1 //====- X86InstrSSE.td - Describe the X86 Instruction Set --*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the X86 SSE instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
17 //===----------------------------------------------------------------------===//
18 // SSE specific DAG Nodes.
19 //===----------------------------------------------------------------------===//
21 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
22 SDTCisFP<0>, SDTCisInt<2> ]>;
24 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
25 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
26 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
27 [SDNPCommutative, SDNPAssociative]>;
28 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
29 [SDNPCommutative, SDNPAssociative]>;
30 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
31 [SDNPCommutative, SDNPAssociative]>;
32 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
33 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
34 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
35 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>;
36 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>;
37 def X86pextrb : SDNode<"X86ISD::PEXTRB",
38 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
39 def X86pextrw : SDNode<"X86ISD::PEXTRW",
40 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
41 def X86pinsrb : SDNode<"X86ISD::PINSRB",
42 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
43 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
44 def X86pinsrw : SDNode<"X86ISD::PINSRW",
45 SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
46 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
47 def X86insrtps : SDNode<"X86ISD::INSERTPS",
48 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
49 SDTCisVT<2, f32>, SDTCisPtrTy<3>]>>;
51 //===----------------------------------------------------------------------===//
52 // SSE 'Special' Instructions
53 //===----------------------------------------------------------------------===//
55 let isImplicitDef = 1 in {
56 def IMPLICIT_DEF_VR128 : I<0, Pseudo, (outs VR128:$dst), (ins),
58 [(set VR128:$dst, (v4f32 (undef)))]>,
60 def IMPLICIT_DEF_FR32 : I<0, Pseudo, (outs FR32:$dst), (ins),
62 [(set FR32:$dst, (undef))]>, Requires<[HasSSE1]>;
63 def IMPLICIT_DEF_FR64 : I<0, Pseudo, (outs FR64:$dst), (ins),
65 [(set FR64:$dst, (undef))]>, Requires<[HasSSE2]>;
68 //===----------------------------------------------------------------------===//
69 // SSE Complex Patterns
70 //===----------------------------------------------------------------------===//
72 // These are 'extloads' from a scalar to the low element of a vector, zeroing
73 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
75 def sse_load_f32 : ComplexPattern<v4f32, 4, "SelectScalarSSELoad", [],
76 [SDNPHasChain, SDNPMayLoad]>;
77 def sse_load_f64 : ComplexPattern<v2f64, 4, "SelectScalarSSELoad", [],
78 [SDNPHasChain, SDNPMayLoad]>;
80 def ssmem : Operand<v4f32> {
81 let PrintMethod = "printf32mem";
82 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
84 def sdmem : Operand<v2f64> {
85 let PrintMethod = "printf64mem";
86 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
89 //===----------------------------------------------------------------------===//
90 // SSE pattern fragments
91 //===----------------------------------------------------------------------===//
93 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
94 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
95 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
96 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
98 // Like 'store', but always requires vector alignment.
99 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
100 (st node:$val, node:$ptr), [{
101 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
102 return !ST->isTruncatingStore() &&
103 ST->getAddressingMode() == ISD::UNINDEXED &&
104 ST->getAlignment() >= 16;
108 // Like 'load', but always requires vector alignment.
109 def alignedload : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
110 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
111 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
112 LD->getAddressingMode() == ISD::UNINDEXED &&
113 LD->getAlignment() >= 16;
117 def alignedloadfsf32 : PatFrag<(ops node:$ptr), (f32 (alignedload node:$ptr))>;
118 def alignedloadfsf64 : PatFrag<(ops node:$ptr), (f64 (alignedload node:$ptr))>;
119 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
120 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
121 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
122 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
124 // Like 'load', but uses special alignment checks suitable for use in
125 // memory operands in most SSE instructions, which are required to
126 // be naturally aligned on some targets but not on others.
127 // FIXME: Actually implement support for targets that don't require the
128 // alignment. This probably wants a subtarget predicate.
129 def memop : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
130 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
131 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
132 LD->getAddressingMode() == ISD::UNINDEXED &&
133 LD->getAlignment() >= 16;
137 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
138 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
139 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
140 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
141 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
142 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
143 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
145 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
147 // FIXME: 8 byte alignment for mmx reads is not required
148 def memop64 : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
149 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
150 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
151 LD->getAddressingMode() == ISD::UNINDEXED &&
152 LD->getAlignment() >= 8;
156 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
157 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
158 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
159 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
161 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
162 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
163 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
164 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
165 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
166 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
168 def fp32imm0 : PatLeaf<(f32 fpimm), [{
169 return N->isExactlyValue(+0.0);
172 def PSxLDQ_imm : SDNodeXForm<imm, [{
173 // Transformation function: imm >> 3
174 return getI32Imm(N->getValue() >> 3);
177 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
179 def SHUFFLE_get_shuf_imm : SDNodeXForm<build_vector, [{
180 return getI8Imm(X86::getShuffleSHUFImmediate(N));
183 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
185 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<build_vector, [{
186 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
189 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
191 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<build_vector, [{
192 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
195 def SSE_splat_mask : PatLeaf<(build_vector), [{
196 return X86::isSplatMask(N);
197 }], SHUFFLE_get_shuf_imm>;
199 def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
200 return X86::isSplatLoMask(N);
203 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
204 return X86::isMOVHLPSMask(N);
207 def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
208 return X86::isMOVHLPS_v_undef_Mask(N);
211 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
212 return X86::isMOVHPMask(N);
215 def MOVLP_shuffle_mask : PatLeaf<(build_vector), [{
216 return X86::isMOVLPMask(N);
219 def MOVL_shuffle_mask : PatLeaf<(build_vector), [{
220 return X86::isMOVLMask(N);
223 def MOVSHDUP_shuffle_mask : PatLeaf<(build_vector), [{
224 return X86::isMOVSHDUPMask(N);
227 def MOVSLDUP_shuffle_mask : PatLeaf<(build_vector), [{
228 return X86::isMOVSLDUPMask(N);
231 def UNPCKL_shuffle_mask : PatLeaf<(build_vector), [{
232 return X86::isUNPCKLMask(N);
235 def UNPCKH_shuffle_mask : PatLeaf<(build_vector), [{
236 return X86::isUNPCKHMask(N);
239 def UNPCKL_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
240 return X86::isUNPCKL_v_undef_Mask(N);
243 def UNPCKH_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
244 return X86::isUNPCKH_v_undef_Mask(N);
247 def PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
248 return X86::isPSHUFDMask(N);
249 }], SHUFFLE_get_shuf_imm>;
251 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
252 return X86::isPSHUFHWMask(N);
253 }], SHUFFLE_get_pshufhw_imm>;
255 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
256 return X86::isPSHUFLWMask(N);
257 }], SHUFFLE_get_pshuflw_imm>;
259 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
260 return X86::isPSHUFDMask(N);
261 }], SHUFFLE_get_shuf_imm>;
263 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
264 return X86::isSHUFPMask(N);
265 }], SHUFFLE_get_shuf_imm>;
267 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
268 return X86::isSHUFPMask(N);
269 }], SHUFFLE_get_shuf_imm>;
271 //===----------------------------------------------------------------------===//
272 // SSE scalar FP Instructions
273 //===----------------------------------------------------------------------===//
275 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
276 // scheduler into a branch sequence.
277 // These are expanded by the scheduler.
278 let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
279 def CMOV_FR32 : I<0, Pseudo,
280 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
281 "#CMOV_FR32 PSEUDO!",
282 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
284 def CMOV_FR64 : I<0, Pseudo,
285 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
286 "#CMOV_FR64 PSEUDO!",
287 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
289 def CMOV_V4F32 : I<0, Pseudo,
290 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
291 "#CMOV_V4F32 PSEUDO!",
293 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
295 def CMOV_V2F64 : I<0, Pseudo,
296 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
297 "#CMOV_V2F64 PSEUDO!",
299 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
301 def CMOV_V2I64 : I<0, Pseudo,
302 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
303 "#CMOV_V2I64 PSEUDO!",
305 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
309 //===----------------------------------------------------------------------===//
311 //===----------------------------------------------------------------------===//
314 let neverHasSideEffects = 1 in
315 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
316 "movss\t{$src, $dst|$dst, $src}", []>;
317 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
318 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
319 "movss\t{$src, $dst|$dst, $src}",
320 [(set FR32:$dst, (loadf32 addr:$src))]>;
321 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
322 "movss\t{$src, $dst|$dst, $src}",
323 [(store FR32:$src, addr:$dst)]>;
325 // Conversion instructions
326 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
327 "cvttss2si\t{$src, $dst|$dst, $src}",
328 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
329 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
330 "cvttss2si\t{$src, $dst|$dst, $src}",
331 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
332 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
333 "cvtsi2ss\t{$src, $dst|$dst, $src}",
334 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
335 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
336 "cvtsi2ss\t{$src, $dst|$dst, $src}",
337 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
339 // Match intrinsics which expect XMM operand(s).
340 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
341 "cvtss2si\t{$src, $dst|$dst, $src}",
342 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
343 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
344 "cvtss2si\t{$src, $dst|$dst, $src}",
345 [(set GR32:$dst, (int_x86_sse_cvtss2si
346 (load addr:$src)))]>;
348 // Match intrinisics which expect MM and XMM operand(s).
349 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
350 "cvtps2pi\t{$src, $dst|$dst, $src}",
351 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
352 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
353 "cvtps2pi\t{$src, $dst|$dst, $src}",
354 [(set VR64:$dst, (int_x86_sse_cvtps2pi
355 (load addr:$src)))]>;
356 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
357 "cvttps2pi\t{$src, $dst|$dst, $src}",
358 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
359 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
360 "cvttps2pi\t{$src, $dst|$dst, $src}",
361 [(set VR64:$dst, (int_x86_sse_cvttps2pi
362 (load addr:$src)))]>;
363 let Constraints = "$src1 = $dst" in {
364 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
365 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
366 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
367 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
369 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
370 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
371 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
372 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
373 (load addr:$src2)))]>;
376 // Aliases for intrinsics
377 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
378 "cvttss2si\t{$src, $dst|$dst, $src}",
380 (int_x86_sse_cvttss2si VR128:$src))]>;
381 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
382 "cvttss2si\t{$src, $dst|$dst, $src}",
384 (int_x86_sse_cvttss2si(load addr:$src)))]>;
386 let Constraints = "$src1 = $dst" in {
387 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
388 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
389 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
390 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
392 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
393 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
394 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
395 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
396 (loadi32 addr:$src2)))]>;
399 // Comparison instructions
400 let Constraints = "$src1 = $dst" in {
401 let neverHasSideEffects = 1 in
402 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
403 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
404 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
405 let neverHasSideEffects = 1, mayLoad = 1 in
406 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
407 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
408 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
411 let Defs = [EFLAGS] in {
412 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
413 "ucomiss\t{$src2, $src1|$src1, $src2}",
414 [(X86cmp FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
415 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
416 "ucomiss\t{$src2, $src1|$src1, $src2}",
417 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
421 // Aliases to match intrinsics which expect XMM operand(s).
