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 X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
25 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
26 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
27 [SDNPCommutative, SDNPAssociative]>;
28 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
29 [SDNPCommutative, SDNPAssociative]>;
30 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
31 [SDNPCommutative, SDNPAssociative]>;
32 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
33 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
34 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
35 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest,
36 [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, (outs VR128:$dst), (ins),
49 [(set VR128:$dst, (v4f32 (undef)))]>,
51 def IMPLICIT_DEF_FR32 : I<0, Pseudo, (outs FR32:$dst), (ins),
53 [(set FR32:$dst, (undef))]>, Requires<[HasSSE2]>;
54 def IMPLICIT_DEF_FR64 : I<0, Pseudo, (outs FR64:$dst), (ins),
56 [(set FR64:$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 loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
84 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
85 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
86 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
88 // Like 'store', but always requires vector alignment.
89 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
90 (st node:$val, node:$ptr), [{
91 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
92 return !ST->isTruncatingStore() &&
93 ST->getAddressingMode() == ISD::UNINDEXED &&
94 ST->getAlignment() >= 16;
98 // Like 'load', but always requires vector alignment.
99 def alignedload : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
100 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
101 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
102 LD->getAddressingMode() == ISD::UNINDEXED &&
103 LD->getAlignment() >= 16;
107 def alignedloadfsf32 : PatFrag<(ops node:$ptr), (f32 (alignedload node:$ptr))>;
108 def alignedloadfsf64 : PatFrag<(ops node:$ptr), (f64 (alignedload node:$ptr))>;
109 def alignedloadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (alignedload node:$ptr))>;
110 def alignedloadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (alignedload node:$ptr))>;
111 def alignedloadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (alignedload node:$ptr))>;
112 def alignedloadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (alignedload node:$ptr))>;
114 // Like 'load', but uses special alignment checks suitable for use in
115 // memory operands in most SSE instructions, which are required to
116 // be naturally aligned on some targets but not on others.
117 // FIXME: Actually implement support for targets that don't require the
118 // alignment. This probably wants a subtarget predicate.
119 def memop : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
120 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
121 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
122 LD->getAddressingMode() == ISD::UNINDEXED &&
123 LD->getAlignment() >= 16;
127 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
128 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
129 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
130 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
131 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
132 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
134 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
136 def memop64 : PatFrag<(ops node:$ptr), (ld node:$ptr), [{
137 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
138 return LD->getExtensionType() == ISD::NON_EXTLOAD &&
139 LD->getAddressingMode() == ISD::UNINDEXED &&
140 LD->getAlignment() >= 8;
144 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
145 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop64 node:$ptr))>;
146 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
147 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
148 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
150 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
151 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
152 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
153 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
154 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
155 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
157 def fp32imm0 : PatLeaf<(f32 fpimm), [{
158 return N->isExactlyValue(+0.0);
161 def PSxLDQ_imm : SDNodeXForm<imm, [{
162 // Transformation function: imm >> 3
163 return getI32Imm(N->getValue() >> 3);
166 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
168 def SHUFFLE_get_shuf_imm : SDNodeXForm<build_vector, [{
169 return getI8Imm(X86::getShuffleSHUFImmediate(N));
172 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
174 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<build_vector, [{
175 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
178 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
180 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<build_vector, [{
181 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
184 def SSE_splat_mask : PatLeaf<(build_vector), [{
185 return X86::isSplatMask(N);
186 }], SHUFFLE_get_shuf_imm>;
188 def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
189 return X86::isSplatLoMask(N);
192 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
193 return X86::isMOVHLPSMask(N);
196 def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
197 return X86::isMOVHLPS_v_undef_Mask(N);
200 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
201 return X86::isMOVHPMask(N);
204 def MOVLP_shuffle_mask : PatLeaf<(build_vector), [{
205 return X86::isMOVLPMask(N);
208 def MOVL_shuffle_mask : PatLeaf<(build_vector), [{
209 return X86::isMOVLMask(N);
212 def MOVSHDUP_shuffle_mask : PatLeaf<(build_vector), [{
213 return X86::isMOVSHDUPMask(N);
216 def MOVSLDUP_shuffle_mask : PatLeaf<(build_vector), [{
217 return X86::isMOVSLDUPMask(N);
220 def UNPCKL_shuffle_mask : PatLeaf<(build_vector), [{
221 return X86::isUNPCKLMask(N);
224 def UNPCKH_shuffle_mask : PatLeaf<(build_vector), [{
225 return X86::isUNPCKHMask(N);
228 def UNPCKL_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
229 return X86::isUNPCKL_v_undef_Mask(N);
232 def UNPCKH_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
233 return X86::isUNPCKH_v_undef_Mask(N);
236 def PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
237 return X86::isPSHUFDMask(N);
238 }], SHUFFLE_get_shuf_imm>;
240 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
241 return X86::isPSHUFHWMask(N);
242 }], SHUFFLE_get_pshufhw_imm>;
244 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
245 return X86::isPSHUFLWMask(N);
246 }], SHUFFLE_get_pshuflw_imm>;
248 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
249 return X86::isPSHUFDMask(N);
250 }], SHUFFLE_get_shuf_imm>;
252 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
253 return X86::isSHUFPMask(N);
254 }], SHUFFLE_get_shuf_imm>;
256 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
257 return X86::isSHUFPMask(N);
258 }], SHUFFLE_get_shuf_imm>;
260 //===----------------------------------------------------------------------===//
261 // SSE scalar FP Instructions
262 //===----------------------------------------------------------------------===//
264 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
265 // scheduler into a branch sequence.
266 let usesCustomDAGSchedInserter = 1 in { // Expanded by the scheduler.
267 def CMOV_FR32 : I<0, Pseudo,
268 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
269 "#CMOV_FR32 PSEUDO!",
270 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond))]>;
271 def CMOV_FR64 : I<0, Pseudo,
272 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
273 "#CMOV_FR64 PSEUDO!",
274 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond))]>;
275 def CMOV_V4F32 : I<0, Pseudo,
276 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
277 "#CMOV_V4F32 PSEUDO!",
279 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
280 def CMOV_V2F64 : I<0, Pseudo,
281 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
282 "#CMOV_V2F64 PSEUDO!",
284 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
285 def CMOV_V2I64 : I<0, Pseudo,
286 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
287 "#CMOV_V2I64 PSEUDO!",
289 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
292 //===----------------------------------------------------------------------===//
294 //===----------------------------------------------------------------------===//
297 def MOVSSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
298 "movss\t{$src, $dst|$dst, $src}", []>;
299 let isLoad = 1, isReMaterializable = 1 in
300 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
301 "movss\t{$src, $dst|$dst, $src}",
302 [(set FR32:$dst, (loadf32 addr:$src))]>;
303 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
304 "movss\t{$src, $dst|$dst, $src}",
305 [(store FR32:$src, addr:$dst)]>;
307 // Conversion instructions
308 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
309 "cvttss2si\t{$src, $dst|$dst, $src}",
310 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
311 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
312 "cvttss2si\t{$src, $dst|$dst, $src}",
313 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
314 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
315 "cvtsi2ss\t{$src, $dst|$dst, $src}",
316 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
317 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
318 "cvtsi2ss\t{$src, $dst|$dst, $src}",
319 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
321 // Match intrinsics which expect XMM operand(s).
322 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
323 "cvtss2si\t{$src, $dst|$dst, $src}",
324 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
325 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
326 "cvtss2si\t{$src, $dst|$dst, $src}",
327 [(set GR32:$dst, (int_x86_sse_cvtss2si
328 (load addr:$src)))]>;
330 // Aliases for intrinsics
331 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
332 "cvttss2si\t{$src, $dst|$dst, $src}",
334 (int_x86_sse_cvttss2si VR128:$src))]>;
335 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
336 "cvttss2si\t{$src, $dst|$dst, $src}",
338 (int_x86_sse_cvttss2si(load addr:$src)))]>;
340 let isTwoAddress = 1 in {
341 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
342 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
343 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
344 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
346 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
347 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
348 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
349 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
350 (loadi32 addr:$src2)))]>;
353 // Comparison instructions
354 let isTwoAddress = 1 in {
355 def CMPSSrr : SSI<0xC2, MRMSrcReg,
356 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
357 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
358 def CMPSSrm : SSI<0xC2, MRMSrcMem,
359 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
360 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
363 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
364 "ucomiss\t{$src2, $src1|$src1, $src2}",
365 [(X86cmp FR32:$src1, FR32:$src2)]>;
366 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
367 "ucomiss\t{$src2, $src1|$src1, $src2}",
368 [(X86cmp FR32:$src1, (loadf32 addr:$src2))]>;
370 // Aliases to match intrinsics which expect XMM operand(s).
371 let isTwoAddress = 1 in {
372 def Int_CMPSSrr : SSI<0xC2, MRMSrcReg,
373 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
374 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
375 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
376 VR128:$src, imm:$cc))]>;
377 def Int_CMPSSrm : SSI<0xC2, MRMSrcMem,
378 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src, SSECC:$cc),
379 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
380 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
381 (load addr:$src), imm:$cc))]>;
384 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
385 "ucomiss\t{$src2, $src1|$src1, $src2}",
386 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2)]>;
387 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
388 "ucomiss\t{$src2, $src1|$src1, $src2}",
389 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2))]>;
391 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
392 "comiss\t{$src2, $src1|$src1, $src2}",
393 [(X86comi (v4f32 VR128:$src1), VR128:$src2)]>;
394 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
395 "comiss\t{$src2, $src1|$src1, $src2}",
396 [(X86comi (v4f32 VR128:$src1), (load addr:$src2))]>;
398 // Aliases of packed SSE1 instructions for scalar use. These all have names that
401 // Alias instructions that map fld0 to pxor for sse.
402 let isReMaterializable = 1 in
403 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
404 "pxor\t$dst, $dst", [(set FR32:$dst, fp32imm0)]>,
405 Requires<[HasSSE1]>, TB, OpSize;
407 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
409 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
410 "movaps\t{$src, $dst|$dst, $src}", []>;
412 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
415 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
416 "movaps\t{$src, $dst|$dst, $src}",
417 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
419 // Alias bitwise logical operations using SSE logical ops on packed FP values.
