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;
37 let Sched = WriteFAdd in {
38 def SSE_ALU_F32S : OpndItins<
39 IIC_SSE_ALU_F32S_RR, IIC_SSE_ALU_F32S_RM
42 def SSE_ALU_F64S : OpndItins<
43 IIC_SSE_ALU_F64S_RR, IIC_SSE_ALU_F64S_RM
47 def SSE_ALU_ITINS_S : SizeItins<
48 SSE_ALU_F32S, SSE_ALU_F64S
51 let Sched = WriteFMul in {
52 def SSE_MUL_F32S : OpndItins<
53 IIC_SSE_MUL_F32S_RR, IIC_SSE_MUL_F64S_RM
56 def SSE_MUL_F64S : OpndItins<
57 IIC_SSE_MUL_F64S_RR, IIC_SSE_MUL_F64S_RM
61 def SSE_MUL_ITINS_S : SizeItins<
62 SSE_MUL_F32S, SSE_MUL_F64S
65 let Sched = WriteFDiv in {
66 def SSE_DIV_F32S : OpndItins<
67 IIC_SSE_DIV_F32S_RR, IIC_SSE_DIV_F64S_RM
70 def SSE_DIV_F64S : OpndItins<
71 IIC_SSE_DIV_F64S_RR, IIC_SSE_DIV_F64S_RM
75 def SSE_DIV_ITINS_S : SizeItins<
76 SSE_DIV_F32S, SSE_DIV_F64S
80 let Sched = WriteFAdd in {
81 def SSE_ALU_F32P : OpndItins<
82 IIC_SSE_ALU_F32P_RR, IIC_SSE_ALU_F32P_RM
85 def SSE_ALU_F64P : OpndItins<
86 IIC_SSE_ALU_F64P_RR, IIC_SSE_ALU_F64P_RM
90 def SSE_ALU_ITINS_P : SizeItins<
91 SSE_ALU_F32P, SSE_ALU_F64P
94 let Sched = WriteFMul in {
95 def SSE_MUL_F32P : OpndItins<
96 IIC_SSE_MUL_F32P_RR, IIC_SSE_MUL_F64P_RM
99 def SSE_MUL_F64P : OpndItins<
100 IIC_SSE_MUL_F64P_RR, IIC_SSE_MUL_F64P_RM
104 def SSE_MUL_ITINS_P : SizeItins<
105 SSE_MUL_F32P, SSE_MUL_F64P
108 let Sched = WriteFDiv in {
109 def SSE_DIV_F32P : OpndItins<
110 IIC_SSE_DIV_F32P_RR, IIC_SSE_DIV_F64P_RM
113 def SSE_DIV_F64P : OpndItins<
114 IIC_SSE_DIV_F64P_RR, IIC_SSE_DIV_F64P_RM
118 def SSE_DIV_ITINS_P : SizeItins<
119 SSE_DIV_F32P, SSE_DIV_F64P
122 let Sched = WriteVecLogic in
123 def SSE_VEC_BIT_ITINS_P : OpndItins<
124 IIC_SSE_BIT_P_RR, IIC_SSE_BIT_P_RM
127 def SSE_BIT_ITINS_P : OpndItins<
128 IIC_SSE_BIT_P_RR, IIC_SSE_BIT_P_RM
131 let Sched = WriteVecALU in {
132 def SSE_INTALU_ITINS_P : OpndItins<
133 IIC_SSE_INTALU_P_RR, IIC_SSE_INTALU_P_RM
136 def SSE_INTALUQ_ITINS_P : OpndItins<
137 IIC_SSE_INTALUQ_P_RR, IIC_SSE_INTALUQ_P_RM
141 let Sched = WriteVecIMul in
142 def SSE_INTMUL_ITINS_P : OpndItins<
143 IIC_SSE_INTMUL_P_RR, IIC_SSE_INTMUL_P_RM
146 def SSE_INTSHIFT_ITINS_P : ShiftOpndItins<
147 IIC_SSE_INTSH_P_RR, IIC_SSE_INTSH_P_RM, IIC_SSE_INTSH_P_RI
150 def SSE_MOVA_ITINS : OpndItins<
151 IIC_SSE_MOVA_P_RR, IIC_SSE_MOVA_P_RM
154 def SSE_MOVU_ITINS : OpndItins<
155 IIC_SSE_MOVU_P_RR, IIC_SSE_MOVU_P_RM
158 def SSE_DPPD_ITINS : OpndItins<
159 IIC_SSE_DPPD_RR, IIC_SSE_DPPD_RM
162 def SSE_DPPS_ITINS : OpndItins<
163 IIC_SSE_DPPS_RR, IIC_SSE_DPPD_RM
166 def DEFAULT_ITINS : OpndItins<
167 IIC_ALU_NONMEM, IIC_ALU_MEM
170 def SSE_EXTRACT_ITINS : OpndItins<
171 IIC_SSE_EXTRACTPS_RR, IIC_SSE_EXTRACTPS_RM
174 def SSE_INSERT_ITINS : OpndItins<
175 IIC_SSE_INSERTPS_RR, IIC_SSE_INSERTPS_RM
178 let Sched = WriteMPSAD in
179 def SSE_MPSADBW_ITINS : OpndItins<
180 IIC_SSE_MPSADBW_RR, IIC_SSE_MPSADBW_RM
183 let Sched = WriteVecIMul in
184 def SSE_PMULLD_ITINS : OpndItins<
185 IIC_SSE_PMULLD_RR, IIC_SSE_PMULLD_RM
188 // Definitions for backward compatibility.
189 // The instructions mapped on these definitions uses a different itinerary
190 // than the actual scheduling model.
191 let Sched = WriteShuffle in
192 def DEFAULT_ITINS_SHUFFLESCHED : OpndItins<
193 IIC_ALU_NONMEM, IIC_ALU_MEM
196 let Sched = WriteVecIMul in
197 def DEFAULT_ITINS_VECIMULSCHED : OpndItins<
198 IIC_ALU_NONMEM, IIC_ALU_MEM
201 let Sched = WriteShuffle in
202 def SSE_INTALU_ITINS_SHUFF_P : OpndItins<
203 IIC_SSE_INTALU_P_RR, IIC_SSE_INTALU_P_RM
206 let Sched = WriteMPSAD in
207 def DEFAULT_ITINS_MPSADSCHED : OpndItins<
208 IIC_ALU_NONMEM, IIC_ALU_MEM
211 let Sched = WriteFBlend in
212 def DEFAULT_ITINS_FBLENDSCHED : OpndItins<
213 IIC_ALU_NONMEM, IIC_ALU_MEM
216 let Sched = WriteBlend in
217 def DEFAULT_ITINS_BLENDSCHED : OpndItins<
218 IIC_ALU_NONMEM, IIC_ALU_MEM
221 let Sched = WriteVarBlend in
222 def DEFAULT_ITINS_VARBLENDSCHED : OpndItins<
223 IIC_ALU_NONMEM, IIC_ALU_MEM
226 let Sched = WriteFBlend in
227 def SSE_INTALU_ITINS_FBLEND_P : OpndItins<
228 IIC_SSE_INTALU_P_RR, IIC_SSE_INTALU_P_RM
231 let Sched = WriteBlend in
232 def SSE_INTALU_ITINS_BLEND_P : OpndItins<
233 IIC_SSE_INTALU_P_RR, IIC_SSE_INTALU_P_RM
236 //===----------------------------------------------------------------------===//
237 // SSE 1 & 2 Instructions Classes
238 //===----------------------------------------------------------------------===//
240 /// sse12_fp_scalar - SSE 1 & 2 scalar instructions class
241 multiclass sse12_fp_scalar<bits<8> opc, string OpcodeStr, SDNode OpNode,
242 RegisterClass RC, X86MemOperand x86memop,
243 Domain d, OpndItins itins, bit Is2Addr = 1> {
244 let isCommutable = 1 in {
245 def rr : SI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
247 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
248 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
249 [(set RC:$dst, (OpNode RC:$src1, RC:$src2))], itins.rr, d>,
250 Sched<[itins.Sched]>;
252 def rm : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
254 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
255 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
256 [(set RC:$dst, (OpNode RC:$src1, (load addr:$src2)))], itins.rm, d>,
257 Sched<[itins.Sched.Folded, ReadAfterLd]>;
260 /// sse12_fp_scalar_int - SSE 1 & 2 scalar instructions intrinsics class
261 multiclass sse12_fp_scalar_int<bits<8> opc, string OpcodeStr,
262 SDPatternOperator OpNode, RegisterClass RC,
263 ValueType VT, string asm, Operand memopr,
264 ComplexPattern mem_cpat, Domain d,
265 OpndItins itins, bit Is2Addr = 1> {
266 let isCodeGenOnly = 1, hasSideEffects = 0 in {
267 def rr_Int : SI_Int<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
269 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
270 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
271 [(set RC:$dst, (VT (OpNode RC:$src1, RC:$src2)))], itins.rr, d>,
272 Sched<[itins.Sched]>;
274 def rm_Int : SI_Int<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memopr:$src2),
276 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
277 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
278 [(set RC:$dst, (VT (OpNode RC:$src1, mem_cpat:$src2)))], itins.rm, d>,
279 Sched<[itins.Sched.Folded, ReadAfterLd]>;
283 /// sse12_fp_packed - SSE 1 & 2 packed instructions class
284 multiclass sse12_fp_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
285 RegisterClass RC, ValueType vt,
286 X86MemOperand x86memop, PatFrag mem_frag,
287 Domain d, OpndItins itins, bit Is2Addr = 1> {
288 let isCommutable = 1 in
289 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
291 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
292 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
293 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], itins.rr, d>,
294 Sched<[itins.Sched]>;
296 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
298 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
299 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
300 [(set RC:$dst, (OpNode RC:$src1, (mem_frag addr:$src2)))],
302 Sched<[itins.Sched.Folded, ReadAfterLd]>;
305 /// sse12_fp_packed_logical_rm - SSE 1 & 2 packed instructions class
306 multiclass sse12_fp_packed_logical_rm<bits<8> opc, RegisterClass RC, Domain d,
307 string OpcodeStr, X86MemOperand x86memop,
308 list<dag> pat_rr, list<dag> pat_rm,
310 let isCommutable = 1, hasSideEffects = 0 in
311 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
313 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
314 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
315 pat_rr, NoItinerary, d>,
316 Sched<[WriteVecLogic]>;
317 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
319 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
320 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
321 pat_rm, NoItinerary, d>,
322 Sched<[WriteVecLogicLd, ReadAfterLd]>;
325 //===----------------------------------------------------------------------===//
326 // Non-instruction patterns
327 //===----------------------------------------------------------------------===//
329 // A vector extract of the first f32/f64 position is a subregister copy
330 def : Pat<(f32 (extractelt (v4f32 VR128:$src), (iPTR 0))),
331 (COPY_TO_REGCLASS (v4f32 VR128:$src), FR32)>;
332 def : Pat<(f64 (extractelt (v2f64 VR128:$src), (iPTR 0))),
333 (COPY_TO_REGCLASS (v2f64 VR128:$src), FR64)>;
335 // A 128-bit subvector extract from the first 256-bit vector position
336 // is a subregister copy that needs no instruction.
337 def : Pat<(v4i32 (extract_subvector (v8i32 VR256:$src), (iPTR 0))),
338 (v4i32 (EXTRACT_SUBREG (v8i32 VR256:$src), sub_xmm))>;
339 def : Pat<(v4f32 (extract_subvector (v8f32 VR256:$src), (iPTR 0))),
340 (v4f32 (EXTRACT_SUBREG (v8f32 VR256:$src), sub_xmm))>;
342 def : Pat<(v2i64 (extract_subvector (v4i64 VR256:$src), (iPTR 0))),
343 (v2i64 (EXTRACT_SUBREG (v4i64 VR256:$src), sub_xmm))>;
344 def : Pat<(v2f64 (extract_subvector (v4f64 VR256:$src), (iPTR 0))),
345 (v2f64 (EXTRACT_SUBREG (v4f64 VR256:$src), sub_xmm))>;
347 def : Pat<(v8i16 (extract_subvector (v16i16 VR256:$src), (iPTR 0))),
348 (v8i16 (EXTRACT_SUBREG (v16i16 VR256:$src), sub_xmm))>;
349 def : Pat<(v16i8 (extract_subvector (v32i8 VR256:$src), (iPTR 0))),
350 (v16i8 (EXTRACT_SUBREG (v32i8 VR256:$src), sub_xmm))>;
352 // A 128-bit subvector insert to the first 256-bit vector position
353 // is a subregister copy that needs no instruction.
354 let AddedComplexity = 25 in { // to give priority over vinsertf128rm
355 def : Pat<(insert_subvector undef, (v2i64 VR128:$src), (iPTR 0)),
356 (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
357 def : Pat<(insert_subvector undef, (v2f64 VR128:$src), (iPTR 0)),
358 (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
359 def : Pat<(insert_subvector undef, (v4i32 VR128:$src), (iPTR 0)),
360 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
361 def : Pat<(insert_subvector undef, (v4f32 VR128:$src), (iPTR 0)),
362 (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
363 def : Pat<(insert_subvector undef, (v8i16 VR128:$src), (iPTR 0)),
364 (INSERT_SUBREG (v16i16 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
365 def : Pat<(insert_subvector undef, (v16i8 VR128:$src), (iPTR 0)),
366 (INSERT_SUBREG (v32i8 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
369 // Implicitly promote a 32-bit scalar to a vector.
370 def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
371 (COPY_TO_REGCLASS FR32:$src, VR128)>;
372 // Implicitly promote a 64-bit scalar to a vector.
