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 X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
35 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest,
36 [SDNPHasChain, SDNPOutFlag]>;
37 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest,
38 [SDNPHasChain, SDNPOutFlag]>;
39 def X86s2vec : SDNode<"X86ISD::S2VEC", SDTypeProfile<1, 1, []>, []>;
40 def X86pextrw : SDNode<"X86ISD::PEXTRW", SDTypeProfile<1, 2, []>, []>;
41 def X86pinsrw : SDNode<"X86ISD::PINSRW", SDTypeProfile<1, 3, []>, []>;
43 //===----------------------------------------------------------------------===//
44 // SSE 'Special' Instructions
45 //===----------------------------------------------------------------------===//
47 def IMPLICIT_DEF_VR128 : I<0, Pseudo, (ops VR128:$dst),
49 [(set VR128:$dst, (v4f32 (undef)))]>,
51 def IMPLICIT_DEF_FR32 : I<0, Pseudo, (ops FR32:$dst),
53 [(set FR32:$dst, (undef))]>, Requires<[HasSSE2]>;
54 def IMPLICIT_DEF_FR64 : I<0, Pseudo, (ops FR64:$dst),
56 [(set FR64:$dst, (undef))]>, Requires<[HasSSE2]>;
58 //===----------------------------------------------------------------------===//
59 // SSE Complex Patterns
60 //===----------------------------------------------------------------------===//
62 // These are 'extloads' from a scalar to the low element of a vector, zeroing
63 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
65 def sse_load_f32 : ComplexPattern<v4f32, 4, "SelectScalarSSELoad", [],
67 def sse_load_f64 : ComplexPattern<v2f64, 4, "SelectScalarSSELoad", [],
70 def ssmem : Operand<v4f32> {
71 let PrintMethod = "printf32mem";
72 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
74 def sdmem : Operand<v2f64> {
75 let PrintMethod = "printf64mem";
76 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
79 //===----------------------------------------------------------------------===//
80 // SSE pattern fragments
81 //===----------------------------------------------------------------------===//
83 def X86loadpf32 : PatFrag<(ops node:$ptr), (f32 (X86loadp node:$ptr))>;
84 def X86loadpf64 : PatFrag<(ops node:$ptr), (f64 (X86loadp node:$ptr))>;
86 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
87 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
88 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
89 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
91 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
92 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
93 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
94 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
95 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
96 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
98 def fp32imm0 : PatLeaf<(f32 fpimm), [{
99 return N->isExactlyValue(+0.0);
102 def PSxLDQ_imm : SDNodeXForm<imm, [{
103 // Transformation function: imm >> 3
104 return getI32Imm(N->getValue() >> 3);
107 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
109 def SHUFFLE_get_shuf_imm : SDNodeXForm<build_vector, [{
110 return getI8Imm(X86::getShuffleSHUFImmediate(N));
113 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
115 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<build_vector, [{
116 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
119 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
121 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<build_vector, [{
122 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
125 def SSE_splat_mask : PatLeaf<(build_vector), [{
126 return X86::isSplatMask(N);
127 }], SHUFFLE_get_shuf_imm>;
129 def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
130 return X86::isSplatLoMask(N);
133 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
134 return X86::isMOVHLPSMask(N);
137 def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
138 return X86::isMOVHLPS_v_undef_Mask(N);
141 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
142 return X86::isMOVHPMask(N);
145 def MOVLP_shuffle_mask : PatLeaf<(build_vector), [{
146 return X86::isMOVLPMask(N);
149 def MOVL_shuffle_mask : PatLeaf<(build_vector), [{
150 return X86::isMOVLMask(N);
153 def MOVSHDUP_shuffle_mask : PatLeaf<(build_vector), [{
154 return X86::isMOVSHDUPMask(N);
157 def MOVSLDUP_shuffle_mask : PatLeaf<(build_vector), [{
158 return X86::isMOVSLDUPMask(N);
161 def UNPCKL_shuffle_mask : PatLeaf<(build_vector), [{
162 return X86::isUNPCKLMask(N);
165 def UNPCKH_shuffle_mask : PatLeaf<(build_vector), [{
166 return X86::isUNPCKHMask(N);
169 def UNPCKL_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
170 return X86::isUNPCKL_v_undef_Mask(N);
173 def UNPCKH_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
174 return X86::isUNPCKH_v_undef_Mask(N);
177 def PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
178 return X86::isPSHUFDMask(N);
179 }], SHUFFLE_get_shuf_imm>;
181 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
182 return X86::isPSHUFHWMask(N);
183 }], SHUFFLE_get_pshufhw_imm>;
185 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
186 return X86::isPSHUFLWMask(N);
187 }], SHUFFLE_get_pshuflw_imm>;
189 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
190 return X86::isPSHUFDMask(N);
191 }], SHUFFLE_get_shuf_imm>;
193 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
194 return X86::isSHUFPMask(N);
195 }], SHUFFLE_get_shuf_imm>;
197 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
198 return X86::isSHUFPMask(N);
199 }], SHUFFLE_get_shuf_imm>;
201 //===----------------------------------------------------------------------===//
202 // SSE scalar FP Instructions
203 //===----------------------------------------------------------------------===//
205 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
206 // scheduler into a branch sequence.
207 let usesCustomDAGSchedInserter = 1 in { // Expanded by the scheduler.
208 def CMOV_FR32 : I<0, Pseudo,
209 (ops FR32:$dst, FR32:$t, FR32:$f, i8imm:$cond),
210 "#CMOV_FR32 PSEUDO!",
211 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond))]>;
212 def CMOV_FR64 : I<0, Pseudo,
213 (ops FR64:$dst, FR64:$t, FR64:$f, i8imm:$cond),
214 "#CMOV_FR64 PSEUDO!",
215 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond))]>;
216 def CMOV_V4F32 : I<0, Pseudo,
217 (ops VR128:$dst, VR128:$t, VR128:$f, i8imm:$cond),
218 "#CMOV_V4F32 PSEUDO!",
220 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
221 def CMOV_V2F64 : I<0, Pseudo,
222 (ops VR128:$dst, VR128:$t, VR128:$f, i8imm:$cond),
223 "#CMOV_V2F64 PSEUDO!",
225 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
226 def CMOV_V2I64 : I<0, Pseudo,
227 (ops VR128:$dst, VR128:$t, VR128:$f, i8imm:$cond),
228 "#CMOV_V2I64 PSEUDO!",
230 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
233 //===----------------------------------------------------------------------===//
235 //===----------------------------------------------------------------------===//
237 // SSE1 Instruction Templates:
239 // SSI - SSE1 instructions with XS prefix.
240 // PSI - SSE1 instructions with TB prefix.
241 // PSIi8 - SSE1 instructions with ImmT == Imm8 and TB prefix.
243 class SSI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
244 : I<o, F, ops, asm, pattern>, XS, Requires<[HasSSE1]>;
245 class PSI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
246 : I<o, F, ops, asm, pattern>, TB, Requires<[HasSSE1]>;
247 class PSIi8<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
248 : Ii8<o, F, ops, asm, pattern>, TB, Requires<[HasSSE1]>;
250 // Helpers for defining instructions that directly correspond to intrinsics.
251 multiclass SS_IntUnary<bits<8> o, string OpcodeStr, Intrinsic IntId> {
252 def r : SSI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
253 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
254 [(set VR128:$dst, (v4f32 (IntId VR128:$src)))]>;
255 def m : SSI<o, MRMSrcMem, (ops VR128:$dst, ssmem:$src),
256 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
257 [(set VR128:$dst, (v4f32 (IntId sse_load_f32:$src)))]>;
261 def MOVSSrr : SSI<0x10, MRMSrcReg, (ops FR32:$dst, FR32:$src),
262 "movss {$src, $dst|$dst, $src}", []>;
263 def MOVSSrm : SSI<0x10, MRMSrcMem, (ops FR32:$dst, f32mem:$src),
264 "movss {$src, $dst|$dst, $src}",
265 [(set FR32:$dst, (loadf32 addr:$src))]>;
266 def MOVSSmr : SSI<0x11, MRMDestMem, (ops f32mem:$dst, FR32:$src),
267 "movss {$src, $dst|$dst, $src}",
268 [(store FR32:$src, addr:$dst)]>;
270 def SQRTSSr : SSI<0x51, MRMSrcReg, (ops FR32:$dst, FR32:$src),
271 "sqrtss {$src, $dst|$dst, $src}",
272 [(set FR32:$dst, (fsqrt FR32:$src))]>;
273 def SQRTSSm : SSI<0x51, MRMSrcMem, (ops FR32:$dst, f32mem:$src),
274 "sqrtss {$src, $dst|$dst, $src}",
275 [(set FR32:$dst, (fsqrt (loadf32 addr:$src)))]>;
277 // Aliases to match intrinsics which expect XMM operand(s).
278 defm SQRTSS_Int : SS_IntUnary<0x51, "sqrtss" , int_x86_sse_sqrt_ss>;
279 defm RSQRTSS_Int : SS_IntUnary<0x52, "rsqrtss", int_x86_sse_rsqrt_ss>;
280 defm RCPSS_Int : SS_IntUnary<0x53, "rcpss" , int_x86_sse_rcp_ss>;
282 // Conversion instructions
283 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (ops GR32:$dst, FR32:$src),
284 "cvttss2si {$src, $dst|$dst, $src}",
285 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
286 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (ops GR32:$dst, f32mem:$src),
287 "cvttss2si {$src, $dst|$dst, $src}",
288 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
289 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (ops FR32:$dst, GR32:$src),
290 "cvtsi2ss {$src, $dst|$dst, $src}",
291 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
292 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (ops FR32:$dst, i32mem:$src),
293 "cvtsi2ss {$src, $dst|$dst, $src}",
294 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
296 // Match intrinsics which expect XMM operand(s).
