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]>;
58 //===----------------------------------------------------------------------===//
59 // SSE Complex Patterns
60 //===----------------------------------------------------------------------===//
62 // These are 'extloads' from a scalar to the low element of a vector, zeroing
63 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
65 def sse_load_f32 : ComplexPattern<v4f32, 4, "SelectScalarSSELoad", [],
66 [SDNPHasChain, SDNPMayLoad]>;
67 def sse_load_f64 : ComplexPattern<v2f64, 4, "SelectScalarSSELoad", [],
68 [SDNPHasChain, SDNPMayLoad]>;
70 def ssmem : Operand<v4f32> {
71 let PrintMethod = "printf32mem";
72 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
74 def sdmem : Operand<v2f64> {
75 let PrintMethod = "printf64mem";
76 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
79 //===----------------------------------------------------------------------===//
80 // SSE pattern fragments
81 //===----------------------------------------------------------------------===//
83 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
84 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
85 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
86 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
88 // Like 'store', but always requires vector alignment.
89 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
90 (st node:$val, node:$ptr), [{
91 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
92 return !ST->isTruncatingStore() &&
93 ST->getAddressingMode() == ISD::UNINDEXED &&
94 ST->getAlignment() >= 16;
98 // Like 'load', but always requires vector alignment.
99 def alignedload : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
100 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
101 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
102 LD->getAddressingMode() == ISD::UNINDEXED &&
103 LD->getAlignment() >= 16;
107 def alignedloadfsf32 : PatFrag<(ops node:$ptr), (f32 (alignedload node:$ptr))>;
108 def alignedloadfsf64 : PatFrag<(ops node:$ptr), (f64 (alignedload node:$ptr))>;
109 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
110 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
111 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
112 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
114 // Like 'load', but uses special alignment checks suitable for use in
115 // memory operands in most SSE instructions, which are required to
116 // be naturally aligned on some targets but not on others.
117 // FIXME: Actually implement support for targets that don't require the
118 // alignment. This probably wants a subtarget predicate.
119 def memop : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
120 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
121 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
122 LD->getAddressingMode() == ISD::UNINDEXED &&
123 LD->getAlignment() >= 16;
127 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
128 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
129 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
130 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
131 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
132 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
134 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
136 def memop64 : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
137 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
138 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
139 LD->getAddressingMode() == ISD::UNINDEXED &&
140 LD->getAlignment() >= 8;
144 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
145 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop64 node:$ptr))>;
146 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
147 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
148 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
150 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
151 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
152 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
153 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
154 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
155 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
157 def fp32imm0 : PatLeaf<(f32 fpimm), [{
158 return N->isExactlyValue(+0.0);
161 def PSxLDQ_imm : SDNodeXForm<imm, [{
162 // Transformation function: imm >> 3
163 return getI32Imm(N->getValue() >> 3);
166 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
168 def SHUFFLE_get_shuf_imm : SDNodeXForm<build_vector, [{
169 return getI8Imm(X86::getShuffleSHUFImmediate(N));
172 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
174 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<build_vector, [{
175 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
178 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
180 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<build_vector, [{
181 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
184 def SSE_splat_mask : PatLeaf<(build_vector), [{
185 return X86::isSplatMask(N);
186 }], SHUFFLE_get_shuf_imm>;
188 def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
189 return X86::isSplatLoMask(N);
192 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
193 return X86::isMOVHLPSMask(N);
196 def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
197 return X86::isMOVHLPS_v_undef_Mask(N);
200 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
201 return X86::isMOVHPMask(N);
204 def MOVLP_shuffle_mask : PatLeaf<(build_vector), [{
205 return X86::isMOVLPMask(N);
208 def MOVL_shuffle_mask : PatLeaf<(build_vector), [{
209 return X86::isMOVLMask(N);
212 def MOVSHDUP_shuffle_mask : PatLeaf<(build_vector), [{
213 return X86::isMOVSHDUPMask(N);
216 def MOVSLDUP_shuffle_mask : PatLeaf<(build_vector), [{
217 return X86::isMOVSLDUPMask(N);
220 def UNPCKL_shuffle_mask : PatLeaf<(build_vector), [{
221 return X86::isUNPCKLMask(N);
224 def UNPCKH_shuffle_mask : PatLeaf<(build_vector), [{
225 return X86::isUNPCKHMask(N);
228 def UNPCKL_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
229 return X86::isUNPCKL_v_undef_Mask(N);
232 def UNPCKH_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
233 return X86::isUNPCKH_v_undef_Mask(N);
236 def PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
237 return X86::isPSHUFDMask(N);
238 }], SHUFFLE_get_shuf_imm>;
240 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
241 return X86::isPSHUFHWMask(N);
242 }], SHUFFLE_get_pshufhw_imm>;
244 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
245 return X86::isPSHUFLWMask(N);
246 }], SHUFFLE_get_pshuflw_imm>;
248 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
249 return X86::isPSHUFDMask(N);
250 }], SHUFFLE_get_shuf_imm>;
252 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
253 return X86::isSHUFPMask(N);
254 }], SHUFFLE_get_shuf_imm>;
256 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
257 return X86::isSHUFPMask(N);
258 }], SHUFFLE_get_shuf_imm>;
260 //===----------------------------------------------------------------------===//
261 // SSE scalar FP Instructions
262 //===----------------------------------------------------------------------===//
264 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
265 // scheduler into a branch sequence.
266 // These are expanded by the scheduler.
267 let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
268 def CMOV_FR32 : I<0, Pseudo,
269 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
270 "#CMOV_FR32 PSEUDO!",
271 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
273 def CMOV_FR64 : I<0, Pseudo,
274 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
275 "#CMOV_FR64 PSEUDO!",
276 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
278 def CMOV_V4F32 : I<0, Pseudo,
279 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
280 "#CMOV_V4F32 PSEUDO!",
282 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
284 def CMOV_V2F64 : I<0, Pseudo,
285 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
286 "#CMOV_V2F64 PSEUDO!",
288 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
290 def CMOV_V2I64 : I<0, Pseudo,
291 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
292 "#CMOV_V2I64 PSEUDO!",
294 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
298 //===----------------------------------------------------------------------===//
300 //===----------------------------------------------------------------------===//
303 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
304 "movss\t{$src, $dst|$dst, $src}", []>;
305 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
306 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
307 "movss\t{$src, $dst|$dst, $src}",
308 [(set FR32:$dst, (loadf32 addr:$src))]>;
309 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
310 "movss\t{$src, $dst|$dst, $src}",
311 [(store FR32:$src, addr:$dst)]>;
313 // Conversion instructions
314 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
315 "cvttss2si\t{$src, $dst|$dst, $src}",
316 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
317 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
318 "cvttss2si\t{$src, $dst|$dst, $src}",
319 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
320 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
321 "cvtsi2ss\t{$src, $dst|$dst, $src}",
322 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
323 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
324 "cvtsi2ss\t{$src, $dst|$dst, $src}",
325 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
327 // Match intrinsics which expect XMM operand(s).
328 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
329 "cvtss2si\t{$src, $dst|$dst, $src}",
330 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
331 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
332 "cvtss2si\t{$src, $dst|$dst, $src}",
333 [(set GR32:$dst, (int_x86_sse_cvtss2si
334 (load addr:$src)))]>;
336 // Match intrinisics which expect MM and XMM operand(s).
337 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
338 "cvtps2pi\t{$src, $dst|$dst, $src}",
339 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
340 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
341 "cvtps2pi\t{$src, $dst|$dst, $src}",
342 [(set VR64:$dst, (int_x86_sse_cvtps2pi
343 (load addr:$src)))]>;
344 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
345 "cvttps2pi\t{$src, $dst|$dst, $src}",
346 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
347 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
348 "cvttps2pi\t{$src, $dst|$dst, $src}",
349 [(set VR64:$dst, (int_x86_sse_cvttps2pi
350 (load addr:$src)))]>;
351 let isTwoAddress = 1 in {
352 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
353 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
354 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
355 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
357 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
358 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
359 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
360 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
361 (load addr:$src2)))]>;
364 // Aliases for intrinsics
365 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
366 "cvttss2si\t{$src, $dst|$dst, $src}",
368 (int_x86_sse_cvttss2si VR128:$src))]>;
369 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
370 "cvttss2si\t{$src, $dst|$dst, $src}",
372 (int_x86_sse_cvttss2si(load addr:$src)))]>;
374 let isTwoAddress = 1 in {
375 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
376 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
377 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
378 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
380 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
381 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
382 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
383 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
384 (loadi32 addr:$src2)))]>;
387 // Comparison instructions
388 let isTwoAddress = 1 in {
389 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
390 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
391 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
392 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
393 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
394 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
397 let Defs = [EFLAGS] in {
398 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
399 "ucomiss\t{$src2, $src1|$src1, $src2}",
400 [(X86cmp FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
401 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
402 "ucomiss\t{$src2, $src1|$src1, $src2}",
403 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
407 // Aliases to match intrinsics which expect XMM operand(s).