420 let isTwoAddress = 1 in {
421 let isCommutable = 1 in {
422 def FsANDPSrr : PSI<0x54, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
423 "andps\t{$src2, $dst|$dst, $src2}",
424 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
425 def FsORPSrr : PSI<0x56, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
426 "orps\t{$src2, $dst|$dst, $src2}",
427 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
428 def FsXORPSrr : PSI<0x57, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
429 "xorps\t{$src2, $dst|$dst, $src2}",
430 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
433 def FsANDPSrm : PSI<0x54, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
434 "andps\t{$src2, $dst|$dst, $src2}",
435 [(set FR32:$dst, (X86fand FR32:$src1,
436 (memopfsf32 addr:$src2)))]>;
437 def FsORPSrm : PSI<0x56, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
438 "orps\t{$src2, $dst|$dst, $src2}",
439 [(set FR32:$dst, (X86for FR32:$src1,
440 (memopfsf32 addr:$src2)))]>;
441 def FsXORPSrm : PSI<0x57, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
442 "xorps\t{$src2, $dst|$dst, $src2}",
443 [(set FR32:$dst, (X86fxor FR32:$src1,
444 (memopfsf32 addr:$src2)))]>;
446 def FsANDNPSrr : PSI<0x55, MRMSrcReg,
447 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
448 "andnps\t{$src2, $dst|$dst, $src2}", []>;
449 def FsANDNPSrm : PSI<0x55, MRMSrcMem,
450 (outs FR32:$dst), (ins FR32:$src1, f128mem:$src2),
451 "andnps\t{$src2, $dst|$dst, $src2}", []>;
454 /// basic_sse1_fp_binop_rm - SSE1 binops come in both scalar and vector forms.
456 /// In addition, we also have a special variant of the scalar form here to
457 /// represent the associated intrinsic operation. This form is unlike the
458 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
459 /// and leaves the top elements undefined.
461 /// These three forms can each be reg+reg or reg+mem, so there are a total of
462 /// six "instructions".
464 let isTwoAddress = 1 in {
465 multiclass basic_sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
466 SDNode OpNode, Intrinsic F32Int,
467 bit Commutable = 0> {
468 // Scalar operation, reg+reg.
469 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
470 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
471 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
472 let isCommutable = Commutable;
475 // Scalar operation, reg+mem.
476 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
477 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
478 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
480 // Vector operation, reg+reg.
481 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
482 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
483 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
484 let isCommutable = Commutable;
487 // Vector operation, reg+mem.
488 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
489 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
490 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
492 // Intrinsic operation, reg+reg.
493 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
494 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
495 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
496 let isCommutable = Commutable;
499 // Intrinsic operation, reg+mem.
500 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
501 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
502 [(set VR128:$dst, (F32Int VR128:$src1,
503 sse_load_f32:$src2))]>;
507 // Arithmetic instructions
508 defm ADD : basic_sse1_fp_binop_rm<0x58, "add", fadd, int_x86_sse_add_ss, 1>;
509 defm MUL : basic_sse1_fp_binop_rm<0x59, "mul", fmul, int_x86_sse_mul_ss, 1>;
510 defm SUB : basic_sse1_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse_sub_ss>;
511 defm DIV : basic_sse1_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse_div_ss>;
513 /// sse1_fp_binop_rm - Other SSE1 binops
515 /// This multiclass is like basic_sse1_fp_binop_rm, with the addition of
516 /// instructions for a full-vector intrinsic form. Operations that map
517 /// onto C operators don't use this form since they just use the plain
518 /// vector form instead of having a separate vector intrinsic form.
520 /// This provides a total of eight "instructions".
522 let isTwoAddress = 1 in {
523 multiclass sse1_fp_binop_rm<bits<8> opc, string OpcodeStr,
527 bit Commutable = 0> {
529 // Scalar operation, reg+reg.
530 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
531 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
532 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
533 let isCommutable = Commutable;
536 // Scalar operation, reg+mem.
537 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
538 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
539 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
541 // Vector operation, reg+reg.
542 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
543 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
544 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
545 let isCommutable = Commutable;
548 // Vector operation, reg+mem.
549 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
550 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
551 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
553 // Intrinsic operation, reg+reg.
554 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
555 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
556 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
557 let isCommutable = Commutable;
560 // Intrinsic operation, reg+mem.
561 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
562 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
563 [(set VR128:$dst, (F32Int VR128:$src1,
564 sse_load_f32:$src2))]>;
566 // Vector intrinsic operation, reg+reg.
567 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
568 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
569 [(set VR128:$dst, (V4F32Int VR128:$src1, VR128:$src2))]> {
570 let isCommutable = Commutable;
573 // Vector intrinsic operation, reg+mem.
574 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
575 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
576 [(set VR128:$dst, (V4F32Int VR128:$src1, (load addr:$src2)))]>;
580 defm MAX : sse1_fp_binop_rm<0x5F, "max", X86fmax,
581 int_x86_sse_max_ss, int_x86_sse_max_ps>;
582 defm MIN : sse1_fp_binop_rm<0x5D, "min", X86fmin,
583 int_x86_sse_min_ss, int_x86_sse_min_ps>;
585 //===----------------------------------------------------------------------===//
586 // SSE packed FP Instructions
589 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
590 "movaps\t{$src, $dst|$dst, $src}", []>;
591 let isLoad = 1, isReMaterializable = 1 in
592 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
593 "movaps\t{$src, $dst|$dst, $src}",
594 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
596 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
597 "movaps\t{$src, $dst|$dst, $src}",
598 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
600 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
601 "movups\t{$src, $dst|$dst, $src}", []>;
603 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
604 "movups\t{$src, $dst|$dst, $src}",
605 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
606 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
607 "movups\t{$src, $dst|$dst, $src}",
608 [(store (v4f32 VR128:$src), addr:$dst)]>;
610 // Intrinsic forms of MOVUPS load and store
612 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
613 "movups\t{$src, $dst|$dst, $src}",
614 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
615 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
616 "movups\t{$src, $dst|$dst, $src}",
617 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
619 let isTwoAddress = 1 in {
620 let AddedComplexity = 20 in {
621 def MOVLPSrm : PSI<0x12, MRMSrcMem,
622 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
623 "movlps\t{$src2, $dst|$dst, $src2}",
625 (v4f32 (vector_shuffle VR128:$src1,
626 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
627 MOVLP_shuffle_mask)))]>;
628 def MOVHPSrm : PSI<0x16, MRMSrcMem,
629 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
630 "movhps\t{$src2, $dst|$dst, $src2}",
632 (v4f32 (vector_shuffle VR128:$src1,
633 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
634 MOVHP_shuffle_mask)))]>;
638 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
639 "movlps\t{$src, $dst|$dst, $src}",
640 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
641 (iPTR 0))), addr:$dst)]>;
643 // v2f64 extract element 1 is always custom lowered to unpack high to low
644 // and extract element 0 so the non-store version isn't too horrible.
645 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
646 "movhps\t{$src, $dst|$dst, $src}",
647 [(store (f64 (vector_extract
648 (v2f64 (vector_shuffle
649 (bc_v2f64 (v4f32 VR128:$src)), (undef),
650 UNPCKH_shuffle_mask)), (iPTR 0))),
653 let isTwoAddress = 1 in {
654 let AddedComplexity = 15 in {
655 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
656 "movlhps\t{$src2, $dst|$dst, $src2}",
658 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
659 MOVHP_shuffle_mask)))]>;
661 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
662 "movhlps\t{$src2, $dst|$dst, $src2}",
664 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
665 MOVHLPS_shuffle_mask)))]>;
673 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
675 /// In addition, we also have a special variant of the scalar form here to
676 /// represent the associated intrinsic operation. This form is unlike the
677 /// plain scalar form, in that it takes an entire vector (instead of a
678 /// scalar) and leaves the top elements undefined.
680 /// And, we have a special variant form for a full-vector intrinsic form.
682 /// These four forms can each have a reg or a mem operand, so there are a
683 /// total of eight "instructions".
685 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
689 bit Commutable = 0> {
690 // Scalar operation, reg.
691 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
692 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
693 [(set FR32:$dst, (OpNode FR32:$src))]> {
694 let isCommutable = Commutable;
697 // Scalar operation, mem.
698 def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
699 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
700 [(set FR32:$dst, (OpNode (load addr:$src)))]>;
702 // Vector operation, reg.
703 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
704 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
705 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
706 let isCommutable = Commutable;
709 // Vector operation, mem.
710 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
711 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
712 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
714 // Intrinsic operation, reg.
715 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
716 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
717 [(set VR128:$dst, (F32Int VR128:$src))]> {
718 let isCommutable = Commutable;
721 // Intrinsic operation, mem.
722 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
723 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
724 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
726 // Vector intrinsic operation, reg
727 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
728 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
729 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
730 let isCommutable = Commutable;
733 // Vector intrinsic operation, mem
734 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
735 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
736 [(set VR128:$dst, (V4F32Int (load addr:$src)))]>;
740 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
741 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
743 // Reciprocal approximations. Note that these typically require refinement
744 // in order to obtain suitable precision.
