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:$src))],
600 IIC_SSE_MOV_S_RM>, Sched<[WriteLoad]>;
603 defm MOVSS : sse12_move<FR32, X86Movss, v4f32, f32mem, "movss">, XS;
604 defm MOVSD : sse12_move<FR64, X86Movsd, v2f64, f64mem, "movsd">, XD;
606 let canFoldAsLoad = 1, isReMaterializable = 1 in {
607 defm MOVSS : sse12_move_rm<FR32, f32mem, loadf32, "movss">, XS;
609 let AddedComplexity = 20 in
610 defm MOVSD : sse12_move_rm<FR64, f64mem, loadf64, "movsd">, XD;
614 let Predicates = [UseAVX] in {
615 let AddedComplexity = 15 in {
616 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
617 // MOVS{S,D} to the lower bits.
618 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
619 (VMOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
620 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
621 (VMOVSSrr (v4f32 (V_SET0)), (COPY_TO_REGCLASS VR128:$src, FR32))>;
622 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
623 (VMOVSSrr (v4i32 (V_SET0)), (COPY_TO_REGCLASS VR128:$src, FR32))>;
624 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
625 (VMOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
627 // Move low f32 and clear high bits.
628 def : Pat<(v8f32 (X86vzmovl (v8f32 VR256:$src))),
629 (SUBREG_TO_REG (i32 0),
630 (VMOVSSrr (v4f32 (V_SET0)),
631 (EXTRACT_SUBREG (v8f32 VR256:$src), sub_xmm)), sub_xmm)>;
632 def : Pat<(v8i32 (X86vzmovl (v8i32 VR256:$src))),
633 (SUBREG_TO_REG (i32 0),
634 (VMOVSSrr (v4i32 (V_SET0)),
635 (EXTRACT_SUBREG (v8i32 VR256:$src), sub_xmm)), sub_xmm)>;
638 let AddedComplexity = 20 in {
639 // MOVSSrm zeros the high parts of the register; represent this
640 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
641 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
642 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
643 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
644 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
645 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
646 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
648 // MOVSDrm zeros the high parts of the register; represent this
649 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
650 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
651 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
652 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
653 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
654 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
655 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
656 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
657 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
658 def : Pat<(v2f64 (X86vzload addr:$src)),
659 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
661 // Represent the same patterns above but in the form they appear for
663 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
664 (v4i32 (scalar_to_vector (loadi32 addr:$src))), (iPTR 0)))),
665 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_xmm)>;
666 def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
667 (v4f32 (scalar_to_vector (loadf32 addr:$src))), (iPTR 0)))),
668 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_xmm)>;
669 def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
670 (v2f64 (scalar_to_vector (loadf64 addr:$src))), (iPTR 0)))),
671 (SUBREG_TO_REG (i32 0), (VMOVSDrm addr:$src), sub_xmm)>;
673 def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
674 (v4f32 (scalar_to_vector FR32:$src)), (iPTR 0)))),
675 (SUBREG_TO_REG (i32 0),
676 (v4f32 (VMOVSSrr (v4f32 (V_SET0)), FR32:$src)),
678 def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
679 (v2f64 (scalar_to_vector FR64:$src)), (iPTR 0)))),
680 (SUBREG_TO_REG (i64 0),
681 (v2f64 (VMOVSDrr (v2f64 (V_SET0)), FR64:$src)),
683 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
684 (v2i64 (scalar_to_vector (loadi64 addr:$src))), (iPTR 0)))),
685 (SUBREG_TO_REG (i64 0), (VMOVSDrm addr:$src), sub_xmm)>;
687 // Move low f64 and clear high bits.
688 def : Pat<(v4f64 (X86vzmovl (v4f64 VR256:$src))),
689 (SUBREG_TO_REG (i32 0),
690 (VMOVSDrr (v2f64 (V_SET0)),
691 (EXTRACT_SUBREG (v4f64 VR256:$src), sub_xmm)), sub_xmm)>;
693 def : Pat<(v4i64 (X86vzmovl (v4i64 VR256:$src))),
694 (SUBREG_TO_REG (i32 0),
695 (VMOVSDrr (v2i64 (V_SET0)),
696 (EXTRACT_SUBREG (v4i64 VR256:$src), sub_xmm)), sub_xmm)>;
698 // Extract and store.
699 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
701 (VMOVSSmr addr:$dst, (COPY_TO_REGCLASS (v4f32 VR128:$src), FR32))>;
702 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
704 (VMOVSDmr addr:$dst, (COPY_TO_REGCLASS (v2f64 VR128:$src), FR64))>;
706 // Shuffle with VMOVSS
707 def : Pat<(v4i32 (X86Movss VR128:$src1, VR128:$src2)),
708 (VMOVSSrr (v4i32 VR128:$src1),
709 (COPY_TO_REGCLASS (v4i32 VR128:$src2), FR32))>;
710 def : Pat<(v4f32 (X86Movss VR128:$src1, VR128:$src2)),
711 (VMOVSSrr (v4f32 VR128:$src1),
712 (COPY_TO_REGCLASS (v4f32 VR128:$src2), FR32))>;
715 def : Pat<(v8i32 (X86Movss VR256:$src1, VR256:$src2)),
716 (SUBREG_TO_REG (i32 0),
717 (VMOVSSrr (EXTRACT_SUBREG (v8i32 VR256:$src1), sub_xmm),
718 (EXTRACT_SUBREG (v8i32 VR256:$src2), sub_xmm)),
720 def : Pat<(v8f32 (X86Movss VR256:$src1, VR256:$src2)),
721 (SUBREG_TO_REG (i32 0),
722 (VMOVSSrr (EXTRACT_SUBREG (v8f32 VR256:$src1), sub_xmm),
723 (EXTRACT_SUBREG (v8f32 VR256:$src2), sub_xmm)),
726 // Shuffle with VMOVSD
727 def : Pat<(v2i64 (X86Movsd VR128:$src1, VR128:$src2)),
728 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
729 def : Pat<(v2f64 (X86Movsd VR128:$src1, VR128:$src2)),
730 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
731 def : Pat<(v4f32 (X86Movsd VR128:$src1, VR128:$src2)),
732 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
733 def : Pat<(v4i32 (X86Movsd VR128:$src1, VR128:$src2)),
734 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
737 def : Pat<(v4i64 (X86Movsd VR256:$src1, VR256:$src2)),
738 (SUBREG_TO_REG (i32 0),
739 (VMOVSDrr (EXTRACT_SUBREG (v4i64 VR256:$src1), sub_xmm),
740 (EXTRACT_SUBREG (v4i64 VR256:$src2), sub_xmm)),
742 def : Pat<(v4f64 (X86Movsd VR256:$src1, VR256:$src2)),
743 (SUBREG_TO_REG (i32 0),
744 (VMOVSDrr (EXTRACT_SUBREG (v4f64 VR256:$src1), sub_xmm),
745 (EXTRACT_SUBREG (v4f64 VR256:$src2), sub_xmm)),
749 // FIXME: Instead of a X86Movlps there should be a X86Movsd here, the problem
750 // is during lowering, where it's not possible to recognize the fold cause
751 // it has two uses through a bitcast. One use disappears at isel time and the
752 // fold opportunity reappears.
753 def : Pat<(v2f64 (X86Movlpd VR128:$src1, VR128:$src2)),
754 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
755 def : Pat<(v2i64 (X86Movlpd VR128:$src1, VR128:$src2)),
756 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
757 def : Pat<(v4f32 (X86Movlps VR128:$src1, VR128:$src2)),
758 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
759 def : Pat<(v4i32 (X86Movlps VR128:$src1, VR128:$src2)),
760 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
763 let Predicates = [UseSSE1] in {
764 let AddedComplexity = 15 in {
765 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
766 // MOVSS to the lower bits.
767 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
768 (MOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
769 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
770 (MOVSSrr (v4f32 (V_SET0)), (COPY_TO_REGCLASS VR128:$src, FR32))>;
771 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
772 (MOVSSrr (v4i32 (V_SET0)), (COPY_TO_REGCLASS VR128:$src, FR32))>;
775 let AddedComplexity = 20 in {
776 // MOVSSrm already zeros the high parts of the register.
777 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
778 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
779 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
780 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
781 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
782 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
785 // Extract and store.
786 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
788 (MOVSSmr addr:$dst, (COPY_TO_REGCLASS VR128:$src, FR32))>;
790 // Shuffle with MOVSS
791 def : Pat<(v4i32 (X86Movss VR128:$src1, VR128:$src2)),
792 (MOVSSrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR32))>;
793 def : Pat<(v4f32 (X86Movss VR128:$src1, VR128:$src2)),
794 (MOVSSrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR32))>;
797 let Predicates = [UseSSE2] in {
798 let AddedComplexity = 15 in {
799 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
800 // MOVSD to the lower bits.
801 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
802 (MOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
805 let AddedComplexity = 20 in {
806 // MOVSDrm already zeros the high parts of the register.
807 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
808 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
809 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
810 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
811 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
812 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
813 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
814 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
815 def : Pat<(v2f64 (X86vzload addr:$src)),
816 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
819 // Extract and store.
820 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
822 (MOVSDmr addr:$dst, (COPY_TO_REGCLASS VR128:$src, FR64))>;
824 // Shuffle with MOVSD
825 def : Pat<(v2i64 (X86Movsd VR128:$src1, VR128:$src2)),
826 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
827 def : Pat<(v2f64 (X86Movsd VR128:$src1, VR128:$src2)),
828 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
829 def : Pat<(v4f32 (X86Movsd VR128:$src1, VR128:$src2)),
830 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
831 def : Pat<(v4i32 (X86Movsd VR128:$src1, VR128:$src2)),
832 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
834 // FIXME: Instead of a X86Movlps there should be a X86Movsd here, the problem
835 // is during lowering, where it's not possible to recognize the fold cause
836 // it has two uses through a bitcast. One use disappears at isel time and the
837 // fold opportunity reappears.
838 def : Pat<(v2f64 (X86Movlpd VR128:$src1, VR128:$src2)),
839 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
840 def : Pat<(v2i64 (X86Movlpd VR128:$src1, VR128:$src2)),
841 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
842 def : Pat<(v4f32 (X86Movlps VR128:$src1, VR128:$src2)),
843 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
844 def : Pat<(v4i32 (X86Movlps VR128:$src1, VR128:$src2)),
845 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
848 //===----------------------------------------------------------------------===//
849 // SSE 1 & 2 - Move Aligned/Unaligned FP Instructions
850 //===----------------------------------------------------------------------===//
852 multiclass sse12_mov_packed<bits<8> opc, RegisterClass RC,
853 X86MemOperand x86memop, PatFrag ld_frag,
854 string asm, Domain d,
856 bit IsReMaterializable = 1> {
857 let neverHasSideEffects = 1 in
858 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
859 !strconcat(asm, "\t{$src, $dst|$dst, $src}"), [], itins.rr, d>,
860 Sched<[WriteFShuffle]>;
861 let canFoldAsLoad = 1, isReMaterializable = IsReMaterializable in
862 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
863 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
864 [(set RC:$dst, (ld_frag addr:$src))], itins.rm, d>,
868 let Predicates = [HasAVX, NoVLX] in {
869 defm VMOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
870 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
872 defm VMOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
873 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
875 defm VMOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
876 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
878 defm VMOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
879 "movupd", SSEPackedDouble, SSE_MOVU_ITINS, 0>,
882 defm VMOVAPSY : sse12_mov_packed<0x28, VR256, f256mem, alignedloadv8f32,
883 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
885 defm VMOVAPDY : sse12_mov_packed<0x28, VR256, f256mem, alignedloadv4f64,
886 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
888 defm VMOVUPSY : sse12_mov_packed<0x10, VR256, f256mem, loadv8f32,
889 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
891 defm VMOVUPDY : sse12_mov_packed<0x10, VR256, f256mem, loadv4f64,
892 "movupd", SSEPackedDouble, SSE_MOVU_ITINS, 0>,
896 let Predicates = [UseSSE1] in {
897 defm MOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
898 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
900 defm MOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
901 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
904 let Predicates = [UseSSE2] in {
905 defm MOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
906 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
908 defm MOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
909 "movupd", SSEPackedDouble, SSE_MOVU_ITINS, 0>,
913 let SchedRW = [WriteStore], Predicates = [HasAVX, NoVLX] in {
914 def VMOVAPSmr : VPSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
915 "movaps\t{$src, $dst|$dst, $src}",
916 [(alignedstore (v4f32 VR128:$src), addr:$dst)],
917 IIC_SSE_MOVA_P_MR>, VEX;
918 def VMOVAPDmr : VPDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
919 "movapd\t{$src, $dst|$dst, $src}",
920 [(alignedstore (v2f64 VR128:$src), addr:$dst)],
921 IIC_SSE_MOVA_P_MR>, VEX;
922 def VMOVUPSmr : VPSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
923 "movups\t{$src, $dst|$dst, $src}",
924 [(store (v4f32 VR128:$src), addr:$dst)],
925 IIC_SSE_MOVU_P_MR>, VEX;
926 def VMOVUPDmr : VPDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
927 "movupd\t{$src, $dst|$dst, $src}",
928 [(store (v2f64 VR128:$src), addr:$dst)],
929 IIC_SSE_MOVU_P_MR>, VEX;
930 def VMOVAPSYmr : VPSI<0x29, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
931 "movaps\t{$src, $dst|$dst, $src}",
932 [(alignedstore256 (v8f32 VR256:$src), addr:$dst)],
933 IIC_SSE_MOVA_P_MR>, VEX, VEX_L;
934 def VMOVAPDYmr : VPDI<0x29, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
935 "movapd\t{$src, $dst|$dst, $src}",
936 [(alignedstore256 (v4f64 VR256:$src), addr:$dst)],
937 IIC_SSE_MOVA_P_MR>, VEX, VEX_L;
938 def VMOVUPSYmr : VPSI<0x11, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
939 "movups\t{$src, $dst|$dst, $src}",
940 [(store (v8f32 VR256:$src), addr:$dst)],
941 IIC_SSE_MOVU_P_MR>, VEX, VEX_L;
942 def VMOVUPDYmr : VPDI<0x11, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
943 "movupd\t{$src, $dst|$dst, $src}",
944 [(store (v4f64 VR256:$src), addr:$dst)],
945 IIC_SSE_MOVU_P_MR>, VEX, VEX_L;
949 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
950 SchedRW = [WriteFShuffle] in {
951 def VMOVAPSrr_REV : VPSI<0x29, MRMDestReg, (outs VR128:$dst),
953 "movaps\t{$src, $dst|$dst, $src}", [],
954 IIC_SSE_MOVA_P_RR>, VEX;
955 def VMOVAPDrr_REV : VPDI<0x29, MRMDestReg, (outs VR128:$dst),
957 "movapd\t{$src, $dst|$dst, $src}", [],
958 IIC_SSE_MOVA_P_RR>, VEX;
959 def VMOVUPSrr_REV : VPSI<0x11, MRMDestReg, (outs VR128:$dst),
961 "movups\t{$src, $dst|$dst, $src}", [],
962 IIC_SSE_MOVU_P_RR>, VEX;
963 def VMOVUPDrr_REV : VPDI<0x11, MRMDestReg, (outs VR128:$dst),
965 "movupd\t{$src, $dst|$dst, $src}", [],
966 IIC_SSE_MOVU_P_RR>, VEX;
967 def VMOVAPSYrr_REV : VPSI<0x29, MRMDestReg, (outs VR256:$dst),
969 "movaps\t{$src, $dst|$dst, $src}", [],
970 IIC_SSE_MOVA_P_RR>, VEX, VEX_L;
971 def VMOVAPDYrr_REV : VPDI<0x29, MRMDestReg, (outs VR256:$dst),
973 "movapd\t{$src, $dst|$dst, $src}", [],
974 IIC_SSE_MOVA_P_RR>, VEX, VEX_L;
975 def VMOVUPSYrr_REV : VPSI<0x11, MRMDestReg, (outs VR256:$dst),
977 "movups\t{$src, $dst|$dst, $src}", [],
978 IIC_SSE_MOVU_P_RR>, VEX, VEX_L;
979 def VMOVUPDYrr_REV : VPDI<0x11, MRMDestReg, (outs VR256:$dst),
981 "movupd\t{$src, $dst|$dst, $src}", [],
982 IIC_SSE_MOVU_P_RR>, VEX, VEX_L;
985 let Predicates = [HasAVX] in {
986 def : Pat<(v8i32 (X86vzmovl
987 (insert_subvector undef, (v4i32 VR128:$src), (iPTR 0)))),
988 (SUBREG_TO_REG (i32 0), (VMOVAPSrr VR128:$src), sub_xmm)>;
989 def : Pat<(v4i64 (X86vzmovl
990 (insert_subvector undef, (v2i64 VR128:$src), (iPTR 0)))),
991 (SUBREG_TO_REG (i32 0), (VMOVAPSrr VR128:$src), sub_xmm)>;
992 def : Pat<(v8f32 (X86vzmovl
993 (insert_subvector undef, (v4f32 VR128:$src), (iPTR 0)))),
994 (SUBREG_TO_REG (i32 0), (VMOVAPSrr VR128:$src), sub_xmm)>;
995 def : Pat<(v4f64 (X86vzmovl
996 (insert_subvector undef, (v2f64 VR128:$src), (iPTR 0)))),
997 (SUBREG_TO_REG (i32 0), (VMOVAPSrr VR128:$src), sub_xmm)>;
1001 def : Pat<(int_x86_avx_storeu_ps_256 addr:$dst, VR256:$src),
1002 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1003 def : Pat<(int_x86_avx_storeu_pd_256 addr:$dst, VR256:$src),
1004 (VMOVUPDYmr addr:$dst, VR256:$src)>;
1006 let SchedRW = [WriteStore] in {
1007 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1008 "movaps\t{$src, $dst|$dst, $src}",
1009 [(alignedstore (v4f32 VR128:$src), addr:$dst)],
1011 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1012 "movapd\t{$src, $dst|$dst, $src}",
1013 [(alignedstore (v2f64 VR128:$src), addr:$dst)],
1015 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1016 "movups\t{$src, $dst|$dst, $src}",
1017 [(store (v4f32 VR128:$src), addr:$dst)],
1019 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1020 "movupd\t{$src, $dst|$dst, $src}",
1021 [(store (v2f64 VR128:$src), addr:$dst)],
1026 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
1027 SchedRW = [WriteFShuffle] in {
1028 def MOVAPSrr_REV : PSI<0x29, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
1029 "movaps\t{$src, $dst|$dst, $src}", [],
1031 def MOVAPDrr_REV : PDI<0x29, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
1032 "movapd\t{$src, $dst|$dst, $src}", [],
1034 def MOVUPSrr_REV : PSI<0x11, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
1035 "movups\t{$src, $dst|$dst, $src}", [],
1037 def MOVUPDrr_REV : PDI<0x11, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
1038 "movupd\t{$src, $dst|$dst, $src}", [],
1042 let Predicates = [HasAVX] in {
1043 def : Pat<(int_x86_sse_storeu_ps addr:$dst, VR128:$src),
1044 (VMOVUPSmr addr:$dst, VR128:$src)>;
1045 def : Pat<(int_x86_sse2_storeu_pd addr:$dst, VR128:$src),
1046 (VMOVUPDmr addr:$dst, VR128:$src)>;
1049 let Predicates = [UseSSE1] in
1050 def : Pat<(int_x86_sse_storeu_ps addr:$dst, VR128:$src),
1051 (MOVUPSmr addr:$dst, VR128:$src)>;
1052 let Predicates = [UseSSE2] in
1053 def : Pat<(int_x86_sse2_storeu_pd addr:$dst, VR128:$src),
1054 (MOVUPDmr addr:$dst, VR128:$src)>;
1056 // Use vmovaps/vmovups for AVX integer load/store.
1057 let Predicates = [HasAVX, NoVLX] in {
1058 // 128-bit load/store
1059 def : Pat<(alignedloadv2i64 addr:$src),
1060 (VMOVAPSrm addr:$src)>;
1061 def : Pat<(loadv2i64 addr:$src),
1062 (VMOVUPSrm addr:$src)>;
1064 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
1065 (VMOVAPSmr addr:$dst, VR128:$src)>;
1066 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
1067 (VMOVAPSmr addr:$dst, VR128:$src)>;
1068 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
1069 (VMOVAPSmr addr:$dst, VR128:$src)>;
1070 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
1071 (VMOVAPSmr addr:$dst, VR128:$src)>;
1072 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
1073 (VMOVUPSmr addr:$dst, VR128:$src)>;
1074 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
1075 (VMOVUPSmr addr:$dst, VR128:$src)>;
1076 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
1077 (VMOVUPSmr addr:$dst, VR128:$src)>;
1078 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
1079 (VMOVUPSmr addr:$dst, VR128:$src)>;
1081 // 256-bit load/store
1082 def : Pat<(alignedloadv4i64 addr:$src),
1083 (VMOVAPSYrm addr:$src)>;
1084 def : Pat<(loadv4i64 addr:$src),
1085 (VMOVUPSYrm addr:$src)>;
1086 def : Pat<(alignedstore256 (v4i64 VR256:$src), addr:$dst),
1087 (VMOVAPSYmr addr:$dst, VR256:$src)>;
1088 def : Pat<(alignedstore256 (v8i32 VR256:$src), addr:$dst),
1089 (VMOVAPSYmr addr:$dst, VR256:$src)>;
1090 def : Pat<(alignedstore256 (v16i16 VR256:$src), addr:$dst),
1091 (VMOVAPSYmr addr:$dst, VR256:$src)>;
1092 def : Pat<(alignedstore256 (v32i8 VR256:$src), addr:$dst),
1093 (VMOVAPSYmr addr:$dst, VR256:$src)>;
1094 def : Pat<(store (v4i64 VR256:$src), addr:$dst),
1095 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1096 def : Pat<(store (v8i32 VR256:$src), addr:$dst),
1097 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1098 def : Pat<(store (v16i16 VR256:$src), addr:$dst),
1099 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1100 def : Pat<(store (v32i8 VR256:$src), addr:$dst),
1101 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1103 // Special patterns for storing subvector extracts of lower 128-bits
1104 // Its cheaper to just use VMOVAPS/VMOVUPS instead of VEXTRACTF128mr
1105 def : Pat<(alignedstore (v2f64 (extract_subvector
1106 (v4f64 VR256:$src), (iPTR 0))), addr:$dst),
1107 (VMOVAPDmr addr:$dst, (v2f64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1108 def : Pat<(alignedstore (v4f32 (extract_subvector
1109 (v8f32 VR256:$src), (iPTR 0))), addr:$dst),
1110 (VMOVAPSmr addr:$dst, (v4f32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1111 def : Pat<(alignedstore (v2i64 (extract_subvector
1112 (v4i64 VR256:$src), (iPTR 0))), addr:$dst),
1113 (VMOVAPDmr addr:$dst, (v2i64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1114 def : Pat<(alignedstore (v4i32 (extract_subvector
1115 (v8i32 VR256:$src), (iPTR 0))), addr:$dst),
1116 (VMOVAPSmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1117 def : Pat<(alignedstore (v8i16 (extract_subvector
1118 (v16i16 VR256:$src), (iPTR 0))), addr:$dst),
1119 (VMOVAPSmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1120 def : Pat<(alignedstore (v16i8 (extract_subvector
1121 (v32i8 VR256:$src), (iPTR 0))), addr:$dst),
1122 (VMOVAPSmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1124 def : Pat<(store (v2f64 (extract_subvector
1125 (v4f64 VR256:$src), (iPTR 0))), addr:$dst),
1126 (VMOVUPDmr addr:$dst, (v2f64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1127 def : Pat<(store (v4f32 (extract_subvector
1128 (v8f32 VR256:$src), (iPTR 0))), addr:$dst),
1129 (VMOVUPSmr addr:$dst, (v4f32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1130 def : Pat<(store (v2i64 (extract_subvector
1131 (v4i64 VR256:$src), (iPTR 0))), addr:$dst),
1132 (VMOVUPDmr addr:$dst, (v2i64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1133 def : Pat<(store (v4i32 (extract_subvector
1134 (v8i32 VR256:$src), (iPTR 0))), addr:$dst),
1135 (VMOVUPSmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1136 def : Pat<(store (v8i16 (extract_subvector
1137 (v16i16 VR256:$src), (iPTR 0))), addr:$dst),
1138 (VMOVUPSmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1139 def : Pat<(store (v16i8 (extract_subvector
1140 (v32i8 VR256:$src), (iPTR 0))), addr:$dst),
1141 (VMOVUPSmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1144 // Use movaps / movups for SSE integer load / store (one byte shorter).
1145 // The instructions selected below are then converted to MOVDQA/MOVDQU
1146 // during the SSE domain pass.
1147 let Predicates = [UseSSE1] in {
1148 def : Pat<(alignedloadv2i64 addr:$src),
1149 (MOVAPSrm addr:$src)>;
1150 def : Pat<(loadv2i64 addr:$src),
1151 (MOVUPSrm addr:$src)>;
1153 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
1154 (MOVAPSmr addr:$dst, VR128:$src)>;
1155 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
1156 (MOVAPSmr addr:$dst, VR128:$src)>;
1157 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
1158 (MOVAPSmr addr:$dst, VR128:$src)>;
1159 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
1160 (MOVAPSmr addr:$dst, VR128:$src)>;
1161 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
1162 (MOVUPSmr addr:$dst, VR128:$src)>;
1163 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
1164 (MOVUPSmr addr:$dst, VR128:$src)>;
1165 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
1166 (MOVUPSmr addr:$dst, VR128:$src)>;
1167 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
1168 (MOVUPSmr addr:$dst, VR128:$src)>;
1171 // Alias instruction to load FR32 or FR64 from f128mem using movaps. Upper
1172 // bits are disregarded. FIXME: Set encoding to pseudo!
1173 let canFoldAsLoad = 1, isReMaterializable = 1, SchedRW = [WriteLoad] in {
1174 let isCodeGenOnly = 1 in {
1175 def FsVMOVAPSrm : VPSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
1176 "movaps\t{$src, $dst|$dst, $src}",
1177 [(set FR32:$dst, (alignedloadfsf32 addr:$src))],
1178 IIC_SSE_MOVA_P_RM>, VEX;
1179 def FsVMOVAPDrm : VPDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1180 "movapd\t{$src, $dst|$dst, $src}",
1181 [(set FR64:$dst, (alignedloadfsf64 addr:$src))],
1182 IIC_SSE_MOVA_P_RM>, VEX;
1183 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
1184 "movaps\t{$src, $dst|$dst, $src}",
1185 [(set FR32:$dst, (alignedloadfsf32 addr:$src))],
1187 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1188 "movapd\t{$src, $dst|$dst, $src}",
1189 [(set FR64:$dst, (alignedloadfsf64 addr:$src))],
1194 //===----------------------------------------------------------------------===//
1195 // SSE 1 & 2 - Move Low packed FP Instructions
1196 //===----------------------------------------------------------------------===//
1198 multiclass sse12_mov_hilo_packed_base<bits<8>opc, SDNode psnode, SDNode pdnode,
1199 string base_opc, string asm_opr,
1200 InstrItinClass itin> {
1201 def PSrm : PI<opc, MRMSrcMem,
1202 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1203 !strconcat(base_opc, "s", asm_opr),
1205 (psnode VR128:$src1,
1206 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))],
1207 itin, SSEPackedSingle>, PS,
1208 Sched<[WriteFShuffleLd, ReadAfterLd]>;
1210 def PDrm : PI<opc, MRMSrcMem,
1211 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1212 !strconcat(base_opc, "d", asm_opr),
1213 [(set VR128:$dst, (v2f64 (pdnode VR128:$src1,
1214 (scalar_to_vector (loadf64 addr:$src2)))))],
1215 itin, SSEPackedDouble>, PD,
1216 Sched<[WriteFShuffleLd, ReadAfterLd]>;
1220 multiclass sse12_mov_hilo_packed<bits<8>opc, SDNode psnode, SDNode pdnode,
1221 string base_opc, InstrItinClass itin> {
1222 defm V#NAME : sse12_mov_hilo_packed_base<opc, psnode, pdnode, base_opc,
1223 "\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1226 let Constraints = "$src1 = $dst" in
1227 defm NAME : sse12_mov_hilo_packed_base<opc, psnode, pdnode, base_opc,
1228 "\t{$src2, $dst|$dst, $src2}",
1232 let AddedComplexity = 20 in {
1233 defm MOVL : sse12_mov_hilo_packed<0x12, X86Movlps, X86Movlpd, "movlp",
1237 let SchedRW = [WriteStore] in {
1238 def VMOVLPSmr : VPSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1239 "movlps\t{$src, $dst|$dst, $src}",
1240 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
1241 (iPTR 0))), addr:$dst)],
1242 IIC_SSE_MOV_LH>, VEX;
1243 def VMOVLPDmr : VPDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1244 "movlpd\t{$src, $dst|$dst, $src}",
1245 [(store (f64 (vector_extract (v2f64 VR128:$src),
1246 (iPTR 0))), addr:$dst)],
1247 IIC_SSE_MOV_LH>, VEX;
1248 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1249 "movlps\t{$src, $dst|$dst, $src}",
1250 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
1251 (iPTR 0))), addr:$dst)],
1253 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1254 "movlpd\t{$src, $dst|$dst, $src}",
1255 [(store (f64 (vector_extract (v2f64 VR128:$src),
1256 (iPTR 0))), addr:$dst)],
1260 let Predicates = [HasAVX] in {
1261 // Shuffle with VMOVLPS
1262 def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))),
1263 (VMOVLPSrm VR128:$src1, addr:$src2)>;
1264 def : Pat<(v4i32 (X86Movlps VR128:$src1, (load addr:$src2))),
1265 (VMOVLPSrm VR128:$src1, addr:$src2)>;
1267 // Shuffle with VMOVLPD
1268 def : Pat<(v2f64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1269 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1270 def : Pat<(v2i64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1271 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1274 def : Pat<(store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)),
1276 (VMOVLPSmr addr:$src1, VR128:$src2)>;
1277 def : Pat<(store (v4i32 (X86Movlps
1278 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)), addr:$src1),
1279 (VMOVLPSmr addr:$src1, VR128:$src2)>;
1280 def : Pat<(store (v2f64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1282 (VMOVLPDmr addr:$src1, VR128:$src2)>;
1283 def : Pat<(store (v2i64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1285 (VMOVLPDmr addr:$src1, VR128:$src2)>;
1288 let Predicates = [UseSSE1] in {
1289 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
1290 def : Pat<(store (i64 (vector_extract (bc_v2i64 (v4f32 VR128:$src2)),
1291 (iPTR 0))), addr:$src1),
1292 (MOVLPSmr addr:$src1, VR128:$src2)>;
1294 // Shuffle with MOVLPS
1295 def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))),
1296 (MOVLPSrm VR128:$src1, addr:$src2)>;
1297 def : Pat<(v4i32 (X86Movlps VR128:$src1, (load addr:$src2))),
1298 (MOVLPSrm VR128:$src1, addr:$src2)>;
1299 def : Pat<(X86Movlps VR128:$src1,
1300 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1301 (MOVLPSrm VR128:$src1, addr:$src2)>;
1304 def : Pat<(store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)),
1306 (MOVLPSmr addr:$src1, VR128:$src2)>;
1307 def : Pat<(store (v4i32 (X86Movlps
1308 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
1310 (MOVLPSmr addr:$src1, VR128:$src2)>;
1313 let Predicates = [UseSSE2] in {
1314 // Shuffle with MOVLPD
1315 def : Pat<(v2f64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1316 (MOVLPDrm VR128:$src1, addr:$src2)>;
1317 def : Pat<(v2i64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1318 (MOVLPDrm VR128:$src1, addr:$src2)>;
1321 def : Pat<(store (v2f64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1323 (MOVLPDmr addr:$src1, VR128:$src2)>;
1324 def : Pat<(store (v2i64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1326 (MOVLPDmr addr:$src1, VR128:$src2)>;
1329 //===----------------------------------------------------------------------===//
1330 // SSE 1 & 2 - Move Hi packed FP Instructions
1331 //===----------------------------------------------------------------------===//
1333 let AddedComplexity = 20 in {
1334 defm MOVH : sse12_mov_hilo_packed<0x16, X86Movlhps, X86Movlhpd, "movhp",
1338 let SchedRW = [WriteStore] in {
1339 // v2f64 extract element 1 is always custom lowered to unpack high to low
1340 // and extract element 0 so the non-store version isn't too horrible.
1341 def VMOVHPSmr : VPSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1342 "movhps\t{$src, $dst|$dst, $src}",
1343 [(store (f64 (vector_extract
1344 (X86Unpckh (bc_v2f64 (v4f32 VR128:$src)),
1345 (bc_v2f64 (v4f32 VR128:$src))),
1346 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>, VEX;
1347 def VMOVHPDmr : VPDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1348 "movhpd\t{$src, $dst|$dst, $src}",
1349 [(store (f64 (vector_extract
1350 (v2f64 (X86Unpckh VR128:$src, VR128:$src)),
1351 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>, VEX;
1352 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1353 "movhps\t{$src, $dst|$dst, $src}",
1354 [(store (f64 (vector_extract
1355 (X86Unpckh (bc_v2f64 (v4f32 VR128:$src)),
1356 (bc_v2f64 (v4f32 VR128:$src))),
1357 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>;
1358 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1359 "movhpd\t{$src, $dst|$dst, $src}",
1360 [(store (f64 (vector_extract
1361 (v2f64 (X86Unpckh VR128:$src, VR128:$src)),
1362 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>;
1365 let Predicates = [HasAVX] in {
1367 def : Pat<(X86Movlhps VR128:$src1,
1368 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1369 (VMOVHPSrm VR128:$src1, addr:$src2)>;
1370 def : Pat<(X86Movlhps VR128:$src1,
1371 (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
1372 (VMOVHPSrm VR128:$src1, addr:$src2)>;
1374 // FIXME: Instead of X86Unpckl, there should be a X86Movlhpd here, the problem
1375 // is during lowering, where it's not possible to recognize the load fold
1376 // cause it has two uses through a bitcast. One use disappears at isel time
1377 // and the fold opportunity reappears.
1378 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1379 (scalar_to_vector (loadf64 addr:$src2)))),
1380 (VMOVHPDrm VR128:$src1, addr:$src2)>;
1383 let Predicates = [UseSSE1] in {
1385 def : Pat<(X86Movlhps VR128:$src1,
1386 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1387 (MOVHPSrm VR128:$src1, addr:$src2)>;
1388 def : Pat<(X86Movlhps VR128:$src1,
1389 (bc_v4f32 (v2i64 (X86vzload addr:$src2)))),
1390 (MOVHPSrm VR128:$src1, addr:$src2)>;
1393 let Predicates = [UseSSE2] in {
1394 // FIXME: Instead of X86Unpckl, there should be a X86Movlhpd here, the problem
1395 // is during lowering, where it's not possible to recognize the load fold
1396 // cause it has two uses through a bitcast. One use disappears at isel time
1397 // and the fold opportunity reappears.
1398 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1399 (scalar_to_vector (loadf64 addr:$src2)))),
1400 (MOVHPDrm VR128:$src1, addr:$src2)>;
1403 //===----------------------------------------------------------------------===//
1404 // SSE 1 & 2 - Move Low to High and High to Low packed FP Instructions
1405 //===----------------------------------------------------------------------===//
1407 let AddedComplexity = 20, Predicates = [UseAVX] in {
1408 def VMOVLHPSrr : VPSI<0x16, MRMSrcReg, (outs VR128:$dst),
1409 (ins VR128:$src1, VR128:$src2),
1410 "movlhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1412 (v4f32 (X86Movlhps VR128:$src1, VR128:$src2)))],
1414 VEX_4V, Sched<[WriteFShuffle]>;
1415 def VMOVHLPSrr : VPSI<0x12, MRMSrcReg, (outs VR128:$dst),
1416 (ins VR128:$src1, VR128:$src2),
1417 "movhlps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1419 (v4f32 (X86Movhlps VR128:$src1, VR128:$src2)))],
1421 VEX_4V, Sched<[WriteFShuffle]>;
1423 let Constraints = "$src1 = $dst", AddedComplexity = 20 in {
1424 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
1425 (ins VR128:$src1, VR128:$src2),
1426 "movlhps\t{$src2, $dst|$dst, $src2}",
1428 (v4f32 (X86Movlhps VR128:$src1, VR128:$src2)))],
1429 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
1430 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
1431 (ins VR128:$src1, VR128:$src2),
1432 "movhlps\t{$src2, $dst|$dst, $src2}",
1434 (v4f32 (X86Movhlps VR128:$src1, VR128:$src2)))],
1435 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
1438 let Predicates = [UseAVX] in {
1440 def : Pat<(v4i32 (X86Movlhps VR128:$src1, VR128:$src2)),
1441 (VMOVLHPSrr VR128:$src1, VR128:$src2)>;
1442 def : Pat<(v2i64 (X86Movlhps VR128:$src1, VR128:$src2)),
1443 (VMOVLHPSrr (v2i64 VR128:$src1), VR128:$src2)>;
1446 def : Pat<(v4i32 (X86Movhlps VR128:$src1, VR128:$src2)),
1447 (VMOVHLPSrr VR128:$src1, VR128:$src2)>;
1450 let Predicates = [UseSSE1] in {
1452 def : Pat<(v4i32 (X86Movlhps VR128:$src1, VR128:$src2)),
1453 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
1454 def : Pat<(v2i64 (X86Movlhps VR128:$src1, VR128:$src2)),
1455 (MOVLHPSrr (v2i64 VR128:$src1), VR128:$src2)>;
1458 def : Pat<(v4i32 (X86Movhlps VR128:$src1, VR128:$src2)),
1459 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
1462 //===----------------------------------------------------------------------===//
1463 // SSE 1 & 2 - Conversion Instructions
1464 //===----------------------------------------------------------------------===//
1466 def SSE_CVT_PD : OpndItins<
1467 IIC_SSE_CVT_PD_RR, IIC_SSE_CVT_PD_RM
1470 let Sched = WriteCvtI2F in
1471 def SSE_CVT_PS : OpndItins<
1472 IIC_SSE_CVT_PS_RR, IIC_SSE_CVT_PS_RM
1475 let Sched = WriteCvtI2F in
1476 def SSE_CVT_Scalar : OpndItins<
1477 IIC_SSE_CVT_Scalar_RR, IIC_SSE_CVT_Scalar_RM
1480 let Sched = WriteCvtF2I in
1481 def SSE_CVT_SS2SI_32 : OpndItins<
1482 IIC_SSE_CVT_SS2SI32_RR, IIC_SSE_CVT_SS2SI32_RM
1485 let Sched = WriteCvtF2I in
1486 def SSE_CVT_SS2SI_64 : OpndItins<
1487 IIC_SSE_CVT_SS2SI64_RR, IIC_SSE_CVT_SS2SI64_RM
1490 let Sched = WriteCvtF2I in
1491 def SSE_CVT_SD2SI : OpndItins<
1492 IIC_SSE_CVT_SD2SI_RR, IIC_SSE_CVT_SD2SI_RM
1495 multiclass sse12_cvt_s<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1496 SDNode OpNode, X86MemOperand x86memop, PatFrag ld_frag,
1497 string asm, OpndItins itins> {
1498 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
1499 [(set DstRC:$dst, (OpNode SrcRC:$src))],
1500 itins.rr>, Sched<[itins.Sched]>;
1501 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
1502 [(set DstRC:$dst, (OpNode (ld_frag addr:$src)))],
1503 itins.rm>, Sched<[itins.Sched.Folded]>;
1506 multiclass sse12_cvt_p<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1507 X86MemOperand x86memop, string asm, Domain d,
1509 let neverHasSideEffects = 1 in {
1510 def rr : I<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
1511 [], itins.rr, d>, Sched<[itins.Sched]>;
1513 def rm : I<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
1514 [], itins.rm, d>, Sched<[itins.Sched.Folded]>;
1518 multiclass sse12_vcvt_avx<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1519 X86MemOperand x86memop, string asm> {
1520 let neverHasSideEffects = 1, Predicates = [UseAVX] in {
1521 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src),
1522 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>,
1523 Sched<[WriteCvtI2F]>;
1525 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
1526 (ins DstRC:$src1, x86memop:$src),
1527 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>,
1528 Sched<[WriteCvtI2FLd, ReadAfterLd]>;
1529 } // neverHasSideEffects = 1
1532 let Predicates = [UseAVX] in {
1533 defm VCVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
1534 "cvttss2si\t{$src, $dst|$dst, $src}",
1537 defm VCVTTSS2SI64 : sse12_cvt_s<0x2C, FR32, GR64, fp_to_sint, f32mem, loadf32,
1538 "cvttss2si\t{$src, $dst|$dst, $src}",
1540 XS, VEX, VEX_W, VEX_LIG;
1541 defm VCVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
1542 "cvttsd2si\t{$src, $dst|$dst, $src}",
1545 defm VCVTTSD2SI64 : sse12_cvt_s<0x2C, FR64, GR64, fp_to_sint, f64mem, loadf64,
1546 "cvttsd2si\t{$src, $dst|$dst, $src}",
1548 XD, VEX, VEX_W, VEX_LIG;
1550 def : InstAlias<"vcvttss2si{l}\t{$src, $dst|$dst, $src}",
1551 (VCVTTSS2SIrr GR32:$dst, FR32:$src), 0>;
1552 def : InstAlias<"vcvttss2si{l}\t{$src, $dst|$dst, $src}",
1553 (VCVTTSS2SIrm GR32:$dst, f32mem:$src), 0>;
1554 def : InstAlias<"vcvttsd2si{l}\t{$src, $dst|$dst, $src}",
1555 (VCVTTSD2SIrr GR32:$dst, FR64:$src), 0>;
1556 def : InstAlias<"vcvttsd2si{l}\t{$src, $dst|$dst, $src}",
1557 (VCVTTSD2SIrm GR32:$dst, f64mem:$src), 0>;
1558 def : InstAlias<"vcvttss2si{q}\t{$src, $dst|$dst, $src}",
1559 (VCVTTSS2SI64rr GR64:$dst, FR32:$src), 0>;
1560 def : InstAlias<"vcvttss2si{q}\t{$src, $dst|$dst, $src}",
1561 (VCVTTSS2SI64rm GR64:$dst, f32mem:$src), 0>;
1562 def : InstAlias<"vcvttsd2si{q}\t{$src, $dst|$dst, $src}",
1563 (VCVTTSD2SI64rr GR64:$dst, FR64:$src), 0>;
1564 def : InstAlias<"vcvttsd2si{q}\t{$src, $dst|$dst, $src}",
1565 (VCVTTSD2SI64rm GR64:$dst, f64mem:$src), 0>;
1567 // The assembler can recognize rr 64-bit instructions by seeing a rxx
1568 // register, but the same isn't true when only using memory operands,
1569 // provide other assembly "l" and "q" forms to address this explicitly
1570 // where appropriate to do so.
1571 defm VCVTSI2SS : sse12_vcvt_avx<0x2A, GR32, FR32, i32mem, "cvtsi2ss{l}">,
1572 XS, VEX_4V, VEX_LIG;
1573 defm VCVTSI2SS64 : sse12_vcvt_avx<0x2A, GR64, FR32, i64mem, "cvtsi2ss{q}">,
1574 XS, VEX_4V, VEX_W, VEX_LIG;
1575 defm VCVTSI2SD : sse12_vcvt_avx<0x2A, GR32, FR64, i32mem, "cvtsi2sd{l}">,
1576 XD, VEX_4V, VEX_LIG;
1577 defm VCVTSI2SD64 : sse12_vcvt_avx<0x2A, GR64, FR64, i64mem, "cvtsi2sd{q}">,
1578 XD, VEX_4V, VEX_W, VEX_LIG;
1580 let Predicates = [UseAVX] in {
1581 def : InstAlias<"vcvtsi2ss\t{$src, $src1, $dst|$dst, $src1, $src}",
1582 (VCVTSI2SSrm FR64:$dst, FR64:$src1, i32mem:$src), 0>;
1583 def : InstAlias<"vcvtsi2sd\t{$src, $src1, $dst|$dst, $src1, $src}",
1584 (VCVTSI2SDrm FR64:$dst, FR64:$src1, i32mem:$src), 0>;
1586 def : Pat<(f32 (sint_to_fp (loadi32 addr:$src))),
1587 (VCVTSI2SSrm (f32 (IMPLICIT_DEF)), addr:$src)>;
1588 def : Pat<(f32 (sint_to_fp (loadi64 addr:$src))),
1589 (VCVTSI2SS64rm (f32 (IMPLICIT_DEF)), addr:$src)>;
1590 def : Pat<(f64 (sint_to_fp (loadi32 addr:$src))),
1591 (VCVTSI2SDrm (f64 (IMPLICIT_DEF)), addr:$src)>;
1592 def : Pat<(f64 (sint_to_fp (loadi64 addr:$src))),
1593 (VCVTSI2SD64rm (f64 (IMPLICIT_DEF)), addr:$src)>;
1595 def : Pat<(f32 (sint_to_fp GR32:$src)),
1596 (VCVTSI2SSrr (f32 (IMPLICIT_DEF)), GR32:$src)>;
1597 def : Pat<(f32 (sint_to_fp GR64:$src)),
1598 (VCVTSI2SS64rr (f32 (IMPLICIT_DEF)), GR64:$src)>;
1599 def : Pat<(f64 (sint_to_fp GR32:$src)),
1600 (VCVTSI2SDrr (f64 (IMPLICIT_DEF)), GR32:$src)>;
1601 def : Pat<(f64 (sint_to_fp GR64:$src)),
1602 (VCVTSI2SD64rr (f64 (IMPLICIT_DEF)), GR64:$src)>;
1605 defm CVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
1606 "cvttss2si\t{$src, $dst|$dst, $src}",
1607 SSE_CVT_SS2SI_32>, XS;
1608 defm CVTTSS2SI64 : sse12_cvt_s<0x2C, FR32, GR64, fp_to_sint, f32mem, loadf32,
1609 "cvttss2si\t{$src, $dst|$dst, $src}",
1610 SSE_CVT_SS2SI_64>, XS, REX_W;
1611 defm CVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
1612 "cvttsd2si\t{$src, $dst|$dst, $src}",
1614 defm CVTTSD2SI64 : sse12_cvt_s<0x2C, FR64, GR64, fp_to_sint, f64mem, loadf64,
1615 "cvttsd2si\t{$src, $dst|$dst, $src}",
1616 SSE_CVT_SD2SI>, XD, REX_W;
1617 defm CVTSI2SS : sse12_cvt_s<0x2A, GR32, FR32, sint_to_fp, i32mem, loadi32,
1618 "cvtsi2ss{l}\t{$src, $dst|$dst, $src}",
1619 SSE_CVT_Scalar>, XS;
1620 defm CVTSI2SS64 : sse12_cvt_s<0x2A, GR64, FR32, sint_to_fp, i64mem, loadi64,
1621 "cvtsi2ss{q}\t{$src, $dst|$dst, $src}",
1622 SSE_CVT_Scalar>, XS, REX_W;
1623 defm CVTSI2SD : sse12_cvt_s<0x2A, GR32, FR64, sint_to_fp, i32mem, loadi32,
1624 "cvtsi2sd{l}\t{$src, $dst|$dst, $src}",
1625 SSE_CVT_Scalar>, XD;
1626 defm CVTSI2SD64 : sse12_cvt_s<0x2A, GR64, FR64, sint_to_fp, i64mem, loadi64,
1627 "cvtsi2sd{q}\t{$src, $dst|$dst, $src}",
1628 SSE_CVT_Scalar>, XD, REX_W;
1630 def : InstAlias<"cvttss2si{l}\t{$src, $dst|$dst, $src}",
1631 (CVTTSS2SIrr GR32:$dst, FR32:$src), 0>;
1632 def : InstAlias<"cvttss2si{l}\t{$src, $dst|$dst, $src}",
1633 (CVTTSS2SIrm GR32:$dst, f32mem:$src), 0>;
1634 def : InstAlias<"cvttsd2si{l}\t{$src, $dst|$dst, $src}",
1635 (CVTTSD2SIrr GR32:$dst, FR64:$src), 0>;
1636 def : InstAlias<"cvttsd2si{l}\t{$src, $dst|$dst, $src}",
1637 (CVTTSD2SIrm GR32:$dst, f64mem:$src), 0>;
1638 def : InstAlias<"cvttss2si{q}\t{$src, $dst|$dst, $src}",
1639 (CVTTSS2SI64rr GR64:$dst, FR32:$src), 0>;
1640 def : InstAlias<"cvttss2si{q}\t{$src, $dst|$dst, $src}",
1641 (CVTTSS2SI64rm GR64:$dst, f32mem:$src), 0>;
1642 def : InstAlias<"cvttsd2si{q}\t{$src, $dst|$dst, $src}",
1643 (CVTTSD2SI64rr GR64:$dst, FR64:$src), 0>;
1644 def : InstAlias<"cvttsd2si{q}\t{$src, $dst|$dst, $src}",
1645 (CVTTSD2SI64rm GR64:$dst, f64mem:$src), 0>;
1647 def : InstAlias<"cvtsi2ss\t{$src, $dst|$dst, $src}",
1648 (CVTSI2SSrm FR64:$dst, i32mem:$src), 0>;
1649 def : InstAlias<"cvtsi2sd\t{$src, $dst|$dst, $src}",
1650 (CVTSI2SDrm FR64:$dst, i32mem:$src), 0>;
1652 // Conversion Instructions Intrinsics - Match intrinsics which expect MM
1653 // and/or XMM operand(s).
