1 //====- X86InstrSSE.td - Describe the X86 Instruction Set -------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Evan Cheng and is distributed under the University
6 // of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the X86 SSE instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
17 //===----------------------------------------------------------------------===//
18 // SSE specific DAG Nodes.
19 //===----------------------------------------------------------------------===//
21 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
22 SDTCisFP<0>, SDTCisInt<2> ]>;
24 def X86loadp : SDNode<"X86ISD::LOAD_PACK", SDTLoad, [SDNPHasChain]>;
25 def X86loadu : SDNode<"X86ISD::LOAD_UA", SDTLoad, [SDNPHasChain]>;
26 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
27 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
28 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
29 [SDNPCommutative, SDNPAssociative]>;
30 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
31 [SDNPCommutative, SDNPAssociative]>;
32 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
33 [SDNPCommutative, SDNPAssociative]>;
34 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
35 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
36 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
37 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest,
38 [SDNPHasChain, SDNPOutFlag]>;
39 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest,
40 [SDNPHasChain, SDNPOutFlag]>;
41 def X86s2vec : SDNode<"X86ISD::S2VEC", SDTypeProfile<1, 1, []>, []>;
42 def X86pextrw : SDNode<"X86ISD::PEXTRW", SDTypeProfile<1, 2, []>, []>;
43 def X86pinsrw : SDNode<"X86ISD::PINSRW", SDTypeProfile<1, 3, []>, []>;
45 //===----------------------------------------------------------------------===//
46 // SSE 'Special' Instructions
47 //===----------------------------------------------------------------------===//
49 def IMPLICIT_DEF_VR128 : I<0, Pseudo, (outs VR128:$dst), (ins),
51 [(set VR128:$dst, (v4f32 (undef)))]>,
53 def IMPLICIT_DEF_FR32 : I<0, Pseudo, (outs FR32:$dst), (ins),
55 [(set FR32:$dst, (undef))]>, Requires<[HasSSE2]>;
56 def IMPLICIT_DEF_FR64 : I<0, Pseudo, (outs FR64:$dst), (ins),
58 [(set FR64:$dst, (undef))]>, Requires<[HasSSE2]>;
60 //===----------------------------------------------------------------------===//
61 // SSE Complex Patterns
62 //===----------------------------------------------------------------------===//
64 // These are 'extloads' from a scalar to the low element of a vector, zeroing
65 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
67 def sse_load_f32 : ComplexPattern<v4f32, 4, "SelectScalarSSELoad", [],
69 def sse_load_f64 : ComplexPattern<v2f64, 4, "SelectScalarSSELoad", [],
72 def ssmem : Operand<v4f32> {
73 let PrintMethod = "printf32mem";
74 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
76 def sdmem : Operand<v2f64> {
77 let PrintMethod = "printf64mem";
78 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
81 //===----------------------------------------------------------------------===//
82 // SSE pattern fragments
83 //===----------------------------------------------------------------------===//
85 def X86loadpf32 : PatFrag<(ops node:$ptr), (f32 (X86loadp node:$ptr))>;
86 def X86loadpf64 : PatFrag<(ops node:$ptr), (f64 (X86loadp node:$ptr))>;
88 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
89 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
90 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
91 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
93 // Like 'store', but always requires natural alignment.
94 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
95 (st node:$val, node:$ptr), [{
96 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
97 return !ST->isTruncatingStore() &&
98 ST->getAddressingMode() == ISD::UNINDEXED &&
99 ST->getAlignment() * 8 >= MVT::getSizeInBits(ST->getStoredVT());
103 // Like 'load', but always requires natural alignment.
104 def alignedload : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
105 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
106 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
107 LD->getAddressingMode() == ISD::UNINDEXED &&
108 LD->getAlignment() * 8 >= MVT::getSizeInBits(LD->getLoadedVT());
112 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
113 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
114 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
115 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
117 // Like 'load', but uses special alignment checks suitable for use in
118 // memory operands in most SSE instructions, which are required to
119 // be naturally aligned on some targets but not on others.
120 // FIXME: Actually implement support for targets that don't require the
121 // alignment. This probably wants a subtarget predicate.
122 def memop : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
123 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
124 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
125 LD->getAddressingMode() == ISD::UNINDEXED &&
126 LD->getAlignment() * 8 >= MVT::getSizeInBits(LD->getLoadedVT());
130 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
131 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
132 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
133 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
135 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
136 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
137 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
138 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
139 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
140 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
142 def fp32imm0 : PatLeaf<(f32 fpimm), [{
143 return N->isExactlyValue(+0.0);
146 def PSxLDQ_imm : SDNodeXForm<imm, [{
147 // Transformation function: imm >> 3
148 return getI32Imm(N->getValue() >> 3);
151 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
153 def SHUFFLE_get_shuf_imm : SDNodeXForm<build_vector, [{
154 return getI8Imm(X86::getShuffleSHUFImmediate(N));
157 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
159 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<build_vector, [{
160 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
163 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
165 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<build_vector, [{
166 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
169 def SSE_splat_mask : PatLeaf<(build_vector), [{
170 return X86::isSplatMask(N);
171 }], SHUFFLE_get_shuf_imm>;
173 def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
174 return X86::isSplatLoMask(N);
177 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
178 return X86::isMOVHLPSMask(N);
181 def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
182 return X86::isMOVHLPS_v_undef_Mask(N);
185 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
186 return X86::isMOVHPMask(N);
189 def MOVLP_shuffle_mask : PatLeaf<(build_vector), [{
190 return X86::isMOVLPMask(N);
193 def MOVL_shuffle_mask : PatLeaf<(build_vector), [{
194 return X86::isMOVLMask(N);
197 def MOVSHDUP_shuffle_mask : PatLeaf<(build_vector), [{
198 return X86::isMOVSHDUPMask(N);
201 def MOVSLDUP_shuffle_mask : PatLeaf<(build_vector), [{
202 return X86::isMOVSLDUPMask(N);
205 def UNPCKL_shuffle_mask : PatLeaf<(build_vector), [{
206 return X86::isUNPCKLMask(N);
209 def UNPCKH_shuffle_mask : PatLeaf<(build_vector), [{
210 return X86::isUNPCKHMask(N);
213 def UNPCKL_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
214 return X86::isUNPCKL_v_undef_Mask(N);
217 def UNPCKH_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
218 return X86::isUNPCKH_v_undef_Mask(N);
221 def PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
222 return X86::isPSHUFDMask(N);
223 }], SHUFFLE_get_shuf_imm>;
225 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
226 return X86::isPSHUFHWMask(N);
227 }], SHUFFLE_get_pshufhw_imm>;
229 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
230 return X86::isPSHUFLWMask(N);
231 }], SHUFFLE_get_pshuflw_imm>;
233 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
234 return X86::isPSHUFDMask(N);
235 }], SHUFFLE_get_shuf_imm>;
237 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
238 return X86::isSHUFPMask(N);
239 }], SHUFFLE_get_shuf_imm>;
241 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
242 return X86::isSHUFPMask(N);
243 }], SHUFFLE_get_shuf_imm>;
245 //===----------------------------------------------------------------------===//
246 // SSE scalar FP Instructions
247 //===----------------------------------------------------------------------===//
249 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
250 // scheduler into a branch sequence.
251 let usesCustomDAGSchedInserter = 1 in { // Expanded by the scheduler.
252 def CMOV_FR32 : I<0, Pseudo,
253 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
254 "#CMOV_FR32 PSEUDO!",
255 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond))]>;
256 def CMOV_FR64 : I<0, Pseudo,
257 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
258 "#CMOV_FR64 PSEUDO!",
259 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond))]>;
260 def CMOV_V4F32 : I<0, Pseudo,
261 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
262 "#CMOV_V4F32 PSEUDO!",
264 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
265 def CMOV_V2F64 : I<0, Pseudo,
266 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
267 "#CMOV_V2F64 PSEUDO!",
269 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
270 def CMOV_V2I64 : I<0, Pseudo,
271 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
272 "#CMOV_V2I64 PSEUDO!",
274 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
277 //===----------------------------------------------------------------------===//
279 //===----------------------------------------------------------------------===//
281 // SSE1 Instruction Templates:
283 // SSI - SSE1 instructions with XS prefix.
284 // PSI - SSE1 instructions with TB prefix.
285 // PSIi8 - SSE1 instructions with ImmT == Imm8 and TB prefix.
287 class SSI<bits<8> o, Format F, dag outs, dag ins, string asm, list<dag> pattern>
288 : I<o, F, outs, ins, asm, pattern>, XS, Requires<[HasSSE1]>;
289 class PSI<bits<8> o, Format F, dag outs, dag ins, string asm, list<dag> pattern>
290 : I<o, F, outs, ins, asm, pattern>, TB, Requires<[HasSSE1]>;
291 class PSIi8<bits<8> o, Format F, dag outs, dag ins, string asm,
293 : Ii8<o, F, outs, ins, asm, pattern>, TB, Requires<[HasSSE1]>;
296 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
297 "movss {$src, $dst|$dst, $src}", []>;
298 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
299 "movss {$src, $dst|$dst, $src}",
300 [(set FR32:$dst, (loadf32 addr:$src))]>;
301 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
302 "movss {$src, $dst|$dst, $src}",
303 [(store FR32:$src, addr:$dst)]>;
305 // Conversion instructions
306 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
307 "cvttss2si {$src, $dst|$dst, $src}",
308 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
309 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
310 "cvttss2si {$src, $dst|$dst, $src}",
311 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
312 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
313 "cvtsi2ss {$src, $dst|$dst, $src}",
314 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
315 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
316 "cvtsi2ss {$src, $dst|$dst, $src}",
317 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
319 // Match intrinsics which expect XMM operand(s).