373 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
374 (COPY_TO_REGCLASS FR64:$src, VR128)>;
376 // Bitcasts between 128-bit vector types. Return the original type since
377 // no instruction is needed for the conversion
378 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
379 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
380 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
381 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
382 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
383 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
384 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
385 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
386 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
387 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
388 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
389 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
390 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
391 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
392 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
393 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
394 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
395 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
396 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
397 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
398 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
399 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
400 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
401 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
402 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
403 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
404 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
405 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
406 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
407 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
408 def : Pat<(f128 (bitconvert (i128 FR128:$src))), (f128 FR128:$src)>;
409 def : Pat<(i128 (bitconvert (f128 FR128:$src))), (i128 FR128:$src)>;
411 // Bitcasts between 256-bit vector types. Return the original type since
412 // no instruction is needed for the conversion
413 def : Pat<(v4i64 (bitconvert (v8i32 VR256:$src))), (v4i64 VR256:$src)>;
414 def : Pat<(v4i64 (bitconvert (v16i16 VR256:$src))), (v4i64 VR256:$src)>;
415 def : Pat<(v4i64 (bitconvert (v32i8 VR256:$src))), (v4i64 VR256:$src)>;
416 def : Pat<(v4i64 (bitconvert (v8f32 VR256:$src))), (v4i64 VR256:$src)>;
417 def : Pat<(v4i64 (bitconvert (v4f64 VR256:$src))), (v4i64 VR256:$src)>;
418 def : Pat<(v8i32 (bitconvert (v4i64 VR256:$src))), (v8i32 VR256:$src)>;
419 def : Pat<(v8i32 (bitconvert (v16i16 VR256:$src))), (v8i32 VR256:$src)>;
420 def : Pat<(v8i32 (bitconvert (v32i8 VR256:$src))), (v8i32 VR256:$src)>;
421 def : Pat<(v8i32 (bitconvert (v4f64 VR256:$src))), (v8i32 VR256:$src)>;
422 def : Pat<(v8i32 (bitconvert (v8f32 VR256:$src))), (v8i32 VR256:$src)>;
423 def : Pat<(v16i16 (bitconvert (v4i64 VR256:$src))), (v16i16 VR256:$src)>;
424 def : Pat<(v16i16 (bitconvert (v8i32 VR256:$src))), (v16i16 VR256:$src)>;
425 def : Pat<(v16i16 (bitconvert (v32i8 VR256:$src))), (v16i16 VR256:$src)>;
426 def : Pat<(v16i16 (bitconvert (v4f64 VR256:$src))), (v16i16 VR256:$src)>;
427 def : Pat<(v16i16 (bitconvert (v8f32 VR256:$src))), (v16i16 VR256:$src)>;
428 def : Pat<(v32i8 (bitconvert (v4i64 VR256:$src))), (v32i8 VR256:$src)>;
429 def : Pat<(v32i8 (bitconvert (v8i32 VR256:$src))), (v32i8 VR256:$src)>;
430 def : Pat<(v32i8 (bitconvert (v16i16 VR256:$src))), (v32i8 VR256:$src)>;
431 def : Pat<(v32i8 (bitconvert (v4f64 VR256:$src))), (v32i8 VR256:$src)>;
432 def : Pat<(v32i8 (bitconvert (v8f32 VR256:$src))), (v32i8 VR256:$src)>;
433 def : Pat<(v8f32 (bitconvert (v4i64 VR256:$src))), (v8f32 VR256:$src)>;
434 def : Pat<(v8f32 (bitconvert (v8i32 VR256:$src))), (v8f32 VR256:$src)>;
435 def : Pat<(v8f32 (bitconvert (v16i16 VR256:$src))), (v8f32 VR256:$src)>;
436 def : Pat<(v8f32 (bitconvert (v32i8 VR256:$src))), (v8f32 VR256:$src)>;
437 def : Pat<(v8f32 (bitconvert (v4f64 VR256:$src))), (v8f32 VR256:$src)>;
438 def : Pat<(v4f64 (bitconvert (v4i64 VR256:$src))), (v4f64 VR256:$src)>;
439 def : Pat<(v4f64 (bitconvert (v8i32 VR256:$src))), (v4f64 VR256:$src)>;
440 def : Pat<(v4f64 (bitconvert (v16i16 VR256:$src))), (v4f64 VR256:$src)>;
441 def : Pat<(v4f64 (bitconvert (v32i8 VR256:$src))), (v4f64 VR256:$src)>;
442 def : Pat<(v4f64 (bitconvert (v8f32 VR256:$src))), (v4f64 VR256:$src)>;
444 // Alias instructions that map fld0 to xorps for sse or vxorps for avx.
445 // This is expanded by ExpandPostRAPseudos.
446 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
447 isPseudo = 1, SchedRW = [WriteZero] in {
448 def FsFLD0SS : I<0, Pseudo, (outs FR32:$dst), (ins), "",
449 [(set FR32:$dst, fp32imm0)]>, Requires<[HasSSE1, NoAVX512]>;
450 def FsFLD0SD : I<0, Pseudo, (outs FR64:$dst), (ins), "",
451 [(set FR64:$dst, fpimm0)]>, Requires<[HasSSE2, NoAVX512]>;
454 //===----------------------------------------------------------------------===//
455 // AVX & SSE - Zero/One Vectors
456 //===----------------------------------------------------------------------===//
458 // Alias instruction that maps zero vector to pxor / xorp* for sse.
459 // This is expanded by ExpandPostRAPseudos to an xorps / vxorps, and then
460 // swizzled by ExecutionDepsFix to pxor.
461 // We set canFoldAsLoad because this can be converted to a constant-pool
462 // load of an all-zeros value if folding it would be beneficial.
463 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
464 isPseudo = 1, SchedRW = [WriteZero] in {
465 def V_SET0 : I<0, Pseudo, (outs VR128:$dst), (ins), "",
466 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
469 let Predicates = [NoAVX512] in
470 def : Pat<(v4i32 immAllZerosV), (V_SET0)>;
473 // The same as done above but for AVX. The 256-bit AVX1 ISA doesn't support PI,
474 // and doesn't need it because on sandy bridge the register is set to zero
475 // at the rename stage without using any execution unit, so SET0PSY
476 // and SET0PDY can be used for vector int instructions without penalty
477 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
478 isPseudo = 1, Predicates = [NoAVX512], SchedRW = [WriteZero] in {
479 def AVX_SET0 : I<0, Pseudo, (outs VR256:$dst), (ins), "",
480 [(set VR256:$dst, (v8i32 immAllZerosV))]>;
483 // We set canFoldAsLoad because this can be converted to a constant-pool
484 // load of an all-ones value if folding it would be beneficial.
485 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
486 isPseudo = 1, SchedRW = [WriteZero] in {
487 def V_SETALLONES : I<0, Pseudo, (outs VR128:$dst), (ins), "",
488 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
489 let Predicates = [HasAVX2] in
490 def AVX2_SETALLONES : I<0, Pseudo, (outs VR256:$dst), (ins), "",
491 [(set VR256:$dst, (v8i32 immAllOnesV))]>;
494 //===----------------------------------------------------------------------===//
495 // SSE 1 & 2 - Move FP Scalar Instructions
497 // Move Instructions. Register-to-register movss/movsd is not used for FR32/64
498 // register copies because it's a partial register update; Register-to-register
499 // movss/movsd is not modeled as an INSERT_SUBREG because INSERT_SUBREG requires
500 // that the insert be implementable in terms of a copy, and just mentioned, we
501 // don't use movss/movsd for copies.
502 //===----------------------------------------------------------------------===//
504 multiclass sse12_move_rr<RegisterClass RC, SDNode OpNode, ValueType vt,
505 X86MemOperand x86memop, string base_opc,
506 string asm_opr, Domain d = GenericDomain> {
507 let isCommutable = 1 in
508 def rr : SI<0x10, MRMSrcReg, (outs VR128:$dst),
509 (ins VR128:$src1, RC:$src2),
510 !strconcat(base_opc, asm_opr),
511 [(set VR128:$dst, (vt (OpNode VR128:$src1,
512 (scalar_to_vector RC:$src2))))],
513 IIC_SSE_MOV_S_RR, d>, Sched<[WriteFShuffle]>;
515 // For the disassembler
516 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in
517 def rr_REV : SI<0x11, MRMDestReg, (outs VR128:$dst),
518 (ins VR128:$src1, RC:$src2),
519 !strconcat(base_opc, asm_opr),
520 [], IIC_SSE_MOV_S_RR>, Sched<[WriteFShuffle]>;
523 multiclass sse12_move<RegisterClass RC, SDNode OpNode, ValueType vt,
524 X86MemOperand x86memop, string OpcodeStr,
525 Domain d = GenericDomain> {
527 defm V#NAME : sse12_move_rr<RC, OpNode, vt, x86memop, OpcodeStr,
528 "\t{$src2, $src1, $dst|$dst, $src1, $src2}", d>,
529 VEX_4V, VEX_LIG, VEX_WIG;
531 def V#NAME#mr : SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, RC:$src),
532 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
533 [(store RC:$src, addr:$dst)], IIC_SSE_MOV_S_MR, d>,
534 VEX, VEX_LIG, Sched<[WriteStore]>, VEX_WIG;
536 let Constraints = "$src1 = $dst" in {
537 defm NAME : sse12_move_rr<RC, OpNode, vt, x86memop, OpcodeStr,
538 "\t{$src2, $dst|$dst, $src2}", d>;
541 def NAME#mr : SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, RC:$src),
542 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
543 [(store RC:$src, addr:$dst)], IIC_SSE_MOV_S_MR, d>,
547 // Loading from memory automatically zeroing upper bits.
548 multiclass sse12_move_rm<RegisterClass RC, X86MemOperand x86memop,
549 PatFrag mem_pat, string OpcodeStr,
550 Domain d = GenericDomain> {
551 def V#NAME#rm : SI<0x10, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
552 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
553 [(set RC:$dst, (mem_pat addr:$src))],
554 IIC_SSE_MOV_S_RM, d>, VEX, VEX_LIG, Sched<[WriteLoad]>, VEX_WIG;
555 def NAME#rm : SI<0x10, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
556 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
557 [(set RC:$dst, (mem_pat addr:$src))],
558 IIC_SSE_MOV_S_RM, d>, Sched<[WriteLoad]>;
561 defm MOVSS : sse12_move<FR32, X86Movss, v4f32, f32mem, "movss",
562 SSEPackedSingle>, XS;
563 defm MOVSD : sse12_move<FR64, X86Movsd, v2f64, f64mem, "movsd",
564 SSEPackedDouble>, XD;
566 let canFoldAsLoad = 1, isReMaterializable = 1 in {
567 defm MOVSS : sse12_move_rm<FR32, f32mem, loadf32, "movss",
568 SSEPackedSingle>, XS;
570 let AddedComplexity = 20 in
571 defm MOVSD : sse12_move_rm<FR64, f64mem, loadf64, "movsd",
572 SSEPackedDouble>, XD;
576 let Predicates = [UseAVX] in {
577 let AddedComplexity = 20 in {
578 // MOVSSrm zeros the high parts of the register; represent this
579 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
580 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
581 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
582 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
583 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
584 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
585 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
586 def : Pat<(v4f32 (X86vzload addr:$src)),
587 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
589 // MOVSDrm zeros the high parts of the register; represent this
590 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
591 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
592 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
593 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
594 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
595 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
596 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
597 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
598 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
599 def : Pat<(v2f64 (X86vzload addr:$src)),
600 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
602 // Represent the same patterns above but in the form they appear for
604 def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
605 (v4f32 (scalar_to_vector (loadf32 addr:$src))), (iPTR 0)))),
606 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_xmm)>;
607 def : Pat<(v8f32 (X86vzload addr:$src)),
608 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_xmm)>;
609 def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
610 (v2f64 (scalar_to_vector (loadf64 addr:$src))), (iPTR 0)))),
611 (SUBREG_TO_REG (i32 0), (VMOVSDrm addr:$src), sub_xmm)>;
612 def : Pat<(v4f64 (X86vzload addr:$src)),
613 (SUBREG_TO_REG (i32 0), (VMOVSDrm addr:$src), sub_xmm)>;
616 // Extract and store.
617 def : Pat<(store (f32 (extractelt (v4f32 VR128:$src), (iPTR 0))),
619 (VMOVSSmr addr:$dst, (COPY_TO_REGCLASS (v4f32 VR128:$src), FR32))>;
621 // Shuffle with VMOVSS
622 def : Pat<(v4i32 (X86Movss VR128:$src1, VR128:$src2)),
623 (VMOVSSrr (v4i32 VR128:$src1),
624 (COPY_TO_REGCLASS (v4i32 VR128:$src2), FR32))>;
625 def : Pat<(v4f32 (X86Movss VR128:$src1, VR128:$src2)),
626 (VMOVSSrr (v4f32 VR128:$src1),
627 (COPY_TO_REGCLASS (v4f32 VR128:$src2), FR32))>;
630 def : Pat<(v8i32 (X86Movss VR256:$src1, VR256:$src2)),
631 (SUBREG_TO_REG (i32 0),
632 (VMOVSSrr (EXTRACT_SUBREG (v8i32 VR256:$src1), sub_xmm),
633 (EXTRACT_SUBREG (v8i32 VR256:$src2), sub_xmm)),
635 def : Pat<(v8f32 (X86Movss VR256:$src1, VR256:$src2)),
636 (SUBREG_TO_REG (i32 0),
637 (VMOVSSrr (EXTRACT_SUBREG (v8f32 VR256:$src1), sub_xmm),
638 (EXTRACT_SUBREG (v8f32 VR256:$src2), sub_xmm)),
641 // Shuffle with VMOVSD
642 def : Pat<(v2i64 (X86Movsd VR128:$src1, VR128:$src2)),
643 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
644 def : Pat<(v2f64 (X86Movsd VR128:$src1, VR128:$src2)),
645 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
648 def : Pat<(v4i64 (X86Movsd VR256:$src1, VR256:$src2)),
649 (SUBREG_TO_REG (i32 0),
650 (VMOVSDrr (EXTRACT_SUBREG (v4i64 VR256:$src1), sub_xmm),
651 (EXTRACT_SUBREG (v4i64 VR256:$src2), sub_xmm)),
653 def : Pat<(v4f64 (X86Movsd VR256:$src1, VR256:$src2)),
654 (SUBREG_TO_REG (i32 0),
655 (VMOVSDrr (EXTRACT_SUBREG (v4f64 VR256:$src1), sub_xmm),
656 (EXTRACT_SUBREG (v4f64 VR256:$src2), sub_xmm)),
659 // FIXME: Instead of a X86Movlps there should be a X86Movsd here, the problem
660 // is during lowering, where it's not possible to recognize the fold cause
661 // it has two uses through a bitcast. One use disappears at isel time and the
662 // fold opportunity reappears.
