",
2290 [(set VR128:$dst, (v4f32 (vector_shuffle
2291 (memopv4f32 addr:$src), (undef),
2292 MOVSHDUP_shuffle_mask)))]>;
2294 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2295 "movsldup\t{$src, $dst|$dst, $src}",
2296 [(set VR128:$dst, (v4f32 (vector_shuffle
2297 VR128:$src, (undef),
2298 MOVSLDUP_shuffle_mask)))]>;
2299 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2300 "movsldup\t{$src, $dst|$dst, $src}",
2301 [(set VR128:$dst, (v4f32 (vector_shuffle
2302 (memopv4f32 addr:$src), (undef),
2303 MOVSLDUP_shuffle_mask)))]>;
2305 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2306 "movddup\t{$src, $dst|$dst, $src}",
2307 [(set VR128:$dst, (v2f64 (vector_shuffle
2308 VR128:$src, (undef),
2309 SSE_splat_lo_mask)))]>;
2310 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2311 "movddup\t{$src, $dst|$dst, $src}",
2313 (v2f64 (vector_shuffle
2314 (scalar_to_vector (loadf64 addr:$src)),
2316 SSE_splat_lo_mask)))]>;
2319 let isTwoAddress = 1 in {
2320 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2321 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2322 "addsubps\t{$src2, $dst|$dst, $src2}",
2323 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2325 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2326 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2327 "addsubps\t{$src2, $dst|$dst, $src2}",
2328 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2329 (load addr:$src2)))]>;
2330 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2331 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2332 "addsubpd\t{$src2, $dst|$dst, $src2}",
2333 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2335 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2336 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2337 "addsubpd\t{$src2, $dst|$dst, $src2}",
2338 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2339 (load addr:$src2)))]>;
2342 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2343 "lddqu\t{$src, $dst|$dst, $src}",
2344 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2347 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2348 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2349 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2350 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2351 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2352 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2353 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2354 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
2355 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2356 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2357 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2358 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2359 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2360 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2361 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2362 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
2364 let isTwoAddress = 1 in {
2365 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2366 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2367 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2368 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2369 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2370 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2371 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2372 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2375 // Thread synchronization
2376 def MONITOR : I<0xC8, RawFrm, (outs), (ins), "monitor",
2377 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2378 def MWAIT : I<0xC9, RawFrm, (outs), (ins), "mwait",
2379 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2381 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2382 let AddedComplexity = 15 in
2383 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2384 MOVSHDUP_shuffle_mask)),
2385 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2386 let AddedComplexity = 20 in
2387 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2388 MOVSHDUP_shuffle_mask)),
2389 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2391 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2392 let AddedComplexity = 15 in
2393 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2394 MOVSLDUP_shuffle_mask)),
2395 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2396 let AddedComplexity = 20 in
2397 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2398 MOVSLDUP_shuffle_mask)),
2399 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2401 //===----------------------------------------------------------------------===//
2402 // SSSE3 Instructions
2403 //===----------------------------------------------------------------------===//
2405 // SSSE3 Instruction Templates:
2407 // SS38I - SSSE3 instructions with T8 prefix.
2408 // SS3AI - SSSE3 instructions with TA prefix.
2410 // Note: SSSE3 instructions have 64-bit and 128-bit versions. The 64-bit version
2411 // uses the MMX registers. We put those instructions here because they better
2412 // fit into the SSSE3 instruction category rather than the MMX category.