320 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
321 "cvtss2si {$src, $dst|$dst, $src}",
322 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
323 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
324 "cvtss2si {$src, $dst|$dst, $src}",
325 [(set GR32:$dst, (int_x86_sse_cvtss2si
326 (load addr:$src)))]>;
328 // Aliases for intrinsics
329 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
330 "cvttss2si {$src, $dst|$dst, $src}",
332 (int_x86_sse_cvttss2si VR128:$src))]>;
333 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
334 "cvttss2si {$src, $dst|$dst, $src}",
336 (int_x86_sse_cvttss2si(load addr:$src)))]>;
338 let isTwoAddress = 1 in {
339 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
340 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
341 "cvtsi2ss {$src2, $dst|$dst, $src2}",
342 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
344 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
345 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
346 "cvtsi2ss {$src2, $dst|$dst, $src2}",
347 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
348 (loadi32 addr:$src2)))]>;
351 // Comparison instructions
352 let isTwoAddress = 1 in {
353 def CMPSSrr : SSI<0xC2, MRMSrcReg,
354 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
355 "cmp${cc}ss {$src, $dst|$dst, $src}",
357 def CMPSSrm : SSI<0xC2, MRMSrcMem,
358 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
359 "cmp${cc}ss {$src, $dst|$dst, $src}", []>;
362 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
363 "ucomiss {$src2, $src1|$src1, $src2}",
364 [(X86cmp FR32:$src1, FR32:$src2)]>;
365 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
366 "ucomiss {$src2, $src1|$src1, $src2}",
367 [(X86cmp FR32:$src1, (loadf32 addr:$src2))]>;
369 // Aliases to match intrinsics which expect XMM operand(s).
370 let isTwoAddress = 1 in {
371 def Int_CMPSSrr : SSI<0xC2, MRMSrcReg,
372 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
373 "cmp${cc}ss {$src, $dst|$dst, $src}",
374 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
375 VR128:$src, imm:$cc))]>;
376 def Int_CMPSSrm : SSI<0xC2, MRMSrcMem,
377 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src, SSECC:$cc),
378 "cmp${cc}ss {$src, $dst|$dst, $src}",
379 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
380 (load addr:$src), imm:$cc))]>;
383 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
384 "ucomiss {$src2, $src1|$src1, $src2}",
385 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2)]>;
386 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
387 "ucomiss {$src2, $src1|$src1, $src2}",
388 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2))]>;
390 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
391 "comiss {$src2, $src1|$src1, $src2}",
392 [(X86comi (v4f32 VR128:$src1), VR128:$src2)]>;
393 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
394 "comiss {$src2, $src1|$src1, $src2}",
395 [(X86comi (v4f32 VR128:$src1), (load addr:$src2))]>;
397 // Aliases of packed SSE1 instructions for scalar use. These all have names that
400 // Alias instructions that map fld0 to pxor for sse.
401 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
402 "pxor $dst, $dst", [(set FR32:$dst, fp32imm0)]>,
403 Requires<[HasSSE1]>, TB, OpSize;
405 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
407 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
408 "movaps {$src, $dst|$dst, $src}", []>;
410 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
412 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
413 "movaps {$src, $dst|$dst, $src}",
414 [(set FR32:$dst, (X86loadpf32 addr:$src))]>;
416 // Alias bitwise logical operations using SSE logical ops on packed FP values.
417 let isTwoAddress = 1 in {
418 let isCommutable = 1 in {
419 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
420 "andps {$src2, $dst|$dst, $src2}",
421 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
422 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
423 "orps {$src2, $dst|$dst, $src2}",
424 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
425 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
426 "xorps {$src2, $dst|$dst, $src2}",
427 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
430 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
431 "andps {$src2, $dst|$dst, $src2}",
432 [(set FR32:$dst, (X86fand FR32:$src1,
433 (X86loadpf32 addr:$src2)))]>;
434 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
435 "orps {$src2, $dst|$dst, $src2}",
436 [(set FR32:$dst, (X86for FR32:$src1,
437 (X86loadpf32 addr:$src2)))]>;
438 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
439 "xorps {$src2, $dst|$dst, $src2}",
440 [(set FR32:$dst, (X86fxor FR32:$src1,
441 (X86loadpf32 addr:$src2)))]>;
443 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
444 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
445 "andnps {$src2, $dst|$dst, $src2}", []>;
446 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
447 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
448 "andnps {$src2, $dst|$dst, $src2}", []>;
451 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
453 /// In addition, we also have a special variant of the scalar form here to
454 /// represent the associated intrinsic operation. This form is unlike the
455 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
456 /// and leaves the top elements undefined.
458 /// These three forms can each be reg+reg or reg+mem, so there are a total of
459 /// six "instructions".
461 let isTwoAddress = 1 in {
462 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
463 SDNode OpNode, Intrinsic F32Int,
464 bit Commutable = 0> {
465 // Scalar operation, reg+reg.
466 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
467 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
468 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
469 let isCommutable = Commutable;
472 // Scalar operation, reg+mem.
473 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
474 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
475 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
477 // Vector operation, reg+reg.
478 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
479 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2}"),
480 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
481 let isCommutable = Commutable;
484 // Vector operation, reg+mem.
485 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
486 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2}"),
487 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
489 // Intrinsic operation, reg+reg.
490 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
491 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
492 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
493 let isCommutable = Commutable;
496 // Intrinsic operation, reg+mem.
497 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
498 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
499 [(set VR128:$dst, (F32Int VR128:$src1,
500 sse_load_f32:$src2))]>;
504 // Arithmetic instructions
505 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
506 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
507 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
508 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
510 /// sse1_fp_binop_rm - Other SSE1 binops
512 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
513 /// instructions for a full-vector intrinsic form. Operations that map
514 /// onto C operators don't use this form since they just use the plain
515 /// vector form instead of having a separate vector intrinsic form.
517 /// This provides a total of eight "instructions".
519 let isTwoAddress = 1 in {
520 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
524 bit Commutable = 0> {
526 // Scalar operation, reg+reg.
527 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
528 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
529 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
530 let isCommutable = Commutable;
533 // Scalar operation, reg+mem.
534 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
535 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
536 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
538 // Vector operation, reg+reg.
539 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
540 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2}"),
541 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
542 let isCommutable = Commutable;
545 // Vector operation, reg+mem.
546 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
547 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2}"),
548 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
550 // Intrinsic operation, reg+reg.
551 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
552 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
553 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
554 let isCommutable = Commutable;
557 // Intrinsic operation, reg+mem.
558 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
559 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
560 [(set VR128:$dst, (F32Int VR128:$src1,
561 sse_load_f32:$src2))]>;
563 // Vector intrinsic operation, reg+reg.
564 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
565 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2}"),
566 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
567 let isCommutable = Commutable;
570 // Vector intrinsic operation, reg+mem.
571 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
572 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2}"),
573 [(set VR128:$dst, (V4F32Int VR128:$src1, (load addr:$src2)))]>;
577 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
578 int_x86_sse_max_ss, int_x86_sse_max_ps>;
579 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
580 int_x86_sse_min_ss, int_x86_sse_min_ps>;
582 //===----------------------------------------------------------------------===//
583 // SSE packed FP Instructions
586 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
587 "movaps {$src, $dst|$dst, $src}", []>;
588 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
589 "movaps {$src, $dst|$dst, $src}",
590 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
592 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
593 "movaps {$src, $dst|$dst, $src}",
594 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
596 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
597 "movups {$src, $dst|$dst, $src}", []>;
598 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
599 "movups {$src, $dst|$dst, $src}",
600 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
601 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
602 "movups {$src, $dst|$dst, $src}",
603 [(store (v4f32 VR128:$src), addr:$dst)]>;
605 // Intrinsic forms of MOVUPS load and store
606 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
607 "movups {$src, $dst|$dst, $src}",
608 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
609 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
610 "movups {$src, $dst|$dst, $src}",
611 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
613 let isTwoAddress = 1 in {
614 let AddedComplexity = 20 in {
615 def MOVLPSrm : PSI<0x12, MRMSrcMem,
616 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
617 "movlps {$src2, $dst|$dst, $src2}",
619 (v4f32 (vector_shuffle VR128:$src1,
620 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
621 MOVLP_shuffle_mask)))]>;
622 def MOVHPSrm : PSI<0x16, MRMSrcMem,
623 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
624 "movhps {$src2, $dst|$dst, $src2}",
626 (v4f32 (vector_shuffle VR128:$src1,
627 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
628 MOVHP_shuffle_mask)))]>;
632 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
633 "movlps {$src, $dst|$dst, $src}",
634 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
635 (iPTR 0))), addr:$dst)]>;
637 // v2f64 extract element 1 is always custom lowered to unpack high to low
638 // and extract element 0 so the non-store version isn't too horrible.
639 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
640 "movhps {$src, $dst|$dst, $src}",
641 [(store (f64 (vector_extract
642 (v2f64 (vector_shuffle
643 (bc_v2f64 (v4f32 VR128:$src)), (undef),
644 UNPCKH_shuffle_mask)), (iPTR 0))),
647 let isTwoAddress = 1 in {
648 let AddedComplexity = 15 in {
649 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
650 "movlhps {$src2, $dst|$dst, $src2}",
652 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
653 MOVHP_shuffle_mask)))]>;
655 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
656 "movhlps {$src2, $dst|$dst, $src2}",
658 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
659 MOVHLPS_shuffle_mask)))]>;
667 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
669 /// In addition, we also have a special variant of the scalar form here to
670 /// represent the associated intrinsic operation. This form is unlike the
671 /// plain scalar form, in that it takes an entire vector (instead of a
672 /// scalar) and leaves the top elements undefined.