422 let Constraints = "$src1 = $dst" in {
423 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
424 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
425 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
426 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
427 VR128:$src, imm:$cc))]>;
428 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
429 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src, SSECC:$cc),
430 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
431 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
432 (load addr:$src), imm:$cc))]>;
435 let Defs = [EFLAGS] in {
436 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs),
437 (ins VR128:$src1, VR128:$src2),
438 "ucomiss\t{$src2, $src1|$src1, $src2}",
439 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
441 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),
442 (ins VR128:$src1, f128mem:$src2),
443 "ucomiss\t{$src2, $src1|$src1, $src2}",
444 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
447 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs),
448 (ins VR128:$src1, VR128:$src2),
449 "comiss\t{$src2, $src1|$src1, $src2}",
450 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
452 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs),
453 (ins VR128:$src1, f128mem:$src2),
454 "comiss\t{$src2, $src1|$src1, $src2}",
455 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
459 // Aliases of packed SSE1 instructions for scalar use. These all have names that
462 // Alias instructions that map fld0 to pxor for sse.
463 let isReMaterializable = 1 in
464 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
465 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
466 Requires<[HasSSE1]>, TB, OpSize;
468 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
470 let neverHasSideEffects = 1 in
471 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
472 "movaps\t{$src, $dst|$dst, $src}", []>;
474 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
476 let isSimpleLoad = 1 in
477 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
478 "movaps\t{$src, $dst|$dst, $src}",
479 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
481 // Alias bitwise logical operations using SSE logical ops on packed FP values.
482 let Constraints = "$src1 = $dst" in {
483 let isCommutable = 1 in {
484 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
485 "andps\t{$src2, $dst|$dst, $src2}",
486 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
487 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
488 "orps\t{$src2, $dst|$dst, $src2}",
489 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
490 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
491 "xorps\t{$src2, $dst|$dst, $src2}",
492 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
495 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
496 "andps\t{$src2, $dst|$dst, $src2}",
497 [(set FR32:$dst, (X86fand FR32:$src1,
498 (memopfsf32 addr:$src2)))]>;
499 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
500 "orps\t{$src2, $dst|$dst, $src2}",
501 [(set FR32:$dst, (X86for FR32:$src1,
502 (memopfsf32 addr:$src2)))]>;
503 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
504 "xorps\t{$src2, $dst|$dst, $src2}",
505 [(set FR32:$dst, (X86fxor FR32:$src1,
506 (memopfsf32 addr:$src2)))]>;
507 let neverHasSideEffects = 1 in {
508 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
509 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
510 "andnps\t{$src2, $dst|$dst, $src2}", []>;
513 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
514 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
515 "andnps\t{$src2, $dst|$dst, $src2}", []>;
519 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
521 /// In addition, we also have a special variant of the scalar form here to
522 /// represent the associated intrinsic operation. This form is unlike the
523 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
524 /// and leaves the top elements undefined.
526 /// These three forms can each be reg+reg or reg+mem, so there are a total of
527 /// six "instructions".
529 let Constraints = "$src1 = $dst" in {
530 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
531 SDNode OpNode, Intrinsic F32Int,
532 bit Commutable = 0> {
533 // Scalar operation, reg+reg.
534 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
535 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
536 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
537 let isCommutable = Commutable;
540 // Scalar operation, reg+mem.
541 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
542 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
543 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
545 // Vector operation, reg+reg.
546 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
547 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
548 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
549 let isCommutable = Commutable;
552 // Vector operation, reg+mem.
553 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
554 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
555 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
557 // Intrinsic operation, reg+reg.
558 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
559 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
560 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
561 let isCommutable = Commutable;
564 // Intrinsic operation, reg+mem.
565 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
566 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
567 [(set VR128:$dst, (F32Int VR128:$src1,
568 sse_load_f32:$src2))]>;
572 // Arithmetic instructions
573 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
574 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
575 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
576 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
578 /// sse1_fp_binop_rm - Other SSE1 binops
580 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
581 /// instructions for a full-vector intrinsic form. Operations that map
582 /// onto C operators don't use this form since they just use the plain
583 /// vector form instead of having a separate vector intrinsic form.
585 /// This provides a total of eight "instructions".
587 let Constraints = "$src1 = $dst" in {
588 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
592 bit Commutable = 0> {
594 // Scalar operation, reg+reg.
595 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
596 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
597 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
598 let isCommutable = Commutable;
601 // Scalar operation, reg+mem.
602 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
603 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
604 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
606 // Vector operation, reg+reg.
607 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
608 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
609 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
610 let isCommutable = Commutable;
613 // Vector operation, reg+mem.
614 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
615 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
616 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
618 // Intrinsic operation, reg+reg.
619 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
620 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
621 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
622 let isCommutable = Commutable;
625 // Intrinsic operation, reg+mem.
626 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
627 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
628 [(set VR128:$dst, (F32Int VR128:$src1,
629 sse_load_f32:$src2))]>;
631 // Vector intrinsic operation, reg+reg.
632 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
633 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
634 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
635 let isCommutable = Commutable;
638 // Vector intrinsic operation, reg+mem.
639 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
640 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
641 [(set VR128:$dst, (V4F32Int VR128:$src1, (load addr:$src2)))]>;
645 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
646 int_x86_sse_max_ss, int_x86_sse_max_ps>;
647 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
648 int_x86_sse_min_ss, int_x86_sse_min_ps>;
650 //===----------------------------------------------------------------------===//
651 // SSE packed FP Instructions
654 let neverHasSideEffects = 1 in
655 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
656 "movaps\t{$src, $dst|$dst, $src}", []>;
657 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
658 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
659 "movaps\t{$src, $dst|$dst, $src}",
660 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
662 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
663 "movaps\t{$src, $dst|$dst, $src}",
664 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
666 let neverHasSideEffects = 1 in
667 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
668 "movups\t{$src, $dst|$dst, $src}", []>;
669 let isSimpleLoad = 1 in
670 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
671 "movups\t{$src, $dst|$dst, $src}",
672 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
673 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
674 "movups\t{$src, $dst|$dst, $src}",
675 [(store (v4f32 VR128:$src), addr:$dst)]>;
677 // Intrinsic forms of MOVUPS load and store
678 let isSimpleLoad = 1 in
679 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
680 "movups\t{$src, $dst|$dst, $src}",
681 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
682 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
683 "movups\t{$src, $dst|$dst, $src}",
684 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
686 let Constraints = "$src1 = $dst" in {
687 let AddedComplexity = 20 in {
688 def MOVLPSrm : PSI<0x12, MRMSrcMem,
689 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
690 "movlps\t{$src2, $dst|$dst, $src2}",
692 (v4f32 (vector_shuffle VR128:$src1,
693 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
694 MOVLP_shuffle_mask)))]>;
695 def MOVHPSrm : PSI<0x16, MRMSrcMem,
696 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
697 "movhps\t{$src2, $dst|$dst, $src2}",
699 (v4f32 (vector_shuffle VR128:$src1,
700 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
701 MOVHP_shuffle_mask)))]>;
703 } // Constraints = "$src1 = $dst"
705 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
706 "movlps\t{$src, $dst|$dst, $src}",
707 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
708 (iPTR 0))), addr:$dst)]>;
710 // v2f64 extract element 1 is always custom lowered to unpack high to low
711 // and extract element 0 so the non-store version isn't too horrible.
712 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
713 "movhps\t{$src, $dst|$dst, $src}",
714 [(store (f64 (vector_extract
715 (v2f64 (vector_shuffle
716 (bc_v2f64 (v4f32 VR128:$src)), (undef),
717 UNPCKH_shuffle_mask)), (iPTR 0))),
720 let Constraints = "$src1 = $dst" in {
721 let AddedComplexity = 15 in {
722 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
723 "movlhps\t{$src2, $dst|$dst, $src2}",
725 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
726 MOVHP_shuffle_mask)))]>;
728 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
729 "movhlps\t{$src2, $dst|$dst, $src2}",
731 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
732 MOVHLPS_shuffle_mask)))]>;
734 } // Constraints = "$src1 = $dst"
740 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
742 /// In addition, we also have a special variant of the scalar form here to
743 /// represent the associated intrinsic operation. This form is unlike the
744 /// plain scalar form, in that it takes an entire vector (instead of a
745 /// scalar) and leaves the top elements undefined.
747 /// And, we have a special variant form for a full-vector intrinsic form.
749 /// These four forms can each have a reg or a mem operand, so there are a
750 /// total of eight "instructions".
752 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
756 bit Commutable = 0> {
757 // Scalar operation, reg.
758 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
759 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
760 [(set FR32:$dst, (OpNode FR32:$src))]> {
761 let isCommutable = Commutable;
764 // Scalar operation, mem.
765 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
766 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
767 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
769 // Vector operation, reg.
770 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
771 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
772 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
773 let isCommutable = Commutable;
776 // Vector operation, mem.
777 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
778 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
779 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
781 // Intrinsic operation, reg.
782 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
783 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
784 [(set VR128:$dst, (F32Int VR128:$src))]> {
785 let isCommutable = Commutable;
788 // Intrinsic operation, mem.
789 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
790 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
791 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
793 // Vector intrinsic operation, reg
794 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
795 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
796 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
797 let isCommutable = Commutable;
800 // Vector intrinsic operation, mem
801 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
802 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
803 [(set VR128:$dst, (V4F32Int (load addr:$src)))]>;
807 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
808 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
810 // Reciprocal approximations. Note that these typically require refinement
811 // in order to obtain suitable precision.
