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 loadv2i32 : PatFrag<(ops node:$ptr), (v2i32 (load node:$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 PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
174 return X86::isPSHUFDMask(N);
175 }], SHUFFLE_get_shuf_imm>;
177 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
178 return X86::isPSHUFHWMask(N);
179 }], SHUFFLE_get_pshufhw_imm>;
181 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
182 return X86::isPSHUFLWMask(N);
183 }], SHUFFLE_get_pshuflw_imm>;
185 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
186 return X86::isPSHUFDMask(N);
187 }], SHUFFLE_get_shuf_imm>;
189 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
190 return X86::isSHUFPMask(N);
191 }], SHUFFLE_get_shuf_imm>;
193 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
194 return X86::isSHUFPMask(N);
195 }], SHUFFLE_get_shuf_imm>;
197 //===----------------------------------------------------------------------===//
198 // SSE scalar FP Instructions
199 //===----------------------------------------------------------------------===//
201 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
202 // scheduler into a branch sequence.
203 let usesCustomDAGSchedInserter = 1 in { // Expanded by the scheduler.
204 def CMOV_FR32 : I<0, Pseudo,
205 (ops FR32:$dst, FR32:$t, FR32:$f, i8imm:$cond),
206 "#CMOV_FR32 PSEUDO!",
207 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond))]>;
208 def CMOV_FR64 : I<0, Pseudo,
209 (ops FR64:$dst, FR64:$t, FR64:$f, i8imm:$cond),
210 "#CMOV_FR64 PSEUDO!",
211 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond))]>;
212 def CMOV_V4F32 : I<0, Pseudo,
213 (ops VR128:$dst, VR128:$t, VR128:$f, i8imm:$cond),
214 "#CMOV_V4F32 PSEUDO!",
216 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
217 def CMOV_V2F64 : I<0, Pseudo,
218 (ops VR128:$dst, VR128:$t, VR128:$f, i8imm:$cond),
219 "#CMOV_V2F64 PSEUDO!",
221 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
222 def CMOV_V2I64 : I<0, Pseudo,
223 (ops VR128:$dst, VR128:$t, VR128:$f, i8imm:$cond),
224 "#CMOV_V2I64 PSEUDO!",
226 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
229 //===----------------------------------------------------------------------===//
231 //===----------------------------------------------------------------------===//
233 // SSE1 Instruction Templates:
235 // SSI - SSE1 instructions with XS prefix.
236 // PSI - SSE1 instructions with TB prefix.
237 // PSIi8 - SSE1 instructions with ImmT == Imm8 and TB prefix.
239 class SSI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
240 : I<o, F, ops, asm, pattern>, XS, Requires<[HasSSE1]>;
241 class PSI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
242 : I<o, F, ops, asm, pattern>, TB, Requires<[HasSSE1]>;
243 class PSIi8<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
244 : Ii8<o, F, ops, asm, pattern>, TB, Requires<[HasSSE1]>;
246 // Helpers for defining instructions that directly correspond to intrinsics.
247 multiclass SS_IntUnary<bits<8> o, string OpcodeStr, Intrinsic IntId> {
248 def r : SSI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
249 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
250 [(set VR128:$dst, (v4f32 (IntId VR128:$src)))]>;
251 def m : SSI<o, MRMSrcMem, (ops VR128:$dst, ssmem:$src),
252 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
253 [(set VR128:$dst, (v4f32 (IntId sse_load_f32:$src)))]>;
257 def MOVSSrr : SSI<0x10, MRMSrcReg, (ops FR32:$dst, FR32:$src),
258 "movss {$src, $dst|$dst, $src}", []>;
259 def MOVSSrm : SSI<0x10, MRMSrcMem, (ops FR32:$dst, f32mem:$src),
260 "movss {$src, $dst|$dst, $src}",
261 [(set FR32:$dst, (loadf32 addr:$src))]>;
262 def MOVSSmr : SSI<0x11, MRMDestMem, (ops f32mem:$dst, FR32:$src),
263 "movss {$src, $dst|$dst, $src}",
264 [(store FR32:$src, addr:$dst)]>;
266 def SQRTSSr : SSI<0x51, MRMSrcReg, (ops FR32:$dst, FR32:$src),
267 "sqrtss {$src, $dst|$dst, $src}",
268 [(set FR32:$dst, (fsqrt FR32:$src))]>;
269 def SQRTSSm : SSI<0x51, MRMSrcMem, (ops FR32:$dst, f32mem:$src),
270 "sqrtss {$src, $dst|$dst, $src}",
271 [(set FR32:$dst, (fsqrt (loadf32 addr:$src)))]>;
273 // Aliases to match intrinsics which expect XMM operand(s).
274 defm SQRTSS_Int : SS_IntUnary<0x51, "sqrtss" , int_x86_sse_sqrt_ss>;
275 defm RSQRTSS_Int : SS_IntUnary<0x52, "rsqrtss", int_x86_sse_rsqrt_ss>;
276 defm RCPSS_Int : SS_IntUnary<0x53, "rcpss" , int_x86_sse_rcp_ss>;
278 // Conversion instructions
279 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (ops GR32:$dst, FR32:$src),
280 "cvttss2si {$src, $dst|$dst, $src}",
281 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
282 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (ops GR32:$dst, f32mem:$src),
283 "cvttss2si {$src, $dst|$dst, $src}",
284 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
285 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (ops FR32:$dst, GR32:$src),
286 "cvtsi2ss {$src, $dst|$dst, $src}",
287 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
288 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (ops FR32:$dst, i32mem:$src),
289 "cvtsi2ss {$src, $dst|$dst, $src}",
290 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
292 // Match intrinsics which expect XMM operand(s).
293 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (ops GR32:$dst, VR128:$src),
294 "cvtss2si {$src, $dst|$dst, $src}",
295 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
296 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (ops GR32:$dst, f32mem:$src),
297 "cvtss2si {$src, $dst|$dst, $src}",
298 [(set GR32:$dst, (int_x86_sse_cvtss2si
299 (load addr:$src)))]>;
301 // Aliases for intrinsics
302 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (ops GR32:$dst, VR128:$src),
303 "cvttss2si {$src, $dst|$dst, $src}",
305 (int_x86_sse_cvttss2si VR128:$src))]>;
306 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (ops GR32:$dst, f32mem:$src),
307 "cvttss2si {$src, $dst|$dst, $src}",
309 (int_x86_sse_cvttss2si(load addr:$src)))]>;
311 let isTwoAddress = 1 in {
312 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
313 (ops VR128:$dst, VR128:$src1, GR32:$src2),
314 "cvtsi2ss {$src2, $dst|$dst, $src2}",
315 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
317 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
318 (ops VR128:$dst, VR128:$src1, i32mem:$src2),
319 "cvtsi2ss {$src2, $dst|$dst, $src2}",
320 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
321 (loadi32 addr:$src2)))]>;
324 // Comparison instructions
325 let isTwoAddress = 1 in {
326 def CMPSSrr : SSI<0xC2, MRMSrcReg,
327 (ops FR32:$dst, FR32:$src1, FR32:$src, SSECC:$cc),
328 "cmp${cc}ss {$src, $dst|$dst, $src}",
330 def CMPSSrm : SSI<0xC2, MRMSrcMem,
331 (ops FR32:$dst, FR32:$src1, f32mem:$src, SSECC:$cc),
332 "cmp${cc}ss {$src, $dst|$dst, $src}", []>;
335 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (ops FR32:$src1, FR32:$src2),
336 "ucomiss {$src2, $src1|$src1, $src2}",
337 [(X86cmp FR32:$src1, FR32:$src2)]>;
338 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (ops FR32:$src1, f32mem:$src2),
339 "ucomiss {$src2, $src1|$src1, $src2}",
340 [(X86cmp FR32:$src1, (loadf32 addr:$src2))]>;
342 // Aliases to match intrinsics which expect XMM operand(s).
343 let isTwoAddress = 1 in {
344 def Int_CMPSSrr : SSI<0xC2, MRMSrcReg,
345 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
346 "cmp${cc}ss {$src, $dst|$dst, $src}",
347 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
348 VR128:$src, imm:$cc))]>;
349 def Int_CMPSSrm : SSI<0xC2, MRMSrcMem,
350 (ops VR128:$dst, VR128:$src1, f32mem:$src, SSECC:$cc),
351 "cmp${cc}ss {$src, $dst|$dst, $src}",
352 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
353 (load addr:$src), imm:$cc))]>;
356 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
357 "ucomiss {$src2, $src1|$src1, $src2}",
358 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2)]>;
359 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
360 "ucomiss {$src2, $src1|$src1, $src2}",
361 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2))]>;
363 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
364 "comiss {$src2, $src1|$src1, $src2}",
365 [(X86comi (v4f32 VR128:$src1), VR128:$src2)]>;
366 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
367 "comiss {$src2, $src1|$src1, $src2}",
368 [(X86comi (v4f32 VR128:$src1), (load addr:$src2))]>;
370 // Aliases of packed SSE1 instructions for scalar use. These all have names that
373 // Alias instructions that map fld0 to pxor for sse.
374 def FsFLD0SS : I<0xEF, MRMInitReg, (ops FR32:$dst),
375 "pxor $dst, $dst", [(set FR32:$dst, fp32imm0)]>,
376 Requires<[HasSSE1]>, TB, OpSize;
378 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
380 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (ops FR32:$dst, FR32:$src),
381 "movaps {$src, $dst|$dst, $src}", []>;
383 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
385 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (ops FR32:$dst, f128mem:$src),
386 "movaps {$src, $dst|$dst, $src}",
387 [(set FR32:$dst, (X86loadpf32 addr:$src))]>;
389 // Alias bitwise logical operations using SSE logical ops on packed FP values.
