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 def V_SET0 : PSI<0x57, MRMInitReg, (ops VR128:$dst),
767 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
769 // FR32 to 128-bit vector conversion.
770 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (ops VR128:$dst, FR32:$src),
771 "movss {$src, $dst|$dst, $src}",
773 (v4f32 (scalar_to_vector FR32:$src)))]>;
774 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
775 "movss {$src, $dst|$dst, $src}",
777 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
779 // FIXME: may not be able to eliminate this movss with coalescing the src and
780 // dest register classes are different. We really want to write this pattern
782 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
784 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (ops FR32:$dst, VR128:$src),
785 "movss {$src, $dst|$dst, $src}",
786 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
788 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (ops f32mem:$dst, VR128:$src),
789 "movss {$src, $dst|$dst, $src}",
790 [(store (f32 (vector_extract (v4f32 VR128:$src),
791 (iPTR 0))), addr:$dst)]>;
794 // Move to lower bits of a VR128, leaving upper bits alone.
795 // Three operand (but two address) aliases.
796 let isTwoAddress = 1 in {
797 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
798 (ops VR128:$dst, VR128:$src1, FR32:$src2),
799 "movss {$src2, $dst|$dst, $src2}", []>;
801 let AddedComplexity = 15 in
802 def MOVLPSrr : SSI<0x10, MRMSrcReg,
803 (ops VR128:$dst, VR128:$src1, VR128:$src2),
804 "movss {$src2, $dst|$dst, $src2}",
806 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
807 MOVL_shuffle_mask)))]>;
810 // Move to lower bits of a VR128 and zeroing upper bits.
811 // Loading from memory automatically zeroing upper bits.
812 let AddedComplexity = 20 in
813 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
814 "movss {$src, $dst|$dst, $src}",
815 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV,
816 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
817 MOVL_shuffle_mask)))]>;
820 //===----------------------------------------------------------------------===//
822 //===----------------------------------------------------------------------===//
824 // SSE2 Instruction Templates:
826 // SDI - SSE2 instructions with XD prefix.
827 // PDI - SSE2 instructions with TB and OpSize prefixes.
828 // PDIi8 - SSE2 instructions with ImmT == Imm8 and TB and OpSize prefixes.
830 class SDI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
831 : I<o, F, ops, asm, pattern>, XD, Requires<[HasSSE2]>;
832 class PDI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
833 : I<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE2]>;
834 class PDIi8<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
835 : Ii8<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE2]>;
837 // Helpers for defining instructions that directly correspond to intrinsics.
838 multiclass SD_IntUnary<bits<8> o, string OpcodeStr, Intrinsic IntId> {
839 def r : SDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
840 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
841 [(set VR128:$dst, (v2f64 (IntId VR128:$src)))]>;
842 def m : SDI<o, MRMSrcMem, (ops VR128:$dst, sdmem:$src),
843 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
844 [(set VR128:$dst, (v2f64 (IntId sse_load_f64:$src)))]>;
848 def MOVSDrr : SDI<0x10, MRMSrcReg, (ops FR64:$dst, FR64:$src),
849 "movsd {$src, $dst|$dst, $src}", []>;
850 def MOVSDrm : SDI<0x10, MRMSrcMem, (ops FR64:$dst, f64mem:$src),
851 "movsd {$src, $dst|$dst, $src}",
852 [(set FR64:$dst, (loadf64 addr:$src))]>;
853 def MOVSDmr : SDI<0x11, MRMDestMem, (ops f64mem:$dst, FR64:$src),
854 "movsd {$src, $dst|$dst, $src}",
855 [(store FR64:$src, addr:$dst)]>;
857 def SQRTSDr : SDI<0x51, MRMSrcReg, (ops FR64:$dst, FR64:$src),
858 "sqrtsd {$src, $dst|$dst, $src}",
859 [(set FR64:$dst, (fsqrt FR64:$src))]>;
860 def SQRTSDm : SDI<0x51, MRMSrcMem, (ops FR64:$dst, f64mem:$src),
861 "sqrtsd {$src, $dst|$dst, $src}",
862 [(set FR64:$dst, (fsqrt (loadf64 addr:$src)))]>;
864 // Aliases to match intrinsics which expect XMM operand(s).
865 defm SQRTSD_Int : SD_IntUnary<0x51, "sqrtsd" , int_x86_sse2_sqrt_sd>;
867 // Conversion instructions
868 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (ops GR32:$dst, FR64:$src),
869 "cvttsd2si {$src, $dst|$dst, $src}",
870 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
871 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (ops GR32:$dst, f64mem:$src),
872 "cvttsd2si {$src, $dst|$dst, $src}",
873 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
874 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (ops FR32:$dst, FR64:$src),
875 "cvtsd2ss {$src, $dst|$dst, $src}",
876 [(set FR32:$dst, (fround FR64:$src))]>;
877 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (ops FR32:$dst, f64mem:$src),
878 "cvtsd2ss {$src, $dst|$dst, $src}",
879 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
880 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (ops FR64:$dst, GR32:$src),
881 "cvtsi2sd {$src, $dst|$dst, $src}",
882 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
883 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (ops FR64:$dst, i32mem:$src),
884 "cvtsi2sd {$src, $dst|$dst, $src}",
885 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
887 // SSE2 instructions with XS prefix
888 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (ops FR64:$dst, FR32:$src),
889 "cvtss2sd {$src, $dst|$dst, $src}",
890 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
892 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (ops FR64:$dst, f32mem:$src),
893 "cvtss2sd {$src, $dst|$dst, $src}",
894 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
897 // Match intrinsics which expect XMM operand(s).
898 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (ops GR32:$dst, VR128:$src),
899 "cvtsd2si {$src, $dst|$dst, $src}",
900 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
901 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (ops GR32:$dst, f128mem:$src),
902 "cvtsd2si {$src, $dst|$dst, $src}",
903 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
904 (load addr:$src)))]>;
906 // Aliases for intrinsics
907 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (ops GR32:$dst, VR128:$src),
908 "cvttsd2si {$src, $dst|$dst, $src}",
910 (int_x86_sse2_cvttsd2si VR128:$src))]>;
911 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (ops GR32:$dst, f128mem:$src),
912 "cvttsd2si {$src, $dst|$dst, $src}",
913 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
914 (load addr:$src)))]>;
916 // Comparison instructions
917 let isTwoAddress = 1 in {
918 def CMPSDrr : SDI<0xC2, MRMSrcReg,
919 (ops FR64:$dst, FR64:$src1, FR64:$src, SSECC:$cc),
920 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
921 def CMPSDrm : SDI<0xC2, MRMSrcMem,
922 (ops FR64:$dst, FR64:$src1, f64mem:$src, SSECC:$cc),
923 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
926 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (ops FR64:$src1, FR64:$src2),
927 "ucomisd {$src2, $src1|$src1, $src2}",
928 [(X86cmp FR64:$src1, FR64:$src2)]>;
929 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (ops FR64:$src1, f64mem:$src2),
930 "ucomisd {$src2, $src1|$src1, $src2}",
931 [(X86cmp FR64:$src1, (loadf64 addr:$src2))]>;
933 // Aliases to match intrinsics which expect XMM operand(s).