812 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
813 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
814 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
815 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
818 let Constraints = "$src1 = $dst" in {
819 let isCommutable = 1 in {
820 def ANDPSrr : PSI<0x54, MRMSrcReg,
821 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
822 "andps\t{$src2, $dst|$dst, $src2}",
823 [(set VR128:$dst, (v2i64
824 (and VR128:$src1, VR128:$src2)))]>;
825 def ORPSrr : PSI<0x56, MRMSrcReg,
826 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
827 "orps\t{$src2, $dst|$dst, $src2}",
828 [(set VR128:$dst, (v2i64
829 (or VR128:$src1, VR128:$src2)))]>;
830 def XORPSrr : PSI<0x57, MRMSrcReg,
831 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
832 "xorps\t{$src2, $dst|$dst, $src2}",
833 [(set VR128:$dst, (v2i64
834 (xor VR128:$src1, VR128:$src2)))]>;
837 def ANDPSrm : PSI<0x54, MRMSrcMem,
838 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
839 "andps\t{$src2, $dst|$dst, $src2}",
840 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
841 (memopv2i64 addr:$src2)))]>;
842 def ORPSrm : PSI<0x56, MRMSrcMem,
843 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
844 "orps\t{$src2, $dst|$dst, $src2}",
845 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
846 (memopv2i64 addr:$src2)))]>;
847 def XORPSrm : PSI<0x57, MRMSrcMem,
848 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
849 "xorps\t{$src2, $dst|$dst, $src2}",
850 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
851 (memopv2i64 addr:$src2)))]>;
852 def ANDNPSrr : PSI<0x55, MRMSrcReg,
853 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
854 "andnps\t{$src2, $dst|$dst, $src2}",
856 (v2i64 (and (xor VR128:$src1,
857 (bc_v2i64 (v4i32 immAllOnesV))),
859 def ANDNPSrm : PSI<0x55, MRMSrcMem,
860 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
861 "andnps\t{$src2, $dst|$dst, $src2}",
863 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
864 (bc_v2i64 (v4i32 immAllOnesV))),
865 (memopv2i64 addr:$src2))))]>;
868 let Constraints = "$src1 = $dst" in {
869 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
870 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
871 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
872 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
873 VR128:$src, imm:$cc))]>;
874 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
875 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
876 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
877 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
878 (load addr:$src), imm:$cc))]>;
881 // Shuffle and unpack instructions
882 let Constraints = "$src1 = $dst" in {
883 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
884 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
885 (outs VR128:$dst), (ins VR128:$src1,
886 VR128:$src2, i32i8imm:$src3),
887 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
889 (v4f32 (vector_shuffle
890 VR128:$src1, VR128:$src2,
891 SHUFP_shuffle_mask:$src3)))]>;
892 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
893 (outs VR128:$dst), (ins VR128:$src1,
894 f128mem:$src2, i32i8imm:$src3),
895 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
897 (v4f32 (vector_shuffle
898 VR128:$src1, (memopv4f32 addr:$src2),
899 SHUFP_shuffle_mask:$src3)))]>;
901 let AddedComplexity = 10 in {
902 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
903 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
904 "unpckhps\t{$src2, $dst|$dst, $src2}",
906 (v4f32 (vector_shuffle
907 VR128:$src1, VR128:$src2,
908 UNPCKH_shuffle_mask)))]>;
909 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
910 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
911 "unpckhps\t{$src2, $dst|$dst, $src2}",
913 (v4f32 (vector_shuffle
914 VR128:$src1, (memopv4f32 addr:$src2),
915 UNPCKH_shuffle_mask)))]>;
917 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
918 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
919 "unpcklps\t{$src2, $dst|$dst, $src2}",
921 (v4f32 (vector_shuffle
922 VR128:$src1, VR128:$src2,
923 UNPCKL_shuffle_mask)))]>;
924 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
925 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
926 "unpcklps\t{$src2, $dst|$dst, $src2}",
928 (v4f32 (vector_shuffle
929 VR128:$src1, (memopv4f32 addr:$src2),
930 UNPCKL_shuffle_mask)))]>;
932 } // Constraints = "$src1 = $dst"
935 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
936 "movmskps\t{$src, $dst|$dst, $src}",
937 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
938 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
939 "movmskpd\t{$src, $dst|$dst, $src}",
940 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
942 // Prefetch intrinsic.
943 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
944 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
945 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
946 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
947 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
948 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
949 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
950 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
952 // Non-temporal stores
953 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
954 "movntps\t{$src, $dst|$dst, $src}",
955 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
957 // Load, store, and memory fence
958 def SFENCE : PSI<0xAE, MRM7m, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
961 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
962 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
963 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
964 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
966 // Alias instructions that map zero vector to pxor / xorp* for sse.
967 let isReMaterializable = 1 in
968 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
970 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
972 // FR32 to 128-bit vector conversion.
973 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
974 "movss\t{$src, $dst|$dst, $src}",
976 (v4f32 (scalar_to_vector FR32:$src)))]>;
977 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
978 "movss\t{$src, $dst|$dst, $src}",
980 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
982 // FIXME: may not be able to eliminate this movss with coalescing the src and
983 // dest register classes are different. We really want to write this pattern
985 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
987 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
988 "movss\t{$src, $dst|$dst, $src}",
989 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
991 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
992 "movss\t{$src, $dst|$dst, $src}",
993 [(store (f32 (vector_extract (v4f32 VR128:$src),
994 (iPTR 0))), addr:$dst)]>;
997 // Move to lower bits of a VR128, leaving upper bits alone.
998 // Three operand (but two address) aliases.
999 let Constraints = "$src1 = $dst" in {
1000 let neverHasSideEffects = 1 in
1001 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
1002 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
1003 "movss\t{$src2, $dst|$dst, $src2}", []>;
1005 let AddedComplexity = 15 in
1006 def MOVLPSrr : SSI<0x10, MRMSrcReg,
1007 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1008 "movss\t{$src2, $dst|$dst, $src2}",
1010 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
1011 MOVL_shuffle_mask)))]>;
1014 // Move to lower bits of a VR128 and zeroing upper bits.
1015 // Loading from memory automatically zeroing upper bits.
1016 let AddedComplexity = 20 in
1017 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1018 "movss\t{$src, $dst|$dst, $src}",
1019 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV_bc,
1020 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
1021 MOVL_shuffle_mask)))]>;
1024 //===----------------------------------------------------------------------===//
1025 // SSE2 Instructions
1026 //===----------------------------------------------------------------------===//
1028 // Move Instructions
1029 let neverHasSideEffects = 1 in
1030 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1031 "movsd\t{$src, $dst|$dst, $src}", []>;
1032 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1033 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1034 "movsd\t{$src, $dst|$dst, $src}",
1035 [(set FR64:$dst, (loadf64 addr:$src))]>;
1036 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1037 "movsd\t{$src, $dst|$dst, $src}",
1038 [(store FR64:$src, addr:$dst)]>;
1040 // Conversion instructions
1041 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1042 "cvttsd2si\t{$src, $dst|$dst, $src}",
1043 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1044 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1045 "cvttsd2si\t{$src, $dst|$dst, $src}",
1046 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1047 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1048 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1049 [(set FR32:$dst, (fround FR64:$src))]>;
1050 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1051 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1052 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1053 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1054 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1055 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1056 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1057 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1058 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1060 // SSE2 instructions with XS prefix
1061 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1062 "cvtss2sd\t{$src, $dst|$dst, $src}",
1063 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1064 Requires<[HasSSE2]>;
1065 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1066 "cvtss2sd\t{$src, $dst|$dst, $src}",
1067 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1068 Requires<[HasSSE2]>;
1070 // Match intrinsics which expect XMM operand(s).
1071 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1072 "cvtsd2si\t{$src, $dst|$dst, $src}",
1073 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1074 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1075 "cvtsd2si\t{$src, $dst|$dst, $src}",
1076 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1077 (load addr:$src)))]>;
1079 // Match intrinisics which expect MM and XMM operand(s).
1080 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1081 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1082 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1083 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1084 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1085 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1086 (load addr:$src)))]>;
1087 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1088 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1089 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1090 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1091 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1092 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1093 (load addr:$src)))]>;
1094 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1095 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1096 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1097 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1098 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1099 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1100 (load addr:$src)))]>;
1102 // Aliases for intrinsics
1103 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1104 "cvttsd2si\t{$src, $dst|$dst, $src}",
1106 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1107 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1108 "cvttsd2si\t{$src, $dst|$dst, $src}",
1109 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1110 (load addr:$src)))]>;
1112 // Comparison instructions
1113 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1114 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1115 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1116 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1118 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1119 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1120 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1123 let Defs = [EFLAGS] in {
1124 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1125 "ucomisd\t{$src2, $src1|$src1, $src2}",
1126 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1127 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1128 "ucomisd\t{$src2, $src1|$src1, $src2}",
1129 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1130 (implicit EFLAGS)]>;
1133 // Aliases to match intrinsics which expect XMM operand(s).
1134 let Constraints = "$src1 = $dst" in {
1135 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1136 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1137 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1138 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1139 VR128:$src, imm:$cc))]>;
1140 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1141 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1142 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1143 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1144 (load addr:$src), imm:$cc))]>;
1147 let Defs = [EFLAGS] in {
1148 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1149 "ucomisd\t{$src2, $src1|$src1, $src2}",
1150 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1151 (implicit EFLAGS)]>;
1152 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1153 "ucomisd\t{$src2, $src1|$src1, $src2}",
1154 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1155 (implicit EFLAGS)]>;
1157 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1158 "comisd\t{$src2, $src1|$src1, $src2}",
1159 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1160 (implicit EFLAGS)]>;
1161 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1162 "comisd\t{$src2, $src1|$src1, $src2}",
1163 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1164 (implicit EFLAGS)]>;
1167 // Aliases of packed SSE2 instructions for scalar use. These all have names that
1170 // Alias instructions that map fld0 to pxor for sse.
1171 let isReMaterializable = 1 in
1172 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1173 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1174 Requires<[HasSSE2]>, TB, OpSize;
1176 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1178 let neverHasSideEffects = 1 in
1179 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1180 "movapd\t{$src, $dst|$dst, $src}", []>;
1182 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1184 let isSimpleLoad = 1 in
1185 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1186 "movapd\t{$src, $dst|$dst, $src}",
1187 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1189 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1190 let Constraints = "$src1 = $dst" in {
1191 let isCommutable = 1 in {
1192 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1193 "andpd\t{$src2, $dst|$dst, $src2}",
1194 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1195 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1196 "orpd\t{$src2, $dst|$dst, $src2}",
1197 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1198 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1199 "xorpd\t{$src2, $dst|$dst, $src2}",
1200 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1203 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1204 "andpd\t{$src2, $dst|$dst, $src2}",
1205 [(set FR64:$dst, (X86fand FR64:$src1,
1206 (memopfsf64 addr:$src2)))]>;
1207 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1208 "orpd\t{$src2, $dst|$dst, $src2}",
1209 [(set FR64:$dst, (X86for FR64:$src1,
1210 (memopfsf64 addr:$src2)))]>;
1211 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1212 "xorpd\t{$src2, $dst|$dst, $src2}",
1213 [(set FR64:$dst, (X86fxor FR64:$src1,
1214 (memopfsf64 addr:$src2)))]>;
1216 let neverHasSideEffects = 1 in {
1217 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1218 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1219 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1221 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1222 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1223 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1227 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1229 /// In addition, we also have a special variant of the scalar form here to
1230 /// represent the associated intrinsic operation. This form is unlike the
1231 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1232 /// and leaves the top elements undefined.
1234 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1235 /// six "instructions".
1237 let Constraints = "$src1 = $dst" in {
1238 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1239 SDNode OpNode, Intrinsic F64Int,
1240 bit Commutable = 0> {
1241 // Scalar operation, reg+reg.
1242 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1243 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1244 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1245 let isCommutable = Commutable;
1248 // Scalar operation, reg+mem.
1249 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1250 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1251 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1253 // Vector operation, reg+reg.
1254 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1255 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1256 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1257 let isCommutable = Commutable;
1260 // Vector operation, reg+mem.
1261 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1262 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1263 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1265 // Intrinsic operation, reg+reg.
1266 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1267 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1268 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1269 let isCommutable = Commutable;
1272 // Intrinsic operation, reg+mem.
1273 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1274 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1275 [(set VR128:$dst, (F64Int VR128:$src1,
1276 sse_load_f64:$src2))]>;
1280 // Arithmetic instructions
1281 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1282 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1283 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1284 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1286 /// sse2_fp_binop_rm - Other SSE2 binops
1288 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1289 /// instructions for a full-vector intrinsic form. Operations that map
1290 /// onto C operators don't use this form since they just use the plain
1291 /// vector form instead of having a separate vector intrinsic form.
1293 /// This provides a total of eight "instructions".
1295 let Constraints = "$src1 = $dst" in {
1296 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1300 bit Commutable = 0> {
1302 // Scalar operation, reg+reg.
1303 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1304 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1305 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1306 let isCommutable = Commutable;
1309 // Scalar operation, reg+mem.
1310 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1311 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1312 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1314 // Vector operation, reg+reg.
