1 //===-- X86InstrSSE.td - SSE Instruction Set ---------------*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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 //===----------------------------------------------------------------------===//
16 class OpndItins<InstrItinClass arg_rr, InstrItinClass arg_rm> {
17 InstrItinClass rr = arg_rr;
18 InstrItinClass rm = arg_rm;
19 // InstrSchedModel info.
20 X86FoldableSchedWrite Sched = WriteFAdd;
23 class SizeItins<OpndItins arg_s, OpndItins arg_d> {
29 class ShiftOpndItins<InstrItinClass arg_rr, InstrItinClass arg_rm,
30 InstrItinClass arg_ri> {
31 InstrItinClass rr = arg_rr;
32 InstrItinClass rm = arg_rm;
33 InstrItinClass ri = arg_ri;
38 let Sched = WriteFAdd in {
39 def SSE_ALU_F32S : OpndItins<
40 IIC_SSE_ALU_F32S_RR, IIC_SSE_ALU_F32S_RM
43 def SSE_ALU_F64S : OpndItins<
44 IIC_SSE_ALU_F64S_RR, IIC_SSE_ALU_F64S_RM
48 def SSE_ALU_ITINS_S : SizeItins<
49 SSE_ALU_F32S, SSE_ALU_F64S
52 let Sched = WriteFMul in {
53 def SSE_MUL_F32S : OpndItins<
54 IIC_SSE_MUL_F32S_RR, IIC_SSE_MUL_F64S_RM
57 def SSE_MUL_F64S : OpndItins<
58 IIC_SSE_MUL_F64S_RR, IIC_SSE_MUL_F64S_RM
62 def SSE_MUL_ITINS_S : SizeItins<
63 SSE_MUL_F32S, SSE_MUL_F64S
66 let Sched = WriteFDiv in {
67 def SSE_DIV_F32S : OpndItins<
68 IIC_SSE_DIV_F32S_RR, IIC_SSE_DIV_F64S_RM
71 def SSE_DIV_F64S : OpndItins<
72 IIC_SSE_DIV_F64S_RR, IIC_SSE_DIV_F64S_RM
76 def SSE_DIV_ITINS_S : SizeItins<
77 SSE_DIV_F32S, SSE_DIV_F64S
81 let Sched = WriteFAdd in {
82 def SSE_ALU_F32P : OpndItins<
83 IIC_SSE_ALU_F32P_RR, IIC_SSE_ALU_F32P_RM
86 def SSE_ALU_F64P : OpndItins<
87 IIC_SSE_ALU_F64P_RR, IIC_SSE_ALU_F64P_RM
91 def SSE_ALU_ITINS_P : SizeItins<
92 SSE_ALU_F32P, SSE_ALU_F64P
95 let Sched = WriteFMul in {
96 def SSE_MUL_F32P : OpndItins<
97 IIC_SSE_MUL_F32P_RR, IIC_SSE_MUL_F64P_RM
100 def SSE_MUL_F64P : OpndItins<
101 IIC_SSE_MUL_F64P_RR, IIC_SSE_MUL_F64P_RM
105 def SSE_MUL_ITINS_P : SizeItins<
106 SSE_MUL_F32P, SSE_MUL_F64P
109 let Sched = WriteFDiv in {
110 def SSE_DIV_F32P : OpndItins<
111 IIC_SSE_DIV_F32P_RR, IIC_SSE_DIV_F64P_RM
114 def SSE_DIV_F64P : OpndItins<
115 IIC_SSE_DIV_F64P_RR, IIC_SSE_DIV_F64P_RM
119 def SSE_DIV_ITINS_P : SizeItins<
120 SSE_DIV_F32P, SSE_DIV_F64P
123 let Sched = WriteVecLogic in
124 def SSE_VEC_BIT_ITINS_P : OpndItins<
125 IIC_SSE_BIT_P_RR, IIC_SSE_BIT_P_RM
128 def SSE_BIT_ITINS_P : OpndItins<
129 IIC_SSE_BIT_P_RR, IIC_SSE_BIT_P_RM
132 let Sched = WriteVecALU in {
133 def SSE_INTALU_ITINS_P : OpndItins<
134 IIC_SSE_INTALU_P_RR, IIC_SSE_INTALU_P_RM
137 def SSE_INTALUQ_ITINS_P : OpndItins<
138 IIC_SSE_INTALUQ_P_RR, IIC_SSE_INTALUQ_P_RM
142 let Sched = WriteVecIMul in
143 def SSE_INTMUL_ITINS_P : OpndItins<
144 IIC_SSE_INTMUL_P_RR, IIC_SSE_INTMUL_P_RM
147 def SSE_INTSHIFT_ITINS_P : ShiftOpndItins<
148 IIC_SSE_INTSH_P_RR, IIC_SSE_INTSH_P_RM, IIC_SSE_INTSH_P_RI
151 def SSE_MOVA_ITINS : OpndItins<
152 IIC_SSE_MOVA_P_RR, IIC_SSE_MOVA_P_RM
155 def SSE_MOVU_ITINS : OpndItins<
156 IIC_SSE_MOVU_P_RR, IIC_SSE_MOVU_P_RM
159 def SSE_DPPD_ITINS : OpndItins<
160 IIC_SSE_DPPD_RR, IIC_SSE_DPPD_RM
163 def SSE_DPPS_ITINS : OpndItins<
164 IIC_SSE_DPPS_RR, IIC_SSE_DPPD_RM
167 def DEFAULT_ITINS : OpndItins<
168 IIC_ALU_NONMEM, IIC_ALU_MEM
171 def SSE_EXTRACT_ITINS : OpndItins<
172 IIC_SSE_EXTRACTPS_RR, IIC_SSE_EXTRACTPS_RM
175 def SSE_INSERT_ITINS : OpndItins<
176 IIC_SSE_INSERTPS_RR, IIC_SSE_INSERTPS_RM
179 let Sched = WriteMPSAD in
180 def SSE_MPSADBW_ITINS : OpndItins<
181 IIC_SSE_MPSADBW_RR, IIC_SSE_MPSADBW_RM
184 let Sched = WriteVecIMul in
185 def SSE_PMULLD_ITINS : OpndItins<
186 IIC_SSE_PMULLD_RR, IIC_SSE_PMULLD_RM
189 // Definitions for backward compatibility.
190 // The instructions mapped on these definitions uses a different itinerary
191 // than the actual scheduling model.
192 let Sched = WriteShuffle in
193 def DEFAULT_ITINS_SHUFFLESCHED : OpndItins<
194 IIC_ALU_NONMEM, IIC_ALU_MEM
197 let Sched = WriteVecIMul in
198 def DEFAULT_ITINS_VECIMULSCHED : OpndItins<
199 IIC_ALU_NONMEM, IIC_ALU_MEM
202 let Sched = WriteShuffle in
203 def SSE_INTALU_ITINS_SHUFF_P : OpndItins<
204 IIC_SSE_INTALU_P_RR, IIC_SSE_INTALU_P_RM
207 let Sched = WriteMPSAD in
208 def DEFAULT_ITINS_MPSADSCHED : OpndItins<
209 IIC_ALU_NONMEM, IIC_ALU_MEM
212 let Sched = WriteFBlend in
213 def DEFAULT_ITINS_FBLENDSCHED : OpndItins<
214 IIC_ALU_NONMEM, IIC_ALU_MEM
217 let Sched = WriteBlend in
218 def DEFAULT_ITINS_BLENDSCHED : OpndItins<
219 IIC_ALU_NONMEM, IIC_ALU_MEM
222 let Sched = WriteVarBlend in
223 def DEFAULT_ITINS_VARBLENDSCHED : OpndItins<
224 IIC_ALU_NONMEM, IIC_ALU_MEM
227 let Sched = WriteFBlend in
228 def SSE_INTALU_ITINS_FBLEND_P : OpndItins<
229 IIC_SSE_INTALU_P_RR, IIC_SSE_INTALU_P_RM
232 let Sched = WriteBlend in
233 def SSE_INTALU_ITINS_BLEND_P : OpndItins<
234 IIC_SSE_INTALU_P_RR, IIC_SSE_INTALU_P_RM
237 //===----------------------------------------------------------------------===//
238 // SSE 1 & 2 Instructions Classes
239 //===----------------------------------------------------------------------===//
241 /// sse12_fp_scalar - SSE 1 & 2 scalar instructions class
242 multiclass sse12_fp_scalar<bits<8> opc, string OpcodeStr, SDNode OpNode,
243 RegisterClass RC, X86MemOperand x86memop,
246 let isCommutable = 1 in {
247 def rr : SI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
249 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
250 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
251 [(set RC:$dst, (OpNode RC:$src1, RC:$src2))], itins.rr>,
252 Sched<[itins.Sched]>;
254 def rm : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
256 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
257 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
258 [(set RC:$dst, (OpNode RC:$src1, (load addr:$src2)))], itins.rm>,
259 Sched<[itins.Sched.Folded, ReadAfterLd]>;
262 /// sse12_fp_scalar_int - SSE 1 & 2 scalar instructions intrinsics class
263 multiclass sse12_fp_scalar_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
264 string asm, string SSEVer, string FPSizeStr,
265 Operand memopr, ComplexPattern mem_cpat,
268 let isCodeGenOnly = 1 in {
269 def rr_Int : SI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
271 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
272 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
273 [(set RC:$dst, (!cast<Intrinsic>(
274 !strconcat("int_x86_sse", SSEVer, "_", OpcodeStr, FPSizeStr))
275 RC:$src1, RC:$src2))], itins.rr>,
276 Sched<[itins.Sched]>;
277 def rm_Int : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memopr:$src2),
279 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
280 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
281 [(set RC:$dst, (!cast<Intrinsic>(!strconcat("int_x86_sse",
282 SSEVer, "_", OpcodeStr, FPSizeStr))
283 RC:$src1, mem_cpat:$src2))], itins.rm>,
284 Sched<[itins.Sched.Folded, ReadAfterLd]>;
288 /// sse12_fp_packed - SSE 1 & 2 packed instructions class
289 multiclass sse12_fp_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
290 RegisterClass RC, ValueType vt,
291 X86MemOperand x86memop, PatFrag mem_frag,
292 Domain d, OpndItins itins, bit Is2Addr = 1> {
293 let isCommutable = 1 in
294 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
296 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
297 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
298 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], itins.rr, d>,
299 Sched<[itins.Sched]>;
301 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
303 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
304 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
305 [(set RC:$dst, (OpNode RC:$src1, (mem_frag addr:$src2)))],
307 Sched<[itins.Sched.Folded, ReadAfterLd]>;
310 /// sse12_fp_packed_logical_rm - SSE 1 & 2 packed instructions class
311 multiclass sse12_fp_packed_logical_rm<bits<8> opc, RegisterClass RC, Domain d,
312 string OpcodeStr, X86MemOperand x86memop,
313 list<dag> pat_rr, list<dag> pat_rm,
315 let isCommutable = 1, hasSideEffects = 0 in
316 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
318 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
319 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
320 pat_rr, NoItinerary, d>,
321 Sched<[WriteVecLogic]>;
322 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
324 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
325 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
326 pat_rm, NoItinerary, d>,
327 Sched<[WriteVecLogicLd, ReadAfterLd]>;
330 //===----------------------------------------------------------------------===//
331 // Non-instruction patterns
332 //===----------------------------------------------------------------------===//
334 // A vector extract of the first f32/f64 position is a subregister copy
335 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
336 (COPY_TO_REGCLASS (v4f32 VR128:$src), FR32)>;
337 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
338 (COPY_TO_REGCLASS (v2f64 VR128:$src), FR64)>;
340 // A 128-bit subvector extract from the first 256-bit vector position
341 // is a subregister copy that needs no instruction.
342 def : Pat<(v4i32 (extract_subvector (v8i32 VR256:$src), (iPTR 0))),
343 (v4i32 (EXTRACT_SUBREG (v8i32 VR256:$src), sub_xmm))>;
344 def : Pat<(v4f32 (extract_subvector (v8f32 VR256:$src), (iPTR 0))),
345 (v4f32 (EXTRACT_SUBREG (v8f32 VR256:$src), sub_xmm))>;
347 def : Pat<(v2i64 (extract_subvector (v4i64 VR256:$src), (iPTR 0))),
348 (v2i64 (EXTRACT_SUBREG (v4i64 VR256:$src), sub_xmm))>;
349 def : Pat<(v2f64 (extract_subvector (v4f64 VR256:$src), (iPTR 0))),
350 (v2f64 (EXTRACT_SUBREG (v4f64 VR256:$src), sub_xmm))>;
352 def : Pat<(v8i16 (extract_subvector (v16i16 VR256:$src), (iPTR 0))),
353 (v8i16 (EXTRACT_SUBREG (v16i16 VR256:$src), sub_xmm))>;
354 def : Pat<(v16i8 (extract_subvector (v32i8 VR256:$src), (iPTR 0))),
355 (v16i8 (EXTRACT_SUBREG (v32i8 VR256:$src), sub_xmm))>;
357 // A 128-bit subvector insert to the first 256-bit vector position
358 // is a subregister copy that needs no instruction.
359 let AddedComplexity = 25 in { // to give priority over vinsertf128rm
360 def : Pat<(insert_subvector undef, (v2i64 VR128:$src), (iPTR 0)),
361 (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
362 def : Pat<(insert_subvector undef, (v2f64 VR128:$src), (iPTR 0)),
363 (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
364 def : Pat<(insert_subvector undef, (v4i32 VR128:$src), (iPTR 0)),
365 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
366 def : Pat<(insert_subvector undef, (v4f32 VR128:$src), (iPTR 0)),
367 (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
368 def : Pat<(insert_subvector undef, (v8i16 VR128:$src), (iPTR 0)),
369 (INSERT_SUBREG (v16i16 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
370 def : Pat<(insert_subvector undef, (v16i8 VR128:$src), (iPTR 0)),
371 (INSERT_SUBREG (v32i8 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
374 // Implicitly promote a 32-bit scalar to a vector.
375 def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
376 (COPY_TO_REGCLASS FR32:$src, VR128)>;
377 def : Pat<(v8f32 (scalar_to_vector FR32:$src)),
378 (COPY_TO_REGCLASS FR32:$src, VR128)>;
379 // Implicitly promote a 64-bit scalar to a vector.
380 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
381 (COPY_TO_REGCLASS FR64:$src, VR128)>;
382 def : Pat<(v4f64 (scalar_to_vector FR64:$src)),
383 (COPY_TO_REGCLASS FR64:$src, VR128)>;
385 // Bitcasts between 128-bit vector types. Return the original type since
386 // no instruction is needed for the conversion
387 let Predicates = [HasSSE2] in {
388 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
389 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
390 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
391 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
392 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
393 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
394 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
395 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
396 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
397 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
398 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
399 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
400 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
401 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
402 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
403 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
404 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
405 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
406 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
407 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
408 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
409 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
410 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
411 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
412 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
413 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
414 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
415 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
416 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
417 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
420 // Bitcasts between 256-bit vector types. Return the original type since
421 // no instruction is needed for the conversion
422 let Predicates = [HasAVX] in {
423 def : Pat<(v4f64 (bitconvert (v8f32 VR256:$src))), (v4f64 VR256:$src)>;
424 def : Pat<(v4f64 (bitconvert (v8i32 VR256:$src))), (v4f64 VR256:$src)>;
425 def : Pat<(v4f64 (bitconvert (v4i64 VR256:$src))), (v4f64 VR256:$src)>;
426 def : Pat<(v4f64 (bitconvert (v16i16 VR256:$src))), (v4f64 VR256:$src)>;
427 def : Pat<(v4f64 (bitconvert (v32i8 VR256:$src))), (v4f64 VR256:$src)>;
428 def : Pat<(v8f32 (bitconvert (v8i32 VR256:$src))), (v8f32 VR256:$src)>;
429 def : Pat<(v8f32 (bitconvert (v4i64 VR256:$src))), (v8f32 VR256:$src)>;
430 def : Pat<(v8f32 (bitconvert (v4f64 VR256:$src))), (v8f32 VR256:$src)>;
431 def : Pat<(v8f32 (bitconvert (v32i8 VR256:$src))), (v8f32 VR256:$src)>;
432 def : Pat<(v8f32 (bitconvert (v16i16 VR256:$src))), (v8f32 VR256:$src)>;
433 def : Pat<(v4i64 (bitconvert (v8f32 VR256:$src))), (v4i64 VR256:$src)>;
434 def : Pat<(v4i64 (bitconvert (v8i32 VR256:$src))), (v4i64 VR256:$src)>;
435 def : Pat<(v4i64 (bitconvert (v4f64 VR256:$src))), (v4i64 VR256:$src)>;
436 def : Pat<(v4i64 (bitconvert (v32i8 VR256:$src))), (v4i64 VR256:$src)>;
437 def : Pat<(v4i64 (bitconvert (v16i16 VR256:$src))), (v4i64 VR256:$src)>;
438 def : Pat<(v32i8 (bitconvert (v4f64 VR256:$src))), (v32i8 VR256:$src)>;
439 def : Pat<(v32i8 (bitconvert (v4i64 VR256:$src))), (v32i8 VR256:$src)>;
440 def : Pat<(v32i8 (bitconvert (v8f32 VR256:$src))), (v32i8 VR256:$src)>;
441 def : Pat<(v32i8 (bitconvert (v8i32 VR256:$src))), (v32i8 VR256:$src)>;
442 def : Pat<(v32i8 (bitconvert (v16i16 VR256:$src))), (v32i8 VR256:$src)>;
443 def : Pat<(v8i32 (bitconvert (v32i8 VR256:$src))), (v8i32 VR256:$src)>;
444 def : Pat<(v8i32 (bitconvert (v16i16 VR256:$src))), (v8i32 VR256:$src)>;
445 def : Pat<(v8i32 (bitconvert (v8f32 VR256:$src))), (v8i32 VR256:$src)>;
446 def : Pat<(v8i32 (bitconvert (v4i64 VR256:$src))), (v8i32 VR256:$src)>;
447 def : Pat<(v8i32 (bitconvert (v4f64 VR256:$src))), (v8i32 VR256:$src)>;
448 def : Pat<(v16i16 (bitconvert (v8f32 VR256:$src))), (v16i16 VR256:$src)>;
449 def : Pat<(v16i16 (bitconvert (v8i32 VR256:$src))), (v16i16 VR256:$src)>;
450 def : Pat<(v16i16 (bitconvert (v4i64 VR256:$src))), (v16i16 VR256:$src)>;
451 def : Pat<(v16i16 (bitconvert (v4f64 VR256:$src))), (v16i16 VR256:$src)>;
452 def : Pat<(v16i16 (bitconvert (v32i8 VR256:$src))), (v16i16 VR256:$src)>;
455 // Alias instructions that map fld0 to xorps for sse or vxorps for avx.
456 // This is expanded by ExpandPostRAPseudos.
457 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
458 isPseudo = 1, SchedRW = [WriteZero] in {
459 def FsFLD0SS : I<0, Pseudo, (outs FR32:$dst), (ins), "",
460 [(set FR32:$dst, fp32imm0)]>, Requires<[HasSSE1]>;
461 def FsFLD0SD : I<0, Pseudo, (outs FR64:$dst), (ins), "",
462 [(set FR64:$dst, fpimm0)]>, Requires<[HasSSE2]>;
465 //===----------------------------------------------------------------------===//
466 // AVX & SSE - Zero/One Vectors
467 //===----------------------------------------------------------------------===//
469 // Alias instruction that maps zero vector to pxor / xorp* for sse.
470 // This is expanded by ExpandPostRAPseudos to an xorps / vxorps, and then
471 // swizzled by ExecutionDepsFix to pxor.
472 // We set canFoldAsLoad because this can be converted to a constant-pool
473 // load of an all-zeros value if folding it would be beneficial.
474 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
475 isPseudo = 1, SchedRW = [WriteZero] in {
476 def V_SET0 : I<0, Pseudo, (outs VR128:$dst), (ins), "",
477 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
480 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
481 def : Pat<(v4i32 immAllZerosV), (V_SET0)>;
482 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
483 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
484 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
487 // The same as done above but for AVX. The 256-bit AVX1 ISA doesn't support PI,
488 // and doesn't need it because on sandy bridge the register is set to zero
489 // at the rename stage without using any execution unit, so SET0PSY
490 // and SET0PDY can be used for vector int instructions without penalty
491 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
492 isPseudo = 1, Predicates = [HasAVX], SchedRW = [WriteZero] in {
493 def AVX_SET0 : I<0, Pseudo, (outs VR256:$dst), (ins), "",
494 [(set VR256:$dst, (v8f32 immAllZerosV))]>;
497 let Predicates = [HasAVX] in
498 def : Pat<(v4f64 immAllZerosV), (AVX_SET0)>;
500 let Predicates = [HasAVX2] in {
501 def : Pat<(v4i64 immAllZerosV), (AVX_SET0)>;
502 def : Pat<(v8i32 immAllZerosV), (AVX_SET0)>;
503 def : Pat<(v16i16 immAllZerosV), (AVX_SET0)>;
504 def : Pat<(v32i8 immAllZerosV), (AVX_SET0)>;
507 // AVX1 has no support for 256-bit integer instructions, but since the 128-bit
508 // VPXOR instruction writes zero to its upper part, it's safe build zeros.
509 let Predicates = [HasAVX1Only] in {
510 def : Pat<(v32i8 immAllZerosV), (SUBREG_TO_REG (i8 0), (V_SET0), sub_xmm)>;
511 def : Pat<(bc_v32i8 (v8f32 immAllZerosV)),
512 (SUBREG_TO_REG (i8 0), (V_SET0), sub_xmm)>;
514 def : Pat<(v16i16 immAllZerosV), (SUBREG_TO_REG (i16 0), (V_SET0), sub_xmm)>;
515 def : Pat<(bc_v16i16 (v8f32 immAllZerosV)),
516 (SUBREG_TO_REG (i16 0), (V_SET0), sub_xmm)>;
518 def : Pat<(v8i32 immAllZerosV), (SUBREG_TO_REG (i32 0), (V_SET0), sub_xmm)>;
519 def : Pat<(bc_v8i32 (v8f32 immAllZerosV)),
520 (SUBREG_TO_REG (i32 0), (V_SET0), sub_xmm)>;
522 def : Pat<(v4i64 immAllZerosV), (SUBREG_TO_REG (i64 0), (V_SET0), sub_xmm)>;
523 def : Pat<(bc_v4i64 (v8f32 immAllZerosV)),
524 (SUBREG_TO_REG (i64 0), (V_SET0), sub_xmm)>;
527 // We set canFoldAsLoad because this can be converted to a constant-pool
528 // load of an all-ones value if folding it would be beneficial.
529 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
530 isPseudo = 1, SchedRW = [WriteZero] in {
531 def V_SETALLONES : I<0, Pseudo, (outs VR128:$dst), (ins), "",
532 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
533 let Predicates = [HasAVX2] in
534 def AVX2_SETALLONES : I<0, Pseudo, (outs VR256:$dst), (ins), "",
535 [(set VR256:$dst, (v8i32 immAllOnesV))]>;
539 //===----------------------------------------------------------------------===//
540 // SSE 1 & 2 - Move FP Scalar Instructions
542 // Move Instructions. Register-to-register movss/movsd is not used for FR32/64
543 // register copies because it's a partial register update; Register-to-register
544 // movss/movsd is not modeled as an INSERT_SUBREG because INSERT_SUBREG requires
545 // that the insert be implementable in terms of a copy, and just mentioned, we
546 // don't use movss/movsd for copies.
547 //===----------------------------------------------------------------------===//
549 multiclass sse12_move_rr<RegisterClass RC, SDNode OpNode, ValueType vt,
550 X86MemOperand x86memop, string base_opc,
552 def rr : SI<0x10, MRMSrcReg, (outs VR128:$dst),
553 (ins VR128:$src1, RC:$src2),
554 !strconcat(base_opc, asm_opr),
555 [(set VR128:$dst, (vt (OpNode VR128:$src1,
556 (scalar_to_vector RC:$src2))))],
557 IIC_SSE_MOV_S_RR>, Sched<[WriteFShuffle]>;
559 // For the disassembler
560 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in
561 def rr_REV : SI<0x11, MRMDestReg, (outs VR128:$dst),
562 (ins VR128:$src1, RC:$src2),
563 !strconcat(base_opc, asm_opr),
564 [], IIC_SSE_MOV_S_RR>, Sched<[WriteFShuffle]>;
567 multiclass sse12_move<RegisterClass RC, SDNode OpNode, ValueType vt,
568 X86MemOperand x86memop, string OpcodeStr> {
570 defm V#NAME : sse12_move_rr<RC, OpNode, vt, x86memop, OpcodeStr,
571 "\t{$src2, $src1, $dst|$dst, $src1, $src2}">,
574 def V#NAME#mr : SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, RC:$src),
575 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
576 [(store RC:$src, addr:$dst)], IIC_SSE_MOV_S_MR>,
577 VEX, VEX_LIG, Sched<[WriteStore]>;
579 let Constraints = "$src1 = $dst" in {
580 defm NAME : sse12_move_rr<RC, OpNode, vt, x86memop, OpcodeStr,
581 "\t{$src2, $dst|$dst, $src2}">;
584 def NAME#mr : SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, RC:$src),
585 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
586 [(store RC:$src, addr:$dst)], IIC_SSE_MOV_S_MR>,
590 // Loading from memory automatically zeroing upper bits.
591 multiclass sse12_move_rm<RegisterClass RC, X86MemOperand x86memop,
592 PatFrag mem_pat, string OpcodeStr> {
593 def V#NAME#rm : SI<0x10, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
594 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
595 [(set RC:$dst, (mem_pat addr:$src))],
596 IIC_SSE_MOV_S_RM>, VEX, VEX_LIG, Sched<[WriteLoad]>;
597 def NAME#rm : SI<0x10, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
598 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
599 [(set RC:$dst, (mem_pat addr:$src))],
600 IIC_SSE_MOV_S_RM>, Sched<[WriteLoad]>;
603 defm MOVSS : sse12_move<FR32, X86Movss, v4f32, f32mem, "movss">, XS;
604 defm MOVSD : sse12_move<FR64, X86Movsd, v2f64, f64mem, "movsd">, XD;
606 let canFoldAsLoad = 1, isReMaterializable = 1 in {
607 defm MOVSS : sse12_move_rm<FR32, f32mem, loadf32, "movss">, XS;
609 let AddedComplexity = 20 in
610 defm MOVSD : sse12_move_rm<FR64, f64mem, loadf64, "movsd">, XD;
614 let Predicates = [UseAVX] in {
615 let AddedComplexity = 20 in {
616 // MOVSSrm zeros the high parts of the register; represent this
617 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
618 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
619 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
620 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
621 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
622 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
623 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
625 // MOVSDrm zeros the high parts of the register; represent this
626 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
627 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
628 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
629 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
630 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
631 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
632 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
633 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
634 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
635 def : Pat<(v2f64 (X86vzload addr:$src)),
636 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
638 // Represent the same patterns above but in the form they appear for
640 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
641 (v4i32 (scalar_to_vector (loadi32 addr:$src))), (iPTR 0)))),
642 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_xmm)>;
643 def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
644 (v4f32 (scalar_to_vector (loadf32 addr:$src))), (iPTR 0)))),
645 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_xmm)>;
646 def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
647 (v2f64 (scalar_to_vector (loadf64 addr:$src))), (iPTR 0)))),
648 (SUBREG_TO_REG (i32 0), (VMOVSDrm addr:$src), sub_xmm)>;
650 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
651 (v2i64 (scalar_to_vector (loadi64 addr:$src))), (iPTR 0)))),
652 (SUBREG_TO_REG (i64 0), (VMOVSDrm addr:$src), sub_xmm)>;
654 // Extract and store.
655 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
657 (VMOVSSmr addr:$dst, (COPY_TO_REGCLASS (v4f32 VR128:$src), FR32))>;
658 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
660 (VMOVSDmr addr:$dst, (COPY_TO_REGCLASS (v2f64 VR128:$src), FR64))>;
662 // Shuffle with VMOVSS
663 def : Pat<(v4i32 (X86Movss VR128:$src1, VR128:$src2)),
664 (VMOVSSrr (v4i32 VR128:$src1),
665 (COPY_TO_REGCLASS (v4i32 VR128:$src2), FR32))>;
666 def : Pat<(v4f32 (X86Movss VR128:$src1, VR128:$src2)),
667 (VMOVSSrr (v4f32 VR128:$src1),
668 (COPY_TO_REGCLASS (v4f32 VR128:$src2), FR32))>;
671 def : Pat<(v8i32 (X86Movss VR256:$src1, VR256:$src2)),
672 (SUBREG_TO_REG (i32 0),
673 (VMOVSSrr (EXTRACT_SUBREG (v8i32 VR256:$src1), sub_xmm),
674 (EXTRACT_SUBREG (v8i32 VR256:$src2), sub_xmm)),
676 def : Pat<(v8f32 (X86Movss VR256:$src1, VR256:$src2)),
677 (SUBREG_TO_REG (i32 0),
678 (VMOVSSrr (EXTRACT_SUBREG (v8f32 VR256:$src1), sub_xmm),
679 (EXTRACT_SUBREG (v8f32 VR256:$src2), sub_xmm)),
682 // Shuffle with VMOVSD
683 def : Pat<(v2i64 (X86Movsd VR128:$src1, VR128:$src2)),
684 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
685 def : Pat<(v2f64 (X86Movsd VR128:$src1, VR128:$src2)),
686 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
687 def : Pat<(v4f32 (X86Movsd VR128:$src1, VR128:$src2)),
688 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
689 def : Pat<(v4i32 (X86Movsd VR128:$src1, VR128:$src2)),
690 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
693 def : Pat<(v4i64 (X86Movsd VR256:$src1, VR256:$src2)),
694 (SUBREG_TO_REG (i32 0),
695 (VMOVSDrr (EXTRACT_SUBREG (v4i64 VR256:$src1), sub_xmm),
696 (EXTRACT_SUBREG (v4i64 VR256:$src2), sub_xmm)),
698 def : Pat<(v4f64 (X86Movsd VR256:$src1, VR256:$src2)),
699 (SUBREG_TO_REG (i32 0),
700 (VMOVSDrr (EXTRACT_SUBREG (v4f64 VR256:$src1), sub_xmm),
701 (EXTRACT_SUBREG (v4f64 VR256:$src2), sub_xmm)),
704 // FIXME: Instead of a X86Movlps there should be a X86Movsd here, the problem
705 // is during lowering, where it's not possible to recognize the fold cause
706 // it has two uses through a bitcast. One use disappears at isel time and the
707 // fold opportunity reappears.
708 def : Pat<(v2f64 (X86Movlpd VR128:$src1, VR128:$src2)),
709 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
710 def : Pat<(v2i64 (X86Movlpd VR128:$src1, VR128:$src2)),
711 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
712 def : Pat<(v4f32 (X86Movlps VR128:$src1, VR128:$src2)),
713 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
714 def : Pat<(v4i32 (X86Movlps VR128:$src1, VR128:$src2)),
715 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
718 let Predicates = [UseSSE1] in {
719 let Predicates = [NoSSE41], AddedComplexity = 15 in {
720 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
721 // MOVSS to the lower bits.
722 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
723 (MOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
724 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
725 (MOVSSrr (v4f32 (V_SET0)), (COPY_TO_REGCLASS VR128:$src, FR32))>;
726 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
727 (MOVSSrr (v4i32 (V_SET0)), (COPY_TO_REGCLASS VR128:$src, FR32))>;
730 let AddedComplexity = 20 in {
731 // MOVSSrm already zeros the high parts of the register.
732 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
733 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
734 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
735 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
736 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
737 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
740 // Extract and store.
741 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
743 (MOVSSmr addr:$dst, (COPY_TO_REGCLASS VR128:$src, FR32))>;
745 // Shuffle with MOVSS
746 def : Pat<(v4i32 (X86Movss VR128:$src1, VR128:$src2)),
747 (MOVSSrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR32))>;
748 def : Pat<(v4f32 (X86Movss VR128:$src1, VR128:$src2)),
749 (MOVSSrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR32))>;
752 let Predicates = [UseSSE2] in {
753 let Predicates = [NoSSE41], AddedComplexity = 15 in {
754 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
755 // MOVSD to the lower bits.
756 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
757 (MOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
760 let AddedComplexity = 20 in {
761 // MOVSDrm already zeros the high parts of the register.
762 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
763 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
764 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
765 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
766 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
767 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
768 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
769 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
770 def : Pat<(v2f64 (X86vzload addr:$src)),
771 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
774 // Extract and store.
775 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
777 (MOVSDmr addr:$dst, (COPY_TO_REGCLASS VR128:$src, FR64))>;
779 // Shuffle with MOVSD
780 def : Pat<(v2i64 (X86Movsd VR128:$src1, VR128:$src2)),
781 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
782 def : Pat<(v2f64 (X86Movsd VR128:$src1, VR128:$src2)),
783 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
784 def : Pat<(v4f32 (X86Movsd VR128:$src1, VR128:$src2)),
785 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
786 def : Pat<(v4i32 (X86Movsd VR128:$src1, VR128:$src2)),
787 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
789 // FIXME: Instead of a X86Movlps there should be a X86Movsd here, the problem
790 // is during lowering, where it's not possible to recognize the fold cause
791 // it has two uses through a bitcast. One use disappears at isel time and the
792 // fold opportunity reappears.
793 def : Pat<(v2f64 (X86Movlpd VR128:$src1, VR128:$src2)),
794 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
795 def : Pat<(v2i64 (X86Movlpd VR128:$src1, VR128:$src2)),
796 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
797 def : Pat<(v4f32 (X86Movlps VR128:$src1, VR128:$src2)),
798 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
799 def : Pat<(v4i32 (X86Movlps VR128:$src1, VR128:$src2)),
800 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
803 //===----------------------------------------------------------------------===//
804 // SSE 1 & 2 - Move Aligned/Unaligned FP Instructions
805 //===----------------------------------------------------------------------===//
807 multiclass sse12_mov_packed<bits<8> opc, RegisterClass RC,
808 X86MemOperand x86memop, PatFrag ld_frag,
809 string asm, Domain d,
811 bit IsReMaterializable = 1> {
812 let neverHasSideEffects = 1 in
813 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
814 !strconcat(asm, "\t{$src, $dst|$dst, $src}"), [], itins.rr, d>,
815 Sched<[WriteFShuffle]>;
816 let canFoldAsLoad = 1, isReMaterializable = IsReMaterializable in
817 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
818 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
819 [(set RC:$dst, (ld_frag addr:$src))], itins.rm, d>,
823 let Predicates = [HasAVX, NoVLX] in {
824 defm VMOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
825 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
827 defm VMOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
828 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
830 defm VMOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
831 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
833 defm VMOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
834 "movupd", SSEPackedDouble, SSE_MOVU_ITINS, 0>,
837 defm VMOVAPSY : sse12_mov_packed<0x28, VR256, f256mem, alignedloadv8f32,
838 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
840 defm VMOVAPDY : sse12_mov_packed<0x28, VR256, f256mem, alignedloadv4f64,
841 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
843 defm VMOVUPSY : sse12_mov_packed<0x10, VR256, f256mem, loadv8f32,
844 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
846 defm VMOVUPDY : sse12_mov_packed<0x10, VR256, f256mem, loadv4f64,
847 "movupd", SSEPackedDouble, SSE_MOVU_ITINS, 0>,
851 let Predicates = [UseSSE1] in {
852 defm MOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
853 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
855 defm MOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
856 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
859 let Predicates = [UseSSE2] in {
860 defm MOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
861 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
863 defm MOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
864 "movupd", SSEPackedDouble, SSE_MOVU_ITINS, 0>,
868 let SchedRW = [WriteStore], Predicates = [HasAVX, NoVLX] in {
869 def VMOVAPSmr : VPSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
870 "movaps\t{$src, $dst|$dst, $src}",
871 [(alignedstore (v4f32 VR128:$src), addr:$dst)],
872 IIC_SSE_MOVA_P_MR>, VEX;
873 def VMOVAPDmr : VPDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
874 "movapd\t{$src, $dst|$dst, $src}",
875 [(alignedstore (v2f64 VR128:$src), addr:$dst)],
876 IIC_SSE_MOVA_P_MR>, VEX;
877 def VMOVUPSmr : VPSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
878 "movups\t{$src, $dst|$dst, $src}",
879 [(store (v4f32 VR128:$src), addr:$dst)],
880 IIC_SSE_MOVU_P_MR>, VEX;
881 def VMOVUPDmr : VPDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
882 "movupd\t{$src, $dst|$dst, $src}",
883 [(store (v2f64 VR128:$src), addr:$dst)],
884 IIC_SSE_MOVU_P_MR>, VEX;
885 def VMOVAPSYmr : VPSI<0x29, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
886 "movaps\t{$src, $dst|$dst, $src}",
887 [(alignedstore256 (v8f32 VR256:$src), addr:$dst)],
888 IIC_SSE_MOVA_P_MR>, VEX, VEX_L;
889 def VMOVAPDYmr : VPDI<0x29, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
890 "movapd\t{$src, $dst|$dst, $src}",
891 [(alignedstore256 (v4f64 VR256:$src), addr:$dst)],
892 IIC_SSE_MOVA_P_MR>, VEX, VEX_L;
893 def VMOVUPSYmr : VPSI<0x11, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
894 "movups\t{$src, $dst|$dst, $src}",
895 [(store (v8f32 VR256:$src), addr:$dst)],
896 IIC_SSE_MOVU_P_MR>, VEX, VEX_L;
897 def VMOVUPDYmr : VPDI<0x11, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
898 "movupd\t{$src, $dst|$dst, $src}",
899 [(store (v4f64 VR256:$src), addr:$dst)],
900 IIC_SSE_MOVU_P_MR>, VEX, VEX_L;
904 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
905 SchedRW = [WriteFShuffle] in {
906 def VMOVAPSrr_REV : VPSI<0x29, MRMDestReg, (outs VR128:$dst),
908 "movaps\t{$src, $dst|$dst, $src}", [],
909 IIC_SSE_MOVA_P_RR>, VEX;
910 def VMOVAPDrr_REV : VPDI<0x29, MRMDestReg, (outs VR128:$dst),
912 "movapd\t{$src, $dst|$dst, $src}", [],
913 IIC_SSE_MOVA_P_RR>, VEX;
914 def VMOVUPSrr_REV : VPSI<0x11, MRMDestReg, (outs VR128:$dst),
916 "movups\t{$src, $dst|$dst, $src}", [],
917 IIC_SSE_MOVU_P_RR>, VEX;
918 def VMOVUPDrr_REV : VPDI<0x11, MRMDestReg, (outs VR128:$dst),
920 "movupd\t{$src, $dst|$dst, $src}", [],
921 IIC_SSE_MOVU_P_RR>, VEX;
922 def VMOVAPSYrr_REV : VPSI<0x29, MRMDestReg, (outs VR256:$dst),
924 "movaps\t{$src, $dst|$dst, $src}", [],
925 IIC_SSE_MOVA_P_RR>, VEX, VEX_L;
926 def VMOVAPDYrr_REV : VPDI<0x29, MRMDestReg, (outs VR256:$dst),
928 "movapd\t{$src, $dst|$dst, $src}", [],
929 IIC_SSE_MOVA_P_RR>, VEX, VEX_L;
930 def VMOVUPSYrr_REV : VPSI<0x11, MRMDestReg, (outs VR256:$dst),
932 "movups\t{$src, $dst|$dst, $src}", [],
933 IIC_SSE_MOVU_P_RR>, VEX, VEX_L;
934 def VMOVUPDYrr_REV : VPDI<0x11, MRMDestReg, (outs VR256:$dst),
936 "movupd\t{$src, $dst|$dst, $src}", [],
937 IIC_SSE_MOVU_P_RR>, VEX, VEX_L;
940 let Predicates = [HasAVX] in {
941 def : Pat<(v8i32 (X86vzmovl
942 (insert_subvector undef, (v4i32 VR128:$src), (iPTR 0)))),
943 (SUBREG_TO_REG (i32 0), (VMOVAPSrr VR128:$src), sub_xmm)>;
944 def : Pat<(v4i64 (X86vzmovl
945 (insert_subvector undef, (v2i64 VR128:$src), (iPTR 0)))),
946 (SUBREG_TO_REG (i32 0), (VMOVAPSrr VR128:$src), sub_xmm)>;
947 def : Pat<(v8f32 (X86vzmovl
948 (insert_subvector undef, (v4f32 VR128:$src), (iPTR 0)))),
949 (SUBREG_TO_REG (i32 0), (VMOVAPSrr VR128:$src), sub_xmm)>;
950 def : Pat<(v4f64 (X86vzmovl
951 (insert_subvector undef, (v2f64 VR128:$src), (iPTR 0)))),
952 (SUBREG_TO_REG (i32 0), (VMOVAPSrr VR128:$src), sub_xmm)>;
956 def : Pat<(int_x86_avx_storeu_ps_256 addr:$dst, VR256:$src),
957 (VMOVUPSYmr addr:$dst, VR256:$src)>;
958 def : Pat<(int_x86_avx_storeu_pd_256 addr:$dst, VR256:$src),
959 (VMOVUPDYmr addr:$dst, VR256:$src)>;
961 let SchedRW = [WriteStore] in {
962 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
963 "movaps\t{$src, $dst|$dst, $src}",
964 [(alignedstore (v4f32 VR128:$src), addr:$dst)],
966 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
967 "movapd\t{$src, $dst|$dst, $src}",
968 [(alignedstore (v2f64 VR128:$src), addr:$dst)],
970 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
971 "movups\t{$src, $dst|$dst, $src}",
972 [(store (v4f32 VR128:$src), addr:$dst)],
974 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
975 "movupd\t{$src, $dst|$dst, $src}",
976 [(store (v2f64 VR128:$src), addr:$dst)],
981 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
982 SchedRW = [WriteFShuffle] in {
983 def MOVAPSrr_REV : PSI<0x29, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
984 "movaps\t{$src, $dst|$dst, $src}", [],
986 def MOVAPDrr_REV : PDI<0x29, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
987 "movapd\t{$src, $dst|$dst, $src}", [],
989 def MOVUPSrr_REV : PSI<0x11, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
990 "movups\t{$src, $dst|$dst, $src}", [],
992 def MOVUPDrr_REV : PDI<0x11, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
993 "movupd\t{$src, $dst|$dst, $src}", [],
997 let Predicates = [HasAVX] in {
998 def : Pat<(int_x86_sse_storeu_ps addr:$dst, VR128:$src),
999 (VMOVUPSmr addr:$dst, VR128:$src)>;
1000 def : Pat<(int_x86_sse2_storeu_pd addr:$dst, VR128:$src),
1001 (VMOVUPDmr addr:$dst, VR128:$src)>;
1004 let Predicates = [UseSSE1] in
1005 def : Pat<(int_x86_sse_storeu_ps addr:$dst, VR128:$src),
1006 (MOVUPSmr addr:$dst, VR128:$src)>;
1007 let Predicates = [UseSSE2] in
1008 def : Pat<(int_x86_sse2_storeu_pd addr:$dst, VR128:$src),
1009 (MOVUPDmr addr:$dst, VR128:$src)>;
1011 // Use vmovaps/vmovups for AVX integer load/store.
1012 let Predicates = [HasAVX, NoVLX] in {
1013 // 128-bit load/store
1014 def : Pat<(alignedloadv2i64 addr:$src),
1015 (VMOVAPSrm addr:$src)>;
1016 def : Pat<(loadv2i64 addr:$src),
1017 (VMOVUPSrm addr:$src)>;
1019 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
1020 (VMOVAPSmr addr:$dst, VR128:$src)>;
1021 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
1022 (VMOVAPSmr addr:$dst, VR128:$src)>;
1023 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
1024 (VMOVAPSmr addr:$dst, VR128:$src)>;
1025 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
1026 (VMOVAPSmr addr:$dst, VR128:$src)>;
1027 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
1028 (VMOVUPSmr addr:$dst, VR128:$src)>;
1029 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
1030 (VMOVUPSmr addr:$dst, VR128:$src)>;
1031 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
1032 (VMOVUPSmr addr:$dst, VR128:$src)>;
1033 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
1034 (VMOVUPSmr addr:$dst, VR128:$src)>;
1036 // 256-bit load/store
1037 def : Pat<(alignedloadv4i64 addr:$src),
1038 (VMOVAPSYrm addr:$src)>;
1039 def : Pat<(loadv4i64 addr:$src),
1040 (VMOVUPSYrm addr:$src)>;
1041 def : Pat<(alignedstore256 (v4i64 VR256:$src), addr:$dst),
1042 (VMOVAPSYmr addr:$dst, VR256:$src)>;
1043 def : Pat<(alignedstore256 (v8i32 VR256:$src), addr:$dst),
1044 (VMOVAPSYmr addr:$dst, VR256:$src)>;
1045 def : Pat<(alignedstore256 (v16i16 VR256:$src), addr:$dst),
1046 (VMOVAPSYmr addr:$dst, VR256:$src)>;
1047 def : Pat<(alignedstore256 (v32i8 VR256:$src), addr:$dst),
1048 (VMOVAPSYmr addr:$dst, VR256:$src)>;
1049 def : Pat<(store (v4i64 VR256:$src), addr:$dst),
1050 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1051 def : Pat<(store (v8i32 VR256:$src), addr:$dst),
1052 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1053 def : Pat<(store (v16i16 VR256:$src), addr:$dst),
1054 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1055 def : Pat<(store (v32i8 VR256:$src), addr:$dst),
1056 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1058 // Special patterns for storing subvector extracts of lower 128-bits
1059 // Its cheaper to just use VMOVAPS/VMOVUPS instead of VEXTRACTF128mr
1060 def : Pat<(alignedstore (v2f64 (extract_subvector
1061 (v4f64 VR256:$src), (iPTR 0))), addr:$dst),
1062 (VMOVAPDmr addr:$dst, (v2f64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1063 def : Pat<(alignedstore (v4f32 (extract_subvector
1064 (v8f32 VR256:$src), (iPTR 0))), addr:$dst),
1065 (VMOVAPSmr addr:$dst, (v4f32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1066 def : Pat<(alignedstore (v2i64 (extract_subvector
1067 (v4i64 VR256:$src), (iPTR 0))), addr:$dst),
1068 (VMOVAPDmr addr:$dst, (v2i64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1069 def : Pat<(alignedstore (v4i32 (extract_subvector
1070 (v8i32 VR256:$src), (iPTR 0))), addr:$dst),
1071 (VMOVAPSmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1072 def : Pat<(alignedstore (v8i16 (extract_subvector
1073 (v16i16 VR256:$src), (iPTR 0))), addr:$dst),
1074 (VMOVAPSmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1075 def : Pat<(alignedstore (v16i8 (extract_subvector
1076 (v32i8 VR256:$src), (iPTR 0))), addr:$dst),
1077 (VMOVAPSmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1079 def : Pat<(store (v2f64 (extract_subvector
1080 (v4f64 VR256:$src), (iPTR 0))), addr:$dst),
1081 (VMOVUPDmr addr:$dst, (v2f64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1082 def : Pat<(store (v4f32 (extract_subvector
1083 (v8f32 VR256:$src), (iPTR 0))), addr:$dst),
1084 (VMOVUPSmr addr:$dst, (v4f32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1085 def : Pat<(store (v2i64 (extract_subvector
1086 (v4i64 VR256:$src), (iPTR 0))), addr:$dst),
1087 (VMOVUPDmr addr:$dst, (v2i64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1088 def : Pat<(store (v4i32 (extract_subvector
1089 (v8i32 VR256:$src), (iPTR 0))), addr:$dst),
1090 (VMOVUPSmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1091 def : Pat<(store (v8i16 (extract_subvector
1092 (v16i16 VR256:$src), (iPTR 0))), addr:$dst),
1093 (VMOVUPSmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1094 def : Pat<(store (v16i8 (extract_subvector
1095 (v32i8 VR256:$src), (iPTR 0))), addr:$dst),
1096 (VMOVUPSmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1099 // Use movaps / movups for SSE integer load / store (one byte shorter).
1100 // The instructions selected below are then converted to MOVDQA/MOVDQU
1101 // during the SSE domain pass.
1102 let Predicates = [UseSSE1] in {
1103 def : Pat<(alignedloadv2i64 addr:$src),
1104 (MOVAPSrm addr:$src)>;
1105 def : Pat<(loadv2i64 addr:$src),
1106 (MOVUPSrm addr:$src)>;
1108 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
1109 (MOVAPSmr addr:$dst, VR128:$src)>;
1110 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
1111 (MOVAPSmr addr:$dst, VR128:$src)>;
1112 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
1113 (MOVAPSmr addr:$dst, VR128:$src)>;
1114 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
1115 (MOVAPSmr addr:$dst, VR128:$src)>;
1116 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
1117 (MOVUPSmr addr:$dst, VR128:$src)>;
1118 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
1119 (MOVUPSmr addr:$dst, VR128:$src)>;
1120 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
1121 (MOVUPSmr addr:$dst, VR128:$src)>;
1122 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
1123 (MOVUPSmr addr:$dst, VR128:$src)>;
1126 // Alias instruction to load FR32 or FR64 from f128mem using movaps. Upper
1127 // bits are disregarded. FIXME: Set encoding to pseudo!
1128 let canFoldAsLoad = 1, isReMaterializable = 1, SchedRW = [WriteLoad] in {
1129 let isCodeGenOnly = 1 in {
1130 def FsVMOVAPSrm : VPSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
1131 "movaps\t{$src, $dst|$dst, $src}",
1132 [(set FR32:$dst, (alignedloadfsf32 addr:$src))],
1133 IIC_SSE_MOVA_P_RM>, VEX;
1134 def FsVMOVAPDrm : VPDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1135 "movapd\t{$src, $dst|$dst, $src}",
1136 [(set FR64:$dst, (alignedloadfsf64 addr:$src))],
1137 IIC_SSE_MOVA_P_RM>, VEX;
1138 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
1139 "movaps\t{$src, $dst|$dst, $src}",
1140 [(set FR32:$dst, (alignedloadfsf32 addr:$src))],
1142 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1143 "movapd\t{$src, $dst|$dst, $src}",
1144 [(set FR64:$dst, (alignedloadfsf64 addr:$src))],
1149 //===----------------------------------------------------------------------===//
1150 // SSE 1 & 2 - Move Low packed FP Instructions
1151 //===----------------------------------------------------------------------===//
1153 multiclass sse12_mov_hilo_packed_base<bits<8>opc, SDNode psnode, SDNode pdnode,
1154 string base_opc, string asm_opr,
1155 InstrItinClass itin> {
1156 def PSrm : PI<opc, MRMSrcMem,
1157 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1158 !strconcat(base_opc, "s", asm_opr),
1160 (psnode VR128:$src1,
1161 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))],
1162 itin, SSEPackedSingle>, PS,
1163 Sched<[WriteFShuffleLd, ReadAfterLd]>;
1165 def PDrm : PI<opc, MRMSrcMem,
1166 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1167 !strconcat(base_opc, "d", asm_opr),
1168 [(set VR128:$dst, (v2f64 (pdnode VR128:$src1,
1169 (scalar_to_vector (loadf64 addr:$src2)))))],
1170 itin, SSEPackedDouble>, PD,
1171 Sched<[WriteFShuffleLd, ReadAfterLd]>;
1175 multiclass sse12_mov_hilo_packed<bits<8>opc, SDNode psnode, SDNode pdnode,
1176 string base_opc, InstrItinClass itin> {
1177 defm V#NAME : sse12_mov_hilo_packed_base<opc, psnode, pdnode, base_opc,
1178 "\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1181 let Constraints = "$src1 = $dst" in
1182 defm NAME : sse12_mov_hilo_packed_base<opc, psnode, pdnode, base_opc,
1183 "\t{$src2, $dst|$dst, $src2}",
1187 let AddedComplexity = 20 in {
1188 defm MOVL : sse12_mov_hilo_packed<0x12, X86Movlps, X86Movlpd, "movlp",
1192 let SchedRW = [WriteStore] in {
1193 def VMOVLPSmr : VPSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1194 "movlps\t{$src, $dst|$dst, $src}",
1195 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
1196 (iPTR 0))), addr:$dst)],
1197 IIC_SSE_MOV_LH>, VEX;
1198 def VMOVLPDmr : VPDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1199 "movlpd\t{$src, $dst|$dst, $src}",
1200 [(store (f64 (vector_extract (v2f64 VR128:$src),
1201 (iPTR 0))), addr:$dst)],
1202 IIC_SSE_MOV_LH>, VEX;
1203 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1204 "movlps\t{$src, $dst|$dst, $src}",
1205 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
1206 (iPTR 0))), addr:$dst)],
1208 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1209 "movlpd\t{$src, $dst|$dst, $src}",
1210 [(store (f64 (vector_extract (v2f64 VR128:$src),
1211 (iPTR 0))), addr:$dst)],
1215 let Predicates = [HasAVX] in {
1216 // Shuffle with VMOVLPS
1217 def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))),
1218 (VMOVLPSrm VR128:$src1, addr:$src2)>;
1219 def : Pat<(v4i32 (X86Movlps VR128:$src1, (load addr:$src2))),
1220 (VMOVLPSrm VR128:$src1, addr:$src2)>;
1222 // Shuffle with VMOVLPD
1223 def : Pat<(v2f64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1224 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1225 def : Pat<(v2i64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1226 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1227 def : Pat<(v2f64 (X86Movsd VR128:$src1,
1228 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))),
1229 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1232 def : Pat<(store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)),
1234 (VMOVLPSmr addr:$src1, VR128:$src2)>;
1235 def : Pat<(store (v4i32 (X86Movlps
1236 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)), addr:$src1),
1237 (VMOVLPSmr addr:$src1, VR128:$src2)>;
1238 def : Pat<(store (v2f64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1240 (VMOVLPDmr addr:$src1, VR128:$src2)>;
1241 def : Pat<(store (v2i64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1243 (VMOVLPDmr addr:$src1, VR128:$src2)>;
1246 let Predicates = [UseSSE1] in {
1247 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
1248 def : Pat<(store (i64 (vector_extract (bc_v2i64 (v4f32 VR128:$src2)),
1249 (iPTR 0))), addr:$src1),
1250 (MOVLPSmr addr:$src1, VR128:$src2)>;
1252 // Shuffle with MOVLPS
1253 def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))),
1254 (MOVLPSrm VR128:$src1, addr:$src2)>;
1255 def : Pat<(v4i32 (X86Movlps VR128:$src1, (load addr:$src2))),
1256 (MOVLPSrm VR128:$src1, addr:$src2)>;
1257 def : Pat<(X86Movlps VR128:$src1,
1258 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1259 (MOVLPSrm VR128:$src1, addr:$src2)>;
1262 def : Pat<(store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)),
1264 (MOVLPSmr addr:$src1, VR128:$src2)>;
1265 def : Pat<(store (v4i32 (X86Movlps
1266 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
1268 (MOVLPSmr addr:$src1, VR128:$src2)>;
1271 let Predicates = [UseSSE2] in {
1272 // Shuffle with MOVLPD
1273 def : Pat<(v2f64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1274 (MOVLPDrm VR128:$src1, addr:$src2)>;
1275 def : Pat<(v2i64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1276 (MOVLPDrm VR128:$src1, addr:$src2)>;
1277 def : Pat<(v2f64 (X86Movsd VR128:$src1,
1278 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))),
1279 (MOVLPDrm VR128:$src1, addr:$src2)>;
1282 def : Pat<(store (v2f64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1284 (MOVLPDmr addr:$src1, VR128:$src2)>;
1285 def : Pat<(store (v2i64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1287 (MOVLPDmr addr:$src1, VR128:$src2)>;
1290 //===----------------------------------------------------------------------===//
1291 // SSE 1 & 2 - Move Hi packed FP Instructions
1292 //===----------------------------------------------------------------------===//
1294 let AddedComplexity = 20 in {
1295 defm MOVH : sse12_mov_hilo_packed<0x16, X86Movlhps, X86Movlhpd, "movhp",
1299 let SchedRW = [WriteStore] in {
1300 // v2f64 extract element 1 is always custom lowered to unpack high to low
1301 // and extract element 0 so the non-store version isn't too horrible.