408 let isTwoAddress = 1 in {
409 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
410 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
411 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
412 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
413 VR128:$src, imm:$cc))]>;
414 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
415 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src, SSECC:$cc),
416 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
417 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
418 (load addr:$src), imm:$cc))]>;
421 let Defs = [EFLAGS] in {
422 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs),
423 (ins VR128:$src1, VR128:$src2),
424 "ucomiss\t{$src2, $src1|$src1, $src2}",
425 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
427 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),
428 (ins VR128:$src1, f128mem:$src2),
429 "ucomiss\t{$src2, $src1|$src1, $src2}",
430 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
433 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs),
434 (ins VR128:$src1, VR128:$src2),
435 "comiss\t{$src2, $src1|$src1, $src2}",
436 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
438 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs),
439 (ins VR128:$src1, f128mem:$src2),
440 "comiss\t{$src2, $src1|$src1, $src2}",
441 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
445 // Aliases of packed SSE1 instructions for scalar use. These all have names that
448 // Alias instructions that map fld0 to pxor for sse.
449 let isReMaterializable = 1 in
450 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
451 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
452 Requires<[HasSSE1]>, TB, OpSize;
454 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
456 let neverHasSideEffects = 1 in
457 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
458 "movaps\t{$src, $dst|$dst, $src}", []>;
460 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
462 let isSimpleLoad = 1 in
463 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
464 "movaps\t{$src, $dst|$dst, $src}",
465 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
467 // Alias bitwise logical operations using SSE logical ops on packed FP values.
468 let isTwoAddress = 1 in {
469 let isCommutable = 1 in {
470 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
471 "andps\t{$src2, $dst|$dst, $src2}",
472 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
473 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
474 "orps\t{$src2, $dst|$dst, $src2}",
475 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
476 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
477 "xorps\t{$src2, $dst|$dst, $src2}",
478 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
481 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
482 "andps\t{$src2, $dst|$dst, $src2}",
483 [(set FR32:$dst, (X86fand FR32:$src1,
484 (memopfsf32 addr:$src2)))]>;
485 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
486 "orps\t{$src2, $dst|$dst, $src2}",
487 [(set FR32:$dst, (X86for FR32:$src1,
488 (memopfsf32 addr:$src2)))]>;
489 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
490 "xorps\t{$src2, $dst|$dst, $src2}",
491 [(set FR32:$dst, (X86fxor FR32:$src1,
492 (memopfsf32 addr:$src2)))]>;
493 let neverHasSideEffects = 1 in {
494 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
495 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
496 "andnps\t{$src2, $dst|$dst, $src2}", []>;
499 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
500 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
501 "andnps\t{$src2, $dst|$dst, $src2}", []>;
505 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
507 /// In addition, we also have a special variant of the scalar form here to
508 /// represent the associated intrinsic operation. This form is unlike the
509 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
510 /// and leaves the top elements undefined.
512 /// These three forms can each be reg+reg or reg+mem, so there are a total of
513 /// six "instructions".
515 let isTwoAddress = 1 in {
516 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
517 SDNode OpNode, Intrinsic F32Int,
518 bit Commutable = 0> {
519 // Scalar operation, reg+reg.
520 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
521 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
522 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
523 let isCommutable = Commutable;
526 // Scalar operation, reg+mem.
527 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
528 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
529 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
531 // Vector operation, reg+reg.
532 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
533 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
534 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
535 let isCommutable = Commutable;
538 // Vector operation, reg+mem.
539 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
540 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
541 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
543 // Intrinsic operation, reg+reg.
544 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
545 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
546 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
547 let isCommutable = Commutable;
550 // Intrinsic operation, reg+mem.
551 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
552 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
553 [(set VR128:$dst, (F32Int VR128:$src1,
554 sse_load_f32:$src2))]>;
558 // Arithmetic instructions
559 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
560 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
561 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
562 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
564 /// sse1_fp_binop_rm - Other SSE1 binops
566 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
567 /// instructions for a full-vector intrinsic form. Operations that map
568 /// onto C operators don't use this form since they just use the plain
569 /// vector form instead of having a separate vector intrinsic form.
571 /// This provides a total of eight "instructions".
573 let isTwoAddress = 1 in {
574 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
578 bit Commutable = 0> {
580 // Scalar operation, reg+reg.
581 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
582 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
583 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
584 let isCommutable = Commutable;
587 // Scalar operation, reg+mem.
588 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
589 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
590 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
592 // Vector operation, reg+reg.
593 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
594 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
595 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
596 let isCommutable = Commutable;
599 // Vector operation, reg+mem.
600 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
601 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
602 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
604 // Intrinsic operation, reg+reg.
605 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
606 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
607 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
608 let isCommutable = Commutable;
611 // Intrinsic operation, reg+mem.
612 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
613 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
614 [(set VR128:$dst, (F32Int VR128:$src1,
615 sse_load_f32:$src2))]>;
617 // Vector intrinsic operation, reg+reg.
618 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
619 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
620 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
621 let isCommutable = Commutable;
624 // Vector intrinsic operation, reg+mem.
625 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
626 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
627 [(set VR128:$dst, (V4F32Int VR128:$src1, (load addr:$src2)))]>;
631 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
632 int_x86_sse_max_ss, int_x86_sse_max_ps>;
633 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
634 int_x86_sse_min_ss, int_x86_sse_min_ps>;
636 //===----------------------------------------------------------------------===//
637 // SSE packed FP Instructions
640 let neverHasSideEffects = 1 in
641 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
642 "movaps\t{$src, $dst|$dst, $src}", []>;
643 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
644 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
645 "movaps\t{$src, $dst|$dst, $src}",
646 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
648 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
649 "movaps\t{$src, $dst|$dst, $src}",
650 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
652 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
653 "movups\t{$src, $dst|$dst, $src}", []>;
654 let isSimpleLoad = 1 in
655 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
656 "movups\t{$src, $dst|$dst, $src}",
657 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
658 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
659 "movups\t{$src, $dst|$dst, $src}",
660 [(store (v4f32 VR128:$src), addr:$dst)]>;
662 // Intrinsic forms of MOVUPS load and store
663 let isSimpleLoad = 1 in
664 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
665 "movups\t{$src, $dst|$dst, $src}",
666 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
667 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
668 "movups\t{$src, $dst|$dst, $src}",
669 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
671 let isTwoAddress = 1 in {
672 let AddedComplexity = 20 in {
673 def MOVLPSrm : PSI<0x12, MRMSrcMem,
674 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
675 "movlps\t{$src2, $dst|$dst, $src2}",
677 (v4f32 (vector_shuffle VR128:$src1,
678 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
679 MOVLP_shuffle_mask)))]>;
680 def MOVHPSrm : PSI<0x16, MRMSrcMem,
681 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
682 "movhps\t{$src2, $dst|$dst, $src2}",
684 (v4f32 (vector_shuffle VR128:$src1,
685 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
686 MOVHP_shuffle_mask)))]>;
690 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
691 "movlps\t{$src, $dst|$dst, $src}",
692 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
693 (iPTR 0))), addr:$dst)]>;
695 // v2f64 extract element 1 is always custom lowered to unpack high to low
696 // and extract element 0 so the non-store version isn't too horrible.
697 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
698 "movhps\t{$src, $dst|$dst, $src}",
699 [(store (f64 (vector_extract
700 (v2f64 (vector_shuffle
701 (bc_v2f64 (v4f32 VR128:$src)), (undef),
702 UNPCKH_shuffle_mask)), (iPTR 0))),
705 let isTwoAddress = 1 in {
706 let AddedComplexity = 15 in {
707 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
708 "movlhps\t{$src2, $dst|$dst, $src2}",
710 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
711 MOVHP_shuffle_mask)))]>;
713 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
714 "movhlps\t{$src2, $dst|$dst, $src2}",
716 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
717 MOVHLPS_shuffle_mask)))]>;
725 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
727 /// In addition, we also have a special variant of the scalar form here to
728 /// represent the associated intrinsic operation. This form is unlike the
729 /// plain scalar form, in that it takes an entire vector (instead of a
730 /// scalar) and leaves the top elements undefined.
732 /// And, we have a special variant form for a full-vector intrinsic form.
734 /// These four forms can each have a reg or a mem operand, so there are a
735 /// total of eight "instructions".
737 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
741 bit Commutable = 0> {
742 // Scalar operation, reg.
743 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
744 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
745 [(set FR32:$dst, (OpNode FR32:$src))]> {
746 let isCommutable = Commutable;
749 // Scalar operation, mem.
750 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
751 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
752 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
754 // Vector operation, reg.
755 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
756 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
757 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
758 let isCommutable = Commutable;
761 // Vector operation, mem.
762 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
763 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
764 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
766 // Intrinsic operation, reg.
767 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
768 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
769 [(set VR128:$dst, (F32Int VR128:$src))]> {
770 let isCommutable = Commutable;
773 // Intrinsic operation, mem.
774 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
775 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
776 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
778 // Vector intrinsic operation, reg
779 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
780 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
781 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
782 let isCommutable = Commutable;
785 // Vector intrinsic operation, mem
786 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
787 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
788 [(set VR128:$dst, (V4F32Int (load addr:$src)))]>;
792 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
793 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
795 // Reciprocal approximations. Note that these typically require refinement
796 // in order to obtain suitable precision.