674 /// And, we have a special variant form for a full-vector intrinsic form.
676 /// These four forms can each have a reg or a mem operand, so there are a
677 /// total of eight "instructions".
679 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
683 bit Commutable = 0> {
684 // Scalar operation, reg.
685 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
686 !strconcat(OpcodeStr, "ss {$src, $dst|$dst, $src}"),
687 [(set FR32:$dst, (OpNode FR32:$src))]> {
688 let isCommutable = Commutable;
691 // Scalar operation, mem.
692 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
693 !strconcat(OpcodeStr, "ss {$src, $dst|$dst, $src}"),
694 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
696 // Vector operation, reg.
697 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
698 !strconcat(OpcodeStr, "ps {$src, $dst|$dst, $src}"),
699 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
700 let isCommutable = Commutable;
703 // Vector operation, mem.
704 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
705 !strconcat(OpcodeStr, "ps {$src, $dst|$dst, $src}"),
706 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
708 // Intrinsic operation, reg.
709 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
710 !strconcat(OpcodeStr, "ss {$src, $dst|$dst, $src}"),
711 [(set VR128:$dst, (F32Int VR128:$src))]> {
712 let isCommutable = Commutable;
715 // Intrinsic operation, mem.
716 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
717 !strconcat(OpcodeStr, "ss {$src, $dst|$dst, $src}"),
718 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
720 // Vector intrinsic operation, reg
721 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
722 !strconcat(OpcodeStr, "ps {$src, $dst|$dst, $src}"),
723 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
724 let isCommutable = Commutable;
727 // Vector intrinsic operation, mem
728 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
729 !strconcat(OpcodeStr, "ps {$src, $dst|$dst, $src}"),
730 [(set VR128:$dst, (V4F32Int (load addr:$src)))]>;
734 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
735 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
737 // Reciprocal approximations. Note that these typically require refinement
738 // in order to obtain suitable precision.
739 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
740 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
741 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
742 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
745 let isTwoAddress = 1 in {
746 let isCommutable = 1 in {
747 def ANDPSrr : PSI<0x54, MRMSrcReg,
748 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
749 "andps {$src2, $dst|$dst, $src2}",
750 [(set VR128:$dst, (v2i64
751 (and VR128:$src1, VR128:$src2)))]>;
752 def ORPSrr : PSI<0x56, MRMSrcReg,
753 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
754 "orps {$src2, $dst|$dst, $src2}",
755 [(set VR128:$dst, (v2i64
756 (or VR128:$src1, VR128:$src2)))]>;
757 def XORPSrr : PSI<0x57, MRMSrcReg,
758 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
759 "xorps {$src2, $dst|$dst, $src2}",
760 [(set VR128:$dst, (v2i64
761 (xor VR128:$src1, VR128:$src2)))]>;
764 def ANDPSrm : PSI<0x54, MRMSrcMem,
765 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
766 "andps {$src2, $dst|$dst, $src2}",
767 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
768 (memopv2i64 addr:$src2)))]>;
769 def ORPSrm : PSI<0x56, MRMSrcMem,
770 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
771 "orps {$src2, $dst|$dst, $src2}",
772 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
773 (memopv2i64 addr:$src2)))]>;
774 def XORPSrm : PSI<0x57, MRMSrcMem,
775 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
776 "xorps {$src2, $dst|$dst, $src2}",
777 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
778 (memopv2i64 addr:$src2)))]>;
779 def ANDNPSrr : PSI<0x55, MRMSrcReg,
780 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
781 "andnps {$src2, $dst|$dst, $src2}",
783 (v2i64 (and (xor VR128:$src1,
784 (bc_v2i64 (v4i32 immAllOnesV))),
786 def ANDNPSrm : PSI<0x55, MRMSrcMem,
787 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
788 "andnps {$src2, $dst|$dst, $src2}",
790 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
791 (bc_v2i64 (v4i32 immAllOnesV))),
792 (memopv2i64 addr:$src2))))]>;
795 let isTwoAddress = 1 in {
796 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
797 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
798 "cmp${cc}ps {$src, $dst|$dst, $src}",
799 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
800 VR128:$src, imm:$cc))]>;
801 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
802 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
803 "cmp${cc}ps {$src, $dst|$dst, $src}",
804 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
805 (load addr:$src), imm:$cc))]>;
808 // Shuffle and unpack instructions
809 let isTwoAddress = 1 in {
810 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
811 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
812 (outs VR128:$dst), (ins VR128:$src1,
813 VR128:$src2, i32i8imm:$src3),
814 "shufps {$src3, $src2, $dst|$dst, $src2, $src3}",
816 (v4f32 (vector_shuffle
817 VR128:$src1, VR128:$src2,
818 SHUFP_shuffle_mask:$src3)))]>;
819 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
820 (outs VR128:$dst), (ins VR128:$src1,
821 f128mem:$src2, i32i8imm:$src3),
822 "shufps {$src3, $src2, $dst|$dst, $src2, $src3}",
824 (v4f32 (vector_shuffle
825 VR128:$src1, (load addr:$src2),
826 SHUFP_shuffle_mask:$src3)))]>;
828 let AddedComplexity = 10 in {
829 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
830 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
831 "unpckhps {$src2, $dst|$dst, $src2}",
833 (v4f32 (vector_shuffle
834 VR128:$src1, VR128:$src2,
835 UNPCKH_shuffle_mask)))]>;
836 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
837 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
838 "unpckhps {$src2, $dst|$dst, $src2}",
840 (v4f32 (vector_shuffle
841 VR128:$src1, (load addr:$src2),
842 UNPCKH_shuffle_mask)))]>;
844 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
845 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
846 "unpcklps {$src2, $dst|$dst, $src2}",
848 (v4f32 (vector_shuffle
849 VR128:$src1, VR128:$src2,
850 UNPCKL_shuffle_mask)))]>;
851 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
852 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
853 "unpcklps {$src2, $dst|$dst, $src2}",
855 (v4f32 (vector_shuffle
856 VR128:$src1, (load addr:$src2),
857 UNPCKL_shuffle_mask)))]>;
862 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
863 "movmskps {$src, $dst|$dst, $src}",
864 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
865 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
866 "movmskpd {$src, $dst|$dst, $src}",
867 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
869 // Prefetching loads.
870 // TODO: no intrinsics for these?
871 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src), "prefetcht0 $src", []>;
872 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src), "prefetcht1 $src", []>;
873 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src), "prefetcht2 $src", []>;
874 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src), "prefetchnta $src", []>;
876 // Non-temporal stores
877 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
878 "movntps {$src, $dst|$dst, $src}",
879 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
881 // Load, store, and memory fence
882 def SFENCE : PSI<0xAE, MRM7m, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
885 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
886 "ldmxcsr $src", [(int_x86_sse_ldmxcsr addr:$src)]>;
887 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
888 "stmxcsr $dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
890 // Alias instructions that map zero vector to pxor / xorp* for sse.
891 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
892 let isReMaterializable = 1 in
893 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
895 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
897 // FR32 to 128-bit vector conversion.
898 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
899 "movss {$src, $dst|$dst, $src}",
901 (v4f32 (scalar_to_vector FR32:$src)))]>;
902 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
903 "movss {$src, $dst|$dst, $src}",
905 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
907 // FIXME: may not be able to eliminate this movss with coalescing the src and
908 // dest register classes are different. We really want to write this pattern
910 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
912 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
913 "movss {$src, $dst|$dst, $src}",
914 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
916 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
917 "movss {$src, $dst|$dst, $src}",
918 [(store (f32 (vector_extract (v4f32 VR128:$src),
919 (iPTR 0))), addr:$dst)]>;
922 // Move to lower bits of a VR128, leaving upper bits alone.
923 // Three operand (but two address) aliases.
924 let isTwoAddress = 1 in {
925 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
926 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
927 "movss {$src2, $dst|$dst, $src2}", []>;
929 let AddedComplexity = 15 in
930 def MOVLPSrr : SSI<0x10, MRMSrcReg,
931 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
932 "movss {$src2, $dst|$dst, $src2}",
934 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
935 MOVL_shuffle_mask)))]>;
938 // Move to lower bits of a VR128 and zeroing upper bits.
939 // Loading from memory automatically zeroing upper bits.
940 let AddedComplexity = 20 in
941 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
942 "movss {$src, $dst|$dst, $src}",
943 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV,
944 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
945 MOVL_shuffle_mask)))]>;
948 //===----------------------------------------------------------------------===//
950 //===----------------------------------------------------------------------===//
952 // SSE2 Instruction Templates:
954 // SDI - SSE2 instructions with XD prefix.
955 // PDI - SSE2 instructions with TB and OpSize prefixes.
956 // PDIi8 - SSE2 instructions with ImmT == Imm8 and TB and OpSize prefixes.