390 let isTwoAddress = 1 in {
392 let isCommutable = 1 in {
393 def FsANDPSrr : PSI<0x54, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
394 "andps {$src2, $dst|$dst, $src2}",
395 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
396 def FsORPSrr : PSI<0x56, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
397 "orps {$src2, $dst|$dst, $src2}",
398 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
399 def FsXORPSrr : PSI<0x57, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
400 "xorps {$src2, $dst|$dst, $src2}",
401 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
404 def FsANDPSrm : PSI<0x54, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
405 "andps {$src2, $dst|$dst, $src2}",
406 [(set FR32:$dst, (X86fand FR32:$src1,
407 (X86loadpf32 addr:$src2)))]>;
408 def FsORPSrm : PSI<0x56, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
409 "orps {$src2, $dst|$dst, $src2}",
410 [(set FR32:$dst, (X86for FR32:$src1,
411 (X86loadpf32 addr:$src2)))]>;
412 def FsXORPSrm : PSI<0x57, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
413 "xorps {$src2, $dst|$dst, $src2}",
414 [(set FR32:$dst, (X86fxor FR32:$src1,
415 (X86loadpf32 addr:$src2)))]>;
417 def FsANDNPSrr : PSI<0x55, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
418 "andnps {$src2, $dst|$dst, $src2}", []>;
419 def FsANDNPSrm : PSI<0x55, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
420 "andnps {$src2, $dst|$dst, $src2}", []>;
423 /// scalar_sse1_fp_binop_rm - Scalar SSE1 binops come in three basic forms:
425 /// 1. f32 - This comes in SSE1 form for floats.
426 /// 2. rr vs rm - They include a reg+reg form and a reg+mem form.
428 /// In addition, scalar SSE ops have an intrinsic form. This form is unlike the
429 /// normal form, in that they take an entire vector (instead of a scalar) and
430 /// leave the top elements undefined. This adds another two variants of the
431 /// above permutations, giving us 8 forms for 'instruction'.
433 let isTwoAddress = 1 in {
434 multiclass scalar_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
435 SDNode OpNode, Intrinsic F32Int,
436 bit Commutable = 0> {
437 // Scalar operation, reg+reg.
438 def SSrr : SSI<opc, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
439 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
440 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
441 let isCommutable = Commutable;
444 // Scalar operation, reg+mem.
445 def SSrm : SSI<opc, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f32mem:$src2),
446 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
447 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
449 // Vector intrinsic operation, reg+reg.
450 def SSrr_Int : SSI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
451 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
452 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
453 let isCommutable = Commutable;
456 // Vector intrinsic operation, reg+mem.
457 def SSrm_Int : SSI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, ssmem:$src2),
458 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
459 [(set VR128:$dst, (F32Int VR128:$src1,
460 sse_load_f32:$src2))]>;
464 // Arithmetic instructions
465 defm ADD : scalar_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
466 defm MUL : scalar_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
467 defm SUB : scalar_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
468 defm DIV : scalar_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
470 defm MAX : scalar_sse1_fp_binop_rm<0x5F, "max", X86fmax, int_x86_sse_max_ss>;
471 defm MIN : scalar_sse1_fp_binop_rm<0x5D, "min", X86fmin, int_x86_sse_min_ss>;
473 //===----------------------------------------------------------------------===//
474 // SSE packed FP Instructions
477 def MOVAPSrr : PSI<0x28, MRMSrcReg, (ops VR128:$dst, VR128:$src),
478 "movaps {$src, $dst|$dst, $src}", []>;
479 def MOVAPSrm : PSI<0x28, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
480 "movaps {$src, $dst|$dst, $src}",
481 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
483 def MOVAPSmr : PSI<0x29, MRMDestMem, (ops f128mem:$dst, VR128:$src),
484 "movaps {$src, $dst|$dst, $src}",
485 [(store (v4f32 VR128:$src), addr:$dst)]>;
487 def MOVUPSrr : PSI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src),
488 "movups {$src, $dst|$dst, $src}", []>;
489 def MOVUPSrm : PSI<0x10, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
490 "movups {$src, $dst|$dst, $src}",
491 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
492 def MOVUPSmr : PSI<0x11, MRMDestMem, (ops f128mem:$dst, VR128:$src),
493 "movups {$src, $dst|$dst, $src}",
494 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
496 let isTwoAddress = 1 in {
497 let AddedComplexity = 20 in {
498 def MOVLPSrm : PSI<0x12, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
499 "movlps {$src2, $dst|$dst, $src2}",
501 (v4f32 (vector_shuffle VR128:$src1,
502 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
503 MOVLP_shuffle_mask)))]>;
504 def MOVHPSrm : PSI<0x16, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
505 "movhps {$src2, $dst|$dst, $src2}",
507 (v4f32 (vector_shuffle VR128:$src1,
508 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
509 MOVHP_shuffle_mask)))]>;
513 def MOVLPSmr : PSI<0x13, MRMDestMem, (ops f64mem:$dst, VR128:$src),
514 "movlps {$src, $dst|$dst, $src}",
515 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
516 (iPTR 0))), addr:$dst)]>;
518 // v2f64 extract element 1 is always custom lowered to unpack high to low
519 // and extract element 0 so the non-store version isn't too horrible.
520 def MOVHPSmr : PSI<0x17, MRMDestMem, (ops f64mem:$dst, VR128:$src),
521 "movhps {$src, $dst|$dst, $src}",
522 [(store (f64 (vector_extract
523 (v2f64 (vector_shuffle
524 (bc_v2f64 (v4f32 VR128:$src)), (undef),
525 UNPCKH_shuffle_mask)), (iPTR 0))),
528 let isTwoAddress = 1 in {
529 let AddedComplexity = 15 in {
530 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
531 "movlhps {$src2, $dst|$dst, $src2}",
533 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
534 MOVHP_shuffle_mask)))]>;
536 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
537 "movhlps {$src2, $dst|$dst, $src2}",
539 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
540 MOVHLPS_shuffle_mask)))]>;
546 /// packed_sse1_fp_binop_rm - Packed SSE binops come in three basic forms:
547 /// 1. v4f32 - This comes in SSE1 form for float.
548 /// 2. rr vs rm - They include a reg+reg form and a ref+mem form.
550 let isTwoAddress = 1 in {
551 multiclass packed_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
552 SDNode OpNode, bit Commutable = 0> {
553 // Packed operation, reg+reg.
554 def PSrr : PSI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
555 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2}"),
556 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
557 let isCommutable = Commutable;
560 // Packed operation, reg+mem.
561 def PSrm : PSI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
562 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2}"),
563 [(set VR128:$dst, (OpNode VR128:$src1, (loadv4f32 addr:$src2)))]>;
567 defm ADD : packed_sse1_fp_binop_rm<0x58, "add", fadd, 1>;
568 defm MUL : packed_sse1_fp_binop_rm<0x59, "mul", fmul, 1>;
569 defm DIV : packed_sse1_fp_binop_rm<0x5E, "div", fdiv>;
570 defm SUB : packed_sse1_fp_binop_rm<0x5C, "sub", fsub>;
574 class PS_Intr<bits<8> o, string OpcodeStr, Intrinsic IntId>
575 : PSI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
576 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
577 [(set VR128:$dst, (IntId VR128:$src))]>;
578 class PS_Intm<bits<8> o, string OpcodeStr, Intrinsic IntId>
579 : PSI<o, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
580 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
581 [(set VR128:$dst, (IntId (load addr:$src)))]>;
583 class PS_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
584 : PSI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
585 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
586 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
587 class PS_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
588 : PSI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f32mem:$src2),
589 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
590 [(set VR128:$dst, (IntId VR128:$src1, (load addr:$src2)))]>;
592 def SQRTPSr : PS_Intr<0x51, "sqrtps", int_x86_sse_sqrt_ps>;
593 def SQRTPSm : PS_Intm<0x51, "sqrtps", int_x86_sse_sqrt_ps>;
595 def RSQRTPSr : PS_Intr<0x52, "rsqrtps", int_x86_sse_rsqrt_ps>;
596 def RSQRTPSm : PS_Intm<0x52, "rsqrtps", int_x86_sse_rsqrt_ps>;
597 def RCPPSr : PS_Intr<0x53, "rcpps", int_x86_sse_rcp_ps>;
598 def RCPPSm : PS_Intm<0x53, "rcpps", int_x86_sse_rcp_ps>;
600 let isTwoAddress = 1 in {
601 let isCommutable = 1 in {
602 def MAXPSrr : PS_Intrr<0x5F, "maxps", int_x86_sse_max_ps>;
603 def MINPSrr : PS_Intrr<0x5D, "minps", int_x86_sse_min_ps>;
606 def MAXPSrm : PS_Intrm<0x5F, "maxps", int_x86_sse_max_ps>;
607 def MINPSrm : PS_Intrm<0x5D, "minps", int_x86_sse_min_ps>;
611 let isTwoAddress = 1 in {
612 let isCommutable = 1 in {
613 def ANDPSrr : PSI<0x54, MRMSrcReg,
614 (ops VR128:$dst, VR128:$src1, VR128:$src2),
615 "andps {$src2, $dst|$dst, $src2}",
616 [(set VR128:$dst, (v2i64
617 (and VR128:$src1, VR128:$src2)))]>;
618 def ORPSrr : PSI<0x56, MRMSrcReg,
619 (ops VR128:$dst, VR128:$src1, VR128:$src2),
620 "orps {$src2, $dst|$dst, $src2}",
621 [(set VR128:$dst, (v2i64
622 (or VR128:$src1, VR128:$src2)))]>;
623 def XORPSrr : PSI<0x57, MRMSrcReg,
624 (ops VR128:$dst, VR128:$src1, VR128:$src2),
625 "xorps {$src2, $dst|$dst, $src2}",
626 [(set VR128:$dst, (v2i64
627 (xor VR128:$src1, VR128:$src2)))]>;
630 def ANDPSrm : PSI<0x54, MRMSrcMem,
631 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
632 "andps {$src2, $dst|$dst, $src2}",
633 [(set VR128:$dst, (and VR128:$src1,
634 (bc_v2i64 (loadv4f32 addr:$src2))))]>;
635 def ORPSrm : PSI<0x56, MRMSrcMem,
636 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
637 "orps {$src2, $dst|$dst, $src2}",
638 [(set VR128:$dst, (or VR128:$src1,