797 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
798 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
799 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
800 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
803 let isTwoAddress = 1 in {
804 let isCommutable = 1 in {
805 def ANDPSrr : PSI<0x54, MRMSrcReg,
806 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
807 "andps\t{$src2, $dst|$dst, $src2}",
808 [(set VR128:$dst, (v2i64
809 (and VR128:$src1, VR128:$src2)))]>;
810 def ORPSrr : PSI<0x56, MRMSrcReg,
811 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
812 "orps\t{$src2, $dst|$dst, $src2}",
813 [(set VR128:$dst, (v2i64
814 (or VR128:$src1, VR128:$src2)))]>;
815 def XORPSrr : PSI<0x57, MRMSrcReg,
816 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
817 "xorps\t{$src2, $dst|$dst, $src2}",
818 [(set VR128:$dst, (v2i64
819 (xor VR128:$src1, VR128:$src2)))]>;
822 def ANDPSrm : PSI<0x54, MRMSrcMem,
823 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
824 "andps\t{$src2, $dst|$dst, $src2}",
825 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
826 (memopv2i64 addr:$src2)))]>;
827 def ORPSrm : PSI<0x56, MRMSrcMem,
828 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
829 "orps\t{$src2, $dst|$dst, $src2}",
830 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
831 (memopv2i64 addr:$src2)))]>;
832 def XORPSrm : PSI<0x57, MRMSrcMem,
833 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
834 "xorps\t{$src2, $dst|$dst, $src2}",
835 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
836 (memopv2i64 addr:$src2)))]>;
837 def ANDNPSrr : PSI<0x55, MRMSrcReg,
838 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
839 "andnps\t{$src2, $dst|$dst, $src2}",
841 (v2i64 (and (xor VR128:$src1,
842 (bc_v2i64 (v4i32 immAllOnesV))),
844 def ANDNPSrm : PSI<0x55, MRMSrcMem,
845 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
846 "andnps\t{$src2, $dst|$dst, $src2}",
848 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
849 (bc_v2i64 (v4i32 immAllOnesV))),
850 (memopv2i64 addr:$src2))))]>;
853 let isTwoAddress = 1 in {
854 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
855 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
856 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
857 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
858 VR128:$src, imm:$cc))]>;
859 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
860 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
861 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
862 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
863 (load addr:$src), imm:$cc))]>;
866 // Shuffle and unpack instructions
867 let isTwoAddress = 1 in {
868 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
869 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
870 (outs VR128:$dst), (ins VR128:$src1,
871 VR128:$src2, i32i8imm:$src3),
872 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
874 (v4f32 (vector_shuffle
875 VR128:$src1, VR128:$src2,
876 SHUFP_shuffle_mask:$src3)))]>;
877 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
878 (outs VR128:$dst), (ins VR128:$src1,
879 f128mem:$src2, i32i8imm:$src3),
880 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
882 (v4f32 (vector_shuffle
883 VR128:$src1, (memopv4f32 addr:$src2),
884 SHUFP_shuffle_mask:$src3)))]>;
886 let AddedComplexity = 10 in {
887 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
888 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
889 "unpckhps\t{$src2, $dst|$dst, $src2}",
891 (v4f32 (vector_shuffle
892 VR128:$src1, VR128:$src2,
893 UNPCKH_shuffle_mask)))]>;
894 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
895 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
896 "unpckhps\t{$src2, $dst|$dst, $src2}",
898 (v4f32 (vector_shuffle
899 VR128:$src1, (memopv4f32 addr:$src2),
900 UNPCKH_shuffle_mask)))]>;
902 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
903 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
904 "unpcklps\t{$src2, $dst|$dst, $src2}",
906 (v4f32 (vector_shuffle
907 VR128:$src1, VR128:$src2,
908 UNPCKL_shuffle_mask)))]>;
909 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
910 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
911 "unpcklps\t{$src2, $dst|$dst, $src2}",
913 (v4f32 (vector_shuffle
914 VR128:$src1, (memopv4f32 addr:$src2),
915 UNPCKL_shuffle_mask)))]>;
920 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
921 "movmskps\t{$src, $dst|$dst, $src}",
922 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
923 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
924 "movmskpd\t{$src, $dst|$dst, $src}",
925 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
927 // Prefetching loads.
928 // TODO: no intrinsics for these?
929 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src), "prefetcht0\t$src", []>;
930 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src), "prefetcht1\t$src", []>;
931 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src), "prefetcht2\t$src", []>;
932 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src), "prefetchnta\t$src", []>;
934 // Non-temporal stores
935 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
936 "movntps\t{$src, $dst|$dst, $src}",
937 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
939 // Load, store, and memory fence
940 def SFENCE : PSI<0xAE, MRM7m, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
943 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
944 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
945 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
946 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
948 // Alias instructions that map zero vector to pxor / xorp* for sse.
949 let isReMaterializable = 1 in
950 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
952 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
954 // FR32 to 128-bit vector conversion.
955 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
956 "movss\t{$src, $dst|$dst, $src}",
958 (v4f32 (scalar_to_vector FR32:$src)))]>;
959 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
960 "movss\t{$src, $dst|$dst, $src}",
962 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
964 // FIXME: may not be able to eliminate this movss with coalescing the src and
965 // dest register classes are different. We really want to write this pattern
967 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
969 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
970 "movss\t{$src, $dst|$dst, $src}",
971 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
973 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
974 "movss\t{$src, $dst|$dst, $src}",
975 [(store (f32 (vector_extract (v4f32 VR128:$src),
976 (iPTR 0))), addr:$dst)]>;
979 // Move to lower bits of a VR128, leaving upper bits alone.
980 // Three operand (but two address) aliases.
981 let isTwoAddress = 1 in {
982 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
983 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
984 "movss\t{$src2, $dst|$dst, $src2}", []>;
986 let AddedComplexity = 15 in
987 def MOVLPSrr : SSI<0x10, MRMSrcReg,
988 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
989 "movss\t{$src2, $dst|$dst, $src2}",
991 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
992 MOVL_shuffle_mask)))]>;
995 // Move to lower bits of a VR128 and zeroing upper bits.
996 // Loading from memory automatically zeroing upper bits.
997 let AddedComplexity = 20 in
998 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
999 "movss\t{$src, $dst|$dst, $src}",
1000 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV_bc,
1001 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
1002 MOVL_shuffle_mask)))]>;
1005 //===----------------------------------------------------------------------===//
1006 // SSE2 Instructions
1007 //===----------------------------------------------------------------------===//
1009 // Move Instructions
1010 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1011 "movsd\t{$src, $dst|$dst, $src}", []>;
1012 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1013 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1014 "movsd\t{$src, $dst|$dst, $src}",
1015 [(set FR64:$dst, (loadf64 addr:$src))]>;
1016 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1017 "movsd\t{$src, $dst|$dst, $src}",
1018 [(store FR64:$src, addr:$dst)]>;
1020 // Conversion instructions
1021 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1022 "cvttsd2si\t{$src, $dst|$dst, $src}",
1023 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1024 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1025 "cvttsd2si\t{$src, $dst|$dst, $src}",
1026 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1027 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1028 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1029 [(set FR32:$dst, (fround FR64:$src))]>;
1030 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1031 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1032 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1033 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1034 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1035 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1036 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1037 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1038 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1040 // SSE2 instructions with XS prefix
1041 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1042 "cvtss2sd\t{$src, $dst|$dst, $src}",
1043 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1044 Requires<[HasSSE2]>;
1045 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1046 "cvtss2sd\t{$src, $dst|$dst, $src}",
1047 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1048 Requires<[HasSSE2]>;
1050 // Match intrinsics which expect XMM operand(s).
1051 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1052 "cvtsd2si\t{$src, $dst|$dst, $src}",
1053 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1054 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1055 "cvtsd2si\t{$src, $dst|$dst, $src}",
1056 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1057 (load addr:$src)))]>;
1059 // Match intrinisics which expect MM and XMM operand(s).
1060 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1061 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1062 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1063 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1064 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1065 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1066 (load addr:$src)))]>;
1067 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1068 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1069 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1070 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1071 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1072 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1073 (load addr:$src)))]>;
1074 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1075 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1076 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1077 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1078 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1079 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1080 (load addr:$src)))]>;
1082 // Aliases for intrinsics
1083 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1084 "cvttsd2si\t{$src, $dst|$dst, $src}",
1086 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1087 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1088 "cvttsd2si\t{$src, $dst|$dst, $src}",
1089 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1090 (load addr:$src)))]>;
1092 // Comparison instructions
1093 let isTwoAddress = 1 in {
1094 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1095 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1096 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1097 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1098 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1099 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1102 let Defs = [EFLAGS] in {
1103 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1104 "ucomisd\t{$src2, $src1|$src1, $src2}",
1105 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1106 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1107 "ucomisd\t{$src2, $src1|$src1, $src2}",
1108 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1109 (implicit EFLAGS)]>;
1112 // Aliases to match intrinsics which expect XMM operand(s).
1113 let isTwoAddress = 1 in {
1114 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1115 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1116 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1117 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1118 VR128:$src, imm:$cc))]>;
1119 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1120 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1121 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1122 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1123 (load addr:$src), imm:$cc))]>;
1126 let Defs = [EFLAGS] in {
1127 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1128 "ucomisd\t{$src2, $src1|$src1, $src2}",
1129 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1130 (implicit EFLAGS)]>;
1131 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1132 "ucomisd\t{$src2, $src1|$src1, $src2}",
1133 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1134 (implicit EFLAGS)]>;
1136 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1137 "comisd\t{$src2, $src1|$src1, $src2}",
1138 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1139 (implicit EFLAGS)]>;
1140 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1141 "comisd\t{$src2, $src1|$src1, $src2}",
1142 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1143 (implicit EFLAGS)]>;
1146 // Aliases of packed SSE2 instructions for scalar use. These all have names that
1149 // Alias instructions that map fld0 to pxor for sse.