745 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
746 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
747 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
748 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
751 let isTwoAddress = 1 in {
752 let isCommutable = 1 in {
753 def ANDPSrr : PSI<0x54, MRMSrcReg,
754 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
755 "andps\t{$src2, $dst|$dst, $src2}",
756 [(set VR128:$dst, (v2i64
757 (and VR128:$src1, VR128:$src2)))]>;
758 def ORPSrr : PSI<0x56, MRMSrcReg,
759 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
760 "orps\t{$src2, $dst|$dst, $src2}",
761 [(set VR128:$dst, (v2i64
762 (or VR128:$src1, VR128:$src2)))]>;
763 def XORPSrr : PSI<0x57, MRMSrcReg,
764 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
765 "xorps\t{$src2, $dst|$dst, $src2}",
766 [(set VR128:$dst, (v2i64
767 (xor VR128:$src1, VR128:$src2)))]>;
770 def ANDPSrm : PSI<0x54, MRMSrcMem,
771 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
772 "andps\t{$src2, $dst|$dst, $src2}",
773 [(set VR128:$dst, (and (bc_v2i64 (v4f32 VR128:$src1)),
774 (memopv2i64 addr:$src2)))]>;
775 def ORPSrm : PSI<0x56, MRMSrcMem,
776 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
777 "orps\t{$src2, $dst|$dst, $src2}",
778 [(set VR128:$dst, (or (bc_v2i64 (v4f32 VR128:$src1)),
779 (memopv2i64 addr:$src2)))]>;
780 def XORPSrm : PSI<0x57, MRMSrcMem,
781 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
782 "xorps\t{$src2, $dst|$dst, $src2}",
783 [(set VR128:$dst, (xor (bc_v2i64 (v4f32 VR128:$src1)),
784 (memopv2i64 addr:$src2)))]>;
785 def ANDNPSrr : PSI<0x55, MRMSrcReg,
786 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
787 "andnps\t{$src2, $dst|$dst, $src2}",
789 (v2i64 (and (xor VR128:$src1,
790 (bc_v2i64 (v4i32 immAllOnesV))),
792 def ANDNPSrm : PSI<0x55, MRMSrcMem,
793 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
794 "andnps\t{$src2, $dst|$dst, $src2}",
796 (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
797 (bc_v2i64 (v4i32 immAllOnesV))),
798 (memopv2i64 addr:$src2))))]>;
801 let isTwoAddress = 1 in {
802 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
803 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
804 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
805 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
806 VR128:$src, imm:$cc))]>;
807 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
808 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
809 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
810 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
811 (load addr:$src), imm:$cc))]>;
814 // Shuffle and unpack instructions
815 let isTwoAddress = 1 in {
816 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
817 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
818 (outs VR128:$dst), (ins VR128:$src1,
819 VR128:$src2, i32i8imm:$src3),
820 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
822 (v4f32 (vector_shuffle
823 VR128:$src1, VR128:$src2,
824 SHUFP_shuffle_mask:$src3)))]>;
825 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
826 (outs VR128:$dst), (ins VR128:$src1,
827 f128mem:$src2, i32i8imm:$src3),
828 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
830 (v4f32 (vector_shuffle
831 VR128:$src1, (memopv4f32 addr:$src2),
832 SHUFP_shuffle_mask:$src3)))]>;
834 let AddedComplexity = 10 in {
835 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
836 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
837 "unpckhps\t{$src2, $dst|$dst, $src2}",
839 (v4f32 (vector_shuffle
840 VR128:$src1, VR128:$src2,
841 UNPCKH_shuffle_mask)))]>;
842 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
843 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
844 "unpckhps\t{$src2, $dst|$dst, $src2}",
846 (v4f32 (vector_shuffle
847 VR128:$src1, (memopv4f32 addr:$src2),
848 UNPCKH_shuffle_mask)))]>;
850 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
851 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
852 "unpcklps\t{$src2, $dst|$dst, $src2}",
854 (v4f32 (vector_shuffle
855 VR128:$src1, VR128:$src2,
856 UNPCKL_shuffle_mask)))]>;
857 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
858 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
859 "unpcklps\t{$src2, $dst|$dst, $src2}",
861 (v4f32 (vector_shuffle
862 VR128:$src1, (memopv4f32 addr:$src2),
863 UNPCKL_shuffle_mask)))]>;
868 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
869 "movmskps\t{$src, $dst|$dst, $src}",
870 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
871 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
872 "movmskpd\t{$src, $dst|$dst, $src}",
873 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
875 // Prefetching loads.
876 // TODO: no intrinsics for these?
877 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src), "prefetcht0\t$src", []>;
878 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src), "prefetcht1\t$src", []>;
879 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src), "prefetcht2\t$src", []>;
880 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src), "prefetchnta\t$src", []>;
882 // Non-temporal stores
883 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
884 "movntps\t{$src, $dst|$dst, $src}",
885 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
887 // Load, store, and memory fence
888 def SFENCE : PSI<0xAE, MRM7m, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>;
891 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
892 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
893 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
894 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
896 // Alias instructions that map zero vector to pxor / xorp* for sse.
897 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
898 let isReMaterializable = 1 in
899 def V_SET0 : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins),
901 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
903 // FR32 to 128-bit vector conversion.
904 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR32:$src),
905 "movss\t{$src, $dst|$dst, $src}",
907 (v4f32 (scalar_to_vector FR32:$src)))]>;
908 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
909 "movss\t{$src, $dst|$dst, $src}",
911 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
913 // FIXME: may not be able to eliminate this movss with coalescing the src and
914 // dest register classes are different. We really want to write this pattern
916 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
918 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (outs FR32:$dst), (ins VR128:$src),
919 "movss\t{$src, $dst|$dst, $src}",
920 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
922 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
923 "movss\t{$src, $dst|$dst, $src}",
924 [(store (f32 (vector_extract (v4f32 VR128:$src),
925 (iPTR 0))), addr:$dst)]>;
928 // Move to lower bits of a VR128, leaving upper bits alone.
929 // Three operand (but two address) aliases.
930 let isTwoAddress = 1 in {
931 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg,
932 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
933 "movss\t{$src2, $dst|$dst, $src2}", []>;
935 let AddedComplexity = 15 in
936 def MOVLPSrr : SSI<0x10, MRMSrcReg,
937 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
938 "movss\t{$src2, $dst|$dst, $src2}",
940 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
941 MOVL_shuffle_mask)))]>;
944 // Move to lower bits of a VR128 and zeroing upper bits.
945 // Loading from memory automatically zeroing upper bits.
946 let AddedComplexity = 20 in
947 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f32mem:$src),
948 "movss\t{$src, $dst|$dst, $src}",
949 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV,
950 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
951 MOVL_shuffle_mask)))]>;
954 //===----------------------------------------------------------------------===//
956 //===----------------------------------------------------------------------===//
959 def MOVSDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
960 "movsd\t{$src, $dst|$dst, $src}", []>;
961 let isLoad = 1, isReMaterializable = 1 in
962 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
963 "movsd\t{$src, $dst|$dst, $src}",
964 [(set FR64:$dst, (loadf64 addr:$src))]>;
965 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
966 "movsd\t{$src, $dst|$dst, $src}",
967 [(store FR64:$src, addr:$dst)]>;
969 // Conversion instructions
970 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
971 "cvttsd2si\t{$src, $dst|$dst, $src}",
972 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
973 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
974 "cvttsd2si\t{$src, $dst|$dst, $src}",
975 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
976 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
977 "cvtsd2ss\t{$src, $dst|$dst, $src}",
978 [(set FR32:$dst, (fround FR64:$src))]>;
979 def CVTSD2SSrm : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
980 "cvtsd2ss\t{$src, $dst|$dst, $src}",
981 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
982 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
983 "cvtsi2sd\t{$src, $dst|$dst, $src}",
984 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
985 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
986 "cvtsi2sd\t{$src, $dst|$dst, $src}",
987 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
989 // SSE2 instructions with XS prefix
990 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
991 "cvtss2sd\t{$src, $dst|$dst, $src}",
992 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
994 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
995 "cvtss2sd\t{$src, $dst|$dst, $src}",
996 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
999 // Match intrinsics which expect XMM operand(s).
1000 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1001 "cvtsd2si\t{$src, $dst|$dst, $src}",
1002 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1003 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1004 "cvtsd2si\t{$src, $dst|$dst, $src}",
1005 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1006 (load addr:$src)))]>;
1008 // Aliases for intrinsics
1009 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1010 "cvttsd2si\t{$src, $dst|$dst, $src}",
1012 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1013 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1014 "cvttsd2si\t{$src, $dst|$dst, $src}",
1015 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1016 (load addr:$src)))]>;
1018 // Comparison instructions
1019 let isTwoAddress = 1 in {
1020 def CMPSDrr : SDI<0xC2, MRMSrcReg,
1021 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1022 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1023 def CMPSDrm : SDI<0xC2, MRMSrcMem,
1024 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1025 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1028 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1029 "ucomisd\t{$src2, $src1|$src1, $src2}",
1030 [(X86cmp FR64:$src1, FR64:$src2)]>;
1031 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1032 "ucomisd\t{$src2, $src1|$src1, $src2}",
1033 [(X86cmp FR64:$src1, (loadf64 addr:$src2))]>;
1035 // Aliases to match intrinsics which expect XMM operand(s).
1036 let isTwoAddress = 1 in {
1037 def Int_CMPSDrr : SDI<0xC2, MRMSrcReg,
1038 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1039 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1040 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1041 VR128:$src, imm:$cc))]>;
1042 def Int_CMPSDrm : SDI<0xC2, MRMSrcMem,
1043 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1044 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1045 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1046 (load addr:$src), imm:$cc))]>;
1049 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1050 "ucomisd\t{$src2, $src1|$src1, $src2}",
1051 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
1052 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1053 "ucomisd\t{$src2, $src1|$src1, $src2}",
1054 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2))]>;
1056 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1057 "comisd\t{$src2, $src1|$src1, $src2}",
1058 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
1059 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1060 "comisd\t{$src2, $src1|$src1, $src2}",
1061 [(X86comi (v2f64 VR128:$src1), (load addr:$src2))]>;
1063 // Aliases of packed SSE2 instructions for scalar use. These all have names that
1066 // Alias instructions that map fld0 to pxor for sse.