958 class SDI<bits<8> o, Format F, dag outs, dag ins, string asm, list<dag> pattern>
959 : I<o, F, outs, ins, asm, pattern>, XD, Requires<[HasSSE2]>;
960 class PDI<bits<8> o, Format F, dag outs, dag ins, string asm, list<dag> pattern>
961 : I<o, F, outs, ins, asm, pattern>, TB, OpSize, Requires<[HasSSE2]>;
962 class PDIi8<bits<8> o, Format F, dag outs, dag ins, string asm,
964 : Ii8<o, F, outs, ins, asm, pattern>, TB, OpSize, Requires<[HasSSE2]>;
967 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
968 "movsd {$src, $dst|$dst, $src}", []>;
969 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
970 "movsd {$src, $dst|$dst, $src}",
971 [(set FR64:$dst, (loadf64 addr:$src))]>;
972 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
973 "movsd {$src, $dst|$dst, $src}",
974 [(store FR64:$src, addr:$dst)]>;
976 // Conversion instructions
977 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
978 "cvttsd2si {$src, $dst|$dst, $src}",
979 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
980 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
981 "cvttsd2si {$src, $dst|$dst, $src}",
982 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
983 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
984 "cvtsd2ss {$src, $dst|$dst, $src}",
985 [(set FR32:$dst, (fround FR64:$src))]>;
986 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
987 "cvtsd2ss {$src, $dst|$dst, $src}",
988 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
989 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
990 "cvtsi2sd {$src, $dst|$dst, $src}",
991 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
992 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
993 "cvtsi2sd {$src, $dst|$dst, $src}",
994 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
996 // SSE2 instructions with XS prefix
997 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
998 "cvtss2sd {$src, $dst|$dst, $src}",
999 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1000 Requires<[HasSSE2]>;
1001 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1002 "cvtss2sd {$src, $dst|$dst, $src}",
1003 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1004 Requires<[HasSSE2]>;
1006 // Match intrinsics which expect XMM operand(s).
1007 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1008 "cvtsd2si {$src, $dst|$dst, $src}",
1009 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1010 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1011 "cvtsd2si {$src, $dst|$dst, $src}",
1012 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1013 (load addr:$src)))]>;
1015 // Aliases for intrinsics
1016 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1017 "cvttsd2si {$src, $dst|$dst, $src}",
1019 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1020 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1021 "cvttsd2si {$src, $dst|$dst, $src}",
1022 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1023 (load addr:$src)))]>;
1025 // Comparison instructions
1026 let isTwoAddress = 1 in {
1027 def CMPSDrr : SDI<0xC2, MRMSrcReg,
1028 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1029 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
1030 def CMPSDrm : SDI<0xC2, MRMSrcMem,
1031 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1032 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
1035 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1036 "ucomisd {$src2, $src1|$src1, $src2}",
1037 [(X86cmp FR64:$src1, FR64:$src2)]>;
1038 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1039 "ucomisd {$src2, $src1|$src1, $src2}",
1040 [(X86cmp FR64:$src1, (loadf64 addr:$src2))]>;
1042 // Aliases to match intrinsics which expect XMM operand(s).
1043 let isTwoAddress = 1 in {
1044 def Int_CMPSDrr : SDI<0xC2, MRMSrcReg,
1045 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1046 "cmp${cc}sd {$src, $dst|$dst, $src}",
1047 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1048 VR128:$src, imm:$cc))]>;
1049 def Int_CMPSDrm : SDI<0xC2, MRMSrcMem,
1050 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1051 "cmp${cc}sd {$src, $dst|$dst, $src}",
1052 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1053 (load addr:$src), imm:$cc))]>;
1056 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1057 "ucomisd {$src2, $src1|$src1, $src2}",
1058 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
1059 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1060 "ucomisd {$src2, $src1|$src1, $src2}",
1061 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2))]>;
1063 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1064 "comisd {$src2, $src1|$src1, $src2}",
1065 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
1066 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1067 "comisd {$src2, $src1|$src1, $src2}",
1068 [(X86comi (v2f64 VR128:$src1), (load addr:$src2))]>;
1070 // Aliases of packed SSE2 instructions for scalar use. These all have names that
1073 // Alias instructions that map fld0 to pxor for sse.
1074 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1075 "pxor $dst, $dst", [(set FR64:$dst, fpimm0)]>,
1076 Requires<[HasSSE2]>, TB, OpSize;
1078 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1080 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1081 "movapd {$src, $dst|$dst, $src}", []>;
1083 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1085 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1086 "movapd {$src, $dst|$dst, $src}",
1087 [(set FR64:$dst, (X86loadpf64 addr:$src))]>;
1089 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1090 let isTwoAddress = 1 in {
1091 let isCommutable = 1 in {
1092 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1093 "andpd {$src2, $dst|$dst, $src2}",
1094 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1095 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1096 "orpd {$src2, $dst|$dst, $src2}",
1097 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1098 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1099 "xorpd {$src2, $dst|$dst, $src2}",
1100 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1103 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1104 "andpd {$src2, $dst|$dst, $src2}",
1105 [(set FR64:$dst, (X86fand FR64:$src1,
1106 (X86loadpf64 addr:$src2)))]>;
1107 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1108 "orpd {$src2, $dst|$dst, $src2}",
1109 [(set FR64:$dst, (X86for FR64:$src1,
1110 (X86loadpf64 addr:$src2)))]>;
1111 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1112 "xorpd {$src2, $dst|$dst, $src2}",
1113 [(set FR64:$dst, (X86fxor FR64:$src1,
1114 (X86loadpf64 addr:$src2)))]>;
1116 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1117 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1118 "andnpd {$src2, $dst|$dst, $src2}", []>;
1119 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1120 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1121 "andnpd {$src2, $dst|$dst, $src2}", []>;
1124 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1126 /// In addition, we also have a special variant of the scalar form here to
1127 /// represent the associated intrinsic operation. This form is unlike the
1128 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1129 /// and leaves the top elements undefined.
1131 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1132 /// six "instructions".
1134 let isTwoAddress = 1 in {
1135 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1136 SDNode OpNode, Intrinsic F64Int,
1137 bit Commutable = 0> {
1138 // Scalar operation, reg+reg.
1139 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1140 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1141 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1142 let isCommutable = Commutable;
1145 // Scalar operation, reg+mem.
1146 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1147 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1148 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1150 // Vector operation, reg+reg.
1151 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1152 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
1153 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1154 let isCommutable = Commutable;
1157 // Vector operation, reg+mem.
1158 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1159 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
1160 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1162 // Intrinsic operation, reg+reg.
1163 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1164 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1165 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1166 let isCommutable = Commutable;
1169 // Intrinsic operation, reg+mem.
1170 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1171 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1172 [(set VR128:$dst, (F64Int VR128:$src1,
1173 sse_load_f64:$src2))]>;
1177 // Arithmetic instructions
1178 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1179 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1180 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1181 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1183 /// sse2_fp_binop_rm - Other SSE2 binops
1185 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1186 /// instructions for a full-vector intrinsic form. Operations that map
1187 /// onto C operators don't use this form since they just use the plain
1188 /// vector form instead of having a separate vector intrinsic form.
1190 /// This provides a total of eight "instructions".
1192 let isTwoAddress = 1 in {
1193 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1197 bit Commutable = 0> {
1199 // Scalar operation, reg+reg.
1200 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1201 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1202 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1203 let isCommutable = Commutable;
1206 // Scalar operation, reg+mem.
1207 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1208 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1209 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1211 // Vector operation, reg+reg.
1212 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1213 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
1214 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1215 let isCommutable = Commutable;
1218 // Vector operation, reg+mem.
1219 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1220 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
1221 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1223 // Intrinsic operation, reg+reg.
1224 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1225 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1226 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1227 let isCommutable = Commutable;
1230 // Intrinsic operation, reg+mem.
1231 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1232 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1233 [(set VR128:$dst, (F64Int VR128:$src1,
1234 sse_load_f64:$src2))]>;
1236 // Vector intrinsic operation, reg+reg.
1237 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1238 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
1239 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1240 let isCommutable = Commutable;
1243 // Vector intrinsic operation, reg+mem.
1244 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1245 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
1246 [(set VR128:$dst, (V2F64Int VR128:$src1, (load addr:$src2)))]>;
1250 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1251 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1252 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1253 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1255 //===----------------------------------------------------------------------===//
1256 // SSE packed FP Instructions
1258 // Move Instructions
1259 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1260 "movapd {$src, $dst|$dst, $src}", []>;
1261 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1262 "movapd {$src, $dst|$dst, $src}",
1263 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1265 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1266 "movapd {$src, $dst|$dst, $src}",
1267 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1269 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1270 "movupd {$src, $dst|$dst, $src}", []>;
1271 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1272 "movupd {$src, $dst|$dst, $src}",
1273 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1274 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1275 "movupd {$src, $dst|$dst, $src}",
1276 [(store (v2f64 VR128:$src), addr:$dst)]>;
1278 // Intrinsic forms of MOVUPD load and store
1279 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1280 "movupd {$src, $dst|$dst, $src}",
1281 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1282 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1283 "movupd {$src, $dst|$dst, $src}",
1284 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1286 let isTwoAddress = 1 in {
1287 let AddedComplexity = 20 in {
1288 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1289 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1290 "movlpd {$src2, $dst|$dst, $src2}",
1292 (v2f64 (vector_shuffle VR128:$src1,
1293 (scalar_to_vector (loadf64 addr:$src2)),
1294 MOVLP_shuffle_mask)))]>;
1295 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1296 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1297 "movhpd {$src2, $dst|$dst, $src2}",
1299 (v2f64 (vector_shuffle VR128:$src1,
1300 (scalar_to_vector (loadf64 addr:$src2)),
1301 MOVHP_shuffle_mask)))]>;
1302 } // AddedComplexity
1305 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1306 "movlpd {$src, $dst|$dst, $src}",
1307 [(store (f64 (vector_extract (v2f64 VR128:$src),
1308 (iPTR 0))), addr:$dst)]>;
1310 // v2f64 extract element 1 is always custom lowered to unpack high to low
1311 // and extract element 0 so the non-store version isn't too horrible.
