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 Complex Patterns
53 //===----------------------------------------------------------------------===//
55 // These are 'extloads' from a scalar to the low element of a vector, zeroing
56 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
58 def sse_load_f32 : ComplexPattern<v4f32, 4, "SelectScalarSSELoad", [],
59 [SDNPHasChain, SDNPMayLoad]>;
60 def sse_load_f64 : ComplexPattern<v2f64, 4, "SelectScalarSSELoad", [],
61 [SDNPHasChain, SDNPMayLoad]>;
63 def ssmem : Operand<v4f32> {
64 let PrintMethod = "printf32mem";
65 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
67 def sdmem : Operand<v2f64> {
68 let PrintMethod = "printf64mem";
69 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
72 //===----------------------------------------------------------------------===//
73 // SSE pattern fragments
74 //===----------------------------------------------------------------------===//
76 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
77 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
78 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
79 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
81 // Like 'store', but always requires vector alignment.
82 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
83 (st node:$val, node:$ptr), [{
84 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
85 return !ST->isTruncatingStore() &&
86 ST->getAddressingMode() == ISD::UNINDEXED &&
87 ST->getAlignment() >= 16;
91 // Like 'load', but always requires vector alignment.
92 def alignedload : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
93 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
94 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
95 LD->getAddressingMode() == ISD::UNINDEXED &&
96 LD->getAlignment() >= 16;
100 def alignedloadfsf32 : PatFrag<(ops node:$ptr), (f32 (alignedload node:$ptr))>;
101 def alignedloadfsf64 : PatFrag<(ops node:$ptr), (f64 (alignedload node:$ptr))>;
102 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
103 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
104 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
105 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
107 // Like 'load', but uses special alignment checks suitable for use in
108 // memory operands in most SSE instructions, which are required to
109 // be naturally aligned on some targets but not on others.
110 // FIXME: Actually implement support for targets that don't require the
111 // alignment. This probably wants a subtarget predicate.
112 def memop : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
113 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
114 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
115 LD->getAddressingMode() == ISD::UNINDEXED &&
116 LD->getAlignment() >= 16;
120 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
121 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
122 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
123 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
124 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
125 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
126 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
128 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
130 // FIXME: 8 byte alignment for mmx reads is not required
131 def memop64 : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
132 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
133 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
134 LD->getAddressingMode() == ISD::UNINDEXED &&
135 LD->getAlignment() >= 8;
139 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
140 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
141 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
142 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
144 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
145 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
146 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
147 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
148 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
149 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
151 def fp32imm0 : PatLeaf<(f32 fpimm), [{
152 return N->isExactlyValue(+0.0);
155 def PSxLDQ_imm : SDNodeXForm<imm, [{
156 // Transformation function: imm >> 3
157 return getI32Imm(N->getValue() >> 3);
160 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
162 def SHUFFLE_get_shuf_imm : SDNodeXForm<build_vector, [{
163 return getI8Imm(X86::getShuffleSHUFImmediate(N));
166 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
168 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<build_vector, [{
169 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
172 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
174 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<build_vector, [{
175 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
178 def SSE_splat_mask : PatLeaf<(build_vector), [{
179 return X86::isSplatMask(N);
180 }], SHUFFLE_get_shuf_imm>;
182 def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
183 return X86::isSplatLoMask(N);
186 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
187 return X86::isMOVHLPSMask(N);
190 def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
191 return X86::isMOVHLPS_v_undef_Mask(N);
194 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
195 return X86::isMOVHPMask(N);
198 def MOVLP_shuffle_mask : PatLeaf<(build_vector), [{
199 return X86::isMOVLPMask(N);
202 def MOVL_shuffle_mask : PatLeaf<(build_vector), [{
203 return X86::isMOVLMask(N);
206 def MOVSHDUP_shuffle_mask : PatLeaf<(build_vector), [{
207 return X86::isMOVSHDUPMask(N);
210 def MOVSLDUP_shuffle_mask : PatLeaf<(build_vector), [{
211 return X86::isMOVSLDUPMask(N);
214 def UNPCKL_shuffle_mask : PatLeaf<(build_vector), [{
215 return X86::isUNPCKLMask(N);
218 def UNPCKH_shuffle_mask : PatLeaf<(build_vector), [{
219 return X86::isUNPCKHMask(N);
222 def UNPCKL_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
223 return X86::isUNPCKL_v_undef_Mask(N);
226 def UNPCKH_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
227 return X86::isUNPCKH_v_undef_Mask(N);
230 def PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
231 return X86::isPSHUFDMask(N);
232 }], SHUFFLE_get_shuf_imm>;
234 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
235 return X86::isPSHUFHWMask(N);
236 }], SHUFFLE_get_pshufhw_imm>;
238 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
239 return X86::isPSHUFLWMask(N);
240 }], SHUFFLE_get_pshuflw_imm>;
242 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
243 return X86::isPSHUFDMask(N);
244 }], SHUFFLE_get_shuf_imm>;
246 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
247 return X86::isSHUFPMask(N);
248 }], SHUFFLE_get_shuf_imm>;
250 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
251 return X86::isSHUFPMask(N);
252 }], SHUFFLE_get_shuf_imm>;
254 //===----------------------------------------------------------------------===//
255 // SSE scalar FP Instructions
256 //===----------------------------------------------------------------------===//
258 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
259 // scheduler into a branch sequence.
260 // These are expanded by the scheduler.
261 let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
262 def CMOV_FR32 : I<0, Pseudo,
263 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
264 "#CMOV_FR32 PSEUDO!",
265 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
267 def CMOV_FR64 : I<0, Pseudo,
268 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
269 "#CMOV_FR64 PSEUDO!",
270 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
272 def CMOV_V4F32 : I<0, Pseudo,
273 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
274 "#CMOV_V4F32 PSEUDO!",
276 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
278 def CMOV_V2F64 : I<0, Pseudo,
279 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
280 "#CMOV_V2F64 PSEUDO!",
282 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
284 def CMOV_V2I64 : I<0, Pseudo,
285 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
286 "#CMOV_V2I64 PSEUDO!",
288 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
292 //===----------------------------------------------------------------------===//
294 //===----------------------------------------------------------------------===//
297 let neverHasSideEffects = 1 in
298 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
299 "movss\t{$src, $dst|$dst, $src}", []>;
300 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
301 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
302 "movss\t{$src, $dst|$dst, $src}",
303 [(set FR32:$dst, (loadf32 addr:$src))]>;
304 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
305 "movss\t{$src, $dst|$dst, $src}",
306 [(store FR32:$src, addr:$dst)]>;
308 // Conversion instructions
309 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
310 "cvttss2si\t{$src, $dst|$dst, $src}",
311 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
312 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
313 "cvttss2si\t{$src, $dst|$dst, $src}",
314 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
315 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
316 "cvtsi2ss\t{$src, $dst|$dst, $src}",
317 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
318 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
319 "cvtsi2ss\t{$src, $dst|$dst, $src}",
320 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
322 // Match intrinsics which expect XMM operand(s).
323 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
324 "cvtss2si\t{$src, $dst|$dst, $src}",
325 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
326 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
327 "cvtss2si\t{$src, $dst|$dst, $src}",
328 [(set GR32:$dst, (int_x86_sse_cvtss2si
329 (load addr:$src)))]>;
331 // Match intrinisics which expect MM and XMM operand(s).
332 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
333 "cvtps2pi\t{$src, $dst|$dst, $src}",
334 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
335 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
336 "cvtps2pi\t{$src, $dst|$dst, $src}",
337 [(set VR64:$dst, (int_x86_sse_cvtps2pi
338 (load addr:$src)))]>;
339 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
340 "cvttps2pi\t{$src, $dst|$dst, $src}",
341 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
342 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
343 "cvttps2pi\t{$src, $dst|$dst, $src}",
344 [(set VR64:$dst, (int_x86_sse_cvttps2pi
345 (load addr:$src)))]>;
346 let Constraints = "$src1 = $dst" in {
347 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
348 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
349 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
350 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
352 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
353 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
354 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
355 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
356 (load addr:$src2)))]>;
359 // Aliases for intrinsics
360 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
361 "cvttss2si\t{$src, $dst|$dst, $src}",
363 (int_x86_sse_cvttss2si VR128:$src))]>;
364 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
365 "cvttss2si\t{$src, $dst|$dst, $src}",
367 (int_x86_sse_cvttss2si(load addr:$src)))]>;
369 let Constraints = "$src1 = $dst" in {
370 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
371 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
372 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
373 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
375 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
376 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
377 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
378 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
379 (loadi32 addr:$src2)))]>;
382 // Comparison instructions
383 let Constraints = "$src1 = $dst" in {
384 let neverHasSideEffects = 1 in
385 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
386 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
387 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
388 let neverHasSideEffects = 1, mayLoad = 1 in
389 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
390 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
391 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
394 let Defs = [EFLAGS] in {
395 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
396 "ucomiss\t{$src2, $src1|$src1, $src2}",
397 [(X86cmp FR32:$src1, FR32:$src2), (implicit EFLAGS)]>;
398 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
399 "ucomiss\t{$src2, $src1|$src1, $src2}",
400 [(X86cmp FR32:$src1, (loadf32 addr:$src2)),
404 // Aliases to match intrinsics which expect XMM operand(s).
405 let Constraints = "$src1 = $dst" in {
406 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
407 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
408 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
409 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
410 VR128:$src, imm:$cc))]>;
411 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
412 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src, SSECC:$cc),
413 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
414 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
415 (load addr:$src), imm:$cc))]>;
418 let Defs = [EFLAGS] in {
419 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs),
420 (ins VR128:$src1, VR128:$src2),
421 "ucomiss\t{$src2, $src1|$src1, $src2}",
422 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2),
424 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),
425 (ins VR128:$src1, f128mem:$src2),
426 "ucomiss\t{$src2, $src1|$src1, $src2}",
427 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2)),
430 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs),
431 (ins VR128:$src1, VR128:$src2),
432 "comiss\t{$src2, $src1|$src1, $src2}",
433 [(X86comi (v4f32 VR128:$src1), VR128:$src2),
435 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs),
436 (ins VR128:$src1, f128mem:$src2),
437 "comiss\t{$src2, $src1|$src1, $src2}",
438 [(X86comi (v4f32 VR128:$src1), (load addr:$src2)),
442 // Aliases of packed SSE1 instructions for scalar use. These all have names that
445 // Alias instructions that map fld0 to pxor for sse.