1067 let isReMaterializable = 1 in
1068 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins),
1069 "pxor\t$dst, $dst", [(set FR64:$dst, fpimm0)]>,
1070 Requires<[HasSSE2]>, TB, OpSize;
1072 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1074 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1075 "movapd\t{$src, $dst|$dst, $src}", []>;
1077 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1080 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1081 "movapd\t{$src, $dst|$dst, $src}",
1082 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1084 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1085 let isTwoAddress = 1 in {
1086 let isCommutable = 1 in {
1087 def FsANDPDrr : PDI<0x54, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1088 "andpd\t{$src2, $dst|$dst, $src2}",
1089 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
1090 def FsORPDrr : PDI<0x56, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1091 "orpd\t{$src2, $dst|$dst, $src2}",
1092 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
1093 def FsXORPDrr : PDI<0x57, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1094 "xorpd\t{$src2, $dst|$dst, $src2}",
1095 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
1098 def FsANDPDrm : PDI<0x54, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1099 "andpd\t{$src2, $dst|$dst, $src2}",
1100 [(set FR64:$dst, (X86fand FR64:$src1,
1101 (memopfsf64 addr:$src2)))]>;
1102 def FsORPDrm : PDI<0x56, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1103 "orpd\t{$src2, $dst|$dst, $src2}",
1104 [(set FR64:$dst, (X86for FR64:$src1,
1105 (memopfsf64 addr:$src2)))]>;
1106 def FsXORPDrm : PDI<0x57, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1107 "xorpd\t{$src2, $dst|$dst, $src2}",
1108 [(set FR64:$dst, (X86fxor FR64:$src1,
1109 (memopfsf64 addr:$src2)))]>;
1111 def FsANDNPDrr : PDI<0x55, MRMSrcReg,
1112 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1113 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1114 def FsANDNPDrm : PDI<0x55, MRMSrcMem,
1115 (outs FR64:$dst), (ins FR64:$src1, f128mem:$src2),
1116 "andnpd\t{$src2, $dst|$dst, $src2}", []>;
1119 /// basic_sse2_fp_binop_rm - SSE2 binops come in both scalar and vector forms.
1121 /// In addition, we also have a special variant of the scalar form here to
1122 /// represent the associated intrinsic operation. This form is unlike the
1123 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1124 /// and leaves the top elements undefined.
1126 /// These three forms can each be reg+reg or reg+mem, so there are a total of
1127 /// six "instructions".
1129 let isTwoAddress = 1 in {
1130 multiclass basic_sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1131 SDNode OpNode, Intrinsic F64Int,
1132 bit Commutable = 0> {
1133 // Scalar operation, reg+reg.
1134 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1135 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1136 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1137 let isCommutable = Commutable;
1140 // Scalar operation, reg+mem.
1141 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1142 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1143 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1145 // Vector operation, reg+reg.
1146 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1147 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1148 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1149 let isCommutable = Commutable;
1152 // Vector operation, reg+mem.
1153 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1154 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1155 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1157 // Intrinsic operation, reg+reg.
1158 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1159 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1160 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1161 let isCommutable = Commutable;
1164 // Intrinsic operation, reg+mem.
1165 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1166 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1167 [(set VR128:$dst, (F64Int VR128:$src1,
1168 sse_load_f64:$src2))]>;
1172 // Arithmetic instructions
1173 defm ADD : basic_sse2_fp_binop_rm<0x58, "add", fadd, int_x86_sse2_add_sd, 1>;
1174 defm MUL : basic_sse2_fp_binop_rm<0x59, "mul", fmul, int_x86_sse2_mul_sd, 1>;
1175 defm SUB : basic_sse2_fp_binop_rm<0x5C, "sub", fsub, int_x86_sse2_sub_sd>;
1176 defm DIV : basic_sse2_fp_binop_rm<0x5E, "div", fdiv, int_x86_sse2_div_sd>;
1178 /// sse2_fp_binop_rm - Other SSE2 binops
1180 /// This multiclass is like basic_sse2_fp_binop_rm, with the addition of
1181 /// instructions for a full-vector intrinsic form. Operations that map
1182 /// onto C operators don't use this form since they just use the plain
1183 /// vector form instead of having a separate vector intrinsic form.
1185 /// This provides a total of eight "instructions".
1187 let isTwoAddress = 1 in {
1188 multiclass sse2_fp_binop_rm<bits<8> opc, string OpcodeStr,
1192 bit Commutable = 0> {
1194 // Scalar operation, reg+reg.
1195 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1196 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1197 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
1198 let isCommutable = Commutable;
1201 // Scalar operation, reg+mem.
1202 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2),
1203 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1204 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
1206 // Vector operation, reg+reg.
1207 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1208 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1209 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
1210 let isCommutable = Commutable;
1213 // Vector operation, reg+mem.
1214 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1215 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1216 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
1218 // Intrinsic operation, reg+reg.
1219 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1220 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1221 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
1222 let isCommutable = Commutable;
1225 // Intrinsic operation, reg+mem.
1226 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1227 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1228 [(set VR128:$dst, (F64Int VR128:$src1,
1229 sse_load_f64:$src2))]>;
1231 // Vector intrinsic operation, reg+reg.
1232 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1233 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1234 [(set VR128:$dst, (V2F64Int VR128:$src1, VR128:$src2))]> {
1235 let isCommutable = Commutable;
1238 // Vector intrinsic operation, reg+mem.
1239 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1240 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1241 [(set VR128:$dst, (V2F64Int VR128:$src1, (load addr:$src2)))]>;
1245 defm MAX : sse2_fp_binop_rm<0x5F, "max", X86fmax,
1246 int_x86_sse2_max_sd, int_x86_sse2_max_pd>;
1247 defm MIN : sse2_fp_binop_rm<0x5D, "min", X86fmin,
1248 int_x86_sse2_min_sd, int_x86_sse2_min_pd>;
1250 //===----------------------------------------------------------------------===//
1251 // SSE packed FP Instructions
1253 // Move Instructions
1254 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1255 "movapd\t{$src, $dst|$dst, $src}", []>;
1256 let isLoad = 1, isReMaterializable = 1 in
1257 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1258 "movapd\t{$src, $dst|$dst, $src}",
1259 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1261 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1262 "movapd\t{$src, $dst|$dst, $src}",
1263 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1265 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1266 "movupd\t{$src, $dst|$dst, $src}", []>;
1268 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1269 "movupd\t{$src, $dst|$dst, $src}",
1270 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1271 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1272 "movupd\t{$src, $dst|$dst, $src}",
1273 [(store (v2f64 VR128:$src), addr:$dst)]>;
1275 // Intrinsic forms of MOVUPD load and store
1276 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1277 "movupd\t{$src, $dst|$dst, $src}",
1278 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1279 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1280 "movupd\t{$src, $dst|$dst, $src}",
1281 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1283 let isTwoAddress = 1 in {
1284 let AddedComplexity = 20 in {
1285 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1286 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1287 "movlpd\t{$src2, $dst|$dst, $src2}",
1289 (v2f64 (vector_shuffle VR128:$src1,
1290 (scalar_to_vector (loadf64 addr:$src2)),
1291 MOVLP_shuffle_mask)))]>;
1292 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1293 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1294 "movhpd\t{$src2, $dst|$dst, $src2}",
1296 (v2f64 (vector_shuffle VR128:$src1,
1297 (scalar_to_vector (loadf64 addr:$src2)),
1298 MOVHP_shuffle_mask)))]>;
1299 } // AddedComplexity
1302 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1303 "movlpd\t{$src, $dst|$dst, $src}",
1304 [(store (f64 (vector_extract (v2f64 VR128:$src),
1305 (iPTR 0))), addr:$dst)]>;
1307 // v2f64 extract element 1 is always custom lowered to unpack high to low
1308 // and extract element 0 so the non-store version isn't too horrible.
1309 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1310 "movhpd\t{$src, $dst|$dst, $src}",
1311 [(store (f64 (vector_extract
1312 (v2f64 (vector_shuffle VR128:$src, (undef),
1313 UNPCKH_shuffle_mask)), (iPTR 0))),
1316 // SSE2 instructions without OpSize prefix
1317 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1318 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1319 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1320 TB, Requires<[HasSSE2]>;
1321 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1322 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1323 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1324 (bitconvert (memopv2i64 addr:$src))))]>,
1325 TB, Requires<[HasSSE2]>;
1327 // SSE2 instructions with XS prefix
1328 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1329 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1330 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1331 XS, Requires<[HasSSE2]>;
1332 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1333 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1334 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1335 (bitconvert (memopv2i64 addr:$src))))]>,
1336 XS, Requires<[HasSSE2]>;
1338 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1339 "cvtps2dq\t{$src, $dst|$dst, $src}",
1340 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1341 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1342 "cvtps2dq\t{$src, $dst|$dst, $src}",
1343 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1344 (load addr:$src)))]>;
1345 // SSE2 packed instructions with XS prefix
1346 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1347 "cvttps2dq\t{$src, $dst|$dst, $src}",
1348 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
1349 XS, Requires<[HasSSE2]>;
1350 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1351 "cvttps2dq\t{$src, $dst|$dst, $src}",
1352 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1353 (load addr:$src)))]>,
1354 XS, Requires<[HasSSE2]>;
1356 // SSE2 packed instructions with XD prefix
1357 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1358 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1359 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1360 XD, Requires<[HasSSE2]>;
1361 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1362 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1363 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1364 (load addr:$src)))]>,
1365 XD, Requires<[HasSSE2]>;
1367 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1368 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1369 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1370 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1371 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1372 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1373 (load addr:$src)))]>;
1375 // SSE2 instructions without OpSize prefix
1376 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1377 "cvtps2pd\t{$src, $dst|$dst, $src}",
1378 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1379 TB, Requires<[HasSSE2]>;
1380 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f64mem:$src),
1381 "cvtps2pd\t{$src, $dst|$dst, $src}",
1382 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1383 (load addr:$src)))]>,
1384 TB, Requires<[HasSSE2]>;
1386 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1387 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1388 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1389 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins f128mem:$src),
1390 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1391 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1392 (load addr:$src)))]>;
1394 // Match intrinsics which expect XMM operand(s).
1395 // Aliases for intrinsics
1396 let isTwoAddress = 1 in {
1397 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1398 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1399 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1400 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1402 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1403 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1404 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1405 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1406 (loadi32 addr:$src2)))]>;
1407 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1408 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1409 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1410 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1412 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1413 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1414 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1415 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1416 (load addr:$src2)))]>;
1417 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1418 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1419 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1420 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1421 VR128:$src2))]>, XS,
1422 Requires<[HasSSE2]>;
1423 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1424 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1425 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1426 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1427 (load addr:$src2)))]>, XS,
1428 Requires<[HasSSE2]>;
1433 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1435 /// In addition, we also have a special variant of the scalar form here to
1436 /// represent the associated intrinsic operation. This form is unlike the
1437 /// plain scalar form, in that it takes an entire vector (instead of a
1438 /// scalar) and leaves the top elements undefined.