639 (bc_v2i64 (loadv4f32 addr:$src2))))]>;
640 def XORPSrm : PSI<0x57, MRMSrcMem,
641 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
642 "xorps {$src2, $dst|$dst, $src2}",
643 [(set VR128:$dst, (xor VR128:$src1,
644 (bc_v2i64 (loadv4f32 addr:$src2))))]>;
645 def ANDNPSrr : PSI<0x55, MRMSrcReg,
646 (ops VR128:$dst, VR128:$src1, VR128:$src2),
647 "andnps {$src2, $dst|$dst, $src2}",
649 (v2i64 (and (xor VR128:$src1,
650 (bc_v2i64 (v4i32 immAllOnesV))),
652 def ANDNPSrm : PSI<0x55, MRMSrcMem,
653 (ops VR128:$dst, VR128:$src1,f128mem:$src2),
654 "andnps {$src2, $dst|$dst, $src2}",
656 (v2i64 (and (xor VR128:$src1,
657 (bc_v2i64 (v4i32 immAllOnesV))),
658 (bc_v2i64 (loadv4f32 addr:$src2)))))]>;
661 let isTwoAddress = 1 in {
662 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
663 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
664 "cmp${cc}ps {$src, $dst|$dst, $src}",
665 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
666 VR128:$src, imm:$cc))]>;
667 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
668 (ops VR128:$dst, VR128:$src1, f128mem:$src, SSECC:$cc),
669 "cmp${cc}ps {$src, $dst|$dst, $src}",
670 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
671 (load addr:$src), imm:$cc))]>;
674 // Shuffle and unpack instructions
675 let isTwoAddress = 1 in {
676 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
677 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
678 (ops VR128:$dst, VR128:$src1,
679 VR128:$src2, i32i8imm:$src3),
680 "shufps {$src3, $src2, $dst|$dst, $src2, $src3}",
682 (v4f32 (vector_shuffle
683 VR128:$src1, VR128:$src2,
684 SHUFP_shuffle_mask:$src3)))]>;
685 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
686 (ops VR128:$dst, VR128:$src1,
687 f128mem:$src2, i32i8imm:$src3),
688 "shufps {$src3, $src2, $dst|$dst, $src2, $src3}",
690 (v4f32 (vector_shuffle
691 VR128:$src1, (load addr:$src2),
692 SHUFP_shuffle_mask:$src3)))]>;
694 let AddedComplexity = 10 in {
695 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
696 (ops VR128:$dst, VR128:$src1, VR128:$src2),
697 "unpckhps {$src2, $dst|$dst, $src2}",
699 (v4f32 (vector_shuffle
700 VR128:$src1, VR128:$src2,
701 UNPCKH_shuffle_mask)))]>;
702 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
703 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
704 "unpckhps {$src2, $dst|$dst, $src2}",
706 (v4f32 (vector_shuffle
707 VR128:$src1, (load addr:$src2),
708 UNPCKH_shuffle_mask)))]>;
710 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
711 (ops VR128:$dst, VR128:$src1, VR128:$src2),
712 "unpcklps {$src2, $dst|$dst, $src2}",
714 (v4f32 (vector_shuffle
715 VR128:$src1, VR128:$src2,
716 UNPCKL_shuffle_mask)))]>;
717 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
718 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
719 "unpcklps {$src2, $dst|$dst, $src2}",
721 (v4f32 (vector_shuffle
722 VR128:$src1, (load addr:$src2),
723 UNPCKL_shuffle_mask)))]>;
728 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (ops GR32:$dst, VR128:$src),
729 "movmskps {$src, $dst|$dst, $src}",
730 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
731 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (ops GR32:$dst, VR128:$src),
732 "movmskpd {$src, $dst|$dst, $src}",
733 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
735 // Prefetching loads.
736 // TODO: no intrinsics for these?
737 def PREFETCHT0 : PSI<0x18, MRM1m, (ops i8mem:$src), "prefetcht0 $src", []>;
738 def PREFETCHT1 : PSI<0x18, MRM2m, (ops i8mem:$src), "prefetcht1 $src", []>;
739 def PREFETCHT2 : PSI<0x18, MRM3m, (ops i8mem:$src), "prefetcht2 $src", []>;
740 def PREFETCHNTA : PSI<0x18, MRM0m, (ops i8mem:$src), "prefetchnta $src", []>;
742 // Non-temporal stores
743 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (ops i128mem:$dst, VR128:$src),
744 "movntps {$src, $dst|$dst, $src}",
745 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
747 // Load, store, and memory fence
748 def SFENCE : PSI<0xAE, MRM7m, (ops), "sfence", [(int_x86_sse_sfence)]>;
751 def LDMXCSR : PSI<0xAE, MRM2m, (ops i32mem:$src),
752 "ldmxcsr $src", [(int_x86_sse_ldmxcsr addr:$src)]>;
753 def STMXCSR : PSI<0xAE, MRM3m, (ops i32mem:$dst),
754 "stmxcsr $dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
756 // Alias instructions that map zero vector to pxor / xorp* for sse.
757 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
758 let isReMaterializable = 1 in
759 def V_SET0 : PSI<0x57, MRMInitReg, (ops VR128:$dst),
761 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
763 // FR32 to 128-bit vector conversion.
764 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (ops VR128:$dst, FR32:$src),
765 "movss {$src, $dst|$dst, $src}",
767 (v4f32 (scalar_to_vector FR32:$src)))]>;
768 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
769 "movss {$src, $dst|$dst, $src}",
771 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
773 // FIXME: may not be able to eliminate this movss with coalescing the src and
774 // dest register classes are different. We really want to write this pattern
776 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
778 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (ops FR32:$dst, VR128:$src),
779 "movss {$src, $dst|$dst, $src}",
780 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
782 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (ops f32mem:$dst, VR128:$src),
783 "movss {$src, $dst|$dst, $src}",
784 [(store (f32 (vector_extract (v4f32 VR128:$src),
785 (iPTR 0))), addr:$dst)]>;
788 // Move to lower bits of a VR128, leaving upper bits alone.
789 // Three operand (but two address) aliases.
790 let isTwoAddress = 1 in {
791 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
792 (ops VR128:$dst, VR128:$src1, FR32:$src2),
793 "movss {$src2, $dst|$dst, $src2}", []>;
795 let AddedComplexity = 15 in
796 def MOVLPSrr : SSI<0x10, MRMSrcReg,
797 (ops VR128:$dst, VR128:$src1, VR128:$src2),
798 "movss {$src2, $dst|$dst, $src2}",
800 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
801 MOVL_shuffle_mask)))]>;
804 // Move to lower bits of a VR128 and zeroing upper bits.
805 // Loading from memory automatically zeroing upper bits.
806 let AddedComplexity = 20 in
807 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
808 "movss {$src, $dst|$dst, $src}",
809 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV,
810 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
811 MOVL_shuffle_mask)))]>;
814 //===----------------------------------------------------------------------===//
816 //===----------------------------------------------------------------------===//
818 // SSE2 Instruction Templates:
820 // SDI - SSE2 instructions with XD prefix.
821 // PDI - SSE2 instructions with TB and OpSize prefixes.
822 // PDIi8 - SSE2 instructions with ImmT == Imm8 and TB and OpSize prefixes.
824 class SDI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
825 : I<o, F, ops, asm, pattern>, XD, Requires<[HasSSE2]>;
826 class PDI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
827 : I<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE2]>;
828 class PDIi8<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
829 : Ii8<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE2]>;
831 // Helpers for defining instructions that directly correspond to intrinsics.
832 multiclass SD_IntUnary<bits<8> o, string OpcodeStr, Intrinsic IntId> {
833 def r : SDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
834 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
835 [(set VR128:$dst, (v2f64 (IntId VR128:$src)))]>;
836 def m : SDI<o, MRMSrcMem, (ops VR128:$dst, sdmem:$src),
837 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
838 [(set VR128:$dst, (v2f64 (IntId sse_load_f64:$src)))]>;
842 def MOVSDrr : SDI<0x10, MRMSrcReg, (ops FR64:$dst, FR64:$src),
843 "movsd {$src, $dst|$dst, $src}", []>;
844 def MOVSDrm : SDI<0x10, MRMSrcMem, (ops FR64:$dst, f64mem:$src),
845 "movsd {$src, $dst|$dst, $src}",
846 [(set FR64:$dst, (loadf64 addr:$src))]>;
847 def MOVSDmr : SDI<0x11, MRMDestMem, (ops f64mem:$dst, FR64:$src),
848 "movsd {$src, $dst|$dst, $src}",
849 [(store FR64:$src, addr:$dst)]>;
851 def SQRTSDr : SDI<0x51, MRMSrcReg, (ops FR64:$dst, FR64:$src),
852 "sqrtsd {$src, $dst|$dst, $src}",
853 [(set FR64:$dst, (fsqrt FR64:$src))]>;
854 def SQRTSDm : SDI<0x51, MRMSrcMem, (ops FR64:$dst, f64mem:$src),
855 "sqrtsd {$src, $dst|$dst, $src}",
856 [(set FR64:$dst, (fsqrt (loadf64 addr:$src)))]>;
858 // Aliases to match intrinsics which expect XMM operand(s).
859 defm SQRTSD_Int : SD_IntUnary<0x51, "sqrtsd" , int_x86_sse2_sqrt_sd>;
861 // Conversion instructions
862 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (ops GR32:$dst, FR64:$src),
863 "cvttsd2si {$src, $dst|$dst, $src}",
864 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
865 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (ops GR32:$dst, f64mem:$src),
866 "cvttsd2si {$src, $dst|$dst, $src}",
867 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
868 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (ops FR32:$dst, FR64:$src),
869 "cvtsd2ss {$src, $dst|$dst, $src}",
870 [(set FR32:$dst, (fround FR64:$src))]>;
871 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (ops FR32:$dst, f64mem:$src),
872 "cvtsd2ss {$src, $dst|$dst, $src}",
873 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
874 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (ops FR64:$dst, GR32:$src),
875 "cvtsi2sd {$src, $dst|$dst, $src}",
876 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
877 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (ops FR64:$dst, i32mem:$src),
878 "cvtsi2sd {$src, $dst|$dst, $src}",
879 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
881 // SSE2 instructions with XS prefix
882 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (ops FR64:$dst, FR32:$src),
883 "cvtss2sd {$src, $dst|$dst, $src}",
884 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
886 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (ops FR64:$dst, f32mem:$src),
887 "cvtss2sd {$src, $dst|$dst, $src}",
888 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
891 // Match intrinsics which expect XMM operand(s).