1302 def VMOVHPSmr : VPSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1303 "movhps\t{$src, $dst|$dst, $src}",
1304 [(store (f64 (vector_extract
1305 (X86Unpckh (bc_v2f64 (v4f32 VR128:$src)),
1306 (bc_v2f64 (v4f32 VR128:$src))),
1307 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>, VEX;
1308 def VMOVHPDmr : VPDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1309 "movhpd\t{$src, $dst|$dst, $src}",
1310 [(store (f64 (vector_extract
1311 (v2f64 (X86Unpckh VR128:$src, VR128:$src)),
1312 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>, VEX;
1313 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1314 "movhps\t{$src, $dst|$dst, $src}",
1315 [(store (f64 (vector_extract
1316 (X86Unpckh (bc_v2f64 (v4f32 VR128:$src)),
1317 (bc_v2f64 (v4f32 VR128:$src))),
1318 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>;
1319 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1320 "movhpd\t{$src, $dst|$dst, $src}",
1321 [(store (f64 (vector_extract
1322 (v2f64 (X86Unpckh VR128:$src, VR128:$src)),
1323 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>;
1326 let Predicates = [HasAVX] in {
1328 def : Pat<(X86Movlhps VR128:$src1,
1329 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1330 (VMOVHPSrm VR128:$src1, addr:$src2)>;
1331 def : Pat<(X86Movlhps VR128:$src1,
1332 (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
1333 (VMOVHPSrm VR128:$src1, addr:$src2)>;
1335 // FIXME: Instead of X86Unpckl, there should be a X86Movlhpd here, the problem
1336 // is during lowering, where it's not possible to recognize the load fold
1337 // cause it has two uses through a bitcast. One use disappears at isel time
1338 // and the fold opportunity reappears.
1339 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1340 (scalar_to_vector (loadf64 addr:$src2)))),
1341 (VMOVHPDrm VR128:$src1, addr:$src2)>;
1342 // Also handle an i64 load because that may get selected as a faster way to
1344 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1345 (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src2)))))),
1346 (VMOVHPDrm VR128:$src1, addr:$src2)>;
1349 let Predicates = [UseSSE1] in {
1351 def : Pat<(X86Movlhps VR128:$src1,
1352 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1353 (MOVHPSrm VR128:$src1, addr:$src2)>;
1354 def : Pat<(X86Movlhps VR128:$src1,
1355 (bc_v4f32 (v2i64 (X86vzload addr:$src2)))),
1356 (MOVHPSrm VR128:$src1, addr:$src2)>;
1359 let Predicates = [UseSSE2] in {
1360 // FIXME: Instead of X86Unpckl, there should be a X86Movlhpd here, the problem
1361 // is during lowering, where it's not possible to recognize the load fold
1362 // cause it has two uses through a bitcast. One use disappears at isel time
1363 // and the fold opportunity reappears.
1364 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1365 (scalar_to_vector (loadf64 addr:$src2)))),
1366 (MOVHPDrm VR128:$src1, addr:$src2)>;
1367 // Also handle an i64 load because that may get selected as a faster way to
1369 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1370 (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src2)))))),
1371 (MOVHPDrm VR128:$src1, addr:$src2)>;
1374 //===----------------------------------------------------------------------===//
1375 // SSE 1 & 2 - Move Low to High and High to Low packed FP Instructions
1376 //===----------------------------------------------------------------------===//
1378 let AddedComplexity = 20, Predicates = [UseAVX] in {
1379 def VMOVLHPSrr : VPSI<0x16, MRMSrcReg, (outs VR128:$dst),
1380 (ins VR128:$src1, VR128:$src2),
1381 "movlhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1383 (v4f32 (X86Movlhps VR128:$src1, VR128:$src2)))],
1385 VEX_4V, Sched<[WriteFShuffle]>;
1386 def VMOVHLPSrr : VPSI<0x12, MRMSrcReg, (outs VR128:$dst),
1387 (ins VR128:$src1, VR128:$src2),
1388 "movhlps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1390 (v4f32 (X86Movhlps VR128:$src1, VR128:$src2)))],
1392 VEX_4V, Sched<[WriteFShuffle]>;
1394 let Constraints = "$src1 = $dst", AddedComplexity = 20 in {
1395 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
1396 (ins VR128:$src1, VR128:$src2),
1397 "movlhps\t{$src2, $dst|$dst, $src2}",
1399 (v4f32 (X86Movlhps VR128:$src1, VR128:$src2)))],
1400 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
1401 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
1402 (ins VR128:$src1, VR128:$src2),
1403 "movhlps\t{$src2, $dst|$dst, $src2}",
1405 (v4f32 (X86Movhlps VR128:$src1, VR128:$src2)))],
1406 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
1409 let Predicates = [UseAVX] in {
1411 def : Pat<(v4i32 (X86Movlhps VR128:$src1, VR128:$src2)),
1412 (VMOVLHPSrr VR128:$src1, VR128:$src2)>;
1413 def : Pat<(v2i64 (X86Movlhps VR128:$src1, VR128:$src2)),
1414 (VMOVLHPSrr (v2i64 VR128:$src1), VR128:$src2)>;
1417 def : Pat<(v4i32 (X86Movhlps VR128:$src1, VR128:$src2)),
1418 (VMOVHLPSrr VR128:$src1, VR128:$src2)>;
1421 let Predicates = [UseSSE1] in {
1423 def : Pat<(v4i32 (X86Movlhps VR128:$src1, VR128:$src2)),
1424 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
1425 def : Pat<(v2i64 (X86Movlhps VR128:$src1, VR128:$src2)),
1426 (MOVLHPSrr (v2i64 VR128:$src1), VR128:$src2)>;
1429 def : Pat<(v4i32 (X86Movhlps VR128:$src1, VR128:$src2)),
1430 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
1433 //===----------------------------------------------------------------------===//
1434 // SSE 1 & 2 - Conversion Instructions
1435 //===----------------------------------------------------------------------===//
1437 def SSE_CVT_PD : OpndItins<
1438 IIC_SSE_CVT_PD_RR, IIC_SSE_CVT_PD_RM
1441 let Sched = WriteCvtI2F in
1442 def SSE_CVT_PS : OpndItins<
1443 IIC_SSE_CVT_PS_RR, IIC_SSE_CVT_PS_RM
1446 let Sched = WriteCvtI2F in
1447 def SSE_CVT_Scalar : OpndItins<
1448 IIC_SSE_CVT_Scalar_RR, IIC_SSE_CVT_Scalar_RM
1451 let Sched = WriteCvtF2I in
1452 def SSE_CVT_SS2SI_32 : OpndItins<
1453 IIC_SSE_CVT_SS2SI32_RR, IIC_SSE_CVT_SS2SI32_RM
1456 let Sched = WriteCvtF2I in
1457 def SSE_CVT_SS2SI_64 : OpndItins<
1458 IIC_SSE_CVT_SS2SI64_RR, IIC_SSE_CVT_SS2SI64_RM
1461 let Sched = WriteCvtF2I in
1462 def SSE_CVT_SD2SI : OpndItins<
1463 IIC_SSE_CVT_SD2SI_RR, IIC_SSE_CVT_SD2SI_RM
1466 multiclass sse12_cvt_s<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1467 SDNode OpNode, X86MemOperand x86memop, PatFrag ld_frag,
1468 string asm, OpndItins itins> {
1469 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
1470 [(set DstRC:$dst, (OpNode SrcRC:$src))],
1471 itins.rr>, Sched<[itins.Sched]>;
1472 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
1473 [(set DstRC:$dst, (OpNode (ld_frag addr:$src)))],
1474 itins.rm>, Sched<[itins.Sched.Folded]>;
1477 multiclass sse12_cvt_p<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1478 X86MemOperand x86memop, string asm, Domain d,
1480 let neverHasSideEffects = 1 in {
1481 def rr : I<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
1482 [], itins.rr, d>, Sched<[itins.Sched]>;
1484 def rm : I<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
1485 [], itins.rm, d>, Sched<[itins.Sched.Folded]>;
1489 multiclass sse12_vcvt_avx<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1490 X86MemOperand x86memop, string asm> {
1491 let neverHasSideEffects = 1, Predicates = [UseAVX] in {
1492 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src),
1493 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>,
1494 Sched<[WriteCvtI2F]>;
1496 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
1497 (ins DstRC:$src1, x86memop:$src),
1498 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>,
1499 Sched<[WriteCvtI2FLd, ReadAfterLd]>;
1500 } // neverHasSideEffects = 1
1503 let Predicates = [UseAVX] in {
1504 defm VCVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
1505 "cvttss2si\t{$src, $dst|$dst, $src}",
1508 defm VCVTTSS2SI64 : sse12_cvt_s<0x2C, FR32, GR64, fp_to_sint, f32mem, loadf32,
1509 "cvttss2si\t{$src, $dst|$dst, $src}",
1511 XS, VEX, VEX_W, VEX_LIG;
1512 defm VCVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
1513 "cvttsd2si\t{$src, $dst|$dst, $src}",
1516 defm VCVTTSD2SI64 : sse12_cvt_s<0x2C, FR64, GR64, fp_to_sint, f64mem, loadf64,
1517 "cvttsd2si\t{$src, $dst|$dst, $src}",
1519 XD, VEX, VEX_W, VEX_LIG;
1521 def : InstAlias<"vcvttss2si{l}\t{$src, $dst|$dst, $src}",
1522 (VCVTTSS2SIrr GR32:$dst, FR32:$src), 0>;
1523 def : InstAlias<"vcvttss2si{l}\t{$src, $dst|$dst, $src}",
1524 (VCVTTSS2SIrm GR32:$dst, f32mem:$src), 0>;
1525 def : InstAlias<"vcvttsd2si{l}\t{$src, $dst|$dst, $src}",
1526 (VCVTTSD2SIrr GR32:$dst, FR64:$src), 0>;
1527 def : InstAlias<"vcvttsd2si{l}\t{$src, $dst|$dst, $src}",
1528 (VCVTTSD2SIrm GR32:$dst, f64mem:$src), 0>;
1529 def : InstAlias<"vcvttss2si{q}\t{$src, $dst|$dst, $src}",
1530 (VCVTTSS2SI64rr GR64:$dst, FR32:$src), 0>;
1531 def : InstAlias<"vcvttss2si{q}\t{$src, $dst|$dst, $src}",
1532 (VCVTTSS2SI64rm GR64:$dst, f32mem:$src), 0>;
1533 def : InstAlias<"vcvttsd2si{q}\t{$src, $dst|$dst, $src}",
1534 (VCVTTSD2SI64rr GR64:$dst, FR64:$src), 0>;
1535 def : InstAlias<"vcvttsd2si{q}\t{$src, $dst|$dst, $src}",
1536 (VCVTTSD2SI64rm GR64:$dst, f64mem:$src), 0>;
1538 // The assembler can recognize rr 64-bit instructions by seeing a rxx
1539 // register, but the same isn't true when only using memory operands,
1540 // provide other assembly "l" and "q" forms to address this explicitly
1541 // where appropriate to do so.
1542 defm VCVTSI2SS : sse12_vcvt_avx<0x2A, GR32, FR32, i32mem, "cvtsi2ss{l}">,
1543 XS, VEX_4V, VEX_LIG;
1544 defm VCVTSI2SS64 : sse12_vcvt_avx<0x2A, GR64, FR32, i64mem, "cvtsi2ss{q}">,
1545 XS, VEX_4V, VEX_W, VEX_LIG;
1546 defm VCVTSI2SD : sse12_vcvt_avx<0x2A, GR32, FR64, i32mem, "cvtsi2sd{l}">,
1547 XD, VEX_4V, VEX_LIG;
1548 defm VCVTSI2SD64 : sse12_vcvt_avx<0x2A, GR64, FR64, i64mem, "cvtsi2sd{q}">,
1549 XD, VEX_4V, VEX_W, VEX_LIG;
1551 let Predicates = [UseAVX] in {
1552 def : InstAlias<"vcvtsi2ss\t{$src, $src1, $dst|$dst, $src1, $src}",
1553 (VCVTSI2SSrm FR64:$dst, FR64:$src1, i32mem:$src), 0>;
1554 def : InstAlias<"vcvtsi2sd\t{$src, $src1, $dst|$dst, $src1, $src}",
1555 (VCVTSI2SDrm FR64:$dst, FR64:$src1, i32mem:$src), 0>;
1557 def : Pat<(f32 (sint_to_fp (loadi32 addr:$src))),
1558 (VCVTSI2SSrm (f32 (IMPLICIT_DEF)), addr:$src)>;
1559 def : Pat<(f32 (sint_to_fp (loadi64 addr:$src))),
1560 (VCVTSI2SS64rm (f32 (IMPLICIT_DEF)), addr:$src)>;
1561 def : Pat<(f64 (sint_to_fp (loadi32 addr:$src))),
1562 (VCVTSI2SDrm (f64 (IMPLICIT_DEF)), addr:$src)>;
1563 def : Pat<(f64 (sint_to_fp (loadi64 addr:$src))),
1564 (VCVTSI2SD64rm (f64 (IMPLICIT_DEF)), addr:$src)>;
1566 def : Pat<(f32 (sint_to_fp GR32:$src)),
1567 (VCVTSI2SSrr (f32 (IMPLICIT_DEF)), GR32:$src)>;
1568 def : Pat<(f32 (sint_to_fp GR64:$src)),
1569 (VCVTSI2SS64rr (f32 (IMPLICIT_DEF)), GR64:$src)>;
1570 def : Pat<(f64 (sint_to_fp GR32:$src)),
1571 (VCVTSI2SDrr (f64 (IMPLICIT_DEF)), GR32:$src)>;
1572 def : Pat<(f64 (sint_to_fp GR64:$src)),
1573 (VCVTSI2SD64rr (f64 (IMPLICIT_DEF)), GR64:$src)>;
1576 defm CVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
1577 "cvttss2si\t{$src, $dst|$dst, $src}",
1578 SSE_CVT_SS2SI_32>, XS;
1579 defm CVTTSS2SI64 : sse12_cvt_s<0x2C, FR32, GR64, fp_to_sint, f32mem, loadf32,
1580 "cvttss2si\t{$src, $dst|$dst, $src}",
1581 SSE_CVT_SS2SI_64>, XS, REX_W;
1582 defm CVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
1583 "cvttsd2si\t{$src, $dst|$dst, $src}",
1585 defm CVTTSD2SI64 : sse12_cvt_s<0x2C, FR64, GR64, fp_to_sint, f64mem, loadf64,
1586 "cvttsd2si\t{$src, $dst|$dst, $src}",
1587 SSE_CVT_SD2SI>, XD, REX_W;
1588 defm CVTSI2SS : sse12_cvt_s<0x2A, GR32, FR32, sint_to_fp, i32mem, loadi32,
1589 "cvtsi2ss{l}\t{$src, $dst|$dst, $src}",
1590 SSE_CVT_Scalar>, XS;
1591 defm CVTSI2SS64 : sse12_cvt_s<0x2A, GR64, FR32, sint_to_fp, i64mem, loadi64,
1592 "cvtsi2ss{q}\t{$src, $dst|$dst, $src}",
1593 SSE_CVT_Scalar>, XS, REX_W;
1594 defm CVTSI2SD : sse12_cvt_s<0x2A, GR32, FR64, sint_to_fp, i32mem, loadi32,
1595 "cvtsi2sd{l}\t{$src, $dst|$dst, $src}",
1596 SSE_CVT_Scalar>, XD;
1597 defm CVTSI2SD64 : sse12_cvt_s<0x2A, GR64, FR64, sint_to_fp, i64mem, loadi64,
1598 "cvtsi2sd{q}\t{$src, $dst|$dst, $src}",
1599 SSE_CVT_Scalar>, XD, REX_W;
1601 def : InstAlias<"cvttss2si{l}\t{$src, $dst|$dst, $src}",
1602 (CVTTSS2SIrr GR32:$dst, FR32:$src), 0>;
1603 def : InstAlias<"cvttss2si{l}\t{$src, $dst|$dst, $src}",
1604 (CVTTSS2SIrm GR32:$dst, f32mem:$src), 0>;
1605 def : InstAlias<"cvttsd2si{l}\t{$src, $dst|$dst, $src}",
1606 (CVTTSD2SIrr GR32:$dst, FR64:$src), 0>;
1607 def : InstAlias<"cvttsd2si{l}\t{$src, $dst|$dst, $src}",
1608 (CVTTSD2SIrm GR32:$dst, f64mem:$src), 0>;
1609 def : InstAlias<"cvttss2si{q}\t{$src, $dst|$dst, $src}",
1610 (CVTTSS2SI64rr GR64:$dst, FR32:$src), 0>;
1611 def : InstAlias<"cvttss2si{q}\t{$src, $dst|$dst, $src}",
1612 (CVTTSS2SI64rm GR64:$dst, f32mem:$src), 0>;
1613 def : InstAlias<"cvttsd2si{q}\t{$src, $dst|$dst, $src}",
1614 (CVTTSD2SI64rr GR64:$dst, FR64:$src), 0>;
1615 def : InstAlias<"cvttsd2si{q}\t{$src, $dst|$dst, $src}",
1616 (CVTTSD2SI64rm GR64:$dst, f64mem:$src), 0>;
1618 def : InstAlias<"cvtsi2ss\t{$src, $dst|$dst, $src}",
1619 (CVTSI2SSrm FR64:$dst, i32mem:$src), 0>;
1620 def : InstAlias<"cvtsi2sd\t{$src, $dst|$dst, $src}",
1621 (CVTSI2SDrm FR64:$dst, i32mem:$src), 0>;
1623 // Conversion Instructions Intrinsics - Match intrinsics which expect MM
1624 // and/or XMM operand(s).
1626 multiclass sse12_cvt_sint<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1627 Intrinsic Int, Operand memop, ComplexPattern mem_cpat,
1628 string asm, OpndItins itins> {
1629 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src),
1630 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1631 [(set DstRC:$dst, (Int SrcRC:$src))], itins.rr>,
1632 Sched<[itins.Sched]>;
1633 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins memop:$src),
1634 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1635 [(set DstRC:$dst, (Int mem_cpat:$src))], itins.rm>,
1636 Sched<[itins.Sched.Folded]>;
1639 multiclass sse12_cvt_sint_3addr<bits<8> opc, RegisterClass SrcRC,
1640 RegisterClass DstRC, Intrinsic Int, X86MemOperand x86memop,
1641 PatFrag ld_frag, string asm, OpndItins itins,
1643 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src2),
1645 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
1646 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
1647 [(set DstRC:$dst, (Int DstRC:$src1, SrcRC:$src2))],
1648 itins.rr>, Sched<[itins.Sched]>;
1649 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
1650 (ins DstRC:$src1, x86memop:$src2),
1652 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
1653 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
1654 [(set DstRC:$dst, (Int DstRC:$src1, (ld_frag addr:$src2)))],
1655 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
1658 let Predicates = [UseAVX] in {
1659 defm VCVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32,
1660 int_x86_sse2_cvtsd2si, sdmem, sse_load_f64, "cvtsd2si",
1661 SSE_CVT_SD2SI>, XD, VEX, VEX_LIG;
1662 defm VCVTSD2SI64 : sse12_cvt_sint<0x2D, VR128, GR64,
1663 int_x86_sse2_cvtsd2si64, sdmem, sse_load_f64, "cvtsd2si",
1664 SSE_CVT_SD2SI>, XD, VEX, VEX_W, VEX_LIG;
1666 defm CVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse2_cvtsd2si,
1667 sdmem, sse_load_f64, "cvtsd2si", SSE_CVT_SD2SI>, XD;
1668 defm CVTSD2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse2_cvtsd2si64,
1669 sdmem, sse_load_f64, "cvtsd2si", SSE_CVT_SD2SI>, XD, REX_W;
1672 let isCodeGenOnly = 1 in {
1673 let Predicates = [UseAVX] in {
1674 defm Int_VCVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1675 int_x86_sse_cvtsi2ss, i32mem, loadi32, "cvtsi2ss{l}",
1676 SSE_CVT_Scalar, 0>, XS, VEX_4V;
1677 defm Int_VCVTSI2SS64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1678 int_x86_sse_cvtsi642ss, i64mem, loadi64, "cvtsi2ss{q}",
1679 SSE_CVT_Scalar, 0>, XS, VEX_4V,
1681 defm Int_VCVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1682 int_x86_sse2_cvtsi2sd, i32mem, loadi32, "cvtsi2sd{l}",
1683 SSE_CVT_Scalar, 0>, XD, VEX_4V;
1684 defm Int_VCVTSI2SD64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1685 int_x86_sse2_cvtsi642sd, i64mem, loadi64, "cvtsi2sd{q}",
1686 SSE_CVT_Scalar, 0>, XD,
1689 let Constraints = "$src1 = $dst" in {
1690 defm Int_CVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1691 int_x86_sse_cvtsi2ss, i32mem, loadi32,
1692 "cvtsi2ss{l}", SSE_CVT_Scalar>, XS;
1693 defm Int_CVTSI2SS64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1694 int_x86_sse_cvtsi642ss, i64mem, loadi64,
1695 "cvtsi2ss{q}", SSE_CVT_Scalar>, XS, REX_W;
1696 defm Int_CVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1697 int_x86_sse2_cvtsi2sd, i32mem, loadi32,
1698 "cvtsi2sd{l}", SSE_CVT_Scalar>, XD;
1699 defm Int_CVTSI2SD64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1700 int_x86_sse2_cvtsi642sd, i64mem, loadi64,
1701 "cvtsi2sd{q}", SSE_CVT_Scalar>, XD, REX_W;
1703 } // isCodeGenOnly = 1
1707 // Aliases for intrinsics
1708 let isCodeGenOnly = 1 in {
1709 let Predicates = [UseAVX] in {
1710 defm Int_VCVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
1711 ssmem, sse_load_f32, "cvttss2si",
1712 SSE_CVT_SS2SI_32>, XS, VEX;
1713 defm Int_VCVTTSS2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1714 int_x86_sse_cvttss2si64, ssmem, sse_load_f32,
1715 "cvttss2si", SSE_CVT_SS2SI_64>,
1717 defm Int_VCVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
1718 sdmem, sse_load_f64, "cvttsd2si",
1719 SSE_CVT_SD2SI>, XD, VEX;
1720 defm Int_VCVTTSD2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1721 int_x86_sse2_cvttsd2si64, sdmem, sse_load_f64,
1722 "cvttsd2si", SSE_CVT_SD2SI>,
1725 defm Int_CVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
1726 ssmem, sse_load_f32, "cvttss2si",
1727 SSE_CVT_SS2SI_32>, XS;
1728 defm Int_CVTTSS2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1729 int_x86_sse_cvttss2si64, ssmem, sse_load_f32,
1730 "cvttss2si", SSE_CVT_SS2SI_64>, XS, REX_W;
1731 defm Int_CVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
1732 sdmem, sse_load_f64, "cvttsd2si",
1734 defm Int_CVTTSD2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1735 int_x86_sse2_cvttsd2si64, sdmem, sse_load_f64,
1736 "cvttsd2si", SSE_CVT_SD2SI>, XD, REX_W;
1737 } // isCodeGenOnly = 1
1739 let Predicates = [UseAVX] in {
1740 defm VCVTSS2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse_cvtss2si,
1741 ssmem, sse_load_f32, "cvtss2si",
1742 SSE_CVT_SS2SI_32>, XS, VEX, VEX_LIG;
1743 defm VCVTSS2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse_cvtss2si64,
1744 ssmem, sse_load_f32, "cvtss2si",
1745 SSE_CVT_SS2SI_64>, XS, VEX, VEX_W, VEX_LIG;
1747 defm CVTSS2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse_cvtss2si,
1748 ssmem, sse_load_f32, "cvtss2si",
1749 SSE_CVT_SS2SI_32>, XS;
1750 defm CVTSS2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse_cvtss2si64,
1751 ssmem, sse_load_f32, "cvtss2si",
1752 SSE_CVT_SS2SI_64>, XS, REX_W;
1754 defm VCVTDQ2PS : sse12_cvt_p<0x5B, VR128, VR128, i128mem,
1755 "vcvtdq2ps\t{$src, $dst|$dst, $src}",
1756 SSEPackedSingle, SSE_CVT_PS>,
1757 PS, VEX, Requires<[HasAVX]>;
1758 defm VCVTDQ2PSY : sse12_cvt_p<0x5B, VR256, VR256, i256mem,
1759 "vcvtdq2ps\t{$src, $dst|$dst, $src}",
1760 SSEPackedSingle, SSE_CVT_PS>,
1761 PS, VEX, VEX_L, Requires<[HasAVX]>;
1763 defm CVTDQ2PS : sse12_cvt_p<0x5B, VR128, VR128, i128mem,
1764 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1765 SSEPackedSingle, SSE_CVT_PS>,
1766 PS, Requires<[UseSSE2]>;
1768 let Predicates = [UseAVX] in {
1769 def : InstAlias<"vcvtss2si{l}\t{$src, $dst|$dst, $src}",
1770 (VCVTSS2SIrr GR32:$dst, VR128:$src), 0>;
1771 def : InstAlias<"vcvtss2si{l}\t{$src, $dst|$dst, $src}",
1772 (VCVTSS2SIrm GR32:$dst, ssmem:$src), 0>;
1773 def : InstAlias<"vcvtsd2si{l}\t{$src, $dst|$dst, $src}",
1774 (VCVTSD2SIrr GR32:$dst, VR128:$src), 0>;
1775 def : InstAlias<"vcvtsd2si{l}\t{$src, $dst|$dst, $src}",
1776 (VCVTSD2SIrm GR32:$dst, sdmem:$src), 0>;
1777 def : InstAlias<"vcvtss2si{q}\t{$src, $dst|$dst, $src}",
1778 (VCVTSS2SI64rr GR64:$dst, VR128:$src), 0>;
1779 def : InstAlias<"vcvtss2si{q}\t{$src, $dst|$dst, $src}",
1780 (VCVTSS2SI64rm GR64:$dst, ssmem:$src), 0>;
1781 def : InstAlias<"vcvtsd2si{q}\t{$src, $dst|$dst, $src}",
1782 (VCVTSD2SI64rr GR64:$dst, VR128:$src), 0>;
1783 def : InstAlias<"vcvtsd2si{q}\t{$src, $dst|$dst, $src}",
1784 (VCVTSD2SI64rm GR64:$dst, sdmem:$src), 0>;
1787 def : InstAlias<"cvtss2si{l}\t{$src, $dst|$dst, $src}",
1788 (CVTSS2SIrr GR32:$dst, VR128:$src), 0>;
1789 def : InstAlias<"cvtss2si{l}\t{$src, $dst|$dst, $src}",
1790 (CVTSS2SIrm GR32:$dst, ssmem:$src), 0>;
1791 def : InstAlias<"cvtsd2si{l}\t{$src, $dst|$dst, $src}",
1792 (CVTSD2SIrr GR32:$dst, VR128:$src), 0>;
1793 def : InstAlias<"cvtsd2si{l}\t{$src, $dst|$dst, $src}",
1794 (CVTSD2SIrm GR32:$dst, sdmem:$src), 0>;
1795 def : InstAlias<"cvtss2si{q}\t{$src, $dst|$dst, $src}",
1796 (CVTSS2SI64rr GR64:$dst, VR128:$src), 0>;
1797 def : InstAlias<"cvtss2si{q}\t{$src, $dst|$dst, $src}",
1798 (CVTSS2SI64rm GR64:$dst, ssmem:$src), 0>;
1799 def : InstAlias<"cvtsd2si{q}\t{$src, $dst|$dst, $src}",
1800 (CVTSD2SI64rr GR64:$dst, VR128:$src), 0>;
1801 def : InstAlias<"cvtsd2si{q}\t{$src, $dst|$dst, $src}",
1802 (CVTSD2SI64rm GR64:$dst, sdmem:$src)>;
1806 // Convert scalar double to scalar single
1807 let neverHasSideEffects = 1, Predicates = [UseAVX] in {
1808 def VCVTSD2SSrr : VSDI<0x5A, MRMSrcReg, (outs FR32:$dst),
1809 (ins FR64:$src1, FR64:$src2),
1810 "cvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}", [],
1811 IIC_SSE_CVT_Scalar_RR>, VEX_4V, VEX_LIG,
1812 Sched<[WriteCvtF2F]>;
1814 def VCVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst),
1815 (ins FR64:$src1, f64mem:$src2),
1816 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1817 [], IIC_SSE_CVT_Scalar_RM>,
1818 XD, Requires<[HasAVX, OptForSize]>, VEX_4V, VEX_LIG,
1819 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1822 def : Pat<(f32 (fround FR64:$src)), (VCVTSD2SSrr FR64:$src, FR64:$src)>,
1825 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1826 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1827 [(set FR32:$dst, (fround FR64:$src))],
1828 IIC_SSE_CVT_Scalar_RR>, Sched<[WriteCvtF2F]>;
1829 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1830 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1831 [(set FR32:$dst, (fround (loadf64 addr:$src)))],
1832 IIC_SSE_CVT_Scalar_RM>,
1834 Requires<[UseSSE2, OptForSize]>, Sched<[WriteCvtF2FLd]>;
1836 let isCodeGenOnly = 1 in {
1837 def Int_VCVTSD2SSrr: I<0x5A, MRMSrcReg,
1838 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1839 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1841 (int_x86_sse2_cvtsd2ss VR128:$src1, VR128:$src2))],
1842 IIC_SSE_CVT_Scalar_RR>, XD, VEX_4V, Requires<[UseAVX]>,
1843 Sched<[WriteCvtF2F]>;
1844 def Int_VCVTSD2SSrm: I<0x5A, MRMSrcReg,
1845 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1846 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1847 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss
1848 VR128:$src1, sse_load_f64:$src2))],
1849 IIC_SSE_CVT_Scalar_RM>, XD, VEX_4V, Requires<[UseAVX]>,
1850 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1852 let Constraints = "$src1 = $dst" in {
1853 def Int_CVTSD2SSrr: I<0x5A, MRMSrcReg,
1854 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1855 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1857 (int_x86_sse2_cvtsd2ss VR128:$src1, VR128:$src2))],
1858 IIC_SSE_CVT_Scalar_RR>, XD, Requires<[UseSSE2]>,
1859 Sched<[WriteCvtF2F]>;
1860 def Int_CVTSD2SSrm: I<0x5A, MRMSrcReg,
1861 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1862 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1863 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss
1864 VR128:$src1, sse_load_f64:$src2))],
1865 IIC_SSE_CVT_Scalar_RM>, XD, Requires<[UseSSE2]>,
1866 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1868 } // isCodeGenOnly = 1
1870 // Convert scalar single to scalar double
1871 // SSE2 instructions with XS prefix
1872 let neverHasSideEffects = 1, Predicates = [UseAVX] in {
1873 def VCVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst),
1874 (ins FR32:$src1, FR32:$src2),
1875 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1876 [], IIC_SSE_CVT_Scalar_RR>,
1877 XS, Requires<[HasAVX]>, VEX_4V, VEX_LIG,
1878 Sched<[WriteCvtF2F]>;
1880 def VCVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst),
1881 (ins FR32:$src1, f32mem:$src2),
1882 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1883 [], IIC_SSE_CVT_Scalar_RM>,
1884 XS, VEX_4V, VEX_LIG, Requires<[HasAVX, OptForSize]>,
1885 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1888 def : Pat<(f64 (fextend FR32:$src)),
1889 (VCVTSS2SDrr FR32:$src, FR32:$src)>, Requires<[UseAVX]>;
1890 def : Pat<(fextend (loadf32 addr:$src)),
1891 (VCVTSS2SDrm (f32 (IMPLICIT_DEF)), addr:$src)>, Requires<[UseAVX]>;
1893 def : Pat<(extloadf32 addr:$src),
1894 (VCVTSS2SDrm (f32 (IMPLICIT_DEF)), addr:$src)>,
1895 Requires<[UseAVX, OptForSize]>;
1896 def : Pat<(extloadf32 addr:$src),
1897 (VCVTSS2SDrr (f32 (IMPLICIT_DEF)), (VMOVSSrm addr:$src))>,
1898 Requires<[UseAVX, OptForSpeed]>;
1900 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1901 "cvtss2sd\t{$src, $dst|$dst, $src}",
1902 [(set FR64:$dst, (fextend FR32:$src))],
1903 IIC_SSE_CVT_Scalar_RR>, XS,
1904 Requires<[UseSSE2]>, Sched<[WriteCvtF2F]>;
1905 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1906 "cvtss2sd\t{$src, $dst|$dst, $src}",
1907 [(set FR64:$dst, (extloadf32 addr:$src))],
1908 IIC_SSE_CVT_Scalar_RM>, XS,
1909 Requires<[UseSSE2, OptForSize]>, Sched<[WriteCvtF2FLd]>;
1911 // extload f32 -> f64. This matches load+fextend because we have a hack in
1912 // the isel (PreprocessForFPConvert) that can introduce loads after dag
1914 // Since these loads aren't folded into the fextend, we have to match it
1916 def : Pat<(fextend (loadf32 addr:$src)),
1917 (CVTSS2SDrm addr:$src)>, Requires<[UseSSE2]>;
1918 def : Pat<(extloadf32 addr:$src),
1919 (CVTSS2SDrr (MOVSSrm addr:$src))>, Requires<[UseSSE2, OptForSpeed]>;
1921 let isCodeGenOnly = 1 in {
1922 def Int_VCVTSS2SDrr: I<0x5A, MRMSrcReg,
1923 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1924 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1926 (int_x86_sse2_cvtss2sd VR128:$src1, VR128:$src2))],
1927 IIC_SSE_CVT_Scalar_RR>, XS, VEX_4V, Requires<[UseAVX]>,
1928 Sched<[WriteCvtF2F]>;
1929 def Int_VCVTSS2SDrm: I<0x5A, MRMSrcMem,
1930 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
1931 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1933 (int_x86_sse2_cvtss2sd VR128:$src1, sse_load_f32:$src2))],
1934 IIC_SSE_CVT_Scalar_RM>, XS, VEX_4V, Requires<[UseAVX]>,
1935 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1936 let Constraints = "$src1 = $dst" in { // SSE2 instructions with XS prefix
1937 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1938 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1939 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1941 (int_x86_sse2_cvtss2sd VR128:$src1, VR128:$src2))],
1942 IIC_SSE_CVT_Scalar_RR>, XS, Requires<[UseSSE2]>,
1943 Sched<[WriteCvtF2F]>;
1944 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1945 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
1946 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1948 (int_x86_sse2_cvtss2sd VR128:$src1, sse_load_f32:$src2))],
1949 IIC_SSE_CVT_Scalar_RM>, XS, Requires<[UseSSE2]>,
1950 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1952 } // isCodeGenOnly = 1
1954 // Convert packed single/double fp to doubleword
1955 def VCVTPS2DQrr : VPDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1956 "cvtps2dq\t{$src, $dst|$dst, $src}",
1957 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))],
1958 IIC_SSE_CVT_PS_RR>, VEX, Sched<[WriteCvtF2I]>;
1959 def VCVTPS2DQrm : VPDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1960 "cvtps2dq\t{$src, $dst|$dst, $src}",
1962 (int_x86_sse2_cvtps2dq (loadv4f32 addr:$src)))],
1963 IIC_SSE_CVT_PS_RM>, VEX, Sched<[WriteCvtF2ILd]>;
1964 def VCVTPS2DQYrr : VPDI<0x5B, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
1965 "cvtps2dq\t{$src, $dst|$dst, $src}",
1967 (int_x86_avx_cvt_ps2dq_256 VR256:$src))],
1968 IIC_SSE_CVT_PS_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>;
1969 def VCVTPS2DQYrm : VPDI<0x5B, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
1970 "cvtps2dq\t{$src, $dst|$dst, $src}",
1972 (int_x86_avx_cvt_ps2dq_256 (loadv8f32 addr:$src)))],
1973 IIC_SSE_CVT_PS_RM>, VEX, VEX_L, Sched<[WriteCvtF2ILd]>;
1974 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1975 "cvtps2dq\t{$src, $dst|$dst, $src}",
1976 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))],
1977 IIC_SSE_CVT_PS_RR>, Sched<[WriteCvtF2I]>;
1978 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1979 "cvtps2dq\t{$src, $dst|$dst, $src}",
1981 (int_x86_sse2_cvtps2dq (memopv4f32 addr:$src)))],
1982 IIC_SSE_CVT_PS_RM>, Sched<[WriteCvtF2ILd]>;
1985 // Convert Packed Double FP to Packed DW Integers
1986 let Predicates = [HasAVX] in {
1987 // The assembler can recognize rr 256-bit instructions by seeing a ymm
1988 // register, but the same isn't true when using memory operands instead.
1989 // Provide other assembly rr and rm forms to address this explicitly.
1990 def VCVTPD2DQrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1991 "vcvtpd2dq\t{$src, $dst|$dst, $src}",
1992 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1993 VEX, Sched<[WriteCvtF2I]>;
1996 def : InstAlias<"vcvtpd2dqx\t{$src, $dst|$dst, $src}",
1997 (VCVTPD2DQrr VR128:$dst, VR128:$src), 0>;
1998 def VCVTPD2DQXrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1999 "vcvtpd2dqx\t{$src, $dst|$dst, $src}",
2001 (int_x86_sse2_cvtpd2dq (loadv2f64 addr:$src)))]>, VEX,
2002 Sched<[WriteCvtF2ILd]>;
2005 def VCVTPD2DQYrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2006 "vcvtpd2dq{y}\t{$src, $dst|$dst, $src}",
2008 (int_x86_avx_cvt_pd2dq_256 VR256:$src))]>, VEX, VEX_L,
2009 Sched<[WriteCvtF2I]>;
2010 def VCVTPD2DQYrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2011 "vcvtpd2dq{y}\t{$src, $dst|$dst, $src}",
2013 (int_x86_avx_cvt_pd2dq_256 (loadv4f64 addr:$src)))]>,
2014 VEX, VEX_L, Sched<[WriteCvtF2ILd]>;
2015 def : InstAlias<"vcvtpd2dq\t{$src, $dst|$dst, $src}",
2016 (VCVTPD2DQYrr VR128:$dst, VR256:$src), 0>;
2019 def CVTPD2DQrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2020 "cvtpd2dq\t{$src, $dst|$dst, $src}",
2022 (int_x86_sse2_cvtpd2dq (memopv2f64 addr:$src)))],
2023 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtF2ILd]>;
2024 def CVTPD2DQrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2025 "cvtpd2dq\t{$src, $dst|$dst, $src}",
2026 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))],
2027 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2I]>;
2029 // Convert with truncation packed single/double fp to doubleword
2030 // SSE2 packed instructions with XS prefix
2031 def VCVTTPS2DQrr : VS2SI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2032 "cvttps2dq\t{$src, $dst|$dst, $src}",
2034 (int_x86_sse2_cvttps2dq VR128:$src))],
2035 IIC_SSE_CVT_PS_RR>, VEX, Sched<[WriteCvtF2I]>;
2036 def VCVTTPS2DQrm : VS2SI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2037 "cvttps2dq\t{$src, $dst|$dst, $src}",
2038 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
2039 (loadv4f32 addr:$src)))],
2040 IIC_SSE_CVT_PS_RM>, VEX, Sched<[WriteCvtF2ILd]>;
2041 def VCVTTPS2DQYrr : VS2SI<0x5B, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
2042 "cvttps2dq\t{$src, $dst|$dst, $src}",
2044 (int_x86_avx_cvtt_ps2dq_256 VR256:$src))],
2045 IIC_SSE_CVT_PS_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>;
2046 def VCVTTPS2DQYrm : VS2SI<0x5B, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
2047 "cvttps2dq\t{$src, $dst|$dst, $src}",
2048 [(set VR256:$dst, (int_x86_avx_cvtt_ps2dq_256
2049 (loadv8f32 addr:$src)))],
2050 IIC_SSE_CVT_PS_RM>, VEX, VEX_L,
2051 Sched<[WriteCvtF2ILd]>;
2053 def CVTTPS2DQrr : S2SI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2054 "cvttps2dq\t{$src, $dst|$dst, $src}",
2055 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))],
2056 IIC_SSE_CVT_PS_RR>, Sched<[WriteCvtF2I]>;
2057 def CVTTPS2DQrm : S2SI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2058 "cvttps2dq\t{$src, $dst|$dst, $src}",
2060 (int_x86_sse2_cvttps2dq (memopv4f32 addr:$src)))],
2061 IIC_SSE_CVT_PS_RM>, Sched<[WriteCvtF2ILd]>;
2063 let Predicates = [HasAVX] in {
2064 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
2065 (VCVTDQ2PSrr VR128:$src)>;
2066 def : Pat<(v4f32 (sint_to_fp (bc_v4i32 (loadv2i64 addr:$src)))),
2067 (VCVTDQ2PSrm addr:$src)>;
2069 def : Pat<(int_x86_sse2_cvtdq2ps VR128:$src),
2070 (VCVTDQ2PSrr VR128:$src)>;
2071 def : Pat<(int_x86_sse2_cvtdq2ps (bc_v4i32 (loadv2i64 addr:$src))),
2072 (VCVTDQ2PSrm addr:$src)>;
2074 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
2075 (VCVTTPS2DQrr VR128:$src)>;
2076 def : Pat<(v4i32 (fp_to_sint (loadv4f32 addr:$src))),
2077 (VCVTTPS2DQrm addr:$src)>;
2079 def : Pat<(v8f32 (sint_to_fp (v8i32 VR256:$src))),
2080 (VCVTDQ2PSYrr VR256:$src)>;
2081 def : Pat<(v8f32 (sint_to_fp (bc_v8i32 (loadv4i64 addr:$src)))),
2082 (VCVTDQ2PSYrm addr:$src)>;
2084 def : Pat<(v8i32 (fp_to_sint (v8f32 VR256:$src))),
2085 (VCVTTPS2DQYrr VR256:$src)>;
2086 def : Pat<(v8i32 (fp_to_sint (loadv8f32 addr:$src))),
2087 (VCVTTPS2DQYrm addr:$src)>;
2090 let Predicates = [UseSSE2] in {
2091 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
2092 (CVTDQ2PSrr VR128:$src)>;
2093 def : Pat<(v4f32 (sint_to_fp (bc_v4i32 (memopv2i64 addr:$src)))),
2094 (CVTDQ2PSrm addr:$src)>;
2096 def : Pat<(int_x86_sse2_cvtdq2ps VR128:$src),
2097 (CVTDQ2PSrr VR128:$src)>;
2098 def : Pat<(int_x86_sse2_cvtdq2ps (bc_v4i32 (memopv2i64 addr:$src))),
2099 (CVTDQ2PSrm addr:$src)>;
2101 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
2102 (CVTTPS2DQrr VR128:$src)>;
2103 def : Pat<(v4i32 (fp_to_sint (memopv4f32 addr:$src))),
2104 (CVTTPS2DQrm addr:$src)>;
2107 def VCVTTPD2DQrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2108 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2110 (int_x86_sse2_cvttpd2dq VR128:$src))],
2111 IIC_SSE_CVT_PD_RR>, VEX, Sched<[WriteCvtF2I]>;
2113 // The assembler can recognize rr 256-bit instructions by seeing a ymm
2114 // register, but the same isn't true when using memory operands instead.
2115 // Provide other assembly rr and rm forms to address this explicitly.
2118 def : InstAlias<"vcvttpd2dqx\t{$src, $dst|$dst, $src}",
2119 (VCVTTPD2DQrr VR128:$dst, VR128:$src), 0>;
2120 def VCVTTPD2DQXrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2121 "cvttpd2dqx\t{$src, $dst|$dst, $src}",
2122 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
2123 (loadv2f64 addr:$src)))],
2124 IIC_SSE_CVT_PD_RM>, VEX, Sched<[WriteCvtF2ILd]>;
2127 def VCVTTPD2DQYrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2128 "cvttpd2dq{y}\t{$src, $dst|$dst, $src}",
2130 (int_x86_avx_cvtt_pd2dq_256 VR256:$src))],
2131 IIC_SSE_CVT_PD_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>;
2132 def VCVTTPD2DQYrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2133 "cvttpd2dq{y}\t{$src, $dst|$dst, $src}",
2135 (int_x86_avx_cvtt_pd2dq_256 (loadv4f64 addr:$src)))],
2136 IIC_SSE_CVT_PD_RM>, VEX, VEX_L, Sched<[WriteCvtF2ILd]>;
2137 def : InstAlias<"vcvttpd2dq\t{$src, $dst|$dst, $src}",
2138 (VCVTTPD2DQYrr VR128:$dst, VR256:$src), 0>;
2140 let Predicates = [HasAVX] in {
2141 def : Pat<(v4i32 (fp_to_sint (v4f64 VR256:$src))),
2142 (VCVTTPD2DQYrr VR256:$src)>;
2143 def : Pat<(v4i32 (fp_to_sint (loadv4f64 addr:$src))),
2144 (VCVTTPD2DQYrm addr:$src)>;
2145 } // Predicates = [HasAVX]
2147 def CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2148 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2149 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))],
2150 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2I]>;
2151 def CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
2152 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2153 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
2154 (memopv2f64 addr:$src)))],
2156 Sched<[WriteCvtF2ILd]>;
2158 // Convert packed single to packed double
2159 let Predicates = [HasAVX] in {
2160 // SSE2 instructions without OpSize prefix
2161 def VCVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2162 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2163 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))],
2164 IIC_SSE_CVT_PD_RR>, PS, VEX, Sched<[WriteCvtF2F]>;
2165 def VCVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2166 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2167 [(set VR128:$dst, (v2f64 (extloadv2f32 addr:$src)))],
2168 IIC_SSE_CVT_PD_RM>, PS, VEX, Sched<[WriteCvtF2FLd]>;
2169 def VCVTPS2PDYrr : I<0x5A, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
2170 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2172 (int_x86_avx_cvt_ps2_pd_256 VR128:$src))],
2173 IIC_SSE_CVT_PD_RR>, PS, VEX, VEX_L, Sched<[WriteCvtF2F]>;
2174 def VCVTPS2PDYrm : I<0x5A, MRMSrcMem, (outs VR256:$dst), (ins f128mem:$src),
2175 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2177 (int_x86_avx_cvt_ps2_pd_256 (loadv4f32 addr:$src)))],
2178 IIC_SSE_CVT_PD_RM>, PS, VEX, VEX_L, Sched<[WriteCvtF2FLd]>;
2181 let Predicates = [UseSSE2] in {
2182 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2183 "cvtps2pd\t{$src, $dst|$dst, $src}",
2184 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))],
2185 IIC_SSE_CVT_PD_RR>, PS, Sched<[WriteCvtF2F]>;
2186 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2187 "cvtps2pd\t{$src, $dst|$dst, $src}",
2188 [(set VR128:$dst, (v2f64 (extloadv2f32 addr:$src)))],
2189 IIC_SSE_CVT_PD_RM>, PS, Sched<[WriteCvtF2FLd]>;
2192 // Convert Packed DW Integers to Packed Double FP
2193 let Predicates = [HasAVX] in {
2194 let neverHasSideEffects = 1, mayLoad = 1 in
2195 def VCVTDQ2PDrm : S2SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2196 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2197 []>, VEX, Sched<[WriteCvtI2FLd]>;
2198 def VCVTDQ2PDrr : S2SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2199 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2201 (int_x86_sse2_cvtdq2pd VR128:$src))]>, VEX,
2202 Sched<[WriteCvtI2F]>;
2203 def VCVTDQ2PDYrm : S2SI<0xE6, MRMSrcMem, (outs VR256:$dst), (ins i128mem:$src),
2204 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2206 (int_x86_avx_cvtdq2_pd_256
2207 (bitconvert (loadv2i64 addr:$src))))]>, VEX, VEX_L,
2208 Sched<[WriteCvtI2FLd]>;
2209 def VCVTDQ2PDYrr : S2SI<0xE6, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
2210 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2212 (int_x86_avx_cvtdq2_pd_256 VR128:$src))]>, VEX, VEX_L,
2213 Sched<[WriteCvtI2F]>;
2216 let neverHasSideEffects = 1, mayLoad = 1 in
2217 def CVTDQ2PDrm : S2SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2218 "cvtdq2pd\t{$src, $dst|$dst, $src}", [],
2219 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtI2FLd]>;
2220 def CVTDQ2PDrr : S2SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2221 "cvtdq2pd\t{$src, $dst|$dst, $src}",
2222 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))],
2223 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtI2F]>;
2225 // AVX 256-bit register conversion intrinsics
2226 let Predicates = [HasAVX] in {
2227 def : Pat<(v4f64 (sint_to_fp (v4i32 VR128:$src))),
2228 (VCVTDQ2PDYrr VR128:$src)>;
2229 def : Pat<(v4f64 (sint_to_fp (bc_v4i32 (loadv2i64 addr:$src)))),
2230 (VCVTDQ2PDYrm addr:$src)>;
2231 } // Predicates = [HasAVX]
2233 // Convert packed double to packed single
2234 // The assembler can recognize rr 256-bit instructions by seeing a ymm
2235 // register, but the same isn't true when using memory operands instead.
2236 // Provide other assembly rr and rm forms to address this explicitly.
2237 def VCVTPD2PSrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2238 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2239 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))],
2240 IIC_SSE_CVT_PD_RR>, VEX, Sched<[WriteCvtF2F]>;
2243 def : InstAlias<"vcvtpd2psx\t{$src, $dst|$dst, $src}",
2244 (VCVTPD2PSrr VR128:$dst, VR128:$src), 0>;
2245 def VCVTPD2PSXrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2246 "cvtpd2psx\t{$src, $dst|$dst, $src}",
2248 (int_x86_sse2_cvtpd2ps (loadv2f64 addr:$src)))],
2249 IIC_SSE_CVT_PD_RM>, VEX, Sched<[WriteCvtF2FLd]>;
2252 def VCVTPD2PSYrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2253 "cvtpd2ps{y}\t{$src, $dst|$dst, $src}",
2255 (int_x86_avx_cvt_pd2_ps_256 VR256:$src))],
2256 IIC_SSE_CVT_PD_RR>, VEX, VEX_L, Sched<[WriteCvtF2F]>;
2257 def VCVTPD2PSYrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2258 "cvtpd2ps{y}\t{$src, $dst|$dst, $src}",
2260 (int_x86_avx_cvt_pd2_ps_256 (loadv4f64 addr:$src)))],
2261 IIC_SSE_CVT_PD_RM>, VEX, VEX_L, Sched<[WriteCvtF2FLd]>;
2262 def : InstAlias<"vcvtpd2ps\t{$src, $dst|$dst, $src}",
2263 (VCVTPD2PSYrr VR128:$dst, VR256:$src), 0>;
2265 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2266 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2267 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))],
2268 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2F]>;
2269 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2270 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2272 (int_x86_sse2_cvtpd2ps (memopv2f64 addr:$src)))],
2273 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtF2FLd]>;
2276 // AVX 256-bit register conversion intrinsics
2277 // FIXME: Migrate SSE conversion intrinsics matching to use patterns as below
2278 // whenever possible to avoid declaring two versions of each one.
2279 let Predicates = [HasAVX] in {
2280 def : Pat<(int_x86_avx_cvtdq2_ps_256 VR256:$src),
2281 (VCVTDQ2PSYrr VR256:$src)>;
2282 def : Pat<(int_x86_avx_cvtdq2_ps_256 (bitconvert (loadv4i64 addr:$src))),
2283 (VCVTDQ2PSYrm addr:$src)>;
2285 // Match fround and fextend for 128/256-bit conversions
2286 def : Pat<(v4f32 (X86vfpround (v2f64 VR128:$src))),
2287 (VCVTPD2PSrr VR128:$src)>;
2288 def : Pat<(v4f32 (X86vfpround (loadv2f64 addr:$src))),
2289 (VCVTPD2PSXrm addr:$src)>;
2290 def : Pat<(v4f32 (fround (v4f64 VR256:$src))),
2291 (VCVTPD2PSYrr VR256:$src)>;
2292 def : Pat<(v4f32 (fround (loadv4f64 addr:$src))),
2293 (VCVTPD2PSYrm addr:$src)>;
2295 def : Pat<(v2f64 (X86vfpext (v4f32 VR128:$src))),
2296 (VCVTPS2PDrr VR128:$src)>;
2297 def : Pat<(v4f64 (fextend (v4f32 VR128:$src))),
2298 (VCVTPS2PDYrr VR128:$src)>;
2299 def : Pat<(v4f64 (extloadv4f32 addr:$src)),
2300 (VCVTPS2PDYrm addr:$src)>;
2303 let Predicates = [UseSSE2] in {
2304 // Match fround and fextend for 128 conversions
2305 def : Pat<(v4f32 (X86vfpround (v2f64 VR128:$src))),
2306 (CVTPD2PSrr VR128:$src)>;
2307 def : Pat<(v4f32 (X86vfpround (memopv2f64 addr:$src))),
2308 (CVTPD2PSrm addr:$src)>;
2310 def : Pat<(v2f64 (X86vfpext (v4f32 VR128:$src))),
2311 (CVTPS2PDrr VR128:$src)>;
2314 //===----------------------------------------------------------------------===//
2315 // SSE 1 & 2 - Compare Instructions
2316 //===----------------------------------------------------------------------===//
2318 // sse12_cmp_scalar - sse 1 & 2 compare scalar instructions
2319 multiclass sse12_cmp_scalar<RegisterClass RC, X86MemOperand x86memop,
2320 Operand CC, SDNode OpNode, ValueType VT,
2321 PatFrag ld_frag, string asm, string asm_alt,
2323 def rr : SIi8<0xC2, MRMSrcReg,
2324 (outs RC:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
2325 [(set RC:$dst, (OpNode (VT RC:$src1), RC:$src2, imm:$cc))],
2326 itins.rr>, Sched<[itins.Sched]>;
2327 def rm : SIi8<0xC2, MRMSrcMem,
2328 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
2329 [(set RC:$dst, (OpNode (VT RC:$src1),
2330 (ld_frag addr:$src2), imm:$cc))],
2332 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2334 // Accept explicit immediate argument form instead of comparison code.