297 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (ops GR32:$dst, VR128:$src),
298 "cvtss2si {$src, $dst|$dst, $src}",
299 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
300 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (ops GR32:$dst, f32mem:$src),
301 "cvtss2si {$src, $dst|$dst, $src}",
302 [(set GR32:$dst, (int_x86_sse_cvtss2si
303 (load addr:$src)))]>;
305 // Aliases for intrinsics
306 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (ops GR32:$dst, VR128:$src),
307 "cvttss2si {$src, $dst|$dst, $src}",
309 (int_x86_sse_cvttss2si VR128:$src))]>;
310 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (ops GR32:$dst, f32mem:$src),
311 "cvttss2si {$src, $dst|$dst, $src}",
313 (int_x86_sse_cvttss2si(load addr:$src)))]>;
315 let isTwoAddress = 1 in {
316 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
317 (ops VR128:$dst, VR128:$src1, GR32:$src2),
318 "cvtsi2ss {$src2, $dst|$dst, $src2}",
319 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
321 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
322 (ops VR128:$dst, VR128:$src1, i32mem:$src2),
323 "cvtsi2ss {$src2, $dst|$dst, $src2}",
324 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
325 (loadi32 addr:$src2)))]>;
328 // Comparison instructions
329 let isTwoAddress = 1 in {
330 def CMPSSrr : SSI<0xC2, MRMSrcReg,
331 (ops FR32:$dst, FR32:$src1, FR32:$src, SSECC:$cc),
332 "cmp${cc}ss {$src, $dst|$dst, $src}",
334 def CMPSSrm : SSI<0xC2, MRMSrcMem,
335 (ops FR32:$dst, FR32:$src1, f32mem:$src, SSECC:$cc),
336 "cmp${cc}ss {$src, $dst|$dst, $src}", []>;
339 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (ops FR32:$src1, FR32:$src2),
340 "ucomiss {$src2, $src1|$src1, $src2}",
341 [(X86cmp FR32:$src1, FR32:$src2)]>;
342 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (ops FR32:$src1, f32mem:$src2),
343 "ucomiss {$src2, $src1|$src1, $src2}",
344 [(X86cmp FR32:$src1, (loadf32 addr:$src2))]>;
346 // Aliases to match intrinsics which expect XMM operand(s).
347 let isTwoAddress = 1 in {
348 def Int_CMPSSrr : SSI<0xC2, MRMSrcReg,
349 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
350 "cmp${cc}ss {$src, $dst|$dst, $src}",
351 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
352 VR128:$src, imm:$cc))]>;
353 def Int_CMPSSrm : SSI<0xC2, MRMSrcMem,
354 (ops VR128:$dst, VR128:$src1, f32mem:$src, SSECC:$cc),
355 "cmp${cc}ss {$src, $dst|$dst, $src}",
356 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
357 (load addr:$src), imm:$cc))]>;
360 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
361 "ucomiss {$src2, $src1|$src1, $src2}",
362 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2)]>;
363 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
364 "ucomiss {$src2, $src1|$src1, $src2}",
365 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2))]>;
367 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
368 "comiss {$src2, $src1|$src1, $src2}",
369 [(X86comi (v4f32 VR128:$src1), VR128:$src2)]>;
370 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
371 "comiss {$src2, $src1|$src1, $src2}",
372 [(X86comi (v4f32 VR128:$src1), (load addr:$src2))]>;
374 // Aliases of packed SSE1 instructions for scalar use. These all have names that
377 // Alias instructions that map fld0 to pxor for sse.
378 def FsFLD0SS : I<0xEF, MRMInitReg, (ops FR32:$dst),
379 "pxor $dst, $dst", [(set FR32:$dst, fp32imm0)]>,
380 Requires<[HasSSE1]>, TB, OpSize;
382 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
384 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (ops FR32:$dst, FR32:$src),
385 "movaps {$src, $dst|$dst, $src}", []>;
387 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
389 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (ops FR32:$dst, f128mem:$src),
390 "movaps {$src, $dst|$dst, $src}",
391 [(set FR32:$dst, (X86loadpf32 addr:$src))]>;
393 // Alias bitwise logical operations using SSE logical ops on packed FP values.
394 let isTwoAddress = 1 in {
395 let isCommutable = 1 in {
396 def FsANDPSrr : PSI<0x54, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
397 "andps {$src2, $dst|$dst, $src2}",
398 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
399 def FsORPSrr : PSI<0x56, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
400 "orps {$src2, $dst|$dst, $src2}",
401 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
402 def FsXORPSrr : PSI<0x57, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
403 "xorps {$src2, $dst|$dst, $src2}",
404 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
407 def FsANDPSrm : PSI<0x54, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
408 "andps {$src2, $dst|$dst, $src2}",
409 [(set FR32:$dst, (X86fand FR32:$src1,
410 (X86loadpf32 addr:$src2)))]>;
411 def FsORPSrm : PSI<0x56, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
412 "orps {$src2, $dst|$dst, $src2}",
413 [(set FR32:$dst, (X86for FR32:$src1,
414 (X86loadpf32 addr:$src2)))]>;
415 def FsXORPSrm : PSI<0x57, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
416 "xorps {$src2, $dst|$dst, $src2}",
417 [(set FR32:$dst, (X86fxor FR32:$src1,
418 (X86loadpf32 addr:$src2)))]>;
420 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
421 (ops FR32:$dst, FR32:$src1, FR32:$src2),
422 "andnps {$src2, $dst|$dst, $src2}", []>;
423 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
424 (ops FR32:$dst, FR32:$src1, f128mem:$src2),
425 "andnps {$src2, $dst|$dst, $src2}", []>;
428 /// scalar_sse1_fp_binop_rm - Scalar SSE1 binops come in three basic forms:
430 /// 1. f32 - This comes in SSE1 form for floats.
431 /// 2. rr vs rm - They include a reg+reg form and a reg+mem form.
433 /// In addition, scalar SSE ops have an intrinsic form. This form is unlike the
434 /// normal form, in that they take an entire vector (instead of a scalar) and
435 /// leave the top elements undefined. This adds another two variants of the
436 /// above permutations, giving us 8 forms for 'instruction'.
438 let isTwoAddress = 1 in {
439 multiclass scalar_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
440 SDNode OpNode, Intrinsic F32Int,
441 bit Commutable = 0> {
442 // Scalar operation, reg+reg.
443 def SSrr : SSI<opc, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
444 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
445 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
446 let isCommutable = Commutable;
449 // Scalar operation, reg+mem.
450 def SSrm : SSI<opc, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f32mem:$src2),
451 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
452 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
454 // Vector intrinsic operation, reg+reg.
455 def SSrr_Int : SSI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
456 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
457 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
458 let isCommutable = Commutable;
461 // Vector intrinsic operation, reg+mem.
462 def SSrm_Int : SSI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, ssmem:$src2),
463 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
464 [(set VR128:$dst, (F32Int VR128:$src1,
465 sse_load_f32:$src2))]>;
469 // Arithmetic instructions
470 defm ADD : scalar_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
471 defm MUL : scalar_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
472 defm SUB : scalar_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
473 defm DIV : scalar_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
475 defm MAX : scalar_sse1_fp_binop_rm<0x5F, "max", X86fmax, int_x86_sse_max_ss>;
476 defm MIN : scalar_sse1_fp_binop_rm<0x5D, "min", X86fmin, int_x86_sse_min_ss>;
478 //===----------------------------------------------------------------------===//
479 // SSE packed FP Instructions
482 def MOVAPSrr : PSI<0x28, MRMSrcReg, (ops VR128:$dst, VR128:$src),
483 "movaps {$src, $dst|$dst, $src}", []>;
484 def MOVAPSrm : PSI<0x28, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
485 "movaps {$src, $dst|$dst, $src}",
486 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
488 def MOVAPSmr : PSI<0x29, MRMDestMem, (ops f128mem:$dst, VR128:$src),
489 "movaps {$src, $dst|$dst, $src}",
490 [(store (v4f32 VR128:$src), addr:$dst)]>;
492 def MOVUPSrr : PSI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src),
493 "movups {$src, $dst|$dst, $src}", []>;
494 def MOVUPSrm : PSI<0x10, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
495 "movups {$src, $dst|$dst, $src}",
496 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
497 def MOVUPSmr : PSI<0x11, MRMDestMem, (ops f128mem:$dst, VR128:$src),
498 "movups {$src, $dst|$dst, $src}",
499 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
501 let isTwoAddress = 1 in {
502 let AddedComplexity = 20 in {
503 def MOVLPSrm : PSI<0x12, MRMSrcMem,
504 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
505 "movlps {$src2, $dst|$dst, $src2}",
507 (v4f32 (vector_shuffle VR128:$src1,
508 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
509 MOVLP_shuffle_mask)))]>;
510 def MOVHPSrm : PSI<0x16, MRMSrcMem,
511 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
512 "movhps {$src2, $dst|$dst, $src2}",
514 (v4f32 (vector_shuffle VR128:$src1,
515 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
516 MOVHP_shuffle_mask)))]>;
520 def MOVLPSmr : PSI<0x13, MRMDestMem, (ops f64mem:$dst, VR128:$src),
521 "movlps {$src, $dst|$dst, $src}",
522 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
523 (iPTR 0))), addr:$dst)]>;
525 // v2f64 extract element 1 is always custom lowered to unpack high to low
526 // and extract element 0 so the non-store version isn't too horrible.
527 def MOVHPSmr : PSI<0x17, MRMDestMem, (ops f64mem:$dst, VR128:$src),
528 "movhps {$src, $dst|$dst, $src}",
529 [(store (f64 (vector_extract
530 (v2f64 (vector_shuffle
531 (bc_v2f64 (v4f32 VR128:$src)), (undef),
532 UNPCKH_shuffle_mask)), (iPTR 0))),
535 let isTwoAddress = 1 in {
536 let AddedComplexity = 15 in {
537 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
538 "movlhps {$src2, $dst|$dst, $src2}",
540 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
541 MOVHP_shuffle_mask)))]>;
543 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
544 "movhlps {$src2, $dst|$dst, $src2}",
546 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
547 MOVHLPS_shuffle_mask)))]>;
553 /// packed_sse1_fp_binop_rm - Packed SSE binops come in three basic forms:
554 /// 1. v4f32 - This comes in SSE1 form for float.
555 /// 2. rr vs rm - They include a reg+reg form and a ref+mem form.
557 let isTwoAddress = 1 in {
558 multiclass packed_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
559 SDNode OpNode, bit Commutable = 0> {
560 // Packed operation, reg+reg.
561 def PSrr : PSI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
562 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2}"),
563 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
564 let isCommutable = Commutable;
567 // Packed operation, reg+mem.