1440 /// And, we have a special variant form for a full-vector intrinsic form.
1442 /// These four forms can each have a reg or a mem operand, so there are a
1443 /// total of eight "instructions".
1445 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1449 bit Commutable = 0> {
1450 // Scalar operation, reg.
1451 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1452 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1453 [(set FR64:$dst, (OpNode FR64:$src))]> {
1454 let isCommutable = Commutable;
1457 // Scalar operation, mem.
1458 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1459 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1460 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1462 // Vector operation, reg.
1463 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1464 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1465 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1466 let isCommutable = Commutable;
1469 // Vector operation, mem.
1470 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1471 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1472 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1474 // Intrinsic operation, reg.
1475 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1476 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1477 [(set VR128:$dst, (F64Int VR128:$src))]> {
1478 let isCommutable = Commutable;
1481 // Intrinsic operation, mem.
1482 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1483 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1484 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1486 // Vector intrinsic operation, reg
1487 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1488 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1489 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1490 let isCommutable = Commutable;
1493 // Vector intrinsic operation, mem
1494 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1495 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1496 [(set VR128:$dst, (V2F64Int (load addr:$src)))]>;
1500 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1501 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1503 // There is no f64 version of the reciprocal approximation instructions.
1506 let isTwoAddress = 1 in {
1507 let isCommutable = 1 in {
1508 def ANDPDrr : PDI<0x54, MRMSrcReg,
1509 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1510 "andpd\t{$src2, $dst|$dst, $src2}",
1512 (and (bc_v2i64 (v2f64 VR128:$src1)),
1513 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1514 def ORPDrr : PDI<0x56, MRMSrcReg,
1515 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1516 "orpd\t{$src2, $dst|$dst, $src2}",
1518 (or (bc_v2i64 (v2f64 VR128:$src1)),
1519 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1520 def XORPDrr : PDI<0x57, MRMSrcReg,
1521 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1522 "xorpd\t{$src2, $dst|$dst, $src2}",
1524 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1525 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1528 def ANDPDrm : PDI<0x54, MRMSrcMem,
1529 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1530 "andpd\t{$src2, $dst|$dst, $src2}",
1532 (and (bc_v2i64 (v2f64 VR128:$src1)),
1533 (memopv2i64 addr:$src2)))]>;
1534 def ORPDrm : PDI<0x56, MRMSrcMem,
1535 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1536 "orpd\t{$src2, $dst|$dst, $src2}",
1538 (or (bc_v2i64 (v2f64 VR128:$src1)),
1539 (memopv2i64 addr:$src2)))]>;
1540 def XORPDrm : PDI<0x57, MRMSrcMem,
1541 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1542 "xorpd\t{$src2, $dst|$dst, $src2}",
1544 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1545 (memopv2i64 addr:$src2)))]>;
1546 def ANDNPDrr : PDI<0x55, MRMSrcReg,
1547 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1548 "andnpd\t{$src2, $dst|$dst, $src2}",
1550 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1551 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1552 def ANDNPDrm : PDI<0x55, MRMSrcMem,
1553 (outs VR128:$dst), (ins VR128:$src1,f128mem:$src2),
1554 "andnpd\t{$src2, $dst|$dst, $src2}",
1556 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1557 (memopv2i64 addr:$src2)))]>;
1560 let isTwoAddress = 1 in {
1561 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1562 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1563 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1564 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1565 VR128:$src, imm:$cc))]>;
1566 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1567 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1568 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1569 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1570 (load addr:$src), imm:$cc))]>;
1573 // Shuffle and unpack instructions
1574 let isTwoAddress = 1 in {
1575 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1576 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1577 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1578 [(set VR128:$dst, (v2f64 (vector_shuffle
1579 VR128:$src1, VR128:$src2,
1580 SHUFP_shuffle_mask:$src3)))]>;
1581 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1582 (outs VR128:$dst), (ins VR128:$src1,
1583 f128mem:$src2, i8imm:$src3),
1584 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1586 (v2f64 (vector_shuffle
1587 VR128:$src1, (memopv2f64 addr:$src2),
1588 SHUFP_shuffle_mask:$src3)))]>;
1590 let AddedComplexity = 10 in {
1591 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1592 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1593 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1595 (v2f64 (vector_shuffle
1596 VR128:$src1, VR128:$src2,
1597 UNPCKH_shuffle_mask)))]>;
1598 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1599 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1600 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1602 (v2f64 (vector_shuffle
1603 VR128:$src1, (memopv2f64 addr:$src2),
1604 UNPCKH_shuffle_mask)))]>;
1606 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1607 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1608 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1610 (v2f64 (vector_shuffle
1611 VR128:$src1, VR128:$src2,
1612 UNPCKL_shuffle_mask)))]>;
1613 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1614 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1615 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1617 (v2f64 (vector_shuffle
1618 VR128:$src1, (memopv2f64 addr:$src2),
1619 UNPCKL_shuffle_mask)))]>;
1620 } // AddedComplexity
1624 //===----------------------------------------------------------------------===//
1625 // SSE integer instructions
1627 // Move Instructions
1628 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1629 "movdqa\t{$src, $dst|$dst, $src}", []>;
1631 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1632 "movdqa\t{$src, $dst|$dst, $src}",
1633 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1634 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1635 "movdqa\t{$src, $dst|$dst, $src}",
1636 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1638 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1639 "movdqu\t{$src, $dst|$dst, $src}",
1640 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1641 XS, Requires<[HasSSE2]>;
1642 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1643 "movdqu\t{$src, $dst|$dst, $src}",
1644 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1645 XS, Requires<[HasSSE2]>;
1647 // Intrinsic forms of MOVDQU load and store
1649 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1650 "movdqu\t{$src, $dst|$dst, $src}",
1651 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1652 XS, Requires<[HasSSE2]>;
1653 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1654 "movdqu\t{$src, $dst|$dst, $src}",
1655 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1656 XS, Requires<[HasSSE2]>;
1658 let isTwoAddress = 1 in {
1660 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1661 bit Commutable = 0> {
1662 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1663 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1664 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1665 let isCommutable = Commutable;
1667 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1668 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1669 [(set VR128:$dst, (IntId VR128:$src1,
1670 (bitconvert (memopv2i64 addr:$src2))))]>;
1673 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1674 string OpcodeStr, Intrinsic IntId> {
1675 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1676 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1677 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1678 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1679 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1680 [(set VR128:$dst, (IntId VR128:$src1,
1681 (bitconvert (memopv2i64 addr:$src2))))]>;
1682 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1683 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1684 [(set VR128:$dst, (IntId VR128:$src1,
1685 (scalar_to_vector (i32 imm:$src2))))]>;
1689 /// PDI_binop_rm - Simple SSE2 binary operator.
1690 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1691 ValueType OpVT, bit Commutable = 0> {
1692 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1693 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1694 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1695 let isCommutable = Commutable;
1697 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1698 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1699 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1700 (bitconvert (memopv2i64 addr:$src2)))))]>;
1703 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1705 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1706 /// to collapse (bitconvert VT to VT) into its operand.
1708 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1709 bit Commutable = 0> {
1710 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1711 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1712 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1713 let isCommutable = Commutable;
1715 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1716 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1717 [(set VR128:$dst, (OpNode VR128:$src1,(memopv2i64 addr:$src2)))]>;
1722 // 128-bit Integer Arithmetic
1724 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1725 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1726 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1727 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1729 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1730 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1731 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1732 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1734 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1735 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1736 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1737 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1739 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1740 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1741 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1742 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1744 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1746 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1747 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1748 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1750 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1752 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1753 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1756 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1757 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1758 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1759 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1760 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1763 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw", int_x86_sse2_psll_w>;
1764 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld", int_x86_sse2_psll_d>;
1765 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq", int_x86_sse2_psll_q>;
1767 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw", int_x86_sse2_psrl_w>;
1768 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld", int_x86_sse2_psrl_d>;
1769 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq", int_x86_sse2_psrl_q>;
1771 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw", int_x86_sse2_psra_w>;
1772 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad", int_x86_sse2_psra_d>;
1773 // PSRAQ doesn't exist in SSE[1-3].
1775 // 128-bit logical shifts.
1776 let isTwoAddress = 1 in {
1777 def PSLLDQri : PDIi8<0x73, MRM7r,
1778 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1779 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
1780 def PSRLDQri : PDIi8<0x73, MRM3r,
1781 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1782 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
1783 // PSRADQri doesn't exist in SSE[1-3].
1786 let Predicates = [HasSSE2] in {
1787 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1788 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1789 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1790 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1791 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1792 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1796 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1797 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1798 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1800 let isTwoAddress = 1 in {
1801 def PANDNrr : PDI<0xDF, MRMSrcReg,
1802 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1803 "pandn\t{$src2, $dst|$dst, $src2}",
1804 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1807 def PANDNrm : PDI<0xDF, MRMSrcMem,
1808 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1809 "pandn\t{$src2, $dst|$dst, $src2}",
1810 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1811 (memopv2i64 addr:$src2))))]>;
1814 // SSE2 Integer comparison
1815 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1816 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1817 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1818 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1819 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1820 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1822 // Pack instructions
1823 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1824 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1825 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1827 // Shuffle and unpack instructions
1828 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1829 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1830 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1831 [(set VR128:$dst, (v4i32 (vector_shuffle
1832 VR128:$src1, (undef),
1833 PSHUFD_shuffle_mask:$src2)))]>;
1834 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1835 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1836 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1837 [(set VR128:$dst, (v4i32 (vector_shuffle
1838 (bc_v4i32(memopv2i64 addr:$src1)),
1840 PSHUFD_shuffle_mask:$src2)))]>;
1842 // SSE2 with ImmT == Imm8 and XS prefix.