2335 let isAsmParserOnly = 1, hasSideEffects = 0 in {
2336 def rr_alt : SIi8<0xC2, MRMSrcReg, (outs RC:$dst),
2337 (ins RC:$src1, RC:$src2, i8imm:$cc), asm_alt, [],
2338 IIC_SSE_ALU_F32S_RR>, Sched<[itins.Sched]>;
2340 def rm_alt : SIi8<0xC2, MRMSrcMem, (outs RC:$dst),
2341 (ins RC:$src1, x86memop:$src2, i8imm:$cc), asm_alt, [],
2342 IIC_SSE_ALU_F32S_RM>,
2343 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2347 defm VCMPSS : sse12_cmp_scalar<FR32, f32mem, AVXCC, X86cmps, f32, loadf32,
2348 "cmp${cc}ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2349 "cmpss\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2351 XS, VEX_4V, VEX_LIG;
2352 defm VCMPSD : sse12_cmp_scalar<FR64, f64mem, AVXCC, X86cmps, f64, loadf64,
2353 "cmp${cc}sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2354 "cmpsd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2355 SSE_ALU_F32S>, // same latency as 32 bit compare
2356 XD, VEX_4V, VEX_LIG;
2358 let Constraints = "$src1 = $dst" in {
2359 defm CMPSS : sse12_cmp_scalar<FR32, f32mem, SSECC, X86cmps, f32, loadf32,
2360 "cmp${cc}ss\t{$src2, $dst|$dst, $src2}",
2361 "cmpss\t{$cc, $src2, $dst|$dst, $src2, $cc}", SSE_ALU_F32S>,
2363 defm CMPSD : sse12_cmp_scalar<FR64, f64mem, SSECC, X86cmps, f64, loadf64,
2364 "cmp${cc}sd\t{$src2, $dst|$dst, $src2}",
2365 "cmpsd\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2370 multiclass sse12_cmp_scalar_int<X86MemOperand x86memop, Operand CC,
2371 Intrinsic Int, string asm, OpndItins itins> {
2372 def rr : SIi8<0xC2, MRMSrcReg, (outs VR128:$dst),
2373 (ins VR128:$src1, VR128:$src, CC:$cc), asm,
2374 [(set VR128:$dst, (Int VR128:$src1,
2375 VR128:$src, imm:$cc))],
2377 Sched<[itins.Sched]>;
2378 def rm : SIi8<0xC2, MRMSrcMem, (outs VR128:$dst),
2379 (ins VR128:$src1, x86memop:$src, CC:$cc), asm,
2380 [(set VR128:$dst, (Int VR128:$src1,
2381 (load addr:$src), imm:$cc))],
2383 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2386 let isCodeGenOnly = 1 in {
2387 // Aliases to match intrinsics which expect XMM operand(s).
2388 defm Int_VCMPSS : sse12_cmp_scalar_int<f32mem, AVXCC, int_x86_sse_cmp_ss,
2389 "cmp${cc}ss\t{$src, $src1, $dst|$dst, $src1, $src}",
2392 defm Int_VCMPSD : sse12_cmp_scalar_int<f64mem, AVXCC, int_x86_sse2_cmp_sd,
2393 "cmp${cc}sd\t{$src, $src1, $dst|$dst, $src1, $src}",
2394 SSE_ALU_F32S>, // same latency as f32
2396 let Constraints = "$src1 = $dst" in {
2397 defm Int_CMPSS : sse12_cmp_scalar_int<f32mem, SSECC, int_x86_sse_cmp_ss,
2398 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
2400 defm Int_CMPSD : sse12_cmp_scalar_int<f64mem, SSECC, int_x86_sse2_cmp_sd,
2401 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
2408 // sse12_ord_cmp - Unordered/Ordered scalar fp compare and set EFLAGS
2409 multiclass sse12_ord_cmp<bits<8> opc, RegisterClass RC, SDNode OpNode,
2410 ValueType vt, X86MemOperand x86memop,
2411 PatFrag ld_frag, string OpcodeStr> {
2412 def rr: SI<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
2413 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2414 [(set EFLAGS, (OpNode (vt RC:$src1), RC:$src2))],
2417 def rm: SI<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
2418 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2419 [(set EFLAGS, (OpNode (vt RC:$src1),
2420 (ld_frag addr:$src2)))],
2422 Sched<[WriteFAddLd, ReadAfterLd]>;
2425 let Defs = [EFLAGS] in {
2426 defm VUCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
2427 "ucomiss">, PS, VEX, VEX_LIG;
2428 defm VUCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
2429 "ucomisd">, PD, VEX, VEX_LIG;
2430 let Pattern = []<dag> in {
2431 defm VCOMISS : sse12_ord_cmp<0x2F, VR128, undef, v4f32, f128mem, load,
2432 "comiss">, PS, VEX, VEX_LIG;
2433 defm VCOMISD : sse12_ord_cmp<0x2F, VR128, undef, v2f64, f128mem, load,
2434 "comisd">, PD, VEX, VEX_LIG;
2437 let isCodeGenOnly = 1 in {
2438 defm Int_VUCOMISS : sse12_ord_cmp<0x2E, VR128, X86ucomi, v4f32, f128mem,
2439 load, "ucomiss">, PS, VEX;
2440 defm Int_VUCOMISD : sse12_ord_cmp<0x2E, VR128, X86ucomi, v2f64, f128mem,
2441 load, "ucomisd">, PD, VEX;
2443 defm Int_VCOMISS : sse12_ord_cmp<0x2F, VR128, X86comi, v4f32, f128mem,
2444 load, "comiss">, PS, VEX;
2445 defm Int_VCOMISD : sse12_ord_cmp<0x2F, VR128, X86comi, v2f64, f128mem,
2446 load, "comisd">, PD, VEX;
2448 defm UCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
2450 defm UCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
2453 let Pattern = []<dag> in {
2454 defm COMISS : sse12_ord_cmp<0x2F, VR128, undef, v4f32, f128mem, load,
2456 defm COMISD : sse12_ord_cmp<0x2F, VR128, undef, v2f64, f128mem, load,
2460 let isCodeGenOnly = 1 in {
2461 defm Int_UCOMISS : sse12_ord_cmp<0x2E, VR128, X86ucomi, v4f32, f128mem,
2462 load, "ucomiss">, PS;
2463 defm Int_UCOMISD : sse12_ord_cmp<0x2E, VR128, X86ucomi, v2f64, f128mem,
2464 load, "ucomisd">, PD;
2466 defm Int_COMISS : sse12_ord_cmp<0x2F, VR128, X86comi, v4f32, f128mem, load,
2468 defm Int_COMISD : sse12_ord_cmp<0x2F, VR128, X86comi, v2f64, f128mem, load,
2471 } // Defs = [EFLAGS]
2473 // sse12_cmp_packed - sse 1 & 2 compare packed instructions
2474 multiclass sse12_cmp_packed<RegisterClass RC, X86MemOperand x86memop,
2475 Operand CC, Intrinsic Int, string asm,
2476 string asm_alt, Domain d,
2477 OpndItins itins = SSE_ALU_F32P> {
2478 def rri : PIi8<0xC2, MRMSrcReg,
2479 (outs RC:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
2480 [(set RC:$dst, (Int RC:$src1, RC:$src2, imm:$cc))],
2483 def rmi : PIi8<0xC2, MRMSrcMem,
2484 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
2485 [(set RC:$dst, (Int RC:$src1, (memop addr:$src2), imm:$cc))],
2487 Sched<[WriteFAddLd, ReadAfterLd]>;
2489 // Accept explicit immediate argument form instead of comparison code.
2490 let isAsmParserOnly = 1, hasSideEffects = 0 in {
2491 def rri_alt : PIi8<0xC2, MRMSrcReg,
2492 (outs RC:$dst), (ins RC:$src1, RC:$src2, i8imm:$cc),
2493 asm_alt, [], itins.rr, d>, Sched<[WriteFAdd]>;
2494 def rmi_alt : PIi8<0xC2, MRMSrcMem,
2495 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, i8imm:$cc),
2496 asm_alt, [], itins.rm, d>,
2497 Sched<[WriteFAddLd, ReadAfterLd]>;
2501 defm VCMPPS : sse12_cmp_packed<VR128, f128mem, AVXCC, int_x86_sse_cmp_ps,
2502 "cmp${cc}ps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2503 "cmpps\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2504 SSEPackedSingle>, PS, VEX_4V;
2505 defm VCMPPD : sse12_cmp_packed<VR128, f128mem, AVXCC, int_x86_sse2_cmp_pd,
2506 "cmp${cc}pd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2507 "cmppd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2508 SSEPackedDouble>, PD, VEX_4V;
2509 defm VCMPPSY : sse12_cmp_packed<VR256, f256mem, AVXCC, int_x86_avx_cmp_ps_256,
2510 "cmp${cc}ps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2511 "cmpps\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2512 SSEPackedSingle>, PS, VEX_4V, VEX_L;
2513 defm VCMPPDY : sse12_cmp_packed<VR256, f256mem, AVXCC, int_x86_avx_cmp_pd_256,
2514 "cmp${cc}pd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2515 "cmppd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2516 SSEPackedDouble>, PD, VEX_4V, VEX_L;
2517 let Constraints = "$src1 = $dst" in {
2518 defm CMPPS : sse12_cmp_packed<VR128, f128mem, SSECC, int_x86_sse_cmp_ps,
2519 "cmp${cc}ps\t{$src2, $dst|$dst, $src2}",
2520 "cmpps\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2521 SSEPackedSingle, SSE_ALU_F32P>, PS;
2522 defm CMPPD : sse12_cmp_packed<VR128, f128mem, SSECC, int_x86_sse2_cmp_pd,
2523 "cmp${cc}pd\t{$src2, $dst|$dst, $src2}",
2524 "cmppd\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2525 SSEPackedDouble, SSE_ALU_F64P>, PD;
2528 let Predicates = [HasAVX] in {
2529 def : Pat<(v4i32 (X86cmpp (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
2530 (VCMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
2531 def : Pat<(v4i32 (X86cmpp (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
2532 (VCMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
2533 def : Pat<(v2i64 (X86cmpp (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
2534 (VCMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
2535 def : Pat<(v2i64 (X86cmpp (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
2536 (VCMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
2538 def : Pat<(v8i32 (X86cmpp (v8f32 VR256:$src1), VR256:$src2, imm:$cc)),
2539 (VCMPPSYrri (v8f32 VR256:$src1), (v8f32 VR256:$src2), imm:$cc)>;
2540 def : Pat<(v8i32 (X86cmpp (v8f32 VR256:$src1), (memop addr:$src2), imm:$cc)),
2541 (VCMPPSYrmi (v8f32 VR256:$src1), addr:$src2, imm:$cc)>;
2542 def : Pat<(v4i64 (X86cmpp (v4f64 VR256:$src1), VR256:$src2, imm:$cc)),
2543 (VCMPPDYrri VR256:$src1, VR256:$src2, imm:$cc)>;
2544 def : Pat<(v4i64 (X86cmpp (v4f64 VR256:$src1), (memop addr:$src2), imm:$cc)),
2545 (VCMPPDYrmi VR256:$src1, addr:$src2, imm:$cc)>;
2548 let Predicates = [UseSSE1] in {
2549 def : Pat<(v4i32 (X86cmpp (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
2550 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
2551 def : Pat<(v4i32 (X86cmpp (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
2552 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
2555 let Predicates = [UseSSE2] in {
2556 def : Pat<(v2i64 (X86cmpp (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
2557 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
2558 def : Pat<(v2i64 (X86cmpp (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
2559 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
2562 //===----------------------------------------------------------------------===//
2563 // SSE 1 & 2 - Shuffle Instructions
2564 //===----------------------------------------------------------------------===//
2566 /// sse12_shuffle - sse 1 & 2 fp shuffle instructions
2567 multiclass sse12_shuffle<RegisterClass RC, X86MemOperand x86memop,
2568 ValueType vt, string asm, PatFrag mem_frag,
2569 Domain d, bit IsConvertibleToThreeAddress = 0> {
2570 def rmi : PIi8<0xC6, MRMSrcMem, (outs RC:$dst),
2571 (ins RC:$src1, x86memop:$src2, i8imm:$src3), asm,
2572 [(set RC:$dst, (vt (X86Shufp RC:$src1, (mem_frag addr:$src2),
2573 (i8 imm:$src3))))], IIC_SSE_SHUFP, d>,
2574 Sched<[WriteFShuffleLd, ReadAfterLd]>;
2575 let isConvertibleToThreeAddress = IsConvertibleToThreeAddress in
2576 def rri : PIi8<0xC6, MRMSrcReg, (outs RC:$dst),
2577 (ins RC:$src1, RC:$src2, i8imm:$src3), asm,
2578 [(set RC:$dst, (vt (X86Shufp RC:$src1, RC:$src2,
2579 (i8 imm:$src3))))], IIC_SSE_SHUFP, d>,
2580 Sched<[WriteFShuffle]>;
2583 defm VSHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
2584 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2585 loadv4f32, SSEPackedSingle>, PS, VEX_4V;
2586 defm VSHUFPSY : sse12_shuffle<VR256, f256mem, v8f32,
2587 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2588 loadv8f32, SSEPackedSingle>, PS, VEX_4V, VEX_L;
2589 defm VSHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
2590 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2591 loadv2f64, SSEPackedDouble>, PD, VEX_4V;
2592 defm VSHUFPDY : sse12_shuffle<VR256, f256mem, v4f64,
2593 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2594 loadv4f64, SSEPackedDouble>, PD, VEX_4V, VEX_L;
2596 let Constraints = "$src1 = $dst" in {
2597 defm SHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
2598 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2599 memopv4f32, SSEPackedSingle, 1 /* cvt to pshufd */>, PS;
2600 defm SHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
2601 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2602 memopv2f64, SSEPackedDouble, 1 /* cvt to pshufd */>, PD;
2605 let Predicates = [HasAVX] in {
2606 def : Pat<(v4i32 (X86Shufp VR128:$src1,
2607 (bc_v4i32 (loadv2i64 addr:$src2)), (i8 imm:$imm))),
2608 (VSHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>;
2609 def : Pat<(v4i32 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2610 (VSHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>;
2612 def : Pat<(v2i64 (X86Shufp VR128:$src1,
2613 (loadv2i64 addr:$src2), (i8 imm:$imm))),
2614 (VSHUFPDrmi VR128:$src1, addr:$src2, imm:$imm)>;
2615 def : Pat<(v2i64 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2616 (VSHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>;
2619 def : Pat<(v8i32 (X86Shufp VR256:$src1, VR256:$src2, (i8 imm:$imm))),
2620 (VSHUFPSYrri VR256:$src1, VR256:$src2, imm:$imm)>;
2621 def : Pat<(v8i32 (X86Shufp VR256:$src1,
2622 (bc_v8i32 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
2623 (VSHUFPSYrmi VR256:$src1, addr:$src2, imm:$imm)>;
2625 def : Pat<(v4i64 (X86Shufp VR256:$src1, VR256:$src2, (i8 imm:$imm))),
2626 (VSHUFPDYrri VR256:$src1, VR256:$src2, imm:$imm)>;
2627 def : Pat<(v4i64 (X86Shufp VR256:$src1,
2628 (loadv4i64 addr:$src2), (i8 imm:$imm))),
2629 (VSHUFPDYrmi VR256:$src1, addr:$src2, imm:$imm)>;
2632 let Predicates = [UseSSE1] in {
2633 def : Pat<(v4i32 (X86Shufp VR128:$src1,
2634 (bc_v4i32 (memopv2i64 addr:$src2)), (i8 imm:$imm))),
2635 (SHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>;
2636 def : Pat<(v4i32 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2637 (SHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>;
2640 let Predicates = [UseSSE2] in {
2641 // Generic SHUFPD patterns
2642 def : Pat<(v2i64 (X86Shufp VR128:$src1,
2643 (memopv2i64 addr:$src2), (i8 imm:$imm))),
2644 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$imm)>;
2645 def : Pat<(v2i64 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2646 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>;
2649 //===----------------------------------------------------------------------===//
2650 // SSE 1 & 2 - Unpack FP Instructions
2651 //===----------------------------------------------------------------------===//
2653 /// sse12_unpack_interleave - sse 1 & 2 fp unpack and interleave
2654 multiclass sse12_unpack_interleave<bits<8> opc, SDNode OpNode, ValueType vt,
2655 PatFrag mem_frag, RegisterClass RC,
2656 X86MemOperand x86memop, string asm,
2658 def rr : PI<opc, MRMSrcReg,
2659 (outs RC:$dst), (ins RC:$src1, RC:$src2),
2661 (vt (OpNode RC:$src1, RC:$src2)))],
2662 IIC_SSE_UNPCK, d>, Sched<[WriteFShuffle]>;
2663 def rm : PI<opc, MRMSrcMem,
2664 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
2666 (vt (OpNode RC:$src1,
2667 (mem_frag addr:$src2))))],
2669 Sched<[WriteFShuffleLd, ReadAfterLd]>;
2672 defm VUNPCKHPS: sse12_unpack_interleave<0x15, X86Unpckh, v4f32, loadv4f32,
2673 VR128, f128mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2674 SSEPackedSingle>, PS, VEX_4V;
2675 defm VUNPCKHPD: sse12_unpack_interleave<0x15, X86Unpckh, v2f64, loadv2f64,
2676 VR128, f128mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2677 SSEPackedDouble>, PD, VEX_4V;
2678 defm VUNPCKLPS: sse12_unpack_interleave<0x14, X86Unpckl, v4f32, loadv4f32,
2679 VR128, f128mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2680 SSEPackedSingle>, PS, VEX_4V;
2681 defm VUNPCKLPD: sse12_unpack_interleave<0x14, X86Unpckl, v2f64, loadv2f64,
2682 VR128, f128mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2683 SSEPackedDouble>, PD, VEX_4V;
2685 defm VUNPCKHPSY: sse12_unpack_interleave<0x15, X86Unpckh, v8f32, loadv8f32,
2686 VR256, f256mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2687 SSEPackedSingle>, PS, VEX_4V, VEX_L;
2688 defm VUNPCKHPDY: sse12_unpack_interleave<0x15, X86Unpckh, v4f64, loadv4f64,
2689 VR256, f256mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2690 SSEPackedDouble>, PD, VEX_4V, VEX_L;
2691 defm VUNPCKLPSY: sse12_unpack_interleave<0x14, X86Unpckl, v8f32, loadv8f32,
2692 VR256, f256mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2693 SSEPackedSingle>, PS, VEX_4V, VEX_L;
2694 defm VUNPCKLPDY: sse12_unpack_interleave<0x14, X86Unpckl, v4f64, loadv4f64,
2695 VR256, f256mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2696 SSEPackedDouble>, PD, VEX_4V, VEX_L;
2698 let Constraints = "$src1 = $dst" in {
2699 defm UNPCKHPS: sse12_unpack_interleave<0x15, X86Unpckh, v4f32, memopv4f32,
2700 VR128, f128mem, "unpckhps\t{$src2, $dst|$dst, $src2}",
2701 SSEPackedSingle>, PS;
2702 defm UNPCKHPD: sse12_unpack_interleave<0x15, X86Unpckh, v2f64, memopv2f64,
2703 VR128, f128mem, "unpckhpd\t{$src2, $dst|$dst, $src2}",
2704 SSEPackedDouble>, PD;
2705 defm UNPCKLPS: sse12_unpack_interleave<0x14, X86Unpckl, v4f32, memopv4f32,
2706 VR128, f128mem, "unpcklps\t{$src2, $dst|$dst, $src2}",
2707 SSEPackedSingle>, PS;
2708 defm UNPCKLPD: sse12_unpack_interleave<0x14, X86Unpckl, v2f64, memopv2f64,
2709 VR128, f128mem, "unpcklpd\t{$src2, $dst|$dst, $src2}",
2710 SSEPackedDouble>, PD;
2711 } // Constraints = "$src1 = $dst"
2713 let Predicates = [HasAVX1Only] in {
2714 def : Pat<(v8i32 (X86Unpckl VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
2715 (VUNPCKLPSYrm VR256:$src1, addr:$src2)>;
2716 def : Pat<(v8i32 (X86Unpckl VR256:$src1, VR256:$src2)),
2717 (VUNPCKLPSYrr VR256:$src1, VR256:$src2)>;
2718 def : Pat<(v8i32 (X86Unpckh VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
2719 (VUNPCKHPSYrm VR256:$src1, addr:$src2)>;
2720 def : Pat<(v8i32 (X86Unpckh VR256:$src1, VR256:$src2)),
2721 (VUNPCKHPSYrr VR256:$src1, VR256:$src2)>;
2723 def : Pat<(v4i64 (X86Unpckl VR256:$src1, (loadv4i64 addr:$src2))),
2724 (VUNPCKLPDYrm VR256:$src1, addr:$src2)>;
2725 def : Pat<(v4i64 (X86Unpckl VR256:$src1, VR256:$src2)),
2726 (VUNPCKLPDYrr VR256:$src1, VR256:$src2)>;
2727 def : Pat<(v4i64 (X86Unpckh VR256:$src1, (loadv4i64 addr:$src2))),
2728 (VUNPCKHPDYrm VR256:$src1, addr:$src2)>;
2729 def : Pat<(v4i64 (X86Unpckh VR256:$src1, VR256:$src2)),
2730 (VUNPCKHPDYrr VR256:$src1, VR256:$src2)>;
2733 let Predicates = [HasAVX] in {
2734 // FIXME: Instead of X86Movddup, there should be a X86Unpckl here, the
2735 // problem is during lowering, where it's not possible to recognize the load
2736 // fold cause it has two uses through a bitcast. One use disappears at isel
2737 // time and the fold opportunity reappears.
2738 def : Pat<(v2f64 (X86Movddup VR128:$src)),
2739 (VUNPCKLPDrr VR128:$src, VR128:$src)>;
2742 let Predicates = [UseSSE2] in {
2743 // FIXME: Instead of X86Movddup, there should be a X86Unpckl here, the
2744 // problem is during lowering, where it's not possible to recognize the load
2745 // fold cause it has two uses through a bitcast. One use disappears at isel
2746 // time and the fold opportunity reappears.
2747 def : Pat<(v2f64 (X86Movddup VR128:$src)),
2748 (UNPCKLPDrr VR128:$src, VR128:$src)>;
2751 //===----------------------------------------------------------------------===//
2752 // SSE 1 & 2 - Extract Floating-Point Sign mask
2753 //===----------------------------------------------------------------------===//
2755 /// sse12_extr_sign_mask - sse 1 & 2 unpack and interleave
2756 multiclass sse12_extr_sign_mask<RegisterClass RC, Intrinsic Int, string asm,
2758 def rr : PI<0x50, MRMSrcReg, (outs GR32orGR64:$dst), (ins RC:$src),
2759 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
2760 [(set GR32orGR64:$dst, (Int RC:$src))], IIC_SSE_MOVMSK, d>,
2761 Sched<[WriteVecLogic]>;
2764 let Predicates = [HasAVX] in {
2765 defm VMOVMSKPS : sse12_extr_sign_mask<VR128, int_x86_sse_movmsk_ps,
2766 "movmskps", SSEPackedSingle>, PS, VEX;
2767 defm VMOVMSKPD : sse12_extr_sign_mask<VR128, int_x86_sse2_movmsk_pd,
2768 "movmskpd", SSEPackedDouble>, PD, VEX;
2769 defm VMOVMSKPSY : sse12_extr_sign_mask<VR256, int_x86_avx_movmsk_ps_256,
2770 "movmskps", SSEPackedSingle>, PS,
2772 defm VMOVMSKPDY : sse12_extr_sign_mask<VR256, int_x86_avx_movmsk_pd_256,
2773 "movmskpd", SSEPackedDouble>, PD,
2776 def : Pat<(i32 (X86fgetsign FR32:$src)),
2777 (VMOVMSKPSrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
2778 def : Pat<(i64 (X86fgetsign FR32:$src)),
2779 (SUBREG_TO_REG (i64 0),
2780 (VMOVMSKPSrr (COPY_TO_REGCLASS FR32:$src, VR128)), sub_32bit)>;
2781 def : Pat<(i32 (X86fgetsign FR64:$src)),
2782 (VMOVMSKPDrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
2783 def : Pat<(i64 (X86fgetsign FR64:$src)),
2784 (SUBREG_TO_REG (i64 0),
2785 (VMOVMSKPDrr (COPY_TO_REGCLASS FR64:$src, VR128)), sub_32bit)>;
2788 defm MOVMSKPS : sse12_extr_sign_mask<VR128, int_x86_sse_movmsk_ps, "movmskps",
2789 SSEPackedSingle>, PS;
2790 defm MOVMSKPD : sse12_extr_sign_mask<VR128, int_x86_sse2_movmsk_pd, "movmskpd",
2791 SSEPackedDouble>, PD;
2793 def : Pat<(i32 (X86fgetsign FR32:$src)),
2794 (MOVMSKPSrr (COPY_TO_REGCLASS FR32:$src, VR128))>,
2795 Requires<[UseSSE1]>;
2796 def : Pat<(i64 (X86fgetsign FR32:$src)),
2797 (SUBREG_TO_REG (i64 0),
2798 (MOVMSKPSrr (COPY_TO_REGCLASS FR32:$src, VR128)), sub_32bit)>,
2799 Requires<[UseSSE1]>;
2800 def : Pat<(i32 (X86fgetsign FR64:$src)),
2801 (MOVMSKPDrr (COPY_TO_REGCLASS FR64:$src, VR128))>,
2802 Requires<[UseSSE2]>;
2803 def : Pat<(i64 (X86fgetsign FR64:$src)),
2804 (SUBREG_TO_REG (i64 0),
2805 (MOVMSKPDrr (COPY_TO_REGCLASS FR64:$src, VR128)), sub_32bit)>,
2806 Requires<[UseSSE2]>;
2808 //===---------------------------------------------------------------------===//
2809 // SSE2 - Packed Integer Logical Instructions
2810 //===---------------------------------------------------------------------===//
2812 let ExeDomain = SSEPackedInt in { // SSE integer instructions
2814 /// PDI_binop_rm - Simple SSE2 binary operator.
2815 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
2816 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
2817 X86MemOperand x86memop, OpndItins itins,
2818 bit IsCommutable, bit Is2Addr> {
2819 let isCommutable = IsCommutable in
2820 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
2821 (ins RC:$src1, RC:$src2),
2823 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2824 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2825 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))], itins.rr>,
2826 Sched<[itins.Sched]>;
2827 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
2828 (ins RC:$src1, x86memop:$src2),
2830 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2831 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2832 [(set RC:$dst, (OpVT (OpNode RC:$src1,
2833 (bitconvert (memop_frag addr:$src2)))))],
2835 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2837 } // ExeDomain = SSEPackedInt
2839 multiclass PDI_binop_all<bits<8> opc, string OpcodeStr, SDNode Opcode,
2840 ValueType OpVT128, ValueType OpVT256,
2841 OpndItins itins, bit IsCommutable = 0> {
2842 let Predicates = [HasAVX] in
2843 defm V#NAME : PDI_binop_rm<opc, !strconcat("v", OpcodeStr), Opcode, OpVT128,
2844 VR128, loadv2i64, i128mem, itins, IsCommutable, 0>, VEX_4V;
2846 let Constraints = "$src1 = $dst" in
2847 defm NAME : PDI_binop_rm<opc, OpcodeStr, Opcode, OpVT128, VR128,
2848 memopv2i64, i128mem, itins, IsCommutable, 1>;
2850 let Predicates = [HasAVX2] in
2851 defm V#NAME#Y : PDI_binop_rm<opc, !strconcat("v", OpcodeStr), Opcode,
2852 OpVT256, VR256, loadv4i64, i256mem, itins,
2853 IsCommutable, 0>, VEX_4V, VEX_L;
2856 // These are ordered here for pattern ordering requirements with the fp versions
2858 defm PAND : PDI_binop_all<0xDB, "pand", and, v2i64, v4i64,
2859 SSE_VEC_BIT_ITINS_P, 1>;
2860 defm POR : PDI_binop_all<0xEB, "por", or, v2i64, v4i64,
2861 SSE_VEC_BIT_ITINS_P, 1>;
2862 defm PXOR : PDI_binop_all<0xEF, "pxor", xor, v2i64, v4i64,
2863 SSE_VEC_BIT_ITINS_P, 1>;
2864 defm PANDN : PDI_binop_all<0xDF, "pandn", X86andnp, v2i64, v4i64,
2865 SSE_VEC_BIT_ITINS_P, 0>;
2867 //===----------------------------------------------------------------------===//
2868 // SSE 1 & 2 - Logical Instructions
2869 //===----------------------------------------------------------------------===//
2871 /// sse12_fp_alias_pack_logical - SSE 1 & 2 aliased packed FP logical ops
2873 multiclass sse12_fp_alias_pack_logical<bits<8> opc, string OpcodeStr,
2874 SDNode OpNode, OpndItins itins> {
2875 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode,
2876 FR32, f32, f128mem, memopfsf32, SSEPackedSingle, itins, 0>,
2879 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode,
2880 FR64, f64, f128mem, memopfsf64, SSEPackedDouble, itins, 0>,
2883 let Constraints = "$src1 = $dst" in {
2884 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode, FR32,
2885 f32, f128mem, memopfsf32, SSEPackedSingle, itins>,
2888 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode, FR64,
2889 f64, f128mem, memopfsf64, SSEPackedDouble, itins>,
2894 // Alias bitwise logical operations using SSE logical ops on packed FP values.
2895 let isCodeGenOnly = 1 in {
2896 defm FsAND : sse12_fp_alias_pack_logical<0x54, "and", X86fand,
2898 defm FsOR : sse12_fp_alias_pack_logical<0x56, "or", X86for,
2900 defm FsXOR : sse12_fp_alias_pack_logical<0x57, "xor", X86fxor,
2903 let isCommutable = 0 in
2904 defm FsANDN : sse12_fp_alias_pack_logical<0x55, "andn", X86fandn,
2908 /// sse12_fp_packed_logical - SSE 1 & 2 packed FP logical ops
2910 multiclass sse12_fp_packed_logical<bits<8> opc, string OpcodeStr,
2912 defm V#NAME#PSY : sse12_fp_packed_logical_rm<opc, VR256, SSEPackedSingle,
2913 !strconcat(OpcodeStr, "ps"), f256mem,
2914 [(set VR256:$dst, (v4i64 (OpNode VR256:$src1, VR256:$src2)))],
2915 [(set VR256:$dst, (OpNode (bc_v4i64 (v8f32 VR256:$src1)),
2916 (loadv4i64 addr:$src2)))], 0>, PS, VEX_4V, VEX_L;
2918 defm V#NAME#PDY : sse12_fp_packed_logical_rm<opc, VR256, SSEPackedDouble,
2919 !strconcat(OpcodeStr, "pd"), f256mem,
2920 [(set VR256:$dst, (OpNode (bc_v4i64 (v4f64 VR256:$src1)),
2921 (bc_v4i64 (v4f64 VR256:$src2))))],
2922 [(set VR256:$dst, (OpNode (bc_v4i64 (v4f64 VR256:$src1)),
2923 (loadv4i64 addr:$src2)))], 0>,
2926 // In AVX no need to add a pattern for 128-bit logical rr ps, because they
2927 // are all promoted to v2i64, and the patterns are covered by the int
2928 // version. This is needed in SSE only, because v2i64 isn't supported on
2929 // SSE1, but only on SSE2.
2930 defm V#NAME#PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
2931 !strconcat(OpcodeStr, "ps"), f128mem, [],
2932 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2933 (loadv2i64 addr:$src2)))], 0>, PS, VEX_4V;
2935 defm V#NAME#PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
2936 !strconcat(OpcodeStr, "pd"), f128mem,
2937 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2938 (bc_v2i64 (v2f64 VR128:$src2))))],
2939 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2940 (loadv2i64 addr:$src2)))], 0>,
2943 let Constraints = "$src1 = $dst" in {
2944 defm PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
2945 !strconcat(OpcodeStr, "ps"), f128mem,
2946 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))],
2947 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2948 (memopv2i64 addr:$src2)))]>, PS;
2950 defm PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
2951 !strconcat(OpcodeStr, "pd"), f128mem,
2952 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2953 (bc_v2i64 (v2f64 VR128:$src2))))],
2954 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2955 (memopv2i64 addr:$src2)))]>, PD;
2959 defm AND : sse12_fp_packed_logical<0x54, "and", and>;
2960 defm OR : sse12_fp_packed_logical<0x56, "or", or>;
2961 defm XOR : sse12_fp_packed_logical<0x57, "xor", xor>;
2962 let isCommutable = 0 in
2963 defm ANDN : sse12_fp_packed_logical<0x55, "andn", X86andnp>;
2965 // AVX1 requires type coercions in order to fold loads directly into logical
2967 let Predicates = [HasAVX1Only] in {
2968 def : Pat<(bc_v8f32 (and VR256:$src1, (loadv4i64 addr:$src2))),
2969 (VANDPSYrm VR256:$src1, addr:$src2)>;
2970 def : Pat<(bc_v8f32 (or VR256:$src1, (loadv4i64 addr:$src2))),
2971 (VORPSYrm VR256:$src1, addr:$src2)>;
2972 def : Pat<(bc_v8f32 (xor VR256:$src1, (loadv4i64 addr:$src2))),
2973 (VXORPSYrm VR256:$src1, addr:$src2)>;
2974 def : Pat<(bc_v8f32 (X86andnp VR256:$src1, (loadv4i64 addr:$src2))),
2975 (VANDNPSYrm VR256:$src1, addr:$src2)>;
2978 //===----------------------------------------------------------------------===//
2979 // SSE 1 & 2 - Arithmetic Instructions
2980 //===----------------------------------------------------------------------===//
2982 /// basic_sse12_fp_binop_xxx - SSE 1 & 2 binops come in both scalar and
2985 /// In addition, we also have a special variant of the scalar form here to
2986 /// represent the associated intrinsic operation. This form is unlike the
2987 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
2988 /// and leaves the top elements unmodified (therefore these cannot be commuted).
2990 /// These three forms can each be reg+reg or reg+mem.
2993 /// FIXME: once all 256-bit intrinsics are matched, cleanup and refactor those
2995 multiclass basic_sse12_fp_binop_p<bits<8> opc, string OpcodeStr,
2996 SDNode OpNode, SizeItins itins> {
2997 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode,
2998 VR128, v4f32, f128mem, loadv4f32,
2999 SSEPackedSingle, itins.s, 0>, PS, VEX_4V;
3000 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode,
3001 VR128, v2f64, f128mem, loadv2f64,
3002 SSEPackedDouble, itins.d, 0>, PD, VEX_4V;
3004 defm V#NAME#PSY : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"),
3005 OpNode, VR256, v8f32, f256mem, loadv8f32,
3006 SSEPackedSingle, itins.s, 0>, PS, VEX_4V, VEX_L;
3007 defm V#NAME#PDY : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"),
3008 OpNode, VR256, v4f64, f256mem, loadv4f64,
3009 SSEPackedDouble, itins.d, 0>, PD, VEX_4V, VEX_L;
3011 let Constraints = "$src1 = $dst" in {
3012 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode, VR128,
3013 v4f32, f128mem, memopv4f32, SSEPackedSingle,
3015 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode, VR128,
3016 v2f64, f128mem, memopv2f64, SSEPackedDouble,
3021 multiclass basic_sse12_fp_binop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
3023 defm V#NAME#SS : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "ss"),
3024 OpNode, FR32, f32mem, itins.s, 0>, XS, VEX_4V, VEX_LIG;
3025 defm V#NAME#SD : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "sd"),
3026 OpNode, FR64, f64mem, itins.d, 0>, XD, VEX_4V, VEX_LIG;
3028 let Constraints = "$src1 = $dst" in {
3029 defm SS : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "ss"),
3030 OpNode, FR32, f32mem, itins.s>, XS;
3031 defm SD : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "sd"),
3032 OpNode, FR64, f64mem, itins.d>, XD;
3036 multiclass basic_sse12_fp_binop_s_int<bits<8> opc, string OpcodeStr,
3038 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
3039 !strconcat(OpcodeStr, "ss"), "", "_ss", ssmem, sse_load_f32,
3040 itins.s, 0>, XS, VEX_4V, VEX_LIG;
3041 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
3042 !strconcat(OpcodeStr, "sd"), "2", "_sd", sdmem, sse_load_f64,
3043 itins.d, 0>, XD, VEX_4V, VEX_LIG;
3045 let Constraints = "$src1 = $dst" in {
3046 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
3047 !strconcat(OpcodeStr, "ss"), "", "_ss", ssmem, sse_load_f32,
3049 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
3050 !strconcat(OpcodeStr, "sd"), "2", "_sd", sdmem, sse_load_f64,
3055 // Binary Arithmetic instructions
3056 defm ADD : basic_sse12_fp_binop_p<0x58, "add", fadd, SSE_ALU_ITINS_P>,
3057 basic_sse12_fp_binop_s<0x58, "add", fadd, SSE_ALU_ITINS_S>,
3058 basic_sse12_fp_binop_s_int<0x58, "add", SSE_ALU_ITINS_S>;
3059 defm MUL : basic_sse12_fp_binop_p<0x59, "mul", fmul, SSE_MUL_ITINS_P>,
3060 basic_sse12_fp_binop_s<0x59, "mul", fmul, SSE_MUL_ITINS_S>,
3061 basic_sse12_fp_binop_s_int<0x59, "mul", SSE_MUL_ITINS_S>;
3062 let isCommutable = 0 in {
3063 defm SUB : basic_sse12_fp_binop_p<0x5C, "sub", fsub, SSE_ALU_ITINS_P>,
3064 basic_sse12_fp_binop_s<0x5C, "sub", fsub, SSE_ALU_ITINS_S>,
3065 basic_sse12_fp_binop_s_int<0x5C, "sub", SSE_ALU_ITINS_S>;
3066 defm DIV : basic_sse12_fp_binop_p<0x5E, "div", fdiv, SSE_DIV_ITINS_P>,
3067 basic_sse12_fp_binop_s<0x5E, "div", fdiv, SSE_DIV_ITINS_S>,
3068 basic_sse12_fp_binop_s_int<0x5E, "div", SSE_DIV_ITINS_S>;
3069 defm MAX : basic_sse12_fp_binop_p<0x5F, "max", X86fmax, SSE_ALU_ITINS_P>,
3070 basic_sse12_fp_binop_s<0x5F, "max", X86fmax, SSE_ALU_ITINS_S>,
3071 basic_sse12_fp_binop_s_int<0x5F, "max", SSE_ALU_ITINS_S>;
3072 defm MIN : basic_sse12_fp_binop_p<0x5D, "min", X86fmin, SSE_ALU_ITINS_P>,
3073 basic_sse12_fp_binop_s<0x5D, "min", X86fmin, SSE_ALU_ITINS_S>,
3074 basic_sse12_fp_binop_s_int<0x5D, "min", SSE_ALU_ITINS_S>;
3077 let isCodeGenOnly = 1 in {
3078 defm MAXC: basic_sse12_fp_binop_p<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_P>,
3079 basic_sse12_fp_binop_s<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_S>;
3080 defm MINC: basic_sse12_fp_binop_p<0x5D, "min", X86fminc, SSE_ALU_ITINS_P>,
3081 basic_sse12_fp_binop_s<0x5D, "min", X86fminc, SSE_ALU_ITINS_S>;
3084 // Patterns used to select SSE scalar fp arithmetic instructions from
3085 // a scalar fp operation followed by a blend.
3087 // These patterns know, for example, how to select an ADDSS from a
3088 // float add plus vector insert.
3090 // The effect is that the backend no longer emits unnecessary vector
3091 // insert instructions immediately after SSE scalar fp instructions
3092 // like addss or mulss.
3094 // For example, given the following code:
3095 // __m128 foo(__m128 A, __m128 B) {
3100 // previously we generated:
3101 // addss %xmm0, %xmm1
3102 // movss %xmm1, %xmm0
3105 // addss %xmm1, %xmm0
3107 let Predicates = [UseSSE1] in {
3108 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector (fadd
3109 (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3111 (ADDSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3112 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector (fsub
3113 (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3115 (SUBSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3116 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector (fmul
3117 (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3119 (MULSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3120 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector (fdiv
3121 (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3123 (DIVSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3126 let Predicates = [UseSSE2] in {
3127 // SSE2 patterns to select scalar double-precision fp arithmetic instructions
3129 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fadd
3130 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3132 (ADDSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3133 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fsub
3134 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3136 (SUBSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3137 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fmul
3138 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3140 (MULSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3141 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fdiv
3142 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3144 (DIVSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3147 let Predicates = [UseSSE41] in {
3148 // If the subtarget has SSE4.1 but not AVX, the vector insert
3149 // instruction is lowered into a X86insertps rather than a X86Movss.
3150 // When selecting SSE scalar single-precision fp arithmetic instructions,
3151 // make sure that we correctly match the X86insertps.
3153 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3154 (fadd (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3155 FR32:$src))), (iPTR 0))),
3156 (ADDSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3157 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3158 (fsub (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3159 FR32:$src))), (iPTR 0))),
3160 (SUBSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3161 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3162 (fmul (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3163 FR32:$src))), (iPTR 0))),
3164 (MULSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3165 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3166 (fdiv (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3167 FR32:$src))), (iPTR 0))),
3168 (DIVSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3171 let Predicates = [HasAVX] in {
3172 // The following patterns select AVX Scalar single/double precision fp
3173 // arithmetic instructions.
3175 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fadd
3176 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3178 (VADDSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3179 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fsub
3180 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3182 (VSUBSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3183 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fmul
3184 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3186 (VMULSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3187 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fdiv
3188 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3190 (VDIVSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3191 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3192 (fadd (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3193 FR32:$src))), (iPTR 0))),
3194 (VADDSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3195 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3196 (fsub (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3197 FR32:$src))), (iPTR 0))),
3198 (VSUBSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3199 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3200 (fmul (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3201 FR32:$src))), (iPTR 0))),
3202 (VMULSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3203 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3204 (fdiv (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3205 FR32:$src))), (iPTR 0))),
3206 (VDIVSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3209 // Patterns used to select SSE scalar fp arithmetic instructions from
3210 // a vector packed single/double fp operation followed by a vector insert.
3212 // The effect is that the backend converts the packed fp instruction
3213 // followed by a vector insert into a single SSE scalar fp instruction.
3215 // For example, given the following code:
3216 // __m128 foo(__m128 A, __m128 B) {
3217 // __m128 C = A + B;
3218 // return (__m128) {c[0], a[1], a[2], a[3]};
3221 // previously we generated:
3222 // addps %xmm0, %xmm1
3223 // movss %xmm1, %xmm0
3226 // addss %xmm1, %xmm0
3228 let Predicates = [UseSSE1] in {
3229 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3230 (fadd (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3231 (ADDSSrr_Int v4f32:$dst, v4f32:$src)>;
3232 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3233 (fsub (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3234 (SUBSSrr_Int v4f32:$dst, v4f32:$src)>;
3235 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3236 (fmul (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3237 (MULSSrr_Int v4f32:$dst, v4f32:$src)>;
3238 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3239 (fdiv (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3240 (DIVSSrr_Int v4f32:$dst, v4f32:$src)>;
3243 let Predicates = [UseSSE2] in {
3244 // SSE2 patterns to select scalar double-precision fp arithmetic instructions
3245 // from a packed double-precision fp instruction plus movsd.
3247 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3248 (fadd (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3249 (ADDSDrr_Int v2f64:$dst, v2f64:$src)>;
3250 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3251 (fsub (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3252 (SUBSDrr_Int v2f64:$dst, v2f64:$src)>;
3253 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3254 (fmul (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3255 (MULSDrr_Int v2f64:$dst, v2f64:$src)>;
3256 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3257 (fdiv (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3258 (DIVSDrr_Int v2f64:$dst, v2f64:$src)>;
3261 let Predicates = [HasAVX] in {
3262 // The following patterns select AVX Scalar single/double precision fp
3263 // arithmetic instructions from a packed single precision fp instruction
3264 // plus movss/movsd.
3266 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3267 (fadd (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3268 (VADDSSrr_Int v4f32:$dst, v4f32:$src)>;
3269 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3270 (fsub (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3271 (VSUBSSrr_Int v4f32:$dst, v4f32:$src)>;
3272 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3273 (fmul (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3274 (VMULSSrr_Int v4f32:$dst, v4f32:$src)>;
3275 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3276 (fdiv (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3277 (VDIVSSrr_Int v4f32:$dst, v4f32:$src)>;
3278 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3279 (fadd (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3280 (VADDSDrr_Int v2f64:$dst, v2f64:$src)>;
3281 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3282 (fsub (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3283 (VSUBSDrr_Int v2f64:$dst, v2f64:$src)>;
3284 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3285 (fmul (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3286 (VMULSDrr_Int v2f64:$dst, v2f64:$src)>;
3287 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3288 (fdiv (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3289 (VDIVSDrr_Int v2f64:$dst, v2f64:$src)>;
3293 /// In addition, we also have a special variant of the scalar form here to
3294 /// represent the associated intrinsic operation. This form is unlike the
3295 /// plain scalar form, in that it takes an entire vector (instead of a
3296 /// scalar) and leaves the top elements undefined.
3298 /// And, we have a special variant form for a full-vector intrinsic form.
3300 let Sched = WriteFSqrt in {
3301 def SSE_SQRTPS : OpndItins<
3302 IIC_SSE_SQRTPS_RR, IIC_SSE_SQRTPS_RM
3305 def SSE_SQRTSS : OpndItins<
3306 IIC_SSE_SQRTSS_RR, IIC_SSE_SQRTSS_RM
3309 def SSE_SQRTPD : OpndItins<
3310 IIC_SSE_SQRTPD_RR, IIC_SSE_SQRTPD_RM
3313 def SSE_SQRTSD : OpndItins<
3314 IIC_SSE_SQRTSD_RR, IIC_SSE_SQRTSD_RM
3318 let Sched = WriteFRsqrt in {
3319 def SSE_RSQRTPS : OpndItins<
3320 IIC_SSE_RSQRTPS_RR, IIC_SSE_RSQRTPS_RM
3323 def SSE_RSQRTSS : OpndItins<
3324 IIC_SSE_RSQRTSS_RR, IIC_SSE_RSQRTSS_RM
3328 let Sched = WriteFRcp in {
3329 def SSE_RCPP : OpndItins<
3330 IIC_SSE_RCPP_RR, IIC_SSE_RCPP_RM
3333 def SSE_RCPS : OpndItins<
3334 IIC_SSE_RCPS_RR, IIC_SSE_RCPS_RM
3338 /// sse1_fp_unop_s - SSE1 unops in scalar form.
3339 multiclass sse1_fp_unop_s<bits<8> opc, string OpcodeStr,
3340 SDNode OpNode, Intrinsic F32Int, OpndItins itins> {
3341 let Predicates = [HasAVX], hasSideEffects = 0 in {
3342 def V#NAME#SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst),
3343 (ins FR32:$src1, FR32:$src2),
3344 !strconcat("v", OpcodeStr,
3345 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3346 []>, VEX_4V, VEX_LIG, Sched<[itins.Sched]>;
3347 let mayLoad = 1 in {
3348 def V#NAME#SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
3349 (ins FR32:$src1,f32mem:$src2),
3350 !strconcat("v", OpcodeStr,
3351 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3352 []>, VEX_4V, VEX_LIG,
3353 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3354 let isCodeGenOnly = 1 in
3355 def V#NAME#SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
3356 (ins VR128:$src1, ssmem:$src2),
3357 !strconcat("v", OpcodeStr,
3358 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3359 []>, VEX_4V, VEX_LIG,
3360 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3364 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
3365 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
3366 [(set FR32:$dst, (OpNode FR32:$src))]>, Sched<[itins.Sched]>;
3367 // For scalar unary operations, fold a load into the operation
3368 // only in OptForSize mode. It eliminates an instruction, but it also
3369 // eliminates a whole-register clobber (the load), so it introduces a
3370 // partial register update condition.
3371 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
3372 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
3373 [(set FR32:$dst, (OpNode (load addr:$src)))], itins.rm>, XS,
3374 Requires<[UseSSE1, OptForSize]>, Sched<[itins.Sched.Folded]>;
3375 let isCodeGenOnly = 1 in {
3376 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3377 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
3378 [(set VR128:$dst, (F32Int VR128:$src))], itins.rr>,
3379 Sched<[itins.Sched]>;
3380 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
3381 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
3382 [(set VR128:$dst, (F32Int sse_load_f32:$src))], itins.rm>,
3383 Sched<[itins.Sched.Folded]>;
3387 /// sse1_fp_unop_s_rw - SSE1 unops where vector form has a read-write operand.
3388 multiclass sse1_fp_unop_rw<bits<8> opc, string OpcodeStr, SDNode OpNode,
3390 let Predicates = [HasAVX], hasSideEffects = 0 in {
3391 def V#NAME#SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst),
3392 (ins FR32:$src1, FR32:$src2),
3393 !strconcat("v", OpcodeStr,
3394 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3395 []>, VEX_4V, VEX_LIG, Sched<[itins.Sched]>;
3396 let mayLoad = 1 in {
3397 def V#NAME#SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
3398 (ins FR32:$src1,f32mem:$src2),
3399 !strconcat("v", OpcodeStr,
3400 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3401 []>, VEX_4V, VEX_LIG,
3402 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3403 let isCodeGenOnly = 1 in
3404 def V#NAME#SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
3405 (ins VR128:$src1, ssmem:$src2),
3406 !strconcat("v", OpcodeStr,
3407 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3408 []>, VEX_4V, VEX_LIG,
3409 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3413 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
3414 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
3415 [(set FR32:$dst, (OpNode FR32:$src))]>, Sched<[itins.Sched]>;
3416 // For scalar unary operations, fold a load into the operation
3417 // only in OptForSize mode. It eliminates an instruction, but it also
3418 // eliminates a whole-register clobber (the load), so it introduces a
3419 // partial register update condition.
3420 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
3421 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
3422 [(set FR32:$dst, (OpNode (load addr:$src)))], itins.rm>, XS,
3423 Requires<[UseSSE1, OptForSize]>, Sched<[itins.Sched.Folded]>;
3424 let isCodeGenOnly = 1, Constraints = "$src1 = $dst" in {
3425 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
3426 (ins VR128:$src1, VR128:$src2),
3427 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
3428 [], itins.rr>, Sched<[itins.Sched]>;
3429 let mayLoad = 1, hasSideEffects = 0 in
3430 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
3431 (ins VR128:$src1, ssmem:$src2),
3432 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
3433 [], itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3437 /// sse1_fp_unop_p - SSE1 unops in packed form.
3438 multiclass sse1_fp_unop_p<bits<8> opc, string OpcodeStr, SDNode OpNode,
3440 let Predicates = [HasAVX] in {
3441 def V#NAME#PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3442 !strconcat("v", OpcodeStr,
3443 "ps\t{$src, $dst|$dst, $src}"),
3444 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))],
3445 itins.rr>, VEX, Sched<[itins.Sched]>;
3446 def V#NAME#PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3447 !strconcat("v", OpcodeStr,
3448 "ps\t{$src, $dst|$dst, $src}"),
3449 [(set VR128:$dst, (OpNode (loadv4f32 addr:$src)))],
3450 itins.rm>, VEX, Sched<[itins.Sched.Folded]>;
3451 def V#NAME#PSYr : PSI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3452 !strconcat("v", OpcodeStr,
3453 "ps\t{$src, $dst|$dst, $src}"),
3454 [(set VR256:$dst, (v8f32 (OpNode VR256:$src)))],
3455 itins.rr>, VEX, VEX_L, Sched<[itins.Sched]>;
3456 def V#NAME#PSYm : PSI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
3457 !strconcat("v", OpcodeStr,
3458 "ps\t{$src, $dst|$dst, $src}"),
3459 [(set VR256:$dst, (OpNode (loadv8f32 addr:$src)))],
3460 itins.rm>, VEX, VEX_L, Sched<[itins.Sched.Folded]>;
3463 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3464 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3465 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))], itins.rr>,
3466 Sched<[itins.Sched]>;
3467 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3468 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3469 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))], itins.rm>,
3470 Sched<[itins.Sched.Folded]>;
3473 /// sse1_fp_unop_p_int - SSE1 intrinsics unops in packed forms.
3474 multiclass sse1_fp_unop_p_int<bits<8> opc, string OpcodeStr,
3475 Intrinsic V4F32Int, Intrinsic V8F32Int,
3477 let isCodeGenOnly = 1 in {
3478 let Predicates = [HasAVX] in {
3479 def V#NAME#PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3480 !strconcat("v", OpcodeStr,
3481 "ps\t{$src, $dst|$dst, $src}"),
3482 [(set VR128:$dst, (V4F32Int VR128:$src))],
3483 itins.rr>, VEX, Sched<[itins.Sched]>;
3484 def V#NAME#PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3485 !strconcat("v", OpcodeStr,
3486 "ps\t{$src, $dst|$dst, $src}"),
3487 [(set VR128:$dst, (V4F32Int (loadv4f32 addr:$src)))],
3488 itins.rm>, VEX, Sched<[itins.Sched.Folded]>;
3489 def V#NAME#PSYr_Int : PSI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3490 !strconcat("v", OpcodeStr,
3491 "ps\t{$src, $dst|$dst, $src}"),
3492 [(set VR256:$dst, (V8F32Int VR256:$src))],
3493 itins.rr>, VEX, VEX_L, Sched<[itins.Sched]>;
3494 def V#NAME#PSYm_Int : PSI<opc, MRMSrcMem, (outs VR256:$dst),
3496 !strconcat("v", OpcodeStr,
3497 "ps\t{$src, $dst|$dst, $src}"),
3498 [(set VR256:$dst, (V8F32Int (loadv8f32 addr:$src)))],
3499 itins.rm>, VEX, VEX_L, Sched<[itins.Sched.Folded]>;
3502 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3503 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3504 [(set VR128:$dst, (V4F32Int VR128:$src))],
3505 itins.rr>, Sched<[itins.Sched]>;
3506 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3507 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3508 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))],
3509 itins.rm>, Sched<[itins.Sched.Folded]>;
3510 } // isCodeGenOnly = 1
3513 /// sse2_fp_unop_s - SSE2 unops in scalar form.