934 let isTwoAddress = 1 in {
935 def Int_CMPSDrr : SDI<0xC2, MRMSrcReg,
936 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
937 "cmp${cc}sd {$src, $dst|$dst, $src}",
938 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
939 VR128:$src, imm:$cc))]>;
940 def Int_CMPSDrm : SDI<0xC2, MRMSrcMem,
941 (ops VR128:$dst, VR128:$src1, f64mem:$src, SSECC:$cc),
942 "cmp${cc}sd {$src, $dst|$dst, $src}",
943 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
944 (load addr:$src), imm:$cc))]>;
947 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
948 "ucomisd {$src2, $src1|$src1, $src2}",
949 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
950 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
951 "ucomisd {$src2, $src1|$src1, $src2}",
952 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2))]>;
954 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
955 "comisd {$src2, $src1|$src1, $src2}",
956 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
957 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
958 "comisd {$src2, $src1|$src1, $src2}",
959 [(X86comi (v2f64 VR128:$src1), (load addr:$src2))]>;
961 // Aliases of packed SSE2 instructions for scalar use. These all have names that
964 // Alias instructions that map fld0 to pxor for sse.
965 def FsFLD0SD : I<0xEF, MRMInitReg, (ops FR64:$dst),
966 "pxor $dst, $dst", [(set FR64:$dst, fp64imm0)]>,
967 Requires<[HasSSE2]>, TB, OpSize;
969 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
971 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (ops FR64:$dst, FR64:$src),
972 "movapd {$src, $dst|$dst, $src}", []>;
974 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
976 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (ops FR64:$dst, f128mem:$src),
977 "movapd {$src, $dst|$dst, $src}",
978 [(set FR64:$dst, (X86loadpf64 addr:$src))]>;
980 // Alias bitwise logical operations using SSE logical ops on packed FP values.
981 let isTwoAddress = 1 in {
982 let isCommutable = 1 in {
983 def FsANDPDrr : PDI<0x54, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
984 "andpd {$src2, $dst|$dst, $src2}",
985 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
986 def FsORPDrr : PDI<0x56, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
987 "orpd {$src2, $dst|$dst, $src2}",
988 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
989 def FsXORPDrr : PDI<0x57, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
990 "xorpd {$src2, $dst|$dst, $src2}",
991 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
994 def FsANDPDrm : PDI<0x54, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
995 "andpd {$src2, $dst|$dst, $src2}",
996 [(set FR64:$dst, (X86fand FR64:$src1,
997 (X86loadpf64 addr:$src2)))]>;
998 def FsORPDrm : PDI<0x56, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
999 "orpd {$src2, $dst|$dst, $src2}",
1000 [(set FR64:$dst, (X86for FR64:$src1,
1001 (X86loadpf64 addr:$src2)))]>;
1002 def FsXORPDrm : PDI<0x57, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
1003 "xorpd {$src2, $dst|$dst, $src2}",
1004 [(set FR64:$dst, (X86fxor FR64:$src1,
1005 (X86loadpf64 addr:$src2)))]>;
1007 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1008 (ops FR64:$dst, FR64:$src1, FR64:$src2),
1009 "andnpd {$src2, $dst|$dst, $src2}", []>;
1010 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1011 (ops FR64:$dst, FR64:$src1, f128mem:$src2),
1012 "andnpd {$src2, $dst|$dst, $src2}", []>;
1015 /// scalar_sse2_fp_binop_rm - Scalar SSE2 binops come in three basic forms:
1017 /// 1. f64 - This comes in SSE2 form for doubles.
1018 /// 2. rr vs rm - They include a reg+reg form and a reg+mem form.
1020 /// In addition, scalar SSE ops have an intrinsic form. This form is unlike the
1021 /// normal form, in that they take an entire vector (instead of a scalar) and
1022 /// leave the top elements undefined. This adds another two variants of the
1023 /// above permutations, giving us 8 forms for 'instruction'.
1025 let isTwoAddress = 1 in {
1026 multiclass scalar_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1027 SDNode OpNode, Intrinsic F64Int,
1028 bit Commutable = 0> {
1029 // Scalar operation, reg+reg.
1030 def SDrr : SDI<opc, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
1031 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1032 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1033 let isCommutable = Commutable;
1036 // Scalar operation, reg+mem.
1037 def SDrm : SDI<opc, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f64mem:$src2),
1038 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1039 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1041 // Vector intrinsic operation, reg+reg.
1042 def SDrr_Int : SDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1043 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1044 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1045 let isCommutable = Commutable;
1048 // Vector intrinsic operation, reg+mem.
1049 def SDrm_Int : SDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, sdmem:$src2),
1050 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1051 [(set VR128:$dst, (F64Int VR128:$src1,
1052 sse_load_f64:$src2))]>;
1056 // Arithmetic instructions
1057 defm ADD : scalar_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1058 defm MUL : scalar_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1059 defm SUB : scalar_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1060 defm DIV : scalar_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1062 defm MAX : scalar_sse2_fp_binop_rm<0x5F, "max", X86fmax, int_x86_sse2_max_sd>;
1063 defm MIN : scalar_sse2_fp_binop_rm<0x5D, "min", X86fmin, int_x86_sse2_min_sd>;
1065 //===----------------------------------------------------------------------===//
1066 // SSE packed FP Instructions
1068 // Move Instructions
1069 def MOVAPDrr : PDI<0x28, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1070 "movapd {$src, $dst|$dst, $src}", []>;
1071 def MOVAPDrm : PDI<0x28, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1072 "movapd {$src, $dst|$dst, $src}",
1073 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1075 def MOVAPDmr : PDI<0x29, MRMDestMem, (ops f128mem:$dst, VR128:$src),
1076 "movapd {$src, $dst|$dst, $src}",
1077 [(store (v2f64 VR128:$src), addr:$dst)]>;
1079 def MOVUPDrr : PDI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1080 "movupd {$src, $dst|$dst, $src}", []>;
1081 def MOVUPDrm : PDI<0x10, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1082 "movupd {$src, $dst|$dst, $src}",
1083 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1084 def MOVUPDmr : PDI<0x11, MRMDestMem, (ops f128mem:$dst, VR128:$src),
1085 "movupd {$src, $dst|$dst, $src}",
1086 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1088 let isTwoAddress = 1 in {
1089 let AddedComplexity = 20 in {
1090 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1091 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
1092 "movlpd {$src2, $dst|$dst, $src2}",
1094 (v2f64 (vector_shuffle VR128:$src1,
1095 (scalar_to_vector (loadf64 addr:$src2)),
1096 MOVLP_shuffle_mask)))]>;
1097 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1098 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
1099 "movhpd {$src2, $dst|$dst, $src2}",
1101 (v2f64 (vector_shuffle VR128:$src1,
1102 (scalar_to_vector (loadf64 addr:$src2)),
1103 MOVHP_shuffle_mask)))]>;
1104 } // AddedComplexity
1107 def MOVLPDmr : PDI<0x13, MRMDestMem, (ops f64mem:$dst, VR128:$src),
1108 "movlpd {$src, $dst|$dst, $src}",
1109 [(store (f64 (vector_extract (v2f64 VR128:$src),
1110 (iPTR 0))), addr:$dst)]>;
1112 // v2f64 extract element 1 is always custom lowered to unpack high to low
1113 // and extract element 0 so the non-store version isn't too horrible.