2414 class SS38I<bits<8> o, Format F, dag outs, dag ins, string asm,
2416 : I<o, F, outs, ins, asm, pattern>, T8, Requires<[HasSSSE3]>;
2417 class SS3AI<bits<8> o, Format F, dag outs, dag ins, string asm,
2419 : I<o, F, outs, ins, asm, pattern>, TA, Requires<[HasSSSE3]>;
2421 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2422 let isTwoAddress = 1 in {
2423 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2424 Intrinsic IntId64, Intrinsic IntId128,
2425 bit Commutable = 0> {
2426 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2427 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2428 [(set VR64:$dst, (IntId64 VR64:$src))]> {
2429 let isCommutable = Commutable;
2431 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2432 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2434 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2436 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2438 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2439 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2441 let isCommutable = Commutable;
2443 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2445 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2448 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2452 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2453 let isTwoAddress = 1 in {
2454 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2455 Intrinsic IntId64, Intrinsic IntId128,
2456 bit Commutable = 0> {
2457 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2459 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2460 [(set VR64:$dst, (IntId64 VR64:$src))]> {
2461 let isCommutable = Commutable;
2463 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2465 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2468 (bitconvert (memopv4i16 addr:$src))))]>;
2470 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2472 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2473 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2475 let isCommutable = Commutable;
2477 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2479 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2482 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2486 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2487 let isTwoAddress = 1 in {
2488 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2489 Intrinsic IntId64, Intrinsic IntId128,
2490 bit Commutable = 0> {
2491 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2493 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2494 [(set VR64:$dst, (IntId64 VR64:$src))]> {
2495 let isCommutable = Commutable;
2497 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2499 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2502 (bitconvert (memopv2i32 addr:$src))))]>;
2504 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2506 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2507 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2509 let isCommutable = Commutable;
2511 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2513 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2516 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2520 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2521 int_x86_ssse3_pabs_b,
2522 int_x86_ssse3_pabs_b_128>;
2523 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2524 int_x86_ssse3_pabs_w,
2525 int_x86_ssse3_pabs_w_128>;
2526 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2527 int_x86_ssse3_pabs_d,
2528 int_x86_ssse3_pabs_d_128>;
2530 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2531 let isTwoAddress = 1 in {
2532 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2533 Intrinsic IntId64, Intrinsic IntId128,
2534 bit Commutable = 0> {
2535 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2536 (ins VR64:$src1, VR64:$src2),
2537 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2538 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2539 let isCommutable = Commutable;
2541 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2542 (ins VR64:$src1, i64mem:$src2),
2543 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2545 (IntId64 VR64:$src1,
2546 (bitconvert (memopv8i8 addr:$src2))))]>;
2548 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2549 (ins VR128:$src1, VR128:$src2),
2550 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2551 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2553 let isCommutable = Commutable;
2555 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2556 (ins VR128:$src1, i128mem:$src2),
2557 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2559 (IntId128 VR128:$src1,
2560 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2564 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2565 let isTwoAddress = 1 in {
2566 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2567 Intrinsic IntId64, Intrinsic IntId128,
2568 bit Commutable = 0> {
2569 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2570 (ins VR64:$src1, VR64:$src2),
2571 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2572 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2573 let isCommutable = Commutable;
2575 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2576 (ins VR64:$src1, i64mem:$src2),
2577 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2579 (IntId64 VR64:$src1,
2580 (bitconvert (memopv4i16 addr:$src2))))]>;
2582 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2583 (ins VR128:$src1, VR128:$src2),
2584 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2585 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2587 let isCommutable = Commutable;
2589 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2590 (ins VR128:$src1, i128mem:$src2),
2591 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2593 (IntId128 VR128:$src1,
2594 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2598 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2599 let isTwoAddress = 1 in {
2600 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2601 Intrinsic IntId64, Intrinsic IntId128,
2602 bit Commutable = 0> {
2603 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2604 (ins VR64:$src1, VR64:$src2),
2605 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2606 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2607 let isCommutable = Commutable;
2609 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2610 (ins VR64:$src1, i64mem:$src2),
2611 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2613 (IntId64 VR64:$src1,
2614 (bitconvert (memopv2i32 addr:$src2))))]>;
2616 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2617 (ins VR128:$src1, VR128:$src2),
2618 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2619 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2621 let isCommutable = Commutable;
2623 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2624 (ins VR128:$src1, i128mem:$src2),
2625 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2627 (IntId128 VR128:$src1,
2628 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2632 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2633 int_x86_ssse3_phadd_w,
2634 int_x86_ssse3_phadd_w_128, 1>;