3514 multiclass sse2_fp_unop_s<bits<8> opc, string OpcodeStr,
3515 SDNode OpNode, Intrinsic F64Int, OpndItins itins> {
3516 let Predicates = [HasAVX], hasSideEffects = 0 in {
3517 def V#NAME#SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst),
3518 (ins FR64:$src1, FR64:$src2),
3519 !strconcat("v", OpcodeStr,
3520 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3521 []>, VEX_4V, VEX_LIG, Sched<[itins.Sched]>;
3522 let mayLoad = 1 in {
3523 def V#NAME#SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
3524 (ins FR64:$src1,f64mem:$src2),
3525 !strconcat("v", OpcodeStr,
3526 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3527 []>, VEX_4V, VEX_LIG,
3528 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3529 let isCodeGenOnly = 1 in
3530 def V#NAME#SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
3531 (ins VR128:$src1, sdmem:$src2),
3532 !strconcat("v", OpcodeStr,
3533 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3534 []>, VEX_4V, VEX_LIG,
3535 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3539 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
3540 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
3541 [(set FR64:$dst, (OpNode FR64:$src))], itins.rr>,
3542 Sched<[itins.Sched]>;
3543 // See the comments in sse1_fp_unop_s for why this is OptForSize.
3544 def SDm : I<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
3545 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
3546 [(set FR64:$dst, (OpNode (load addr:$src)))], itins.rm>, XD,
3547 Requires<[UseSSE2, OptForSize]>, Sched<[itins.Sched.Folded]>;
3548 let isCodeGenOnly = 1 in {
3549 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3550 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
3551 [(set VR128:$dst, (F64Int VR128:$src))], itins.rr>,
3552 Sched<[itins.Sched]>;
3553 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
3554 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
3555 [(set VR128:$dst, (F64Int sse_load_f64:$src))], itins.rm>,
3556 Sched<[itins.Sched.Folded]>;
3560 /// sse2_fp_unop_p - SSE2 unops in vector forms.
3561 multiclass sse2_fp_unop_p<bits<8> opc, string OpcodeStr,
3562 SDNode OpNode, OpndItins itins> {
3563 let Predicates = [HasAVX] in {
3564 def V#NAME#PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3565 !strconcat("v", OpcodeStr,
3566 "pd\t{$src, $dst|$dst, $src}"),
3567 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))],
3568 itins.rr>, VEX, Sched<[itins.Sched]>;
3569 def V#NAME#PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3570 !strconcat("v", OpcodeStr,
3571 "pd\t{$src, $dst|$dst, $src}"),
3572 [(set VR128:$dst, (OpNode (loadv2f64 addr:$src)))],
3573 itins.rm>, VEX, Sched<[itins.Sched.Folded]>;
3574 def V#NAME#PDYr : PDI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3575 !strconcat("v", OpcodeStr,
3576 "pd\t{$src, $dst|$dst, $src}"),
3577 [(set VR256:$dst, (v4f64 (OpNode VR256:$src)))],
3578 itins.rr>, VEX, VEX_L, Sched<[itins.Sched]>;
3579 def V#NAME#PDYm : PDI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
3580 !strconcat("v", OpcodeStr,
3581 "pd\t{$src, $dst|$dst, $src}"),
3582 [(set VR256:$dst, (OpNode (loadv4f64 addr:$src)))],
3583 itins.rm>, VEX, VEX_L, Sched<[itins.Sched.Folded]>;
3586 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3587 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
3588 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))], itins.rr>,
3589 Sched<[itins.Sched]>;
3590 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3591 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
3592 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))], itins.rm>,
3593 Sched<[itins.Sched.Folded]>;
3597 defm SQRT : sse1_fp_unop_s<0x51, "sqrt", fsqrt, int_x86_sse_sqrt_ss,
3599 sse1_fp_unop_p<0x51, "sqrt", fsqrt, SSE_SQRTPS>,
3600 sse2_fp_unop_s<0x51, "sqrt", fsqrt, int_x86_sse2_sqrt_sd,
3602 sse2_fp_unop_p<0x51, "sqrt", fsqrt, SSE_SQRTPD>;
3604 // Reciprocal approximations. Note that these typically require refinement
3605 // in order to obtain suitable precision.
3606 defm RSQRT : sse1_fp_unop_rw<0x52, "rsqrt", X86frsqrt, SSE_RSQRTSS>,
3607 sse1_fp_unop_p<0x52, "rsqrt", X86frsqrt, SSE_RSQRTPS>,
3608 sse1_fp_unop_p_int<0x52, "rsqrt", int_x86_sse_rsqrt_ps,
3609 int_x86_avx_rsqrt_ps_256, SSE_RSQRTPS>;
3610 defm RCP : sse1_fp_unop_rw<0x53, "rcp", X86frcp, SSE_RCPS>,
3611 sse1_fp_unop_p<0x53, "rcp", X86frcp, SSE_RCPP>,
3612 sse1_fp_unop_p_int<0x53, "rcp", int_x86_sse_rcp_ps,
3613 int_x86_avx_rcp_ps_256, SSE_RCPP>;
3615 let Predicates = [UseAVX] in {
3616 def : Pat<(f32 (fsqrt FR32:$src)),
3617 (VSQRTSSr (f32 (IMPLICIT_DEF)), FR32:$src)>, Requires<[HasAVX]>;
3618 def : Pat<(f32 (fsqrt (load addr:$src))),
3619 (VSQRTSSm (f32 (IMPLICIT_DEF)), addr:$src)>,
3620 Requires<[HasAVX, OptForSize]>;
3621 def : Pat<(f64 (fsqrt FR64:$src)),
3622 (VSQRTSDr (f64 (IMPLICIT_DEF)), FR64:$src)>, Requires<[HasAVX]>;
3623 def : Pat<(f64 (fsqrt (load addr:$src))),
3624 (VSQRTSDm (f64 (IMPLICIT_DEF)), addr:$src)>,
3625 Requires<[HasAVX, OptForSize]>;
3627 def : Pat<(f32 (X86frsqrt FR32:$src)),
3628 (VRSQRTSSr (f32 (IMPLICIT_DEF)), FR32:$src)>, Requires<[HasAVX]>;
3629 def : Pat<(f32 (X86frsqrt (load addr:$src))),
3630 (VRSQRTSSm (f32 (IMPLICIT_DEF)), addr:$src)>,
3631 Requires<[HasAVX, OptForSize]>;
3633 def : Pat<(f32 (X86frcp FR32:$src)),
3634 (VRCPSSr (f32 (IMPLICIT_DEF)), FR32:$src)>, Requires<[HasAVX]>;
3635 def : Pat<(f32 (X86frcp (load addr:$src))),
3636 (VRCPSSm (f32 (IMPLICIT_DEF)), addr:$src)>,
3637 Requires<[HasAVX, OptForSize]>;
3639 let Predicates = [UseAVX] in {
3640 def : Pat<(int_x86_sse_sqrt_ss VR128:$src),
3641 (COPY_TO_REGCLASS (VSQRTSSr (f32 (IMPLICIT_DEF)),
3642 (COPY_TO_REGCLASS VR128:$src, FR32)),
3644 def : Pat<(int_x86_sse_sqrt_ss sse_load_f32:$src),
3645 (VSQRTSSm_Int (v4f32 (IMPLICIT_DEF)), sse_load_f32:$src)>;
3647 def : Pat<(int_x86_sse2_sqrt_sd VR128:$src),
3648 (COPY_TO_REGCLASS (VSQRTSDr (f64 (IMPLICIT_DEF)),
3649 (COPY_TO_REGCLASS VR128:$src, FR64)),
3651 def : Pat<(int_x86_sse2_sqrt_sd sse_load_f64:$src),
3652 (VSQRTSDm_Int (v2f64 (IMPLICIT_DEF)), sse_load_f64:$src)>;
3655 let Predicates = [HasAVX] in {
3656 def : Pat<(int_x86_sse_rsqrt_ss VR128:$src),
3657 (COPY_TO_REGCLASS (VRSQRTSSr (f32 (IMPLICIT_DEF)),
3658 (COPY_TO_REGCLASS VR128:$src, FR32)),
3660 def : Pat<(int_x86_sse_rsqrt_ss sse_load_f32:$src),
3661 (VRSQRTSSm_Int (v4f32 (IMPLICIT_DEF)), sse_load_f32:$src)>;
3663 def : Pat<(int_x86_sse_rcp_ss VR128:$src),
3664 (COPY_TO_REGCLASS (VRCPSSr (f32 (IMPLICIT_DEF)),
3665 (COPY_TO_REGCLASS VR128:$src, FR32)),
3667 def : Pat<(int_x86_sse_rcp_ss sse_load_f32:$src),
3668 (VRCPSSm_Int (v4f32 (IMPLICIT_DEF)), sse_load_f32:$src)>;
3671 // Reciprocal approximations. Note that these typically require refinement
3672 // in order to obtain suitable precision.
3673 let Predicates = [UseSSE1] in {
3674 def : Pat<(int_x86_sse_rsqrt_ss VR128:$src),
3675 (RSQRTSSr_Int VR128:$src, VR128:$src)>;
3676 def : Pat<(int_x86_sse_rcp_ss VR128:$src),
3677 (RCPSSr_Int VR128:$src, VR128:$src)>;
3680 // There is no f64 version of the reciprocal approximation instructions.
3682 //===----------------------------------------------------------------------===//
3683 // SSE 1 & 2 - Non-temporal stores
3684 //===----------------------------------------------------------------------===//
3686 let AddedComplexity = 400 in { // Prefer non-temporal versions
3687 let SchedRW = [WriteStore] in {
3688 let Predicates = [HasAVX, NoVLX] in {
3689 def VMOVNTPSmr : VPSI<0x2B, MRMDestMem, (outs),
3690 (ins f128mem:$dst, VR128:$src),
3691 "movntps\t{$src, $dst|$dst, $src}",
3692 [(alignednontemporalstore (v4f32 VR128:$src),
3694 IIC_SSE_MOVNT>, VEX;
3695 def VMOVNTPDmr : VPDI<0x2B, MRMDestMem, (outs),
3696 (ins f128mem:$dst, VR128:$src),
3697 "movntpd\t{$src, $dst|$dst, $src}",
3698 [(alignednontemporalstore (v2f64 VR128:$src),
3700 IIC_SSE_MOVNT>, VEX;
3702 let ExeDomain = SSEPackedInt in
3703 def VMOVNTDQmr : VPDI<0xE7, MRMDestMem, (outs),
3704 (ins f128mem:$dst, VR128:$src),
3705 "movntdq\t{$src, $dst|$dst, $src}",
3706 [(alignednontemporalstore (v2i64 VR128:$src),
3708 IIC_SSE_MOVNT>, VEX;
3710 def VMOVNTPSYmr : VPSI<0x2B, MRMDestMem, (outs),
3711 (ins f256mem:$dst, VR256:$src),
3712 "movntps\t{$src, $dst|$dst, $src}",
3713 [(alignednontemporalstore (v8f32 VR256:$src),
3715 IIC_SSE_MOVNT>, VEX, VEX_L;
3716 def VMOVNTPDYmr : VPDI<0x2B, MRMDestMem, (outs),
3717 (ins f256mem:$dst, VR256:$src),
3718 "movntpd\t{$src, $dst|$dst, $src}",
3719 [(alignednontemporalstore (v4f64 VR256:$src),
3721 IIC_SSE_MOVNT>, VEX, VEX_L;
3722 let ExeDomain = SSEPackedInt in
3723 def VMOVNTDQYmr : VPDI<0xE7, MRMDestMem, (outs),
3724 (ins f256mem:$dst, VR256:$src),
3725 "movntdq\t{$src, $dst|$dst, $src}",
3726 [(alignednontemporalstore (v4i64 VR256:$src),
3728 IIC_SSE_MOVNT>, VEX, VEX_L;
3731 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3732 "movntps\t{$src, $dst|$dst, $src}",
3733 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)],
3735 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3736 "movntpd\t{$src, $dst|$dst, $src}",
3737 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)],
3740 let ExeDomain = SSEPackedInt in
3741 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3742 "movntdq\t{$src, $dst|$dst, $src}",
3743 [(alignednontemporalstore (v2i64 VR128:$src), addr:$dst)],
3746 // There is no AVX form for instructions below this point
3747 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
3748 "movnti{l}\t{$src, $dst|$dst, $src}",
3749 [(nontemporalstore (i32 GR32:$src), addr:$dst)],
3751 PS, Requires<[HasSSE2]>;
3752 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
3753 "movnti{q}\t{$src, $dst|$dst, $src}",
3754 [(nontemporalstore (i64 GR64:$src), addr:$dst)],
3756 PS, Requires<[HasSSE2]>;
3757 } // SchedRW = [WriteStore]
3759 } // AddedComplexity
3761 //===----------------------------------------------------------------------===//
3762 // SSE 1 & 2 - Prefetch and memory fence
3763 //===----------------------------------------------------------------------===//
3765 // Prefetch intrinsic.
3766 let Predicates = [HasSSE1], SchedRW = [WriteLoad] in {
3767 def PREFETCHT0 : I<0x18, MRM1m, (outs), (ins i8mem:$src),
3768 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3), (i32 1))],
3769 IIC_SSE_PREFETCH>, TB;
3770 def PREFETCHT1 : I<0x18, MRM2m, (outs), (ins i8mem:$src),
3771 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2), (i32 1))],
3772 IIC_SSE_PREFETCH>, TB;
3773 def PREFETCHT2 : I<0x18, MRM3m, (outs), (ins i8mem:$src),
3774 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1), (i32 1))],
3775 IIC_SSE_PREFETCH>, TB;
3776 def PREFETCHNTA : I<0x18, MRM0m, (outs), (ins i8mem:$src),
3777 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0), (i32 1))],
3778 IIC_SSE_PREFETCH>, TB;
3781 // FIXME: How should flush instruction be modeled?
3782 let SchedRW = [WriteLoad] in {
3784 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
3785 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)],
3786 IIC_SSE_PREFETCH>, TB, Requires<[HasSSE2]>;
3789 let SchedRW = [WriteNop] in {
3790 // Pause. This "instruction" is encoded as "rep; nop", so even though it
3791 // was introduced with SSE2, it's backward compatible.
3792 def PAUSE : I<0x90, RawFrm, (outs), (ins),
3793 "pause", [(int_x86_sse2_pause)], IIC_SSE_PAUSE>,
3794 OBXS, Requires<[HasSSE2]>;
3797 let SchedRW = [WriteFence] in {
3798 // Load, store, and memory fence
3799 def SFENCE : I<0xAE, MRM_F8, (outs), (ins),
3800 "sfence", [(int_x86_sse_sfence)], IIC_SSE_SFENCE>,
3801 TB, Requires<[HasSSE1]>;
3802 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
3803 "lfence", [(int_x86_sse2_lfence)], IIC_SSE_LFENCE>,
3804 TB, Requires<[HasSSE2]>;
3805 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
3806 "mfence", [(int_x86_sse2_mfence)], IIC_SSE_MFENCE>,
3807 TB, Requires<[HasSSE2]>;
3810 def : Pat<(X86SFence), (SFENCE)>;
3811 def : Pat<(X86LFence), (LFENCE)>;
3812 def : Pat<(X86MFence), (MFENCE)>;
3814 //===----------------------------------------------------------------------===//
3815 // SSE 1 & 2 - Load/Store XCSR register
3816 //===----------------------------------------------------------------------===//
3818 def VLDMXCSR : VPSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
3819 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)],
3820 IIC_SSE_LDMXCSR>, VEX, Sched<[WriteLoad]>;
3821 def VSTMXCSR : VPSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
3822 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)],
3823 IIC_SSE_STMXCSR>, VEX, Sched<[WriteStore]>;
3825 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
3826 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)],
3827 IIC_SSE_LDMXCSR>, Sched<[WriteLoad]>;
3828 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
3829 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)],
3830 IIC_SSE_STMXCSR>, Sched<[WriteStore]>;
3832 //===---------------------------------------------------------------------===//
3833 // SSE2 - Move Aligned/Unaligned Packed Integer Instructions
3834 //===---------------------------------------------------------------------===//
3836 let ExeDomain = SSEPackedInt in { // SSE integer instructions
3838 let neverHasSideEffects = 1, SchedRW = [WriteMove] in {
3839 def VMOVDQArr : VPDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3840 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>,
3842 def VMOVDQAYrr : VPDI<0x6F, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3843 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>,
3845 def VMOVDQUrr : VSSI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3846 "movdqu\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVU_P_RR>,
3848 def VMOVDQUYrr : VSSI<0x6F, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3849 "movdqu\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVU_P_RR>,
3854 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
3855 SchedRW = [WriteMove] in {
3856 def VMOVDQArr_REV : VPDI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3857 "movdqa\t{$src, $dst|$dst, $src}", [],
3860 def VMOVDQAYrr_REV : VPDI<0x7F, MRMDestReg, (outs VR256:$dst), (ins VR256:$src),
3861 "movdqa\t{$src, $dst|$dst, $src}", [],
3862 IIC_SSE_MOVA_P_RR>, VEX, VEX_L;
3863 def VMOVDQUrr_REV : VSSI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3864 "movdqu\t{$src, $dst|$dst, $src}", [],
3867 def VMOVDQUYrr_REV : VSSI<0x7F, MRMDestReg, (outs VR256:$dst), (ins VR256:$src),
3868 "movdqu\t{$src, $dst|$dst, $src}", [],
3869 IIC_SSE_MOVU_P_RR>, VEX, VEX_L;
3872 let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1,
3873 neverHasSideEffects = 1, SchedRW = [WriteLoad] in {
3874 def VMOVDQArm : VPDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3875 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RM>,
3877 def VMOVDQAYrm : VPDI<0x6F, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
3878 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RM>,
3880 let Predicates = [HasAVX] in {
3881 def VMOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3882 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_RM>,
3884 def VMOVDQUYrm : I<0x6F, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
3885 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_RM>,
3890 let mayStore = 1, neverHasSideEffects = 1, SchedRW = [WriteStore] in {
3891 def VMOVDQAmr : VPDI<0x7F, MRMDestMem, (outs),
3892 (ins i128mem:$dst, VR128:$src),
3893 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_MR>,
3895 def VMOVDQAYmr : VPDI<0x7F, MRMDestMem, (outs),
3896 (ins i256mem:$dst, VR256:$src),
3897 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_MR>,
3899 let Predicates = [HasAVX] in {
3900 def VMOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3901 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_MR>,
3903 def VMOVDQUYmr : I<0x7F, MRMDestMem, (outs), (ins i256mem:$dst, VR256:$src),
3904 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_MR>,
3909 let SchedRW = [WriteMove] in {
3910 let neverHasSideEffects = 1 in
3911 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3912 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>;
3914 def MOVDQUrr : I<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3915 "movdqu\t{$src, $dst|$dst, $src}",
3916 [], IIC_SSE_MOVU_P_RR>, XS, Requires<[UseSSE2]>;
3919 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in {
3920 def MOVDQArr_REV : PDI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3921 "movdqa\t{$src, $dst|$dst, $src}", [],
3924 def MOVDQUrr_REV : I<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3925 "movdqu\t{$src, $dst|$dst, $src}",
3926 [], IIC_SSE_MOVU_P_RR>, XS, Requires<[UseSSE2]>;
3930 let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1,
3931 neverHasSideEffects = 1, SchedRW = [WriteLoad] in {
3932 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3933 "movdqa\t{$src, $dst|$dst, $src}",
3934 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/],
3936 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3937 "movdqu\t{$src, $dst|$dst, $src}",
3938 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/],
3940 XS, Requires<[UseSSE2]>;
3943 let mayStore = 1, neverHasSideEffects = 1, SchedRW = [WriteStore] in {
3944 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3945 "movdqa\t{$src, $dst|$dst, $src}",
3946 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/],
3948 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3949 "movdqu\t{$src, $dst|$dst, $src}",
3950 [/*(store (v2i64 VR128:$src), addr:$dst)*/],
3952 XS, Requires<[UseSSE2]>;
3955 } // ExeDomain = SSEPackedInt
3957 let Predicates = [HasAVX] in {
3958 def : Pat<(int_x86_sse2_storeu_dq addr:$dst, VR128:$src),
3959 (VMOVDQUmr addr:$dst, VR128:$src)>;
3960 def : Pat<(int_x86_avx_storeu_dq_256 addr:$dst, VR256:$src),
3961 (VMOVDQUYmr addr:$dst, VR256:$src)>;
3963 let Predicates = [UseSSE2] in
3964 def : Pat<(int_x86_sse2_storeu_dq addr:$dst, VR128:$src),
3965 (MOVDQUmr addr:$dst, VR128:$src)>;
3967 //===---------------------------------------------------------------------===//
3968 // SSE2 - Packed Integer Arithmetic Instructions
3969 //===---------------------------------------------------------------------===//
3971 let Sched = WriteVecIMul in
3972 def SSE_PMADD : OpndItins<
3973 IIC_SSE_PMADD, IIC_SSE_PMADD
3976 let ExeDomain = SSEPackedInt in { // SSE integer instructions
3978 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
3979 RegisterClass RC, PatFrag memop_frag,
3980 X86MemOperand x86memop,
3982 bit IsCommutable = 0,
3984 let isCommutable = IsCommutable in
3985 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
3986 (ins RC:$src1, RC:$src2),
3988 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3989 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3990 [(set RC:$dst, (IntId RC:$src1, RC:$src2))], itins.rr>,
3991 Sched<[itins.Sched]>;
3992 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
3993 (ins RC:$src1, x86memop:$src2),
3995 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3996 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3997 [(set RC:$dst, (IntId RC:$src1, (bitconvert (memop_frag addr:$src2))))],
3998 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
4001 multiclass PDI_binop_all_int<bits<8> opc, string OpcodeStr, Intrinsic IntId128,
4002 Intrinsic IntId256, OpndItins itins,
4003 bit IsCommutable = 0> {
4004 let Predicates = [HasAVX] in
4005 defm V#NAME : PDI_binop_rm_int<opc, !strconcat("v", OpcodeStr), IntId128,
4006 VR128, loadv2i64, i128mem, itins,
4007 IsCommutable, 0>, VEX_4V;
4009 let Constraints = "$src1 = $dst" in
4010 defm NAME : PDI_binop_rm_int<opc, OpcodeStr, IntId128, VR128, memopv2i64,
4011 i128mem, itins, IsCommutable, 1>;
4013 let Predicates = [HasAVX2] in
4014 defm V#NAME#Y : PDI_binop_rm_int<opc, !strconcat("v", OpcodeStr), IntId256,
4015 VR256, loadv4i64, i256mem, itins,
4016 IsCommutable, 0>, VEX_4V, VEX_L;
4019 multiclass PDI_binop_rmi<bits<8> opc, bits<8> opc2, Format ImmForm,
4020 string OpcodeStr, SDNode OpNode,
4021 SDNode OpNode2, RegisterClass RC,
4022 ValueType DstVT, ValueType SrcVT, PatFrag bc_frag,
4023 ShiftOpndItins itins,
4025 // src2 is always 128-bit
4026 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
4027 (ins RC:$src1, VR128:$src2),
4029 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4030 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4031 [(set RC:$dst, (DstVT (OpNode RC:$src1, (SrcVT VR128:$src2))))],
4032 itins.rr>, Sched<[WriteVecShift]>;
4033 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
4034 (ins RC:$src1, i128mem:$src2),
4036 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4037 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4038 [(set RC:$dst, (DstVT (OpNode RC:$src1,
4039 (bc_frag (memopv2i64 addr:$src2)))))], itins.rm>,
4040 Sched<[WriteVecShiftLd, ReadAfterLd]>;
4041 def ri : PDIi8<opc2, ImmForm, (outs RC:$dst),
4042 (ins RC:$src1, i8imm:$src2),
4044 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4045 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4046 [(set RC:$dst, (DstVT (OpNode2 RC:$src1, (i8 imm:$src2))))], itins.ri>,
4047 Sched<[WriteVecShift]>;
4050 /// PDI_binop_rm2 - Simple SSE2 binary operator with different src and dst types
4051 multiclass PDI_binop_rm2<bits<8> opc, string OpcodeStr, SDNode OpNode,
4052 ValueType DstVT, ValueType SrcVT, RegisterClass RC,
4053 PatFrag memop_frag, X86MemOperand x86memop,
4055 bit IsCommutable = 0, bit Is2Addr = 1> {
4056 let isCommutable = IsCommutable in
4057 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
4058 (ins RC:$src1, RC:$src2),
4060 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4061 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4062 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1), RC:$src2)))]>,
4063 Sched<[itins.Sched]>;
4064 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
4065 (ins RC:$src1, x86memop:$src2),
4067 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4068 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4069 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1),
4070 (bitconvert (memop_frag addr:$src2)))))]>,
4071 Sched<[itins.Sched.Folded, ReadAfterLd]>;
4073 } // ExeDomain = SSEPackedInt
4075 defm PADDB : PDI_binop_all<0xFC, "paddb", add, v16i8, v32i8,
4076 SSE_INTALU_ITINS_P, 1>;
4077 defm PADDW : PDI_binop_all<0xFD, "paddw", add, v8i16, v16i16,
4078 SSE_INTALU_ITINS_P, 1>;
4079 defm PADDD : PDI_binop_all<0xFE, "paddd", add, v4i32, v8i32,
4080 SSE_INTALU_ITINS_P, 1>;
4081 defm PADDQ : PDI_binop_all<0xD4, "paddq", add, v2i64, v4i64,
4082 SSE_INTALUQ_ITINS_P, 1>;
4083 defm PMULLW : PDI_binop_all<0xD5, "pmullw", mul, v8i16, v16i16,
4084 SSE_INTMUL_ITINS_P, 1>;
4085 defm PMULHUW : PDI_binop_all<0xE4, "pmulhuw", mulhu, v8i16, v16i16,
4086 SSE_INTMUL_ITINS_P, 1>;
4087 defm PMULHW : PDI_binop_all<0xE5, "pmulhw", mulhs, v8i16, v16i16,
4088 SSE_INTMUL_ITINS_P, 1>;
4089 defm PSUBB : PDI_binop_all<0xF8, "psubb", sub, v16i8, v32i8,
4090 SSE_INTALU_ITINS_P, 0>;
4091 defm PSUBW : PDI_binop_all<0xF9, "psubw", sub, v8i16, v16i16,
4092 SSE_INTALU_ITINS_P, 0>;
4093 defm PSUBD : PDI_binop_all<0xFA, "psubd", sub, v4i32, v8i32,
4094 SSE_INTALU_ITINS_P, 0>;
4095 defm PSUBQ : PDI_binop_all<0xFB, "psubq", sub, v2i64, v4i64,
4096 SSE_INTALUQ_ITINS_P, 0>;
4097 defm PSUBUSB : PDI_binop_all<0xD8, "psubusb", X86subus, v16i8, v32i8,
4098 SSE_INTALU_ITINS_P, 0>;
4099 defm PSUBUSW : PDI_binop_all<0xD9, "psubusw", X86subus, v8i16, v16i16,
4100 SSE_INTALU_ITINS_P, 0>;
4101 defm PMINUB : PDI_binop_all<0xDA, "pminub", X86umin, v16i8, v32i8,
4102 SSE_INTALU_ITINS_P, 1>;
4103 defm PMINSW : PDI_binop_all<0xEA, "pminsw", X86smin, v8i16, v16i16,
4104 SSE_INTALU_ITINS_P, 1>;
4105 defm PMAXUB : PDI_binop_all<0xDE, "pmaxub", X86umax, v16i8, v32i8,
4106 SSE_INTALU_ITINS_P, 1>;
4107 defm PMAXSW : PDI_binop_all<0xEE, "pmaxsw", X86smax, v8i16, v16i16,
4108 SSE_INTALU_ITINS_P, 1>;
4111 defm PSUBSB : PDI_binop_all_int<0xE8, "psubsb", int_x86_sse2_psubs_b,
4112 int_x86_avx2_psubs_b, SSE_INTALU_ITINS_P, 0>;
4113 defm PSUBSW : PDI_binop_all_int<0xE9, "psubsw" , int_x86_sse2_psubs_w,
4114 int_x86_avx2_psubs_w, SSE_INTALU_ITINS_P, 0>;
4115 defm PADDSB : PDI_binop_all_int<0xEC, "paddsb" , int_x86_sse2_padds_b,
4116 int_x86_avx2_padds_b, SSE_INTALU_ITINS_P, 1>;
4117 defm PADDSW : PDI_binop_all_int<0xED, "paddsw" , int_x86_sse2_padds_w,
4118 int_x86_avx2_padds_w, SSE_INTALU_ITINS_P, 1>;
4119 defm PADDUSB : PDI_binop_all_int<0xDC, "paddusb", int_x86_sse2_paddus_b,
4120 int_x86_avx2_paddus_b, SSE_INTALU_ITINS_P, 1>;
4121 defm PADDUSW : PDI_binop_all_int<0xDD, "paddusw", int_x86_sse2_paddus_w,
4122 int_x86_avx2_paddus_w, SSE_INTALU_ITINS_P, 1>;
4123 defm PMADDWD : PDI_binop_all_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd,
4124 int_x86_avx2_pmadd_wd, SSE_PMADD, 1>;
4125 defm PAVGB : PDI_binop_all_int<0xE0, "pavgb", int_x86_sse2_pavg_b,
4126 int_x86_avx2_pavg_b, SSE_INTALU_ITINS_P, 1>;
4127 defm PAVGW : PDI_binop_all_int<0xE3, "pavgw", int_x86_sse2_pavg_w,
4128 int_x86_avx2_pavg_w, SSE_INTALU_ITINS_P, 1>;
4129 defm PSADBW : PDI_binop_all_int<0xF6, "psadbw", int_x86_sse2_psad_bw,
4130 int_x86_avx2_psad_bw, SSE_PMADD, 1>;
4132 let Predicates = [HasAVX] in
4133 defm VPMULUDQ : PDI_binop_rm2<0xF4, "vpmuludq", X86pmuludq, v2i64, v4i32, VR128,
4134 loadv2i64, i128mem, SSE_INTMUL_ITINS_P, 1, 0>,
4136 let Predicates = [HasAVX2] in
4137 defm VPMULUDQY : PDI_binop_rm2<0xF4, "vpmuludq", X86pmuludq, v4i64, v8i32,
4138 VR256, loadv4i64, i256mem,
4139 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V, VEX_L;
4140 let Constraints = "$src1 = $dst" in
4141 defm PMULUDQ : PDI_binop_rm2<0xF4, "pmuludq", X86pmuludq, v2i64, v4i32, VR128,
4142 memopv2i64, i128mem, SSE_INTMUL_ITINS_P, 1>;
4144 //===---------------------------------------------------------------------===//
4145 // SSE2 - Packed Integer Logical Instructions
4146 //===---------------------------------------------------------------------===//
4148 let Predicates = [HasAVX] in {
4149 defm VPSLLW : PDI_binop_rmi<0xF1, 0x71, MRM6r, "vpsllw", X86vshl, X86vshli,
4150 VR128, v8i16, v8i16, bc_v8i16,
4151 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4152 defm VPSLLD : PDI_binop_rmi<0xF2, 0x72, MRM6r, "vpslld", X86vshl, X86vshli,
4153 VR128, v4i32, v4i32, bc_v4i32,
4154 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4155 defm VPSLLQ : PDI_binop_rmi<0xF3, 0x73, MRM6r, "vpsllq", X86vshl, X86vshli,
4156 VR128, v2i64, v2i64, bc_v2i64,
4157 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4159 defm VPSRLW : PDI_binop_rmi<0xD1, 0x71, MRM2r, "vpsrlw", X86vsrl, X86vsrli,
4160 VR128, v8i16, v8i16, bc_v8i16,
4161 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4162 defm VPSRLD : PDI_binop_rmi<0xD2, 0x72, MRM2r, "vpsrld", X86vsrl, X86vsrli,
4163 VR128, v4i32, v4i32, bc_v4i32,
4164 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4165 defm VPSRLQ : PDI_binop_rmi<0xD3, 0x73, MRM2r, "vpsrlq", X86vsrl, X86vsrli,
4166 VR128, v2i64, v2i64, bc_v2i64,
4167 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4169 defm VPSRAW : PDI_binop_rmi<0xE1, 0x71, MRM4r, "vpsraw", X86vsra, X86vsrai,
4170 VR128, v8i16, v8i16, bc_v8i16,
4171 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4172 defm VPSRAD : PDI_binop_rmi<0xE2, 0x72, MRM4r, "vpsrad", X86vsra, X86vsrai,
4173 VR128, v4i32, v4i32, bc_v4i32,
4174 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4176 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift] in {
4177 // 128-bit logical shifts.
4178 def VPSLLDQri : PDIi8<0x73, MRM7r,
4179 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
4180 "vpslldq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4182 (int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2))]>,
4184 def VPSRLDQri : PDIi8<0x73, MRM3r,
4185 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
4186 "vpsrldq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4188 (int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2))]>,
4190 // PSRADQri doesn't exist in SSE[1-3].
4192 } // Predicates = [HasAVX]
4194 let Predicates = [HasAVX2] in {
4195 defm VPSLLWY : PDI_binop_rmi<0xF1, 0x71, MRM6r, "vpsllw", X86vshl, X86vshli,
4196 VR256, v16i16, v8i16, bc_v8i16,
4197 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4198 defm VPSLLDY : PDI_binop_rmi<0xF2, 0x72, MRM6r, "vpslld", X86vshl, X86vshli,
4199 VR256, v8i32, v4i32, bc_v4i32,
4200 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4201 defm VPSLLQY : PDI_binop_rmi<0xF3, 0x73, MRM6r, "vpsllq", X86vshl, X86vshli,
4202 VR256, v4i64, v2i64, bc_v2i64,
4203 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4205 defm VPSRLWY : PDI_binop_rmi<0xD1, 0x71, MRM2r, "vpsrlw", X86vsrl, X86vsrli,
4206 VR256, v16i16, v8i16, bc_v8i16,
4207 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4208 defm VPSRLDY : PDI_binop_rmi<0xD2, 0x72, MRM2r, "vpsrld", X86vsrl, X86vsrli,
4209 VR256, v8i32, v4i32, bc_v4i32,
4210 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4211 defm VPSRLQY : PDI_binop_rmi<0xD3, 0x73, MRM2r, "vpsrlq", X86vsrl, X86vsrli,
4212 VR256, v4i64, v2i64, bc_v2i64,
4213 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4215 defm VPSRAWY : PDI_binop_rmi<0xE1, 0x71, MRM4r, "vpsraw", X86vsra, X86vsrai,
4216 VR256, v16i16, v8i16, bc_v8i16,
4217 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4218 defm VPSRADY : PDI_binop_rmi<0xE2, 0x72, MRM4r, "vpsrad", X86vsra, X86vsrai,
4219 VR256, v8i32, v4i32, bc_v4i32,
4220 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4222 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift] in {
4223 // 256-bit logical shifts.
4224 def VPSLLDQYri : PDIi8<0x73, MRM7r,
4225 (outs VR256:$dst), (ins VR256:$src1, i32i8imm:$src2),
4226 "vpslldq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4228 (int_x86_avx2_psll_dq_bs VR256:$src1, imm:$src2))]>,
4230 def VPSRLDQYri : PDIi8<0x73, MRM3r,
4231 (outs VR256:$dst), (ins VR256:$src1, i32i8imm:$src2),
4232 "vpsrldq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4234 (int_x86_avx2_psrl_dq_bs VR256:$src1, imm:$src2))]>,
4236 // PSRADQYri doesn't exist in SSE[1-3].
4238 } // Predicates = [HasAVX2]
4240 let Constraints = "$src1 = $dst" in {
4241 defm PSLLW : PDI_binop_rmi<0xF1, 0x71, MRM6r, "psllw", X86vshl, X86vshli,
4242 VR128, v8i16, v8i16, bc_v8i16,
4243 SSE_INTSHIFT_ITINS_P>;
4244 defm PSLLD : PDI_binop_rmi<0xF2, 0x72, MRM6r, "pslld", X86vshl, X86vshli,
4245 VR128, v4i32, v4i32, bc_v4i32,
4246 SSE_INTSHIFT_ITINS_P>;
4247 defm PSLLQ : PDI_binop_rmi<0xF3, 0x73, MRM6r, "psllq", X86vshl, X86vshli,
4248 VR128, v2i64, v2i64, bc_v2i64,
4249 SSE_INTSHIFT_ITINS_P>;
4251 defm PSRLW : PDI_binop_rmi<0xD1, 0x71, MRM2r, "psrlw", X86vsrl, X86vsrli,
4252 VR128, v8i16, v8i16, bc_v8i16,
4253 SSE_INTSHIFT_ITINS_P>;
4254 defm PSRLD : PDI_binop_rmi<0xD2, 0x72, MRM2r, "psrld", X86vsrl, X86vsrli,
4255 VR128, v4i32, v4i32, bc_v4i32,
4256 SSE_INTSHIFT_ITINS_P>;
4257 defm PSRLQ : PDI_binop_rmi<0xD3, 0x73, MRM2r, "psrlq", X86vsrl, X86vsrli,
4258 VR128, v2i64, v2i64, bc_v2i64,
4259 SSE_INTSHIFT_ITINS_P>;
4261 defm PSRAW : PDI_binop_rmi<0xE1, 0x71, MRM4r, "psraw", X86vsra, X86vsrai,
4262 VR128, v8i16, v8i16, bc_v8i16,
4263 SSE_INTSHIFT_ITINS_P>;
4264 defm PSRAD : PDI_binop_rmi<0xE2, 0x72, MRM4r, "psrad", X86vsra, X86vsrai,
4265 VR128, v4i32, v4i32, bc_v4i32,
4266 SSE_INTSHIFT_ITINS_P>;
4268 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift] in {
4269 // 128-bit logical shifts.
4270 def PSLLDQri : PDIi8<0x73, MRM7r,
4271 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
4272 "pslldq\t{$src2, $dst|$dst, $src2}",
4274 (int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2))],
4275 IIC_SSE_INTSHDQ_P_RI>;
4276 def PSRLDQri : PDIi8<0x73, MRM3r,
4277 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
4278 "psrldq\t{$src2, $dst|$dst, $src2}",
4280 (int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2))],
4281 IIC_SSE_INTSHDQ_P_RI>;
4282 // PSRADQri doesn't exist in SSE[1-3].
4284 } // Constraints = "$src1 = $dst"
4286 let Predicates = [HasAVX] in {
4287 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
4288 (VPSLLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4289 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
4290 (VPSRLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4291 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
4292 (VPSRLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4294 // Shift up / down and insert zero's.
4295 def : Pat<(v2i64 (X86vshldq VR128:$src, (i8 imm:$amt))),
4296 (VPSLLDQri VR128:$src, (BYTE_imm imm:$amt))>;
4297 def : Pat<(v2i64 (X86vshrdq VR128:$src, (i8 imm:$amt))),
4298 (VPSRLDQri VR128:$src, (BYTE_imm imm:$amt))>;
4301 let Predicates = [HasAVX2] in {
4302 def : Pat<(int_x86_avx2_psll_dq VR256:$src1, imm:$src2),
4303 (VPSLLDQYri VR256:$src1, (BYTE_imm imm:$src2))>;
4304 def : Pat<(int_x86_avx2_psrl_dq VR256:$src1, imm:$src2),
4305 (VPSRLDQYri VR256:$src1, (BYTE_imm imm:$src2))>;
4308 let Predicates = [UseSSE2] in {
4309 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
4310 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4311 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
4312 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4313 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
4314 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4316 // Shift up / down and insert zero's.