1315 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1316 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1317 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1318 let isCommutable = Commutable;
1321 // Vector operation, reg+mem.
1322 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1323 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1324 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1326 // Intrinsic operation, reg+reg.
1327 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1328 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1329 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1330 let isCommutable = Commutable;
1333 // Intrinsic operation, reg+mem.
1334 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1335 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1336 [(set VR128:$dst, (F64Int VR128:$src1,
1337 sse_load_f64:$src2))]>;
1339 // Vector intrinsic operation, reg+reg.
1340 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1341 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1342 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1343 let isCommutable = Commutable;
1346 // Vector intrinsic operation, reg+mem.
1347 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1348 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1349 [(set VR128:$dst, (V2F64Int VR128:$src1, (load addr:$src2)))]>;
1353 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1354 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1355 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1356 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1358 //===----------------------------------------------------------------------===//
1359 // SSE packed FP Instructions
1361 // Move Instructions
1362 let neverHasSideEffects = 1 in
1363 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1364 "movapd\t{$src, $dst|$dst, $src}", []>;
1365 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1366 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1367 "movapd\t{$src, $dst|$dst, $src}",
1368 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1370 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1371 "movapd\t{$src, $dst|$dst, $src}",
1372 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1374 let neverHasSideEffects = 1 in
1375 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1376 "movupd\t{$src, $dst|$dst, $src}", []>;
1377 let isSimpleLoad = 1 in
1378 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1379 "movupd\t{$src, $dst|$dst, $src}",
1380 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1381 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1382 "movupd\t{$src, $dst|$dst, $src}",
1383 [(store (v2f64 VR128:$src), addr:$dst)]>;
1385 // Intrinsic forms of MOVUPD load and store
1386 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1387 "movupd\t{$src, $dst|$dst, $src}",
1388 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1389 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1390 "movupd\t{$src, $dst|$dst, $src}",
1391 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1393 let Constraints = "$src1 = $dst" in {
1394 let AddedComplexity = 20 in {
1395 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1396 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1397 "movlpd\t{$src2, $dst|$dst, $src2}",
1399 (v2f64 (vector_shuffle VR128:$src1,
1400 (scalar_to_vector (loadf64 addr:$src2)),
1401 MOVLP_shuffle_mask)))]>;
1402 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1403 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1404 "movhpd\t{$src2, $dst|$dst, $src2}",
1406 (v2f64 (vector_shuffle VR128:$src1,
1407 (scalar_to_vector (loadf64 addr:$src2)),
1408 MOVHP_shuffle_mask)))]>;
1409 } // AddedComplexity
1410 } // Constraints = "$src1 = $dst"
1412 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1413 "movlpd\t{$src, $dst|$dst, $src}",
1414 [(store (f64 (vector_extract (v2f64 VR128:$src),
1415 (iPTR 0))), addr:$dst)]>;
1417 // v2f64 extract element 1 is always custom lowered to unpack high to low
1418 // and extract element 0 so the non-store version isn't too horrible.
1419 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1420 "movhpd\t{$src, $dst|$dst, $src}",
1421 [(store (f64 (vector_extract
1422 (v2f64 (vector_shuffle VR128:$src, (undef),
1423 UNPCKH_shuffle_mask)), (iPTR 0))),
1426 // SSE2 instructions without OpSize prefix
1427 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1428 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1429 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1430 TB, Requires<[HasSSE2]>;
1431 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1432 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1433 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1434 (bitconvert (memopv2i64 addr:$src))))]>,
1435 TB, Requires<[HasSSE2]>;
1437 // SSE2 instructions with XS prefix
1438 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1439 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1440 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1441 XS, Requires<[HasSSE2]>;
1442 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1443 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1444 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1445 (bitconvert (memopv2i64 addr:$src))))]>,
1446 XS, Requires<[HasSSE2]>;
1448 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1449 "cvtps2dq\t{$src, $dst|$dst, $src}",
1450 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1451 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1452 "cvtps2dq\t{$src, $dst|$dst, $src}",
1453 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1454 (load addr:$src)))]>;
1455 // SSE2 packed instructions with XS prefix
1456 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1457 "cvttps2dq\t{$src, $dst|$dst, $src}",
1458 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1459 XS, Requires<[HasSSE2]>;
1460 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1461 "cvttps2dq\t{$src, $dst|$dst, $src}",
1462 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1463 (load addr:$src)))]>,
1464 XS, Requires<[HasSSE2]>;
1466 // SSE2 packed instructions with XD prefix
1467 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1468 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1469 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1470 XD, Requires<[HasSSE2]>;
1471 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1472 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1473 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1474 (load addr:$src)))]>,
1475 XD, Requires<[HasSSE2]>;
1477 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1478 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1479 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1480 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1481 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1482 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1483 (load addr:$src)))]>;
1485 // SSE2 instructions without OpSize prefix
1486 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1487 "cvtps2pd\t{$src, $dst|$dst, $src}",
1488 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1489 TB, Requires<[HasSSE2]>;
1490 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f64mem:$src),
1491 "cvtps2pd\t{$src, $dst|$dst, $src}",
1492 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1493 (load addr:$src)))]>,
1494 TB, Requires<[HasSSE2]>;
1496 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1497 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1498 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1499 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f128mem:$src),
1500 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1501 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1502 (load addr:$src)))]>;
1504 // Match intrinsics which expect XMM operand(s).
1505 // Aliases for intrinsics
1506 let Constraints = "$src1 = $dst" in {
1507 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1508 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1509 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1510 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1512 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1513 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1514 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1515 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1516 (loadi32 addr:$src2)))]>;
1517 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1518 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1519 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1520 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1522 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1523 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1524 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1525 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1526 (load addr:$src2)))]>;
1527 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1528 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1529 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1530 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1531 VR128:$src2))]>, XS,
1532 Requires<[HasSSE2]>;
1533 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1534 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1535 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1536 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1537 (load addr:$src2)))]>, XS,
1538 Requires<[HasSSE2]>;
1543 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1545 /// In addition, we also have a special variant of the scalar form here to
1546 /// represent the associated intrinsic operation. This form is unlike the
1547 /// plain scalar form, in that it takes an entire vector (instead of a
1548 /// scalar) and leaves the top elements undefined.
1550 /// And, we have a special variant form for a full-vector intrinsic form.
1552 /// These four forms can each have a reg or a mem operand, so there are a
1553 /// total of eight "instructions".
1555 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1559 bit Commutable = 0> {
1560 // Scalar operation, reg.
1561 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1562 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1563 [(set FR64:$dst, (OpNode FR64:$src))]> {
1564 let isCommutable = Commutable;
1567 // Scalar operation, mem.
1568 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1569 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1570 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1572 // Vector operation, reg.
1573 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1574 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1575 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1576 let isCommutable = Commutable;
1579 // Vector operation, mem.
1580 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1581 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1582 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1584 // Intrinsic operation, reg.
1585 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1586 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1587 [(set VR128:$dst, (F64Int VR128:$src))]> {
1588 let isCommutable = Commutable;
1591 // Intrinsic operation, mem.
1592 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1593 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1594 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1596 // Vector intrinsic operation, reg
1597 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1598 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1599 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1600 let isCommutable = Commutable;
1603 // Vector intrinsic operation, mem
1604 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1605 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1606 [(set VR128:$dst, (V2F64Int (load addr:$src)))]>;
1610 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1611 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1613 // There is no f64 version of the reciprocal approximation instructions.
1616 let Constraints = "$src1 = $dst" in {
1617 let isCommutable = 1 in {
1618 def ANDPDrr : PDI<0x54, MRMSrcReg,
1619 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1620 "andpd\t{$src2, $dst|$dst, $src2}",
1622 (and (bc_v2i64 (v2f64 VR128:$src1)),
1623 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1624 def ORPDrr : PDI<0x56, MRMSrcReg,
1625 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1626 "orpd\t{$src2, $dst|$dst, $src2}",
1628 (or (bc_v2i64 (v2f64 VR128:$src1)),
1629 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1630 def XORPDrr : PDI<0x57, MRMSrcReg,
1631 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1632 "xorpd\t{$src2, $dst|$dst, $src2}",
1634 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1635 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1638 def ANDPDrm : PDI<0x54, MRMSrcMem,
1639 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1640 "andpd\t{$src2, $dst|$dst, $src2}",
1642 (and (bc_v2i64 (v2f64 VR128:$src1)),
1643 (memopv2i64 addr:$src2)))]>;
1644 def ORPDrm : PDI<0x56, MRMSrcMem,
1645 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1646 "orpd\t{$src2, $dst|$dst, $src2}",
1648 (or (bc_v2i64 (v2f64 VR128:$src1)),
1649 (memopv2i64 addr:$src2)))]>;
1650 def XORPDrm : PDI<0x57, MRMSrcMem,
1651 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1652 "xorpd\t{$src2, $dst|$dst, $src2}",
1654 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1655 (memopv2i64 addr:$src2)))]>;
1656 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1657 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1658 "andnpd\t{$src2, $dst|$dst, $src2}",
1660 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1661 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1662 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1663 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1664 "andnpd\t{$src2, $dst|$dst, $src2}",
1666 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1667 (memopv2i64 addr:$src2)))]>;
1670 let Constraints = "$src1 = $dst" in {
1671 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1672 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1673 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1674 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1675 VR128:$src, imm:$cc))]>;
1676 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1677 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1678 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1679 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1680 (load addr:$src), imm:$cc))]>;
1683 // Shuffle and unpack instructions
1684 let Constraints = "$src1 = $dst" in {
1685 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1686 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1687 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1688 [(set VR128:$dst, (v2f64 (vector_shuffle
1689 VR128:$src1, VR128:$src2,
1690 SHUFP_shuffle_mask:$src3)))]>;
1691 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1692 (outs VR128:$dst), (ins VR128:$src1,
1693 f128mem:$src2, i8imm:$src3),
1694 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1696 (v2f64 (vector_shuffle
1697 VR128:$src1, (memopv2f64 addr:$src2),
1698 SHUFP_shuffle_mask:$src3)))]>;
1700 let AddedComplexity = 10 in {
1701 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1702 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1703 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1705 (v2f64 (vector_shuffle
1706 VR128:$src1, VR128:$src2,
1707 UNPCKH_shuffle_mask)))]>;
1708 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1709 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1710 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1712 (v2f64 (vector_shuffle
1713 VR128:$src1, (memopv2f64 addr:$src2),
1714 UNPCKH_shuffle_mask)))]>;
1716 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1717 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1718 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1720 (v2f64 (vector_shuffle
1721 VR128:$src1, VR128:$src2,
1722 UNPCKL_shuffle_mask)))]>;
1723 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1724 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1725 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1727 (v2f64 (vector_shuffle
1728 VR128:$src1, (memopv2f64 addr:$src2),
1729 UNPCKL_shuffle_mask)))]>;
1730 } // AddedComplexity
1731 } // Constraints = "$src1 = $dst"
1734 //===----------------------------------------------------------------------===//
1735 // SSE integer instructions
1737 // Move Instructions
1738 let neverHasSideEffects = 1 in
1739 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1740 "movdqa\t{$src, $dst|$dst, $src}", []>;
1741 let isSimpleLoad = 1, mayLoad = 1 in
1742 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1743 "movdqa\t{$src, $dst|$dst, $src}",
1744 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1746 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1747 "movdqa\t{$src, $dst|$dst, $src}",
1748 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1749 let isSimpleLoad = 1, mayLoad = 1 in
1750 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1751 "movdqu\t{$src, $dst|$dst, $src}",
1752 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1753 XS, Requires<[HasSSE2]>;
1755 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1756 "movdqu\t{$src, $dst|$dst, $src}",
1757 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1758 XS, Requires<[HasSSE2]>;
1760 // Intrinsic forms of MOVDQU load and store
1761 let isSimpleLoad = 1 in
1762 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1763 "movdqu\t{$src, $dst|$dst, $src}",
1764 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1765 XS, Requires<[HasSSE2]>;
1766 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1767 "movdqu\t{$src, $dst|$dst, $src}",
1768 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1769 XS, Requires<[HasSSE2]>;
1771 let Constraints = "$src1 = $dst" in {
1773 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1774 bit Commutable = 0> {
1775 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1776 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1777 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1778 let isCommutable = Commutable;
1780 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1781 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1782 [(set VR128:$dst, (IntId VR128:$src1,
1783 (bitconvert (memopv2i64 addr:$src2))))]>;
1786 /// PDI_binop_rm - Simple SSE2 binary operator.