663 def : Pat<(v2f64 (X86Movlpd VR128:$src1, VR128:$src2)),
664 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
665 def : Pat<(v2i64 (X86Movlpd VR128:$src1, VR128:$src2)),
666 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
667 def : Pat<(v4f32 (X86Movlps VR128:$src1, VR128:$src2)),
668 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
669 def : Pat<(v4i32 (X86Movlps VR128:$src1, VR128:$src2)),
670 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
673 let Predicates = [UseSSE1] in {
674 let Predicates = [NoSSE41], AddedComplexity = 15 in {
675 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
676 // MOVSS to the lower bits.
677 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
678 (MOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
679 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
680 (MOVSSrr (v4f32 (V_SET0)), (COPY_TO_REGCLASS VR128:$src, FR32))>;
681 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
682 (MOVSSrr (v4i32 (V_SET0)), (COPY_TO_REGCLASS VR128:$src, FR32))>;
685 let AddedComplexity = 20 in {
686 // MOVSSrm already zeros the high parts of the register.
687 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
688 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
689 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
690 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
691 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
692 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
693 def : Pat<(v4f32 (X86vzload addr:$src)),
694 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
697 // Extract and store.
698 def : Pat<(store (f32 (extractelt (v4f32 VR128:$src), (iPTR 0))),
700 (MOVSSmr addr:$dst, (COPY_TO_REGCLASS VR128:$src, FR32))>;
702 // Shuffle with MOVSS
703 def : Pat<(v4i32 (X86Movss VR128:$src1, VR128:$src2)),
704 (MOVSSrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR32))>;
705 def : Pat<(v4f32 (X86Movss VR128:$src1, VR128:$src2)),
706 (MOVSSrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR32))>;
709 let Predicates = [UseSSE2] in {
710 let Predicates = [NoSSE41], AddedComplexity = 15 in {
711 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
712 // MOVSD to the lower bits.
713 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
714 (MOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
717 let AddedComplexity = 20 in {
718 // MOVSDrm already zeros the high parts of the register.
719 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
720 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
721 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
722 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
723 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
724 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
725 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
726 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
727 def : Pat<(v2f64 (X86vzload addr:$src)),
728 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
731 // Shuffle with MOVSD
732 def : Pat<(v2i64 (X86Movsd VR128:$src1, VR128:$src2)),
733 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
734 def : Pat<(v2f64 (X86Movsd VR128:$src1, VR128:$src2)),
735 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
737 // FIXME: Instead of a X86Movlps there should be a X86Movsd here, the problem
738 // is during lowering, where it's not possible to recognize the fold because
739 // it has two uses through a bitcast. One use disappears at isel time and the
740 // fold opportunity reappears.
741 def : Pat<(v2f64 (X86Movlpd VR128:$src1, VR128:$src2)),
742 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
743 def : Pat<(v2i64 (X86Movlpd VR128:$src1, VR128:$src2)),
744 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
745 def : Pat<(v4f32 (X86Movlps VR128:$src1, VR128:$src2)),
746 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
747 def : Pat<(v4i32 (X86Movlps VR128:$src1, VR128:$src2)),
748 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
751 // Aliases to help the assembler pick two byte VEX encodings by swapping the
752 // operands relative to the normal instructions to use VEX.R instead of VEX.B.
753 def : InstAlias<"vmovss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
754 (VMOVSSrr_REV VR128L:$dst, VR128:$src1, VR128H:$src2), 0>;
755 def : InstAlias<"vmovsd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
756 (VMOVSDrr_REV VR128L:$dst, VR128:$src1, VR128H:$src2), 0>;
758 //===----------------------------------------------------------------------===//
759 // SSE 1 & 2 - Move Aligned/Unaligned FP Instructions
760 //===----------------------------------------------------------------------===//
762 multiclass sse12_mov_packed<bits<8> opc, RegisterClass RC,
763 X86MemOperand x86memop, PatFrag ld_frag,
764 string asm, Domain d,
766 let hasSideEffects = 0 in
767 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
768 !strconcat(asm, "\t{$src, $dst|$dst, $src}"), [], itins.rr, d>,
769 Sched<[WriteFShuffle]>;
770 let canFoldAsLoad = 1, isReMaterializable = 1 in
771 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
772 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
773 [(set RC:$dst, (ld_frag addr:$src))], itins.rm, d>,
777 let Predicates = [HasAVX, NoVLX] in {
778 defm VMOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
779 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
781 defm VMOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
782 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
784 defm VMOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
785 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
787 defm VMOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
788 "movupd", SSEPackedDouble, SSE_MOVU_ITINS>,
791 defm VMOVAPSY : sse12_mov_packed<0x28, VR256, f256mem, alignedloadv8f32,
792 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
793 PS, VEX, VEX_L, VEX_WIG;
794 defm VMOVAPDY : sse12_mov_packed<0x28, VR256, f256mem, alignedloadv4f64,
795 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
796 PD, VEX, VEX_L, VEX_WIG;
797 defm VMOVUPSY : sse12_mov_packed<0x10, VR256, f256mem, loadv8f32,
798 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
799 PS, VEX, VEX_L, VEX_WIG;
800 defm VMOVUPDY : sse12_mov_packed<0x10, VR256, f256mem, loadv4f64,
801 "movupd", SSEPackedDouble, SSE_MOVU_ITINS>,
802 PD, VEX, VEX_L, VEX_WIG;
805 let Predicates = [UseSSE1] in {
806 defm MOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
807 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
809 defm MOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
810 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
813 let Predicates = [UseSSE2] in {
814 defm MOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
815 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
817 defm MOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
818 "movupd", SSEPackedDouble, SSE_MOVU_ITINS>,
822 let SchedRW = [WriteStore], Predicates = [HasAVX, NoVLX] in {
823 def VMOVAPSmr : VPSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
824 "movaps\t{$src, $dst|$dst, $src}",
825 [(alignedstore (v4f32 VR128:$src), addr:$dst)],
826 IIC_SSE_MOVA_P_MR>, VEX, VEX_WIG;
827 def VMOVAPDmr : VPDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
828 "movapd\t{$src, $dst|$dst, $src}",
829 [(alignedstore (v2f64 VR128:$src), addr:$dst)],
830 IIC_SSE_MOVA_P_MR>, VEX, VEX_WIG;
831 def VMOVUPSmr : VPSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
832 "movups\t{$src, $dst|$dst, $src}",
833 [(store (v4f32 VR128:$src), addr:$dst)],
834 IIC_SSE_MOVU_P_MR>, VEX, VEX_WIG;
835 def VMOVUPDmr : VPDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
836 "movupd\t{$src, $dst|$dst, $src}",
837 [(store (v2f64 VR128:$src), addr:$dst)],
838 IIC_SSE_MOVU_P_MR>, VEX, VEX_WIG;
839 def VMOVAPSYmr : VPSI<0x29, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
840 "movaps\t{$src, $dst|$dst, $src}",
841 [(alignedstore256 (v8f32 VR256:$src), addr:$dst)],
842 IIC_SSE_MOVA_P_MR>, VEX, VEX_L, VEX_WIG;
843 def VMOVAPDYmr : VPDI<0x29, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
844 "movapd\t{$src, $dst|$dst, $src}",
845 [(alignedstore256 (v4f64 VR256:$src), addr:$dst)],
846 IIC_SSE_MOVA_P_MR>, VEX, VEX_L, VEX_WIG;
847 def VMOVUPSYmr : VPSI<0x11, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
848 "movups\t{$src, $dst|$dst, $src}",
849 [(store (v8f32 VR256:$src), addr:$dst)],
850 IIC_SSE_MOVU_P_MR>, VEX, VEX_L, VEX_WIG;
851 def VMOVUPDYmr : VPDI<0x11, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
852 "movupd\t{$src, $dst|$dst, $src}",
853 [(store (v4f64 VR256:$src), addr:$dst)],
854 IIC_SSE_MOVU_P_MR>, VEX, VEX_L, VEX_WIG;
858 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
859 SchedRW = [WriteFShuffle] in {
860 def VMOVAPSrr_REV : VPSI<0x29, MRMDestReg, (outs VR128:$dst),
862 "movaps\t{$src, $dst|$dst, $src}", [],
863 IIC_SSE_MOVA_P_RR>, VEX, VEX_WIG;
864 def VMOVAPDrr_REV : VPDI<0x29, MRMDestReg, (outs VR128:$dst),
866 "movapd\t{$src, $dst|$dst, $src}", [],
867 IIC_SSE_MOVA_P_RR>, VEX, VEX_WIG;
868 def VMOVUPSrr_REV : VPSI<0x11, MRMDestReg, (outs VR128:$dst),
870 "movups\t{$src, $dst|$dst, $src}", [],
871 IIC_SSE_MOVU_P_RR>, VEX, VEX_WIG;
872 def VMOVUPDrr_REV : VPDI<0x11, MRMDestReg, (outs VR128:$dst),
874 "movupd\t{$src, $dst|$dst, $src}", [],
875 IIC_SSE_MOVU_P_RR>, VEX, VEX_WIG;
876 def VMOVAPSYrr_REV : VPSI<0x29, MRMDestReg, (outs VR256:$dst),
878 "movaps\t{$src, $dst|$dst, $src}", [],
879 IIC_SSE_MOVA_P_RR>, VEX, VEX_L, VEX_WIG;
880 def VMOVAPDYrr_REV : VPDI<0x29, MRMDestReg, (outs VR256:$dst),
882 "movapd\t{$src, $dst|$dst, $src}", [],
883 IIC_SSE_MOVA_P_RR>, VEX, VEX_L, VEX_WIG;
884 def VMOVUPSYrr_REV : VPSI<0x11, MRMDestReg, (outs VR256:$dst),
886 "movups\t{$src, $dst|$dst, $src}", [],
887 IIC_SSE_MOVU_P_RR>, VEX, VEX_L, VEX_WIG;
888 def VMOVUPDYrr_REV : VPDI<0x11, MRMDestReg, (outs VR256:$dst),
890 "movupd\t{$src, $dst|$dst, $src}", [],
891 IIC_SSE_MOVU_P_RR>, VEX, VEX_L, VEX_WIG;
894 // Aliases to help the assembler pick two byte VEX encodings by swapping the
895 // operands relative to the normal instructions to use VEX.R instead of VEX.B.
896 def : InstAlias<"vmovaps\t{$src, $dst|$dst, $src}",
897 (VMOVAPSrr_REV VR128L:$dst, VR128H:$src), 0>;
898 def : InstAlias<"vmovapd\t{$src, $dst|$dst, $src}",
899 (VMOVAPDrr_REV VR128L:$dst, VR128H:$src), 0>;
900 def : InstAlias<"vmovups\t{$src, $dst|$dst, $src}",
901 (VMOVUPSrr_REV VR128L:$dst, VR128H:$src), 0>;
902 def : InstAlias<"vmovupd\t{$src, $dst|$dst, $src}",
903 (VMOVUPDrr_REV VR128L:$dst, VR128H:$src), 0>;
904 def : InstAlias<"vmovaps\t{$src, $dst|$dst, $src}",
905 (VMOVAPSYrr_REV VR256L:$dst, VR256H:$src), 0>;
906 def : InstAlias<"vmovapd\t{$src, $dst|$dst, $src}",
907 (VMOVAPDYrr_REV VR256L:$dst, VR256H:$src), 0>;
908 def : InstAlias<"vmovups\t{$src, $dst|$dst, $src}",
909 (VMOVUPSYrr_REV VR256L:$dst, VR256H:$src), 0>;
910 def : InstAlias<"vmovupd\t{$src, $dst|$dst, $src}",
911 (VMOVUPDYrr_REV VR256L:$dst, VR256H:$src), 0>;
913 let SchedRW = [WriteStore] in {
914 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
915 "movaps\t{$src, $dst|$dst, $src}",
916 [(alignedstore (v4f32 VR128:$src), addr:$dst)],
918 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
919 "movapd\t{$src, $dst|$dst, $src}",
920 [(alignedstore (v2f64 VR128:$src), addr:$dst)],
922 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
923 "movups\t{$src, $dst|$dst, $src}",
924 [(store (v4f32 VR128:$src), addr:$dst)],
926 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
927 "movupd\t{$src, $dst|$dst, $src}",
928 [(store (v2f64 VR128:$src), addr:$dst)],
933 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
934 SchedRW = [WriteFShuffle] in {
935 def MOVAPSrr_REV : PSI<0x29, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
936 "movaps\t{$src, $dst|$dst, $src}", [],
938 def MOVAPDrr_REV : PDI<0x29, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
939 "movapd\t{$src, $dst|$dst, $src}", [],
941 def MOVUPSrr_REV : PSI<0x11, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
942 "movups\t{$src, $dst|$dst, $src}", [],
944 def MOVUPDrr_REV : PDI<0x11, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
945 "movupd\t{$src, $dst|$dst, $src}", [],
949 let Predicates = [HasAVX, NoVLX] in {
950 // 256-bit load/store need to use floating point load/store in case we don't
951 // have AVX2. Execution domain fixing will convert to integer if AVX2 is
952 // available and changing the domain is beneficial.