568 def PSrm : PSI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
569 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2}"),
570 [(set VR128:$dst, (OpNode VR128:$src1, (loadv4f32 addr:$src2)))]>;
574 defm ADD : packed_sse1_fp_binop_rm<0x58, "add", fadd, 1>;
575 defm MUL : packed_sse1_fp_binop_rm<0x59, "mul", fmul, 1>;
576 defm DIV : packed_sse1_fp_binop_rm<0x5E, "div", fdiv>;
577 defm SUB : packed_sse1_fp_binop_rm<0x5C, "sub", fsub>;
581 class PS_Intr<bits<8> o, string OpcodeStr, Intrinsic IntId>
582 : PSI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
583 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
584 [(set VR128:$dst, (IntId VR128:$src))]>;
585 class PS_Intm<bits<8> o, string OpcodeStr, Intrinsic IntId>
586 : PSI<o, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
587 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
588 [(set VR128:$dst, (IntId (load addr:$src)))]>;
590 class PS_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
591 : PSI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
592 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
593 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
594 class PS_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
595 : PSI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f32mem:$src2),
596 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
597 [(set VR128:$dst, (IntId VR128:$src1, (load addr:$src2)))]>;
599 def SQRTPSr : PS_Intr<0x51, "sqrtps", int_x86_sse_sqrt_ps>;
600 def SQRTPSm : PS_Intm<0x51, "sqrtps", int_x86_sse_sqrt_ps>;
602 def RSQRTPSr : PS_Intr<0x52, "rsqrtps", int_x86_sse_rsqrt_ps>;
603 def RSQRTPSm : PS_Intm<0x52, "rsqrtps", int_x86_sse_rsqrt_ps>;
604 def RCPPSr : PS_Intr<0x53, "rcpps", int_x86_sse_rcp_ps>;
605 def RCPPSm : PS_Intm<0x53, "rcpps", int_x86_sse_rcp_ps>;
607 let isTwoAddress = 1 in {
608 let isCommutable = 1 in {
609 def MAXPSrr : PS_Intrr<0x5F, "maxps", int_x86_sse_max_ps>;
610 def MINPSrr : PS_Intrr<0x5D, "minps", int_x86_sse_min_ps>;
613 def MAXPSrm : PS_Intrm<0x5F, "maxps", int_x86_sse_max_ps>;
614 def MINPSrm : PS_Intrm<0x5D, "minps", int_x86_sse_min_ps>;
618 let isTwoAddress = 1 in {
619 let isCommutable = 1 in {
620 def ANDPSrr : PSI<0x54, MRMSrcReg,
621 (ops VR128:$dst, VR128:$src1, VR128:$src2),
622 "andps {$src2, $dst|$dst, $src2}",
623 [(set VR128:$dst, (v2i64
624 (and VR128:$src1, VR128:$src2)))]>;
625 def ORPSrr : PSI<0x56, MRMSrcReg,
626 (ops VR128:$dst, VR128:$src1, VR128:$src2),
627 "orps {$src2, $dst|$dst, $src2}",
628 [(set VR128:$dst, (v2i64
629 (or VR128:$src1, VR128:$src2)))]>;
630 def XORPSrr : PSI<0x57, MRMSrcReg,
631 (ops VR128:$dst, VR128:$src1, VR128:$src2),
632 "xorps {$src2, $dst|$dst, $src2}",
633 [(set VR128:$dst, (v2i64
634 (xor VR128:$src1, VR128:$src2)))]>;
637 def ANDPSrm : PSI<0x54, MRMSrcMem,
638 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
639 "andps {$src2, $dst|$dst, $src2}",
640 [(set VR128:$dst, (and VR128:$src1,
641 (bc_v2i64 (loadv4f32 addr:$src2))))]>;
642 def ORPSrm : PSI<0x56, MRMSrcMem,
643 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
644 "orps {$src2, $dst|$dst, $src2}",
645 [(set VR128:$dst, (or VR128:$src1,
646 (bc_v2i64 (loadv4f32 addr:$src2))))]>;
647 def XORPSrm : PSI<0x57, MRMSrcMem,
648 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
649 "xorps {$src2, $dst|$dst, $src2}",
650 [(set VR128:$dst, (xor VR128:$src1,
651 (bc_v2i64 (loadv4f32 addr:$src2))))]>;
652 def ANDNPSrr : PSI<0x55, MRMSrcReg,
653 (ops VR128:$dst, VR128:$src1, VR128:$src2),
654 "andnps {$src2, $dst|$dst, $src2}",
656 (v2i64 (and (xor VR128:$src1,
657 (bc_v2i64 (v4i32 immAllOnesV))),
659 def ANDNPSrm : PSI<0x55, MRMSrcMem,
660 (ops VR128:$dst, VR128:$src1,f128mem:$src2),
661 "andnps {$src2, $dst|$dst, $src2}",
663 (v2i64 (and (xor VR128:$src1,
664 (bc_v2i64 (v4i32 immAllOnesV))),
665 (bc_v2i64 (loadv4f32 addr:$src2)))))]>;
668 let isTwoAddress = 1 in {
669 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
670 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
671 "cmp${cc}ps {$src, $dst|$dst, $src}",
672 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
673 VR128:$src, imm:$cc))]>;
674 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
675 (ops VR128:$dst, VR128:$src1, f128mem:$src, SSECC:$cc),
676 "cmp${cc}ps {$src, $dst|$dst, $src}",
677 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
678 (load addr:$src), imm:$cc))]>;
681 // Shuffle and unpack instructions
682 let isTwoAddress = 1 in {
683 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
684 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
685 (ops VR128:$dst, VR128:$src1,
686 VR128:$src2, i32i8imm:$src3),
687 "shufps {$src3, $src2, $dst|$dst, $src2, $src3}",
689 (v4f32 (vector_shuffle
690 VR128:$src1, VR128:$src2,
691 SHUFP_shuffle_mask:$src3)))]>;
692 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
693 (ops VR128:$dst, VR128:$src1,
694 f128mem:$src2, i32i8imm:$src3),
695 "shufps {$src3, $src2, $dst|$dst, $src2, $src3}",
697 (v4f32 (vector_shuffle
698 VR128:$src1, (load addr:$src2),
699 SHUFP_shuffle_mask:$src3)))]>;
701 let AddedComplexity = 10 in {
702 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
703 (ops VR128:$dst, VR128:$src1, VR128:$src2),
704 "unpckhps {$src2, $dst|$dst, $src2}",
706 (v4f32 (vector_shuffle
707 VR128:$src1, VR128:$src2,
708 UNPCKH_shuffle_mask)))]>;
709 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
710 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
711 "unpckhps {$src2, $dst|$dst, $src2}",
713 (v4f32 (vector_shuffle
714 VR128:$src1, (load addr:$src2),
715 UNPCKH_shuffle_mask)))]>;
717 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
718 (ops VR128:$dst, VR128:$src1, VR128:$src2),
719 "unpcklps {$src2, $dst|$dst, $src2}",
721 (v4f32 (vector_shuffle
722 VR128:$src1, VR128:$src2,
723 UNPCKL_shuffle_mask)))]>;
724 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
725 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
726 "unpcklps {$src2, $dst|$dst, $src2}",
728 (v4f32 (vector_shuffle
729 VR128:$src1, (load addr:$src2),
730 UNPCKL_shuffle_mask)))]>;
735 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (ops GR32:$dst, VR128:$src),
736 "movmskps {$src, $dst|$dst, $src}",
737 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
738 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (ops GR32:$dst, VR128:$src),
739 "movmskpd {$src, $dst|$dst, $src}",
740 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
742 // Prefetching loads.
743 // TODO: no intrinsics for these?
744 def PREFETCHT0 : PSI<0x18, MRM1m, (ops i8mem:$src), "prefetcht0 $src", []>;
745 def PREFETCHT1 : PSI<0x18, MRM2m, (ops i8mem:$src), "prefetcht1 $src", []>;
746 def PREFETCHT2 : PSI<0x18, MRM3m, (ops i8mem:$src), "prefetcht2 $src", []>;
747 def PREFETCHNTA : PSI<0x18, MRM0m, (ops i8mem:$src), "prefetchnta $src", []>;
749 // Non-temporal stores
750 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (ops i128mem:$dst, VR128:$src),
751 "movntps {$src, $dst|$dst, $src}",
752 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
754 // Load, store, and memory fence
755 def SFENCE : PSI<0xAE, MRM7m, (ops), "sfence", [(int_x86_sse_sfence)]>;
758 def LDMXCSR : PSI<0xAE, MRM2m, (ops i32mem:$src),
759 "ldmxcsr $src", [(int_x86_sse_ldmxcsr addr:$src)]>;
760 def STMXCSR : PSI<0xAE, MRM3m, (ops i32mem:$dst),
761 "stmxcsr $dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
763 // Alias instructions that map zero vector to pxor / xorp* for sse.
764 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
765 let isReMaterializable = 1 in
766 def V_SET0 : PSI<0x57, MRMInitReg, (ops VR128:$dst),
768 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
770 // FR32 to 128-bit vector conversion.
771 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (ops VR128:$dst, FR32:$src),
772 "movss {$src, $dst|$dst, $src}",
774 (v4f32 (scalar_to_vector FR32:$src)))]>;
775 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
776 "movss {$src, $dst|$dst, $src}",
778 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
780 // FIXME: may not be able to eliminate this movss with coalescing the src and
781 // dest register classes are different. We really want to write this pattern
783 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
785 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (ops FR32:$dst, VR128:$src),
786 "movss {$src, $dst|$dst, $src}",
787 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
789 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (ops f32mem:$dst, VR128:$src),
790 "movss {$src, $dst|$dst, $src}",
791 [(store (f32 (vector_extract (v4f32 VR128:$src),
792 (iPTR 0))), addr:$dst)]>;
795 // Move to lower bits of a VR128, leaving upper bits alone.
796 // Three operand (but two address) aliases.
797 let isTwoAddress = 1 in {
798 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
799 (ops VR128:$dst, VR128:$src1, FR32:$src2),
800 "movss {$src2, $dst|$dst, $src2}", []>;
802 let AddedComplexity = 15 in
803 def MOVLPSrr : SSI<0x10, MRMSrcReg,
804 (ops VR128:$dst, VR128:$src1, VR128:$src2),
805 "movss {$src2, $dst|$dst, $src2}",
807 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
808 MOVL_shuffle_mask)))]>;
811 // Move to lower bits of a VR128 and zeroing upper bits.
812 // Loading from memory automatically zeroing upper bits.
813 let AddedComplexity = 20 in
814 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
815 "movss {$src, $dst|$dst, $src}",
816 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV,
817 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
818 MOVL_shuffle_mask)))]>;
821 //===----------------------------------------------------------------------===//
823 //===----------------------------------------------------------------------===//
825 // SSE2 Instruction Templates:
827 // SDI - SSE2 instructions with XD prefix.
828 // PDI - SSE2 instructions with TB and OpSize prefixes.
829 // PDIi8 - SSE2 instructions with ImmT == Imm8 and TB and OpSize prefixes.
831 class SDI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
832 : I<o, F, ops, asm, pattern>, XD, Requires<[HasSSE2]>;
833 class PDI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
834 : I<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE2]>;
835 class PDIi8<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
836 : Ii8<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE2]>;
838 // Helpers for defining instructions that directly correspond to intrinsics.