1787 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1788 ValueType OpVT, bit Commutable = 0> {
1789 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1790 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1791 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1792 let isCommutable = Commutable;
1794 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1795 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1796 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1797 (bitconvert (memopv2i64 addr:$src2)))))]>;
1800 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1802 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1803 /// to collapse (bitconvert VT to VT) into its operand.
1805 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1806 bit Commutable = 0> {
1807 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1808 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1809 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1810 let isCommutable = Commutable;
1812 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1813 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1814 [(set VR128:$dst, (OpNode VR128:$src1,(memopv2i64 addr:$src2)))]>;
1817 } // Constraints = "$src1 = $dst"
1819 // 128-bit Integer Arithmetic
1821 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1822 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1823 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1824 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1826 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1827 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1828 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1829 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1831 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1832 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1833 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1834 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1836 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1837 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1838 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1839 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1841 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1843 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1844 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1845 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1847 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1849 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1850 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1853 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1854 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1855 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1856 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1857 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1860 defm PSLLW : PDI_binop_rm_int<0xF1, "psllw", int_x86_sse2_psll_w>;
1861 defm PSLLD : PDI_binop_rm_int<0xF2, "pslld", int_x86_sse2_psll_d>;
1862 defm PSLLQ : PDI_binop_rm_int<0xF3, "psllq", int_x86_sse2_psll_q>;
1864 defm PSRLW : PDI_binop_rm_int<0xD1, "psrlw", int_x86_sse2_psrl_w>;
1865 defm PSRLD : PDI_binop_rm_int<0xD2, "psrld", int_x86_sse2_psrl_d>;
1866 defm PSRLQ : PDI_binop_rm_int<0xD3, "psrlq", int_x86_sse2_psrl_q>;
1868 defm PSRAW : PDI_binop_rm_int<0xE1, "psraw", int_x86_sse2_psra_w>;
1869 defm PSRAD : PDI_binop_rm_int<0xE2, "psrad", int_x86_sse2_psra_d>;
1871 // Some immediate variants need to match a bit_convert.
1872 let Constraints = "$src1 = $dst" in {
1873 def PSLLWri : PDIi8<0x71, MRM6r, (outs VR128:$dst),
1874 (ins VR128:$src1, i32i8imm:$src2),
1875 "psllw\t{$src2, $dst|$dst, $src2}",
1876 [(set VR128:$dst, (int_x86_sse2_psll_w VR128:$src1,
1877 (bc_v8i16 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
1878 def PSLLDri : PDIi8<0x72, MRM6r, (outs VR128:$dst),
1879 (ins VR128:$src1, i32i8imm:$src2),
1880 "pslld\t{$src2, $dst|$dst, $src2}",
1881 [(set VR128:$dst, (int_x86_sse2_psll_d VR128:$src1,
1882 (scalar_to_vector (i32 imm:$src2))))]>;
1883 def PSLLQri : PDIi8<0x73, MRM6r, (outs VR128:$dst),
1884 (ins VR128:$src1, i32i8imm:$src2),
1885 "psllq\t{$src2, $dst|$dst, $src2}",
1886 [(set VR128:$dst, (int_x86_sse2_psll_q VR128:$src1,
1887 (bc_v2i64 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
1889 def PSRLWri : PDIi8<0x71, MRM2r, (outs VR128:$dst),
1890 (ins VR128:$src1, i32i8imm:$src2),
1891 "psrlw\t{$src2, $dst|$dst, $src2}",
1892 [(set VR128:$dst, (int_x86_sse2_psrl_w VR128:$src1,
1893 (bc_v8i16 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
1894 def PSRLDri : PDIi8<0x72, MRM2r, (outs VR128:$dst),
1895 (ins VR128:$src1, i32i8imm:$src2),
1896 "psrld\t{$src2, $dst|$dst, $src2}",
1897 [(set VR128:$dst, (int_x86_sse2_psrl_d VR128:$src1,
1898 (scalar_to_vector (i32 imm:$src2))))]>;
1899 def PSRLQri : PDIi8<0x73, MRM2r, (outs VR128:$dst),
1900 (ins VR128:$src1, i32i8imm:$src2),
1901 "psrlq\t{$src2, $dst|$dst, $src2}",
1902 [(set VR128:$dst, (int_x86_sse2_psrl_q VR128:$src1,
1903 (bc_v2i64 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
1905 def PSRAWri : PDIi8<0x71, MRM4r, (outs VR128:$dst),
1906 (ins VR128:$src1, i32i8imm:$src2),
1907 "psraw\t{$src2, $dst|$dst, $src2}",
1908 [(set VR128:$dst, (int_x86_sse2_psra_w VR128:$src1,
1909 (bc_v8i16 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
1910 def PSRADri : PDIi8<0x72, MRM4r, (outs VR128:$dst),
1911 (ins VR128:$src1, i32i8imm:$src2),
1912 "psrad\t{$src2, $dst|$dst, $src2}",
1913 [(set VR128:$dst, (int_x86_sse2_psra_d VR128:$src1,
1914 (scalar_to_vector (i32 imm:$src2))))]>;
1917 // PSRAQ doesn't exist in SSE[1-3].
1919 // 128-bit logical shifts.
1920 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1921 def PSLLDQri : PDIi8<0x73, MRM7r,
1922 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1923 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1924 def PSRLDQri : PDIi8<0x73, MRM3r,
1925 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1926 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1927 // PSRADQri doesn't exist in SSE[1-3].
1930 let Predicates = [HasSSE2] in {
1931 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1932 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1933 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1934 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1935 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1936 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1940 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1941 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1942 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1944 let Constraints = "$src1 = $dst" in {
1945 def PANDNrr : PDI<0xDF, MRMSrcReg,
1946 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1947 "pandn\t{$src2, $dst|$dst, $src2}",
1948 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1951 def PANDNrm : PDI<0xDF, MRMSrcMem,
1952 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1953 "pandn\t{$src2, $dst|$dst, $src2}",
1954 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1955 (memopv2i64 addr:$src2))))]>;
1958 // SSE2 Integer comparison
1959 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1960 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1961 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1962 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1963 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1964 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1966 // Pack instructions
1967 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1968 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1969 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1971 // Shuffle and unpack instructions
1972 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1973 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1974 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1975 [(set VR128:$dst, (v4i32 (vector_shuffle
1976 VR128:$src1, (undef),
1977 PSHUFD_shuffle_mask:$src2)))]>;
1978 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1979 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1980 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1981 [(set VR128:$dst, (v4i32 (vector_shuffle
1982 (bc_v4i32(memopv2i64 addr:$src1)),
1984 PSHUFD_shuffle_mask:$src2)))]>;
1986 // SSE2 with ImmT == Imm8 and XS prefix.
1987 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1988 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1989 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1990 [(set VR128:$dst, (v8i16 (vector_shuffle
1991 VR128:$src1, (undef),
1992 PSHUFHW_shuffle_mask:$src2)))]>,
1993 XS, Requires<[HasSSE2]>;
1994 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
1995 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1996 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1997 [(set VR128:$dst, (v8i16 (vector_shuffle
1998 (bc_v8i16 (memopv2i64 addr:$src1)),
2000 PSHUFHW_shuffle_mask:$src2)))]>,
2001 XS, Requires<[HasSSE2]>;
2003 // SSE2 with ImmT == Imm8 and XD prefix.
2004 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2005 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2006 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2007 [(set VR128:$dst, (v8i16 (vector_shuffle
2008 VR128:$src1, (undef),
2009 PSHUFLW_shuffle_mask:$src2)))]>,
2010 XD, Requires<[HasSSE2]>;
2011 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2012 (outs VR128:$dst), (ins i128mem:$src1, i32i8imm:$src2),
2013 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2014 [(set VR128:$dst, (v8i16 (vector_shuffle
2015 (bc_v8i16 (memopv2i64 addr:$src1)),
2017 PSHUFLW_shuffle_mask:$src2)))]>,
2018 XD, Requires<[HasSSE2]>;
2021 let Constraints = "$src1 = $dst" in {
2022 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
2023 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2024 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2026 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
2027 UNPCKL_shuffle_mask)))]>;
2028 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2029 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2030 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2032 (v16i8 (vector_shuffle VR128:$src1,
2033 (bc_v16i8 (memopv2i64 addr:$src2)),
2034 UNPCKL_shuffle_mask)))]>;
2035 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2036 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2037 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2039 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2040 UNPCKL_shuffle_mask)))]>;
2041 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2042 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2043 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2045 (v8i16 (vector_shuffle VR128:$src1,
2046 (bc_v8i16 (memopv2i64 addr:$src2)),
2047 UNPCKL_shuffle_mask)))]>;
2048 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2049 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2050 "punpckldq\t{$src2, $dst|$dst, $src2}",
2052 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2053 UNPCKL_shuffle_mask)))]>;
2054 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2055 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2056 "punpckldq\t{$src2, $dst|$dst, $src2}",
2058 (v4i32 (vector_shuffle VR128:$src1,
2059 (bc_v4i32 (memopv2i64 addr:$src2)),
2060 UNPCKL_shuffle_mask)))]>;
2061 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2062 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2063 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2065 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2066 UNPCKL_shuffle_mask)))]>;
2067 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2068 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2069 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2071 (v2i64 (vector_shuffle VR128:$src1,
2072 (memopv2i64 addr:$src2),
2073 UNPCKL_shuffle_mask)))]>;
2075 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2076 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2077 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2079 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
2080 UNPCKH_shuffle_mask)))]>;
2081 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2082 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2083 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2085 (v16i8 (vector_shuffle VR128:$src1,
2086 (bc_v16i8 (memopv2i64 addr:$src2)),
2087 UNPCKH_shuffle_mask)))]>;
2088 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2089 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2090 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2092 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2093 UNPCKH_shuffle_mask)))]>;
2094 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2095 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2096 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2098 (v8i16 (vector_shuffle VR128:$src1,
2099 (bc_v8i16 (memopv2i64 addr:$src2)),
2100 UNPCKH_shuffle_mask)))]>;
2101 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2102 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2103 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2105 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2106 UNPCKH_shuffle_mask)))]>;
2107 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2108 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2109 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2111 (v4i32 (vector_shuffle VR128:$src1,
2112 (bc_v4i32 (memopv2i64 addr:$src2)),
2113 UNPCKH_shuffle_mask)))]>;
2114 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2115 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2116 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2118 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2119 UNPCKH_shuffle_mask)))]>;
2120 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2121 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2122 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2124 (v2i64 (vector_shuffle VR128:$src1,
2125 (memopv2i64 addr:$src2),
2126 UNPCKH_shuffle_mask)))]>;
2130 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2131 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2132 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2133 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2135 let Constraints = "$src1 = $dst" in {
2136 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2137 (outs VR128:$dst), (ins VR128:$src1,
2138 GR32:$src2, i32i8imm:$src3),
2139 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2141 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2142 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2143 (outs VR128:$dst), (ins VR128:$src1,
2144 i16mem:$src2, i32i8imm:$src3),
2145 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2147 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2152 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2153 "pmovmskb\t{$src, $dst|$dst, $src}",
2154 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2156 // Conditional store
2158 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2159 "maskmovdqu\t{$mask, $src|$src, $mask}",
2160 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2162 // Non-temporal stores
2163 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2164 "movntpd\t{$src, $dst|$dst, $src}",
2165 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2166 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2167 "movntdq\t{$src, $dst|$dst, $src}",
2168 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2169 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2170 "movnti\t{$src, $dst|$dst, $src}",
2171 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2172 TB, Requires<[HasSSE2]>;
2175 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2176 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2177 TB, Requires<[HasSSE2]>;
2179 // Load, store, and memory fence
2180 def LFENCE : I<0xAE, MRM5m, (outs), (ins),
2181 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2182 def MFENCE : I<0xAE, MRM6m, (outs), (ins),
2183 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2185 //TODO: custom lower this so as to never even generate the noop
2186 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2188 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2189 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2190 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2193 // Alias instructions that map zero vector to pxor / xorp* for sse.