446 let isReMaterializable = 1 in
447 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
448 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
449 Requires<[HasSSE1]>, TB, OpSize;
451 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
453 let neverHasSideEffects = 1 in
454 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
455 "movaps\t{$src, $dst|$dst, $src}", []>;
457 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
459 let isSimpleLoad = 1 in
460 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
461 "movaps\t{$src, $dst|$dst, $src}",
462 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
464 // Alias bitwise logical operations using SSE logical ops on packed FP values.
465 let Constraints = "$src1 = $dst" in {
466 let isCommutable = 1 in {
467 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
468 "andps\t{$src2, $dst|$dst, $src2}",
469 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
470 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
471 "orps\t{$src2, $dst|$dst, $src2}",
472 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
473 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
474 "xorps\t{$src2, $dst|$dst, $src2}",
475 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
478 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
479 "andps\t{$src2, $dst|$dst, $src2}",
480 [(set FR32:$dst, (X86fand FR32:$src1,
481 (memopfsf32 addr:$src2)))]>;
482 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
483 "orps\t{$src2, $dst|$dst, $src2}",
484 [(set FR32:$dst, (X86for FR32:$src1,
485 (memopfsf32 addr:$src2)))]>;
486 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
487 "xorps\t{$src2, $dst|$dst, $src2}",
488 [(set FR32:$dst, (X86fxor FR32:$src1,
489 (memopfsf32 addr:$src2)))]>;
490 let neverHasSideEffects = 1 in {
491 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
492 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
493 "andnps\t{$src2, $dst|$dst, $src2}", []>;
496 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
497 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
498 "andnps\t{$src2, $dst|$dst, $src2}", []>;
502 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
504 /// In addition, we also have a special variant of the scalar form here to
505 /// represent the associated intrinsic operation. This form is unlike the
506 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
507 /// and leaves the top elements undefined.
509 /// These three forms can each be reg+reg or reg+mem, so there are a total of
510 /// six "instructions".
512 let Constraints = "$src1 = $dst" in {
513 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
514 SDNode OpNode, Intrinsic F32Int,
515 bit Commutable = 0> {
516 // Scalar operation, reg+reg.
517 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
518 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
519 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
520 let isCommutable = Commutable;
523 // Scalar operation, reg+mem.
524 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
525 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
526 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
528 // Vector operation, reg+reg.
529 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
530 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
531 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
532 let isCommutable = Commutable;
535 // Vector operation, reg+mem.
536 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
537 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
538 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
540 // Intrinsic operation, reg+reg.
541 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
542 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
543 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
544 let isCommutable = Commutable;
547 // Intrinsic operation, reg+mem.
548 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
549 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
550 [(set VR128:$dst, (F32Int VR128:$src1,
551 sse_load_f32:$src2))]>;
555 // Arithmetic instructions
556 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
557 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
558 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
559 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
561 /// sse1_fp_binop_rm - Other SSE1 binops
563 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
564 /// instructions for a full-vector intrinsic form. Operations that map
565 /// onto C operators don't use this form since they just use the plain
566 /// vector form instead of having a separate vector intrinsic form.
568 /// This provides a total of eight "instructions".
570 let Constraints = "$src1 = $dst" in {
571 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
575 bit Commutable = 0> {
577 // Scalar operation, reg+reg.
578 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
579 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
580 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
581 let isCommutable = Commutable;
584 // Scalar operation, reg+mem.
585 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
586 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
587 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
589 // Vector operation, reg+reg.
590 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
591 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
592 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
593 let isCommutable = Commutable;
596 // Vector operation, reg+mem.
597 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
598 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
599 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
601 // Intrinsic operation, reg+reg.
602 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
603 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
604 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
605 let isCommutable = Commutable;
608 // Intrinsic operation, reg+mem.
609 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
610 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
611 [(set VR128:$dst, (F32Int VR128:$src1,
612 sse_load_f32:$src2))]>;
614 // Vector intrinsic operation, reg+reg.
615 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
616 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
617 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
618 let isCommutable = Commutable;
621 // Vector intrinsic operation, reg+mem.
622 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
623 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
624 [(set VR128:$dst, (V4F32Int VR128:$src1, (load addr:$src2)))]>;
628 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
629 int_x86_sse_max_ss, int_x86_sse_max_ps>;
630 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
631 int_x86_sse_min_ss, int_x86_sse_min_ps>;
633 //===----------------------------------------------------------------------===//
634 // SSE packed FP Instructions
637 let neverHasSideEffects = 1 in
638 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
639 "movaps\t{$src, $dst|$dst, $src}", []>;
640 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
641 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
642 "movaps\t{$src, $dst|$dst, $src}",
643 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
645 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
646 "movaps\t{$src, $dst|$dst, $src}",
647 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
649 let neverHasSideEffects = 1 in
650 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
651 "movups\t{$src, $dst|$dst, $src}", []>;
652 let isSimpleLoad = 1 in
653 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
654 "movups\t{$src, $dst|$dst, $src}",
655 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
656 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
657 "movups\t{$src, $dst|$dst, $src}",
658 [(store (v4f32 VR128:$src), addr:$dst)]>;
660 // Intrinsic forms of MOVUPS load and store
661 let isSimpleLoad = 1 in
662 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
663 "movups\t{$src, $dst|$dst, $src}",
664 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
665 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
666 "movups\t{$src, $dst|$dst, $src}",
667 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
669 let Constraints = "$src1 = $dst" in {
670 let AddedComplexity = 20 in {
671 def MOVLPSrm : PSI<0x12, MRMSrcMem,
672 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
673 "movlps\t{$src2, $dst|$dst, $src2}",
675 (v4f32 (vector_shuffle VR128:$src1,
676 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
677 MOVLP_shuffle_mask)))]>;
678 def MOVHPSrm : PSI<0x16, MRMSrcMem,
679 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
680 "movhps\t{$src2, $dst|$dst, $src2}",
682 (v4f32 (vector_shuffle VR128:$src1,
683 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
684 MOVHP_shuffle_mask)))]>;
686 } // Constraints = "$src1 = $dst"
688 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
689 "movlps\t{$src, $dst|$dst, $src}",
690 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
691 (iPTR 0))), addr:$dst)]>;
693 // v2f64 extract element 1 is always custom lowered to unpack high to low
694 // and extract element 0 so the non-store version isn't too horrible.
695 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
696 "movhps\t{$src, $dst|$dst, $src}",
697 [(store (f64 (vector_extract
698 (v2f64 (vector_shuffle
699 (bc_v2f64 (v4f32 VR128:$src)), (undef),
700 UNPCKH_shuffle_mask)), (iPTR 0))),
703 let Constraints = "$src1 = $dst" in {
704 let AddedComplexity = 15 in {
705 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
706 "movlhps\t{$src2, $dst|$dst, $src2}",
708 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
709 MOVHP_shuffle_mask)))]>;
711 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
712 "movhlps\t{$src2, $dst|$dst, $src2}",
714 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
715 MOVHLPS_shuffle_mask)))]>;
717 } // Constraints = "$src1 = $dst"
723 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
725 /// In addition, we also have a special variant of the scalar form here to
726 /// represent the associated intrinsic operation. This form is unlike the
727 /// plain scalar form, in that it takes an entire vector (instead of a
728 /// scalar) and leaves the top elements undefined.
730 /// And, we have a special variant form for a full-vector intrinsic form.
732 /// These four forms can each have a reg or a mem operand, so there are a
733 /// total of eight "instructions".
735 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
739 bit Commutable = 0> {
740 // Scalar operation, reg.
741 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
742 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
743 [(set FR32:$dst, (OpNode FR32:$src))]> {
744 let isCommutable = Commutable;
747 // Scalar operation, mem.
748 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
749 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
750 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
752 // Vector operation, reg.
753 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
754 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
755 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
756 let isCommutable = Commutable;
759 // Vector operation, mem.
760 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
761 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
762 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
764 // Intrinsic operation, reg.
765 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
766 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
767 [(set VR128:$dst, (F32Int VR128:$src))]> {
768 let isCommutable = Commutable;
771 // Intrinsic operation, mem.
772 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
773 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
774 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
776 // Vector intrinsic operation, reg
777 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
778 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
779 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
780 let isCommutable = Commutable;
783 // Vector intrinsic operation, mem
784 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
785 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
786 [(set VR128:$dst, (V4F32Int (load addr:$src)))]>;
790 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
791 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
793 // Reciprocal approximations. Note that these typically require refinement
794 // in order to obtain suitable precision.
795 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
796 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
797 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
798 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
801 let Constraints = "$src1 = $dst" in {
802 let isCommutable = 1 in {
803 def ANDPSrr : PSI<0x54, MRMSrcReg,
804 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
805 "andps\t{$src2, $dst|$dst, $src2}",
806 [(set VR128:$dst, (v2i64
807 (and VR128:$src1, VR128:$src2)))]>;
808 def ORPSrr : PSI<0x56, MRMSrcReg,
809 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
810 "orps\t{$src2, $dst|$dst, $src2}",
811 [(set VR128:$dst, (v2i64
812 (or VR128:$src1, VR128:$src2)))]>;
813 def XORPSrr : PSI<0x57, MRMSrcReg,
814 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
815 "xorps\t{$src2, $dst|$dst, $src2}",
816 [(set VR128:$dst, (v2i64
817 (xor VR128:$src1, VR128:$src2)))]>;
820 def ANDPSrm : PSI<0x54, MRMSrcMem,
821 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
822 "andps\t{$src2, $dst|$dst, $src2}",
823 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
824 (memopv2i64 addr:$src2)))]>;
825 def ORPSrm : PSI<0x56, MRMSrcMem,
826 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
827 "orps\t{$src2, $dst|$dst, $src2}",
828 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
829 (memopv2i64 addr:$src2)))]>;
830 def XORPSrm : PSI<0x57, MRMSrcMem,
831 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
832 "xorps\t{$src2, $dst|$dst, $src2}",
833 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
834 (memopv2i64 addr:$src2)))]>;
835 def ANDNPSrr : PSI<0x55, MRMSrcReg,
836 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
837 "andnps\t{$src2, $dst|$dst, $src2}",
839 (v2i64 (and (xor VR128:$src1,
840 (bc_v2i64 (v4i32 immAllOnesV))),
842 def ANDNPSrm : PSI<0x55, MRMSrcMem,
843 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
844 "andnps\t{$src2, $dst|$dst, $src2}",
846 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
847 (bc_v2i64 (v4i32 immAllOnesV))),
848 (memopv2i64 addr:$src2))))]>;
851 let Constraints = "$src1 = $dst" in {
852 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
853 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
854 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
855 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
856 VR128:$src, imm:$cc))]>;
857 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
858 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
859 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
860 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
861 (load addr:$src), imm:$cc))]>;
864 // Shuffle and unpack instructions
865 let Constraints = "$src1 = $dst" in {
866 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
867 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
868 (outs VR128:$dst), (ins VR128:$src1,
869 VR128:$src2, i32i8imm:$src3),
870 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
872 (v4f32 (vector_shuffle
873 VR128:$src1, VR128:$src2,
874 SHUFP_shuffle_mask:$src3)))]>;
875 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
876 (outs VR128:$dst), (ins VR128:$src1,
877 f128mem:$src2, i32i8imm:$src3),
878 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
880 (v4f32 (vector_shuffle
881 VR128:$src1, (memopv4f32 addr:$src2),
882 SHUFP_shuffle_mask:$src3)))]>;
884 let AddedComplexity = 10 in {
885 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
886 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
887 "unpckhps\t{$src2, $dst|$dst, $src2}",
889 (v4f32 (vector_shuffle
890 VR128:$src1, VR128:$src2,
891 UNPCKH_shuffle_mask)))]>;
892 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
893 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
894 "unpckhps\t{$src2, $dst|$dst, $src2}",
896 (v4f32 (vector_shuffle
897 VR128:$src1, (memopv4f32 addr:$src2),
898 UNPCKH_shuffle_mask)))]>;
900 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
901 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
902 "unpcklps\t{$src2, $dst|$dst, $src2}",
904 (v4f32 (vector_shuffle
905 VR128:$src1, VR128:$src2,
906 UNPCKL_shuffle_mask)))]>;
907 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
908 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
909 "unpcklps\t{$src2, $dst|$dst, $src2}",
911 (v4f32 (vector_shuffle
912 VR128:$src1, (memopv4f32 addr:$src2),
913 UNPCKL_shuffle_mask)))]>;
915 } // Constraints = "$src1 = $dst"
918 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
919 "movmskps\t{$src, $dst|$dst, $src}",
920 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
921 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
922 "movmskpd\t{$src, $dst|$dst, $src}",
923 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
925 // Prefetch intrinsic.