892 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (ops GR32:$dst, VR128:$src),
893 "cvtsd2si {$src, $dst|$dst, $src}",
894 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
895 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (ops GR32:$dst, f128mem:$src),
896 "cvtsd2si {$src, $dst|$dst, $src}",
897 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
898 (load addr:$src)))]>;
900 // Aliases for intrinsics
901 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (ops GR32:$dst, VR128:$src),
902 "cvttsd2si {$src, $dst|$dst, $src}",
904 (int_x86_sse2_cvttsd2si VR128:$src))]>;
905 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (ops GR32:$dst, f128mem:$src),
906 "cvttsd2si {$src, $dst|$dst, $src}",
907 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
908 (load addr:$src)))]>;
910 // Comparison instructions
911 let isTwoAddress = 1 in {
912 def CMPSDrr : SDI<0xC2, MRMSrcReg,
913 (ops FR64:$dst, FR64:$src1, FR64:$src, SSECC:$cc),
914 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
915 def CMPSDrm : SDI<0xC2, MRMSrcMem,
916 (ops FR64:$dst, FR64:$src1, f64mem:$src, SSECC:$cc),
917 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
920 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (ops FR64:$src1, FR64:$src2),
921 "ucomisd {$src2, $src1|$src1, $src2}",
922 [(X86cmp FR64:$src1, FR64:$src2)]>;
923 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (ops FR64:$src1, f64mem:$src2),
924 "ucomisd {$src2, $src1|$src1, $src2}",
925 [(X86cmp FR64:$src1, (loadf64 addr:$src2))]>;
927 // Aliases to match intrinsics which expect XMM operand(s).
928 let isTwoAddress = 1 in {
929 def Int_CMPSDrr : SDI<0xC2, MRMSrcReg,
930 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
931 "cmp${cc}sd {$src, $dst|$dst, $src}",
932 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
933 VR128:$src, imm:$cc))]>;
934 def Int_CMPSDrm : SDI<0xC2, MRMSrcMem,
935 (ops VR128:$dst, VR128:$src1, f64mem:$src, SSECC:$cc),
936 "cmp${cc}sd {$src, $dst|$dst, $src}",
937 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
938 (load addr:$src), imm:$cc))]>;
941 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
942 "ucomisd {$src2, $src1|$src1, $src2}",
943 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
944 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
945 "ucomisd {$src2, $src1|$src1, $src2}",
946 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2))]>;
948 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
949 "comisd {$src2, $src1|$src1, $src2}",
950 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
951 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
952 "comisd {$src2, $src1|$src1, $src2}",
953 [(X86comi (v2f64 VR128:$src1), (load addr:$src2))]>;
955 // Aliases of packed instructions for scalar use. These all have names that
958 // Alias instructions that map fld0 to pxor for sse.
959 def FsFLD0SD : I<0xEF, MRMInitReg, (ops FR64:$dst),
960 "pxor $dst, $dst", [(set FR64:$dst, fp64imm0)]>,
961 Requires<[HasSSE2]>, TB, OpSize;
963 // Alias instructions to do FR64 reg-to-reg copy using movapd. Upper bits are
965 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (ops FR64:$dst, FR64:$src),
966 "movapd {$src, $dst|$dst, $src}", []>;
968 // Alias instructions to load FR64 from f128mem using movapd. Upper bits are
970 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (ops FR64:$dst, f128mem:$src),
971 "movapd {$src, $dst|$dst, $src}",
972 [(set FR64:$dst, (X86loadpf64 addr:$src))]>;
974 // Alias bitwise logical operations using SSE logical ops on packed FP values.
975 let isTwoAddress = 1 in {
976 let isCommutable = 1 in {
977 def FsANDPDrr : PDI<0x54, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
978 "andpd {$src2, $dst|$dst, $src2}",
979 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
980 def FsORPDrr : PDI<0x56, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
981 "orpd {$src2, $dst|$dst, $src2}",
982 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
983 def FsXORPDrr : PDI<0x57, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
984 "xorpd {$src2, $dst|$dst, $src2}",
985 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
988 def FsANDPDrm : PDI<0x54, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
989 "andpd {$src2, $dst|$dst, $src2}",
990 [(set FR64:$dst, (X86fand FR64:$src1,
991 (X86loadpf64 addr:$src2)))]>;
992 def FsORPDrm : PDI<0x56, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
993 "orpd {$src2, $dst|$dst, $src2}",
994 [(set FR64:$dst, (X86for FR64:$src1,
995 (X86loadpf64 addr:$src2)))]>;
996 def FsXORPDrm : PDI<0x57, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
997 "xorpd {$src2, $dst|$dst, $src2}",
998 [(set FR64:$dst, (X86fxor FR64:$src1,
999 (X86loadpf64 addr:$src2)))]>;
1001 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1002 (ops FR64:$dst, FR64:$src1, FR64:$src2),
1003 "andnpd {$src2, $dst|$dst, $src2}", []>;
1004 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1005 (ops FR64:$dst, FR64:$src1, f128mem:$src2),
1006 "andnpd {$src2, $dst|$dst, $src2}", []>;
1009 /// scalar_sse2_fp_binop_rm - Scalar SSE2 binops come in three basic forms:
1011 /// 1. f64 - This comes in SSE2 form for doubles.
1012 /// 2. rr vs rm - They include a reg+reg form and a reg+mem form.
1014 /// In addition, scalar SSE ops have an intrinsic form. This form is unlike the
1015 /// normal form, in that they take an entire vector (instead of a scalar) and
1016 /// leave the top elements undefined. This adds another two variants of the
1017 /// above permutations, giving us 8 forms for 'instruction'.
1019 let isTwoAddress = 1 in {
1020 multiclass scalar_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1021 SDNode OpNode, Intrinsic F64Int,
1022 bit Commutable = 0> {
1023 // Scalar operation, reg+reg.
1024 def SDrr : SDI<opc, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
1025 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1026 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1027 let isCommutable = Commutable;
1030 // Scalar operation, reg+mem.
1031 def SDrm : SDI<opc, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f64mem:$src2),
1032 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1033 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1035 // Vector intrinsic operation, reg+reg.
1036 def SDrr_Int : SDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1037 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1038 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1039 let isCommutable = Commutable;
1042 // Vector intrinsic operation, reg+mem.
1043 def SDrm_Int : SDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, sdmem:$src2),
1044 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
1045 [(set VR128:$dst, (F64Int VR128:$src1,
1046 sse_load_f64:$src2))]>;
1050 // Arithmetic instructions
1051 defm ADD : scalar_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1052 defm MUL : scalar_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1053 defm SUB : scalar_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1054 defm DIV : scalar_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1056 defm MAX : scalar_sse2_fp_binop_rm<0x5F, "max", X86fmax, int_x86_sse2_max_sd>;
1057 defm MIN : scalar_sse2_fp_binop_rm<0x5D, "min", X86fmin, int_x86_sse2_min_sd>;
1059 //===----------------------------------------------------------------------===//
1060 // SSE packed FP Instructions
1062 // Move Instructions
1063 def MOVAPDrr : PDI<0x28, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1064 "movapd {$src, $dst|$dst, $src}", []>;
1065 def MOVAPDrm : PDI<0x28, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1066 "movapd {$src, $dst|$dst, $src}",
1067 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1069 def MOVAPDmr : PDI<0x29, MRMDestMem, (ops f128mem:$dst, VR128:$src),
1070 "movapd {$src, $dst|$dst, $src}",
1071 [(store (v2f64 VR128:$src), addr:$dst)]>;
1073 def MOVUPDrr : PDI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1074 "movupd {$src, $dst|$dst, $src}", []>;
1075 def MOVUPDrm : PDI<0x10, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1076 "movupd {$src, $dst|$dst, $src}",
1077 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1078 def MOVUPDmr : PDI<0x11, MRMDestMem, (ops f128mem:$dst, VR128:$src),
1079 "movupd {$src, $dst|$dst, $src}",
1080 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1082 let isTwoAddress = 1 in {
1083 let AddedComplexity = 20 in {
1084 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1085 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
1086 "movlpd {$src2, $dst|$dst, $src2}",
1088 (v2f64 (vector_shuffle VR128:$src1,
1089 (scalar_to_vector (loadf64 addr:$src2)),
1090 MOVLP_shuffle_mask)))]>;
1091 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1092 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
1093 "movhpd {$src2, $dst|$dst, $src2}",
1095 (v2f64 (vector_shuffle VR128:$src1,
1096 (scalar_to_vector (loadf64 addr:$src2)),
1097 MOVHP_shuffle_mask)))]>;
1098 } // AddedComplexity
1101 def MOVLPDmr : PDI<0x13, MRMDestMem, (ops f64mem:$dst, VR128:$src),
1102 "movlpd {$src, $dst|$dst, $src}",
1103 [(store (f64 (vector_extract (v2f64 VR128:$src),
1104 (iPTR 0))), addr:$dst)]>;
1106 // v2f64 extract element 1 is always custom lowered to unpack high to low
1107 // and extract element 0 so the non-store version isn't too horrible.