1114 def MOVHPDmr : PDI<0x17, MRMDestMem, (ops f64mem:$dst, VR128:$src),
1115 "movhpd {$src, $dst|$dst, $src}",
1116 [(store (f64 (vector_extract
1117 (v2f64 (vector_shuffle VR128:$src, (undef),
1118 UNPCKH_shuffle_mask)), (iPTR 0))),
1121 // SSE2 instructions without OpSize prefix
1122 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1123 "cvtdq2ps {$src, $dst|$dst, $src}",
1124 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1125 TB, Requires<[HasSSE2]>;
1126 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1127 "cvtdq2ps {$src, $dst|$dst, $src}",
1128 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1129 (bitconvert (loadv2i64 addr:$src))))]>,
1130 TB, Requires<[HasSSE2]>;
1132 // SSE2 instructions with XS prefix
1133 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1134 "cvtdq2pd {$src, $dst|$dst, $src}",
1135 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1136 XS, Requires<[HasSSE2]>;
1137 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1138 "cvtdq2pd {$src, $dst|$dst, $src}",
1139 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1140 (bitconvert (loadv2i64 addr:$src))))]>,
1141 XS, Requires<[HasSSE2]>;
1143 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1144 "cvtps2dq {$src, $dst|$dst, $src}",
1145 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1146 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1147 "cvtps2dq {$src, $dst|$dst, $src}",
1148 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1149 (load addr:$src)))]>;
1150 // SSE2 packed instructions with XS prefix
1151 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1152 "cvttps2dq {$src, $dst|$dst, $src}",
1153 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1154 XS, Requires<[HasSSE2]>;
1155 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1156 "cvttps2dq {$src, $dst|$dst, $src}",
1157 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1158 (load addr:$src)))]>,
1159 XS, Requires<[HasSSE2]>;
1161 // SSE2 packed instructions with XD prefix
1162 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1163 "cvtpd2dq {$src, $dst|$dst, $src}",
1164 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1165 XD, Requires<[HasSSE2]>;
1166 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1167 "cvtpd2dq {$src, $dst|$dst, $src}",
1168 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1169 (load addr:$src)))]>,
1170 XD, Requires<[HasSSE2]>;
1172 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1173 "cvttpd2dq {$src, $dst|$dst, $src}",
1174 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1175 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1176 "cvttpd2dq {$src, $dst|$dst, $src}",
1177 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1178 (load addr:$src)))]>;
1180 // SSE2 instructions without OpSize prefix
1181 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1182 "cvtps2pd {$src, $dst|$dst, $src}",
1183 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1184 TB, Requires<[HasSSE2]>;
1185 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (ops VR128:$dst, f64mem:$src),
1186 "cvtps2pd {$src, $dst|$dst, $src}",
1187 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1188 (load addr:$src)))]>,
1189 TB, Requires<[HasSSE2]>;
1191 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1192 "cvtpd2ps {$src, $dst|$dst, $src}",
1193 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1194 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (ops VR128:$dst, f128mem:$src),
1195 "cvtpd2ps {$src, $dst|$dst, $src}",
1196 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1197 (load addr:$src)))]>;
1199 // Match intrinsics which expect XMM operand(s).
1200 // Aliases for intrinsics
1201 let isTwoAddress = 1 in {
1202 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1203 (ops VR128:$dst, VR128:$src1, GR32:$src2),
1204 "cvtsi2sd {$src2, $dst|$dst, $src2}",
1205 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1207 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1208 (ops VR128:$dst, VR128:$src1, i32mem:$src2),
1209 "cvtsi2sd {$src2, $dst|$dst, $src2}",
1210 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1211 (loadi32 addr:$src2)))]>;
1212 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1213 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1214 "cvtsd2ss {$src2, $dst|$dst, $src2}",
1215 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1217 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1218 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
1219 "cvtsd2ss {$src2, $dst|$dst, $src2}",
1220 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1221 (load addr:$src2)))]>;
1222 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1223 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1224 "cvtss2sd {$src2, $dst|$dst, $src2}",
1225 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1226 VR128:$src2))]>, XS,
1227 Requires<[HasSSE2]>;
1228 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1229 (ops VR128:$dst, VR128:$src1, f32mem:$src2),
1230 "cvtss2sd {$src2, $dst|$dst, $src2}",
1231 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1232 (load addr:$src2)))]>, XS,
1233 Requires<[HasSSE2]>;
1236 /// packed_sse2_fp_binop_rm - Packed SSE binops come in three basic forms:
1237 /// 1. v2f64 - This comes in SSE2 form for doubles.
1238 /// 2. rr vs rm - They include a reg+reg form and a ref+mem form.
1240 let isTwoAddress = 1 in {
1241 multiclass packed_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1242 SDNode OpNode, bit Commutable = 0> {
1243 // Packed operation, reg+reg.
1244 def PDrr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1245 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
1246 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1247 let isCommutable = Commutable;
1250 // Packed operation, reg+mem.