926 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
927 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
928 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
929 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
930 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
931 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
932 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
933 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
935 // Non-temporal stores
936 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
937 "movntps\t{$src, $dst|$dst, $src}",
938 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
940 // Load, store, and memory fence
941 def SFENCE : PSI<0xAE, MRM7m, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
944 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
945 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
946 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
947 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
949 // Alias instructions that map zero vector to pxor / xorp* for sse.
950 let isReMaterializable = 1 in
951 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
953 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
955 let Predicates = [HasSSE1] in {
956 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
957 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
958 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
959 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
960 def : Pat<(v4f32 immAllZerosV), (V_SET0)>;
963 // FR32 to 128-bit vector conversion.
964 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
965 "movss\t{$src, $dst|$dst, $src}",
967 (v4f32 (scalar_to_vector FR32:$src)))]>;
968 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
969 "movss\t{$src, $dst|$dst, $src}",
971 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
973 // FIXME: may not be able to eliminate this movss with coalescing the src and
974 // dest register classes are different. We really want to write this pattern
976 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
978 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
979 "movss\t{$src, $dst|$dst, $src}",
980 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
982 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
983 "movss\t{$src, $dst|$dst, $src}",
984 [(store (f32 (vector_extract (v4f32 VR128:$src),
985 (iPTR 0))), addr:$dst)]>;
988 // Move to lower bits of a VR128, leaving upper bits alone.
989 // Three operand (but two address) aliases.
990 let Constraints = "$src1 = $dst" in {
991 let neverHasSideEffects = 1 in
992 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
993 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
994 "movss\t{$src2, $dst|$dst, $src2}", []>;
996 let AddedComplexity = 15 in
997 def MOVLPSrr : SSI<0x10, MRMSrcReg,
998 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
999 "movss\t{$src2, $dst|$dst, $src2}",
1001 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
1002 MOVL_shuffle_mask)))]>;
1005 // Move to lower bits of a VR128 and zeroing upper bits.
1006 // Loading from memory automatically zeroing upper bits.
1007 let AddedComplexity = 20 in
1008 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
1009 "movss\t{$src, $dst|$dst, $src}",
1010 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV_bc,
1011 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
1012 MOVL_shuffle_mask)))]>;
1015 //===----------------------------------------------------------------------===//
1016 // SSE2 Instructions
1017 //===----------------------------------------------------------------------===//
1019 // Move Instructions
1020 let neverHasSideEffects = 1 in
1021 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1022 "movsd\t{$src, $dst|$dst, $src}", []>;
1023 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1024 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1025 "movsd\t{$src, $dst|$dst, $src}",
1026 [(set FR64:$dst, (loadf64 addr:$src))]>;
1027 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1028 "movsd\t{$src, $dst|$dst, $src}",
1029 [(store FR64:$src, addr:$dst)]>;
1031 // Conversion instructions
1032 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1033 "cvttsd2si\t{$src, $dst|$dst, $src}",
1034 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1035 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1036 "cvttsd2si\t{$src, $dst|$dst, $src}",
1037 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1038 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1039 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1040 [(set FR32:$dst, (fround FR64:$src))]>;
1041 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1042 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1043 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
1044 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1045 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1046 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1047 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1048 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1049 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1051 // SSE2 instructions with XS prefix
1052 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1053 "cvtss2sd\t{$src, $dst|$dst, $src}",
1054 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1055 Requires<[HasSSE2]>;
1056 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1057 "cvtss2sd\t{$src, $dst|$dst, $src}",
1058 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1059 Requires<[HasSSE2]>;
1061 // Match intrinsics which expect XMM operand(s).
1062 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1063 "cvtsd2si\t{$src, $dst|$dst, $src}",
1064 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1065 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1066 "cvtsd2si\t{$src, $dst|$dst, $src}",
1067 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1068 (load addr:$src)))]>;
1070 // Match intrinisics which expect MM and XMM operand(s).
1071 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1072 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1073 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1074 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1075 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1076 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1077 (load addr:$src)))]>;
1078 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1079 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1080 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1081 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1082 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1083 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1084 (load addr:$src)))]>;
1085 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1086 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1087 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1088 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1089 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1090 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1091 (load addr:$src)))]>;
1093 // Aliases for intrinsics
1094 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1095 "cvttsd2si\t{$src, $dst|$dst, $src}",
1097 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1098 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1099 "cvttsd2si\t{$src, $dst|$dst, $src}",
1100 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1101 (load addr:$src)))]>;
1103 // Comparison instructions
1104 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1105 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1106 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1107 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1109 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1110 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1111 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1114 let Defs = [EFLAGS] in {
1115 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1116 "ucomisd\t{$src2, $src1|$src1, $src2}",
1117 [(X86cmp FR64:$src1, FR64:$src2), (implicit EFLAGS)]>;
1118 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1119 "ucomisd\t{$src2, $src1|$src1, $src2}",
1120 [(X86cmp FR64:$src1, (loadf64 addr:$src2)),
1121 (implicit EFLAGS)]>;
1124 // Aliases to match intrinsics which expect XMM operand(s).
1125 let Constraints = "$src1 = $dst" in {
1126 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1127 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1128 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1129 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1130 VR128:$src, imm:$cc))]>;
1131 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1132 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1133 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1134 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1135 (load addr:$src), imm:$cc))]>;
1138 let Defs = [EFLAGS] in {
1139 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1140 "ucomisd\t{$src2, $src1|$src1, $src2}",
1141 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1142 (implicit EFLAGS)]>;
1143 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1144 "ucomisd\t{$src2, $src1|$src1, $src2}",
1145 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2)),
1146 (implicit EFLAGS)]>;
1148 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1149 "comisd\t{$src2, $src1|$src1, $src2}",
1150 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2)),
1151 (implicit EFLAGS)]>;
1152 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1153 "comisd\t{$src2, $src1|$src1, $src2}",
1154 [(X86comi (v2f64 VR128:$src1), (load addr:$src2)),
1155 (implicit EFLAGS)]>;
1158 // Aliases of packed SSE2 instructions for scalar use. These all have names that
1161 // Alias instructions that map fld0 to pxor for sse.
1162 let isReMaterializable = 1 in
1163 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1164 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1165 Requires<[HasSSE2]>, TB, OpSize;
1167 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1169 let neverHasSideEffects = 1 in
1170 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1171 "movapd\t{$src, $dst|$dst, $src}", []>;
1173 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1175 let isSimpleLoad = 1 in
1176 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1177 "movapd\t{$src, $dst|$dst, $src}",
1178 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1180 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1181 let Constraints = "$src1 = $dst" in {
1182 let isCommutable = 1 in {
1183 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst),
1184 (ins FR64:$src1, FR64:$src2),
1185 "andpd\t{$src2, $dst|$dst, $src2}",
1186 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1187 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst),
1188 (ins FR64:$src1, FR64:$src2),
1189 "orpd\t{$src2, $dst|$dst, $src2}",
1190 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1191 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst),
1192 (ins FR64:$src1, FR64:$src2),
1193 "xorpd\t{$src2, $dst|$dst, $src2}",
1194 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1197 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst),
1198 (ins FR64:$src1, f128mem:$src2),
1199 "andpd\t{$src2, $dst|$dst, $src2}",
1200 [(set FR64:$dst, (X86fand FR64:$src1,
1201 (memopfsf64 addr:$src2)))]>;
1202 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst),
1203 (ins FR64:$src1, f128mem:$src2),
1204 "orpd\t{$src2, $dst|$dst, $src2}",
1205 [(set FR64:$dst, (X86for FR64:$src1,
1206 (memopfsf64 addr:$src2)))]>;
1207 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst),
1208 (ins FR64:$src1, f128mem:$src2),
1209 "xorpd\t{$src2, $dst|$dst, $src2}",
1210 [(set FR64:$dst, (X86fxor FR64:$src1,
1211 (memopfsf64 addr:$src2)))]>;
1213 let neverHasSideEffects = 1 in {
1214 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1215 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1216 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1218 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1219 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1220 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1224 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1226 /// In addition, we also have a special variant of the scalar form here to
1227 /// represent the associated intrinsic operation. This form is unlike the
1228 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1229 /// and leaves the top elements undefined.
1231 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1232 /// six "instructions".
1234 let Constraints = "$src1 = $dst" in {
1235 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1236 SDNode OpNode, Intrinsic F64Int,
1237 bit Commutable = 0> {
1238 // Scalar operation, reg+reg.
1239 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1240 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1241 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1242 let isCommutable = Commutable;
1245 // Scalar operation, reg+mem.
1246 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1247 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1248 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1250 // Vector operation, reg+reg.