1150 let isReMaterializable = 1 in
1151 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1152 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1153 Requires<[HasSSE2]>, TB, OpSize;
1155 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1157 let neverHasSideEffects = 1 in
1158 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1159 "movapd\t{$src, $dst|$dst, $src}", []>;
1161 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1163 let isSimpleLoad = 1 in
1164 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1165 "movapd\t{$src, $dst|$dst, $src}",
1166 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1168 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1169 let isTwoAddress = 1 in {
1170 let isCommutable = 1 in {
1171 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1172 "andpd\t{$src2, $dst|$dst, $src2}",
1173 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1174 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1175 "orpd\t{$src2, $dst|$dst, $src2}",
1176 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1177 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1178 "xorpd\t{$src2, $dst|$dst, $src2}",
1179 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1182 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1183 "andpd\t{$src2, $dst|$dst, $src2}",
1184 [(set FR64:$dst, (X86fand FR64:$src1,
1185 (memopfsf64 addr:$src2)))]>;
1186 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1187 "orpd\t{$src2, $dst|$dst, $src2}",
1188 [(set FR64:$dst, (X86for FR64:$src1,
1189 (memopfsf64 addr:$src2)))]>;
1190 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1191 "xorpd\t{$src2, $dst|$dst, $src2}",
1192 [(set FR64:$dst, (X86fxor FR64:$src1,
1193 (memopfsf64 addr:$src2)))]>;
1195 let neverHasSideEffects = 1 in {
1196 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1197 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1198 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1200 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1201 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1202 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1206 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1208 /// In addition, we also have a special variant of the scalar form here to
1209 /// represent the associated intrinsic operation. This form is unlike the
1210 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1211 /// and leaves the top elements undefined.
1213 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1214 /// six "instructions".
1216 let isTwoAddress = 1 in {
1217 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1218 SDNode OpNode, Intrinsic F64Int,
1219 bit Commutable = 0> {
1220 // Scalar operation, reg+reg.
1221 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1222 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1223 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1224 let isCommutable = Commutable;
1227 // Scalar operation, reg+mem.
1228 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1229 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1230 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1232 // Vector operation, reg+reg.
1233 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1234 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1235 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1236 let isCommutable = Commutable;
1239 // Vector operation, reg+mem.
1240 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1241 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1242 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1244 // Intrinsic operation, reg+reg.
1245 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1246 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1247 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1248 let isCommutable = Commutable;
1251 // Intrinsic operation, reg+mem.
1252 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1253 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1254 [(set VR128:$dst, (F64Int VR128:$src1,
1255 sse_load_f64:$src2))]>;
1259 // Arithmetic instructions
1260 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1261 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1262 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1263 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1265 /// sse2_fp_binop_rm - Other SSE2 binops
1267 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1268 /// instructions for a full-vector intrinsic form. Operations that map
1269 /// onto C operators don't use this form since they just use the plain
1270 /// vector form instead of having a separate vector intrinsic form.
1272 /// This provides a total of eight "instructions".
1274 let isTwoAddress = 1 in {
1275 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1279 bit Commutable = 0> {
1281 // Scalar operation, reg+reg.
1282 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1283 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1284 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1285 let isCommutable = Commutable;
1288 // Scalar operation, reg+mem.
1289 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1290 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1291 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1293 // Vector operation, reg+reg.
1294 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1295 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1296 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1297 let isCommutable = Commutable;
1300 // Vector operation, reg+mem.
1301 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1302 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1303 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1305 // Intrinsic operation, reg+reg.
1306 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1307 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1308 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1309 let isCommutable = Commutable;
1312 // Intrinsic operation, reg+mem.
1313 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1314 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1315 [(set VR128:$dst, (F64Int VR128:$src1,
1316 sse_load_f64:$src2))]>;
1318 // Vector intrinsic operation, reg+reg.
1319 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1320 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1321 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1322 let isCommutable = Commutable;
1325 // Vector intrinsic operation, reg+mem.
1326 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1327 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1328 [(set VR128:$dst, (V2F64Int VR128:$src1, (load addr:$src2)))]>;
1332 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1333 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1334 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1335 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1337 //===----------------------------------------------------------------------===//
1338 // SSE packed FP Instructions
1340 // Move Instructions
1341 let neverHasSideEffects = 1 in
1342 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1343 "movapd\t{$src, $dst|$dst, $src}", []>;
1344 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1345 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1346 "movapd\t{$src, $dst|$dst, $src}",
1347 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1349 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1350 "movapd\t{$src, $dst|$dst, $src}",
1351 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1353 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1354 "movupd\t{$src, $dst|$dst, $src}", []>;
1355 let isSimpleLoad = 1 in
1356 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1357 "movupd\t{$src, $dst|$dst, $src}",
1358 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1359 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1360 "movupd\t{$src, $dst|$dst, $src}",
1361 [(store (v2f64 VR128:$src), addr:$dst)]>;
1363 // Intrinsic forms of MOVUPD load and store
1364 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1365 "movupd\t{$src, $dst|$dst, $src}",
1366 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1367 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1368 "movupd\t{$src, $dst|$dst, $src}",
1369 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1371 let isTwoAddress = 1 in {
1372 let AddedComplexity = 20 in {
1373 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1374 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1375 "movlpd\t{$src2, $dst|$dst, $src2}",
1377 (v2f64 (vector_shuffle VR128:$src1,
1378 (scalar_to_vector (loadf64 addr:$src2)),
1379 MOVLP_shuffle_mask)))]>;
1380 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1381 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1382 "movhpd\t{$src2, $dst|$dst, $src2}",
1384 (v2f64 (vector_shuffle VR128:$src1,
1385 (scalar_to_vector (loadf64 addr:$src2)),
1386 MOVHP_shuffle_mask)))]>;
1387 } // AddedComplexity
1390 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1391 "movlpd\t{$src, $dst|$dst, $src}",
1392 [(store (f64 (vector_extract (v2f64 VR128:$src),
1393 (iPTR 0))), addr:$dst)]>;
1395 // v2f64 extract element 1 is always custom lowered to unpack high to low
1396 // and extract element 0 so the non-store version isn't too horrible.
1397 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1398 "movhpd\t{$src, $dst|$dst, $src}",
1399 [(store (f64 (vector_extract
1400 (v2f64 (vector_shuffle VR128:$src, (undef),
1401 UNPCKH_shuffle_mask)), (iPTR 0))),
1404 // SSE2 instructions without OpSize prefix
1405 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1406 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1407 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1408 TB, Requires<[HasSSE2]>;
1409 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1410 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1411 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1412 (bitconvert (memopv2i64 addr:$src))))]>,
1413 TB, Requires<[HasSSE2]>;
1415 // SSE2 instructions with XS prefix
1416 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1417 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1418 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1419 XS, Requires<[HasSSE2]>;
1420 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1421 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1422 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1423 (bitconvert (memopv2i64 addr:$src))))]>,
1424 XS, Requires<[HasSSE2]>;
1426 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1427 "cvtps2dq\t{$src, $dst|$dst, $src}",
1428 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1429 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1430 "cvtps2dq\t{$src, $dst|$dst, $src}",
1431 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1432 (load addr:$src)))]>;
1433 // SSE2 packed instructions with XS prefix
1434 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1435 "cvttps2dq\t{$src, $dst|$dst, $src}",
1436 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1437 XS, Requires<[HasSSE2]>;
1438 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1439 "cvttps2dq\t{$src, $dst|$dst, $src}",
1440 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1441 (load addr:$src)))]>,
1442 XS, Requires<[HasSSE2]>;
1444 // SSE2 packed instructions with XD prefix
1445 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1446 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1447 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1448 XD, Requires<[HasSSE2]>;
1449 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1450 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1451 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1452 (load addr:$src)))]>,
1453 XD, Requires<[HasSSE2]>;
1455 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1456 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1457 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1458 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1459 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1460 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1461 (load addr:$src)))]>;
1463 // SSE2 instructions without OpSize prefix
1464 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1465 "cvtps2pd\t{$src, $dst|$dst, $src}",
1466 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1467 TB, Requires<[HasSSE2]>;
1468 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f64mem:$src),
1469 "cvtps2pd\t{$src, $dst|$dst, $src}",
1470 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1471 (load addr:$src)))]>,
1472 TB, Requires<[HasSSE2]>;
1474 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1475 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1476 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1477 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f128mem:$src),
1478 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1479 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1480 (load addr:$src)))]>;
1482 // Match intrinsics which expect XMM operand(s).
1483 // Aliases for intrinsics
1484 let isTwoAddress = 1 in {
1485 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1486 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1487 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1488 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1490 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1491 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1492 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1493 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1494 (loadi32 addr:$src2)))]>;
1495 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1496 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1497 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1498 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1500 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1501 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1502 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1503 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1504 (load addr:$src2)))]>;
1505 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1506 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1507 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1508 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1509 VR128:$src2))]>, XS,
1510 Requires<[HasSSE2]>;
1511 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1512 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1513 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1514 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1515 (load addr:$src2)))]>, XS,
1516 Requires<[HasSSE2]>;
1521 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1523 /// In addition, we also have a special variant of the scalar form here to
1524 /// represent the associated intrinsic operation. This form is unlike the
1525 /// plain scalar form, in that it takes an entire vector (instead of a
1526 /// scalar) and leaves the top elements undefined.
1528 /// And, we have a special variant form for a full-vector intrinsic form.