1251 def PDrm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1252 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
1253 [(set VR128:$dst, (OpNode VR128:$src1, (loadv2f64 addr:$src2)))]>;
1257 defm ADD : packed_sse2_fp_binop_rm<0x58, "add", fadd, 1>;
1258 defm MUL : packed_sse2_fp_binop_rm<0x59, "mul", fmul, 1>;
1259 defm DIV : packed_sse2_fp_binop_rm<0x5E, "div", fdiv>;
1260 defm SUB : packed_sse2_fp_binop_rm<0x5C, "sub", fsub>;
1264 class PD_Intr<bits<8> o, string OpcodeStr, Intrinsic IntId>
1265 : PDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1266 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
1267 [(set VR128:$dst, (IntId VR128:$src))]>;
1268 class PD_Intm<bits<8> o, string OpcodeStr, Intrinsic IntId>
1269 : PDI<o, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1270 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
1271 [(set VR128:$dst, (IntId (load addr:$src)))]>;
1273 class PD_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
1274 : PDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1275 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1276 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1277 class PD_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
1278 : PDI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
1279 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1280 [(set VR128:$dst, (IntId VR128:$src1, (load addr:$src2)))]>;
1282 def SQRTPDr : PD_Intr<0x51, "sqrtpd", int_x86_sse2_sqrt_pd>;
1283 def SQRTPDm : PD_Intm<0x51, "sqrtpd", int_x86_sse2_sqrt_pd>;
1285 let isTwoAddress = 1 in {
1286 let isCommutable = 1 in {
1287 def MAXPDrr : PD_Intrr<0x5F, "maxpd", int_x86_sse2_max_pd>;
1288 def MINPDrr : PD_Intrr<0x5D, "minpd", int_x86_sse2_min_pd>;
1291 def MAXPDrm : PD_Intrm<0x5F, "maxpd", int_x86_sse2_max_pd>;
1292 def MINPDrm : PD_Intrm<0x5D, "minpd", int_x86_sse2_min_pd>;
1296 let isTwoAddress = 1 in {
1297 let isCommutable = 1 in {
1298 def ANDPDrr : PDI<0x54, MRMSrcReg,
1299 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1300 "andpd {$src2, $dst|$dst, $src2}",
1302 (and (bc_v2i64 (v2f64 VR128:$src1)),
1303 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1304 def ORPDrr : PDI<0x56, MRMSrcReg,
1305 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1306 "orpd {$src2, $dst|$dst, $src2}",
1308 (or (bc_v2i64 (v2f64 VR128:$src1)),
1309 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1310 def XORPDrr : PDI<0x57, MRMSrcReg,
1311 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1312 "xorpd {$src2, $dst|$dst, $src2}",
1314 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1315 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1318 def ANDPDrm : PDI<0x54, MRMSrcMem,
1319 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1320 "andpd {$src2, $dst|$dst, $src2}",
1322 (and (bc_v2i64 (v2f64 VR128:$src1)),
1323 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1324 def ORPDrm : PDI<0x56, MRMSrcMem,
1325 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1326 "orpd {$src2, $dst|$dst, $src2}",
1328 (or (bc_v2i64 (v2f64 VR128:$src1)),
1329 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1330 def XORPDrm : PDI<0x57, MRMSrcMem,
1331 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1332 "xorpd {$src2, $dst|$dst, $src2}",
1334 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1335 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1336 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1337 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1338 "andnpd {$src2, $dst|$dst, $src2}",
1340 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1341 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1342 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1343 (ops VR128:$dst, VR128:$src1,f128mem:$src2),
1344 "andnpd {$src2, $dst|$dst, $src2}",
1346 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1347 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1350 let isTwoAddress = 1 in {
1351 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1352 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
1353 "cmp${cc}pd {$src, $dst|$dst, $src}",
1354 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1355 VR128:$src, imm:$cc))]>;
1356 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1357 (ops VR128:$dst, VR128:$src1, f128mem:$src, SSECC:$cc),
1358 "cmp${cc}pd {$src, $dst|$dst, $src}",
1359 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1360 (load addr:$src), imm:$cc))]>;
1363 // Shuffle and unpack instructions
1364 let isTwoAddress = 1 in {
1365 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1366 (ops VR128:$dst, VR128:$src1, VR128:$src2, i8imm:$src3),
1367 "shufpd {$src3, $src2, $dst|$dst, $src2, $src3}",
1368 [(set VR128:$dst, (v2f64 (vector_shuffle
1369 VR128:$src1, VR128:$src2,
1370 SHUFP_shuffle_mask:$src3)))]>;
1371 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1372 (ops VR128:$dst, VR128:$src1,
1373 f128mem:$src2, i8imm:$src3),
1374 "shufpd {$src3, $src2, $dst|$dst, $src2, $src3}",
1376 (v2f64 (vector_shuffle
1377 VR128:$src1, (load addr:$src2),
1378 SHUFP_shuffle_mask:$src3)))]>;
1380 let AddedComplexity = 10 in {
1381 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1382 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1383 "unpckhpd {$src2, $dst|$dst, $src2}",
1385 (v2f64 (vector_shuffle
1386 VR128:$src1, VR128:$src2,
1387 UNPCKH_shuffle_mask)))]>;
1388 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1389 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1390 "unpckhpd {$src2, $dst|$dst, $src2}",
1392 (v2f64 (vector_shuffle
1393 VR128:$src1, (load addr:$src2),
1394 UNPCKH_shuffle_mask)))]>;
1396 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1397 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1398 "unpcklpd {$src2, $dst|$dst, $src2}",
1400 (v2f64 (vector_shuffle
1401 VR128:$src1, VR128:$src2,
1402 UNPCKL_shuffle_mask)))]>;
1403 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1404 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1405 "unpcklpd {$src2, $dst|$dst, $src2}",
1407 (v2f64 (vector_shuffle
1408 VR128:$src1, (load addr:$src2),
1409 UNPCKL_shuffle_mask)))]>;
1410 } // AddedComplexity
1414 //===----------------------------------------------------------------------===//
1415 // SSE integer instructions
1417 // Move Instructions
1418 def MOVDQArr : PDI<0x6F, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1419 "movdqa {$src, $dst|$dst, $src}", []>;
1420 def MOVDQArm : PDI<0x6F, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1421 "movdqa {$src, $dst|$dst, $src}",
1422 [(set VR128:$dst, (loadv2i64 addr:$src))]>;
1423 def MOVDQAmr : PDI<0x7F, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1424 "movdqa {$src, $dst|$dst, $src}",
1425 [(store (v2i64 VR128:$src), addr:$dst)]>;
1426 def MOVDQUrm : I<0x6F, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1427 "movdqu {$src, $dst|$dst, $src}",
1428 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1429 XS, Requires<[HasSSE2]>;
1430 def MOVDQUmr : I<0x7F, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1431 "movdqu {$src, $dst|$dst, $src}",
1432 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1433 XS, Requires<[HasSSE2]>;
1436 let isTwoAddress = 1 in {
1438 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1439 bit Commutable = 0> {
1440 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1441 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1442 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1443 let isCommutable = Commutable;
1445 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1446 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1447 [(set VR128:$dst, (IntId VR128:$src1,
1448 (bitconvert (loadv2i64 addr:$src2))))]>;
1451 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1452 string OpcodeStr, Intrinsic IntId> {
1453 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1454 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1455 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1456 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1457 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1458 [(set VR128:$dst, (IntId VR128:$src1,
1459 (bitconvert (loadv2i64 addr:$src2))))]>;
1460 def ri : PDIi8<opc2, ImmForm, (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1461 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1462 [(set VR128:$dst, (IntId VR128:$src1,
1463 (scalar_to_vector (i32 imm:$src2))))]>;
1467 /// PDI_binop_rm - Simple SSE2 binary operator.
1468 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1469 ValueType OpVT, bit Commutable = 0> {
1470 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1471 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1472 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1473 let isCommutable = Commutable;
1475 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1476 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1477 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1478 (bitconvert (loadv2i64 addr:$src2)))))]>;
1481 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1483 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1484 /// to collapse (bitconvert VT to VT) into its operand.