1251 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1252 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1253 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1254 let isCommutable = Commutable;
1257 // Vector operation, reg+mem.
1258 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1259 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1260 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1262 // Intrinsic operation, reg+reg.
1263 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1264 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1265 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1266 let isCommutable = Commutable;
1269 // Intrinsic operation, reg+mem.
1270 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1271 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1272 [(set VR128:$dst, (F64Int VR128:$src1,
1273 sse_load_f64:$src2))]>;
1277 // Arithmetic instructions
1278 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1279 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1280 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1281 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1283 /// sse2_fp_binop_rm - Other SSE2 binops
1285 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1286 /// instructions for a full-vector intrinsic form. Operations that map
1287 /// onto C operators don't use this form since they just use the plain
1288 /// vector form instead of having a separate vector intrinsic form.
1290 /// This provides a total of eight "instructions".
1292 let Constraints = "$src1 = $dst" in {
1293 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1297 bit Commutable = 0> {
1299 // Scalar operation, reg+reg.
1300 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1301 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1302 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1303 let isCommutable = Commutable;
1306 // Scalar operation, reg+mem.
1307 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1308 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1309 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1311 // Vector operation, reg+reg.
1312 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1313 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1314 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1315 let isCommutable = Commutable;
1318 // Vector operation, reg+mem.
1319 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1320 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1321 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1323 // Intrinsic operation, reg+reg.
1324 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1325 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1326 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1327 let isCommutable = Commutable;
1330 // Intrinsic operation, reg+mem.
1331 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1332 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1333 [(set VR128:$dst, (F64Int VR128:$src1,
1334 sse_load_f64:$src2))]>;
1336 // Vector intrinsic operation, reg+reg.
1337 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1338 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1339 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1340 let isCommutable = Commutable;
1343 // Vector intrinsic operation, reg+mem.
1344 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1345 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1346 [(set VR128:$dst, (V2F64Int VR128:$src1, (load addr:$src2)))]>;
1350 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1351 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1352 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1353 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1355 //===----------------------------------------------------------------------===//
1356 // SSE packed FP Instructions
1358 // Move Instructions
1359 let neverHasSideEffects = 1 in
1360 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1361 "movapd\t{$src, $dst|$dst, $src}", []>;
1362 let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
1363 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1364 "movapd\t{$src, $dst|$dst, $src}",
1365 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1367 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1368 "movapd\t{$src, $dst|$dst, $src}",
1369 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1371 let neverHasSideEffects = 1 in
1372 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1373 "movupd\t{$src, $dst|$dst, $src}", []>;
1374 let isSimpleLoad = 1 in
1375 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1376 "movupd\t{$src, $dst|$dst, $src}",
1377 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1378 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1379 "movupd\t{$src, $dst|$dst, $src}",
1380 [(store (v2f64 VR128:$src), addr:$dst)]>;
1382 // Intrinsic forms of MOVUPD load and store
1383 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1384 "movupd\t{$src, $dst|$dst, $src}",
1385 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1386 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1387 "movupd\t{$src, $dst|$dst, $src}",
1388 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1390 let Constraints = "$src1 = $dst" in {
1391 let AddedComplexity = 20 in {
1392 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1393 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1394 "movlpd\t{$src2, $dst|$dst, $src2}",
1396 (v2f64 (vector_shuffle VR128:$src1,
1397 (scalar_to_vector (loadf64 addr:$src2)),
1398 MOVLP_shuffle_mask)))]>;
1399 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1400 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1401 "movhpd\t{$src2, $dst|$dst, $src2}",
1403 (v2f64 (vector_shuffle VR128:$src1,
1404 (scalar_to_vector (loadf64 addr:$src2)),
1405 MOVHP_shuffle_mask)))]>;
1406 } // AddedComplexity
1407 } // Constraints = "$src1 = $dst"
1409 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1410 "movlpd\t{$src, $dst|$dst, $src}",
1411 [(store (f64 (vector_extract (v2f64 VR128:$src),
1412 (iPTR 0))), addr:$dst)]>;
1414 // v2f64 extract element 1 is always custom lowered to unpack high to low
1415 // and extract element 0 so the non-store version isn't too horrible.
1416 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1417 "movhpd\t{$src, $dst|$dst, $src}",
1418 [(store (f64 (vector_extract
1419 (v2f64 (vector_shuffle VR128:$src, (undef),
1420 UNPCKH_shuffle_mask)), (iPTR 0))),
1423 // SSE2 instructions without OpSize prefix
1424 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1425 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1426 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1427 TB, Requires<[HasSSE2]>;
1428 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1429 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1430 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1431 (bitconvert (memopv2i64 addr:$src))))]>,
1432 TB, Requires<[HasSSE2]>;
1434 // SSE2 instructions with XS prefix
1435 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1436 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1437 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1438 XS, Requires<[HasSSE2]>;
1439 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1440 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1441 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1442 (bitconvert (memopv2i64 addr:$src))))]>,
1443 XS, Requires<[HasSSE2]>;
1445 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1446 "cvtps2dq\t{$src, $dst|$dst, $src}",
1447 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1448 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1449 "cvtps2dq\t{$src, $dst|$dst, $src}",
1450 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1451 (load addr:$src)))]>;
1452 // SSE2 packed instructions with XS prefix
1453 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1454 "cvttps2dq\t{$src, $dst|$dst, $src}",
1455 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1456 XS, Requires<[HasSSE2]>;
1457 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1458 "cvttps2dq\t{$src, $dst|$dst, $src}",
1459 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1460 (load addr:$src)))]>,
1461 XS, Requires<[HasSSE2]>;
1463 // SSE2 packed instructions with XD prefix
1464 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1465 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1466 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1467 XD, Requires<[HasSSE2]>;
1468 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1469 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1470 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1471 (load addr:$src)))]>,
1472 XD, Requires<[HasSSE2]>;
1474 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1475 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1476 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1477 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1478 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1479 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1480 (load addr:$src)))]>;
1482 // SSE2 instructions without OpSize prefix
1483 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1484 "cvtps2pd\t{$src, $dst|$dst, $src}",
1485 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1486 TB, Requires<[HasSSE2]>;
1487 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f64mem:$src),
1488 "cvtps2pd\t{$src, $dst|$dst, $src}",
1489 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1490 (load addr:$src)))]>,
1491 TB, Requires<[HasSSE2]>;
1493 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1494 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1495 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1496 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f128mem:$src),
1497 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1498 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1499 (load addr:$src)))]>;
1501 // Match intrinsics which expect XMM operand(s).
1502 // Aliases for intrinsics
1503 let Constraints = "$src1 = $dst" in {
1504 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1505 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1506 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1507 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1509 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1510 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1511 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1512 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1513 (loadi32 addr:$src2)))]>;
1514 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1515 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1516 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1517 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1519 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1520 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1521 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1522 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1523 (load addr:$src2)))]>;
1524 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1525 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1526 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1527 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1528 VR128:$src2))]>, XS,
1529 Requires<[HasSSE2]>;
1530 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1531 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1532 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1533 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1534 (load addr:$src2)))]>, XS,
1535 Requires<[HasSSE2]>;
1540 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1542 /// In addition, we also have a special variant of the scalar form here to
1543 /// represent the associated intrinsic operation. This form is unlike the
1544 /// plain scalar form, in that it takes an entire vector (instead of a
1545 /// scalar) and leaves the top elements undefined.
1547 /// And, we have a special variant form for a full-vector intrinsic form.
1549 /// These four forms can each have a reg or a mem operand, so there are a
1550 /// total of eight "instructions".
1552 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1556 bit Commutable = 0> {
1557 // Scalar operation, reg.
1558 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1559 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1560 [(set FR64:$dst, (OpNode FR64:$src))]> {
1561 let isCommutable = Commutable;
1564 // Scalar operation, mem.
1565 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1566 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1567 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1569 // Vector operation, reg.
1570 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1571 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1572 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1573 let isCommutable = Commutable;
1576 // Vector operation, mem.
1577 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1578 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1579 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1581 // Intrinsic operation, reg.
1582 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1583 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1584 [(set VR128:$dst, (F64Int VR128:$src))]> {
1585 let isCommutable = Commutable;
1588 // Intrinsic operation, mem.
1589 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1590 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1591 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1593 // Vector intrinsic operation, reg
1594 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1595 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1596 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1597 let isCommutable = Commutable;
1600 // Vector intrinsic operation, mem
1601 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1602 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1603 [(set VR128:$dst, (V2F64Int (load addr:$src)))]>;
1607 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1608 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1610 // There is no f64 version of the reciprocal approximation instructions.