1312 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1313 "movhpd {$src, $dst|$dst, $src}",
1314 [(store (f64 (vector_extract
1315 (v2f64 (vector_shuffle VR128:$src, (undef),
1316 UNPCKH_shuffle_mask)), (iPTR 0))),
1319 // SSE2 instructions without OpSize prefix
1320 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1321 "cvtdq2ps {$src, $dst|$dst, $src}",
1322 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1323 TB, Requires<[HasSSE2]>;
1324 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1325 "cvtdq2ps {$src, $dst|$dst, $src}",
1326 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1327 (bitconvert (memopv2i64 addr:$src))))]>,
1328 TB, Requires<[HasSSE2]>;
1330 // SSE2 instructions with XS prefix
1331 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1332 "cvtdq2pd {$src, $dst|$dst, $src}",
1333 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1334 XS, Requires<[HasSSE2]>;
1335 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1336 "cvtdq2pd {$src, $dst|$dst, $src}",
1337 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1338 (bitconvert (memopv2i64 addr:$src))))]>,
1339 XS, Requires<[HasSSE2]>;
1341 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1342 "cvtps2dq {$src, $dst|$dst, $src}",
1343 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1344 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1345 "cvtps2dq {$src, $dst|$dst, $src}",
1346 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1347 (load addr:$src)))]>;
1348 // SSE2 packed instructions with XS prefix
1349 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1350 "cvttps2dq {$src, $dst|$dst, $src}",
1351 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1352 XS, Requires<[HasSSE2]>;
1353 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1354 "cvttps2dq {$src, $dst|$dst, $src}",
1355 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1356 (load addr:$src)))]>,
1357 XS, Requires<[HasSSE2]>;
1359 // SSE2 packed instructions with XD prefix
1360 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1361 "cvtpd2dq {$src, $dst|$dst, $src}",
1362 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1363 XD, Requires<[HasSSE2]>;
1364 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1365 "cvtpd2dq {$src, $dst|$dst, $src}",
1366 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1367 (load addr:$src)))]>,
1368 XD, Requires<[HasSSE2]>;
1370 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1371 "cvttpd2dq {$src, $dst|$dst, $src}",
1372 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1373 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1374 "cvttpd2dq {$src, $dst|$dst, $src}",
1375 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1376 (load addr:$src)))]>;
1378 // SSE2 instructions without OpSize prefix
1379 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1380 "cvtps2pd {$src, $dst|$dst, $src}",
1381 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1382 TB, Requires<[HasSSE2]>;
1383 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f64mem:$src),
1384 "cvtps2pd {$src, $dst|$dst, $src}",
1385 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1386 (load addr:$src)))]>,
1387 TB, Requires<[HasSSE2]>;
1389 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1390 "cvtpd2ps {$src, $dst|$dst, $src}",
1391 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1392 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f128mem:$src),
1393 "cvtpd2ps {$src, $dst|$dst, $src}",
1394 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1395 (load addr:$src)))]>;
1397 // Match intrinsics which expect XMM operand(s).
1398 // Aliases for intrinsics
1399 let isTwoAddress = 1 in {
1400 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1401 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1402 "cvtsi2sd {$src2, $dst|$dst, $src2}",
1403 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1405 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1406 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1407 "cvtsi2sd {$src2, $dst|$dst, $src2}",
1408 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1409 (loadi32 addr:$src2)))]>;
1410 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1411 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1412 "cvtsd2ss {$src2, $dst|$dst, $src2}",
1413 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1415 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1416 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1417 "cvtsd2ss {$src2, $dst|$dst, $src2}",
1418 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1419 (load addr:$src2)))]>;
1420 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1421 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1422 "cvtss2sd {$src2, $dst|$dst, $src2}",
1423 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1424 VR128:$src2))]>, XS,
1425 Requires<[HasSSE2]>;
1426 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1427 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1428 "cvtss2sd {$src2, $dst|$dst, $src2}",
1429 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1430 (load addr:$src2)))]>, XS,
1431 Requires<[HasSSE2]>;
1436 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1438 /// In addition, we also have a special variant of the scalar form here to
1439 /// represent the associated intrinsic operation. This form is unlike the
1440 /// plain scalar form, in that it takes an entire vector (instead of a
1441 /// scalar) and leaves the top elements undefined.
1443 /// And, we have a special variant form for a full-vector intrinsic form.
1445 /// These four forms can each have a reg or a mem operand, so there are a
1446 /// total of eight "instructions".
1448 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1452 bit Commutable = 0> {
1453 // Scalar operation, reg.
1454 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1455 !strconcat(OpcodeStr, "sd {$src, $dst|$dst, $src}"),
1456 [(set FR64:$dst, (OpNode FR64:$src))]> {
1457 let isCommutable = Commutable;
1460 // Scalar operation, mem.
1461 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1462 !strconcat(OpcodeStr, "sd {$src, $dst|$dst, $src}"),
1463 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1465 // Vector operation, reg.
1466 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1467 !strconcat(OpcodeStr, "pd {$src, $dst|$dst, $src}"),
1468 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1469 let isCommutable = Commutable;
1472 // Vector operation, mem.
1473 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1474 !strconcat(OpcodeStr, "pd {$src, $dst|$dst, $src}"),
1475 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1477 // Intrinsic operation, reg.
1478 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1479 !strconcat(OpcodeStr, "sd {$src, $dst|$dst, $src}"),
1480 [(set VR128:$dst, (F64Int VR128:$src))]> {
1481 let isCommutable = Commutable;
1484 // Intrinsic operation, mem.
1485 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1486 !strconcat(OpcodeStr, "sd {$src, $dst|$dst, $src}"),
1487 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1489 // Vector intrinsic operation, reg
1490 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1491 !strconcat(OpcodeStr, "pd {$src, $dst|$dst, $src}"),
1492 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1493 let isCommutable = Commutable;
1496 // Vector intrinsic operation, mem
1497 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1498 !strconcat(OpcodeStr, "pd {$src, $dst|$dst, $src}"),
1499 [(set VR128:$dst, (V2F64Int (load addr:$src)))]>;
1503 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1504 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1506 // There is no f64 version of the reciprocal approximation instructions.
1509 let isTwoAddress = 1 in {
1510 let isCommutable = 1 in {
1511 def ANDPDrr : PDI<0x54, MRMSrcReg,
1512 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1513 "andpd {$src2, $dst|$dst, $src2}",
1515 (and (bc_v2i64 (v2f64 VR128:$src1)),
1516 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1517 def ORPDrr : PDI<0x56, MRMSrcReg,
1518 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1519 "orpd {$src2, $dst|$dst, $src2}",
1521 (or (bc_v2i64 (v2f64 VR128:$src1)),
1522 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1523 def XORPDrr : PDI<0x57, MRMSrcReg,
1524 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1525 "xorpd {$src2, $dst|$dst, $src2}",
1527 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1528 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1531 def ANDPDrm : PDI<0x54, MRMSrcMem,
1532 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1533 "andpd {$src2, $dst|$dst, $src2}",
1535 (and (bc_v2i64 (v2f64 VR128:$src1)),
1536 (memopv2i64 addr:$src2)))]>;
1537 def ORPDrm : PDI<0x56, MRMSrcMem,
1538 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1539 "orpd {$src2, $dst|$dst, $src2}",
1541 (or (bc_v2i64 (v2f64 VR128:$src1)),
1542 (memopv2i64 addr:$src2)))]>;
1543 def XORPDrm : PDI<0x57, MRMSrcMem,
1544 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1545 "xorpd {$src2, $dst|$dst, $src2}",
1547 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1548 (memopv2i64 addr:$src2)))]>;
1549 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1550 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1551 "andnpd {$src2, $dst|$dst, $src2}",
1553 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1554 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1555 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1556 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1557 "andnpd {$src2, $dst|$dst, $src2}",
1559 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1560 (memopv2i64 addr:$src2)))]>;
1563 let isTwoAddress = 1 in {
1564 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1565 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1566 "cmp${cc}pd {$src, $dst|$dst, $src}",
1567 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1568 VR128:$src, imm:$cc))]>;
1569 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1570 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1571 "cmp${cc}pd {$src, $dst|$dst, $src}",
1572 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1573 (load addr:$src), imm:$cc))]>;
1576 // Shuffle and unpack instructions
1577 let isTwoAddress = 1 in {
1578 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1579 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1580 "shufpd {$src3, $src2, $dst|$dst, $src2, $src3}",
1581 [(set VR128:$dst, (v2f64 (vector_shuffle
1582 VR128:$src1, VR128:$src2,
1583 SHUFP_shuffle_mask:$src3)))]>;
1584 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1585 (outs VR128:$dst), (ins VR128:$src1,
1586 f128mem:$src2, i8imm:$src3),
1587 "shufpd {$src3, $src2, $dst|$dst, $src2, $src3}",
1589 (v2f64 (vector_shuffle
1590 VR128:$src1, (load addr:$src2),
1591 SHUFP_shuffle_mask:$src3)))]>;
1593 let AddedComplexity = 10 in {
1594 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1595 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1596 "unpckhpd {$src2, $dst|$dst, $src2}",
1598 (v2f64 (vector_shuffle
1599 VR128:$src1, VR128:$src2,
1600 UNPCKH_shuffle_mask)))]>;
1601 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1602 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1603 "unpckhpd {$src2, $dst|$dst, $src2}",
1605 (v2f64 (vector_shuffle
1606 VR128:$src1, (load addr:$src2),
1607 UNPCKH_shuffle_mask)))]>;
1609 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1610 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1611 "unpcklpd {$src2, $dst|$dst, $src2}",
1613 (v2f64 (vector_shuffle
1614 VR128:$src1, VR128:$src2,
1615 UNPCKL_shuffle_mask)))]>;
1616 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1617 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1618 "unpcklpd {$src2, $dst|$dst, $src2}",
1620 (v2f64 (vector_shuffle
1621 VR128:$src1, (load addr:$src2),
1622 UNPCKL_shuffle_mask)))]>;
1623 } // AddedComplexity
1627 //===----------------------------------------------------------------------===//
1628 // SSE integer instructions
1630 // Move Instructions
1631 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1632 "movdqa {$src, $dst|$dst, $src}", []>;
1633 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1634 "movdqa {$src, $dst|$dst, $src}",
1635 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1636 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1637 "movdqa {$src, $dst|$dst, $src}",
1638 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1639 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1640 "movdqu {$src, $dst|$dst, $src}",
1641 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1642 XS, Requires<[HasSSE2]>;
1643 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1644 "movdqu {$src, $dst|$dst, $src}",
1645 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1646 XS, Requires<[HasSSE2]>;
1648 // Intrinsic forms of MOVDQU load and store
1649 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1650 "movdqu {$src, $dst|$dst, $src}",
1651 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1652 XS, Requires<[HasSSE2]>;
1653 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1654 "movdqu {$src, $dst|$dst, $src}",
1655 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1656 XS, Requires<[HasSSE2]>;
1658 let isTwoAddress = 1 in {
1660 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1661 bit Commutable = 0> {
1662 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1663 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1664 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1665 let isCommutable = Commutable;
1667 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1668 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1669 [(set VR128:$dst, (IntId VR128:$src1,
1670 (bitconvert (memopv2i64 addr:$src2))))]>;
1673 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1674 string OpcodeStr, Intrinsic IntId> {
1675 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1676 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1677 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1678 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1679 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1680 [(set VR128:$dst, (IntId VR128:$src1,
1681 (bitconvert (memopv2i64 addr:$src2))))]>;
1682 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1683 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1684 [(set VR128:$dst, (IntId VR128:$src1,
1685 (scalar_to_vector (i32 imm:$src2))))]>;
1689 /// PDI_binop_rm - Simple SSE2 binary operator.