1486 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1487 bit Commutable = 0> {
1488 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1489 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1490 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1491 let isCommutable = Commutable;
1493 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1494 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1495 [(set VR128:$dst, (OpNode VR128:$src1,(loadv2i64 addr:$src2)))]>;
1500 // 128-bit Integer Arithmetic
1502 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1503 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1504 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1505 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1507 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1508 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1509 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1510 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1512 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1513 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1514 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1515 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1517 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1518 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1519 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1520 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1522 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1524 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1525 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1526 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1528 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1530 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1531 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1534 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1535 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1536 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1537 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1538 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1541 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw", int_x86_sse2_psll_w>;
1542 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld", int_x86_sse2_psll_d>;
1543 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq", int_x86_sse2_psll_q>;
1545 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw", int_x86_sse2_psrl_w>;
1546 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld", int_x86_sse2_psrl_d>;
1547 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq", int_x86_sse2_psrl_q>;
1549 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw", int_x86_sse2_psra_w>;
1550 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad", int_x86_sse2_psra_d>;
1551 // PSRAQ doesn't exist in SSE[1-3].
1553 // 128-bit logical shifts.
1554 let isTwoAddress = 1 in {
1555 def PSLLDQri : PDIi8<0x73, MRM7r,
1556 (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1557 "pslldq {$src2, $dst|$dst, $src2}", []>;
1558 def PSRLDQri : PDIi8<0x73, MRM3r,
1559 (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1560 "psrldq {$src2, $dst|$dst, $src2}", []>;
1561 // PSRADQri doesn't exist in SSE[1-3].
1564 let Predicates = [HasSSE2] in {
1565 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1566 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1567 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1568 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1569 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1570 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1574 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1575 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1576 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1578 let isTwoAddress = 1 in {
1579 def PANDNrr : PDI<0xDF, MRMSrcReg,
1580 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1581 "pandn {$src2, $dst|$dst, $src2}",
1582 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1585 def PANDNrm : PDI<0xDF, MRMSrcMem,
1586 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1587 "pandn {$src2, $dst|$dst, $src2}",
1588 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1589 (load addr:$src2))))]>;
1592 // SSE2 Integer comparison
1593 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1594 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1595 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1596 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1597 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1598 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1600 // Pack instructions
1601 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1602 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1603 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1605 // Shuffle and unpack instructions
1606 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1607 (ops VR128:$dst, VR128:$src1, i8imm:$src2),
1608 "pshufd {$src2, $src1, $dst|$dst, $src1, $src2}",
1609 [(set VR128:$dst, (v4i32 (vector_shuffle
1610 VR128:$src1, (undef),
1611 PSHUFD_shuffle_mask:$src2)))]>;
1612 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1613 (ops VR128:$dst, i128mem:$src1, i8imm:$src2),
1614 "pshufd {$src2, $src1, $dst|$dst, $src1, $src2}",
1615 [(set VR128:$dst, (v4i32 (vector_shuffle
1616 (bc_v4i32(loadv2i64 addr:$src1)),
1618 PSHUFD_shuffle_mask:$src2)))]>;
1620 // SSE2 with ImmT == Imm8 and XS prefix.
1621 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1622 (ops VR128:$dst, VR128:$src1, i8imm:$src2),
1623 "pshufhw {$src2, $src1, $dst|$dst, $src1, $src2}",
1624 [(set VR128:$dst, (v8i16 (vector_shuffle
1625 VR128:$src1, (undef),
1626 PSHUFHW_shuffle_mask:$src2)))]>,
1627 XS, Requires<[HasSSE2]>;
1628 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
1629 (ops VR128:$dst, i128mem:$src1, i8imm:$src2),
1630 "pshufhw {$src2, $src1, $dst|$dst, $src1, $src2}",
1631 [(set VR128:$dst, (v8i16 (vector_shuffle
1632 (bc_v8i16 (loadv2i64 addr:$src1)),
1634 PSHUFHW_shuffle_mask:$src2)))]>,
1635 XS, Requires<[HasSSE2]>;
1637 // SSE2 with ImmT == Imm8 and XD prefix.
1638 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
1639 (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1640 "pshuflw {$src2, $src1, $dst|$dst, $src1, $src2}",
1641 [(set VR128:$dst, (v8i16 (vector_shuffle
1642 VR128:$src1, (undef),
1643 PSHUFLW_shuffle_mask:$src2)))]>,
1644 XD, Requires<[HasSSE2]>;
1645 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
1646 (ops VR128:$dst, i128mem:$src1, i32i8imm:$src2),
1647 "pshuflw {$src2, $src1, $dst|$dst, $src1, $src2}",
1648 [(set VR128:$dst, (v8i16 (vector_shuffle
1649 (bc_v8i16 (loadv2i64 addr:$src1)),
1651 PSHUFLW_shuffle_mask:$src2)))]>,
1652 XD, Requires<[HasSSE2]>;
1655 let isTwoAddress = 1 in {
1656 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
1657 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1658 "punpcklbw {$src2, $dst|$dst, $src2}",
1660 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1661 UNPCKL_shuffle_mask)))]>;
1662 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
1663 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1664 "punpcklbw {$src2, $dst|$dst, $src2}",
1666 (v16i8 (vector_shuffle VR128:$src1,
1667 (bc_v16i8 (loadv2i64 addr:$src2)),
1668 UNPCKL_shuffle_mask)))]>;
1669 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
1670 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1671 "punpcklwd {$src2, $dst|$dst, $src2}",
1673 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1674 UNPCKL_shuffle_mask)))]>;
1675 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
1676 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1677 "punpcklwd {$src2, $dst|$dst, $src2}",
1679 (v8i16 (vector_shuffle VR128:$src1,
1680 (bc_v8i16 (loadv2i64 addr:$src2)),
1681 UNPCKL_shuffle_mask)))]>;
1682 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
1683 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1684 "punpckldq {$src2, $dst|$dst, $src2}",
1686 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1687 UNPCKL_shuffle_mask)))]>;
1688 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
1689 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1690 "punpckldq {$src2, $dst|$dst, $src2}",
1692 (v4i32 (vector_shuffle VR128:$src1,
1693 (bc_v4i32 (loadv2i64 addr:$src2)),
1694 UNPCKL_shuffle_mask)))]>;
1695 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
1696 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1697 "punpcklqdq {$src2, $dst|$dst, $src2}",
1699 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1700 UNPCKL_shuffle_mask)))]>;