1613 let Constraints = "$src1 = $dst" in {
1614 let isCommutable = 1 in {
1615 def ANDPDrr : PDI<0x54, MRMSrcReg,
1616 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1617 "andpd\t{$src2, $dst|$dst, $src2}",
1619 (and (bc_v2i64 (v2f64 VR128:$src1)),
1620 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1621 def ORPDrr : PDI<0x56, MRMSrcReg,
1622 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1623 "orpd\t{$src2, $dst|$dst, $src2}",
1625 (or (bc_v2i64 (v2f64 VR128:$src1)),
1626 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1627 def XORPDrr : PDI<0x57, MRMSrcReg,
1628 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1629 "xorpd\t{$src2, $dst|$dst, $src2}",
1631 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1632 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1635 def ANDPDrm : PDI<0x54, MRMSrcMem,
1636 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1637 "andpd\t{$src2, $dst|$dst, $src2}",
1639 (and (bc_v2i64 (v2f64 VR128:$src1)),
1640 (memopv2i64 addr:$src2)))]>;
1641 def ORPDrm : PDI<0x56, MRMSrcMem,
1642 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1643 "orpd\t{$src2, $dst|$dst, $src2}",
1645 (or (bc_v2i64 (v2f64 VR128:$src1)),
1646 (memopv2i64 addr:$src2)))]>;
1647 def XORPDrm : PDI<0x57, MRMSrcMem,
1648 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1649 "xorpd\t{$src2, $dst|$dst, $src2}",
1651 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1652 (memopv2i64 addr:$src2)))]>;
1653 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1654 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1655 "andnpd\t{$src2, $dst|$dst, $src2}",
1657 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1658 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1659 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1660 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1661 "andnpd\t{$src2, $dst|$dst, $src2}",
1663 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1664 (memopv2i64 addr:$src2)))]>;
1667 let Constraints = "$src1 = $dst" in {
1668 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1669 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1670 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1671 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1672 VR128:$src, imm:$cc))]>;
1673 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1674 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1675 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1676 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1677 (load addr:$src), imm:$cc))]>;
1680 // Shuffle and unpack instructions
1681 let Constraints = "$src1 = $dst" in {
1682 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1683 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1684 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1685 [(set VR128:$dst, (v2f64 (vector_shuffle
1686 VR128:$src1, VR128:$src2,
1687 SHUFP_shuffle_mask:$src3)))]>;
1688 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1689 (outs VR128:$dst), (ins VR128:$src1,
1690 f128mem:$src2, i8imm:$src3),
1691 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1693 (v2f64 (vector_shuffle
1694 VR128:$src1, (memopv2f64 addr:$src2),
1695 SHUFP_shuffle_mask:$src3)))]>;
1697 let AddedComplexity = 10 in {
1698 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1699 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1700 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1702 (v2f64 (vector_shuffle
1703 VR128:$src1, VR128:$src2,
1704 UNPCKH_shuffle_mask)))]>;
1705 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1706 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1707 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1709 (v2f64 (vector_shuffle
1710 VR128:$src1, (memopv2f64 addr:$src2),
1711 UNPCKH_shuffle_mask)))]>;
1713 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1714 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1715 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1717 (v2f64 (vector_shuffle
1718 VR128:$src1, VR128:$src2,
1719 UNPCKL_shuffle_mask)))]>;
1720 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1721 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1722 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1724 (v2f64 (vector_shuffle
1725 VR128:$src1, (memopv2f64 addr:$src2),
1726 UNPCKL_shuffle_mask)))]>;
1727 } // AddedComplexity
1728 } // Constraints = "$src1 = $dst"
1731 //===----------------------------------------------------------------------===//
1732 // SSE integer instructions
1734 // Move Instructions
1735 let neverHasSideEffects = 1 in
1736 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1737 "movdqa\t{$src, $dst|$dst, $src}", []>;
1738 let isSimpleLoad = 1, mayLoad = 1 in
1739 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1740 "movdqa\t{$src, $dst|$dst, $src}",
1741 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1743 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1744 "movdqa\t{$src, $dst|$dst, $src}",
1745 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1746 let isSimpleLoad = 1, mayLoad = 1 in
1747 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1748 "movdqu\t{$src, $dst|$dst, $src}",
1749 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1750 XS, Requires<[HasSSE2]>;
1752 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1753 "movdqu\t{$src, $dst|$dst, $src}",
1754 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1755 XS, Requires<[HasSSE2]>;
1757 // Intrinsic forms of MOVDQU load and store
1758 let isSimpleLoad = 1 in
1759 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1760 "movdqu\t{$src, $dst|$dst, $src}",
1761 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1762 XS, Requires<[HasSSE2]>;
1763 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1764 "movdqu\t{$src, $dst|$dst, $src}",
1765 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1766 XS, Requires<[HasSSE2]>;
1768 let Constraints = "$src1 = $dst" in {
1770 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1771 bit Commutable = 0> {
1772 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1773 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1774 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1775 let isCommutable = Commutable;
1777 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1778 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1779 [(set VR128:$dst, (IntId VR128:$src1,
1780 (bitconvert (memopv2i64 addr:$src2))))]>;
1783 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1785 Intrinsic IntId, Intrinsic IntId2> {
1786 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1787 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1788 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1789 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1790 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1791 [(set VR128:$dst, (IntId VR128:$src1,
1792 (bitconvert (memopv2i64 addr:$src2))))]>;
1793 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1794 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1795 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1798 /// PDI_binop_rm - Simple SSE2 binary operator.
1799 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1800 ValueType OpVT, bit Commutable = 0> {
1801 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1802 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1803 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1804 let isCommutable = Commutable;
1806 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1807 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1808 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1809 (bitconvert (memopv2i64 addr:$src2)))))]>;
1812 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1814 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1815 /// to collapse (bitconvert VT to VT) into its operand.
1817 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1818 bit Commutable = 0> {
1819 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1820 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1821 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1822 let isCommutable = Commutable;
1824 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1825 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1826 [(set VR128:$dst, (OpNode VR128:$src1,(memopv2i64 addr:$src2)))]>;
1829 } // Constraints = "$src1 = $dst"
1831 // 128-bit Integer Arithmetic
1833 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1834 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1835 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1836 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1838 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1839 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1840 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1841 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1843 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1844 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1845 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1846 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1848 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1849 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1850 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1851 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1853 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1855 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1856 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1857 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1859 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1861 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1862 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1865 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1866 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1867 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1868 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1869 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1872 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
1873 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
1874 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
1875 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
1876 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
1877 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
1879 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
1880 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
1881 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
1882 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
1883 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x72, MRM2r, "psrlq",
1884 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
1886 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
1887 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
1888 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x71, MRM4r, "psrad",
1889 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
1891 // 128-bit logical shifts.
1892 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1893 def PSLLDQri : PDIi8<0x73, MRM7r,
1894 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1895 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1896 def PSRLDQri : PDIi8<0x73, MRM3r,
1897 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1898 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1899 // PSRADQri doesn't exist in SSE[1-3].
1902 let Predicates = [HasSSE2] in {
1903 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1904 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1905 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1906 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1907 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1908 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1912 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1913 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1914 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1916 let Constraints = "$src1 = $dst" in {
1917 def PANDNrr : PDI<0xDF, MRMSrcReg,
1918 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1919 "pandn\t{$src2, $dst|$dst, $src2}",
1920 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1923 def PANDNrm : PDI<0xDF, MRMSrcMem,
1924 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1925 "pandn\t{$src2, $dst|$dst, $src2}",
1926 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1927 (memopv2i64 addr:$src2))))]>;
1930 // SSE2 Integer comparison
1931 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1932 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1933 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1934 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1935 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1936 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1938 // Pack instructions
1939 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1940 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1941 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1943 // Shuffle and unpack instructions
1944 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1945 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1946 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1947 [(set VR128:$dst, (v4i32 (vector_shuffle
1948 VR128:$src1, (undef),
1949 PSHUFD_shuffle_mask:$src2)))]>;
1950 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1951 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1952 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1953 [(set VR128:$dst, (v4i32 (vector_shuffle
1954 (bc_v4i32(memopv2i64 addr:$src1)),
1956 PSHUFD_shuffle_mask:$src2)))]>;
1958 // SSE2 with ImmT == Imm8 and XS prefix.
1959 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1960 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1961 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1962 [(set VR128:$dst, (v8i16 (vector_shuffle
1963 VR128:$src1, (undef),
1964 PSHUFHW_shuffle_mask:$src2)))]>,
1965 XS, Requires<[HasSSE2]>;
1966 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
1967 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1968 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1969 [(set VR128:$dst, (v8i16 (vector_shuffle
1970 (bc_v8i16 (memopv2i64 addr:$src1)),
1972 PSHUFHW_shuffle_mask:$src2)))]>,
1973 XS, Requires<[HasSSE2]>;
1975 // SSE2 with ImmT == Imm8 and XD prefix.
1976 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
1977 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1978 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1979 [(set VR128:$dst, (v8i16 (vector_shuffle
1980 VR128:$src1, (undef),
1981 PSHUFLW_shuffle_mask:$src2)))]>,
1982 XD, Requires<[HasSSE2]>;
1983 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
1984 (outs VR128:$dst), (ins i128mem:$src1, i32i8imm:$src2),
1985 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1986 [(set VR128:$dst, (v8i16 (vector_shuffle
1987 (bc_v8i16 (memopv2i64 addr:$src1)),
1989 PSHUFLW_shuffle_mask:$src2)))]>,
1990 XD, Requires<[HasSSE2]>;
1993 let Constraints = "$src1 = $dst" in {
1994 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
1995 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1996 "punpcklbw\t{$src2, $dst|$dst, $src2}",
1998 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1999 UNPCKL_shuffle_mask)))]>;
2000 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
2001 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2002 "punpcklbw\t{$src2, $dst|$dst, $src2}",
2004 (v16i8 (vector_shuffle VR128:$src1,
2005 (bc_v16i8 (memopv2i64 addr:$src2)),
2006 UNPCKL_shuffle_mask)))]>;
2007 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
2008 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2009 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2011 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2012 UNPCKL_shuffle_mask)))]>;
2013 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
2014 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2015 "punpcklwd\t{$src2, $dst|$dst, $src2}",
2017 (v8i16 (vector_shuffle VR128:$src1,
2018 (bc_v8i16 (memopv2i64 addr:$src2)),
2019 UNPCKL_shuffle_mask)))]>;
2020 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
2021 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2022 "punpckldq\t{$src2, $dst|$dst, $src2}",
2024 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2025 UNPCKL_shuffle_mask)))]>;
2026 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
2027 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2028 "punpckldq\t{$src2, $dst|$dst, $src2}",
2030 (v4i32 (vector_shuffle VR128:$src1,
2031 (bc_v4i32 (memopv2i64 addr:$src2)),
2032 UNPCKL_shuffle_mask)))]>;
2033 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2034 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2035 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2037 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2038 UNPCKL_shuffle_mask)))]>;
2039 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2040 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2041 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2043 (v2i64 (vector_shuffle VR128:$src1,
2044 (memopv2i64 addr:$src2),
2045 UNPCKL_shuffle_mask)))]>;
2047 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
2048 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2049 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2051 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
2052 UNPCKH_shuffle_mask)))]>;
2053 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
2054 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2055 "punpckhbw\t{$src2, $dst|$dst, $src2}",
2057 (v16i8 (vector_shuffle VR128:$src1,
2058 (bc_v16i8 (memopv2i64 addr:$src2)),
2059 UNPCKH_shuffle_mask)))]>;
2060 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
2061 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2062 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2064 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
2065 UNPCKH_shuffle_mask)))]>;
2066 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
2067 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2068 "punpckhwd\t{$src2, $dst|$dst, $src2}",
2070 (v8i16 (vector_shuffle VR128:$src1,
2071 (bc_v8i16 (memopv2i64 addr:$src2)),
2072 UNPCKH_shuffle_mask)))]>;
2073 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
2074 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2075 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2077 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2078 UNPCKH_shuffle_mask)))]>;
2079 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
2080 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2081 "punpckhdq\t{$src2, $dst|$dst, $src2}",
2083 (v4i32 (vector_shuffle VR128:$src1,
2084 (bc_v4i32 (memopv2i64 addr:$src2)),
2085 UNPCKH_shuffle_mask)))]>;
2086 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2087 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2088 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2090 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2091 UNPCKH_shuffle_mask)))]>;
2092 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2093 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2094 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2096 (v2i64 (vector_shuffle VR128:$src1,
2097 (memopv2i64 addr:$src2),
2098 UNPCKH_shuffle_mask)))]>;
2102 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2103 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2104 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2105 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2107 let Constraints = "$src1 = $dst" in {
2108 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2109 (outs VR128:$dst), (ins VR128:$src1,
2110 GR32:$src2, i32i8imm:$src3),
2111 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2113 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2114 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2115 (outs VR128:$dst), (ins VR128:$src1,
2116 i16mem:$src2, i32i8imm:$src3),
2117 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2119 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2124 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2125 "pmovmskb\t{$src, $dst|$dst, $src}",
2126 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2128 // Conditional store
2130 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2131 "maskmovdqu\t{$mask, $src|$src, $mask}",
2132 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2134 // Non-temporal stores
2135 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2136 "movntpd\t{$src, $dst|$dst, $src}",
2137 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2138 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2139 "movntdq\t{$src, $dst|$dst, $src}",
2140 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2141 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2142 "movnti\t{$src, $dst|$dst, $src}",
2143 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2144 TB, Requires<[HasSSE2]>;
2147 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2148 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2149 TB, Requires<[HasSSE2]>;
2151 // Load, store, and memory fence
2152 def LFENCE : I<0xAE, MRM5m, (outs), (ins),
2153 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2154 def MFENCE : I<0xAE, MRM6m, (outs), (ins),
2155 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2157 //TODO: custom lower this so as to never even generate the noop
2158 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2160 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2161 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2162 def : Pat<(membarrier (i8 imm:$ll), (i8 imm:$ls), (i8 imm:$sl), (i8 imm:$ss),
2165 // Alias instructions that map zero vector to pxor / xorp* for sse.