1108 def MOVHPDmr : PDI<0x17, MRMDestMem, (ops f64mem:$dst, VR128:$src),
1109 "movhpd {$src, $dst|$dst, $src}",
1110 [(store (f64 (vector_extract
1111 (v2f64 (vector_shuffle VR128:$src, (undef),
1112 UNPCKH_shuffle_mask)), (iPTR 0))),
1115 // SSE2 instructions without OpSize prefix
1116 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1117 "cvtdq2ps {$src, $dst|$dst, $src}",
1118 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1119 TB, Requires<[HasSSE2]>;
1120 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1121 "cvtdq2ps {$src, $dst|$dst, $src}",
1122 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1123 (bitconvert (loadv2i64 addr:$src))))]>,
1124 TB, Requires<[HasSSE2]>;
1126 // SSE2 instructions with XS prefix
1127 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1128 "cvtdq2pd {$src, $dst|$dst, $src}",
1129 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1130 XS, Requires<[HasSSE2]>;
1131 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1132 "cvtdq2pd {$src, $dst|$dst, $src}",
1133 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1134 (bitconvert (loadv2i64 addr:$src))))]>,
1135 XS, Requires<[HasSSE2]>;
1137 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1138 "cvtps2dq {$src, $dst|$dst, $src}",
1139 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1140 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1141 "cvtps2dq {$src, $dst|$dst, $src}",
1142 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1143 (load addr:$src)))]>;
1144 // SSE2 packed instructions with XS prefix
1145 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1146 "cvttps2dq {$src, $dst|$dst, $src}",
1147 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1148 XS, Requires<[HasSSE2]>;
1149 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1150 "cvttps2dq {$src, $dst|$dst, $src}",
1151 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1152 (load addr:$src)))]>,
1153 XS, Requires<[HasSSE2]>;
1155 // SSE2 packed instructions with XD prefix
1156 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1157 "cvtpd2dq {$src, $dst|$dst, $src}",
1158 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1159 XD, Requires<[HasSSE2]>;
1160 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1161 "cvtpd2dq {$src, $dst|$dst, $src}",
1162 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1163 (load addr:$src)))]>,
1164 XD, Requires<[HasSSE2]>;
1166 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1167 "cvttpd2dq {$src, $dst|$dst, $src}",
1168 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1169 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1170 "cvttpd2dq {$src, $dst|$dst, $src}",
1171 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1172 (load addr:$src)))]>;
1174 // SSE2 instructions without OpSize prefix
1175 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1176 "cvtps2pd {$src, $dst|$dst, $src}",
1177 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1178 TB, Requires<[HasSSE2]>;
1179 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (ops VR128:$dst, f64mem:$src),
1180 "cvtps2pd {$src, $dst|$dst, $src}",
1181 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1182 (load addr:$src)))]>,
1183 TB, Requires<[HasSSE2]>;
1185 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1186 "cvtpd2ps {$src, $dst|$dst, $src}",
1187 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1188 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (ops VR128:$dst, f128mem:$src),
1189 "cvtpd2ps {$src, $dst|$dst, $src}",
1190 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1191 (load addr:$src)))]>;
1193 // Match intrinsics which expect XMM operand(s).
1194 // Aliases for intrinsics
1195 let isTwoAddress = 1 in {
1196 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1197 (ops VR128:$dst, VR128:$src1, GR32:$src2),
1198 "cvtsi2sd {$src2, $dst|$dst, $src2}",
1199 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1201 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1202 (ops VR128:$dst, VR128:$src1, i32mem:$src2),
1203 "cvtsi2sd {$src2, $dst|$dst, $src2}",
1204 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1205 (loadi32 addr:$src2)))]>;
1206 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1207 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1208 "cvtsd2ss {$src2, $dst|$dst, $src2}",
1209 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1211 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1212 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
1213 "cvtsd2ss {$src2, $dst|$dst, $src2}",
1214 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1215 (load addr:$src2)))]>;
1216 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1217 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1218 "cvtss2sd {$src2, $dst|$dst, $src2}",
1219 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1220 VR128:$src2))]>, XS,
1221 Requires<[HasSSE2]>;
1222 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1223 (ops VR128:$dst, VR128:$src1, f32mem:$src2),
1224 "cvtss2sd {$src2, $dst|$dst, $src2}",
1225 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1226 (load addr:$src2)))]>, XS,
1227 Requires<[HasSSE2]>;
1230 /// packed_sse2_fp_binop_rm - Packed SSE binops come in three basic forms:
1231 /// 1. v2f64 - This comes in SSE2 form for doubles.
1232 /// 2. rr vs rm - They include a reg+reg form and a ref+mem form.
1234 let isTwoAddress = 1 in {
1235 multiclass packed_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1236 SDNode OpNode, bit Commutable = 0> {
1237 // Packed operation, reg+reg.
1238 def PDrr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1239 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
1240 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1241 let isCommutable = Commutable;
1244 // Packed operation, reg+mem.
1245 def PDrm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1246 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
1247 [(set VR128:$dst, (OpNode VR128:$src1, (loadv2f64 addr:$src2)))]>;
1251 defm ADD : packed_sse2_fp_binop_rm<0x58, "add", fadd, 1>;
1252 defm MUL : packed_sse2_fp_binop_rm<0x59, "mul", fmul, 1>;
1253 defm DIV : packed_sse2_fp_binop_rm<0x5E, "div", fdiv>;
1254 defm SUB : packed_sse2_fp_binop_rm<0x5C, "sub", fsub>;
1258 class PD_Intr<bits<8> o, string OpcodeStr, Intrinsic IntId>
1259 : PDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1260 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
1261 [(set VR128:$dst, (IntId VR128:$src))]>;
1262 class PD_Intm<bits<8> o, string OpcodeStr, Intrinsic IntId>
1263 : PDI<o, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1264 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
1265 [(set VR128:$dst, (IntId (load addr:$src)))]>;
1267 class PD_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
1268 : PDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1269 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1270 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1271 class PD_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
1272 : PDI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
1273 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1274 [(set VR128:$dst, (IntId VR128:$src1, (load addr:$src2)))]>;
1276 def SQRTPDr : PD_Intr<0x51, "sqrtpd", int_x86_sse2_sqrt_pd>;
1277 def SQRTPDm : PD_Intm<0x51, "sqrtpd", int_x86_sse2_sqrt_pd>;
1279 let isTwoAddress = 1 in {
1280 let isCommutable = 1 in {
1281 def MAXPDrr : PD_Intrr<0x5F, "maxpd", int_x86_sse2_max_pd>;
1282 def MINPDrr : PD_Intrr<0x5D, "minpd", int_x86_sse2_min_pd>;
1285 def MAXPDrm : PD_Intrm<0x5F, "maxpd", int_x86_sse2_max_pd>;
1286 def MINPDrm : PD_Intrm<0x5D, "minpd", int_x86_sse2_min_pd>;
1290 let isTwoAddress = 1 in {
1291 let isCommutable = 1 in {
1292 def ANDPDrr : PDI<0x54, MRMSrcReg,
1293 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1294 "andpd {$src2, $dst|$dst, $src2}",
1296 (and (bc_v2i64 (v2f64 VR128:$src1)),
1297 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1298 def ORPDrr : PDI<0x56, MRMSrcReg,
1299 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1300 "orpd {$src2, $dst|$dst, $src2}",
1302 (or (bc_v2i64 (v2f64 VR128:$src1)),
1303 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1304 def XORPDrr : PDI<0x57, MRMSrcReg,
1305 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1306 "xorpd {$src2, $dst|$dst, $src2}",
1308 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1309 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1312 def ANDPDrm : PDI<0x54, MRMSrcMem,
1313 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1314 "andpd {$src2, $dst|$dst, $src2}",
1316 (and (bc_v2i64 (v2f64 VR128:$src1)),
1317 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1318 def ORPDrm : PDI<0x56, MRMSrcMem,
1319 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1320 "orpd {$src2, $dst|$dst, $src2}",
1322 (or (bc_v2i64 (v2f64 VR128:$src1)),
1323 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1324 def XORPDrm : PDI<0x57, MRMSrcMem,
1325 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1326 "xorpd {$src2, $dst|$dst, $src2}",
1328 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1329 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1330 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1331 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1332 "andnpd {$src2, $dst|$dst, $src2}",
1334 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1335 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1336 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1337 (ops VR128:$dst, VR128:$src1,f128mem:$src2),
1338 "andnpd {$src2, $dst|$dst, $src2}",
1340 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1341 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1344 let isTwoAddress = 1 in {
1345 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1346 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
1347 "cmp${cc}pd {$src, $dst|$dst, $src}",
1348 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1349 VR128:$src, imm:$cc))]>;
1350 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1351 (ops VR128:$dst, VR128:$src1, f128mem:$src, SSECC:$cc),
1352 "cmp${cc}pd {$src, $dst|$dst, $src}",
1353 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1354 (load addr:$src), imm:$cc))]>;
1357 // Shuffle and unpack instructions
1358 let isTwoAddress = 1 in {
1359 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1360 (ops VR128:$dst, VR128:$src1, VR128:$src2, i8imm:$src3),
1361 "shufpd {$src3, $src2, $dst|$dst, $src2, $src3}",
1362 [(set VR128:$dst, (v2f64 (vector_shuffle
1363 VR128:$src1, VR128:$src2,
1364 SHUFP_shuffle_mask:$src3)))]>;
1365 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1366 (ops VR128:$dst, VR128:$src1,
1367 f128mem:$src2, i8imm:$src3),
1368 "shufpd {$src3, $src2, $dst|$dst, $src2, $src3}",
1370 (v2f64 (vector_shuffle
1371 VR128:$src1, (load addr:$src2),
1372 SHUFP_shuffle_mask:$src3)))]>;
1374 let AddedComplexity = 10 in {
1375 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1376 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1377 "unpckhpd {$src2, $dst|$dst, $src2}",
1379 (v2f64 (vector_shuffle
1380 VR128:$src1, VR128:$src2,
1381 UNPCKH_shuffle_mask)))]>;
1382 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1383 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1384 "unpckhpd {$src2, $dst|$dst, $src2}",
1386 (v2f64 (vector_shuffle
1387 VR128:$src1, (load addr:$src2),
1388 UNPCKH_shuffle_mask)))]>;
1390 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1391 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1392 "unpcklpd {$src2, $dst|$dst, $src2}",
1394 (v2f64 (vector_shuffle
1395 VR128:$src1, VR128:$src2,
1396 UNPCKL_shuffle_mask)))]>;
1397 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1398 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1399 "unpcklpd {$src2, $dst|$dst, $src2}",
1401 (v2f64 (vector_shuffle
1402 VR128:$src1, (load addr:$src2),
1403 UNPCKL_shuffle_mask)))]>;
1404 } // AddedComplexity
1408 //===----------------------------------------------------------------------===//
1409 // SSE integer instructions
1411 // Move Instructions
1412 def MOVDQArr : PDI<0x6F, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1413 "movdqa {$src, $dst|$dst, $src}", []>;
1414 def MOVDQArm : PDI<0x6F, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1415 "movdqa {$src, $dst|$dst, $src}",
1416 [(set VR128:$dst, (loadv2i64 addr:$src))]>;
1417 def MOVDQAmr : PDI<0x7F, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1418 "movdqa {$src, $dst|$dst, $src}",
1419 [(store (v2i64 VR128:$src), addr:$dst)]>;
1420 def MOVDQUrm : I<0x6F, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1421 "movdqu {$src, $dst|$dst, $src}",
1422 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1423 XS, Requires<[HasSSE2]>;
1424 def MOVDQUmr : I<0x7F, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1425 "movdqu {$src, $dst|$dst, $src}",
1426 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1427 XS, Requires<[HasSSE2]>;
1430 let isTwoAddress = 1 in {
1432 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1433 bit Commutable = 0> {
1434 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1435 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1436 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1437 let isCommutable = Commutable;
1439 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1440 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1441 [(set VR128:$dst, (IntId VR128:$src1,
1442 (bitconvert (loadv2i64 addr:$src2))))]>;
1445 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1446 string OpcodeStr, Intrinsic IntId> {
1447 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1448 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1449 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1450 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1451 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1452 [(set VR128:$dst, (IntId VR128:$src1,
1453 (bitconvert (loadv2i64 addr:$src2))))]>;
1454 def ri : PDIi8<opc2, ImmForm, (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1455 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1456 [(set VR128:$dst, (IntId VR128:$src1,
1457 (scalar_to_vector (i32 imm:$src2))))]>;
1461 /// PDI_binop_rm - Simple SSE2 binary operator.