953 def : Pat<(alignedloadv4i64 addr:$src),
954 (VMOVAPSYrm addr:$src)>;
955 def : Pat<(loadv4i64 addr:$src),
956 (VMOVUPSYrm addr:$src)>;
957 def : Pat<(alignedstore256 (v4i64 VR256:$src), addr:$dst),
958 (VMOVAPSYmr addr:$dst, VR256:$src)>;
959 def : Pat<(alignedstore256 (v8i32 VR256:$src), addr:$dst),
960 (VMOVAPSYmr addr:$dst, VR256:$src)>;
961 def : Pat<(alignedstore256 (v16i16 VR256:$src), addr:$dst),
962 (VMOVAPSYmr addr:$dst, VR256:$src)>;
963 def : Pat<(alignedstore256 (v32i8 VR256:$src), addr:$dst),
964 (VMOVAPSYmr addr:$dst, VR256:$src)>;
965 def : Pat<(store (v4i64 VR256:$src), addr:$dst),
966 (VMOVUPSYmr addr:$dst, VR256:$src)>;
967 def : Pat<(store (v8i32 VR256:$src), addr:$dst),
968 (VMOVUPSYmr addr:$dst, VR256:$src)>;
969 def : Pat<(store (v16i16 VR256:$src), addr:$dst),
970 (VMOVUPSYmr addr:$dst, VR256:$src)>;
971 def : Pat<(store (v32i8 VR256:$src), addr:$dst),
972 (VMOVUPSYmr addr:$dst, VR256:$src)>;
974 // Special patterns for storing subvector extracts of lower 128-bits
975 // Its cheaper to just use VMOVAPS/VMOVUPS instead of VEXTRACTF128mr
976 def : Pat<(alignedstore (v2f64 (extract_subvector
977 (v4f64 VR256:$src), (iPTR 0))), addr:$dst),
978 (VMOVAPDmr addr:$dst, (v2f64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
979 def : Pat<(alignedstore (v4f32 (extract_subvector
980 (v8f32 VR256:$src), (iPTR 0))), addr:$dst),
981 (VMOVAPSmr addr:$dst, (v4f32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
983 def : Pat<(store (v2f64 (extract_subvector
984 (v4f64 VR256:$src), (iPTR 0))), addr:$dst),
985 (VMOVUPDmr addr:$dst, (v2f64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
986 def : Pat<(store (v4f32 (extract_subvector
987 (v8f32 VR256:$src), (iPTR 0))), addr:$dst),
988 (VMOVUPSmr addr:$dst, (v4f32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
991 // Use movaps / movups for SSE integer load / store (one byte shorter).
992 // The instructions selected below are then converted to MOVDQA/MOVDQU
993 // during the SSE domain pass.
994 let Predicates = [UseSSE1] in {
995 def : Pat<(alignedloadv2i64 addr:$src),
996 (MOVAPSrm addr:$src)>;
997 def : Pat<(loadv2i64 addr:$src),
998 (MOVUPSrm addr:$src)>;
1000 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
1001 (MOVAPSmr addr:$dst, VR128:$src)>;
1002 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
1003 (MOVAPSmr addr:$dst, VR128:$src)>;
1004 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
1005 (MOVAPSmr addr:$dst, VR128:$src)>;
1006 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
1007 (MOVAPSmr addr:$dst, VR128:$src)>;
1008 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
1009 (MOVUPSmr addr:$dst, VR128:$src)>;
1010 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
1011 (MOVUPSmr addr:$dst, VR128:$src)>;
1012 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
1013 (MOVUPSmr addr:$dst, VR128:$src)>;
1014 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
1015 (MOVUPSmr addr:$dst, VR128:$src)>;
1018 //===----------------------------------------------------------------------===//
1019 // SSE 1 & 2 - Move Low packed FP Instructions
1020 //===----------------------------------------------------------------------===//
1022 multiclass sse12_mov_hilo_packed_base<bits<8>opc, SDNode psnode, SDNode pdnode,
1023 string base_opc, string asm_opr,
1024 InstrItinClass itin> {
1025 def PSrm : PI<opc, MRMSrcMem,
1026 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1027 !strconcat(base_opc, "s", asm_opr),
1029 (psnode VR128:$src1,
1030 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))],
1031 itin, SSEPackedSingle>, PS,
1032 Sched<[WriteFShuffleLd, ReadAfterLd]>;
1034 def PDrm : PI<opc, MRMSrcMem,
1035 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1036 !strconcat(base_opc, "d", asm_opr),
1037 [(set VR128:$dst, (v2f64 (pdnode VR128:$src1,
1038 (scalar_to_vector (loadf64 addr:$src2)))))],
1039 itin, SSEPackedDouble>, PD,
1040 Sched<[WriteFShuffleLd, ReadAfterLd]>;
1044 multiclass sse12_mov_hilo_packed<bits<8>opc, SDNode psnode, SDNode pdnode,
1045 string base_opc, InstrItinClass itin> {
1046 let Predicates = [UseAVX] in
1047 defm V#NAME : sse12_mov_hilo_packed_base<opc, psnode, pdnode, base_opc,
1048 "\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1049 itin>, VEX_4V, VEX_WIG;
1051 let Constraints = "$src1 = $dst" in
1052 defm NAME : sse12_mov_hilo_packed_base<opc, psnode, pdnode, base_opc,
1053 "\t{$src2, $dst|$dst, $src2}",
1057 let AddedComplexity = 20 in {
1058 defm MOVL : sse12_mov_hilo_packed<0x12, X86Movlps, X86Movlpd, "movlp",
1062 let SchedRW = [WriteStore] in {
1063 let Predicates = [UseAVX] in {
1064 def VMOVLPSmr : VPSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1065 "movlps\t{$src, $dst|$dst, $src}",
1066 [(store (f64 (extractelt (bc_v2f64 (v4f32 VR128:$src)),
1067 (iPTR 0))), addr:$dst)],
1068 IIC_SSE_MOV_LH>, VEX, VEX_WIG;
1069 def VMOVLPDmr : VPDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1070 "movlpd\t{$src, $dst|$dst, $src}",
1071 [(store (f64 (extractelt (v2f64 VR128:$src),
1072 (iPTR 0))), addr:$dst)],
1073 IIC_SSE_MOV_LH>, VEX, VEX_WIG;
1075 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1076 "movlps\t{$src, $dst|$dst, $src}",
1077 [(store (f64 (extractelt (bc_v2f64 (v4f32 VR128:$src)),
1078 (iPTR 0))), addr:$dst)],
1080 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1081 "movlpd\t{$src, $dst|$dst, $src}",
1082 [(store (f64 (extractelt (v2f64 VR128:$src),
1083 (iPTR 0))), addr:$dst)],
1087 let Predicates = [UseAVX] in {
1088 // Shuffle with VMOVLPS
1089 def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))),
1090 (VMOVLPSrm VR128:$src1, addr:$src2)>;
1091 def : Pat<(v4i32 (X86Movlps VR128:$src1, (load addr:$src2))),
1092 (VMOVLPSrm VR128:$src1, addr:$src2)>;
1094 // Shuffle with VMOVLPD
1095 def : Pat<(v2f64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1096 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1097 def : Pat<(v2i64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1098 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1099 def : Pat<(v2f64 (X86Movsd VR128:$src1,
1100 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))),
1101 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1104 def : Pat<(store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)),
1106 (VMOVLPSmr addr:$src1, VR128:$src2)>;
1107 def : Pat<(store (v4i32 (X86Movlps
1108 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)), addr:$src1),
1109 (VMOVLPSmr addr:$src1, VR128:$src2)>;
1110 def : Pat<(store (v2f64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1112 (VMOVLPDmr addr:$src1, VR128:$src2)>;
1113 def : Pat<(store (v2i64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1115 (VMOVLPDmr addr:$src1, VR128:$src2)>;
1118 let Predicates = [UseSSE1] in {
1119 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
1120 def : Pat<(store (i64 (extractelt (bc_v2i64 (v4f32 VR128:$src2)),
1121 (iPTR 0))), addr:$src1),
1122 (MOVLPSmr addr:$src1, VR128:$src2)>;
1124 // Shuffle with MOVLPS
1125 def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))),
1126 (MOVLPSrm VR128:$src1, addr:$src2)>;
1127 def : Pat<(v4i32 (X86Movlps VR128:$src1, (load addr:$src2))),
1128 (MOVLPSrm VR128:$src1, addr:$src2)>;
1129 def : Pat<(X86Movlps VR128:$src1,
1130 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1131 (MOVLPSrm VR128:$src1, addr:$src2)>;
1134 def : Pat<(store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)),
1136 (MOVLPSmr addr:$src1, VR128:$src2)>;
1137 def : Pat<(store (v4i32 (X86Movlps
1138 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
1140 (MOVLPSmr addr:$src1, VR128:$src2)>;
1143 let Predicates = [UseSSE2] in {
1144 // Shuffle with MOVLPD
1145 def : Pat<(v2f64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1146 (MOVLPDrm VR128:$src1, addr:$src2)>;
1147 def : Pat<(v2i64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1148 (MOVLPDrm VR128:$src1, addr:$src2)>;
1149 def : Pat<(v2f64 (X86Movsd VR128:$src1,
1150 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))),
1151 (MOVLPDrm VR128:$src1, addr:$src2)>;
1154 def : Pat<(store (v2f64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1156 (MOVLPDmr addr:$src1, VR128:$src2)>;
1157 def : Pat<(store (v2i64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1159 (MOVLPDmr addr:$src1, VR128:$src2)>;
1162 //===----------------------------------------------------------------------===//
1163 // SSE 1 & 2 - Move Hi packed FP Instructions
1164 //===----------------------------------------------------------------------===//
1166 let AddedComplexity = 20 in {
1167 defm MOVH : sse12_mov_hilo_packed<0x16, X86Movlhps, X86Movlhpd, "movhp",
1171 let SchedRW = [WriteStore] in {
1172 // v2f64 extract element 1 is always custom lowered to unpack high to low
1173 // and extract element 0 so the non-store version isn't too horrible.
1174 let Predicates = [UseAVX] in {
1175 def VMOVHPSmr : VPSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1176 "movhps\t{$src, $dst|$dst, $src}",
1177 [(store (f64 (extractelt
1178 (X86Unpckh (bc_v2f64 (v4f32 VR128:$src)),
1179 (bc_v2f64 (v4f32 VR128:$src))),
1180 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>, VEX, VEX_WIG;
1181 def VMOVHPDmr : VPDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1182 "movhpd\t{$src, $dst|$dst, $src}",
1183 [(store (f64 (extractelt
1184 (v2f64 (X86Unpckh VR128:$src, VR128:$src)),
1185 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>, VEX, VEX_WIG;
1187 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1188 "movhps\t{$src, $dst|$dst, $src}",
1189 [(store (f64 (extractelt
1190 (X86Unpckh (bc_v2f64 (v4f32 VR128:$src)),
1191 (bc_v2f64 (v4f32 VR128:$src))),
1192 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>;
1193 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1194 "movhpd\t{$src, $dst|$dst, $src}",
1195 [(store (f64 (extractelt
1196 (v2f64 (X86Unpckh VR128:$src, VR128:$src)),
1197 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>;
1200 let Predicates = [UseAVX] in {
1202 def : Pat<(X86Movlhps VR128:$src1,
1203 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1204 (VMOVHPSrm VR128:$src1, addr:$src2)>;
1205 def : Pat<(X86Movlhps VR128:$src1,
1206 (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
1207 (VMOVHPSrm VR128:$src1, addr:$src2)>;
1211 // FIXME: Instead of X86Unpckl, there should be a X86Movlhpd here, the problem
1212 // is during lowering, where it's not possible to recognize the load fold
1213 // cause it has two uses through a bitcast. One use disappears at isel time
1214 // and the fold opportunity reappears.
1215 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1216 (scalar_to_vector (loadf64 addr:$src2)))),
1217 (VMOVHPDrm VR128:$src1, addr:$src2)>;
1219 // Also handle an i64 load because that may get selected as a faster way to
1221 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1222 (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src2)))))),
1223 (VMOVHPDrm VR128:$src1, addr:$src2)>;
1225 def : Pat<(store (f64 (extractelt
1226 (bc_v2f64 (v4f32 (X86Movhlps VR128:$src, VR128:$src))),
1227 (iPTR 0))), addr:$dst),
1228 (VMOVHPDmr addr:$dst, VR128:$src)>;
1230 def : Pat<(store (f64 (extractelt
1231 (v2f64 (X86VPermilpi VR128:$src, (i8 1))),
1232 (iPTR 0))), addr:$dst),
1233 (VMOVHPDmr addr:$dst, VR128:$src)>;
1236 let Predicates = [UseSSE1] in {
1238 def : Pat<(X86Movlhps VR128:$src1,
1239 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1240 (MOVHPSrm VR128:$src1, addr:$src2)>;
1241 def : Pat<(X86Movlhps VR128:$src1,
1242 (bc_v4f32 (v2i64 (X86vzload addr:$src2)))),
1243 (MOVHPSrm VR128:$src1, addr:$src2)>;
1246 let Predicates = [UseSSE2] in {
1249 // FIXME: Instead of X86Unpckl, there should be a X86Movlhpd here, the problem
1250 // is during lowering, where it's not possible to recognize the load fold
1251 // cause it has two uses through a bitcast. One use disappears at isel time
1252 // and the fold opportunity reappears.