2194 let isReMaterializable = 1 in
2195 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2196 "pcmpeqd\t$dst, $dst",
2197 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2199 // FR64 to 128-bit vector conversion.
2200 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2201 "movsd\t{$src, $dst|$dst, $src}",
2203 (v2f64 (scalar_to_vector FR64:$src)))]>;
2204 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2205 "movsd\t{$src, $dst|$dst, $src}",
2207 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2209 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2210 "movd\t{$src, $dst|$dst, $src}",
2212 (v4i32 (scalar_to_vector GR32:$src)))]>;
2213 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2214 "movd\t{$src, $dst|$dst, $src}",
2216 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2218 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2219 "movd\t{$src, $dst|$dst, $src}",
2220 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2222 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2223 "movd\t{$src, $dst|$dst, $src}",
2224 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2226 // SSE2 instructions with XS prefix
2227 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2228 "movq\t{$src, $dst|$dst, $src}",
2230 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2231 Requires<[HasSSE2]>;
2232 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2233 "movq\t{$src, $dst|$dst, $src}",
2234 [(store (i64 (vector_extract (v2i64 VR128:$src),
2235 (iPTR 0))), addr:$dst)]>;
2237 // FIXME: may not be able to eliminate this movss with coalescing the src and
2238 // dest register classes are different. We really want to write this pattern
2240 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2241 // (f32 FR32:$src)>;
2242 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2243 "movsd\t{$src, $dst|$dst, $src}",
2244 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2246 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2247 "movsd\t{$src, $dst|$dst, $src}",
2248 [(store (f64 (vector_extract (v2f64 VR128:$src),
2249 (iPTR 0))), addr:$dst)]>;
2250 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2251 "movd\t{$src, $dst|$dst, $src}",
2252 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2254 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2255 "movd\t{$src, $dst|$dst, $src}",
2256 [(store (i32 (vector_extract (v4i32 VR128:$src),
2257 (iPTR 0))), addr:$dst)]>;
2259 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2260 "movd\t{$src, $dst|$dst, $src}",
2261 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2262 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2263 "movd\t{$src, $dst|$dst, $src}",
2264 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2267 // Move to lower bits of a VR128, leaving upper bits alone.
2268 // Three operand (but two address) aliases.
2269 let Constraints = "$src1 = $dst" in {
2270 let neverHasSideEffects = 1 in
2271 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2272 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2273 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2275 let AddedComplexity = 15 in
2276 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2277 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2278 "movsd\t{$src2, $dst|$dst, $src2}",
2280 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
2281 MOVL_shuffle_mask)))]>;
2284 // Store / copy lower 64-bits of a XMM register.
2285 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2286 "movq\t{$src, $dst|$dst, $src}",
2287 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2289 // Move to lower bits of a VR128 and zeroing upper bits.
2290 // Loading from memory automatically zeroing upper bits.
2291 let AddedComplexity = 20 in
2292 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2293 "movsd\t{$src, $dst|$dst, $src}",
2295 (v2f64 (vector_shuffle immAllZerosV_bc,
2296 (v2f64 (scalar_to_vector
2297 (loadf64 addr:$src))),
2298 MOVL_shuffle_mask)))]>;
2300 // movd / movq to XMM register zero-extends
2301 let AddedComplexity = 15 in {
2302 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2303 "movd\t{$src, $dst|$dst, $src}",
2305 (v4i32 (vector_shuffle immAllZerosV,
2306 (v4i32 (scalar_to_vector GR32:$src)),
2307 MOVL_shuffle_mask)))]>;
2308 // This is X86-64 only.
2309 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2310 "mov{d|q}\t{$src, $dst|$dst, $src}",
2312 (v2i64 (vector_shuffle immAllZerosV_bc,
2313 (v2i64 (scalar_to_vector GR64:$src)),
2314 MOVL_shuffle_mask)))]>;
2317 let AddedComplexity = 20 in {
2318 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2319 "movd\t{$src, $dst|$dst, $src}",
2321 (v4i32 (vector_shuffle immAllZerosV,
2322 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
2323 MOVL_shuffle_mask)))]>;
2324 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2325 "movq\t{$src, $dst|$dst, $src}",
2327 (v2i64 (vector_shuffle immAllZerosV_bc,
2328 (v2i64 (scalar_to_vector (loadi64 addr:$src))),
2329 MOVL_shuffle_mask)))]>, XS,
2330 Requires<[HasSSE2]>;
2333 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2334 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2335 let AddedComplexity = 15 in
2336 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2337 "movq\t{$src, $dst|$dst, $src}",
2338 [(set VR128:$dst, (v2i64 (vector_shuffle immAllZerosV_bc,
2340 MOVL_shuffle_mask)))]>,
2341 XS, Requires<[HasSSE2]>;
2343 let AddedComplexity = 20 in
2344 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2345 "movq\t{$src, $dst|$dst, $src}",
2346 [(set VR128:$dst, (v2i64 (vector_shuffle immAllZerosV_bc,
2347 (memopv2i64 addr:$src),
2348 MOVL_shuffle_mask)))]>,
2349 XS, Requires<[HasSSE2]>;
2351 //===----------------------------------------------------------------------===//
2352 // SSE3 Instructions
2353 //===----------------------------------------------------------------------===//
2355 // Move Instructions
2356 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2357 "movshdup\t{$src, $dst|$dst, $src}",
2358 [(set VR128:$dst, (v4f32 (vector_shuffle
2359 VR128:$src, (undef),
2360 MOVSHDUP_shuffle_mask)))]>;
2361 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2362 "movshdup\t{$src, $dst|$dst, $src}",
2363 [(set VR128:$dst, (v4f32 (vector_shuffle
2364 (memopv4f32 addr:$src), (undef),
2365 MOVSHDUP_shuffle_mask)))]>;
2367 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2368 "movsldup\t{$src, $dst|$dst, $src}",
2369 [(set VR128:$dst, (v4f32 (vector_shuffle
2370 VR128:$src, (undef),
2371 MOVSLDUP_shuffle_mask)))]>;
2372 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2373 "movsldup\t{$src, $dst|$dst, $src}",
2374 [(set VR128:$dst, (v4f32 (vector_shuffle
2375 (memopv4f32 addr:$src), (undef),
2376 MOVSLDUP_shuffle_mask)))]>;
2378 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2379 "movddup\t{$src, $dst|$dst, $src}",
2380 [(set VR128:$dst, (v2f64 (vector_shuffle
2381 VR128:$src, (undef),
2382 SSE_splat_lo_mask)))]>;
2383 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2384 "movddup\t{$src, $dst|$dst, $src}",
2386 (v2f64 (vector_shuffle
2387 (scalar_to_vector (loadf64 addr:$src)),
2389 SSE_splat_lo_mask)))]>;
2392 let Constraints = "$src1 = $dst" in {
2393 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2394 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2395 "addsubps\t{$src2, $dst|$dst, $src2}",
2396 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2398 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2399 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2400 "addsubps\t{$src2, $dst|$dst, $src2}",
2401 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2402 (load addr:$src2)))]>;
2403 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2404 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2405 "addsubpd\t{$src2, $dst|$dst, $src2}",
2406 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2408 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2409 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2410 "addsubpd\t{$src2, $dst|$dst, $src2}",
2411 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2412 (load addr:$src2)))]>;
2415 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2416 "lddqu\t{$src, $dst|$dst, $src}",
2417 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2420 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2421 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2422 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2423 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2424 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2425 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2426 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2427 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
2428 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2429 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2430 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2431 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2432 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2433 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2434 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2435 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
2437 let Constraints = "$src1 = $dst" in {
2438 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2439 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2440 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2441 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2442 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2443 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2444 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2445 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2448 // Thread synchronization
2449 def MONITOR : I<0xC8, RawFrm, (outs), (ins), "monitor",
2450 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2451 def MWAIT : I<0xC9, RawFrm, (outs), (ins), "mwait",
2452 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2454 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2455 let AddedComplexity = 15 in
2456 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2457 MOVSHDUP_shuffle_mask)),
2458 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2459 let AddedComplexity = 20 in
2460 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2461 MOVSHDUP_shuffle_mask)),
2462 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2464 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2465 let AddedComplexity = 15 in
2466 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2467 MOVSLDUP_shuffle_mask)),
2468 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2469 let AddedComplexity = 20 in
2470 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2471 MOVSLDUP_shuffle_mask)),
2472 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2474 //===----------------------------------------------------------------------===//
2475 // SSSE3 Instructions
2476 //===----------------------------------------------------------------------===//
2478 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2479 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2480 Intrinsic IntId64, Intrinsic IntId128> {
2481 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2482 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2483 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2485 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2486 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2488 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2490 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2492 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2493 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2496 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2498 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2501 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2504 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2505 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2506 Intrinsic IntId64, Intrinsic IntId128> {
2507 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2509 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2510 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2512 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2514 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2517 (bitconvert (memopv4i16 addr:$src))))]>;
2519 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2521 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2522 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2525 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2527 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2530 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2533 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2534 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2535 Intrinsic IntId64, Intrinsic IntId128> {
2536 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2538 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2539 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2541 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2543 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2546 (bitconvert (memopv2i32 addr:$src))))]>;
2548 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2550 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2551 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2554 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2556 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2559 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2562 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2563 int_x86_ssse3_pabs_b,
2564 int_x86_ssse3_pabs_b_128>;
2565 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2566 int_x86_ssse3_pabs_w,
2567 int_x86_ssse3_pabs_w_128>;
2568 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2569 int_x86_ssse3_pabs_d,
2570 int_x86_ssse3_pabs_d_128>;
2572 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2573 let Constraints = "$src1 = $dst" in {
2574 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2575 Intrinsic IntId64, Intrinsic IntId128,
2576 bit Commutable = 0> {
2577 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2578 (ins VR64:$src1, VR64:$src2),
2579 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2580 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2581 let isCommutable = Commutable;
2583 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2584 (ins VR64:$src1, i64mem:$src2),
2585 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2587 (IntId64 VR64:$src1,
2588 (bitconvert (memopv8i8 addr:$src2))))]>;
2590 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2591 (ins VR128:$src1, VR128:$src2),
2592 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2593 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2595 let isCommutable = Commutable;
2597 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2598 (ins VR128:$src1, i128mem:$src2),
2599 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2601 (IntId128 VR128:$src1,
2602 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2606 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2607 let Constraints = "$src1 = $dst" in {
2608 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2609 Intrinsic IntId64, Intrinsic IntId128,
2610 bit Commutable = 0> {
2611 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2612 (ins VR64:$src1, VR64:$src2),
2613 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2614 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2615 let isCommutable = Commutable;
2617 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2618 (ins VR64:$src1, i64mem:$src2),
2619 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2621 (IntId64 VR64:$src1,
2622 (bitconvert (memopv4i16 addr:$src2))))]>;
2624 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2625 (ins VR128:$src1, VR128:$src2),
2626 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2627 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2629 let isCommutable = Commutable;
2631 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2632 (ins VR128:$src1, i128mem:$src2),
2633 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2635 (IntId128 VR128:$src1,
2636 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2640 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2641 let Constraints = "$src1 = $dst" in {
2642 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2643 Intrinsic IntId64, Intrinsic IntId128,
2644 bit Commutable = 0> {