2635 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2636 int_x86_ssse3_phadd_d,
2637 int_x86_ssse3_phadd_d_128, 1>;
2638 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2639 int_x86_ssse3_phadd_sw,
2640 int_x86_ssse3_phadd_sw_128, 1>;
2641 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2642 int_x86_ssse3_phsub_w,
2643 int_x86_ssse3_phsub_w_128>;
2644 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2645 int_x86_ssse3_phsub_d,
2646 int_x86_ssse3_phsub_d_128>;
2647 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2648 int_x86_ssse3_phsub_sw,
2649 int_x86_ssse3_phsub_sw_128>;
2650 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2651 int_x86_ssse3_pmadd_ub_sw,
2652 int_x86_ssse3_pmadd_ub_sw_128, 1>;
2653 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2654 int_x86_ssse3_pmul_hr_sw,
2655 int_x86_ssse3_pmul_hr_sw_128, 1>;
2656 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2657 int_x86_ssse3_pshuf_b,
2658 int_x86_ssse3_pshuf_b_128>;
2659 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2660 int_x86_ssse3_psign_b,
2661 int_x86_ssse3_psign_b_128>;
2662 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2663 int_x86_ssse3_psign_w,
2664 int_x86_ssse3_psign_w_128>;
2665 defm PSIGND : SS3I_binop_rm_int_32<0x09, "psignd",
2666 int_x86_ssse3_psign_d,
2667 int_x86_ssse3_psign_d_128>;
2669 let isTwoAddress = 1 in {
2670 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2671 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2672 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2674 (int_x86_ssse3_palign_r
2675 VR64:$src1, VR64:$src2,
2677 def PALIGNR64rm : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2678 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2679 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2681 (int_x86_ssse3_palign_r
2683 (bitconvert (memopv2i32 addr:$src2)),
2686 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2687 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2688 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2690 (int_x86_ssse3_palign_r_128
2691 VR128:$src1, VR128:$src2,
2692 imm:$src3))]>, OpSize;
2693 def PALIGNR128rm : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2694 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2695 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2697 (int_x86_ssse3_palign_r_128
2699 (bitconvert (memopv4i32 addr:$src2)),
2700 imm:$src3))]>, OpSize;
2703 //===----------------------------------------------------------------------===//
2704 // Non-Instruction Patterns
2705 //===----------------------------------------------------------------------===//
2707 // 128-bit vector undef's.
2708 def : Pat<(v4f32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2709 def : Pat<(v2f64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2710 def : Pat<(v16i8 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2711 def : Pat<(v8i16 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2712 def : Pat<(v4i32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2713 def : Pat<(v2i64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2715 // Scalar to v8i16 / v16i8. The source may be a GR32, but only the lower 8 or
2717 def : Pat<(v8i16 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2718 Requires<[HasSSE2]>;
2719 def : Pat<(v16i8 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2720 Requires<[HasSSE2]>;
2723 let Predicates = [HasSSE2] in {
2724 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2725 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2726 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2727 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2728 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2729 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2730 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2731 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2732 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2733 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2734 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2735 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2736 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2737 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2738 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2739 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2740 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2741 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2742 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2743 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2744 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2745 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2746 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2747 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2748 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2749 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2750 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2751 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2752 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2753 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2756 // Move scalar to XMM zero-extended
2757 // movd to XMM register zero-extends
2758 let AddedComplexity = 15 in {
2759 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2760 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc,
2761 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
2762 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2763 def : Pat<(v4f32 (vector_shuffle immAllZerosV_bc,
2764 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
2765 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
2768 // Splat v2f64 / v2i64
2769 let AddedComplexity = 10 in {
2770 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2771 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2772 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2773 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2774 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2775 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2776 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2777 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2781 def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SSE_splat_mask:$sm),
2782 (SHUFPSrri VR128:$src, VR128:$src, SSE_splat_mask:$sm)>,
2783 Requires<[HasSSE1]>;
2785 // Special unary SHUFPSrri case.
2786 // FIXME: when we want non two-address code, then we should use PSHUFD?
2787 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2788 SHUFP_unary_shuffle_mask:$sm)),
2789 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2790 Requires<[HasSSE1]>;
2791 // Special unary SHUFPDrri case.
2792 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (undef),
2793 SHUFP_unary_shuffle_mask:$sm)),
2794 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2795 Requires<[HasSSE2]>;
2796 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2797 def : Pat<(vector_shuffle (memopv4f32 addr:$src1), (undef),
2798 SHUFP_unary_shuffle_mask:$sm),
2799 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2800 Requires<[HasSSE2]>;
2801 // Special binary v4i32 shuffle cases with SHUFPS.