1655 multiclass sse12_cvt_sint<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1656 Intrinsic Int, Operand memop, ComplexPattern mem_cpat,
1657 string asm, OpndItins itins> {
1658 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src),
1659 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1660 [(set DstRC:$dst, (Int SrcRC:$src))], itins.rr>,
1661 Sched<[itins.Sched]>;
1662 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins memop:$src),
1663 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1664 [(set DstRC:$dst, (Int mem_cpat:$src))], itins.rm>,
1665 Sched<[itins.Sched.Folded]>;
1668 multiclass sse12_cvt_sint_3addr<bits<8> opc, RegisterClass SrcRC,
1669 RegisterClass DstRC, Intrinsic Int, X86MemOperand x86memop,
1670 PatFrag ld_frag, string asm, OpndItins itins,
1672 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src2),
1674 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
1675 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
1676 [(set DstRC:$dst, (Int DstRC:$src1, SrcRC:$src2))],
1677 itins.rr>, Sched<[itins.Sched]>;
1678 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
1679 (ins DstRC:$src1, x86memop:$src2),
1681 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
1682 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
1683 [(set DstRC:$dst, (Int DstRC:$src1, (ld_frag addr:$src2)))],
1684 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
1687 let Predicates = [UseAVX] in {
1688 defm VCVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32,
1689 int_x86_sse2_cvtsd2si, sdmem, sse_load_f64, "cvtsd2si",
1690 SSE_CVT_SD2SI>, XD, VEX, VEX_LIG;
1691 defm VCVTSD2SI64 : sse12_cvt_sint<0x2D, VR128, GR64,
1692 int_x86_sse2_cvtsd2si64, sdmem, sse_load_f64, "cvtsd2si",
1693 SSE_CVT_SD2SI>, XD, VEX, VEX_W, VEX_LIG;
1695 defm CVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse2_cvtsd2si,
1696 sdmem, sse_load_f64, "cvtsd2si", SSE_CVT_SD2SI>, XD;
1697 defm CVTSD2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse2_cvtsd2si64,
1698 sdmem, sse_load_f64, "cvtsd2si", SSE_CVT_SD2SI>, XD, REX_W;
1701 let isCodeGenOnly = 1 in {
1702 let Predicates = [UseAVX] in {
1703 defm Int_VCVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1704 int_x86_sse_cvtsi2ss, i32mem, loadi32, "cvtsi2ss{l}",
1705 SSE_CVT_Scalar, 0>, XS, VEX_4V;
1706 defm Int_VCVTSI2SS64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1707 int_x86_sse_cvtsi642ss, i64mem, loadi64, "cvtsi2ss{q}",
1708 SSE_CVT_Scalar, 0>, XS, VEX_4V,
1710 defm Int_VCVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1711 int_x86_sse2_cvtsi2sd, i32mem, loadi32, "cvtsi2sd{l}",
1712 SSE_CVT_Scalar, 0>, XD, VEX_4V;
1713 defm Int_VCVTSI2SD64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1714 int_x86_sse2_cvtsi642sd, i64mem, loadi64, "cvtsi2sd{q}",
1715 SSE_CVT_Scalar, 0>, XD,
1718 let Constraints = "$src1 = $dst" in {
1719 defm Int_CVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1720 int_x86_sse_cvtsi2ss, i32mem, loadi32,
1721 "cvtsi2ss{l}", SSE_CVT_Scalar>, XS;
1722 defm Int_CVTSI2SS64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1723 int_x86_sse_cvtsi642ss, i64mem, loadi64,
1724 "cvtsi2ss{q}", SSE_CVT_Scalar>, XS, REX_W;
1725 defm Int_CVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1726 int_x86_sse2_cvtsi2sd, i32mem, loadi32,
1727 "cvtsi2sd{l}", SSE_CVT_Scalar>, XD;
1728 defm Int_CVTSI2SD64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1729 int_x86_sse2_cvtsi642sd, i64mem, loadi64,
1730 "cvtsi2sd{q}", SSE_CVT_Scalar>, XD, REX_W;
1732 } // isCodeGenOnly = 1
1736 // Aliases for intrinsics
1737 let isCodeGenOnly = 1 in {
1738 let Predicates = [UseAVX] in {
1739 defm Int_VCVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
1740 ssmem, sse_load_f32, "cvttss2si",
1741 SSE_CVT_SS2SI_32>, XS, VEX;
1742 defm Int_VCVTTSS2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1743 int_x86_sse_cvttss2si64, ssmem, sse_load_f32,
1744 "cvttss2si", SSE_CVT_SS2SI_64>,
1746 defm Int_VCVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
1747 sdmem, sse_load_f64, "cvttsd2si",
1748 SSE_CVT_SD2SI>, XD, VEX;
1749 defm Int_VCVTTSD2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1750 int_x86_sse2_cvttsd2si64, sdmem, sse_load_f64,
1751 "cvttsd2si", SSE_CVT_SD2SI>,
1754 defm Int_CVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
1755 ssmem, sse_load_f32, "cvttss2si",
1756 SSE_CVT_SS2SI_32>, XS;
1757 defm Int_CVTTSS2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1758 int_x86_sse_cvttss2si64, ssmem, sse_load_f32,
1759 "cvttss2si", SSE_CVT_SS2SI_64>, XS, REX_W;
1760 defm Int_CVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
1761 sdmem, sse_load_f64, "cvttsd2si",
1763 defm Int_CVTTSD2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1764 int_x86_sse2_cvttsd2si64, sdmem, sse_load_f64,
1765 "cvttsd2si", SSE_CVT_SD2SI>, XD, REX_W;
1766 } // isCodeGenOnly = 1
1768 let Predicates = [UseAVX] in {
1769 defm VCVTSS2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse_cvtss2si,
1770 ssmem, sse_load_f32, "cvtss2si",
1771 SSE_CVT_SS2SI_32>, XS, VEX, VEX_LIG;
1772 defm VCVTSS2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse_cvtss2si64,
1773 ssmem, sse_load_f32, "cvtss2si",
1774 SSE_CVT_SS2SI_64>, XS, VEX, VEX_W, VEX_LIG;
1776 defm CVTSS2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse_cvtss2si,
1777 ssmem, sse_load_f32, "cvtss2si",
1778 SSE_CVT_SS2SI_32>, XS;
1779 defm CVTSS2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse_cvtss2si64,
1780 ssmem, sse_load_f32, "cvtss2si",
1781 SSE_CVT_SS2SI_64>, XS, REX_W;
1783 defm VCVTDQ2PS : sse12_cvt_p<0x5B, VR128, VR128, i128mem,
1784 "vcvtdq2ps\t{$src, $dst|$dst, $src}",
1785 SSEPackedSingle, SSE_CVT_PS>,
1786 PS, VEX, Requires<[HasAVX]>;
1787 defm VCVTDQ2PSY : sse12_cvt_p<0x5B, VR256, VR256, i256mem,
1788 "vcvtdq2ps\t{$src, $dst|$dst, $src}",
1789 SSEPackedSingle, SSE_CVT_PS>,
1790 PS, VEX, VEX_L, Requires<[HasAVX]>;
1792 defm CVTDQ2PS : sse12_cvt_p<0x5B, VR128, VR128, i128mem,
1793 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1794 SSEPackedSingle, SSE_CVT_PS>,
1795 PS, Requires<[UseSSE2]>;
1797 let Predicates = [UseAVX] in {
1798 def : InstAlias<"vcvtss2si{l}\t{$src, $dst|$dst, $src}",
1799 (VCVTSS2SIrr GR32:$dst, VR128:$src), 0>;
1800 def : InstAlias<"vcvtss2si{l}\t{$src, $dst|$dst, $src}",
1801 (VCVTSS2SIrm GR32:$dst, ssmem:$src), 0>;
1802 def : InstAlias<"vcvtsd2si{l}\t{$src, $dst|$dst, $src}",
1803 (VCVTSD2SIrr GR32:$dst, VR128:$src), 0>;
1804 def : InstAlias<"vcvtsd2si{l}\t{$src, $dst|$dst, $src}",
1805 (VCVTSD2SIrm GR32:$dst, sdmem:$src), 0>;
1806 def : InstAlias<"vcvtss2si{q}\t{$src, $dst|$dst, $src}",
1807 (VCVTSS2SI64rr GR64:$dst, VR128:$src), 0>;
1808 def : InstAlias<"vcvtss2si{q}\t{$src, $dst|$dst, $src}",
1809 (VCVTSS2SI64rm GR64:$dst, ssmem:$src), 0>;
1810 def : InstAlias<"vcvtsd2si{q}\t{$src, $dst|$dst, $src}",
1811 (VCVTSD2SI64rr GR64:$dst, VR128:$src), 0>;
1812 def : InstAlias<"vcvtsd2si{q}\t{$src, $dst|$dst, $src}",
1813 (VCVTSD2SI64rm GR64:$dst, sdmem:$src), 0>;
1816 def : InstAlias<"cvtss2si{l}\t{$src, $dst|$dst, $src}",
1817 (CVTSS2SIrr GR32:$dst, VR128:$src), 0>;
1818 def : InstAlias<"cvtss2si{l}\t{$src, $dst|$dst, $src}",
1819 (CVTSS2SIrm GR32:$dst, ssmem:$src), 0>;
1820 def : InstAlias<"cvtsd2si{l}\t{$src, $dst|$dst, $src}",
1821 (CVTSD2SIrr GR32:$dst, VR128:$src), 0>;
1822 def : InstAlias<"cvtsd2si{l}\t{$src, $dst|$dst, $src}",
1823 (CVTSD2SIrm GR32:$dst, sdmem:$src), 0>;
1824 def : InstAlias<"cvtss2si{q}\t{$src, $dst|$dst, $src}",
1825 (CVTSS2SI64rr GR64:$dst, VR128:$src), 0>;
1826 def : InstAlias<"cvtss2si{q}\t{$src, $dst|$dst, $src}",
1827 (CVTSS2SI64rm GR64:$dst, ssmem:$src), 0>;
1828 def : InstAlias<"cvtsd2si{q}\t{$src, $dst|$dst, $src}",
1829 (CVTSD2SI64rr GR64:$dst, VR128:$src), 0>;
1830 def : InstAlias<"cvtsd2si{q}\t{$src, $dst|$dst, $src}",
1831 (CVTSD2SI64rm GR64:$dst, sdmem:$src)>;
1835 // Convert scalar double to scalar single
1836 let neverHasSideEffects = 1, Predicates = [UseAVX] in {
1837 def VCVTSD2SSrr : VSDI<0x5A, MRMSrcReg, (outs FR32:$dst),
1838 (ins FR64:$src1, FR64:$src2),
1839 "cvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}", [],
1840 IIC_SSE_CVT_Scalar_RR>, VEX_4V, VEX_LIG,
1841 Sched<[WriteCvtF2F]>;
1843 def VCVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst),
1844 (ins FR64:$src1, f64mem:$src2),
1845 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1846 [], IIC_SSE_CVT_Scalar_RM>,
1847 XD, Requires<[HasAVX, OptForSize]>, VEX_4V, VEX_LIG,
1848 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1851 def : Pat<(f32 (fround FR64:$src)), (VCVTSD2SSrr FR64:$src, FR64:$src)>,
1854 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1855 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1856 [(set FR32:$dst, (fround FR64:$src))],
1857 IIC_SSE_CVT_Scalar_RR>, Sched<[WriteCvtF2F]>;
1858 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1859 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1860 [(set FR32:$dst, (fround (loadf64 addr:$src)))],
1861 IIC_SSE_CVT_Scalar_RM>,
1863 Requires<[UseSSE2, OptForSize]>, Sched<[WriteCvtF2FLd]>;
1865 let isCodeGenOnly = 1 in {
1866 def Int_VCVTSD2SSrr: I<0x5A, MRMSrcReg,
1867 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1868 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1870 (int_x86_sse2_cvtsd2ss VR128:$src1, VR128:$src2))],
1871 IIC_SSE_CVT_Scalar_RR>, XD, VEX_4V, Requires<[UseAVX]>,
1872 Sched<[WriteCvtF2F]>;
1873 def Int_VCVTSD2SSrm: I<0x5A, MRMSrcReg,
1874 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1875 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1876 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss
1877 VR128:$src1, sse_load_f64:$src2))],
1878 IIC_SSE_CVT_Scalar_RM>, XD, VEX_4V, Requires<[UseAVX]>,
1879 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1881 let Constraints = "$src1 = $dst" in {
1882 def Int_CVTSD2SSrr: I<0x5A, MRMSrcReg,
1883 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1884 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1886 (int_x86_sse2_cvtsd2ss VR128:$src1, VR128:$src2))],
1887 IIC_SSE_CVT_Scalar_RR>, XD, Requires<[UseSSE2]>,
1888 Sched<[WriteCvtF2F]>;
1889 def Int_CVTSD2SSrm: I<0x5A, MRMSrcReg,
1890 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1891 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1892 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss
1893 VR128:$src1, sse_load_f64:$src2))],
1894 IIC_SSE_CVT_Scalar_RM>, XD, Requires<[UseSSE2]>,
1895 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1897 } // isCodeGenOnly = 1
1899 // Convert scalar single to scalar double
1900 // SSE2 instructions with XS prefix
1901 let neverHasSideEffects = 1, Predicates = [UseAVX] in {
1902 def VCVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst),
1903 (ins FR32:$src1, FR32:$src2),
1904 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1905 [], IIC_SSE_CVT_Scalar_RR>,
1906 XS, Requires<[HasAVX]>, VEX_4V, VEX_LIG,
1907 Sched<[WriteCvtF2F]>;
1909 def VCVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst),
1910 (ins FR32:$src1, f32mem:$src2),
1911 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1912 [], IIC_SSE_CVT_Scalar_RM>,
1913 XS, VEX_4V, VEX_LIG, Requires<[HasAVX, OptForSize]>,
1914 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1917 def : Pat<(f64 (fextend FR32:$src)),
1918 (VCVTSS2SDrr FR32:$src, FR32:$src)>, Requires<[UseAVX]>;
1919 def : Pat<(fextend (loadf32 addr:$src)),
1920 (VCVTSS2SDrm (f32 (IMPLICIT_DEF)), addr:$src)>, Requires<[UseAVX]>;
1922 def : Pat<(extloadf32 addr:$src),
1923 (VCVTSS2SDrm (f32 (IMPLICIT_DEF)), addr:$src)>,
1924 Requires<[UseAVX, OptForSize]>;
1925 def : Pat<(extloadf32 addr:$src),
1926 (VCVTSS2SDrr (f32 (IMPLICIT_DEF)), (VMOVSSrm addr:$src))>,
1927 Requires<[UseAVX, OptForSpeed]>;
1929 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1930 "cvtss2sd\t{$src, $dst|$dst, $src}",
1931 [(set FR64:$dst, (fextend FR32:$src))],
1932 IIC_SSE_CVT_Scalar_RR>, XS,
1933 Requires<[UseSSE2]>, Sched<[WriteCvtF2F]>;
1934 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1935 "cvtss2sd\t{$src, $dst|$dst, $src}",
1936 [(set FR64:$dst, (extloadf32 addr:$src))],
1937 IIC_SSE_CVT_Scalar_RM>, XS,
1938 Requires<[UseSSE2, OptForSize]>, Sched<[WriteCvtF2FLd]>;
1940 // extload f32 -> f64. This matches load+fextend because we have a hack in
1941 // the isel (PreprocessForFPConvert) that can introduce loads after dag
1943 // Since these loads aren't folded into the fextend, we have to match it
1945 def : Pat<(fextend (loadf32 addr:$src)),
1946 (CVTSS2SDrm addr:$src)>, Requires<[UseSSE2]>;
1947 def : Pat<(extloadf32 addr:$src),
1948 (CVTSS2SDrr (MOVSSrm addr:$src))>, Requires<[UseSSE2, OptForSpeed]>;
1950 let isCodeGenOnly = 1 in {
1951 def Int_VCVTSS2SDrr: I<0x5A, MRMSrcReg,
1952 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1953 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1955 (int_x86_sse2_cvtss2sd VR128:$src1, VR128:$src2))],
1956 IIC_SSE_CVT_Scalar_RR>, XS, VEX_4V, Requires<[UseAVX]>,
1957 Sched<[WriteCvtF2F]>;
1958 def Int_VCVTSS2SDrm: I<0x5A, MRMSrcMem,
1959 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
1960 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1962 (int_x86_sse2_cvtss2sd VR128:$src1, sse_load_f32:$src2))],
1963 IIC_SSE_CVT_Scalar_RM>, XS, VEX_4V, Requires<[UseAVX]>,
1964 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1965 let Constraints = "$src1 = $dst" in { // SSE2 instructions with XS prefix
1966 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1967 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1968 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1970 (int_x86_sse2_cvtss2sd VR128:$src1, VR128:$src2))],
1971 IIC_SSE_CVT_Scalar_RR>, XS, Requires<[UseSSE2]>,
1972 Sched<[WriteCvtF2F]>;
1973 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1974 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
1975 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1977 (int_x86_sse2_cvtss2sd VR128:$src1, sse_load_f32:$src2))],
1978 IIC_SSE_CVT_Scalar_RM>, XS, Requires<[UseSSE2]>,
1979 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1981 } // isCodeGenOnly = 1
1983 // Convert packed single/double fp to doubleword
1984 def VCVTPS2DQrr : VPDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1985 "cvtps2dq\t{$src, $dst|$dst, $src}",
1986 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))],
1987 IIC_SSE_CVT_PS_RR>, VEX, Sched<[WriteCvtF2I]>;
1988 def VCVTPS2DQrm : VPDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1989 "cvtps2dq\t{$src, $dst|$dst, $src}",
1991 (int_x86_sse2_cvtps2dq (loadv4f32 addr:$src)))],
1992 IIC_SSE_CVT_PS_RM>, VEX, Sched<[WriteCvtF2ILd]>;
1993 def VCVTPS2DQYrr : VPDI<0x5B, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
1994 "cvtps2dq\t{$src, $dst|$dst, $src}",
1996 (int_x86_avx_cvt_ps2dq_256 VR256:$src))],
1997 IIC_SSE_CVT_PS_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>;
1998 def VCVTPS2DQYrm : VPDI<0x5B, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
1999 "cvtps2dq\t{$src, $dst|$dst, $src}",
2001 (int_x86_avx_cvt_ps2dq_256 (loadv8f32 addr:$src)))],
2002 IIC_SSE_CVT_PS_RM>, VEX, VEX_L, Sched<[WriteCvtF2ILd]>;
2003 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2004 "cvtps2dq\t{$src, $dst|$dst, $src}",
2005 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))],
2006 IIC_SSE_CVT_PS_RR>, Sched<[WriteCvtF2I]>;
2007 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2008 "cvtps2dq\t{$src, $dst|$dst, $src}",
2010 (int_x86_sse2_cvtps2dq (memopv4f32 addr:$src)))],
2011 IIC_SSE_CVT_PS_RM>, Sched<[WriteCvtF2ILd]>;
2014 // Convert Packed Double FP to Packed DW Integers
2015 let Predicates = [HasAVX] in {
2016 // The assembler can recognize rr 256-bit instructions by seeing a ymm
2017 // register, but the same isn't true when using memory operands instead.
2018 // Provide other assembly rr and rm forms to address this explicitly.
2019 def VCVTPD2DQrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2020 "vcvtpd2dq\t{$src, $dst|$dst, $src}",
2021 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
2022 VEX, Sched<[WriteCvtF2I]>;
2025 def : InstAlias<"vcvtpd2dqx\t{$src, $dst|$dst, $src}",
2026 (VCVTPD2DQrr VR128:$dst, VR128:$src), 0>;
2027 def VCVTPD2DQXrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2028 "vcvtpd2dqx\t{$src, $dst|$dst, $src}",
2030 (int_x86_sse2_cvtpd2dq (loadv2f64 addr:$src)))]>, VEX,
2031 Sched<[WriteCvtF2ILd]>;
2034 def VCVTPD2DQYrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2035 "vcvtpd2dq{y}\t{$src, $dst|$dst, $src}",
2037 (int_x86_avx_cvt_pd2dq_256 VR256:$src))]>, VEX, VEX_L,
2038 Sched<[WriteCvtF2I]>;
2039 def VCVTPD2DQYrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2040 "vcvtpd2dq{y}\t{$src, $dst|$dst, $src}",
2042 (int_x86_avx_cvt_pd2dq_256 (loadv4f64 addr:$src)))]>,
2043 VEX, VEX_L, Sched<[WriteCvtF2ILd]>;
2044 def : InstAlias<"vcvtpd2dq\t{$src, $dst|$dst, $src}",
2045 (VCVTPD2DQYrr VR128:$dst, VR256:$src), 0>;
2048 def CVTPD2DQrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2049 "cvtpd2dq\t{$src, $dst|$dst, $src}",
2051 (int_x86_sse2_cvtpd2dq (memopv2f64 addr:$src)))],
2052 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtF2ILd]>;
2053 def CVTPD2DQrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2054 "cvtpd2dq\t{$src, $dst|$dst, $src}",
2055 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))],
2056 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2I]>;
2058 // Convert with truncation packed single/double fp to doubleword
2059 // SSE2 packed instructions with XS prefix
2060 def VCVTTPS2DQrr : VS2SI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2061 "cvttps2dq\t{$src, $dst|$dst, $src}",
2063 (int_x86_sse2_cvttps2dq VR128:$src))],
2064 IIC_SSE_CVT_PS_RR>, VEX, Sched<[WriteCvtF2I]>;
2065 def VCVTTPS2DQrm : VS2SI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2066 "cvttps2dq\t{$src, $dst|$dst, $src}",
2067 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
2068 (loadv4f32 addr:$src)))],
2069 IIC_SSE_CVT_PS_RM>, VEX, Sched<[WriteCvtF2ILd]>;
2070 def VCVTTPS2DQYrr : VS2SI<0x5B, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
2071 "cvttps2dq\t{$src, $dst|$dst, $src}",
2073 (int_x86_avx_cvtt_ps2dq_256 VR256:$src))],
2074 IIC_SSE_CVT_PS_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>;
2075 def VCVTTPS2DQYrm : VS2SI<0x5B, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
2076 "cvttps2dq\t{$src, $dst|$dst, $src}",
2077 [(set VR256:$dst, (int_x86_avx_cvtt_ps2dq_256
2078 (loadv8f32 addr:$src)))],
2079 IIC_SSE_CVT_PS_RM>, VEX, VEX_L,
2080 Sched<[WriteCvtF2ILd]>;
2082 def CVTTPS2DQrr : S2SI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2083 "cvttps2dq\t{$src, $dst|$dst, $src}",
2084 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))],
2085 IIC_SSE_CVT_PS_RR>, Sched<[WriteCvtF2I]>;
2086 def CVTTPS2DQrm : S2SI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2087 "cvttps2dq\t{$src, $dst|$dst, $src}",
2089 (int_x86_sse2_cvttps2dq (memopv4f32 addr:$src)))],
2090 IIC_SSE_CVT_PS_RM>, Sched<[WriteCvtF2ILd]>;
2092 let Predicates = [HasAVX] in {
2093 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
2094 (VCVTDQ2PSrr VR128:$src)>;
2095 def : Pat<(v4f32 (sint_to_fp (bc_v4i32 (loadv2i64 addr:$src)))),
2096 (VCVTDQ2PSrm addr:$src)>;
2098 def : Pat<(int_x86_sse2_cvtdq2ps VR128:$src),
2099 (VCVTDQ2PSrr VR128:$src)>;
2100 def : Pat<(int_x86_sse2_cvtdq2ps (bc_v4i32 (loadv2i64 addr:$src))),
2101 (VCVTDQ2PSrm addr:$src)>;
2103 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
2104 (VCVTTPS2DQrr VR128:$src)>;
2105 def : Pat<(v4i32 (fp_to_sint (loadv4f32 addr:$src))),
2106 (VCVTTPS2DQrm addr:$src)>;
2108 def : Pat<(v8f32 (sint_to_fp (v8i32 VR256:$src))),
2109 (VCVTDQ2PSYrr VR256:$src)>;
2110 def : Pat<(v8f32 (sint_to_fp (bc_v8i32 (loadv4i64 addr:$src)))),
2111 (VCVTDQ2PSYrm addr:$src)>;
2113 def : Pat<(v8i32 (fp_to_sint (v8f32 VR256:$src))),
2114 (VCVTTPS2DQYrr VR256:$src)>;
2115 def : Pat<(v8i32 (fp_to_sint (loadv8f32 addr:$src))),
2116 (VCVTTPS2DQYrm addr:$src)>;
2119 let Predicates = [UseSSE2] in {
2120 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
2121 (CVTDQ2PSrr VR128:$src)>;
2122 def : Pat<(v4f32 (sint_to_fp (bc_v4i32 (memopv2i64 addr:$src)))),
2123 (CVTDQ2PSrm addr:$src)>;
2125 def : Pat<(int_x86_sse2_cvtdq2ps VR128:$src),
2126 (CVTDQ2PSrr VR128:$src)>;
2127 def : Pat<(int_x86_sse2_cvtdq2ps (bc_v4i32 (memopv2i64 addr:$src))),
2128 (CVTDQ2PSrm addr:$src)>;
2130 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
2131 (CVTTPS2DQrr VR128:$src)>;
2132 def : Pat<(v4i32 (fp_to_sint (memopv4f32 addr:$src))),
2133 (CVTTPS2DQrm addr:$src)>;
2136 def VCVTTPD2DQrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2137 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2139 (int_x86_sse2_cvttpd2dq VR128:$src))],
2140 IIC_SSE_CVT_PD_RR>, VEX, Sched<[WriteCvtF2I]>;
2142 // The assembler can recognize rr 256-bit instructions by seeing a ymm
2143 // register, but the same isn't true when using memory operands instead.
2144 // Provide other assembly rr and rm forms to address this explicitly.
2147 def : InstAlias<"vcvttpd2dqx\t{$src, $dst|$dst, $src}",
2148 (VCVTTPD2DQrr VR128:$dst, VR128:$src), 0>;
2149 def VCVTTPD2DQXrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2150 "cvttpd2dqx\t{$src, $dst|$dst, $src}",
2151 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
2152 (loadv2f64 addr:$src)))],
2153 IIC_SSE_CVT_PD_RM>, VEX, Sched<[WriteCvtF2ILd]>;
2156 def VCVTTPD2DQYrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2157 "cvttpd2dq{y}\t{$src, $dst|$dst, $src}",
2159 (int_x86_avx_cvtt_pd2dq_256 VR256:$src))],
2160 IIC_SSE_CVT_PD_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>;
2161 def VCVTTPD2DQYrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2162 "cvttpd2dq{y}\t{$src, $dst|$dst, $src}",
2164 (int_x86_avx_cvtt_pd2dq_256 (loadv4f64 addr:$src)))],
2165 IIC_SSE_CVT_PD_RM>, VEX, VEX_L, Sched<[WriteCvtF2ILd]>;
2166 def : InstAlias<"vcvttpd2dq\t{$src, $dst|$dst, $src}",
2167 (VCVTTPD2DQYrr VR128:$dst, VR256:$src), 0>;
2169 let Predicates = [HasAVX] in {
2170 def : Pat<(v4i32 (fp_to_sint (v4f64 VR256:$src))),
2171 (VCVTTPD2DQYrr VR256:$src)>;
2172 def : Pat<(v4i32 (fp_to_sint (loadv4f64 addr:$src))),
2173 (VCVTTPD2DQYrm addr:$src)>;
2174 } // Predicates = [HasAVX]
2176 def CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2177 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2178 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))],
2179 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2I]>;
2180 def CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
2181 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2182 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
2183 (memopv2f64 addr:$src)))],
2185 Sched<[WriteCvtF2ILd]>;
2187 // Convert packed single to packed double
2188 let Predicates = [HasAVX] in {
2189 // SSE2 instructions without OpSize prefix
2190 def VCVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2191 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2192 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))],
2193 IIC_SSE_CVT_PD_RR>, PS, VEX, Sched<[WriteCvtF2F]>;
2194 def VCVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2195 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2196 [(set VR128:$dst, (v2f64 (extloadv2f32 addr:$src)))],
2197 IIC_SSE_CVT_PD_RM>, PS, VEX, Sched<[WriteCvtF2FLd]>;
2198 def VCVTPS2PDYrr : I<0x5A, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
2199 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2201 (int_x86_avx_cvt_ps2_pd_256 VR128:$src))],
2202 IIC_SSE_CVT_PD_RR>, PS, VEX, VEX_L, Sched<[WriteCvtF2F]>;
2203 def VCVTPS2PDYrm : I<0x5A, MRMSrcMem, (outs VR256:$dst), (ins f128mem:$src),
2204 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2206 (int_x86_avx_cvt_ps2_pd_256 (loadv4f32 addr:$src)))],
2207 IIC_SSE_CVT_PD_RM>, PS, VEX, VEX_L, Sched<[WriteCvtF2FLd]>;
2210 let Predicates = [UseSSE2] in {
2211 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2212 "cvtps2pd\t{$src, $dst|$dst, $src}",
2213 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))],
2214 IIC_SSE_CVT_PD_RR>, PS, Sched<[WriteCvtF2F]>;
2215 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2216 "cvtps2pd\t{$src, $dst|$dst, $src}",
2217 [(set VR128:$dst, (v2f64 (extloadv2f32 addr:$src)))],
2218 IIC_SSE_CVT_PD_RM>, PS, Sched<[WriteCvtF2FLd]>;
2221 // Convert Packed DW Integers to Packed Double FP
2222 let Predicates = [HasAVX] in {
2223 let neverHasSideEffects = 1, mayLoad = 1 in
2224 def VCVTDQ2PDrm : S2SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2225 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2226 []>, VEX, Sched<[WriteCvtI2FLd]>;
2227 def VCVTDQ2PDrr : S2SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2228 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2230 (int_x86_sse2_cvtdq2pd VR128:$src))]>, VEX,
2231 Sched<[WriteCvtI2F]>;
2232 def VCVTDQ2PDYrm : S2SI<0xE6, MRMSrcMem, (outs VR256:$dst), (ins i128mem:$src),
2233 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2235 (int_x86_avx_cvtdq2_pd_256
2236 (bitconvert (loadv2i64 addr:$src))))]>, VEX, VEX_L,
2237 Sched<[WriteCvtI2FLd]>;
2238 def VCVTDQ2PDYrr : S2SI<0xE6, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
2239 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2241 (int_x86_avx_cvtdq2_pd_256 VR128:$src))]>, VEX, VEX_L,
2242 Sched<[WriteCvtI2F]>;
2245 let neverHasSideEffects = 1, mayLoad = 1 in
2246 def CVTDQ2PDrm : S2SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2247 "cvtdq2pd\t{$src, $dst|$dst, $src}", [],
2248 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtI2FLd]>;
2249 def CVTDQ2PDrr : S2SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2250 "cvtdq2pd\t{$src, $dst|$dst, $src}",
2251 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))],
2252 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtI2F]>;
2254 // AVX 256-bit register conversion intrinsics
2255 let Predicates = [HasAVX] in {
2256 def : Pat<(v4f64 (sint_to_fp (v4i32 VR128:$src))),
2257 (VCVTDQ2PDYrr VR128:$src)>;
2258 def : Pat<(v4f64 (sint_to_fp (bc_v4i32 (loadv2i64 addr:$src)))),
2259 (VCVTDQ2PDYrm addr:$src)>;
2260 } // Predicates = [HasAVX]
2262 // Convert packed double to packed single
2263 // The assembler can recognize rr 256-bit instructions by seeing a ymm
2264 // register, but the same isn't true when using memory operands instead.
2265 // Provide other assembly rr and rm forms to address this explicitly.
2266 def VCVTPD2PSrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2267 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2268 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))],
2269 IIC_SSE_CVT_PD_RR>, VEX, Sched<[WriteCvtF2F]>;
2272 def : InstAlias<"vcvtpd2psx\t{$src, $dst|$dst, $src}",
2273 (VCVTPD2PSrr VR128:$dst, VR128:$src), 0>;
2274 def VCVTPD2PSXrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2275 "cvtpd2psx\t{$src, $dst|$dst, $src}",
2277 (int_x86_sse2_cvtpd2ps (loadv2f64 addr:$src)))],
2278 IIC_SSE_CVT_PD_RM>, VEX, Sched<[WriteCvtF2FLd]>;
2281 def VCVTPD2PSYrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2282 "cvtpd2ps{y}\t{$src, $dst|$dst, $src}",
2284 (int_x86_avx_cvt_pd2_ps_256 VR256:$src))],
2285 IIC_SSE_CVT_PD_RR>, VEX, VEX_L, Sched<[WriteCvtF2F]>;
2286 def VCVTPD2PSYrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2287 "cvtpd2ps{y}\t{$src, $dst|$dst, $src}",
2289 (int_x86_avx_cvt_pd2_ps_256 (loadv4f64 addr:$src)))],
2290 IIC_SSE_CVT_PD_RM>, VEX, VEX_L, Sched<[WriteCvtF2FLd]>;
2291 def : InstAlias<"vcvtpd2ps\t{$src, $dst|$dst, $src}",
2292 (VCVTPD2PSYrr VR128:$dst, VR256:$src), 0>;
2294 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2295 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2296 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))],
2297 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2F]>;
2298 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2299 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2301 (int_x86_sse2_cvtpd2ps (memopv2f64 addr:$src)))],
2302 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtF2FLd]>;
2305 // AVX 256-bit register conversion intrinsics
2306 // FIXME: Migrate SSE conversion intrinsics matching to use patterns as below
2307 // whenever possible to avoid declaring two versions of each one.
2308 let Predicates = [HasAVX] in {
2309 def : Pat<(int_x86_avx_cvtdq2_ps_256 VR256:$src),
2310 (VCVTDQ2PSYrr VR256:$src)>;
2311 def : Pat<(int_x86_avx_cvtdq2_ps_256 (bitconvert (loadv4i64 addr:$src))),
2312 (VCVTDQ2PSYrm addr:$src)>;
2314 // Match fround and fextend for 128/256-bit conversions
2315 def : Pat<(v4f32 (X86vfpround (v2f64 VR128:$src))),
2316 (VCVTPD2PSrr VR128:$src)>;
2317 def : Pat<(v4f32 (X86vfpround (loadv2f64 addr:$src))),
2318 (VCVTPD2PSXrm addr:$src)>;
2319 def : Pat<(v4f32 (fround (v4f64 VR256:$src))),
2320 (VCVTPD2PSYrr VR256:$src)>;
2321 def : Pat<(v4f32 (fround (loadv4f64 addr:$src))),
2322 (VCVTPD2PSYrm addr:$src)>;
2324 def : Pat<(v2f64 (X86vfpext (v4f32 VR128:$src))),
2325 (VCVTPS2PDrr VR128:$src)>;
2326 def : Pat<(v4f64 (fextend (v4f32 VR128:$src))),
2327 (VCVTPS2PDYrr VR128:$src)>;
2328 def : Pat<(v4f64 (extloadv4f32 addr:$src)),
2329 (VCVTPS2PDYrm addr:$src)>;
2332 let Predicates = [UseSSE2] in {
2333 // Match fround and fextend for 128 conversions
2334 def : Pat<(v4f32 (X86vfpround (v2f64 VR128:$src))),
2335 (CVTPD2PSrr VR128:$src)>;
2336 def : Pat<(v4f32 (X86vfpround (memopv2f64 addr:$src))),
2337 (CVTPD2PSrm addr:$src)>;
2339 def : Pat<(v2f64 (X86vfpext (v4f32 VR128:$src))),
2340 (CVTPS2PDrr VR128:$src)>;
2343 //===----------------------------------------------------------------------===//
2344 // SSE 1 & 2 - Compare Instructions
2345 //===----------------------------------------------------------------------===//
2347 // sse12_cmp_scalar - sse 1 & 2 compare scalar instructions
2348 multiclass sse12_cmp_scalar<RegisterClass RC, X86MemOperand x86memop,
2349 Operand CC, SDNode OpNode, ValueType VT,
2350 PatFrag ld_frag, string asm, string asm_alt,
2352 def rr : SIi8<0xC2, MRMSrcReg,
2353 (outs RC:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
2354 [(set RC:$dst, (OpNode (VT RC:$src1), RC:$src2, imm:$cc))],
2355 itins.rr>, Sched<[itins.Sched]>;
2356 def rm : SIi8<0xC2, MRMSrcMem,
2357 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
2358 [(set RC:$dst, (OpNode (VT RC:$src1),
2359 (ld_frag addr:$src2), imm:$cc))],
2361 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2363 // Accept explicit immediate argument form instead of comparison code.
2364 let isAsmParserOnly = 1, hasSideEffects = 0 in {
2365 def rr_alt : SIi8<0xC2, MRMSrcReg, (outs RC:$dst),
2366 (ins RC:$src1, RC:$src2, i8imm:$cc), asm_alt, [],
2367 IIC_SSE_ALU_F32S_RR>, Sched<[itins.Sched]>;
2369 def rm_alt : SIi8<0xC2, MRMSrcMem, (outs RC:$dst),
2370 (ins RC:$src1, x86memop:$src2, i8imm:$cc), asm_alt, [],
2371 IIC_SSE_ALU_F32S_RM>,
2372 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2376 defm VCMPSS : sse12_cmp_scalar<FR32, f32mem, AVXCC, X86cmps, f32, loadf32,
2377 "cmp${cc}ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2378 "cmpss\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2380 XS, VEX_4V, VEX_LIG;
2381 defm VCMPSD : sse12_cmp_scalar<FR64, f64mem, AVXCC, X86cmps, f64, loadf64,
2382 "cmp${cc}sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2383 "cmpsd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2384 SSE_ALU_F32S>, // same latency as 32 bit compare
2385 XD, VEX_4V, VEX_LIG;
2387 let Constraints = "$src1 = $dst" in {
2388 defm CMPSS : sse12_cmp_scalar<FR32, f32mem, SSECC, X86cmps, f32, loadf32,
2389 "cmp${cc}ss\t{$src2, $dst|$dst, $src2}",
2390 "cmpss\t{$cc, $src2, $dst|$dst, $src2, $cc}", SSE_ALU_F32S>,
2392 defm CMPSD : sse12_cmp_scalar<FR64, f64mem, SSECC, X86cmps, f64, loadf64,
2393 "cmp${cc}sd\t{$src2, $dst|$dst, $src2}",
2394 "cmpsd\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2399 multiclass sse12_cmp_scalar_int<X86MemOperand x86memop, Operand CC,
2400 Intrinsic Int, string asm, OpndItins itins> {
2401 def rr : SIi8<0xC2, MRMSrcReg, (outs VR128:$dst),
2402 (ins VR128:$src1, VR128:$src, CC:$cc), asm,
2403 [(set VR128:$dst, (Int VR128:$src1,
2404 VR128:$src, imm:$cc))],
2406 Sched<[itins.Sched]>;
2407 def rm : SIi8<0xC2, MRMSrcMem, (outs VR128:$dst),
2408 (ins VR128:$src1, x86memop:$src, CC:$cc), asm,
2409 [(set VR128:$dst, (Int VR128:$src1,
2410 (load addr:$src), imm:$cc))],
2412 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2415 let isCodeGenOnly = 1 in {
2416 // Aliases to match intrinsics which expect XMM operand(s).
2417 defm Int_VCMPSS : sse12_cmp_scalar_int<f32mem, AVXCC, int_x86_sse_cmp_ss,
2418 "cmp${cc}ss\t{$src, $src1, $dst|$dst, $src1, $src}",
2421 defm Int_VCMPSD : sse12_cmp_scalar_int<f64mem, AVXCC, int_x86_sse2_cmp_sd,
2422 "cmp${cc}sd\t{$src, $src1, $dst|$dst, $src1, $src}",
2423 SSE_ALU_F32S>, // same latency as f32
2425 let Constraints = "$src1 = $dst" in {
2426 defm Int_CMPSS : sse12_cmp_scalar_int<f32mem, SSECC, int_x86_sse_cmp_ss,
2427 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
2429 defm Int_CMPSD : sse12_cmp_scalar_int<f64mem, SSECC, int_x86_sse2_cmp_sd,
2430 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
2437 // sse12_ord_cmp - Unordered/Ordered scalar fp compare and set EFLAGS
2438 multiclass sse12_ord_cmp<bits<8> opc, RegisterClass RC, SDNode OpNode,
2439 ValueType vt, X86MemOperand x86memop,
2440 PatFrag ld_frag, string OpcodeStr> {
2441 def rr: SI<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
2442 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2443 [(set EFLAGS, (OpNode (vt RC:$src1), RC:$src2))],
2446 def rm: SI<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
2447 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2448 [(set EFLAGS, (OpNode (vt RC:$src1),
2449 (ld_frag addr:$src2)))],
2451 Sched<[WriteFAddLd, ReadAfterLd]>;
2454 let Defs = [EFLAGS] in {
2455 defm VUCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
2456 "ucomiss">, PS, VEX, VEX_LIG;
2457 defm VUCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
2458 "ucomisd">, PD, VEX, VEX_LIG;
2459 let Pattern = []<dag> in {
2460 defm VCOMISS : sse12_ord_cmp<0x2F, VR128, undef, v4f32, f128mem, load,
2461 "comiss">, PS, VEX, VEX_LIG;
2462 defm VCOMISD : sse12_ord_cmp<0x2F, VR128, undef, v2f64, f128mem, load,
2463 "comisd">, PD, VEX, VEX_LIG;
2466 let isCodeGenOnly = 1 in {
2467 defm Int_VUCOMISS : sse12_ord_cmp<0x2E, VR128, X86ucomi, v4f32, f128mem,
2468 load, "ucomiss">, PS, VEX;
2469 defm Int_VUCOMISD : sse12_ord_cmp<0x2E, VR128, X86ucomi, v2f64, f128mem,
2470 load, "ucomisd">, PD, VEX;
2472 defm Int_VCOMISS : sse12_ord_cmp<0x2F, VR128, X86comi, v4f32, f128mem,
2473 load, "comiss">, PS, VEX;
2474 defm Int_VCOMISD : sse12_ord_cmp<0x2F, VR128, X86comi, v2f64, f128mem,
2475 load, "comisd">, PD, VEX;
2477 defm UCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
2479 defm UCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
2482 let Pattern = []<dag> in {
2483 defm COMISS : sse12_ord_cmp<0x2F, VR128, undef, v4f32, f128mem, load,
2485 defm COMISD : sse12_ord_cmp<0x2F, VR128, undef, v2f64, f128mem, load,
2489 let isCodeGenOnly = 1 in {
2490 defm Int_UCOMISS : sse12_ord_cmp<0x2E, VR128, X86ucomi, v4f32, f128mem,
2491 load, "ucomiss">, PS;
2492 defm Int_UCOMISD : sse12_ord_cmp<0x2E, VR128, X86ucomi, v2f64, f128mem,
2493 load, "ucomisd">, PD;
2495 defm Int_COMISS : sse12_ord_cmp<0x2F, VR128, X86comi, v4f32, f128mem, load,
2497 defm Int_COMISD : sse12_ord_cmp<0x2F, VR128, X86comi, v2f64, f128mem, load,
2500 } // Defs = [EFLAGS]
2502 // sse12_cmp_packed - sse 1 & 2 compare packed instructions
2503 multiclass sse12_cmp_packed<RegisterClass RC, X86MemOperand x86memop,
2504 Operand CC, Intrinsic Int, string asm,
2505 string asm_alt, Domain d,
2506 OpndItins itins = SSE_ALU_F32P> {
2507 def rri : PIi8<0xC2, MRMSrcReg,
2508 (outs RC:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
2509 [(set RC:$dst, (Int RC:$src1, RC:$src2, imm:$cc))],
2512 def rmi : PIi8<0xC2, MRMSrcMem,
2513 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
2514 [(set RC:$dst, (Int RC:$src1, (memop addr:$src2), imm:$cc))],
2516 Sched<[WriteFAddLd, ReadAfterLd]>;
2518 // Accept explicit immediate argument form instead of comparison code.
2519 let isAsmParserOnly = 1, hasSideEffects = 0 in {
2520 def rri_alt : PIi8<0xC2, MRMSrcReg,
2521 (outs RC:$dst), (ins RC:$src1, RC:$src2, i8imm:$cc),
2522 asm_alt, [], itins.rr, d>, Sched<[WriteFAdd]>;
2523 def rmi_alt : PIi8<0xC2, MRMSrcMem,
2524 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, i8imm:$cc),
2525 asm_alt, [], itins.rm, d>,
2526 Sched<[WriteFAddLd, ReadAfterLd]>;
2530 defm VCMPPS : sse12_cmp_packed<VR128, f128mem, AVXCC, int_x86_sse_cmp_ps,
2531 "cmp${cc}ps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2532 "cmpps\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2533 SSEPackedSingle>, PS, VEX_4V;
2534 defm VCMPPD : sse12_cmp_packed<VR128, f128mem, AVXCC, int_x86_sse2_cmp_pd,
2535 "cmp${cc}pd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2536 "cmppd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2537 SSEPackedDouble>, PD, VEX_4V;
2538 defm VCMPPSY : sse12_cmp_packed<VR256, f256mem, AVXCC, int_x86_avx_cmp_ps_256,
2539 "cmp${cc}ps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2540 "cmpps\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2541 SSEPackedSingle>, PS, VEX_4V, VEX_L;
2542 defm VCMPPDY : sse12_cmp_packed<VR256, f256mem, AVXCC, int_x86_avx_cmp_pd_256,
2543 "cmp${cc}pd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2544 "cmppd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2545 SSEPackedDouble>, PD, VEX_4V, VEX_L;
2546 let Constraints = "$src1 = $dst" in {
2547 defm CMPPS : sse12_cmp_packed<VR128, f128mem, SSECC, int_x86_sse_cmp_ps,
2548 "cmp${cc}ps\t{$src2, $dst|$dst, $src2}",
2549 "cmpps\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2550 SSEPackedSingle, SSE_ALU_F32P>, PS;
2551 defm CMPPD : sse12_cmp_packed<VR128, f128mem, SSECC, int_x86_sse2_cmp_pd,
2552 "cmp${cc}pd\t{$src2, $dst|$dst, $src2}",
2553 "cmppd\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2554 SSEPackedDouble, SSE_ALU_F64P>, PD;
2557 let Predicates = [HasAVX] in {
2558 def : Pat<(v4i32 (X86cmpp (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
2559 (VCMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
2560 def : Pat<(v4i32 (X86cmpp (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
2561 (VCMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
2562 def : Pat<(v2i64 (X86cmpp (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
2563 (VCMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
2564 def : Pat<(v2i64 (X86cmpp (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
2565 (VCMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
2567 def : Pat<(v8i32 (X86cmpp (v8f32 VR256:$src1), VR256:$src2, imm:$cc)),
2568 (VCMPPSYrri (v8f32 VR256:$src1), (v8f32 VR256:$src2), imm:$cc)>;
2569 def : Pat<(v8i32 (X86cmpp (v8f32 VR256:$src1), (memop addr:$src2), imm:$cc)),
2570 (VCMPPSYrmi (v8f32 VR256:$src1), addr:$src2, imm:$cc)>;
2571 def : Pat<(v4i64 (X86cmpp (v4f64 VR256:$src1), VR256:$src2, imm:$cc)),
2572 (VCMPPDYrri VR256:$src1, VR256:$src2, imm:$cc)>;
2573 def : Pat<(v4i64 (X86cmpp (v4f64 VR256:$src1), (memop addr:$src2), imm:$cc)),
2574 (VCMPPDYrmi VR256:$src1, addr:$src2, imm:$cc)>;
2577 let Predicates = [UseSSE1] in {
2578 def : Pat<(v4i32 (X86cmpp (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
2579 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
2580 def : Pat<(v4i32 (X86cmpp (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
2581 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
2584 let Predicates = [UseSSE2] in {
2585 def : Pat<(v2i64 (X86cmpp (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
2586 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
2587 def : Pat<(v2i64 (X86cmpp (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
2588 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
2591 //===----------------------------------------------------------------------===//
2592 // SSE 1 & 2 - Shuffle Instructions
2593 //===----------------------------------------------------------------------===//
2595 /// sse12_shuffle - sse 1 & 2 fp shuffle instructions
2596 multiclass sse12_shuffle<RegisterClass RC, X86MemOperand x86memop,
2597 ValueType vt, string asm, PatFrag mem_frag,
2598 Domain d, bit IsConvertibleToThreeAddress = 0> {
2599 def rmi : PIi8<0xC6, MRMSrcMem, (outs RC:$dst),
2600 (ins RC:$src1, x86memop:$src2, i8imm:$src3), asm,
2601 [(set RC:$dst, (vt (X86Shufp RC:$src1, (mem_frag addr:$src2),
2602 (i8 imm:$src3))))], IIC_SSE_SHUFP, d>,
2603 Sched<[WriteFShuffleLd, ReadAfterLd]>;
2604 let isConvertibleToThreeAddress = IsConvertibleToThreeAddress in
2605 def rri : PIi8<0xC6, MRMSrcReg, (outs RC:$dst),
2606 (ins RC:$src1, RC:$src2, i8imm:$src3), asm,
2607 [(set RC:$dst, (vt (X86Shufp RC:$src1, RC:$src2,
2608 (i8 imm:$src3))))], IIC_SSE_SHUFP, d>,
2609 Sched<[WriteFShuffle]>;
2612 defm VSHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
2613 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2614 loadv4f32, SSEPackedSingle>, PS, VEX_4V;
2615 defm VSHUFPSY : sse12_shuffle<VR256, f256mem, v8f32,
2616 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2617 loadv8f32, SSEPackedSingle>, PS, VEX_4V, VEX_L;
2618 defm VSHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
2619 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2620 loadv2f64, SSEPackedDouble>, PD, VEX_4V;
2621 defm VSHUFPDY : sse12_shuffle<VR256, f256mem, v4f64,
2622 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2623 loadv4f64, SSEPackedDouble>, PD, VEX_4V, VEX_L;
2625 let Constraints = "$src1 = $dst" in {
2626 defm SHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
2627 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2628 memopv4f32, SSEPackedSingle, 1 /* cvt to pshufd */>, PS;
2629 defm SHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
2630 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2631 memopv2f64, SSEPackedDouble, 1 /* cvt to pshufd */>, PD;
2634 let Predicates = [HasAVX] in {
2635 def : Pat<(v4i32 (X86Shufp VR128:$src1,
2636 (bc_v4i32 (loadv2i64 addr:$src2)), (i8 imm:$imm))),
2637 (VSHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>;
2638 def : Pat<(v4i32 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2639 (VSHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>;
2641 def : Pat<(v2i64 (X86Shufp VR128:$src1,
2642 (loadv2i64 addr:$src2), (i8 imm:$imm))),
2643 (VSHUFPDrmi VR128:$src1, addr:$src2, imm:$imm)>;
2644 def : Pat<(v2i64 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2645 (VSHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>;
2648 def : Pat<(v8i32 (X86Shufp VR256:$src1, VR256:$src2, (i8 imm:$imm))),
2649 (VSHUFPSYrri VR256:$src1, VR256:$src2, imm:$imm)>;
2650 def : Pat<(v8i32 (X86Shufp VR256:$src1,
2651 (bc_v8i32 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
2652 (VSHUFPSYrmi VR256:$src1, addr:$src2, imm:$imm)>;
2654 def : Pat<(v4i64 (X86Shufp VR256:$src1, VR256:$src2, (i8 imm:$imm))),
2655 (VSHUFPDYrri VR256:$src1, VR256:$src2, imm:$imm)>;
2656 def : Pat<(v4i64 (X86Shufp VR256:$src1,
2657 (loadv4i64 addr:$src2), (i8 imm:$imm))),
2658 (VSHUFPDYrmi VR256:$src1, addr:$src2, imm:$imm)>;
2661 let Predicates = [UseSSE1] in {
2662 def : Pat<(v4i32 (X86Shufp VR128:$src1,
2663 (bc_v4i32 (memopv2i64 addr:$src2)), (i8 imm:$imm))),
2664 (SHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>;
2665 def : Pat<(v4i32 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2666 (SHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>;
2669 let Predicates = [UseSSE2] in {
2670 // Generic SHUFPD patterns
2671 def : Pat<(v2i64 (X86Shufp VR128:$src1,
2672 (memopv2i64 addr:$src2), (i8 imm:$imm))),
2673 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$imm)>;
2674 def : Pat<(v2i64 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2675 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>;
2678 //===----------------------------------------------------------------------===//
2679 // SSE 1 & 2 - Unpack FP Instructions
2680 //===----------------------------------------------------------------------===//
2682 /// sse12_unpack_interleave - sse 1 & 2 fp unpack and interleave
2683 multiclass sse12_unpack_interleave<bits<8> opc, SDNode OpNode, ValueType vt,
2684 PatFrag mem_frag, RegisterClass RC,
2685 X86MemOperand x86memop, string asm,
2687 def rr : PI<opc, MRMSrcReg,
2688 (outs RC:$dst), (ins RC:$src1, RC:$src2),
2690 (vt (OpNode RC:$src1, RC:$src2)))],
2691 IIC_SSE_UNPCK, d>, Sched<[WriteFShuffle]>;
2692 def rm : PI<opc, MRMSrcMem,
2693 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
2695 (vt (OpNode RC:$src1,
2696 (mem_frag addr:$src2))))],
2698 Sched<[WriteFShuffleLd, ReadAfterLd]>;
2701 defm VUNPCKHPS: sse12_unpack_interleave<0x15, X86Unpckh, v4f32, loadv4f32,
2702 VR128, f128mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2703 SSEPackedSingle>, PS, VEX_4V;
2704 defm VUNPCKHPD: sse12_unpack_interleave<0x15, X86Unpckh, v2f64, loadv2f64,
2705 VR128, f128mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2706 SSEPackedDouble>, PD, VEX_4V;
2707 defm VUNPCKLPS: sse12_unpack_interleave<0x14, X86Unpckl, v4f32, loadv4f32,
2708 VR128, f128mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2709 SSEPackedSingle>, PS, VEX_4V;
2710 defm VUNPCKLPD: sse12_unpack_interleave<0x14, X86Unpckl, v2f64, loadv2f64,
2711 VR128, f128mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2712 SSEPackedDouble>, PD, VEX_4V;
2714 defm VUNPCKHPSY: sse12_unpack_interleave<0x15, X86Unpckh, v8f32, loadv8f32,
2715 VR256, f256mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2716 SSEPackedSingle>, PS, VEX_4V, VEX_L;
2717 defm VUNPCKHPDY: sse12_unpack_interleave<0x15, X86Unpckh, v4f64, loadv4f64,
2718 VR256, f256mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2719 SSEPackedDouble>, PD, VEX_4V, VEX_L;
2720 defm VUNPCKLPSY: sse12_unpack_interleave<0x14, X86Unpckl, v8f32, loadv8f32,
2721 VR256, f256mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2722 SSEPackedSingle>, PS, VEX_4V, VEX_L;
2723 defm VUNPCKLPDY: sse12_unpack_interleave<0x14, X86Unpckl, v4f64, loadv4f64,
2724 VR256, f256mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2725 SSEPackedDouble>, PD, VEX_4V, VEX_L;
2727 let Constraints = "$src1 = $dst" in {
2728 defm UNPCKHPS: sse12_unpack_interleave<0x15, X86Unpckh, v4f32, memopv4f32,
2729 VR128, f128mem, "unpckhps\t{$src2, $dst|$dst, $src2}",
2730 SSEPackedSingle>, PS;
2731 defm UNPCKHPD: sse12_unpack_interleave<0x15, X86Unpckh, v2f64, memopv2f64,
2732 VR128, f128mem, "unpckhpd\t{$src2, $dst|$dst, $src2}",
2733 SSEPackedDouble>, PD;
2734 defm UNPCKLPS: sse12_unpack_interleave<0x14, X86Unpckl, v4f32, memopv4f32,
2735 VR128, f128mem, "unpcklps\t{$src2, $dst|$dst, $src2}",
2736 SSEPackedSingle>, PS;
2737 defm UNPCKLPD: sse12_unpack_interleave<0x14, X86Unpckl, v2f64, memopv2f64,
2738 VR128, f128mem, "unpcklpd\t{$src2, $dst|$dst, $src2}",
2739 SSEPackedDouble>, PD;
2740 } // Constraints = "$src1 = $dst"
2742 let Predicates = [HasAVX1Only] in {
2743 def : Pat<(v8i32 (X86Unpckl VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
2744 (VUNPCKLPSYrm VR256:$src1, addr:$src2)>;
2745 def : Pat<(v8i32 (X86Unpckl VR256:$src1, VR256:$src2)),
2746 (VUNPCKLPSYrr VR256:$src1, VR256:$src2)>;
2747 def : Pat<(v8i32 (X86Unpckh VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
2748 (VUNPCKHPSYrm VR256:$src1, addr:$src2)>;
2749 def : Pat<(v8i32 (X86Unpckh VR256:$src1, VR256:$src2)),
2750 (VUNPCKHPSYrr VR256:$src1, VR256:$src2)>;
2752 def : Pat<(v4i64 (X86Unpckl VR256:$src1, (loadv4i64 addr:$src2))),
2753 (VUNPCKLPDYrm VR256:$src1, addr:$src2)>;
2754 def : Pat<(v4i64 (X86Unpckl VR256:$src1, VR256:$src2)),
2755 (VUNPCKLPDYrr VR256:$src1, VR256:$src2)>;
2756 def : Pat<(v4i64 (X86Unpckh VR256:$src1, (loadv4i64 addr:$src2))),
2757 (VUNPCKHPDYrm VR256:$src1, addr:$src2)>;
2758 def : Pat<(v4i64 (X86Unpckh VR256:$src1, VR256:$src2)),
2759 (VUNPCKHPDYrr VR256:$src1, VR256:$src2)>;
2762 let Predicates = [HasAVX] in {
2763 // FIXME: Instead of X86Movddup, there should be a X86Unpckl here, the
2764 // problem is during lowering, where it's not possible to recognize the load
2765 // fold cause it has two uses through a bitcast. One use disappears at isel
2766 // time and the fold opportunity reappears.