2166 let isReMaterializable = 1 in
2167 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2168 "pcmpeqd\t$dst, $dst",
2169 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2171 // FR64 to 128-bit vector conversion.
2172 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2173 "movsd\t{$src, $dst|$dst, $src}",
2175 (v2f64 (scalar_to_vector FR64:$src)))]>;
2176 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2177 "movsd\t{$src, $dst|$dst, $src}",
2179 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2181 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2182 "movd\t{$src, $dst|$dst, $src}",
2184 (v4i32 (scalar_to_vector GR32:$src)))]>;
2185 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2186 "movd\t{$src, $dst|$dst, $src}",
2188 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2190 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2191 "movd\t{$src, $dst|$dst, $src}",
2192 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2194 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2195 "movd\t{$src, $dst|$dst, $src}",
2196 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2198 // SSE2 instructions with XS prefix
2199 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2200 "movq\t{$src, $dst|$dst, $src}",
2202 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2203 Requires<[HasSSE2]>;
2204 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2205 "movq\t{$src, $dst|$dst, $src}",
2206 [(store (i64 (vector_extract (v2i64 VR128:$src),
2207 (iPTR 0))), addr:$dst)]>;
2209 // FIXME: may not be able to eliminate this movss with coalescing the src and
2210 // dest register classes are different. We really want to write this pattern
2212 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2213 // (f32 FR32:$src)>;
2214 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2215 "movsd\t{$src, $dst|$dst, $src}",
2216 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2218 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2219 "movsd\t{$src, $dst|$dst, $src}",
2220 [(store (f64 (vector_extract (v2f64 VR128:$src),
2221 (iPTR 0))), addr:$dst)]>;
2222 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2223 "movd\t{$src, $dst|$dst, $src}",
2224 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2226 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2227 "movd\t{$src, $dst|$dst, $src}",
2228 [(store (i32 (vector_extract (v4i32 VR128:$src),
2229 (iPTR 0))), addr:$dst)]>;
2231 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2232 "movd\t{$src, $dst|$dst, $src}",
2233 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2234 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2235 "movd\t{$src, $dst|$dst, $src}",
2236 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2239 // Move to lower bits of a VR128, leaving upper bits alone.
2240 // Three operand (but two address) aliases.
2241 let Constraints = "$src1 = $dst" in {
2242 let neverHasSideEffects = 1 in
2243 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2244 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2245 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2247 let AddedComplexity = 15 in
2248 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2249 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2250 "movsd\t{$src2, $dst|$dst, $src2}",
2252 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
2253 MOVL_shuffle_mask)))]>;
2256 // Store / copy lower 64-bits of a XMM register.
2257 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2258 "movq\t{$src, $dst|$dst, $src}",
2259 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2261 // Move to lower bits of a VR128 and zeroing upper bits.
2262 // Loading from memory automatically zeroing upper bits.
2263 let AddedComplexity = 20 in
2264 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2265 "movsd\t{$src, $dst|$dst, $src}",
2267 (v2f64 (vector_shuffle immAllZerosV_bc,
2268 (v2f64 (scalar_to_vector
2269 (loadf64 addr:$src))),
2270 MOVL_shuffle_mask)))]>;
2272 // movd / movq to XMM register zero-extends
2273 let AddedComplexity = 15 in {
2274 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2275 "movd\t{$src, $dst|$dst, $src}",
2277 (v4i32 (vector_shuffle immAllZerosV,
2278 (v4i32 (scalar_to_vector GR32:$src)),
2279 MOVL_shuffle_mask)))]>;
2280 // This is X86-64 only.
2281 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2282 "mov{d|q}\t{$src, $dst|$dst, $src}",
2284 (v2i64 (vector_shuffle immAllZerosV_bc,
2285 (v2i64 (scalar_to_vector GR64:$src)),
2286 MOVL_shuffle_mask)))]>;
2289 // Handle the v2f64 form of 'MOVZQI2PQIrr' for PR2108. FIXME: this would be
2290 // better written as a dag combine xform.
2291 let AddedComplexity = 15 in
2292 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc,
2293 (v2f64 (scalar_to_vector
2294 (f64 (bitconvert GR64:$src)))),
2295 MOVL_shuffle_mask)),
2296 (MOVZQI2PQIrr GR64:$src)>, Requires<[HasSSE2]>;
2299 let AddedComplexity = 20 in {
2300 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2301 "movd\t{$src, $dst|$dst, $src}",
2303 (v4i32 (vector_shuffle immAllZerosV,
2304 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
2305 MOVL_shuffle_mask)))]>;
2306 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2307 "movq\t{$src, $dst|$dst, $src}",
2309 (v2i64 (vector_shuffle immAllZerosV_bc,
2310 (v2i64 (scalar_to_vector (loadi64 addr:$src))),
2311 MOVL_shuffle_mask)))]>, XS,
2312 Requires<[HasSSE2]>;
2315 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2316 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2317 let AddedComplexity = 15 in
2318 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2319 "movq\t{$src, $dst|$dst, $src}",
2320 [(set VR128:$dst, (v2i64 (vector_shuffle immAllZerosV_bc,
2322 MOVL_shuffle_mask)))]>,
2323 XS, Requires<[HasSSE2]>;
2325 let AddedComplexity = 20 in
2326 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2327 "movq\t{$src, $dst|$dst, $src}",
2328 [(set VR128:$dst, (v2i64 (vector_shuffle immAllZerosV_bc,
2329 (memopv2i64 addr:$src),
2330 MOVL_shuffle_mask)))]>,
2331 XS, Requires<[HasSSE2]>;
2333 //===----------------------------------------------------------------------===//
2334 // SSE3 Instructions
2335 //===----------------------------------------------------------------------===//
2337 // Move Instructions
2338 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2339 "movshdup\t{$src, $dst|$dst, $src}",
2340 [(set VR128:$dst, (v4f32 (vector_shuffle
2341 VR128:$src, (undef),
2342 MOVSHDUP_shuffle_mask)))]>;
2343 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2344 "movshdup\t{$src, $dst|$dst, $src}",
2345 [(set VR128:$dst, (v4f32 (vector_shuffle
2346 (memopv4f32 addr:$src), (undef),
2347 MOVSHDUP_shuffle_mask)))]>;
2349 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2350 "movsldup\t{$src, $dst|$dst, $src}",
2351 [(set VR128:$dst, (v4f32 (vector_shuffle
2352 VR128:$src, (undef),
2353 MOVSLDUP_shuffle_mask)))]>;
2354 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2355 "movsldup\t{$src, $dst|$dst, $src}",
2356 [(set VR128:$dst, (v4f32 (vector_shuffle
2357 (memopv4f32 addr:$src), (undef),
2358 MOVSLDUP_shuffle_mask)))]>;
2360 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2361 "movddup\t{$src, $dst|$dst, $src}",
2362 [(set VR128:$dst, (v2f64 (vector_shuffle
2363 VR128:$src, (undef),
2364 SSE_splat_lo_mask)))]>;
2365 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2366 "movddup\t{$src, $dst|$dst, $src}",
2368 (v2f64 (vector_shuffle
2369 (scalar_to_vector (loadf64 addr:$src)),
2371 SSE_splat_lo_mask)))]>;
2374 let Constraints = "$src1 = $dst" in {
2375 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2376 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2377 "addsubps\t{$src2, $dst|$dst, $src2}",
2378 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2380 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2381 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2382 "addsubps\t{$src2, $dst|$dst, $src2}",
2383 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2384 (load addr:$src2)))]>;
2385 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2386 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2387 "addsubpd\t{$src2, $dst|$dst, $src2}",
2388 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2390 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2391 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2392 "addsubpd\t{$src2, $dst|$dst, $src2}",
2393 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2394 (load addr:$src2)))]>;
2397 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2398 "lddqu\t{$src, $dst|$dst, $src}",
2399 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2402 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2403 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2404 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2405 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2406 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2407 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2408 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2409 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
2410 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2411 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2412 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2413 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2414 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2415 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2416 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2417 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
2419 let Constraints = "$src1 = $dst" in {
2420 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2421 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2422 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2423 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2424 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2425 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2426 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2427 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2430 // Thread synchronization
2431 def MONITOR : I<0xC8, RawFrm, (outs), (ins), "monitor",
2432 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2433 def MWAIT : I<0xC9, RawFrm, (outs), (ins), "mwait",
2434 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2436 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2437 let AddedComplexity = 15 in
2438 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2439 MOVSHDUP_shuffle_mask)),
2440 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2441 let AddedComplexity = 20 in
2442 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2443 MOVSHDUP_shuffle_mask)),
2444 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2446 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2447 let AddedComplexity = 15 in
2448 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2449 MOVSLDUP_shuffle_mask)),
2450 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2451 let AddedComplexity = 20 in
2452 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2453 MOVSLDUP_shuffle_mask)),
2454 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2456 //===----------------------------------------------------------------------===//
2457 // SSSE3 Instructions
2458 //===----------------------------------------------------------------------===//
2460 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2461 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2462 Intrinsic IntId64, Intrinsic IntId128> {
2463 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2464 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2465 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2467 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2468 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2470 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2472 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2474 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2475 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2478 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2480 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2483 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2486 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2487 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2488 Intrinsic IntId64, Intrinsic IntId128> {
2489 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2491 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2492 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2494 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2496 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2499 (bitconvert (memopv4i16 addr:$src))))]>;
2501 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2503 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2504 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2507 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2509 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2512 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2515 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2516 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2517 Intrinsic IntId64, Intrinsic IntId128> {
2518 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2520 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2521 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2523 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2525 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2528 (bitconvert (memopv2i32 addr:$src))))]>;
2530 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2532 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2533 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2536 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2538 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2541 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2544 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2545 int_x86_ssse3_pabs_b,
2546 int_x86_ssse3_pabs_b_128>;
2547 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2548 int_x86_ssse3_pabs_w,
2549 int_x86_ssse3_pabs_w_128>;
2550 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2551 int_x86_ssse3_pabs_d,
2552 int_x86_ssse3_pabs_d_128>;
2554 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2555 let Constraints = "$src1 = $dst" in {
2556 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2557 Intrinsic IntId64, Intrinsic IntId128,
2558 bit Commutable = 0> {
2559 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2560 (ins VR64:$src1, VR64:$src2),
2561 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2562 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2563 let isCommutable = Commutable;
2565 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2566 (ins VR64:$src1, i64mem:$src2),