1253 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1254 (scalar_to_vector (loadf64 addr:$src2)))),
1255 (MOVHPDrm VR128:$src1, addr:$src2)>;
1257 // Also handle an i64 load because that may get selected as a faster way to
1259 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1260 (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src2)))))),
1261 (MOVHPDrm VR128:$src1, addr:$src2)>;
1263 def : Pat<(store (f64 (extractelt
1264 (bc_v2f64 (v4f32 (X86Movhlps VR128:$src, VR128:$src))),
1265 (iPTR 0))), addr:$dst),
1266 (MOVHPDmr addr:$dst, VR128:$src)>;
1268 def : Pat<(store (f64 (extractelt
1269 (v2f64 (X86Shufp VR128:$src, VR128:$src, (i8 1))),
1270 (iPTR 0))), addr:$dst),
1271 (MOVHPDmr addr:$dst, VR128:$src)>;
1274 //===----------------------------------------------------------------------===//
1275 // SSE 1 & 2 - Move Low to High and High to Low packed FP Instructions
1276 //===----------------------------------------------------------------------===//
1278 let AddedComplexity = 20, Predicates = [UseAVX] in {
1279 def VMOVLHPSrr : VPSI<0x16, MRMSrcReg, (outs VR128:$dst),
1280 (ins VR128:$src1, VR128:$src2),
1281 "movlhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1283 (v4f32 (X86Movlhps VR128:$src1, VR128:$src2)))],
1285 VEX_4V, Sched<[WriteFShuffle]>, VEX_WIG;
1286 def VMOVHLPSrr : VPSI<0x12, MRMSrcReg, (outs VR128:$dst),
1287 (ins VR128:$src1, VR128:$src2),
1288 "movhlps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1290 (v4f32 (X86Movhlps VR128:$src1, VR128:$src2)))],
1292 VEX_4V, Sched<[WriteFShuffle]>, VEX_WIG;
1294 let Constraints = "$src1 = $dst", AddedComplexity = 20 in {
1295 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
1296 (ins VR128:$src1, VR128:$src2),
1297 "movlhps\t{$src2, $dst|$dst, $src2}",
1299 (v4f32 (X86Movlhps VR128:$src1, VR128:$src2)))],
1300 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
1301 let isCommutable = 1 in
1302 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
1303 (ins VR128:$src1, VR128:$src2),
1304 "movhlps\t{$src2, $dst|$dst, $src2}",
1306 (v4f32 (X86Movhlps VR128:$src1, VR128:$src2)))],
1307 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
1310 let Predicates = [UseAVX] in {
1312 def : Pat<(v4i32 (X86Movlhps VR128:$src1, VR128:$src2)),
1313 (VMOVLHPSrr VR128:$src1, VR128:$src2)>;
1314 def : Pat<(v2i64 (X86Movlhps VR128:$src1, VR128:$src2)),
1315 (VMOVLHPSrr (v2i64 VR128:$src1), VR128:$src2)>;
1318 def : Pat<(v4i32 (X86Movhlps VR128:$src1, VR128:$src2)),
1319 (VMOVHLPSrr VR128:$src1, VR128:$src2)>;
1322 let Predicates = [UseSSE1] in {
1324 def : Pat<(v4i32 (X86Movlhps VR128:$src1, VR128:$src2)),
1325 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
1326 def : Pat<(v2i64 (X86Movlhps VR128:$src1, VR128:$src2)),
1327 (MOVLHPSrr (v2i64 VR128:$src1), VR128:$src2)>;
1330 def : Pat<(v4i32 (X86Movhlps VR128:$src1, VR128:$src2)),
1331 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
1334 //===----------------------------------------------------------------------===//
1335 // SSE 1 & 2 - Conversion Instructions
1336 //===----------------------------------------------------------------------===//
1338 def SSE_CVT_PD : OpndItins<
1339 IIC_SSE_CVT_PD_RR, IIC_SSE_CVT_PD_RM
1342 let Sched = WriteCvtI2F in
1343 def SSE_CVT_PS : OpndItins<
1344 IIC_SSE_CVT_PS_RR, IIC_SSE_CVT_PS_RM
1347 let Sched = WriteCvtI2F in
1348 def SSE_CVT_Scalar : OpndItins<
1349 IIC_SSE_CVT_Scalar_RR, IIC_SSE_CVT_Scalar_RM
1352 let Sched = WriteCvtF2I in
1353 def SSE_CVT_SS2SI_32 : OpndItins<
1354 IIC_SSE_CVT_SS2SI32_RR, IIC_SSE_CVT_SS2SI32_RM
1357 let Sched = WriteCvtF2I in
1358 def SSE_CVT_SS2SI_64 : OpndItins<
1359 IIC_SSE_CVT_SS2SI64_RR, IIC_SSE_CVT_SS2SI64_RM
1362 let Sched = WriteCvtF2I in
1363 def SSE_CVT_SD2SI : OpndItins<
1364 IIC_SSE_CVT_SD2SI_RR, IIC_SSE_CVT_SD2SI_RM
1367 // FIXME: We probably want to match the rm form only when optimizing for
1368 // size, to avoid false depenendecies (see sse_fp_unop_s for details)
1369 multiclass sse12_cvt_s<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1370 SDNode OpNode, X86MemOperand x86memop, PatFrag ld_frag,
1371 string asm, OpndItins itins> {
1372 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
1373 [(set DstRC:$dst, (OpNode SrcRC:$src))],
1374 itins.rr>, Sched<[itins.Sched]>;
1375 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
1376 [(set DstRC:$dst, (OpNode (ld_frag addr:$src)))],
1377 itins.rm>, Sched<[itins.Sched.Folded]>;
1380 multiclass sse12_cvt_p<bits<8> opc, RegisterClass RC, X86MemOperand x86memop,
1381 ValueType DstTy, ValueType SrcTy, PatFrag ld_frag,
1382 string asm, Domain d, OpndItins itins> {
1383 let hasSideEffects = 0 in {
1384 def rr : I<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src), asm,
1385 [(set RC:$dst, (DstTy (sint_to_fp (SrcTy RC:$src))))],
1386 itins.rr, d>, Sched<[itins.Sched]>;
1388 def rm : I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src), asm,
1389 [(set RC:$dst, (DstTy (sint_to_fp
1390 (SrcTy (bitconvert (ld_frag addr:$src))))))],
1391 itins.rm, d>, Sched<[itins.Sched.Folded]>;
1395 // FIXME: We probably want to match the rm form only when optimizing for
1396 // size, to avoid false depenendecies (see sse_fp_unop_s for details)
1397 multiclass sse12_vcvt_avx<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1398 X86MemOperand x86memop, string asm> {
1399 let hasSideEffects = 0, Predicates = [UseAVX] in {
1400 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src),
1401 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>,
1402 Sched<[WriteCvtI2F]>;
1404 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
1405 (ins DstRC:$src1, x86memop:$src),
1406 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>,
1407 Sched<[WriteCvtI2FLd, ReadAfterLd]>;
1408 } // hasSideEffects = 0
1411 let Predicates = [UseAVX] in {
1412 defm VCVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
1413 "cvttss2si\t{$src, $dst|$dst, $src}",
1416 defm VCVTTSS2SI64 : sse12_cvt_s<0x2C, FR32, GR64, fp_to_sint, f32mem, loadf32,
1417 "cvttss2si\t{$src, $dst|$dst, $src}",
1419 XS, VEX, VEX_W, VEX_LIG;
1420 defm VCVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
1421 "cvttsd2si\t{$src, $dst|$dst, $src}",
1424 defm VCVTTSD2SI64 : sse12_cvt_s<0x2C, FR64, GR64, fp_to_sint, f64mem, loadf64,
1425 "cvttsd2si\t{$src, $dst|$dst, $src}",
1427 XD, VEX, VEX_W, VEX_LIG;
1429 def : InstAlias<"vcvttss2si{l}\t{$src, $dst|$dst, $src}",
1430 (VCVTTSS2SIrr GR32:$dst, FR32:$src), 0>;
1431 def : InstAlias<"vcvttss2si{l}\t{$src, $dst|$dst, $src}",
1432 (VCVTTSS2SIrm GR32:$dst, f32mem:$src), 0>;
1433 def : InstAlias<"vcvttsd2si{l}\t{$src, $dst|$dst, $src}",
1434 (VCVTTSD2SIrr GR32:$dst, FR64:$src), 0>;
1435 def : InstAlias<"vcvttsd2si{l}\t{$src, $dst|$dst, $src}",
1436 (VCVTTSD2SIrm GR32:$dst, f64mem:$src), 0>;
1437 def : InstAlias<"vcvttss2si{q}\t{$src, $dst|$dst, $src}",
1438 (VCVTTSS2SI64rr GR64:$dst, FR32:$src), 0>;
1439 def : InstAlias<"vcvttss2si{q}\t{$src, $dst|$dst, $src}",
1440 (VCVTTSS2SI64rm GR64:$dst, f32mem:$src), 0>;
1441 def : InstAlias<"vcvttsd2si{q}\t{$src, $dst|$dst, $src}",
1442 (VCVTTSD2SI64rr GR64:$dst, FR64:$src), 0>;
1443 def : InstAlias<"vcvttsd2si{q}\t{$src, $dst|$dst, $src}",
1444 (VCVTTSD2SI64rm GR64:$dst, f64mem:$src), 0>;
1446 // The assembler can recognize rr 64-bit instructions by seeing a rxx
1447 // register, but the same isn't true when only using memory operands,
1448 // provide other assembly "l" and "q" forms to address this explicitly
1449 // where appropriate to do so.
1450 defm VCVTSI2SS : sse12_vcvt_avx<0x2A, GR32, FR32, i32mem, "cvtsi2ss{l}">,
1451 XS, VEX_4V, VEX_LIG;
1452 defm VCVTSI2SS64 : sse12_vcvt_avx<0x2A, GR64, FR32, i64mem, "cvtsi2ss{q}">,
1453 XS, VEX_4V, VEX_W, VEX_LIG;
1454 defm VCVTSI2SD : sse12_vcvt_avx<0x2A, GR32, FR64, i32mem, "cvtsi2sd{l}">,
1455 XD, VEX_4V, VEX_LIG;
1456 defm VCVTSI2SD64 : sse12_vcvt_avx<0x2A, GR64, FR64, i64mem, "cvtsi2sd{q}">,
1457 XD, VEX_4V, VEX_W, VEX_LIG;
1459 let Predicates = [UseAVX] in {
1460 def : InstAlias<"vcvtsi2ss\t{$src, $src1, $dst|$dst, $src1, $src}",
1461 (VCVTSI2SSrm FR64:$dst, FR64:$src1, i32mem:$src), 0>;
1462 def : InstAlias<"vcvtsi2sd\t{$src, $src1, $dst|$dst, $src1, $src}",
1463 (VCVTSI2SDrm FR64:$dst, FR64:$src1, i32mem:$src), 0>;
1465 def : Pat<(f32 (sint_to_fp (loadi32 addr:$src))),
1466 (VCVTSI2SSrm (f32 (IMPLICIT_DEF)), addr:$src)>;
1467 def : Pat<(f32 (sint_to_fp (loadi64 addr:$src))),
1468 (VCVTSI2SS64rm (f32 (IMPLICIT_DEF)), addr:$src)>;
1469 def : Pat<(f64 (sint_to_fp (loadi32 addr:$src))),
1470 (VCVTSI2SDrm (f64 (IMPLICIT_DEF)), addr:$src)>;
1471 def : Pat<(f64 (sint_to_fp (loadi64 addr:$src))),
1472 (VCVTSI2SD64rm (f64 (IMPLICIT_DEF)), addr:$src)>;
1474 def : Pat<(f32 (sint_to_fp GR32:$src)),
1475 (VCVTSI2SSrr (f32 (IMPLICIT_DEF)), GR32:$src)>;
1476 def : Pat<(f32 (sint_to_fp GR64:$src)),
1477 (VCVTSI2SS64rr (f32 (IMPLICIT_DEF)), GR64:$src)>;
1478 def : Pat<(f64 (sint_to_fp GR32:$src)),
1479 (VCVTSI2SDrr (f64 (IMPLICIT_DEF)), GR32:$src)>;
1480 def : Pat<(f64 (sint_to_fp GR64:$src)),
1481 (VCVTSI2SD64rr (f64 (IMPLICIT_DEF)), GR64:$src)>;
1484 defm CVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
1485 "cvttss2si\t{$src, $dst|$dst, $src}",
1486 SSE_CVT_SS2SI_32>, XS;
1487 defm CVTTSS2SI64 : sse12_cvt_s<0x2C, FR32, GR64, fp_to_sint, f32mem, loadf32,
1488 "cvttss2si\t{$src, $dst|$dst, $src}",
1489 SSE_CVT_SS2SI_64>, XS, REX_W;
1490 defm CVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
1491 "cvttsd2si\t{$src, $dst|$dst, $src}",
1493 defm CVTTSD2SI64 : sse12_cvt_s<0x2C, FR64, GR64, fp_to_sint, f64mem, loadf64,
1494 "cvttsd2si\t{$src, $dst|$dst, $src}",
1495 SSE_CVT_SD2SI>, XD, REX_W;
1496 defm CVTSI2SS : sse12_cvt_s<0x2A, GR32, FR32, sint_to_fp, i32mem, loadi32,
1497 "cvtsi2ss{l}\t{$src, $dst|$dst, $src}",
1498 SSE_CVT_Scalar>, XS;
1499 defm CVTSI2SS64 : sse12_cvt_s<0x2A, GR64, FR32, sint_to_fp, i64mem, loadi64,
1500 "cvtsi2ss{q}\t{$src, $dst|$dst, $src}",
1501 SSE_CVT_Scalar>, XS, REX_W;
1502 defm CVTSI2SD : sse12_cvt_s<0x2A, GR32, FR64, sint_to_fp, i32mem, loadi32,
1503 "cvtsi2sd{l}\t{$src, $dst|$dst, $src}",
1504 SSE_CVT_Scalar>, XD;
1505 defm CVTSI2SD64 : sse12_cvt_s<0x2A, GR64, FR64, sint_to_fp, i64mem, loadi64,
1506 "cvtsi2sd{q}\t{$src, $dst|$dst, $src}",
1507 SSE_CVT_Scalar>, XD, REX_W;
1509 def : InstAlias<"cvttss2si{l}\t{$src, $dst|$dst, $src}",
1510 (CVTTSS2SIrr GR32:$dst, FR32:$src), 0>;
1511 def : InstAlias<"cvttss2si{l}\t{$src, $dst|$dst, $src}",
1512 (CVTTSS2SIrm GR32:$dst, f32mem:$src), 0>;
1513 def : InstAlias<"cvttsd2si{l}\t{$src, $dst|$dst, $src}",
1514 (CVTTSD2SIrr GR32:$dst, FR64:$src), 0>;
1515 def : InstAlias<"cvttsd2si{l}\t{$src, $dst|$dst, $src}",
1516 (CVTTSD2SIrm GR32:$dst, f64mem:$src), 0>;
1517 def : InstAlias<"cvttss2si{q}\t{$src, $dst|$dst, $src}",
1518 (CVTTSS2SI64rr GR64:$dst, FR32:$src), 0>;
1519 def : InstAlias<"cvttss2si{q}\t{$src, $dst|$dst, $src}",
1520 (CVTTSS2SI64rm GR64:$dst, f32mem:$src), 0>;
1521 def : InstAlias<"cvttsd2si{q}\t{$src, $dst|$dst, $src}",
1522 (CVTTSD2SI64rr GR64:$dst, FR64:$src), 0>;
1523 def : InstAlias<"cvttsd2si{q}\t{$src, $dst|$dst, $src}",
1524 (CVTTSD2SI64rm GR64:$dst, f64mem:$src), 0>;
1526 def : InstAlias<"cvtsi2ss\t{$src, $dst|$dst, $src}",
1527 (CVTSI2SSrm FR64:$dst, i32mem:$src), 0>;
1528 def : InstAlias<"cvtsi2sd\t{$src, $dst|$dst, $src}",
1529 (CVTSI2SDrm FR64:$dst, i32mem:$src), 0>;
1531 // Conversion Instructions Intrinsics - Match intrinsics which expect MM
1532 // and/or XMM operand(s).