2767 def : Pat<(v2f64 (X86Movddup VR128:$src)),
2768 (VUNPCKLPDrr VR128:$src, VR128:$src)>;
2771 let Predicates = [UseSSE2] in {
2772 // FIXME: Instead of X86Movddup, there should be a X86Unpckl here, the
2773 // problem is during lowering, where it's not possible to recognize the load
2774 // fold cause it has two uses through a bitcast. One use disappears at isel
2775 // time and the fold opportunity reappears.
2776 def : Pat<(v2f64 (X86Movddup VR128:$src)),
2777 (UNPCKLPDrr VR128:$src, VR128:$src)>;
2780 //===----------------------------------------------------------------------===//
2781 // SSE 1 & 2 - Extract Floating-Point Sign mask
2782 //===----------------------------------------------------------------------===//
2784 /// sse12_extr_sign_mask - sse 1 & 2 unpack and interleave
2785 multiclass sse12_extr_sign_mask<RegisterClass RC, Intrinsic Int, string asm,
2787 def rr : PI<0x50, MRMSrcReg, (outs GR32orGR64:$dst), (ins RC:$src),
2788 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
2789 [(set GR32orGR64:$dst, (Int RC:$src))], IIC_SSE_MOVMSK, d>,
2790 Sched<[WriteVecLogic]>;
2793 let Predicates = [HasAVX] in {
2794 defm VMOVMSKPS : sse12_extr_sign_mask<VR128, int_x86_sse_movmsk_ps,
2795 "movmskps", SSEPackedSingle>, PS, VEX;
2796 defm VMOVMSKPD : sse12_extr_sign_mask<VR128, int_x86_sse2_movmsk_pd,
2797 "movmskpd", SSEPackedDouble>, PD, VEX;
2798 defm VMOVMSKPSY : sse12_extr_sign_mask<VR256, int_x86_avx_movmsk_ps_256,
2799 "movmskps", SSEPackedSingle>, PS,
2801 defm VMOVMSKPDY : sse12_extr_sign_mask<VR256, int_x86_avx_movmsk_pd_256,
2802 "movmskpd", SSEPackedDouble>, PD,
2805 def : Pat<(i32 (X86fgetsign FR32:$src)),
2806 (VMOVMSKPSrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
2807 def : Pat<(i64 (X86fgetsign FR32:$src)),
2808 (SUBREG_TO_REG (i64 0),
2809 (VMOVMSKPSrr (COPY_TO_REGCLASS FR32:$src, VR128)), sub_32bit)>;
2810 def : Pat<(i32 (X86fgetsign FR64:$src)),
2811 (VMOVMSKPDrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
2812 def : Pat<(i64 (X86fgetsign FR64:$src)),
2813 (SUBREG_TO_REG (i64 0),
2814 (VMOVMSKPDrr (COPY_TO_REGCLASS FR64:$src, VR128)), sub_32bit)>;
2817 defm MOVMSKPS : sse12_extr_sign_mask<VR128, int_x86_sse_movmsk_ps, "movmskps",
2818 SSEPackedSingle>, PS;
2819 defm MOVMSKPD : sse12_extr_sign_mask<VR128, int_x86_sse2_movmsk_pd, "movmskpd",
2820 SSEPackedDouble>, PD;
2822 def : Pat<(i32 (X86fgetsign FR32:$src)),
2823 (MOVMSKPSrr (COPY_TO_REGCLASS FR32:$src, VR128))>,
2824 Requires<[UseSSE1]>;
2825 def : Pat<(i64 (X86fgetsign FR32:$src)),
2826 (SUBREG_TO_REG (i64 0),
2827 (MOVMSKPSrr (COPY_TO_REGCLASS FR32:$src, VR128)), sub_32bit)>,
2828 Requires<[UseSSE1]>;
2829 def : Pat<(i32 (X86fgetsign FR64:$src)),
2830 (MOVMSKPDrr (COPY_TO_REGCLASS FR64:$src, VR128))>,
2831 Requires<[UseSSE2]>;
2832 def : Pat<(i64 (X86fgetsign FR64:$src)),
2833 (SUBREG_TO_REG (i64 0),
2834 (MOVMSKPDrr (COPY_TO_REGCLASS FR64:$src, VR128)), sub_32bit)>,
2835 Requires<[UseSSE2]>;
2837 //===---------------------------------------------------------------------===//
2838 // SSE2 - Packed Integer Logical Instructions
2839 //===---------------------------------------------------------------------===//
2841 let ExeDomain = SSEPackedInt in { // SSE integer instructions
2843 /// PDI_binop_rm - Simple SSE2 binary operator.
2844 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
2845 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
2846 X86MemOperand x86memop, OpndItins itins,
2847 bit IsCommutable, bit Is2Addr> {
2848 let isCommutable = IsCommutable in
2849 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
2850 (ins RC:$src1, RC:$src2),
2852 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2853 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2854 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))], itins.rr>,
2855 Sched<[itins.Sched]>;
2856 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
2857 (ins RC:$src1, x86memop:$src2),
2859 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2860 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2861 [(set RC:$dst, (OpVT (OpNode RC:$src1,
2862 (bitconvert (memop_frag addr:$src2)))))],
2864 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2866 } // ExeDomain = SSEPackedInt
2868 multiclass PDI_binop_all<bits<8> opc, string OpcodeStr, SDNode Opcode,
2869 ValueType OpVT128, ValueType OpVT256,
2870 OpndItins itins, bit IsCommutable = 0> {
2871 let Predicates = [HasAVX] in
2872 defm V#NAME : PDI_binop_rm<opc, !strconcat("v", OpcodeStr), Opcode, OpVT128,
2873 VR128, loadv2i64, i128mem, itins, IsCommutable, 0>, VEX_4V;
2875 let Constraints = "$src1 = $dst" in
2876 defm NAME : PDI_binop_rm<opc, OpcodeStr, Opcode, OpVT128, VR128,
2877 memopv2i64, i128mem, itins, IsCommutable, 1>;
2879 let Predicates = [HasAVX2] in
2880 defm V#NAME#Y : PDI_binop_rm<opc, !strconcat("v", OpcodeStr), Opcode,
2881 OpVT256, VR256, loadv4i64, i256mem, itins,
2882 IsCommutable, 0>, VEX_4V, VEX_L;
2885 // These are ordered here for pattern ordering requirements with the fp versions
2887 defm PAND : PDI_binop_all<0xDB, "pand", and, v2i64, v4i64,
2888 SSE_VEC_BIT_ITINS_P, 1>;
2889 defm POR : PDI_binop_all<0xEB, "por", or, v2i64, v4i64,
2890 SSE_VEC_BIT_ITINS_P, 1>;
2891 defm PXOR : PDI_binop_all<0xEF, "pxor", xor, v2i64, v4i64,
2892 SSE_VEC_BIT_ITINS_P, 1>;
2893 defm PANDN : PDI_binop_all<0xDF, "pandn", X86andnp, v2i64, v4i64,
2894 SSE_VEC_BIT_ITINS_P, 0>;
2896 //===----------------------------------------------------------------------===//
2897 // SSE 1 & 2 - Logical Instructions
2898 //===----------------------------------------------------------------------===//
2900 /// sse12_fp_alias_pack_logical - SSE 1 & 2 aliased packed FP logical ops
2902 multiclass sse12_fp_alias_pack_logical<bits<8> opc, string OpcodeStr,
2903 SDNode OpNode, OpndItins itins> {
2904 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode,
2905 FR32, f32, f128mem, memopfsf32, SSEPackedSingle, itins, 0>,
2908 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode,
2909 FR64, f64, f128mem, memopfsf64, SSEPackedDouble, itins, 0>,
2912 let Constraints = "$src1 = $dst" in {
2913 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode, FR32,
2914 f32, f128mem, memopfsf32, SSEPackedSingle, itins>,
2917 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode, FR64,
2918 f64, f128mem, memopfsf64, SSEPackedDouble, itins>,
2923 // Alias bitwise logical operations using SSE logical ops on packed FP values.
2924 let isCodeGenOnly = 1 in {
2925 defm FsAND : sse12_fp_alias_pack_logical<0x54, "and", X86fand,
2927 defm FsOR : sse12_fp_alias_pack_logical<0x56, "or", X86for,
2929 defm FsXOR : sse12_fp_alias_pack_logical<0x57, "xor", X86fxor,
2932 let isCommutable = 0 in
2933 defm FsANDN : sse12_fp_alias_pack_logical<0x55, "andn", X86fandn,
2937 /// sse12_fp_packed_logical - SSE 1 & 2 packed FP logical ops
2939 multiclass sse12_fp_packed_logical<bits<8> opc, string OpcodeStr,
2941 defm V#NAME#PSY : sse12_fp_packed_logical_rm<opc, VR256, SSEPackedSingle,
2942 !strconcat(OpcodeStr, "ps"), f256mem,
2943 [(set VR256:$dst, (v4i64 (OpNode VR256:$src1, VR256:$src2)))],
2944 [(set VR256:$dst, (OpNode (bc_v4i64 (v8f32 VR256:$src1)),
2945 (loadv4i64 addr:$src2)))], 0>, PS, VEX_4V, VEX_L;
2947 defm V#NAME#PDY : sse12_fp_packed_logical_rm<opc, VR256, SSEPackedDouble,
2948 !strconcat(OpcodeStr, "pd"), f256mem,
2949 [(set VR256:$dst, (OpNode (bc_v4i64 (v4f64 VR256:$src1)),
2950 (bc_v4i64 (v4f64 VR256:$src2))))],
2951 [(set VR256:$dst, (OpNode (bc_v4i64 (v4f64 VR256:$src1)),
2952 (loadv4i64 addr:$src2)))], 0>,
2955 // In AVX no need to add a pattern for 128-bit logical rr ps, because they
2956 // are all promoted to v2i64, and the patterns are covered by the int
2957 // version. This is needed in SSE only, because v2i64 isn't supported on
2958 // SSE1, but only on SSE2.
2959 defm V#NAME#PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
2960 !strconcat(OpcodeStr, "ps"), f128mem, [],
2961 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2962 (loadv2i64 addr:$src2)))], 0>, PS, VEX_4V;
2964 defm V#NAME#PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
2965 !strconcat(OpcodeStr, "pd"), f128mem,
2966 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2967 (bc_v2i64 (v2f64 VR128:$src2))))],
2968 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2969 (loadv2i64 addr:$src2)))], 0>,
2972 let Constraints = "$src1 = $dst" in {
2973 defm PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
2974 !strconcat(OpcodeStr, "ps"), f128mem,
2975 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))],
2976 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2977 (memopv2i64 addr:$src2)))]>, PS;
2979 defm PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
2980 !strconcat(OpcodeStr, "pd"), f128mem,
2981 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2982 (bc_v2i64 (v2f64 VR128:$src2))))],
2983 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2984 (memopv2i64 addr:$src2)))]>, PD;
2988 defm AND : sse12_fp_packed_logical<0x54, "and", and>;
2989 defm OR : sse12_fp_packed_logical<0x56, "or", or>;
2990 defm XOR : sse12_fp_packed_logical<0x57, "xor", xor>;
2991 let isCommutable = 0 in
2992 defm ANDN : sse12_fp_packed_logical<0x55, "andn", X86andnp>;
2994 // AVX1 requires type coercions in order to fold loads directly into logical
2996 let Predicates = [HasAVX1Only] in {
2997 def : Pat<(bc_v8f32 (and VR256:$src1, (loadv4i64 addr:$src2))),
2998 (VANDPSYrm VR256:$src1, addr:$src2)>;
2999 def : Pat<(bc_v8f32 (or VR256:$src1, (loadv4i64 addr:$src2))),
3000 (VORPSYrm VR256:$src1, addr:$src2)>;
3001 def : Pat<(bc_v8f32 (xor VR256:$src1, (loadv4i64 addr:$src2))),
3002 (VXORPSYrm VR256:$src1, addr:$src2)>;
3003 def : Pat<(bc_v8f32 (X86andnp VR256:$src1, (loadv4i64 addr:$src2))),
3004 (VANDNPSYrm VR256:$src1, addr:$src2)>;
3007 //===----------------------------------------------------------------------===//
3008 // SSE 1 & 2 - Arithmetic Instructions
3009 //===----------------------------------------------------------------------===//
3011 /// basic_sse12_fp_binop_xxx - SSE 1 & 2 binops come in both scalar and
3014 /// In addition, we also have a special variant of the scalar form here to
3015 /// represent the associated intrinsic operation. This form is unlike the
3016 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
3017 /// and leaves the top elements unmodified (therefore these cannot be commuted).
3019 /// These three forms can each be reg+reg or reg+mem.
3022 /// FIXME: once all 256-bit intrinsics are matched, cleanup and refactor those
3024 multiclass basic_sse12_fp_binop_p<bits<8> opc, string OpcodeStr,
3025 SDNode OpNode, SizeItins itins> {
3026 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode,
3027 VR128, v4f32, f128mem, loadv4f32,
3028 SSEPackedSingle, itins.s, 0>, PS, VEX_4V;
3029 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode,
3030 VR128, v2f64, f128mem, loadv2f64,
3031 SSEPackedDouble, itins.d, 0>, PD, VEX_4V;
3033 defm V#NAME#PSY : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"),
3034 OpNode, VR256, v8f32, f256mem, loadv8f32,
3035 SSEPackedSingle, itins.s, 0>, PS, VEX_4V, VEX_L;
3036 defm V#NAME#PDY : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"),
3037 OpNode, VR256, v4f64, f256mem, loadv4f64,
3038 SSEPackedDouble, itins.d, 0>, PD, VEX_4V, VEX_L;
3040 let Constraints = "$src1 = $dst" in {
3041 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode, VR128,
3042 v4f32, f128mem, memopv4f32, SSEPackedSingle,
3044 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode, VR128,
3045 v2f64, f128mem, memopv2f64, SSEPackedDouble,
3050 multiclass basic_sse12_fp_binop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
3052 defm V#NAME#SS : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "ss"),
3053 OpNode, FR32, f32mem, itins.s, 0>, XS, VEX_4V, VEX_LIG;
3054 defm V#NAME#SD : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "sd"),
3055 OpNode, FR64, f64mem, itins.d, 0>, XD, VEX_4V, VEX_LIG;
3057 let Constraints = "$src1 = $dst" in {
3058 defm SS : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "ss"),
3059 OpNode, FR32, f32mem, itins.s>, XS;
3060 defm SD : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "sd"),
3061 OpNode, FR64, f64mem, itins.d>, XD;
3065 multiclass basic_sse12_fp_binop_s_int<bits<8> opc, string OpcodeStr,
3067 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
3068 !strconcat(OpcodeStr, "ss"), "", "_ss", ssmem, sse_load_f32,
3069 itins.s, 0>, XS, VEX_4V, VEX_LIG;
3070 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
3071 !strconcat(OpcodeStr, "sd"), "2", "_sd", sdmem, sse_load_f64,
3072 itins.d, 0>, XD, VEX_4V, VEX_LIG;
3074 let Constraints = "$src1 = $dst" in {
3075 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
3076 !strconcat(OpcodeStr, "ss"), "", "_ss", ssmem, sse_load_f32,
3078 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
3079 !strconcat(OpcodeStr, "sd"), "2", "_sd", sdmem, sse_load_f64,
3084 // Binary Arithmetic instructions
3085 defm ADD : basic_sse12_fp_binop_p<0x58, "add", fadd, SSE_ALU_ITINS_P>,
3086 basic_sse12_fp_binop_s<0x58, "add", fadd, SSE_ALU_ITINS_S>,
3087 basic_sse12_fp_binop_s_int<0x58, "add", SSE_ALU_ITINS_S>;
3088 defm MUL : basic_sse12_fp_binop_p<0x59, "mul", fmul, SSE_MUL_ITINS_P>,
3089 basic_sse12_fp_binop_s<0x59, "mul", fmul, SSE_MUL_ITINS_S>,
3090 basic_sse12_fp_binop_s_int<0x59, "mul", SSE_MUL_ITINS_S>;
3091 let isCommutable = 0 in {
3092 defm SUB : basic_sse12_fp_binop_p<0x5C, "sub", fsub, SSE_ALU_ITINS_P>,
3093 basic_sse12_fp_binop_s<0x5C, "sub", fsub, SSE_ALU_ITINS_S>,
3094 basic_sse12_fp_binop_s_int<0x5C, "sub", SSE_ALU_ITINS_S>;
3095 defm DIV : basic_sse12_fp_binop_p<0x5E, "div", fdiv, SSE_DIV_ITINS_P>,
3096 basic_sse12_fp_binop_s<0x5E, "div", fdiv, SSE_DIV_ITINS_S>,
3097 basic_sse12_fp_binop_s_int<0x5E, "div", SSE_DIV_ITINS_S>;
3098 defm MAX : basic_sse12_fp_binop_p<0x5F, "max", X86fmax, SSE_ALU_ITINS_P>,
3099 basic_sse12_fp_binop_s<0x5F, "max", X86fmax, SSE_ALU_ITINS_S>,
3100 basic_sse12_fp_binop_s_int<0x5F, "max", SSE_ALU_ITINS_S>;
3101 defm MIN : basic_sse12_fp_binop_p<0x5D, "min", X86fmin, SSE_ALU_ITINS_P>,
3102 basic_sse12_fp_binop_s<0x5D, "min", X86fmin, SSE_ALU_ITINS_S>,
3103 basic_sse12_fp_binop_s_int<0x5D, "min", SSE_ALU_ITINS_S>;
3106 let isCodeGenOnly = 1 in {
3107 defm MAXC: basic_sse12_fp_binop_p<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_P>,
3108 basic_sse12_fp_binop_s<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_S>;
3109 defm MINC: basic_sse12_fp_binop_p<0x5D, "min", X86fminc, SSE_ALU_ITINS_P>,
3110 basic_sse12_fp_binop_s<0x5D, "min", X86fminc, SSE_ALU_ITINS_S>;
3113 // Patterns used to select SSE scalar fp arithmetic instructions from
3114 // a scalar fp operation followed by a blend.
3116 // These patterns know, for example, how to select an ADDSS from a
3117 // float add plus vector insert.
3119 // The effect is that the backend no longer emits unnecessary vector
3120 // insert instructions immediately after SSE scalar fp instructions
3121 // like addss or mulss.
3123 // For example, given the following code:
3124 // __m128 foo(__m128 A, __m128 B) {
3129 // previously we generated:
3130 // addss %xmm0, %xmm1
3131 // movss %xmm1, %xmm0
3134 // addss %xmm1, %xmm0
3136 let Predicates = [UseSSE1] in {
3137 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector (fadd
3138 (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3140 (ADDSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3141 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector (fsub
3142 (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3144 (SUBSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3145 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector (fmul
3146 (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3148 (MULSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3149 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector (fdiv
3150 (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3152 (DIVSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3155 let Predicates = [UseSSE2] in {
3156 // SSE2 patterns to select scalar double-precision fp arithmetic instructions
3158 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fadd
3159 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3161 (ADDSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3162 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fsub
3163 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3165 (SUBSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3166 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fmul
3167 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3169 (MULSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3170 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fdiv
3171 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3173 (DIVSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3176 let Predicates = [UseSSE41] in {
3177 // If the subtarget has SSE4.1 but not AVX, the vector insert
3178 // instruction is lowered into a X86insertps rather than a X86Movss.
3179 // When selecting SSE scalar single-precision fp arithmetic instructions,
3180 // make sure that we correctly match the X86insertps.
3182 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3183 (fadd (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3184 FR32:$src))), (iPTR 0))),
3185 (ADDSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3186 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3187 (fsub (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3188 FR32:$src))), (iPTR 0))),
3189 (SUBSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3190 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3191 (fmul (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3192 FR32:$src))), (iPTR 0))),
3193 (MULSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3194 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3195 (fdiv (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3196 FR32:$src))), (iPTR 0))),
3197 (DIVSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3200 let Predicates = [HasAVX] in {
3201 // The following patterns select AVX Scalar single/double precision fp
3202 // arithmetic instructions.
3204 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fadd
3205 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3207 (VADDSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3208 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fsub
3209 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3211 (VSUBSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3212 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fmul
3213 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3215 (VMULSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3216 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector (fdiv
3217 (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3219 (VDIVSDrr_Int v2f64:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
3220 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3221 (fadd (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3222 FR32:$src))), (iPTR 0))),
3223 (VADDSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3224 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3225 (fsub (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3226 FR32:$src))), (iPTR 0))),
3227 (VSUBSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3228 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3229 (fmul (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3230 FR32:$src))), (iPTR 0))),
3231 (VMULSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3232 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3233 (fdiv (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3234 FR32:$src))), (iPTR 0))),
3235 (VDIVSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
3238 // Patterns used to select SSE scalar fp arithmetic instructions from
3239 // a vector packed single/double fp operation followed by a vector insert.
3241 // The effect is that the backend converts the packed fp instruction
3242 // followed by a vector insert into a single SSE scalar fp instruction.
3244 // For example, given the following code:
3245 // __m128 foo(__m128 A, __m128 B) {
3246 // __m128 C = A + B;
3247 // return (__m128) {c[0], a[1], a[2], a[3]};
3250 // previously we generated:
3251 // addps %xmm0, %xmm1
3252 // movss %xmm1, %xmm0
3255 // addss %xmm1, %xmm0
3257 let Predicates = [UseSSE1] in {
3258 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3259 (fadd (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3260 (ADDSSrr_Int v4f32:$dst, v4f32:$src)>;
3261 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3262 (fsub (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3263 (SUBSSrr_Int v4f32:$dst, v4f32:$src)>;
3264 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3265 (fmul (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3266 (MULSSrr_Int v4f32:$dst, v4f32:$src)>;
3267 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3268 (fdiv (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3269 (DIVSSrr_Int v4f32:$dst, v4f32:$src)>;
3272 let Predicates = [UseSSE2] in {
3273 // SSE2 patterns to select scalar double-precision fp arithmetic instructions
3274 // from a packed double-precision fp instruction plus movsd.
3276 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3277 (fadd (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3278 (ADDSDrr_Int v2f64:$dst, v2f64:$src)>;
3279 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3280 (fsub (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3281 (SUBSDrr_Int v2f64:$dst, v2f64:$src)>;
3282 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3283 (fmul (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3284 (MULSDrr_Int v2f64:$dst, v2f64:$src)>;
3285 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3286 (fdiv (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3287 (DIVSDrr_Int v2f64:$dst, v2f64:$src)>;
3290 let Predicates = [HasAVX] in {
3291 // The following patterns select AVX Scalar single/double precision fp
3292 // arithmetic instructions from a packed single precision fp instruction
3293 // plus movss/movsd.
3295 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3296 (fadd (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3297 (VADDSSrr_Int v4f32:$dst, v4f32:$src)>;
3298 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3299 (fsub (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3300 (VSUBSSrr_Int v4f32:$dst, v4f32:$src)>;
3301 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3302 (fmul (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3303 (VMULSSrr_Int v4f32:$dst, v4f32:$src)>;
3304 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3305 (fdiv (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3306 (VDIVSSrr_Int v4f32:$dst, v4f32:$src)>;
3307 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3308 (fadd (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3309 (VADDSDrr_Int v2f64:$dst, v2f64:$src)>;
3310 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3311 (fsub (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3312 (VSUBSDrr_Int v2f64:$dst, v2f64:$src)>;
3313 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3314 (fmul (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3315 (VMULSDrr_Int v2f64:$dst, v2f64:$src)>;
3316 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3317 (fdiv (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3318 (VDIVSDrr_Int v2f64:$dst, v2f64:$src)>;
3322 /// In addition, we also have a special variant of the scalar form here to
3323 /// represent the associated intrinsic operation. This form is unlike the
3324 /// plain scalar form, in that it takes an entire vector (instead of a
3325 /// scalar) and leaves the top elements undefined.
3327 /// And, we have a special variant form for a full-vector intrinsic form.
3329 let Sched = WriteFSqrt in {
3330 def SSE_SQRTPS : OpndItins<
3331 IIC_SSE_SQRTPS_RR, IIC_SSE_SQRTPS_RM
3334 def SSE_SQRTSS : OpndItins<
3335 IIC_SSE_SQRTSS_RR, IIC_SSE_SQRTSS_RM
3338 def SSE_SQRTPD : OpndItins<
3339 IIC_SSE_SQRTPD_RR, IIC_SSE_SQRTPD_RM
3342 def SSE_SQRTSD : OpndItins<
3343 IIC_SSE_SQRTSD_RR, IIC_SSE_SQRTSD_RM
3347 let Sched = WriteFRcp in {
3348 def SSE_RCPP : OpndItins<
3349 IIC_SSE_RCPP_RR, IIC_SSE_RCPP_RM
3352 def SSE_RCPS : OpndItins<
3353 IIC_SSE_RCPS_RR, IIC_SSE_RCPS_RM
3357 /// sse1_fp_unop_s - SSE1 unops in scalar form.
3358 multiclass sse1_fp_unop_s<bits<8> opc, string OpcodeStr,
3359 SDNode OpNode, Intrinsic F32Int, OpndItins itins> {
3360 let Predicates = [HasAVX], hasSideEffects = 0 in {
3361 def V#NAME#SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst),
3362 (ins FR32:$src1, FR32:$src2),
3363 !strconcat("v", OpcodeStr,
3364 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3365 []>, VEX_4V, VEX_LIG, Sched<[itins.Sched]>;
3366 let mayLoad = 1 in {
3367 def V#NAME#SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
3368 (ins FR32:$src1,f32mem:$src2),
3369 !strconcat("v", OpcodeStr,
3370 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3371 []>, VEX_4V, VEX_LIG,
3372 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3373 let isCodeGenOnly = 1 in
3374 def V#NAME#SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
3375 (ins VR128:$src1, ssmem:$src2),
3376 !strconcat("v", OpcodeStr,
3377 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3378 []>, VEX_4V, VEX_LIG,
3379 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3383 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
3384 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
3385 [(set FR32:$dst, (OpNode FR32:$src))]>, Sched<[itins.Sched]>;
3386 // For scalar unary operations, fold a load into the operation
3387 // only in OptForSize mode. It eliminates an instruction, but it also
3388 // eliminates a whole-register clobber (the load), so it introduces a
3389 // partial register update condition.
3390 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
3391 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
3392 [(set FR32:$dst, (OpNode (load addr:$src)))], itins.rm>, XS,
3393 Requires<[UseSSE1, OptForSize]>, Sched<[itins.Sched.Folded]>;
3394 let isCodeGenOnly = 1 in {
3395 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3396 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
3397 [(set VR128:$dst, (F32Int VR128:$src))], itins.rr>,
3398 Sched<[itins.Sched]>;
3399 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
3400 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
3401 [(set VR128:$dst, (F32Int sse_load_f32:$src))], itins.rm>,
3402 Sched<[itins.Sched.Folded]>;
3406 /// sse1_fp_unop_s_rw - SSE1 unops where vector form has a read-write operand.
3407 multiclass sse1_fp_unop_rw<bits<8> opc, string OpcodeStr, SDNode OpNode,
3409 let Predicates = [HasAVX], hasSideEffects = 0 in {
3410 def V#NAME#SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst),
3411 (ins FR32:$src1, FR32:$src2),
3412 !strconcat("v", OpcodeStr,
3413 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3414 []>, VEX_4V, VEX_LIG, Sched<[itins.Sched]>;
3415 let mayLoad = 1 in {
3416 def V#NAME#SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
3417 (ins FR32:$src1,f32mem:$src2),
3418 !strconcat("v", OpcodeStr,
3419 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3420 []>, VEX_4V, VEX_LIG,
3421 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3422 let isCodeGenOnly = 1 in
3423 def V#NAME#SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
3424 (ins VR128:$src1, ssmem:$src2),
3425 !strconcat("v", OpcodeStr,
3426 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3427 []>, VEX_4V, VEX_LIG,
3428 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3432 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
3433 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
3434 [(set FR32:$dst, (OpNode FR32:$src))]>, Sched<[itins.Sched]>;
3435 // For scalar unary operations, fold a load into the operation
3436 // only in OptForSize mode. It eliminates an instruction, but it also
3437 // eliminates a whole-register clobber (the load), so it introduces a
3438 // partial register update condition.
3439 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
3440 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
3441 [(set FR32:$dst, (OpNode (load addr:$src)))], itins.rm>, XS,
3442 Requires<[UseSSE1, OptForSize]>, Sched<[itins.Sched.Folded]>;
3443 let isCodeGenOnly = 1, Constraints = "$src1 = $dst" in {
3444 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
3445 (ins VR128:$src1, VR128:$src2),
3446 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
3447 [], itins.rr>, Sched<[itins.Sched]>;
3448 let mayLoad = 1, hasSideEffects = 0 in
3449 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
3450 (ins VR128:$src1, ssmem:$src2),
3451 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
3452 [], itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3456 /// sse1_fp_unop_p - SSE1 unops in packed form.
3457 multiclass sse1_fp_unop_p<bits<8> opc, string OpcodeStr, SDNode OpNode,
3459 let Predicates = [HasAVX] in {
3460 def V#NAME#PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3461 !strconcat("v", OpcodeStr,
3462 "ps\t{$src, $dst|$dst, $src}"),
3463 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))],
3464 itins.rr>, VEX, Sched<[itins.Sched]>;
3465 def V#NAME#PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3466 !strconcat("v", OpcodeStr,
3467 "ps\t{$src, $dst|$dst, $src}"),
3468 [(set VR128:$dst, (OpNode (loadv4f32 addr:$src)))],
3469 itins.rm>, VEX, Sched<[itins.Sched.Folded]>;
3470 def V#NAME#PSYr : PSI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3471 !strconcat("v", OpcodeStr,
3472 "ps\t{$src, $dst|$dst, $src}"),
3473 [(set VR256:$dst, (v8f32 (OpNode VR256:$src)))],
3474 itins.rr>, VEX, VEX_L, Sched<[itins.Sched]>;
3475 def V#NAME#PSYm : PSI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
3476 !strconcat("v", OpcodeStr,
3477 "ps\t{$src, $dst|$dst, $src}"),
3478 [(set VR256:$dst, (OpNode (loadv8f32 addr:$src)))],
3479 itins.rm>, VEX, VEX_L, Sched<[itins.Sched.Folded]>;
3482 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3483 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3484 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))], itins.rr>,
3485 Sched<[itins.Sched]>;
3486 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3487 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3488 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))], itins.rm>,
3489 Sched<[itins.Sched.Folded]>;
3492 /// sse1_fp_unop_p_int - SSE1 intrinsics unops in packed forms.
3493 multiclass sse1_fp_unop_p_int<bits<8> opc, string OpcodeStr,
3494 Intrinsic V4F32Int, Intrinsic V8F32Int,
3496 let isCodeGenOnly = 1 in {
3497 let Predicates = [HasAVX] in {
3498 def V#NAME#PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3499 !strconcat("v", OpcodeStr,
3500 "ps\t{$src, $dst|$dst, $src}"),
3501 [(set VR128:$dst, (V4F32Int VR128:$src))],
3502 itins.rr>, VEX, Sched<[itins.Sched]>;
3503 def V#NAME#PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3504 !strconcat("v", OpcodeStr,
3505 "ps\t{$src, $dst|$dst, $src}"),
3506 [(set VR128:$dst, (V4F32Int (loadv4f32 addr:$src)))],
3507 itins.rm>, VEX, Sched<[itins.Sched.Folded]>;
3508 def V#NAME#PSYr_Int : PSI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3509 !strconcat("v", OpcodeStr,
3510 "ps\t{$src, $dst|$dst, $src}"),
3511 [(set VR256:$dst, (V8F32Int VR256:$src))],
3512 itins.rr>, VEX, VEX_L, Sched<[itins.Sched]>;
3513 def V#NAME#PSYm_Int : PSI<opc, MRMSrcMem, (outs VR256:$dst),
3515 !strconcat("v", OpcodeStr,
3516 "ps\t{$src, $dst|$dst, $src}"),
3517 [(set VR256:$dst, (V8F32Int (loadv8f32 addr:$src)))],
3518 itins.rm>, VEX, VEX_L, Sched<[itins.Sched.Folded]>;
3521 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3522 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3523 [(set VR128:$dst, (V4F32Int VR128:$src))],
3524 itins.rr>, Sched<[itins.Sched]>;
3525 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3526 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3527 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))],
3528 itins.rm>, Sched<[itins.Sched.Folded]>;
3529 } // isCodeGenOnly = 1
3532 /// sse2_fp_unop_s - SSE2 unops in scalar form.
3533 multiclass sse2_fp_unop_s<bits<8> opc, string OpcodeStr,
3534 SDNode OpNode, Intrinsic F64Int, OpndItins itins> {
3535 let Predicates = [HasAVX], hasSideEffects = 0 in {
3536 def V#NAME#SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst),
3537 (ins FR64:$src1, FR64:$src2),
3538 !strconcat("v", OpcodeStr,
3539 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3540 []>, VEX_4V, VEX_LIG, Sched<[itins.Sched]>;
3541 let mayLoad = 1 in {
3542 def V#NAME#SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
3543 (ins FR64:$src1,f64mem:$src2),
3544 !strconcat("v", OpcodeStr,
3545 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3546 []>, VEX_4V, VEX_LIG,
3547 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3548 let isCodeGenOnly = 1 in
3549 def V#NAME#SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
3550 (ins VR128:$src1, sdmem:$src2),
3551 !strconcat("v", OpcodeStr,
3552 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3553 []>, VEX_4V, VEX_LIG,
3554 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3558 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
3559 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
3560 [(set FR64:$dst, (OpNode FR64:$src))], itins.rr>,
3561 Sched<[itins.Sched]>;
3562 // See the comments in sse1_fp_unop_s for why this is OptForSize.
3563 def SDm : I<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
3564 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
3565 [(set FR64:$dst, (OpNode (load addr:$src)))], itins.rm>, XD,
3566 Requires<[UseSSE2, OptForSize]>, Sched<[itins.Sched.Folded]>;
3567 let isCodeGenOnly = 1 in {
3568 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3569 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
3570 [(set VR128:$dst, (F64Int VR128:$src))], itins.rr>,
3571 Sched<[itins.Sched]>;
3572 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
3573 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
3574 [(set VR128:$dst, (F64Int sse_load_f64:$src))], itins.rm>,
3575 Sched<[itins.Sched.Folded]>;
3579 /// sse2_fp_unop_p - SSE2 unops in vector forms.
3580 multiclass sse2_fp_unop_p<bits<8> opc, string OpcodeStr,
3581 SDNode OpNode, OpndItins itins> {
3582 let Predicates = [HasAVX] in {
3583 def V#NAME#PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3584 !strconcat("v", OpcodeStr,
3585 "pd\t{$src, $dst|$dst, $src}"),
3586 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))],
3587 itins.rr>, VEX, Sched<[itins.Sched]>;
3588 def V#NAME#PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3589 !strconcat("v", OpcodeStr,
3590 "pd\t{$src, $dst|$dst, $src}"),
3591 [(set VR128:$dst, (OpNode (loadv2f64 addr:$src)))],
3592 itins.rm>, VEX, Sched<[itins.Sched.Folded]>;
3593 def V#NAME#PDYr : PDI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3594 !strconcat("v", OpcodeStr,
3595 "pd\t{$src, $dst|$dst, $src}"),
3596 [(set VR256:$dst, (v4f64 (OpNode VR256:$src)))],
3597 itins.rr>, VEX, VEX_L, Sched<[itins.Sched]>;
3598 def V#NAME#PDYm : PDI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
3599 !strconcat("v", OpcodeStr,
3600 "pd\t{$src, $dst|$dst, $src}"),
3601 [(set VR256:$dst, (OpNode (loadv4f64 addr:$src)))],
3602 itins.rm>, VEX, VEX_L, Sched<[itins.Sched.Folded]>;
3605 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3606 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
3607 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))], itins.rr>,
3608 Sched<[itins.Sched]>;
3609 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3610 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
3611 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))], itins.rm>,
3612 Sched<[itins.Sched.Folded]>;
3616 defm SQRT : sse1_fp_unop_s<0x51, "sqrt", fsqrt, int_x86_sse_sqrt_ss,
3618 sse1_fp_unop_p<0x51, "sqrt", fsqrt, SSE_SQRTPS>,
3619 sse2_fp_unop_s<0x51, "sqrt", fsqrt, int_x86_sse2_sqrt_sd,
3621 sse2_fp_unop_p<0x51, "sqrt", fsqrt, SSE_SQRTPD>;
3623 // Reciprocal approximations. Note that these typically require refinement
3624 // in order to obtain suitable precision.
3625 defm RSQRT : sse1_fp_unop_rw<0x52, "rsqrt", X86frsqrt, SSE_SQRTSS>,
3626 sse1_fp_unop_p<0x52, "rsqrt", X86frsqrt, SSE_SQRTPS>,
3627 sse1_fp_unop_p_int<0x52, "rsqrt", int_x86_sse_rsqrt_ps,
3628 int_x86_avx_rsqrt_ps_256, SSE_SQRTPS>;
3629 defm RCP : sse1_fp_unop_rw<0x53, "rcp", X86frcp, SSE_RCPS>,
3630 sse1_fp_unop_p<0x53, "rcp", X86frcp, SSE_RCPP>,
3631 sse1_fp_unop_p_int<0x53, "rcp", int_x86_sse_rcp_ps,
3632 int_x86_avx_rcp_ps_256, SSE_RCPP>;
3634 let Predicates = [UseAVX] in {
3635 def : Pat<(f32 (fsqrt FR32:$src)),
3636 (VSQRTSSr (f32 (IMPLICIT_DEF)), FR32:$src)>, Requires<[HasAVX]>;
3637 def : Pat<(f32 (fsqrt (load addr:$src))),
3638 (VSQRTSSm (f32 (IMPLICIT_DEF)), addr:$src)>,
3639 Requires<[HasAVX, OptForSize]>;
3640 def : Pat<(f64 (fsqrt FR64:$src)),
3641 (VSQRTSDr (f64 (IMPLICIT_DEF)), FR64:$src)>, Requires<[HasAVX]>;
3642 def : Pat<(f64 (fsqrt (load addr:$src))),
3643 (VSQRTSDm (f64 (IMPLICIT_DEF)), addr:$src)>,
3644 Requires<[HasAVX, OptForSize]>;
3646 def : Pat<(f32 (X86frsqrt FR32:$src)),
3647 (VRSQRTSSr (f32 (IMPLICIT_DEF)), FR32:$src)>, Requires<[HasAVX]>;
3648 def : Pat<(f32 (X86frsqrt (load addr:$src))),
3649 (VRSQRTSSm (f32 (IMPLICIT_DEF)), addr:$src)>,
3650 Requires<[HasAVX, OptForSize]>;
3652 def : Pat<(f32 (X86frcp FR32:$src)),
3653 (VRCPSSr (f32 (IMPLICIT_DEF)), FR32:$src)>, Requires<[HasAVX]>;
3654 def : Pat<(f32 (X86frcp (load addr:$src))),
3655 (VRCPSSm (f32 (IMPLICIT_DEF)), addr:$src)>,
3656 Requires<[HasAVX, OptForSize]>;
3658 let Predicates = [UseAVX] in {
3659 def : Pat<(int_x86_sse_sqrt_ss VR128:$src),
3660 (COPY_TO_REGCLASS (VSQRTSSr (f32 (IMPLICIT_DEF)),
3661 (COPY_TO_REGCLASS VR128:$src, FR32)),
3663 def : Pat<(int_x86_sse_sqrt_ss sse_load_f32:$src),
3664 (VSQRTSSm_Int (v4f32 (IMPLICIT_DEF)), sse_load_f32:$src)>;
3666 def : Pat<(int_x86_sse2_sqrt_sd VR128:$src),
3667 (COPY_TO_REGCLASS (VSQRTSDr (f64 (IMPLICIT_DEF)),
3668 (COPY_TO_REGCLASS VR128:$src, FR64)),
3670 def : Pat<(int_x86_sse2_sqrt_sd sse_load_f64:$src),
3671 (VSQRTSDm_Int (v2f64 (IMPLICIT_DEF)), sse_load_f64:$src)>;
3674 let Predicates = [HasAVX] in {
3675 def : Pat<(int_x86_sse_rsqrt_ss VR128:$src),
3676 (COPY_TO_REGCLASS (VRSQRTSSr (f32 (IMPLICIT_DEF)),
3677 (COPY_TO_REGCLASS VR128:$src, FR32)),
3679 def : Pat<(int_x86_sse_rsqrt_ss sse_load_f32:$src),
3680 (VRSQRTSSm_Int (v4f32 (IMPLICIT_DEF)), sse_load_f32:$src)>;
3682 def : Pat<(int_x86_sse_rcp_ss VR128:$src),
3683 (COPY_TO_REGCLASS (VRCPSSr (f32 (IMPLICIT_DEF)),
3684 (COPY_TO_REGCLASS VR128:$src, FR32)),
3686 def : Pat<(int_x86_sse_rcp_ss sse_load_f32:$src),
3687 (VRCPSSm_Int (v4f32 (IMPLICIT_DEF)), sse_load_f32:$src)>;
3690 // Reciprocal approximations. Note that these typically require refinement
3691 // in order to obtain suitable precision.
3692 let Predicates = [UseSSE1] in {
3693 def : Pat<(int_x86_sse_rsqrt_ss VR128:$src),
3694 (RSQRTSSr_Int VR128:$src, VR128:$src)>;
3695 def : Pat<(int_x86_sse_rcp_ss VR128:$src),
3696 (RCPSSr_Int VR128:$src, VR128:$src)>;
3699 // There is no f64 version of the reciprocal approximation instructions.
3701 //===----------------------------------------------------------------------===//
3702 // SSE 1 & 2 - Non-temporal stores
3703 //===----------------------------------------------------------------------===//
3705 let AddedComplexity = 400 in { // Prefer non-temporal versions
3706 let SchedRW = [WriteStore] in {
3707 let Predicates = [HasAVX, NoVLX] in {
3708 def VMOVNTPSmr : VPSI<0x2B, MRMDestMem, (outs),
3709 (ins f128mem:$dst, VR128:$src),
3710 "movntps\t{$src, $dst|$dst, $src}",
3711 [(alignednontemporalstore (v4f32 VR128:$src),
3713 IIC_SSE_MOVNT>, VEX;
3714 def VMOVNTPDmr : VPDI<0x2B, MRMDestMem, (outs),
3715 (ins f128mem:$dst, VR128:$src),
3716 "movntpd\t{$src, $dst|$dst, $src}",
3717 [(alignednontemporalstore (v2f64 VR128:$src),
3719 IIC_SSE_MOVNT>, VEX;
3721 let ExeDomain = SSEPackedInt in
3722 def VMOVNTDQmr : VPDI<0xE7, MRMDestMem, (outs),
3723 (ins f128mem:$dst, VR128:$src),
3724 "movntdq\t{$src, $dst|$dst, $src}",
3725 [(alignednontemporalstore (v2i64 VR128:$src),
3727 IIC_SSE_MOVNT>, VEX;
3729 def VMOVNTPSYmr : VPSI<0x2B, MRMDestMem, (outs),
3730 (ins f256mem:$dst, VR256:$src),
3731 "movntps\t{$src, $dst|$dst, $src}",
3732 [(alignednontemporalstore (v8f32 VR256:$src),
3734 IIC_SSE_MOVNT>, VEX, VEX_L;
3735 def VMOVNTPDYmr : VPDI<0x2B, MRMDestMem, (outs),
3736 (ins f256mem:$dst, VR256:$src),
3737 "movntpd\t{$src, $dst|$dst, $src}",
3738 [(alignednontemporalstore (v4f64 VR256:$src),
3740 IIC_SSE_MOVNT>, VEX, VEX_L;
3741 let ExeDomain = SSEPackedInt in
3742 def VMOVNTDQYmr : VPDI<0xE7, MRMDestMem, (outs),
3743 (ins f256mem:$dst, VR256:$src),
3744 "movntdq\t{$src, $dst|$dst, $src}",
3745 [(alignednontemporalstore (v4i64 VR256:$src),
3747 IIC_SSE_MOVNT>, VEX, VEX_L;
3750 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3751 "movntps\t{$src, $dst|$dst, $src}",
3752 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)],
3754 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3755 "movntpd\t{$src, $dst|$dst, $src}",
3756 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)],
3759 let ExeDomain = SSEPackedInt in
3760 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3761 "movntdq\t{$src, $dst|$dst, $src}",
3762 [(alignednontemporalstore (v2i64 VR128:$src), addr:$dst)],
3765 // There is no AVX form for instructions below this point
3766 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
3767 "movnti{l}\t{$src, $dst|$dst, $src}",
3768 [(nontemporalstore (i32 GR32:$src), addr:$dst)],
3770 PS, Requires<[HasSSE2]>;
3771 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
3772 "movnti{q}\t{$src, $dst|$dst, $src}",
3773 [(nontemporalstore (i64 GR64:$src), addr:$dst)],
3775 PS, Requires<[HasSSE2]>;
3776 } // SchedRW = [WriteStore]
3778 } // AddedComplexity
3780 //===----------------------------------------------------------------------===//
3781 // SSE 1 & 2 - Prefetch and memory fence
3782 //===----------------------------------------------------------------------===//
3784 // Prefetch intrinsic.