2645 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2646 (ins VR64:$src1, VR64:$src2),
2647 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2648 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2649 let isCommutable = Commutable;
2651 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2652 (ins VR64:$src1, i64mem:$src2),
2653 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2655 (IntId64 VR64:$src1,
2656 (bitconvert (memopv2i32 addr:$src2))))]>;
2658 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2659 (ins VR128:$src1, VR128:$src2),
2660 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2661 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2663 let isCommutable = Commutable;
2665 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2666 (ins VR128:$src1, i128mem:$src2),
2667 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2669 (IntId128 VR128:$src1,
2670 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2674 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2675 int_x86_ssse3_phadd_w,
2676 int_x86_ssse3_phadd_w_128, 1>;
2677 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2678 int_x86_ssse3_phadd_d,
2679 int_x86_ssse3_phadd_d_128, 1>;
2680 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2681 int_x86_ssse3_phadd_sw,
2682 int_x86_ssse3_phadd_sw_128, 1>;
2683 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2684 int_x86_ssse3_phsub_w,
2685 int_x86_ssse3_phsub_w_128>;
2686 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2687 int_x86_ssse3_phsub_d,
2688 int_x86_ssse3_phsub_d_128>;
2689 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2690 int_x86_ssse3_phsub_sw,
2691 int_x86_ssse3_phsub_sw_128>;
2692 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2693 int_x86_ssse3_pmadd_ub_sw,
2694 int_x86_ssse3_pmadd_ub_sw_128, 1>;
2695 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2696 int_x86_ssse3_pmul_hr_sw,
2697 int_x86_ssse3_pmul_hr_sw_128, 1>;
2698 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2699 int_x86_ssse3_pshuf_b,
2700 int_x86_ssse3_pshuf_b_128>;
2701 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2702 int_x86_ssse3_psign_b,
2703 int_x86_ssse3_psign_b_128>;
2704 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2705 int_x86_ssse3_psign_w,
2706 int_x86_ssse3_psign_w_128>;
2707 defm PSIGND : SS3I_binop_rm_int_32<0x09, "psignd",
2708 int_x86_ssse3_psign_d,
2709 int_x86_ssse3_psign_d_128>;
2711 let Constraints = "$src1 = $dst" in {
2712 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2713 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2714 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2716 (int_x86_ssse3_palign_r
2717 VR64:$src1, VR64:$src2,
2719 def PALIGNR64rm : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2720 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2721 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2723 (int_x86_ssse3_palign_r
2725 (bitconvert (memopv2i32 addr:$src2)),
2728 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2729 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2730 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2732 (int_x86_ssse3_palign_r_128
2733 VR128:$src1, VR128:$src2,
2734 imm:$src3))]>, OpSize;
2735 def PALIGNR128rm : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2736 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2737 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2739 (int_x86_ssse3_palign_r_128
2741 (bitconvert (memopv4i32 addr:$src2)),
2742 imm:$src3))]>, OpSize;
2745 //===----------------------------------------------------------------------===//
2746 // Non-Instruction Patterns
2747 //===----------------------------------------------------------------------===//
2749 // 128-bit vector undef's.
2750 def : Pat<(v4f32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2751 def : Pat<(v2f64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2752 def : Pat<(v16i8 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2753 def : Pat<(v8i16 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2754 def : Pat<(v4i32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2755 def : Pat<(v2i64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2757 // extload f32 -> f64. This matches load+fextend because we have a hack in
2758 // the isel (PreprocessForFPConvert) that can introduce loads after dag combine.
2759 // Since these loads aren't folded into the fextend, we have to match it
2761 let Predicates = [HasSSE2] in
2762 def : Pat<(fextend (loadf32 addr:$src)),
2763 (CVTSS2SDrm addr:$src)>;
2766 let Predicates = [HasSSE2] in {
2767 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2768 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2769 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2770 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2771 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2772 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2773 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2774 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2775 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2776 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2777 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2778 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2779 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2780 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2781 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2782 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2783 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2784 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2785 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2786 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2787 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2788 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2789 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2790 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2791 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2792 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2793 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2794 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2795 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2796 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2799 // Move scalar to XMM zero-extended
2800 // movd to XMM register zero-extends
2801 let AddedComplexity = 15 in {
2802 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2803 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc,
2804 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
2805 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2806 def : Pat<(v4f32 (vector_shuffle immAllZerosV_bc,
2807 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
2808 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
2811 // Splat v2f64 / v2i64
2812 let AddedComplexity = 10 in {
2813 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2814 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2815 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2816 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2817 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2818 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2819 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2820 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2824 def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SSE_splat_mask:$sm),
2825 (SHUFPSrri VR128:$src, VR128:$src, SSE_splat_mask:$sm)>,
2826 Requires<[HasSSE1]>;
2828 // Special unary SHUFPSrri case.
2829 // FIXME: when we want non two-address code, then we should use PSHUFD?
2830 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2831 SHUFP_unary_shuffle_mask:$sm)),
2832 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2833 Requires<[HasSSE1]>;
2834 // Special unary SHUFPDrri case.
2835 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (undef),
2836 SHUFP_unary_shuffle_mask:$sm)),
2837 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2838 Requires<[HasSSE2]>;
2839 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2840 def : Pat<(vector_shuffle (memopv4f32 addr:$src1), (undef),
2841 SHUFP_unary_shuffle_mask:$sm),
2842 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2843 Requires<[HasSSE2]>;
2844 // Special binary v4i32 shuffle cases with SHUFPS.
2845 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (v4i32 VR128:$src2),
2846 PSHUFD_binary_shuffle_mask:$sm)),
2847 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2848 Requires<[HasSSE2]>;
2849 def : Pat<(v4i32 (vector_shuffle VR128:$src1,
2850 (bc_v4i32 (memopv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm)),
2851 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2852 Requires<[HasSSE2]>;
2853 // Special binary v2i64 shuffle cases using SHUFPDrri.
2854 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2855 SHUFP_shuffle_mask:$sm)),
2856 (SHUFPDrri VR128:$src1, VR128:$src2, SHUFP_shuffle_mask:$sm)>,
2857 Requires<[HasSSE2]>;
2858 // Special unary SHUFPDrri case.
2859 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (undef),
2860 SHUFP_unary_shuffle_mask:$sm)),
2861 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2862 Requires<[HasSSE2]>;
2864 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2865 let AddedComplexity = 10 in {
2866 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2867 UNPCKL_v_undef_shuffle_mask)),
2868 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2869 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2870 UNPCKL_v_undef_shuffle_mask)),
2871 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2872 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2873 UNPCKL_v_undef_shuffle_mask)),
2874 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2875 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2876 UNPCKL_v_undef_shuffle_mask)),
2877 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2880 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2881 let AddedComplexity = 10 in {
2882 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2883 UNPCKH_v_undef_shuffle_mask)),
2884 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2885 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2886 UNPCKH_v_undef_shuffle_mask)),
2887 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2888 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2889 UNPCKH_v_undef_shuffle_mask)),
2890 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2891 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2892 UNPCKH_v_undef_shuffle_mask)),
2893 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2896 let AddedComplexity = 15 in {
2897 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2898 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2899 MOVHP_shuffle_mask)),
2900 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2902 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2903 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2904 MOVHLPS_shuffle_mask)),
2905 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2907 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2908 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2909 MOVHLPS_v_undef_shuffle_mask)),
2910 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2911 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2912 MOVHLPS_v_undef_shuffle_mask)),
2913 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2916 let AddedComplexity = 20 in {
2917 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2918 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2919 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2920 MOVLP_shuffle_mask)),
2921 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2922 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2923 MOVLP_shuffle_mask)),
2924 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2925 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2926 MOVHP_shuffle_mask)),
2927 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2928 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2929 MOVHP_shuffle_mask)),
2930 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2932 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2933 MOVLP_shuffle_mask)),
2934 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2935 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2936 MOVLP_shuffle_mask)),
2937 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2938 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2939 MOVHP_shuffle_mask)),
2940 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2941 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2942 MOVLP_shuffle_mask)),
2943 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2946 let AddedComplexity = 15 in {
2947 // Setting the lowest element in the vector.
2948 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2949 MOVL_shuffle_mask)),
2950 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2951 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2952 MOVL_shuffle_mask)),
2953 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2955 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2956 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2957 MOVLP_shuffle_mask)),
2958 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2959 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2960 MOVLP_shuffle_mask)),
2961 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2964 // Set lowest element and zero upper elements.
2965 let AddedComplexity = 15 in
2966 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc, VR128:$src,
2967 MOVL_shuffle_mask)),
2968 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
2971 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2972 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2973 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2974 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2975 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2976 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2977 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2978 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2979 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2980 Requires<[HasSSE2]>;
2981 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2982 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2983 Requires<[HasSSE2]>;
2984 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2985 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2986 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2987 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2988 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2989 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2990 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2991 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2992 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
2993 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2994 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
2995 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2996 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
2997 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2998 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
2999 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3001 // Some special case pandn patterns.
3002 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3004 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3005 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3007 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3008 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3010 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3012 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3013 (memopv2i64 addr:$src2))),
3014 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3015 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3016 (memopv2i64 addr:$src2))),
3017 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3018 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3019 (memopv2i64 addr:$src2))),
3020 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3022 // vector -> vector casts
3023 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3024 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3025 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3026 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3028 // Use movaps / movups for SSE integer load / store (one byte shorter).