4317 def : Pat<(v2i64 (X86vshldq VR128:$src, (i8 imm:$amt))),
4318 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt))>;
4319 def : Pat<(v2i64 (X86vshrdq VR128:$src, (i8 imm:$amt))),
4320 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt))>;
4323 //===---------------------------------------------------------------------===//
4324 // SSE2 - Packed Integer Comparison Instructions
4325 //===---------------------------------------------------------------------===//
4327 defm PCMPEQB : PDI_binop_all<0x74, "pcmpeqb", X86pcmpeq, v16i8, v32i8,
4328 SSE_INTALU_ITINS_P, 1>;
4329 defm PCMPEQW : PDI_binop_all<0x75, "pcmpeqw", X86pcmpeq, v8i16, v16i16,
4330 SSE_INTALU_ITINS_P, 1>;
4331 defm PCMPEQD : PDI_binop_all<0x76, "pcmpeqd", X86pcmpeq, v4i32, v8i32,
4332 SSE_INTALU_ITINS_P, 1>;
4333 defm PCMPGTB : PDI_binop_all<0x64, "pcmpgtb", X86pcmpgt, v16i8, v32i8,
4334 SSE_INTALU_ITINS_P, 0>;
4335 defm PCMPGTW : PDI_binop_all<0x65, "pcmpgtw", X86pcmpgt, v8i16, v16i16,
4336 SSE_INTALU_ITINS_P, 0>;
4337 defm PCMPGTD : PDI_binop_all<0x66, "pcmpgtd", X86pcmpgt, v4i32, v8i32,
4338 SSE_INTALU_ITINS_P, 0>;
4340 //===---------------------------------------------------------------------===//
4341 // SSE2 - Packed Integer Shuffle Instructions
4342 //===---------------------------------------------------------------------===//
4344 let ExeDomain = SSEPackedInt in {
4345 multiclass sse2_pshuffle<string OpcodeStr, ValueType vt128, ValueType vt256,
4347 let Predicates = [HasAVX] in {
4348 def V#NAME#ri : Ii8<0x70, MRMSrcReg, (outs VR128:$dst),
4349 (ins VR128:$src1, i8imm:$src2),
4350 !strconcat("v", OpcodeStr,
4351 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4353 (vt128 (OpNode VR128:$src1, (i8 imm:$src2))))],
4354 IIC_SSE_PSHUF_RI>, VEX, Sched<[WriteShuffle]>;
4355 def V#NAME#mi : Ii8<0x70, MRMSrcMem, (outs VR128:$dst),
4356 (ins i128mem:$src1, i8imm:$src2),
4357 !strconcat("v", OpcodeStr,
4358 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4360 (vt128 (OpNode (bitconvert (loadv2i64 addr:$src1)),
4361 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>, VEX,
4362 Sched<[WriteShuffleLd]>;
4365 let Predicates = [HasAVX2] in {
4366 def V#NAME#Yri : Ii8<0x70, MRMSrcReg, (outs VR256:$dst),
4367 (ins VR256:$src1, i8imm:$src2),
4368 !strconcat("v", OpcodeStr,
4369 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4371 (vt256 (OpNode VR256:$src1, (i8 imm:$src2))))],
4372 IIC_SSE_PSHUF_RI>, VEX, VEX_L, Sched<[WriteShuffle]>;
4373 def V#NAME#Ymi : Ii8<0x70, MRMSrcMem, (outs VR256:$dst),
4374 (ins i256mem:$src1, i8imm:$src2),
4375 !strconcat("v", OpcodeStr,
4376 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4378 (vt256 (OpNode (bitconvert (loadv4i64 addr:$src1)),
4379 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>, VEX, VEX_L,
4380 Sched<[WriteShuffleLd]>;
4383 let Predicates = [UseSSE2] in {
4384 def ri : Ii8<0x70, MRMSrcReg,
4385 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
4386 !strconcat(OpcodeStr,
4387 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4389 (vt128 (OpNode VR128:$src1, (i8 imm:$src2))))],
4390 IIC_SSE_PSHUF_RI>, Sched<[WriteShuffle]>;
4391 def mi : Ii8<0x70, MRMSrcMem,
4392 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
4393 !strconcat(OpcodeStr,
4394 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4396 (vt128 (OpNode (bitconvert (memopv2i64 addr:$src1)),
4397 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>,
4398 Sched<[WriteShuffleLd, ReadAfterLd]>;
4401 } // ExeDomain = SSEPackedInt
4403 defm PSHUFD : sse2_pshuffle<"pshufd", v4i32, v8i32, X86PShufd>, PD;
4404 defm PSHUFHW : sse2_pshuffle<"pshufhw", v8i16, v16i16, X86PShufhw>, XS;
4405 defm PSHUFLW : sse2_pshuffle<"pshuflw", v8i16, v16i16, X86PShuflw>, XD;
4407 let Predicates = [HasAVX] in {
4408 def : Pat<(v4f32 (X86PShufd (loadv4f32 addr:$src1), (i8 imm:$imm))),
4409 (VPSHUFDmi addr:$src1, imm:$imm)>;
4410 def : Pat<(v4f32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
4411 (VPSHUFDri VR128:$src1, imm:$imm)>;
4414 let Predicates = [UseSSE2] in {
4415 def : Pat<(v4f32 (X86PShufd (memopv4f32 addr:$src1), (i8 imm:$imm))),
4416 (PSHUFDmi addr:$src1, imm:$imm)>;
4417 def : Pat<(v4f32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
4418 (PSHUFDri VR128:$src1, imm:$imm)>;
4421 //===---------------------------------------------------------------------===//
4422 // Packed Integer Pack Instructions (SSE & AVX)
4423 //===---------------------------------------------------------------------===//
4425 let ExeDomain = SSEPackedInt in {
4426 multiclass sse2_pack<bits<8> opc, string OpcodeStr, ValueType OutVT,
4427 ValueType ArgVT, SDNode OpNode, PatFrag bc_frag,
4429 def rr : PDI<opc, MRMSrcReg,
4430 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4432 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4433 !strconcat(OpcodeStr,
4434 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4436 (OutVT (OpNode (ArgVT VR128:$src1), VR128:$src2)))]>,
4437 Sched<[WriteShuffle]>;
4438 def rm : PDI<opc, MRMSrcMem,
4439 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4441 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4442 !strconcat(OpcodeStr,
4443 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4445 (OutVT (OpNode VR128:$src1,
4446 (bc_frag (memopv2i64 addr:$src2)))))]>,
4447 Sched<[WriteShuffleLd, ReadAfterLd]>;
4450 multiclass sse2_pack_y<bits<8> opc, string OpcodeStr, ValueType OutVT,
4451 ValueType ArgVT, SDNode OpNode, PatFrag bc_frag> {
4452 def Yrr : PDI<opc, MRMSrcReg,
4453 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4454 !strconcat(OpcodeStr,
4455 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4457 (OutVT (OpNode (ArgVT VR256:$src1), VR256:$src2)))]>,
4458 Sched<[WriteShuffle]>;
4459 def Yrm : PDI<opc, MRMSrcMem,
4460 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4461 !strconcat(OpcodeStr,
4462 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4464 (OutVT (OpNode VR256:$src1,
4465 (bc_frag (memopv4i64 addr:$src2)))))]>,
4466 Sched<[WriteShuffleLd, ReadAfterLd]>;
4469 multiclass sse4_pack<bits<8> opc, string OpcodeStr, ValueType OutVT,
4470 ValueType ArgVT, SDNode OpNode, PatFrag bc_frag,
4472 def rr : SS48I<opc, MRMSrcReg,
4473 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4475 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4476 !strconcat(OpcodeStr,
4477 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4479 (OutVT (OpNode (ArgVT VR128:$src1), VR128:$src2)))]>,
4480 Sched<[WriteShuffle]>;
4481 def rm : SS48I<opc, MRMSrcMem,
4482 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4484 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4485 !strconcat(OpcodeStr,
4486 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4488 (OutVT (OpNode VR128:$src1,
4489 (bc_frag (memopv2i64 addr:$src2)))))]>,
4490 Sched<[WriteShuffleLd, ReadAfterLd]>;
4493 multiclass sse4_pack_y<bits<8> opc, string OpcodeStr, ValueType OutVT,
4494 ValueType ArgVT, SDNode OpNode, PatFrag bc_frag> {
4495 def Yrr : SS48I<opc, MRMSrcReg,
4496 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4497 !strconcat(OpcodeStr,
4498 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4500 (OutVT (OpNode (ArgVT VR256:$src1), VR256:$src2)))]>,
4501 Sched<[WriteShuffle]>;
4502 def Yrm : SS48I<opc, MRMSrcMem,
4503 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4504 !strconcat(OpcodeStr,
4505 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4507 (OutVT (OpNode VR256:$src1,
4508 (bc_frag (memopv4i64 addr:$src2)))))]>,
4509 Sched<[WriteShuffleLd, ReadAfterLd]>;
4512 let Predicates = [HasAVX] in {
4513 defm VPACKSSWB : sse2_pack<0x63, "vpacksswb", v16i8, v8i16, X86Packss,
4514 bc_v8i16, 0>, VEX_4V;
4515 defm VPACKSSDW : sse2_pack<0x6B, "vpackssdw", v8i16, v4i32, X86Packss,
4516 bc_v4i32, 0>, VEX_4V;
4518 defm VPACKUSWB : sse2_pack<0x67, "vpackuswb", v16i8, v8i16, X86Packus,
4519 bc_v8i16, 0>, VEX_4V;
4520 defm VPACKUSDW : sse4_pack<0x2B, "vpackusdw", v8i16, v4i32, X86Packus,
4521 bc_v4i32, 0>, VEX_4V;
4524 let Predicates = [HasAVX2] in {
4525 defm VPACKSSWB : sse2_pack_y<0x63, "vpacksswb", v32i8, v16i16, X86Packss,
4526 bc_v16i16>, VEX_4V, VEX_L;
4527 defm VPACKSSDW : sse2_pack_y<0x6B, "vpackssdw", v16i16, v8i32, X86Packss,
4528 bc_v8i32>, VEX_4V, VEX_L;
4530 defm VPACKUSWB : sse2_pack_y<0x67, "vpackuswb", v32i8, v16i16, X86Packus,
4531 bc_v16i16>, VEX_4V, VEX_L;
4532 defm VPACKUSDW : sse4_pack_y<0x2B, "vpackusdw", v16i16, v8i32, X86Packus,
4533 bc_v8i32>, VEX_4V, VEX_L;
4536 let Constraints = "$src1 = $dst" in {
4537 defm PACKSSWB : sse2_pack<0x63, "packsswb", v16i8, v8i16, X86Packss,
4539 defm PACKSSDW : sse2_pack<0x6B, "packssdw", v8i16, v4i32, X86Packss,
4542 defm PACKUSWB : sse2_pack<0x67, "packuswb", v16i8, v8i16, X86Packus,
4545 let Predicates = [HasSSE41] in
4546 defm PACKUSDW : sse4_pack<0x2B, "packusdw", v8i16, v4i32, X86Packus,
4549 } // ExeDomain = SSEPackedInt
4551 //===---------------------------------------------------------------------===//
4552 // SSE2 - Packed Integer Unpack Instructions
4553 //===---------------------------------------------------------------------===//
4555 let ExeDomain = SSEPackedInt in {
4556 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
4557 SDNode OpNode, PatFrag bc_frag, bit Is2Addr = 1> {
4558 def rr : PDI<opc, MRMSrcReg,
4559 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4561 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
4562 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4563 [(set VR128:$dst, (vt (OpNode VR128:$src1, VR128:$src2)))],
4564 IIC_SSE_UNPCK>, Sched<[WriteShuffle]>;
4565 def rm : PDI<opc, MRMSrcMem,
4566 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4568 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
4569 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4570 [(set VR128:$dst, (OpNode VR128:$src1,
4571 (bc_frag (memopv2i64
4574 Sched<[WriteShuffleLd, ReadAfterLd]>;
4577 multiclass sse2_unpack_y<bits<8> opc, string OpcodeStr, ValueType vt,
4578 SDNode OpNode, PatFrag bc_frag> {
4579 def Yrr : PDI<opc, MRMSrcReg,
4580 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4581 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4582 [(set VR256:$dst, (vt (OpNode VR256:$src1, VR256:$src2)))]>,
4583 Sched<[WriteShuffle]>;
4584 def Yrm : PDI<opc, MRMSrcMem,
4585 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4586 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4587 [(set VR256:$dst, (OpNode VR256:$src1,
4588 (bc_frag (memopv4i64 addr:$src2))))]>,
4589 Sched<[WriteShuffleLd, ReadAfterLd]>;
4592 let Predicates = [HasAVX] in {
4593 defm VPUNPCKLBW : sse2_unpack<0x60, "vpunpcklbw", v16i8, X86Unpckl,
4594 bc_v16i8, 0>, VEX_4V;
4595 defm VPUNPCKLWD : sse2_unpack<0x61, "vpunpcklwd", v8i16, X86Unpckl,
4596 bc_v8i16, 0>, VEX_4V;
4597 defm VPUNPCKLDQ : sse2_unpack<0x62, "vpunpckldq", v4i32, X86Unpckl,
4598 bc_v4i32, 0>, VEX_4V;
4599 defm VPUNPCKLQDQ : sse2_unpack<0x6C, "vpunpcklqdq", v2i64, X86Unpckl,
4600 bc_v2i64, 0>, VEX_4V;
4602 defm VPUNPCKHBW : sse2_unpack<0x68, "vpunpckhbw", v16i8, X86Unpckh,
4603 bc_v16i8, 0>, VEX_4V;
4604 defm VPUNPCKHWD : sse2_unpack<0x69, "vpunpckhwd", v8i16, X86Unpckh,
4605 bc_v8i16, 0>, VEX_4V;
4606 defm VPUNPCKHDQ : sse2_unpack<0x6A, "vpunpckhdq", v4i32, X86Unpckh,
4607 bc_v4i32, 0>, VEX_4V;
4608 defm VPUNPCKHQDQ : sse2_unpack<0x6D, "vpunpckhqdq", v2i64, X86Unpckh,
4609 bc_v2i64, 0>, VEX_4V;
4612 let Predicates = [HasAVX2] in {
4613 defm VPUNPCKLBW : sse2_unpack_y<0x60, "vpunpcklbw", v32i8, X86Unpckl,
4614 bc_v32i8>, VEX_4V, VEX_L;
4615 defm VPUNPCKLWD : sse2_unpack_y<0x61, "vpunpcklwd", v16i16, X86Unpckl,
4616 bc_v16i16>, VEX_4V, VEX_L;
4617 defm VPUNPCKLDQ : sse2_unpack_y<0x62, "vpunpckldq", v8i32, X86Unpckl,
4618 bc_v8i32>, VEX_4V, VEX_L;
4619 defm VPUNPCKLQDQ : sse2_unpack_y<0x6C, "vpunpcklqdq", v4i64, X86Unpckl,
4620 bc_v4i64>, VEX_4V, VEX_L;
4622 defm VPUNPCKHBW : sse2_unpack_y<0x68, "vpunpckhbw", v32i8, X86Unpckh,
4623 bc_v32i8>, VEX_4V, VEX_L;
4624 defm VPUNPCKHWD : sse2_unpack_y<0x69, "vpunpckhwd", v16i16, X86Unpckh,
4625 bc_v16i16>, VEX_4V, VEX_L;
4626 defm VPUNPCKHDQ : sse2_unpack_y<0x6A, "vpunpckhdq", v8i32, X86Unpckh,
4627 bc_v8i32>, VEX_4V, VEX_L;
4628 defm VPUNPCKHQDQ : sse2_unpack_y<0x6D, "vpunpckhqdq", v4i64, X86Unpckh,
4629 bc_v4i64>, VEX_4V, VEX_L;
4632 let Constraints = "$src1 = $dst" in {
4633 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, X86Unpckl,
4635 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, X86Unpckl,
4637 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, X86Unpckl,
4639 defm PUNPCKLQDQ : sse2_unpack<0x6C, "punpcklqdq", v2i64, X86Unpckl,
4642 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, X86Unpckh,
4644 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, X86Unpckh,
4646 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, X86Unpckh,
4648 defm PUNPCKHQDQ : sse2_unpack<0x6D, "punpckhqdq", v2i64, X86Unpckh,
4651 } // ExeDomain = SSEPackedInt
4653 //===---------------------------------------------------------------------===//
4654 // SSE2 - Packed Integer Extract and Insert
4655 //===---------------------------------------------------------------------===//
4657 let ExeDomain = SSEPackedInt in {
4658 multiclass sse2_pinsrw<bit Is2Addr = 1> {
4659 def rri : Ii8<0xC4, MRMSrcReg,
4660 (outs VR128:$dst), (ins VR128:$src1,
4661 GR32orGR64:$src2, i32i8imm:$src3),
4663 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
4664 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4666 (X86pinsrw VR128:$src1, GR32orGR64:$src2, imm:$src3))],
4667 IIC_SSE_PINSRW>, Sched<[WriteShuffle]>;
4668 def rmi : Ii8<0xC4, MRMSrcMem,
4669 (outs VR128:$dst), (ins VR128:$src1,
4670 i16mem:$src2, i32i8imm:$src3),
4672 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
4673 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4675 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
4676 imm:$src3))], IIC_SSE_PINSRW>,
4677 Sched<[WriteShuffleLd, ReadAfterLd]>;
4681 let Predicates = [HasAVX] in
4682 def VPEXTRWri : Ii8<0xC5, MRMSrcReg,
4683 (outs GR32orGR64:$dst), (ins VR128:$src1, i32i8imm:$src2),
4684 "vpextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4685 [(set GR32orGR64:$dst, (X86pextrw (v8i16 VR128:$src1),
4686 imm:$src2))]>, PD, VEX,
4687 Sched<[WriteShuffle]>;
4688 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
4689 (outs GR32orGR64:$dst), (ins VR128:$src1, i32i8imm:$src2),
4690 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4691 [(set GR32orGR64:$dst, (X86pextrw (v8i16 VR128:$src1),
4692 imm:$src2))], IIC_SSE_PEXTRW>,
4693 Sched<[WriteShuffleLd, ReadAfterLd]>;
4696 let Predicates = [HasAVX] in
4697 defm VPINSRW : sse2_pinsrw<0>, PD, VEX_4V;
4699 let Predicates = [UseSSE2], Constraints = "$src1 = $dst" in
4700 defm PINSRW : sse2_pinsrw, PD;
4702 } // ExeDomain = SSEPackedInt
4704 //===---------------------------------------------------------------------===//
4705 // SSE2 - Packed Mask Creation
4706 //===---------------------------------------------------------------------===//
4708 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecLogic] in {
4710 def VPMOVMSKBrr : VPDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst),
4712 "pmovmskb\t{$src, $dst|$dst, $src}",
4713 [(set GR32orGR64:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))],
4714 IIC_SSE_MOVMSK>, VEX;
4716 let Predicates = [HasAVX2] in {
4717 def VPMOVMSKBYrr : VPDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst),
4719 "pmovmskb\t{$src, $dst|$dst, $src}",
4720 [(set GR32orGR64:$dst, (int_x86_avx2_pmovmskb VR256:$src))]>,
4724 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst), (ins VR128:$src),
4725 "pmovmskb\t{$src, $dst|$dst, $src}",
4726 [(set GR32orGR64:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))],
4729 } // ExeDomain = SSEPackedInt
4731 //===---------------------------------------------------------------------===//
4732 // SSE2 - Conditional Store
4733 //===---------------------------------------------------------------------===//
4735 let ExeDomain = SSEPackedInt, SchedRW = [WriteStore] in {
4737 let Uses = [EDI], Predicates = [HasAVX,Not64BitMode] in
4738 def VMASKMOVDQU : VPDI<0xF7, MRMSrcReg, (outs),
4739 (ins VR128:$src, VR128:$mask),
4740 "maskmovdqu\t{$mask, $src|$src, $mask}",
4741 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)],
4742 IIC_SSE_MASKMOV>, VEX;
4743 let Uses = [RDI], Predicates = [HasAVX,In64BitMode] in
4744 def VMASKMOVDQU64 : VPDI<0xF7, MRMSrcReg, (outs),
4745 (ins VR128:$src, VR128:$mask),
4746 "maskmovdqu\t{$mask, $src|$src, $mask}",
4747 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)],
4748 IIC_SSE_MASKMOV>, VEX;
4750 let Uses = [EDI], Predicates = [UseSSE2,Not64BitMode] in
4751 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
4752 "maskmovdqu\t{$mask, $src|$src, $mask}",
4753 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)],
4755 let Uses = [RDI], Predicates = [UseSSE2,In64BitMode] in
4756 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
4757 "maskmovdqu\t{$mask, $src|$src, $mask}",
4758 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)],
4761 } // ExeDomain = SSEPackedInt
4763 //===---------------------------------------------------------------------===//
4764 // SSE2 - Move Doubleword
4765 //===---------------------------------------------------------------------===//
4767 //===---------------------------------------------------------------------===//
4768 // Move Int Doubleword to Packed Double Int
4770 def VMOVDI2PDIrr : VS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
4771 "movd\t{$src, $dst|$dst, $src}",
4773 (v4i32 (scalar_to_vector GR32:$src)))], IIC_SSE_MOVDQ>,
4774 VEX, Sched<[WriteMove]>;
4775 def VMOVDI2PDIrm : VS2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
4776 "movd\t{$src, $dst|$dst, $src}",
4778 (v4i32 (scalar_to_vector (loadi32 addr:$src))))],
4780 VEX, Sched<[WriteLoad]>;
4781 def VMOV64toPQIrr : VRS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4782 "movq\t{$src, $dst|$dst, $src}",
4784 (v2i64 (scalar_to_vector GR64:$src)))],
4785 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4786 let isCodeGenOnly = 1 in
4787 def VMOV64toSDrr : VRS2I<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
4788 "movq\t{$src, $dst|$dst, $src}",
4789 [(set FR64:$dst, (bitconvert GR64:$src))],
4790 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4792 def MOVDI2PDIrr : S2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
4793 "movd\t{$src, $dst|$dst, $src}",
4795 (v4i32 (scalar_to_vector GR32:$src)))], IIC_SSE_MOVDQ>,
4797 def MOVDI2PDIrm : S2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
4798 "movd\t{$src, $dst|$dst, $src}",
4800 (v4i32 (scalar_to_vector (loadi32 addr:$src))))],
4801 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4802 def MOV64toPQIrr : RS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4803 "mov{d|q}\t{$src, $dst|$dst, $src}",
4805 (v2i64 (scalar_to_vector GR64:$src)))],
4806 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4807 let isCodeGenOnly = 1 in
4808 def MOV64toSDrr : RS2I<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
4809 "mov{d|q}\t{$src, $dst|$dst, $src}",
4810 [(set FR64:$dst, (bitconvert GR64:$src))],
4811 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4813 //===---------------------------------------------------------------------===//
4814 // Move Int Doubleword to Single Scalar
4816 let isCodeGenOnly = 1 in {
4817 def VMOVDI2SSrr : VS2I<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
4818 "movd\t{$src, $dst|$dst, $src}",
4819 [(set FR32:$dst, (bitconvert GR32:$src))],
4820 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4822 def VMOVDI2SSrm : VS2I<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
4823 "movd\t{$src, $dst|$dst, $src}",
4824 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))],
4826 VEX, Sched<[WriteLoad]>;
4827 def MOVDI2SSrr : S2I<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
4828 "movd\t{$src, $dst|$dst, $src}",
4829 [(set FR32:$dst, (bitconvert GR32:$src))],
4830 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4832 def MOVDI2SSrm : S2I<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
4833 "movd\t{$src, $dst|$dst, $src}",
4834 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))],
4835 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4838 //===---------------------------------------------------------------------===//
4839 // Move Packed Doubleword Int to Packed Double Int
4841 def VMOVPDI2DIrr : VS2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
4842 "movd\t{$src, $dst|$dst, $src}",
4843 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
4844 (iPTR 0)))], IIC_SSE_MOVD_ToGP>, VEX,
4846 def VMOVPDI2DImr : VS2I<0x7E, MRMDestMem, (outs),
4847 (ins i32mem:$dst, VR128:$src),
4848 "movd\t{$src, $dst|$dst, $src}",
4849 [(store (i32 (vector_extract (v4i32 VR128:$src),
4850 (iPTR 0))), addr:$dst)], IIC_SSE_MOVDQ>,
4851 VEX, Sched<[WriteStore]>;
4852 def MOVPDI2DIrr : S2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
4853 "movd\t{$src, $dst|$dst, $src}",
4854 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
4855 (iPTR 0)))], IIC_SSE_MOVD_ToGP>,
4857 def MOVPDI2DImr : S2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
4858 "movd\t{$src, $dst|$dst, $src}",
4859 [(store (i32 (vector_extract (v4i32 VR128:$src),
4860 (iPTR 0))), addr:$dst)],
4861 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4863 def : Pat<(v8i32 (X86Vinsert (v8i32 immAllZerosV), GR32:$src2, (iPTR 0))),
4864 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrr GR32:$src2), sub_xmm)>;
4866 def : Pat<(v4i64 (X86Vinsert (bc_v4i64 (v8i32 immAllZerosV)), GR64:$src2, (iPTR 0))),
4867 (SUBREG_TO_REG (i32 0), (VMOV64toPQIrr GR64:$src2), sub_xmm)>;
4869 def : Pat<(v8i32 (X86Vinsert undef, GR32:$src2, (iPTR 0))),
4870 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrr GR32:$src2), sub_xmm)>;
4872 def : Pat<(v4i64 (X86Vinsert undef, GR64:$src2, (iPTR 0))),
4873 (SUBREG_TO_REG (i32 0), (VMOV64toPQIrr GR64:$src2), sub_xmm)>;
4875 //===---------------------------------------------------------------------===//
4876 // Move Packed Doubleword Int first element to Doubleword Int
4878 let SchedRW = [WriteMove] in {
4879 def VMOVPQIto64rr : VRS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
4880 "movq\t{$src, $dst|$dst, $src}",
4881 [(set GR64:$dst, (vector_extract (v2i64 VR128:$src),
4886 def MOVPQIto64rr : RS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
4887 "mov{d|q}\t{$src, $dst|$dst, $src}",
4888 [(set GR64:$dst, (vector_extract (v2i64 VR128:$src),
4893 //===---------------------------------------------------------------------===//
4894 // Bitcast FR64 <-> GR64
4896 let isCodeGenOnly = 1 in {
4897 let Predicates = [UseAVX] in
4898 def VMOV64toSDrm : VS2SI<0x7E, MRMSrcMem, (outs FR64:$dst), (ins i64mem:$src),
4899 "movq\t{$src, $dst|$dst, $src}",
4900 [(set FR64:$dst, (bitconvert (loadi64 addr:$src)))]>,
4901 VEX, Sched<[WriteLoad]>;
4902 def VMOVSDto64rr : VRS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins FR64:$src),
4903 "movq\t{$src, $dst|$dst, $src}",
4904 [(set GR64:$dst, (bitconvert FR64:$src))],
4905 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4906 def VMOVSDto64mr : VRS2I<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, FR64:$src),
4907 "movq\t{$src, $dst|$dst, $src}",
4908 [(store (i64 (bitconvert FR64:$src)), addr:$dst)],
4909 IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4911 def MOV64toSDrm : S2SI<0x7E, MRMSrcMem, (outs FR64:$dst), (ins i64mem:$src),
4912 "movq\t{$src, $dst|$dst, $src}",
4913 [(set FR64:$dst, (bitconvert (loadi64 addr:$src)))],
4914 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4915 def MOVSDto64rr : RS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins FR64:$src),
4916 "mov{d|q}\t{$src, $dst|$dst, $src}",
4917 [(set GR64:$dst, (bitconvert FR64:$src))],
4918 IIC_SSE_MOVD_ToGP>, Sched<[WriteMove]>;
4919 def MOVSDto64mr : RS2I<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, FR64:$src),
4920 "movq\t{$src, $dst|$dst, $src}",
4921 [(store (i64 (bitconvert FR64:$src)), addr:$dst)],
4922 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4925 //===---------------------------------------------------------------------===//
4926 // Move Scalar Single to Double Int
4928 let isCodeGenOnly = 1 in {
4929 def VMOVSS2DIrr : VS2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
4930 "movd\t{$src, $dst|$dst, $src}",
4931 [(set GR32:$dst, (bitconvert FR32:$src))],
4932 IIC_SSE_MOVD_ToGP>, VEX, Sched<[WriteMove]>;
4933 def VMOVSS2DImr : VS2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
4934 "movd\t{$src, $dst|$dst, $src}",
4935 [(store (i32 (bitconvert FR32:$src)), addr:$dst)],
4936 IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4937 def MOVSS2DIrr : S2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
4938 "movd\t{$src, $dst|$dst, $src}",
4939 [(set GR32:$dst, (bitconvert FR32:$src))],
4940 IIC_SSE_MOVD_ToGP>, Sched<[WriteMove]>;
4941 def MOVSS2DImr : S2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
4942 "movd\t{$src, $dst|$dst, $src}",
4943 [(store (i32 (bitconvert FR32:$src)), addr:$dst)],
4944 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4947 //===---------------------------------------------------------------------===//
4948 // Patterns and instructions to describe movd/movq to XMM register zero-extends
4950 let isCodeGenOnly = 1, SchedRW = [WriteMove] in {
4951 let AddedComplexity = 15 in {
4952 def VMOVZQI2PQIrr : VS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4953 "movq\t{$src, $dst|$dst, $src}", // X86-64 only
4954 [(set VR128:$dst, (v2i64 (X86vzmovl
4955 (v2i64 (scalar_to_vector GR64:$src)))))],
4958 def MOVZQI2PQIrr : RS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4959 "mov{d|q}\t{$src, $dst|$dst, $src}", // X86-64 only
4960 [(set VR128:$dst, (v2i64 (X86vzmovl
4961 (v2i64 (scalar_to_vector GR64:$src)))))],
4964 } // isCodeGenOnly, SchedRW
4966 let Predicates = [UseAVX] in {
4967 let AddedComplexity = 15 in
4968 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector GR32:$src)))),
4969 (VMOVDI2PDIrr GR32:$src)>;
4971 // AVX 128-bit movd/movq instruction write zeros in the high 128-bit part.
4972 let AddedComplexity = 20 in {
4973 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
4974 (VMOVDI2PDIrm addr:$src)>;
4975 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
4976 (VMOVDI2PDIrm addr:$src)>;
4977 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
4978 (VMOVDI2PDIrm addr:$src)>;
4980 // Use regular 128-bit instructions to match 256-bit scalar_to_vec+zext.
4981 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
4982 (v4i32 (scalar_to_vector GR32:$src)),(iPTR 0)))),
4983 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrr GR32:$src), sub_xmm)>;
4984 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
4985 (v2i64 (scalar_to_vector GR64:$src)),(iPTR 0)))),
4986 (SUBREG_TO_REG (i64 0), (VMOVZQI2PQIrr GR64:$src), sub_xmm)>;
4989 let Predicates = [UseSSE2] in {
4990 let AddedComplexity = 15 in
4991 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector GR32:$src)))),
4992 (MOVDI2PDIrr GR32:$src)>;
4994 let AddedComplexity = 20 in {
4995 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
4996 (MOVDI2PDIrm addr:$src)>;
4997 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
4998 (MOVDI2PDIrm addr:$src)>;
4999 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
5000 (MOVDI2PDIrm addr:$src)>;
5004 // These are the correct encodings of the instructions so that we know how to
5005 // read correct assembly, even though we continue to emit the wrong ones for
5006 // compatibility with Darwin's buggy assembler.
5007 def : InstAlias<"movq\t{$src, $dst|$dst, $src}",
5008 (MOV64toPQIrr VR128:$dst, GR64:$src), 0>;
5009 def : InstAlias<"movq\t{$src, $dst|$dst, $src}",
5010 (MOVPQIto64rr GR64:$dst, VR128:$src), 0>;
5011 // Allow "vmovd" but print "vmovq" since we don't need compatibility for AVX.
5012 def : InstAlias<"vmovd\t{$src, $dst|$dst, $src}",
5013 (VMOV64toPQIrr VR128:$dst, GR64:$src), 0>;
5014 def : InstAlias<"vmovd\t{$src, $dst|$dst, $src}",
5015 (VMOVPQIto64rr GR64:$dst, VR128:$src), 0>;
5017 //===---------------------------------------------------------------------===//
5018 // SSE2 - Move Quadword
5019 //===---------------------------------------------------------------------===//
5021 //===---------------------------------------------------------------------===//
5022 // Move Quadword Int to Packed Quadword Int
5025 let SchedRW = [WriteLoad] in {
5026 def VMOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
5027 "vmovq\t{$src, $dst|$dst, $src}",
5029 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
5030 VEX, Requires<[UseAVX]>;
5031 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
5032 "movq\t{$src, $dst|$dst, $src}",
5034 (v2i64 (scalar_to_vector (loadi64 addr:$src))))],
5036 Requires<[UseSSE2]>; // SSE2 instruction with XS Prefix
5039 //===---------------------------------------------------------------------===//
5040 // Move Packed Quadword Int to Quadword Int
5042 let SchedRW = [WriteStore] in {
5043 def VMOVPQI2QImr : VS2I<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
5044 "movq\t{$src, $dst|$dst, $src}",
5045 [(store (i64 (vector_extract (v2i64 VR128:$src),
5046 (iPTR 0))), addr:$dst)],
5047 IIC_SSE_MOVDQ>, VEX;
5048 def MOVPQI2QImr : S2I<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
5049 "movq\t{$src, $dst|$dst, $src}",
5050 [(store (i64 (vector_extract (v2i64 VR128:$src),
5051 (iPTR 0))), addr:$dst)],
5055 // For disassembler only
5056 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
5057 SchedRW = [WriteVecLogic] in {
5058 def VMOVPQI2QIrr : VS2I<0xD6, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
5059 "movq\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVQ_RR>, VEX;
5060 def MOVPQI2QIrr : S2I<0xD6, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
5061 "movq\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVQ_RR>;
5064 //===---------------------------------------------------------------------===//
5065 // Store / copy lower 64-bits of a XMM register.
5067 let Predicates = [UseAVX] in
5068 def : Pat<(int_x86_sse2_storel_dq addr:$dst, VR128:$src),
5069 (VMOVPQI2QImr addr:$dst, VR128:$src)>;
5070 let Predicates = [UseSSE2] in
5071 def : Pat<(int_x86_sse2_storel_dq addr:$dst, VR128:$src),
5072 (MOVPQI2QImr addr:$dst, VR128:$src)>;
5074 let isCodeGenOnly = 1, AddedComplexity = 20 in {
5075 def VMOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
5076 "vmovq\t{$src, $dst|$dst, $src}",
5078 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
5079 (loadi64 addr:$src))))))],
5081 XS, VEX, Requires<[UseAVX]>, Sched<[WriteLoad]>;
5083 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
5084 "movq\t{$src, $dst|$dst, $src}",
5086 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
5087 (loadi64 addr:$src))))))],
5089 XS, Requires<[UseSSE2]>, Sched<[WriteLoad]>;
5092 let Predicates = [UseAVX], AddedComplexity = 20 in {
5093 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
5094 (VMOVZQI2PQIrm addr:$src)>;
5095 def : Pat<(v2i64 (X86vzload addr:$src)),
5096 (VMOVZQI2PQIrm addr:$src)>;
5099 let Predicates = [UseSSE2], AddedComplexity = 20 in {
5100 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
5101 (MOVZQI2PQIrm addr:$src)>;
5102 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
5105 let Predicates = [HasAVX] in {
5106 def : Pat<(v4i64 (alignedX86vzload addr:$src)),
5107 (SUBREG_TO_REG (i32 0), (VMOVAPSrm addr:$src), sub_xmm)>;
5108 def : Pat<(v4i64 (X86vzload addr:$src)),
5109 (SUBREG_TO_REG (i32 0), (VMOVUPSrm addr:$src), sub_xmm)>;
5112 //===---------------------------------------------------------------------===//
5113 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
5114 // IA32 document. movq xmm1, xmm2 does clear the high bits.
5116 let SchedRW = [WriteVecLogic] in {
5117 let AddedComplexity = 15 in
5118 def VMOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5119 "vmovq\t{$src, $dst|$dst, $src}",
5120 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))],
5122 XS, VEX, Requires<[UseAVX]>;
5123 let AddedComplexity = 15 in
5124 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5125 "movq\t{$src, $dst|$dst, $src}",
5126 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))],
5128 XS, Requires<[UseSSE2]>;
5131 let isCodeGenOnly = 1, SchedRW = [WriteVecLogicLd] in {
5132 let AddedComplexity = 20 in
5133 def VMOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5134 "vmovq\t{$src, $dst|$dst, $src}",
5135 [(set VR128:$dst, (v2i64 (X86vzmovl
5136 (loadv2i64 addr:$src))))],
5138 XS, VEX, Requires<[UseAVX]>;
5139 let AddedComplexity = 20 in {
5140 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5141 "movq\t{$src, $dst|$dst, $src}",
5142 [(set VR128:$dst, (v2i64 (X86vzmovl
5143 (loadv2i64 addr:$src))))],
5145 XS, Requires<[UseSSE2]>;
5147 } // isCodeGenOnly, SchedRW
5149 let AddedComplexity = 20 in {
5150 let Predicates = [UseAVX] in {
5151 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
5152 (VMOVZPQILo2PQIrr VR128:$src)>;
5154 let Predicates = [UseSSE2] in {
5155 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
5156 (MOVZPQILo2PQIrr VR128:$src)>;
5160 //===---------------------------------------------------------------------===//
5161 // SSE3 - Replicate Single FP - MOVSHDUP and MOVSLDUP
5162 //===---------------------------------------------------------------------===//
5163 multiclass sse3_replicate_sfp<bits<8> op, SDNode OpNode, string OpcodeStr,
5164 ValueType vt, RegisterClass RC, PatFrag mem_frag,
5165 X86MemOperand x86memop> {
5166 def rr : S3SI<op, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
5167 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5168 [(set RC:$dst, (vt (OpNode RC:$src)))],
5169 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
5170 def rm : S3SI<op, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
5171 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5172 [(set RC:$dst, (OpNode (mem_frag addr:$src)))],
5173 IIC_SSE_MOV_LH>, Sched<[WriteLoad]>;
5176 let Predicates = [HasAVX] in {
5177 defm VMOVSHDUP : sse3_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
5178 v4f32, VR128, loadv4f32, f128mem>, VEX;
5179 defm VMOVSLDUP : sse3_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
5180 v4f32, VR128, loadv4f32, f128mem>, VEX;
5181 defm VMOVSHDUPY : sse3_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
5182 v8f32, VR256, loadv8f32, f256mem>, VEX, VEX_L;
5183 defm VMOVSLDUPY : sse3_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
5184 v8f32, VR256, loadv8f32, f256mem>, VEX, VEX_L;
5186 defm MOVSHDUP : sse3_replicate_sfp<0x16, X86Movshdup, "movshdup", v4f32, VR128,
5187 memopv4f32, f128mem>;
5188 defm MOVSLDUP : sse3_replicate_sfp<0x12, X86Movsldup, "movsldup", v4f32, VR128,
5189 memopv4f32, f128mem>;
5191 let Predicates = [HasAVX] in {
5192 def : Pat<(v4i32 (X86Movshdup VR128:$src)),
5193 (VMOVSHDUPrr VR128:$src)>;
5194 def : Pat<(v4i32 (X86Movshdup (bc_v4i32 (loadv2i64 addr:$src)))),
5195 (VMOVSHDUPrm addr:$src)>;
5196 def : Pat<(v4i32 (X86Movsldup VR128:$src)),
5197 (VMOVSLDUPrr VR128:$src)>;
5198 def : Pat<(v4i32 (X86Movsldup (bc_v4i32 (loadv2i64 addr:$src)))),
5199 (VMOVSLDUPrm addr:$src)>;
5200 def : Pat<(v8i32 (X86Movshdup VR256:$src)),
5201 (VMOVSHDUPYrr VR256:$src)>;
5202 def : Pat<(v8i32 (X86Movshdup (bc_v8i32 (loadv4i64 addr:$src)))),
5203 (VMOVSHDUPYrm addr:$src)>;
5204 def : Pat<(v8i32 (X86Movsldup VR256:$src)),
5205 (VMOVSLDUPYrr VR256:$src)>;
5206 def : Pat<(v8i32 (X86Movsldup (bc_v8i32 (loadv4i64 addr:$src)))),
5207 (VMOVSLDUPYrm addr:$src)>;
5210 let Predicates = [UseSSE3] in {
5211 def : Pat<(v4i32 (X86Movshdup VR128:$src)),
5212 (MOVSHDUPrr VR128:$src)>;
5213 def : Pat<(v4i32 (X86Movshdup (bc_v4i32 (memopv2i64 addr:$src)))),
5214 (MOVSHDUPrm addr:$src)>;
5215 def : Pat<(v4i32 (X86Movsldup VR128:$src)),
5216 (MOVSLDUPrr VR128:$src)>;
5217 def : Pat<(v4i32 (X86Movsldup (bc_v4i32 (memopv2i64 addr:$src)))),
5218 (MOVSLDUPrm addr:$src)>;
5221 //===---------------------------------------------------------------------===//
5222 // SSE3 - Replicate Double FP - MOVDDUP
5223 //===---------------------------------------------------------------------===//
5225 multiclass sse3_replicate_dfp<string OpcodeStr> {
5226 let neverHasSideEffects = 1 in
5227 def rr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5228 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5229 [], IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
5230 def rm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
5231 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5234 (scalar_to_vector (loadf64 addr:$src)))))],
5235 IIC_SSE_MOV_LH>, Sched<[WriteLoad]>;
5238 // FIXME: Merge with above classe when there're patterns for the ymm version
5239 multiclass sse3_replicate_dfp_y<string OpcodeStr> {
5240 def rr : S3DI<0x12, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
5241 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5242 [(set VR256:$dst, (v4f64 (X86Movddup VR256:$src)))]>,
5243 Sched<[WriteFShuffle]>;
5244 def rm : S3DI<0x12, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
5245 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5248 (scalar_to_vector (loadf64 addr:$src)))))]>,
5252 let Predicates = [HasAVX] in {
5253 defm VMOVDDUP : sse3_replicate_dfp<"vmovddup">, VEX;
5254 defm VMOVDDUPY : sse3_replicate_dfp_y<"vmovddup">, VEX, VEX_L;
5257 defm MOVDDUP : sse3_replicate_dfp<"movddup">;
5259 let Predicates = [HasAVX] in {
5260 def : Pat<(X86Movddup (loadv2f64 addr:$src)),
5261 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5262 def : Pat<(X86Movddup (bc_v2f64 (loadv4f32 addr:$src))),
5263 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5264 def : Pat<(X86Movddup (bc_v2f64 (loadv2i64 addr:$src))),
5265 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5266 def : Pat<(X86Movddup (bc_v2f64
5267 (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
5268 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5271 def : Pat<(X86Movddup (loadv4f64 addr:$src)),
5272 (VMOVDDUPYrm addr:$src)>;
5273 def : Pat<(X86Movddup (loadv4i64 addr:$src)),
5274 (VMOVDDUPYrm addr:$src)>;
5275 def : Pat<(X86Movddup (v4i64 (scalar_to_vector (loadi64 addr:$src)))),
5276 (VMOVDDUPYrm addr:$src)>;
5277 def : Pat<(X86Movddup (v4i64 VR256:$src)),
5278 (VMOVDDUPYrr VR256:$src)>;
5281 let Predicates = [UseAVX, OptForSize] in {
5282 def : Pat<(v2f64 (X86VBroadcast (loadf64 addr:$src))),
5283 (VMOVDDUPrm addr:$src)>;
5284 def : Pat<(v2i64 (X86VBroadcast (loadi64 addr:$src))),
5285 (VMOVDDUPrm addr:$src)>;
5288 let Predicates = [UseSSE3] in {
5289 def : Pat<(X86Movddup (memopv2f64 addr:$src)),
5290 (MOVDDUPrm addr:$src)>;
5291 def : Pat<(X86Movddup (bc_v2f64 (memopv4f32 addr:$src))),
5292 (MOVDDUPrm addr:$src)>;
5293 def : Pat<(X86Movddup (bc_v2f64 (memopv2i64 addr:$src))),
5294 (MOVDDUPrm addr:$src)>;
5295 def : Pat<(X86Movddup (bc_v2f64
5296 (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
5297 (MOVDDUPrm addr:$src)>;
5300 //===---------------------------------------------------------------------===//
5301 // SSE3 - Move Unaligned Integer
5302 //===---------------------------------------------------------------------===//
5304 let SchedRW = [WriteLoad] in {
5305 let Predicates = [HasAVX] in {
5306 def VLDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5307 "vlddqu\t{$src, $dst|$dst, $src}",
5308 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>, VEX;
5309 def VLDDQUYrm : S3DI<0xF0, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
5310 "vlddqu\t{$src, $dst|$dst, $src}",
5311 [(set VR256:$dst, (int_x86_avx_ldu_dq_256 addr:$src))]>,
5314 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5315 "lddqu\t{$src, $dst|$dst, $src}",
5316 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))],
5320 //===---------------------------------------------------------------------===//
5321 // SSE3 - Arithmetic
5322 //===---------------------------------------------------------------------===//
5324 multiclass sse3_addsub<Intrinsic Int, string OpcodeStr, RegisterClass RC,
5325 X86MemOperand x86memop, OpndItins itins,
5327 def rr : I<0xD0, MRMSrcReg,
5328 (outs RC:$dst), (ins RC:$src1, RC:$src2),
5330 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5331 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5332 [(set RC:$dst, (Int RC:$src1, RC:$src2))], itins.rr>,
5333 Sched<[itins.Sched]>;
5334 def rm : I<0xD0, MRMSrcMem,
5335 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5337 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5338 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5339 [(set RC:$dst, (Int RC:$src1, (memop addr:$src2)))], itins.rr>,
5340 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5343 let Predicates = [HasAVX] in {
5344 let ExeDomain = SSEPackedSingle in {
5345 defm VADDSUBPS : sse3_addsub<int_x86_sse3_addsub_ps, "vaddsubps", VR128,
5346 f128mem, SSE_ALU_F32P, 0>, XD, VEX_4V;
5347 defm VADDSUBPSY : sse3_addsub<int_x86_avx_addsub_ps_256, "vaddsubps", VR256,
5348 f256mem, SSE_ALU_F32P, 0>, XD, VEX_4V, VEX_L;
5350 let ExeDomain = SSEPackedDouble in {
5351 defm VADDSUBPD : sse3_addsub<int_x86_sse3_addsub_pd, "vaddsubpd", VR128,
5352 f128mem, SSE_ALU_F64P, 0>, PD, VEX_4V;
5353 defm VADDSUBPDY : sse3_addsub<int_x86_avx_addsub_pd_256, "vaddsubpd", VR256,
5354 f256mem, SSE_ALU_F64P, 0>, PD, VEX_4V, VEX_L;
5357 let Constraints = "$src1 = $dst", Predicates = [UseSSE3] in {
5358 let ExeDomain = SSEPackedSingle in
5359 defm ADDSUBPS : sse3_addsub<int_x86_sse3_addsub_ps, "addsubps", VR128,
5360 f128mem, SSE_ALU_F32P>, XD;
5361 let ExeDomain = SSEPackedDouble in
5362 defm ADDSUBPD : sse3_addsub<int_x86_sse3_addsub_pd, "addsubpd", VR128,
5363 f128mem, SSE_ALU_F64P>, PD;
5366 // Patterns used to select 'addsub' instructions.
5367 let Predicates = [HasAVX] in {
5368 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 VR128:$rhs))),
5369 (VADDSUBPSrr VR128:$lhs, VR128:$rhs)>;
5370 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 (memop addr:$rhs)))),
5371 (VADDSUBPSrm VR128:$lhs, f128mem:$rhs)>;
5372 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 VR128:$rhs))),
5373 (VADDSUBPDrr VR128:$lhs, VR128:$rhs)>;
5374 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 (memop addr:$rhs)))),
5375 (VADDSUBPDrm VR128:$lhs, f128mem:$rhs)>;
5377 def : Pat<(v8f32 (X86Addsub (v8f32 VR256:$lhs), (v8f32 VR256:$rhs))),
5378 (VADDSUBPSYrr VR256:$lhs, VR256:$rhs)>;
5379 def : Pat<(v8f32 (X86Addsub (v8f32 VR256:$lhs), (v8f32 (memop addr:$rhs)))),
5380 (VADDSUBPSYrm VR256:$lhs, f256mem:$rhs)>;
5381 def : Pat<(v4f64 (X86Addsub (v4f64 VR256:$lhs), (v4f64 VR256:$rhs))),
5382 (VADDSUBPDYrr VR256:$lhs, VR256:$rhs)>;
5383 def : Pat<(v4f64 (X86Addsub (v4f64 VR256:$lhs), (v4f64 (memop addr:$rhs)))),
5384 (VADDSUBPDYrm VR256:$lhs, f256mem:$rhs)>;
5387 let Predicates = [UseSSE3] in {
5388 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 VR128:$rhs))),
5389 (ADDSUBPSrr VR128:$lhs, VR128:$rhs)>;
5390 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 (memop addr:$rhs)))),
5391 (ADDSUBPSrm VR128:$lhs, f128mem:$rhs)>;
5392 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 VR128:$rhs))),
5393 (ADDSUBPDrr VR128:$lhs, VR128:$rhs)>;
5394 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 (memop addr:$rhs)))),
5395 (ADDSUBPDrm VR128:$lhs, f128mem:$rhs)>;
5398 //===---------------------------------------------------------------------===//
5399 // SSE3 Instructions
5400 //===---------------------------------------------------------------------===//
5403 multiclass S3D_Int<bits<8> o, string OpcodeStr, ValueType vt, RegisterClass RC,
5404 X86MemOperand x86memop, SDNode OpNode, bit Is2Addr = 1> {
5405 def rr : S3DI<o, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
5407 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5408 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5409 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], IIC_SSE_HADDSUB_RR>,
5412 def rm : S3DI<o, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5414 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5415 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5416 [(set RC:$dst, (vt (OpNode RC:$src1, (memop addr:$src2))))],
5417 IIC_SSE_HADDSUB_RM>, Sched<[WriteFAddLd, ReadAfterLd]>;
5419 multiclass S3_Int<bits<8> o, string OpcodeStr, ValueType vt, RegisterClass RC,
5420 X86MemOperand x86memop, SDNode OpNode, bit Is2Addr = 1> {
5421 def rr : S3I<o, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
5423 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5424 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5425 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], IIC_SSE_HADDSUB_RR>,
5428 def rm : S3I<o, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5430 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5431 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5432 [(set RC:$dst, (vt (OpNode RC:$src1, (memop addr:$src2))))],
5433 IIC_SSE_HADDSUB_RM>, Sched<[WriteFAddLd, ReadAfterLd]>;
5436 let Predicates = [HasAVX] in {
5437 let ExeDomain = SSEPackedSingle in {
5438 defm VHADDPS : S3D_Int<0x7C, "vhaddps", v4f32, VR128, f128mem,
5439 X86fhadd, 0>, VEX_4V;
5440 defm VHSUBPS : S3D_Int<0x7D, "vhsubps", v4f32, VR128, f128mem,
5441 X86fhsub, 0>, VEX_4V;
5442 defm VHADDPSY : S3D_Int<0x7C, "vhaddps", v8f32, VR256, f256mem,
5443 X86fhadd, 0>, VEX_4V, VEX_L;
5444 defm VHSUBPSY : S3D_Int<0x7D, "vhsubps", v8f32, VR256, f256mem,
5445 X86fhsub, 0>, VEX_4V, VEX_L;
5447 let ExeDomain = SSEPackedDouble in {
5448 defm VHADDPD : S3_Int <0x7C, "vhaddpd", v2f64, VR128, f128mem,
5449 X86fhadd, 0>, VEX_4V;
5450 defm VHSUBPD : S3_Int <0x7D, "vhsubpd", v2f64, VR128, f128mem,
5451 X86fhsub, 0>, VEX_4V;
5452 defm VHADDPDY : S3_Int <0x7C, "vhaddpd", v4f64, VR256, f256mem,
5453 X86fhadd, 0>, VEX_4V, VEX_L;
5454 defm VHSUBPDY : S3_Int <0x7D, "vhsubpd", v4f64, VR256, f256mem,
5455 X86fhsub, 0>, VEX_4V, VEX_L;
5459 let Constraints = "$src1 = $dst" in {
5460 let ExeDomain = SSEPackedSingle in {
5461 defm HADDPS : S3D_Int<0x7C, "haddps", v4f32, VR128, f128mem, X86fhadd>;
5462 defm HSUBPS : S3D_Int<0x7D, "hsubps", v4f32, VR128, f128mem, X86fhsub>;
5464 let ExeDomain = SSEPackedDouble in {
5465 defm HADDPD : S3_Int<0x7C, "haddpd", v2f64, VR128, f128mem, X86fhadd>;
5466 defm HSUBPD : S3_Int<0x7D, "hsubpd", v2f64, VR128, f128mem, X86fhsub>;
5470 //===---------------------------------------------------------------------===//
5471 // SSSE3 - Packed Absolute Instructions
5472 //===---------------------------------------------------------------------===//
5475 /// SS3I_unop_rm_int - Simple SSSE3 unary op whose type can be v*{i8,i16,i32}.
5476 multiclass SS3I_unop_rm_int<bits<8> opc, string OpcodeStr,
5477 Intrinsic IntId128> {
5478 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
5480 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5481 [(set VR128:$dst, (IntId128 VR128:$src))], IIC_SSE_PABS_RR>,
5482 Sched<[WriteVecALU]>;
5484 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
5486 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5489 (bitconvert (memopv2i64 addr:$src))))], IIC_SSE_PABS_RM>,
5490 Sched<[WriteVecALULd]>;
5493 /// SS3I_unop_rm_int_y - Simple SSSE3 unary op whose type can be v*{i8,i16,i32}.
5494 multiclass SS3I_unop_rm_int_y<bits<8> opc, string OpcodeStr,
5495 Intrinsic IntId256> {
5496 def rr256 : SS38I<opc, MRMSrcReg, (outs VR256:$dst),
5498 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5499 [(set VR256:$dst, (IntId256 VR256:$src))]>,
5500 Sched<[WriteVecALU]>;
5502 def rm256 : SS38I<opc, MRMSrcMem, (outs VR256:$dst),
5504 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5507 (bitconvert (memopv4i64 addr:$src))))]>,
5508 Sched<[WriteVecALULd]>;
5511 // Helper fragments to match sext vXi1 to vXiY.
5512 def v16i1sextv16i8 : PatLeaf<(v16i8 (X86pcmpgt (bc_v16i8 (v4i32 immAllZerosV)),
5514 def v8i1sextv8i16 : PatLeaf<(v8i16 (X86vsrai VR128:$src, (i8 15)))>;
5515 def v4i1sextv4i32 : PatLeaf<(v4i32 (X86vsrai VR128:$src, (i8 31)))>;
5516 def v32i1sextv32i8 : PatLeaf<(v32i8 (X86pcmpgt (bc_v32i8 (v8i32 immAllZerosV)),
5518 def v16i1sextv16i16: PatLeaf<(v16i16 (X86vsrai VR256:$src, (i8 15)))>;
5519 def v8i1sextv8i32 : PatLeaf<(v8i32 (X86vsrai VR256:$src, (i8 31)))>;
5521 let Predicates = [HasAVX] in {
5522 defm VPABSB : SS3I_unop_rm_int<0x1C, "vpabsb",
5523 int_x86_ssse3_pabs_b_128>, VEX;
5524 defm VPABSW : SS3I_unop_rm_int<0x1D, "vpabsw",
5525 int_x86_ssse3_pabs_w_128>, VEX;
5526 defm VPABSD : SS3I_unop_rm_int<0x1E, "vpabsd",
5527 int_x86_ssse3_pabs_d_128>, VEX;
5530 (bc_v2i64 (v16i1sextv16i8)),
5531 (bc_v2i64 (add (v16i8 VR128:$src), (v16i1sextv16i8)))),
5532 (VPABSBrr128 VR128:$src)>;
5534 (bc_v2i64 (v8i1sextv8i16)),
5535 (bc_v2i64 (add (v8i16 VR128:$src), (v8i1sextv8i16)))),
5536 (VPABSWrr128 VR128:$src)>;
5538 (bc_v2i64 (v4i1sextv4i32)),
5539 (bc_v2i64 (add (v4i32 VR128:$src), (v4i1sextv4i32)))),
5540 (VPABSDrr128 VR128:$src)>;
5543 let Predicates = [HasAVX2] in {
5544 defm VPABSB : SS3I_unop_rm_int_y<0x1C, "vpabsb",
5545 int_x86_avx2_pabs_b>, VEX, VEX_L;
5546 defm VPABSW : SS3I_unop_rm_int_y<0x1D, "vpabsw",
5547 int_x86_avx2_pabs_w>, VEX, VEX_L;
5548 defm VPABSD : SS3I_unop_rm_int_y<0x1E, "vpabsd",
5549 int_x86_avx2_pabs_d>, VEX, VEX_L;
5552 (bc_v4i64 (v32i1sextv32i8)),
5553 (bc_v4i64 (add (v32i8 VR256:$src), (v32i1sextv32i8)))),
5554 (VPABSBrr256 VR256:$src)>;
5556 (bc_v4i64 (v16i1sextv16i16)),
5557 (bc_v4i64 (add (v16i16 VR256:$src), (v16i1sextv16i16)))),
5558 (VPABSWrr256 VR256:$src)>;
5560 (bc_v4i64 (v8i1sextv8i32)),
5561 (bc_v4i64 (add (v8i32 VR256:$src), (v8i1sextv8i32)))),
5562 (VPABSDrr256 VR256:$src)>;
5565 defm PABSB : SS3I_unop_rm_int<0x1C, "pabsb",
5566 int_x86_ssse3_pabs_b_128>;
5567 defm PABSW : SS3I_unop_rm_int<0x1D, "pabsw",
5568 int_x86_ssse3_pabs_w_128>;
5569 defm PABSD : SS3I_unop_rm_int<0x1E, "pabsd",
5570 int_x86_ssse3_pabs_d_128>;
5572 let Predicates = [HasSSSE3] in {
5574 (bc_v2i64 (v16i1sextv16i8)),
5575 (bc_v2i64 (add (v16i8 VR128:$src), (v16i1sextv16i8)))),
5576 (PABSBrr128 VR128:$src)>;
5578 (bc_v2i64 (v8i1sextv8i16)),
5579 (bc_v2i64 (add (v8i16 VR128:$src), (v8i1sextv8i16)))),
5580 (PABSWrr128 VR128:$src)>;
5582 (bc_v2i64 (v4i1sextv4i32)),
5583 (bc_v2i64 (add (v4i32 VR128:$src), (v4i1sextv4i32)))),
5584 (PABSDrr128 VR128:$src)>;
5587 //===---------------------------------------------------------------------===//
5588 // SSSE3 - Packed Binary Operator Instructions
5589 //===---------------------------------------------------------------------===//
5591 let Sched = WriteVecALU in {
5592 def SSE_PHADDSUBD : OpndItins<
5593 IIC_SSE_PHADDSUBD_RR, IIC_SSE_PHADDSUBD_RM
5595 def SSE_PHADDSUBSW : OpndItins<
5596 IIC_SSE_PHADDSUBSW_RR, IIC_SSE_PHADDSUBSW_RM
5598 def SSE_PHADDSUBW : OpndItins<
5599 IIC_SSE_PHADDSUBW_RR, IIC_SSE_PHADDSUBW_RM
5602 let Sched = WriteShuffle in
5603 def SSE_PSHUFB : OpndItins<
5604 IIC_SSE_PSHUFB_RR, IIC_SSE_PSHUFB_RM
5606 let Sched = WriteVecALU in
5607 def SSE_PSIGN : OpndItins<
5608 IIC_SSE_PSIGN_RR, IIC_SSE_PSIGN_RM
5610 let Sched = WriteVecIMul in
5611 def SSE_PMULHRSW : OpndItins<
5612 IIC_SSE_PMULHRSW, IIC_SSE_PMULHRSW
5615 /// SS3I_binop_rm - Simple SSSE3 bin op
5616 multiclass SS3I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
5617 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
5618 X86MemOperand x86memop, OpndItins itins,
5620 let isCommutable = 1 in
5621 def rr : SS38I<opc, MRMSrcReg, (outs RC:$dst),
5622 (ins RC:$src1, RC:$src2),
5624 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5625 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5626 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))], itins.rr>,
5627 Sched<[itins.Sched]>;
5628 def rm : SS38I<opc, MRMSrcMem, (outs RC:$dst),
5629 (ins RC:$src1, x86memop:$src2),
5631 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5632 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5634 (OpVT (OpNode RC:$src1,
5635 (bitconvert (memop_frag addr:$src2)))))], itins.rm>,
5636 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5639 /// SS3I_binop_rm_int - Simple SSSE3 bin op whose type can be v*{i8,i16,i32}.
5640 multiclass SS3I_binop_rm_int<bits<8> opc, string OpcodeStr,
5641 Intrinsic IntId128, OpndItins itins,
5643 let isCommutable = 1 in
5644 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
5645 (ins VR128:$src1, VR128:$src2),
5647 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5648 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5649 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
5650 Sched<[itins.Sched]>;
5651 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
5652 (ins VR128:$src1, i128mem:$src2),
5654 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5655 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5657 (IntId128 VR128:$src1,
5658 (bitconvert (memopv2i64 addr:$src2))))]>,
5659 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5662 multiclass SS3I_binop_rm_int_y<bits<8> opc, string OpcodeStr,
5664 X86FoldableSchedWrite Sched> {
5665 let isCommutable = 1 in
5666 def rr256 : SS38I<opc, MRMSrcReg, (outs VR256:$dst),
5667 (ins VR256:$src1, VR256:$src2),
5668 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5669 [(set VR256:$dst, (IntId256 VR256:$src1, VR256:$src2))]>,
5671 def rm256 : SS38I<opc, MRMSrcMem, (outs VR256:$dst),
5672 (ins VR256:$src1, i256mem:$src2),
5673 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5675 (IntId256 VR256:$src1, (bitconvert (loadv4i64 addr:$src2))))]>,
5676 Sched<[Sched.Folded, ReadAfterLd]>;
5679 let ImmT = NoImm, Predicates = [HasAVX] in {
5680 let isCommutable = 0 in {
5681 defm VPHADDW : SS3I_binop_rm<0x01, "vphaddw", X86hadd, v8i16, VR128,
5683 SSE_PHADDSUBW, 0>, VEX_4V;
5684 defm VPHADDD : SS3I_binop_rm<0x02, "vphaddd", X86hadd, v4i32, VR128,
5686 SSE_PHADDSUBD, 0>, VEX_4V;
5687 defm VPHSUBW : SS3I_binop_rm<0x05, "vphsubw", X86hsub, v8i16, VR128,
5689 SSE_PHADDSUBW, 0>, VEX_4V;
5690 defm VPHSUBD : SS3I_binop_rm<0x06, "vphsubd", X86hsub, v4i32, VR128,
5692 SSE_PHADDSUBD, 0>, VEX_4V;
5693 defm VPSIGNB : SS3I_binop_rm<0x08, "vpsignb", X86psign, v16i8, VR128,
5695 SSE_PSIGN, 0>, VEX_4V;
5696 defm VPSIGNW : SS3I_binop_rm<0x09, "vpsignw", X86psign, v8i16, VR128,
5698 SSE_PSIGN, 0>, VEX_4V;
5699 defm VPSIGND : SS3I_binop_rm<0x0A, "vpsignd", X86psign, v4i32, VR128,
5701 SSE_PSIGN, 0>, VEX_4V;
5702 defm VPSHUFB : SS3I_binop_rm<0x00, "vpshufb", X86pshufb, v16i8, VR128,
5704 SSE_PSHUFB, 0>, VEX_4V;
5705 defm VPHADDSW : SS3I_binop_rm_int<0x03, "vphaddsw",
5706 int_x86_ssse3_phadd_sw_128,
5707 SSE_PHADDSUBSW, 0>, VEX_4V;
5708 defm VPHSUBSW : SS3I_binop_rm_int<0x07, "vphsubsw",
5709 int_x86_ssse3_phsub_sw_128,
5710 SSE_PHADDSUBSW, 0>, VEX_4V;
5711 defm VPMADDUBSW : SS3I_binop_rm_int<0x04, "vpmaddubsw",
5712 int_x86_ssse3_pmadd_ub_sw_128,
5713 SSE_PMADD, 0>, VEX_4V;
5715 defm VPMULHRSW : SS3I_binop_rm_int<0x0B, "vpmulhrsw",
5716 int_x86_ssse3_pmul_hr_sw_128,
5717 SSE_PMULHRSW, 0>, VEX_4V;
5720 let ImmT = NoImm, Predicates = [HasAVX2] in {
5721 let isCommutable = 0 in {
5722 defm VPHADDWY : SS3I_binop_rm<0x01, "vphaddw", X86hadd, v16i16, VR256,
5724 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5725 defm VPHADDDY : SS3I_binop_rm<0x02, "vphaddd", X86hadd, v8i32, VR256,
5727 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5728 defm VPHSUBWY : SS3I_binop_rm<0x05, "vphsubw", X86hsub, v16i16, VR256,
5730 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5731 defm VPHSUBDY : SS3I_binop_rm<0x06, "vphsubd", X86hsub, v8i32, VR256,
5733 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5734 defm VPSIGNBY : SS3I_binop_rm<0x08, "vpsignb", X86psign, v32i8, VR256,
5736 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5737 defm VPSIGNWY : SS3I_binop_rm<0x09, "vpsignw", X86psign, v16i16, VR256,
5739 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5740 defm VPSIGNDY : SS3I_binop_rm<0x0A, "vpsignd", X86psign, v8i32, VR256,
5742 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5743 defm VPSHUFBY : SS3I_binop_rm<0x00, "vpshufb", X86pshufb, v32i8, VR256,
5745 SSE_PSHUFB, 0>, VEX_4V, VEX_L;
5746 defm VPHADDSW : SS3I_binop_rm_int_y<0x03, "vphaddsw",
5747 int_x86_avx2_phadd_sw,
5748 WriteVecALU>, VEX_4V, VEX_L;
5749 defm VPHSUBSW : SS3I_binop_rm_int_y<0x07, "vphsubsw",
5750 int_x86_avx2_phsub_sw,
5751 WriteVecALU>, VEX_4V, VEX_L;
5752 defm VPMADDUBSW : SS3I_binop_rm_int_y<0x04, "vpmaddubsw",
5753 int_x86_avx2_pmadd_ub_sw,
5754 WriteVecIMul>, VEX_4V, VEX_L;
5756 defm VPMULHRSW : SS3I_binop_rm_int_y<0x0B, "vpmulhrsw",
5757 int_x86_avx2_pmul_hr_sw,
5758 WriteVecIMul>, VEX_4V, VEX_L;
5761 // None of these have i8 immediate fields.