3785 let Predicates = [HasSSE1], SchedRW = [WriteLoad] in {
3786 def PREFETCHT0 : I<0x18, MRM1m, (outs), (ins i8mem:$src),
3787 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3), (i32 1))],
3788 IIC_SSE_PREFETCH>, TB;
3789 def PREFETCHT1 : I<0x18, MRM2m, (outs), (ins i8mem:$src),
3790 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2), (i32 1))],
3791 IIC_SSE_PREFETCH>, TB;
3792 def PREFETCHT2 : I<0x18, MRM3m, (outs), (ins i8mem:$src),
3793 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1), (i32 1))],
3794 IIC_SSE_PREFETCH>, TB;
3795 def PREFETCHNTA : I<0x18, MRM0m, (outs), (ins i8mem:$src),
3796 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0), (i32 1))],
3797 IIC_SSE_PREFETCH>, TB;
3800 // FIXME: How should flush instruction be modeled?
3801 let SchedRW = [WriteLoad] in {
3803 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
3804 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)],
3805 IIC_SSE_PREFETCH>, TB, Requires<[HasSSE2]>;
3808 let SchedRW = [WriteNop] in {
3809 // Pause. This "instruction" is encoded as "rep; nop", so even though it
3810 // was introduced with SSE2, it's backward compatible.
3811 def PAUSE : I<0x90, RawFrm, (outs), (ins),
3812 "pause", [(int_x86_sse2_pause)], IIC_SSE_PAUSE>,
3813 OBXS, Requires<[HasSSE2]>;
3816 let SchedRW = [WriteFence] in {
3817 // Load, store, and memory fence
3818 def SFENCE : I<0xAE, MRM_F8, (outs), (ins),
3819 "sfence", [(int_x86_sse_sfence)], IIC_SSE_SFENCE>,
3820 TB, Requires<[HasSSE1]>;
3821 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
3822 "lfence", [(int_x86_sse2_lfence)], IIC_SSE_LFENCE>,
3823 TB, Requires<[HasSSE2]>;
3824 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
3825 "mfence", [(int_x86_sse2_mfence)], IIC_SSE_MFENCE>,
3826 TB, Requires<[HasSSE2]>;
3829 def : Pat<(X86SFence), (SFENCE)>;
3830 def : Pat<(X86LFence), (LFENCE)>;
3831 def : Pat<(X86MFence), (MFENCE)>;
3833 //===----------------------------------------------------------------------===//
3834 // SSE 1 & 2 - Load/Store XCSR register
3835 //===----------------------------------------------------------------------===//
3837 def VLDMXCSR : VPSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
3838 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)],
3839 IIC_SSE_LDMXCSR>, VEX, Sched<[WriteLoad]>;
3840 def VSTMXCSR : VPSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
3841 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)],
3842 IIC_SSE_STMXCSR>, VEX, Sched<[WriteStore]>;
3844 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
3845 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)],
3846 IIC_SSE_LDMXCSR>, Sched<[WriteLoad]>;
3847 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
3848 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)],
3849 IIC_SSE_STMXCSR>, Sched<[WriteStore]>;
3851 //===---------------------------------------------------------------------===//
3852 // SSE2 - Move Aligned/Unaligned Packed Integer Instructions
3853 //===---------------------------------------------------------------------===//
3855 let ExeDomain = SSEPackedInt in { // SSE integer instructions
3857 let neverHasSideEffects = 1, SchedRW = [WriteMove] in {
3858 def VMOVDQArr : VPDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3859 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>,
3861 def VMOVDQAYrr : VPDI<0x6F, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3862 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>,
3864 def VMOVDQUrr : VSSI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3865 "movdqu\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVU_P_RR>,
3867 def VMOVDQUYrr : VSSI<0x6F, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3868 "movdqu\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVU_P_RR>,
3873 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
3874 SchedRW = [WriteMove] in {
3875 def VMOVDQArr_REV : VPDI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3876 "movdqa\t{$src, $dst|$dst, $src}", [],
3879 def VMOVDQAYrr_REV : VPDI<0x7F, MRMDestReg, (outs VR256:$dst), (ins VR256:$src),
3880 "movdqa\t{$src, $dst|$dst, $src}", [],
3881 IIC_SSE_MOVA_P_RR>, VEX, VEX_L;
3882 def VMOVDQUrr_REV : VSSI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3883 "movdqu\t{$src, $dst|$dst, $src}", [],
3886 def VMOVDQUYrr_REV : VSSI<0x7F, MRMDestReg, (outs VR256:$dst), (ins VR256:$src),
3887 "movdqu\t{$src, $dst|$dst, $src}", [],
3888 IIC_SSE_MOVU_P_RR>, VEX, VEX_L;
3891 let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1,
3892 neverHasSideEffects = 1, SchedRW = [WriteLoad] in {
3893 def VMOVDQArm : VPDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3894 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RM>,
3896 def VMOVDQAYrm : VPDI<0x6F, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
3897 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RM>,
3899 let Predicates = [HasAVX] in {
3900 def VMOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3901 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_RM>,
3903 def VMOVDQUYrm : I<0x6F, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
3904 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_RM>,
3909 let mayStore = 1, neverHasSideEffects = 1, SchedRW = [WriteStore] in {
3910 def VMOVDQAmr : VPDI<0x7F, MRMDestMem, (outs),
3911 (ins i128mem:$dst, VR128:$src),
3912 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_MR>,
3914 def VMOVDQAYmr : VPDI<0x7F, MRMDestMem, (outs),
3915 (ins i256mem:$dst, VR256:$src),
3916 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_MR>,
3918 let Predicates = [HasAVX] in {
3919 def VMOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3920 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_MR>,
3922 def VMOVDQUYmr : I<0x7F, MRMDestMem, (outs), (ins i256mem:$dst, VR256:$src),
3923 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_MR>,
3928 let SchedRW = [WriteMove] in {
3929 let neverHasSideEffects = 1 in
3930 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3931 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>;
3933 def MOVDQUrr : I<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3934 "movdqu\t{$src, $dst|$dst, $src}",
3935 [], IIC_SSE_MOVU_P_RR>, XS, Requires<[UseSSE2]>;
3938 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in {
3939 def MOVDQArr_REV : PDI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3940 "movdqa\t{$src, $dst|$dst, $src}", [],
3943 def MOVDQUrr_REV : I<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3944 "movdqu\t{$src, $dst|$dst, $src}",
3945 [], IIC_SSE_MOVU_P_RR>, XS, Requires<[UseSSE2]>;
3949 let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1,
3950 neverHasSideEffects = 1, SchedRW = [WriteLoad] in {
3951 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3952 "movdqa\t{$src, $dst|$dst, $src}",
3953 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/],
3955 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3956 "movdqu\t{$src, $dst|$dst, $src}",
3957 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/],
3959 XS, Requires<[UseSSE2]>;
3962 let mayStore = 1, neverHasSideEffects = 1, SchedRW = [WriteStore] in {
3963 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3964 "movdqa\t{$src, $dst|$dst, $src}",
3965 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/],
3967 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3968 "movdqu\t{$src, $dst|$dst, $src}",
3969 [/*(store (v2i64 VR128:$src), addr:$dst)*/],
3971 XS, Requires<[UseSSE2]>;
3974 } // ExeDomain = SSEPackedInt
3976 let Predicates = [HasAVX] in {
3977 def : Pat<(int_x86_sse2_storeu_dq addr:$dst, VR128:$src),
3978 (VMOVDQUmr addr:$dst, VR128:$src)>;
3979 def : Pat<(int_x86_avx_storeu_dq_256 addr:$dst, VR256:$src),
3980 (VMOVDQUYmr addr:$dst, VR256:$src)>;
3982 let Predicates = [UseSSE2] in
3983 def : Pat<(int_x86_sse2_storeu_dq addr:$dst, VR128:$src),
3984 (MOVDQUmr addr:$dst, VR128:$src)>;
3986 //===---------------------------------------------------------------------===//
3987 // SSE2 - Packed Integer Arithmetic Instructions
3988 //===---------------------------------------------------------------------===//
3990 let Sched = WriteVecIMul in
3991 def SSE_PMADD : OpndItins<
3992 IIC_SSE_PMADD, IIC_SSE_PMADD
3995 let ExeDomain = SSEPackedInt in { // SSE integer instructions
3997 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
3998 RegisterClass RC, PatFrag memop_frag,
3999 X86MemOperand x86memop,
4001 bit IsCommutable = 0,
4003 let isCommutable = IsCommutable in
4004 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
4005 (ins RC:$src1, RC:$src2),
4007 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4008 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4009 [(set RC:$dst, (IntId RC:$src1, RC:$src2))], itins.rr>,
4010 Sched<[itins.Sched]>;
4011 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
4012 (ins RC:$src1, x86memop:$src2),
4014 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4015 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4016 [(set RC:$dst, (IntId RC:$src1, (bitconvert (memop_frag addr:$src2))))],
4017 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
4020 multiclass PDI_binop_all_int<bits<8> opc, string OpcodeStr, Intrinsic IntId128,
4021 Intrinsic IntId256, OpndItins itins,
4022 bit IsCommutable = 0> {
4023 let Predicates = [HasAVX] in
4024 defm V#NAME : PDI_binop_rm_int<opc, !strconcat("v", OpcodeStr), IntId128,
4025 VR128, loadv2i64, i128mem, itins,
4026 IsCommutable, 0>, VEX_4V;
4028 let Constraints = "$src1 = $dst" in
4029 defm NAME : PDI_binop_rm_int<opc, OpcodeStr, IntId128, VR128, memopv2i64,
4030 i128mem, itins, IsCommutable, 1>;
4032 let Predicates = [HasAVX2] in
4033 defm V#NAME#Y : PDI_binop_rm_int<opc, !strconcat("v", OpcodeStr), IntId256,
4034 VR256, loadv4i64, i256mem, itins,
4035 IsCommutable, 0>, VEX_4V, VEX_L;
4038 multiclass PDI_binop_rmi<bits<8> opc, bits<8> opc2, Format ImmForm,
4039 string OpcodeStr, SDNode OpNode,
4040 SDNode OpNode2, RegisterClass RC,
4041 ValueType DstVT, ValueType SrcVT, PatFrag bc_frag,
4042 ShiftOpndItins itins,
4044 // src2 is always 128-bit
4045 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
4046 (ins RC:$src1, VR128:$src2),
4048 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4049 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4050 [(set RC:$dst, (DstVT (OpNode RC:$src1, (SrcVT VR128:$src2))))],
4051 itins.rr>, Sched<[WriteVecShift]>;
4052 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
4053 (ins RC:$src1, i128mem:$src2),
4055 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4056 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4057 [(set RC:$dst, (DstVT (OpNode RC:$src1,
4058 (bc_frag (memopv2i64 addr:$src2)))))], itins.rm>,
4059 Sched<[WriteVecShiftLd, ReadAfterLd]>;
4060 def ri : PDIi8<opc2, ImmForm, (outs RC:$dst),
4061 (ins RC:$src1, i8imm:$src2),
4063 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4064 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4065 [(set RC:$dst, (DstVT (OpNode2 RC:$src1, (i8 imm:$src2))))], itins.ri>,
4066 Sched<[WriteVecShift]>;
4069 /// PDI_binop_rm2 - Simple SSE2 binary operator with different src and dst types
4070 multiclass PDI_binop_rm2<bits<8> opc, string OpcodeStr, SDNode OpNode,
4071 ValueType DstVT, ValueType SrcVT, RegisterClass RC,
4072 PatFrag memop_frag, X86MemOperand x86memop,
4074 bit IsCommutable = 0, bit Is2Addr = 1> {
4075 let isCommutable = IsCommutable in
4076 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
4077 (ins RC:$src1, RC:$src2),
4079 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4080 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4081 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1), RC:$src2)))]>,
4082 Sched<[itins.Sched]>;
4083 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
4084 (ins RC:$src1, x86memop:$src2),
4086 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4087 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4088 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1),
4089 (bitconvert (memop_frag addr:$src2)))))]>,
4090 Sched<[itins.Sched.Folded, ReadAfterLd]>;
4092 } // ExeDomain = SSEPackedInt
4094 defm PADDB : PDI_binop_all<0xFC, "paddb", add, v16i8, v32i8,
4095 SSE_INTALU_ITINS_P, 1>;
4096 defm PADDW : PDI_binop_all<0xFD, "paddw", add, v8i16, v16i16,
4097 SSE_INTALU_ITINS_P, 1>;
4098 defm PADDD : PDI_binop_all<0xFE, "paddd", add, v4i32, v8i32,
4099 SSE_INTALU_ITINS_P, 1>;
4100 defm PADDQ : PDI_binop_all<0xD4, "paddq", add, v2i64, v4i64,
4101 SSE_INTALUQ_ITINS_P, 1>;
4102 defm PMULLW : PDI_binop_all<0xD5, "pmullw", mul, v8i16, v16i16,
4103 SSE_INTMUL_ITINS_P, 1>;
4104 defm PMULHUW : PDI_binop_all<0xE4, "pmulhuw", mulhu, v8i16, v16i16,
4105 SSE_INTMUL_ITINS_P, 1>;
4106 defm PMULHW : PDI_binop_all<0xE5, "pmulhw", mulhs, v8i16, v16i16,
4107 SSE_INTMUL_ITINS_P, 1>;
4108 defm PSUBB : PDI_binop_all<0xF8, "psubb", sub, v16i8, v32i8,
4109 SSE_INTALU_ITINS_P, 0>;
4110 defm PSUBW : PDI_binop_all<0xF9, "psubw", sub, v8i16, v16i16,
4111 SSE_INTALU_ITINS_P, 0>;
4112 defm PSUBD : PDI_binop_all<0xFA, "psubd", sub, v4i32, v8i32,
4113 SSE_INTALU_ITINS_P, 0>;
4114 defm PSUBQ : PDI_binop_all<0xFB, "psubq", sub, v2i64, v4i64,
4115 SSE_INTALUQ_ITINS_P, 0>;
4116 defm PSUBUSB : PDI_binop_all<0xD8, "psubusb", X86subus, v16i8, v32i8,
4117 SSE_INTALU_ITINS_P, 0>;
4118 defm PSUBUSW : PDI_binop_all<0xD9, "psubusw", X86subus, v8i16, v16i16,
4119 SSE_INTALU_ITINS_P, 0>;
4120 defm PMINUB : PDI_binop_all<0xDA, "pminub", X86umin, v16i8, v32i8,
4121 SSE_INTALU_ITINS_P, 1>;
4122 defm PMINSW : PDI_binop_all<0xEA, "pminsw", X86smin, v8i16, v16i16,
4123 SSE_INTALU_ITINS_P, 1>;
4124 defm PMAXUB : PDI_binop_all<0xDE, "pmaxub", X86umax, v16i8, v32i8,
4125 SSE_INTALU_ITINS_P, 1>;
4126 defm PMAXSW : PDI_binop_all<0xEE, "pmaxsw", X86smax, v8i16, v16i16,
4127 SSE_INTALU_ITINS_P, 1>;
4130 defm PSUBSB : PDI_binop_all_int<0xE8, "psubsb", int_x86_sse2_psubs_b,
4131 int_x86_avx2_psubs_b, SSE_INTALU_ITINS_P, 0>;
4132 defm PSUBSW : PDI_binop_all_int<0xE9, "psubsw" , int_x86_sse2_psubs_w,
4133 int_x86_avx2_psubs_w, SSE_INTALU_ITINS_P, 0>;
4134 defm PADDSB : PDI_binop_all_int<0xEC, "paddsb" , int_x86_sse2_padds_b,
4135 int_x86_avx2_padds_b, SSE_INTALU_ITINS_P, 1>;
4136 defm PADDSW : PDI_binop_all_int<0xED, "paddsw" , int_x86_sse2_padds_w,
4137 int_x86_avx2_padds_w, SSE_INTALU_ITINS_P, 1>;
4138 defm PADDUSB : PDI_binop_all_int<0xDC, "paddusb", int_x86_sse2_paddus_b,
4139 int_x86_avx2_paddus_b, SSE_INTALU_ITINS_P, 1>;
4140 defm PADDUSW : PDI_binop_all_int<0xDD, "paddusw", int_x86_sse2_paddus_w,
4141 int_x86_avx2_paddus_w, SSE_INTALU_ITINS_P, 1>;
4142 defm PMADDWD : PDI_binop_all_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd,
4143 int_x86_avx2_pmadd_wd, SSE_PMADD, 1>;
4144 defm PAVGB : PDI_binop_all_int<0xE0, "pavgb", int_x86_sse2_pavg_b,
4145 int_x86_avx2_pavg_b, SSE_INTALU_ITINS_P, 1>;
4146 defm PAVGW : PDI_binop_all_int<0xE3, "pavgw", int_x86_sse2_pavg_w,
4147 int_x86_avx2_pavg_w, SSE_INTALU_ITINS_P, 1>;
4148 defm PSADBW : PDI_binop_all_int<0xF6, "psadbw", int_x86_sse2_psad_bw,
4149 int_x86_avx2_psad_bw, SSE_PMADD, 1>;
4151 let Predicates = [HasAVX] in
4152 defm VPMULUDQ : PDI_binop_rm2<0xF4, "vpmuludq", X86pmuludq, v2i64, v4i32, VR128,
4153 loadv2i64, i128mem, SSE_INTMUL_ITINS_P, 1, 0>,
4155 let Predicates = [HasAVX2] in
4156 defm VPMULUDQY : PDI_binop_rm2<0xF4, "vpmuludq", X86pmuludq, v4i64, v8i32,
4157 VR256, loadv4i64, i256mem,
4158 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V, VEX_L;
4159 let Constraints = "$src1 = $dst" in
4160 defm PMULUDQ : PDI_binop_rm2<0xF4, "pmuludq", X86pmuludq, v2i64, v4i32, VR128,
4161 memopv2i64, i128mem, SSE_INTMUL_ITINS_P, 1>;
4163 //===---------------------------------------------------------------------===//
4164 // SSE2 - Packed Integer Logical Instructions
4165 //===---------------------------------------------------------------------===//
4167 let Predicates = [HasAVX] in {
4168 defm VPSLLW : PDI_binop_rmi<0xF1, 0x71, MRM6r, "vpsllw", X86vshl, X86vshli,
4169 VR128, v8i16, v8i16, bc_v8i16,
4170 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4171 defm VPSLLD : PDI_binop_rmi<0xF2, 0x72, MRM6r, "vpslld", X86vshl, X86vshli,
4172 VR128, v4i32, v4i32, bc_v4i32,
4173 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4174 defm VPSLLQ : PDI_binop_rmi<0xF3, 0x73, MRM6r, "vpsllq", X86vshl, X86vshli,
4175 VR128, v2i64, v2i64, bc_v2i64,
4176 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4178 defm VPSRLW : PDI_binop_rmi<0xD1, 0x71, MRM2r, "vpsrlw", X86vsrl, X86vsrli,
4179 VR128, v8i16, v8i16, bc_v8i16,
4180 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4181 defm VPSRLD : PDI_binop_rmi<0xD2, 0x72, MRM2r, "vpsrld", X86vsrl, X86vsrli,
4182 VR128, v4i32, v4i32, bc_v4i32,
4183 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4184 defm VPSRLQ : PDI_binop_rmi<0xD3, 0x73, MRM2r, "vpsrlq", X86vsrl, X86vsrli,
4185 VR128, v2i64, v2i64, bc_v2i64,
4186 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4188 defm VPSRAW : PDI_binop_rmi<0xE1, 0x71, MRM4r, "vpsraw", X86vsra, X86vsrai,
4189 VR128, v8i16, v8i16, bc_v8i16,
4190 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4191 defm VPSRAD : PDI_binop_rmi<0xE2, 0x72, MRM4r, "vpsrad", X86vsra, X86vsrai,
4192 VR128, v4i32, v4i32, bc_v4i32,
4193 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4195 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift] in {
4196 // 128-bit logical shifts.
4197 def VPSLLDQri : PDIi8<0x73, MRM7r,
4198 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
4199 "vpslldq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4201 (int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2))]>,
4203 def VPSRLDQri : PDIi8<0x73, MRM3r,
4204 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
4205 "vpsrldq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4207 (int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2))]>,
4209 // PSRADQri doesn't exist in SSE[1-3].
4211 } // Predicates = [HasAVX]
4213 let Predicates = [HasAVX2] in {
4214 defm VPSLLWY : PDI_binop_rmi<0xF1, 0x71, MRM6r, "vpsllw", X86vshl, X86vshli,
4215 VR256, v16i16, v8i16, bc_v8i16,
4216 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4217 defm VPSLLDY : PDI_binop_rmi<0xF2, 0x72, MRM6r, "vpslld", X86vshl, X86vshli,
4218 VR256, v8i32, v4i32, bc_v4i32,
4219 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4220 defm VPSLLQY : PDI_binop_rmi<0xF3, 0x73, MRM6r, "vpsllq", X86vshl, X86vshli,
4221 VR256, v4i64, v2i64, bc_v2i64,
4222 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4224 defm VPSRLWY : PDI_binop_rmi<0xD1, 0x71, MRM2r, "vpsrlw", X86vsrl, X86vsrli,
4225 VR256, v16i16, v8i16, bc_v8i16,
4226 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4227 defm VPSRLDY : PDI_binop_rmi<0xD2, 0x72, MRM2r, "vpsrld", X86vsrl, X86vsrli,
4228 VR256, v8i32, v4i32, bc_v4i32,
4229 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4230 defm VPSRLQY : PDI_binop_rmi<0xD3, 0x73, MRM2r, "vpsrlq", X86vsrl, X86vsrli,
4231 VR256, v4i64, v2i64, bc_v2i64,
4232 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4234 defm VPSRAWY : PDI_binop_rmi<0xE1, 0x71, MRM4r, "vpsraw", X86vsra, X86vsrai,
4235 VR256, v16i16, v8i16, bc_v8i16,
4236 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4237 defm VPSRADY : PDI_binop_rmi<0xE2, 0x72, MRM4r, "vpsrad", X86vsra, X86vsrai,
4238 VR256, v8i32, v4i32, bc_v4i32,
4239 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4241 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift] in {
4242 // 256-bit logical shifts.
4243 def VPSLLDQYri : PDIi8<0x73, MRM7r,
4244 (outs VR256:$dst), (ins VR256:$src1, i32i8imm:$src2),
4245 "vpslldq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4247 (int_x86_avx2_psll_dq_bs VR256:$src1, imm:$src2))]>,
4249 def VPSRLDQYri : PDIi8<0x73, MRM3r,
4250 (outs VR256:$dst), (ins VR256:$src1, i32i8imm:$src2),
4251 "vpsrldq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4253 (int_x86_avx2_psrl_dq_bs VR256:$src1, imm:$src2))]>,
4255 // PSRADQYri doesn't exist in SSE[1-3].
4257 } // Predicates = [HasAVX2]
4259 let Constraints = "$src1 = $dst" in {
4260 defm PSLLW : PDI_binop_rmi<0xF1, 0x71, MRM6r, "psllw", X86vshl, X86vshli,
4261 VR128, v8i16, v8i16, bc_v8i16,
4262 SSE_INTSHIFT_ITINS_P>;
4263 defm PSLLD : PDI_binop_rmi<0xF2, 0x72, MRM6r, "pslld", X86vshl, X86vshli,
4264 VR128, v4i32, v4i32, bc_v4i32,
4265 SSE_INTSHIFT_ITINS_P>;
4266 defm PSLLQ : PDI_binop_rmi<0xF3, 0x73, MRM6r, "psllq", X86vshl, X86vshli,
4267 VR128, v2i64, v2i64, bc_v2i64,
4268 SSE_INTSHIFT_ITINS_P>;
4270 defm PSRLW : PDI_binop_rmi<0xD1, 0x71, MRM2r, "psrlw", X86vsrl, X86vsrli,
4271 VR128, v8i16, v8i16, bc_v8i16,
4272 SSE_INTSHIFT_ITINS_P>;
4273 defm PSRLD : PDI_binop_rmi<0xD2, 0x72, MRM2r, "psrld", X86vsrl, X86vsrli,
4274 VR128, v4i32, v4i32, bc_v4i32,
4275 SSE_INTSHIFT_ITINS_P>;
4276 defm PSRLQ : PDI_binop_rmi<0xD3, 0x73, MRM2r, "psrlq", X86vsrl, X86vsrli,
4277 VR128, v2i64, v2i64, bc_v2i64,
4278 SSE_INTSHIFT_ITINS_P>;
4280 defm PSRAW : PDI_binop_rmi<0xE1, 0x71, MRM4r, "psraw", X86vsra, X86vsrai,
4281 VR128, v8i16, v8i16, bc_v8i16,
4282 SSE_INTSHIFT_ITINS_P>;
4283 defm PSRAD : PDI_binop_rmi<0xE2, 0x72, MRM4r, "psrad", X86vsra, X86vsrai,
4284 VR128, v4i32, v4i32, bc_v4i32,
4285 SSE_INTSHIFT_ITINS_P>;
4287 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift] in {
4288 // 128-bit logical shifts.
4289 def PSLLDQri : PDIi8<0x73, MRM7r,
4290 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
4291 "pslldq\t{$src2, $dst|$dst, $src2}",
4293 (int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2))],
4294 IIC_SSE_INTSHDQ_P_RI>;
4295 def PSRLDQri : PDIi8<0x73, MRM3r,
4296 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
4297 "psrldq\t{$src2, $dst|$dst, $src2}",
4299 (int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2))],
4300 IIC_SSE_INTSHDQ_P_RI>;
4301 // PSRADQri doesn't exist in SSE[1-3].
4303 } // Constraints = "$src1 = $dst"
4305 let Predicates = [HasAVX] in {
4306 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
4307 (VPSLLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4308 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
4309 (VPSRLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4310 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
4311 (VPSRLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4313 // Shift up / down and insert zero's.
4314 def : Pat<(v2i64 (X86vshldq VR128:$src, (i8 imm:$amt))),
4315 (VPSLLDQri VR128:$src, (BYTE_imm imm:$amt))>;
4316 def : Pat<(v2i64 (X86vshrdq VR128:$src, (i8 imm:$amt))),
4317 (VPSRLDQri VR128:$src, (BYTE_imm imm:$amt))>;
4320 let Predicates = [HasAVX2] in {
4321 def : Pat<(int_x86_avx2_psll_dq VR256:$src1, imm:$src2),
4322 (VPSLLDQYri VR256:$src1, (BYTE_imm imm:$src2))>;
4323 def : Pat<(int_x86_avx2_psrl_dq VR256:$src1, imm:$src2),
4324 (VPSRLDQYri VR256:$src1, (BYTE_imm imm:$src2))>;
4327 let Predicates = [UseSSE2] in {
4328 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
4329 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4330 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
4331 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4332 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
4333 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4335 // Shift up / down and insert zero's.
4336 def : Pat<(v2i64 (X86vshldq VR128:$src, (i8 imm:$amt))),
4337 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt))>;
4338 def : Pat<(v2i64 (X86vshrdq VR128:$src, (i8 imm:$amt))),
4339 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt))>;
4342 //===---------------------------------------------------------------------===//
4343 // SSE2 - Packed Integer Comparison Instructions
4344 //===---------------------------------------------------------------------===//
4346 defm PCMPEQB : PDI_binop_all<0x74, "pcmpeqb", X86pcmpeq, v16i8, v32i8,
4347 SSE_INTALU_ITINS_P, 1>;
4348 defm PCMPEQW : PDI_binop_all<0x75, "pcmpeqw", X86pcmpeq, v8i16, v16i16,
4349 SSE_INTALU_ITINS_P, 1>;
4350 defm PCMPEQD : PDI_binop_all<0x76, "pcmpeqd", X86pcmpeq, v4i32, v8i32,
4351 SSE_INTALU_ITINS_P, 1>;
4352 defm PCMPGTB : PDI_binop_all<0x64, "pcmpgtb", X86pcmpgt, v16i8, v32i8,
4353 SSE_INTALU_ITINS_P, 0>;
4354 defm PCMPGTW : PDI_binop_all<0x65, "pcmpgtw", X86pcmpgt, v8i16, v16i16,
4355 SSE_INTALU_ITINS_P, 0>;
4356 defm PCMPGTD : PDI_binop_all<0x66, "pcmpgtd", X86pcmpgt, v4i32, v8i32,
4357 SSE_INTALU_ITINS_P, 0>;
4359 //===---------------------------------------------------------------------===//
4360 // SSE2 - Packed Integer Shuffle Instructions
4361 //===---------------------------------------------------------------------===//
4363 let ExeDomain = SSEPackedInt in {
4364 multiclass sse2_pshuffle<string OpcodeStr, ValueType vt128, ValueType vt256,
4366 let Predicates = [HasAVX] in {
4367 def V#NAME#ri : Ii8<0x70, MRMSrcReg, (outs VR128:$dst),
4368 (ins VR128:$src1, i8imm:$src2),
4369 !strconcat("v", OpcodeStr,
4370 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4372 (vt128 (OpNode VR128:$src1, (i8 imm:$src2))))],
4373 IIC_SSE_PSHUF_RI>, VEX, Sched<[WriteShuffle]>;
4374 def V#NAME#mi : Ii8<0x70, MRMSrcMem, (outs VR128:$dst),
4375 (ins i128mem:$src1, i8imm:$src2),
4376 !strconcat("v", OpcodeStr,
4377 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4379 (vt128 (OpNode (bitconvert (loadv2i64 addr:$src1)),
4380 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>, VEX,
4381 Sched<[WriteShuffleLd]>;
4384 let Predicates = [HasAVX2] in {
4385 def V#NAME#Yri : Ii8<0x70, MRMSrcReg, (outs VR256:$dst),
4386 (ins VR256:$src1, i8imm:$src2),
4387 !strconcat("v", OpcodeStr,
4388 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4390 (vt256 (OpNode VR256:$src1, (i8 imm:$src2))))],
4391 IIC_SSE_PSHUF_RI>, VEX, VEX_L, Sched<[WriteShuffle]>;
4392 def V#NAME#Ymi : Ii8<0x70, MRMSrcMem, (outs VR256:$dst),
4393 (ins i256mem:$src1, i8imm:$src2),
4394 !strconcat("v", OpcodeStr,
4395 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4397 (vt256 (OpNode (bitconvert (loadv4i64 addr:$src1)),
4398 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>, VEX, VEX_L,
4399 Sched<[WriteShuffleLd]>;
4402 let Predicates = [UseSSE2] in {
4403 def ri : Ii8<0x70, MRMSrcReg,
4404 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
4405 !strconcat(OpcodeStr,
4406 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4408 (vt128 (OpNode VR128:$src1, (i8 imm:$src2))))],
4409 IIC_SSE_PSHUF_RI>, Sched<[WriteShuffle]>;
4410 def mi : Ii8<0x70, MRMSrcMem,
4411 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
4412 !strconcat(OpcodeStr,
4413 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4415 (vt128 (OpNode (bitconvert (memopv2i64 addr:$src1)),
4416 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>,
4417 Sched<[WriteShuffleLd, ReadAfterLd]>;
4420 } // ExeDomain = SSEPackedInt
4422 defm PSHUFD : sse2_pshuffle<"pshufd", v4i32, v8i32, X86PShufd>, PD;
4423 defm PSHUFHW : sse2_pshuffle<"pshufhw", v8i16, v16i16, X86PShufhw>, XS;
4424 defm PSHUFLW : sse2_pshuffle<"pshuflw", v8i16, v16i16, X86PShuflw>, XD;
4426 let Predicates = [HasAVX] in {
4427 def : Pat<(v4f32 (X86PShufd (loadv4f32 addr:$src1), (i8 imm:$imm))),
4428 (VPSHUFDmi addr:$src1, imm:$imm)>;
4429 def : Pat<(v4f32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
4430 (VPSHUFDri VR128:$src1, imm:$imm)>;
4433 let Predicates = [UseSSE2] in {
4434 def : Pat<(v4f32 (X86PShufd (memopv4f32 addr:$src1), (i8 imm:$imm))),
4435 (PSHUFDmi addr:$src1, imm:$imm)>;
4436 def : Pat<(v4f32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
4437 (PSHUFDri VR128:$src1, imm:$imm)>;
4440 //===---------------------------------------------------------------------===//
4441 // Packed Integer Pack Instructions (SSE & AVX)
4442 //===---------------------------------------------------------------------===//
4444 let ExeDomain = SSEPackedInt in {
4445 multiclass sse2_pack<bits<8> opc, string OpcodeStr, ValueType OutVT,
4446 ValueType ArgVT, SDNode OpNode, PatFrag bc_frag,
4448 def rr : PDI<opc, MRMSrcReg,
4449 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4451 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4452 !strconcat(OpcodeStr,
4453 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4455 (OutVT (OpNode (ArgVT VR128:$src1), VR128:$src2)))]>,
4456 Sched<[WriteShuffle]>;
4457 def rm : PDI<opc, MRMSrcMem,
4458 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4460 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4461 !strconcat(OpcodeStr,
4462 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4464 (OutVT (OpNode VR128:$src1,
4465 (bc_frag (memopv2i64 addr:$src2)))))]>,
4466 Sched<[WriteShuffleLd, ReadAfterLd]>;
4469 multiclass sse2_pack_y<bits<8> opc, string OpcodeStr, ValueType OutVT,
4470 ValueType ArgVT, SDNode OpNode, PatFrag bc_frag> {
4471 def Yrr : PDI<opc, MRMSrcReg,
4472 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4473 !strconcat(OpcodeStr,
4474 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4476 (OutVT (OpNode (ArgVT VR256:$src1), VR256:$src2)))]>,
4477 Sched<[WriteShuffle]>;
4478 def Yrm : PDI<opc, MRMSrcMem,
4479 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4480 !strconcat(OpcodeStr,
4481 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4483 (OutVT (OpNode VR256:$src1,
4484 (bc_frag (memopv4i64 addr:$src2)))))]>,
4485 Sched<[WriteShuffleLd, ReadAfterLd]>;
4488 multiclass sse4_pack<bits<8> opc, string OpcodeStr, ValueType OutVT,
4489 ValueType ArgVT, SDNode OpNode, PatFrag bc_frag,
4491 def rr : SS48I<opc, MRMSrcReg,
4492 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4494 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4495 !strconcat(OpcodeStr,
4496 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4498 (OutVT (OpNode (ArgVT VR128:$src1), VR128:$src2)))]>,
4499 Sched<[WriteShuffle]>;
4500 def rm : SS48I<opc, MRMSrcMem,
4501 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4503 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4504 !strconcat(OpcodeStr,
4505 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4507 (OutVT (OpNode VR128:$src1,
4508 (bc_frag (memopv2i64 addr:$src2)))))]>,
4509 Sched<[WriteShuffleLd, ReadAfterLd]>;
4512 multiclass sse4_pack_y<bits<8> opc, string OpcodeStr, ValueType OutVT,
4513 ValueType ArgVT, SDNode OpNode, PatFrag bc_frag> {
4514 def Yrr : SS48I<opc, MRMSrcReg,
4515 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4516 !strconcat(OpcodeStr,
4517 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4519 (OutVT (OpNode (ArgVT VR256:$src1), VR256:$src2)))]>,
4520 Sched<[WriteShuffle]>;
4521 def Yrm : SS48I<opc, MRMSrcMem,
4522 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4523 !strconcat(OpcodeStr,
4524 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4526 (OutVT (OpNode VR256:$src1,
4527 (bc_frag (memopv4i64 addr:$src2)))))]>,
4528 Sched<[WriteShuffleLd, ReadAfterLd]>;
4531 let Predicates = [HasAVX] in {
4532 defm VPACKSSWB : sse2_pack<0x63, "vpacksswb", v16i8, v8i16, X86Packss,
4533 bc_v8i16, 0>, VEX_4V;
4534 defm VPACKSSDW : sse2_pack<0x6B, "vpackssdw", v8i16, v4i32, X86Packss,
4535 bc_v4i32, 0>, VEX_4V;
4537 defm VPACKUSWB : sse2_pack<0x67, "vpackuswb", v16i8, v8i16, X86Packus,
4538 bc_v8i16, 0>, VEX_4V;
4539 defm VPACKUSDW : sse4_pack<0x2B, "vpackusdw", v8i16, v4i32, X86Packus,
4540 bc_v4i32, 0>, VEX_4V;
4543 let Predicates = [HasAVX2] in {
4544 defm VPACKSSWB : sse2_pack_y<0x63, "vpacksswb", v32i8, v16i16, X86Packss,
4545 bc_v16i16>, VEX_4V, VEX_L;
4546 defm VPACKSSDW : sse2_pack_y<0x6B, "vpackssdw", v16i16, v8i32, X86Packss,
4547 bc_v8i32>, VEX_4V, VEX_L;
4549 defm VPACKUSWB : sse2_pack_y<0x67, "vpackuswb", v32i8, v16i16, X86Packus,
4550 bc_v16i16>, VEX_4V, VEX_L;
4551 defm VPACKUSDW : sse4_pack_y<0x2B, "vpackusdw", v16i16, v8i32, X86Packus,
4552 bc_v8i32>, VEX_4V, VEX_L;
4555 let Constraints = "$src1 = $dst" in {
4556 defm PACKSSWB : sse2_pack<0x63, "packsswb", v16i8, v8i16, X86Packss,
4558 defm PACKSSDW : sse2_pack<0x6B, "packssdw", v8i16, v4i32, X86Packss,
4561 defm PACKUSWB : sse2_pack<0x67, "packuswb", v16i8, v8i16, X86Packus,
4564 let Predicates = [HasSSE41] in
4565 defm PACKUSDW : sse4_pack<0x2B, "packusdw", v8i16, v4i32, X86Packus,
4568 } // ExeDomain = SSEPackedInt
4570 //===---------------------------------------------------------------------===//
4571 // SSE2 - Packed Integer Unpack Instructions
4572 //===---------------------------------------------------------------------===//
4574 let ExeDomain = SSEPackedInt in {
4575 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
4576 SDNode OpNode, PatFrag bc_frag, bit Is2Addr = 1> {
4577 def rr : PDI<opc, MRMSrcReg,
4578 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4580 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
4581 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4582 [(set VR128:$dst, (vt (OpNode VR128:$src1, VR128:$src2)))],
4583 IIC_SSE_UNPCK>, Sched<[WriteShuffle]>;
4584 def rm : PDI<opc, MRMSrcMem,
4585 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4587 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
4588 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4589 [(set VR128:$dst, (OpNode VR128:$src1,
4590 (bc_frag (memopv2i64
4593 Sched<[WriteShuffleLd, ReadAfterLd]>;
4596 multiclass sse2_unpack_y<bits<8> opc, string OpcodeStr, ValueType vt,
4597 SDNode OpNode, PatFrag bc_frag> {
4598 def Yrr : PDI<opc, MRMSrcReg,
4599 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4600 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4601 [(set VR256:$dst, (vt (OpNode VR256:$src1, VR256:$src2)))]>,
4602 Sched<[WriteShuffle]>;
4603 def Yrm : PDI<opc, MRMSrcMem,
4604 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4605 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4606 [(set VR256:$dst, (OpNode VR256:$src1,
4607 (bc_frag (memopv4i64 addr:$src2))))]>,
4608 Sched<[WriteShuffleLd, ReadAfterLd]>;
4611 let Predicates = [HasAVX] in {
4612 defm VPUNPCKLBW : sse2_unpack<0x60, "vpunpcklbw", v16i8, X86Unpckl,
4613 bc_v16i8, 0>, VEX_4V;
4614 defm VPUNPCKLWD : sse2_unpack<0x61, "vpunpcklwd", v8i16, X86Unpckl,
4615 bc_v8i16, 0>, VEX_4V;
4616 defm VPUNPCKLDQ : sse2_unpack<0x62, "vpunpckldq", v4i32, X86Unpckl,
4617 bc_v4i32, 0>, VEX_4V;
4618 defm VPUNPCKLQDQ : sse2_unpack<0x6C, "vpunpcklqdq", v2i64, X86Unpckl,
4619 bc_v2i64, 0>, VEX_4V;
4621 defm VPUNPCKHBW : sse2_unpack<0x68, "vpunpckhbw", v16i8, X86Unpckh,
4622 bc_v16i8, 0>, VEX_4V;
4623 defm VPUNPCKHWD : sse2_unpack<0x69, "vpunpckhwd", v8i16, X86Unpckh,
4624 bc_v8i16, 0>, VEX_4V;
4625 defm VPUNPCKHDQ : sse2_unpack<0x6A, "vpunpckhdq", v4i32, X86Unpckh,
4626 bc_v4i32, 0>, VEX_4V;
4627 defm VPUNPCKHQDQ : sse2_unpack<0x6D, "vpunpckhqdq", v2i64, X86Unpckh,
4628 bc_v2i64, 0>, VEX_4V;
4631 let Predicates = [HasAVX2] in {
4632 defm VPUNPCKLBW : sse2_unpack_y<0x60, "vpunpcklbw", v32i8, X86Unpckl,
4633 bc_v32i8>, VEX_4V, VEX_L;
4634 defm VPUNPCKLWD : sse2_unpack_y<0x61, "vpunpcklwd", v16i16, X86Unpckl,
4635 bc_v16i16>, VEX_4V, VEX_L;
4636 defm VPUNPCKLDQ : sse2_unpack_y<0x62, "vpunpckldq", v8i32, X86Unpckl,
4637 bc_v8i32>, VEX_4V, VEX_L;
4638 defm VPUNPCKLQDQ : sse2_unpack_y<0x6C, "vpunpcklqdq", v4i64, X86Unpckl,
4639 bc_v4i64>, VEX_4V, VEX_L;
4641 defm VPUNPCKHBW : sse2_unpack_y<0x68, "vpunpckhbw", v32i8, X86Unpckh,
4642 bc_v32i8>, VEX_4V, VEX_L;
4643 defm VPUNPCKHWD : sse2_unpack_y<0x69, "vpunpckhwd", v16i16, X86Unpckh,
4644 bc_v16i16>, VEX_4V, VEX_L;
4645 defm VPUNPCKHDQ : sse2_unpack_y<0x6A, "vpunpckhdq", v8i32, X86Unpckh,
4646 bc_v8i32>, VEX_4V, VEX_L;
4647 defm VPUNPCKHQDQ : sse2_unpack_y<0x6D, "vpunpckhqdq", v4i64, X86Unpckh,
4648 bc_v4i64>, VEX_4V, VEX_L;
4651 let Constraints = "$src1 = $dst" in {
4652 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, X86Unpckl,
4654 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, X86Unpckl,
4656 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, X86Unpckl,
4658 defm PUNPCKLQDQ : sse2_unpack<0x6C, "punpcklqdq", v2i64, X86Unpckl,
4661 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, X86Unpckh,
4663 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, X86Unpckh,
4665 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, X86Unpckh,
4667 defm PUNPCKHQDQ : sse2_unpack<0x6D, "punpckhqdq", v2i64, X86Unpckh,
4670 } // ExeDomain = SSEPackedInt
4672 //===---------------------------------------------------------------------===//
4673 // SSE2 - Packed Integer Extract and Insert
4674 //===---------------------------------------------------------------------===//
4676 let ExeDomain = SSEPackedInt in {
4677 multiclass sse2_pinsrw<bit Is2Addr = 1> {
4678 def rri : Ii8<0xC4, MRMSrcReg,
4679 (outs VR128:$dst), (ins VR128:$src1,
4680 GR32orGR64:$src2, i32i8imm:$src3),
4682 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
4683 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4685 (X86pinsrw VR128:$src1, GR32orGR64:$src2, imm:$src3))],
4686 IIC_SSE_PINSRW>, Sched<[WriteShuffle]>;
4687 def rmi : Ii8<0xC4, MRMSrcMem,
4688 (outs VR128:$dst), (ins VR128:$src1,
4689 i16mem:$src2, i32i8imm:$src3),
4691 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
4692 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4694 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
4695 imm:$src3))], IIC_SSE_PINSRW>,
4696 Sched<[WriteShuffleLd, ReadAfterLd]>;
4700 let Predicates = [HasAVX] in
4701 def VPEXTRWri : Ii8<0xC5, MRMSrcReg,
4702 (outs GR32orGR64:$dst), (ins VR128:$src1, i32i8imm:$src2),
4703 "vpextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4704 [(set GR32orGR64:$dst, (X86pextrw (v8i16 VR128:$src1),
4705 imm:$src2))]>, PD, VEX,
4706 Sched<[WriteShuffle]>;
4707 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
4708 (outs GR32orGR64:$dst), (ins VR128:$src1, i32i8imm:$src2),
4709 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4710 [(set GR32orGR64:$dst, (X86pextrw (v8i16 VR128:$src1),
4711 imm:$src2))], IIC_SSE_PEXTRW>,
4712 Sched<[WriteShuffleLd, ReadAfterLd]>;
4715 let Predicates = [HasAVX] in
4716 defm VPINSRW : sse2_pinsrw<0>, PD, VEX_4V;
4718 let Predicates = [UseSSE2], Constraints = "$src1 = $dst" in
4719 defm PINSRW : sse2_pinsrw, PD;
4721 } // ExeDomain = SSEPackedInt
4723 //===---------------------------------------------------------------------===//
4724 // SSE2 - Packed Mask Creation
4725 //===---------------------------------------------------------------------===//
4727 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecLogic] in {
4729 def VPMOVMSKBrr : VPDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst),
4731 "pmovmskb\t{$src, $dst|$dst, $src}",
4732 [(set GR32orGR64:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))],
4733 IIC_SSE_MOVMSK>, VEX;
4735 let Predicates = [HasAVX2] in {
4736 def VPMOVMSKBYrr : VPDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst),
4738 "pmovmskb\t{$src, $dst|$dst, $src}",
4739 [(set GR32orGR64:$dst, (int_x86_avx2_pmovmskb VR256:$src))]>,
4743 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst), (ins VR128:$src),
4744 "pmovmskb\t{$src, $dst|$dst, $src}",
4745 [(set GR32orGR64:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))],
4748 } // ExeDomain = SSEPackedInt
4750 //===---------------------------------------------------------------------===//
4751 // SSE2 - Conditional Store
4752 //===---------------------------------------------------------------------===//
4754 let ExeDomain = SSEPackedInt, SchedRW = [WriteStore] in {
4756 let Uses = [EDI], Predicates = [HasAVX,Not64BitMode] in
4757 def VMASKMOVDQU : VPDI<0xF7, MRMSrcReg, (outs),
4758 (ins VR128:$src, VR128:$mask),
4759 "maskmovdqu\t{$mask, $src|$src, $mask}",
4760 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)],
4761 IIC_SSE_MASKMOV>, VEX;
4762 let Uses = [RDI], Predicates = [HasAVX,In64BitMode] in
4763 def VMASKMOVDQU64 : VPDI<0xF7, MRMSrcReg, (outs),
4764 (ins VR128:$src, VR128:$mask),
4765 "maskmovdqu\t{$mask, $src|$src, $mask}",
4766 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)],
4767 IIC_SSE_MASKMOV>, VEX;
4769 let Uses = [EDI], Predicates = [UseSSE2,Not64BitMode] in
4770 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
4771 "maskmovdqu\t{$mask, $src|$src, $mask}",
4772 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)],
4774 let Uses = [RDI], Predicates = [UseSSE2,In64BitMode] in
4775 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
4776 "maskmovdqu\t{$mask, $src|$src, $mask}",
4777 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)],
4780 } // ExeDomain = SSEPackedInt
4782 //===---------------------------------------------------------------------===//
4783 // SSE2 - Move Doubleword
4784 //===---------------------------------------------------------------------===//
4786 //===---------------------------------------------------------------------===//
4787 // Move Int Doubleword to Packed Double Int
4789 def VMOVDI2PDIrr : VS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
4790 "movd\t{$src, $dst|$dst, $src}",
4792 (v4i32 (scalar_to_vector GR32:$src)))], IIC_SSE_MOVDQ>,
4793 VEX, Sched<[WriteMove]>;
4794 def VMOVDI2PDIrm : VS2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
4795 "movd\t{$src, $dst|$dst, $src}",
4797 (v4i32 (scalar_to_vector (loadi32 addr:$src))))],
4799 VEX, Sched<[WriteLoad]>;
4800 def VMOV64toPQIrr : VRS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4801 "movq\t{$src, $dst|$dst, $src}",
4803 (v2i64 (scalar_to_vector GR64:$src)))],
4804 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4805 let isCodeGenOnly = 1 in
4806 def VMOV64toSDrr : VRS2I<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
4807 "movq\t{$src, $dst|$dst, $src}",
4808 [(set FR64:$dst, (bitconvert GR64:$src))],
4809 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4811 def MOVDI2PDIrr : S2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
4812 "movd\t{$src, $dst|$dst, $src}",
4814 (v4i32 (scalar_to_vector GR32:$src)))], IIC_SSE_MOVDQ>,
4816 def MOVDI2PDIrm : S2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
4817 "movd\t{$src, $dst|$dst, $src}",
4819 (v4i32 (scalar_to_vector (loadi32 addr:$src))))],
4820 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4821 def MOV64toPQIrr : RS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4822 "mov{d|q}\t{$src, $dst|$dst, $src}",
4824 (v2i64 (scalar_to_vector GR64:$src)))],
4825 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4826 let isCodeGenOnly = 1 in
4827 def MOV64toSDrr : RS2I<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
4828 "mov{d|q}\t{$src, $dst|$dst, $src}",
4829 [(set FR64:$dst, (bitconvert GR64:$src))],
4830 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4832 //===---------------------------------------------------------------------===//
4833 // Move Int Doubleword to Single Scalar
4835 let isCodeGenOnly = 1 in {
4836 def VMOVDI2SSrr : VS2I<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
4837 "movd\t{$src, $dst|$dst, $src}",
4838 [(set FR32:$dst, (bitconvert GR32:$src))],
4839 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4841 def VMOVDI2SSrm : VS2I<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
4842 "movd\t{$src, $dst|$dst, $src}",
4843 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))],
4845 VEX, Sched<[WriteLoad]>;
4846 def MOVDI2SSrr : S2I<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
4847 "movd\t{$src, $dst|$dst, $src}",
4848 [(set FR32:$dst, (bitconvert GR32:$src))],
4849 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4851 def MOVDI2SSrm : S2I<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
4852 "movd\t{$src, $dst|$dst, $src}",
4853 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))],
4854 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4857 //===---------------------------------------------------------------------===//
4858 // Move Packed Doubleword Int to Packed Double Int
4860 def VMOVPDI2DIrr : VS2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
4861 "movd\t{$src, $dst|$dst, $src}",
4862 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
4863 (iPTR 0)))], IIC_SSE_MOVD_ToGP>, VEX,
4865 def VMOVPDI2DImr : VS2I<0x7E, MRMDestMem, (outs),
4866 (ins i32mem:$dst, VR128:$src),
4867 "movd\t{$src, $dst|$dst, $src}",
4868 [(store (i32 (vector_extract (v4i32 VR128:$src),
4869 (iPTR 0))), addr:$dst)], IIC_SSE_MOVDQ>,
4870 VEX, Sched<[WriteStore]>;
4871 def MOVPDI2DIrr : S2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
4872 "movd\t{$src, $dst|$dst, $src}",
4873 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
4874 (iPTR 0)))], IIC_SSE_MOVD_ToGP>,
4876 def MOVPDI2DImr : S2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
4877 "movd\t{$src, $dst|$dst, $src}",
4878 [(store (i32 (vector_extract (v4i32 VR128:$src),
4879 (iPTR 0))), addr:$dst)],
4880 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4882 def : Pat<(v8i32 (X86Vinsert (v8i32 immAllZerosV), GR32:$src2, (iPTR 0))),
4883 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrr GR32:$src2), sub_xmm)>;
4885 def : Pat<(v4i64 (X86Vinsert (bc_v4i64 (v8i32 immAllZerosV)), GR64:$src2, (iPTR 0))),
4886 (SUBREG_TO_REG (i32 0), (VMOV64toPQIrr GR64:$src2), sub_xmm)>;
4888 def : Pat<(v8i32 (X86Vinsert undef, GR32:$src2, (iPTR 0))),
4889 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrr GR32:$src2), sub_xmm)>;
4891 def : Pat<(v4i64 (X86Vinsert undef, GR64:$src2, (iPTR 0))),
4892 (SUBREG_TO_REG (i32 0), (VMOV64toPQIrr GR64:$src2), sub_xmm)>;
4894 //===---------------------------------------------------------------------===//
4895 // Move Packed Doubleword Int first element to Doubleword Int
4897 let SchedRW = [WriteMove] in {
4898 def VMOVPQIto64rr : VRS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
4899 "movq\t{$src, $dst|$dst, $src}",
4900 [(set GR64:$dst, (vector_extract (v2i64 VR128:$src),
4905 def MOVPQIto64rr : RS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
4906 "mov{d|q}\t{$src, $dst|$dst, $src}",
4907 [(set GR64:$dst, (vector_extract (v2i64 VR128:$src),
4912 //===---------------------------------------------------------------------===//
4913 // Bitcast FR64 <-> GR64
4915 let isCodeGenOnly = 1 in {
4916 let Predicates = [UseAVX] in
4917 def VMOV64toSDrm : VS2SI<0x7E, MRMSrcMem, (outs FR64:$dst), (ins i64mem:$src),
4918 "movq\t{$src, $dst|$dst, $src}",
4919 [(set FR64:$dst, (bitconvert (loadi64 addr:$src)))]>,
4920 VEX, Sched<[WriteLoad]>;
4921 def VMOVSDto64rr : VRS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins FR64:$src),
4922 "movq\t{$src, $dst|$dst, $src}",
4923 [(set GR64:$dst, (bitconvert FR64:$src))],
4924 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4925 def VMOVSDto64mr : VRS2I<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, FR64:$src),
4926 "movq\t{$src, $dst|$dst, $src}",
4927 [(store (i64 (bitconvert FR64:$src)), addr:$dst)],
4928 IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4930 def MOV64toSDrm : S2SI<0x7E, MRMSrcMem, (outs FR64:$dst), (ins i64mem:$src),
4931 "movq\t{$src, $dst|$dst, $src}",
4932 [(set FR64:$dst, (bitconvert (loadi64 addr:$src)))],
4933 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4934 def MOVSDto64rr : RS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins FR64:$src),
4935 "mov{d|q}\t{$src, $dst|$dst, $src}",
4936 [(set GR64:$dst, (bitconvert FR64:$src))],
4937 IIC_SSE_MOVD_ToGP>, Sched<[WriteMove]>;
4938 def MOVSDto64mr : RS2I<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, FR64:$src),
4939 "movq\t{$src, $dst|$dst, $src}",
4940 [(store (i64 (bitconvert FR64:$src)), addr:$dst)],
4941 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4944 //===---------------------------------------------------------------------===//
4945 // Move Scalar Single to Double Int
4947 let isCodeGenOnly = 1 in {
4948 def VMOVSS2DIrr : VS2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
4949 "movd\t{$src, $dst|$dst, $src}",
4950 [(set GR32:$dst, (bitconvert FR32:$src))],
4951 IIC_SSE_MOVD_ToGP>, VEX, Sched<[WriteMove]>;
4952 def VMOVSS2DImr : VS2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
4953 "movd\t{$src, $dst|$dst, $src}",
4954 [(store (i32 (bitconvert FR32:$src)), addr:$dst)],
4955 IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4956 def MOVSS2DIrr : S2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
4957 "movd\t{$src, $dst|$dst, $src}",
4958 [(set GR32:$dst, (bitconvert FR32:$src))],
4959 IIC_SSE_MOVD_ToGP>, Sched<[WriteMove]>;
4960 def MOVSS2DImr : S2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
4961 "movd\t{$src, $dst|$dst, $src}",
4962 [(store (i32 (bitconvert FR32:$src)), addr:$dst)],
4963 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4966 //===---------------------------------------------------------------------===//
4967 // Patterns and instructions to describe movd/movq to XMM register zero-extends
4969 let isCodeGenOnly = 1, SchedRW = [WriteMove] in {
4970 let AddedComplexity = 15 in {
4971 def VMOVZQI2PQIrr : VS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4972 "movq\t{$src, $dst|$dst, $src}", // X86-64 only
4973 [(set VR128:$dst, (v2i64 (X86vzmovl
4974 (v2i64 (scalar_to_vector GR64:$src)))))],
4977 def MOVZQI2PQIrr : RS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4978 "mov{d|q}\t{$src, $dst|$dst, $src}", // X86-64 only
4979 [(set VR128:$dst, (v2i64 (X86vzmovl
4980 (v2i64 (scalar_to_vector GR64:$src)))))],
4983 } // isCodeGenOnly, SchedRW
4985 let Predicates = [UseAVX] in {
4986 let AddedComplexity = 15 in
4987 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector GR32:$src)))),
4988 (VMOVDI2PDIrr GR32:$src)>;
4990 // AVX 128-bit movd/movq instruction write zeros in the high 128-bit part.