1534 // FIXME: We probably want to match the rm form only when optimizing for
1535 // size, to avoid false depenendecies (see sse_fp_unop_s for details)
1536 multiclass sse12_cvt_sint<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1537 Intrinsic Int, Operand memop, ComplexPattern mem_cpat,
1538 string asm, OpndItins itins> {
1539 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src),
1540 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1541 [(set DstRC:$dst, (Int SrcRC:$src))], itins.rr>,
1542 Sched<[itins.Sched]>;
1543 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins memop:$src),
1544 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1545 [(set DstRC:$dst, (Int mem_cpat:$src))], itins.rm>,
1546 Sched<[itins.Sched.Folded]>;
1549 multiclass sse12_cvt_sint_3addr<bits<8> opc, RegisterClass SrcRC,
1550 RegisterClass DstRC, Intrinsic Int, X86MemOperand x86memop,
1551 PatFrag ld_frag, string asm, OpndItins itins,
1553 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src2),
1555 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
1556 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
1557 [(set DstRC:$dst, (Int DstRC:$src1, SrcRC:$src2))],
1558 itins.rr>, Sched<[itins.Sched]>;
1559 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
1560 (ins DstRC:$src1, x86memop:$src2),
1562 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
1563 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
1564 [(set DstRC:$dst, (Int DstRC:$src1, (ld_frag addr:$src2)))],
1565 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
1568 let Predicates = [UseAVX] in {
1569 defm VCVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32,
1570 int_x86_sse2_cvtsd2si, sdmem, sse_load_f64, "cvtsd2si",
1571 SSE_CVT_SD2SI>, XD, VEX, VEX_LIG;
1572 defm VCVTSD2SI64 : sse12_cvt_sint<0x2D, VR128, GR64,
1573 int_x86_sse2_cvtsd2si64, sdmem, sse_load_f64, "cvtsd2si",
1574 SSE_CVT_SD2SI>, XD, VEX, VEX_W, VEX_LIG;
1576 defm CVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse2_cvtsd2si,
1577 sdmem, sse_load_f64, "cvtsd2si", SSE_CVT_SD2SI>, XD;
1578 defm CVTSD2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse2_cvtsd2si64,
1579 sdmem, sse_load_f64, "cvtsd2si", SSE_CVT_SD2SI>, XD, REX_W;
1582 let isCodeGenOnly = 1 in {
1583 let Predicates = [UseAVX] in {
1584 defm Int_VCVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1585 int_x86_sse_cvtsi2ss, i32mem, loadi32, "cvtsi2ss{l}",
1586 SSE_CVT_Scalar, 0>, XS, VEX_4V;
1587 defm Int_VCVTSI2SS64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1588 int_x86_sse_cvtsi642ss, i64mem, loadi64, "cvtsi2ss{q}",
1589 SSE_CVT_Scalar, 0>, XS, VEX_4V,
1591 defm Int_VCVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1592 int_x86_sse2_cvtsi2sd, i32mem, loadi32, "cvtsi2sd{l}",
1593 SSE_CVT_Scalar, 0>, XD, VEX_4V;
1594 defm Int_VCVTSI2SD64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1595 int_x86_sse2_cvtsi642sd, i64mem, loadi64, "cvtsi2sd{q}",
1596 SSE_CVT_Scalar, 0>, XD,
1599 let Constraints = "$src1 = $dst" in {
1600 defm Int_CVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1601 int_x86_sse_cvtsi2ss, i32mem, loadi32,
1602 "cvtsi2ss{l}", SSE_CVT_Scalar>, XS;
1603 defm Int_CVTSI2SS64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1604 int_x86_sse_cvtsi642ss, i64mem, loadi64,
1605 "cvtsi2ss{q}", SSE_CVT_Scalar>, XS, REX_W;
1606 defm Int_CVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1607 int_x86_sse2_cvtsi2sd, i32mem, loadi32,
1608 "cvtsi2sd{l}", SSE_CVT_Scalar>, XD;
1609 defm Int_CVTSI2SD64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1610 int_x86_sse2_cvtsi642sd, i64mem, loadi64,
1611 "cvtsi2sd{q}", SSE_CVT_Scalar>, XD, REX_W;
1613 } // isCodeGenOnly = 1
1617 // Aliases for intrinsics
1618 let isCodeGenOnly = 1 in {
1619 let Predicates = [UseAVX] in {
1620 defm Int_VCVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
1621 ssmem, sse_load_f32, "cvttss2si",
1622 SSE_CVT_SS2SI_32>, XS, VEX;
1623 defm Int_VCVTTSS2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1624 int_x86_sse_cvttss2si64, ssmem, sse_load_f32,
1625 "cvttss2si", SSE_CVT_SS2SI_64>,
1627 defm Int_VCVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
1628 sdmem, sse_load_f64, "cvttsd2si",
1629 SSE_CVT_SD2SI>, XD, VEX;
1630 defm Int_VCVTTSD2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1631 int_x86_sse2_cvttsd2si64, sdmem, sse_load_f64,
1632 "cvttsd2si", SSE_CVT_SD2SI>,
1635 defm Int_CVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
1636 ssmem, sse_load_f32, "cvttss2si",
1637 SSE_CVT_SS2SI_32>, XS;
1638 defm Int_CVTTSS2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1639 int_x86_sse_cvttss2si64, ssmem, sse_load_f32,
1640 "cvttss2si", SSE_CVT_SS2SI_64>, XS, REX_W;
1641 defm Int_CVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
1642 sdmem, sse_load_f64, "cvttsd2si",
1644 defm Int_CVTTSD2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1645 int_x86_sse2_cvttsd2si64, sdmem, sse_load_f64,
1646 "cvttsd2si", SSE_CVT_SD2SI>, XD, REX_W;
1647 } // isCodeGenOnly = 1
1649 let Predicates = [UseAVX] in {
1650 defm VCVTSS2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse_cvtss2si,
1651 ssmem, sse_load_f32, "cvtss2si",
1652 SSE_CVT_SS2SI_32>, XS, VEX, VEX_LIG;
1653 defm VCVTSS2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse_cvtss2si64,
1654 ssmem, sse_load_f32, "cvtss2si",
1655 SSE_CVT_SS2SI_64>, XS, VEX, VEX_W, VEX_LIG;
1657 defm CVTSS2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse_cvtss2si,
1658 ssmem, sse_load_f32, "cvtss2si",
1659 SSE_CVT_SS2SI_32>, XS;
1660 defm CVTSS2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse_cvtss2si64,
1661 ssmem, sse_load_f32, "cvtss2si",
1662 SSE_CVT_SS2SI_64>, XS, REX_W;
1664 defm VCVTDQ2PS : sse12_cvt_p<0x5B, VR128, i128mem, v4f32, v4i32, loadv2i64,
1665 "vcvtdq2ps\t{$src, $dst|$dst, $src}",
1666 SSEPackedSingle, SSE_CVT_PS>,
1667 PS, VEX, Requires<[HasAVX, NoVLX]>, VEX_WIG;
1668 defm VCVTDQ2PSY : sse12_cvt_p<0x5B, VR256, i256mem, v8f32, v8i32, loadv4i64,
1669 "vcvtdq2ps\t{$src, $dst|$dst, $src}",
1670 SSEPackedSingle, SSE_CVT_PS>,
1671 PS, VEX, VEX_L, Requires<[HasAVX, NoVLX]>, VEX_WIG;
1673 defm CVTDQ2PS : sse12_cvt_p<0x5B, VR128, i128mem, v4f32, v4i32, memopv2i64,
1674 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1675 SSEPackedSingle, SSE_CVT_PS>,
1676 PS, Requires<[UseSSE2]>;
1678 let Predicates = [UseAVX] in {
1679 def : InstAlias<"vcvtss2si{l}\t{$src, $dst|$dst, $src}",
1680 (VCVTSS2SIrr GR32:$dst, VR128:$src), 0>;
1681 def : InstAlias<"vcvtss2si{l}\t{$src, $dst|$dst, $src}",
1682 (VCVTSS2SIrm GR32:$dst, ssmem:$src), 0>;
1683 def : InstAlias<"vcvtsd2si{l}\t{$src, $dst|$dst, $src}",
1684 (VCVTSD2SIrr GR32:$dst, VR128:$src), 0>;
1685 def : InstAlias<"vcvtsd2si{l}\t{$src, $dst|$dst, $src}",
1686 (VCVTSD2SIrm GR32:$dst, sdmem:$src), 0>;
1687 def : InstAlias<"vcvtss2si{q}\t{$src, $dst|$dst, $src}",
1688 (VCVTSS2SI64rr GR64:$dst, VR128:$src), 0>;
1689 def : InstAlias<"vcvtss2si{q}\t{$src, $dst|$dst, $src}",
1690 (VCVTSS2SI64rm GR64:$dst, ssmem:$src), 0>;
1691 def : InstAlias<"vcvtsd2si{q}\t{$src, $dst|$dst, $src}",
1692 (VCVTSD2SI64rr GR64:$dst, VR128:$src), 0>;
1693 def : InstAlias<"vcvtsd2si{q}\t{$src, $dst|$dst, $src}",
1694 (VCVTSD2SI64rm GR64:$dst, sdmem:$src), 0>;
1697 def : InstAlias<"cvtss2si{l}\t{$src, $dst|$dst, $src}",
1698 (CVTSS2SIrr GR32:$dst, VR128:$src), 0>;
1699 def : InstAlias<"cvtss2si{l}\t{$src, $dst|$dst, $src}",
1700 (CVTSS2SIrm GR32:$dst, ssmem:$src), 0>;
1701 def : InstAlias<"cvtsd2si{l}\t{$src, $dst|$dst, $src}",
1702 (CVTSD2SIrr GR32:$dst, VR128:$src), 0>;
1703 def : InstAlias<"cvtsd2si{l}\t{$src, $dst|$dst, $src}",
1704 (CVTSD2SIrm GR32:$dst, sdmem:$src), 0>;
1705 def : InstAlias<"cvtss2si{q}\t{$src, $dst|$dst, $src}",
1706 (CVTSS2SI64rr GR64:$dst, VR128:$src), 0>;
1707 def : InstAlias<"cvtss2si{q}\t{$src, $dst|$dst, $src}",
1708 (CVTSS2SI64rm GR64:$dst, ssmem:$src), 0>;
1709 def : InstAlias<"cvtsd2si{q}\t{$src, $dst|$dst, $src}",
1710 (CVTSD2SI64rr GR64:$dst, VR128:$src), 0>;
1711 def : InstAlias<"cvtsd2si{q}\t{$src, $dst|$dst, $src}",
1712 (CVTSD2SI64rm GR64:$dst, sdmem:$src), 0>;
1716 // Convert scalar double to scalar single
1717 let hasSideEffects = 0, Predicates = [UseAVX] in {
1718 def VCVTSD2SSrr : VSDI<0x5A, MRMSrcReg, (outs FR32:$dst),
1719 (ins FR32:$src1, FR64:$src2),
1720 "cvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}", [],
1721 IIC_SSE_CVT_Scalar_RR>, VEX_4V, VEX_LIG,
1722 Sched<[WriteCvtF2F]>, VEX_WIG;
1724 def VCVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst),
1725 (ins FR32:$src1, f64mem:$src2),
1726 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1727 [], IIC_SSE_CVT_Scalar_RM>,
1728 XD, Requires<[HasAVX, OptForSize]>, VEX_4V, VEX_LIG,
1729 Sched<[WriteCvtF2FLd, ReadAfterLd]>, VEX_WIG;
1732 def : Pat<(f32 (fpround FR64:$src)),
1733 (VCVTSD2SSrr (COPY_TO_REGCLASS FR64:$src, FR32), FR64:$src)>,
1736 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1737 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1738 [(set FR32:$dst, (fpround FR64:$src))],
1739 IIC_SSE_CVT_Scalar_RR>, Sched<[WriteCvtF2F]>;
1740 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1741 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1742 [(set FR32:$dst, (fpround (loadf64 addr:$src)))],