1701 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
1702 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1703 "punpcklqdq {$src2, $dst|$dst, $src2}",
1705 (v2i64 (vector_shuffle VR128:$src1,
1706 (loadv2i64 addr:$src2),
1707 UNPCKL_shuffle_mask)))]>;
1709 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
1710 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1711 "punpckhbw {$src2, $dst|$dst, $src2}",
1713 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1714 UNPCKH_shuffle_mask)))]>;
1715 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
1716 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1717 "punpckhbw {$src2, $dst|$dst, $src2}",
1719 (v16i8 (vector_shuffle VR128:$src1,
1720 (bc_v16i8 (loadv2i64 addr:$src2)),
1721 UNPCKH_shuffle_mask)))]>;
1722 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
1723 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1724 "punpckhwd {$src2, $dst|$dst, $src2}",
1726 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1727 UNPCKH_shuffle_mask)))]>;
1728 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
1729 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1730 "punpckhwd {$src2, $dst|$dst, $src2}",
1732 (v8i16 (vector_shuffle VR128:$src1,
1733 (bc_v8i16 (loadv2i64 addr:$src2)),
1734 UNPCKH_shuffle_mask)))]>;
1735 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
1736 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1737 "punpckhdq {$src2, $dst|$dst, $src2}",
1739 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1740 UNPCKH_shuffle_mask)))]>;
1741 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
1742 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1743 "punpckhdq {$src2, $dst|$dst, $src2}",
1745 (v4i32 (vector_shuffle VR128:$src1,
1746 (bc_v4i32 (loadv2i64 addr:$src2)),
1747 UNPCKH_shuffle_mask)))]>;
1748 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
1749 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1750 "punpckhqdq {$src2, $dst|$dst, $src2}",
1752 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1753 UNPCKH_shuffle_mask)))]>;
1754 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
1755 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1756 "punpckhqdq {$src2, $dst|$dst, $src2}",
1758 (v2i64 (vector_shuffle VR128:$src1,
1759 (loadv2i64 addr:$src2),
1760 UNPCKH_shuffle_mask)))]>;
1764 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
1765 (ops GR32:$dst, VR128:$src1, i32i8imm:$src2),
1766 "pextrw {$src2, $src1, $dst|$dst, $src1, $src2}",
1767 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
1768 (iPTR imm:$src2)))]>;
1769 let isTwoAddress = 1 in {
1770 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
1771 (ops VR128:$dst, VR128:$src1,
1772 GR32:$src2, i32i8imm:$src3),
1773 "pinsrw {$src3, $src2, $dst|$dst, $src2, $src3}",
1775 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
1776 GR32:$src2, (iPTR imm:$src3))))]>;
1777 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
1778 (ops VR128:$dst, VR128:$src1,
1779 i16mem:$src2, i32i8imm:$src3),
1780 "pinsrw {$src3, $src2, $dst|$dst, $src2, $src3}",
1782 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
1783 (i32 (anyext (loadi16 addr:$src2))),
1784 (iPTR imm:$src3))))]>;
1788 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (ops GR32:$dst, VR128:$src),
1789 "pmovmskb {$src, $dst|$dst, $src}",
1790 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
1792 // Conditional store
1793 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (ops VR128:$src, VR128:$mask),
1794 "maskmovdqu {$mask, $src|$src, $mask}",
1795 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>,
1798 // Non-temporal stores
1799 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1800 "movntpd {$src, $dst|$dst, $src}",
1801 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
1802 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (ops f128mem:$dst, VR128:$src),
1803 "movntdq {$src, $dst|$dst, $src}",
1804 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
1805 def MOVNTImr : I<0xC3, MRMDestMem, (ops i32mem:$dst, GR32:$src),
1806 "movnti {$src, $dst|$dst, $src}",
1807 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
1808 TB, Requires<[HasSSE2]>;
1811 def CLFLUSH : I<0xAE, MRM7m, (ops i8mem:$src),
1812 "clflush $src", [(int_x86_sse2_clflush addr:$src)]>,
1813 TB, Requires<[HasSSE2]>;
1815 // Load, store, and memory fence
1816 def LFENCE : I<0xAE, MRM5m, (ops),
1817 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
1818 def MFENCE : I<0xAE, MRM6m, (ops),
1819 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
1822 // Alias instructions that map zero vector to pxor / xorp* for sse.
1823 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
1824 def V_SETALLONES : PDI<0x76, MRMInitReg, (ops VR128:$dst),
1825 "pcmpeqd $dst, $dst",
1826 [(set VR128:$dst, (v2f64 immAllOnesV))]>;
1828 // FR64 to 128-bit vector conversion.
1829 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (ops VR128:$dst, FR64:$src),
1830 "movsd {$src, $dst|$dst, $src}",
1832 (v2f64 (scalar_to_vector FR64:$src)))]>;
1833 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1834 "movsd {$src, $dst|$dst, $src}",
1836 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
1838 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (ops VR128:$dst, GR32:$src),
1839 "movd {$src, $dst|$dst, $src}",
1841 (v4i32 (scalar_to_vector GR32:$src)))]>;
1842 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (ops VR128:$dst, i32mem:$src),
1843 "movd {$src, $dst|$dst, $src}",
1845 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
1847 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (ops FR32:$dst, GR32:$src),
1848 "movd {$src, $dst|$dst, $src}",
1849 [(set FR32:$dst, (bitconvert GR32:$src))]>;
1851 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (ops FR32:$dst, i32mem:$src),
1852 "movd {$src, $dst|$dst, $src}",
1853 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
1855 // SSE2 instructions with XS prefix
1856 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1857 "movq {$src, $dst|$dst, $src}",
1859 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
1860 Requires<[HasSSE2]>;
1861 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (ops i64mem:$dst, VR128:$src),
1862 "movq {$src, $dst|$dst, $src}",
1863 [(store (i64 (vector_extract (v2i64 VR128:$src),
1864 (iPTR 0))), addr:$dst)]>;
1866 // FIXME: may not be able to eliminate this movss with coalescing the src and
1867 // dest register classes are different. We really want to write this pattern
1869 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1870 // (f32 FR32:$src)>;
1871 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (ops FR64:$dst, VR128:$src),
1872 "movsd {$src, $dst|$dst, $src}",
1873 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
1875 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (ops f64mem:$dst, VR128:$src),
1876 "movsd {$src, $dst|$dst, $src}",
1877 [(store (f64 (vector_extract (v2f64 VR128:$src),
1878 (iPTR 0))), addr:$dst)]>;
1879 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (ops GR32:$dst, VR128:$src),
1880 "movd {$src, $dst|$dst, $src}",
1881 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
1883 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (ops i32mem:$dst, VR128:$src),
1884 "movd {$src, $dst|$dst, $src}",
1885 [(store (i32 (vector_extract (v4i32 VR128:$src),
1886 (iPTR 0))), addr:$dst)]>;
1888 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (ops GR32:$dst, FR32:$src),
1889 "movd {$src, $dst|$dst, $src}",
1890 [(set GR32:$dst, (bitconvert FR32:$src))]>;
1891 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (ops i32mem:$dst, FR32:$src),
1892 "movd {$src, $dst|$dst, $src}",
1893 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
1896 // Move to lower bits of a VR128, leaving upper bits alone.
1897 // Three operand (but two address) aliases.
1898 let isTwoAddress = 1 in {
1899 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
1900 (ops VR128:$dst, VR128:$src1, FR64:$src2),
1901 "movsd {$src2, $dst|$dst, $src2}", []>;
1903 let AddedComplexity = 15 in
1904 def MOVLPDrr : SDI<0x10, MRMSrcReg,
1905 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1906 "movsd {$src2, $dst|$dst, $src2}",
1908 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
1909 MOVL_shuffle_mask)))]>;
1912 // Store / copy lower 64-bits of a XMM register.
1913 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (ops i64mem:$dst, VR128:$src),
1914 "movq {$src, $dst|$dst, $src}",
1915 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
1917 // Move to lower bits of a VR128 and zeroing upper bits.
1918 // Loading from memory automatically zeroing upper bits.