4991 let AddedComplexity = 20 in {
4992 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
4993 (VMOVDI2PDIrm addr:$src)>;
4994 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
4995 (VMOVDI2PDIrm addr:$src)>;
4996 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
4997 (VMOVDI2PDIrm addr:$src)>;
4999 // Use regular 128-bit instructions to match 256-bit scalar_to_vec+zext.
5000 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
5001 (v4i32 (scalar_to_vector GR32:$src)),(iPTR 0)))),
5002 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrr GR32:$src), sub_xmm)>;
5003 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
5004 (v2i64 (scalar_to_vector GR64:$src)),(iPTR 0)))),
5005 (SUBREG_TO_REG (i64 0), (VMOVZQI2PQIrr GR64:$src), sub_xmm)>;
5008 let Predicates = [UseSSE2] in {
5009 let AddedComplexity = 15 in
5010 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector GR32:$src)))),
5011 (MOVDI2PDIrr GR32:$src)>;
5013 let AddedComplexity = 20 in {
5014 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
5015 (MOVDI2PDIrm addr:$src)>;
5016 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
5017 (MOVDI2PDIrm addr:$src)>;
5018 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
5019 (MOVDI2PDIrm addr:$src)>;
5023 // These are the correct encodings of the instructions so that we know how to
5024 // read correct assembly, even though we continue to emit the wrong ones for
5025 // compatibility with Darwin's buggy assembler.
5026 def : InstAlias<"movq\t{$src, $dst|$dst, $src}",
5027 (MOV64toPQIrr VR128:$dst, GR64:$src), 0>;
5028 def : InstAlias<"movq\t{$src, $dst|$dst, $src}",
5029 (MOVPQIto64rr GR64:$dst, VR128:$src), 0>;
5030 // Allow "vmovd" but print "vmovq" since we don't need compatibility for AVX.
5031 def : InstAlias<"vmovd\t{$src, $dst|$dst, $src}",
5032 (VMOV64toPQIrr VR128:$dst, GR64:$src), 0>;
5033 def : InstAlias<"vmovd\t{$src, $dst|$dst, $src}",
5034 (VMOVPQIto64rr GR64:$dst, VR128:$src), 0>;
5036 //===---------------------------------------------------------------------===//
5037 // SSE2 - Move Quadword
5038 //===---------------------------------------------------------------------===//
5040 //===---------------------------------------------------------------------===//
5041 // Move Quadword Int to Packed Quadword Int
5044 let SchedRW = [WriteLoad] in {
5045 def VMOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
5046 "vmovq\t{$src, $dst|$dst, $src}",
5048 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
5049 VEX, Requires<[UseAVX]>;
5050 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
5051 "movq\t{$src, $dst|$dst, $src}",
5053 (v2i64 (scalar_to_vector (loadi64 addr:$src))))],
5055 Requires<[UseSSE2]>; // SSE2 instruction with XS Prefix
5058 //===---------------------------------------------------------------------===//
5059 // Move Packed Quadword Int to Quadword Int
5061 let SchedRW = [WriteStore] in {
5062 def VMOVPQI2QImr : VS2I<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
5063 "movq\t{$src, $dst|$dst, $src}",
5064 [(store (i64 (vector_extract (v2i64 VR128:$src),
5065 (iPTR 0))), addr:$dst)],
5066 IIC_SSE_MOVDQ>, VEX;
5067 def MOVPQI2QImr : S2I<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
5068 "movq\t{$src, $dst|$dst, $src}",
5069 [(store (i64 (vector_extract (v2i64 VR128:$src),
5070 (iPTR 0))), addr:$dst)],
5074 // For disassembler only
5075 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
5076 SchedRW = [WriteVecLogic] in {
5077 def VMOVPQI2QIrr : VS2I<0xD6, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
5078 "movq\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVQ_RR>, VEX;
5079 def MOVPQI2QIrr : S2I<0xD6, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
5080 "movq\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVQ_RR>;
5083 //===---------------------------------------------------------------------===//
5084 // Store / copy lower 64-bits of a XMM register.
5086 let Predicates = [UseAVX] in
5087 def : Pat<(int_x86_sse2_storel_dq addr:$dst, VR128:$src),
5088 (VMOVPQI2QImr addr:$dst, VR128:$src)>;
5089 let Predicates = [UseSSE2] in
5090 def : Pat<(int_x86_sse2_storel_dq addr:$dst, VR128:$src),
5091 (MOVPQI2QImr addr:$dst, VR128:$src)>;
5093 let isCodeGenOnly = 1, AddedComplexity = 20 in {
5094 def VMOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
5095 "vmovq\t{$src, $dst|$dst, $src}",
5097 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
5098 (loadi64 addr:$src))))))],
5100 XS, VEX, Requires<[UseAVX]>, Sched<[WriteLoad]>;
5102 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
5103 "movq\t{$src, $dst|$dst, $src}",
5105 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
5106 (loadi64 addr:$src))))))],
5108 XS, Requires<[UseSSE2]>, Sched<[WriteLoad]>;
5111 let Predicates = [UseAVX], AddedComplexity = 20 in {
5112 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
5113 (VMOVZQI2PQIrm addr:$src)>;
5114 def : Pat<(v2i64 (X86vzload addr:$src)),
5115 (VMOVZQI2PQIrm addr:$src)>;
5118 let Predicates = [UseSSE2], AddedComplexity = 20 in {
5119 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
5120 (MOVZQI2PQIrm addr:$src)>;
5121 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
5124 let Predicates = [HasAVX] in {
5125 def : Pat<(v4i64 (alignedX86vzload addr:$src)),
5126 (SUBREG_TO_REG (i32 0), (VMOVAPSrm addr:$src), sub_xmm)>;
5127 def : Pat<(v4i64 (X86vzload addr:$src)),
5128 (SUBREG_TO_REG (i32 0), (VMOVUPSrm addr:$src), sub_xmm)>;
5131 //===---------------------------------------------------------------------===//
5132 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
5133 // IA32 document. movq xmm1, xmm2 does clear the high bits.
5135 let SchedRW = [WriteVecLogic] in {
5136 let AddedComplexity = 15 in
5137 def VMOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5138 "vmovq\t{$src, $dst|$dst, $src}",
5139 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))],
5141 XS, VEX, Requires<[UseAVX]>;
5142 let AddedComplexity = 15 in
5143 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5144 "movq\t{$src, $dst|$dst, $src}",
5145 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))],
5147 XS, Requires<[UseSSE2]>;
5150 let isCodeGenOnly = 1, SchedRW = [WriteVecLogicLd] in {
5151 let AddedComplexity = 20 in
5152 def VMOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5153 "vmovq\t{$src, $dst|$dst, $src}",
5154 [(set VR128:$dst, (v2i64 (X86vzmovl
5155 (loadv2i64 addr:$src))))],
5157 XS, VEX, Requires<[UseAVX]>;
5158 let AddedComplexity = 20 in {
5159 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5160 "movq\t{$src, $dst|$dst, $src}",
5161 [(set VR128:$dst, (v2i64 (X86vzmovl
5162 (loadv2i64 addr:$src))))],
5164 XS, Requires<[UseSSE2]>;
5166 } // isCodeGenOnly, SchedRW
5168 let AddedComplexity = 20 in {
5169 let Predicates = [UseAVX] in {
5170 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
5171 (VMOVZPQILo2PQIrr VR128:$src)>;
5173 let Predicates = [UseSSE2] in {
5174 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
5175 (MOVZPQILo2PQIrr VR128:$src)>;
5179 //===---------------------------------------------------------------------===//
5180 // SSE3 - Replicate Single FP - MOVSHDUP and MOVSLDUP
5181 //===---------------------------------------------------------------------===//
5182 multiclass sse3_replicate_sfp<bits<8> op, SDNode OpNode, string OpcodeStr,
5183 ValueType vt, RegisterClass RC, PatFrag mem_frag,
5184 X86MemOperand x86memop> {
5185 def rr : S3SI<op, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
5186 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5187 [(set RC:$dst, (vt (OpNode RC:$src)))],
5188 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
5189 def rm : S3SI<op, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
5190 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5191 [(set RC:$dst, (OpNode (mem_frag addr:$src)))],
5192 IIC_SSE_MOV_LH>, Sched<[WriteLoad]>;
5195 let Predicates = [HasAVX] in {
5196 defm VMOVSHDUP : sse3_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
5197 v4f32, VR128, loadv4f32, f128mem>, VEX;
5198 defm VMOVSLDUP : sse3_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
5199 v4f32, VR128, loadv4f32, f128mem>, VEX;
5200 defm VMOVSHDUPY : sse3_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
5201 v8f32, VR256, loadv8f32, f256mem>, VEX, VEX_L;
5202 defm VMOVSLDUPY : sse3_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
5203 v8f32, VR256, loadv8f32, f256mem>, VEX, VEX_L;
5205 defm MOVSHDUP : sse3_replicate_sfp<0x16, X86Movshdup, "movshdup", v4f32, VR128,
5206 memopv4f32, f128mem>;
5207 defm MOVSLDUP : sse3_replicate_sfp<0x12, X86Movsldup, "movsldup", v4f32, VR128,
5208 memopv4f32, f128mem>;
5210 let Predicates = [HasAVX] in {
5211 def : Pat<(v4i32 (X86Movshdup VR128:$src)),
5212 (VMOVSHDUPrr VR128:$src)>;
5213 def : Pat<(v4i32 (X86Movshdup (bc_v4i32 (loadv2i64 addr:$src)))),
5214 (VMOVSHDUPrm addr:$src)>;
5215 def : Pat<(v4i32 (X86Movsldup VR128:$src)),
5216 (VMOVSLDUPrr VR128:$src)>;
5217 def : Pat<(v4i32 (X86Movsldup (bc_v4i32 (loadv2i64 addr:$src)))),
5218 (VMOVSLDUPrm addr:$src)>;
5219 def : Pat<(v8i32 (X86Movshdup VR256:$src)),
5220 (VMOVSHDUPYrr VR256:$src)>;
5221 def : Pat<(v8i32 (X86Movshdup (bc_v8i32 (loadv4i64 addr:$src)))),
5222 (VMOVSHDUPYrm addr:$src)>;
5223 def : Pat<(v8i32 (X86Movsldup VR256:$src)),
5224 (VMOVSLDUPYrr VR256:$src)>;
5225 def : Pat<(v8i32 (X86Movsldup (bc_v8i32 (loadv4i64 addr:$src)))),
5226 (VMOVSLDUPYrm addr:$src)>;
5229 let Predicates = [UseSSE3] in {
5230 def : Pat<(v4i32 (X86Movshdup VR128:$src)),
5231 (MOVSHDUPrr VR128:$src)>;
5232 def : Pat<(v4i32 (X86Movshdup (bc_v4i32 (memopv2i64 addr:$src)))),
5233 (MOVSHDUPrm addr:$src)>;
5234 def : Pat<(v4i32 (X86Movsldup VR128:$src)),
5235 (MOVSLDUPrr VR128:$src)>;
5236 def : Pat<(v4i32 (X86Movsldup (bc_v4i32 (memopv2i64 addr:$src)))),
5237 (MOVSLDUPrm addr:$src)>;
5240 //===---------------------------------------------------------------------===//
5241 // SSE3 - Replicate Double FP - MOVDDUP
5242 //===---------------------------------------------------------------------===//
5244 multiclass sse3_replicate_dfp<string OpcodeStr> {
5245 let neverHasSideEffects = 1 in
5246 def rr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5247 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5248 [], IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
5249 def rm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
5250 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5253 (scalar_to_vector (loadf64 addr:$src)))))],
5254 IIC_SSE_MOV_LH>, Sched<[WriteLoad]>;
5257 // FIXME: Merge with above classe when there're patterns for the ymm version
5258 multiclass sse3_replicate_dfp_y<string OpcodeStr> {
5259 def rr : S3DI<0x12, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
5260 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5261 [(set VR256:$dst, (v4f64 (X86Movddup VR256:$src)))]>,
5262 Sched<[WriteFShuffle]>;
5263 def rm : S3DI<0x12, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
5264 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5267 (scalar_to_vector (loadf64 addr:$src)))))]>,
5271 let Predicates = [HasAVX] in {
5272 defm VMOVDDUP : sse3_replicate_dfp<"vmovddup">, VEX;
5273 defm VMOVDDUPY : sse3_replicate_dfp_y<"vmovddup">, VEX, VEX_L;
5276 defm MOVDDUP : sse3_replicate_dfp<"movddup">;
5278 let Predicates = [HasAVX] in {
5279 def : Pat<(X86Movddup (loadv2f64 addr:$src)),
5280 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5281 def : Pat<(X86Movddup (bc_v2f64 (loadv4f32 addr:$src))),
5282 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5283 def : Pat<(X86Movddup (bc_v2f64 (loadv2i64 addr:$src))),
5284 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5285 def : Pat<(X86Movddup (bc_v2f64
5286 (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
5287 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5290 def : Pat<(X86Movddup (loadv4f64 addr:$src)),
5291 (VMOVDDUPYrm addr:$src)>;
5292 def : Pat<(X86Movddup (loadv4i64 addr:$src)),
5293 (VMOVDDUPYrm addr:$src)>;
5294 def : Pat<(X86Movddup (v4i64 (scalar_to_vector (loadi64 addr:$src)))),
5295 (VMOVDDUPYrm addr:$src)>;
5296 def : Pat<(X86Movddup (v4i64 VR256:$src)),
5297 (VMOVDDUPYrr VR256:$src)>;
5300 let Predicates = [UseAVX, OptForSize] in {
5301 def : Pat<(v2f64 (X86VBroadcast (loadf64 addr:$src))),
5302 (VMOVDDUPrm addr:$src)>;
5303 def : Pat<(v2i64 (X86VBroadcast (loadi64 addr:$src))),
5304 (VMOVDDUPrm addr:$src)>;
5307 let Predicates = [UseSSE3] in {
5308 def : Pat<(X86Movddup (memopv2f64 addr:$src)),
5309 (MOVDDUPrm addr:$src)>;
5310 def : Pat<(X86Movddup (bc_v2f64 (memopv4f32 addr:$src))),
5311 (MOVDDUPrm addr:$src)>;
5312 def : Pat<(X86Movddup (bc_v2f64 (memopv2i64 addr:$src))),
5313 (MOVDDUPrm addr:$src)>;
5314 def : Pat<(X86Movddup (bc_v2f64
5315 (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
5316 (MOVDDUPrm addr:$src)>;
5319 //===---------------------------------------------------------------------===//
5320 // SSE3 - Move Unaligned Integer
5321 //===---------------------------------------------------------------------===//
5323 let SchedRW = [WriteLoad] in {
5324 let Predicates = [HasAVX] in {
5325 def VLDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5326 "vlddqu\t{$src, $dst|$dst, $src}",
5327 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>, VEX;
5328 def VLDDQUYrm : S3DI<0xF0, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
5329 "vlddqu\t{$src, $dst|$dst, $src}",
5330 [(set VR256:$dst, (int_x86_avx_ldu_dq_256 addr:$src))]>,
5333 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5334 "lddqu\t{$src, $dst|$dst, $src}",
5335 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))],
5339 //===---------------------------------------------------------------------===//
5340 // SSE3 - Arithmetic
5341 //===---------------------------------------------------------------------===//
5343 multiclass sse3_addsub<Intrinsic Int, string OpcodeStr, RegisterClass RC,
5344 X86MemOperand x86memop, OpndItins itins,
5346 def rr : I<0xD0, MRMSrcReg,
5347 (outs RC:$dst), (ins RC:$src1, RC:$src2),
5349 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5350 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5351 [(set RC:$dst, (Int RC:$src1, RC:$src2))], itins.rr>,
5352 Sched<[itins.Sched]>;
5353 def rm : I<0xD0, MRMSrcMem,
5354 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5356 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5357 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5358 [(set RC:$dst, (Int RC:$src1, (memop addr:$src2)))], itins.rr>,
5359 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5362 let Predicates = [HasAVX] in {
5363 let ExeDomain = SSEPackedSingle in {
5364 defm VADDSUBPS : sse3_addsub<int_x86_sse3_addsub_ps, "vaddsubps", VR128,
5365 f128mem, SSE_ALU_F32P, 0>, XD, VEX_4V;
5366 defm VADDSUBPSY : sse3_addsub<int_x86_avx_addsub_ps_256, "vaddsubps", VR256,
5367 f256mem, SSE_ALU_F32P, 0>, XD, VEX_4V, VEX_L;
5369 let ExeDomain = SSEPackedDouble in {
5370 defm VADDSUBPD : sse3_addsub<int_x86_sse3_addsub_pd, "vaddsubpd", VR128,
5371 f128mem, SSE_ALU_F64P, 0>, PD, VEX_4V;
5372 defm VADDSUBPDY : sse3_addsub<int_x86_avx_addsub_pd_256, "vaddsubpd", VR256,
5373 f256mem, SSE_ALU_F64P, 0>, PD, VEX_4V, VEX_L;
5376 let Constraints = "$src1 = $dst", Predicates = [UseSSE3] in {
5377 let ExeDomain = SSEPackedSingle in
5378 defm ADDSUBPS : sse3_addsub<int_x86_sse3_addsub_ps, "addsubps", VR128,
5379 f128mem, SSE_ALU_F32P>, XD;
5380 let ExeDomain = SSEPackedDouble in
5381 defm ADDSUBPD : sse3_addsub<int_x86_sse3_addsub_pd, "addsubpd", VR128,
5382 f128mem, SSE_ALU_F64P>, PD;
5385 // Patterns used to select 'addsub' instructions.
5386 let Predicates = [HasAVX] in {
5387 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 VR128:$rhs))),
5388 (VADDSUBPSrr VR128:$lhs, VR128:$rhs)>;
5389 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 (memop addr:$rhs)))),
5390 (VADDSUBPSrm VR128:$lhs, f128mem:$rhs)>;
5391 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 VR128:$rhs))),
5392 (VADDSUBPDrr VR128:$lhs, VR128:$rhs)>;
5393 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 (memop addr:$rhs)))),
5394 (VADDSUBPDrm VR128:$lhs, f128mem:$rhs)>;
5396 def : Pat<(v8f32 (X86Addsub (v8f32 VR256:$lhs), (v8f32 VR256:$rhs))),
5397 (VADDSUBPSYrr VR256:$lhs, VR256:$rhs)>;
5398 def : Pat<(v8f32 (X86Addsub (v8f32 VR256:$lhs), (v8f32 (memop addr:$rhs)))),
5399 (VADDSUBPSYrm VR256:$lhs, f256mem:$rhs)>;
5400 def : Pat<(v4f64 (X86Addsub (v4f64 VR256:$lhs), (v4f64 VR256:$rhs))),
5401 (VADDSUBPDYrr VR256:$lhs, VR256:$rhs)>;
5402 def : Pat<(v4f64 (X86Addsub (v4f64 VR256:$lhs), (v4f64 (memop addr:$rhs)))),
5403 (VADDSUBPDYrm VR256:$lhs, f256mem:$rhs)>;
5406 let Predicates = [UseSSE3] in {
5407 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 VR128:$rhs))),
5408 (ADDSUBPSrr VR128:$lhs, VR128:$rhs)>;
5409 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 (memop addr:$rhs)))),
5410 (ADDSUBPSrm VR128:$lhs, f128mem:$rhs)>;
5411 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 VR128:$rhs))),
5412 (ADDSUBPDrr VR128:$lhs, VR128:$rhs)>;
5413 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 (memop addr:$rhs)))),
5414 (ADDSUBPDrm VR128:$lhs, f128mem:$rhs)>;
5417 //===---------------------------------------------------------------------===//
5418 // SSE3 Instructions
5419 //===---------------------------------------------------------------------===//
5422 multiclass S3D_Int<bits<8> o, string OpcodeStr, ValueType vt, RegisterClass RC,
5423 X86MemOperand x86memop, SDNode OpNode, bit Is2Addr = 1> {
5424 def rr : S3DI<o, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
5426 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5427 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5428 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], IIC_SSE_HADDSUB_RR>,
5431 def rm : S3DI<o, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5433 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5434 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5435 [(set RC:$dst, (vt (OpNode RC:$src1, (memop addr:$src2))))],
5436 IIC_SSE_HADDSUB_RM>, Sched<[WriteFAddLd, ReadAfterLd]>;
5438 multiclass S3_Int<bits<8> o, string OpcodeStr, ValueType vt, RegisterClass RC,
5439 X86MemOperand x86memop, SDNode OpNode, bit Is2Addr = 1> {
5440 def rr : S3I<o, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
5442 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5443 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5444 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], IIC_SSE_HADDSUB_RR>,
5447 def rm : S3I<o, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5449 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5450 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5451 [(set RC:$dst, (vt (OpNode RC:$src1, (memop addr:$src2))))],
5452 IIC_SSE_HADDSUB_RM>, Sched<[WriteFAddLd, ReadAfterLd]>;
5455 let Predicates = [HasAVX] in {
5456 let ExeDomain = SSEPackedSingle in {
5457 defm VHADDPS : S3D_Int<0x7C, "vhaddps", v4f32, VR128, f128mem,
5458 X86fhadd, 0>, VEX_4V;
5459 defm VHSUBPS : S3D_Int<0x7D, "vhsubps", v4f32, VR128, f128mem,
5460 X86fhsub, 0>, VEX_4V;
5461 defm VHADDPSY : S3D_Int<0x7C, "vhaddps", v8f32, VR256, f256mem,
5462 X86fhadd, 0>, VEX_4V, VEX_L;
5463 defm VHSUBPSY : S3D_Int<0x7D, "vhsubps", v8f32, VR256, f256mem,
5464 X86fhsub, 0>, VEX_4V, VEX_L;
5466 let ExeDomain = SSEPackedDouble in {
5467 defm VHADDPD : S3_Int <0x7C, "vhaddpd", v2f64, VR128, f128mem,
5468 X86fhadd, 0>, VEX_4V;
5469 defm VHSUBPD : S3_Int <0x7D, "vhsubpd", v2f64, VR128, f128mem,
5470 X86fhsub, 0>, VEX_4V;
5471 defm VHADDPDY : S3_Int <0x7C, "vhaddpd", v4f64, VR256, f256mem,
5472 X86fhadd, 0>, VEX_4V, VEX_L;
5473 defm VHSUBPDY : S3_Int <0x7D, "vhsubpd", v4f64, VR256, f256mem,
5474 X86fhsub, 0>, VEX_4V, VEX_L;
5478 let Constraints = "$src1 = $dst" in {
5479 let ExeDomain = SSEPackedSingle in {
5480 defm HADDPS : S3D_Int<0x7C, "haddps", v4f32, VR128, f128mem, X86fhadd>;
5481 defm HSUBPS : S3D_Int<0x7D, "hsubps", v4f32, VR128, f128mem, X86fhsub>;
5483 let ExeDomain = SSEPackedDouble in {
5484 defm HADDPD : S3_Int<0x7C, "haddpd", v2f64, VR128, f128mem, X86fhadd>;
5485 defm HSUBPD : S3_Int<0x7D, "hsubpd", v2f64, VR128, f128mem, X86fhsub>;
5489 //===---------------------------------------------------------------------===//
5490 // SSSE3 - Packed Absolute Instructions
5491 //===---------------------------------------------------------------------===//
5494 /// SS3I_unop_rm_int - Simple SSSE3 unary op whose type can be v*{i8,i16,i32}.
5495 multiclass SS3I_unop_rm_int<bits<8> opc, string OpcodeStr,
5496 Intrinsic IntId128> {
5497 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
5499 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5500 [(set VR128:$dst, (IntId128 VR128:$src))], IIC_SSE_PABS_RR>,
5501 Sched<[WriteVecALU]>;
5503 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
5505 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5508 (bitconvert (memopv2i64 addr:$src))))], IIC_SSE_PABS_RM>,
5509 Sched<[WriteVecALULd]>;
5512 /// SS3I_unop_rm_int_y - Simple SSSE3 unary op whose type can be v*{i8,i16,i32}.
5513 multiclass SS3I_unop_rm_int_y<bits<8> opc, string OpcodeStr,
5514 Intrinsic IntId256> {
5515 def rr256 : SS38I<opc, MRMSrcReg, (outs VR256:$dst),
5517 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5518 [(set VR256:$dst, (IntId256 VR256:$src))]>,
5519 Sched<[WriteVecALU]>;
5521 def rm256 : SS38I<opc, MRMSrcMem, (outs VR256:$dst),
5523 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5526 (bitconvert (memopv4i64 addr:$src))))]>,
5527 Sched<[WriteVecALULd]>;
5530 // Helper fragments to match sext vXi1 to vXiY.
5531 def v16i1sextv16i8 : PatLeaf<(v16i8 (X86pcmpgt (bc_v16i8 (v4i32 immAllZerosV)),
5533 def v8i1sextv8i16 : PatLeaf<(v8i16 (X86vsrai VR128:$src, (i8 15)))>;
5534 def v4i1sextv4i32 : PatLeaf<(v4i32 (X86vsrai VR128:$src, (i8 31)))>;
5535 def v32i1sextv32i8 : PatLeaf<(v32i8 (X86pcmpgt (bc_v32i8 (v8i32 immAllZerosV)),
5537 def v16i1sextv16i16: PatLeaf<(v16i16 (X86vsrai VR256:$src, (i8 15)))>;
5538 def v8i1sextv8i32 : PatLeaf<(v8i32 (X86vsrai VR256:$src, (i8 31)))>;
5540 let Predicates = [HasAVX] in {
5541 defm VPABSB : SS3I_unop_rm_int<0x1C, "vpabsb",
5542 int_x86_ssse3_pabs_b_128>, VEX;
5543 defm VPABSW : SS3I_unop_rm_int<0x1D, "vpabsw",
5544 int_x86_ssse3_pabs_w_128>, VEX;
5545 defm VPABSD : SS3I_unop_rm_int<0x1E, "vpabsd",
5546 int_x86_ssse3_pabs_d_128>, VEX;
5549 (bc_v2i64 (v16i1sextv16i8)),
5550 (bc_v2i64 (add (v16i8 VR128:$src), (v16i1sextv16i8)))),
5551 (VPABSBrr128 VR128:$src)>;
5553 (bc_v2i64 (v8i1sextv8i16)),
5554 (bc_v2i64 (add (v8i16 VR128:$src), (v8i1sextv8i16)))),
5555 (VPABSWrr128 VR128:$src)>;
5557 (bc_v2i64 (v4i1sextv4i32)),
5558 (bc_v2i64 (add (v4i32 VR128:$src), (v4i1sextv4i32)))),
5559 (VPABSDrr128 VR128:$src)>;
5562 let Predicates = [HasAVX2] in {
5563 defm VPABSB : SS3I_unop_rm_int_y<0x1C, "vpabsb",
5564 int_x86_avx2_pabs_b>, VEX, VEX_L;
5565 defm VPABSW : SS3I_unop_rm_int_y<0x1D, "vpabsw",
5566 int_x86_avx2_pabs_w>, VEX, VEX_L;
5567 defm VPABSD : SS3I_unop_rm_int_y<0x1E, "vpabsd",
5568 int_x86_avx2_pabs_d>, VEX, VEX_L;
5571 (bc_v4i64 (v32i1sextv32i8)),
5572 (bc_v4i64 (add (v32i8 VR256:$src), (v32i1sextv32i8)))),
5573 (VPABSBrr256 VR256:$src)>;
5575 (bc_v4i64 (v16i1sextv16i16)),
5576 (bc_v4i64 (add (v16i16 VR256:$src), (v16i1sextv16i16)))),
5577 (VPABSWrr256 VR256:$src)>;
5579 (bc_v4i64 (v8i1sextv8i32)),
5580 (bc_v4i64 (add (v8i32 VR256:$src), (v8i1sextv8i32)))),
5581 (VPABSDrr256 VR256:$src)>;
5584 defm PABSB : SS3I_unop_rm_int<0x1C, "pabsb",
5585 int_x86_ssse3_pabs_b_128>;
5586 defm PABSW : SS3I_unop_rm_int<0x1D, "pabsw",
5587 int_x86_ssse3_pabs_w_128>;
5588 defm PABSD : SS3I_unop_rm_int<0x1E, "pabsd",
5589 int_x86_ssse3_pabs_d_128>;
5591 let Predicates = [HasSSSE3] in {
5593 (bc_v2i64 (v16i1sextv16i8)),
5594 (bc_v2i64 (add (v16i8 VR128:$src), (v16i1sextv16i8)))),
5595 (PABSBrr128 VR128:$src)>;
5597 (bc_v2i64 (v8i1sextv8i16)),
5598 (bc_v2i64 (add (v8i16 VR128:$src), (v8i1sextv8i16)))),
5599 (PABSWrr128 VR128:$src)>;
5601 (bc_v2i64 (v4i1sextv4i32)),
5602 (bc_v2i64 (add (v4i32 VR128:$src), (v4i1sextv4i32)))),
5603 (PABSDrr128 VR128:$src)>;
5606 //===---------------------------------------------------------------------===//
5607 // SSSE3 - Packed Binary Operator Instructions
5608 //===---------------------------------------------------------------------===//
5610 let Sched = WriteVecALU in {
5611 def SSE_PHADDSUBD : OpndItins<
5612 IIC_SSE_PHADDSUBD_RR, IIC_SSE_PHADDSUBD_RM
5614 def SSE_PHADDSUBSW : OpndItins<
5615 IIC_SSE_PHADDSUBSW_RR, IIC_SSE_PHADDSUBSW_RM
5617 def SSE_PHADDSUBW : OpndItins<
5618 IIC_SSE_PHADDSUBW_RR, IIC_SSE_PHADDSUBW_RM
5621 let Sched = WriteShuffle in
5622 def SSE_PSHUFB : OpndItins<
5623 IIC_SSE_PSHUFB_RR, IIC_SSE_PSHUFB_RM
5625 let Sched = WriteVecALU in
5626 def SSE_PSIGN : OpndItins<
5627 IIC_SSE_PSIGN_RR, IIC_SSE_PSIGN_RM
5629 let Sched = WriteVecIMul in
5630 def SSE_PMULHRSW : OpndItins<
5631 IIC_SSE_PMULHRSW, IIC_SSE_PMULHRSW
5634 /// SS3I_binop_rm - Simple SSSE3 bin op
5635 multiclass SS3I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
5636 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
5637 X86MemOperand x86memop, OpndItins itins,
5639 let isCommutable = 1 in
5640 def rr : SS38I<opc, MRMSrcReg, (outs RC:$dst),
5641 (ins RC:$src1, RC:$src2),
5643 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5644 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5645 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))], itins.rr>,
5646 Sched<[itins.Sched]>;
5647 def rm : SS38I<opc, MRMSrcMem, (outs RC:$dst),
5648 (ins RC:$src1, x86memop:$src2),
5650 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5651 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5653 (OpVT (OpNode RC:$src1,
5654 (bitconvert (memop_frag addr:$src2)))))], itins.rm>,
5655 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5658 /// SS3I_binop_rm_int - Simple SSSE3 bin op whose type can be v*{i8,i16,i32}.
5659 multiclass SS3I_binop_rm_int<bits<8> opc, string OpcodeStr,
5660 Intrinsic IntId128, OpndItins itins,
5662 let isCommutable = 1 in
5663 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
5664 (ins VR128:$src1, VR128:$src2),
5666 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5667 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5668 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
5669 Sched<[itins.Sched]>;
5670 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
5671 (ins VR128:$src1, i128mem:$src2),
5673 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5674 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5676 (IntId128 VR128:$src1,
5677 (bitconvert (memopv2i64 addr:$src2))))]>,
5678 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5681 multiclass SS3I_binop_rm_int_y<bits<8> opc, string OpcodeStr,
5683 X86FoldableSchedWrite Sched> {
5684 let isCommutable = 1 in
5685 def rr256 : SS38I<opc, MRMSrcReg, (outs VR256:$dst),
5686 (ins VR256:$src1, VR256:$src2),
5687 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5688 [(set VR256:$dst, (IntId256 VR256:$src1, VR256:$src2))]>,
5690 def rm256 : SS38I<opc, MRMSrcMem, (outs VR256:$dst),
5691 (ins VR256:$src1, i256mem:$src2),
5692 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5694 (IntId256 VR256:$src1, (bitconvert (loadv4i64 addr:$src2))))]>,
5695 Sched<[Sched.Folded, ReadAfterLd]>;
5698 let ImmT = NoImm, Predicates = [HasAVX] in {
5699 let isCommutable = 0 in {
5700 defm VPHADDW : SS3I_binop_rm<0x01, "vphaddw", X86hadd, v8i16, VR128,
5702 SSE_PHADDSUBW, 0>, VEX_4V;
5703 defm VPHADDD : SS3I_binop_rm<0x02, "vphaddd", X86hadd, v4i32, VR128,
5705 SSE_PHADDSUBD, 0>, VEX_4V;
5706 defm VPHSUBW : SS3I_binop_rm<0x05, "vphsubw", X86hsub, v8i16, VR128,
5708 SSE_PHADDSUBW, 0>, VEX_4V;
5709 defm VPHSUBD : SS3I_binop_rm<0x06, "vphsubd", X86hsub, v4i32, VR128,
5711 SSE_PHADDSUBD, 0>, VEX_4V;
5712 defm VPSIGNB : SS3I_binop_rm<0x08, "vpsignb", X86psign, v16i8, VR128,
5714 SSE_PSIGN, 0>, VEX_4V;
5715 defm VPSIGNW : SS3I_binop_rm<0x09, "vpsignw", X86psign, v8i16, VR128,
5717 SSE_PSIGN, 0>, VEX_4V;
5718 defm VPSIGND : SS3I_binop_rm<0x0A, "vpsignd", X86psign, v4i32, VR128,
5720 SSE_PSIGN, 0>, VEX_4V;
5721 defm VPSHUFB : SS3I_binop_rm<0x00, "vpshufb", X86pshufb, v16i8, VR128,
5723 SSE_PSHUFB, 0>, VEX_4V;
5724 defm VPHADDSW : SS3I_binop_rm_int<0x03, "vphaddsw",
5725 int_x86_ssse3_phadd_sw_128,
5726 SSE_PHADDSUBSW, 0>, VEX_4V;
5727 defm VPHSUBSW : SS3I_binop_rm_int<0x07, "vphsubsw",
5728 int_x86_ssse3_phsub_sw_128,
5729 SSE_PHADDSUBSW, 0>, VEX_4V;
5730 defm VPMADDUBSW : SS3I_binop_rm_int<0x04, "vpmaddubsw",
5731 int_x86_ssse3_pmadd_ub_sw_128,
5732 SSE_PMADD, 0>, VEX_4V;
5734 defm VPMULHRSW : SS3I_binop_rm_int<0x0B, "vpmulhrsw",
5735 int_x86_ssse3_pmul_hr_sw_128,
5736 SSE_PMULHRSW, 0>, VEX_4V;
5739 let ImmT = NoImm, Predicates = [HasAVX2] in {
5740 let isCommutable = 0 in {
5741 defm VPHADDWY : SS3I_binop_rm<0x01, "vphaddw", X86hadd, v16i16, VR256,
5743 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5744 defm VPHADDDY : SS3I_binop_rm<0x02, "vphaddd", X86hadd, v8i32, VR256,
5746 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5747 defm VPHSUBWY : SS3I_binop_rm<0x05, "vphsubw", X86hsub, v16i16, VR256,
5749 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5750 defm VPHSUBDY : SS3I_binop_rm<0x06, "vphsubd", X86hsub, v8i32, VR256,
5752 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5753 defm VPSIGNBY : SS3I_binop_rm<0x08, "vpsignb", X86psign, v32i8, VR256,
5755 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5756 defm VPSIGNWY : SS3I_binop_rm<0x09, "vpsignw", X86psign, v16i16, VR256,
5758 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5759 defm VPSIGNDY : SS3I_binop_rm<0x0A, "vpsignd", X86psign, v8i32, VR256,
5761 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5762 defm VPSHUFBY : SS3I_binop_rm<0x00, "vpshufb", X86pshufb, v32i8, VR256,
5764 SSE_PSHUFB, 0>, VEX_4V, VEX_L;
5765 defm VPHADDSW : SS3I_binop_rm_int_y<0x03, "vphaddsw",
5766 int_x86_avx2_phadd_sw,
5767 WriteVecALU>, VEX_4V, VEX_L;
5768 defm VPHSUBSW : SS3I_binop_rm_int_y<0x07, "vphsubsw",
5769 int_x86_avx2_phsub_sw,
5770 WriteVecALU>, VEX_4V, VEX_L;
5771 defm VPMADDUBSW : SS3I_binop_rm_int_y<0x04, "vpmaddubsw",
5772 int_x86_avx2_pmadd_ub_sw,
5773 WriteVecIMul>, VEX_4V, VEX_L;
5775 defm VPMULHRSW : SS3I_binop_rm_int_y<0x0B, "vpmulhrsw",
5776 int_x86_avx2_pmul_hr_sw,
5777 WriteVecIMul>, VEX_4V, VEX_L;
5780 // None of these have i8 immediate fields.
5781 let ImmT = NoImm, Constraints = "$src1 = $dst" in {
5782 let isCommutable = 0 in {
5783 defm PHADDW : SS3I_binop_rm<0x01, "phaddw", X86hadd, v8i16, VR128,
5784 memopv2i64, i128mem, SSE_PHADDSUBW>;
5785 defm PHADDD : SS3I_binop_rm<0x02, "phaddd", X86hadd, v4i32, VR128,
5786 memopv2i64, i128mem, SSE_PHADDSUBD>;
5787 defm PHSUBW : SS3I_binop_rm<0x05, "phsubw", X86hsub, v8i16, VR128,
5788 memopv2i64, i128mem, SSE_PHADDSUBW>;
5789 defm PHSUBD : SS3I_binop_rm<0x06, "phsubd", X86hsub, v4i32, VR128,
5790 memopv2i64, i128mem, SSE_PHADDSUBD>;
5791 defm PSIGNB : SS3I_binop_rm<0x08, "psignb", X86psign, v16i8, VR128,
5792 memopv2i64, i128mem, SSE_PSIGN>;
5793 defm PSIGNW : SS3I_binop_rm<0x09, "psignw", X86psign, v8i16, VR128,
5794 memopv2i64, i128mem, SSE_PSIGN>;
5795 defm PSIGND : SS3I_binop_rm<0x0A, "psignd", X86psign, v4i32, VR128,
5796 memopv2i64, i128mem, SSE_PSIGN>;
5797 defm PSHUFB : SS3I_binop_rm<0x00, "pshufb", X86pshufb, v16i8, VR128,
5798 memopv2i64, i128mem, SSE_PSHUFB>;
5799 defm PHADDSW : SS3I_binop_rm_int<0x03, "phaddsw",
5800 int_x86_ssse3_phadd_sw_128,
5802 defm PHSUBSW : SS3I_binop_rm_int<0x07, "phsubsw",
5803 int_x86_ssse3_phsub_sw_128,
5805 defm PMADDUBSW : SS3I_binop_rm_int<0x04, "pmaddubsw",
5806 int_x86_ssse3_pmadd_ub_sw_128, SSE_PMADD>;
5808 defm PMULHRSW : SS3I_binop_rm_int<0x0B, "pmulhrsw",
5809 int_x86_ssse3_pmul_hr_sw_128,
5813 //===---------------------------------------------------------------------===//
5814 // SSSE3 - Packed Align Instruction Patterns
5815 //===---------------------------------------------------------------------===//
5817 multiclass ssse3_palignr<string asm, bit Is2Addr = 1> {
5818 let neverHasSideEffects = 1 in {
5819 def R128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
5820 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
5822 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
5824 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
5825 [], IIC_SSE_PALIGNRR>, Sched<[WriteShuffle]>;
5827 def R128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
5828 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
5830 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
5832 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
5833 [], IIC_SSE_PALIGNRM>, Sched<[WriteShuffleLd, ReadAfterLd]>;
5837 multiclass ssse3_palignr_y<string asm, bit Is2Addr = 1> {
5838 let neverHasSideEffects = 1 in {
5839 def R256rr : SS3AI<0x0F, MRMSrcReg, (outs VR256:$dst),
5840 (ins VR256:$src1, VR256:$src2, i8imm:$src3),
5842 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
5843 []>, Sched<[WriteShuffle]>;
5845 def R256rm : SS3AI<0x0F, MRMSrcMem, (outs VR256:$dst),
5846 (ins VR256:$src1, i256mem:$src2, i8imm:$src3),
5848 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
5849 []>, Sched<[WriteShuffleLd, ReadAfterLd]>;
5853 let Predicates = [HasAVX] in
5854 defm VPALIGN : ssse3_palignr<"vpalignr", 0>, VEX_4V;
5855 let Predicates = [HasAVX2] in
5856 defm VPALIGN : ssse3_palignr_y<"vpalignr", 0>, VEX_4V, VEX_L;
5857 let Constraints = "$src1 = $dst", Predicates = [UseSSSE3] in
5858 defm PALIGN : ssse3_palignr<"palignr">;
5860 let Predicates = [HasAVX2] in {
5861 def : Pat<(v8i32 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5862 (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
5863 def : Pat<(v8f32 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5864 (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
5865 def : Pat<(v16i16 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5866 (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
5867 def : Pat<(v32i8 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5868 (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
5871 let Predicates = [HasAVX] in {
5872 def : Pat<(v4i32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5873 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5874 def : Pat<(v4f32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5875 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5876 def : Pat<(v8i16 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5877 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5878 def : Pat<(v16i8 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5879 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5882 let Predicates = [UseSSSE3] in {
5883 def : Pat<(v4i32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5884 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5885 def : Pat<(v4f32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5886 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5887 def : Pat<(v8i16 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5888 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5889 def : Pat<(v16i8 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5890 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5893 //===---------------------------------------------------------------------===//
5894 // SSSE3 - Thread synchronization
5895 //===---------------------------------------------------------------------===//
5897 let SchedRW = [WriteSystem] in {
5898 let usesCustomInserter = 1 in {
5899 def MONITOR : PseudoI<(outs), (ins i32mem:$src1, GR32:$src2, GR32:$src3),
5900 [(int_x86_sse3_monitor addr:$src1, GR32:$src2, GR32:$src3)]>,
5901 Requires<[HasSSE3]>;
5904 let Uses = [EAX, ECX, EDX] in
5905 def MONITORrrr : I<0x01, MRM_C8, (outs), (ins), "monitor", [], IIC_SSE_MONITOR>,
5906 TB, Requires<[HasSSE3]>;
5907 let Uses = [ECX, EAX] in
5908 def MWAITrr : I<0x01, MRM_C9, (outs), (ins), "mwait",
5909 [(int_x86_sse3_mwait ECX, EAX)], IIC_SSE_MWAIT>,
5910 TB, Requires<[HasSSE3]>;
5913 def : InstAlias<"mwait\t{%eax, %ecx|ecx, eax}", (MWAITrr)>, Requires<[Not64BitMode]>;
5914 def : InstAlias<"mwait\t{%rax, %rcx|rcx, rax}", (MWAITrr)>, Requires<[In64BitMode]>;
5916 def : InstAlias<"monitor\t{%eax, %ecx, %edx|edx, ecx, eax}", (MONITORrrr)>,
5917 Requires<[Not64BitMode]>;
5918 def : InstAlias<"monitor\t{%rax, %rcx, %rdx|rdx, rcx, rax}", (MONITORrrr)>,
5919 Requires<[In64BitMode]>;
5921 //===----------------------------------------------------------------------===//
5922 // SSE4.1 - Packed Move with Sign/Zero Extend
5923 //===----------------------------------------------------------------------===//
5925 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId,
5926 OpndItins itins = DEFAULT_ITINS> {
5927 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5928 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5929 [(set VR128:$dst, (IntId VR128:$src))], itins.rr>,
5930 Sched<[itins.Sched]>;
5932 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
5933 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5935 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))],
5936 itins.rm>, Sched<[itins.Sched.Folded]>;
5939 multiclass SS41I_binop_rm_int16_y<bits<8> opc, string OpcodeStr,
5940 Intrinsic IntId, X86FoldableSchedWrite Sched> {
5941 def Yrr : SS48I<opc, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
5942 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5943 [(set VR256:$dst, (IntId VR128:$src))]>, Sched<[Sched]>;
5945 def Yrm : SS48I<opc, MRMSrcMem, (outs VR256:$dst), (ins i128mem:$src),
5946 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5947 [(set VR256:$dst, (IntId (load addr:$src)))]>,
5948 Sched<[Sched.Folded]>;
5951 let Predicates = [HasAVX] in {
5952 defm VPMOVSXBW : SS41I_binop_rm_int8<0x20, "vpmovsxbw",
5953 int_x86_sse41_pmovsxbw,
5954 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
5955 defm VPMOVSXWD : SS41I_binop_rm_int8<0x23, "vpmovsxwd",
5956 int_x86_sse41_pmovsxwd,
5957 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
5958 defm VPMOVSXDQ : SS41I_binop_rm_int8<0x25, "vpmovsxdq",
5959 int_x86_sse41_pmovsxdq,
5960 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
5961 defm VPMOVZXBW : SS41I_binop_rm_int8<0x30, "vpmovzxbw",
5962 int_x86_sse41_pmovzxbw,
5963 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
5964 defm VPMOVZXWD : SS41I_binop_rm_int8<0x33, "vpmovzxwd",
5965 int_x86_sse41_pmovzxwd,
5966 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
5967 defm VPMOVZXDQ : SS41I_binop_rm_int8<0x35, "vpmovzxdq",
5968 int_x86_sse41_pmovzxdq,
5969 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
5972 let Predicates = [HasAVX2] in {
5973 defm VPMOVSXBW : SS41I_binop_rm_int16_y<0x20, "vpmovsxbw",
5974 int_x86_avx2_pmovsxbw,
5975 WriteShuffle>, VEX, VEX_L;
5976 defm VPMOVSXWD : SS41I_binop_rm_int16_y<0x23, "vpmovsxwd",
5977 int_x86_avx2_pmovsxwd,
5978 WriteShuffle>, VEX, VEX_L;
5979 defm VPMOVSXDQ : SS41I_binop_rm_int16_y<0x25, "vpmovsxdq",
5980 int_x86_avx2_pmovsxdq,
5981 WriteShuffle>, VEX, VEX_L;
5982 defm VPMOVZXBW : SS41I_binop_rm_int16_y<0x30, "vpmovzxbw",
5983 int_x86_avx2_pmovzxbw,
5984 WriteShuffle>, VEX, VEX_L;
5985 defm VPMOVZXWD : SS41I_binop_rm_int16_y<0x33, "vpmovzxwd",
5986 int_x86_avx2_pmovzxwd,
5987 WriteShuffle>, VEX, VEX_L;
5988 defm VPMOVZXDQ : SS41I_binop_rm_int16_y<0x35, "vpmovzxdq",
5989 int_x86_avx2_pmovzxdq,
5990 WriteShuffle>, VEX, VEX_L;
5993 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw,
5994 SSE_INTALU_ITINS_SHUFF_P>;
5995 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd,
5996 SSE_INTALU_ITINS_SHUFF_P>;
5997 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq,
5998 SSE_INTALU_ITINS_SHUFF_P>;
5999 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw,
6000 SSE_INTALU_ITINS_SHUFF_P>;
6001 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd,
6002 SSE_INTALU_ITINS_SHUFF_P>;
6003 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq,
6004 SSE_INTALU_ITINS_SHUFF_P>;
6006 let Predicates = [HasAVX] in {
6007 // Common patterns involving scalar load.