1690 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1691 ValueType OpVT, bit Commutable = 0> {
1692 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1693 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1694 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1695 let isCommutable = Commutable;
1697 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1698 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1699 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1700 (bitconvert (memopv2i64 addr:$src2)))))]>;
1703 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1705 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1706 /// to collapse (bitconvert VT to VT) into its operand.
1708 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1709 bit Commutable = 0> {
1710 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1711 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1712 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1713 let isCommutable = Commutable;
1715 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1716 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1717 [(set VR128:$dst, (OpNode VR128:$src1,(memopv2i64 addr:$src2)))]>;
1722 // 128-bit Integer Arithmetic
1724 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1725 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1726 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1727 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1729 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1730 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1731 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1732 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1734 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1735 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1736 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1737 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1739 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1740 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1741 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1742 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1744 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1746 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1747 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1748 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1750 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1752 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1753 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1756 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1757 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1758 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1759 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1760 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1763 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw", int_x86_sse2_psll_w>;
1764 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld", int_x86_sse2_psll_d>;
1765 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq", int_x86_sse2_psll_q>;
1767 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw", int_x86_sse2_psrl_w>;
1768 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld", int_x86_sse2_psrl_d>;
1769 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq", int_x86_sse2_psrl_q>;
1771 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw", int_x86_sse2_psra_w>;
1772 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad", int_x86_sse2_psra_d>;
1773 // PSRAQ doesn't exist in SSE[1-3].
1775 // 128-bit logical shifts.
1776 let isTwoAddress = 1 in {
1777 def PSLLDQri : PDIi8<0x73, MRM7r,
1778 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1779 "pslldq {$src2, $dst|$dst, $src2}", []>;
1780 def PSRLDQri : PDIi8<0x73, MRM3r,
1781 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1782 "psrldq {$src2, $dst|$dst, $src2}", []>;
1783 // PSRADQri doesn't exist in SSE[1-3].
1786 let Predicates = [HasSSE2] in {
1787 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1788 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1789 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1790 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1791 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1792 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1796 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1797 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1798 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1800 let isTwoAddress = 1 in {
1801 def PANDNrr : PDI<0xDF, MRMSrcReg,
1802 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1803 "pandn {$src2, $dst|$dst, $src2}",
1804 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1807 def PANDNrm : PDI<0xDF, MRMSrcMem,
1808 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1809 "pandn {$src2, $dst|$dst, $src2}",
1810 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1811 (load addr:$src2))))]>;
1814 // SSE2 Integer comparison
1815 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1816 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1817 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1818 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1819 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1820 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1822 // Pack instructions
1823 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1824 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1825 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1827 // Shuffle and unpack instructions
1828 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1829 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1830 "pshufd {$src2, $src1, $dst|$dst, $src1, $src2}",
1831 [(set VR128:$dst, (v4i32 (vector_shuffle
1832 VR128:$src1, (undef),
1833 PSHUFD_shuffle_mask:$src2)))]>;
1834 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1835 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1836 "pshufd {$src2, $src1, $dst|$dst, $src1, $src2}",
1837 [(set VR128:$dst, (v4i32 (vector_shuffle
1838 (bc_v4i32(memopv2i64 addr:$src1)),
1840 PSHUFD_shuffle_mask:$src2)))]>;
1842 // SSE2 with ImmT == Imm8 and XS prefix.
1843 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1844 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1845 "pshufhw {$src2, $src1, $dst|$dst, $src1, $src2}",
1846 [(set VR128:$dst, (v8i16 (vector_shuffle
1847 VR128:$src1, (undef),
1848 PSHUFHW_shuffle_mask:$src2)))]>,
1849 XS, Requires<[HasSSE2]>;
1850 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
1851 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1852 "pshufhw {$src2, $src1, $dst|$dst, $src1, $src2}",
1853 [(set VR128:$dst, (v8i16 (vector_shuffle
1854 (bc_v8i16 (memopv2i64 addr:$src1)),
1856 PSHUFHW_shuffle_mask:$src2)))]>,
1857 XS, Requires<[HasSSE2]>;
1859 // SSE2 with ImmT == Imm8 and XD prefix.
1860 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
1861 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1862 "pshuflw {$src2, $src1, $dst|$dst, $src1, $src2}",
1863 [(set VR128:$dst, (v8i16 (vector_shuffle
1864 VR128:$src1, (undef),
1865 PSHUFLW_shuffle_mask:$src2)))]>,
1866 XD, Requires<[HasSSE2]>;
1867 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
1868 (outs VR128:$dst), (ins i128mem:$src1, i32i8imm:$src2),
1869 "pshuflw {$src2, $src1, $dst|$dst, $src1, $src2}",
1870 [(set VR128:$dst, (v8i16 (vector_shuffle
1871 (bc_v8i16 (memopv2i64 addr:$src1)),
1873 PSHUFLW_shuffle_mask:$src2)))]>,
1874 XD, Requires<[HasSSE2]>;
1877 let isTwoAddress = 1 in {
1878 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
1879 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1880 "punpcklbw {$src2, $dst|$dst, $src2}",
1882 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1883 UNPCKL_shuffle_mask)))]>;
1884 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
1885 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1886 "punpcklbw {$src2, $dst|$dst, $src2}",
1888 (v16i8 (vector_shuffle VR128:$src1,
1889 (bc_v16i8 (memopv2i64 addr:$src2)),
1890 UNPCKL_shuffle_mask)))]>;
1891 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
1892 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1893 "punpcklwd {$src2, $dst|$dst, $src2}",
1895 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1896 UNPCKL_shuffle_mask)))]>;
1897 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
1898 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1899 "punpcklwd {$src2, $dst|$dst, $src2}",
1901 (v8i16 (vector_shuffle VR128:$src1,
1902 (bc_v8i16 (memopv2i64 addr:$src2)),
1903 UNPCKL_shuffle_mask)))]>;
1904 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
1905 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1906 "punpckldq {$src2, $dst|$dst, $src2}",
1908 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1909 UNPCKL_shuffle_mask)))]>;
1910 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
1911 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1912 "punpckldq {$src2, $dst|$dst, $src2}",
1914 (v4i32 (vector_shuffle VR128:$src1,
1915 (bc_v4i32 (memopv2i64 addr:$src2)),
1916 UNPCKL_shuffle_mask)))]>;
1917 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
1918 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1919 "punpcklqdq {$src2, $dst|$dst, $src2}",
1921 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1922 UNPCKL_shuffle_mask)))]>;
1923 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
1924 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1925 "punpcklqdq {$src2, $dst|$dst, $src2}",
1927 (v2i64 (vector_shuffle VR128:$src1,
1928 (memopv2i64 addr:$src2),
1929 UNPCKL_shuffle_mask)))]>;
1931 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
1932 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1933 "punpckhbw {$src2, $dst|$dst, $src2}",
1935 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1936 UNPCKH_shuffle_mask)))]>;
1937 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
1938 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1939 "punpckhbw {$src2, $dst|$dst, $src2}",
1941 (v16i8 (vector_shuffle VR128:$src1,
1942 (bc_v16i8 (memopv2i64 addr:$src2)),
1943 UNPCKH_shuffle_mask)))]>;
1944 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
1945 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1946 "punpckhwd {$src2, $dst|$dst, $src2}",
1948 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1949 UNPCKH_shuffle_mask)))]>;
1950 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
1951 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1952 "punpckhwd {$src2, $dst|$dst, $src2}",
1954 (v8i16 (vector_shuffle VR128:$src1,
1955 (bc_v8i16 (memopv2i64 addr:$src2)),
1956 UNPCKH_shuffle_mask)))]>;
1957 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
1958 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1959 "punpckhdq {$src2, $dst|$dst, $src2}",
1961 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1962 UNPCKH_shuffle_mask)))]>;
1963 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
1964 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1965 "punpckhdq {$src2, $dst|$dst, $src2}",
1967 (v4i32 (vector_shuffle VR128:$src1,
1968 (bc_v4i32 (memopv2i64 addr:$src2)),
1969 UNPCKH_shuffle_mask)))]>;
1970 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
1971 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1972 "punpckhqdq {$src2, $dst|$dst, $src2}",
1974 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1975 UNPCKH_shuffle_mask)))]>;
1976 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
1977 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1978 "punpckhqdq {$src2, $dst|$dst, $src2}",
1980 (v2i64 (vector_shuffle VR128:$src1,
1981 (memopv2i64 addr:$src2),
1982 UNPCKH_shuffle_mask)))]>;
1986 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
1987 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
1988 "pextrw {$src2, $src1, $dst|$dst, $src1, $src2}",
1989 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
1990 (iPTR imm:$src2)))]>;
1991 let isTwoAddress = 1 in {
1992 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
1993 (outs VR128:$dst), (ins VR128:$src1,
1994 GR32:$src2, i32i8imm:$src3),
1995 "pinsrw {$src3, $src2, $dst|$dst, $src2, $src3}",
1997 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
1998 GR32:$src2, (iPTR imm:$src3))))]>;
1999 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2000 (outs VR128:$dst), (ins VR128:$src1,
2001 i16mem:$src2, i32i8imm:$src3),
2002 "pinsrw {$src3, $src2, $dst|$dst, $src2, $src3}",
2004 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
2005 (i32 (anyext (loadi16 addr:$src2))),
2006 (iPTR imm:$src3))))]>;
2010 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2011 "pmovmskb {$src, $dst|$dst, $src}",
2012 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2014 // Conditional store
2015 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2016 "maskmovdqu {$mask, $src|$src, $mask}",
2017 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>,
2020 // Non-temporal stores
2021 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2022 "movntpd {$src, $dst|$dst, $src}",
2023 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2024 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2025 "movntdq {$src, $dst|$dst, $src}",
2026 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2027 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2028 "movnti {$src, $dst|$dst, $src}",
2029 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2030 TB, Requires<[HasSSE2]>;
2033 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2034 "clflush $src", [(int_x86_sse2_clflush addr:$src)]>,
2035 TB, Requires<[HasSSE2]>;
2037 // Load, store, and memory fence
2038 def LFENCE : I<0xAE, MRM5m, (outs), (ins),
2039 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2040 def MFENCE : I<0xAE, MRM6m, (outs), (ins),
2041 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2044 // Alias instructions that map zero vector to pxor / xorp* for sse.