1530 /// These four forms can each have a reg or a mem operand, so there are a
1531 /// total of eight "instructions".
1533 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1537 bit Commutable = 0> {
1538 // Scalar operation, reg.
1539 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1540 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1541 [(set FR64:$dst, (OpNode FR64:$src))]> {
1542 let isCommutable = Commutable;
1545 // Scalar operation, mem.
1546 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1547 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1548 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1550 // Vector operation, reg.
1551 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1552 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1553 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1554 let isCommutable = Commutable;
1557 // Vector operation, mem.
1558 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1559 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1560 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1562 // Intrinsic operation, reg.
1563 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1564 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1565 [(set VR128:$dst, (F64Int VR128:$src))]> {
1566 let isCommutable = Commutable;
1569 // Intrinsic operation, mem.
1570 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1571 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1572 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1574 // Vector intrinsic operation, reg
1575 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1576 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1577 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1578 let isCommutable = Commutable;
1581 // Vector intrinsic operation, mem
1582 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1583 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1584 [(set VR128:$dst, (V2F64Int (load addr:$src)))]>;
1588 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1589 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1591 // There is no f64 version of the reciprocal approximation instructions.
1594 let isTwoAddress = 1 in {
1595 let isCommutable = 1 in {
1596 def ANDPDrr : PDI<0x54, MRMSrcReg,
1597 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1598 "andpd\t{$src2, $dst|$dst, $src2}",
1600 (and (bc_v2i64 (v2f64 VR128:$src1)),
1601 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1602 def ORPDrr : PDI<0x56, MRMSrcReg,
1603 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1604 "orpd\t{$src2, $dst|$dst, $src2}",
1606 (or (bc_v2i64 (v2f64 VR128:$src1)),
1607 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1608 def XORPDrr : PDI<0x57, MRMSrcReg,
1609 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1610 "xorpd\t{$src2, $dst|$dst, $src2}",
1612 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1613 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1616 def ANDPDrm : PDI<0x54, MRMSrcMem,
1617 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1618 "andpd\t{$src2, $dst|$dst, $src2}",
1620 (and (bc_v2i64 (v2f64 VR128:$src1)),
1621 (memopv2i64 addr:$src2)))]>;
1622 def ORPDrm : PDI<0x56, MRMSrcMem,
1623 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1624 "orpd\t{$src2, $dst|$dst, $src2}",
1626 (or (bc_v2i64 (v2f64 VR128:$src1)),
1627 (memopv2i64 addr:$src2)))]>;
1628 def XORPDrm : PDI<0x57, MRMSrcMem,
1629 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1630 "xorpd\t{$src2, $dst|$dst, $src2}",
1632 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1633 (memopv2i64 addr:$src2)))]>;
1634 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1635 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1636 "andnpd\t{$src2, $dst|$dst, $src2}",
1638 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1639 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1640 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1641 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1642 "andnpd\t{$src2, $dst|$dst, $src2}",
1644 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1645 (memopv2i64 addr:$src2)))]>;
1648 let isTwoAddress = 1 in {
1649 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1650 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1651 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1652 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1653 VR128:$src, imm:$cc))]>;
1654 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1655 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1656 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1657 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1658 (load addr:$src), imm:$cc))]>;
1661 // Shuffle and unpack instructions
1662 let isTwoAddress = 1 in {
1663 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1664 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1665 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1666 [(set VR128:$dst, (v2f64 (vector_shuffle
1667 VR128:$src1, VR128:$src2,
1668 SHUFP_shuffle_mask:$src3)))]>;
1669 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1670 (outs VR128:$dst), (ins VR128:$src1,
1671 f128mem:$src2, i8imm:$src3),
1672 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1674 (v2f64 (vector_shuffle
1675 VR128:$src1, (memopv2f64 addr:$src2),
1676 SHUFP_shuffle_mask:$src3)))]>;
1678 let AddedComplexity = 10 in {
1679 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1680 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1681 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1683 (v2f64 (vector_shuffle
1684 VR128:$src1, VR128:$src2,
1685 UNPCKH_shuffle_mask)))]>;
1686 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1687 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1688 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1690 (v2f64 (vector_shuffle
1691 VR128:$src1, (memopv2f64 addr:$src2),
1692 UNPCKH_shuffle_mask)))]>;
1694 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1695 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1696 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1698 (v2f64 (vector_shuffle
1699 VR128:$src1, VR128:$src2,
1700 UNPCKL_shuffle_mask)))]>;
1701 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1702 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1703 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1705 (v2f64 (vector_shuffle
1706 VR128:$src1, (memopv2f64 addr:$src2),
1707 UNPCKL_shuffle_mask)))]>;
1708 } // AddedComplexity
1712 //===----------------------------------------------------------------------===//
1713 // SSE integer instructions
1715 // Move Instructions
1716 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1717 "movdqa\t{$src, $dst|$dst, $src}", []>;
1718 let isSimpleLoad = 1 in
1719 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1720 "movdqa\t{$src, $dst|$dst, $src}",
1721 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1722 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1723 "movdqa\t{$src, $dst|$dst, $src}",
1724 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1725 let isSimpleLoad = 1 in
1726 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1727 "movdqu\t{$src, $dst|$dst, $src}",
1728 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1729 XS, Requires<[HasSSE2]>;
1730 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1731 "movdqu\t{$src, $dst|$dst, $src}",
1732 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1733 XS, Requires<[HasSSE2]>;
1735 // Intrinsic forms of MOVDQU load and store
1736 let isSimpleLoad = 1 in
1737 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1738 "movdqu\t{$src, $dst|$dst, $src}",
1739 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1740 XS, Requires<[HasSSE2]>;
1741 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1742 "movdqu\t{$src, $dst|$dst, $src}",
1743 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1744 XS, Requires<[HasSSE2]>;
1746 let isTwoAddress = 1 in {
1748 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1749 bit Commutable = 0> {
1750 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1751 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1752 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1753 let isCommutable = Commutable;
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))))]>;
1761 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1762 string OpcodeStr, Intrinsic IntId> {
1763 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1764 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1765 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1766 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1767 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1768 [(set VR128:$dst, (IntId VR128:$src1,
1769 (bitconvert (memopv2i64 addr:$src2))))]>;
1770 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1771 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1772 [(set VR128:$dst, (IntId VR128:$src1,
1773 (scalar_to_vector (i32 imm:$src2))))]>;
1777 /// PDI_binop_rm - Simple SSE2 binary operator.
1778 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1779 ValueType OpVT, bit Commutable = 0> {
1780 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1781 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1782 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1783 let isCommutable = Commutable;
1785 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1786 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1787 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1788 (bitconvert (memopv2i64 addr:$src2)))))]>;
1791 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1793 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1794 /// to collapse (bitconvert VT to VT) into its operand.
1796 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1797 bit Commutable = 0> {
1798 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1799 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1800 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1801 let isCommutable = Commutable;
1803 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1804 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1805 [(set VR128:$dst, (OpNode VR128:$src1,(memopv2i64 addr:$src2)))]>;
1810 // 128-bit Integer Arithmetic
1812 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1813 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1814 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1815 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1817 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1818 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1819 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1820 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1822 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1823 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1824 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1825 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1827 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1828 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1829 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1830 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1832 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1834 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1835 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1836 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1838 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1840 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1841 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1844 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1845 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1846 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1847 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1848 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1851 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw", int_x86_sse2_psll_w>;
1852 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld", int_x86_sse2_psll_d>;
1853 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq", int_x86_sse2_psll_q>;
1855 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw", int_x86_sse2_psrl_w>;
1856 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld", int_x86_sse2_psrl_d>;
1857 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq", int_x86_sse2_psrl_q>;
1859 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw", int_x86_sse2_psra_w>;
1860 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad", int_x86_sse2_psra_d>;
1861 // PSRAQ doesn't exist in SSE[1-3].
1863 // 128-bit logical shifts.
1864 let isTwoAddress = 1 in {
1865 def PSLLDQri : PDIi8<0x73, MRM7r,
1866 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1867 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1868 def PSRLDQri : PDIi8<0x73, MRM3r,
1869 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1870 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1871 // PSRADQri doesn't exist in SSE[1-3].
1874 let Predicates = [HasSSE2] in {
1875 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1876 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1877 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1878 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1879 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1880 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1884 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1885 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1886 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1888 let isTwoAddress = 1 in {
1889 def PANDNrr : PDI<0xDF, MRMSrcReg,
1890 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1891 "pandn\t{$src2, $dst|$dst, $src2}",
1892 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1895 def PANDNrm : PDI<0xDF, MRMSrcMem,
1896 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1897 "pandn\t{$src2, $dst|$dst, $src2}",
1898 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1899 (memopv2i64 addr:$src2))))]>;
1902 // SSE2 Integer comparison
1903 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1904 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1905 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1906 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1907 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1908 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1910 // Pack instructions
1911 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1912 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1913 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1915 // Shuffle and unpack instructions
1916 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1917 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1918 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1919 [(set VR128:$dst, (v4i32 (vector_shuffle
1920 VR128:$src1, (undef),
1921 PSHUFD_shuffle_mask:$src2)))]>;
1922 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1923 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1924 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1925 [(set VR128:$dst, (v4i32 (vector_shuffle
1926 (bc_v4i32(memopv2i64 addr:$src1)),
1928 PSHUFD_shuffle_mask:$src2)))]>;
1930 // SSE2 with ImmT == Imm8 and XS prefix.