839 multiclass SD_IntUnary<bits<8> o, string OpcodeStr, Intrinsic IntId> {
840 def r : SDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
841 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
842 [(set VR128:$dst, (v2f64 (IntId VR128:$src)))]>;
843 def m : SDI<o, MRMSrcMem, (ops VR128:$dst, sdmem:$src),
844 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
845 [(set VR128:$dst, (v2f64 (IntId sse_load_f64:$src)))]>;
849 def MOVSDrr : SDI<0x10, MRMSrcReg, (ops FR64:$dst, FR64:$src),
850 "movsd {$src, $dst|$dst, $src}", []>;
851 def MOVSDrm : SDI<0x10, MRMSrcMem, (ops FR64:$dst, f64mem:$src),
852 "movsd {$src, $dst|$dst, $src}",
853 [(set FR64:$dst, (loadf64 addr:$src))]>;
854 def MOVSDmr : SDI<0x11, MRMDestMem, (ops f64mem:$dst, FR64:$src),
855 "movsd {$src, $dst|$dst, $src}",
856 [(store FR64:$src, addr:$dst)]>;
858 def SQRTSDr : SDI<0x51, MRMSrcReg, (ops FR64:$dst, FR64:$src),
859 "sqrtsd {$src, $dst|$dst, $src}",
860 [(set FR64:$dst, (fsqrt FR64:$src))]>;
861 def SQRTSDm : SDI<0x51, MRMSrcMem, (ops FR64:$dst, f64mem:$src),
862 "sqrtsd {$src, $dst|$dst, $src}",
863 [(set FR64:$dst, (fsqrt (loadf64 addr:$src)))]>;
865 // Aliases to match intrinsics which expect XMM operand(s).
866 defm SQRTSD_Int : SD_IntUnary<0x51, "sqrtsd" , int_x86_sse2_sqrt_sd>;
868 // Conversion instructions
869 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (ops GR32:$dst, FR64:$src),
870 "cvttsd2si {$src, $dst|$dst, $src}",
871 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
872 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (ops GR32:$dst, f64mem:$src),
873 "cvttsd2si {$src, $dst|$dst, $src}",
874 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
875 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (ops FR32:$dst, FR64:$src),
876 "cvtsd2ss {$src, $dst|$dst, $src}",
877 [(set FR32:$dst, (fround FR64:$src))]>;
878 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (ops FR32:$dst, f64mem:$src),
879 "cvtsd2ss {$src, $dst|$dst, $src}",
880 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
881 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (ops FR64:$dst, GR32:$src),
882 "cvtsi2sd {$src, $dst|$dst, $src}",
883 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
884 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (ops FR64:$dst, i32mem:$src),
885 "cvtsi2sd {$src, $dst|$dst, $src}",
886 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
888 // SSE2 instructions with XS prefix
889 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (ops FR64:$dst, FR32:$src),
890 "cvtss2sd {$src, $dst|$dst, $src}",
891 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
893 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (ops FR64:$dst, f32mem:$src),
894 "cvtss2sd {$src, $dst|$dst, $src}",
895 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
898 // Match intrinsics which expect XMM operand(s).
899 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (ops GR32:$dst, VR128:$src),
900 "cvtsd2si {$src, $dst|$dst, $src}",
901 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
902 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (ops GR32:$dst, f128mem:$src),
903 "cvtsd2si {$src, $dst|$dst, $src}",
904 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
905 (load addr:$src)))]>;
907 // Aliases for intrinsics
908 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (ops GR32:$dst, VR128:$src),
909 "cvttsd2si {$src, $dst|$dst, $src}",
911 (int_x86_sse2_cvttsd2si VR128:$src))]>;
912 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (ops GR32:$dst, f128mem:$src),
913 "cvttsd2si {$src, $dst|$dst, $src}",
914 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
915 (load addr:$src)))]>;
917 // Comparison instructions
918 let isTwoAddress = 1 in {
919 def CMPSDrr : SDI<0xC2, MRMSrcReg,
920 (ops FR64:$dst, FR64:$src1, FR64:$src, SSECC:$cc),
921 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
922 def CMPSDrm : SDI<0xC2, MRMSrcMem,
923 (ops FR64:$dst, FR64:$src1, f64mem:$src, SSECC:$cc),
924 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
927 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (ops FR64:$src1, FR64:$src2),
928 "ucomisd {$src2, $src1|$src1, $src2}",
929 [(X86cmp FR64:$src1, FR64:$src2)]>;
930 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (ops FR64:$src1, f64mem:$src2),
931 "ucomisd {$src2, $src1|$src1, $src2}",
932 [(X86cmp FR64:$src1, (loadf64 addr:$src2))]>;
934 // Aliases to match intrinsics which expect XMM operand(s).
935 let isTwoAddress = 1 in {
936 def Int_CMPSDrr : SDI<0xC2, MRMSrcReg,
937 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
938 "cmp${cc}sd {$src, $dst|$dst, $src}",
939 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
940 VR128:$src, imm:$cc))]>;
941 def Int_CMPSDrm : SDI<0xC2, MRMSrcMem,
942 (ops VR128:$dst, VR128:$src1, f64mem:$src, SSECC:$cc),
943 "cmp${cc}sd {$src, $dst|$dst, $src}",
944 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
945 (load addr:$src), imm:$cc))]>;
948 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
949 "ucomisd {$src2, $src1|$src1, $src2}",
950 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
951 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
952 "ucomisd {$src2, $src1|$src1, $src2}",
953 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2))]>;
955 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
956 "comisd {$src2, $src1|$src1, $src2}",
957 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
958 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
959 "comisd {$src2, $src1|$src1, $src2}",
960 [(X86comi (v2f64 VR128:$src1), (load addr:$src2))]>;
962 // Aliases of packed SSE2 instructions for scalar use. These all have names that
965 // Alias instructions that map fld0 to pxor for sse.
966 def FsFLD0SD : I<0xEF, MRMInitReg, (ops FR64:$dst),
967 "pxor $dst, $dst", [(set FR64:$dst, fpimm0)]>,
968 Requires<[HasSSE2]>, TB, OpSize;
970 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
972 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (ops FR64:$dst, FR64:$src),
973 "movapd {$src, $dst|$dst, $src}", []>;
975 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
977 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (ops FR64:$dst, f128mem:$src),
978 "movapd {$src, $dst|$dst, $src}",
979 [(set FR64:$dst, (X86loadpf64 addr:$src))]>;
981 // Alias bitwise logical operations using SSE logical ops on packed FP values.
982 let isTwoAddress = 1 in {
983 let isCommutable = 1 in {
984 def FsANDPDrr : PDI<0x54, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
985 "andpd {$src2, $dst|$dst, $src2}",
986 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
987 def FsORPDrr : PDI<0x56, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
988 "orpd {$src2, $dst|$dst, $src2}",
989 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
990 def FsXORPDrr : PDI<0x57, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
991 "xorpd {$src2, $dst|$dst, $src2}",
992 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
995 def FsANDPDrm : PDI<0x54, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
996 "andpd {$src2, $dst|$dst, $src2}",
997 [(set FR64:$dst, (X86fand FR64:$src1,
998 (X86loadpf64 addr:$src2)))]>;
999 def FsORPDrm : PDI<0x56, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
1000 "orpd {$src2, $dst|$dst, $src2}",
1001 [(set FR64:$dst, (X86for FR64:$src1,
1002 (X86loadpf64 addr:$src2)))]>;
1003 def FsXORPDrm : PDI<0x57, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
1004 "xorpd {$src2, $dst|$dst, $src2}",
1005 [(set FR64:$dst, (X86fxor FR64:$src1,
1006 (X86loadpf64 addr:$src2)))]>;
1008 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1009 (ops FR64:$dst, FR64:$src1, FR64:$src2),
1010 "andnpd {$src2, $dst|$dst, $src2}", []>;
1011 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1012 (ops FR64:$dst, FR64:$src1, f128mem:$src2),
1013 "andnpd {$src2, $dst|$dst, $src2}", []>;
1016 /// scalar_sse2_fp_binop_rm - Scalar SSE2 binops come in three basic forms:
1018 /// 1. f64 - This comes in SSE2 form for doubles.
1019 /// 2. rr vs rm - They include a reg+reg form and a reg+mem form.
1021 /// In addition, scalar SSE ops have an intrinsic form. This form is unlike the
1022 /// normal form, in that they take an entire vector (instead of a scalar) and
1023 /// leave the top elements undefined. This adds another two variants of the
1024 /// above permutations, giving us 8 forms for 'instruction'.
1026 let isTwoAddress = 1 in {
1027 multiclass scalar_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1028 SDNode OpNode, Intrinsic F64Int,
1029 bit Commutable = 0> {
1030 // Scalar operation, reg+reg.
1031 def SDrr : SDI<opc, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
1032 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1033 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1034 let isCommutable = Commutable;
1037 // Scalar operation, reg+mem.
1038 def SDrm : SDI<opc, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f64mem:$src2),
1039 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1040 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1042 // Vector intrinsic operation, reg+reg.
1043 def SDrr_Int : SDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1044 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1045 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1046 let isCommutable = Commutable;
1049 // Vector intrinsic operation, reg+mem.
1050 def SDrm_Int : SDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, sdmem:$src2),
1051 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1052 [(set VR128:$dst, (F64Int VR128:$src1,
1053 sse_load_f64:$src2))]>;
1057 // Arithmetic instructions
1058 defm ADD : scalar_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1059 defm MUL : scalar_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1060 defm SUB : scalar_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1061 defm DIV : scalar_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1063 defm MAX : scalar_sse2_fp_binop_rm<0x5F, "max", X86fmax, int_x86_sse2_max_sd>;
1064 defm MIN : scalar_sse2_fp_binop_rm<0x5D, "min", X86fmin, int_x86_sse2_min_sd>;
1066 //===----------------------------------------------------------------------===//
1067 // SSE packed FP Instructions
1069 // Move Instructions
1070 def MOVAPDrr : PDI<0x28, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1071 "movapd {$src, $dst|$dst, $src}", []>;
1072 def MOVAPDrm : PDI<0x28, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1073 "movapd {$src, $dst|$dst, $src}",
1074 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1076 def MOVAPDmr : PDI<0x29, MRMDestMem, (ops f128mem:$dst, VR128:$src),
1077 "movapd {$src, $dst|$dst, $src}",
1078 [(store (v2f64 VR128:$src), addr:$dst)]>;
1080 def MOVUPDrr : PDI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1081 "movupd {$src, $dst|$dst, $src}", []>;
1082 def MOVUPDrm : PDI<0x10, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1083 "movupd {$src, $dst|$dst, $src}",
1084 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1085 def MOVUPDmr : PDI<0x11, MRMDestMem, (ops f128mem:$dst, VR128:$src),
1086 "movupd {$src, $dst|$dst, $src}",
1087 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1089 let isTwoAddress = 1 in {
1090 let AddedComplexity = 20 in {
1091 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1092 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
1093 "movlpd {$src2, $dst|$dst, $src2}",
1095 (v2f64 (vector_shuffle VR128:$src1,
1096 (scalar_to_vector (loadf64 addr:$src2)),
1097 MOVLP_shuffle_mask)))]>;
1098 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1099 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
1100 "movhpd {$src2, $dst|$dst, $src2}",
1102 (v2f64 (vector_shuffle VR128:$src1,
1103 (scalar_to_vector (loadf64 addr:$src2)),
1104 MOVHP_shuffle_mask)))]>;
1105 } // AddedComplexity
1108 def MOVLPDmr : PDI<0x13, MRMDestMem, (ops f64mem:$dst, VR128:$src),
1109 "movlpd {$src, $dst|$dst, $src}",
1110 [(store (f64 (vector_extract (v2f64 VR128:$src),
1111 (iPTR 0))), addr:$dst)]>;
1113 // v2f64 extract element 1 is always custom lowered to unpack high to low
1114 // and extract element 0 so the non-store version isn't too horrible.