5762 let ImmT = NoImm, Constraints = "$src1 = $dst" in {
5763 let isCommutable = 0 in {
5764 defm PHADDW : SS3I_binop_rm<0x01, "phaddw", X86hadd, v8i16, VR128,
5765 memopv2i64, i128mem, SSE_PHADDSUBW>;
5766 defm PHADDD : SS3I_binop_rm<0x02, "phaddd", X86hadd, v4i32, VR128,
5767 memopv2i64, i128mem, SSE_PHADDSUBD>;
5768 defm PHSUBW : SS3I_binop_rm<0x05, "phsubw", X86hsub, v8i16, VR128,
5769 memopv2i64, i128mem, SSE_PHADDSUBW>;
5770 defm PHSUBD : SS3I_binop_rm<0x06, "phsubd", X86hsub, v4i32, VR128,
5771 memopv2i64, i128mem, SSE_PHADDSUBD>;
5772 defm PSIGNB : SS3I_binop_rm<0x08, "psignb", X86psign, v16i8, VR128,
5773 memopv2i64, i128mem, SSE_PSIGN>;
5774 defm PSIGNW : SS3I_binop_rm<0x09, "psignw", X86psign, v8i16, VR128,
5775 memopv2i64, i128mem, SSE_PSIGN>;
5776 defm PSIGND : SS3I_binop_rm<0x0A, "psignd", X86psign, v4i32, VR128,
5777 memopv2i64, i128mem, SSE_PSIGN>;
5778 defm PSHUFB : SS3I_binop_rm<0x00, "pshufb", X86pshufb, v16i8, VR128,
5779 memopv2i64, i128mem, SSE_PSHUFB>;
5780 defm PHADDSW : SS3I_binop_rm_int<0x03, "phaddsw",
5781 int_x86_ssse3_phadd_sw_128,
5783 defm PHSUBSW : SS3I_binop_rm_int<0x07, "phsubsw",
5784 int_x86_ssse3_phsub_sw_128,
5786 defm PMADDUBSW : SS3I_binop_rm_int<0x04, "pmaddubsw",
5787 int_x86_ssse3_pmadd_ub_sw_128, SSE_PMADD>;
5789 defm PMULHRSW : SS3I_binop_rm_int<0x0B, "pmulhrsw",
5790 int_x86_ssse3_pmul_hr_sw_128,
5794 //===---------------------------------------------------------------------===//
5795 // SSSE3 - Packed Align Instruction Patterns
5796 //===---------------------------------------------------------------------===//
5798 multiclass ssse3_palignr<string asm, bit Is2Addr = 1> {
5799 let neverHasSideEffects = 1 in {
5800 def R128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
5801 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
5803 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
5805 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
5806 [], IIC_SSE_PALIGNRR>, Sched<[WriteShuffle]>;
5808 def R128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
5809 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
5811 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
5813 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
5814 [], IIC_SSE_PALIGNRM>, Sched<[WriteShuffleLd, ReadAfterLd]>;
5818 multiclass ssse3_palignr_y<string asm, bit Is2Addr = 1> {
5819 let neverHasSideEffects = 1 in {
5820 def R256rr : SS3AI<0x0F, MRMSrcReg, (outs VR256:$dst),
5821 (ins VR256:$src1, VR256:$src2, i8imm:$src3),
5823 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
5824 []>, Sched<[WriteShuffle]>;
5826 def R256rm : SS3AI<0x0F, MRMSrcMem, (outs VR256:$dst),
5827 (ins VR256:$src1, i256mem:$src2, i8imm:$src3),
5829 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
5830 []>, Sched<[WriteShuffleLd, ReadAfterLd]>;
5834 let Predicates = [HasAVX] in
5835 defm VPALIGN : ssse3_palignr<"vpalignr", 0>, VEX_4V;
5836 let Predicates = [HasAVX2] in
5837 defm VPALIGN : ssse3_palignr_y<"vpalignr", 0>, VEX_4V, VEX_L;
5838 let Constraints = "$src1 = $dst", Predicates = [UseSSSE3] in
5839 defm PALIGN : ssse3_palignr<"palignr">;
5841 let Predicates = [HasAVX2] in {
5842 def : Pat<(v8i32 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5843 (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
5844 def : Pat<(v8f32 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5845 (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
5846 def : Pat<(v16i16 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5847 (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
5848 def : Pat<(v32i8 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5849 (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
5852 let Predicates = [HasAVX] in {
5853 def : Pat<(v4i32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5854 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5855 def : Pat<(v4f32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5856 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5857 def : Pat<(v8i16 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5858 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5859 def : Pat<(v16i8 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5860 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5863 let Predicates = [UseSSSE3] in {
5864 def : Pat<(v4i32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5865 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5866 def : Pat<(v4f32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5867 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5868 def : Pat<(v8i16 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5869 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5870 def : Pat<(v16i8 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5871 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5874 //===---------------------------------------------------------------------===//
5875 // SSSE3 - Thread synchronization
5876 //===---------------------------------------------------------------------===//
5878 let SchedRW = [WriteSystem] in {
5879 let usesCustomInserter = 1 in {
5880 def MONITOR : PseudoI<(outs), (ins i32mem:$src1, GR32:$src2, GR32:$src3),
5881 [(int_x86_sse3_monitor addr:$src1, GR32:$src2, GR32:$src3)]>,
5882 Requires<[HasSSE3]>;
5885 let Uses = [EAX, ECX, EDX] in
5886 def MONITORrrr : I<0x01, MRM_C8, (outs), (ins), "monitor", [], IIC_SSE_MONITOR>,
5887 TB, Requires<[HasSSE3]>;
5888 let Uses = [ECX, EAX] in
5889 def MWAITrr : I<0x01, MRM_C9, (outs), (ins), "mwait",
5890 [(int_x86_sse3_mwait ECX, EAX)], IIC_SSE_MWAIT>,
5891 TB, Requires<[HasSSE3]>;
5894 def : InstAlias<"mwait\t{%eax, %ecx|ecx, eax}", (MWAITrr)>, Requires<[Not64BitMode]>;
5895 def : InstAlias<"mwait\t{%rax, %rcx|rcx, rax}", (MWAITrr)>, Requires<[In64BitMode]>;
5897 def : InstAlias<"monitor\t{%eax, %ecx, %edx|edx, ecx, eax}", (MONITORrrr)>,
5898 Requires<[Not64BitMode]>;
5899 def : InstAlias<"monitor\t{%rax, %rcx, %rdx|rdx, rcx, rax}", (MONITORrrr)>,
5900 Requires<[In64BitMode]>;
5902 //===----------------------------------------------------------------------===//
5903 // SSE4.1 - Packed Move with Sign/Zero Extend
5904 //===----------------------------------------------------------------------===//
5906 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId,
5907 OpndItins itins = DEFAULT_ITINS> {
5908 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5909 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5910 [(set VR128:$dst, (IntId VR128:$src))], itins.rr>,
5911 Sched<[itins.Sched]>;
5913 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
5914 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5916 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))],
5917 itins.rm>, Sched<[itins.Sched.Folded]>;
5920 multiclass SS41I_binop_rm_int16_y<bits<8> opc, string OpcodeStr,
5921 Intrinsic IntId, X86FoldableSchedWrite Sched> {
5922 def Yrr : SS48I<opc, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
5923 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5924 [(set VR256:$dst, (IntId VR128:$src))]>, Sched<[Sched]>;
5926 def Yrm : SS48I<opc, MRMSrcMem, (outs VR256:$dst), (ins i128mem:$src),
5927 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5928 [(set VR256:$dst, (IntId (load addr:$src)))]>,
5929 Sched<[Sched.Folded]>;
5932 let Predicates = [HasAVX] in {
5933 defm VPMOVSXBW : SS41I_binop_rm_int8<0x20, "vpmovsxbw",
5934 int_x86_sse41_pmovsxbw,
5935 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
5936 defm VPMOVSXWD : SS41I_binop_rm_int8<0x23, "vpmovsxwd",
5937 int_x86_sse41_pmovsxwd,
5938 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
5939 defm VPMOVSXDQ : SS41I_binop_rm_int8<0x25, "vpmovsxdq",
5940 int_x86_sse41_pmovsxdq,
5941 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
5942 defm VPMOVZXBW : SS41I_binop_rm_int8<0x30, "vpmovzxbw",
5943 int_x86_sse41_pmovzxbw,
5944 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
5945 defm VPMOVZXWD : SS41I_binop_rm_int8<0x33, "vpmovzxwd",
5946 int_x86_sse41_pmovzxwd,
5947 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
5948 defm VPMOVZXDQ : SS41I_binop_rm_int8<0x35, "vpmovzxdq",
5949 int_x86_sse41_pmovzxdq,
5950 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
5953 let Predicates = [HasAVX2] in {
5954 defm VPMOVSXBW : SS41I_binop_rm_int16_y<0x20, "vpmovsxbw",
5955 int_x86_avx2_pmovsxbw,
5956 WriteShuffle>, VEX, VEX_L;
5957 defm VPMOVSXWD : SS41I_binop_rm_int16_y<0x23, "vpmovsxwd",
5958 int_x86_avx2_pmovsxwd,
5959 WriteShuffle>, VEX, VEX_L;
5960 defm VPMOVSXDQ : SS41I_binop_rm_int16_y<0x25, "vpmovsxdq",
5961 int_x86_avx2_pmovsxdq,
5962 WriteShuffle>, VEX, VEX_L;
5963 defm VPMOVZXBW : SS41I_binop_rm_int16_y<0x30, "vpmovzxbw",
5964 int_x86_avx2_pmovzxbw,
5965 WriteShuffle>, VEX, VEX_L;
5966 defm VPMOVZXWD : SS41I_binop_rm_int16_y<0x33, "vpmovzxwd",
5967 int_x86_avx2_pmovzxwd,
5968 WriteShuffle>, VEX, VEX_L;
5969 defm VPMOVZXDQ : SS41I_binop_rm_int16_y<0x35, "vpmovzxdq",
5970 int_x86_avx2_pmovzxdq,
5971 WriteShuffle>, VEX, VEX_L;
5974 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw,
5975 SSE_INTALU_ITINS_SHUFF_P>;
5976 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd,
5977 SSE_INTALU_ITINS_SHUFF_P>;
5978 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq,
5979 SSE_INTALU_ITINS_SHUFF_P>;
5980 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw,
5981 SSE_INTALU_ITINS_SHUFF_P>;
5982 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd,
5983 SSE_INTALU_ITINS_SHUFF_P>;
5984 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq,
5985 SSE_INTALU_ITINS_SHUFF_P>;
5987 let Predicates = [HasAVX] in {
5988 // Common patterns involving scalar load.
5989 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
5990 (VPMOVSXBWrm addr:$src)>;
5991 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
5992 (VPMOVSXBWrm addr:$src)>;
5993 def : Pat<(int_x86_sse41_pmovsxbw (bc_v16i8 (loadv2i64 addr:$src))),
5994 (VPMOVSXBWrm addr:$src)>;
5996 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
5997 (VPMOVSXWDrm addr:$src)>;
5998 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
5999 (VPMOVSXWDrm addr:$src)>;
6000 def : Pat<(int_x86_sse41_pmovsxwd (bc_v8i16 (loadv2i64 addr:$src))),
6001 (VPMOVSXWDrm addr:$src)>;
6003 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
6004 (VPMOVSXDQrm addr:$src)>;
6005 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
6006 (VPMOVSXDQrm addr:$src)>;
6007 def : Pat<(int_x86_sse41_pmovsxdq (bc_v4i32 (loadv2i64 addr:$src))),
6008 (VPMOVSXDQrm addr:$src)>;
6010 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
6011 (VPMOVZXBWrm addr:$src)>;
6012 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
6013 (VPMOVZXBWrm addr:$src)>;
6014 def : Pat<(int_x86_sse41_pmovzxbw (bc_v16i8 (loadv2i64 addr:$src))),
6015 (VPMOVZXBWrm addr:$src)>;
6017 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
6018 (VPMOVZXWDrm addr:$src)>;
6019 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
6020 (VPMOVZXWDrm addr:$src)>;
6021 def : Pat<(int_x86_sse41_pmovzxwd (bc_v8i16 (loadv2i64 addr:$src))),
6022 (VPMOVZXWDrm addr:$src)>;
6024 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
6025 (VPMOVZXDQrm addr:$src)>;
6026 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
6027 (VPMOVZXDQrm addr:$src)>;
6028 def : Pat<(int_x86_sse41_pmovzxdq (bc_v4i32 (loadv2i64 addr:$src))),
6029 (VPMOVZXDQrm addr:$src)>;
6032 let Predicates = [UseSSE41] in {
6033 // Common patterns involving scalar load.
6034 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
6035 (PMOVSXBWrm addr:$src)>;
6036 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
6037 (PMOVSXBWrm addr:$src)>;
6038 def : Pat<(int_x86_sse41_pmovsxbw (bc_v16i8 (loadv2i64 addr:$src))),
6039 (PMOVSXBWrm addr:$src)>;
6041 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
6042 (PMOVSXWDrm addr:$src)>;
6043 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
6044 (PMOVSXWDrm addr:$src)>;
6045 def : Pat<(int_x86_sse41_pmovsxwd (bc_v8i16 (loadv2i64 addr:$src))),
6046 (PMOVSXWDrm addr:$src)>;
6048 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
6049 (PMOVSXDQrm addr:$src)>;
6050 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
6051 (PMOVSXDQrm addr:$src)>;
6052 def : Pat<(int_x86_sse41_pmovsxdq (bc_v4i32 (loadv2i64 addr:$src))),
6053 (PMOVSXDQrm addr:$src)>;
6055 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
6056 (PMOVZXBWrm addr:$src)>;
6057 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
6058 (PMOVZXBWrm addr:$src)>;
6059 def : Pat<(int_x86_sse41_pmovzxbw (bc_v16i8 (loadv2i64 addr:$src))),
6060 (PMOVZXBWrm addr:$src)>;
6062 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
6063 (PMOVZXWDrm addr:$src)>;
6064 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
6065 (PMOVZXWDrm addr:$src)>;
6066 def : Pat<(int_x86_sse41_pmovzxwd (bc_v8i16 (loadv2i64 addr:$src))),
6067 (PMOVZXWDrm addr:$src)>;
6069 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
6070 (PMOVZXDQrm addr:$src)>;
6071 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
6072 (PMOVZXDQrm addr:$src)>;
6073 def : Pat<(int_x86_sse41_pmovzxdq (bc_v4i32 (loadv2i64 addr:$src))),
6074 (PMOVZXDQrm addr:$src)>;
6077 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId,
6078 OpndItins itins = DEFAULT_ITINS> {
6079 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
6080 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6081 [(set VR128:$dst, (IntId VR128:$src))], itins.rr>,
6082 Sched<[itins.Sched]>;
6084 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
6085 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6087 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))],
6088 itins.rm>, Sched<[itins.Sched.Folded]>;
6091 multiclass SS41I_binop_rm_int8_y<bits<8> opc, string OpcodeStr,
6092 Intrinsic IntId, X86FoldableSchedWrite Sched> {
6093 def Yrr : SS48I<opc, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
6094 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6095 [(set VR256:$dst, (IntId VR128:$src))]>, Sched<[Sched]>;
6097 def Yrm : SS48I<opc, MRMSrcMem, (outs VR256:$dst), (ins i32mem:$src),
6098 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6100 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
6101 Sched<[Sched.Folded]>;
6104 let Predicates = [HasAVX] in {
6105 defm VPMOVSXBD : SS41I_binop_rm_int4<0x21, "vpmovsxbd", int_x86_sse41_pmovsxbd,
6106 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
6107 defm VPMOVSXWQ : SS41I_binop_rm_int4<0x24, "vpmovsxwq", int_x86_sse41_pmovsxwq,
6108 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
6109 defm VPMOVZXBD : SS41I_binop_rm_int4<0x31, "vpmovzxbd", int_x86_sse41_pmovzxbd,
6110 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
6111 defm VPMOVZXWQ : SS41I_binop_rm_int4<0x34, "vpmovzxwq", int_x86_sse41_pmovzxwq,
6112 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
6115 let Predicates = [HasAVX2] in {
6116 defm VPMOVSXBD : SS41I_binop_rm_int8_y<0x21, "vpmovsxbd",
6117 int_x86_avx2_pmovsxbd, WriteShuffle>,
6119 defm VPMOVSXWQ : SS41I_binop_rm_int8_y<0x24, "vpmovsxwq",
6120 int_x86_avx2_pmovsxwq, WriteShuffle>,
6122 defm VPMOVZXBD : SS41I_binop_rm_int8_y<0x31, "vpmovzxbd",
6123 int_x86_avx2_pmovzxbd, WriteShuffle>,
6125 defm VPMOVZXWQ : SS41I_binop_rm_int8_y<0x34, "vpmovzxwq",
6126 int_x86_avx2_pmovzxwq, WriteShuffle>,
6130 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd,
6131 SSE_INTALU_ITINS_SHUFF_P>;
6132 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq,
6133 SSE_INTALU_ITINS_SHUFF_P>;
6134 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd,
6135 SSE_INTALU_ITINS_SHUFF_P>;
6136 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq,
6137 SSE_INTALU_ITINS_SHUFF_P>;
6139 let Predicates = [HasAVX] in {
6140 // Common patterns involving scalar load
6141 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
6142 (VPMOVSXBDrm addr:$src)>;
6143 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
6144 (VPMOVSXWQrm addr:$src)>;
6146 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
6147 (VPMOVZXBDrm addr:$src)>;
6148 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
6149 (VPMOVZXWQrm addr:$src)>;
6152 let Predicates = [UseSSE41] in {
6153 // Common patterns involving scalar load
6154 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
6155 (PMOVSXBDrm addr:$src)>;
6156 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
6157 (PMOVSXWQrm addr:$src)>;
6159 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
6160 (PMOVZXBDrm addr:$src)>;
6161 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
6162 (PMOVZXWQrm addr:$src)>;
6165 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId,
6166 X86FoldableSchedWrite Sched> {
6167 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
6168 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6169 [(set VR128:$dst, (IntId VR128:$src))]>, Sched<[Sched]>;
6171 // Expecting a i16 load any extended to i32 value.
6172 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
6173 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6174 [(set VR128:$dst, (IntId (bitconvert
6175 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
6176 Sched<[Sched.Folded]>;
6179 multiclass SS41I_binop_rm_int4_y<bits<8> opc, string OpcodeStr,
6180 Intrinsic IntId, X86FoldableSchedWrite Sched> {
6181 def Yrr : SS48I<opc, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
6182 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6183 [(set VR256:$dst, (IntId VR128:$src))]>, Sched<[Sched]>;
6185 // Expecting a i16 load any extended to i32 value.
6186 def Yrm : SS48I<opc, MRMSrcMem, (outs VR256:$dst), (ins i16mem:$src),
6187 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6188 [(set VR256:$dst, (IntId (bitconvert
6189 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
6190 Sched<[Sched.Folded]>;
6193 let Predicates = [HasAVX] in {
6194 defm VPMOVSXBQ : SS41I_binop_rm_int2<0x22, "vpmovsxbq", int_x86_sse41_pmovsxbq,
6196 defm VPMOVZXBQ : SS41I_binop_rm_int2<0x32, "vpmovzxbq", int_x86_sse41_pmovzxbq,
6199 let Predicates = [HasAVX2] in {
6200 defm VPMOVSXBQ : SS41I_binop_rm_int4_y<0x22, "vpmovsxbq", int_x86_avx2_pmovsxbq,
6201 WriteShuffle>, VEX, VEX_L;
6202 defm VPMOVZXBQ : SS41I_binop_rm_int4_y<0x32, "vpmovzxbq", int_x86_avx2_pmovzxbq,
6203 WriteShuffle>, VEX, VEX_L;
6205 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq,
6207 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq,
6210 let Predicates = [HasAVX2] in {
6211 def : Pat<(v16i16 (X86vsext (v16i8 VR128:$src))), (VPMOVSXBWYrr VR128:$src)>;
6212 def : Pat<(v8i32 (X86vsext (v16i8 VR128:$src))), (VPMOVSXBDYrr VR128:$src)>;
6213 def : Pat<(v4i64 (X86vsext (v16i8 VR128:$src))), (VPMOVSXBQYrr VR128:$src)>;
6215 def : Pat<(v8i32 (X86vsext (v8i16 VR128:$src))), (VPMOVSXWDYrr VR128:$src)>;
6216 def : Pat<(v4i64 (X86vsext (v8i16 VR128:$src))), (VPMOVSXWQYrr VR128:$src)>;
6218 def : Pat<(v4i64 (X86vsext (v4i32 VR128:$src))), (VPMOVSXDQYrr VR128:$src)>;
6220 def : Pat<(v16i16 (X86vsext (v32i8 VR256:$src))),
6221 (VPMOVSXBWYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6222 def : Pat<(v8i32 (X86vsext (v32i8 VR256:$src))),
6223 (VPMOVSXBDYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6224 def : Pat<(v4i64 (X86vsext (v32i8 VR256:$src))),
6225 (VPMOVSXBQYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6227 def : Pat<(v8i32 (X86vsext (v16i16 VR256:$src))),
6228 (VPMOVSXWDYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6229 def : Pat<(v4i64 (X86vsext (v16i16 VR256:$src))),
6230 (VPMOVSXWQYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6232 def : Pat<(v4i64 (X86vsext (v8i32 VR256:$src))),
6233 (VPMOVSXDQYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6235 def : Pat<(v8i32 (X86vsext (v8i16 (bitconvert (v2i64 (load addr:$src)))))),
6236 (VPMOVSXWDYrm addr:$src)>;
6237 def : Pat<(v4i64 (X86vsext (v4i32 (bitconvert (v2i64 (load addr:$src)))))),
6238 (VPMOVSXDQYrm addr:$src)>;
6240 def : Pat<(v8i32 (X86vsext (v16i8 (bitconvert (v2i64
6241 (scalar_to_vector (loadi64 addr:$src))))))),
6242 (VPMOVSXBDYrm addr:$src)>;
6243 def : Pat<(v8i32 (X86vsext (v16i8 (bitconvert (v2f64
6244 (scalar_to_vector (loadf64 addr:$src))))))),
6245 (VPMOVSXBDYrm addr:$src)>;
6247 def : Pat<(v4i64 (X86vsext (v8i16 (bitconvert (v2i64
6248 (scalar_to_vector (loadi64 addr:$src))))))),
6249 (VPMOVSXWQYrm addr:$src)>;
6250 def : Pat<(v4i64 (X86vsext (v8i16 (bitconvert (v2f64
6251 (scalar_to_vector (loadf64 addr:$src))))))),
6252 (VPMOVSXWQYrm addr:$src)>;
6254 def : Pat<(v4i64 (X86vsext (v16i8 (bitconvert (v4i32
6255 (scalar_to_vector (loadi32 addr:$src))))))),
6256 (VPMOVSXBQYrm addr:$src)>;
6259 let Predicates = [HasAVX] in {
6260 // Common patterns involving scalar load
6261 def : Pat<(int_x86_sse41_pmovsxbq
6262 (bitconvert (v4i32 (X86vzmovl
6263 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6264 (VPMOVSXBQrm addr:$src)>;
6266 def : Pat<(int_x86_sse41_pmovzxbq
6267 (bitconvert (v4i32 (X86vzmovl
6268 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6269 (VPMOVZXBQrm addr:$src)>;
6272 let Predicates = [UseSSE41] in {
6273 def : Pat<(v8i16 (X86vsext (v16i8 VR128:$src))), (PMOVSXBWrr VR128:$src)>;
6274 def : Pat<(v4i32 (X86vsext (v16i8 VR128:$src))), (PMOVSXBDrr VR128:$src)>;
6275 def : Pat<(v2i64 (X86vsext (v16i8 VR128:$src))), (PMOVSXBQrr VR128:$src)>;
6277 def : Pat<(v4i32 (X86vsext (v8i16 VR128:$src))), (PMOVSXWDrr VR128:$src)>;
6278 def : Pat<(v2i64 (X86vsext (v8i16 VR128:$src))), (PMOVSXWQrr VR128:$src)>;
6280 def : Pat<(v2i64 (X86vsext (v4i32 VR128:$src))), (PMOVSXDQrr VR128:$src)>;
6282 // Common patterns involving scalar load
6283 def : Pat<(int_x86_sse41_pmovsxbq
6284 (bitconvert (v4i32 (X86vzmovl
6285 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6286 (PMOVSXBQrm addr:$src)>;
6288 def : Pat<(int_x86_sse41_pmovzxbq
6289 (bitconvert (v4i32 (X86vzmovl
6290 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6291 (PMOVZXBQrm addr:$src)>;
6293 def : Pat<(v4i32 (X86vsext (v8i16 (bitconvert (v2i64
6294 (scalar_to_vector (loadi64 addr:$src))))))),
6295 (PMOVSXWDrm addr:$src)>;
6296 def : Pat<(v4i32 (X86vsext (v8i16 (bitconvert (v2f64
6297 (scalar_to_vector (loadf64 addr:$src))))))),
6298 (PMOVSXWDrm addr:$src)>;
6299 def : Pat<(v4i32 (X86vsext (v16i8 (bitconvert (v4i32
6300 (scalar_to_vector (loadi32 addr:$src))))))),
6301 (PMOVSXBDrm addr:$src)>;
6302 def : Pat<(v2i64 (X86vsext (v8i16 (bitconvert (v4i32
6303 (scalar_to_vector (loadi32 addr:$src))))))),
6304 (PMOVSXWQrm addr:$src)>;
6305 def : Pat<(v2i64 (X86vsext (v16i8 (bitconvert (v4i32
6306 (scalar_to_vector (extloadi32i16 addr:$src))))))),
6307 (PMOVSXBQrm addr:$src)>;
6308 def : Pat<(v2i64 (X86vsext (v4i32 (bitconvert (v2i64
6309 (scalar_to_vector (loadi64 addr:$src))))))),
6310 (PMOVSXDQrm addr:$src)>;
6311 def : Pat<(v2i64 (X86vsext (v4i32 (bitconvert (v2f64
6312 (scalar_to_vector (loadf64 addr:$src))))))),
6313 (PMOVSXDQrm addr:$src)>;
6314 def : Pat<(v8i16 (X86vsext (v16i8 (bitconvert (v2i64
6315 (scalar_to_vector (loadi64 addr:$src))))))),
6316 (PMOVSXBWrm addr:$src)>;
6317 def : Pat<(v8i16 (X86vsext (v16i8 (bitconvert (v2f64
6318 (scalar_to_vector (loadf64 addr:$src))))))),
6319 (PMOVSXBWrm addr:$src)>;
6322 let Predicates = [HasAVX2] in {
6323 def : Pat<(v16i16 (X86vzext (v16i8 VR128:$src))), (VPMOVZXBWYrr VR128:$src)>;
6324 def : Pat<(v8i32 (X86vzext (v16i8 VR128:$src))), (VPMOVZXBDYrr VR128:$src)>;
6325 def : Pat<(v4i64 (X86vzext (v16i8 VR128:$src))), (VPMOVZXBQYrr VR128:$src)>;
6327 def : Pat<(v8i32 (X86vzext (v8i16 VR128:$src))), (VPMOVZXWDYrr VR128:$src)>;
6328 def : Pat<(v4i64 (X86vzext (v8i16 VR128:$src))), (VPMOVZXWQYrr VR128:$src)>;
6330 def : Pat<(v4i64 (X86vzext (v4i32 VR128:$src))), (VPMOVZXDQYrr VR128:$src)>;
6332 def : Pat<(v16i16 (X86vzext (v32i8 VR256:$src))),
6333 (VPMOVZXBWYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6334 def : Pat<(v8i32 (X86vzext (v32i8 VR256:$src))),
6335 (VPMOVZXBDYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6336 def : Pat<(v4i64 (X86vzext (v32i8 VR256:$src))),
6337 (VPMOVZXBQYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6339 def : Pat<(v8i32 (X86vzext (v16i16 VR256:$src))),
6340 (VPMOVZXWDYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6341 def : Pat<(v4i64 (X86vzext (v16i16 VR256:$src))),
6342 (VPMOVZXWQYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6344 def : Pat<(v4i64 (X86vzext (v8i32 VR256:$src))),
6345 (VPMOVZXDQYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6348 let Predicates = [HasAVX] in {
6349 def : Pat<(v8i16 (X86vzext (v16i8 VR128:$src))), (VPMOVZXBWrr VR128:$src)>;
6350 def : Pat<(v4i32 (X86vzext (v16i8 VR128:$src))), (VPMOVZXBDrr VR128:$src)>;
6351 def : Pat<(v2i64 (X86vzext (v16i8 VR128:$src))), (VPMOVZXBQrr VR128:$src)>;
6353 def : Pat<(v4i32 (X86vzext (v8i16 VR128:$src))), (VPMOVZXWDrr VR128:$src)>;
6354 def : Pat<(v2i64 (X86vzext (v8i16 VR128:$src))), (VPMOVZXWQrr VR128:$src)>;
6356 def : Pat<(v2i64 (X86vzext (v4i32 VR128:$src))), (VPMOVZXDQrr VR128:$src)>;
6358 def : Pat<(v8i16 (X86vzext (v16i8 (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
6359 (VPMOVZXBWrm addr:$src)>;
6360 def : Pat<(v8i16 (X86vzext (v16i8 (bitconvert (v2f64 (scalar_to_vector (loadf64 addr:$src))))))),
6361 (VPMOVZXBWrm addr:$src)>;
6362 def : Pat<(v4i32 (X86vzext (v16i8 (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6363 (VPMOVZXBDrm addr:$src)>;
6364 def : Pat<(v2i64 (X86vzext (v16i8 (bitconvert (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))),
6365 (VPMOVZXBQrm addr:$src)>;
6367 def : Pat<(v4i32 (X86vzext (v8i16 (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
6368 (VPMOVZXWDrm addr:$src)>;
6369 def : Pat<(v4i32 (X86vzext (v8i16 (bitconvert (v2f64 (scalar_to_vector (loadf64 addr:$src))))))),
6370 (VPMOVZXWDrm addr:$src)>;
6371 def : Pat<(v2i64 (X86vzext (v8i16 (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6372 (VPMOVZXWQrm addr:$src)>;
6374 def : Pat<(v2i64 (X86vzext (v4i32 (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
6375 (VPMOVZXDQrm addr:$src)>;
6376 def : Pat<(v2i64 (X86vzext (v4i32 (bitconvert (v2f64 (scalar_to_vector (loadf64 addr:$src))))))),
6377 (VPMOVZXDQrm addr:$src)>;
6378 def : Pat<(v2i64 (X86vzext (v4i32 (bitconvert (v2i64 (X86vzload addr:$src)))))),
6379 (VPMOVZXDQrm addr:$src)>;
6381 def : Pat<(v8i16 (X86vsext (v16i8 VR128:$src))), (VPMOVSXBWrr VR128:$src)>;
6382 def : Pat<(v4i32 (X86vsext (v16i8 VR128:$src))), (VPMOVSXBDrr VR128:$src)>;
6383 def : Pat<(v2i64 (X86vsext (v16i8 VR128:$src))), (VPMOVSXBQrr VR128:$src)>;
6385 def : Pat<(v4i32 (X86vsext (v8i16 VR128:$src))), (VPMOVSXWDrr VR128:$src)>;
6386 def : Pat<(v2i64 (X86vsext (v8i16 VR128:$src))), (VPMOVSXWQrr VR128:$src)>;
6388 def : Pat<(v2i64 (X86vsext (v4i32 VR128:$src))), (VPMOVSXDQrr VR128:$src)>;
6390 def : Pat<(v4i32 (X86vsext (v8i16 (bitconvert (v2i64
6391 (scalar_to_vector (loadi64 addr:$src))))))),
6392 (VPMOVSXWDrm addr:$src)>;
6393 def : Pat<(v2i64 (X86vsext (v4i32 (bitconvert (v2i64
6394 (scalar_to_vector (loadi64 addr:$src))))))),
6395 (VPMOVSXDQrm addr:$src)>;
6396 def : Pat<(v4i32 (X86vsext (v8i16 (bitconvert (v2f64
6397 (scalar_to_vector (loadf64 addr:$src))))))),
6398 (VPMOVSXWDrm addr:$src)>;
6399 def : Pat<(v2i64 (X86vsext (v4i32 (bitconvert (v2f64
6400 (scalar_to_vector (loadf64 addr:$src))))))),
6401 (VPMOVSXDQrm addr:$src)>;
6402 def : Pat<(v8i16 (X86vsext (v16i8 (bitconvert (v2i64
6403 (scalar_to_vector (loadi64 addr:$src))))))),
6404 (VPMOVSXBWrm addr:$src)>;
6405 def : Pat<(v8i16 (X86vsext (v16i8 (bitconvert (v2f64
6406 (scalar_to_vector (loadf64 addr:$src))))))),
6407 (VPMOVSXBWrm addr:$src)>;
6409 def : Pat<(v4i32 (X86vsext (v16i8 (bitconvert (v4i32
6410 (scalar_to_vector (loadi32 addr:$src))))))),
6411 (VPMOVSXBDrm addr:$src)>;
6412 def : Pat<(v2i64 (X86vsext (v8i16 (bitconvert (v4i32
6413 (scalar_to_vector (loadi32 addr:$src))))))),
6414 (VPMOVSXWQrm addr:$src)>;
6415 def : Pat<(v2i64 (X86vsext (v16i8 (bitconvert (v4i32
6416 (scalar_to_vector (extloadi32i16 addr:$src))))))),
6417 (VPMOVSXBQrm addr:$src)>;
6420 let Predicates = [UseSSE41] in {
6421 def : Pat<(v8i16 (X86vzext (v16i8 VR128:$src))), (PMOVZXBWrr VR128:$src)>;
6422 def : Pat<(v4i32 (X86vzext (v16i8 VR128:$src))), (PMOVZXBDrr VR128:$src)>;
6423 def : Pat<(v2i64 (X86vzext (v16i8 VR128:$src))), (PMOVZXBQrr VR128:$src)>;
6425 def : Pat<(v4i32 (X86vzext (v8i16 VR128:$src))), (PMOVZXWDrr VR128:$src)>;
6426 def : Pat<(v2i64 (X86vzext (v8i16 VR128:$src))), (PMOVZXWQrr VR128:$src)>;
6428 def : Pat<(v2i64 (X86vzext (v4i32 VR128:$src))), (PMOVZXDQrr VR128:$src)>;
6430 def : Pat<(v8i16 (X86vzext (v16i8 (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
6431 (PMOVZXBWrm addr:$src)>;
6432 def : Pat<(v8i16 (X86vzext (v16i8 (bitconvert (v2f64 (scalar_to_vector (loadf64 addr:$src))))))),
6433 (PMOVZXBWrm addr:$src)>;
6434 def : Pat<(v4i32 (X86vzext (v16i8 (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6435 (PMOVZXBDrm addr:$src)>;
6436 def : Pat<(v2i64 (X86vzext (v16i8 (bitconvert (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))),
6437 (PMOVZXBQrm addr:$src)>;
6439 def : Pat<(v4i32 (X86vzext (v8i16 (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
6440 (PMOVZXWDrm addr:$src)>;
6441 def : Pat<(v4i32 (X86vzext (v8i16 (bitconvert (v2f64 (scalar_to_vector (loadf64 addr:$src))))))),
6442 (PMOVZXWDrm addr:$src)>;
6443 def : Pat<(v2i64 (X86vzext (v8i16 (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6444 (PMOVZXWQrm addr:$src)>;
6446 def : Pat<(v2i64 (X86vzext (v4i32 (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
6447 (PMOVZXDQrm addr:$src)>;
6448 def : Pat<(v2i64 (X86vzext (v4i32 (bitconvert (v2f64 (scalar_to_vector (loadf64 addr:$src))))))),
6449 (PMOVZXDQrm addr:$src)>;
6450 def : Pat<(v2i64 (X86vzext (v4i32 (bitconvert (v2i64 (X86vzload addr:$src)))))),
6451 (PMOVZXDQrm addr:$src)>;
6454 //===----------------------------------------------------------------------===//
6455 // SSE4.1 - Extract Instructions
6456 //===----------------------------------------------------------------------===//
6458 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
6459 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
6460 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
6461 (ins VR128:$src1, i32i8imm:$src2),
6462 !strconcat(OpcodeStr,
6463 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6464 [(set GR32orGR64:$dst, (X86pextrb (v16i8 VR128:$src1),
6466 Sched<[WriteShuffle]>;
6467 let neverHasSideEffects = 1, mayStore = 1,
6468 SchedRW = [WriteShuffleLd, WriteRMW] in
6469 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6470 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
6471 !strconcat(OpcodeStr,
6472 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6473 [(store (i8 (trunc (assertzext (X86pextrb (v16i8 VR128:$src1),
6474 imm:$src2)))), addr:$dst)]>;
6477 let Predicates = [HasAVX] in
6478 defm VPEXTRB : SS41I_extract8<0x14, "vpextrb">, VEX;
6480 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
6483 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
6484 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
6485 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in
6486 def rr_REV : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
6487 (ins VR128:$src1, i32i8imm:$src2),
6488 !strconcat(OpcodeStr,
6489 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6490 []>, Sched<[WriteShuffle]>;
6492 let neverHasSideEffects = 1, mayStore = 1,
6493 SchedRW = [WriteShuffleLd, WriteRMW] in
6494 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6495 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
6496 !strconcat(OpcodeStr,
6497 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6498 [(store (i16 (trunc (assertzext (X86pextrw (v8i16 VR128:$src1),
6499 imm:$src2)))), addr:$dst)]>;
6502 let Predicates = [HasAVX] in
6503 defm VPEXTRW : SS41I_extract16<0x15, "vpextrw">, VEX;
6505 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
6508 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
6509 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
6510 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
6511 (ins VR128:$src1, i32i8imm:$src2),
6512 !strconcat(OpcodeStr,
6513 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6515 (extractelt (v4i32 VR128:$src1), imm:$src2))]>,
6516 Sched<[WriteShuffle]>;
6517 let SchedRW = [WriteShuffleLd, WriteRMW] in
6518 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6519 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
6520 !strconcat(OpcodeStr,
6521 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6522 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
6526 let Predicates = [HasAVX] in
6527 defm VPEXTRD : SS41I_extract32<0x16, "vpextrd">, VEX;
6529 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
6531 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
6532 multiclass SS41I_extract64<bits<8> opc, string OpcodeStr> {
6533 def rr : SS4AIi8<opc, MRMDestReg, (outs GR64:$dst),
6534 (ins VR128:$src1, i32i8imm:$src2),
6535 !strconcat(OpcodeStr,
6536 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6538 (extractelt (v2i64 VR128:$src1), imm:$src2))]>,
6539 Sched<[WriteShuffle]>, REX_W;
6540 let SchedRW = [WriteShuffleLd, WriteRMW] in
6541 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6542 (ins i64mem:$dst, VR128:$src1, i32i8imm:$src2),
6543 !strconcat(OpcodeStr,
6544 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6545 [(store (extractelt (v2i64 VR128:$src1), imm:$src2),
6546 addr:$dst)]>, REX_W;
6549 let Predicates = [HasAVX] in
6550 defm VPEXTRQ : SS41I_extract64<0x16, "vpextrq">, VEX, VEX_W;
6552 defm PEXTRQ : SS41I_extract64<0x16, "pextrq">;
6554 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
6556 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr,
6557 OpndItins itins = DEFAULT_ITINS> {
6558 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
6559 (ins VR128:$src1, i32i8imm:$src2),
6560 !strconcat(OpcodeStr,
6561 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6562 [(set GR32orGR64:$dst,
6563 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))],
6564 itins.rr>, Sched<[WriteFBlend]>;
6565 let SchedRW = [WriteFBlendLd, WriteRMW] in
6566 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6567 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
6568 !strconcat(OpcodeStr,
6569 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6570 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
6571 addr:$dst)], itins.rm>;
6574 let ExeDomain = SSEPackedSingle in {
6575 let Predicates = [UseAVX] in
6576 defm VEXTRACTPS : SS41I_extractf32<0x17, "vextractps">, VEX;
6577 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps", SSE_EXTRACT_ITINS>;
6580 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
6581 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
6584 (VEXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
6586 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
6589 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
6590 Requires<[UseSSE41]>;
6592 //===----------------------------------------------------------------------===//
6593 // SSE4.1 - Insert Instructions
6594 //===----------------------------------------------------------------------===//
6596 multiclass SS41I_insert8<bits<8> opc, string asm, bit Is2Addr = 1> {
6597 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6598 (ins VR128:$src1, GR32orGR64:$src2, i32i8imm:$src3),
6600 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6602 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6604 (X86pinsrb VR128:$src1, GR32orGR64:$src2, imm:$src3))]>,
6605 Sched<[WriteShuffle]>;
6606 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6607 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
6609 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6611 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6613 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
6614 imm:$src3))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6617 let Predicates = [HasAVX] in
6618 defm VPINSRB : SS41I_insert8<0x20, "vpinsrb", 0>, VEX_4V;
6619 let Constraints = "$src1 = $dst" in
6620 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
6622 multiclass SS41I_insert32<bits<8> opc, string asm, bit Is2Addr = 1> {
6623 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6624 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
6626 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6628 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6630 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
6631 Sched<[WriteShuffle]>;
6632 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6633 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
6635 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6637 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6639 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
6640 imm:$src3)))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6643 let Predicates = [HasAVX] in
6644 defm VPINSRD : SS41I_insert32<0x22, "vpinsrd", 0>, VEX_4V;
6645 let Constraints = "$src1 = $dst" in
6646 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
6648 multiclass SS41I_insert64<bits<8> opc, string asm, bit Is2Addr = 1> {
6649 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6650 (ins VR128:$src1, GR64:$src2, i32i8imm:$src3),
6652 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6654 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6656 (v2i64 (insertelt VR128:$src1, GR64:$src2, imm:$src3)))]>,
6657 Sched<[WriteShuffle]>;
6658 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6659 (ins VR128:$src1, i64mem:$src2, i32i8imm:$src3),
6661 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6663 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6665 (v2i64 (insertelt VR128:$src1, (loadi64 addr:$src2),
6666 imm:$src3)))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6669 let Predicates = [HasAVX] in
6670 defm VPINSRQ : SS41I_insert64<0x22, "vpinsrq", 0>, VEX_4V, VEX_W;
6671 let Constraints = "$src1 = $dst" in
6672 defm PINSRQ : SS41I_insert64<0x22, "pinsrq">, REX_W;
6674 // insertps has a few different modes, there's the first two here below which
6675 // are optimized inserts that won't zero arbitrary elements in the destination
6676 // vector. The next one matches the intrinsic and could zero arbitrary elements
6677 // in the target vector.
6678 multiclass SS41I_insertf32<bits<8> opc, string asm, bit Is2Addr = 1,
6679 OpndItins itins = DEFAULT_ITINS> {
6680 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6681 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
6683 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6685 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6687 (X86insertps VR128:$src1, VR128:$src2, imm:$src3))], itins.rr>,
6688 Sched<[WriteFShuffle]>;
6689 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6690 (ins VR128:$src1, f32mem:$src2, i8imm:$src3),
6692 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6694 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6696 (X86insertps VR128:$src1,
6697 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
6698 imm:$src3))], itins.rm>,
6699 Sched<[WriteFShuffleLd, ReadAfterLd]>;
6702 let ExeDomain = SSEPackedSingle in {
6703 let Predicates = [UseAVX] in
6704 defm VINSERTPS : SS41I_insertf32<0x21, "vinsertps", 0>, VEX_4V;
6705 let Constraints = "$src1 = $dst" in
6706 defm INSERTPS : SS41I_insertf32<0x21, "insertps", 1, SSE_INSERT_ITINS>;
6709 let Predicates = [UseSSE41] in {
6710 // If we're inserting an element from a load or a null pshuf of a load,
6711 // fold the load into the insertps instruction.
6712 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd (v4f32
6713 (scalar_to_vector (loadf32 addr:$src2))), (i8 0)),
6715 (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6716 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd
6717 (loadv4f32 addr:$src2), (i8 0)), imm:$src3)),
6718 (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6721 let Predicates = [UseAVX] in {
6722 // If we're inserting an element from a vbroadcast of a load, fold the
6723 // load into the X86insertps instruction.
6724 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
6725 (X86VBroadcast (loadf32 addr:$src2)), imm:$src3)),
6726 (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6727 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
6728 (X86VBroadcast (loadv4f32 addr:$src2)), imm:$src3)),
6729 (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6732 //===----------------------------------------------------------------------===//
6733 // SSE4.1 - Round Instructions
6734 //===----------------------------------------------------------------------===//
6736 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd, string OpcodeStr,
6737 X86MemOperand x86memop, RegisterClass RC,
6738 PatFrag mem_frag32, PatFrag mem_frag64,
6739 Intrinsic V4F32Int, Intrinsic V2F64Int> {
6740 let ExeDomain = SSEPackedSingle in {
6741 // Intrinsic operation, reg.
6742 // Vector intrinsic operation, reg
6743 def PSr : SS4AIi8<opcps, MRMSrcReg,
6744 (outs RC:$dst), (ins RC:$src1, i32i8imm:$src2),
6745 !strconcat(OpcodeStr,
6746 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6747 [(set RC:$dst, (V4F32Int RC:$src1, imm:$src2))],
6748 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAdd]>;
6750 // Vector intrinsic operation, mem
6751 def PSm : SS4AIi8<opcps, MRMSrcMem,
6752 (outs RC:$dst), (ins x86memop:$src1, i32i8imm:$src2),
6753 !strconcat(OpcodeStr,
6754 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6756 (V4F32Int (mem_frag32 addr:$src1),imm:$src2))],
6757 IIC_SSE_ROUNDPS_MEM>, Sched<[WriteFAddLd]>;
6758 } // ExeDomain = SSEPackedSingle
6760 let ExeDomain = SSEPackedDouble in {
6761 // Vector intrinsic operation, reg
6762 def PDr : SS4AIi8<opcpd, MRMSrcReg,
6763 (outs RC:$dst), (ins RC:$src1, i32i8imm:$src2),
6764 !strconcat(OpcodeStr,
6765 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6766 [(set RC:$dst, (V2F64Int RC:$src1, imm:$src2))],
6767 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAdd]>;
6769 // Vector intrinsic operation, mem
6770 def PDm : SS4AIi8<opcpd, MRMSrcMem,
6771 (outs RC:$dst), (ins x86memop:$src1, i32i8imm:$src2),
6772 !strconcat(OpcodeStr,
6773 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6775 (V2F64Int (mem_frag64 addr:$src1),imm:$src2))],
6776 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAddLd]>;
6777 } // ExeDomain = SSEPackedDouble
6780 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
6783 Intrinsic F64Int, bit Is2Addr = 1> {
6784 let ExeDomain = GenericDomain in {
6786 let hasSideEffects = 0 in
6787 def SSr : SS4AIi8<opcss, MRMSrcReg,
6788 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2, i32i8imm:$src3),
6790 !strconcat(OpcodeStr,
6791 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6792 !strconcat(OpcodeStr,
6793 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6794 []>, Sched<[WriteFAdd]>;
6796 // Intrinsic operation, reg.
6797 let isCodeGenOnly = 1 in
6798 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
6799 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
6801 !strconcat(OpcodeStr,
6802 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6803 !strconcat(OpcodeStr,
6804 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6805 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
6808 // Intrinsic operation, mem.
6809 def SSm : SS4AIi8<opcss, MRMSrcMem,
6810 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
6812 !strconcat(OpcodeStr,
6813 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6814 !strconcat(OpcodeStr,
6815 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6817 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
6818 Sched<[WriteFAddLd, ReadAfterLd]>;
6821 let hasSideEffects = 0 in
6822 def SDr : SS4AIi8<opcsd, MRMSrcReg,
6823 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2, i32i8imm:$src3),
6825 !strconcat(OpcodeStr,
6826 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6827 !strconcat(OpcodeStr,
6828 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6829 []>, Sched<[WriteFAdd]>;
6831 // Intrinsic operation, reg.
6832 let isCodeGenOnly = 1 in
6833 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
6834 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
6836 !strconcat(OpcodeStr,
6837 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6838 !strconcat(OpcodeStr,
6839 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6840 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
6843 // Intrinsic operation, mem.
6844 def SDm : SS4AIi8<opcsd, MRMSrcMem,
6845 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
6847 !strconcat(OpcodeStr,
6848 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6849 !strconcat(OpcodeStr,
6850 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6852 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
6853 Sched<[WriteFAddLd, ReadAfterLd]>;
6854 } // ExeDomain = GenericDomain
6857 // FP round - roundss, roundps, roundsd, roundpd
6858 let Predicates = [HasAVX] in {
6860 defm VROUND : sse41_fp_unop_rm<0x08, 0x09, "vround", f128mem, VR128,
6861 loadv4f32, loadv2f64,
6862 int_x86_sse41_round_ps,
6863 int_x86_sse41_round_pd>, VEX;
6864 defm VROUNDY : sse41_fp_unop_rm<0x08, 0x09, "vround", f256mem, VR256,
6865 loadv8f32, loadv4f64,
6866 int_x86_avx_round_ps_256,
6867 int_x86_avx_round_pd_256>, VEX, VEX_L;
6868 defm VROUND : sse41_fp_binop_rm<0x0A, 0x0B, "vround",
6869 int_x86_sse41_round_ss,
6870 int_x86_sse41_round_sd, 0>, VEX_4V, VEX_LIG;
6872 def : Pat<(ffloor FR32:$src),
6873 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x1))>;
6874 def : Pat<(f64 (ffloor FR64:$src)),
6875 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x1))>;
6876 def : Pat<(f32 (fnearbyint FR32:$src)),
6877 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0xC))>;
6878 def : Pat<(f64 (fnearbyint FR64:$src)),
6879 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0xC))>;
6880 def : Pat<(f32 (fceil FR32:$src)),
6881 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x2))>;
6882 def : Pat<(f64 (fceil FR64:$src)),
6883 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x2))>;
6884 def : Pat<(f32 (frint FR32:$src)),
6885 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x4))>;
6886 def : Pat<(f64 (frint FR64:$src)),
6887 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x4))>;
6888 def : Pat<(f32 (ftrunc FR32:$src)),
6889 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x3))>;
6890 def : Pat<(f64 (ftrunc FR64:$src)),
6891 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x3))>;
6893 def : Pat<(v4f32 (ffloor VR128:$src)),
6894 (VROUNDPSr VR128:$src, (i32 0x1))>;
6895 def : Pat<(v4f32 (fnearbyint VR128:$src)),
6896 (VROUNDPSr VR128:$src, (i32 0xC))>;
6897 def : Pat<(v4f32 (fceil VR128:$src)),
6898 (VROUNDPSr VR128:$src, (i32 0x2))>;
6899 def : Pat<(v4f32 (frint VR128:$src)),
6900 (VROUNDPSr VR128:$src, (i32 0x4))>;
6901 def : Pat<(v4f32 (ftrunc VR128:$src)),
6902 (VROUNDPSr VR128:$src, (i32 0x3))>;
6904 def : Pat<(v2f64 (ffloor VR128:$src)),
6905 (VROUNDPDr VR128:$src, (i32 0x1))>;
6906 def : Pat<(v2f64 (fnearbyint VR128:$src)),
6907 (VROUNDPDr VR128:$src, (i32 0xC))>;
6908 def : Pat<(v2f64 (fceil VR128:$src)),
6909 (VROUNDPDr VR128:$src, (i32 0x2))>;
6910 def : Pat<(v2f64 (frint VR128:$src)),
6911 (VROUNDPDr VR128:$src, (i32 0x4))>;
6912 def : Pat<(v2f64 (ftrunc VR128:$src)),
6913 (VROUNDPDr VR128:$src, (i32 0x3))>;
6915 def : Pat<(v8f32 (ffloor VR256:$src)),
6916 (VROUNDYPSr VR256:$src, (i32 0x1))>;
6917 def : Pat<(v8f32 (fnearbyint VR256:$src)),
6918 (VROUNDYPSr VR256:$src, (i32 0xC))>;
6919 def : Pat<(v8f32 (fceil VR256:$src)),
6920 (VROUNDYPSr VR256:$src, (i32 0x2))>;
6921 def : Pat<(v8f32 (frint VR256:$src)),
6922 (VROUNDYPSr VR256:$src, (i32 0x4))>;
6923 def : Pat<(v8f32 (ftrunc VR256:$src)),
6924 (VROUNDYPSr VR256:$src, (i32 0x3))>;
6926 def : Pat<(v4f64 (ffloor VR256:$src)),
6927 (VROUNDYPDr VR256:$src, (i32 0x1))>;
6928 def : Pat<(v4f64 (fnearbyint VR256:$src)),
6929 (VROUNDYPDr VR256:$src, (i32 0xC))>;
6930 def : Pat<(v4f64 (fceil VR256:$src)),
6931 (VROUNDYPDr VR256:$src, (i32 0x2))>;
6932 def : Pat<(v4f64 (frint VR256:$src)),
6933 (VROUNDYPDr VR256:$src, (i32 0x4))>;
6934 def : Pat<(v4f64 (ftrunc VR256:$src)),
6935 (VROUNDYPDr VR256:$src, (i32 0x3))>;
6938 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round", f128mem, VR128,
6939 memopv4f32, memopv2f64,
6940 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
6941 let Constraints = "$src1 = $dst" in
6942 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
6943 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
6945 let Predicates = [UseSSE41] in {
6946 def : Pat<(ffloor FR32:$src),
6947 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x1))>;
6948 def : Pat<(f64 (ffloor FR64:$src)),
6949 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x1))>;
6950 def : Pat<(f32 (fnearbyint FR32:$src)),
6951 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0xC))>;
6952 def : Pat<(f64 (fnearbyint FR64:$src)),
6953 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0xC))>;
6954 def : Pat<(f32 (fceil FR32:$src)),
6955 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x2))>;
6956 def : Pat<(f64 (fceil FR64:$src)),
6957 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x2))>;
6958 def : Pat<(f32 (frint FR32:$src)),
6959 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x4))>;
6960 def : Pat<(f64 (frint FR64:$src)),
6961 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x4))>;
6962 def : Pat<(f32 (ftrunc FR32:$src)),
6963 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x3))>;
6964 def : Pat<(f64 (ftrunc FR64:$src)),
6965 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x3))>;
6967 def : Pat<(v4f32 (ffloor VR128:$src)),
6968 (ROUNDPSr VR128:$src, (i32 0x1))>;
6969 def : Pat<(v4f32 (fnearbyint VR128:$src)),
6970 (ROUNDPSr VR128:$src, (i32 0xC))>;
6971 def : Pat<(v4f32 (fceil VR128:$src)),
6972 (ROUNDPSr VR128:$src, (i32 0x2))>;
6973 def : Pat<(v4f32 (frint VR128:$src)),
6974 (ROUNDPSr VR128:$src, (i32 0x4))>;
6975 def : Pat<(v4f32 (ftrunc VR128:$src)),
6976 (ROUNDPSr VR128:$src, (i32 0x3))>;
6978 def : Pat<(v2f64 (ffloor VR128:$src)),
6979 (ROUNDPDr VR128:$src, (i32 0x1))>;
6980 def : Pat<(v2f64 (fnearbyint VR128:$src)),
6981 (ROUNDPDr VR128:$src, (i32 0xC))>;
6982 def : Pat<(v2f64 (fceil VR128:$src)),
6983 (ROUNDPDr VR128:$src, (i32 0x2))>;
6984 def : Pat<(v2f64 (frint VR128:$src)),
6985 (ROUNDPDr VR128:$src, (i32 0x4))>;
6986 def : Pat<(v2f64 (ftrunc VR128:$src)),
6987 (ROUNDPDr VR128:$src, (i32 0x3))>;
6990 //===----------------------------------------------------------------------===//
6991 // SSE4.1 - Packed Bit Test
6992 //===----------------------------------------------------------------------===//
6994 // ptest instruction we'll lower to this in X86ISelLowering primarily from
6995 // the intel intrinsic that corresponds to this.