2802 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (v4i32 VR128:$src2),
2803 PSHUFD_binary_shuffle_mask:$sm)),
2804 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2805 Requires<[HasSSE2]>;
2806 def : Pat<(v4i32 (vector_shuffle VR128:$src1,
2807 (bc_v4i32 (memopv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm)),
2808 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2809 Requires<[HasSSE2]>;
2810 // Special binary v2i64 shuffle cases using SHUFPDrri.
2811 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2812 SHUFP_shuffle_mask:$sm)),
2813 (SHUFPDrri VR128:$src1, VR128:$src2, SHUFP_shuffle_mask:$sm)>,
2814 Requires<[HasSSE2]>;
2815 // Special unary SHUFPDrri case.
2816 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (undef),
2817 SHUFP_unary_shuffle_mask:$sm)),
2818 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2819 Requires<[HasSSE2]>;
2821 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2822 let AddedComplexity = 10 in {
2823 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2824 UNPCKL_v_undef_shuffle_mask)),
2825 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2826 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2827 UNPCKL_v_undef_shuffle_mask)),
2828 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2829 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2830 UNPCKL_v_undef_shuffle_mask)),
2831 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2832 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2833 UNPCKL_v_undef_shuffle_mask)),
2834 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2837 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2838 let AddedComplexity = 10 in {
2839 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2840 UNPCKH_v_undef_shuffle_mask)),
2841 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2842 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2843 UNPCKH_v_undef_shuffle_mask)),
2844 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2845 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2846 UNPCKH_v_undef_shuffle_mask)),
2847 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2848 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2849 UNPCKH_v_undef_shuffle_mask)),
2850 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2853 let AddedComplexity = 15 in {
2854 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2855 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2856 MOVHP_shuffle_mask)),
2857 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2859 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2860 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2861 MOVHLPS_shuffle_mask)),
2862 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2864 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2865 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2866 MOVHLPS_v_undef_shuffle_mask)),
2867 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2868 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2869 MOVHLPS_v_undef_shuffle_mask)),
2870 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2873 let AddedComplexity = 20 in {
2874 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2875 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2876 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2877 MOVLP_shuffle_mask)),
2878 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2879 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2880 MOVLP_shuffle_mask)),
2881 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2882 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2883 MOVHP_shuffle_mask)),
2884 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2885 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2886 MOVHP_shuffle_mask)),
2887 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2889 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2890 MOVLP_shuffle_mask)),
2891 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2892 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2893 MOVLP_shuffle_mask)),
2894 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2895 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2896 MOVHP_shuffle_mask)),
2897 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2898 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2899 MOVLP_shuffle_mask)),
2900 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2903 let AddedComplexity = 15 in {
2904 // Setting the lowest element in the vector.
2905 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2906 MOVL_shuffle_mask)),
2907 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2908 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2909 MOVL_shuffle_mask)),
2910 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2912 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2913 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2914 MOVLP_shuffle_mask)),
2915 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2916 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2917 MOVLP_shuffle_mask)),
2918 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2921 // Set lowest element and zero upper elements.
2922 let AddedComplexity = 15 in
2923 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc, VR128:$src,
2924 MOVL_shuffle_mask)),
2925 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
2928 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2929 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2930 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2931 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2932 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2933 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2934 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2935 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2936 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2937 Requires<[HasSSE2]>;
2938 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2939 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2940 Requires<[HasSSE2]>;
2941 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2942 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2943 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2944 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2945 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2946 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2947 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2948 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2949 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
2950 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2951 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
2952 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2953 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
2954 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2955 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
2956 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2958 // Some special case pandn patterns.
2959 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2961 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2962 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2964 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2965 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2967 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2969 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2970 (memopv2i64 addr:$src2))),
2971 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2972 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2973 (memopv2i64 addr:$src2))),
2974 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2975 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2976 (memopv2i64 addr:$src2))),
2977 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2979 // vector -> vector casts
2980 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
2981 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
2982 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
2983 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
2985 // Use movaps / movups for SSE integer load / store (one byte shorter).
2986 def : Pat<(alignedloadv4i32 addr:$src),
2987 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
2988 def : Pat<(loadv4i32 addr:$src),
2989 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
2990 def : Pat<(alignedloadv2i64 addr:$src),
2991 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
2992 def : Pat<(loadv2i64 addr:$src),
2993 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
2995 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
2996 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2997 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
2998 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2999 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3000 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3001 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3002 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3003 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3004 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3005 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3006 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3007 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3008 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3009 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3010 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;