1919 let AddedComplexity = 20 in
1920 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1921 "movsd {$src, $dst|$dst, $src}",
1923 (v2f64 (vector_shuffle immAllZerosV,
1924 (v2f64 (scalar_to_vector
1925 (loadf64 addr:$src))),
1926 MOVL_shuffle_mask)))]>;
1928 let AddedComplexity = 15 in
1929 // movd / movq to XMM register zero-extends
1930 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (ops VR128:$dst, GR32:$src),
1931 "movd {$src, $dst|$dst, $src}",
1933 (v4i32 (vector_shuffle immAllZerosV,
1934 (v4i32 (scalar_to_vector GR32:$src)),
1935 MOVL_shuffle_mask)))]>;
1936 let AddedComplexity = 20 in
1937 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (ops VR128:$dst, i32mem:$src),
1938 "movd {$src, $dst|$dst, $src}",
1940 (v4i32 (vector_shuffle immAllZerosV,
1941 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
1942 MOVL_shuffle_mask)))]>;
1944 // Moving from XMM to XMM but still clear upper 64 bits.
1945 let AddedComplexity = 15 in
1946 def MOVZQI2PQIrr : I<0x7E, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1947 "movq {$src, $dst|$dst, $src}",
1948 [(set VR128:$dst, (int_x86_sse2_movl_dq VR128:$src))]>,
1949 XS, Requires<[HasSSE2]>;
1950 let AddedComplexity = 20 in
1951 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1952 "movq {$src, $dst|$dst, $src}",
1953 [(set VR128:$dst, (int_x86_sse2_movl_dq
1954 (bitconvert (loadv2i64 addr:$src))))]>,
1955 XS, Requires<[HasSSE2]>;
1958 //===----------------------------------------------------------------------===//
1959 // SSE3 Instructions
1960 //===----------------------------------------------------------------------===//
1962 // SSE3 Instruction Templates:
1964 // S3I - SSE3 instructions with TB and OpSize prefixes.
1965 // S3SI - SSE3 instructions with XS prefix.
1966 // S3DI - SSE3 instructions with XD prefix.
1968 class S3SI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
1969 : I<o, F, ops, asm, pattern>, XS, Requires<[HasSSE3]>;
1970 class S3DI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
1971 : I<o, F, ops, asm, pattern>, XD, Requires<[HasSSE3]>;
1972 class S3I<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
1973 : I<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE3]>;
1975 // Move Instructions
1976 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1977 "movshdup {$src, $dst|$dst, $src}",
1978 [(set VR128:$dst, (v4f32 (vector_shuffle
1979 VR128:$src, (undef),
1980 MOVSHDUP_shuffle_mask)))]>;
1981 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1982 "movshdup {$src, $dst|$dst, $src}",
1983 [(set VR128:$dst, (v4f32 (vector_shuffle
1984 (loadv4f32 addr:$src), (undef),
1985 MOVSHDUP_shuffle_mask)))]>;
1987 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1988 "movsldup {$src, $dst|$dst, $src}",
1989 [(set VR128:$dst, (v4f32 (vector_shuffle
1990 VR128:$src, (undef),
1991 MOVSLDUP_shuffle_mask)))]>;
1992 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1993 "movsldup {$src, $dst|$dst, $src}",
1994 [(set VR128:$dst, (v4f32 (vector_shuffle
1995 (loadv4f32 addr:$src), (undef),
1996 MOVSLDUP_shuffle_mask)))]>;
1998 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1999 "movddup {$src, $dst|$dst, $src}",
2000 [(set VR128:$dst, (v2f64 (vector_shuffle
2001 VR128:$src, (undef),
2002 SSE_splat_lo_mask)))]>;
2003 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
2004 "movddup {$src, $dst|$dst, $src}",
2006 (v2f64 (vector_shuffle
2007 (scalar_to_vector (loadf64 addr:$src)),
2009 SSE_splat_lo_mask)))]>;
2012 let isTwoAddress = 1 in {
2013 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2014 (ops VR128:$dst, VR128:$src1, VR128:$src2),
2015 "addsubps {$src2, $dst|$dst, $src2}",
2016 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2018 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2019 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
2020 "addsubps {$src2, $dst|$dst, $src2}",
2021 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2022 (load addr:$src2)))]>;
2023 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2024 (ops VR128:$dst, VR128:$src1, VR128:$src2),
2025 "addsubpd {$src2, $dst|$dst, $src2}",
2026 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2028 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2029 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
2030 "addsubpd {$src2, $dst|$dst, $src2}",
2031 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2032 (load addr:$src2)))]>;
2035 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
2036 "lddqu {$src, $dst|$dst, $src}",
2037 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2040 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2041 : S3DI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
2042 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2043 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2044 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2045 : S3DI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
2046 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2047 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
2048 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2049 : S3I<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
2050 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2051 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2052 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2053 : S3I<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
2054 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2055 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
2057 let isTwoAddress = 1 in {
2058 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2059 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2060 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2061 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2062 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2063 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2064 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2065 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2068 // Thread synchronization
2069 def MONITOR : I<0xC8, RawFrm, (ops), "monitor",
2070 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2071 def MWAIT : I<0xC9, RawFrm, (ops), "mwait",
2072 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2074 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2075 let AddedComplexity = 15 in
2076 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2077 MOVSHDUP_shuffle_mask)),
2078 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2079 let AddedComplexity = 20 in
2080 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (loadv2i64 addr:$src)), (undef),
2081 MOVSHDUP_shuffle_mask)),
2082 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2084 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2085 let AddedComplexity = 15 in
2086 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2087 MOVSLDUP_shuffle_mask)),
2088 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2089 let AddedComplexity = 20 in
2090 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (loadv2i64 addr:$src)), (undef),
2091 MOVSLDUP_shuffle_mask)),
2092 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2094 //===----------------------------------------------------------------------===//
2095 // SSSE3 Instructions
2096 //===----------------------------------------------------------------------===//
2098 // SSE3 Instruction Templates:
2100 // SS38I - SSSE3 instructions with T8 and OpSize prefixes.
2101 // SS3AI - SSSE3 instructions with TA and OpSize prefixes.
2103 class SS38I<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
2104 : I<o, F, ops, asm, pattern>, T8, OpSize, Requires<[HasSSSE3]>;
2105 class SS3AI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
2106 : I<o, F, ops, asm, pattern>, TA, OpSize, Requires<[HasSSSE3]>;
2108 /// SS3I_binop_rm_int - Simple SSSE3 binary operatr whose type is v2i64.
2109 let isTwoAddress = 1 in {
2110 multiclass SS3I_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
2111 bit Commutable = 0> {
2112 def rr : SS38I<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
2113 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2114 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
2115 let isCommutable = Commutable;
2117 def rm : SS38I<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
2118 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2121 (bitconvert (loadv2i64 addr:$src2))))]>;
2125 defm PMULHRSW128 : SS3I_binop_rm_int<0x0B, "pmulhrsw",
2126 int_x86_ssse3_pmulhrsw_128, 1>;
2128 //===----------------------------------------------------------------------===//
2129 // Non-Instruction Patterns
2130 //===----------------------------------------------------------------------===//
2132 // 128-bit vector undef's.