1115 def MOVHPDmr : PDI<0x17, MRMDestMem, (ops f64mem:$dst, VR128:$src),
1116 "movhpd {$src, $dst|$dst, $src}",
1117 [(store (f64 (vector_extract
1118 (v2f64 (vector_shuffle VR128:$src, (undef),
1119 UNPCKH_shuffle_mask)), (iPTR 0))),
1122 // SSE2 instructions without OpSize prefix
1123 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1124 "cvtdq2ps {$src, $dst|$dst, $src}",
1125 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1126 TB, Requires<[HasSSE2]>;
1127 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1128 "cvtdq2ps {$src, $dst|$dst, $src}",
1129 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1130 (bitconvert (loadv2i64 addr:$src))))]>,
1131 TB, Requires<[HasSSE2]>;
1133 // SSE2 instructions with XS prefix
1134 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1135 "cvtdq2pd {$src, $dst|$dst, $src}",
1136 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1137 XS, Requires<[HasSSE2]>;
1138 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1139 "cvtdq2pd {$src, $dst|$dst, $src}",
1140 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1141 (bitconvert (loadv2i64 addr:$src))))]>,
1142 XS, Requires<[HasSSE2]>;
1144 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1145 "cvtps2dq {$src, $dst|$dst, $src}",
1146 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1147 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1148 "cvtps2dq {$src, $dst|$dst, $src}",
1149 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1150 (load addr:$src)))]>;
1151 // SSE2 packed instructions with XS prefix
1152 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1153 "cvttps2dq {$src, $dst|$dst, $src}",
1154 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1155 XS, Requires<[HasSSE2]>;
1156 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1157 "cvttps2dq {$src, $dst|$dst, $src}",
1158 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1159 (load addr:$src)))]>,
1160 XS, Requires<[HasSSE2]>;
1162 // SSE2 packed instructions with XD prefix
1163 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1164 "cvtpd2dq {$src, $dst|$dst, $src}",
1165 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1166 XD, Requires<[HasSSE2]>;
1167 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1168 "cvtpd2dq {$src, $dst|$dst, $src}",
1169 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1170 (load addr:$src)))]>,
1171 XD, Requires<[HasSSE2]>;
1173 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1174 "cvttpd2dq {$src, $dst|$dst, $src}",
1175 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1176 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1177 "cvttpd2dq {$src, $dst|$dst, $src}",
1178 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1179 (load addr:$src)))]>;
1181 // SSE2 instructions without OpSize prefix
1182 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1183 "cvtps2pd {$src, $dst|$dst, $src}",
1184 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1185 TB, Requires<[HasSSE2]>;
1186 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (ops VR128:$dst, f64mem:$src),
1187 "cvtps2pd {$src, $dst|$dst, $src}",
1188 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1189 (load addr:$src)))]>,
1190 TB, Requires<[HasSSE2]>;
1192 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1193 "cvtpd2ps {$src, $dst|$dst, $src}",
1194 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1195 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (ops VR128:$dst, f128mem:$src),
1196 "cvtpd2ps {$src, $dst|$dst, $src}",
1197 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1198 (load addr:$src)))]>;
1200 // Match intrinsics which expect XMM operand(s).
1201 // Aliases for intrinsics
1202 let isTwoAddress = 1 in {
1203 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1204 (ops VR128:$dst, VR128:$src1, GR32:$src2),
1205 "cvtsi2sd {$src2, $dst|$dst, $src2}",
1206 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1208 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1209 (ops VR128:$dst, VR128:$src1, i32mem:$src2),
1210 "cvtsi2sd {$src2, $dst|$dst, $src2}",
1211 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1212 (loadi32 addr:$src2)))]>;
1213 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1214 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1215 "cvtsd2ss {$src2, $dst|$dst, $src2}",
1216 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1218 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1219 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
1220 "cvtsd2ss {$src2, $dst|$dst, $src2}",
1221 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1222 (load addr:$src2)))]>;
1223 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1224 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1225 "cvtss2sd {$src2, $dst|$dst, $src2}",
1226 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1227 VR128:$src2))]>, XS,
1228 Requires<[HasSSE2]>;
1229 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1230 (ops VR128:$dst, VR128:$src1, f32mem:$src2),
1231 "cvtss2sd {$src2, $dst|$dst, $src2}",
1232 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1233 (load addr:$src2)))]>, XS,
1234 Requires<[HasSSE2]>;
1237 /// packed_sse2_fp_binop_rm - Packed SSE binops come in three basic forms:
1238 /// 1. v2f64 - This comes in SSE2 form for doubles.
1239 /// 2. rr vs rm - They include a reg+reg form and a ref+mem form.
1241 let isTwoAddress = 1 in {
1242 multiclass packed_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1243 SDNode OpNode, bit Commutable = 0> {
1244 // Packed operation, reg+reg.
1245 def PDrr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1246 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
1247 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1248 let isCommutable = Commutable;
1251 // Packed operation, reg+mem.
1252 def PDrm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1253 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
1254 [(set VR128:$dst, (OpNode VR128:$src1, (loadv2f64 addr:$src2)))]>;
1258 defm ADD : packed_sse2_fp_binop_rm<0x58, "add", fadd, 1>;
1259 defm MUL : packed_sse2_fp_binop_rm<0x59, "mul", fmul, 1>;
1260 defm DIV : packed_sse2_fp_binop_rm<0x5E, "div", fdiv>;
1261 defm SUB : packed_sse2_fp_binop_rm<0x5C, "sub", fsub>;
1265 class PD_Intr<bits<8> o, string OpcodeStr, Intrinsic IntId>
1266 : PDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1267 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
1268 [(set VR128:$dst, (IntId VR128:$src))]>;
1269 class PD_Intm<bits<8> o, string OpcodeStr, Intrinsic IntId>
1270 : PDI<o, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1271 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
1272 [(set VR128:$dst, (IntId (load addr:$src)))]>;
1274 class PD_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
1275 : PDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1276 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1277 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1278 class PD_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
1279 : PDI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
1280 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1281 [(set VR128:$dst, (IntId VR128:$src1, (load addr:$src2)))]>;
1283 def SQRTPDr : PD_Intr<0x51, "sqrtpd", int_x86_sse2_sqrt_pd>;
1284 def SQRTPDm : PD_Intm<0x51, "sqrtpd", int_x86_sse2_sqrt_pd>;
1286 let isTwoAddress = 1 in {
1287 let isCommutable = 1 in {
1288 def MAXPDrr : PD_Intrr<0x5F, "maxpd", int_x86_sse2_max_pd>;
1289 def MINPDrr : PD_Intrr<0x5D, "minpd", int_x86_sse2_min_pd>;
1292 def MAXPDrm : PD_Intrm<0x5F, "maxpd", int_x86_sse2_max_pd>;
1293 def MINPDrm : PD_Intrm<0x5D, "minpd", int_x86_sse2_min_pd>;
1297 let isTwoAddress = 1 in {
1298 let isCommutable = 1 in {
1299 def ANDPDrr : PDI<0x54, MRMSrcReg,
1300 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1301 "andpd {$src2, $dst|$dst, $src2}",
1303 (and (bc_v2i64 (v2f64 VR128:$src1)),
1304 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1305 def ORPDrr : PDI<0x56, MRMSrcReg,
1306 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1307 "orpd {$src2, $dst|$dst, $src2}",
1309 (or (bc_v2i64 (v2f64 VR128:$src1)),
1310 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1311 def XORPDrr : PDI<0x57, MRMSrcReg,
1312 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1313 "xorpd {$src2, $dst|$dst, $src2}",
1315 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1316 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1319 def ANDPDrm : PDI<0x54, MRMSrcMem,
1320 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1321 "andpd {$src2, $dst|$dst, $src2}",
1323 (and (bc_v2i64 (v2f64 VR128:$src1)),
1324 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1325 def ORPDrm : PDI<0x56, MRMSrcMem,
1326 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1327 "orpd {$src2, $dst|$dst, $src2}",
1329 (or (bc_v2i64 (v2f64 VR128:$src1)),
1330 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1331 def XORPDrm : PDI<0x57, MRMSrcMem,
1332 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1333 "xorpd {$src2, $dst|$dst, $src2}",
1335 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1336 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1337 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1338 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1339 "andnpd {$src2, $dst|$dst, $src2}",
1341 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1342 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1343 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1344 (ops VR128:$dst, VR128:$src1,f128mem:$src2),
1345 "andnpd {$src2, $dst|$dst, $src2}",
1347 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1348 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1351 let isTwoAddress = 1 in {
1352 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1353 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
1354 "cmp${cc}pd {$src, $dst|$dst, $src}",
1355 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1356 VR128:$src, imm:$cc))]>;
1357 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1358 (ops VR128:$dst, VR128:$src1, f128mem:$src, SSECC:$cc),
1359 "cmp${cc}pd {$src, $dst|$dst, $src}",
1360 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1361 (load addr:$src), imm:$cc))]>;
1364 // Shuffle and unpack instructions
1365 let isTwoAddress = 1 in {
1366 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1367 (ops VR128:$dst, VR128:$src1, VR128:$src2, i8imm:$src3),
1368 "shufpd {$src3, $src2, $dst|$dst, $src2, $src3}",
1369 [(set VR128:$dst, (v2f64 (vector_shuffle
1370 VR128:$src1, VR128:$src2,
1371 SHUFP_shuffle_mask:$src3)))]>;
1372 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1373 (ops VR128:$dst, VR128:$src1,
1374 f128mem:$src2, i8imm:$src3),
1375 "shufpd {$src3, $src2, $dst|$dst, $src2, $src3}",
1377 (v2f64 (vector_shuffle
1378 VR128:$src1, (load addr:$src2),
1379 SHUFP_shuffle_mask:$src3)))]>;
1381 let AddedComplexity = 10 in {
1382 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1383 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1384 "unpckhpd {$src2, $dst|$dst, $src2}",
1386 (v2f64 (vector_shuffle
1387 VR128:$src1, VR128:$src2,
1388 UNPCKH_shuffle_mask)))]>;
1389 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1390 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1391 "unpckhpd {$src2, $dst|$dst, $src2}",
1393 (v2f64 (vector_shuffle
1394 VR128:$src1, (load addr:$src2),
1395 UNPCKH_shuffle_mask)))]>;
1397 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1398 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1399 "unpcklpd {$src2, $dst|$dst, $src2}",
1401 (v2f64 (vector_shuffle
1402 VR128:$src1, VR128:$src2,
1403 UNPCKL_shuffle_mask)))]>;
1404 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1405 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1406 "unpcklpd {$src2, $dst|$dst, $src2}",
1408 (v2f64 (vector_shuffle
1409 VR128:$src1, (load addr:$src2),
1410 UNPCKL_shuffle_mask)))]>;
1411 } // AddedComplexity
1415 //===----------------------------------------------------------------------===//
1416 // SSE integer instructions
1418 // Move Instructions
1419 def MOVDQArr : PDI<0x6F, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1420 "movdqa {$src, $dst|$dst, $src}", []>;
1421 def MOVDQArm : PDI<0x6F, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1422 "movdqa {$src, $dst|$dst, $src}",
1423 [(set VR128:$dst, (loadv2i64 addr:$src))]>;
1424 def MOVDQAmr : PDI<0x7F, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1425 "movdqa {$src, $dst|$dst, $src}",
1426 [(store (v2i64 VR128:$src), addr:$dst)]>;
1427 def MOVDQUrm : I<0x6F, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1428 "movdqu {$src, $dst|$dst, $src}",
1429 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1430 XS, Requires<[HasSSE2]>;
1431 def MOVDQUmr : I<0x7F, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1432 "movdqu {$src, $dst|$dst, $src}",
1433 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1434 XS, Requires<[HasSSE2]>;
1437 let isTwoAddress = 1 in {
1439 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1440 bit Commutable = 0> {
1441 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1442 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1443 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1444 let isCommutable = Commutable;
1446 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1447 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1448 [(set VR128:$dst, (IntId VR128:$src1,
1449 (bitconvert (loadv2i64 addr:$src2))))]>;
1452 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1453 string OpcodeStr, Intrinsic IntId> {
1454 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1455 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1456 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1457 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1458 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1459 [(set VR128:$dst, (IntId VR128:$src1,
1460 (bitconvert (loadv2i64 addr:$src2))))]>;
1461 def ri : PDIi8<opc2, ImmForm, (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1462 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1463 [(set VR128:$dst, (IntId VR128:$src1,
1464 (scalar_to_vector (i32 imm:$src2))))]>;
1468 /// PDI_binop_rm - Simple SSE2 binary operator.