6008 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
6009 (VPMOVSXBWrm addr:$src)>;
6010 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
6011 (VPMOVSXBWrm addr:$src)>;
6012 def : Pat<(int_x86_sse41_pmovsxbw (bc_v16i8 (loadv2i64 addr:$src))),
6013 (VPMOVSXBWrm addr:$src)>;
6015 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
6016 (VPMOVSXWDrm addr:$src)>;
6017 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
6018 (VPMOVSXWDrm addr:$src)>;
6019 def : Pat<(int_x86_sse41_pmovsxwd (bc_v8i16 (loadv2i64 addr:$src))),
6020 (VPMOVSXWDrm addr:$src)>;
6022 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
6023 (VPMOVSXDQrm addr:$src)>;
6024 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
6025 (VPMOVSXDQrm addr:$src)>;
6026 def : Pat<(int_x86_sse41_pmovsxdq (bc_v4i32 (loadv2i64 addr:$src))),
6027 (VPMOVSXDQrm addr:$src)>;
6029 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
6030 (VPMOVZXBWrm addr:$src)>;
6031 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
6032 (VPMOVZXBWrm addr:$src)>;
6033 def : Pat<(int_x86_sse41_pmovzxbw (bc_v16i8 (loadv2i64 addr:$src))),
6034 (VPMOVZXBWrm addr:$src)>;
6036 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
6037 (VPMOVZXWDrm addr:$src)>;
6038 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
6039 (VPMOVZXWDrm addr:$src)>;
6040 def : Pat<(int_x86_sse41_pmovzxwd (bc_v8i16 (loadv2i64 addr:$src))),
6041 (VPMOVZXWDrm addr:$src)>;
6043 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
6044 (VPMOVZXDQrm addr:$src)>;
6045 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
6046 (VPMOVZXDQrm addr:$src)>;
6047 def : Pat<(int_x86_sse41_pmovzxdq (bc_v4i32 (loadv2i64 addr:$src))),
6048 (VPMOVZXDQrm addr:$src)>;
6051 let Predicates = [UseSSE41] in {
6052 // Common patterns involving scalar load.
6053 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
6054 (PMOVSXBWrm addr:$src)>;
6055 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
6056 (PMOVSXBWrm addr:$src)>;
6057 def : Pat<(int_x86_sse41_pmovsxbw (bc_v16i8 (loadv2i64 addr:$src))),
6058 (PMOVSXBWrm addr:$src)>;
6060 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
6061 (PMOVSXWDrm addr:$src)>;
6062 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
6063 (PMOVSXWDrm addr:$src)>;
6064 def : Pat<(int_x86_sse41_pmovsxwd (bc_v8i16 (loadv2i64 addr:$src))),
6065 (PMOVSXWDrm addr:$src)>;
6067 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
6068 (PMOVSXDQrm addr:$src)>;
6069 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
6070 (PMOVSXDQrm addr:$src)>;
6071 def : Pat<(int_x86_sse41_pmovsxdq (bc_v4i32 (loadv2i64 addr:$src))),
6072 (PMOVSXDQrm addr:$src)>;
6074 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
6075 (PMOVZXBWrm addr:$src)>;
6076 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
6077 (PMOVZXBWrm addr:$src)>;
6078 def : Pat<(int_x86_sse41_pmovzxbw (bc_v16i8 (loadv2i64 addr:$src))),
6079 (PMOVZXBWrm addr:$src)>;
6081 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
6082 (PMOVZXWDrm addr:$src)>;
6083 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
6084 (PMOVZXWDrm addr:$src)>;
6085 def : Pat<(int_x86_sse41_pmovzxwd (bc_v8i16 (loadv2i64 addr:$src))),
6086 (PMOVZXWDrm addr:$src)>;
6088 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
6089 (PMOVZXDQrm addr:$src)>;
6090 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
6091 (PMOVZXDQrm addr:$src)>;
6092 def : Pat<(int_x86_sse41_pmovzxdq (bc_v4i32 (loadv2i64 addr:$src))),
6093 (PMOVZXDQrm addr:$src)>;
6096 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId,
6097 OpndItins itins = DEFAULT_ITINS> {
6098 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
6099 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6100 [(set VR128:$dst, (IntId VR128:$src))], itins.rr>,
6101 Sched<[itins.Sched]>;
6103 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
6104 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6106 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))],
6107 itins.rm>, Sched<[itins.Sched.Folded]>;
6110 multiclass SS41I_binop_rm_int8_y<bits<8> opc, string OpcodeStr,
6111 Intrinsic IntId, X86FoldableSchedWrite Sched> {
6112 def Yrr : SS48I<opc, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
6113 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6114 [(set VR256:$dst, (IntId VR128:$src))]>, Sched<[Sched]>;
6116 def Yrm : SS48I<opc, MRMSrcMem, (outs VR256:$dst), (ins i32mem:$src),
6117 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6119 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
6120 Sched<[Sched.Folded]>;
6123 let Predicates = [HasAVX] in {
6124 defm VPMOVSXBD : SS41I_binop_rm_int4<0x21, "vpmovsxbd", int_x86_sse41_pmovsxbd,
6125 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
6126 defm VPMOVSXWQ : SS41I_binop_rm_int4<0x24, "vpmovsxwq", int_x86_sse41_pmovsxwq,
6127 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
6128 defm VPMOVZXBD : SS41I_binop_rm_int4<0x31, "vpmovzxbd", int_x86_sse41_pmovzxbd,
6129 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
6130 defm VPMOVZXWQ : SS41I_binop_rm_int4<0x34, "vpmovzxwq", int_x86_sse41_pmovzxwq,
6131 DEFAULT_ITINS_SHUFFLESCHED>, VEX;
6134 let Predicates = [HasAVX2] in {
6135 defm VPMOVSXBD : SS41I_binop_rm_int8_y<0x21, "vpmovsxbd",
6136 int_x86_avx2_pmovsxbd, WriteShuffle>,
6138 defm VPMOVSXWQ : SS41I_binop_rm_int8_y<0x24, "vpmovsxwq",
6139 int_x86_avx2_pmovsxwq, WriteShuffle>,
6141 defm VPMOVZXBD : SS41I_binop_rm_int8_y<0x31, "vpmovzxbd",
6142 int_x86_avx2_pmovzxbd, WriteShuffle>,
6144 defm VPMOVZXWQ : SS41I_binop_rm_int8_y<0x34, "vpmovzxwq",
6145 int_x86_avx2_pmovzxwq, WriteShuffle>,
6149 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd,
6150 SSE_INTALU_ITINS_SHUFF_P>;
6151 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq,
6152 SSE_INTALU_ITINS_SHUFF_P>;
6153 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd,
6154 SSE_INTALU_ITINS_SHUFF_P>;
6155 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq,
6156 SSE_INTALU_ITINS_SHUFF_P>;
6158 let Predicates = [HasAVX] in {
6159 // Common patterns involving scalar load
6160 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
6161 (VPMOVSXBDrm addr:$src)>;
6162 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
6163 (VPMOVSXWQrm addr:$src)>;
6165 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
6166 (VPMOVZXBDrm addr:$src)>;
6167 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
6168 (VPMOVZXWQrm addr:$src)>;
6171 let Predicates = [UseSSE41] in {
6172 // Common patterns involving scalar load
6173 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
6174 (PMOVSXBDrm addr:$src)>;
6175 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
6176 (PMOVSXWQrm addr:$src)>;
6178 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
6179 (PMOVZXBDrm addr:$src)>;
6180 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
6181 (PMOVZXWQrm addr:$src)>;
6184 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId,
6185 X86FoldableSchedWrite Sched> {
6186 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
6187 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6188 [(set VR128:$dst, (IntId VR128:$src))]>, Sched<[Sched]>;
6190 // Expecting a i16 load any extended to i32 value.
6191 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
6192 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6193 [(set VR128:$dst, (IntId (bitconvert
6194 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
6195 Sched<[Sched.Folded]>;
6198 multiclass SS41I_binop_rm_int4_y<bits<8> opc, string OpcodeStr,
6199 Intrinsic IntId, X86FoldableSchedWrite Sched> {
6200 def Yrr : SS48I<opc, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
6201 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6202 [(set VR256:$dst, (IntId VR128:$src))]>, Sched<[Sched]>;
6204 // Expecting a i16 load any extended to i32 value.
6205 def Yrm : SS48I<opc, MRMSrcMem, (outs VR256:$dst), (ins i16mem:$src),
6206 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6207 [(set VR256:$dst, (IntId (bitconvert
6208 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
6209 Sched<[Sched.Folded]>;
6212 let Predicates = [HasAVX] in {
6213 defm VPMOVSXBQ : SS41I_binop_rm_int2<0x22, "vpmovsxbq", int_x86_sse41_pmovsxbq,
6215 defm VPMOVZXBQ : SS41I_binop_rm_int2<0x32, "vpmovzxbq", int_x86_sse41_pmovzxbq,
6218 let Predicates = [HasAVX2] in {
6219 defm VPMOVSXBQ : SS41I_binop_rm_int4_y<0x22, "vpmovsxbq", int_x86_avx2_pmovsxbq,
6220 WriteShuffle>, VEX, VEX_L;
6221 defm VPMOVZXBQ : SS41I_binop_rm_int4_y<0x32, "vpmovzxbq", int_x86_avx2_pmovzxbq,
6222 WriteShuffle>, VEX, VEX_L;
6224 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq,
6226 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq,
6229 let Predicates = [HasAVX2] in {
6230 def : Pat<(v16i16 (X86vsext (v16i8 VR128:$src))), (VPMOVSXBWYrr VR128:$src)>;
6231 def : Pat<(v8i32 (X86vsext (v16i8 VR128:$src))), (VPMOVSXBDYrr VR128:$src)>;
6232 def : Pat<(v4i64 (X86vsext (v16i8 VR128:$src))), (VPMOVSXBQYrr VR128:$src)>;
6234 def : Pat<(v8i32 (X86vsext (v8i16 VR128:$src))), (VPMOVSXWDYrr VR128:$src)>;
6235 def : Pat<(v4i64 (X86vsext (v8i16 VR128:$src))), (VPMOVSXWQYrr VR128:$src)>;
6237 def : Pat<(v4i64 (X86vsext (v4i32 VR128:$src))), (VPMOVSXDQYrr VR128:$src)>;
6239 def : Pat<(v16i16 (X86vsext (v32i8 VR256:$src))),
6240 (VPMOVSXBWYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6241 def : Pat<(v8i32 (X86vsext (v32i8 VR256:$src))),
6242 (VPMOVSXBDYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6243 def : Pat<(v4i64 (X86vsext (v32i8 VR256:$src))),
6244 (VPMOVSXBQYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6246 def : Pat<(v8i32 (X86vsext (v16i16 VR256:$src))),
6247 (VPMOVSXWDYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6248 def : Pat<(v4i64 (X86vsext (v16i16 VR256:$src))),
6249 (VPMOVSXWQYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6251 def : Pat<(v4i64 (X86vsext (v8i32 VR256:$src))),
6252 (VPMOVSXDQYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6254 def : Pat<(v8i32 (X86vsext (v8i16 (bitconvert (v2i64 (load addr:$src)))))),
6255 (VPMOVSXWDYrm addr:$src)>;
6256 def : Pat<(v4i64 (X86vsext (v4i32 (bitconvert (v2i64 (load addr:$src)))))),
6257 (VPMOVSXDQYrm addr:$src)>;
6259 def : Pat<(v8i32 (X86vsext (v16i8 (bitconvert (v2i64
6260 (scalar_to_vector (loadi64 addr:$src))))))),
6261 (VPMOVSXBDYrm addr:$src)>;
6262 def : Pat<(v8i32 (X86vsext (v16i8 (bitconvert (v2f64
6263 (scalar_to_vector (loadf64 addr:$src))))))),
6264 (VPMOVSXBDYrm addr:$src)>;
6266 def : Pat<(v4i64 (X86vsext (v8i16 (bitconvert (v2i64
6267 (scalar_to_vector (loadi64 addr:$src))))))),
6268 (VPMOVSXWQYrm addr:$src)>;
6269 def : Pat<(v4i64 (X86vsext (v8i16 (bitconvert (v2f64
6270 (scalar_to_vector (loadf64 addr:$src))))))),
6271 (VPMOVSXWQYrm addr:$src)>;
6273 def : Pat<(v4i64 (X86vsext (v16i8 (bitconvert (v4i32
6274 (scalar_to_vector (loadi32 addr:$src))))))),
6275 (VPMOVSXBQYrm addr:$src)>;
6278 let Predicates = [HasAVX] in {
6279 // Common patterns involving scalar load
6280 def : Pat<(int_x86_sse41_pmovsxbq
6281 (bitconvert (v4i32 (X86vzmovl
6282 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6283 (VPMOVSXBQrm addr:$src)>;
6285 def : Pat<(int_x86_sse41_pmovzxbq
6286 (bitconvert (v4i32 (X86vzmovl
6287 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6288 (VPMOVZXBQrm addr:$src)>;
6291 let Predicates = [UseSSE41] in {
6292 def : Pat<(v8i16 (X86vsext (v16i8 VR128:$src))), (PMOVSXBWrr VR128:$src)>;
6293 def : Pat<(v4i32 (X86vsext (v16i8 VR128:$src))), (PMOVSXBDrr VR128:$src)>;
6294 def : Pat<(v2i64 (X86vsext (v16i8 VR128:$src))), (PMOVSXBQrr VR128:$src)>;
6296 def : Pat<(v4i32 (X86vsext (v8i16 VR128:$src))), (PMOVSXWDrr VR128:$src)>;
6297 def : Pat<(v2i64 (X86vsext (v8i16 VR128:$src))), (PMOVSXWQrr VR128:$src)>;
6299 def : Pat<(v2i64 (X86vsext (v4i32 VR128:$src))), (PMOVSXDQrr VR128:$src)>;
6301 // Common patterns involving scalar load
6302 def : Pat<(int_x86_sse41_pmovsxbq
6303 (bitconvert (v4i32 (X86vzmovl
6304 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6305 (PMOVSXBQrm addr:$src)>;
6307 def : Pat<(int_x86_sse41_pmovzxbq
6308 (bitconvert (v4i32 (X86vzmovl
6309 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6310 (PMOVZXBQrm addr:$src)>;
6312 def : Pat<(v4i32 (X86vsext (v8i16 (bitconvert (v2i64
6313 (scalar_to_vector (loadi64 addr:$src))))))),
6314 (PMOVSXWDrm addr:$src)>;
6315 def : Pat<(v4i32 (X86vsext (v8i16 (bitconvert (v2f64
6316 (scalar_to_vector (loadf64 addr:$src))))))),
6317 (PMOVSXWDrm addr:$src)>;
6318 def : Pat<(v4i32 (X86vsext (v16i8 (bitconvert (v4i32
6319 (scalar_to_vector (loadi32 addr:$src))))))),
6320 (PMOVSXBDrm addr:$src)>;
6321 def : Pat<(v2i64 (X86vsext (v8i16 (bitconvert (v4i32
6322 (scalar_to_vector (loadi32 addr:$src))))))),
6323 (PMOVSXWQrm addr:$src)>;
6324 def : Pat<(v2i64 (X86vsext (v16i8 (bitconvert (v4i32
6325 (scalar_to_vector (extloadi32i16 addr:$src))))))),
6326 (PMOVSXBQrm addr:$src)>;
6327 def : Pat<(v2i64 (X86vsext (v4i32 (bitconvert (v2i64
6328 (scalar_to_vector (loadi64 addr:$src))))))),
6329 (PMOVSXDQrm addr:$src)>;
6330 def : Pat<(v2i64 (X86vsext (v4i32 (bitconvert (v2f64
6331 (scalar_to_vector (loadf64 addr:$src))))))),
6332 (PMOVSXDQrm addr:$src)>;
6333 def : Pat<(v8i16 (X86vsext (v16i8 (bitconvert (v2i64
6334 (scalar_to_vector (loadi64 addr:$src))))))),
6335 (PMOVSXBWrm addr:$src)>;
6336 def : Pat<(v8i16 (X86vsext (v16i8 (bitconvert (v2f64
6337 (scalar_to_vector (loadf64 addr:$src))))))),
6338 (PMOVSXBWrm addr:$src)>;
6341 let Predicates = [HasAVX2] in {
6342 def : Pat<(v16i16 (X86vzext (v16i8 VR128:$src))), (VPMOVZXBWYrr VR128:$src)>;
6343 def : Pat<(v8i32 (X86vzext (v16i8 VR128:$src))), (VPMOVZXBDYrr VR128:$src)>;
6344 def : Pat<(v4i64 (X86vzext (v16i8 VR128:$src))), (VPMOVZXBQYrr VR128:$src)>;
6346 def : Pat<(v8i32 (X86vzext (v8i16 VR128:$src))), (VPMOVZXWDYrr VR128:$src)>;
6347 def : Pat<(v4i64 (X86vzext (v8i16 VR128:$src))), (VPMOVZXWQYrr VR128:$src)>;
6349 def : Pat<(v4i64 (X86vzext (v4i32 VR128:$src))), (VPMOVZXDQYrr VR128:$src)>;
6351 def : Pat<(v16i16 (X86vzext (v32i8 VR256:$src))),
6352 (VPMOVZXBWYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6353 def : Pat<(v8i32 (X86vzext (v32i8 VR256:$src))),
6354 (VPMOVZXBDYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6355 def : Pat<(v4i64 (X86vzext (v32i8 VR256:$src))),
6356 (VPMOVZXBQYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6358 def : Pat<(v8i32 (X86vzext (v16i16 VR256:$src))),
6359 (VPMOVZXWDYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6360 def : Pat<(v4i64 (X86vzext (v16i16 VR256:$src))),
6361 (VPMOVZXWQYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6363 def : Pat<(v4i64 (X86vzext (v8i32 VR256:$src))),
6364 (VPMOVZXDQYrr (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
6367 let Predicates = [HasAVX] in {
6368 def : Pat<(v8i16 (X86vzext (v16i8 VR128:$src))), (VPMOVZXBWrr VR128:$src)>;
6369 def : Pat<(v4i32 (X86vzext (v16i8 VR128:$src))), (VPMOVZXBDrr VR128:$src)>;
6370 def : Pat<(v2i64 (X86vzext (v16i8 VR128:$src))), (VPMOVZXBQrr VR128:$src)>;
6372 def : Pat<(v4i32 (X86vzext (v8i16 VR128:$src))), (VPMOVZXWDrr VR128:$src)>;
6373 def : Pat<(v2i64 (X86vzext (v8i16 VR128:$src))), (VPMOVZXWQrr VR128:$src)>;
6375 def : Pat<(v2i64 (X86vzext (v4i32 VR128:$src))), (VPMOVZXDQrr VR128:$src)>;
6377 def : Pat<(v8i16 (X86vzext (v16i8 (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
6378 (VPMOVZXBWrm addr:$src)>;
6379 def : Pat<(v8i16 (X86vzext (v16i8 (bitconvert (v2f64 (scalar_to_vector (loadf64 addr:$src))))))),
6380 (VPMOVZXBWrm addr:$src)>;
6381 def : Pat<(v4i32 (X86vzext (v16i8 (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6382 (VPMOVZXBDrm addr:$src)>;
6383 def : Pat<(v2i64 (X86vzext (v16i8 (bitconvert (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))),
6384 (VPMOVZXBQrm addr:$src)>;
6386 def : Pat<(v4i32 (X86vzext (v8i16 (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
6387 (VPMOVZXWDrm addr:$src)>;
6388 def : Pat<(v4i32 (X86vzext (v8i16 (bitconvert (v2f64 (scalar_to_vector (loadf64 addr:$src))))))),
6389 (VPMOVZXWDrm addr:$src)>;
6390 def : Pat<(v2i64 (X86vzext (v8i16 (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6391 (VPMOVZXWQrm addr:$src)>;
6393 def : Pat<(v2i64 (X86vzext (v4i32 (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
6394 (VPMOVZXDQrm addr:$src)>;
6395 def : Pat<(v2i64 (X86vzext (v4i32 (bitconvert (v2f64 (scalar_to_vector (loadf64 addr:$src))))))),
6396 (VPMOVZXDQrm addr:$src)>;
6397 def : Pat<(v2i64 (X86vzext (v4i32 (bitconvert (v2i64 (X86vzload addr:$src)))))),
6398 (VPMOVZXDQrm addr:$src)>;
6400 def : Pat<(v8i16 (X86vsext (v16i8 VR128:$src))), (VPMOVSXBWrr VR128:$src)>;
6401 def : Pat<(v4i32 (X86vsext (v16i8 VR128:$src))), (VPMOVSXBDrr VR128:$src)>;
6402 def : Pat<(v2i64 (X86vsext (v16i8 VR128:$src))), (VPMOVSXBQrr VR128:$src)>;
6404 def : Pat<(v4i32 (X86vsext (v8i16 VR128:$src))), (VPMOVSXWDrr VR128:$src)>;
6405 def : Pat<(v2i64 (X86vsext (v8i16 VR128:$src))), (VPMOVSXWQrr VR128:$src)>;
6407 def : Pat<(v2i64 (X86vsext (v4i32 VR128:$src))), (VPMOVSXDQrr VR128:$src)>;
6409 def : Pat<(v4i32 (X86vsext (v8i16 (bitconvert (v2i64
6410 (scalar_to_vector (loadi64 addr:$src))))))),
6411 (VPMOVSXWDrm addr:$src)>;
6412 def : Pat<(v2i64 (X86vsext (v4i32 (bitconvert (v2i64
6413 (scalar_to_vector (loadi64 addr:$src))))))),
6414 (VPMOVSXDQrm addr:$src)>;
6415 def : Pat<(v4i32 (X86vsext (v8i16 (bitconvert (v2f64
6416 (scalar_to_vector (loadf64 addr:$src))))))),
6417 (VPMOVSXWDrm addr:$src)>;
6418 def : Pat<(v2i64 (X86vsext (v4i32 (bitconvert (v2f64
6419 (scalar_to_vector (loadf64 addr:$src))))))),
6420 (VPMOVSXDQrm addr:$src)>;
6421 def : Pat<(v8i16 (X86vsext (v16i8 (bitconvert (v2i64
6422 (scalar_to_vector (loadi64 addr:$src))))))),
6423 (VPMOVSXBWrm addr:$src)>;
6424 def : Pat<(v8i16 (X86vsext (v16i8 (bitconvert (v2f64
6425 (scalar_to_vector (loadf64 addr:$src))))))),
6426 (VPMOVSXBWrm addr:$src)>;
6428 def : Pat<(v4i32 (X86vsext (v16i8 (bitconvert (v4i32
6429 (scalar_to_vector (loadi32 addr:$src))))))),
6430 (VPMOVSXBDrm addr:$src)>;
6431 def : Pat<(v2i64 (X86vsext (v8i16 (bitconvert (v4i32
6432 (scalar_to_vector (loadi32 addr:$src))))))),
6433 (VPMOVSXWQrm addr:$src)>;
6434 def : Pat<(v2i64 (X86vsext (v16i8 (bitconvert (v4i32
6435 (scalar_to_vector (extloadi32i16 addr:$src))))))),
6436 (VPMOVSXBQrm addr:$src)>;
6439 let Predicates = [UseSSE41] in {
6440 def : Pat<(v8i16 (X86vzext (v16i8 VR128:$src))), (PMOVZXBWrr VR128:$src)>;
6441 def : Pat<(v4i32 (X86vzext (v16i8 VR128:$src))), (PMOVZXBDrr VR128:$src)>;
6442 def : Pat<(v2i64 (X86vzext (v16i8 VR128:$src))), (PMOVZXBQrr VR128:$src)>;
6444 def : Pat<(v4i32 (X86vzext (v8i16 VR128:$src))), (PMOVZXWDrr VR128:$src)>;
6445 def : Pat<(v2i64 (X86vzext (v8i16 VR128:$src))), (PMOVZXWQrr VR128:$src)>;
6447 def : Pat<(v2i64 (X86vzext (v4i32 VR128:$src))), (PMOVZXDQrr VR128:$src)>;
6449 def : Pat<(v8i16 (X86vzext (v16i8 (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
6450 (PMOVZXBWrm addr:$src)>;
6451 def : Pat<(v8i16 (X86vzext (v16i8 (bitconvert (v2f64 (scalar_to_vector (loadf64 addr:$src))))))),
6452 (PMOVZXBWrm addr:$src)>;
6453 def : Pat<(v4i32 (X86vzext (v16i8 (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6454 (PMOVZXBDrm addr:$src)>;
6455 def : Pat<(v2i64 (X86vzext (v16i8 (bitconvert (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))),
6456 (PMOVZXBQrm addr:$src)>;
6458 def : Pat<(v4i32 (X86vzext (v8i16 (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
6459 (PMOVZXWDrm addr:$src)>;
6460 def : Pat<(v4i32 (X86vzext (v8i16 (bitconvert (v2f64 (scalar_to_vector (loadf64 addr:$src))))))),
6461 (PMOVZXWDrm addr:$src)>;
6462 def : Pat<(v2i64 (X86vzext (v8i16 (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
6463 (PMOVZXWQrm addr:$src)>;
6465 def : Pat<(v2i64 (X86vzext (v4i32 (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
6466 (PMOVZXDQrm addr:$src)>;
6467 def : Pat<(v2i64 (X86vzext (v4i32 (bitconvert (v2f64 (scalar_to_vector (loadf64 addr:$src))))))),
6468 (PMOVZXDQrm addr:$src)>;
6469 def : Pat<(v2i64 (X86vzext (v4i32 (bitconvert (v2i64 (X86vzload addr:$src)))))),
6470 (PMOVZXDQrm addr:$src)>;
6473 //===----------------------------------------------------------------------===//
6474 // SSE4.1 - Extract Instructions
6475 //===----------------------------------------------------------------------===//
6477 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
6478 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
6479 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
6480 (ins VR128:$src1, i32i8imm:$src2),
6481 !strconcat(OpcodeStr,
6482 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6483 [(set GR32orGR64:$dst, (X86pextrb (v16i8 VR128:$src1),
6485 Sched<[WriteShuffle]>;
6486 let neverHasSideEffects = 1, mayStore = 1,
6487 SchedRW = [WriteShuffleLd, WriteRMW] in
6488 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6489 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
6490 !strconcat(OpcodeStr,
6491 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6492 [(store (i8 (trunc (assertzext (X86pextrb (v16i8 VR128:$src1),
6493 imm:$src2)))), addr:$dst)]>;
6496 let Predicates = [HasAVX] in
6497 defm VPEXTRB : SS41I_extract8<0x14, "vpextrb">, VEX;
6499 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
6502 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
6503 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
6504 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in
6505 def rr_REV : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
6506 (ins VR128:$src1, i32i8imm:$src2),
6507 !strconcat(OpcodeStr,
6508 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6509 []>, Sched<[WriteShuffle]>;
6511 let neverHasSideEffects = 1, mayStore = 1,
6512 SchedRW = [WriteShuffleLd, WriteRMW] in
6513 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6514 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
6515 !strconcat(OpcodeStr,
6516 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6517 [(store (i16 (trunc (assertzext (X86pextrw (v8i16 VR128:$src1),
6518 imm:$src2)))), addr:$dst)]>;
6521 let Predicates = [HasAVX] in
6522 defm VPEXTRW : SS41I_extract16<0x15, "vpextrw">, VEX;
6524 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
6527 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
6528 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
6529 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
6530 (ins VR128:$src1, i32i8imm:$src2),
6531 !strconcat(OpcodeStr,
6532 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6534 (extractelt (v4i32 VR128:$src1), imm:$src2))]>,
6535 Sched<[WriteShuffle]>;
6536 let SchedRW = [WriteShuffleLd, WriteRMW] in
6537 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6538 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
6539 !strconcat(OpcodeStr,
6540 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6541 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
6545 let Predicates = [HasAVX] in
6546 defm VPEXTRD : SS41I_extract32<0x16, "vpextrd">, VEX;
6548 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
6550 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
6551 multiclass SS41I_extract64<bits<8> opc, string OpcodeStr> {
6552 def rr : SS4AIi8<opc, MRMDestReg, (outs GR64:$dst),
6553 (ins VR128:$src1, i32i8imm:$src2),
6554 !strconcat(OpcodeStr,
6555 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6557 (extractelt (v2i64 VR128:$src1), imm:$src2))]>,
6558 Sched<[WriteShuffle]>, REX_W;
6559 let SchedRW = [WriteShuffleLd, WriteRMW] in
6560 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6561 (ins i64mem:$dst, VR128:$src1, i32i8imm:$src2),
6562 !strconcat(OpcodeStr,
6563 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6564 [(store (extractelt (v2i64 VR128:$src1), imm:$src2),
6565 addr:$dst)]>, REX_W;
6568 let Predicates = [HasAVX] in
6569 defm VPEXTRQ : SS41I_extract64<0x16, "vpextrq">, VEX, VEX_W;
6571 defm PEXTRQ : SS41I_extract64<0x16, "pextrq">;
6573 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
6575 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr,
6576 OpndItins itins = DEFAULT_ITINS> {
6577 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
6578 (ins VR128:$src1, i32i8imm:$src2),
6579 !strconcat(OpcodeStr,
6580 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6581 [(set GR32orGR64:$dst,
6582 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))],
6583 itins.rr>, Sched<[WriteFBlend]>;
6584 let SchedRW = [WriteFBlendLd, WriteRMW] in
6585 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6586 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
6587 !strconcat(OpcodeStr,
6588 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6589 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
6590 addr:$dst)], itins.rm>;
6593 let ExeDomain = SSEPackedSingle in {
6594 let Predicates = [UseAVX] in
6595 defm VEXTRACTPS : SS41I_extractf32<0x17, "vextractps">, VEX;
6596 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps", SSE_EXTRACT_ITINS>;
6599 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
6600 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
6603 (VEXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
6605 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
6608 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
6609 Requires<[UseSSE41]>;
6611 //===----------------------------------------------------------------------===//
6612 // SSE4.1 - Insert Instructions
6613 //===----------------------------------------------------------------------===//
6615 multiclass SS41I_insert8<bits<8> opc, string asm, bit Is2Addr = 1> {
6616 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6617 (ins VR128:$src1, GR32orGR64:$src2, i32i8imm:$src3),
6619 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6621 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6623 (X86pinsrb VR128:$src1, GR32orGR64:$src2, imm:$src3))]>,
6624 Sched<[WriteShuffle]>;
6625 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6626 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
6628 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6630 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6632 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
6633 imm:$src3))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6636 let Predicates = [HasAVX] in
6637 defm VPINSRB : SS41I_insert8<0x20, "vpinsrb", 0>, VEX_4V;
6638 let Constraints = "$src1 = $dst" in
6639 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
6641 multiclass SS41I_insert32<bits<8> opc, string asm, bit Is2Addr = 1> {
6642 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6643 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
6645 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6647 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6649 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
6650 Sched<[WriteShuffle]>;
6651 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6652 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
6654 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6656 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6658 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
6659 imm:$src3)))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6662 let Predicates = [HasAVX] in
6663 defm VPINSRD : SS41I_insert32<0x22, "vpinsrd", 0>, VEX_4V;
6664 let Constraints = "$src1 = $dst" in
6665 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
6667 multiclass SS41I_insert64<bits<8> opc, string asm, bit Is2Addr = 1> {
6668 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6669 (ins VR128:$src1, GR64:$src2, i32i8imm:$src3),
6671 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6673 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6675 (v2i64 (insertelt VR128:$src1, GR64:$src2, imm:$src3)))]>,
6676 Sched<[WriteShuffle]>;
6677 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6678 (ins VR128:$src1, i64mem:$src2, i32i8imm:$src3),
6680 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6682 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6684 (v2i64 (insertelt VR128:$src1, (loadi64 addr:$src2),
6685 imm:$src3)))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6688 let Predicates = [HasAVX] in
6689 defm VPINSRQ : SS41I_insert64<0x22, "vpinsrq", 0>, VEX_4V, VEX_W;
6690 let Constraints = "$src1 = $dst" in
6691 defm PINSRQ : SS41I_insert64<0x22, "pinsrq">, REX_W;
6693 // insertps has a few different modes, there's the first two here below which
6694 // are optimized inserts that won't zero arbitrary elements in the destination
6695 // vector. The next one matches the intrinsic and could zero arbitrary elements
6696 // in the target vector.
6697 multiclass SS41I_insertf32<bits<8> opc, string asm, bit Is2Addr = 1,
6698 OpndItins itins = DEFAULT_ITINS> {
6699 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6700 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
6702 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6704 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6706 (X86insertps VR128:$src1, VR128:$src2, imm:$src3))], itins.rr>,
6707 Sched<[WriteFShuffle]>;
6708 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6709 (ins VR128:$src1, f32mem:$src2, i8imm:$src3),
6711 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6713 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6715 (X86insertps VR128:$src1,
6716 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
6717 imm:$src3))], itins.rm>,
6718 Sched<[WriteFShuffleLd, ReadAfterLd]>;
6721 let ExeDomain = SSEPackedSingle in {
6722 let Predicates = [UseAVX] in
6723 defm VINSERTPS : SS41I_insertf32<0x21, "vinsertps", 0>, VEX_4V;
6724 let Constraints = "$src1 = $dst" in
6725 defm INSERTPS : SS41I_insertf32<0x21, "insertps", 1, SSE_INSERT_ITINS>;
6728 let Predicates = [UseSSE41] in {
6729 // If we're inserting an element from a load or a null pshuf of a load,
6730 // fold the load into the insertps instruction.
6731 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd (v4f32
6732 (scalar_to_vector (loadf32 addr:$src2))), (i8 0)),
6734 (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6735 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd
6736 (loadv4f32 addr:$src2), (i8 0)), imm:$src3)),
6737 (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6740 let Predicates = [UseAVX] in {
6741 // If we're inserting an element from a vbroadcast of a load, fold the
6742 // load into the X86insertps instruction.
6743 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
6744 (X86VBroadcast (loadf32 addr:$src2)), imm:$src3)),
6745 (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6746 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
6747 (X86VBroadcast (loadv4f32 addr:$src2)), imm:$src3)),
6748 (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6751 //===----------------------------------------------------------------------===//
6752 // SSE4.1 - Round Instructions
6753 //===----------------------------------------------------------------------===//
6755 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd, string OpcodeStr,
6756 X86MemOperand x86memop, RegisterClass RC,
6757 PatFrag mem_frag32, PatFrag mem_frag64,
6758 Intrinsic V4F32Int, Intrinsic V2F64Int> {
6759 let ExeDomain = SSEPackedSingle in {
6760 // Intrinsic operation, reg.
6761 // Vector intrinsic operation, reg
6762 def PSr : SS4AIi8<opcps, MRMSrcReg,
6763 (outs RC:$dst), (ins RC:$src1, i32i8imm:$src2),
6764 !strconcat(OpcodeStr,
6765 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6766 [(set RC:$dst, (V4F32Int RC:$src1, imm:$src2))],
6767 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAdd]>;
6769 // Vector intrinsic operation, mem
6770 def PSm : SS4AIi8<opcps, MRMSrcMem,
6771 (outs RC:$dst), (ins x86memop:$src1, i32i8imm:$src2),
6772 !strconcat(OpcodeStr,
6773 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6775 (V4F32Int (mem_frag32 addr:$src1),imm:$src2))],
6776 IIC_SSE_ROUNDPS_MEM>, Sched<[WriteFAddLd]>;
6777 } // ExeDomain = SSEPackedSingle
6779 let ExeDomain = SSEPackedDouble in {
6780 // Vector intrinsic operation, reg
6781 def PDr : SS4AIi8<opcpd, MRMSrcReg,
6782 (outs RC:$dst), (ins RC:$src1, i32i8imm:$src2),
6783 !strconcat(OpcodeStr,
6784 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6785 [(set RC:$dst, (V2F64Int RC:$src1, imm:$src2))],
6786 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAdd]>;
6788 // Vector intrinsic operation, mem
6789 def PDm : SS4AIi8<opcpd, MRMSrcMem,
6790 (outs RC:$dst), (ins x86memop:$src1, i32i8imm:$src2),
6791 !strconcat(OpcodeStr,
6792 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6794 (V2F64Int (mem_frag64 addr:$src1),imm:$src2))],
6795 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAddLd]>;
6796 } // ExeDomain = SSEPackedDouble
6799 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
6802 Intrinsic F64Int, bit Is2Addr = 1> {
6803 let ExeDomain = GenericDomain in {
6805 let hasSideEffects = 0 in
6806 def SSr : SS4AIi8<opcss, MRMSrcReg,
6807 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2, i32i8imm:$src3),
6809 !strconcat(OpcodeStr,
6810 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6811 !strconcat(OpcodeStr,
6812 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6813 []>, Sched<[WriteFAdd]>;
6815 // Intrinsic operation, reg.
6816 let isCodeGenOnly = 1 in
6817 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
6818 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
6820 !strconcat(OpcodeStr,
6821 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6822 !strconcat(OpcodeStr,
6823 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6824 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
6827 // Intrinsic operation, mem.
6828 def SSm : SS4AIi8<opcss, MRMSrcMem,
6829 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
6831 !strconcat(OpcodeStr,
6832 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6833 !strconcat(OpcodeStr,
6834 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6836 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
6837 Sched<[WriteFAddLd, ReadAfterLd]>;
6840 let hasSideEffects = 0 in
6841 def SDr : SS4AIi8<opcsd, MRMSrcReg,
6842 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2, i32i8imm:$src3),
6844 !strconcat(OpcodeStr,
6845 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6846 !strconcat(OpcodeStr,
6847 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6848 []>, Sched<[WriteFAdd]>;
6850 // Intrinsic operation, reg.
6851 let isCodeGenOnly = 1 in
6852 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
6853 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
6855 !strconcat(OpcodeStr,
6856 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6857 !strconcat(OpcodeStr,
6858 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6859 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
6862 // Intrinsic operation, mem.
6863 def SDm : SS4AIi8<opcsd, MRMSrcMem,
6864 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
6866 !strconcat(OpcodeStr,
6867 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6868 !strconcat(OpcodeStr,
6869 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6871 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
6872 Sched<[WriteFAddLd, ReadAfterLd]>;
6873 } // ExeDomain = GenericDomain
6876 // FP round - roundss, roundps, roundsd, roundpd
6877 let Predicates = [HasAVX] in {
6879 defm VROUND : sse41_fp_unop_rm<0x08, 0x09, "vround", f128mem, VR128,
6880 loadv4f32, loadv2f64,
6881 int_x86_sse41_round_ps,
6882 int_x86_sse41_round_pd>, VEX;
6883 defm VROUNDY : sse41_fp_unop_rm<0x08, 0x09, "vround", f256mem, VR256,
6884 loadv8f32, loadv4f64,
6885 int_x86_avx_round_ps_256,
6886 int_x86_avx_round_pd_256>, VEX, VEX_L;
6887 defm VROUND : sse41_fp_binop_rm<0x0A, 0x0B, "vround",
6888 int_x86_sse41_round_ss,
6889 int_x86_sse41_round_sd, 0>, VEX_4V, VEX_LIG;
6891 def : Pat<(ffloor FR32:$src),
6892 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x1))>;
6893 def : Pat<(f64 (ffloor FR64:$src)),
6894 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x1))>;
6895 def : Pat<(f32 (fnearbyint FR32:$src)),
6896 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0xC))>;
6897 def : Pat<(f64 (fnearbyint FR64:$src)),
6898 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0xC))>;
6899 def : Pat<(f32 (fceil FR32:$src)),
6900 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x2))>;
6901 def : Pat<(f64 (fceil FR64:$src)),
6902 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x2))>;
6903 def : Pat<(f32 (frint FR32:$src)),
6904 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x4))>;
6905 def : Pat<(f64 (frint FR64:$src)),
6906 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x4))>;
6907 def : Pat<(f32 (ftrunc FR32:$src)),
6908 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x3))>;
6909 def : Pat<(f64 (ftrunc FR64:$src)),
6910 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x3))>;
6912 def : Pat<(v4f32 (ffloor VR128:$src)),
6913 (VROUNDPSr VR128:$src, (i32 0x1))>;
6914 def : Pat<(v4f32 (fnearbyint VR128:$src)),
6915 (VROUNDPSr VR128:$src, (i32 0xC))>;
6916 def : Pat<(v4f32 (fceil VR128:$src)),
6917 (VROUNDPSr VR128:$src, (i32 0x2))>;
6918 def : Pat<(v4f32 (frint VR128:$src)),
6919 (VROUNDPSr VR128:$src, (i32 0x4))>;
6920 def : Pat<(v4f32 (ftrunc VR128:$src)),
6921 (VROUNDPSr VR128:$src, (i32 0x3))>;
6923 def : Pat<(v2f64 (ffloor VR128:$src)),
6924 (VROUNDPDr VR128:$src, (i32 0x1))>;
6925 def : Pat<(v2f64 (fnearbyint VR128:$src)),
6926 (VROUNDPDr VR128:$src, (i32 0xC))>;
6927 def : Pat<(v2f64 (fceil VR128:$src)),
6928 (VROUNDPDr VR128:$src, (i32 0x2))>;
6929 def : Pat<(v2f64 (frint VR128:$src)),
6930 (VROUNDPDr VR128:$src, (i32 0x4))>;
6931 def : Pat<(v2f64 (ftrunc VR128:$src)),
6932 (VROUNDPDr VR128:$src, (i32 0x3))>;
6934 def : Pat<(v8f32 (ffloor VR256:$src)),
6935 (VROUNDYPSr VR256:$src, (i32 0x1))>;
6936 def : Pat<(v8f32 (fnearbyint VR256:$src)),
6937 (VROUNDYPSr VR256:$src, (i32 0xC))>;
6938 def : Pat<(v8f32 (fceil VR256:$src)),
6939 (VROUNDYPSr VR256:$src, (i32 0x2))>;
6940 def : Pat<(v8f32 (frint VR256:$src)),
6941 (VROUNDYPSr VR256:$src, (i32 0x4))>;
6942 def : Pat<(v8f32 (ftrunc VR256:$src)),
6943 (VROUNDYPSr VR256:$src, (i32 0x3))>;
6945 def : Pat<(v4f64 (ffloor VR256:$src)),
6946 (VROUNDYPDr VR256:$src, (i32 0x1))>;
6947 def : Pat<(v4f64 (fnearbyint VR256:$src)),
6948 (VROUNDYPDr VR256:$src, (i32 0xC))>;
6949 def : Pat<(v4f64 (fceil VR256:$src)),
6950 (VROUNDYPDr VR256:$src, (i32 0x2))>;
6951 def : Pat<(v4f64 (frint VR256:$src)),
6952 (VROUNDYPDr VR256:$src, (i32 0x4))>;
6953 def : Pat<(v4f64 (ftrunc VR256:$src)),
6954 (VROUNDYPDr VR256:$src, (i32 0x3))>;
6957 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round", f128mem, VR128,
6958 memopv4f32, memopv2f64,
6959 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
6960 let Constraints = "$src1 = $dst" in
6961 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
6962 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
6964 let Predicates = [UseSSE41] in {
6965 def : Pat<(ffloor FR32:$src),
6966 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x1))>;
6967 def : Pat<(f64 (ffloor FR64:$src)),
6968 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x1))>;
6969 def : Pat<(f32 (fnearbyint FR32:$src)),
6970 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0xC))>;
6971 def : Pat<(f64 (fnearbyint FR64:$src)),
6972 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0xC))>;
6973 def : Pat<(f32 (fceil FR32:$src)),
6974 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x2))>;
6975 def : Pat<(f64 (fceil FR64:$src)),
6976 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x2))>;
6977 def : Pat<(f32 (frint FR32:$src)),
6978 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x4))>;
6979 def : Pat<(f64 (frint FR64:$src)),
6980 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x4))>;
6981 def : Pat<(f32 (ftrunc FR32:$src)),
6982 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x3))>;
6983 def : Pat<(f64 (ftrunc FR64:$src)),
6984 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x3))>;
6986 def : Pat<(v4f32 (ffloor VR128:$src)),
6987 (ROUNDPSr VR128:$src, (i32 0x1))>;
6988 def : Pat<(v4f32 (fnearbyint VR128:$src)),
6989 (ROUNDPSr VR128:$src, (i32 0xC))>;
6990 def : Pat<(v4f32 (fceil VR128:$src)),
6991 (ROUNDPSr VR128:$src, (i32 0x2))>;
6992 def : Pat<(v4f32 (frint VR128:$src)),
6993 (ROUNDPSr VR128:$src, (i32 0x4))>;
6994 def : Pat<(v4f32 (ftrunc VR128:$src)),
6995 (ROUNDPSr VR128:$src, (i32 0x3))>;
6997 def : Pat<(v2f64 (ffloor VR128:$src)),
6998 (ROUNDPDr VR128:$src, (i32 0x1))>;
6999 def : Pat<(v2f64 (fnearbyint VR128:$src)),
7000 (ROUNDPDr VR128:$src, (i32 0xC))>;
7001 def : Pat<(v2f64 (fceil VR128:$src)),
7002 (ROUNDPDr VR128:$src, (i32 0x2))>;
7003 def : Pat<(v2f64 (frint VR128:$src)),
7004 (ROUNDPDr VR128:$src, (i32 0x4))>;
7005 def : Pat<(v2f64 (ftrunc VR128:$src)),
7006 (ROUNDPDr VR128:$src, (i32 0x3))>;
7009 //===----------------------------------------------------------------------===//
7010 // SSE4.1 - Packed Bit Test
7011 //===----------------------------------------------------------------------===//
7013 // ptest instruction we'll lower to this in X86ISelLowering primarily from
7014 // the intel intrinsic that corresponds to this.
7015 let Defs = [EFLAGS], Predicates = [HasAVX] in {
7016 def VPTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
7017 "vptest\t{$src2, $src1|$src1, $src2}",
7018 [(set EFLAGS, (X86ptest VR128:$src1, (v2i64 VR128:$src2)))]>,
7019 Sched<[WriteVecLogic]>, VEX;
7020 def VPTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
7021 "vptest\t{$src2, $src1|$src1, $src2}",
7022 [(set EFLAGS,(X86ptest VR128:$src1, (loadv2i64 addr:$src2)))]>,
7023 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX;
7025 def VPTESTYrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR256:$src1, VR256:$src2),
7026 "vptest\t{$src2, $src1|$src1, $src2}",
7027 [(set EFLAGS, (X86ptest VR256:$src1, (v4i64 VR256:$src2)))]>,
7028 Sched<[WriteVecLogic]>, VEX, VEX_L;
7029 def VPTESTYrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR256:$src1, i256mem:$src2),
7030 "vptest\t{$src2, $src1|$src1, $src2}",
7031 [(set EFLAGS,(X86ptest VR256:$src1, (loadv4i64 addr:$src2)))]>,
7032 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX, VEX_L;
7035 let Defs = [EFLAGS] in {
7036 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
7037 "ptest\t{$src2, $src1|$src1, $src2}",
7038 [(set EFLAGS, (X86ptest VR128:$src1, (v2i64 VR128:$src2)))]>,
7039 Sched<[WriteVecLogic]>;
7040 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
7041 "ptest\t{$src2, $src1|$src1, $src2}",
7042 [(set EFLAGS, (X86ptest VR128:$src1, (memopv2i64 addr:$src2)))]>,
7043 Sched<[WriteVecLogicLd, ReadAfterLd]>;
7046 // The bit test instructions below are AVX only
7047 multiclass avx_bittest<bits<8> opc, string OpcodeStr, RegisterClass RC,
7048 X86MemOperand x86memop, PatFrag mem_frag, ValueType vt> {
7049 def rr : SS48I<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
7050 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
7051 [(set EFLAGS, (X86testp RC:$src1, (vt RC:$src2)))]>,
7052 Sched<[WriteVecLogic]>, VEX;
7053 def rm : SS48I<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
7054 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
7055 [(set EFLAGS, (X86testp RC:$src1, (mem_frag addr:$src2)))]>,
7056 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX;
7059 let Defs = [EFLAGS], Predicates = [HasAVX] in {
7060 let ExeDomain = SSEPackedSingle in {
7061 defm VTESTPS : avx_bittest<0x0E, "vtestps", VR128, f128mem, loadv4f32, v4f32>;
7062 defm VTESTPSY : avx_bittest<0x0E, "vtestps", VR256, f256mem, loadv8f32, v8f32>,
7065 let ExeDomain = SSEPackedDouble in {
7066 defm VTESTPD : avx_bittest<0x0F, "vtestpd", VR128, f128mem, loadv2f64, v2f64>;
7067 defm VTESTPDY : avx_bittest<0x0F, "vtestpd", VR256, f256mem, loadv4f64, v4f64>,
7072 //===----------------------------------------------------------------------===//
7073 // SSE4.1 - Misc Instructions
7074 //===----------------------------------------------------------------------===//
7076 let Defs = [EFLAGS], Predicates = [HasPOPCNT] in {
7077 def POPCNT16rr : I<0xB8, MRMSrcReg, (outs GR16:$dst), (ins GR16:$src),
7078 "popcnt{w}\t{$src, $dst|$dst, $src}",
7079 [(set GR16:$dst, (ctpop GR16:$src)), (implicit EFLAGS)],
7080 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>,
7082 def POPCNT16rm : I<0xB8, MRMSrcMem, (outs GR16:$dst), (ins i16mem:$src),
7083 "popcnt{w}\t{$src, $dst|$dst, $src}",
7084 [(set GR16:$dst, (ctpop (loadi16 addr:$src))),
7085 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
7086 Sched<[WriteFAddLd]>, OpSize16, XS;
7088 def POPCNT32rr : I<0xB8, MRMSrcReg, (outs GR32:$dst), (ins GR32:$src),
7089 "popcnt{l}\t{$src, $dst|$dst, $src}",
7090 [(set GR32:$dst, (ctpop GR32:$src)), (implicit EFLAGS)],
7091 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>,
7094 def POPCNT32rm : I<0xB8, MRMSrcMem, (outs GR32:$dst), (ins i32mem:$src),
7095 "popcnt{l}\t{$src, $dst|$dst, $src}",
7096 [(set GR32:$dst, (ctpop (loadi32 addr:$src))),
7097 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
7098 Sched<[WriteFAddLd]>, OpSize32, XS;
7100 def POPCNT64rr : RI<0xB8, MRMSrcReg, (outs GR64:$dst), (ins GR64:$src),
7101 "popcnt{q}\t{$src, $dst|$dst, $src}",
7102 [(set GR64:$dst, (ctpop GR64:$src)), (implicit EFLAGS)],
7103 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>, XS;
7104 def POPCNT64rm : RI<0xB8, MRMSrcMem, (outs GR64:$dst), (ins i64mem:$src),
7105 "popcnt{q}\t{$src, $dst|$dst, $src}",
7106 [(set GR64:$dst, (ctpop (loadi64 addr:$src))),
7107 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
7108 Sched<[WriteFAddLd]>, XS;
7113 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
7114 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
7116 X86FoldableSchedWrite Sched> {
7117 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
7119 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
7120 [(set VR128:$dst, (IntId128 VR128:$src))]>,
7122 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
7124 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
7126 (IntId128 (bitconvert (memopv2i64 addr:$src))))]>,
7127 Sched<[Sched.Folded]>;
7130 // PHMIN has the same profile as PSAD, thus we use the same scheduling
7131 // model, although the naming is misleading.