1743 IIC_SSE_CVT_Scalar_RM>,
1745 Requires<[UseSSE2, OptForSize]>, Sched<[WriteCvtF2FLd]>;
1747 let isCodeGenOnly = 1 in {
1748 def Int_VCVTSD2SSrr: I<0x5A, MRMSrcReg,
1749 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1750 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1752 (int_x86_sse2_cvtsd2ss VR128:$src1, VR128:$src2))],
1753 IIC_SSE_CVT_Scalar_RR>, XD, VEX_4V, VEX_WIG,
1754 Requires<[HasAVX]>, Sched<[WriteCvtF2F]>;
1755 def Int_VCVTSD2SSrm: I<0x5A, MRMSrcMem,
1756 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1757 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1758 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss
1759 VR128:$src1, sse_load_f64:$src2))],
1760 IIC_SSE_CVT_Scalar_RM>, XD, VEX_4V, VEX_WIG,
1761 Requires<[HasAVX]>, Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1763 let Constraints = "$src1 = $dst" in {
1764 def Int_CVTSD2SSrr: I<0x5A, MRMSrcReg,
1765 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1766 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1768 (int_x86_sse2_cvtsd2ss VR128:$src1, VR128:$src2))],
1769 IIC_SSE_CVT_Scalar_RR>, XD, Requires<[UseSSE2]>,
1770 Sched<[WriteCvtF2F]>;
1771 def Int_CVTSD2SSrm: I<0x5A, MRMSrcMem,
1772 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1773 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1774 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss
1775 VR128:$src1, sse_load_f64:$src2))],
1776 IIC_SSE_CVT_Scalar_RM>, XD, Requires<[UseSSE2]>,
1777 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1779 } // isCodeGenOnly = 1
1781 // Convert scalar single to scalar double
1782 // SSE2 instructions with XS prefix
1783 let hasSideEffects = 0, Predicates = [UseAVX] in {
1784 def VCVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst),
1785 (ins FR64:$src1, FR32:$src2),
1786 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1787 [], IIC_SSE_CVT_Scalar_RR>,
1788 XS, Requires<[HasAVX]>, VEX_4V, VEX_LIG,
1789 Sched<[WriteCvtF2F]>, VEX_WIG;
1791 def VCVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst),
1792 (ins FR64:$src1, f32mem:$src2),
1793 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1794 [], IIC_SSE_CVT_Scalar_RM>,
1795 XS, VEX_4V, VEX_LIG, Requires<[HasAVX, OptForSize]>,
1796 Sched<[WriteCvtF2FLd, ReadAfterLd]>, VEX_WIG;
1799 def : Pat<(f64 (fpextend FR32:$src)),
1800 (VCVTSS2SDrr (COPY_TO_REGCLASS FR32:$src, FR64), FR32:$src)>, Requires<[UseAVX]>;
1801 def : Pat<(fpextend (loadf32 addr:$src)),
1802 (VCVTSS2SDrm (f64 (IMPLICIT_DEF)), addr:$src)>, Requires<[UseAVX]>;
1804 def : Pat<(extloadf32 addr:$src),
1805 (VCVTSS2SDrm (f64 (IMPLICIT_DEF)), addr:$src)>,
1806 Requires<[UseAVX, OptForSize]>;
1807 def : Pat<(extloadf32 addr:$src),
1808 (VCVTSS2SDrr (f64 (IMPLICIT_DEF)), (VMOVSSrm addr:$src))>,
1809 Requires<[UseAVX, OptForSpeed]>;
1811 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1812 "cvtss2sd\t{$src, $dst|$dst, $src}",
1813 [(set FR64:$dst, (fpextend FR32:$src))],
1814 IIC_SSE_CVT_Scalar_RR>, XS,
1815 Requires<[UseSSE2]>, Sched<[WriteCvtF2F]>;
1816 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1817 "cvtss2sd\t{$src, $dst|$dst, $src}",
1818 [(set FR64:$dst, (extloadf32 addr:$src))],
1819 IIC_SSE_CVT_Scalar_RM>, XS,
1820 Requires<[UseSSE2, OptForSize]>, Sched<[WriteCvtF2FLd]>;
1822 // extload f32 -> f64. This matches load+fpextend because we have a hack in
1823 // the isel (PreprocessForFPConvert) that can introduce loads after dag
1825 // Since these loads aren't folded into the fpextend, we have to match it
1827 def : Pat<(fpextend (loadf32 addr:$src)),
1828 (CVTSS2SDrm addr:$src)>, Requires<[UseSSE2]>;
1829 def : Pat<(extloadf32 addr:$src),
1830 (CVTSS2SDrr (MOVSSrm addr:$src))>, Requires<[UseSSE2, OptForSpeed]>;
1832 let isCodeGenOnly = 1 in {
1833 def Int_VCVTSS2SDrr: I<0x5A, MRMSrcReg,
1834 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1835 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1837 (int_x86_sse2_cvtss2sd VR128:$src1, VR128:$src2))],
1838 IIC_SSE_CVT_Scalar_RR>, XS, VEX_4V, VEX_WIG,
1839 Requires<[HasAVX]>, Sched<[WriteCvtF2F]>;
1840 def Int_VCVTSS2SDrm: I<0x5A, MRMSrcMem,
1841 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
1842 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1844 (int_x86_sse2_cvtss2sd VR128:$src1, sse_load_f32:$src2))],
1845 IIC_SSE_CVT_Scalar_RM>, XS, VEX_4V, VEX_WIG,
1846 Requires<[HasAVX]>, Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1847 let Constraints = "$src1 = $dst" in { // SSE2 instructions with XS prefix
1848 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1849 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1850 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1852 (int_x86_sse2_cvtss2sd VR128:$src1, VR128:$src2))],
1853 IIC_SSE_CVT_Scalar_RR>, XS, Requires<[UseSSE2]>,
1854 Sched<[WriteCvtF2F]>;
1855 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1856 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
1857 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1859 (int_x86_sse2_cvtss2sd VR128:$src1, sse_load_f32:$src2))],
1860 IIC_SSE_CVT_Scalar_RM>, XS, Requires<[UseSSE2]>,
1861 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1863 } // isCodeGenOnly = 1
1865 // Patterns used for matching (v)cvtsi2ss, (v)cvtsi2sd, (v)cvtsd2ss and
1866 // (v)cvtss2sd intrinsic sequences from clang which produce unnecessary
1867 // vmovs{s,d} instructions
1868 let Predicates = [UseAVX] in {
1869 def : Pat<(v4f32 (X86Movss
1871 (v4f32 (scalar_to_vector
1872 (f32 (fpround (f64 (extractelt VR128:$src, (iPTR 0))))))))),
1873 (Int_VCVTSD2SSrr VR128:$dst, VR128:$src)>;
1875 def : Pat<(v2f64 (X86Movsd
1877 (v2f64 (scalar_to_vector
1878 (f64 (fpextend (f32 (extractelt VR128:$src, (iPTR 0))))))))),
1879 (Int_VCVTSS2SDrr VR128:$dst, VR128:$src)>;
1881 def : Pat<(v4f32 (X86Movss
1883 (v4f32 (scalar_to_vector (f32 (sint_to_fp GR64:$src)))))),
1884 (Int_VCVTSI2SS64rr VR128:$dst, GR64:$src)>;
1886 def : Pat<(v4f32 (X86Movss
1888 (v4f32 (scalar_to_vector (f32 (sint_to_fp GR32:$src)))))),
1889 (Int_VCVTSI2SSrr VR128:$dst, GR32:$src)>;
1891 def : Pat<(v2f64 (X86Movsd
1893 (v2f64 (scalar_to_vector (f64 (sint_to_fp GR64:$src)))))),
1894 (Int_VCVTSI2SD64rr VR128:$dst, GR64:$src)>;
1896 def : Pat<(v2f64 (X86Movsd
1898 (v2f64 (scalar_to_vector (f64 (sint_to_fp GR32:$src)))))),
1899 (Int_VCVTSI2SDrr VR128:$dst, GR32:$src)>;
1900 } // Predicates = [UseAVX]
1902 let Predicates = [UseSSE2] in {
1903 def : Pat<(v4f32 (X86Movss
1905 (v4f32 (scalar_to_vector
1906 (f32 (fpround (f64 (extractelt VR128:$src, (iPTR 0))))))))),
1907 (Int_CVTSD2SSrr VR128:$dst, VR128:$src)>;
1909 def : Pat<(v2f64 (X86Movsd
1911 (v2f64 (scalar_to_vector
1912 (f64 (fpextend (f32 (extractelt VR128:$src, (iPTR 0))))))))),
1913 (Int_CVTSS2SDrr VR128:$dst, VR128:$src)>;
1915 def : Pat<(v2f64 (X86Movsd
1917 (v2f64 (scalar_to_vector (f64 (sint_to_fp GR64:$src)))))),
1918 (Int_CVTSI2SD64rr VR128:$dst, GR64:$src)>;
1920 def : Pat<(v2f64 (X86Movsd
1922 (v2f64 (scalar_to_vector (f64 (sint_to_fp GR32:$src)))))),
1923 (Int_CVTSI2SDrr VR128:$dst, GR32:$src)>;
1924 } // Predicates = [UseSSE2]
1926 let Predicates = [UseSSE1] in {
1927 def : Pat<(v4f32 (X86Movss
1929 (v4f32 (scalar_to_vector (f32 (sint_to_fp GR64:$src)))))),
1930 (Int_CVTSI2SS64rr VR128:$dst, GR64:$src)>;
1932 def : Pat<(v4f32 (X86Movss
1934 (v4f32 (scalar_to_vector (f32 (sint_to_fp GR32:$src)))))),
1935 (Int_CVTSI2SSrr VR128:$dst, GR32:$src)>;
1936 } // Predicates = [UseSSE1]
1938 // Convert packed single/double fp to doubleword
1939 def VCVTPS2DQrr : VPDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1940 "cvtps2dq\t{$src, $dst|$dst, $src}",
1941 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))],
1942 IIC_SSE_CVT_PS_RR>, VEX, Sched<[WriteCvtF2I]>, VEX_WIG;
1943 def VCVTPS2DQrm : VPDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1944 "cvtps2dq\t{$src, $dst|$dst, $src}",
1946 (int_x86_sse2_cvtps2dq (loadv4f32 addr:$src)))],
1947 IIC_SSE_CVT_PS_RM>, VEX, Sched<[WriteCvtF2ILd]>, VEX_WIG;
1948 def VCVTPS2DQYrr : VPDI<0x5B, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
1949 "cvtps2dq\t{$src, $dst|$dst, $src}",
1951 (int_x86_avx_cvt_ps2dq_256 VR256:$src))],
1952 IIC_SSE_CVT_PS_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>, VEX_WIG;
1953 def VCVTPS2DQYrm : VPDI<0x5B, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
1954 "cvtps2dq\t{$src, $dst|$dst, $src}",
1956 (int_x86_avx_cvt_ps2dq_256 (loadv8f32 addr:$src)))],
1957 IIC_SSE_CVT_PS_RM>, VEX, VEX_L, Sched<[WriteCvtF2ILd]>, VEX_WIG;
1958 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1959 "cvtps2dq\t{$src, $dst|$dst, $src}",
1960 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))],
1961 IIC_SSE_CVT_PS_RR>, Sched<[WriteCvtF2I]>;
1962 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1963 "cvtps2dq\t{$src, $dst|$dst, $src}",
1965 (int_x86_sse2_cvtps2dq (memopv4f32 addr:$src)))],
1966 IIC_SSE_CVT_PS_RM>, Sched<[WriteCvtF2ILd]>;
1969 // Convert Packed Double FP to Packed DW Integers
1970 let Predicates = [HasAVX, NoVLX] in {
1971 // The assembler can recognize rr 256-bit instructions by seeing a ymm
1972 // register, but the same isn't true when using memory operands instead.
1973 // Provide other assembly rr and rm forms to address this explicitly.