2567 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2569 (IntId64 VR64:$src1,
2570 (bitconvert (memopv8i8 addr:$src2))))]>;
2572 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2573 (ins VR128:$src1, VR128:$src2),
2574 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2575 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2577 let isCommutable = Commutable;
2579 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2580 (ins VR128:$src1, i128mem:$src2),
2581 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2583 (IntId128 VR128:$src1,
2584 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2588 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2589 let Constraints = "$src1 = $dst" in {
2590 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2591 Intrinsic IntId64, Intrinsic IntId128,
2592 bit Commutable = 0> {
2593 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2594 (ins VR64:$src1, VR64:$src2),
2595 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2596 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2597 let isCommutable = Commutable;
2599 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2600 (ins VR64:$src1, i64mem:$src2),
2601 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2603 (IntId64 VR64:$src1,
2604 (bitconvert (memopv4i16 addr:$src2))))]>;
2606 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2607 (ins VR128:$src1, VR128:$src2),
2608 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2609 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2611 let isCommutable = Commutable;
2613 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2614 (ins VR128:$src1, i128mem:$src2),
2615 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2617 (IntId128 VR128:$src1,
2618 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2622 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2623 let Constraints = "$src1 = $dst" in {
2624 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2625 Intrinsic IntId64, Intrinsic IntId128,
2626 bit Commutable = 0> {
2627 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2628 (ins VR64:$src1, VR64:$src2),
2629 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2630 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2631 let isCommutable = Commutable;
2633 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2634 (ins VR64:$src1, i64mem:$src2),
2635 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2637 (IntId64 VR64:$src1,
2638 (bitconvert (memopv2i32 addr:$src2))))]>;
2640 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2641 (ins VR128:$src1, VR128:$src2),
2642 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2643 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2645 let isCommutable = Commutable;
2647 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2648 (ins VR128:$src1, i128mem:$src2),
2649 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2651 (IntId128 VR128:$src1,
2652 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2656 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2657 int_x86_ssse3_phadd_w,
2658 int_x86_ssse3_phadd_w_128, 1>;
2659 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2660 int_x86_ssse3_phadd_d,
2661 int_x86_ssse3_phadd_d_128, 1>;
2662 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2663 int_x86_ssse3_phadd_sw,
2664 int_x86_ssse3_phadd_sw_128, 1>;
2665 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2666 int_x86_ssse3_phsub_w,
2667 int_x86_ssse3_phsub_w_128>;
2668 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2669 int_x86_ssse3_phsub_d,
2670 int_x86_ssse3_phsub_d_128>;
2671 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2672 int_x86_ssse3_phsub_sw,
2673 int_x86_ssse3_phsub_sw_128>;
2674 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2675 int_x86_ssse3_pmadd_ub_sw,
2676 int_x86_ssse3_pmadd_ub_sw_128, 1>;
2677 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2678 int_x86_ssse3_pmul_hr_sw,
2679 int_x86_ssse3_pmul_hr_sw_128, 1>;
2680 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2681 int_x86_ssse3_pshuf_b,
2682 int_x86_ssse3_pshuf_b_128>;
2683 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2684 int_x86_ssse3_psign_b,
2685 int_x86_ssse3_psign_b_128>;
2686 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2687 int_x86_ssse3_psign_w,
2688 int_x86_ssse3_psign_w_128>;
2689 defm PSIGND : SS3I_binop_rm_int_32<0x09, "psignd",
2690 int_x86_ssse3_psign_d,
2691 int_x86_ssse3_psign_d_128>;
2693 let Constraints = "$src1 = $dst" in {
2694 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2695 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2696 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2698 (int_x86_ssse3_palign_r
2699 VR64:$src1, VR64:$src2,
2701 def PALIGNR64rm : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2702 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2703 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2705 (int_x86_ssse3_palign_r
2707 (bitconvert (memopv2i32 addr:$src2)),
2710 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2711 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2712 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2714 (int_x86_ssse3_palign_r_128
2715 VR128:$src1, VR128:$src2,
2716 imm:$src3))]>, OpSize;
2717 def PALIGNR128rm : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2718 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2719 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2721 (int_x86_ssse3_palign_r_128
2723 (bitconvert (memopv4i32 addr:$src2)),
2724 imm:$src3))]>, OpSize;
2727 //===----------------------------------------------------------------------===//
2728 // Non-Instruction Patterns
2729 //===----------------------------------------------------------------------===//
2731 // extload f32 -> f64. This matches load+fextend because we have a hack in
2732 // the isel (PreprocessForFPConvert) that can introduce loads after dag combine.
2733 // Since these loads aren't folded into the fextend, we have to match it
2735 let Predicates = [HasSSE2] in
2736 def : Pat<(fextend (loadf32 addr:$src)),
2737 (CVTSS2SDrm addr:$src)>;
2740 let Predicates = [HasSSE2] in {
2741 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2742 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2743 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2744 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2745 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2746 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2747 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2748 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2749 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2750 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2751 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2752 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2753 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2754 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2755 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2756 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2757 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2758 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2759 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2760 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2761 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2762 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2763 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2764 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2765 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2766 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2767 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2768 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2769 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2770 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2773 // Move scalar to XMM zero-extended
2774 // movd to XMM register zero-extends
2775 let AddedComplexity = 15 in {
2776 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2777 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc,
2778 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
2779 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2780 def : Pat<(v4f32 (vector_shuffle immAllZerosV_bc,
2781 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
2782 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
2785 // Splat v2f64 / v2i64
2786 let AddedComplexity = 10 in {
2787 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2788 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2789 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2790 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2791 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2792 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2793 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2794 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2797 // Special unary SHUFPSrri case.
2798 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2799 SHUFP_unary_shuffle_mask:$sm)),
2800 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2801 Requires<[HasSSE1]>;
2802 // Special unary SHUFPDrri case.
2803 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (undef),
2804 SHUFP_unary_shuffle_mask:$sm)),
2805 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2806 Requires<[HasSSE2]>;
2807 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2808 def : Pat<(vector_shuffle (bc_v4i32 (memopv4f32 addr:$src1)), (undef),
2809 SHUFP_unary_shuffle_mask:$sm),
2810 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2811 Requires<[HasSSE2]>;
2812 // Special binary v4i32 shuffle cases with SHUFPS.
2813 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (v4i32 VR128:$src2),
2814 PSHUFD_binary_shuffle_mask:$sm)),
2815 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2816 Requires<[HasSSE2]>;
2817 def : Pat<(v4i32 (vector_shuffle VR128:$src1,
2818 (bc_v4i32 (memopv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm)),
2819 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2820 Requires<[HasSSE2]>;
2821 // Special binary v2i64 shuffle cases using SHUFPDrri.
2822 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2823 SHUFP_shuffle_mask:$sm)),
2824 (SHUFPDrri VR128:$src1, VR128:$src2, SHUFP_shuffle_mask:$sm)>,
2825 Requires<[HasSSE2]>;
2826 // Special unary SHUFPDrri case.
2827 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (undef),
2828 SHUFP_unary_shuffle_mask:$sm)),
2829 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2830 Requires<[HasSSE2]>;
2832 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2833 let AddedComplexity = 10 in {
2834 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2835 UNPCKL_v_undef_shuffle_mask)),
2836 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2837 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2838 UNPCKL_v_undef_shuffle_mask)),
2839 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2840 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2841 UNPCKL_v_undef_shuffle_mask)),
2842 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2843 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2844 UNPCKL_v_undef_shuffle_mask)),
2845 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2848 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2849 let AddedComplexity = 10 in {
2850 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2851 UNPCKH_v_undef_shuffle_mask)),
2852 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2853 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2854 UNPCKH_v_undef_shuffle_mask)),
2855 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2856 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2857 UNPCKH_v_undef_shuffle_mask)),
2858 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2859 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2860 UNPCKH_v_undef_shuffle_mask)),
2861 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2864 let AddedComplexity = 15 in {
2865 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2866 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2867 MOVHP_shuffle_mask)),
2868 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2870 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2871 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2872 MOVHLPS_shuffle_mask)),
2873 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2875 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2876 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2877 MOVHLPS_v_undef_shuffle_mask)),
2878 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2879 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2880 MOVHLPS_v_undef_shuffle_mask)),
2881 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2884 let AddedComplexity = 20 in {
2885 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2886 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2887 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2888 MOVLP_shuffle_mask)),
2889 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2890 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2891 MOVLP_shuffle_mask)),
2892 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2893 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2894 MOVHP_shuffle_mask)),
2895 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2896 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2897 MOVHP_shuffle_mask)),
2898 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2900 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2901 MOVLP_shuffle_mask)),
2902 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2903 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2904 MOVLP_shuffle_mask)),
2905 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2906 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2907 MOVHP_shuffle_mask)),
2908 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2909 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2910 MOVLP_shuffle_mask)),
2911 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2914 let AddedComplexity = 15 in {
2915 // Setting the lowest element in the vector.
2916 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2917 MOVL_shuffle_mask)),
2918 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2919 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2920 MOVL_shuffle_mask)),
2921 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2923 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2924 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2925 MOVLP_shuffle_mask)),
2926 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2927 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2928 MOVLP_shuffle_mask)),
2929 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2932 // Set lowest element and zero upper elements.