1462 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1463 ValueType OpVT, bit Commutable = 0> {
1464 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1465 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1466 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1467 let isCommutable = Commutable;
1469 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1470 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1471 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1472 (bitconvert (loadv2i64 addr:$src2)))))]>;
1475 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1477 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1478 /// to collapse (bitconvert VT to VT) into its operand.
1480 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1481 bit Commutable = 0> {
1482 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1483 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1484 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1485 let isCommutable = Commutable;
1487 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1488 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1489 [(set VR128:$dst, (OpNode VR128:$src1,(loadv2i64 addr:$src2)))]>;
1494 // 128-bit Integer Arithmetic
1496 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1497 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1498 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1499 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1501 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1502 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1503 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1504 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1506 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1507 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1508 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1509 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1511 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1512 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1513 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1514 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1516 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1518 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1519 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1520 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1522 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1524 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1525 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1528 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1529 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1530 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1531 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1532 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1535 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw", int_x86_sse2_psll_w>;
1536 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld", int_x86_sse2_psll_d>;
1537 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq", int_x86_sse2_psll_q>;
1539 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw", int_x86_sse2_psrl_w>;
1540 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld", int_x86_sse2_psrl_d>;
1541 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq", int_x86_sse2_psrl_q>;
1543 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw", int_x86_sse2_psra_w>;
1544 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad", int_x86_sse2_psra_d>;
1545 // PSRAQ doesn't exist in SSE[1-3].
1547 // 128-bit logical shifts.
1548 let isTwoAddress = 1 in {
1549 def PSLLDQri : PDIi8<0x73, MRM7r,
1550 (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1551 "pslldq {$src2, $dst|$dst, $src2}", []>;
1552 def PSRLDQri : PDIi8<0x73, MRM3r,
1553 (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1554 "psrldq {$src2, $dst|$dst, $src2}", []>;
1555 // PSRADQri doesn't exist in SSE[1-3].
1558 let Predicates = [HasSSE2] in {
1559 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1560 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1561 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1562 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1563 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1564 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1568 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1569 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1570 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1572 let isTwoAddress = 1 in {
1573 def PANDNrr : PDI<0xDF, MRMSrcReg,
1574 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1575 "pandn {$src2, $dst|$dst, $src2}",
1576 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1579 def PANDNrm : PDI<0xDF, MRMSrcMem,
1580 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1581 "pandn {$src2, $dst|$dst, $src2}",
1582 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1583 (load addr:$src2))))]>;
1586 // SSE2 Integer comparison
1587 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1588 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1589 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1590 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1591 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1592 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1594 // Pack instructions
1595 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1596 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1597 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1599 // Shuffle and unpack instructions
1600 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1601 (ops VR128:$dst, VR128:$src1, i8imm:$src2),
1602 "pshufd {$src2, $src1, $dst|$dst, $src1, $src2}",
1603 [(set VR128:$dst, (v4i32 (vector_shuffle
1604 VR128:$src1, (undef),
1605 PSHUFD_shuffle_mask:$src2)))]>;
1606 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1607 (ops VR128:$dst, i128mem:$src1, i8imm:$src2),
1608 "pshufd {$src2, $src1, $dst|$dst, $src1, $src2}",
1609 [(set VR128:$dst, (v4i32 (vector_shuffle
1610 (bc_v4i32(loadv2i64 addr:$src1)),
1612 PSHUFD_shuffle_mask:$src2)))]>;
1614 // SSE2 with ImmT == Imm8 and XS prefix.
1615 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1616 (ops VR128:$dst, VR128:$src1, i8imm:$src2),
1617 "pshufhw {$src2, $src1, $dst|$dst, $src1, $src2}",
1618 [(set VR128:$dst, (v8i16 (vector_shuffle
1619 VR128:$src1, (undef),
1620 PSHUFHW_shuffle_mask:$src2)))]>,
1621 XS, Requires<[HasSSE2]>;
1622 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
1623 (ops VR128:$dst, i128mem:$src1, i8imm:$src2),
1624 "pshufhw {$src2, $src1, $dst|$dst, $src1, $src2}",
1625 [(set VR128:$dst, (v8i16 (vector_shuffle
1626 (bc_v8i16 (loadv2i64 addr:$src1)),
1628 PSHUFHW_shuffle_mask:$src2)))]>,
1629 XS, Requires<[HasSSE2]>;
1631 // SSE2 with ImmT == Imm8 and XD prefix.
1632 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
1633 (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1634 "pshuflw {$src2, $src1, $dst|$dst, $src1, $src2}",
1635 [(set VR128:$dst, (v8i16 (vector_shuffle
1636 VR128:$src1, (undef),
1637 PSHUFLW_shuffle_mask:$src2)))]>,
1638 XD, Requires<[HasSSE2]>;
1639 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
1640 (ops VR128:$dst, i128mem:$src1, i32i8imm:$src2),
1641 "pshuflw {$src2, $src1, $dst|$dst, $src1, $src2}",
1642 [(set VR128:$dst, (v8i16 (vector_shuffle
1643 (bc_v8i16 (loadv2i64 addr:$src1)),
1645 PSHUFLW_shuffle_mask:$src2)))]>,
1646 XD, Requires<[HasSSE2]>;
1649 let isTwoAddress = 1 in {
1650 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
1651 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1652 "punpcklbw {$src2, $dst|$dst, $src2}",
1654 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1655 UNPCKL_shuffle_mask)))]>;
1656 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
1657 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1658 "punpcklbw {$src2, $dst|$dst, $src2}",
1660 (v16i8 (vector_shuffle VR128:$src1,
1661 (bc_v16i8 (loadv2i64 addr:$src2)),
1662 UNPCKL_shuffle_mask)))]>;
1663 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
1664 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1665 "punpcklwd {$src2, $dst|$dst, $src2}",
1667 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1668 UNPCKL_shuffle_mask)))]>;
1669 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
1670 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1671 "punpcklwd {$src2, $dst|$dst, $src2}",
1673 (v8i16 (vector_shuffle VR128:$src1,
1674 (bc_v8i16 (loadv2i64 addr:$src2)),
1675 UNPCKL_shuffle_mask)))]>;
1676 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
1677 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1678 "punpckldq {$src2, $dst|$dst, $src2}",
1680 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1681 UNPCKL_shuffle_mask)))]>;
1682 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
1683 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1684 "punpckldq {$src2, $dst|$dst, $src2}",
1686 (v4i32 (vector_shuffle VR128:$src1,
1687 (bc_v4i32 (loadv2i64 addr:$src2)),
1688 UNPCKL_shuffle_mask)))]>;
1689 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
1690 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1691 "punpcklqdq {$src2, $dst|$dst, $src2}",
1693 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1694 UNPCKL_shuffle_mask)))]>;
1695 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
1696 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1697 "punpcklqdq {$src2, $dst|$dst, $src2}",
1699 (v2i64 (vector_shuffle VR128:$src1,
1700 (loadv2i64 addr:$src2),
1701 UNPCKL_shuffle_mask)))]>;
1703 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
1704 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1705 "punpckhbw {$src2, $dst|$dst, $src2}",
1707 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1708 UNPCKH_shuffle_mask)))]>;
1709 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
1710 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1711 "punpckhbw {$src2, $dst|$dst, $src2}",
1713 (v16i8 (vector_shuffle VR128:$src1,
1714 (bc_v16i8 (loadv2i64 addr:$src2)),
1715 UNPCKH_shuffle_mask)))]>;
1716 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
1717 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1718 "punpckhwd {$src2, $dst|$dst, $src2}",
1720 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1721 UNPCKH_shuffle_mask)))]>;
1722 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
1723 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1724 "punpckhwd {$src2, $dst|$dst, $src2}",
1726 (v8i16 (vector_shuffle VR128:$src1,
1727 (bc_v8i16 (loadv2i64 addr:$src2)),
1728 UNPCKH_shuffle_mask)))]>;
1729 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
1730 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1731 "punpckhdq {$src2, $dst|$dst, $src2}",
1733 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1734 UNPCKH_shuffle_mask)))]>;
1735 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
1736 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1737 "punpckhdq {$src2, $dst|$dst, $src2}",
1739 (v4i32 (vector_shuffle VR128:$src1,
1740 (bc_v4i32 (loadv2i64 addr:$src2)),
1741 UNPCKH_shuffle_mask)))]>;
1742 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
1743 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1744 "punpckhqdq {$src2, $dst|$dst, $src2}",
1746 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1747 UNPCKH_shuffle_mask)))]>;
1748 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
1749 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1750 "punpckhqdq {$src2, $dst|$dst, $src2}",
1752 (v2i64 (vector_shuffle VR128:$src1,
1753 (loadv2i64 addr:$src2),
1754 UNPCKH_shuffle_mask)))]>;
1758 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
1759 (ops GR32:$dst, VR128:$src1, i32i8imm:$src2),
1760 "pextrw {$src2, $src1, $dst|$dst, $src1, $src2}",
1761 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
1762 (iPTR imm:$src2)))]>;
1763 let isTwoAddress = 1 in {
1764 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
1765 (ops VR128:$dst, VR128:$src1,
1766 GR32:$src2, i32i8imm:$src3),
1767 "pinsrw {$src3, $src2, $dst|$dst, $src2, $src3}",
1769 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
1770 GR32:$src2, (iPTR imm:$src3))))]>;
1771 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
1772 (ops VR128:$dst, VR128:$src1,
1773 i16mem:$src2, i32i8imm:$src3),
1774 "pinsrw {$src3, $src2, $dst|$dst, $src2, $src3}",
1776 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
1777 (i32 (anyext (loadi16 addr:$src2))),
1778 (iPTR imm:$src3))))]>;
1782 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (ops GR32:$dst, VR128:$src),
1783 "pmovmskb {$src, $dst|$dst, $src}",
1784 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
1786 // Conditional store
1787 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (ops VR128:$src, VR128:$mask),
1788 "maskmovdqu {$mask, $src|$src, $mask}",
1789 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>,
1792 // Non-temporal stores
1793 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1794 "movntpd {$src, $dst|$dst, $src}",
1795 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
1796 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (ops f128mem:$dst, VR128:$src),
1797 "movntdq {$src, $dst|$dst, $src}",
1798 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
1799 def MOVNTImr : I<0xC3, MRMDestMem, (ops i32mem:$dst, GR32:$src),
1800 "movnti {$src, $dst|$dst, $src}",
1801 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
1802 TB, Requires<[HasSSE2]>;
1805 def CLFLUSH : I<0xAE, MRM7m, (ops i8mem:$src),
1806 "clflush $src", [(int_x86_sse2_clflush addr:$src)]>,
1807 TB, Requires<[HasSSE2]>;
1809 // Load, store, and memory fence
1810 def LFENCE : I<0xAE, MRM5m, (ops),
1811 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
1812 def MFENCE : I<0xAE, MRM6m, (ops),
1813 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
1816 // Alias instructions that map zero vector to pxor / xorp* for sse.