1931 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1932 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1933 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1934 [(set VR128:$dst, (v8i16 (vector_shuffle
1935 VR128:$src1, (undef),
1936 PSHUFHW_shuffle_mask:$src2)))]>,
1937 XS, Requires<[HasSSE2]>;
1938 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
1939 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1940 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1941 [(set VR128:$dst, (v8i16 (vector_shuffle
1942 (bc_v8i16 (memopv2i64 addr:$src1)),
1944 PSHUFHW_shuffle_mask:$src2)))]>,
1945 XS, Requires<[HasSSE2]>;
1947 // SSE2 with ImmT == Imm8 and XD prefix.
1948 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
1949 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1950 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1951 [(set VR128:$dst, (v8i16 (vector_shuffle
1952 VR128:$src1, (undef),
1953 PSHUFLW_shuffle_mask:$src2)))]>,
1954 XD, Requires<[HasSSE2]>;
1955 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
1956 (outs VR128:$dst), (ins i128mem:$src1, i32i8imm:$src2),
1957 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1958 [(set VR128:$dst, (v8i16 (vector_shuffle
1959 (bc_v8i16 (memopv2i64 addr:$src1)),
1961 PSHUFLW_shuffle_mask:$src2)))]>,
1962 XD, Requires<[HasSSE2]>;
1965 let isTwoAddress = 1 in {
1966 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
1967 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1968 "punpcklbw\t{$src2, $dst|$dst, $src2}",
1970 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1971 UNPCKL_shuffle_mask)))]>;
1972 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
1973 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1974 "punpcklbw\t{$src2, $dst|$dst, $src2}",
1976 (v16i8 (vector_shuffle VR128:$src1,
1977 (bc_v16i8 (memopv2i64 addr:$src2)),
1978 UNPCKL_shuffle_mask)))]>;
1979 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
1980 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1981 "punpcklwd\t{$src2, $dst|$dst, $src2}",
1983 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1984 UNPCKL_shuffle_mask)))]>;
1985 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
1986 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1987 "punpcklwd\t{$src2, $dst|$dst, $src2}",
1989 (v8i16 (vector_shuffle VR128:$src1,
1990 (bc_v8i16 (memopv2i64 addr:$src2)),
1991 UNPCKL_shuffle_mask)))]>;
1992 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
1993 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1994 "punpckldq\t{$src2, $dst|$dst, $src2}",
1996 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1997 UNPCKL_shuffle_mask)))]>;
1998 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
1999 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2000 "punpckldq\t{$src2, $dst|$dst, $src2}",
2002 (v4i32 (vector_shuffle VR128:$src1,
2003 (bc_v4i32 (memopv2i64 addr:$src2)),
2004 UNPCKL_shuffle_mask)))]>;
2005 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2006 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2007 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2009 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2010 UNPCKL_shuffle_mask)))]>;
2011 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2012 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2013 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2015 (v2i64 (vector_shuffle VR128:$src1,
2016 (memopv2i64 addr:$src2),
2017 UNPCKL_shuffle_mask)))]>;
2019 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2020 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2021 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2023 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
2024 UNPCKH_shuffle_mask)))]>;
2025 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2026 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2027 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2029 (v16i8 (vector_shuffle VR128:$src1,
2030 (bc_v16i8 (memopv2i64 addr:$src2)),
2031 UNPCKH_shuffle_mask)))]>;
2032 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2033 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2034 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2036 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2037 UNPCKH_shuffle_mask)))]>;
2038 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2039 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2040 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2042 (v8i16 (vector_shuffle VR128:$src1,
2043 (bc_v8i16 (memopv2i64 addr:$src2)),
2044 UNPCKH_shuffle_mask)))]>;
2045 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2046 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2047 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2049 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2050 UNPCKH_shuffle_mask)))]>;
2051 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2052 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2053 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2055 (v4i32 (vector_shuffle VR128:$src1,
2056 (bc_v4i32 (memopv2i64 addr:$src2)),
2057 UNPCKH_shuffle_mask)))]>;
2058 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2059 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2060 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2062 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2063 UNPCKH_shuffle_mask)))]>;
2064 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2065 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2066 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2068 (v2i64 (vector_shuffle VR128:$src1,
2069 (memopv2i64 addr:$src2),
2070 UNPCKH_shuffle_mask)))]>;
2074 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2075 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2076 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2077 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2078 (iPTR imm:$src2)))]>;
2079 let isTwoAddress = 1 in {
2080 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2081 (outs VR128:$dst), (ins VR128:$src1,
2082 GR32:$src2, i32i8imm:$src3),
2083 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2085 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
2086 GR32:$src2, (iPTR imm:$src3))))]>;
2087 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2088 (outs VR128:$dst), (ins VR128:$src1,
2089 i16mem:$src2, i32i8imm:$src3),
2090 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2092 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
2093 (i32 (anyext (loadi16 addr:$src2))),
2094 (iPTR imm:$src3))))]>;
2098 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2099 "pmovmskb\t{$src, $dst|$dst, $src}",
2100 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2102 // Conditional store
2104 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2105 "maskmovdqu\t{$mask, $src|$src, $mask}",
2106 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2108 // Non-temporal stores
2109 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2110 "movntpd\t{$src, $dst|$dst, $src}",
2111 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2112 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2113 "movntdq\t{$src, $dst|$dst, $src}",
2114 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2115 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2116 "movnti\t{$src, $dst|$dst, $src}",
2117 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2118 TB, Requires<[HasSSE2]>;
2121 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2122 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2123 TB, Requires<[HasSSE2]>;
2125 // Load, store, and memory fence
2126 def LFENCE : I<0xAE, MRM5m, (outs), (ins),
2127 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2128 def MFENCE : I<0xAE, MRM6m, (outs), (ins),
2129 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2131 // Alias instructions that map zero vector to pxor / xorp* for sse.
2132 let isReMaterializable = 1 in
2133 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2134 "pcmpeqd\t$dst, $dst",
2135 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2137 // FR64 to 128-bit vector conversion.
2138 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2139 "movsd\t{$src, $dst|$dst, $src}",
2141 (v2f64 (scalar_to_vector FR64:$src)))]>;
2142 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2143 "movsd\t{$src, $dst|$dst, $src}",
2145 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2147 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2148 "movd\t{$src, $dst|$dst, $src}",
2150 (v4i32 (scalar_to_vector GR32:$src)))]>;
2151 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2152 "movd\t{$src, $dst|$dst, $src}",
2154 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2156 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2157 "movd\t{$src, $dst|$dst, $src}",
2158 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2160 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2161 "movd\t{$src, $dst|$dst, $src}",
2162 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2164 // SSE2 instructions with XS prefix
2165 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2166 "movq\t{$src, $dst|$dst, $src}",
2168 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2169 Requires<[HasSSE2]>;
2170 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2171 "movq\t{$src, $dst|$dst, $src}",
2172 [(store (i64 (vector_extract (v2i64 VR128:$src),
2173 (iPTR 0))), addr:$dst)]>;
2175 // FIXME: may not be able to eliminate this movss with coalescing the src and
2176 // dest register classes are different. We really want to write this pattern
2178 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2179 // (f32 FR32:$src)>;
2180 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2181 "movsd\t{$src, $dst|$dst, $src}",
2182 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2184 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2185 "movsd\t{$src, $dst|$dst, $src}",
2186 [(store (f64 (vector_extract (v2f64 VR128:$src),
2187 (iPTR 0))), addr:$dst)]>;
2188 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2189 "movd\t{$src, $dst|$dst, $src}",
2190 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2192 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2193 "movd\t{$src, $dst|$dst, $src}",
2194 [(store (i32 (vector_extract (v4i32 VR128:$src),
2195 (iPTR 0))), addr:$dst)]>;
2197 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2198 "movd\t{$src, $dst|$dst, $src}",
2199 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2200 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2201 "movd\t{$src, $dst|$dst, $src}",
2202 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2205 // Move to lower bits of a VR128, leaving upper bits alone.
2206 // Three operand (but two address) aliases.
2207 let isTwoAddress = 1 in {
2208 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2209 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2210 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2212 let AddedComplexity = 15 in
2213 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2214 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2215 "movsd\t{$src2, $dst|$dst, $src2}",
2217 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
2218 MOVL_shuffle_mask)))]>;
2221 // Store / copy lower 64-bits of a XMM register.
2222 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2223 "movq\t{$src, $dst|$dst, $src}",
2224 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2226 // Move to lower bits of a VR128 and zeroing upper bits.
2227 // Loading from memory automatically zeroing upper bits.
2228 let AddedComplexity = 20 in
2229 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2230 "movsd\t{$src, $dst|$dst, $src}",
2232 (v2f64 (vector_shuffle immAllZerosV_bc,
2233 (v2f64 (scalar_to_vector
2234 (loadf64 addr:$src))),
2235 MOVL_shuffle_mask)))]>;
2237 // movd / movq to XMM register zero-extends
2238 let AddedComplexity = 15 in {
2239 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2240 "movd\t{$src, $dst|$dst, $src}",
2242 (v4i32 (vector_shuffle immAllZerosV,
2243 (v4i32 (scalar_to_vector GR32:$src)),
2244 MOVL_shuffle_mask)))]>;
2245 // This is X86-64 only.