1469 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1470 ValueType OpVT, bit Commutable = 0> {
1471 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1472 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1473 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1474 let isCommutable = Commutable;
1476 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1477 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1478 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1479 (bitconvert (loadv2i64 addr:$src2)))))]>;
1482 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1484 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1485 /// to collapse (bitconvert VT to VT) into its operand.
1487 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1488 bit Commutable = 0> {
1489 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1490 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1491 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1492 let isCommutable = Commutable;
1494 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1495 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1496 [(set VR128:$dst, (OpNode VR128:$src1,(loadv2i64 addr:$src2)))]>;
1501 // 128-bit Integer Arithmetic
1503 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1504 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1505 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1506 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1508 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1509 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1510 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1511 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1513 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1514 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1515 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1516 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1518 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1519 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1520 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1521 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1523 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1525 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1526 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1527 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1529 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1531 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1532 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1535 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1536 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1537 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1538 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1539 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1542 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw", int_x86_sse2_psll_w>;
1543 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld", int_x86_sse2_psll_d>;
1544 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq", int_x86_sse2_psll_q>;
1546 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw", int_x86_sse2_psrl_w>;
1547 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld", int_x86_sse2_psrl_d>;
1548 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq", int_x86_sse2_psrl_q>;
1550 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw", int_x86_sse2_psra_w>;
1551 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad", int_x86_sse2_psra_d>;
1552 // PSRAQ doesn't exist in SSE[1-3].
1554 // 128-bit logical shifts.
1555 let isTwoAddress = 1 in {
1556 def PSLLDQri : PDIi8<0x73, MRM7r,
1557 (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1558 "pslldq {$src2, $dst|$dst, $src2}", []>;
1559 def PSRLDQri : PDIi8<0x73, MRM3r,
1560 (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1561 "psrldq {$src2, $dst|$dst, $src2}", []>;
1562 // PSRADQri doesn't exist in SSE[1-3].
1565 let Predicates = [HasSSE2] in {
1566 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1567 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1568 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1569 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1570 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1571 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1575 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1576 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1577 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1579 let isTwoAddress = 1 in {
1580 def PANDNrr : PDI<0xDF, MRMSrcReg,
1581 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1582 "pandn {$src2, $dst|$dst, $src2}",
1583 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1586 def PANDNrm : PDI<0xDF, MRMSrcMem,
1587 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1588 "pandn {$src2, $dst|$dst, $src2}",
1589 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1590 (load addr:$src2))))]>;
1593 // SSE2 Integer comparison
1594 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1595 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1596 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1597 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1598 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1599 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1601 // Pack instructions
1602 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1603 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1604 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1606 // Shuffle and unpack instructions
1607 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1608 (ops VR128:$dst, VR128:$src1, i8imm:$src2),
1609 "pshufd {$src2, $src1, $dst|$dst, $src1, $src2}",
1610 [(set VR128:$dst, (v4i32 (vector_shuffle
1611 VR128:$src1, (undef),
1612 PSHUFD_shuffle_mask:$src2)))]>;
1613 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1614 (ops VR128:$dst, i128mem:$src1, i8imm:$src2),
1615 "pshufd {$src2, $src1, $dst|$dst, $src1, $src2}",
1616 [(set VR128:$dst, (v4i32 (vector_shuffle
1617 (bc_v4i32(loadv2i64 addr:$src1)),
1619 PSHUFD_shuffle_mask:$src2)))]>;
1621 // SSE2 with ImmT == Imm8 and XS prefix.
1622 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1623 (ops VR128:$dst, VR128:$src1, i8imm:$src2),
1624 "pshufhw {$src2, $src1, $dst|$dst, $src1, $src2}",
1625 [(set VR128:$dst, (v8i16 (vector_shuffle
1626 VR128:$src1, (undef),
1627 PSHUFHW_shuffle_mask:$src2)))]>,
1628 XS, Requires<[HasSSE2]>;
1629 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
1630 (ops VR128:$dst, i128mem:$src1, i8imm:$src2),
1631 "pshufhw {$src2, $src1, $dst|$dst, $src1, $src2}",
1632 [(set VR128:$dst, (v8i16 (vector_shuffle
1633 (bc_v8i16 (loadv2i64 addr:$src1)),
1635 PSHUFHW_shuffle_mask:$src2)))]>,
1636 XS, Requires<[HasSSE2]>;
1638 // SSE2 with ImmT == Imm8 and XD prefix.