1843 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1844 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
1845 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1846 [(set VR128:$dst, (v8i16 (vector_shuffle
1847 VR128:$src1, (undef),
1848 PSHUFHW_shuffle_mask:$src2)))]>,
1849 XS, Requires<[HasSSE2]>;
1850 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
1851 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
1852 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1853 [(set VR128:$dst, (v8i16 (vector_shuffle
1854 (bc_v8i16 (memopv2i64 addr:$src1)),
1856 PSHUFHW_shuffle_mask:$src2)))]>,
1857 XS, Requires<[HasSSE2]>;
1859 // SSE2 with ImmT == Imm8 and XD prefix.
1860 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
1861 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
1862 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1863 [(set VR128:$dst, (v8i16 (vector_shuffle
1864 VR128:$src1, (undef),
1865 PSHUFLW_shuffle_mask:$src2)))]>,
1866 XD, Requires<[HasSSE2]>;
1867 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
1868 (outs VR128:$dst), (ins i128mem:$src1, i32i8imm:$src2),
1869 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1870 [(set VR128:$dst, (v8i16 (vector_shuffle
1871 (bc_v8i16 (memopv2i64 addr:$src1)),
1873 PSHUFLW_shuffle_mask:$src2)))]>,
1874 XD, Requires<[HasSSE2]>;
1877 let isTwoAddress = 1 in {
1878 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
1879 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1880 "punpcklbw\t{$src2, $dst|$dst, $src2}",
1882 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1883 UNPCKL_shuffle_mask)))]>;
1884 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
1885 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1886 "punpcklbw\t{$src2, $dst|$dst, $src2}",
1888 (v16i8 (vector_shuffle VR128:$src1,
1889 (bc_v16i8 (memopv2i64 addr:$src2)),
1890 UNPCKL_shuffle_mask)))]>;
1891 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
1892 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1893 "punpcklwd\t{$src2, $dst|$dst, $src2}",
1895 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1896 UNPCKL_shuffle_mask)))]>;
1897 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
1898 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1899 "punpcklwd\t{$src2, $dst|$dst, $src2}",
1901 (v8i16 (vector_shuffle VR128:$src1,
1902 (bc_v8i16 (memopv2i64 addr:$src2)),
1903 UNPCKL_shuffle_mask)))]>;
1904 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
1905 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1906 "punpckldq\t{$src2, $dst|$dst, $src2}",
1908 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1909 UNPCKL_shuffle_mask)))]>;
1910 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
1911 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1912 "punpckldq\t{$src2, $dst|$dst, $src2}",
1914 (v4i32 (vector_shuffle VR128:$src1,
1915 (bc_v4i32 (memopv2i64 addr:$src2)),
1916 UNPCKL_shuffle_mask)))]>;
1917 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
1918 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1919 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
1921 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1922 UNPCKL_shuffle_mask)))]>;
1923 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
1924 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1925 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
1927 (v2i64 (vector_shuffle VR128:$src1,
1928 (memopv2i64 addr:$src2),
1929 UNPCKL_shuffle_mask)))]>;
1931 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
1932 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1933 "punpckhbw\t{$src2, $dst|$dst, $src2}",
1935 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1936 UNPCKH_shuffle_mask)))]>;
1937 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
1938 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1939 "punpckhbw\t{$src2, $dst|$dst, $src2}",
1941 (v16i8 (vector_shuffle VR128:$src1,
1942 (bc_v16i8 (memopv2i64 addr:$src2)),
1943 UNPCKH_shuffle_mask)))]>;
1944 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
1945 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1946 "punpckhwd\t{$src2, $dst|$dst, $src2}",
1948 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1949 UNPCKH_shuffle_mask)))]>;
1950 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
1951 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1952 "punpckhwd\t{$src2, $dst|$dst, $src2}",
1954 (v8i16 (vector_shuffle VR128:$src1,
1955 (bc_v8i16 (memopv2i64 addr:$src2)),
1956 UNPCKH_shuffle_mask)))]>;
1957 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
1958 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1959 "punpckhdq\t{$src2, $dst|$dst, $src2}",
1961 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1962 UNPCKH_shuffle_mask)))]>;
1963 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
1964 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1965 "punpckhdq\t{$src2, $dst|$dst, $src2}",
1967 (v4i32 (vector_shuffle VR128:$src1,
1968 (bc_v4i32 (memopv2i64 addr:$src2)),
1969 UNPCKH_shuffle_mask)))]>;
1970 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
1971 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1972 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
1974 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1975 UNPCKH_shuffle_mask)))]>;
1976 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
1977 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
1978 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
1980 (v2i64 (vector_shuffle VR128:$src1,
1981 (memopv2i64 addr:$src2),
1982 UNPCKH_shuffle_mask)))]>;
1986 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
1987 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
1988 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1989 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
1990 (iPTR imm:$src2)))]>;
1991 let isTwoAddress = 1 in {
1992 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
1993 (outs VR128:$dst), (ins VR128:$src1,
1994 GR32:$src2, i32i8imm:$src3),
1995 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1997 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
1998 GR32:$src2, (iPTR imm:$src3))))]>;
1999 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2000 (outs VR128:$dst), (ins VR128:$src1,
2001 i16mem:$src2, i32i8imm:$src3),
2002 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2004 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
2005 (i32 (anyext (loadi16 addr:$src2))),
2006 (iPTR imm:$src3))))]>;
2010 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2011 "pmovmskb\t{$src, $dst|$dst, $src}",
2012 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2014 // Conditional store
2015 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2016 "maskmovdqu\t{$mask, $src|$src, $mask}",
2017 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>,
2020 // Non-temporal stores
2021 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2022 "movntpd\t{$src, $dst|$dst, $src}",
2023 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2024 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2025 "movntdq\t{$src, $dst|$dst, $src}",
2026 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2027 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2028 "movnti\t{$src, $dst|$dst, $src}",
2029 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2030 TB, Requires<[HasSSE2]>;
2033 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2034 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2035 TB, Requires<[HasSSE2]>;
2037 // Load, store, and memory fence
2038 def LFENCE : I<0xAE, MRM5m, (outs), (ins),
2039 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2040 def MFENCE : I<0xAE, MRM6m, (outs), (ins),
2041 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2044 // Alias instructions that map zero vector to pxor / xorp* for sse.
2045 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
2046 let isReMaterializable = 1 in
2047 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins),
2048 "pcmpeqd\t$dst, $dst",
2049 [(set VR128:$dst, (v2f64 immAllOnesV))]>;
2051 // FR64 to 128-bit vector conversion.
2052 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins FR64:$src),
2053 "movsd\t{$src, $dst|$dst, $src}",
2055 (v2f64 (scalar_to_vector FR64:$src)))]>;
2056 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2057 "movsd\t{$src, $dst|$dst, $src}",
2059 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
2061 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2062 "movd\t{$src, $dst|$dst, $src}",
2064 (v4i32 (scalar_to_vector GR32:$src)))]>;
2065 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2066 "movd\t{$src, $dst|$dst, $src}",
2068 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2070 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2071 "movd\t{$src, $dst|$dst, $src}",
2072 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2074 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2075 "movd\t{$src, $dst|$dst, $src}",
2076 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2078 // SSE2 instructions with XS prefix
2079 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2080 "movq\t{$src, $dst|$dst, $src}",
2082 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2083 Requires<[HasSSE2]>;
2084 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2085 "movq\t{$src, $dst|$dst, $src}",
2086 [(store (i64 (vector_extract (v2i64 VR128:$src),
2087 (iPTR 0))), addr:$dst)]>;
2089 // FIXME: may not be able to eliminate this movss with coalescing the src and
2090 // dest register classes are different. We really want to write this pattern
2092 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2093 // (f32 FR32:$src)>;
2094 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (outs FR64:$dst), (ins VR128:$src),
2095 "movsd\t{$src, $dst|$dst, $src}",
2096 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
2098 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
2099 "movsd\t{$src, $dst|$dst, $src}",
2100 [(store (f64 (vector_extract (v2f64 VR128:$src),
2101 (iPTR 0))), addr:$dst)]>;
2102 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2103 "movd\t{$src, $dst|$dst, $src}",
2104 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2106 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2107 "movd\t{$src, $dst|$dst, $src}",
2108 [(store (i32 (vector_extract (v4i32 VR128:$src),
2109 (iPTR 0))), addr:$dst)]>;
2111 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2112 "movd\t{$src, $dst|$dst, $src}",
2113 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2114 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2115 "movd\t{$src, $dst|$dst, $src}",
2116 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2119 // Move to lower bits of a VR128, leaving upper bits alone.
2120 // Three operand (but two address) aliases.
2121 let isTwoAddress = 1 in {
2122 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg,
2123 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
2124 "movsd\t{$src2, $dst|$dst, $src2}", []>;
2126 let AddedComplexity = 15 in
2127 def MOVLPDrr : SDI<0x10, MRMSrcReg,
2128 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2129 "movsd\t{$src2, $dst|$dst, $src2}",
2131 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
2132 MOVL_shuffle_mask)))]>;
2135 // Store / copy lower 64-bits of a XMM register.
2136 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2137 "movq\t{$src, $dst|$dst, $src}",
2138 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2140 // Move to lower bits of a VR128 and zeroing upper bits.
2141 // Loading from memory automatically zeroing upper bits.
2142 let AddedComplexity = 20 in
2143 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2144 "movsd\t{$src, $dst|$dst, $src}",
2146 (v2f64 (vector_shuffle immAllZerosV,
2147 (v2f64 (scalar_to_vector
2148 (loadf64 addr:$src))),
2149 MOVL_shuffle_mask)))]>;
2151 let AddedComplexity = 15 in
2152 // movd / movq to XMM register zero-extends
2153 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2154 "movd\t{$src, $dst|$dst, $src}",
2156 (v4i32 (vector_shuffle immAllZerosV,
2157 (v4i32 (scalar_to_vector GR32:$src)),
2158 MOVL_shuffle_mask)))]>;
2159 let AddedComplexity = 20 in
2160 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2161 "movd\t{$src, $dst|$dst, $src}",
2163 (v4i32 (vector_shuffle immAllZerosV,
2164 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
2165 MOVL_shuffle_mask)))]>;
2167 // Moving from XMM to XMM but still clear upper 64 bits.