2246 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2247 "mov{d|q}\t{$src, $dst|$dst, $src}",
2249 (v2i64 (vector_shuffle immAllZerosV_bc,
2250 (v2i64 (scalar_to_vector GR64:$src)),
2251 MOVL_shuffle_mask)))]>;
2254 let AddedComplexity = 20 in {
2255 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2256 "movd\t{$src, $dst|$dst, $src}",
2258 (v4i32 (vector_shuffle immAllZerosV,
2259 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
2260 MOVL_shuffle_mask)))]>;
2261 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2262 "movq\t{$src, $dst|$dst, $src}",
2264 (v2i64 (vector_shuffle immAllZerosV_bc,
2265 (v2i64 (scalar_to_vector (loadi64 addr:$src))),
2266 MOVL_shuffle_mask)))]>, XS,
2267 Requires<[HasSSE2]>;
2270 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2271 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2272 let AddedComplexity = 15 in
2273 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2274 "movq\t{$src, $dst|$dst, $src}",
2275 [(set VR128:$dst, (v2i64 (vector_shuffle immAllZerosV_bc,
2277 MOVL_shuffle_mask)))]>,
2278 XS, Requires<[HasSSE2]>;
2280 let AddedComplexity = 20 in
2281 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2282 "movq\t{$src, $dst|$dst, $src}",
2283 [(set VR128:$dst, (v2i64 (vector_shuffle immAllZerosV_bc,
2284 (memopv2i64 addr:$src),
2285 MOVL_shuffle_mask)))]>,
2286 XS, Requires<[HasSSE2]>;
2288 //===----------------------------------------------------------------------===//
2289 // SSE3 Instructions
2290 //===----------------------------------------------------------------------===//
2292 // Move Instructions
2293 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2294 "movshdup\t{$src, $dst|$dst, $src}",
2295 [(set VR128:$dst, (v4f32 (vector_shuffle
2296 VR128:$src, (undef),
2297 MOVSHDUP_shuffle_mask)))]>;
2298 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2299 "movshdup\t{$src, $dst|$dst, $src}",
2300 [(set VR128:$dst, (v4f32 (vector_shuffle
2301 (memopv4f32 addr:$src), (undef),
2302 MOVSHDUP_shuffle_mask)))]>;
2304 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2305 "movsldup\t{$src, $dst|$dst, $src}",
2306 [(set VR128:$dst, (v4f32 (vector_shuffle
2307 VR128:$src, (undef),
2308 MOVSLDUP_shuffle_mask)))]>;
2309 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2310 "movsldup\t{$src, $dst|$dst, $src}",
2311 [(set VR128:$dst, (v4f32 (vector_shuffle
2312 (memopv4f32 addr:$src), (undef),
2313 MOVSLDUP_shuffle_mask)))]>;
2315 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2316 "movddup\t{$src, $dst|$dst, $src}",
2317 [(set VR128:$dst, (v2f64 (vector_shuffle
2318 VR128:$src, (undef),
2319 SSE_splat_lo_mask)))]>;
2320 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2321 "movddup\t{$src, $dst|$dst, $src}",
2323 (v2f64 (vector_shuffle
2324 (scalar_to_vector (loadf64 addr:$src)),
2326 SSE_splat_lo_mask)))]>;
2329 let isTwoAddress = 1 in {
2330 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2331 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2332 "addsubps\t{$src2, $dst|$dst, $src2}",
2333 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2335 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2336 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2337 "addsubps\t{$src2, $dst|$dst, $src2}",
2338 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2339 (load addr:$src2)))]>;
2340 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2341 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2342 "addsubpd\t{$src2, $dst|$dst, $src2}",
2343 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2345 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2346 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2347 "addsubpd\t{$src2, $dst|$dst, $src2}",
2348 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2349 (load addr:$src2)))]>;
2352 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2353 "lddqu\t{$src, $dst|$dst, $src}",
2354 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2357 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2358 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2359 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2360 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2361 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2362 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2363 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2364 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
2365 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2366 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2367 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2368 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2369 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2370 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2371 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2372 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
2374 let isTwoAddress = 1 in {
2375 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2376 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2377 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2378 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2379 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2380 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2381 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2382 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2385 // Thread synchronization
2386 def MONITOR : I<0xC8, RawFrm, (outs), (ins), "monitor",
2387 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2388 def MWAIT : I<0xC9, RawFrm, (outs), (ins), "mwait",
2389 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2391 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2392 let AddedComplexity = 15 in
2393 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2394 MOVSHDUP_shuffle_mask)),
2395 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2396 let AddedComplexity = 20 in
2397 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2398 MOVSHDUP_shuffle_mask)),
2399 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2401 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2402 let AddedComplexity = 15 in
2403 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2404 MOVSLDUP_shuffle_mask)),
2405 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2406 let AddedComplexity = 20 in
2407 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2408 MOVSLDUP_shuffle_mask)),
2409 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2411 //===----------------------------------------------------------------------===//
2412 // SSSE3 Instructions
2413 //===----------------------------------------------------------------------===//
2415 // SSSE3 Instruction Templates:
2417 // SS38I - SSSE3 instructions with T8 prefix.
2418 // SS3AI - SSSE3 instructions with TA prefix.
2420 // Note: SSSE3 instructions have 64-bit and 128-bit versions. The 64-bit version
2421 // uses the MMX registers. We put those instructions here because they better
2422 // fit into the SSSE3 instruction category rather than the MMX category.
2424 class SS38I<bits<8> o, Format F, dag outs, dag ins, string asm,
2426 : I<o, F, outs, ins, asm, pattern>, T8, Requires<[HasSSSE3]>;
2427 class SS3AI<bits<8> o, Format F, dag outs, dag ins, string asm,
2429 : I<o, F, outs, ins, asm, pattern>, TA, Requires<[HasSSSE3]>;
2431 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2432 let isTwoAddress = 1 in {
2433 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2434 Intrinsic IntId64, Intrinsic IntId128,
2435 bit Commutable = 0> {
2436 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2437 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2438 [(set VR64:$dst, (IntId64 VR64:$src))]> {
2439 let isCommutable = Commutable;
2441 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2442 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2444 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2446 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2448 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2449 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2451 let isCommutable = Commutable;
2453 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2455 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2458 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2462 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2463 let isTwoAddress = 1 in {
2464 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2465 Intrinsic IntId64, Intrinsic IntId128,
2466 bit Commutable = 0> {
2467 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2469 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2470 [(set VR64:$dst, (IntId64 VR64:$src))]> {
2471 let isCommutable = Commutable;
2473 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2475 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2478 (bitconvert (memopv4i16 addr:$src))))]>;
2480 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2482 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2483 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2485 let isCommutable = Commutable;
2487 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2489 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2492 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2496 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2497 let isTwoAddress = 1 in {
2498 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2499 Intrinsic IntId64, Intrinsic IntId128,
2500 bit Commutable = 0> {
2501 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2503 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2504 [(set VR64:$dst, (IntId64 VR64:$src))]> {
2505 let isCommutable = Commutable;
2507 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2509 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2512 (bitconvert (memopv2i32 addr:$src))))]>;
2514 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2516 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2517 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2519 let isCommutable = Commutable;
2521 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2523 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2526 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2530 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2531 int_x86_ssse3_pabs_b,
2532 int_x86_ssse3_pabs_b_128>;
2533 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2534 int_x86_ssse3_pabs_w,
2535 int_x86_ssse3_pabs_w_128>;
2536 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2537 int_x86_ssse3_pabs_d,
2538 int_x86_ssse3_pabs_d_128>;
2540 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2541 let isTwoAddress = 1 in {
2542 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2543 Intrinsic IntId64, Intrinsic IntId128,
2544 bit Commutable = 0> {
2545 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2546 (ins VR64:$src1, VR64:$src2),
2547 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2548 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2549 let isCommutable = Commutable;
2551 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2552 (ins VR64:$src1, i64mem:$src2),
2553 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2555 (IntId64 VR64:$src1,
2556 (bitconvert (memopv8i8 addr:$src2))))]>;
2558 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2559 (ins VR128:$src1, VR128:$src2),
2560 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2561 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2563 let isCommutable = Commutable;
2565 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2566 (ins VR128:$src1, i128mem:$src2),
2567 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2569 (IntId128 VR128:$src1,
2570 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2574 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2575 let isTwoAddress = 1 in {
2576 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2577 Intrinsic IntId64, Intrinsic IntId128,
2578 bit Commutable = 0> {
2579 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2580 (ins VR64:$src1, VR64:$src2),
2581 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2582 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2583 let isCommutable = Commutable;
2585 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2586 (ins VR64:$src1, i64mem:$src2),
2587 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2589 (IntId64 VR64:$src1,
2590 (bitconvert (memopv4i16 addr:$src2))))]>;
2592 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2593 (ins VR128:$src1, VR128:$src2),
2594 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2595 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2597 let isCommutable = Commutable;
2599 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2600 (ins VR128:$src1, i128mem:$src2),
2601 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2603 (IntId128 VR128:$src1,
2604 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2608 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2609 let isTwoAddress = 1 in {
2610 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2611 Intrinsic IntId64, Intrinsic IntId128,
2612 bit Commutable = 0> {
2613 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2614 (ins VR64:$src1, VR64:$src2),
2615 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2616 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2617 let isCommutable = Commutable;
2619 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2620 (ins VR64:$src1, i64mem:$src2),
2621 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2623 (IntId64 VR64:$src1,
2624 (bitconvert (memopv2i32 addr:$src2))))]>;
2626 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2627 (ins VR128:$src1, VR128:$src2),
2628 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2629 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2631 let isCommutable = Commutable;
2633 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2634 (ins VR128:$src1, i128mem:$src2),
2635 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2637 (IntId128 VR128:$src1,
2638 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2642 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2643 int_x86_ssse3_phadd_w,
2644 int_x86_ssse3_phadd_w_128, 1>;
2645 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2646 int_x86_ssse3_phadd_d,
2647 int_x86_ssse3_phadd_d_128, 1>;
2648 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2649 int_x86_ssse3_phadd_sw,
2650 int_x86_ssse3_phadd_sw_128, 1>;
2651 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2652 int_x86_ssse3_phsub_w,
2653 int_x86_ssse3_phsub_w_128>;
2654 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2655 int_x86_ssse3_phsub_d,
2656 int_x86_ssse3_phsub_d_128>;
2657 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2658 int_x86_ssse3_phsub_sw,
2659 int_x86_ssse3_phsub_sw_128>;
2660 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2661 int_x86_ssse3_pmadd_ub_sw,
2662 int_x86_ssse3_pmadd_ub_sw_128, 1>;
2663 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2664 int_x86_ssse3_pmul_hr_sw,
2665 int_x86_ssse3_pmul_hr_sw_128, 1>;
2666 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2667 int_x86_ssse3_pshuf_b,
2668 int_x86_ssse3_pshuf_b_128>;
2669 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2670 int_x86_ssse3_psign_b,
2671 int_x86_ssse3_psign_b_128>;
2672 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2673 int_x86_ssse3_psign_w,
2674 int_x86_ssse3_psign_w_128>;
2675 defm PSIGND : SS3I_binop_rm_int_32<0x09, "psignd",
2676 int_x86_ssse3_psign_d,
2677 int_x86_ssse3_psign_d_128>;
2679 let isTwoAddress = 1 in {
2680 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2681 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2682 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2684 (int_x86_ssse3_palign_r
2685 VR64:$src1, VR64:$src2,
2687 def PALIGNR64rm : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2688 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2689 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2691 (int_x86_ssse3_palign_r
2693 (bitconvert (memopv2i32 addr:$src2)),
2696 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2697 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2698 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2700 (int_x86_ssse3_palign_r_128
2701 VR128:$src1, VR128:$src2,
2702 imm:$src3))]>, OpSize;
2703 def PALIGNR128rm : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2704 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2705 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2707 (int_x86_ssse3_palign_r_128
2709 (bitconvert (memopv4i32 addr:$src2)),
2710 imm:$src3))]>, OpSize;
2713 //===----------------------------------------------------------------------===//
2714 // Non-Instruction Patterns
2715 //===----------------------------------------------------------------------===//
2717 // 128-bit vector undef's.