1639 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
1640 (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1641 "pshuflw {$src2, $src1, $dst|$dst, $src1, $src2}",
1642 [(set VR128:$dst, (v8i16 (vector_shuffle
1643 VR128:$src1, (undef),
1644 PSHUFLW_shuffle_mask:$src2)))]>,
1645 XD, Requires<[HasSSE2]>;
1646 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
1647 (ops VR128:$dst, i128mem:$src1, i32i8imm:$src2),
1648 "pshuflw {$src2, $src1, $dst|$dst, $src1, $src2}",
1649 [(set VR128:$dst, (v8i16 (vector_shuffle
1650 (bc_v8i16 (loadv2i64 addr:$src1)),
1652 PSHUFLW_shuffle_mask:$src2)))]>,
1653 XD, Requires<[HasSSE2]>;
1656 let isTwoAddress = 1 in {
1657 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
1658 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1659 "punpcklbw {$src2, $dst|$dst, $src2}",
1661 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1662 UNPCKL_shuffle_mask)))]>;
1663 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
1664 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1665 "punpcklbw {$src2, $dst|$dst, $src2}",
1667 (v16i8 (vector_shuffle VR128:$src1,
1668 (bc_v16i8 (loadv2i64 addr:$src2)),
1669 UNPCKL_shuffle_mask)))]>;
1670 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
1671 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1672 "punpcklwd {$src2, $dst|$dst, $src2}",
1674 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1675 UNPCKL_shuffle_mask)))]>;
1676 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
1677 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1678 "punpcklwd {$src2, $dst|$dst, $src2}",
1680 (v8i16 (vector_shuffle VR128:$src1,
1681 (bc_v8i16 (loadv2i64 addr:$src2)),
1682 UNPCKL_shuffle_mask)))]>;
1683 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
1684 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1685 "punpckldq {$src2, $dst|$dst, $src2}",
1687 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1688 UNPCKL_shuffle_mask)))]>;
1689 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
1690 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1691 "punpckldq {$src2, $dst|$dst, $src2}",
1693 (v4i32 (vector_shuffle VR128:$src1,
1694 (bc_v4i32 (loadv2i64 addr:$src2)),
1695 UNPCKL_shuffle_mask)))]>;
1696 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
1697 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1698 "punpcklqdq {$src2, $dst|$dst, $src2}",
1700 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1701 UNPCKL_shuffle_mask)))]>;
1702 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
1703 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1704 "punpcklqdq {$src2, $dst|$dst, $src2}",
1706 (v2i64 (vector_shuffle VR128:$src1,
1707 (loadv2i64 addr:$src2),
1708 UNPCKL_shuffle_mask)))]>;
1710 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
1711 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1712 "punpckhbw {$src2, $dst|$dst, $src2}",
1714 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1715 UNPCKH_shuffle_mask)))]>;
1716 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
1717 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1718 "punpckhbw {$src2, $dst|$dst, $src2}",
1720 (v16i8 (vector_shuffle VR128:$src1,
1721 (bc_v16i8 (loadv2i64 addr:$src2)),
1722 UNPCKH_shuffle_mask)))]>;
1723 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
1724 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1725 "punpckhwd {$src2, $dst|$dst, $src2}",
1727 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1728 UNPCKH_shuffle_mask)))]>;
1729 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
1730 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1731 "punpckhwd {$src2, $dst|$dst, $src2}",
1733 (v8i16 (vector_shuffle VR128:$src1,
1734 (bc_v8i16 (loadv2i64 addr:$src2)),
1735 UNPCKH_shuffle_mask)))]>;
1736 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
1737 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1738 "punpckhdq {$src2, $dst|$dst, $src2}",
1740 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1741 UNPCKH_shuffle_mask)))]>;
1742 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
1743 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1744 "punpckhdq {$src2, $dst|$dst, $src2}",
1746 (v4i32 (vector_shuffle VR128:$src1,
1747 (bc_v4i32 (loadv2i64 addr:$src2)),
1748 UNPCKH_shuffle_mask)))]>;
1749 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
1750 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1751 "punpckhqdq {$src2, $dst|$dst, $src2}",
1753 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1754 UNPCKH_shuffle_mask)))]>;
1755 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
1756 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1757 "punpckhqdq {$src2, $dst|$dst, $src2}",
1759 (v2i64 (vector_shuffle VR128:$src1,
1760 (loadv2i64 addr:$src2),
1761 UNPCKH_shuffle_mask)))]>;
1765 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
1766 (ops GR32:$dst, VR128:$src1, i32i8imm:$src2),
1767 "pextrw {$src2, $src1, $dst|$dst, $src1, $src2}",
1768 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
1769 (iPTR imm:$src2)))]>;
1770 let isTwoAddress = 1 in {
1771 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
1772 (ops VR128:$dst, VR128:$src1,
1773 GR32:$src2, i32i8imm:$src3),
1774 "pinsrw {$src3, $src2, $dst|$dst, $src2, $src3}",
1776 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
1777 GR32:$src2, (iPTR imm:$src3))))]>;
1778 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
1779 (ops VR128:$dst, VR128:$src1,
1780 i16mem:$src2, i32i8imm:$src3),
1781 "pinsrw {$src3, $src2, $dst|$dst, $src2, $src3}",
1783 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
1784 (i32 (anyext (loadi16 addr:$src2))),
1785 (iPTR imm:$src3))))]>;
1789 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (ops GR32:$dst, VR128:$src),
1790 "pmovmskb {$src, $dst|$dst, $src}",
1791 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
1793 // Conditional store
1794 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (ops VR128:$src, VR128:$mask),
1795 "maskmovdqu {$mask, $src|$src, $mask}",
1796 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>,
1799 // Non-temporal stores
1800 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1801 "movntpd {$src, $dst|$dst, $src}",
1802 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
1803 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (ops f128mem:$dst, VR128:$src),
1804 "movntdq {$src, $dst|$dst, $src}",
1805 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
1806 def MOVNTImr : I<0xC3, MRMDestMem, (ops i32mem:$dst, GR32:$src),
1807 "movnti {$src, $dst|$dst, $src}",
1808 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
1809 TB, Requires<[HasSSE2]>;
1812 def CLFLUSH : I<0xAE, MRM7m, (ops i8mem:$src),
1813 "clflush $src", [(int_x86_sse2_clflush addr:$src)]>,
1814 TB, Requires<[HasSSE2]>;
1816 // Load, store, and memory fence
1817 def LFENCE : I<0xAE, MRM5m, (ops),
1818 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
1819 def MFENCE : I<0xAE, MRM6m, (ops),
1820 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
1823 // Alias instructions that map zero vector to pxor / xorp* for sse.
1824 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
1825 let isReMaterializable = 1 in
1826 def V_SETALLONES : PDI<0x76, MRMInitReg, (ops VR128:$dst),
1827 "pcmpeqd $dst, $dst",
1828 [(set VR128:$dst, (v2f64 immAllOnesV))]>;
1830 // FR64 to 128-bit vector conversion.
1831 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (ops VR128:$dst, FR64:$src),
1832 "movsd {$src, $dst|$dst, $src}",
1834 (v2f64 (scalar_to_vector FR64:$src)))]>;
1835 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1836 "movsd {$src, $dst|$dst, $src}",
1838 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
1840 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (ops VR128:$dst, GR32:$src),
1841 "movd {$src, $dst|$dst, $src}",
1843 (v4i32 (scalar_to_vector GR32:$src)))]>;
1844 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (ops VR128:$dst, i32mem:$src),
1845 "movd {$src, $dst|$dst, $src}",
1847 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
1849 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (ops FR32:$dst, GR32:$src),
1850 "movd {$src, $dst|$dst, $src}",
1851 [(set FR32:$dst, (bitconvert GR32:$src))]>;
1853 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (ops FR32:$dst, i32mem:$src),
1854 "movd {$src, $dst|$dst, $src}",
1855 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
1857 // SSE2 instructions with XS prefix
1858 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1859 "movq {$src, $dst|$dst, $src}",
1861 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
1862 Requires<[HasSSE2]>;
1863 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (ops i64mem:$dst, VR128:$src),
1864 "movq {$src, $dst|$dst, $src}",
1865 [(store (i64 (vector_extract (v2i64 VR128:$src),
1866 (iPTR 0))), addr:$dst)]>;
1868 // FIXME: may not be able to eliminate this movss with coalescing the src and
1869 // dest register classes are different. We really want to write this pattern
1871 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1872 // (f32 FR32:$src)>;
1873 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (ops FR64:$dst, VR128:$src),
1874 "movsd {$src, $dst|$dst, $src}",
1875 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
1877 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (ops f64mem:$dst, VR128:$src),
1878 "movsd {$src, $dst|$dst, $src}",
1879 [(store (f64 (vector_extract (v2f64 VR128:$src),
1880 (iPTR 0))), addr:$dst)]>;
1881 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (ops GR32:$dst, VR128:$src),
1882 "movd {$src, $dst|$dst, $src}",
1883 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
1885 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (ops i32mem:$dst, VR128:$src),
1886 "movd {$src, $dst|$dst, $src}",
1887 [(store (i32 (vector_extract (v4i32 VR128:$src),
1888 (iPTR 0))), addr:$dst)]>;
1890 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (ops GR32:$dst, FR32:$src),
1891 "movd {$src, $dst|$dst, $src}",
1892 [(set GR32:$dst, (bitconvert FR32:$src))]>;
1893 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (ops i32mem:$dst, FR32:$src),
1894 "movd {$src, $dst|$dst, $src}",
1895 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
1898 // Move to lower bits of a VR128, leaving upper bits alone.
1899 // Three operand (but two address) aliases.
1900 let isTwoAddress = 1 in {
1901 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
1902 (ops VR128:$dst, VR128:$src1, FR64:$src2),
1903 "movsd {$src2, $dst|$dst, $src2}", []>;
1905 let AddedComplexity = 15 in
1906 def MOVLPDrr : SDI<0x10, MRMSrcReg,
1907 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1908 "movsd {$src2, $dst|$dst, $src2}",
1910 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
1911 MOVL_shuffle_mask)))]>;
1914 // Store / copy lower 64-bits of a XMM register.
1915 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (ops i64mem:$dst, VR128:$src),
1916 "movq {$src, $dst|$dst, $src}",
1917 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
1919 // Move to lower bits of a VR128 and zeroing upper bits.
1920 // Loading from memory automatically zeroing upper bits.
1921 let AddedComplexity = 20 in
1922 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1923 "movsd {$src, $dst|$dst, $src}",
1925 (v2f64 (vector_shuffle immAllZerosV,
1926 (v2f64 (scalar_to_vector
1927 (loadf64 addr:$src))),
1928 MOVL_shuffle_mask)))]>;
1930 let AddedComplexity = 15 in
1931 // movd / movq to XMM register zero-extends
1932 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (ops VR128:$dst, GR32:$src),
1933 "movd {$src, $dst|$dst, $src}",
1935 (v4i32 (vector_shuffle immAllZerosV,
1936 (v4i32 (scalar_to_vector GR32:$src)),
1937 MOVL_shuffle_mask)))]>;
1938 let AddedComplexity = 20 in
1939 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (ops VR128:$dst, i32mem:$src),
1940 "movd {$src, $dst|$dst, $src}",
1942 (v4i32 (vector_shuffle immAllZerosV,
1943 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
1944 MOVL_shuffle_mask)))]>;
1946 // Moving from XMM to XMM but still clear upper 64 bits.
1947 let AddedComplexity = 15 in
1948 def MOVZQI2PQIrr : I<0x7E, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1949 "movq {$src, $dst|$dst, $src}",
1950 [(set VR128:$dst, (int_x86_sse2_movl_dq VR128:$src))]>,
1951 XS, Requires<[HasSSE2]>;
1952 let AddedComplexity = 20 in
1953 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1954 "movq {$src, $dst|$dst, $src}",
1955 [(set VR128:$dst, (int_x86_sse2_movl_dq
1956 (bitconvert (loadv2i64 addr:$src))))]>,
1957 XS, Requires<[HasSSE2]>;
1960 //===----------------------------------------------------------------------===//
1961 // SSE3 Instructions
1962 //===----------------------------------------------------------------------===//
1964 // SSE3 Instruction Templates:
1966 // S3I - SSE3 instructions with TB and OpSize prefixes.
1967 // S3SI - SSE3 instructions with XS prefix.
1968 // S3DI - SSE3 instructions with XD prefix.