1817 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
1818 let isReMaterializable = 1 in
1819 def V_SETALLONES : PDI<0x76, MRMInitReg, (ops VR128:$dst),
1820 "pcmpeqd $dst, $dst",
1821 [(set VR128:$dst, (v2f64 immAllOnesV))]>;
1823 // FR64 to 128-bit vector conversion.
1824 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (ops VR128:$dst, FR64:$src),
1825 "movsd {$src, $dst|$dst, $src}",
1827 (v2f64 (scalar_to_vector FR64:$src)))]>;
1828 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1829 "movsd {$src, $dst|$dst, $src}",
1831 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
1833 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (ops VR128:$dst, GR32:$src),
1834 "movd {$src, $dst|$dst, $src}",
1836 (v4i32 (scalar_to_vector GR32:$src)))]>;
1837 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (ops VR128:$dst, i32mem:$src),
1838 "movd {$src, $dst|$dst, $src}",
1840 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
1842 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (ops FR32:$dst, GR32:$src),
1843 "movd {$src, $dst|$dst, $src}",
1844 [(set FR32:$dst, (bitconvert GR32:$src))]>;
1846 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (ops FR32:$dst, i32mem:$src),
1847 "movd {$src, $dst|$dst, $src}",
1848 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
1850 // SSE2 instructions with XS prefix
1851 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1852 "movq {$src, $dst|$dst, $src}",
1854 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
1855 Requires<[HasSSE2]>;
1856 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (ops i64mem:$dst, VR128:$src),
1857 "movq {$src, $dst|$dst, $src}",
1858 [(store (i64 (vector_extract (v2i64 VR128:$src),
1859 (iPTR 0))), addr:$dst)]>;
1861 // FIXME: may not be able to eliminate this movss with coalescing the src and
1862 // dest register classes are different. We really want to write this pattern
1864 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1865 // (f32 FR32:$src)>;
1866 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (ops FR64:$dst, VR128:$src),
1867 "movsd {$src, $dst|$dst, $src}",
1868 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
1870 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (ops f64mem:$dst, VR128:$src),
1871 "movsd {$src, $dst|$dst, $src}",
1872 [(store (f64 (vector_extract (v2f64 VR128:$src),
1873 (iPTR 0))), addr:$dst)]>;
1874 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (ops GR32:$dst, VR128:$src),
1875 "movd {$src, $dst|$dst, $src}",
1876 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
1878 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (ops i32mem:$dst, VR128:$src),
1879 "movd {$src, $dst|$dst, $src}",
1880 [(store (i32 (vector_extract (v4i32 VR128:$src),
1881 (iPTR 0))), addr:$dst)]>;
1883 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (ops GR32:$dst, FR32:$src),
1884 "movd {$src, $dst|$dst, $src}",
1885 [(set GR32:$dst, (bitconvert FR32:$src))]>;
1886 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (ops i32mem:$dst, FR32:$src),
1887 "movd {$src, $dst|$dst, $src}",
1888 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
1891 // Move to lower bits of a VR128, leaving upper bits alone.
1892 // Three operand (but two address) aliases.
1893 let isTwoAddress = 1 in {
1894 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
1895 (ops VR128:$dst, VR128:$src1, FR64:$src2),
1896 "movsd {$src2, $dst|$dst, $src2}", []>;
1898 let AddedComplexity = 15 in
1899 def MOVLPDrr : SDI<0x10, MRMSrcReg,
1900 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1901 "movsd {$src2, $dst|$dst, $src2}",
1903 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
1904 MOVL_shuffle_mask)))]>;
1907 // Store / copy lower 64-bits of a XMM register.
1908 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (ops i64mem:$dst, VR128:$src),
1909 "movq {$src, $dst|$dst, $src}",
1910 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
1912 // Move to lower bits of a VR128 and zeroing upper bits.
1913 // Loading from memory automatically zeroing upper bits.
1914 let AddedComplexity = 20 in
1915 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1916 "movsd {$src, $dst|$dst, $src}",
1918 (v2f64 (vector_shuffle immAllZerosV,
1919 (v2f64 (scalar_to_vector
1920 (loadf64 addr:$src))),
1921 MOVL_shuffle_mask)))]>;
1923 let AddedComplexity = 15 in
1924 // movd / movq to XMM register zero-extends
1925 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (ops VR128:$dst, GR32:$src),
1926 "movd {$src, $dst|$dst, $src}",
1928 (v4i32 (vector_shuffle immAllZerosV,
1929 (v4i32 (scalar_to_vector GR32:$src)),
1930 MOVL_shuffle_mask)))]>;
1931 let AddedComplexity = 20 in
1932 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (ops VR128:$dst, i32mem:$src),
1933 "movd {$src, $dst|$dst, $src}",
1935 (v4i32 (vector_shuffle immAllZerosV,
1936 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
1937 MOVL_shuffle_mask)))]>;
1939 // Moving from XMM to XMM but still clear upper 64 bits.
1940 let AddedComplexity = 15 in
1941 def MOVZQI2PQIrr : I<0x7E, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1942 "movq {$src, $dst|$dst, $src}",
1943 [(set VR128:$dst, (int_x86_sse2_movl_dq VR128:$src))]>,
1944 XS, Requires<[HasSSE2]>;
1945 let AddedComplexity = 20 in
1946 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1947 "movq {$src, $dst|$dst, $src}",
1948 [(set VR128:$dst, (int_x86_sse2_movl_dq
1949 (bitconvert (loadv2i64 addr:$src))))]>,
1950 XS, Requires<[HasSSE2]>;
1953 //===----------------------------------------------------------------------===//
1954 // SSE3 Instructions
1955 //===----------------------------------------------------------------------===//
1957 // SSE3 Instruction Templates:
1959 // S3I - SSE3 instructions with TB and OpSize prefixes.
1960 // S3SI - SSE3 instructions with XS prefix.
1961 // S3DI - SSE3 instructions with XD prefix.