6996 let Defs = [EFLAGS], Predicates = [HasAVX] in {
6997 def VPTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
6998 "vptest\t{$src2, $src1|$src1, $src2}",
6999 [(set EFLAGS, (X86ptest VR128:$src1, (v2i64 VR128:$src2)))]>,
7000 Sched<[WriteVecLogic]>, VEX;
7001 def VPTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
7002 "vptest\t{$src2, $src1|$src1, $src2}",
7003 [(set EFLAGS,(X86ptest VR128:$src1, (loadv2i64 addr:$src2)))]>,
7004 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX;
7006 def VPTESTYrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR256:$src1, VR256:$src2),
7007 "vptest\t{$src2, $src1|$src1, $src2}",
7008 [(set EFLAGS, (X86ptest VR256:$src1, (v4i64 VR256:$src2)))]>,
7009 Sched<[WriteVecLogic]>, VEX, VEX_L;
7010 def VPTESTYrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR256:$src1, i256mem:$src2),
7011 "vptest\t{$src2, $src1|$src1, $src2}",
7012 [(set EFLAGS,(X86ptest VR256:$src1, (loadv4i64 addr:$src2)))]>,
7013 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX, VEX_L;
7016 let Defs = [EFLAGS] in {
7017 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
7018 "ptest\t{$src2, $src1|$src1, $src2}",
7019 [(set EFLAGS, (X86ptest VR128:$src1, (v2i64 VR128:$src2)))]>,
7020 Sched<[WriteVecLogic]>;
7021 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
7022 "ptest\t{$src2, $src1|$src1, $src2}",
7023 [(set EFLAGS, (X86ptest VR128:$src1, (memopv2i64 addr:$src2)))]>,
7024 Sched<[WriteVecLogicLd, ReadAfterLd]>;
7027 // The bit test instructions below are AVX only
7028 multiclass avx_bittest<bits<8> opc, string OpcodeStr, RegisterClass RC,
7029 X86MemOperand x86memop, PatFrag mem_frag, ValueType vt> {
7030 def rr : SS48I<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
7031 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
7032 [(set EFLAGS, (X86testp RC:$src1, (vt RC:$src2)))]>,
7033 Sched<[WriteVecLogic]>, VEX;
7034 def rm : SS48I<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
7035 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
7036 [(set EFLAGS, (X86testp RC:$src1, (mem_frag addr:$src2)))]>,
7037 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX;
7040 let Defs = [EFLAGS], Predicates = [HasAVX] in {
7041 let ExeDomain = SSEPackedSingle in {
7042 defm VTESTPS : avx_bittest<0x0E, "vtestps", VR128, f128mem, loadv4f32, v4f32>;
7043 defm VTESTPSY : avx_bittest<0x0E, "vtestps", VR256, f256mem, loadv8f32, v8f32>,
7046 let ExeDomain = SSEPackedDouble in {
7047 defm VTESTPD : avx_bittest<0x0F, "vtestpd", VR128, f128mem, loadv2f64, v2f64>;
7048 defm VTESTPDY : avx_bittest<0x0F, "vtestpd", VR256, f256mem, loadv4f64, v4f64>,
7053 //===----------------------------------------------------------------------===//
7054 // SSE4.1 - Misc Instructions
7055 //===----------------------------------------------------------------------===//
7057 let Defs = [EFLAGS], Predicates = [HasPOPCNT] in {
7058 def POPCNT16rr : I<0xB8, MRMSrcReg, (outs GR16:$dst), (ins GR16:$src),
7059 "popcnt{w}\t{$src, $dst|$dst, $src}",
7060 [(set GR16:$dst, (ctpop GR16:$src)), (implicit EFLAGS)],
7061 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>,
7063 def POPCNT16rm : I<0xB8, MRMSrcMem, (outs GR16:$dst), (ins i16mem:$src),
7064 "popcnt{w}\t{$src, $dst|$dst, $src}",
7065 [(set GR16:$dst, (ctpop (loadi16 addr:$src))),
7066 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
7067 Sched<[WriteFAddLd]>, OpSize16, XS;
7069 def POPCNT32rr : I<0xB8, MRMSrcReg, (outs GR32:$dst), (ins GR32:$src),
7070 "popcnt{l}\t{$src, $dst|$dst, $src}",
7071 [(set GR32:$dst, (ctpop GR32:$src)), (implicit EFLAGS)],
7072 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>,
7075 def POPCNT32rm : I<0xB8, MRMSrcMem, (outs GR32:$dst), (ins i32mem:$src),
7076 "popcnt{l}\t{$src, $dst|$dst, $src}",
7077 [(set GR32:$dst, (ctpop (loadi32 addr:$src))),
7078 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
7079 Sched<[WriteFAddLd]>, OpSize32, XS;
7081 def POPCNT64rr : RI<0xB8, MRMSrcReg, (outs GR64:$dst), (ins GR64:$src),
7082 "popcnt{q}\t{$src, $dst|$dst, $src}",
7083 [(set GR64:$dst, (ctpop GR64:$src)), (implicit EFLAGS)],
7084 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>, XS;
7085 def POPCNT64rm : RI<0xB8, MRMSrcMem, (outs GR64:$dst), (ins i64mem:$src),
7086 "popcnt{q}\t{$src, $dst|$dst, $src}",
7087 [(set GR64:$dst, (ctpop (loadi64 addr:$src))),
7088 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
7089 Sched<[WriteFAddLd]>, XS;
7094 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
7095 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
7097 X86FoldableSchedWrite Sched> {
7098 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
7100 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
7101 [(set VR128:$dst, (IntId128 VR128:$src))]>,
7103 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
7105 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
7107 (IntId128 (bitconvert (memopv2i64 addr:$src))))]>,
7108 Sched<[Sched.Folded]>;
7111 // PHMIN has the same profile as PSAD, thus we use the same scheduling
7112 // model, although the naming is misleading.
7113 let Predicates = [HasAVX] in
7114 defm VPHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "vphminposuw",
7115 int_x86_sse41_phminposuw,
7117 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
7118 int_x86_sse41_phminposuw,
7121 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
7122 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
7123 Intrinsic IntId128, bit Is2Addr = 1,
7124 OpndItins itins = DEFAULT_ITINS> {
7125 let isCommutable = 1 in
7126 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
7127 (ins VR128:$src1, VR128:$src2),
7129 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7130 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7131 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))],
7132 itins.rr>, Sched<[itins.Sched]>;
7133 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
7134 (ins VR128:$src1, i128mem:$src2),
7136 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7137 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7139 (IntId128 VR128:$src1, (bitconvert (memopv2i64 addr:$src2))))],
7140 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
7143 /// SS41I_binop_rm_int_y - Simple SSE 4.1 binary operator
7144 multiclass SS41I_binop_rm_int_y<bits<8> opc, string OpcodeStr,
7146 X86FoldableSchedWrite Sched> {
7147 let isCommutable = 1 in
7148 def Yrr : SS48I<opc, MRMSrcReg, (outs VR256:$dst),
7149 (ins VR256:$src1, VR256:$src2),
7150 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7151 [(set VR256:$dst, (IntId256 VR256:$src1, VR256:$src2))]>,
7153 def Yrm : SS48I<opc, MRMSrcMem, (outs VR256:$dst),
7154 (ins VR256:$src1, i256mem:$src2),
7155 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7157 (IntId256 VR256:$src1, (bitconvert (loadv4i64 addr:$src2))))]>,
7158 Sched<[Sched.Folded, ReadAfterLd]>;
7162 /// SS48I_binop_rm - Simple SSE41 binary operator.
7163 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
7164 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
7165 X86MemOperand x86memop, bit Is2Addr = 1,
7166 OpndItins itins = SSE_INTALU_ITINS_P> {
7167 let isCommutable = 1 in
7168 def rr : SS48I<opc, MRMSrcReg, (outs RC:$dst),
7169 (ins RC:$src1, RC:$src2),
7171 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7172 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7173 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))]>,
7174 Sched<[itins.Sched]>;
7175 def rm : SS48I<opc, MRMSrcMem, (outs RC:$dst),
7176 (ins RC:$src1, x86memop:$src2),
7178 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7179 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7181 (OpVT (OpNode RC:$src1, (bitconvert (memop_frag addr:$src2)))))]>,
7182 Sched<[itins.Sched.Folded, ReadAfterLd]>;
7185 /// SS48I_binop_rm2 - Simple SSE41 binary operator with different src and dst
7187 multiclass SS48I_binop_rm2<bits<8> opc, string OpcodeStr, SDNode OpNode,
7188 ValueType DstVT, ValueType SrcVT, RegisterClass RC,
7189 PatFrag memop_frag, X86MemOperand x86memop,
7191 bit IsCommutable = 0, bit Is2Addr = 1> {
7192 let isCommutable = IsCommutable in
7193 def rr : SS48I<opc, MRMSrcReg, (outs RC:$dst),
7194 (ins RC:$src1, RC:$src2),
7196 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7197 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7198 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1), RC:$src2)))]>,
7199 Sched<[itins.Sched]>;
7200 def rm : SS48I<opc, MRMSrcMem, (outs RC:$dst),
7201 (ins RC:$src1, x86memop:$src2),
7203 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7204 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7205 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1),
7206 (bitconvert (memop_frag addr:$src2)))))]>,
7207 Sched<[itins.Sched.Folded, ReadAfterLd]>;
7210 let Predicates = [HasAVX] in {
7211 let isCommutable = 0 in
7212 defm VPMINSB : SS48I_binop_rm<0x38, "vpminsb", X86smin, v16i8, VR128,
7213 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7215 defm VPMINSD : SS48I_binop_rm<0x39, "vpminsd", X86smin, v4i32, VR128,
7216 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7218 defm VPMINUD : SS48I_binop_rm<0x3B, "vpminud", X86umin, v4i32, VR128,
7219 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7221 defm VPMINUW : SS48I_binop_rm<0x3A, "vpminuw", X86umin, v8i16, VR128,
7222 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7224 defm VPMAXSB : SS48I_binop_rm<0x3C, "vpmaxsb", X86smax, v16i8, VR128,
7225 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7227 defm VPMAXSD : SS48I_binop_rm<0x3D, "vpmaxsd", X86smax, v4i32, VR128,
7228 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7230 defm VPMAXUD : SS48I_binop_rm<0x3F, "vpmaxud", X86umax, v4i32, VR128,
7231 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7233 defm VPMAXUW : SS48I_binop_rm<0x3E, "vpmaxuw", X86umax, v8i16, VR128,
7234 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7236 defm VPMULDQ : SS48I_binop_rm2<0x28, "vpmuldq", X86pmuldq, v2i64, v4i32,
7237 VR128, loadv2i64, i128mem,
7238 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V;
7241 let Predicates = [HasAVX2] in {
7242 let isCommutable = 0 in
7243 defm VPMINSBY : SS48I_binop_rm<0x38, "vpminsb", X86smin, v32i8, VR256,
7244 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7246 defm VPMINSDY : SS48I_binop_rm<0x39, "vpminsd", X86smin, v8i32, VR256,
7247 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7249 defm VPMINUDY : SS48I_binop_rm<0x3B, "vpminud", X86umin, v8i32, VR256,
7250 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7252 defm VPMINUWY : SS48I_binop_rm<0x3A, "vpminuw", X86umin, v16i16, VR256,
7253 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7255 defm VPMAXSBY : SS48I_binop_rm<0x3C, "vpmaxsb", X86smax, v32i8, VR256,
7256 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7258 defm VPMAXSDY : SS48I_binop_rm<0x3D, "vpmaxsd", X86smax, v8i32, VR256,
7259 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7261 defm VPMAXUDY : SS48I_binop_rm<0x3F, "vpmaxud", X86umax, v8i32, VR256,
7262 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7264 defm VPMAXUWY : SS48I_binop_rm<0x3E, "vpmaxuw", X86umax, v16i16, VR256,
7265 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7267 defm VPMULDQY : SS48I_binop_rm2<0x28, "vpmuldq", X86pmuldq, v4i64, v8i32,
7268 VR256, loadv4i64, i256mem,
7269 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V, VEX_L;
7272 let Constraints = "$src1 = $dst" in {
7273 let isCommutable = 0 in
7274 defm PMINSB : SS48I_binop_rm<0x38, "pminsb", X86smin, v16i8, VR128,
7275 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7276 defm PMINSD : SS48I_binop_rm<0x39, "pminsd", X86smin, v4i32, VR128,
7277 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7278 defm PMINUD : SS48I_binop_rm<0x3B, "pminud", X86umin, v4i32, VR128,
7279 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7280 defm PMINUW : SS48I_binop_rm<0x3A, "pminuw", X86umin, v8i16, VR128,
7281 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7282 defm PMAXSB : SS48I_binop_rm<0x3C, "pmaxsb", X86smax, v16i8, VR128,
7283 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7284 defm PMAXSD : SS48I_binop_rm<0x3D, "pmaxsd", X86smax, v4i32, VR128,
7285 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7286 defm PMAXUD : SS48I_binop_rm<0x3F, "pmaxud", X86umax, v4i32, VR128,
7287 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7288 defm PMAXUW : SS48I_binop_rm<0x3E, "pmaxuw", X86umax, v8i16, VR128,
7289 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7290 defm PMULDQ : SS48I_binop_rm2<0x28, "pmuldq", X86pmuldq, v2i64, v4i32,
7291 VR128, memopv2i64, i128mem,
7292 SSE_INTMUL_ITINS_P, 1>;
7295 let Predicates = [HasAVX] in {
7296 defm VPMULLD : SS48I_binop_rm<0x40, "vpmulld", mul, v4i32, VR128,
7297 memopv2i64, i128mem, 0, SSE_PMULLD_ITINS>,
7299 defm VPCMPEQQ : SS48I_binop_rm<0x29, "vpcmpeqq", X86pcmpeq, v2i64, VR128,
7300 memopv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7303 let Predicates = [HasAVX2] in {
7304 defm VPMULLDY : SS48I_binop_rm<0x40, "vpmulld", mul, v8i32, VR256,
7305 memopv4i64, i256mem, 0, SSE_PMULLD_ITINS>,
7307 defm VPCMPEQQY : SS48I_binop_rm<0x29, "vpcmpeqq", X86pcmpeq, v4i64, VR256,
7308 memopv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7312 let Constraints = "$src1 = $dst" in {
7313 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, VR128,
7314 memopv2i64, i128mem, 1, SSE_PMULLD_ITINS>;
7315 defm PCMPEQQ : SS48I_binop_rm<0x29, "pcmpeqq", X86pcmpeq, v2i64, VR128,
7316 memopv2i64, i128mem, 1, SSE_INTALUQ_ITINS_P>;
7319 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
7320 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
7321 Intrinsic IntId, RegisterClass RC, PatFrag memop_frag,
7322 X86MemOperand x86memop, bit Is2Addr = 1,
7323 OpndItins itins = DEFAULT_ITINS> {
7324 let isCommutable = 1 in
7325 def rri : SS4AIi8<opc, MRMSrcReg, (outs RC:$dst),
7326 (ins RC:$src1, RC:$src2, i8imm:$src3),
7328 !strconcat(OpcodeStr,
7329 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
7330 !strconcat(OpcodeStr,
7331 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
7332 [(set RC:$dst, (IntId RC:$src1, RC:$src2, imm:$src3))], itins.rr>,
7333 Sched<[itins.Sched]>;
7334 def rmi : SS4AIi8<opc, MRMSrcMem, (outs RC:$dst),
7335 (ins RC:$src1, x86memop:$src2, i8imm:$src3),
7337 !strconcat(OpcodeStr,
7338 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
7339 !strconcat(OpcodeStr,
7340 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
7343 (bitconvert (memop_frag addr:$src2)), imm:$src3))], itins.rm>,
7344 Sched<[itins.Sched.Folded, ReadAfterLd]>;
7347 let Predicates = [HasAVX] in {
7348 let isCommutable = 0 in {
7349 let ExeDomain = SSEPackedSingle in {
7350 defm VBLENDPS : SS41I_binop_rmi_int<0x0C, "vblendps", int_x86_sse41_blendps,
7351 VR128, loadv4f32, f128mem, 0,
7352 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V;
7353 defm VBLENDPSY : SS41I_binop_rmi_int<0x0C, "vblendps",
7354 int_x86_avx_blend_ps_256, VR256, loadv8f32,
7355 f256mem, 0, DEFAULT_ITINS_FBLENDSCHED>,
7358 let ExeDomain = SSEPackedDouble in {
7359 defm VBLENDPD : SS41I_binop_rmi_int<0x0D, "vblendpd", int_x86_sse41_blendpd,
7360 VR128, loadv2f64, f128mem, 0,
7361 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V;
7362 defm VBLENDPDY : SS41I_binop_rmi_int<0x0D, "vblendpd",
7363 int_x86_avx_blend_pd_256,VR256, loadv4f64,
7364 f256mem, 0, DEFAULT_ITINS_FBLENDSCHED>,
7367 defm VPBLENDW : SS41I_binop_rmi_int<0x0E, "vpblendw", int_x86_sse41_pblendw,
7368 VR128, loadv2i64, i128mem, 0,
7369 DEFAULT_ITINS_BLENDSCHED>, VEX_4V;
7370 defm VMPSADBW : SS41I_binop_rmi_int<0x42, "vmpsadbw", int_x86_sse41_mpsadbw,
7371 VR128, loadv2i64, i128mem, 0,
7372 DEFAULT_ITINS_MPSADSCHED>, VEX_4V;
7374 let ExeDomain = SSEPackedSingle in
7375 defm VDPPS : SS41I_binop_rmi_int<0x40, "vdpps", int_x86_sse41_dpps,
7376 VR128, loadv4f32, f128mem, 0,
7377 SSE_DPPS_ITINS>, VEX_4V;
7378 let ExeDomain = SSEPackedDouble in
7379 defm VDPPD : SS41I_binop_rmi_int<0x41, "vdppd", int_x86_sse41_dppd,
7380 VR128, loadv2f64, f128mem, 0,
7381 SSE_DPPS_ITINS>, VEX_4V;
7382 let ExeDomain = SSEPackedSingle in
7383 defm VDPPSY : SS41I_binop_rmi_int<0x40, "vdpps", int_x86_avx_dp_ps_256,
7384 VR256, loadv8f32, i256mem, 0,
7385 SSE_DPPS_ITINS>, VEX_4V, VEX_L;
7388 let Predicates = [HasAVX2] in {
7389 let isCommutable = 0 in {
7390 defm VPBLENDWY : SS41I_binop_rmi_int<0x0E, "vpblendw", int_x86_avx2_pblendw,
7391 VR256, loadv4i64, i256mem, 0,
7392 DEFAULT_ITINS_BLENDSCHED>, VEX_4V, VEX_L;
7393 defm VMPSADBWY : SS41I_binop_rmi_int<0x42, "vmpsadbw", int_x86_avx2_mpsadbw,
7394 VR256, loadv4i64, i256mem, 0,
7395 DEFAULT_ITINS_MPSADSCHED>, VEX_4V, VEX_L;
7399 let Constraints = "$src1 = $dst" in {
7400 let isCommutable = 0 in {
7401 let ExeDomain = SSEPackedSingle in
7402 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps", int_x86_sse41_blendps,
7403 VR128, memopv4f32, f128mem,
7404 1, SSE_INTALU_ITINS_FBLEND_P>;
7405 let ExeDomain = SSEPackedDouble in
7406 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd", int_x86_sse41_blendpd,
7407 VR128, memopv2f64, f128mem,
7408 1, SSE_INTALU_ITINS_FBLEND_P>;
7409 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw", int_x86_sse41_pblendw,
7410 VR128, memopv2i64, i128mem,
7411 1, SSE_INTALU_ITINS_BLEND_P>;
7412 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw", int_x86_sse41_mpsadbw,
7413 VR128, memopv2i64, i128mem,
7414 1, SSE_MPSADBW_ITINS>;
7416 let ExeDomain = SSEPackedSingle in
7417 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps", int_x86_sse41_dpps,
7418 VR128, memopv4f32, f128mem, 1,
7420 let ExeDomain = SSEPackedDouble in
7421 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd", int_x86_sse41_dppd,
7422 VR128, memopv2f64, f128mem, 1,
7426 /// SS41I_quaternary_int_avx - AVX SSE 4.1 with 4 operators
7427 multiclass SS41I_quaternary_int_avx<bits<8> opc, string OpcodeStr,
7428 RegisterClass RC, X86MemOperand x86memop,
7429 PatFrag mem_frag, Intrinsic IntId,
7430 X86FoldableSchedWrite Sched> {
7431 def rr : Ii8<opc, MRMSrcReg, (outs RC:$dst),
7432 (ins RC:$src1, RC:$src2, RC:$src3),
7433 !strconcat(OpcodeStr,
7434 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
7435 [(set RC:$dst, (IntId RC:$src1, RC:$src2, RC:$src3))],
7436 NoItinerary, SSEPackedInt>, TAPD, VEX_4V, VEX_I8IMM,
7439 def rm : Ii8<opc, MRMSrcMem, (outs RC:$dst),
7440 (ins RC:$src1, x86memop:$src2, RC:$src3),
7441 !strconcat(OpcodeStr,
7442 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
7444 (IntId RC:$src1, (bitconvert (mem_frag addr:$src2)),
7446 NoItinerary, SSEPackedInt>, TAPD, VEX_4V, VEX_I8IMM,
7447 Sched<[Sched.Folded, ReadAfterLd]>;
7450 let Predicates = [HasAVX] in {
7451 let ExeDomain = SSEPackedDouble in {
7452 defm VBLENDVPD : SS41I_quaternary_int_avx<0x4B, "vblendvpd", VR128, f128mem,
7453 loadv2f64, int_x86_sse41_blendvpd,
7455 defm VBLENDVPDY : SS41I_quaternary_int_avx<0x4B, "vblendvpd", VR256, f256mem,
7456 loadv4f64, int_x86_avx_blendv_pd_256,
7457 WriteFVarBlend>, VEX_L;
7458 } // ExeDomain = SSEPackedDouble
7459 let ExeDomain = SSEPackedSingle in {
7460 defm VBLENDVPS : SS41I_quaternary_int_avx<0x4A, "vblendvps", VR128, f128mem,
7461 loadv4f32, int_x86_sse41_blendvps,
7463 defm VBLENDVPSY : SS41I_quaternary_int_avx<0x4A, "vblendvps", VR256, f256mem,
7464 loadv8f32, int_x86_avx_blendv_ps_256,
7465 WriteFVarBlend>, VEX_L;
7466 } // ExeDomain = SSEPackedSingle
7467 defm VPBLENDVB : SS41I_quaternary_int_avx<0x4C, "vpblendvb", VR128, i128mem,
7468 loadv2i64, int_x86_sse41_pblendvb,
7472 let Predicates = [HasAVX2] in {
7473 defm VPBLENDVBY : SS41I_quaternary_int_avx<0x4C, "vpblendvb", VR256, i256mem,
7474 loadv4i64, int_x86_avx2_pblendvb,
7475 WriteVarBlend>, VEX_L;
7478 let Predicates = [HasAVX] in {
7479 def : Pat<(v16i8 (vselect (v16i8 VR128:$mask), (v16i8 VR128:$src1),
7480 (v16i8 VR128:$src2))),
7481 (VPBLENDVBrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7482 def : Pat<(v4i32 (vselect (v4i32 VR128:$mask), (v4i32 VR128:$src1),
7483 (v4i32 VR128:$src2))),
7484 (VBLENDVPSrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7485 def : Pat<(v4f32 (vselect (v4i32 VR128:$mask), (v4f32 VR128:$src1),
7486 (v4f32 VR128:$src2))),
7487 (VBLENDVPSrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7488 def : Pat<(v2i64 (vselect (v2i64 VR128:$mask), (v2i64 VR128:$src1),
7489 (v2i64 VR128:$src2))),
7490 (VBLENDVPDrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7491 def : Pat<(v2f64 (vselect (v2i64 VR128:$mask), (v2f64 VR128:$src1),
7492 (v2f64 VR128:$src2))),
7493 (VBLENDVPDrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7494 def : Pat<(v8i32 (vselect (v8i32 VR256:$mask), (v8i32 VR256:$src1),
7495 (v8i32 VR256:$src2))),
7496 (VBLENDVPSYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7497 def : Pat<(v8f32 (vselect (v8i32 VR256:$mask), (v8f32 VR256:$src1),
7498 (v8f32 VR256:$src2))),
7499 (VBLENDVPSYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7500 def : Pat<(v4i64 (vselect (v4i64 VR256:$mask), (v4i64 VR256:$src1),
7501 (v4i64 VR256:$src2))),
7502 (VBLENDVPDYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7503 def : Pat<(v4f64 (vselect (v4i64 VR256:$mask), (v4f64 VR256:$src1),
7504 (v4f64 VR256:$src2))),
7505 (VBLENDVPDYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7507 def : Pat<(v8f32 (X86Blendi (v8f32 VR256:$src1), (v8f32 VR256:$src2),
7509 (VBLENDPSYrri VR256:$src1, VR256:$src2, imm:$mask)>;
7510 def : Pat<(v4f64 (X86Blendi (v4f64 VR256:$src1), (v4f64 VR256:$src2),
7512 (VBLENDPDYrri VR256:$src1, VR256:$src2, imm:$mask)>;
7514 def : Pat<(v8i16 (X86Blendi (v8i16 VR128:$src1), (v8i16 VR128:$src2),
7516 (VPBLENDWrri VR128:$src1, VR128:$src2, imm:$mask)>;
7517 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$src1), (v4f32 VR128:$src2),
7519 (VBLENDPSrri VR128:$src1, VR128:$src2, imm:$mask)>;
7520 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$src1), (v2f64 VR128:$src2),
7522 (VBLENDPDrri VR128:$src1, VR128:$src2, imm:$mask)>;
7525 let Predicates = [HasAVX2] in {
7526 def : Pat<(v32i8 (vselect (v32i8 VR256:$mask), (v32i8 VR256:$src1),
7527 (v32i8 VR256:$src2))),
7528 (VPBLENDVBYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7529 def : Pat<(v16i16 (X86Blendi (v16i16 VR256:$src1), (v16i16 VR256:$src2),
7531 (VPBLENDWYrri VR256:$src1, VR256:$src2, imm:$mask)>;
7535 let Predicates = [UseAVX] in {
7536 let AddedComplexity = 15 in {
7537 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
7538 // MOVS{S,D} to the lower bits.
7539 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
7540 (VMOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
7541 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
7542 (VBLENDPSrri (v4f32 (V_SET0)), VR128:$src, (i8 1))>;
7543 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
7544 (VBLENDPSrri (v4i32 (V_SET0)), VR128:$src, (i8 1))>;
7545 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
7546 (VMOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
7548 // Move low f32 and clear high bits.
7549 def : Pat<(v8f32 (X86vzmovl (v8f32 VR256:$src))),
7550 (VBLENDPSYrri (v8f32 (AVX_SET0)), VR256:$src, (i8 1))>;
7551 def : Pat<(v8i32 (X86vzmovl (v8i32 VR256:$src))),
7552 (VBLENDPSYrri (v8i32 (AVX_SET0)), VR256:$src, (i8 1))>;
7555 def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
7556 (v4f32 (scalar_to_vector FR32:$src)), (iPTR 0)))),
7557 (SUBREG_TO_REG (i32 0),
7558 (v4f32 (VMOVSSrr (v4f32 (V_SET0)), FR32:$src)),
7560 def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
7561 (v2f64 (scalar_to_vector FR64:$src)), (iPTR 0)))),
7562 (SUBREG_TO_REG (i64 0),
7563 (v2f64 (VMOVSDrr (v2f64 (V_SET0)), FR64:$src)),
7566 // Move low f64 and clear high bits.
7567 def : Pat<(v4f64 (X86vzmovl (v4f64 VR256:$src))),
7568 (VBLENDPDYrri (v4f64 (AVX_SET0)), VR256:$src, (i8 1))>;
7570 def : Pat<(v4i64 (X86vzmovl (v4i64 VR256:$src))),
7571 (VBLENDPDYrri (v4i64 (AVX_SET0)), VR256:$src, (i8 1))>;
7574 let Predicates = [UseSSE41] in {
7575 // With SSE41 we can use blends for these patterns.
7576 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
7577 (BLENDPSrri (v4f32 (V_SET0)), VR128:$src, (i8 1))>;
7578 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
7579 (BLENDPSrri (v4f32 (V_SET0)), VR128:$src, (i8 1))>;
7580 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
7581 (BLENDPDrri (v2f64 (V_SET0)), VR128:$src, (i8 1))>;
7585 /// SS41I_ternary_int - SSE 4.1 ternary operator
7586 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
7587 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
7588 X86MemOperand x86memop, Intrinsic IntId,
7589 OpndItins itins = DEFAULT_ITINS> {
7590 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
7591 (ins VR128:$src1, VR128:$src2),
7592 !strconcat(OpcodeStr,
7593 "\t{$src2, $dst|$dst, $src2}"),
7594 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))],
7595 itins.rr>, Sched<[itins.Sched]>;
7597 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
7598 (ins VR128:$src1, x86memop:$src2),
7599 !strconcat(OpcodeStr,
7600 "\t{$src2, $dst|$dst, $src2}"),
7603 (bitconvert (mem_frag addr:$src2)), XMM0))],
7604 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
7608 let ExeDomain = SSEPackedDouble in
7609 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", memopv2f64, f128mem,
7610 int_x86_sse41_blendvpd,
7611 DEFAULT_ITINS_FBLENDSCHED>;
7612 let ExeDomain = SSEPackedSingle in
7613 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", memopv4f32, f128mem,
7614 int_x86_sse41_blendvps,
7615 DEFAULT_ITINS_FBLENDSCHED>;
7616 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", memopv2i64, i128mem,
7617 int_x86_sse41_pblendvb,
7618 DEFAULT_ITINS_VARBLENDSCHED>;
7620 // Aliases with the implicit xmm0 argument
7621 def : InstAlias<"blendvpd\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7622 (BLENDVPDrr0 VR128:$dst, VR128:$src2)>;
7623 def : InstAlias<"blendvpd\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7624 (BLENDVPDrm0 VR128:$dst, f128mem:$src2)>;
7625 def : InstAlias<"blendvps\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7626 (BLENDVPSrr0 VR128:$dst, VR128:$src2)>;
7627 def : InstAlias<"blendvps\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7628 (BLENDVPSrm0 VR128:$dst, f128mem:$src2)>;
7629 def : InstAlias<"pblendvb\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7630 (PBLENDVBrr0 VR128:$dst, VR128:$src2)>;
7631 def : InstAlias<"pblendvb\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7632 (PBLENDVBrm0 VR128:$dst, i128mem:$src2)>;
7634 let Predicates = [UseSSE41] in {
7635 def : Pat<(v16i8 (vselect (v16i8 XMM0), (v16i8 VR128:$src1),
7636 (v16i8 VR128:$src2))),
7637 (PBLENDVBrr0 VR128:$src2, VR128:$src1)>;
7638 def : Pat<(v4i32 (vselect (v4i32 XMM0), (v4i32 VR128:$src1),
7639 (v4i32 VR128:$src2))),
7640 (BLENDVPSrr0 VR128:$src2, VR128:$src1)>;
7641 def : Pat<(v4f32 (vselect (v4i32 XMM0), (v4f32 VR128:$src1),
7642 (v4f32 VR128:$src2))),
7643 (BLENDVPSrr0 VR128:$src2, VR128:$src1)>;
7644 def : Pat<(v2i64 (vselect (v2i64 XMM0), (v2i64 VR128:$src1),
7645 (v2i64 VR128:$src2))),
7646 (BLENDVPDrr0 VR128:$src2, VR128:$src1)>;
7647 def : Pat<(v2f64 (vselect (v2i64 XMM0), (v2f64 VR128:$src1),
7648 (v2f64 VR128:$src2))),
7649 (BLENDVPDrr0 VR128:$src2, VR128:$src1)>;
7651 def : Pat<(v8i16 (X86Blendi (v8i16 VR128:$src1), (v8i16 VR128:$src2),
7653 (PBLENDWrri VR128:$src1, VR128:$src2, imm:$mask)>;
7654 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$src1), (v4f32 VR128:$src2),
7656 (BLENDPSrri VR128:$src1, VR128:$src2, imm:$mask)>;
7657 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$src1), (v2f64 VR128:$src2),
7659 (BLENDPDrri VR128:$src1, VR128:$src2, imm:$mask)>;
7663 let SchedRW = [WriteLoad] in {
7664 let Predicates = [HasAVX] in
7665 def VMOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
7666 "vmovntdqa\t{$src, $dst|$dst, $src}",
7667 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
7669 let Predicates = [HasAVX2] in
7670 def VMOVNTDQAYrm : SS48I<0x2A, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
7671 "vmovntdqa\t{$src, $dst|$dst, $src}",
7672 [(set VR256:$dst, (int_x86_avx2_movntdqa addr:$src))]>,
7674 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
7675 "movntdqa\t{$src, $dst|$dst, $src}",
7676 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;
7679 //===----------------------------------------------------------------------===//
7680 // SSE4.2 - Compare Instructions
7681 //===----------------------------------------------------------------------===//
7683 /// SS42I_binop_rm - Simple SSE 4.2 binary operator
7684 multiclass SS42I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
7685 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
7686 X86MemOperand x86memop, bit Is2Addr = 1> {
7687 def rr : SS428I<opc, MRMSrcReg, (outs RC:$dst),
7688 (ins RC:$src1, RC:$src2),
7690 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7691 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7692 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))]>;
7693 def rm : SS428I<opc, MRMSrcMem, (outs RC:$dst),
7694 (ins RC:$src1, x86memop:$src2),
7696 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7697 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7699 (OpVT (OpNode RC:$src1, (memop_frag addr:$src2))))]>;
7702 let Predicates = [HasAVX] in
7703 defm VPCMPGTQ : SS42I_binop_rm<0x37, "vpcmpgtq", X86pcmpgt, v2i64, VR128,
7704 loadv2i64, i128mem, 0>, VEX_4V;
7706 let Predicates = [HasAVX2] in
7707 defm VPCMPGTQY : SS42I_binop_rm<0x37, "vpcmpgtq", X86pcmpgt, v4i64, VR256,
7708 loadv4i64, i256mem, 0>, VEX_4V, VEX_L;
7710 let Constraints = "$src1 = $dst" in
7711 defm PCMPGTQ : SS42I_binop_rm<0x37, "pcmpgtq", X86pcmpgt, v2i64, VR128,
7712 memopv2i64, i128mem>;
7714 //===----------------------------------------------------------------------===//
7715 // SSE4.2 - String/text Processing Instructions
7716 //===----------------------------------------------------------------------===//
7718 // Packed Compare Implicit Length Strings, Return Mask
7719 multiclass pseudo_pcmpistrm<string asm> {
7720 def REG : PseudoI<(outs VR128:$dst),
7721 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
7722 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
7724 def MEM : PseudoI<(outs VR128:$dst),
7725 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
7726 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1,
7727 (bc_v16i8 (memopv2i64 addr:$src2)), imm:$src3))]>;
7730 let Defs = [EFLAGS], usesCustomInserter = 1 in {
7731 defm VPCMPISTRM128 : pseudo_pcmpistrm<"#VPCMPISTRM128">, Requires<[HasAVX]>;
7732 defm PCMPISTRM128 : pseudo_pcmpistrm<"#PCMPISTRM128">, Requires<[UseSSE42]>;
7735 multiclass pcmpistrm_SS42AI<string asm> {
7736 def rr : SS42AI<0x62, MRMSrcReg, (outs),
7737 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
7738 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7739 []>, Sched<[WritePCmpIStrM]>;
7741 def rm :SS42AI<0x62, MRMSrcMem, (outs),
7742 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
7743 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7744 []>, Sched<[WritePCmpIStrMLd, ReadAfterLd]>;
7747 let Defs = [XMM0, EFLAGS], neverHasSideEffects = 1 in {
7748 let Predicates = [HasAVX] in
7749 defm VPCMPISTRM128 : pcmpistrm_SS42AI<"vpcmpistrm">, VEX;
7750 defm PCMPISTRM128 : pcmpistrm_SS42AI<"pcmpistrm"> ;
7753 // Packed Compare Explicit Length Strings, Return Mask
7754 multiclass pseudo_pcmpestrm<string asm> {
7755 def REG : PseudoI<(outs VR128:$dst),
7756 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
7757 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
7758 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>;
7759 def MEM : PseudoI<(outs VR128:$dst),
7760 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
7761 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128 VR128:$src1, EAX,
7762 (bc_v16i8 (memopv2i64 addr:$src3)), EDX, imm:$src5))]>;
7765 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
7766 defm VPCMPESTRM128 : pseudo_pcmpestrm<"#VPCMPESTRM128">, Requires<[HasAVX]>;
7767 defm PCMPESTRM128 : pseudo_pcmpestrm<"#PCMPESTRM128">, Requires<[UseSSE42]>;
7770 multiclass SS42AI_pcmpestrm<string asm> {
7771 def rr : SS42AI<0x60, MRMSrcReg, (outs),
7772 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
7773 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7774 []>, Sched<[WritePCmpEStrM]>;
7776 def rm : SS42AI<0x60, MRMSrcMem, (outs),
7777 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
7778 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7779 []>, Sched<[WritePCmpEStrMLd, ReadAfterLd]>;
7782 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX], neverHasSideEffects = 1 in {
7783 let Predicates = [HasAVX] in
7784 defm VPCMPESTRM128 : SS42AI_pcmpestrm<"vpcmpestrm">, VEX;
7785 defm PCMPESTRM128 : SS42AI_pcmpestrm<"pcmpestrm">;
7788 // Packed Compare Implicit Length Strings, Return Index
7789 multiclass pseudo_pcmpistri<string asm> {
7790 def REG : PseudoI<(outs GR32:$dst),
7791 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
7792 [(set GR32:$dst, EFLAGS,
7793 (X86pcmpistri VR128:$src1, VR128:$src2, imm:$src3))]>;
7794 def MEM : PseudoI<(outs GR32:$dst),
7795 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
7796 [(set GR32:$dst, EFLAGS, (X86pcmpistri VR128:$src1,
7797 (bc_v16i8 (memopv2i64 addr:$src2)), imm:$src3))]>;
7800 let Defs = [EFLAGS], usesCustomInserter = 1 in {
7801 defm VPCMPISTRI : pseudo_pcmpistri<"#VPCMPISTRI">, Requires<[HasAVX]>;
7802 defm PCMPISTRI : pseudo_pcmpistri<"#PCMPISTRI">, Requires<[UseSSE42]>;
7805 multiclass SS42AI_pcmpistri<string asm> {
7806 def rr : SS42AI<0x63, MRMSrcReg, (outs),
7807 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
7808 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7809 []>, Sched<[WritePCmpIStrI]>;
7811 def rm : SS42AI<0x63, MRMSrcMem, (outs),
7812 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
7813 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7814 []>, Sched<[WritePCmpIStrILd, ReadAfterLd]>;
7817 let Defs = [ECX, EFLAGS], neverHasSideEffects = 1 in {
7818 let Predicates = [HasAVX] in
7819 defm VPCMPISTRI : SS42AI_pcmpistri<"vpcmpistri">, VEX;
7820 defm PCMPISTRI : SS42AI_pcmpistri<"pcmpistri">;
7823 // Packed Compare Explicit Length Strings, Return Index
7824 multiclass pseudo_pcmpestri<string asm> {
7825 def REG : PseudoI<(outs GR32:$dst),
7826 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
7827 [(set GR32:$dst, EFLAGS,
7828 (X86pcmpestri VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>;
7829 def MEM : PseudoI<(outs GR32:$dst),
7830 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
7831 [(set GR32:$dst, EFLAGS,
7832 (X86pcmpestri VR128:$src1, EAX, (bc_v16i8 (memopv2i64 addr:$src3)), EDX,
7836 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
7837 defm VPCMPESTRI : pseudo_pcmpestri<"#VPCMPESTRI">, Requires<[HasAVX]>;
7838 defm PCMPESTRI : pseudo_pcmpestri<"#PCMPESTRI">, Requires<[UseSSE42]>;
7841 multiclass SS42AI_pcmpestri<string asm> {
7842 def rr : SS42AI<0x61, MRMSrcReg, (outs),
7843 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
7844 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7845 []>, Sched<[WritePCmpEStrI]>;
7847 def rm : SS42AI<0x61, MRMSrcMem, (outs),
7848 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
7849 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7850 []>, Sched<[WritePCmpEStrILd, ReadAfterLd]>;
7853 let Defs = [ECX, EFLAGS], Uses = [EAX, EDX], neverHasSideEffects = 1 in {
7854 let Predicates = [HasAVX] in
7855 defm VPCMPESTRI : SS42AI_pcmpestri<"vpcmpestri">, VEX;
7856 defm PCMPESTRI : SS42AI_pcmpestri<"pcmpestri">;
7859 //===----------------------------------------------------------------------===//
7860 // SSE4.2 - CRC Instructions
7861 //===----------------------------------------------------------------------===//
7863 // No CRC instructions have AVX equivalents
7865 // crc intrinsic instruction
7866 // This set of instructions are only rm, the only difference is the size
7868 class SS42I_crc32r<bits<8> opc, string asm, RegisterClass RCOut,
7869 RegisterClass RCIn, SDPatternOperator Int> :
7870 SS42FI<opc, MRMSrcReg, (outs RCOut:$dst), (ins RCOut:$src1, RCIn:$src2),
7871 !strconcat(asm, "\t{$src2, $src1|$src1, $src2}"),
7872 [(set RCOut:$dst, (Int RCOut:$src1, RCIn:$src2))], IIC_CRC32_REG>,
7875 class SS42I_crc32m<bits<8> opc, string asm, RegisterClass RCOut,
7876 X86MemOperand x86memop, SDPatternOperator Int> :
7877 SS42FI<opc, MRMSrcMem, (outs RCOut:$dst), (ins RCOut:$src1, x86memop:$src2),
7878 !strconcat(asm, "\t{$src2, $src1|$src1, $src2}"),
7879 [(set RCOut:$dst, (Int RCOut:$src1, (load addr:$src2)))],
7880 IIC_CRC32_MEM>, Sched<[WriteFAddLd, ReadAfterLd]>;
7882 let Constraints = "$src1 = $dst" in {
7883 def CRC32r32m8 : SS42I_crc32m<0xF0, "crc32{b}", GR32, i8mem,
7884 int_x86_sse42_crc32_32_8>;
7885 def CRC32r32r8 : SS42I_crc32r<0xF0, "crc32{b}", GR32, GR8,
7886 int_x86_sse42_crc32_32_8>;
7887 def CRC32r32m16 : SS42I_crc32m<0xF1, "crc32{w}", GR32, i16mem,
7888 int_x86_sse42_crc32_32_16>, OpSize16;
7889 def CRC32r32r16 : SS42I_crc32r<0xF1, "crc32{w}", GR32, GR16,
7890 int_x86_sse42_crc32_32_16>, OpSize16;
7891 def CRC32r32m32 : SS42I_crc32m<0xF1, "crc32{l}", GR32, i32mem,
7892 int_x86_sse42_crc32_32_32>, OpSize32;
7893 def CRC32r32r32 : SS42I_crc32r<0xF1, "crc32{l}", GR32, GR32,
7894 int_x86_sse42_crc32_32_32>, OpSize32;
7895 def CRC32r64m64 : SS42I_crc32m<0xF1, "crc32{q}", GR64, i64mem,
7896 int_x86_sse42_crc32_64_64>, REX_W;
7897 def CRC32r64r64 : SS42I_crc32r<0xF1, "crc32{q}", GR64, GR64,
7898 int_x86_sse42_crc32_64_64>, REX_W;
7899 let hasSideEffects = 0 in {
7901 def CRC32r64m8 : SS42I_crc32m<0xF0, "crc32{b}", GR64, i8mem,
7903 def CRC32r64r8 : SS42I_crc32r<0xF0, "crc32{b}", GR64, GR8,
7908 //===----------------------------------------------------------------------===//
7909 // SHA-NI Instructions
7910 //===----------------------------------------------------------------------===//
7912 multiclass SHAI_binop<bits<8> Opc, string OpcodeStr, Intrinsic IntId,
7914 def rr : I<Opc, MRMSrcReg, (outs VR128:$dst),
7915 (ins VR128:$src1, VR128:$src2),
7916 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7918 (set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0)),
7919 (set VR128:$dst, (IntId VR128:$src1, VR128:$src2)))]>, T8;
7921 def rm : I<Opc, MRMSrcMem, (outs VR128:$dst),
7922 (ins VR128:$src1, i128mem:$src2),
7923 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7925 (set VR128:$dst, (IntId VR128:$src1,
7926 (bc_v4i32 (memopv2i64 addr:$src2)), XMM0)),
7927 (set VR128:$dst, (IntId VR128:$src1,
7928 (bc_v4i32 (memopv2i64 addr:$src2)))))]>, T8;
7931 let Constraints = "$src1 = $dst", Predicates = [HasSHA] in {
7932 def SHA1RNDS4rri : Ii8<0xCC, MRMSrcReg, (outs VR128:$dst),
7933 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
7934 "sha1rnds4\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7936 (int_x86_sha1rnds4 VR128:$src1, VR128:$src2,
7937 (i8 imm:$src3)))]>, TA;
7938 def SHA1RNDS4rmi : Ii8<0xCC, MRMSrcMem, (outs VR128:$dst),
7939 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
7940 "sha1rnds4\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7942 (int_x86_sha1rnds4 VR128:$src1,
7943 (bc_v4i32 (memopv2i64 addr:$src2)),
7944 (i8 imm:$src3)))]>, TA;
7946 defm SHA1NEXTE : SHAI_binop<0xC8, "sha1nexte", int_x86_sha1nexte>;
7947 defm SHA1MSG1 : SHAI_binop<0xC9, "sha1msg1", int_x86_sha1msg1>;
7948 defm SHA1MSG2 : SHAI_binop<0xCA, "sha1msg2", int_x86_sha1msg2>;
7951 defm SHA256RNDS2 : SHAI_binop<0xCB, "sha256rnds2", int_x86_sha256rnds2, 1>;
7953 defm SHA256MSG1 : SHAI_binop<0xCC, "sha256msg1", int_x86_sha256msg1>;
7954 defm SHA256MSG2 : SHAI_binop<0xCD, "sha256msg2", int_x86_sha256msg2>;
7957 // Aliases with explicit %xmm0
7958 def : InstAlias<"sha256rnds2\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7959 (SHA256RNDS2rr VR128:$dst, VR128:$src2)>;
7960 def : InstAlias<"sha256rnds2\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7961 (SHA256RNDS2rm VR128:$dst, i128mem:$src2)>;
7963 //===----------------------------------------------------------------------===//
7964 // AES-NI Instructions
7965 //===----------------------------------------------------------------------===//
7967 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr,
7968 Intrinsic IntId128, bit Is2Addr = 1> {
7969 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
7970 (ins VR128:$src1, VR128:$src2),
7972 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7973 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7974 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
7975 Sched<[WriteAESDecEnc]>;
7976 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
7977 (ins VR128:$src1, i128mem:$src2),
7979 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7980 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7982 (IntId128 VR128:$src1, (memopv2i64 addr:$src2)))]>,
7983 Sched<[WriteAESDecEncLd, ReadAfterLd]>;
7986 // Perform One Round of an AES Encryption/Decryption Flow
7987 let Predicates = [HasAVX, HasAES] in {
7988 defm VAESENC : AESI_binop_rm_int<0xDC, "vaesenc",
7989 int_x86_aesni_aesenc, 0>, VEX_4V;
7990 defm VAESENCLAST : AESI_binop_rm_int<0xDD, "vaesenclast",
7991 int_x86_aesni_aesenclast, 0>, VEX_4V;
7992 defm VAESDEC : AESI_binop_rm_int<0xDE, "vaesdec",
7993 int_x86_aesni_aesdec, 0>, VEX_4V;
7994 defm VAESDECLAST : AESI_binop_rm_int<0xDF, "vaesdeclast",
7995 int_x86_aesni_aesdeclast, 0>, VEX_4V;
7998 let Constraints = "$src1 = $dst" in {
7999 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
8000 int_x86_aesni_aesenc>;
8001 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
8002 int_x86_aesni_aesenclast>;
8003 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
8004 int_x86_aesni_aesdec>;
8005 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
8006 int_x86_aesni_aesdeclast>;
8009 // Perform the AES InvMixColumn Transformation
8010 let Predicates = [HasAVX, HasAES] in {
8011 def VAESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
8013 "vaesimc\t{$src1, $dst|$dst, $src1}",
8015 (int_x86_aesni_aesimc VR128:$src1))]>, Sched<[WriteAESIMC]>,
8017 def VAESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
8018 (ins i128mem:$src1),
8019 "vaesimc\t{$src1, $dst|$dst, $src1}",
8020 [(set VR128:$dst, (int_x86_aesni_aesimc (loadv2i64 addr:$src1)))]>,
8021 Sched<[WriteAESIMCLd]>, VEX;
8023 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
8025 "aesimc\t{$src1, $dst|$dst, $src1}",
8027 (int_x86_aesni_aesimc VR128:$src1))]>, Sched<[WriteAESIMC]>;
8028 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
8029 (ins i128mem:$src1),
8030 "aesimc\t{$src1, $dst|$dst, $src1}",
8031 [(set VR128:$dst, (int_x86_aesni_aesimc (memopv2i64 addr:$src1)))]>,
8032 Sched<[WriteAESIMCLd]>;
8034 // AES Round Key Generation Assist
8035 let Predicates = [HasAVX, HasAES] in {
8036 def VAESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
8037 (ins VR128:$src1, i8imm:$src2),
8038 "vaeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8040 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
8041 Sched<[WriteAESKeyGen]>, VEX;
8042 def VAESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
8043 (ins i128mem:$src1, i8imm:$src2),
8044 "vaeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8046 (int_x86_aesni_aeskeygenassist (loadv2i64 addr:$src1), imm:$src2))]>,
8047 Sched<[WriteAESKeyGenLd]>, VEX;
8049 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
8050 (ins VR128:$src1, i8imm:$src2),
8051 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8053 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
8054 Sched<[WriteAESKeyGen]>;
8055 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
8056 (ins i128mem:$src1, i8imm:$src2),
8057 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8059 (int_x86_aesni_aeskeygenassist (memopv2i64 addr:$src1), imm:$src2))]>,
8060 Sched<[WriteAESKeyGenLd]>;
8062 //===----------------------------------------------------------------------===//
8063 // PCLMUL Instructions
8064 //===----------------------------------------------------------------------===//
8066 // AVX carry-less Multiplication instructions
8067 def VPCLMULQDQrr : AVXPCLMULIi8<0x44, MRMSrcReg, (outs VR128:$dst),
8068 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
8069 "vpclmulqdq\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8071 (int_x86_pclmulqdq VR128:$src1, VR128:$src2, imm:$src3))]>,
8072 Sched<[WriteCLMul]>;
8074 def VPCLMULQDQrm : AVXPCLMULIi8<0x44, MRMSrcMem, (outs VR128:$dst),
8075 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
8076 "vpclmulqdq\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8077 [(set VR128:$dst, (int_x86_pclmulqdq VR128:$src1,
8078 (loadv2i64 addr:$src2), imm:$src3))]>,
8079 Sched<[WriteCLMulLd, ReadAfterLd]>;
8081 // Carry-less Multiplication instructions
8082 let Constraints = "$src1 = $dst" in {
8083 def PCLMULQDQrr : PCLMULIi8<0x44, MRMSrcReg, (outs VR128:$dst),
8084 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
8085 "pclmulqdq\t{$src3, $src2, $dst|$dst, $src2, $src3}",
8087 (int_x86_pclmulqdq VR128:$src1, VR128:$src2, imm:$src3))],
8088 IIC_SSE_PCLMULQDQ_RR>, Sched<[WriteCLMul]>;
8090 def PCLMULQDQrm : PCLMULIi8<0x44, MRMSrcMem, (outs VR128:$dst),