3029 def : Pat<(alignedloadv4i32 addr:$src),
3030 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
3031 def : Pat<(loadv4i32 addr:$src),
3032 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3033 def : Pat<(alignedloadv2i64 addr:$src),
3034 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3035 def : Pat<(loadv2i64 addr:$src),
3036 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3038 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3039 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3040 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3041 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3042 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3043 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3044 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3045 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3046 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3047 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3048 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3049 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3050 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3051 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3052 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3053 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3055 //===----------------------------------------------------------------------===//
3056 // SSE4.1 Instructions
3057 //===----------------------------------------------------------------------===//
3059 multiclass sse41_fp_unop_rm<bits<8> opcss, bits<8> opcps,
3060 bits<8> opcsd, bits<8> opcpd,
3065 Intrinsic V2F64Int> {
3066 // Intrinsic operation, reg.
3067 def SSr_Int : SS4AI<opcss, MRMSrcReg,
3068 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3069 !strconcat(OpcodeStr,
3070 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3071 [(set VR128:$dst, (F32Int VR128:$src1, imm:$src2))]>,
3074 // Intrinsic operation, mem.
3075 def SSm_Int : SS4AI<opcss, MRMSrcMem,
3076 (outs VR128:$dst), (ins ssmem:$src1, i32i8imm:$src2),
3077 !strconcat(OpcodeStr,
3078 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3079 [(set VR128:$dst, (F32Int sse_load_f32:$src1, imm:$src2))]>,
3082 // Vector intrinsic operation, reg
3083 def PSr_Int : SS4AI<opcps, MRMSrcReg,
3084 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3085 !strconcat(OpcodeStr,
3086 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3087 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3090 // Vector intrinsic operation, mem
3091 def PSm_Int : SS4AI<opcps, MRMSrcMem,
3092 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3093 !strconcat(OpcodeStr,
3094 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3095 [(set VR128:$dst, (V4F32Int (load addr:$src1),imm:$src2))]>,
3098 // Intrinsic operation, reg.
3099 def SDr_Int : SS4AI<opcsd, MRMSrcReg,
3100 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3101 !strconcat(OpcodeStr,
3102 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3103 [(set VR128:$dst, (F64Int VR128:$src1, imm:$src2))]>,
3106 // Intrinsic operation, mem.
3107 def SDm_Int : SS4AI<opcsd, MRMSrcMem,
3108 (outs VR128:$dst), (ins sdmem:$src1, i32i8imm:$src2),
3109 !strconcat(OpcodeStr,
3110 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3111 [(set VR128:$dst, (F64Int sse_load_f64:$src1, imm:$src2))]>,
3114 // Vector intrinsic operation, reg
3115 def PDr_Int : SS4AI<opcpd, MRMSrcReg,
3116 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3117 !strconcat(OpcodeStr,
3118 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3119 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3122 // Vector intrinsic operation, mem
3123 def PDm_Int : SS4AI<opcpd, MRMSrcMem,
3124 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3125 !strconcat(OpcodeStr,
3126 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3127 [(set VR128:$dst, (V2F64Int (load addr:$src1),imm:$src2))]>,
3131 // FP round - roundss, roundps, roundsd, roundpd
3132 defm ROUND : sse41_fp_unop_rm<0x0A, 0x08, 0x0B, 0x09, "round",
3133 int_x86_sse41_round_ss, int_x86_sse41_round_ps,
3134 int_x86_sse41_round_sd, int_x86_sse41_round_pd>;
3136 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3137 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3138 Intrinsic IntId128> {
3139 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3141 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3142 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3143 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3145 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3148 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3151 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3152 int_x86_sse41_phminposuw>;
3154 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3155 let Constraints = "$src1 = $dst" in {
3156 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3157 Intrinsic IntId128, bit Commutable = 0> {
3158 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3159 (ins VR128:$src1, VR128:$src2),
3160 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3161 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3163 let isCommutable = Commutable;
3165 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3166 (ins VR128:$src1, i128mem:$src2),
3167 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3169 (IntId128 VR128:$src1,
3170 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3174 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3175 int_x86_sse41_pcmpeqq, 1>;
3176 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3177 int_x86_sse41_packusdw, 0>;
3178 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3179 int_x86_sse41_pminsb, 1>;
3180 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3181 int_x86_sse41_pminsd, 1>;
3182 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3183 int_x86_sse41_pminud, 1>;
3184 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3185 int_x86_sse41_pminuw, 1>;
3186 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3187 int_x86_sse41_pmaxsb, 1>;
3188 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3189 int_x86_sse41_pmaxsd, 1>;
3190 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3191 int_x86_sse41_pmaxud, 1>;
3192 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3193 int_x86_sse41_pmaxuw, 1>;
3194 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq",
3195 int_x86_sse41_pmuldq, 1>;
3198 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3199 let Constraints = "$src1 = $dst" in {
3200 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, SDNode OpNode,
3201 Intrinsic IntId128, bit Commutable = 0> {
3202 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3203 (ins VR128:$src1, VR128:$src2),
3204 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3205 [(set VR128:$dst, (OpNode (v4i32 VR128:$src1),
3206 VR128:$src2))]>, OpSize {
3207 let isCommutable = Commutable;
3209 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3210 (ins VR128:$src1, VR128:$src2),
3211 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3212 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3214 let isCommutable = Commutable;
3216 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3217 (ins VR128:$src1, i128mem:$src2),
3218 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3220 (OpNode VR128:$src1, (memopv4i32 addr:$src2)))]>, OpSize;
3221 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3222 (ins VR128:$src1, i128mem:$src2),
3223 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3225 (IntId128 VR128:$src1, (memopv4i32 addr:$src2)))]>,
3229 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", mul,
3230 int_x86_sse41_pmulld, 1>;
3233 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with immediate
3234 let Constraints = "$src1 = $dst" in {
3235 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3236 Intrinsic IntId128, bit Commutable = 0> {
3237 def rri : SS4AI<opc, MRMSrcReg, (outs VR128:$dst),
3238 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3239 !strconcat(OpcodeStr,
3240 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3242 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3244 let isCommutable = Commutable;
3246 def rmi : SS4AI<opc, MRMSrcMem, (outs VR128:$dst),
3247 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3248 !strconcat(OpcodeStr,
3249 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3251 (IntId128 VR128:$src1,
3252 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3257 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3258 int_x86_sse41_blendps, 0>;
3259 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3260 int_x86_sse41_blendpd, 0>;
3261 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3262 int_x86_sse41_pblendw, 0>;
3263 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3264 int_x86_sse41_dpps, 1>;
3265 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3266 int_x86_sse41_dppd, 1>;
3267 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3268 int_x86_sse41_mpsadbw, 0>;
3271 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with immediate
3272 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3273 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3274 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3275 (ins VR128:$src1, VR128:$src2),
3276 !strconcat(OpcodeStr,
3277 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3278 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3281 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3282 (ins VR128:$src1, i128mem:$src2),
3283 !strconcat(OpcodeStr,
3284 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3287 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3291 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3292 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3293 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3296 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3297 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3298 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3299 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3301 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3302 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3304 (IntId (bitconvert (v4i32 (load addr:$src)))))]>, OpSize;
3307 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3308 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3309 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3310 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3311 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3312 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3314 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3315 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3316 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3317 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3319 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3320 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3322 (IntId (bitconvert (v4i32 (load addr:$src)))))]>, OpSize;
3325 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3326 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3327 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3328 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3330 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3331 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3332 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3333 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3335 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3336 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3338 (IntId (bitconvert (v4i32 (load addr:$src)))))]>, OpSize;
3341 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3342 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovsxbq", int_x86_sse41_pmovzxbq>;
3345 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3346 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3347 def rr : SS4AI<opc, MRMSrcReg, (outs GR32:$dst),
3348 (ins VR128:$src1, i32i8imm:$src2),
3349 !strconcat(OpcodeStr,
3350 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3351 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3353 def mr : SS4AI<opc, MRMDestMem, (outs),
3354 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3355 !strconcat(OpcodeStr,
3356 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3359 // There's an AssertZext in the way of writing the store pattern
3360 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3363 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3366 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3367 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3368 def mr : SS4AI<opc, MRMDestMem, (outs),
3369 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3370 !strconcat(OpcodeStr,
3371 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3374 // There's an AssertZext in the way of writing the store pattern
3375 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3378 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3381 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3382 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3383 def rr : SS4AI<opc, MRMSrcReg, (outs GR32:$dst),
3384 (ins VR128:$src1, i32i8imm:$src2),
3385 !strconcat(OpcodeStr,
3386 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3388 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3389 def mr : SS4AI<opc, MRMDestMem, (outs),
3390 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3391 !strconcat(OpcodeStr,
3392 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3393 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3394 addr:$dst)]>, OpSize;
3397 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3400 /// SS41I_extractf32 - SSE 4.1 extract 32 bits to fp reg or memory destination
3401 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3402 def rr : SS4AI<opc, MRMSrcReg, (outs FR32:$dst),
3403 (ins VR128:$src1, i32i8imm:$src2),
3404 !strconcat(OpcodeStr,
3405 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3407 (extractelt (v4f32 VR128:$src1), imm:$src2))]>, OpSize;
3408 def mr : SS4AI<opc, MRMDestMem, (outs),
3409 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3410 !strconcat(OpcodeStr,
3411 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3412 [(store (extractelt (v4f32 VR128:$src1), imm:$src2),
3413 addr:$dst)]>, OpSize;
3416 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3418 let Constraints = "$src1 = $dst" in {
3419 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3420 def rr : SS4AI<opc, MRMSrcReg, (outs VR128:$dst),
3421 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3422 !strconcat(OpcodeStr,
3423 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3425 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3426 def rm : SS4AI<opc, MRMSrcMem, (outs VR128:$dst),
3427 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3428 !strconcat(OpcodeStr,
3429 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3431 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3432 imm:$src3))]>, OpSize;
3436 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3438 let Constraints = "$src1 = $dst" in {
3439 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3440 def rr : SS4AI<opc, MRMSrcReg, (outs VR128:$dst),
3441 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3442 !strconcat(OpcodeStr,
3443 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3445 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3447 def rm : SS4AI<opc, MRMSrcMem, (outs VR128:$dst),
3448 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3449 !strconcat(OpcodeStr,
3450 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3452 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3453 imm:$src3)))]>, OpSize;
3457 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3459 let Constraints = "$src1 = $dst" in {
3460 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3461 def rr : SS4AI<opc, MRMSrcReg, (outs VR128:$dst),
3462 (ins VR128:$src1, FR32:$src2, i32i8imm:$src3),
3463 !strconcat(OpcodeStr,
3464 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3466 (X86insrtps VR128:$src1, FR32:$src2, imm:$src3))]>, OpSize;
3467 def rm : SS4AI<opc, MRMSrcMem, (outs VR128:$dst),
3468 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3469 !strconcat(OpcodeStr,
3470 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3472 (X86insrtps VR128:$src1, (loadf32 addr:$src2),
3473 imm:$src3))]>, OpSize;
3477 defm INSERTPS : SS41I_insertf32<0x31, "insertps">;