2045 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
2046 let isReMaterializable = 1 in
2047 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2048 "pcmpeqd $dst, $dst",
2049 [(set VR128:$dst, (v2f64 immAllOnesV))]>;
2051 // FR64 to 128-bit vector conversion.
2052 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2053 "movsd {$src, $dst|$dst, $src}",
2055 (v2f64 (scalar_to_vector FR64:$src)))]>;
2056 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2057 "movsd {$src, $dst|$dst, $src}",
2059 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2061 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2062 "movd {$src, $dst|$dst, $src}",
2064 (v4i32 (scalar_to_vector GR32:$src)))]>;
2065 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2066 "movd {$src, $dst|$dst, $src}",
2068 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2070 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2071 "movd {$src, $dst|$dst, $src}",
2072 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2074 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2075 "movd {$src, $dst|$dst, $src}",
2076 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2078 // SSE2 instructions with XS prefix
2079 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2080 "movq {$src, $dst|$dst, $src}",
2082 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2083 Requires<[HasSSE2]>;
2084 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2085 "movq {$src, $dst|$dst, $src}",
2086 [(store (i64 (vector_extract (v2i64 VR128:$src),
2087 (iPTR 0))), addr:$dst)]>;
2089 // FIXME: may not be able to eliminate this movss with coalescing the src and
2090 // dest register classes are different. We really want to write this pattern
2092 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2093 // (f32 FR32:$src)>;
2094 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2095 "movsd {$src, $dst|$dst, $src}",
2096 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2098 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2099 "movsd {$src, $dst|$dst, $src}",
2100 [(store (f64 (vector_extract (v2f64 VR128:$src),
2101 (iPTR 0))), addr:$dst)]>;
2102 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2103 "movd {$src, $dst|$dst, $src}",
2104 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2106 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2107 "movd {$src, $dst|$dst, $src}",
2108 [(store (i32 (vector_extract (v4i32 VR128:$src),
2109 (iPTR 0))), addr:$dst)]>;
2111 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2112 "movd {$src, $dst|$dst, $src}",
2113 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2114 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2115 "movd {$src, $dst|$dst, $src}",
2116 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2119 // Move to lower bits of a VR128, leaving upper bits alone.
2120 // Three operand (but two address) aliases.
2121 let isTwoAddress = 1 in {
2122 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2123 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2124 "movsd {$src2, $dst|$dst, $src2}", []>;
2126 let AddedComplexity = 15 in
2127 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2128 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2129 "movsd {$src2, $dst|$dst, $src2}",
2131 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
2132 MOVL_shuffle_mask)))]>;
2135 // Store / copy lower 64-bits of a XMM register.
2136 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2137 "movq {$src, $dst|$dst, $src}",
2138 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2140 // Move to lower bits of a VR128 and zeroing upper bits.
2141 // Loading from memory automatically zeroing upper bits.
2142 let AddedComplexity = 20 in
2143 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2144 "movsd {$src, $dst|$dst, $src}",
2146 (v2f64 (vector_shuffle immAllZerosV,
2147 (v2f64 (scalar_to_vector
2148 (loadf64 addr:$src))),
2149 MOVL_shuffle_mask)))]>;
2151 let AddedComplexity = 15 in
2152 // movd / movq to XMM register zero-extends
2153 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2154 "movd {$src, $dst|$dst, $src}",
2156 (v4i32 (vector_shuffle immAllZerosV,
2157 (v4i32 (scalar_to_vector GR32:$src)),
2158 MOVL_shuffle_mask)))]>;
2159 let AddedComplexity = 20 in
2160 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2161 "movd {$src, $dst|$dst, $src}",
2163 (v4i32 (vector_shuffle immAllZerosV,
2164 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
2165 MOVL_shuffle_mask)))]>;
2167 // Moving from XMM to XMM but still clear upper 64 bits.
2168 let AddedComplexity = 15 in
2169 def MOVZQI2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2170 "movq {$src, $dst|$dst, $src}",
2171 [(set VR128:$dst, (int_x86_sse2_movl_dq VR128:$src))]>,
2172 XS, Requires<[HasSSE2]>;
2173 let AddedComplexity = 20 in
2174 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2175 "movq {$src, $dst|$dst, $src}",
2176 [(set VR128:$dst, (int_x86_sse2_movl_dq
2177 (bitconvert (memopv2i64 addr:$src))))]>,
2178 XS, Requires<[HasSSE2]>;
2181 //===----------------------------------------------------------------------===//
2182 // SSE3 Instructions
2183 //===----------------------------------------------------------------------===//
2185 // SSE3 Instruction Templates:
2187 // S3I - SSE3 instructions with TB and OpSize prefixes.
2188 // S3SI - SSE3 instructions with XS prefix.
2189 // S3DI - SSE3 instructions with XD prefix.
2191 class S3SI<bits<8> o, Format F, dag outs, dag ins, string asm, list<dag> pattern>
2192 : I<o, F, outs, ins, asm, pattern>, XS, Requires<[HasSSE3]>;
2193 class S3DI<bits<8> o, Format F, dag outs, dag ins, string asm, list<dag> pattern>
2194 : I<o, F, outs, ins, asm, pattern>, XD, Requires<[HasSSE3]>;
2195 class S3I<bits<8> o, Format F, dag outs, dag ins, string asm, list<dag> pattern>
2196 : I<o, F, outs, ins, asm, pattern>, TB, OpSize, Requires<[HasSSE3]>;
2198 // Move Instructions
2199 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2200 "movshdup {$src, $dst|$dst, $src}",
2201 [(set VR128:$dst, (v4f32 (vector_shuffle
2202 VR128:$src, (undef),
2203 MOVSHDUP_shuffle_mask)))]>;
2204 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2205 "movshdup {$src, $dst|$dst, $src}",
2206 [(set VR128:$dst, (v4f32 (vector_shuffle
2207 (memopv4f32 addr:$src), (undef),
2208 MOVSHDUP_shuffle_mask)))]>;
2210 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2211 "movsldup {$src, $dst|$dst, $src}",
2212 [(set VR128:$dst, (v4f32 (vector_shuffle
2213 VR128:$src, (undef),
2214 MOVSLDUP_shuffle_mask)))]>;
2215 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2216 "movsldup {$src, $dst|$dst, $src}",
2217 [(set VR128:$dst, (v4f32 (vector_shuffle
2218 (memopv4f32 addr:$src), (undef),
2219 MOVSLDUP_shuffle_mask)))]>;
2221 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2222 "movddup {$src, $dst|$dst, $src}",
2223 [(set VR128:$dst, (v2f64 (vector_shuffle
2224 VR128:$src, (undef),
2225 SSE_splat_lo_mask)))]>;
2226 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2227 "movddup {$src, $dst|$dst, $src}",
2229 (v2f64 (vector_shuffle
2230 (scalar_to_vector (loadf64 addr:$src)),
2232 SSE_splat_lo_mask)))]>;
2235 let isTwoAddress = 1 in {
2236 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2237 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2238 "addsubps {$src2, $dst|$dst, $src2}",
2239 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2241 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2242 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2243 "addsubps {$src2, $dst|$dst, $src2}",
2244 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2245 (load addr:$src2)))]>;
2246 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2247 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2248 "addsubpd {$src2, $dst|$dst, $src2}",
2249 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2251 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2252 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2253 "addsubpd {$src2, $dst|$dst, $src2}",
2254 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2255 (load addr:$src2)))]>;
2258 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2259 "lddqu {$src, $dst|$dst, $src}",
2260 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2263 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2264 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2265 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2266 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2267 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2268 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2269 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2270 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
2271 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2272 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2273 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2274 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2275 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2276 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2277 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2278 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
2280 let isTwoAddress = 1 in {
2281 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2282 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2283 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2284 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2285 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2286 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2287 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2288 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2291 // Thread synchronization
2292 def MONITOR : I<0xC8, RawFrm, (outs), (ins), "monitor",
2293 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2294 def MWAIT : I<0xC9, RawFrm, (outs), (ins), "mwait",
2295 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2297 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2298 let AddedComplexity = 15 in
2299 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2300 MOVSHDUP_shuffle_mask)),
2301 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2302 let AddedComplexity = 20 in
2303 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2304 MOVSHDUP_shuffle_mask)),
2305 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2307 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2308 let AddedComplexity = 15 in
2309 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2310 MOVSLDUP_shuffle_mask)),
2311 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2312 let AddedComplexity = 20 in
2313 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2314 MOVSLDUP_shuffle_mask)),
2315 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2317 //===----------------------------------------------------------------------===//
2318 // SSSE3 Instructions
2319 //===----------------------------------------------------------------------===//
2321 // SSE3 Instruction Templates:
2323 // SS38I - SSSE3 instructions with T8 and OpSize prefixes.