7132 let Predicates = [HasAVX] in
7133 defm VPHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "vphminposuw",
7134 int_x86_sse41_phminposuw,
7136 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
7137 int_x86_sse41_phminposuw,
7140 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
7141 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
7142 Intrinsic IntId128, bit Is2Addr = 1,
7143 OpndItins itins = DEFAULT_ITINS> {
7144 let isCommutable = 1 in
7145 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
7146 (ins VR128:$src1, VR128:$src2),
7148 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7149 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7150 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))],
7151 itins.rr>, Sched<[itins.Sched]>;
7152 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
7153 (ins VR128:$src1, i128mem:$src2),
7155 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7156 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7158 (IntId128 VR128:$src1, (bitconvert (memopv2i64 addr:$src2))))],
7159 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
7162 /// SS41I_binop_rm_int_y - Simple SSE 4.1 binary operator
7163 multiclass SS41I_binop_rm_int_y<bits<8> opc, string OpcodeStr,
7165 X86FoldableSchedWrite Sched> {
7166 let isCommutable = 1 in
7167 def Yrr : SS48I<opc, MRMSrcReg, (outs VR256:$dst),
7168 (ins VR256:$src1, VR256:$src2),
7169 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7170 [(set VR256:$dst, (IntId256 VR256:$src1, VR256:$src2))]>,
7172 def Yrm : SS48I<opc, MRMSrcMem, (outs VR256:$dst),
7173 (ins VR256:$src1, i256mem:$src2),
7174 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7176 (IntId256 VR256:$src1, (bitconvert (loadv4i64 addr:$src2))))]>,
7177 Sched<[Sched.Folded, ReadAfterLd]>;
7181 /// SS48I_binop_rm - Simple SSE41 binary operator.
7182 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
7183 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
7184 X86MemOperand x86memop, bit Is2Addr = 1,
7185 OpndItins itins = SSE_INTALU_ITINS_P> {
7186 let isCommutable = 1 in
7187 def rr : SS48I<opc, MRMSrcReg, (outs RC:$dst),
7188 (ins RC:$src1, RC:$src2),
7190 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7191 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7192 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))]>,
7193 Sched<[itins.Sched]>;
7194 def rm : SS48I<opc, MRMSrcMem, (outs RC:$dst),
7195 (ins RC:$src1, x86memop:$src2),
7197 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7198 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7200 (OpVT (OpNode RC:$src1, (bitconvert (memop_frag addr:$src2)))))]>,
7201 Sched<[itins.Sched.Folded, ReadAfterLd]>;
7204 /// SS48I_binop_rm2 - Simple SSE41 binary operator with different src and dst
7206 multiclass SS48I_binop_rm2<bits<8> opc, string OpcodeStr, SDNode OpNode,
7207 ValueType DstVT, ValueType SrcVT, RegisterClass RC,
7208 PatFrag memop_frag, X86MemOperand x86memop,
7210 bit IsCommutable = 0, bit Is2Addr = 1> {
7211 let isCommutable = IsCommutable in
7212 def rr : SS48I<opc, MRMSrcReg, (outs RC:$dst),
7213 (ins RC:$src1, RC:$src2),
7215 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7216 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7217 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1), RC:$src2)))]>,
7218 Sched<[itins.Sched]>;
7219 def rm : SS48I<opc, MRMSrcMem, (outs RC:$dst),
7220 (ins RC:$src1, x86memop:$src2),
7222 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7223 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7224 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1),
7225 (bitconvert (memop_frag addr:$src2)))))]>,
7226 Sched<[itins.Sched.Folded, ReadAfterLd]>;
7229 let Predicates = [HasAVX] in {
7230 let isCommutable = 0 in
7231 defm VPMINSB : SS48I_binop_rm<0x38, "vpminsb", X86smin, v16i8, VR128,
7232 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7234 defm VPMINSD : SS48I_binop_rm<0x39, "vpminsd", X86smin, v4i32, VR128,
7235 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7237 defm VPMINUD : SS48I_binop_rm<0x3B, "vpminud", X86umin, v4i32, VR128,
7238 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7240 defm VPMINUW : SS48I_binop_rm<0x3A, "vpminuw", X86umin, v8i16, VR128,
7241 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7243 defm VPMAXSB : SS48I_binop_rm<0x3C, "vpmaxsb", X86smax, v16i8, VR128,
7244 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7246 defm VPMAXSD : SS48I_binop_rm<0x3D, "vpmaxsd", X86smax, v4i32, VR128,
7247 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7249 defm VPMAXUD : SS48I_binop_rm<0x3F, "vpmaxud", X86umax, v4i32, VR128,
7250 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7252 defm VPMAXUW : SS48I_binop_rm<0x3E, "vpmaxuw", X86umax, v8i16, VR128,
7253 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7255 defm VPMULDQ : SS48I_binop_rm2<0x28, "vpmuldq", X86pmuldq, v2i64, v4i32,
7256 VR128, loadv2i64, i128mem,
7257 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V;
7260 let Predicates = [HasAVX2] in {
7261 let isCommutable = 0 in
7262 defm VPMINSBY : SS48I_binop_rm<0x38, "vpminsb", X86smin, v32i8, VR256,
7263 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7265 defm VPMINSDY : SS48I_binop_rm<0x39, "vpminsd", X86smin, v8i32, VR256,
7266 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7268 defm VPMINUDY : SS48I_binop_rm<0x3B, "vpminud", X86umin, v8i32, VR256,
7269 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7271 defm VPMINUWY : SS48I_binop_rm<0x3A, "vpminuw", X86umin, v16i16, VR256,
7272 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7274 defm VPMAXSBY : SS48I_binop_rm<0x3C, "vpmaxsb", X86smax, v32i8, VR256,
7275 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7277 defm VPMAXSDY : SS48I_binop_rm<0x3D, "vpmaxsd", X86smax, v8i32, VR256,
7278 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7280 defm VPMAXUDY : SS48I_binop_rm<0x3F, "vpmaxud", X86umax, v8i32, VR256,
7281 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7283 defm VPMAXUWY : SS48I_binop_rm<0x3E, "vpmaxuw", X86umax, v16i16, VR256,
7284 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7286 defm VPMULDQY : SS48I_binop_rm2<0x28, "vpmuldq", X86pmuldq, v4i64, v8i32,
7287 VR256, loadv4i64, i256mem,
7288 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V, VEX_L;
7291 let Constraints = "$src1 = $dst" in {
7292 let isCommutable = 0 in
7293 defm PMINSB : SS48I_binop_rm<0x38, "pminsb", X86smin, v16i8, VR128,
7294 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7295 defm PMINSD : SS48I_binop_rm<0x39, "pminsd", X86smin, v4i32, VR128,
7296 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7297 defm PMINUD : SS48I_binop_rm<0x3B, "pminud", X86umin, v4i32, VR128,
7298 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7299 defm PMINUW : SS48I_binop_rm<0x3A, "pminuw", X86umin, v8i16, VR128,
7300 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7301 defm PMAXSB : SS48I_binop_rm<0x3C, "pmaxsb", X86smax, v16i8, VR128,
7302 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7303 defm PMAXSD : SS48I_binop_rm<0x3D, "pmaxsd", X86smax, v4i32, VR128,
7304 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7305 defm PMAXUD : SS48I_binop_rm<0x3F, "pmaxud", X86umax, v4i32, VR128,
7306 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7307 defm PMAXUW : SS48I_binop_rm<0x3E, "pmaxuw", X86umax, v8i16, VR128,
7308 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
7309 defm PMULDQ : SS48I_binop_rm2<0x28, "pmuldq", X86pmuldq, v2i64, v4i32,
7310 VR128, memopv2i64, i128mem,
7311 SSE_INTMUL_ITINS_P, 1>;
7314 let Predicates = [HasAVX] in {
7315 defm VPMULLD : SS48I_binop_rm<0x40, "vpmulld", mul, v4i32, VR128,
7316 memopv2i64, i128mem, 0, SSE_PMULLD_ITINS>,
7318 defm VPCMPEQQ : SS48I_binop_rm<0x29, "vpcmpeqq", X86pcmpeq, v2i64, VR128,
7319 memopv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
7322 let Predicates = [HasAVX2] in {
7323 defm VPMULLDY : SS48I_binop_rm<0x40, "vpmulld", mul, v8i32, VR256,
7324 memopv4i64, i256mem, 0, SSE_PMULLD_ITINS>,
7326 defm VPCMPEQQY : SS48I_binop_rm<0x29, "vpcmpeqq", X86pcmpeq, v4i64, VR256,
7327 memopv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
7331 let Constraints = "$src1 = $dst" in {
7332 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, VR128,
7333 memopv2i64, i128mem, 1, SSE_PMULLD_ITINS>;
7334 defm PCMPEQQ : SS48I_binop_rm<0x29, "pcmpeqq", X86pcmpeq, v2i64, VR128,
7335 memopv2i64, i128mem, 1, SSE_INTALUQ_ITINS_P>;
7338 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
7339 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
7340 Intrinsic IntId, RegisterClass RC, PatFrag memop_frag,
7341 X86MemOperand x86memop, bit Is2Addr = 1,
7342 OpndItins itins = DEFAULT_ITINS> {
7343 let isCommutable = 1 in
7344 def rri : SS4AIi8<opc, MRMSrcReg, (outs RC:$dst),
7345 (ins RC:$src1, RC:$src2, i8imm:$src3),
7347 !strconcat(OpcodeStr,
7348 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
7349 !strconcat(OpcodeStr,
7350 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
7351 [(set RC:$dst, (IntId RC:$src1, RC:$src2, imm:$src3))], itins.rr>,
7352 Sched<[itins.Sched]>;
7353 def rmi : SS4AIi8<opc, MRMSrcMem, (outs RC:$dst),
7354 (ins RC:$src1, x86memop:$src2, i8imm:$src3),
7356 !strconcat(OpcodeStr,
7357 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
7358 !strconcat(OpcodeStr,
7359 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
7362 (bitconvert (memop_frag addr:$src2)), imm:$src3))], itins.rm>,
7363 Sched<[itins.Sched.Folded, ReadAfterLd]>;
7366 let Predicates = [HasAVX] in {
7367 let isCommutable = 0 in {
7368 let ExeDomain = SSEPackedSingle in {
7369 defm VBLENDPS : SS41I_binop_rmi_int<0x0C, "vblendps", int_x86_sse41_blendps,
7370 VR128, loadv4f32, f128mem, 0,
7371 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V;
7372 defm VBLENDPSY : SS41I_binop_rmi_int<0x0C, "vblendps",
7373 int_x86_avx_blend_ps_256, VR256, loadv8f32,
7374 f256mem, 0, DEFAULT_ITINS_FBLENDSCHED>,
7377 let ExeDomain = SSEPackedDouble in {
7378 defm VBLENDPD : SS41I_binop_rmi_int<0x0D, "vblendpd", int_x86_sse41_blendpd,
7379 VR128, loadv2f64, f128mem, 0,
7380 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V;
7381 defm VBLENDPDY : SS41I_binop_rmi_int<0x0D, "vblendpd",
7382 int_x86_avx_blend_pd_256,VR256, loadv4f64,
7383 f256mem, 0, DEFAULT_ITINS_FBLENDSCHED>,
7386 defm VPBLENDW : SS41I_binop_rmi_int<0x0E, "vpblendw", int_x86_sse41_pblendw,
7387 VR128, loadv2i64, i128mem, 0,
7388 DEFAULT_ITINS_BLENDSCHED>, VEX_4V;
7389 defm VMPSADBW : SS41I_binop_rmi_int<0x42, "vmpsadbw", int_x86_sse41_mpsadbw,
7390 VR128, loadv2i64, i128mem, 0,
7391 DEFAULT_ITINS_MPSADSCHED>, VEX_4V;
7393 let ExeDomain = SSEPackedSingle in
7394 defm VDPPS : SS41I_binop_rmi_int<0x40, "vdpps", int_x86_sse41_dpps,
7395 VR128, loadv4f32, f128mem, 0,
7396 SSE_DPPS_ITINS>, VEX_4V;
7397 let ExeDomain = SSEPackedDouble in
7398 defm VDPPD : SS41I_binop_rmi_int<0x41, "vdppd", int_x86_sse41_dppd,
7399 VR128, loadv2f64, f128mem, 0,
7400 SSE_DPPS_ITINS>, VEX_4V;
7401 let ExeDomain = SSEPackedSingle in
7402 defm VDPPSY : SS41I_binop_rmi_int<0x40, "vdpps", int_x86_avx_dp_ps_256,
7403 VR256, loadv8f32, i256mem, 0,
7404 SSE_DPPS_ITINS>, VEX_4V, VEX_L;
7407 let Predicates = [HasAVX2] in {
7408 let isCommutable = 0 in {
7409 defm VPBLENDWY : SS41I_binop_rmi_int<0x0E, "vpblendw", int_x86_avx2_pblendw,
7410 VR256, loadv4i64, i256mem, 0,
7411 DEFAULT_ITINS_BLENDSCHED>, VEX_4V, VEX_L;
7412 defm VMPSADBWY : SS41I_binop_rmi_int<0x42, "vmpsadbw", int_x86_avx2_mpsadbw,
7413 VR256, loadv4i64, i256mem, 0,
7414 DEFAULT_ITINS_MPSADSCHED>, VEX_4V, VEX_L;
7418 let Constraints = "$src1 = $dst" in {
7419 let isCommutable = 0 in {
7420 let ExeDomain = SSEPackedSingle in
7421 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps", int_x86_sse41_blendps,
7422 VR128, memopv4f32, f128mem,
7423 1, SSE_INTALU_ITINS_FBLEND_P>;
7424 let ExeDomain = SSEPackedDouble in
7425 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd", int_x86_sse41_blendpd,
7426 VR128, memopv2f64, f128mem,
7427 1, SSE_INTALU_ITINS_FBLEND_P>;
7428 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw", int_x86_sse41_pblendw,
7429 VR128, memopv2i64, i128mem,
7430 1, SSE_INTALU_ITINS_BLEND_P>;
7431 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw", int_x86_sse41_mpsadbw,
7432 VR128, memopv2i64, i128mem,
7433 1, SSE_MPSADBW_ITINS>;
7435 let ExeDomain = SSEPackedSingle in
7436 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps", int_x86_sse41_dpps,
7437 VR128, memopv4f32, f128mem, 1,
7439 let ExeDomain = SSEPackedDouble in
7440 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd", int_x86_sse41_dppd,
7441 VR128, memopv2f64, f128mem, 1,
7445 /// SS41I_quaternary_int_avx - AVX SSE 4.1 with 4 operators
7446 multiclass SS41I_quaternary_int_avx<bits<8> opc, string OpcodeStr,
7447 RegisterClass RC, X86MemOperand x86memop,
7448 PatFrag mem_frag, Intrinsic IntId,
7449 X86FoldableSchedWrite Sched> {
7450 def rr : Ii8<opc, MRMSrcReg, (outs RC:$dst),
7451 (ins RC:$src1, RC:$src2, RC:$src3),
7452 !strconcat(OpcodeStr,
7453 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
7454 [(set RC:$dst, (IntId RC:$src1, RC:$src2, RC:$src3))],
7455 NoItinerary, SSEPackedInt>, TAPD, VEX_4V, VEX_I8IMM,
7458 def rm : Ii8<opc, MRMSrcMem, (outs RC:$dst),
7459 (ins RC:$src1, x86memop:$src2, RC:$src3),
7460 !strconcat(OpcodeStr,
7461 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
7463 (IntId RC:$src1, (bitconvert (mem_frag addr:$src2)),
7465 NoItinerary, SSEPackedInt>, TAPD, VEX_4V, VEX_I8IMM,
7466 Sched<[Sched.Folded, ReadAfterLd]>;
7469 let Predicates = [HasAVX] in {
7470 let ExeDomain = SSEPackedDouble in {
7471 defm VBLENDVPD : SS41I_quaternary_int_avx<0x4B, "vblendvpd", VR128, f128mem,
7472 loadv2f64, int_x86_sse41_blendvpd,
7474 defm VBLENDVPDY : SS41I_quaternary_int_avx<0x4B, "vblendvpd", VR256, f256mem,
7475 loadv4f64, int_x86_avx_blendv_pd_256,
7476 WriteFVarBlend>, VEX_L;
7477 } // ExeDomain = SSEPackedDouble
7478 let ExeDomain = SSEPackedSingle in {
7479 defm VBLENDVPS : SS41I_quaternary_int_avx<0x4A, "vblendvps", VR128, f128mem,
7480 loadv4f32, int_x86_sse41_blendvps,
7482 defm VBLENDVPSY : SS41I_quaternary_int_avx<0x4A, "vblendvps", VR256, f256mem,
7483 loadv8f32, int_x86_avx_blendv_ps_256,
7484 WriteFVarBlend>, VEX_L;
7485 } // ExeDomain = SSEPackedSingle
7486 defm VPBLENDVB : SS41I_quaternary_int_avx<0x4C, "vpblendvb", VR128, i128mem,
7487 loadv2i64, int_x86_sse41_pblendvb,
7491 let Predicates = [HasAVX2] in {
7492 defm VPBLENDVBY : SS41I_quaternary_int_avx<0x4C, "vpblendvb", VR256, i256mem,
7493 loadv4i64, int_x86_avx2_pblendvb,
7494 WriteVarBlend>, VEX_L;
7497 let Predicates = [HasAVX] in {
7498 def : Pat<(v16i8 (vselect (v16i8 VR128:$mask), (v16i8 VR128:$src1),
7499 (v16i8 VR128:$src2))),
7500 (VPBLENDVBrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7501 def : Pat<(v4i32 (vselect (v4i32 VR128:$mask), (v4i32 VR128:$src1),
7502 (v4i32 VR128:$src2))),
7503 (VBLENDVPSrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7504 def : Pat<(v4f32 (vselect (v4i32 VR128:$mask), (v4f32 VR128:$src1),
7505 (v4f32 VR128:$src2))),
7506 (VBLENDVPSrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7507 def : Pat<(v2i64 (vselect (v2i64 VR128:$mask), (v2i64 VR128:$src1),
7508 (v2i64 VR128:$src2))),
7509 (VBLENDVPDrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7510 def : Pat<(v2f64 (vselect (v2i64 VR128:$mask), (v2f64 VR128:$src1),
7511 (v2f64 VR128:$src2))),
7512 (VBLENDVPDrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7513 def : Pat<(v8i32 (vselect (v8i32 VR256:$mask), (v8i32 VR256:$src1),
7514 (v8i32 VR256:$src2))),
7515 (VBLENDVPSYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7516 def : Pat<(v8f32 (vselect (v8i32 VR256:$mask), (v8f32 VR256:$src1),
7517 (v8f32 VR256:$src2))),
7518 (VBLENDVPSYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7519 def : Pat<(v4i64 (vselect (v4i64 VR256:$mask), (v4i64 VR256:$src1),
7520 (v4i64 VR256:$src2))),
7521 (VBLENDVPDYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7522 def : Pat<(v4f64 (vselect (v4i64 VR256:$mask), (v4f64 VR256:$src1),
7523 (v4f64 VR256:$src2))),
7524 (VBLENDVPDYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7526 def : Pat<(v8f32 (X86Blendi (v8f32 VR256:$src1), (v8f32 VR256:$src2),
7528 (VBLENDPSYrri VR256:$src1, VR256:$src2, imm:$mask)>;
7529 def : Pat<(v4f64 (X86Blendi (v4f64 VR256:$src1), (v4f64 VR256:$src2),
7531 (VBLENDPDYrri VR256:$src1, VR256:$src2, imm:$mask)>;
7533 def : Pat<(v8i16 (X86Blendi (v8i16 VR128:$src1), (v8i16 VR128:$src2),
7535 (VPBLENDWrri VR128:$src1, VR128:$src2, imm:$mask)>;
7536 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$src1), (v4f32 VR128:$src2),
7538 (VBLENDPSrri VR128:$src1, VR128:$src2, imm:$mask)>;
7539 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$src1), (v2f64 VR128:$src2),
7541 (VBLENDPDrri VR128:$src1, VR128:$src2, imm:$mask)>;
7544 let Predicates = [HasAVX2] in {
7545 def : Pat<(v32i8 (vselect (v32i8 VR256:$mask), (v32i8 VR256:$src1),
7546 (v32i8 VR256:$src2))),
7547 (VPBLENDVBYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7548 def : Pat<(v16i16 (X86Blendi (v16i16 VR256:$src1), (v16i16 VR256:$src2),
7550 (VPBLENDWYrri VR256:$src1, VR256:$src2, imm:$mask)>;
7553 /// SS41I_ternary_int - SSE 4.1 ternary operator
7554 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
7555 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
7556 X86MemOperand x86memop, Intrinsic IntId,
7557 OpndItins itins = DEFAULT_ITINS> {
7558 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
7559 (ins VR128:$src1, VR128:$src2),
7560 !strconcat(OpcodeStr,
7561 "\t{$src2, $dst|$dst, $src2}"),
7562 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))],
7563 itins.rr>, Sched<[itins.Sched]>;
7565 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
7566 (ins VR128:$src1, x86memop:$src2),
7567 !strconcat(OpcodeStr,
7568 "\t{$src2, $dst|$dst, $src2}"),
7571 (bitconvert (mem_frag addr:$src2)), XMM0))],
7572 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
7576 let ExeDomain = SSEPackedDouble in
7577 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", memopv2f64, f128mem,
7578 int_x86_sse41_blendvpd,
7579 DEFAULT_ITINS_FBLENDSCHED>;
7580 let ExeDomain = SSEPackedSingle in
7581 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", memopv4f32, f128mem,
7582 int_x86_sse41_blendvps,
7583 DEFAULT_ITINS_FBLENDSCHED>;
7584 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", memopv2i64, i128mem,
7585 int_x86_sse41_pblendvb,
7586 DEFAULT_ITINS_VARBLENDSCHED>;
7588 // Aliases with the implicit xmm0 argument
7589 def : InstAlias<"blendvpd\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7590 (BLENDVPDrr0 VR128:$dst, VR128:$src2)>;
7591 def : InstAlias<"blendvpd\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7592 (BLENDVPDrm0 VR128:$dst, f128mem:$src2)>;
7593 def : InstAlias<"blendvps\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7594 (BLENDVPSrr0 VR128:$dst, VR128:$src2)>;
7595 def : InstAlias<"blendvps\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7596 (BLENDVPSrm0 VR128:$dst, f128mem:$src2)>;
7597 def : InstAlias<"pblendvb\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7598 (PBLENDVBrr0 VR128:$dst, VR128:$src2)>;
7599 def : InstAlias<"pblendvb\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7600 (PBLENDVBrm0 VR128:$dst, i128mem:$src2)>;
7602 let Predicates = [UseSSE41] in {
7603 def : Pat<(v16i8 (vselect (v16i8 XMM0), (v16i8 VR128:$src1),
7604 (v16i8 VR128:$src2))),
7605 (PBLENDVBrr0 VR128:$src2, VR128:$src1)>;
7606 def : Pat<(v4i32 (vselect (v4i32 XMM0), (v4i32 VR128:$src1),
7607 (v4i32 VR128:$src2))),
7608 (BLENDVPSrr0 VR128:$src2, VR128:$src1)>;
7609 def : Pat<(v4f32 (vselect (v4i32 XMM0), (v4f32 VR128:$src1),
7610 (v4f32 VR128:$src2))),
7611 (BLENDVPSrr0 VR128:$src2, VR128:$src1)>;
7612 def : Pat<(v2i64 (vselect (v2i64 XMM0), (v2i64 VR128:$src1),
7613 (v2i64 VR128:$src2))),
7614 (BLENDVPDrr0 VR128:$src2, VR128:$src1)>;
7615 def : Pat<(v2f64 (vselect (v2i64 XMM0), (v2f64 VR128:$src1),
7616 (v2f64 VR128:$src2))),
7617 (BLENDVPDrr0 VR128:$src2, VR128:$src1)>;
7619 def : Pat<(v8i16 (X86Blendi (v8i16 VR128:$src1), (v8i16 VR128:$src2),
7621 (PBLENDWrri VR128:$src1, VR128:$src2, imm:$mask)>;
7622 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$src1), (v4f32 VR128:$src2),
7624 (BLENDPSrri VR128:$src1, VR128:$src2, imm:$mask)>;
7625 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$src1), (v2f64 VR128:$src2),
7627 (BLENDPDrri VR128:$src1, VR128:$src2, imm:$mask)>;
7631 let SchedRW = [WriteLoad] in {
7632 let Predicates = [HasAVX] in
7633 def VMOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
7634 "vmovntdqa\t{$src, $dst|$dst, $src}",
7635 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
7637 let Predicates = [HasAVX2] in
7638 def VMOVNTDQAYrm : SS48I<0x2A, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
7639 "vmovntdqa\t{$src, $dst|$dst, $src}",
7640 [(set VR256:$dst, (int_x86_avx2_movntdqa addr:$src))]>,
7642 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
7643 "movntdqa\t{$src, $dst|$dst, $src}",
7644 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;
7647 //===----------------------------------------------------------------------===//
7648 // SSE4.2 - Compare Instructions
7649 //===----------------------------------------------------------------------===//
7651 /// SS42I_binop_rm - Simple SSE 4.2 binary operator
7652 multiclass SS42I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
7653 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
7654 X86MemOperand x86memop, bit Is2Addr = 1> {
7655 def rr : SS428I<opc, MRMSrcReg, (outs RC:$dst),
7656 (ins RC:$src1, RC:$src2),
7658 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7659 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7660 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))]>;
7661 def rm : SS428I<opc, MRMSrcMem, (outs RC:$dst),
7662 (ins RC:$src1, x86memop:$src2),
7664 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7665 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7667 (OpVT (OpNode RC:$src1, (memop_frag addr:$src2))))]>;
7670 let Predicates = [HasAVX] in
7671 defm VPCMPGTQ : SS42I_binop_rm<0x37, "vpcmpgtq", X86pcmpgt, v2i64, VR128,
7672 loadv2i64, i128mem, 0>, VEX_4V;
7674 let Predicates = [HasAVX2] in
7675 defm VPCMPGTQY : SS42I_binop_rm<0x37, "vpcmpgtq", X86pcmpgt, v4i64, VR256,
7676 loadv4i64, i256mem, 0>, VEX_4V, VEX_L;
7678 let Constraints = "$src1 = $dst" in
7679 defm PCMPGTQ : SS42I_binop_rm<0x37, "pcmpgtq", X86pcmpgt, v2i64, VR128,
7680 memopv2i64, i128mem>;
7682 //===----------------------------------------------------------------------===//
7683 // SSE4.2 - String/text Processing Instructions
7684 //===----------------------------------------------------------------------===//
7686 // Packed Compare Implicit Length Strings, Return Mask
7687 multiclass pseudo_pcmpistrm<string asm> {
7688 def REG : PseudoI<(outs VR128:$dst),
7689 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
7690 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
7692 def MEM : PseudoI<(outs VR128:$dst),
7693 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
7694 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1,
7695 (bc_v16i8 (memopv2i64 addr:$src2)), imm:$src3))]>;
7698 let Defs = [EFLAGS], usesCustomInserter = 1 in {
7699 defm VPCMPISTRM128 : pseudo_pcmpistrm<"#VPCMPISTRM128">, Requires<[HasAVX]>;
7700 defm PCMPISTRM128 : pseudo_pcmpistrm<"#PCMPISTRM128">, Requires<[UseSSE42]>;
7703 multiclass pcmpistrm_SS42AI<string asm> {
7704 def rr : SS42AI<0x62, MRMSrcReg, (outs),
7705 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
7706 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7707 []>, Sched<[WritePCmpIStrM]>;
7709 def rm :SS42AI<0x62, MRMSrcMem, (outs),
7710 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
7711 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7712 []>, Sched<[WritePCmpIStrMLd, ReadAfterLd]>;
7715 let Defs = [XMM0, EFLAGS], neverHasSideEffects = 1 in {
7716 let Predicates = [HasAVX] in
7717 defm VPCMPISTRM128 : pcmpistrm_SS42AI<"vpcmpistrm">, VEX;
7718 defm PCMPISTRM128 : pcmpistrm_SS42AI<"pcmpistrm"> ;
7721 // Packed Compare Explicit Length Strings, Return Mask
7722 multiclass pseudo_pcmpestrm<string asm> {
7723 def REG : PseudoI<(outs VR128:$dst),
7724 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
7725 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
7726 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>;
7727 def MEM : PseudoI<(outs VR128:$dst),
7728 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
7729 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128 VR128:$src1, EAX,
7730 (bc_v16i8 (memopv2i64 addr:$src3)), EDX, imm:$src5))]>;
7733 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
7734 defm VPCMPESTRM128 : pseudo_pcmpestrm<"#VPCMPESTRM128">, Requires<[HasAVX]>;
7735 defm PCMPESTRM128 : pseudo_pcmpestrm<"#PCMPESTRM128">, Requires<[UseSSE42]>;
7738 multiclass SS42AI_pcmpestrm<string asm> {
7739 def rr : SS42AI<0x60, MRMSrcReg, (outs),
7740 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
7741 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7742 []>, Sched<[WritePCmpEStrM]>;
7744 def rm : SS42AI<0x60, MRMSrcMem, (outs),
7745 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
7746 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7747 []>, Sched<[WritePCmpEStrMLd, ReadAfterLd]>;
7750 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX], neverHasSideEffects = 1 in {
7751 let Predicates = [HasAVX] in
7752 defm VPCMPESTRM128 : SS42AI_pcmpestrm<"vpcmpestrm">, VEX;
7753 defm PCMPESTRM128 : SS42AI_pcmpestrm<"pcmpestrm">;
7756 // Packed Compare Implicit Length Strings, Return Index
7757 multiclass pseudo_pcmpistri<string asm> {
7758 def REG : PseudoI<(outs GR32:$dst),
7759 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
7760 [(set GR32:$dst, EFLAGS,
7761 (X86pcmpistri VR128:$src1, VR128:$src2, imm:$src3))]>;
7762 def MEM : PseudoI<(outs GR32:$dst),
7763 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
7764 [(set GR32:$dst, EFLAGS, (X86pcmpistri VR128:$src1,
7765 (bc_v16i8 (memopv2i64 addr:$src2)), imm:$src3))]>;
7768 let Defs = [EFLAGS], usesCustomInserter = 1 in {
7769 defm VPCMPISTRI : pseudo_pcmpistri<"#VPCMPISTRI">, Requires<[HasAVX]>;
7770 defm PCMPISTRI : pseudo_pcmpistri<"#PCMPISTRI">, Requires<[UseSSE42]>;
7773 multiclass SS42AI_pcmpistri<string asm> {
7774 def rr : SS42AI<0x63, MRMSrcReg, (outs),
7775 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
7776 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7777 []>, Sched<[WritePCmpIStrI]>;
7779 def rm : SS42AI<0x63, MRMSrcMem, (outs),
7780 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
7781 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7782 []>, Sched<[WritePCmpIStrILd, ReadAfterLd]>;
7785 let Defs = [ECX, EFLAGS], neverHasSideEffects = 1 in {
7786 let Predicates = [HasAVX] in
7787 defm VPCMPISTRI : SS42AI_pcmpistri<"vpcmpistri">, VEX;
7788 defm PCMPISTRI : SS42AI_pcmpistri<"pcmpistri">;
7791 // Packed Compare Explicit Length Strings, Return Index
7792 multiclass pseudo_pcmpestri<string asm> {
7793 def REG : PseudoI<(outs GR32:$dst),
7794 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
7795 [(set GR32:$dst, EFLAGS,
7796 (X86pcmpestri VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>;
7797 def MEM : PseudoI<(outs GR32:$dst),
7798 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
7799 [(set GR32:$dst, EFLAGS,
7800 (X86pcmpestri VR128:$src1, EAX, (bc_v16i8 (memopv2i64 addr:$src3)), EDX,
7804 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
7805 defm VPCMPESTRI : pseudo_pcmpestri<"#VPCMPESTRI">, Requires<[HasAVX]>;
7806 defm PCMPESTRI : pseudo_pcmpestri<"#PCMPESTRI">, Requires<[UseSSE42]>;
7809 multiclass SS42AI_pcmpestri<string asm> {
7810 def rr : SS42AI<0x61, MRMSrcReg, (outs),
7811 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
7812 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7813 []>, Sched<[WritePCmpEStrI]>;
7815 def rm : SS42AI<0x61, MRMSrcMem, (outs),
7816 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
7817 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7818 []>, Sched<[WritePCmpEStrILd, ReadAfterLd]>;
7821 let Defs = [ECX, EFLAGS], Uses = [EAX, EDX], neverHasSideEffects = 1 in {
7822 let Predicates = [HasAVX] in
7823 defm VPCMPESTRI : SS42AI_pcmpestri<"vpcmpestri">, VEX;
7824 defm PCMPESTRI : SS42AI_pcmpestri<"pcmpestri">;
7827 //===----------------------------------------------------------------------===//
7828 // SSE4.2 - CRC Instructions
7829 //===----------------------------------------------------------------------===//
7831 // No CRC instructions have AVX equivalents
7833 // crc intrinsic instruction
7834 // This set of instructions are only rm, the only difference is the size
7836 class SS42I_crc32r<bits<8> opc, string asm, RegisterClass RCOut,
7837 RegisterClass RCIn, SDPatternOperator Int> :
7838 SS42FI<opc, MRMSrcReg, (outs RCOut:$dst), (ins RCOut:$src1, RCIn:$src2),
7839 !strconcat(asm, "\t{$src2, $src1|$src1, $src2}"),
7840 [(set RCOut:$dst, (Int RCOut:$src1, RCIn:$src2))], IIC_CRC32_REG>,
7843 class SS42I_crc32m<bits<8> opc, string asm, RegisterClass RCOut,
7844 X86MemOperand x86memop, SDPatternOperator Int> :
7845 SS42FI<opc, MRMSrcMem, (outs RCOut:$dst), (ins RCOut:$src1, x86memop:$src2),
7846 !strconcat(asm, "\t{$src2, $src1|$src1, $src2}"),
7847 [(set RCOut:$dst, (Int RCOut:$src1, (load addr:$src2)))],
7848 IIC_CRC32_MEM>, Sched<[WriteFAddLd, ReadAfterLd]>;
7850 let Constraints = "$src1 = $dst" in {
7851 def CRC32r32m8 : SS42I_crc32m<0xF0, "crc32{b}", GR32, i8mem,
7852 int_x86_sse42_crc32_32_8>;
7853 def CRC32r32r8 : SS42I_crc32r<0xF0, "crc32{b}", GR32, GR8,
7854 int_x86_sse42_crc32_32_8>;
7855 def CRC32r32m16 : SS42I_crc32m<0xF1, "crc32{w}", GR32, i16mem,
7856 int_x86_sse42_crc32_32_16>, OpSize16;
7857 def CRC32r32r16 : SS42I_crc32r<0xF1, "crc32{w}", GR32, GR16,
7858 int_x86_sse42_crc32_32_16>, OpSize16;
7859 def CRC32r32m32 : SS42I_crc32m<0xF1, "crc32{l}", GR32, i32mem,
7860 int_x86_sse42_crc32_32_32>, OpSize32;
7861 def CRC32r32r32 : SS42I_crc32r<0xF1, "crc32{l}", GR32, GR32,
7862 int_x86_sse42_crc32_32_32>, OpSize32;
7863 def CRC32r64m64 : SS42I_crc32m<0xF1, "crc32{q}", GR64, i64mem,
7864 int_x86_sse42_crc32_64_64>, REX_W;
7865 def CRC32r64r64 : SS42I_crc32r<0xF1, "crc32{q}", GR64, GR64,
7866 int_x86_sse42_crc32_64_64>, REX_W;
7867 let hasSideEffects = 0 in {
7869 def CRC32r64m8 : SS42I_crc32m<0xF0, "crc32{b}", GR64, i8mem,
7871 def CRC32r64r8 : SS42I_crc32r<0xF0, "crc32{b}", GR64, GR8,
7876 //===----------------------------------------------------------------------===//
7877 // SHA-NI Instructions
7878 //===----------------------------------------------------------------------===//
7880 multiclass SHAI_binop<bits<8> Opc, string OpcodeStr, Intrinsic IntId,
7882 def rr : I<Opc, MRMSrcReg, (outs VR128:$dst),
7883 (ins VR128:$src1, VR128:$src2),
7884 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7886 (set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0)),
7887 (set VR128:$dst, (IntId VR128:$src1, VR128:$src2)))]>, T8;
7889 def rm : I<Opc, MRMSrcMem, (outs VR128:$dst),
7890 (ins VR128:$src1, i128mem:$src2),
7891 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7893 (set VR128:$dst, (IntId VR128:$src1,
7894 (bc_v4i32 (memopv2i64 addr:$src2)), XMM0)),
7895 (set VR128:$dst, (IntId VR128:$src1,
7896 (bc_v4i32 (memopv2i64 addr:$src2)))))]>, T8;
7899 let Constraints = "$src1 = $dst", Predicates = [HasSHA] in {
7900 def SHA1RNDS4rri : Ii8<0xCC, MRMSrcReg, (outs VR128:$dst),
7901 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
7902 "sha1rnds4\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7904 (int_x86_sha1rnds4 VR128:$src1, VR128:$src2,
7905 (i8 imm:$src3)))]>, TA;
7906 def SHA1RNDS4rmi : Ii8<0xCC, MRMSrcMem, (outs VR128:$dst),
7907 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
7908 "sha1rnds4\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7910 (int_x86_sha1rnds4 VR128:$src1,
7911 (bc_v4i32 (memopv2i64 addr:$src2)),
7912 (i8 imm:$src3)))]>, TA;
7914 defm SHA1NEXTE : SHAI_binop<0xC8, "sha1nexte", int_x86_sha1nexte>;
7915 defm SHA1MSG1 : SHAI_binop<0xC9, "sha1msg1", int_x86_sha1msg1>;
7916 defm SHA1MSG2 : SHAI_binop<0xCA, "sha1msg2", int_x86_sha1msg2>;
7919 defm SHA256RNDS2 : SHAI_binop<0xCB, "sha256rnds2", int_x86_sha256rnds2, 1>;
7921 defm SHA256MSG1 : SHAI_binop<0xCC, "sha256msg1", int_x86_sha256msg1>;
7922 defm SHA256MSG2 : SHAI_binop<0xCD, "sha256msg2", int_x86_sha256msg2>;
7925 // Aliases with explicit %xmm0
7926 def : InstAlias<"sha256rnds2\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7927 (SHA256RNDS2rr VR128:$dst, VR128:$src2)>;
7928 def : InstAlias<"sha256rnds2\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7929 (SHA256RNDS2rm VR128:$dst, i128mem:$src2)>;
7931 //===----------------------------------------------------------------------===//
7932 // AES-NI Instructions
7933 //===----------------------------------------------------------------------===//
7935 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr,
7936 Intrinsic IntId128, bit Is2Addr = 1> {
7937 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
7938 (ins VR128:$src1, VR128:$src2),
7940 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7941 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7942 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
7943 Sched<[WriteAESDecEnc]>;
7944 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
7945 (ins VR128:$src1, i128mem:$src2),
7947 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7948 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7950 (IntId128 VR128:$src1, (memopv2i64 addr:$src2)))]>,
7951 Sched<[WriteAESDecEncLd, ReadAfterLd]>;
7954 // Perform One Round of an AES Encryption/Decryption Flow
7955 let Predicates = [HasAVX, HasAES] in {
7956 defm VAESENC : AESI_binop_rm_int<0xDC, "vaesenc",
7957 int_x86_aesni_aesenc, 0>, VEX_4V;
7958 defm VAESENCLAST : AESI_binop_rm_int<0xDD, "vaesenclast",
7959 int_x86_aesni_aesenclast, 0>, VEX_4V;
7960 defm VAESDEC : AESI_binop_rm_int<0xDE, "vaesdec",
7961 int_x86_aesni_aesdec, 0>, VEX_4V;
7962 defm VAESDECLAST : AESI_binop_rm_int<0xDF, "vaesdeclast",
7963 int_x86_aesni_aesdeclast, 0>, VEX_4V;
7966 let Constraints = "$src1 = $dst" in {
7967 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
7968 int_x86_aesni_aesenc>;
7969 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
7970 int_x86_aesni_aesenclast>;
7971 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
7972 int_x86_aesni_aesdec>;
7973 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
7974 int_x86_aesni_aesdeclast>;
7977 // Perform the AES InvMixColumn Transformation
7978 let Predicates = [HasAVX, HasAES] in {
7979 def VAESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
7981 "vaesimc\t{$src1, $dst|$dst, $src1}",
7983 (int_x86_aesni_aesimc VR128:$src1))]>, Sched<[WriteAESIMC]>,
7985 def VAESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
7986 (ins i128mem:$src1),
7987 "vaesimc\t{$src1, $dst|$dst, $src1}",
7988 [(set VR128:$dst, (int_x86_aesni_aesimc (loadv2i64 addr:$src1)))]>,
7989 Sched<[WriteAESIMCLd]>, VEX;
7991 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
7993 "aesimc\t{$src1, $dst|$dst, $src1}",
7995 (int_x86_aesni_aesimc VR128:$src1))]>, Sched<[WriteAESIMC]>;
7996 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
7997 (ins i128mem:$src1),
7998 "aesimc\t{$src1, $dst|$dst, $src1}",
7999 [(set VR128:$dst, (int_x86_aesni_aesimc (memopv2i64 addr:$src1)))]>,
8000 Sched<[WriteAESIMCLd]>;
8002 // AES Round Key Generation Assist
8003 let Predicates = [HasAVX, HasAES] in {
8004 def VAESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
8005 (ins VR128:$src1, i8imm:$src2),
8006 "vaeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8008 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
8009 Sched<[WriteAESKeyGen]>, VEX;
8010 def VAESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
8011 (ins i128mem:$src1, i8imm:$src2),
8012 "vaeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8014 (int_x86_aesni_aeskeygenassist (loadv2i64 addr:$src1), imm:$src2))]>,
8015 Sched<[WriteAESKeyGenLd]>, VEX;
8017 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
8018 (ins VR128:$src1, i8imm:$src2),
8019 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8021 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
8022 Sched<[WriteAESKeyGen]>;
8023 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
8024 (ins i128mem:$src1, i8imm:$src2),
8025 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8027 (int_x86_aesni_aeskeygenassist (memopv2i64 addr:$src1), imm:$src2))]>,
8028 Sched<[WriteAESKeyGenLd]>;
8030 //===----------------------------------------------------------------------===//
8031 // PCLMUL Instructions
8032 //===----------------------------------------------------------------------===//
8034 // AVX carry-less Multiplication instructions
8035 def VPCLMULQDQrr : AVXPCLMULIi8<0x44, MRMSrcReg, (outs VR128:$dst),
8036 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
8037 "vpclmulqdq\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8039 (int_x86_pclmulqdq VR128:$src1, VR128:$src2, imm:$src3))]>,
8040 Sched<[WriteCLMul]>;
8042 def VPCLMULQDQrm : AVXPCLMULIi8<0x44, MRMSrcMem, (outs VR128:$dst),
8043 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
8044 "vpclmulqdq\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8045 [(set VR128:$dst, (int_x86_pclmulqdq VR128:$src1,
8046 (loadv2i64 addr:$src2), imm:$src3))]>,
8047 Sched<[WriteCLMulLd, ReadAfterLd]>;
8049 // Carry-less Multiplication instructions
8050 let Constraints = "$src1 = $dst" in {
8051 def PCLMULQDQrr : PCLMULIi8<0x44, MRMSrcReg, (outs VR128:$dst),
8052 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
8053 "pclmulqdq\t{$src3, $src2, $dst|$dst, $src2, $src3}",
8055 (int_x86_pclmulqdq VR128:$src1, VR128:$src2, imm:$src3))],
8056 IIC_SSE_PCLMULQDQ_RR>, Sched<[WriteCLMul]>;
8058 def PCLMULQDQrm : PCLMULIi8<0x44, MRMSrcMem, (outs VR128:$dst),
8059 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
8060 "pclmulqdq\t{$src3, $src2, $dst|$dst, $src2, $src3}",
8061 [(set VR128:$dst, (int_x86_pclmulqdq VR128:$src1,
8062 (memopv2i64 addr:$src2), imm:$src3))],
8063 IIC_SSE_PCLMULQDQ_RM>,
8064 Sched<[WriteCLMulLd, ReadAfterLd]>;
8065 } // Constraints = "$src1 = $dst"
8068 multiclass pclmul_alias<string asm, int immop> {
8069 def : InstAlias<!strconcat("pclmul", asm, "dq {$src, $dst|$dst, $src}"),
8070 (PCLMULQDQrr VR128:$dst, VR128:$src, immop), 0>;
8072 def : InstAlias<!strconcat("pclmul", asm, "dq {$src, $dst|$dst, $src}"),
8073 (PCLMULQDQrm VR128:$dst, i128mem:$src, immop), 0>;
8075 def : InstAlias<!strconcat("vpclmul", asm,
8076 "dq {$src2, $src1, $dst|$dst, $src1, $src2}"),
8077 (VPCLMULQDQrr VR128:$dst, VR128:$src1, VR128:$src2, immop),
8080 def : InstAlias<!strconcat("vpclmul", asm,
8081 "dq {$src2, $src1, $dst|$dst, $src1, $src2}"),
8082 (VPCLMULQDQrm VR128:$dst, VR128:$src1, i128mem:$src2, immop),
8085 defm : pclmul_alias<"hqhq", 0x11>;
8086 defm : pclmul_alias<"hqlq", 0x01>;
8087 defm : pclmul_alias<"lqhq", 0x10>;
8088 defm : pclmul_alias<"lqlq", 0x00>;
8090 //===----------------------------------------------------------------------===//
8091 // SSE4A Instructions
8092 //===----------------------------------------------------------------------===//
8094 let Predicates = [HasSSE4A] in {
8096 let Constraints = "$src = $dst" in {
8097 def EXTRQI : Ii8<0x78, MRMXr, (outs VR128:$dst),
8098 (ins VR128:$src, i8imm:$len, i8imm:$idx),
8099 "extrq\t{$idx, $len, $src|$src, $len, $idx}",
8100 [(set VR128:$dst, (int_x86_sse4a_extrqi VR128:$src, imm:$len,
8102 def EXTRQ : I<0x79, MRMSrcReg, (outs VR128:$dst),
8103 (ins VR128:$src, VR128:$mask),
8104 "extrq\t{$mask, $src|$src, $mask}",
8105 [(set VR128:$dst, (int_x86_sse4a_extrq VR128:$src,
8106 VR128:$mask))]>, PD;
8108 def INSERTQI : Ii8<0x78, MRMSrcReg, (outs VR128:$dst),
8109 (ins VR128:$src, VR128:$src2, i8imm:$len, i8imm:$idx),
8110 "insertq\t{$idx, $len, $src2, $src|$src, $src2, $len, $idx}",
8111 [(set VR128:$dst, (int_x86_sse4a_insertqi VR128:$src,
8112 VR128:$src2, imm:$len, imm:$idx))]>, XD;
8113 def INSERTQ : I<0x79, MRMSrcReg, (outs VR128:$dst),
8114 (ins VR128:$src, VR128:$mask),
8115 "insertq\t{$mask, $src|$src, $mask}",
8116 [(set VR128:$dst, (int_x86_sse4a_insertq VR128:$src,
8117 VR128:$mask))]>, XD;
8120 def MOVNTSS : I<0x2B, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
8121 "movntss\t{$src, $dst|$dst, $src}",
8122 [(int_x86_sse4a_movnt_ss addr:$dst, VR128:$src)]>, XS;
8124 def MOVNTSD : I<0x2B, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
8125 "movntsd\t{$src, $dst|$dst, $src}",
8126 [(int_x86_sse4a_movnt_sd addr:$dst, VR128:$src)]>, XD;
8129 //===----------------------------------------------------------------------===//
8131 //===----------------------------------------------------------------------===//
8133 //===----------------------------------------------------------------------===//
8134 // VBROADCAST - Load from memory and broadcast to all elements of the
8135 // destination operand
8137 class avx_broadcast<bits<8> opc, string OpcodeStr, RegisterClass RC,