1970 class S3SI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
1971 : I<o, F, ops, asm, pattern>, XS, Requires<[HasSSE3]>;
1972 class S3DI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
1973 : I<o, F, ops, asm, pattern>, XD, Requires<[HasSSE3]>;
1974 class S3I<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
1975 : I<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE3]>;
1977 // Move Instructions
1978 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1979 "movshdup {$src, $dst|$dst, $src}",
1980 [(set VR128:$dst, (v4f32 (vector_shuffle
1981 VR128:$src, (undef),
1982 MOVSHDUP_shuffle_mask)))]>;
1983 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1984 "movshdup {$src, $dst|$dst, $src}",
1985 [(set VR128:$dst, (v4f32 (vector_shuffle
1986 (loadv4f32 addr:$src), (undef),
1987 MOVSHDUP_shuffle_mask)))]>;
1989 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1990 "movsldup {$src, $dst|$dst, $src}",
1991 [(set VR128:$dst, (v4f32 (vector_shuffle
1992 VR128:$src, (undef),
1993 MOVSLDUP_shuffle_mask)))]>;
1994 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1995 "movsldup {$src, $dst|$dst, $src}",
1996 [(set VR128:$dst, (v4f32 (vector_shuffle
1997 (loadv4f32 addr:$src), (undef),
1998 MOVSLDUP_shuffle_mask)))]>;
2000 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src),
2001 "movddup {$src, $dst|$dst, $src}",
2002 [(set VR128:$dst, (v2f64 (vector_shuffle
2003 VR128:$src, (undef),
2004 SSE_splat_lo_mask)))]>;
2005 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
2006 "movddup {$src, $dst|$dst, $src}",
2008 (v2f64 (vector_shuffle
2009 (scalar_to_vector (loadf64 addr:$src)),
2011 SSE_splat_lo_mask)))]>;
2014 let isTwoAddress = 1 in {
2015 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2016 (ops VR128:$dst, VR128:$src1, VR128:$src2),
2017 "addsubps {$src2, $dst|$dst, $src2}",
2018 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2020 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2021 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
2022 "addsubps {$src2, $dst|$dst, $src2}",
2023 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2024 (load addr:$src2)))]>;
2025 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2026 (ops VR128:$dst, VR128:$src1, VR128:$src2),
2027 "addsubpd {$src2, $dst|$dst, $src2}",
2028 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2030 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2031 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
2032 "addsubpd {$src2, $dst|$dst, $src2}",
2033 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2034 (load addr:$src2)))]>;
2037 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
2038 "lddqu {$src, $dst|$dst, $src}",
2039 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2042 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2043 : S3DI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
2044 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2045 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2046 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2047 : S3DI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
2048 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2049 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
2050 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2051 : S3I<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
2052 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2053 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2054 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2055 : S3I<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
2056 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2057 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
2059 let isTwoAddress = 1 in {
2060 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2061 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2062 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2063 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2064 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2065 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2066 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2067 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2070 // Thread synchronization
2071 def MONITOR : I<0xC8, RawFrm, (ops), "monitor",
2072 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2073 def MWAIT : I<0xC9, RawFrm, (ops), "mwait",
2074 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2076 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2077 let AddedComplexity = 15 in
2078 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2079 MOVSHDUP_shuffle_mask)),
2080 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2081 let AddedComplexity = 20 in
2082 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (loadv2i64 addr:$src)), (undef),
2083 MOVSHDUP_shuffle_mask)),
2084 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2086 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2087 let AddedComplexity = 15 in
2088 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2089 MOVSLDUP_shuffle_mask)),
2090 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2091 let AddedComplexity = 20 in
2092 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (loadv2i64 addr:$src)), (undef),
2093 MOVSLDUP_shuffle_mask)),
2094 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2096 //===----------------------------------------------------------------------===//
2097 // SSSE3 Instructions
2098 //===----------------------------------------------------------------------===//
2100 // SSE3 Instruction Templates:
2102 // SS38I - SSSE3 instructions with T8 and OpSize prefixes.
2103 // SS3AI - SSSE3 instructions with TA and OpSize prefixes.
2105 class SS38I<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
2106 : I<o, F, ops, asm, pattern>, T8, OpSize, Requires<[HasSSSE3]>;
2107 class SS3AI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
2108 : I<o, F, ops, asm, pattern>, TA, OpSize, Requires<[HasSSSE3]>;
2110 /// SS3I_binop_rm_int - Simple SSSE3 binary operatr whose type is v2i64.
2111 let isTwoAddress = 1 in {
2112 multiclass SS3I_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
2113 bit Commutable = 0> {
2114 def rr : SS38I<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
2115 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2116 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
2117 let isCommutable = Commutable;
2119 def rm : SS38I<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
2120 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2123 (bitconvert (loadv2i64 addr:$src2))))]>;
2127 defm PMULHRSW128 : SS3I_binop_rm_int<0x0B, "pmulhrsw",
2128 int_x86_ssse3_pmulhrsw_128, 1>;
2130 //===----------------------------------------------------------------------===//
2131 // Non-Instruction Patterns
2132 //===----------------------------------------------------------------------===//
2134 // 128-bit vector undef's.
2135 def : Pat<(v2f64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2136 def : Pat<(v16i8 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2137 def : Pat<(v8i16 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2138 def : Pat<(v4i32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2139 def : Pat<(v2i64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2141 // 128-bit vector all zero's.
2142 def : Pat<(v16i8 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2143 def : Pat<(v8i16 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2144 def : Pat<(v4i32 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2145 def : Pat<(v2i64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2146 def : Pat<(v2f64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2148 // 128-bit vector all one's.
2149 def : Pat<(v16i8 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2150 def : Pat<(v8i16 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2151 def : Pat<(v4i32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2152 def : Pat<(v2i64 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2153 def : Pat<(v4f32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE1]>;
2155 // Store 128-bit integer vector values.
2156 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
2157 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2158 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
2159 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2160 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
2161 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2163 // Scalar to v8i16 / v16i8. The source may be a GR32, but only the lower 8 or
2165 def : Pat<(v8i16 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2166 Requires<[HasSSE2]>;
2167 def : Pat<(v16i8 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2168 Requires<[HasSSE2]>;
2171 let Predicates = [HasSSE2] in {
2172 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2173 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2174 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2175 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2176 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2177 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2178 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2179 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2180 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2181 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2182 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2183 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2184 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2185 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2186 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2187 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2188 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2189 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2190 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2191 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2192 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2193 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2194 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2195 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2196 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2197 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2198 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2199 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2200 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2201 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2204 // Move scalar to XMM zero-extended
2205 // movd to XMM register zero-extends
2206 let AddedComplexity = 15 in {
2207 def : Pat<(v8i16 (vector_shuffle immAllZerosV,
2208 (v8i16 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
2209 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
2210 def : Pat<(v16i8 (vector_shuffle immAllZerosV,
2211 (v16i8 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
2212 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
2213 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2214 def : Pat<(v2f64 (vector_shuffle immAllZerosV,
2215 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
2216 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2217 def : Pat<(v4f32 (vector_shuffle immAllZerosV,
2218 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
2219 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
2222 // Splat v2f64 / v2i64
2223 let AddedComplexity = 10 in {
2224 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2225 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2226 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2227 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2228 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2229 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2230 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2231 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2235 def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SSE_splat_mask:$sm),
2236 (SHUFPSrri VR128:$src, VR128:$src, SSE_splat_mask:$sm)>,
2237 Requires<[HasSSE1]>;
2239 // Special unary SHUFPSrri case.
2240 // FIXME: when we want non two-address code, then we should use PSHUFD?
2241 def : Pat<(vector_shuffle (v4f32 VR128:$src1), (undef),
2242 SHUFP_unary_shuffle_mask:$sm),
2243 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2244 Requires<[HasSSE1]>;
2245 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2246 def : Pat<(vector_shuffle (loadv4f32 addr:$src1), (undef),
2247 SHUFP_unary_shuffle_mask:$sm),
2248 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2249 Requires<[HasSSE2]>;
2250 // Special binary v4i32 shuffle cases with SHUFPS.
2251 def : Pat<(vector_shuffle (v4i32 VR128:$src1), (v4i32 VR128:$src2),
2252 PSHUFD_binary_shuffle_mask:$sm),
2253 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2254 Requires<[HasSSE2]>;
2255 def : Pat<(vector_shuffle (v4i32 VR128:$src1),
2256 (bc_v4i32 (loadv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm),
2257 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2258 Requires<[HasSSE2]>;
2260 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2261 let AddedComplexity = 10 in {
2262 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2263 UNPCKL_v_undef_shuffle_mask)),
2264 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2265 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2266 UNPCKL_v_undef_shuffle_mask)),
2267 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2268 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2269 UNPCKL_v_undef_shuffle_mask)),
2270 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2271 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2272 UNPCKL_v_undef_shuffle_mask)),
2273 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2276 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2277 let AddedComplexity = 10 in {
2278 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2279 UNPCKH_v_undef_shuffle_mask)),
2280 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2281 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2282 UNPCKH_v_undef_shuffle_mask)),
2283 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2284 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2285 UNPCKH_v_undef_shuffle_mask)),
2286 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2287 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2288 UNPCKH_v_undef_shuffle_mask)),
2289 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2292 let AddedComplexity = 15 in {
2293 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2294 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2295 MOVHP_shuffle_mask)),
2296 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2298 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2299 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2300 MOVHLPS_shuffle_mask)),
2301 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2303 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2304 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2305 MOVHLPS_v_undef_shuffle_mask)),
2306 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2307 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2308 MOVHLPS_v_undef_shuffle_mask)),
2309 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2312 let AddedComplexity = 20 in {
2313 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2314 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2315 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (loadv4f32 addr:$src2),
2316 MOVLP_shuffle_mask)),
2317 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2318 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (loadv2f64 addr:$src2),
2319 MOVLP_shuffle_mask)),
2320 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2321 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (loadv4f32 addr:$src2),
2322 MOVHP_shuffle_mask)),
2323 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2324 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (loadv2f64 addr:$src2),
2325 MOVHP_shuffle_mask)),
2326 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2328 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)),
2329 MOVLP_shuffle_mask)),
2330 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2331 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (loadv2i64 addr:$src2),
2332 MOVLP_shuffle_mask)),
2333 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2334 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)),
2335 MOVHP_shuffle_mask)),
2336 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2337 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (loadv2i64 addr:$src2),
2338 MOVLP_shuffle_mask)),
2339 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2342 let AddedComplexity = 15 in {
2343 // Setting the lowest element in the vector.
2344 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2345 MOVL_shuffle_mask)),
2346 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2347 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2348 MOVL_shuffle_mask)),
2349 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2351 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2352 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2353 MOVLP_shuffle_mask)),
2354 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2355 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2356 MOVLP_shuffle_mask)),
2357 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2360 // Set lowest element and zero upper elements.
2361 let AddedComplexity = 20 in
2362 def : Pat<(bc_v2i64 (vector_shuffle immAllZerosV,
2363 (v2f64 (scalar_to_vector (loadf64 addr:$src))),
2364 MOVL_shuffle_mask)),
2365 (MOVZQI2PQIrm addr:$src)>, Requires<[HasSSE2]>;
2367 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2368 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2369 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2370 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2371 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2372 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2373 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2374 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2375 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2376 Requires<[HasSSE2]>;
2377 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2378 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2379 Requires<[HasSSE2]>;
2380 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2381 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2382 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2383 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2384 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2385 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2386 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2387 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2388 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
2389 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2390 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
2391 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2392 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
2393 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2394 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
2395 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2397 // Some special case pandn patterns.
2398 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2400 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2401 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2403 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2404 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2406 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2408 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2409 (load addr:$src2))),
2410 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2411 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2412 (load addr:$src2))),
2413 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2414 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2415 (load addr:$src2))),
2416 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2419 def : Pat<(v4f32 (X86loadu addr:$src)), (MOVUPSrm addr:$src)>,
2420 Requires<[HasSSE1]>;