2324 // SS3AI - SSSE3 instructions with TA and OpSize prefixes.
2326 class SS38I<bits<8> o, Format F, dag outs, dag ins, string asm,
2328 : I<o, F, outs, ins, asm, pattern>, T8, OpSize, Requires<[HasSSSE3]>;
2329 class SS3AI<bits<8> o, Format F, dag outs, dag ins, string asm,
2331 : I<o, F, outs, ins, asm, pattern>, TA, OpSize, Requires<[HasSSSE3]>;
2333 /// SS3I_binop_rm_int - Simple SSSE3 binary operatr whose type is v2i64.
2334 let isTwoAddress = 1 in {
2335 multiclass SS3I_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
2336 bit Commutable = 0> {
2337 def rr : SS38I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2338 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2339 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
2340 let isCommutable = Commutable;
2342 def rm : SS38I<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2343 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2346 (bitconvert (memopv2i64 addr:$src2))))]>;
2350 defm PMULHRSW128 : SS3I_binop_rm_int<0x0B, "pmulhrsw",
2351 int_x86_ssse3_pmulhrsw_128, 1>;
2353 //===----------------------------------------------------------------------===//
2354 // Non-Instruction Patterns
2355 //===----------------------------------------------------------------------===//
2357 // 128-bit vector undef's.
2358 def : Pat<(v2f64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2359 def : Pat<(v16i8 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2360 def : Pat<(v8i16 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2361 def : Pat<(v4i32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2362 def : Pat<(v2i64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2364 // 128-bit vector all zero's.
2365 def : Pat<(v16i8 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2366 def : Pat<(v8i16 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2367 def : Pat<(v4i32 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2368 def : Pat<(v2i64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2369 def : Pat<(v2f64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2371 // 128-bit vector all one's.
2372 def : Pat<(v16i8 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2373 def : Pat<(v8i16 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2374 def : Pat<(v4i32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2375 def : Pat<(v2i64 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2376 def : Pat<(v4f32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE1]>;
2379 // Scalar to v8i16 / v16i8. The source may be a GR32, but only the lower 8 or
2381 def : Pat<(v8i16 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2382 Requires<[HasSSE2]>;
2383 def : Pat<(v16i8 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2384 Requires<[HasSSE2]>;
2387 let Predicates = [HasSSE2] in {
2388 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2389 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2390 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2391 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2392 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2393 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2394 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2395 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2396 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2397 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2398 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2399 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2400 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2401 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2402 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2403 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2404 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2405 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2406 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2407 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2408 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2409 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2410 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2411 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2412 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2413 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2414 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2415 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2416 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2417 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2420 // Move scalar to XMM zero-extended
2421 // movd to XMM register zero-extends
2422 let AddedComplexity = 15 in {
2423 def : Pat<(v8i16 (vector_shuffle immAllZerosV,
2424 (v8i16 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
2425 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
2426 def : Pat<(v16i8 (vector_shuffle immAllZerosV,
2427 (v16i8 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
2428 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
2429 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2430 def : Pat<(v2f64 (vector_shuffle immAllZerosV,
2431 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
2432 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2433 def : Pat<(v4f32 (vector_shuffle immAllZerosV,
2434 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
2435 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
2438 // Splat v2f64 / v2i64
2439 let AddedComplexity = 10 in {
2440 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2441 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2442 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2443 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2444 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2445 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2446 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2447 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2451 def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SSE_splat_mask:$sm),
2452 (SHUFPSrri VR128:$src, VR128:$src, SSE_splat_mask:$sm)>,
2453 Requires<[HasSSE1]>;
2455 // Special unary SHUFPSrri case.
2456 // FIXME: when we want non two-address code, then we should use PSHUFD?
2457 def : Pat<(vector_shuffle (v4f32 VR128:$src1), (undef),
2458 SHUFP_unary_shuffle_mask:$sm),
2459 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2460 Requires<[HasSSE1]>;
2461 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2462 def : Pat<(vector_shuffle (memopv4f32 addr:$src1), (undef),
2463 SHUFP_unary_shuffle_mask:$sm),
2464 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2465 Requires<[HasSSE2]>;
2466 // Special binary v4i32 shuffle cases with SHUFPS.
2467 def : Pat<(vector_shuffle (v4i32 VR128:$src1), (v4i32 VR128:$src2),
2468 PSHUFD_binary_shuffle_mask:$sm),
2469 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2470 Requires<[HasSSE2]>;
2471 def : Pat<(vector_shuffle (v4i32 VR128:$src1),
2472 (bc_v4i32 (memopv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm),
2473 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2474 Requires<[HasSSE2]>;
2476 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2477 let AddedComplexity = 10 in {
2478 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2479 UNPCKL_v_undef_shuffle_mask)),
2480 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2481 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2482 UNPCKL_v_undef_shuffle_mask)),
2483 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2484 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2485 UNPCKL_v_undef_shuffle_mask)),
2486 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2487 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2488 UNPCKL_v_undef_shuffle_mask)),
2489 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2492 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2493 let AddedComplexity = 10 in {
2494 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2495 UNPCKH_v_undef_shuffle_mask)),
2496 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2497 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2498 UNPCKH_v_undef_shuffle_mask)),
2499 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2500 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2501 UNPCKH_v_undef_shuffle_mask)),
2502 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2503 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2504 UNPCKH_v_undef_shuffle_mask)),
2505 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2508 let AddedComplexity = 15 in {
2509 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2510 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2511 MOVHP_shuffle_mask)),
2512 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2514 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2515 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2516 MOVHLPS_shuffle_mask)),
2517 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2519 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2520 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2521 MOVHLPS_v_undef_shuffle_mask)),
2522 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2523 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2524 MOVHLPS_v_undef_shuffle_mask)),
2525 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2528 let AddedComplexity = 20 in {
2529 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2530 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2531 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2532 MOVLP_shuffle_mask)),
2533 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2534 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2535 MOVLP_shuffle_mask)),
2536 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2537 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2538 MOVHP_shuffle_mask)),
2539 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2540 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2541 MOVHP_shuffle_mask)),
2542 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2544 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2545 MOVLP_shuffle_mask)),
2546 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2547 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2548 MOVLP_shuffle_mask)),
2549 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2550 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2551 MOVHP_shuffle_mask)),
2552 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2553 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2554 MOVLP_shuffle_mask)),
2555 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2558 let AddedComplexity = 15 in {
2559 // Setting the lowest element in the vector.
2560 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2561 MOVL_shuffle_mask)),
2562 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2563 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2564 MOVL_shuffle_mask)),
2565 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2567 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2568 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2569 MOVLP_shuffle_mask)),
2570 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2571 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2572 MOVLP_shuffle_mask)),
2573 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2576 // Set lowest element and zero upper elements.
2577 let AddedComplexity = 20 in
2578 def : Pat<(bc_v2i64 (vector_shuffle immAllZerosV,
2579 (v2f64 (scalar_to_vector (loadf64 addr:$src))),
2580 MOVL_shuffle_mask)),
2581 (MOVZQI2PQIrm addr:$src)>, Requires<[HasSSE2]>;
2583 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2584 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2585 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2586 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2587 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2588 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2589 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2590 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2591 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2592 Requires<[HasSSE2]>;
2593 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2594 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2595 Requires<[HasSSE2]>;
2596 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2597 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2598 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2599 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2600 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2601 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2602 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2603 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2604 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
2605 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2606 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
2607 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2608 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
2609 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2610 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
2611 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2613 // Some special case pandn patterns.
2614 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2616 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2617 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2619 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2620 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2622 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2624 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2625 (load addr:$src2))),
2626 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2627 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2628 (load addr:$src2))),
2629 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2630 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2631 (load addr:$src2))),
2632 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2635 def : Pat<(v4f32 (X86loadu addr:$src)), (MOVUPSrm addr:$src)>,
2636 Requires<[HasSSE1]>;
2638 // Use movaps / movups for SSE integer load / store (one byte shorter).
2639 def : Pat<(alignedloadv2i64 addr:$src),
2640 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
2641 def : Pat<(loadv2i64 addr:$src),
2642 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
2644 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
2645 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2646 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
2647 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2648 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
2649 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2650 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
2651 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2652 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
2653 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2654 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
2655 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2656 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
2657 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2658 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
2659 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;