2168 let AddedComplexity = 15 in
2169 def MOVZQI2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2170 "movq\t{$src, $dst|$dst, $src}",
2171 [(set VR128:$dst, (int_x86_sse2_movl_dq VR128:$src))]>,
2172 XS, Requires<[HasSSE2]>;
2173 let AddedComplexity = 20 in
2174 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2175 "movq\t{$src, $dst|$dst, $src}",
2176 [(set VR128:$dst, (int_x86_sse2_movl_dq
2177 (bitconvert (memopv2i64 addr:$src))))]>,
2178 XS, Requires<[HasSSE2]>;
2181 //===----------------------------------------------------------------------===//
2182 // SSE3 Instructions
2183 //===----------------------------------------------------------------------===//
2185 // Move Instructions
2186 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2187 "movshdup\t{$src, $dst|$dst, $src}",
2188 [(set VR128:$dst, (v4f32 (vector_shuffle
2189 VR128:$src, (undef),
2190 MOVSHDUP_shuffle_mask)))]>;
2191 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2192 "movshdup\t{$src, $dst|$dst, $src}",
2193 [(set VR128:$dst, (v4f32 (vector_shuffle
2194 (memopv4f32 addr:$src), (undef),
2195 MOVSHDUP_shuffle_mask)))]>;
2197 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2198 "movsldup\t{$src, $dst|$dst, $src}",
2199 [(set VR128:$dst, (v4f32 (vector_shuffle
2200 VR128:$src, (undef),
2201 MOVSLDUP_shuffle_mask)))]>;
2202 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2203 "movsldup\t{$src, $dst|$dst, $src}",
2204 [(set VR128:$dst, (v4f32 (vector_shuffle
2205 (memopv4f32 addr:$src), (undef),
2206 MOVSLDUP_shuffle_mask)))]>;
2208 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2209 "movddup\t{$src, $dst|$dst, $src}",
2210 [(set VR128:$dst, (v2f64 (vector_shuffle
2211 VR128:$src, (undef),
2212 SSE_splat_lo_mask)))]>;
2213 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2214 "movddup\t{$src, $dst|$dst, $src}",
2216 (v2f64 (vector_shuffle
2217 (scalar_to_vector (loadf64 addr:$src)),
2219 SSE_splat_lo_mask)))]>;
2222 let isTwoAddress = 1 in {
2223 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2224 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2225 "addsubps\t{$src2, $dst|$dst, $src2}",
2226 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2228 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2229 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2230 "addsubps\t{$src2, $dst|$dst, $src2}",
2231 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2232 (load addr:$src2)))]>;
2233 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2234 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2235 "addsubpd\t{$src2, $dst|$dst, $src2}",
2236 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2238 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2239 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2240 "addsubpd\t{$src2, $dst|$dst, $src2}",
2241 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2242 (load addr:$src2)))]>;
2245 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2246 "lddqu\t{$src, $dst|$dst, $src}",
2247 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2250 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2251 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2252 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2253 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2254 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2255 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2256 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2257 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
2258 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2259 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2260 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2261 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2262 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2263 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2264 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2265 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
2267 let isTwoAddress = 1 in {
2268 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2269 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2270 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2271 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2272 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2273 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2274 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2275 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2278 // Thread synchronization
2279 def MONITOR : I<0xC8, RawFrm, (outs), (ins), "monitor",
2280 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2281 def MWAIT : I<0xC9, RawFrm, (outs), (ins), "mwait",
2282 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2284 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2285 let AddedComplexity = 15 in
2286 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2287 MOVSHDUP_shuffle_mask)),
2288 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2289 let AddedComplexity = 20 in
2290 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2291 MOVSHDUP_shuffle_mask)),
2292 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2294 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2295 let AddedComplexity = 15 in
2296 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2297 MOVSLDUP_shuffle_mask)),
2298 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2299 let AddedComplexity = 20 in
2300 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (memopv2i64 addr:$src)), (undef),
2301 MOVSLDUP_shuffle_mask)),
2302 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2304 //===----------------------------------------------------------------------===//
2305 // SSSE3 Instructions
2306 //===----------------------------------------------------------------------===//
2308 // SSSE3 Instruction Templates:
2310 // SS38I - SSSE3 instructions with T8 prefix.
2311 // SS3AI - SSSE3 instructions with TA prefix.
2313 // Note: SSSE3 instructions have 64-bit and 128-bit versions. The 64-bit version
2314 // uses the MMX registers. We put those instructions here because they better
2315 // fit into the SSSE3 instruction category rather than the MMX category.
2317 class SS38I<bits<8> o, Format F, dag outs, dag ins, string asm,
2319 : I<o, F, outs, ins, asm, pattern>, T8, Requires<[HasSSSE3]>;
2320 class SS3AI<bits<8> o, Format F, dag outs, dag ins, string asm,
2322 : I<o, F, outs, ins, asm, pattern>, TA, Requires<[HasSSSE3]>;
2324 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2325 let isTwoAddress = 1 in {
2326 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2327 Intrinsic IntId64, Intrinsic IntId128,
2328 bit Commutable = 0> {
2329 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2330 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2331 [(set VR64:$dst, (IntId64 VR64:$src))]> {
2332 let isCommutable = Commutable;
2334 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2335 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2337 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2339 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2341 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2342 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2344 let isCommutable = Commutable;
2346 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2348 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2351 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2355 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2356 let isTwoAddress = 1 in {
2357 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2358 Intrinsic IntId64, Intrinsic IntId128,
2359 bit Commutable = 0> {
2360 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2362 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2363 [(set VR64:$dst, (IntId64 VR64:$src))]> {
2364 let isCommutable = Commutable;
2366 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2368 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2371 (bitconvert (memopv4i16 addr:$src))))]>;
2373 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2375 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2376 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2378 let isCommutable = Commutable;
2380 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2382 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2385 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2389 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2390 let isTwoAddress = 1 in {
2391 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2392 Intrinsic IntId64, Intrinsic IntId128,
2393 bit Commutable = 0> {
2394 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2396 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2397 [(set VR64:$dst, (IntId64 VR64:$src))]> {
2398 let isCommutable = Commutable;
2400 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2402 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2405 (bitconvert (memopv2i32 addr:$src))))]>;
2407 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2409 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2410 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2412 let isCommutable = Commutable;
2414 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2416 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2419 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2423 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2424 int_x86_ssse3_pabs_b,
2425 int_x86_ssse3_pabs_b_128>;
2426 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2427 int_x86_ssse3_pabs_w,
2428 int_x86_ssse3_pabs_w_128>;
2429 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2430 int_x86_ssse3_pabs_d,
2431 int_x86_ssse3_pabs_d_128>;
2433 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2434 let isTwoAddress = 1 in {
2435 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2436 Intrinsic IntId64, Intrinsic IntId128,
2437 bit Commutable = 0> {
2438 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2439 (ins VR64:$src1, VR64:$src2),
2440 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2441 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2442 let isCommutable = Commutable;
2444 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2445 (ins VR64:$src1, i64mem:$src2),
2446 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2448 (IntId64 VR64:$src1,
2449 (bitconvert (memopv8i8 addr:$src2))))]>;
2451 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2452 (ins VR128:$src1, VR128:$src2),
2453 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2454 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2456 let isCommutable = Commutable;
2458 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2459 (ins VR128:$src1, i128mem:$src2),
2460 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2462 (IntId128 VR128:$src1,
2463 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2467 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2468 let isTwoAddress = 1 in {
2469 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2470 Intrinsic IntId64, Intrinsic IntId128,
2471 bit Commutable = 0> {
2472 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2473 (ins VR64:$src1, VR64:$src2),
2474 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2475 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2476 let isCommutable = Commutable;
2478 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2479 (ins VR64:$src1, i64mem:$src2),
2480 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2482 (IntId64 VR64:$src1,
2483 (bitconvert (memopv4i16 addr:$src2))))]>;
2485 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2486 (ins VR128:$src1, VR128:$src2),
2487 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2488 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2490 let isCommutable = Commutable;
2492 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2493 (ins VR128:$src1, i128mem:$src2),
2494 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2496 (IntId128 VR128:$src1,
2497 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2501 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2502 let isTwoAddress = 1 in {
2503 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2504 Intrinsic IntId64, Intrinsic IntId128,
2505 bit Commutable = 0> {
2506 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2507 (ins VR64:$src1, VR64:$src2),
2508 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2509 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2510 let isCommutable = Commutable;
2512 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2513 (ins VR64:$src1, i64mem:$src2),
2514 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2516 (IntId64 VR64:$src1,
2517 (bitconvert (memopv2i32 addr:$src2))))]>;
2519 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2520 (ins VR128:$src1, VR128:$src2),
2521 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2522 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2524 let isCommutable = Commutable;
2526 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2527 (ins VR128:$src1, i128mem:$src2),
2528 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2530 (IntId128 VR128:$src1,
2531 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2535 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2536 int_x86_ssse3_phadd_w,
2537 int_x86_ssse3_phadd_w_128, 1>;
2538 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2539 int_x86_ssse3_phadd_d,
2540 int_x86_ssse3_phadd_d_128, 1>;
2541 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2542 int_x86_ssse3_phadd_sw,
2543 int_x86_ssse3_phadd_sw_128, 1>;
2544 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2545 int_x86_ssse3_phsub_w,
2546 int_x86_ssse3_phsub_w_128>;
2547 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2548 int_x86_ssse3_phsub_d,
2549 int_x86_ssse3_phsub_d_128>;
2550 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2551 int_x86_ssse3_phsub_sw,
2552 int_x86_ssse3_phsub_sw_128>;
2553 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2554 int_x86_ssse3_pmadd_ub_sw,
2555 int_x86_ssse3_pmadd_ub_sw_128, 1>;
2556 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2557 int_x86_ssse3_pmul_hr_sw,
2558 int_x86_ssse3_pmul_hr_sw_128, 1>;
2559 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2560 int_x86_ssse3_pshuf_b,
2561 int_x86_ssse3_pshuf_b_128>;
2562 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2563 int_x86_ssse3_psign_b,
2564 int_x86_ssse3_psign_b_128>;
2565 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2566 int_x86_ssse3_psign_w,
2567 int_x86_ssse3_psign_w_128>;
2568 defm PSIGND : SS3I_binop_rm_int_32<0x09, "psignd",
2569 int_x86_ssse3_psign_d,
2570 int_x86_ssse3_psign_d_128>;
2572 let isTwoAddress = 1 in {
2573 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2574 (ins VR64:$src1, VR64:$src2, i16imm:$src3),
2575 "palignr\t{$src2, $dst|$dst, $src2}",
2577 (int_x86_ssse3_palign_r
2578 VR64:$src1, VR64:$src2,
2580 def PALIGNR64rm : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2581 (ins VR64:$src1, i64mem:$src2, i16imm:$src3),
2582 "palignr\t{$src2, $dst|$dst, $src2}",
2584 (int_x86_ssse3_palign_r
2586 (bitconvert (memopv2i32 addr:$src2)),
2589 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2590 (ins VR128:$src1, VR128:$src2, i32imm:$src3),
2591 "palignr\t{$src2, $dst|$dst, $src2}",
2593 (int_x86_ssse3_palign_r_128
2594 VR128:$src1, VR128:$src2,
2595 imm:$src3))]>, OpSize;
2596 def PALIGNR128rm : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2597 (ins VR128:$src1, i128mem:$src2, i32imm:$src3),
2598 "palignr\t{$src2, $dst|$dst, $src2}",
2600 (int_x86_ssse3_palign_r_128
2602 (bitconvert (memopv4i32 addr:$src2)),
2603 imm:$src3))]>, OpSize;
2606 //===----------------------------------------------------------------------===//
2607 // Non-Instruction Patterns
2608 //===----------------------------------------------------------------------===//
2610 // 128-bit vector undef's.