1963 class S3SI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
1964 : I<o, F, ops, asm, pattern>, XS, Requires<[HasSSE3]>;
1965 class S3DI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
1966 : I<o, F, ops, asm, pattern>, XD, Requires<[HasSSE3]>;
1967 class S3I<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
1968 : I<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE3]>;
1970 // Move Instructions
1971 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1972 "movshdup {$src, $dst|$dst, $src}",
1973 [(set VR128:$dst, (v4f32 (vector_shuffle
1974 VR128:$src, (undef),
1975 MOVSHDUP_shuffle_mask)))]>;
1976 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1977 "movshdup {$src, $dst|$dst, $src}",
1978 [(set VR128:$dst, (v4f32 (vector_shuffle
1979 (loadv4f32 addr:$src), (undef),
1980 MOVSHDUP_shuffle_mask)))]>;
1982 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1983 "movsldup {$src, $dst|$dst, $src}",
1984 [(set VR128:$dst, (v4f32 (vector_shuffle
1985 VR128:$src, (undef),
1986 MOVSLDUP_shuffle_mask)))]>;
1987 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
1988 "movsldup {$src, $dst|$dst, $src}",
1989 [(set VR128:$dst, (v4f32 (vector_shuffle
1990 (loadv4f32 addr:$src), (undef),
1991 MOVSLDUP_shuffle_mask)))]>;
1993 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1994 "movddup {$src, $dst|$dst, $src}",
1995 [(set VR128:$dst, (v2f64 (vector_shuffle
1996 VR128:$src, (undef),
1997 SSE_splat_lo_mask)))]>;
1998 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1999 "movddup {$src, $dst|$dst, $src}",
2001 (v2f64 (vector_shuffle
2002 (scalar_to_vector (loadf64 addr:$src)),
2004 SSE_splat_lo_mask)))]>;
2007 let isTwoAddress = 1 in {
2008 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2009 (ops VR128:$dst, VR128:$src1, VR128:$src2),
2010 "addsubps {$src2, $dst|$dst, $src2}",
2011 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2013 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2014 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
2015 "addsubps {$src2, $dst|$dst, $src2}",
2016 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2017 (load addr:$src2)))]>;
2018 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2019 (ops VR128:$dst, VR128:$src1, VR128:$src2),
2020 "addsubpd {$src2, $dst|$dst, $src2}",
2021 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2023 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2024 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
2025 "addsubpd {$src2, $dst|$dst, $src2}",
2026 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2027 (load addr:$src2)))]>;
2030 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
2031 "lddqu {$src, $dst|$dst, $src}",
2032 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2035 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2036 : S3DI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
2037 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2038 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2039 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2040 : S3DI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
2041 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2042 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
2043 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2044 : S3I<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
2045 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2046 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2047 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2048 : S3I<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
2049 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2050 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
2052 let isTwoAddress = 1 in {
2053 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2054 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2055 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2056 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2057 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2058 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2059 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2060 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2063 // Thread synchronization
2064 def MONITOR : I<0xC8, RawFrm, (ops), "monitor",
2065 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2066 def MWAIT : I<0xC9, RawFrm, (ops), "mwait",
2067 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2069 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2070 let AddedComplexity = 15 in
2071 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2072 MOVSHDUP_shuffle_mask)),
2073 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2074 let AddedComplexity = 20 in
2075 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (loadv2i64 addr:$src)), (undef),
2076 MOVSHDUP_shuffle_mask)),
2077 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2079 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2080 let AddedComplexity = 15 in
2081 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2082 MOVSLDUP_shuffle_mask)),
2083 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2084 let AddedComplexity = 20 in
2085 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (loadv2i64 addr:$src)), (undef),
2086 MOVSLDUP_shuffle_mask)),
2087 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2089 //===----------------------------------------------------------------------===//
2090 // SSSE3 Instructions
2091 //===----------------------------------------------------------------------===//
2093 // SSE3 Instruction Templates:
2095 // SS38I - SSSE3 instructions with T8 and OpSize prefixes.
2096 // SS3AI - SSSE3 instructions with TA and OpSize prefixes.
2098 class SS38I<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
2099 : I<o, F, ops, asm, pattern>, T8, OpSize, Requires<[HasSSSE3]>;
2100 class SS3AI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
2101 : I<o, F, ops, asm, pattern>, TA, OpSize, Requires<[HasSSSE3]>;
2103 /// SS3I_binop_rm_int - Simple SSSE3 binary operatr whose type is v2i64.
2104 let isTwoAddress = 1 in {
2105 multiclass SS3I_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
2106 bit Commutable = 0> {
2107 def rr : SS38I<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
2108 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2109 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
2110 let isCommutable = Commutable;
2112 def rm : SS38I<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
2113 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
2116 (bitconvert (loadv2i64 addr:$src2))))]>;
2120 defm PMULHRSW128 : SS3I_binop_rm_int<0x0B, "pmulhrsw",
2121 int_x86_ssse3_pmulhrsw_128, 1>;
2123 //===----------------------------------------------------------------------===//
2124 // Non-Instruction Patterns
2125 //===----------------------------------------------------------------------===//
2127 // 128-bit vector undef's.
2128 def : Pat<(v2f64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2129 def : Pat<(v16i8 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2130 def : Pat<(v8i16 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2131 def : Pat<(v4i32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2132 def : Pat<(v2i64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2134 // 128-bit vector all zero's.
2135 def : Pat<(v16i8 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2136 def : Pat<(v8i16 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2137 def : Pat<(v4i32 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2138 def : Pat<(v2i64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2139 def : Pat<(v2f64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2141 // 128-bit vector all one's.
2142 def : Pat<(v16i8 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2143 def : Pat<(v8i16 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2144 def : Pat<(v4i32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2145 def : Pat<(v2i64 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2146 def : Pat<(v4f32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE1]>;
2148 // Store 128-bit integer vector values.
2149 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
2150 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2151 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
2152 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2153 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
2154 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2156 // Scalar to v8i16 / v16i8. The source may be a GR32, but only the lower 8 or
2158 def : Pat<(v8i16 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2159 Requires<[HasSSE2]>;
2160 def : Pat<(v16i8 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2161 Requires<[HasSSE2]>;
2164 let Predicates = [HasSSE2] in {
2165 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2166 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2167 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2168 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2169 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2170 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2171 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2172 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2173 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2174 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2175 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2176 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2177 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2178 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2179 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2180 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2181 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2182 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2183 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2184 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2185 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2186 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2187 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2188 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2189 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2190 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2191 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2192 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2193 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2194 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2197 // Move scalar to XMM zero-extended
2198 // movd to XMM register zero-extends
2199 let AddedComplexity = 15 in {
2200 def : Pat<(v8i16 (vector_shuffle immAllZerosV,
2201 (v8i16 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
2202 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
2203 def : Pat<(v16i8 (vector_shuffle immAllZerosV,
2204 (v16i8 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
2205 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
2206 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2207 def : Pat<(v2f64 (vector_shuffle immAllZerosV,
2208 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
2209 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2210 def : Pat<(v4f32 (vector_shuffle immAllZerosV,
2211 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
2212 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
2215 // Splat v2f64 / v2i64
2216 let AddedComplexity = 10 in {
2217 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2218 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2219 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2220 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2221 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2222 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2223 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2224 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2228 def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SSE_splat_mask:$sm),
2229 (SHUFPSrri VR128:$src, VR128:$src, SSE_splat_mask:$sm)>,
2230 Requires<[HasSSE1]>;
2232 // Special unary SHUFPSrri case.
2233 // FIXME: when we want non two-address code, then we should use PSHUFD?
2234 def : Pat<(vector_shuffle (v4f32 VR128:$src1), (undef),
2235 SHUFP_unary_shuffle_mask:$sm),
2236 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2237 Requires<[HasSSE1]>;
2238 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2239 def : Pat<(vector_shuffle (loadv4f32 addr:$src1), (undef),
2240 SHUFP_unary_shuffle_mask:$sm),
2241 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2242 Requires<[HasSSE2]>;
2243 // Special binary v4i32 shuffle cases with SHUFPS.
2244 def : Pat<(vector_shuffle (v4i32 VR128:$src1), (v4i32 VR128:$src2),
2245 PSHUFD_binary_shuffle_mask:$sm),
2246 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2247 Requires<[HasSSE2]>;
2248 def : Pat<(vector_shuffle (v4i32 VR128:$src1),
2249 (bc_v4i32 (loadv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm),
2250 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2251 Requires<[HasSSE2]>;
2253 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2254 let AddedComplexity = 10 in {
2255 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2256 UNPCKL_v_undef_shuffle_mask)),
2257 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2258 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2259 UNPCKL_v_undef_shuffle_mask)),
2260 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2261 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2262 UNPCKL_v_undef_shuffle_mask)),
2263 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2264 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2265 UNPCKL_v_undef_shuffle_mask)),
2266 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2269 let AddedComplexity = 15 in {
2270 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2271 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2272 MOVHP_shuffle_mask)),
2273 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2275 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2276 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2277 MOVHLPS_shuffle_mask)),
2278 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2280 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2281 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2282 MOVHLPS_v_undef_shuffle_mask)),
2283 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2284 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2285 MOVHLPS_v_undef_shuffle_mask)),
2286 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2289 let AddedComplexity = 20 in {
2290 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2291 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2292 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (loadv4f32 addr:$src2),
2293 MOVLP_shuffle_mask)),
2294 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2295 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (loadv2f64 addr:$src2),
2296 MOVLP_shuffle_mask)),
2297 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2298 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (loadv4f32 addr:$src2),
2299 MOVHP_shuffle_mask)),
2300 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2301 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (loadv2f64 addr:$src2),
2302 MOVHP_shuffle_mask)),
2303 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2305 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)),
2306 MOVLP_shuffle_mask)),
2307 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2308 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (loadv2i64 addr:$src2),
2309 MOVLP_shuffle_mask)),
2310 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2311 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)),
2312 MOVHP_shuffle_mask)),
2313 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2314 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (loadv2i64 addr:$src2),
2315 MOVLP_shuffle_mask)),
2316 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2319 let AddedComplexity = 15 in {
2320 // Setting the lowest element in the vector.
2321 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2322 MOVL_shuffle_mask)),
2323 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2324 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2325 MOVL_shuffle_mask)),
2326 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2328 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2329 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2330 MOVLP_shuffle_mask)),
2331 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2332 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2333 MOVLP_shuffle_mask)),
2334 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2337 // Set lowest element and zero upper elements.
2338 let AddedComplexity = 20 in
2339 def : Pat<(bc_v2i64 (vector_shuffle immAllZerosV,
2340 (v2f64 (scalar_to_vector (loadf64 addr:$src))),
2341 MOVL_shuffle_mask)),
2342 (MOVZQI2PQIrm addr:$src)>, Requires<[HasSSE2]>;
2344 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2345 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2346 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2347 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2348 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2349 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2350 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2351 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2352 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2353 Requires<[HasSSE2]>;
2354 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2355 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2356 Requires<[HasSSE2]>;
2357 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2358 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2359 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2360 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2361 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2362 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2363 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2364 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2365 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
2366 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2367 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
2368 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2369 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
2370 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2371 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
2372 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2374 // Some special case pandn patterns.
2375 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2377 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2378 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2380 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2381 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2383 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2385 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2386 (load addr:$src2))),
2387 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2388 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2389 (load addr:$src2))),
2390 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2391 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2392 (load addr:$src2))),
2393 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2396 def : Pat<(v4f32 (X86loadu addr:$src)), (MOVUPSrm addr:$src)>,
2397 Requires<[HasSSE1]>;