2133 def : Pat<(v2f64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2134 def : Pat<(v16i8 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2135 def : Pat<(v8i16 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2136 def : Pat<(v4i32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2137 def : Pat<(v2i64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2139 // 128-bit vector all zero's.
2140 def : Pat<(v16i8 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2141 def : Pat<(v8i16 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2142 def : Pat<(v4i32 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2143 def : Pat<(v2i64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2144 def : Pat<(v2f64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2146 // 128-bit vector all one's.
2147 def : Pat<(v16i8 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2148 def : Pat<(v8i16 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2149 def : Pat<(v4i32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2150 def : Pat<(v2i64 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2151 def : Pat<(v4f32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE1]>;
2153 // Store 128-bit integer vector values.
2154 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
2155 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2156 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
2157 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2158 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
2159 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2161 // Scalar to v8i16 / v16i8. The source may be a GR32, but only the lower 8 or
2163 def : Pat<(v8i16 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2164 Requires<[HasSSE2]>;
2165 def : Pat<(v16i8 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2166 Requires<[HasSSE2]>;
2169 let Predicates = [HasSSE2] in {
2170 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2171 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2172 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2173 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2174 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2175 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2176 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2177 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2178 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2179 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2180 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2181 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2182 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2183 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2184 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2185 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2186 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2187 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2188 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2189 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2190 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2191 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2192 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2193 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2194 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2195 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2196 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2197 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2198 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2199 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2202 // Move scalar to XMM zero-extended
2203 // movd to XMM register zero-extends
2204 let AddedComplexity = 15 in {
2205 def : Pat<(v8i16 (vector_shuffle immAllZerosV,
2206 (v8i16 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
2207 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
2208 def : Pat<(v16i8 (vector_shuffle immAllZerosV,
2209 (v16i8 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
2210 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
2211 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2212 def : Pat<(v2f64 (vector_shuffle immAllZerosV,
2213 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
2214 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2215 def : Pat<(v4f32 (vector_shuffle immAllZerosV,
2216 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
2217 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
2220 // Splat v2f64 / v2i64
2221 let AddedComplexity = 10 in {
2222 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2223 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2224 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2225 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2226 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2227 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2228 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2229 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2233 def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SSE_splat_mask:$sm),
2234 (SHUFPSrri VR128:$src, VR128:$src, SSE_splat_mask:$sm)>,
2235 Requires<[HasSSE1]>;
2237 // Special unary SHUFPSrri case.
2238 // FIXME: when we want non two-address code, then we should use PSHUFD?
2239 def : Pat<(vector_shuffle (v4f32 VR128:$src1), (undef),
2240 SHUFP_unary_shuffle_mask:$sm),
2241 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2242 Requires<[HasSSE1]>;
2243 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2244 def : Pat<(vector_shuffle (loadv4f32 addr:$src1), (undef),
2245 SHUFP_unary_shuffle_mask:$sm),
2246 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2247 Requires<[HasSSE2]>;
2248 // Special binary v4i32 shuffle cases with SHUFPS.
2249 def : Pat<(vector_shuffle (v4i32 VR128:$src1), (v4i32 VR128:$src2),
2250 PSHUFD_binary_shuffle_mask:$sm),
2251 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2252 Requires<[HasSSE2]>;
2253 def : Pat<(vector_shuffle (v4i32 VR128:$src1),
2254 (bc_v4i32 (loadv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm),
2255 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2256 Requires<[HasSSE2]>;
2258 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2259 let AddedComplexity = 10 in {
2260 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2261 UNPCKL_v_undef_shuffle_mask)),
2262 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2263 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2264 UNPCKL_v_undef_shuffle_mask)),
2265 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2266 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2267 UNPCKL_v_undef_shuffle_mask)),
2268 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2269 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2270 UNPCKL_v_undef_shuffle_mask)),
2271 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2274 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2275 let AddedComplexity = 10 in {
2276 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2277 UNPCKH_v_undef_shuffle_mask)),
2278 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2279 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2280 UNPCKH_v_undef_shuffle_mask)),
2281 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2282 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2283 UNPCKH_v_undef_shuffle_mask)),
2284 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2285 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2286 UNPCKH_v_undef_shuffle_mask)),
2287 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2290 let AddedComplexity = 15 in {
2291 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2292 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2293 MOVHP_shuffle_mask)),
2294 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2296 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2297 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2298 MOVHLPS_shuffle_mask)),
2299 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2301 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2302 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2303 MOVHLPS_v_undef_shuffle_mask)),
2304 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2305 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2306 MOVHLPS_v_undef_shuffle_mask)),
2307 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2310 let AddedComplexity = 20 in {
2311 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2312 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2313 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (loadv4f32 addr:$src2),
2314 MOVLP_shuffle_mask)),
2315 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2316 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (loadv2f64 addr:$src2),
2317 MOVLP_shuffle_mask)),
2318 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2319 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (loadv4f32 addr:$src2),
2320 MOVHP_shuffle_mask)),
2321 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2322 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (loadv2f64 addr:$src2),
2323 MOVHP_shuffle_mask)),
2324 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2326 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)),
2327 MOVLP_shuffle_mask)),
2328 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2329 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (loadv2i64 addr:$src2),
2330 MOVLP_shuffle_mask)),
2331 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2332 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)),
2333 MOVHP_shuffle_mask)),
2334 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2335 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (loadv2i64 addr:$src2),
2336 MOVLP_shuffle_mask)),
2337 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2340 let AddedComplexity = 15 in {
2341 // Setting the lowest element in the vector.
2342 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2343 MOVL_shuffle_mask)),
2344 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2345 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2346 MOVL_shuffle_mask)),
2347 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2349 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2350 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2351 MOVLP_shuffle_mask)),
2352 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2353 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2354 MOVLP_shuffle_mask)),
2355 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2358 // Set lowest element and zero upper elements.
2359 let AddedComplexity = 20 in
2360 def : Pat<(bc_v2i64 (vector_shuffle immAllZerosV,
2361 (v2f64 (scalar_to_vector (loadf64 addr:$src))),
2362 MOVL_shuffle_mask)),
2363 (MOVZQI2PQIrm addr:$src)>, Requires<[HasSSE2]>;
2365 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2366 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2367 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2368 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2369 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2370 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2371 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2372 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2373 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2374 Requires<[HasSSE2]>;
2375 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2376 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2377 Requires<[HasSSE2]>;
2378 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2379 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2380 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2381 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2382 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2383 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2384 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2385 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2386 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
2387 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2388 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
2389 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2390 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
2391 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2392 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
2393 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2395 // Some special case pandn patterns.
2396 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2398 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2399 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2401 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2402 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2404 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2406 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2407 (load addr:$src2))),
2408 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2409 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2410 (load addr:$src2))),
2411 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2412 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2413 (load addr:$src2))),
2414 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2417 def : Pat<(v4f32 (X86loadu addr:$src)), (MOVUPSrm addr:$src)>,
2418 Requires<[HasSSE1]>;