2611 def : Pat<(v4f32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2612 def : Pat<(v2f64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2613 def : Pat<(v16i8 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2614 def : Pat<(v8i16 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2615 def : Pat<(v4i32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2616 def : Pat<(v2i64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
2618 // 128-bit vector all zero's.
2619 def : Pat<(v16i8 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2620 def : Pat<(v8i16 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2621 def : Pat<(v4i32 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2622 def : Pat<(v2i64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2623 def : Pat<(v2f64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
2625 // 128-bit vector all one's.
2626 def : Pat<(v16i8 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2627 def : Pat<(v8i16 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2628 def : Pat<(v4i32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2629 def : Pat<(v2i64 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
2630 def : Pat<(v4f32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE1]>;
2633 // Scalar to v8i16 / v16i8. The source may be a GR32, but only the lower 8 or
2635 def : Pat<(v8i16 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2636 Requires<[HasSSE2]>;
2637 def : Pat<(v16i8 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
2638 Requires<[HasSSE2]>;
2641 let Predicates = [HasSSE2] in {
2642 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2643 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2644 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2645 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2646 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2647 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2648 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2649 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2650 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2651 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2652 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2653 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2654 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2655 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2656 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2657 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2658 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2659 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2660 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2661 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2662 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2663 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2664 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2665 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2666 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2667 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2668 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2669 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2670 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2671 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2674 // Move scalar to XMM zero-extended
2675 // movd to XMM register zero-extends
2676 let AddedComplexity = 15 in {
2677 def : Pat<(v8i16 (vector_shuffle immAllZerosV,
2678 (v8i16 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
2679 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
2680 def : Pat<(v16i8 (vector_shuffle immAllZerosV,
2681 (v16i8 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
2682 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
2683 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2684 def : Pat<(v2f64 (vector_shuffle immAllZerosV,
2685 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
2686 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
2687 def : Pat<(v4f32 (vector_shuffle immAllZerosV,
2688 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
2689 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
2692 // Splat v2f64 / v2i64
2693 let AddedComplexity = 10 in {
2694 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2695 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2696 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2697 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2698 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
2699 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2700 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
2701 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2705 def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SSE_splat_mask:$sm),
2706 (SHUFPSrri VR128:$src, VR128:$src, SSE_splat_mask:$sm)>,
2707 Requires<[HasSSE1]>;
2709 // Special unary SHUFPSrri case.
2710 // FIXME: when we want non two-address code, then we should use PSHUFD?
2711 def : Pat<(vector_shuffle (v4f32 VR128:$src1), (undef),
2712 SHUFP_unary_shuffle_mask:$sm),
2713 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2714 Requires<[HasSSE1]>;
2715 // Special unary SHUFPDrri case.
2716 def : Pat<(vector_shuffle (v2f64 VR128:$src1), (undef),
2717 SHUFP_unary_shuffle_mask:$sm),
2718 (SHUFPDrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2719 Requires<[HasSSE2]>;
2720 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2721 def : Pat<(vector_shuffle (memopv4f32 addr:$src1), (undef),
2722 SHUFP_unary_shuffle_mask:$sm),
2723 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
2724 Requires<[HasSSE2]>;
2725 // Special binary v4i32 shuffle cases with SHUFPS.
2726 def : Pat<(vector_shuffle (v4i32 VR128:$src1), (v4i32 VR128:$src2),
2727 PSHUFD_binary_shuffle_mask:$sm),
2728 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2729 Requires<[HasSSE2]>;
2730 def : Pat<(vector_shuffle (v4i32 VR128:$src1),
2731 (bc_v4i32 (memopv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm),
2732 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
2733 Requires<[HasSSE2]>;
2735 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2736 let AddedComplexity = 10 in {
2737 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2738 UNPCKL_v_undef_shuffle_mask)),
2739 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2740 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2741 UNPCKL_v_undef_shuffle_mask)),
2742 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2743 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2744 UNPCKL_v_undef_shuffle_mask)),
2745 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2746 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2747 UNPCKL_v_undef_shuffle_mask)),
2748 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2751 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2752 let AddedComplexity = 10 in {
2753 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
2754 UNPCKH_v_undef_shuffle_mask)),
2755 (UNPCKHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2756 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
2757 UNPCKH_v_undef_shuffle_mask)),
2758 (PUNPCKHBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2759 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
2760 UNPCKH_v_undef_shuffle_mask)),
2761 (PUNPCKHWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2762 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2763 UNPCKH_v_undef_shuffle_mask)),
2764 (PUNPCKHDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2767 let AddedComplexity = 15 in {
2768 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2769 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2770 MOVHP_shuffle_mask)),
2771 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2773 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2774 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2775 MOVHLPS_shuffle_mask)),
2776 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2778 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2779 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2780 MOVHLPS_v_undef_shuffle_mask)),
2781 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2782 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2783 MOVHLPS_v_undef_shuffle_mask)),
2784 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2787 let AddedComplexity = 20 in {
2788 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2789 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2790 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2791 MOVLP_shuffle_mask)),
2792 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2793 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2794 MOVLP_shuffle_mask)),
2795 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2796 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (memopv4f32 addr:$src2),
2797 MOVHP_shuffle_mask)),
2798 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2799 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (memopv2f64 addr:$src2),
2800 MOVHP_shuffle_mask)),
2801 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2803 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2804 MOVLP_shuffle_mask)),
2805 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2806 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2807 MOVLP_shuffle_mask)),
2808 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2809 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)),
2810 MOVHP_shuffle_mask)),
2811 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2812 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (memopv2i64 addr:$src2),
2813 MOVLP_shuffle_mask)),
2814 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2817 let AddedComplexity = 15 in {
2818 // Setting the lowest element in the vector.
2819 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2820 MOVL_shuffle_mask)),
2821 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2822 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2823 MOVL_shuffle_mask)),
2824 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2826 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2827 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2828 MOVLP_shuffle_mask)),
2829 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2830 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2831 MOVLP_shuffle_mask)),
2832 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2835 // Set lowest element and zero upper elements.
2836 let AddedComplexity = 20 in
2837 def : Pat<(bc_v2i64 (vector_shuffle immAllZerosV,
2838 (v2f64 (scalar_to_vector (loadf64 addr:$src))),
2839 MOVL_shuffle_mask)),
2840 (MOVZQI2PQIrm addr:$src)>, Requires<[HasSSE2]>;
2842 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2843 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2844 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2845 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2846 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2847 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2848 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2849 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2850 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2851 Requires<[HasSSE2]>;
2852 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2853 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2854 Requires<[HasSSE2]>;
2855 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2856 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2857 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2858 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2859 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2860 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2861 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2862 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2863 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
2864 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2865 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
2866 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2867 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
2868 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2869 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
2870 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2872 // Some special case pandn patterns.
2873 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2875 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2876 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2878 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2879 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2881 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2883 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2884 (memopv2i64 addr:$src2))),
2885 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2886 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2887 (memopv2i64 addr:$src2))),
2888 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2889 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2890 (memopv2i64 addr:$src2))),
2891 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2893 // Use movaps / movups for SSE integer load / store (one byte shorter).
2894 def : Pat<(alignedloadv4i32 addr:$src),
2895 (MOVAPSrm addr:$src)>, Requires<[HasSSE1]>;
2896 def : Pat<(loadv4i32 addr:$src),
2897 (MOVUPSrm addr:$src)>, Requires<[HasSSE1]>;
2898 def : Pat<(alignedloadv2i64 addr:$src),
2899 (MOVAPSrm addr:$src)>, Requires<[HasSSE2]>;
2900 def : Pat<(loadv2i64 addr:$src),
2901 (MOVUPSrm addr:$src)>, Requires<[HasSSE2]>;
2903 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
2904 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2905 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
2906 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2907 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
2908 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2909 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
2910 (MOVAPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2911 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
2912 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2913 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
2914 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2915 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
2916 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2917 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
2918 (MOVUPSmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
2920 // (vextract (v4i32 bc (v4f32 s2v (f32 load $addr))), 0) -> (i32 load $addr)
2921 def : Pat<(vector_extract
2922 (bc_v4i32 (v4f32 (scalar_to_vector (loadf32 addr:$src)))), (iPTR 0)),
2923 (MOV32rm addr:$src)>, Requires<[HasSSE2]>;
2924 def : Pat<(vector_extract
2925 (bc_v2i64 (v2f64 (scalar_to_vector (loadf64 addr:$src)))), (iPTR 0)),
2926 (MOV64rm addr:$src)>, Requires<[HasSSE2, In64BitMode]>;