2718 def : Pat<(v4f32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2719 def : Pat<(v2f64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2720 def : Pat<(v16i8 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2721 def : Pat<(v8i16 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2722 def : Pat<(v4i32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2723 def : Pat<(v2i64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2725 // Scalar to v8i16 / v16i8. The source may be a GR32, but only the lower 8 or
2727 def : Pat<(v8i16 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2728 Requires<[HasSSE2]>;
2729 def : Pat<(v16i8 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2730 Requires<[HasSSE2]>;
2733 let Predicates = [HasSSE2] in {
2734 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2735 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2736 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2737 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2738 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2739 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2740 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2741 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2742 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2743 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2744 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2745 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2746 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2747 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2748 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2749 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2750 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2751 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2752 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2753 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2754 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2755 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2756 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2757 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2758 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2759 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2760 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2761 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2762 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2763 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2766 // Move scalar to XMM zero-extended
2767 // movd to XMM register zero-extends
2768 let AddedComplexity = 15 in {
2769 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2770 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc,
2771 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
2772 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2773 def : Pat<(v4f32 (vector_shuffle immAllZerosV_bc,
2774 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
2775 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
2778 // Splat v2f64 / v2i64
2779 let AddedComplexity = 10 in {
2780 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2781 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2782 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2783 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2784 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2785 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2786 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2787 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2791 def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SSE_splat_mask:$sm),
2792 (SHUFPSrri VR128:$src, VR128:$src, SSE_splat_mask:$sm)>,
2793 Requires<[HasSSE1]>;
2795 // Special unary SHUFPSrri case.
2796 // FIXME: when we want non two-address code, then we should use PSHUFD?
2797 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2798 SHUFP_unary_shuffle_mask:$sm)),
2799 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2800 Requires<[HasSSE1]>;
2801 // Special unary SHUFPDrri case.
2802 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (undef),
2803 SHUFP_unary_shuffle_mask:$sm)),
2804 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2805 Requires<[HasSSE2]>;
2806 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2807 def : Pat<(vector_shuffle (memopv4f32 addr:$src1), (undef),
2808 SHUFP_unary_shuffle_mask:$sm),
2809 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2810 Requires<[HasSSE2]>;
2811 // Special binary v4i32 shuffle cases with SHUFPS.
2812 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (v4i32 VR128:$src2),
2813 PSHUFD_binary_shuffle_mask:$sm)),
2814 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2815 Requires<[HasSSE2]>;
2816 def : Pat<(v4i32 (vector_shuffle VR128:$src1,
2817 (bc_v4i32 (memopv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm)),
2818 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2819 Requires<[HasSSE2]>;
2820 // Special binary v2i64 shuffle cases using SHUFPDrri.
2821 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2822 SHUFP_shuffle_mask:$sm)),
2823 (SHUFPDrri VR128:$src1, VR128:$src2, SHUFP_shuffle_mask:$sm)>,
2824 Requires<[HasSSE2]>;
2825 // Special unary SHUFPDrri case.
2826 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (undef),
2827 SHUFP_unary_shuffle_mask:$sm)),
2828 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2829 Requires<[HasSSE2]>;
2831 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2832 let AddedComplexity = 10 in {
2833 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2834 UNPCKL_v_undef_shuffle_mask)),
2835 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2836 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2837 UNPCKL_v_undef_shuffle_mask)),
2838 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2839 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2840 UNPCKL_v_undef_shuffle_mask)),
2841 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2842 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2843 UNPCKL_v_undef_shuffle_mask)),
2844 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2847 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2848 let AddedComplexity = 10 in {
2849 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2850 UNPCKH_v_undef_shuffle_mask)),
2851 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2852 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2853 UNPCKH_v_undef_shuffle_mask)),
2854 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2855 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2856 UNPCKH_v_undef_shuffle_mask)),
2857 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2858 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2859 UNPCKH_v_undef_shuffle_mask)),
2860 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2863 let AddedComplexity = 15 in {
2864 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2865 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2866 MOVHP_shuffle_mask)),
2867 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2869 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2870 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2871 MOVHLPS_shuffle_mask)),
2872 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2874 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2875 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2876 MOVHLPS_v_undef_shuffle_mask)),
2877 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2878 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2879 MOVHLPS_v_undef_shuffle_mask)),
2880 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2883 let AddedComplexity = 20 in {
2884 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2885 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2886 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2887 MOVLP_shuffle_mask)),
2888 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2889 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2890 MOVLP_shuffle_mask)),
2891 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2892 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2893 MOVHP_shuffle_mask)),
2894 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2895 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2896 MOVHP_shuffle_mask)),
2897 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2899 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2900 MOVLP_shuffle_mask)),
2901 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2902 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2903 MOVLP_shuffle_mask)),
2904 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2905 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2906 MOVHP_shuffle_mask)),
2907 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2908 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2909 MOVLP_shuffle_mask)),
2910 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2913 let AddedComplexity = 15 in {
2914 // Setting the lowest element in the vector.
2915 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2916 MOVL_shuffle_mask)),
2917 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2918 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2919 MOVL_shuffle_mask)),
2920 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2922 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2923 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2924 MOVLP_shuffle_mask)),
2925 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2926 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2927 MOVLP_shuffle_mask)),
2928 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2931 // Set lowest element and zero upper elements.
2932 let AddedComplexity = 15 in
2933 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc, VR128:$src,
2934 MOVL_shuffle_mask)),
2935 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
2938 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2939 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2940 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2941 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2942 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2943 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2944 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2945 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2946 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2947 Requires<[HasSSE2]>;
2948 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2949 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2950 Requires<[HasSSE2]>;
2951 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2952 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2953 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2954 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2955 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2956 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2957 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2958 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2959 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
2960 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2961 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
2962 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2963 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
2964 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2965 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
2966 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2968 // Some special case pandn patterns.
2969 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2971 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2972 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2974 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2975 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2977 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2979 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2980 (memopv2i64 addr:$src2))),
2981 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2982 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2983 (memopv2i64 addr:$src2))),
2984 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2985 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2986 (memopv2i64 addr:$src2))),
2987 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2989 // vector -> vector casts
2990 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
2991 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
2992 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
2993 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
2995 // Use movaps / movups for SSE integer load / store (one byte shorter).
2996 def : Pat<(alignedloadv4i32 addr:$src),
2997 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
2998 def : Pat<(loadv4i32 addr:$src),
2999 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3000 def : Pat<(alignedloadv2i64 addr:$src),
3001 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3002 def : Pat<(loadv2i64 addr:$src),
3003 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3005 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3006 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3007 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3008 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3009 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3010 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3011 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3012 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3013 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3014 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3015 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3016 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3017 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3018 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3019 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3020 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;