8138 X86MemOperand x86memop, Intrinsic Int, SchedWrite Sched> :
8139 AVX8I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
8140 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8141 [(set RC:$dst, (Int addr:$src))]>, Sched<[Sched]>, VEX;
8143 class avx_broadcast_no_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
8144 X86MemOperand x86memop, ValueType VT,
8145 PatFrag ld_frag, SchedWrite Sched> :
8146 AVX8I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
8147 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8148 [(set RC:$dst, (VT (X86VBroadcast (ld_frag addr:$src))))]>,
8149 Sched<[Sched]>, VEX {
8153 // AVX2 adds register forms
8154 class avx2_broadcast_reg<bits<8> opc, string OpcodeStr, RegisterClass RC,
8155 Intrinsic Int, SchedWrite Sched> :
8156 AVX28I<opc, MRMSrcReg, (outs RC:$dst), (ins VR128:$src),
8157 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8158 [(set RC:$dst, (Int VR128:$src))]>, Sched<[Sched]>, VEX;
8160 let ExeDomain = SSEPackedSingle in {
8161 def VBROADCASTSSrm : avx_broadcast_no_int<0x18, "vbroadcastss", VR128,
8162 f32mem, v4f32, loadf32, WriteLoad>;
8163 def VBROADCASTSSYrm : avx_broadcast_no_int<0x18, "vbroadcastss", VR256,
8164 f32mem, v8f32, loadf32,
8165 WriteFShuffleLd>, VEX_L;
8167 let ExeDomain = SSEPackedDouble in
8168 def VBROADCASTSDYrm : avx_broadcast_no_int<0x19, "vbroadcastsd", VR256, f64mem,
8169 v4f64, loadf64, WriteFShuffleLd>, VEX_L;
8170 def VBROADCASTF128 : avx_broadcast<0x1A, "vbroadcastf128", VR256, f128mem,
8171 int_x86_avx_vbroadcastf128_pd_256,
8172 WriteFShuffleLd>, VEX_L;
8174 let ExeDomain = SSEPackedSingle in {
8175 def VBROADCASTSSrr : avx2_broadcast_reg<0x18, "vbroadcastss", VR128,
8176 int_x86_avx2_vbroadcast_ss_ps,
8178 def VBROADCASTSSYrr : avx2_broadcast_reg<0x18, "vbroadcastss", VR256,
8179 int_x86_avx2_vbroadcast_ss_ps_256,
8180 WriteFShuffle256>, VEX_L;
8182 let ExeDomain = SSEPackedDouble in
8183 def VBROADCASTSDYrr : avx2_broadcast_reg<0x19, "vbroadcastsd", VR256,
8184 int_x86_avx2_vbroadcast_sd_pd_256,
8185 WriteFShuffle256>, VEX_L;
8187 let Predicates = [HasAVX2] in
8188 def VBROADCASTI128 : avx_broadcast<0x5A, "vbroadcasti128", VR256, i128mem,
8189 int_x86_avx2_vbroadcasti128, WriteLoad>,
8192 let Predicates = [HasAVX] in
8193 def : Pat<(int_x86_avx_vbroadcastf128_ps_256 addr:$src),
8194 (VBROADCASTF128 addr:$src)>;
8197 //===----------------------------------------------------------------------===//
8198 // VINSERTF128 - Insert packed floating-point values
8200 let neverHasSideEffects = 1, ExeDomain = SSEPackedSingle in {
8201 def VINSERTF128rr : AVXAIi8<0x18, MRMSrcReg, (outs VR256:$dst),
8202 (ins VR256:$src1, VR128:$src2, i8imm:$src3),
8203 "vinsertf128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8204 []>, Sched<[WriteFShuffle]>, VEX_4V, VEX_L;
8206 def VINSERTF128rm : AVXAIi8<0x18, MRMSrcMem, (outs VR256:$dst),
8207 (ins VR256:$src1, f128mem:$src2, i8imm:$src3),
8208 "vinsertf128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8209 []>, Sched<[WriteFShuffleLd, ReadAfterLd]>, VEX_4V, VEX_L;
8212 let Predicates = [HasAVX] in {
8213 def : Pat<(vinsert128_insert:$ins (v8f32 VR256:$src1), (v4f32 VR128:$src2),
8215 (VINSERTF128rr VR256:$src1, VR128:$src2,
8216 (INSERT_get_vinsert128_imm VR256:$ins))>;
8217 def : Pat<(vinsert128_insert:$ins (v4f64 VR256:$src1), (v2f64 VR128:$src2),
8219 (VINSERTF128rr VR256:$src1, VR128:$src2,
8220 (INSERT_get_vinsert128_imm VR256:$ins))>;
8222 def : Pat<(vinsert128_insert:$ins (v8f32 VR256:$src1), (loadv4f32 addr:$src2),
8224 (VINSERTF128rm VR256:$src1, addr:$src2,
8225 (INSERT_get_vinsert128_imm VR256:$ins))>;
8226 def : Pat<(vinsert128_insert:$ins (v4f64 VR256:$src1), (loadv2f64 addr:$src2),
8228 (VINSERTF128rm VR256:$src1, addr:$src2,
8229 (INSERT_get_vinsert128_imm VR256:$ins))>;
8232 let Predicates = [HasAVX1Only] in {
8233 def : Pat<(vinsert128_insert:$ins (v4i64 VR256:$src1), (v2i64 VR128:$src2),
8235 (VINSERTF128rr VR256:$src1, VR128:$src2,
8236 (INSERT_get_vinsert128_imm VR256:$ins))>;
8237 def : Pat<(vinsert128_insert:$ins (v8i32 VR256:$src1), (v4i32 VR128:$src2),
8239 (VINSERTF128rr VR256:$src1, VR128:$src2,
8240 (INSERT_get_vinsert128_imm VR256:$ins))>;
8241 def : Pat<(vinsert128_insert:$ins (v32i8 VR256:$src1), (v16i8 VR128:$src2),
8243 (VINSERTF128rr VR256:$src1, VR128:$src2,
8244 (INSERT_get_vinsert128_imm VR256:$ins))>;
8245 def : Pat<(vinsert128_insert:$ins (v16i16 VR256:$src1), (v8i16 VR128:$src2),
8247 (VINSERTF128rr VR256:$src1, VR128:$src2,
8248 (INSERT_get_vinsert128_imm VR256:$ins))>;
8250 def : Pat<(vinsert128_insert:$ins (v4i64 VR256:$src1), (loadv2i64 addr:$src2),
8252 (VINSERTF128rm VR256:$src1, addr:$src2,
8253 (INSERT_get_vinsert128_imm VR256:$ins))>;
8254 def : Pat<(vinsert128_insert:$ins (v8i32 VR256:$src1),
8255 (bc_v4i32 (loadv2i64 addr:$src2)),
8257 (VINSERTF128rm VR256:$src1, addr:$src2,
8258 (INSERT_get_vinsert128_imm VR256:$ins))>;
8259 def : Pat<(vinsert128_insert:$ins (v32i8 VR256:$src1),
8260 (bc_v16i8 (loadv2i64 addr:$src2)),
8262 (VINSERTF128rm VR256:$src1, addr:$src2,
8263 (INSERT_get_vinsert128_imm VR256:$ins))>;
8264 def : Pat<(vinsert128_insert:$ins (v16i16 VR256:$src1),
8265 (bc_v8i16 (loadv2i64 addr:$src2)),
8267 (VINSERTF128rm VR256:$src1, addr:$src2,
8268 (INSERT_get_vinsert128_imm VR256:$ins))>;
8271 //===----------------------------------------------------------------------===//
8272 // VEXTRACTF128 - Extract packed floating-point values
8274 let neverHasSideEffects = 1, ExeDomain = SSEPackedSingle in {
8275 def VEXTRACTF128rr : AVXAIi8<0x19, MRMDestReg, (outs VR128:$dst),
8276 (ins VR256:$src1, i8imm:$src2),
8277 "vextractf128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8278 []>, Sched<[WriteFShuffle]>, VEX, VEX_L;
8280 def VEXTRACTF128mr : AVXAIi8<0x19, MRMDestMem, (outs),
8281 (ins f128mem:$dst, VR256:$src1, i8imm:$src2),
8282 "vextractf128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8283 []>, Sched<[WriteStore]>, VEX, VEX_L;
8287 let Predicates = [HasAVX] in {
8288 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8289 (v4f32 (VEXTRACTF128rr
8290 (v8f32 VR256:$src1),
8291 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8292 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8293 (v2f64 (VEXTRACTF128rr
8294 (v4f64 VR256:$src1),
8295 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8297 def : Pat<(store (v4f32 (vextract128_extract:$ext (v8f32 VR256:$src1),
8298 (iPTR imm))), addr:$dst),
8299 (VEXTRACTF128mr addr:$dst, VR256:$src1,
8300 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8301 def : Pat<(store (v2f64 (vextract128_extract:$ext (v4f64 VR256:$src1),
8302 (iPTR imm))), addr:$dst),
8303 (VEXTRACTF128mr addr:$dst, VR256:$src1,
8304 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8307 let Predicates = [HasAVX1Only] in {
8308 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8309 (v2i64 (VEXTRACTF128rr
8310 (v4i64 VR256:$src1),
8311 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8312 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8313 (v4i32 (VEXTRACTF128rr
8314 (v8i32 VR256:$src1),
8315 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8316 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8317 (v8i16 (VEXTRACTF128rr
8318 (v16i16 VR256:$src1),
8319 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8320 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8321 (v16i8 (VEXTRACTF128rr
8322 (v32i8 VR256:$src1),
8323 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8325 def : Pat<(alignedstore (v2i64 (vextract128_extract:$ext (v4i64 VR256:$src1),
8326 (iPTR imm))), addr:$dst),
8327 (VEXTRACTF128mr addr:$dst, VR256:$src1,
8328 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8329 def : Pat<(alignedstore (v4i32 (vextract128_extract:$ext (v8i32 VR256:$src1),
8330 (iPTR imm))), addr:$dst),
8331 (VEXTRACTF128mr addr:$dst, VR256:$src1,
8332 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8333 def : Pat<(alignedstore (v8i16 (vextract128_extract:$ext (v16i16 VR256:$src1),
8334 (iPTR imm))), addr:$dst),
8335 (VEXTRACTF128mr addr:$dst, VR256:$src1,
8336 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8337 def : Pat<(alignedstore (v16i8 (vextract128_extract:$ext (v32i8 VR256:$src1),
8338 (iPTR imm))), addr:$dst),
8339 (VEXTRACTF128mr addr:$dst, VR256:$src1,
8340 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8343 //===----------------------------------------------------------------------===//
8344 // VMASKMOV - Conditional SIMD Packed Loads and Stores
8346 multiclass avx_movmask_rm<bits<8> opc_rm, bits<8> opc_mr, string OpcodeStr,
8347 Intrinsic IntLd, Intrinsic IntLd256,
8348 Intrinsic IntSt, Intrinsic IntSt256> {
8349 def rm : AVX8I<opc_rm, MRMSrcMem, (outs VR128:$dst),
8350 (ins VR128:$src1, f128mem:$src2),
8351 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8352 [(set VR128:$dst, (IntLd addr:$src2, VR128:$src1))]>,
8354 def Yrm : AVX8I<opc_rm, MRMSrcMem, (outs VR256:$dst),
8355 (ins VR256:$src1, f256mem:$src2),
8356 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8357 [(set VR256:$dst, (IntLd256 addr:$src2, VR256:$src1))]>,
8359 def mr : AVX8I<opc_mr, MRMDestMem, (outs),
8360 (ins f128mem:$dst, VR128:$src1, VR128:$src2),
8361 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8362 [(IntSt addr:$dst, VR128:$src1, VR128:$src2)]>, VEX_4V;
8363 def Ymr : AVX8I<opc_mr, MRMDestMem, (outs),
8364 (ins f256mem:$dst, VR256:$src1, VR256:$src2),
8365 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8366 [(IntSt256 addr:$dst, VR256:$src1, VR256:$src2)]>, VEX_4V, VEX_L;
8369 let ExeDomain = SSEPackedSingle in
8370 defm VMASKMOVPS : avx_movmask_rm<0x2C, 0x2E, "vmaskmovps",
8371 int_x86_avx_maskload_ps,
8372 int_x86_avx_maskload_ps_256,
8373 int_x86_avx_maskstore_ps,
8374 int_x86_avx_maskstore_ps_256>;
8375 let ExeDomain = SSEPackedDouble in
8376 defm VMASKMOVPD : avx_movmask_rm<0x2D, 0x2F, "vmaskmovpd",
8377 int_x86_avx_maskload_pd,
8378 int_x86_avx_maskload_pd_256,
8379 int_x86_avx_maskstore_pd,
8380 int_x86_avx_maskstore_pd_256>;
8382 //===----------------------------------------------------------------------===//
8383 // VPERMIL - Permute Single and Double Floating-Point Values
8385 multiclass avx_permil<bits<8> opc_rm, bits<8> opc_rmi, string OpcodeStr,
8386 RegisterClass RC, X86MemOperand x86memop_f,
8387 X86MemOperand x86memop_i, PatFrag i_frag,
8388 Intrinsic IntVar, ValueType vt> {
8389 def rr : AVX8I<opc_rm, MRMSrcReg, (outs RC:$dst),
8390 (ins RC:$src1, RC:$src2),
8391 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8392 [(set RC:$dst, (IntVar RC:$src1, RC:$src2))]>, VEX_4V,
8393 Sched<[WriteFShuffle]>;
8394 def rm : AVX8I<opc_rm, MRMSrcMem, (outs RC:$dst),
8395 (ins RC:$src1, x86memop_i:$src2),
8396 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8397 [(set RC:$dst, (IntVar RC:$src1,
8398 (bitconvert (i_frag addr:$src2))))]>, VEX_4V,
8399 Sched<[WriteFShuffleLd, ReadAfterLd]>;
8401 def ri : AVXAIi8<opc_rmi, MRMSrcReg, (outs RC:$dst),
8402 (ins RC:$src1, i8imm:$src2),
8403 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8404 [(set RC:$dst, (vt (X86VPermilpi RC:$src1, (i8 imm:$src2))))]>, VEX,
8405 Sched<[WriteFShuffle]>;
8406 def mi : AVXAIi8<opc_rmi, MRMSrcMem, (outs RC:$dst),
8407 (ins x86memop_f:$src1, i8imm:$src2),
8408 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8410 (vt (X86VPermilpi (memop addr:$src1), (i8 imm:$src2))))]>, VEX,
8411 Sched<[WriteFShuffleLd]>;
8414 let ExeDomain = SSEPackedSingle in {
8415 defm VPERMILPS : avx_permil<0x0C, 0x04, "vpermilps", VR128, f128mem, i128mem,
8416 loadv2i64, int_x86_avx_vpermilvar_ps, v4f32>;
8417 defm VPERMILPSY : avx_permil<0x0C, 0x04, "vpermilps", VR256, f256mem, i256mem,
8418 loadv4i64, int_x86_avx_vpermilvar_ps_256, v8f32>, VEX_L;
8420 let ExeDomain = SSEPackedDouble in {
8421 defm VPERMILPD : avx_permil<0x0D, 0x05, "vpermilpd", VR128, f128mem, i128mem,
8422 loadv2i64, int_x86_avx_vpermilvar_pd, v2f64>;
8423 defm VPERMILPDY : avx_permil<0x0D, 0x05, "vpermilpd", VR256, f256mem, i256mem,
8424 loadv4i64, int_x86_avx_vpermilvar_pd_256, v4f64>, VEX_L;
8427 let Predicates = [HasAVX] in {
8428 def : Pat<(v8f32 (X86VPermilpv VR256:$src1, (v8i32 VR256:$src2))),
8429 (VPERMILPSYrr VR256:$src1, VR256:$src2)>;
8430 def : Pat<(v8f32 (X86VPermilpv VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
8431 (VPERMILPSYrm VR256:$src1, addr:$src2)>;
8432 def : Pat<(v4f64 (X86VPermilpv VR256:$src1, (v4i64 VR256:$src2))),
8433 (VPERMILPDYrr VR256:$src1, VR256:$src2)>;
8434 def : Pat<(v4f64 (X86VPermilpv VR256:$src1, (loadv4i64 addr:$src2))),
8435 (VPERMILPDYrm VR256:$src1, addr:$src2)>;
8437 def : Pat<(v8i32 (X86VPermilpi VR256:$src1, (i8 imm:$imm))),
8438 (VPERMILPSYri VR256:$src1, imm:$imm)>;
8439 def : Pat<(v4i64 (X86VPermilpi VR256:$src1, (i8 imm:$imm))),
8440 (VPERMILPDYri VR256:$src1, imm:$imm)>;
8441 def : Pat<(v8i32 (X86VPermilpi (bc_v8i32 (loadv4i64 addr:$src1)),
8443 (VPERMILPSYmi addr:$src1, imm:$imm)>;
8444 def : Pat<(v4i64 (X86VPermilpi (loadv4i64 addr:$src1), (i8 imm:$imm))),
8445 (VPERMILPDYmi addr:$src1, imm:$imm)>;
8447 def : Pat<(v4f32 (X86VPermilpv VR128:$src1, (v4i32 VR128:$src2))),
8448 (VPERMILPSrr VR128:$src1, VR128:$src2)>;
8449 def : Pat<(v4f32 (X86VPermilpv VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)))),
8450 (VPERMILPSrm VR128:$src1, addr:$src2)>;
8451 def : Pat<(v2f64 (X86VPermilpv VR128:$src1, (v2i64 VR128:$src2))),
8452 (VPERMILPDrr VR128:$src1, VR128:$src2)>;
8453 def : Pat<(v2f64 (X86VPermilpv VR128:$src1, (loadv2i64 addr:$src2))),
8454 (VPERMILPDrm VR128:$src1, addr:$src2)>;
8456 def : Pat<(v2i64 (X86VPermilpi VR128:$src1, (i8 imm:$imm))),
8457 (VPERMILPDri VR128:$src1, imm:$imm)>;
8458 def : Pat<(v2i64 (X86VPermilpi (loadv2i64 addr:$src1), (i8 imm:$imm))),
8459 (VPERMILPDmi addr:$src1, imm:$imm)>;
8462 //===----------------------------------------------------------------------===//
8463 // VPERM2F128 - Permute Floating-Point Values in 128-bit chunks
8465 let ExeDomain = SSEPackedSingle in {
8466 def VPERM2F128rr : AVXAIi8<0x06, MRMSrcReg, (outs VR256:$dst),
8467 (ins VR256:$src1, VR256:$src2, i8imm:$src3),
8468 "vperm2f128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8469 [(set VR256:$dst, (v8f32 (X86VPerm2x128 VR256:$src1, VR256:$src2,
8470 (i8 imm:$src3))))]>, VEX_4V, VEX_L,
8471 Sched<[WriteFShuffle]>;
8472 def VPERM2F128rm : AVXAIi8<0x06, MRMSrcMem, (outs VR256:$dst),
8473 (ins VR256:$src1, f256mem:$src2, i8imm:$src3),
8474 "vperm2f128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8475 [(set VR256:$dst, (X86VPerm2x128 VR256:$src1, (loadv8f32 addr:$src2),
8476 (i8 imm:$src3)))]>, VEX_4V, VEX_L,
8477 Sched<[WriteFShuffleLd, ReadAfterLd]>;
8480 let Predicates = [HasAVX] in {
8481 def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8482 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8483 def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1,
8484 (loadv4f64 addr:$src2), (i8 imm:$imm))),
8485 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8488 let Predicates = [HasAVX1Only] in {
8489 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8490 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8491 def : Pat<(v4i64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8492 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8493 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8494 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8495 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8496 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8498 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1,
8499 (bc_v8i32 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8500 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8501 def : Pat<(v4i64 (X86VPerm2x128 VR256:$src1,
8502 (loadv4i64 addr:$src2), (i8 imm:$imm))),
8503 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8504 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1,
8505 (bc_v32i8 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8506 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8507 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1,
8508 (bc_v16i16 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8509 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8512 //===----------------------------------------------------------------------===//
8513 // VZERO - Zero YMM registers
8515 let Defs = [YMM0, YMM1, YMM2, YMM3, YMM4, YMM5, YMM6, YMM7,
8516 YMM8, YMM9, YMM10, YMM11, YMM12, YMM13, YMM14, YMM15] in {
8517 // Zero All YMM registers
8518 def VZEROALL : I<0x77, RawFrm, (outs), (ins), "vzeroall",
8519 [(int_x86_avx_vzeroall)]>, PS, VEX, VEX_L, Requires<[HasAVX]>;
8521 // Zero Upper bits of YMM registers
8522 def VZEROUPPER : I<0x77, RawFrm, (outs), (ins), "vzeroupper",
8523 [(int_x86_avx_vzeroupper)]>, PS, VEX, Requires<[HasAVX]>;
8526 //===----------------------------------------------------------------------===//
8527 // Half precision conversion instructions
8528 //===----------------------------------------------------------------------===//
8529 multiclass f16c_ph2ps<RegisterClass RC, X86MemOperand x86memop, Intrinsic Int> {
8530 def rr : I<0x13, MRMSrcReg, (outs RC:$dst), (ins VR128:$src),
8531 "vcvtph2ps\t{$src, $dst|$dst, $src}",
8532 [(set RC:$dst, (Int VR128:$src))]>,
8533 T8PD, VEX, Sched<[WriteCvtF2F]>;
8534 let neverHasSideEffects = 1, mayLoad = 1 in
8535 def rm : I<0x13, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
8536 "vcvtph2ps\t{$src, $dst|$dst, $src}", []>, T8PD, VEX,
8537 Sched<[WriteCvtF2FLd]>;
8540 multiclass f16c_ps2ph<RegisterClass RC, X86MemOperand x86memop, Intrinsic Int> {
8541 def rr : Ii8<0x1D, MRMDestReg, (outs VR128:$dst),
8542 (ins RC:$src1, i32i8imm:$src2),
8543 "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8544 [(set VR128:$dst, (Int RC:$src1, imm:$src2))]>,
8545 TAPD, VEX, Sched<[WriteCvtF2F]>;
8546 let neverHasSideEffects = 1, mayStore = 1,
8547 SchedRW = [WriteCvtF2FLd, WriteRMW] in
8548 def mr : Ii8<0x1D, MRMDestMem, (outs),
8549 (ins x86memop:$dst, RC:$src1, i32i8imm:$src2),
8550 "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8554 let Predicates = [HasF16C] in {
8555 defm VCVTPH2PS : f16c_ph2ps<VR128, f64mem, int_x86_vcvtph2ps_128>;
8556 defm VCVTPH2PSY : f16c_ph2ps<VR256, f128mem, int_x86_vcvtph2ps_256>, VEX_L;
8557 defm VCVTPS2PH : f16c_ps2ph<VR128, f64mem, int_x86_vcvtps2ph_128>;
8558 defm VCVTPS2PHY : f16c_ps2ph<VR256, f128mem, int_x86_vcvtps2ph_256>, VEX_L;
8560 // Pattern match vcvtph2ps of a scalar i64 load.
8561 def : Pat<(int_x86_vcvtph2ps_128 (vzmovl_v2i64 addr:$src)),
8562 (VCVTPH2PSrm addr:$src)>;
8563 def : Pat<(int_x86_vcvtph2ps_128 (vzload_v2i64 addr:$src)),
8564 (VCVTPH2PSrm addr:$src)>;
8567 // Patterns for matching conversions from float to half-float and vice versa.
8568 let Predicates = [HasF16C] in {
8569 def : Pat<(fp_to_f16 FR32:$src),
8570 (i16 (EXTRACT_SUBREG (VMOVPDI2DIrr (VCVTPS2PHrr
8571 (COPY_TO_REGCLASS FR32:$src, VR128), 0)), sub_16bit))>;
8573 def : Pat<(f16_to_fp GR16:$src),
8574 (f32 (COPY_TO_REGCLASS (VCVTPH2PSrr
8575 (COPY_TO_REGCLASS (MOVSX32rr16 GR16:$src), VR128)), FR32)) >;
8577 def : Pat<(f16_to_fp (i16 (fp_to_f16 FR32:$src))),
8578 (f32 (COPY_TO_REGCLASS (VCVTPH2PSrr
8579 (VCVTPS2PHrr (COPY_TO_REGCLASS FR32:$src, VR128), 0)), FR32)) >;
8582 //===----------------------------------------------------------------------===//
8583 // AVX2 Instructions
8584 //===----------------------------------------------------------------------===//
8586 /// AVX2_binop_rmi_int - AVX2 binary operator with 8-bit immediate
8587 multiclass AVX2_binop_rmi_int<bits<8> opc, string OpcodeStr,
8588 Intrinsic IntId, RegisterClass RC, PatFrag memop_frag,
8589 X86MemOperand x86memop> {
8590 let isCommutable = 1 in
8591 def rri : AVX2AIi8<opc, MRMSrcReg, (outs RC:$dst),
8592 (ins RC:$src1, RC:$src2, i8imm:$src3),
8593 !strconcat(OpcodeStr,
8594 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
8595 [(set RC:$dst, (IntId RC:$src1, RC:$src2, imm:$src3))]>,
8596 Sched<[WriteBlend]>, VEX_4V;
8597 def rmi : AVX2AIi8<opc, MRMSrcMem, (outs RC:$dst),
8598 (ins RC:$src1, x86memop:$src2, i8imm:$src3),
8599 !strconcat(OpcodeStr,
8600 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
8603 (bitconvert (memop_frag addr:$src2)), imm:$src3))]>,
8604 Sched<[WriteBlendLd, ReadAfterLd]>, VEX_4V;
8607 let isCommutable = 0 in {
8608 defm VPBLENDD : AVX2_binop_rmi_int<0x02, "vpblendd", int_x86_avx2_pblendd_128,
8609 VR128, loadv2i64, i128mem>;
8610 defm VPBLENDDY : AVX2_binop_rmi_int<0x02, "vpblendd", int_x86_avx2_pblendd_256,
8611 VR256, loadv4i64, i256mem>, VEX_L;
8614 def : Pat<(v4i32 (X86Blendi (v4i32 VR128:$src1), (v4i32 VR128:$src2),
8616 (VPBLENDDrri VR128:$src1, VR128:$src2, imm:$mask)>;
8617 def : Pat<(v8i32 (X86Blendi (v8i32 VR256:$src1), (v8i32 VR256:$src2),
8619 (VPBLENDDYrri VR256:$src1, VR256:$src2, imm:$mask)>;
8621 //===----------------------------------------------------------------------===//
8622 // VPBROADCAST - Load from memory and broadcast to all elements of the
8623 // destination operand
8625 multiclass avx2_broadcast<bits<8> opc, string OpcodeStr,
8626 X86MemOperand x86memop, PatFrag ld_frag,
8627 Intrinsic Int128, Intrinsic Int256> {
8628 def rr : AVX28I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
8629 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8630 [(set VR128:$dst, (Int128 VR128:$src))]>,
8631 Sched<[WriteShuffle]>, VEX;
8632 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst), (ins x86memop:$src),
8633 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8635 (Int128 (scalar_to_vector (ld_frag addr:$src))))]>,
8636 Sched<[WriteLoad]>, VEX;
8637 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
8638 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8639 [(set VR256:$dst, (Int256 VR128:$src))]>,
8640 Sched<[WriteShuffle256]>, VEX, VEX_L;
8641 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst), (ins x86memop:$src),
8642 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8644 (Int256 (scalar_to_vector (ld_frag addr:$src))))]>,
8645 Sched<[WriteLoad]>, VEX, VEX_L;
8648 defm VPBROADCASTB : avx2_broadcast<0x78, "vpbroadcastb", i8mem, loadi8,
8649 int_x86_avx2_pbroadcastb_128,
8650 int_x86_avx2_pbroadcastb_256>;
8651 defm VPBROADCASTW : avx2_broadcast<0x79, "vpbroadcastw", i16mem, loadi16,
8652 int_x86_avx2_pbroadcastw_128,
8653 int_x86_avx2_pbroadcastw_256>;
8654 defm VPBROADCASTD : avx2_broadcast<0x58, "vpbroadcastd", i32mem, loadi32,
8655 int_x86_avx2_pbroadcastd_128,
8656 int_x86_avx2_pbroadcastd_256>;
8657 defm VPBROADCASTQ : avx2_broadcast<0x59, "vpbroadcastq", i64mem, loadi64,
8658 int_x86_avx2_pbroadcastq_128,
8659 int_x86_avx2_pbroadcastq_256>;
8661 let Predicates = [HasAVX2] in {
8662 def : Pat<(v16i8 (X86VBroadcast (loadi8 addr:$src))),
8663 (VPBROADCASTBrm addr:$src)>;
8664 def : Pat<(v32i8 (X86VBroadcast (loadi8 addr:$src))),
8665 (VPBROADCASTBYrm addr:$src)>;
8666 def : Pat<(v8i16 (X86VBroadcast (loadi16 addr:$src))),
8667 (VPBROADCASTWrm addr:$src)>;
8668 def : Pat<(v16i16 (X86VBroadcast (loadi16 addr:$src))),
8669 (VPBROADCASTWYrm addr:$src)>;
8670 def : Pat<(v4i32 (X86VBroadcast (loadi32 addr:$src))),
8671 (VPBROADCASTDrm addr:$src)>;
8672 def : Pat<(v8i32 (X86VBroadcast (loadi32 addr:$src))),
8673 (VPBROADCASTDYrm addr:$src)>;
8674 def : Pat<(v2i64 (X86VBroadcast (loadi64 addr:$src))),
8675 (VPBROADCASTQrm addr:$src)>;
8676 def : Pat<(v4i64 (X86VBroadcast (loadi64 addr:$src))),
8677 (VPBROADCASTQYrm addr:$src)>;
8679 def : Pat<(v16i8 (X86VBroadcast (v16i8 VR128:$src))),
8680 (VPBROADCASTBrr VR128:$src)>;
8681 def : Pat<(v32i8 (X86VBroadcast (v16i8 VR128:$src))),
8682 (VPBROADCASTBYrr VR128:$src)>;
8683 def : Pat<(v8i16 (X86VBroadcast (v8i16 VR128:$src))),
8684 (VPBROADCASTWrr VR128:$src)>;
8685 def : Pat<(v16i16 (X86VBroadcast (v8i16 VR128:$src))),
8686 (VPBROADCASTWYrr VR128:$src)>;
8687 def : Pat<(v4i32 (X86VBroadcast (v4i32 VR128:$src))),
8688 (VPBROADCASTDrr VR128:$src)>;
8689 def : Pat<(v8i32 (X86VBroadcast (v4i32 VR128:$src))),
8690 (VPBROADCASTDYrr VR128:$src)>;
8691 def : Pat<(v2i64 (X86VBroadcast (v2i64 VR128:$src))),
8692 (VPBROADCASTQrr VR128:$src)>;
8693 def : Pat<(v4i64 (X86VBroadcast (v2i64 VR128:$src))),
8694 (VPBROADCASTQYrr VR128:$src)>;
8695 def : Pat<(v4f32 (X86VBroadcast (v4f32 VR128:$src))),
8696 (VBROADCASTSSrr VR128:$src)>;
8697 def : Pat<(v8f32 (X86VBroadcast (v4f32 VR128:$src))),
8698 (VBROADCASTSSYrr VR128:$src)>;
8699 def : Pat<(v2f64 (X86VBroadcast (v2f64 VR128:$src))),
8700 (VPBROADCASTQrr VR128:$src)>;
8701 def : Pat<(v4f64 (X86VBroadcast (v2f64 VR128:$src))),
8702 (VBROADCASTSDYrr VR128:$src)>;
8704 // Provide fallback in case the load node that is used in the patterns above
8705 // is used by additional users, which prevents the pattern selection.
8706 let AddedComplexity = 20 in {
8707 def : Pat<(v4f32 (X86VBroadcast FR32:$src)),
8708 (VBROADCASTSSrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
8709 def : Pat<(v8f32 (X86VBroadcast FR32:$src)),
8710 (VBROADCASTSSYrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
8711 def : Pat<(v4f64 (X86VBroadcast FR64:$src)),
8712 (VBROADCASTSDYrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
8714 def : Pat<(v4i32 (X86VBroadcast GR32:$src)),
8715 (VBROADCASTSSrr (COPY_TO_REGCLASS GR32:$src, VR128))>;
8716 def : Pat<(v8i32 (X86VBroadcast GR32:$src)),
8717 (VBROADCASTSSYrr (COPY_TO_REGCLASS GR32:$src, VR128))>;
8718 def : Pat<(v4i64 (X86VBroadcast GR64:$src)),
8719 (VBROADCASTSDYrr (COPY_TO_REGCLASS GR64:$src, VR128))>;
8721 def : Pat<(v16i8 (X86VBroadcast GR8:$src)),
8722 (VPBROADCASTBrr (COPY_TO_REGCLASS
8723 (i32 (SUBREG_TO_REG (i32 0), GR8:$src, sub_8bit)),
8725 def : Pat<(v32i8 (X86VBroadcast GR8:$src)),
8726 (VPBROADCASTBYrr (COPY_TO_REGCLASS
8727 (i32 (SUBREG_TO_REG (i32 0), GR8:$src, sub_8bit)),
8730 def : Pat<(v8i16 (X86VBroadcast GR16:$src)),
8731 (VPBROADCASTWrr (COPY_TO_REGCLASS
8732 (i32 (SUBREG_TO_REG (i32 0), GR16:$src, sub_16bit)),
8734 def : Pat<(v16i16 (X86VBroadcast GR16:$src)),
8735 (VPBROADCASTWYrr (COPY_TO_REGCLASS
8736 (i32 (SUBREG_TO_REG (i32 0), GR16:$src, sub_16bit)),
8739 // The patterns for VPBROADCASTD are not needed because they would match
8740 // the exact same thing as VBROADCASTSS patterns.
8742 def : Pat<(v2i64 (X86VBroadcast GR64:$src)),
8743 (VPBROADCASTQrr (COPY_TO_REGCLASS GR64:$src, VR128))>;
8744 // The v4i64 pattern is not needed because VBROADCASTSDYrr already match.
8748 // AVX1 broadcast patterns
8749 let Predicates = [HasAVX1Only] in {
8750 def : Pat<(v8i32 (X86VBroadcast (loadi32 addr:$src))),
8751 (VBROADCASTSSYrm addr:$src)>;
8752 def : Pat<(v4i64 (X86VBroadcast (loadi64 addr:$src))),
8753 (VBROADCASTSDYrm addr:$src)>;
8754 def : Pat<(v4i32 (X86VBroadcast (loadi32 addr:$src))),
8755 (VBROADCASTSSrm addr:$src)>;
8758 let Predicates = [HasAVX] in {
8759 // Provide fallback in case the load node that is used in the patterns above
8760 // is used by additional users, which prevents the pattern selection.
8761 let AddedComplexity = 20 in {
8762 // 128bit broadcasts:
8763 def : Pat<(v4f32 (X86VBroadcast FR32:$src)),
8764 (VPSHUFDri (COPY_TO_REGCLASS FR32:$src, VR128), 0)>;
8765 def : Pat<(v8f32 (X86VBroadcast FR32:$src)),
8766 (VINSERTF128rr (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)),
8767 (VPSHUFDri (COPY_TO_REGCLASS FR32:$src, VR128), 0), sub_xmm),
8768 (VPSHUFDri (COPY_TO_REGCLASS FR32:$src, VR128), 0), 1)>;
8769 def : Pat<(v4f64 (X86VBroadcast FR64:$src)),
8770 (VINSERTF128rr (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)),
8771 (VPSHUFDri (COPY_TO_REGCLASS FR64:$src, VR128), 0x44), sub_xmm),
8772 (VPSHUFDri (COPY_TO_REGCLASS FR64:$src, VR128), 0x44), 1)>;
8774 def : Pat<(v4i32 (X86VBroadcast GR32:$src)),
8775 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0)>;
8776 def : Pat<(v8i32 (X86VBroadcast GR32:$src)),
8777 (VINSERTF128rr (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8778 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0), sub_xmm),
8779 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0), 1)>;
8780 def : Pat<(v4i64 (X86VBroadcast GR64:$src)),
8781 (VINSERTF128rr (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)),
8782 (VPSHUFDri (COPY_TO_REGCLASS GR64:$src, VR128), 0x44), sub_xmm),
8783 (VPSHUFDri (COPY_TO_REGCLASS GR64:$src, VR128), 0x44), 1)>;
8786 def : Pat<(v2f64 (X86VBroadcast f64:$src)),
8787 (VMOVDDUPrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
8790 //===----------------------------------------------------------------------===//
8791 // VPERM - Permute instructions
8794 multiclass avx2_perm<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
8795 ValueType OpVT, X86FoldableSchedWrite Sched> {
8796 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst),
8797 (ins VR256:$src1, VR256:$src2),
8798 !strconcat(OpcodeStr,
8799 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8801 (OpVT (X86VPermv VR256:$src1, VR256:$src2)))]>,
8802 Sched<[Sched]>, VEX_4V, VEX_L;
8803 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst),
8804 (ins VR256:$src1, i256mem:$src2),
8805 !strconcat(OpcodeStr,
8806 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8808 (OpVT (X86VPermv VR256:$src1,
8809 (bitconvert (mem_frag addr:$src2)))))]>,
8810 Sched<[Sched.Folded, ReadAfterLd]>, VEX_4V, VEX_L;
8813 defm VPERMD : avx2_perm<0x36, "vpermd", loadv4i64, v8i32, WriteShuffle256>;
8814 let ExeDomain = SSEPackedSingle in
8815 defm VPERMPS : avx2_perm<0x16, "vpermps", loadv8f32, v8f32, WriteFShuffle256>;
8817 multiclass avx2_perm_imm<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
8818 ValueType OpVT, X86FoldableSchedWrite Sched> {
8819 def Yri : AVX2AIi8<opc, MRMSrcReg, (outs VR256:$dst),
8820 (ins VR256:$src1, i8imm:$src2),
8821 !strconcat(OpcodeStr,
8822 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8824 (OpVT (X86VPermi VR256:$src1, (i8 imm:$src2))))]>,
8825 Sched<[Sched]>, VEX, VEX_L;
8826 def Ymi : AVX2AIi8<opc, MRMSrcMem, (outs VR256:$dst),
8827 (ins i256mem:$src1, i8imm:$src2),
8828 !strconcat(OpcodeStr,
8829 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8831 (OpVT (X86VPermi (mem_frag addr:$src1),
8832 (i8 imm:$src2))))]>,
8833 Sched<[Sched.Folded, ReadAfterLd]>, VEX, VEX_L;
8836 defm VPERMQ : avx2_perm_imm<0x00, "vpermq", loadv4i64, v4i64,
8837 WriteShuffle256>, VEX_W;
8838 let ExeDomain = SSEPackedDouble in
8839 defm VPERMPD : avx2_perm_imm<0x01, "vpermpd", loadv4f64, v4f64,
8840 WriteFShuffle256>, VEX_W;
8842 //===----------------------------------------------------------------------===//
8843 // VPERM2I128 - Permute Floating-Point Values in 128-bit chunks
8845 def VPERM2I128rr : AVX2AIi8<0x46, MRMSrcReg, (outs VR256:$dst),
8846 (ins VR256:$src1, VR256:$src2, i8imm:$src3),
8847 "vperm2i128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8848 [(set VR256:$dst, (v4i64 (X86VPerm2x128 VR256:$src1, VR256:$src2,
8849 (i8 imm:$src3))))]>, Sched<[WriteShuffle256]>,
8851 def VPERM2I128rm : AVX2AIi8<0x46, MRMSrcMem, (outs VR256:$dst),
8852 (ins VR256:$src1, f256mem:$src2, i8imm:$src3),
8853 "vperm2i128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8854 [(set VR256:$dst, (X86VPerm2x128 VR256:$src1, (loadv4i64 addr:$src2),
8856 Sched<[WriteShuffle256Ld, ReadAfterLd]>, VEX_4V, VEX_L;
8858 let Predicates = [HasAVX2] in {
8859 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8860 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8861 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8862 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8863 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8864 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8866 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, (bc_v32i8 (loadv4i64 addr:$src2)),
8868 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8869 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1,
8870 (bc_v16i16 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8871 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8872 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)),
8874 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8878 //===----------------------------------------------------------------------===//
8879 // VINSERTI128 - Insert packed integer values
8881 let neverHasSideEffects = 1 in {
8882 def VINSERTI128rr : AVX2AIi8<0x38, MRMSrcReg, (outs VR256:$dst),
8883 (ins VR256:$src1, VR128:$src2, i8imm:$src3),
8884 "vinserti128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8885 []>, Sched<[WriteShuffle256]>, VEX_4V, VEX_L;
8887 def VINSERTI128rm : AVX2AIi8<0x38, MRMSrcMem, (outs VR256:$dst),
8888 (ins VR256:$src1, i128mem:$src2, i8imm:$src3),
8889 "vinserti128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8890 []>, Sched<[WriteShuffle256Ld, ReadAfterLd]>, VEX_4V, VEX_L;
8893 let Predicates = [HasAVX2] in {
8894 def : Pat<(vinsert128_insert:$ins (v4i64 VR256:$src1), (v2i64 VR128:$src2),
8896 (VINSERTI128rr VR256:$src1, VR128:$src2,
8897 (INSERT_get_vinsert128_imm VR256:$ins))>;
8898 def : Pat<(vinsert128_insert:$ins (v8i32 VR256:$src1), (v4i32 VR128:$src2),
8900 (VINSERTI128rr VR256:$src1, VR128:$src2,
8901 (INSERT_get_vinsert128_imm VR256:$ins))>;
8902 def : Pat<(vinsert128_insert:$ins (v32i8 VR256:$src1), (v16i8 VR128:$src2),
8904 (VINSERTI128rr VR256:$src1, VR128:$src2,
8905 (INSERT_get_vinsert128_imm VR256:$ins))>;
8906 def : Pat<(vinsert128_insert:$ins (v16i16 VR256:$src1), (v8i16 VR128:$src2),
8908 (VINSERTI128rr VR256:$src1, VR128:$src2,
8909 (INSERT_get_vinsert128_imm VR256:$ins))>;
8911 def : Pat<(vinsert128_insert:$ins (v4i64 VR256:$src1), (loadv2i64 addr:$src2),
8913 (VINSERTI128rm VR256:$src1, addr:$src2,
8914 (INSERT_get_vinsert128_imm VR256:$ins))>;
8915 def : Pat<(vinsert128_insert:$ins (v8i32 VR256:$src1),
8916 (bc_v4i32 (loadv2i64 addr:$src2)),
8918 (VINSERTI128rm VR256:$src1, addr:$src2,
8919 (INSERT_get_vinsert128_imm VR256:$ins))>;
8920 def : Pat<(vinsert128_insert:$ins (v32i8 VR256:$src1),
8921 (bc_v16i8 (loadv2i64 addr:$src2)),
8923 (VINSERTI128rm VR256:$src1, addr:$src2,
8924 (INSERT_get_vinsert128_imm VR256:$ins))>;
8925 def : Pat<(vinsert128_insert:$ins (v16i16 VR256:$src1),
8926 (bc_v8i16 (loadv2i64 addr:$src2)),
8928 (VINSERTI128rm VR256:$src1, addr:$src2,
8929 (INSERT_get_vinsert128_imm VR256:$ins))>;
8932 //===----------------------------------------------------------------------===//
8933 // VEXTRACTI128 - Extract packed integer values
8935 def VEXTRACTI128rr : AVX2AIi8<0x39, MRMDestReg, (outs VR128:$dst),
8936 (ins VR256:$src1, i8imm:$src2),
8937 "vextracti128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8939 (int_x86_avx2_vextracti128 VR256:$src1, imm:$src2))]>,
8940 Sched<[WriteShuffle256]>, VEX, VEX_L;
8941 let neverHasSideEffects = 1, mayStore = 1 in
8942 def VEXTRACTI128mr : AVX2AIi8<0x39, MRMDestMem, (outs),
8943 (ins i128mem:$dst, VR256:$src1, i8imm:$src2),
8944 "vextracti128\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8945 Sched<[WriteStore]>, VEX, VEX_L;
8947 let Predicates = [HasAVX2] in {
8948 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8949 (v2i64 (VEXTRACTI128rr
8950 (v4i64 VR256:$src1),
8951 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8952 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8953 (v4i32 (VEXTRACTI128rr
8954 (v8i32 VR256:$src1),
8955 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8956 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8957 (v8i16 (VEXTRACTI128rr
8958 (v16i16 VR256:$src1),
8959 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8960 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8961 (v16i8 (VEXTRACTI128rr
8962 (v32i8 VR256:$src1),
8963 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8965 def : Pat<(store (v2i64 (vextract128_extract:$ext (v4i64 VR256:$src1),
8966 (iPTR imm))), addr:$dst),
8967 (VEXTRACTI128mr addr:$dst, VR256:$src1,
8968 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8969 def : Pat<(store (v4i32 (vextract128_extract:$ext (v8i32 VR256:$src1),
8970 (iPTR imm))), addr:$dst),
8971 (VEXTRACTI128mr addr:$dst, VR256:$src1,
8972 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8973 def : Pat<(store (v8i16 (vextract128_extract:$ext (v16i16 VR256:$src1),
8974 (iPTR imm))), addr:$dst),
8975 (VEXTRACTI128mr addr:$dst, VR256:$src1,
8976 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8977 def : Pat<(store (v16i8 (vextract128_extract:$ext (v32i8 VR256:$src1),
8978 (iPTR imm))), addr:$dst),
8979 (VEXTRACTI128mr addr:$dst, VR256:$src1,
8980 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8983 //===----------------------------------------------------------------------===//
8984 // VPMASKMOV - Conditional SIMD Integer Packed Loads and Stores
8986 multiclass avx2_pmovmask<string OpcodeStr,
8987 Intrinsic IntLd128, Intrinsic IntLd256,
8988 Intrinsic IntSt128, Intrinsic IntSt256> {
8989 def rm : AVX28I<0x8c, MRMSrcMem, (outs VR128:$dst),
8990 (ins VR128:$src1, i128mem:$src2),
8991 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8992 [(set VR128:$dst, (IntLd128 addr:$src2, VR128:$src1))]>, VEX_4V;
8993 def Yrm : AVX28I<0x8c, MRMSrcMem, (outs VR256:$dst),
8994 (ins VR256:$src1, i256mem:$src2),
8995 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8996 [(set VR256:$dst, (IntLd256 addr:$src2, VR256:$src1))]>,
8998 def mr : AVX28I<0x8e, MRMDestMem, (outs),
8999 (ins i128mem:$dst, VR128:$src1, VR128:$src2),
9000 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9001 [(IntSt128 addr:$dst, VR128:$src1, VR128:$src2)]>, VEX_4V;
9002 def Ymr : AVX28I<0x8e, MRMDestMem, (outs),
9003 (ins i256mem:$dst, VR256:$src1, VR256:$src2),
9004 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9005 [(IntSt256 addr:$dst, VR256:$src1, VR256:$src2)]>, VEX_4V, VEX_L;
9008 defm VPMASKMOVD : avx2_pmovmask<"vpmaskmovd",
9009 int_x86_avx2_maskload_d,
9010 int_x86_avx2_maskload_d_256,
9011 int_x86_avx2_maskstore_d,
9012 int_x86_avx2_maskstore_d_256>;
9013 defm VPMASKMOVQ : avx2_pmovmask<"vpmaskmovq",
9014 int_x86_avx2_maskload_q,
9015 int_x86_avx2_maskload_q_256,
9016 int_x86_avx2_maskstore_q,
9017 int_x86_avx2_maskstore_q_256>, VEX_W;
9020 //===----------------------------------------------------------------------===//
9021 // Variable Bit Shifts
9023 multiclass avx2_var_shift<bits<8> opc, string OpcodeStr, SDNode OpNode,
9024 ValueType vt128, ValueType vt256> {
9025 def rr : AVX28I<opc, MRMSrcReg, (outs VR128:$dst),
9026 (ins VR128:$src1, VR128:$src2),
9027 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9029 (vt128 (OpNode VR128:$src1, (vt128 VR128:$src2))))]>,
9030 VEX_4V, Sched<[WriteVarVecShift]>;
9031 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst),
9032 (ins VR128:$src1, i128mem:$src2),
9033 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9035 (vt128 (OpNode VR128:$src1,
9036 (vt128 (bitconvert (loadv2i64 addr:$src2))))))]>,
9037 VEX_4V, Sched<[WriteVarVecShiftLd, ReadAfterLd]>;
9038 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst),
9039 (ins VR256:$src1, VR256:$src2),
9040 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9042 (vt256 (OpNode VR256:$src1, (vt256 VR256:$src2))))]>,
9043 VEX_4V, VEX_L, Sched<[WriteVarVecShift]>;
9044 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst),
9045 (ins VR256:$src1, i256mem:$src2),
9046 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
9048 (vt256 (OpNode VR256:$src1,
9049 (vt256 (bitconvert (loadv4i64 addr:$src2))))))]>,
9050 VEX_4V, VEX_L, Sched<[WriteVarVecShiftLd, ReadAfterLd]>;
9053 defm VPSLLVD : avx2_var_shift<0x47, "vpsllvd", shl, v4i32, v8i32>;
9054 defm VPSLLVQ : avx2_var_shift<0x47, "vpsllvq", shl, v2i64, v4i64>, VEX_W;
9055 defm VPSRLVD : avx2_var_shift<0x45, "vpsrlvd", srl, v4i32, v8i32>;
9056 defm VPSRLVQ : avx2_var_shift<0x45, "vpsrlvq", srl, v2i64, v4i64>, VEX_W;
9057 defm VPSRAVD : avx2_var_shift<0x46, "vpsravd", sra, v4i32, v8i32>;
9059 //===----------------------------------------------------------------------===//
9060 // VGATHER - GATHER Operations
9061 multiclass avx2_gather<bits<8> opc, string OpcodeStr, RegisterClass RC256,
9062 X86MemOperand memop128, X86MemOperand memop256> {
9063 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst, VR128:$mask_wb),
9064 (ins VR128:$src1, memop128:$src2, VR128:$mask),
9065 !strconcat(OpcodeStr,
9066 "\t{$mask, $src2, $dst|$dst, $src2, $mask}"),
9068 def Yrm : AVX28I<opc, MRMSrcMem, (outs RC256:$dst, RC256:$mask_wb),
9069 (ins RC256:$src1, memop256:$src2, RC256:$mask),
9070 !strconcat(OpcodeStr,
9071 "\t{$mask, $src2, $dst|$dst, $src2, $mask}"),
9072 []>, VEX_4VOp3, VEX_L;
9075 let mayLoad = 1, Constraints
9076 = "@earlyclobber $dst,@earlyclobber $mask_wb, $src1 = $dst, $mask = $mask_wb"
9078 defm VPGATHERDQ : avx2_gather<0x90, "vpgatherdq", VR256, vx64mem, vx64mem>, VEX_W;
9079 defm VPGATHERQQ : avx2_gather<0x91, "vpgatherqq", VR256, vx64mem, vy64mem>, VEX_W;
9080 defm VPGATHERDD : avx2_gather<0x90, "vpgatherdd", VR256, vx32mem, vy32mem>;
9081 defm VPGATHERQD : avx2_gather<0x91, "vpgatherqd", VR128, vx32mem, vy32mem>;
9083 let ExeDomain = SSEPackedDouble in {
9084 defm VGATHERDPD : avx2_gather<0x92, "vgatherdpd", VR256, vx64mem, vx64mem>, VEX_W;
9085 defm VGATHERQPD : avx2_gather<0x93, "vgatherqpd", VR256, vx64mem, vy64mem>, VEX_W;
9088 let ExeDomain = SSEPackedSingle in {
9089 defm VGATHERDPS : avx2_gather<0x92, "vgatherdps", VR256, vx32mem, vy32mem>;
9090 defm VGATHERQPS : avx2_gather<0x93, "vgatherqps", VR128, vx32mem, vy32mem>;