2933 let AddedComplexity = 15 in
2934 def : Pat<(v2f64 (vector_shuffle immAllZerosV_bc, VR128:$src,
2935 MOVL_shuffle_mask)),
2936 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
2939 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2940 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2941 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2942 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2943 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2944 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2945 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2946 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2947 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2948 Requires<[HasSSE2]>;
2949 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2950 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2951 Requires<[HasSSE2]>;
2952 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2953 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2954 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2955 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2956 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2957 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2958 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2959 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2960 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
2961 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2962 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
2963 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2964 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
2965 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2966 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
2967 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2969 // Some special case pandn patterns.
2970 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2972 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2973 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2975 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2976 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2978 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2980 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2981 (memopv2i64 addr:$src2))),
2982 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2983 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2984 (memopv2i64 addr:$src2))),
2985 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2986 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2987 (memopv2i64 addr:$src2))),
2988 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2990 // vector -> vector casts
2991 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
2992 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
2993 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
2994 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
2996 // Use movaps / movups for SSE integer load / store (one byte shorter).
2997 def : Pat<(alignedloadv4i32 addr:$src),
2998 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
2999 def : Pat<(loadv4i32 addr:$src),
3000 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
3001 def : Pat<(alignedloadv2i64 addr:$src),
3002 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
3003 def : Pat<(loadv2i64 addr:$src),
3004 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
3006 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3007 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3008 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3009 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3010 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3011 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3012 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3013 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3014 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3015 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3016 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3017 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3018 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3019 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3020 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3021 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
3023 //===----------------------------------------------------------------------===//
3024 // SSE4.1 Instructions
3025 //===----------------------------------------------------------------------===//
3027 multiclass sse41_fp_unop_rm<bits<8> opcss, bits<8> opcps,
3028 bits<8> opcsd, bits<8> opcpd,
3033 Intrinsic V2F64Int> {
3034 // Intrinsic operation, reg.
3035 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3036 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3037 !strconcat(OpcodeStr,
3038 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3039 [(set VR128:$dst, (F32Int VR128:$src1, imm:$src2))]>,
3042 // Intrinsic operation, mem.
3043 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3044 (outs VR128:$dst), (ins ssmem:$src1, i32i8imm:$src2),
3045 !strconcat(OpcodeStr,
3046 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3047 [(set VR128:$dst, (F32Int sse_load_f32:$src1, imm:$src2))]>,
3050 // Vector intrinsic operation, reg
3051 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3052 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3053 !strconcat(OpcodeStr,
3054 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3055 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3058 // Vector intrinsic operation, mem
3059 def PSm_Int : SS4AIi8<opcps, MRMSrcMem,
3060 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3061 !strconcat(OpcodeStr,
3062 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3063 [(set VR128:$dst, (V4F32Int (load addr:$src1),imm:$src2))]>,
3066 // Intrinsic operation, reg.
3067 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3068 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3069 !strconcat(OpcodeStr,
3070 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3071 [(set VR128:$dst, (F64Int VR128:$src1, imm:$src2))]>,
3074 // Intrinsic operation, mem.
3075 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3076 (outs VR128:$dst), (ins sdmem:$src1, i32i8imm:$src2),
3077 !strconcat(OpcodeStr,
3078 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3079 [(set VR128:$dst, (F64Int sse_load_f64:$src1, imm:$src2))]>,
3082 // Vector intrinsic operation, reg
3083 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3084 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3085 !strconcat(OpcodeStr,
3086 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3087 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3090 // Vector intrinsic operation, mem
3091 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3092 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3093 !strconcat(OpcodeStr,
3094 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3095 [(set VR128:$dst, (V2F64Int (load addr:$src1),imm:$src2))]>,
3099 // FP round - roundss, roundps, roundsd, roundpd
3100 defm ROUND : sse41_fp_unop_rm<0x0A, 0x08, 0x0B, 0x09, "round",
3101 int_x86_sse41_round_ss, int_x86_sse41_round_ps,
3102 int_x86_sse41_round_sd, int_x86_sse41_round_pd>;
3104 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3105 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3106 Intrinsic IntId128> {
3107 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3109 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3110 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3111 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3113 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3116 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3119 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3120 int_x86_sse41_phminposuw>;
3122 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3123 let Constraints = "$src1 = $dst" in {
3124 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3125 Intrinsic IntId128, bit Commutable = 0> {
3126 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3127 (ins VR128:$src1, VR128:$src2),
3128 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3129 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3131 let isCommutable = Commutable;
3133 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3134 (ins VR128:$src1, i128mem:$src2),
3135 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3137 (IntId128 VR128:$src1,
3138 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3142 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3143 int_x86_sse41_pcmpeqq, 1>;
3144 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3145 int_x86_sse41_packusdw, 0>;
3146 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3147 int_x86_sse41_pminsb, 1>;
3148 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3149 int_x86_sse41_pminsd, 1>;
3150 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3151 int_x86_sse41_pminud, 1>;
3152 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3153 int_x86_sse41_pminuw, 1>;
3154 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3155 int_x86_sse41_pmaxsb, 1>;
3156 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3157 int_x86_sse41_pmaxsd, 1>;
3158 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3159 int_x86_sse41_pmaxud, 1>;
3160 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3161 int_x86_sse41_pmaxuw, 1>;
3162 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq",
3163 int_x86_sse41_pmuldq, 1>;
3166 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3167 let Constraints = "$src1 = $dst" in {
3168 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, SDNode OpNode,
3169 Intrinsic IntId128, bit Commutable = 0> {
3170 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3171 (ins VR128:$src1, VR128:$src2),
3172 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3173 [(set VR128:$dst, (OpNode (v4i32 VR128:$src1),
3174 VR128:$src2))]>, OpSize {
3175 let isCommutable = Commutable;
3177 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3178 (ins VR128:$src1, VR128:$src2),
3179 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3180 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3182 let isCommutable = Commutable;
3184 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3185 (ins VR128:$src1, i128mem:$src2),
3186 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3188 (OpNode VR128:$src1, (memopv4i32 addr:$src2)))]>, OpSize;
3189 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3190 (ins VR128:$src1, i128mem:$src2),
3191 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3193 (IntId128 VR128:$src1, (memopv4i32 addr:$src2)))]>,
3197 defm PMULLD : SS41I_binop_patint<0x40, "pmulld", mul,
3198 int_x86_sse41_pmulld, 1>;
3201 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3202 let Constraints = "$src1 = $dst" in {
3203 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3204 Intrinsic IntId128, bit Commutable = 0> {
3205 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3206 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3207 !strconcat(OpcodeStr,
3208 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3210 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3212 let isCommutable = Commutable;
3214 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3215 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3216 !strconcat(OpcodeStr,
3217 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3219 (IntId128 VR128:$src1,
3220 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3225 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3226 int_x86_sse41_blendps, 0>;
3227 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3228 int_x86_sse41_blendpd, 0>;
3229 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3230 int_x86_sse41_pblendw, 0>;
3231 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3232 int_x86_sse41_dpps, 1>;
3233 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3234 int_x86_sse41_dppd, 1>;
3235 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3236 int_x86_sse41_mpsadbw, 0>;
3239 /// SS41I_ternary_int - SSE 4.1 ternary operator
3240 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3241 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3242 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3243 (ins VR128:$src1, VR128:$src2),
3244 !strconcat(OpcodeStr,
3245 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3246 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3249 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3250 (ins VR128:$src1, i128mem:$src2),
3251 !strconcat(OpcodeStr,
3252 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3255 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3259 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3260 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3261 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3264 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3265 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3266 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3267 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3269 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3270 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3272 (IntId (bitconvert (v4i32 (load addr:$src)))))]>, OpSize;
3275 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3276 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3277 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3278 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3279 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3280 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3282 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3283 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3284 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3285 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3287 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3288 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3290 (IntId (bitconvert (v4i32 (load addr:$src)))))]>, OpSize;
3293 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3294 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3295 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3296 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3298 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3299 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3300 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3301 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3303 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3304 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3306 (IntId (bitconvert (v4i32 (load addr:$src)))))]>, OpSize;
3309 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3310 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovsxbq", int_x86_sse41_pmovzxbq>;
3313 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3314 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3315 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3316 (ins VR128:$src1, i32i8imm:$src2),
3317 !strconcat(OpcodeStr,
3318 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3319 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3321 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3322 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3323 !strconcat(OpcodeStr,
3324 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3327 // There's an AssertZext in the way of writing the store pattern
3328 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3331 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3334 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3335 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3336 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3337 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3338 !strconcat(OpcodeStr,
3339 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3342 // There's an AssertZext in the way of writing the store pattern
3343 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3346 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3349 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3350 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3351 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3352 (ins VR128:$src1, i32i8imm:$src2),
3353 !strconcat(OpcodeStr,
3354 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3356 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3357 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3358 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3359 !strconcat(OpcodeStr,
3360 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3361 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3362 addr:$dst)]>, OpSize;
3365 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3368 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3370 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3371 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3372 (ins VR128:$src1, i32i8imm:$src2),
3373 !strconcat(OpcodeStr,
3374 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3376 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3378 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3379 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3380 !strconcat(OpcodeStr,
3381 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3382 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3383 addr:$dst)]>, OpSize;
3386 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3388 let Constraints = "$src1 = $dst" in {
3389 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3390 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3391 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3392 !strconcat(OpcodeStr,
3393 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3395 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3396 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3397 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3398 !strconcat(OpcodeStr,
3399 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3401 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3402 imm:$src3))]>, OpSize;
3406 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3408 let Constraints = "$src1 = $dst" in {
3409 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3410 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3411 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3412 !strconcat(OpcodeStr,
3413 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3415 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3417 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3418 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3419 !strconcat(OpcodeStr,
3420 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3422 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3423 imm:$src3)))]>, OpSize;
3427 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3429 let Constraints = "$src1 = $dst" in {
3430 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3431 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3432 (ins VR128:$src1, FR32:$src2, i32i8imm:$src3),
3433 !strconcat(OpcodeStr,
3434 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3436 (X86insrtps VR128:$src1, FR32:$src2, imm:$src3))]>, OpSize;
3437 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3438 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3439 !strconcat(OpcodeStr,
3440 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3442 (X86insrtps VR128:$src1, (loadf32 addr:$src2),
3443 imm:$src3))]>, OpSize;
3447 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3449 let Defs = [EFLAGS] in {
3450 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3451 "ptest \t{$src2, $src1|$src1, $src2}", []>, OpSize;
3452 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3453 "ptest \t{$src2, $src1|$src1, $src2}", []>, OpSize;
3456 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3457 "movntdqa\t{$src, $dst|$dst, $src}",
3458 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;