8091 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
8092 "pclmulqdq\t{$src3, $src2, $dst|$dst, $src2, $src3}",
8093 [(set VR128:$dst, (int_x86_pclmulqdq VR128:$src1,
8094 (memopv2i64 addr:$src2), imm:$src3))],
8095 IIC_SSE_PCLMULQDQ_RM>,
8096 Sched<[WriteCLMulLd, ReadAfterLd]>;
8097 } // Constraints = "$src1 = $dst"
8100 multiclass pclmul_alias<string asm, int immop> {
8101 def : InstAlias<!strconcat("pclmul", asm, "dq {$src, $dst|$dst, $src}"),
8102 (PCLMULQDQrr VR128:$dst, VR128:$src, immop), 0>;
8104 def : InstAlias<!strconcat("pclmul", asm, "dq {$src, $dst|$dst, $src}"),
8105 (PCLMULQDQrm VR128:$dst, i128mem:$src, immop), 0>;
8107 def : InstAlias<!strconcat("vpclmul", asm,
8108 "dq {$src2, $src1, $dst|$dst, $src1, $src2}"),
8109 (VPCLMULQDQrr VR128:$dst, VR128:$src1, VR128:$src2, immop),
8112 def : InstAlias<!strconcat("vpclmul", asm,
8113 "dq {$src2, $src1, $dst|$dst, $src1, $src2}"),
8114 (VPCLMULQDQrm VR128:$dst, VR128:$src1, i128mem:$src2, immop),
8117 defm : pclmul_alias<"hqhq", 0x11>;
8118 defm : pclmul_alias<"hqlq", 0x01>;
8119 defm : pclmul_alias<"lqhq", 0x10>;
8120 defm : pclmul_alias<"lqlq", 0x00>;
8122 //===----------------------------------------------------------------------===//
8123 // SSE4A Instructions
8124 //===----------------------------------------------------------------------===//
8126 let Predicates = [HasSSE4A] in {
8128 let Constraints = "$src = $dst" in {
8129 def EXTRQI : Ii8<0x78, MRMXr, (outs VR128:$dst),
8130 (ins VR128:$src, i8imm:$len, i8imm:$idx),
8131 "extrq\t{$idx, $len, $src|$src, $len, $idx}",
8132 [(set VR128:$dst, (int_x86_sse4a_extrqi VR128:$src, imm:$len,
8134 def EXTRQ : I<0x79, MRMSrcReg, (outs VR128:$dst),
8135 (ins VR128:$src, VR128:$mask),
8136 "extrq\t{$mask, $src|$src, $mask}",
8137 [(set VR128:$dst, (int_x86_sse4a_extrq VR128:$src,
8138 VR128:$mask))]>, PD;
8140 def INSERTQI : Ii8<0x78, MRMSrcReg, (outs VR128:$dst),
8141 (ins VR128:$src, VR128:$src2, i8imm:$len, i8imm:$idx),
8142 "insertq\t{$idx, $len, $src2, $src|$src, $src2, $len, $idx}",
8143 [(set VR128:$dst, (int_x86_sse4a_insertqi VR128:$src,
8144 VR128:$src2, imm:$len, imm:$idx))]>, XD;
8145 def INSERTQ : I<0x79, MRMSrcReg, (outs VR128:$dst),
8146 (ins VR128:$src, VR128:$mask),
8147 "insertq\t{$mask, $src|$src, $mask}",
8148 [(set VR128:$dst, (int_x86_sse4a_insertq VR128:$src,
8149 VR128:$mask))]>, XD;
8152 def MOVNTSS : I<0x2B, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
8153 "movntss\t{$src, $dst|$dst, $src}",
8154 [(int_x86_sse4a_movnt_ss addr:$dst, VR128:$src)]>, XS;
8156 def MOVNTSD : I<0x2B, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
8157 "movntsd\t{$src, $dst|$dst, $src}",
8158 [(int_x86_sse4a_movnt_sd addr:$dst, VR128:$src)]>, XD;
8161 //===----------------------------------------------------------------------===//
8163 //===----------------------------------------------------------------------===//
8165 //===----------------------------------------------------------------------===//
8166 // VBROADCAST - Load from memory and broadcast to all elements of the
8167 // destination operand
8169 class avx_broadcast<bits<8> opc, string OpcodeStr, RegisterClass RC,
8170 X86MemOperand x86memop, Intrinsic Int, SchedWrite Sched> :
8171 AVX8I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
8172 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8173 [(set RC:$dst, (Int addr:$src))]>, Sched<[Sched]>, VEX;
8175 class avx_broadcast_no_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
8176 X86MemOperand x86memop, ValueType VT,
8177 PatFrag ld_frag, SchedWrite Sched> :
8178 AVX8I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
8179 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8180 [(set RC:$dst, (VT (X86VBroadcast (ld_frag addr:$src))))]>,
8181 Sched<[Sched]>, VEX {
8185 // AVX2 adds register forms
8186 class avx2_broadcast_reg<bits<8> opc, string OpcodeStr, RegisterClass RC,
8187 Intrinsic Int, SchedWrite Sched> :
8188 AVX28I<opc, MRMSrcReg, (outs RC:$dst), (ins VR128:$src),
8189 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8190 [(set RC:$dst, (Int VR128:$src))]>, Sched<[Sched]>, VEX;
8192 let ExeDomain = SSEPackedSingle in {
8193 def VBROADCASTSSrm : avx_broadcast_no_int<0x18, "vbroadcastss", VR128,
8194 f32mem, v4f32, loadf32, WriteLoad>;
8195 def VBROADCASTSSYrm : avx_broadcast_no_int<0x18, "vbroadcastss", VR256,
8196 f32mem, v8f32, loadf32,
8197 WriteFShuffleLd>, VEX_L;
8199 let ExeDomain = SSEPackedDouble in
8200 def VBROADCASTSDYrm : avx_broadcast_no_int<0x19, "vbroadcastsd", VR256, f64mem,
8201 v4f64, loadf64, WriteFShuffleLd>, VEX_L;
8202 def VBROADCASTF128 : avx_broadcast<0x1A, "vbroadcastf128", VR256, f128mem,
8203 int_x86_avx_vbroadcastf128_pd_256,
8204 WriteFShuffleLd>, VEX_L;
8206 let ExeDomain = SSEPackedSingle in {
8207 def VBROADCASTSSrr : avx2_broadcast_reg<0x18, "vbroadcastss", VR128,
8208 int_x86_avx2_vbroadcast_ss_ps,
8210 def VBROADCASTSSYrr : avx2_broadcast_reg<0x18, "vbroadcastss", VR256,
8211 int_x86_avx2_vbroadcast_ss_ps_256,
8212 WriteFShuffle256>, VEX_L;
8214 let ExeDomain = SSEPackedDouble in
8215 def VBROADCASTSDYrr : avx2_broadcast_reg<0x19, "vbroadcastsd", VR256,
8216 int_x86_avx2_vbroadcast_sd_pd_256,
8217 WriteFShuffle256>, VEX_L;
8219 let Predicates = [HasAVX2] in
8220 def VBROADCASTI128 : avx_broadcast<0x5A, "vbroadcasti128", VR256, i128mem,
8221 int_x86_avx2_vbroadcasti128, WriteLoad>,
8224 let Predicates = [HasAVX] in
8225 def : Pat<(int_x86_avx_vbroadcastf128_ps_256 addr:$src),
8226 (VBROADCASTF128 addr:$src)>;
8229 //===----------------------------------------------------------------------===//
8230 // VINSERTF128 - Insert packed floating-point values
8232 let neverHasSideEffects = 1, ExeDomain = SSEPackedSingle in {
8233 def VINSERTF128rr : AVXAIi8<0x18, MRMSrcReg, (outs VR256:$dst),
8234 (ins VR256:$src1, VR128:$src2, i8imm:$src3),
8235 "vinsertf128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8236 []>, Sched<[WriteFShuffle]>, VEX_4V, VEX_L;
8238 def VINSERTF128rm : AVXAIi8<0x18, MRMSrcMem, (outs VR256:$dst),
8239 (ins VR256:$src1, f128mem:$src2, i8imm:$src3),
8240 "vinsertf128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8241 []>, Sched<[WriteFShuffleLd, ReadAfterLd]>, VEX_4V, VEX_L;
8244 let Predicates = [HasAVX] in {
8245 def : Pat<(vinsert128_insert:$ins (v8f32 VR256:$src1), (v4f32 VR128:$src2),
8247 (VINSERTF128rr VR256:$src1, VR128:$src2,
8248 (INSERT_get_vinsert128_imm VR256:$ins))>;
8249 def : Pat<(vinsert128_insert:$ins (v4f64 VR256:$src1), (v2f64 VR128:$src2),
8251 (VINSERTF128rr VR256:$src1, VR128:$src2,
8252 (INSERT_get_vinsert128_imm VR256:$ins))>;
8254 def : Pat<(vinsert128_insert:$ins (v8f32 VR256:$src1), (loadv4f32 addr:$src2),
8256 (VINSERTF128rm VR256:$src1, addr:$src2,
8257 (INSERT_get_vinsert128_imm VR256:$ins))>;
8258 def : Pat<(vinsert128_insert:$ins (v4f64 VR256:$src1), (loadv2f64 addr:$src2),
8260 (VINSERTF128rm VR256:$src1, addr:$src2,
8261 (INSERT_get_vinsert128_imm VR256:$ins))>;
8264 let Predicates = [HasAVX1Only] in {
8265 def : Pat<(vinsert128_insert:$ins (v4i64 VR256:$src1), (v2i64 VR128:$src2),
8267 (VINSERTF128rr VR256:$src1, VR128:$src2,
8268 (INSERT_get_vinsert128_imm VR256:$ins))>;
8269 def : Pat<(vinsert128_insert:$ins (v8i32 VR256:$src1), (v4i32 VR128:$src2),
8271 (VINSERTF128rr VR256:$src1, VR128:$src2,
8272 (INSERT_get_vinsert128_imm VR256:$ins))>;
8273 def : Pat<(vinsert128_insert:$ins (v32i8 VR256:$src1), (v16i8 VR128:$src2),
8275 (VINSERTF128rr VR256:$src1, VR128:$src2,
8276 (INSERT_get_vinsert128_imm VR256:$ins))>;
8277 def : Pat<(vinsert128_insert:$ins (v16i16 VR256:$src1), (v8i16 VR128:$src2),
8279 (VINSERTF128rr VR256:$src1, VR128:$src2,
8280 (INSERT_get_vinsert128_imm VR256:$ins))>;
8282 def : Pat<(vinsert128_insert:$ins (v4i64 VR256:$src1), (loadv2i64 addr:$src2),
8284 (VINSERTF128rm VR256:$src1, addr:$src2,
8285 (INSERT_get_vinsert128_imm VR256:$ins))>;
8286 def : Pat<(vinsert128_insert:$ins (v8i32 VR256:$src1),
8287 (bc_v4i32 (loadv2i64 addr:$src2)),
8289 (VINSERTF128rm VR256:$src1, addr:$src2,
8290 (INSERT_get_vinsert128_imm VR256:$ins))>;
8291 def : Pat<(vinsert128_insert:$ins (v32i8 VR256:$src1),
8292 (bc_v16i8 (loadv2i64 addr:$src2)),
8294 (VINSERTF128rm VR256:$src1, addr:$src2,
8295 (INSERT_get_vinsert128_imm VR256:$ins))>;
8296 def : Pat<(vinsert128_insert:$ins (v16i16 VR256:$src1),
8297 (bc_v8i16 (loadv2i64 addr:$src2)),
8299 (VINSERTF128rm VR256:$src1, addr:$src2,
8300 (INSERT_get_vinsert128_imm VR256:$ins))>;
8303 //===----------------------------------------------------------------------===//
8304 // VEXTRACTF128 - Extract packed floating-point values
8306 let neverHasSideEffects = 1, ExeDomain = SSEPackedSingle in {
8307 def VEXTRACTF128rr : AVXAIi8<0x19, MRMDestReg, (outs VR128:$dst),
8308 (ins VR256:$src1, i8imm:$src2),
8309 "vextractf128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8310 []>, Sched<[WriteFShuffle]>, VEX, VEX_L;
8312 def VEXTRACTF128mr : AVXAIi8<0x19, MRMDestMem, (outs),
8313 (ins f128mem:$dst, VR256:$src1, i8imm:$src2),
8314 "vextractf128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8315 []>, Sched<[WriteStore]>, VEX, VEX_L;
8319 let Predicates = [HasAVX] in {
8320 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8321 (v4f32 (VEXTRACTF128rr
8322 (v8f32 VR256:$src1),
8323 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8324 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8325 (v2f64 (VEXTRACTF128rr
8326 (v4f64 VR256:$src1),
8327 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8329 def : Pat<(store (v4f32 (vextract128_extract:$ext (v8f32 VR256:$src1),
8330 (iPTR imm))), addr:$dst),
8331 (VEXTRACTF128mr addr:$dst, VR256:$src1,
8332 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8333 def : Pat<(store (v2f64 (vextract128_extract:$ext (v4f64 VR256:$src1),
8334 (iPTR imm))), addr:$dst),
8335 (VEXTRACTF128mr addr:$dst, VR256:$src1,
8336 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8339 let Predicates = [HasAVX1Only] in {
8340 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8341 (v2i64 (VEXTRACTF128rr
8342 (v4i64 VR256:$src1),
8343 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8344 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8345 (v4i32 (VEXTRACTF128rr
8346 (v8i32 VR256:$src1),
8347 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8348 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8349 (v8i16 (VEXTRACTF128rr
8350 (v16i16 VR256:$src1),
8351 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8352 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8353 (v16i8 (VEXTRACTF128rr
8354 (v32i8 VR256:$src1),
8355 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8357 def : Pat<(alignedstore (v2i64 (vextract128_extract:$ext (v4i64 VR256:$src1),
8358 (iPTR imm))), addr:$dst),
8359 (VEXTRACTF128mr addr:$dst, VR256:$src1,
8360 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8361 def : Pat<(alignedstore (v4i32 (vextract128_extract:$ext (v8i32 VR256:$src1),
8362 (iPTR imm))), addr:$dst),
8363 (VEXTRACTF128mr addr:$dst, VR256:$src1,
8364 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8365 def : Pat<(alignedstore (v8i16 (vextract128_extract:$ext (v16i16 VR256:$src1),
8366 (iPTR imm))), addr:$dst),
8367 (VEXTRACTF128mr addr:$dst, VR256:$src1,
8368 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8369 def : Pat<(alignedstore (v16i8 (vextract128_extract:$ext (v32i8 VR256:$src1),
8370 (iPTR imm))), addr:$dst),
8371 (VEXTRACTF128mr addr:$dst, VR256:$src1,
8372 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8375 //===----------------------------------------------------------------------===//
8376 // VMASKMOV - Conditional SIMD Packed Loads and Stores
8378 multiclass avx_movmask_rm<bits<8> opc_rm, bits<8> opc_mr, string OpcodeStr,
8379 Intrinsic IntLd, Intrinsic IntLd256,
8380 Intrinsic IntSt, Intrinsic IntSt256> {
8381 def rm : AVX8I<opc_rm, MRMSrcMem, (outs VR128:$dst),
8382 (ins VR128:$src1, f128mem:$src2),
8383 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8384 [(set VR128:$dst, (IntLd addr:$src2, VR128:$src1))]>,
8386 def Yrm : AVX8I<opc_rm, MRMSrcMem, (outs VR256:$dst),
8387 (ins VR256:$src1, f256mem:$src2),
8388 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8389 [(set VR256:$dst, (IntLd256 addr:$src2, VR256:$src1))]>,
8391 def mr : AVX8I<opc_mr, MRMDestMem, (outs),
8392 (ins f128mem:$dst, VR128:$src1, VR128:$src2),
8393 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8394 [(IntSt addr:$dst, VR128:$src1, VR128:$src2)]>, VEX_4V;
8395 def Ymr : AVX8I<opc_mr, MRMDestMem, (outs),
8396 (ins f256mem:$dst, VR256:$src1, VR256:$src2),
8397 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8398 [(IntSt256 addr:$dst, VR256:$src1, VR256:$src2)]>, VEX_4V, VEX_L;
8401 let ExeDomain = SSEPackedSingle in
8402 defm VMASKMOVPS : avx_movmask_rm<0x2C, 0x2E, "vmaskmovps",
8403 int_x86_avx_maskload_ps,
8404 int_x86_avx_maskload_ps_256,
8405 int_x86_avx_maskstore_ps,
8406 int_x86_avx_maskstore_ps_256>;
8407 let ExeDomain = SSEPackedDouble in
8408 defm VMASKMOVPD : avx_movmask_rm<0x2D, 0x2F, "vmaskmovpd",
8409 int_x86_avx_maskload_pd,
8410 int_x86_avx_maskload_pd_256,
8411 int_x86_avx_maskstore_pd,
8412 int_x86_avx_maskstore_pd_256>;
8414 //===----------------------------------------------------------------------===//
8415 // VPERMIL - Permute Single and Double Floating-Point Values
8417 multiclass avx_permil<bits<8> opc_rm, bits<8> opc_rmi, string OpcodeStr,
8418 RegisterClass RC, X86MemOperand x86memop_f,
8419 X86MemOperand x86memop_i, PatFrag i_frag,
8420 Intrinsic IntVar, ValueType vt> {
8421 def rr : AVX8I<opc_rm, MRMSrcReg, (outs RC:$dst),
8422 (ins RC:$src1, RC:$src2),
8423 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8424 [(set RC:$dst, (IntVar RC:$src1, RC:$src2))]>, VEX_4V,
8425 Sched<[WriteFShuffle]>;
8426 def rm : AVX8I<opc_rm, MRMSrcMem, (outs RC:$dst),
8427 (ins RC:$src1, x86memop_i:$src2),
8428 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8429 [(set RC:$dst, (IntVar RC:$src1,
8430 (bitconvert (i_frag addr:$src2))))]>, VEX_4V,
8431 Sched<[WriteFShuffleLd, ReadAfterLd]>;
8433 def ri : AVXAIi8<opc_rmi, MRMSrcReg, (outs RC:$dst),
8434 (ins RC:$src1, i8imm:$src2),
8435 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8436 [(set RC:$dst, (vt (X86VPermilpi RC:$src1, (i8 imm:$src2))))]>, VEX,
8437 Sched<[WriteFShuffle]>;
8438 def mi : AVXAIi8<opc_rmi, MRMSrcMem, (outs RC:$dst),
8439 (ins x86memop_f:$src1, i8imm:$src2),
8440 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8442 (vt (X86VPermilpi (memop addr:$src1), (i8 imm:$src2))))]>, VEX,
8443 Sched<[WriteFShuffleLd]>;
8446 let ExeDomain = SSEPackedSingle in {
8447 defm VPERMILPS : avx_permil<0x0C, 0x04, "vpermilps", VR128, f128mem, i128mem,
8448 loadv2i64, int_x86_avx_vpermilvar_ps, v4f32>;
8449 defm VPERMILPSY : avx_permil<0x0C, 0x04, "vpermilps", VR256, f256mem, i256mem,
8450 loadv4i64, int_x86_avx_vpermilvar_ps_256, v8f32>, VEX_L;
8452 let ExeDomain = SSEPackedDouble in {
8453 defm VPERMILPD : avx_permil<0x0D, 0x05, "vpermilpd", VR128, f128mem, i128mem,
8454 loadv2i64, int_x86_avx_vpermilvar_pd, v2f64>;
8455 defm VPERMILPDY : avx_permil<0x0D, 0x05, "vpermilpd", VR256, f256mem, i256mem,
8456 loadv4i64, int_x86_avx_vpermilvar_pd_256, v4f64>, VEX_L;
8459 let Predicates = [HasAVX] in {
8460 def : Pat<(v8f32 (X86VPermilpv VR256:$src1, (v8i32 VR256:$src2))),
8461 (VPERMILPSYrr VR256:$src1, VR256:$src2)>;
8462 def : Pat<(v8f32 (X86VPermilpv VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
8463 (VPERMILPSYrm VR256:$src1, addr:$src2)>;
8464 def : Pat<(v4f64 (X86VPermilpv VR256:$src1, (v4i64 VR256:$src2))),
8465 (VPERMILPDYrr VR256:$src1, VR256:$src2)>;
8466 def : Pat<(v4f64 (X86VPermilpv VR256:$src1, (loadv4i64 addr:$src2))),
8467 (VPERMILPDYrm VR256:$src1, addr:$src2)>;
8469 def : Pat<(v8i32 (X86VPermilpi VR256:$src1, (i8 imm:$imm))),
8470 (VPERMILPSYri VR256:$src1, imm:$imm)>;
8471 def : Pat<(v4i64 (X86VPermilpi VR256:$src1, (i8 imm:$imm))),
8472 (VPERMILPDYri VR256:$src1, imm:$imm)>;
8473 def : Pat<(v8i32 (X86VPermilpi (bc_v8i32 (loadv4i64 addr:$src1)),
8475 (VPERMILPSYmi addr:$src1, imm:$imm)>;
8476 def : Pat<(v4i64 (X86VPermilpi (loadv4i64 addr:$src1), (i8 imm:$imm))),
8477 (VPERMILPDYmi addr:$src1, imm:$imm)>;
8479 def : Pat<(v4f32 (X86VPermilpv VR128:$src1, (v4i32 VR128:$src2))),
8480 (VPERMILPSrr VR128:$src1, VR128:$src2)>;
8481 def : Pat<(v4f32 (X86VPermilpv VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)))),
8482 (VPERMILPSrm VR128:$src1, addr:$src2)>;
8483 def : Pat<(v2f64 (X86VPermilpv VR128:$src1, (v2i64 VR128:$src2))),
8484 (VPERMILPDrr VR128:$src1, VR128:$src2)>;
8485 def : Pat<(v2f64 (X86VPermilpv VR128:$src1, (loadv2i64 addr:$src2))),
8486 (VPERMILPDrm VR128:$src1, addr:$src2)>;
8488 def : Pat<(v2i64 (X86VPermilpi VR128:$src1, (i8 imm:$imm))),
8489 (VPERMILPDri VR128:$src1, imm:$imm)>;
8490 def : Pat<(v2i64 (X86VPermilpi (loadv2i64 addr:$src1), (i8 imm:$imm))),
8491 (VPERMILPDmi addr:$src1, imm:$imm)>;
8494 //===----------------------------------------------------------------------===//
8495 // VPERM2F128 - Permute Floating-Point Values in 128-bit chunks
8497 let ExeDomain = SSEPackedSingle in {
8498 def VPERM2F128rr : AVXAIi8<0x06, MRMSrcReg, (outs VR256:$dst),
8499 (ins VR256:$src1, VR256:$src2, i8imm:$src3),
8500 "vperm2f128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8501 [(set VR256:$dst, (v8f32 (X86VPerm2x128 VR256:$src1, VR256:$src2,
8502 (i8 imm:$src3))))]>, VEX_4V, VEX_L,
8503 Sched<[WriteFShuffle]>;
8504 def VPERM2F128rm : AVXAIi8<0x06, MRMSrcMem, (outs VR256:$dst),
8505 (ins VR256:$src1, f256mem:$src2, i8imm:$src3),
8506 "vperm2f128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8507 [(set VR256:$dst, (X86VPerm2x128 VR256:$src1, (loadv8f32 addr:$src2),
8508 (i8 imm:$src3)))]>, VEX_4V, VEX_L,
8509 Sched<[WriteFShuffleLd, ReadAfterLd]>;
8512 let Predicates = [HasAVX] in {
8513 def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8514 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8515 def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1,
8516 (loadv4f64 addr:$src2), (i8 imm:$imm))),
8517 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8520 let Predicates = [HasAVX1Only] in {
8521 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8522 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8523 def : Pat<(v4i64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8524 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8525 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8526 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8527 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8528 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8530 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1,
8531 (bc_v8i32 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8532 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8533 def : Pat<(v4i64 (X86VPerm2x128 VR256:$src1,
8534 (loadv4i64 addr:$src2), (i8 imm:$imm))),
8535 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8536 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1,
8537 (bc_v32i8 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8538 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8539 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1,
8540 (bc_v16i16 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8541 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8544 //===----------------------------------------------------------------------===//
8545 // VZERO - Zero YMM registers
8547 let Defs = [YMM0, YMM1, YMM2, YMM3, YMM4, YMM5, YMM6, YMM7,
8548 YMM8, YMM9, YMM10, YMM11, YMM12, YMM13, YMM14, YMM15] in {
8549 // Zero All YMM registers
8550 def VZEROALL : I<0x77, RawFrm, (outs), (ins), "vzeroall",
8551 [(int_x86_avx_vzeroall)]>, PS, VEX, VEX_L, Requires<[HasAVX]>;
8553 // Zero Upper bits of YMM registers
8554 def VZEROUPPER : I<0x77, RawFrm, (outs), (ins), "vzeroupper",
8555 [(int_x86_avx_vzeroupper)]>, PS, VEX, Requires<[HasAVX]>;
8558 //===----------------------------------------------------------------------===//
8559 // Half precision conversion instructions
8560 //===----------------------------------------------------------------------===//
8561 multiclass f16c_ph2ps<RegisterClass RC, X86MemOperand x86memop, Intrinsic Int> {
8562 def rr : I<0x13, MRMSrcReg, (outs RC:$dst), (ins VR128:$src),
8563 "vcvtph2ps\t{$src, $dst|$dst, $src}",
8564 [(set RC:$dst, (Int VR128:$src))]>,
8565 T8PD, VEX, Sched<[WriteCvtF2F]>;
8566 let neverHasSideEffects = 1, mayLoad = 1 in
8567 def rm : I<0x13, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
8568 "vcvtph2ps\t{$src, $dst|$dst, $src}", []>, T8PD, VEX,
8569 Sched<[WriteCvtF2FLd]>;
8572 multiclass f16c_ps2ph<RegisterClass RC, X86MemOperand x86memop, Intrinsic Int> {
8573 def rr : Ii8<0x1D, MRMDestReg, (outs VR128:$dst),
8574 (ins RC:$src1, i32i8imm:$src2),
8575 "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8576 [(set VR128:$dst, (Int RC:$src1, imm:$src2))]>,
8577 TAPD, VEX, Sched<[WriteCvtF2F]>;
8578 let neverHasSideEffects = 1, mayStore = 1,
8579 SchedRW = [WriteCvtF2FLd, WriteRMW] in
8580 def mr : Ii8<0x1D, MRMDestMem, (outs),
8581 (ins x86memop:$dst, RC:$src1, i32i8imm:$src2),
8582 "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8586 let Predicates = [HasF16C] in {
8587 defm VCVTPH2PS : f16c_ph2ps<VR128, f64mem, int_x86_vcvtph2ps_128>;
8588 defm VCVTPH2PSY : f16c_ph2ps<VR256, f128mem, int_x86_vcvtph2ps_256>, VEX_L;
8589 defm VCVTPS2PH : f16c_ps2ph<VR128, f64mem, int_x86_vcvtps2ph_128>;
8590 defm VCVTPS2PHY : f16c_ps2ph<VR256, f128mem, int_x86_vcvtps2ph_256>, VEX_L;
8592 // Pattern match vcvtph2ps of a scalar i64 load.
8593 def : Pat<(int_x86_vcvtph2ps_128 (vzmovl_v2i64 addr:$src)),
8594 (VCVTPH2PSrm addr:$src)>;
8595 def : Pat<(int_x86_vcvtph2ps_128 (vzload_v2i64 addr:$src)),
8596 (VCVTPH2PSrm addr:$src)>;
8599 // Patterns for matching conversions from float to half-float and vice versa.
8600 let Predicates = [HasF16C] in {
8601 def : Pat<(fp_to_f16 FR32:$src),
8602 (i16 (EXTRACT_SUBREG (VMOVPDI2DIrr (VCVTPS2PHrr
8603 (COPY_TO_REGCLASS FR32:$src, VR128), 0)), sub_16bit))>;
8605 def : Pat<(f16_to_fp GR16:$src),
8606 (f32 (COPY_TO_REGCLASS (VCVTPH2PSrr
8607 (COPY_TO_REGCLASS (MOVSX32rr16 GR16:$src), VR128)), FR32)) >;
8609 def : Pat<(f16_to_fp (i16 (fp_to_f16 FR32:$src))),
8610 (f32 (COPY_TO_REGCLASS (VCVTPH2PSrr
8611 (VCVTPS2PHrr (COPY_TO_REGCLASS FR32:$src, VR128), 0)), FR32)) >;
8614 //===----------------------------------------------------------------------===//
8615 // AVX2 Instructions
8616 //===----------------------------------------------------------------------===//
8618 /// AVX2_binop_rmi_int - AVX2 binary operator with 8-bit immediate
8619 multiclass AVX2_binop_rmi_int<bits<8> opc, string OpcodeStr,
8620 Intrinsic IntId, RegisterClass RC, PatFrag memop_frag,
8621 X86MemOperand x86memop> {
8622 let isCommutable = 1 in
8623 def rri : AVX2AIi8<opc, MRMSrcReg, (outs RC:$dst),
8624 (ins RC:$src1, RC:$src2, i8imm:$src3),
8625 !strconcat(OpcodeStr,
8626 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
8627 [(set RC:$dst, (IntId RC:$src1, RC:$src2, imm:$src3))]>,
8628 Sched<[WriteBlend]>, VEX_4V;
8629 def rmi : AVX2AIi8<opc, MRMSrcMem, (outs RC:$dst),
8630 (ins RC:$src1, x86memop:$src2, i8imm:$src3),
8631 !strconcat(OpcodeStr,
8632 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
8635 (bitconvert (memop_frag addr:$src2)), imm:$src3))]>,
8636 Sched<[WriteBlendLd, ReadAfterLd]>, VEX_4V;
8639 let isCommutable = 0 in {
8640 defm VPBLENDD : AVX2_binop_rmi_int<0x02, "vpblendd", int_x86_avx2_pblendd_128,
8641 VR128, loadv2i64, i128mem>;
8642 defm VPBLENDDY : AVX2_binop_rmi_int<0x02, "vpblendd", int_x86_avx2_pblendd_256,
8643 VR256, loadv4i64, i256mem>, VEX_L;
8646 def : Pat<(v4i32 (X86Blendi (v4i32 VR128:$src1), (v4i32 VR128:$src2),
8648 (VPBLENDDrri VR128:$src1, VR128:$src2, imm:$mask)>;
8649 def : Pat<(v8i32 (X86Blendi (v8i32 VR256:$src1), (v8i32 VR256:$src2),
8651 (VPBLENDDYrri VR256:$src1, VR256:$src2, imm:$mask)>;
8653 //===----------------------------------------------------------------------===//
8654 // VPBROADCAST - Load from memory and broadcast to all elements of the
8655 // destination operand
8657 multiclass avx2_broadcast<bits<8> opc, string OpcodeStr,
8658 X86MemOperand x86memop, PatFrag ld_frag,
8659 Intrinsic Int128, Intrinsic Int256> {
8660 def rr : AVX28I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
8661 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8662 [(set VR128:$dst, (Int128 VR128:$src))]>,
8663 Sched<[WriteShuffle]>, VEX;
8664 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst), (ins x86memop:$src),
8665 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8667 (Int128 (scalar_to_vector (ld_frag addr:$src))))]>,
8668 Sched<[WriteLoad]>, VEX;
8669 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
8670 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8671 [(set VR256:$dst, (Int256 VR128:$src))]>,
8672 Sched<[WriteShuffle256]>, VEX, VEX_L;
8673 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst), (ins x86memop:$src),
8674 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8676 (Int256 (scalar_to_vector (ld_frag addr:$src))))]>,
8677 Sched<[WriteLoad]>, VEX, VEX_L;
8680 defm VPBROADCASTB : avx2_broadcast<0x78, "vpbroadcastb", i8mem, loadi8,
8681 int_x86_avx2_pbroadcastb_128,
8682 int_x86_avx2_pbroadcastb_256>;
8683 defm VPBROADCASTW : avx2_broadcast<0x79, "vpbroadcastw", i16mem, loadi16,
8684 int_x86_avx2_pbroadcastw_128,
8685 int_x86_avx2_pbroadcastw_256>;
8686 defm VPBROADCASTD : avx2_broadcast<0x58, "vpbroadcastd", i32mem, loadi32,
8687 int_x86_avx2_pbroadcastd_128,
8688 int_x86_avx2_pbroadcastd_256>;
8689 defm VPBROADCASTQ : avx2_broadcast<0x59, "vpbroadcastq", i64mem, loadi64,
8690 int_x86_avx2_pbroadcastq_128,
8691 int_x86_avx2_pbroadcastq_256>;
8693 let Predicates = [HasAVX2] in {
8694 def : Pat<(v16i8 (X86VBroadcast (loadi8 addr:$src))),
8695 (VPBROADCASTBrm addr:$src)>;
8696 def : Pat<(v32i8 (X86VBroadcast (loadi8 addr:$src))),
8697 (VPBROADCASTBYrm addr:$src)>;
8698 def : Pat<(v8i16 (X86VBroadcast (loadi16 addr:$src))),
8699 (VPBROADCASTWrm addr:$src)>;
8700 def : Pat<(v16i16 (X86VBroadcast (loadi16 addr:$src))),
8701 (VPBROADCASTWYrm addr:$src)>;
8702 def : Pat<(v4i32 (X86VBroadcast (loadi32 addr:$src))),
8703 (VPBROADCASTDrm addr:$src)>;
8704 def : Pat<(v8i32 (X86VBroadcast (loadi32 addr:$src))),
8705 (VPBROADCASTDYrm addr:$src)>;
8706 def : Pat<(v2i64 (X86VBroadcast (loadi64 addr:$src))),
8707 (VPBROADCASTQrm addr:$src)>;
8708 def : Pat<(v4i64 (X86VBroadcast (loadi64 addr:$src))),
8709 (VPBROADCASTQYrm addr:$src)>;
8711 def : Pat<(v16i8 (X86VBroadcast (v16i8 VR128:$src))),
8712 (VPBROADCASTBrr VR128:$src)>;
8713 def : Pat<(v32i8 (X86VBroadcast (v16i8 VR128:$src))),
8714 (VPBROADCASTBYrr VR128:$src)>;
8715 def : Pat<(v8i16 (X86VBroadcast (v8i16 VR128:$src))),
8716 (VPBROADCASTWrr VR128:$src)>;
8717 def : Pat<(v16i16 (X86VBroadcast (v8i16 VR128:$src))),
8718 (VPBROADCASTWYrr VR128:$src)>;
8719 def : Pat<(v4i32 (X86VBroadcast (v4i32 VR128:$src))),
8720 (VPBROADCASTDrr VR128:$src)>;
8721 def : Pat<(v8i32 (X86VBroadcast (v4i32 VR128:$src))),
8722 (VPBROADCASTDYrr VR128:$src)>;
8723 def : Pat<(v2i64 (X86VBroadcast (v2i64 VR128:$src))),
8724 (VPBROADCASTQrr VR128:$src)>;
8725 def : Pat<(v4i64 (X86VBroadcast (v2i64 VR128:$src))),
8726 (VPBROADCASTQYrr VR128:$src)>;
8727 def : Pat<(v4f32 (X86VBroadcast (v4f32 VR128:$src))),
8728 (VBROADCASTSSrr VR128:$src)>;
8729 def : Pat<(v8f32 (X86VBroadcast (v4f32 VR128:$src))),
8730 (VBROADCASTSSYrr VR128:$src)>;
8731 def : Pat<(v2f64 (X86VBroadcast (v2f64 VR128:$src))),
8732 (VPBROADCASTQrr VR128:$src)>;
8733 def : Pat<(v4f64 (X86VBroadcast (v2f64 VR128:$src))),
8734 (VBROADCASTSDYrr VR128:$src)>;
8736 // Provide aliases for broadcast from the same regitser class that
8737 // automatically does the extract.
8738 def : Pat<(v32i8 (X86VBroadcast (v32i8 VR256:$src))),
8739 (VPBROADCASTBYrr (v16i8 (EXTRACT_SUBREG (v32i8 VR256:$src),
8741 def : Pat<(v16i16 (X86VBroadcast (v16i16 VR256:$src))),
8742 (VPBROADCASTWYrr (v8i16 (EXTRACT_SUBREG (v16i16 VR256:$src),
8744 def : Pat<(v8i32 (X86VBroadcast (v8i32 VR256:$src))),
8745 (VPBROADCASTDYrr (v4i32 (EXTRACT_SUBREG (v8i32 VR256:$src),
8747 def : Pat<(v4i64 (X86VBroadcast (v4i64 VR256:$src))),
8748 (VPBROADCASTQYrr (v2i64 (EXTRACT_SUBREG (v4i64 VR256:$src),
8750 def : Pat<(v8f32 (X86VBroadcast (v8f32 VR256:$src))),
8751 (VBROADCASTSSYrr (v4f32 (EXTRACT_SUBREG (v8f32 VR256:$src),
8753 def : Pat<(v4f64 (X86VBroadcast (v4f64 VR256:$src))),
8754 (VBROADCASTSDYrr (v2f64 (EXTRACT_SUBREG (v4f64 VR256:$src),
8757 // Provide fallback in case the load node that is used in the patterns above
8758 // is used by additional users, which prevents the pattern selection.
8759 let AddedComplexity = 20 in {
8760 def : Pat<(v4f32 (X86VBroadcast FR32:$src)),
8761 (VBROADCASTSSrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
8762 def : Pat<(v8f32 (X86VBroadcast FR32:$src)),
8763 (VBROADCASTSSYrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
8764 def : Pat<(v4f64 (X86VBroadcast FR64:$src)),
8765 (VBROADCASTSDYrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
8767 def : Pat<(v4i32 (X86VBroadcast GR32:$src)),
8768 (VBROADCASTSSrr (COPY_TO_REGCLASS GR32:$src, VR128))>;
8769 def : Pat<(v8i32 (X86VBroadcast GR32:$src)),
8770 (VBROADCASTSSYrr (COPY_TO_REGCLASS GR32:$src, VR128))>;
8771 def : Pat<(v4i64 (X86VBroadcast GR64:$src)),
8772 (VBROADCASTSDYrr (COPY_TO_REGCLASS GR64:$src, VR128))>;
8774 def : Pat<(v16i8 (X86VBroadcast GR8:$src)),
8775 (VPBROADCASTBrr (COPY_TO_REGCLASS
8776 (i32 (SUBREG_TO_REG (i32 0), GR8:$src, sub_8bit)),
8778 def : Pat<(v32i8 (X86VBroadcast GR8:$src)),
8779 (VPBROADCASTBYrr (COPY_TO_REGCLASS
8780 (i32 (SUBREG_TO_REG (i32 0), GR8:$src, sub_8bit)),
8783 def : Pat<(v8i16 (X86VBroadcast GR16:$src)),
8784 (VPBROADCASTWrr (COPY_TO_REGCLASS
8785 (i32 (SUBREG_TO_REG (i32 0), GR16:$src, sub_16bit)),
8787 def : Pat<(v16i16 (X86VBroadcast GR16:$src)),
8788 (VPBROADCASTWYrr (COPY_TO_REGCLASS
8789 (i32 (SUBREG_TO_REG (i32 0), GR16:$src, sub_16bit)),
8792 // The patterns for VPBROADCASTD are not needed because they would match
8793 // the exact same thing as VBROADCASTSS patterns.
8795 def : Pat<(v2i64 (X86VBroadcast GR64:$src)),
8796 (VPBROADCASTQrr (COPY_TO_REGCLASS GR64:$src, VR128))>;
8797 // The v4i64 pattern is not needed because VBROADCASTSDYrr already match.
8801 // AVX1 broadcast patterns
8802 let Predicates = [HasAVX1Only] in {
8803 def : Pat<(v8i32 (X86VBroadcast (loadi32 addr:$src))),
8804 (VBROADCASTSSYrm addr:$src)>;
8805 def : Pat<(v4i64 (X86VBroadcast (loadi64 addr:$src))),
8806 (VBROADCASTSDYrm addr:$src)>;
8807 def : Pat<(v4i32 (X86VBroadcast (loadi32 addr:$src))),
8808 (VBROADCASTSSrm addr:$src)>;
8811 let Predicates = [HasAVX] in {
8812 // Provide fallback in case the load node that is used in the patterns above
8813 // is used by additional users, which prevents the pattern selection.
8814 let AddedComplexity = 20 in {
8815 // 128bit broadcasts:
8816 def : Pat<(v4f32 (X86VBroadcast FR32:$src)),
8817 (VPSHUFDri (COPY_TO_REGCLASS FR32:$src, VR128), 0)>;
8818 def : Pat<(v8f32 (X86VBroadcast FR32:$src)),
8819 (VINSERTF128rr (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)),
8820 (VPSHUFDri (COPY_TO_REGCLASS FR32:$src, VR128), 0), sub_xmm),
8821 (VPSHUFDri (COPY_TO_REGCLASS FR32:$src, VR128), 0), 1)>;
8822 def : Pat<(v4f64 (X86VBroadcast FR64:$src)),
8823 (VINSERTF128rr (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)),
8824 (VPSHUFDri (COPY_TO_REGCLASS FR64:$src, VR128), 0x44), sub_xmm),
8825 (VPSHUFDri (COPY_TO_REGCLASS FR64:$src, VR128), 0x44), 1)>;
8827 def : Pat<(v4i32 (X86VBroadcast GR32:$src)),
8828 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0)>;
8829 def : Pat<(v8i32 (X86VBroadcast GR32:$src)),
8830 (VINSERTF128rr (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8831 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0), sub_xmm),
8832 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0), 1)>;
8833 def : Pat<(v4i64 (X86VBroadcast GR64:$src)),
8834 (VINSERTF128rr (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)),
8835 (VPSHUFDri (COPY_TO_REGCLASS GR64:$src, VR128), 0x44), sub_xmm),
8836 (VPSHUFDri (COPY_TO_REGCLASS GR64:$src, VR128), 0x44), 1)>;
8839 def : Pat<(v2f64 (X86VBroadcast f64:$src)),
8840 (VMOVDDUPrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
8843 //===----------------------------------------------------------------------===//
8844 // VPERM - Permute instructions
8847 multiclass avx2_perm<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
8848 ValueType OpVT, X86FoldableSchedWrite Sched> {
8849 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst),
8850 (ins VR256:$src1, VR256:$src2),
8851 !strconcat(OpcodeStr,
8852 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8854 (OpVT (X86VPermv VR256:$src1, VR256:$src2)))]>,
8855 Sched<[Sched]>, VEX_4V, VEX_L;
8856 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst),
8857 (ins VR256:$src1, i256mem:$src2),
8858 !strconcat(OpcodeStr,
8859 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8861 (OpVT (X86VPermv VR256:$src1,
8862 (bitconvert (mem_frag addr:$src2)))))]>,
8863 Sched<[Sched.Folded, ReadAfterLd]>, VEX_4V, VEX_L;
8866 defm VPERMD : avx2_perm<0x36, "vpermd", loadv4i64, v8i32, WriteShuffle256>;
8867 let ExeDomain = SSEPackedSingle in
8868 defm VPERMPS : avx2_perm<0x16, "vpermps", loadv8f32, v8f32, WriteFShuffle256>;
8870 multiclass avx2_perm_imm<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
8871 ValueType OpVT, X86FoldableSchedWrite Sched> {
8872 def Yri : AVX2AIi8<opc, MRMSrcReg, (outs VR256:$dst),
8873 (ins VR256:$src1, i8imm:$src2),
8874 !strconcat(OpcodeStr,
8875 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8877 (OpVT (X86VPermi VR256:$src1, (i8 imm:$src2))))]>,
8878 Sched<[Sched]>, VEX, VEX_L;
8879 def Ymi : AVX2AIi8<opc, MRMSrcMem, (outs VR256:$dst),
8880 (ins i256mem:$src1, i8imm:$src2),
8881 !strconcat(OpcodeStr,
8882 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8884 (OpVT (X86VPermi (mem_frag addr:$src1),
8885 (i8 imm:$src2))))]>,
8886 Sched<[Sched.Folded, ReadAfterLd]>, VEX, VEX_L;
8889 defm VPERMQ : avx2_perm_imm<0x00, "vpermq", loadv4i64, v4i64,
8890 WriteShuffle256>, VEX_W;
8891 let ExeDomain = SSEPackedDouble in
8892 defm VPERMPD : avx2_perm_imm<0x01, "vpermpd", loadv4f64, v4f64,
8893 WriteFShuffle256>, VEX_W;
8895 //===----------------------------------------------------------------------===//
8896 // VPERM2I128 - Permute Floating-Point Values in 128-bit chunks
8898 def VPERM2I128rr : AVX2AIi8<0x46, MRMSrcReg, (outs VR256:$dst),
8899 (ins VR256:$src1, VR256:$src2, i8imm:$src3),
8900 "vperm2i128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8901 [(set VR256:$dst, (v4i64 (X86VPerm2x128 VR256:$src1, VR256:$src2,
8902 (i8 imm:$src3))))]>, Sched<[WriteShuffle256]>,
8904 def VPERM2I128rm : AVX2AIi8<0x46, MRMSrcMem, (outs VR256:$dst),
8905 (ins VR256:$src1, f256mem:$src2, i8imm:$src3),
8906 "vperm2i128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8907 [(set VR256:$dst, (X86VPerm2x128 VR256:$src1, (loadv4i64 addr:$src2),
8909 Sched<[WriteShuffle256Ld, ReadAfterLd]>, VEX_4V, VEX_L;
8911 let Predicates = [HasAVX2] in {
8912 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8913 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8914 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8915 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8916 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8917 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8919 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, (bc_v32i8 (loadv4i64 addr:$src2)),
8921 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8922 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1,
8923 (bc_v16i16 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8924 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8925 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)),
8927 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8931 //===----------------------------------------------------------------------===//
8932 // VINSERTI128 - Insert packed integer values
8934 let neverHasSideEffects = 1 in {
8935 def VINSERTI128rr : AVX2AIi8<0x38, MRMSrcReg, (outs VR256:$dst),
8936 (ins VR256:$src1, VR128:$src2, i8imm:$src3),
8937 "vinserti128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8938 []>, Sched<[WriteShuffle256]>, VEX_4V, VEX_L;
8940 def VINSERTI128rm : AVX2AIi8<0x38, MRMSrcMem, (outs VR256:$dst),
8941 (ins VR256:$src1, i128mem:$src2, i8imm:$src3),
8942 "vinserti128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8943 []>, Sched<[WriteShuffle256Ld, ReadAfterLd]>, VEX_4V, VEX_L;
8946 let Predicates = [HasAVX2] in {
8947 def : Pat<(vinsert128_insert:$ins (v4i64 VR256:$src1), (v2i64 VR128:$src2),
8949 (VINSERTI128rr VR256:$src1, VR128:$src2,
8950 (INSERT_get_vinsert128_imm VR256:$ins))>;
8951 def : Pat<(vinsert128_insert:$ins (v8i32 VR256:$src1), (v4i32 VR128:$src2),
8953 (VINSERTI128rr VR256:$src1, VR128:$src2,
8954 (INSERT_get_vinsert128_imm VR256:$ins))>;
8955 def : Pat<(vinsert128_insert:$ins (v32i8 VR256:$src1), (v16i8 VR128:$src2),
8957 (VINSERTI128rr VR256:$src1, VR128:$src2,
8958 (INSERT_get_vinsert128_imm VR256:$ins))>;
8959 def : Pat<(vinsert128_insert:$ins (v16i16 VR256:$src1), (v8i16 VR128:$src2),
8961 (VINSERTI128rr VR256:$src1, VR128:$src2,
8962 (INSERT_get_vinsert128_imm VR256:$ins))>;
8964 def : Pat<(vinsert128_insert:$ins (v4i64 VR256:$src1), (loadv2i64 addr:$src2),
8966 (VINSERTI128rm VR256:$src1, addr:$src2,
8967 (INSERT_get_vinsert128_imm VR256:$ins))>;
8968 def : Pat<(vinsert128_insert:$ins (v8i32 VR256:$src1),
8969 (bc_v4i32 (loadv2i64 addr:$src2)),
8971 (VINSERTI128rm VR256:$src1, addr:$src2,
8972 (INSERT_get_vinsert128_imm VR256:$ins))>;
8973 def : Pat<(vinsert128_insert:$ins (v32i8 VR256:$src1),
8974 (bc_v16i8 (loadv2i64 addr:$src2)),
8976 (VINSERTI128rm VR256:$src1, addr:$src2,
8977 (INSERT_get_vinsert128_imm VR256:$ins))>;
8978 def : Pat<(vinsert128_insert:$ins (v16i16 VR256:$src1),
8979 (bc_v8i16 (loadv2i64 addr:$src2)),
8981 (VINSERTI128rm VR256:$src1, addr:$src2,
8982 (INSERT_get_vinsert128_imm VR256:$ins))>;
8985 //===----------------------------------------------------------------------===//
8986 // VEXTRACTI128 - Extract packed integer values
8988 def VEXTRACTI128rr : AVX2AIi8<0x39, MRMDestReg, (outs VR128:$dst),
8989 (ins VR256:$src1, i8imm:$src2),
8990 "vextracti128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8992 (int_x86_avx2_vextracti128 VR256:$src1, imm:$src2))]>,
8993 Sched<[WriteShuffle256]>, VEX, VEX_L;
8994 let neverHasSideEffects = 1, mayStore = 1 in
8995 def VEXTRACTI128mr : AVX2AIi8<0x39, MRMDestMem, (outs),
8996 (ins i128mem:$dst, VR256:$src1, i8imm:$src2),
8997 "vextracti128\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8998 Sched<[WriteStore]>, VEX, VEX_L;
9000 let Predicates = [HasAVX2] in {
9001 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
9002 (v2i64 (VEXTRACTI128rr
9003 (v4i64 VR256:$src1),
9004 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
9005 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
9006 (v4i32 (VEXTRACTI128rr
9007 (v8i32 VR256:$src1),
9008 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
9009 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
9010 (v8i16 (VEXTRACTI128rr
9011 (v16i16 VR256:$src1),
9012 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
9013 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
9014 (v16i8 (VEXTRACTI128rr
9015 (v32i8 VR256:$src1),
9016 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
9018 def : Pat<(store (v2i64 (vextract128_extract:$ext (v4i64 VR256:$src1),
9019 (iPTR imm))), addr:$dst),
9020 (VEXTRACTI128mr addr:$dst, VR256:$src1,
9021 (EXTRACT_get_vextract128_imm VR128:$ext))>;
9022 def : Pat<(store (v4i32 (vextract128_extract:$ext (v8i32 VR256:$src1),
9023 (iPTR imm))), addr:$dst),
9024 (VEXTRACTI128mr addr:$dst, VR256:$src1,
9025 (EXTRACT_get_vextract128_imm VR128:$ext))>;
9026 def : Pat<(store (v8i16 (vextract128_extract:$ext (v16i16 VR256:$src1),
9027 (iPTR imm))), addr:$dst),
9028 (VEXTRACTI128mr addr:$dst, VR256:$src1,
9029 (EXTRACT_get_vextract128_imm VR128:$ext))>;
9030 def : Pat<(store (v16i8 (vextract128_extract:$ext (v32i8 VR256:$src1),
9031 (iPTR imm))), addr:$dst),
9032 (VEXTRACTI128mr addr:$dst, VR256:$src1,
9033 (EXTRACT_get_vextract128_imm VR128:$ext))>;
9036 //===----------------------------------------------------------------------===//
9037 // VPMASKMOV - Conditional SIMD Integer Packed Loads and Stores
9039 multiclass avx2_pmovmask<string OpcodeStr,
9040 Intrinsic IntLd128, Intrinsic IntLd256,
9041 Intrinsic IntSt128, Intrinsic IntSt256> {
9042 def rm : AVX28I<0x8c, MRMSrcMem, (outs VR128:$dst),
9043 (ins VR128:$src1, i128mem:$src2),
9044 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9045 [(set VR128:$dst, (IntLd128 addr:$src2, VR128:$src1))]>, VEX_4V;
9046 def Yrm : AVX28I<0x8c, MRMSrcMem, (outs VR256:$dst),
9047 (ins VR256:$src1, i256mem:$src2),
9048 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9049 [(set VR256:$dst, (IntLd256 addr:$src2, VR256:$src1))]>,
9051 def mr : AVX28I<0x8e, MRMDestMem, (outs),
9052 (ins i128mem:$dst, VR128:$src1, VR128:$src2),
9053 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9054 [(IntSt128 addr:$dst, VR128:$src1, VR128:$src2)]>, VEX_4V;
9055 def Ymr : AVX28I<0x8e, MRMDestMem, (outs),
9056 (ins i256mem:$dst, VR256:$src1, VR256:$src2),
9057 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9058 [(IntSt256 addr:$dst, VR256:$src1, VR256:$src2)]>, VEX_4V, VEX_L;
9061 defm VPMASKMOVD : avx2_pmovmask<"vpmaskmovd",
9062 int_x86_avx2_maskload_d,
9063 int_x86_avx2_maskload_d_256,
9064 int_x86_avx2_maskstore_d,
9065 int_x86_avx2_maskstore_d_256>;
9066 defm VPMASKMOVQ : avx2_pmovmask<"vpmaskmovq",
9067 int_x86_avx2_maskload_q,
9068 int_x86_avx2_maskload_q_256,
9069 int_x86_avx2_maskstore_q,
9070 int_x86_avx2_maskstore_q_256>, VEX_W;
9073 //===----------------------------------------------------------------------===//
9074 // Variable Bit Shifts
9076 multiclass avx2_var_shift<bits<8> opc, string OpcodeStr, SDNode OpNode,
9077 ValueType vt128, ValueType vt256> {
9078 def rr : AVX28I<opc, MRMSrcReg, (outs VR128:$dst),
9079 (ins VR128:$src1, VR128:$src2),
9080 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9082 (vt128 (OpNode VR128:$src1, (vt128 VR128:$src2))))]>,
9083 VEX_4V, Sched<[WriteVarVecShift]>;
9084 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst),
9085 (ins VR128:$src1, i128mem:$src2),
9086 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9088 (vt128 (OpNode VR128:$src1,
9089 (vt128 (bitconvert (loadv2i64 addr:$src2))))))]>,
9090 VEX_4V, Sched<[WriteVarVecShiftLd, ReadAfterLd]>;
9091 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst),
9092 (ins VR256:$src1, VR256:$src2),
9093 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9095 (vt256 (OpNode VR256:$src1, (vt256 VR256:$src2))))]>,
9096 VEX_4V, VEX_L, Sched<[WriteVarVecShift]>;
9097 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst),
9098 (ins VR256:$src1, i256mem:$src2),
9099 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9101 (vt256 (OpNode VR256:$src1,
9102 (vt256 (bitconvert (loadv4i64 addr:$src2))))))]>,
9103 VEX_4V, VEX_L, Sched<[WriteVarVecShiftLd, ReadAfterLd]>;
9106 defm VPSLLVD : avx2_var_shift<0x47, "vpsllvd", shl, v4i32, v8i32>;
9107 defm VPSLLVQ : avx2_var_shift<0x47, "vpsllvq", shl, v2i64, v4i64>, VEX_W;
9108 defm VPSRLVD : avx2_var_shift<0x45, "vpsrlvd", srl, v4i32, v8i32>;
9109 defm VPSRLVQ : avx2_var_shift<0x45, "vpsrlvq", srl, v2i64, v4i64>, VEX_W;
9110 defm VPSRAVD : avx2_var_shift<0x46, "vpsravd", sra, v4i32, v8i32>;
9112 //===----------------------------------------------------------------------===//
9113 // VGATHER - GATHER Operations
9114 multiclass avx2_gather<bits<8> opc, string OpcodeStr, RegisterClass RC256,
9115 X86MemOperand memop128, X86MemOperand memop256> {
9116 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst, VR128:$mask_wb),
9117 (ins VR128:$src1, memop128:$src2, VR128:$mask),
9118 !strconcat(OpcodeStr,
9119 "\t{$mask, $src2, $dst|$dst, $src2, $mask}"),
9121 def Yrm : AVX28I<opc, MRMSrcMem, (outs RC256:$dst, RC256:$mask_wb),
9122 (ins RC256:$src1, memop256:$src2, RC256:$mask),
9123 !strconcat(OpcodeStr,
9124 "\t{$mask, $src2, $dst|$dst, $src2, $mask}"),
9125 []>, VEX_4VOp3, VEX_L;
9128 let mayLoad = 1, Constraints
9129 = "@earlyclobber $dst,@earlyclobber $mask_wb, $src1 = $dst, $mask = $mask_wb"
9131 defm VPGATHERDQ : avx2_gather<0x90, "vpgatherdq", VR256, vx64mem, vx64mem>, VEX_W;
9132 defm VPGATHERQQ : avx2_gather<0x91, "vpgatherqq", VR256, vx64mem, vy64mem>, VEX_W;
9133 defm VPGATHERDD : avx2_gather<0x90, "vpgatherdd", VR256, vx32mem, vy32mem>;
9134 defm VPGATHERQD : avx2_gather<0x91, "vpgatherqd", VR128, vx32mem, vy32mem>;
9136 let ExeDomain = SSEPackedDouble in {
9137 defm VGATHERDPD : avx2_gather<0x92, "vgatherdpd", VR256, vx64mem, vx64mem>, VEX_W;
9138 defm VGATHERQPD : avx2_gather<0x93, "vgatherqpd", VR256, vx64mem, vy64mem>, VEX_W;
9141 let ExeDomain = SSEPackedSingle in {
9142 defm VGATHERDPS : avx2_gather<0x92, "vgatherdps", VR256, vx32mem, vy32mem>;
9143 defm VGATHERQPS : avx2_gather<0x93, "vgatherqps", VR128, vx32mem, vy32mem>;