1974 def VCVTPD2DQrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1975 "vcvtpd2dq\t{$src, $dst|$dst, $src}",
1977 (v4i32 (X86cvtp2Int (v2f64 VR128:$src))))]>,
1978 VEX, Sched<[WriteCvtF2I]>, VEX_WIG;
1981 def : InstAlias<"vcvtpd2dqx\t{$src, $dst|$dst, $src}",
1982 (VCVTPD2DQrr VR128:$dst, VR128:$src), 0>;
1983 def VCVTPD2DQrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1984 "vcvtpd2dq{x}\t{$src, $dst|$dst, $src}",
1986 (v4i32 (X86cvtp2Int (loadv2f64 addr:$src))))]>, VEX,
1987 Sched<[WriteCvtF2ILd]>, VEX_WIG;
1988 def : InstAlias<"vcvtpd2dqx\t{$src, $dst|$dst, $src}",
1989 (VCVTPD2DQrm VR128:$dst, f128mem:$src), 0>;
1992 def VCVTPD2DQYrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
1993 "vcvtpd2dq\t{$src, $dst|$dst, $src}",
1995 (v4i32 (X86cvtp2Int (v4f64 VR256:$src))))]>,
1996 VEX, VEX_L, Sched<[WriteCvtF2I]>, VEX_WIG;
1997 def VCVTPD2DQYrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
1998 "vcvtpd2dq{y}\t{$src, $dst|$dst, $src}",
2000 (v4i32 (X86cvtp2Int (loadv4f64 addr:$src))))]>,
2001 VEX, VEX_L, Sched<[WriteCvtF2ILd]>, VEX_WIG;
2002 def : InstAlias<"vcvtpd2dqy\t{$src, $dst|$dst, $src}",
2003 (VCVTPD2DQYrr VR128:$dst, VR256:$src), 0>;
2004 def : InstAlias<"vcvtpd2dqy\t{$src, $dst|$dst, $src}",
2005 (VCVTPD2DQYrm VR128:$dst, f256mem:$src), 0>;
2008 def CVTPD2DQrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2009 "cvtpd2dq\t{$src, $dst|$dst, $src}",
2011 (v4i32 (X86cvtp2Int (memopv2f64 addr:$src))))],
2012 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtF2ILd]>;
2013 def CVTPD2DQrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2014 "cvtpd2dq\t{$src, $dst|$dst, $src}",
2016 (v4i32 (X86cvtp2Int (v2f64 VR128:$src))))],
2017 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2I]>;
2019 // Convert with truncation packed single/double fp to doubleword
2020 // SSE2 packed instructions with XS prefix
2021 let Predicates = [HasAVX, NoVLX] in {
2022 def VCVTTPS2DQrr : VS2SI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2023 "cvttps2dq\t{$src, $dst|$dst, $src}",
2025 (v4i32 (fp_to_sint (v4f32 VR128:$src))))],
2026 IIC_SSE_CVT_PS_RR>, VEX, Sched<[WriteCvtF2I]>, VEX_WIG;
2027 def VCVTTPS2DQrm : VS2SI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2028 "cvttps2dq\t{$src, $dst|$dst, $src}",
2030 (v4i32 (fp_to_sint (loadv4f32 addr:$src))))],
2031 IIC_SSE_CVT_PS_RM>, VEX, Sched<[WriteCvtF2ILd]>, VEX_WIG;
2032 def VCVTTPS2DQYrr : VS2SI<0x5B, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
2033 "cvttps2dq\t{$src, $dst|$dst, $src}",
2035 (v8i32 (fp_to_sint (v8f32 VR256:$src))))],
2036 IIC_SSE_CVT_PS_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>, VEX_WIG;
2037 def VCVTTPS2DQYrm : VS2SI<0x5B, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
2038 "cvttps2dq\t{$src, $dst|$dst, $src}",
2040 (v8i32 (fp_to_sint (loadv8f32 addr:$src))))],
2041 IIC_SSE_CVT_PS_RM>, VEX, VEX_L,
2042 Sched<[WriteCvtF2ILd]>, VEX_WIG;
2045 def CVTTPS2DQrr : S2SI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2046 "cvttps2dq\t{$src, $dst|$dst, $src}",
2048 (v4i32 (fp_to_sint (v4f32 VR128:$src))))],
2049 IIC_SSE_CVT_PS_RR>, Sched<[WriteCvtF2I]>;
2050 def CVTTPS2DQrm : S2SI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2051 "cvttps2dq\t{$src, $dst|$dst, $src}",
2053 (v4i32 (fp_to_sint (memopv4f32 addr:$src))))],
2054 IIC_SSE_CVT_PS_RM>, Sched<[WriteCvtF2ILd]>;
2056 let Predicates = [HasAVX, NoVLX] in
2057 def VCVTTPD2DQrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2058 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2060 (v4i32 (X86cvttp2si (v2f64 VR128:$src))))],
2061 IIC_SSE_CVT_PD_RR>, VEX, Sched<[WriteCvtF2I]>, VEX_WIG;
2063 // The assembler can recognize rr 256-bit instructions by seeing a ymm
2064 // register, but the same isn't true when using memory operands instead.
2065 // Provide other assembly rr and rm forms to address this explicitly.
2068 def : InstAlias<"vcvttpd2dqx\t{$src, $dst|$dst, $src}",
2069 (VCVTTPD2DQrr VR128:$dst, VR128:$src), 0>;
2070 let Predicates = [HasAVX, NoVLX] in
2071 def VCVTTPD2DQrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2072 "cvttpd2dq{x}\t{$src, $dst|$dst, $src}",
2074 (v4i32 (X86cvttp2si (loadv2f64 addr:$src))))],
2075 IIC_SSE_CVT_PD_RM>, VEX, Sched<[WriteCvtF2ILd]>, VEX_WIG;
2076 def : InstAlias<"vcvttpd2dqx\t{$src, $dst|$dst, $src}",
2077 (VCVTTPD2DQrm VR128:$dst, f128mem:$src), 0>;
2080 let Predicates = [HasAVX, NoVLX] in {
2081 def VCVTTPD2DQYrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2082 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2084 (v4i32 (fp_to_sint (v4f64 VR256:$src))))],
2085 IIC_SSE_CVT_PD_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>, VEX_WIG;
2086 def VCVTTPD2DQYrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2087 "cvttpd2dq{y}\t{$src, $dst|$dst, $src}",
2089 (v4i32 (fp_to_sint (loadv4f64 addr:$src))))],
2090 IIC_SSE_CVT_PD_RM>, VEX, VEX_L, Sched<[WriteCvtF2ILd]>, VEX_WIG;
2092 def : InstAlias<"vcvttpd2dqy\t{$src, $dst|$dst, $src}",
2093 (VCVTTPD2DQYrr VR128:$dst, VR256:$src), 0>;
2094 def : InstAlias<"vcvttpd2dqy\t{$src, $dst|$dst, $src}",
2095 (VCVTTPD2DQYrm VR128:$dst, f256mem:$src), 0>;
2097 let Predicates = [HasAVX, NoVLX] in {
2098 let AddedComplexity = 15 in {
2099 def : Pat<(X86vzmovl (v2i64 (bitconvert
2100 (v4i32 (X86cvtp2Int (v2f64 VR128:$src)))))),
2101 (VCVTPD2DQrr VR128:$src)>;
2102 def : Pat<(X86vzmovl (v2i64 (bitconvert
2103 (v4i32 (X86cvttp2si (v2f64 VR128:$src)))))),
2104 (VCVTTPD2DQrr VR128:$src)>;
2106 } // Predicates = [HasAVX]
2108 def CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2109 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2111 (v4i32 (X86cvttp2si (v2f64 VR128:$src))))],
2112 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2I]>;
2113 def CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
2114 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2116 (v4i32 (X86cvttp2si (memopv2f64 addr:$src))))],
2117 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtF2ILd]>;
2119 let Predicates = [UseSSE2] in {
2120 let AddedComplexity = 15 in {
2121 def : Pat<(X86vzmovl (v2i64 (bitconvert
2122 (v4i32 (X86cvtp2Int (v2f64 VR128:$src)))))),
2123 (CVTPD2DQrr VR128:$src)>;
2124 def : Pat<(X86vzmovl (v2i64 (bitconvert
2125 (v4i32 (X86cvttp2si (v2f64 VR128:$src)))))),
2126 (CVTTPD2DQrr VR128:$src)>;
2128 } // Predicates = [UseSSE2]
2130 // Convert packed single to packed double
2131 let Predicates = [HasAVX, NoVLX] in {
2132 // SSE2 instructions without OpSize prefix
2133 def VCVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2134 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2135 [(set VR128:$dst, (v2f64 (X86vfpext (v4f32 VR128:$src))))],
2136 IIC_SSE_CVT_PD_RR>, PS, VEX, Sched<[WriteCvtF2F]>, VEX_WIG;
2137 def VCVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2138 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2139 [(set VR128:$dst, (v2f64 (extloadv2f32 addr:$src)))],
2140 IIC_SSE_CVT_PD_RM>, PS, VEX, Sched<[WriteCvtF2FLd]>, VEX_WIG;
2141 def VCVTPS2PDYrr : I<0x5A, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
2142 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2143 [(set VR256:$dst, (v4f64 (fpextend (v4f32 VR128:$src))))],
2144 IIC_SSE_CVT_PD_RR>, PS, VEX, VEX_L, Sched<[WriteCvtF2F]>, VEX_WIG;
2145 def VCVTPS2PDYrm : I<0x5A, MRMSrcMem, (outs VR256:$dst), (ins f128mem:$src),
2146 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2147 [(set VR256:$dst, (v4f64 (extloadv4f32 addr:$src)))],
2148 IIC_SSE_CVT_PD_RM>, PS, VEX, VEX_L, Sched<[WriteCvtF2FLd]>, VEX_WIG;
2151 let Predicates = [UseSSE2] in {
2152 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2153 "cvtps2pd\t{$src, $dst|$dst, $src}",
2154 [(set VR128:$dst, (v2f64 (X86vfpext (v4f32 VR128:$src))))],
2155 IIC_SSE_CVT_PD_RR>, PS, Sched<[WriteCvtF2F]>;
2156 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2157 "cvtps2pd\t{$src, $dst|$dst, $src}",
2158 [(set VR128:$dst, (v2f64 (extloadv2f32 addr:$src)))],
2159 IIC_SSE_CVT_PD_RM>, PS, Sched<[WriteCvtF2FLd]>;
2162 // Convert Packed DW Integers to Packed Double FP
2163 let Predicates = [HasAVX, NoVLX] in {
2164 let hasSideEffects = 0, mayLoad = 1 in
2165 def VCVTDQ2PDrm : S2SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2166 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2168 (v2f64 (X86VSintToFP (bc_v4i32 (v2i64 (X86vzload addr:$src))))))]>,
2169 VEX, Sched<[WriteCvtI2FLd]>, VEX_WIG;
2170 def VCVTDQ2PDrr : S2SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2171 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2173 (v2f64 (X86VSintToFP (v4i32 VR128:$src))))]>,
2174 VEX, Sched<[WriteCvtI2F]>, VEX_WIG;
2175 def VCVTDQ2PDYrm : S2SI<0xE6, MRMSrcMem, (outs VR256:$dst), (ins i128mem:$src),
2176 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2178 (v4f64 (sint_to_fp (bc_v4i32 (loadv2i64 addr:$src)))))]>,
2179 VEX, VEX_L, Sched<[WriteCvtI2FLd]>, VEX_WIG;
2180 def VCVTDQ2PDYrr : S2SI<0xE6, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
2181 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2183 (v4f64 (sint_to_fp (v4i32 VR128:$src))))]>,
2184 VEX, VEX_L, Sched<[WriteCvtI2F]>, VEX_WIG;
2187 let hasSideEffects = 0, mayLoad = 1 in
2188 def CVTDQ2PDrm : S2SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2189 "cvtdq2pd\t{$src, $dst|$dst, $src}",
2191 (v2f64 (X86VSintToFP (bc_v4i32 (v2i64 (X86vzload addr:$src))))))],
2192 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtI2FLd]>;
2193 def CVTDQ2PDrr : S2SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2194 "cvtdq2pd\t{$src, $dst|$dst, $src}",
2196 (v2f64 (X86VSintToFP (v4i32 VR128:$src))))],
2197 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtI2F]>;
2199 // AVX register conversion intrinsics
2200 let Predicates = [HasAVX, NoVLX] in {
2201 def : Pat<(v2f64 (X86VSintToFP (bc_v4i32 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
2202 (VCVTDQ2PDrm addr:$src)>;
2203 } // Predicates = [HasAVX, NoVLX]
2205 // SSE2 register conversion intrinsics
2206 let Predicates = [UseSSE2] in {
2207 def : Pat<(v2f64 (X86VSintToFP (bc_v4i32 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
2208 (CVTDQ2PDrm addr:$src)>;
2209 } // Predicates = [UseSSE2]
2211 // Convert packed double to packed single
2212 // The assembler can recognize rr 256-bit instructions by seeing a ymm
2213 // register, but the same isn't true when using memory operands instead.
2214 // Provide other assembly rr and rm forms to address this explicitly.
2215 let Predicates = [HasAVX, NoVLX] in
2216 def VCVTPD2PSrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2217 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2218 [(set VR128:$dst, (X86vfpround (v2f64 VR128:$src)))],
2219 IIC_SSE_CVT_PD_RR>, VEX, Sched<[WriteCvtF2F]>, VEX_WIG;
2222 def : InstAlias<"vcvtpd2psx\t{$src, $dst|$dst, $src}",
2223 (VCVTPD2PSrr VR128:$dst, VR128:$src), 0>;
2224 let Predicates = [HasAVX, NoVLX] in
2225 def VCVTPD2PSrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2226 "cvtpd2ps{x}\t{$src, $dst|$dst, $src}",
2227 [(set VR128:$dst, (X86vfpround (loadv2f64 addr:$src)))],
2228 IIC_SSE_CVT_PD_RM>, VEX, Sched<[WriteCvtF2FLd]>, VEX_WIG;
2229 def : InstAlias<"vcvtpd2psx\t{$src, $dst|$dst, $src}",
2230 (VCVTPD2PSrm VR128:$dst, f128mem:$src), 0>;
2233 let Predicates = [HasAVX, NoVLX] in {
2234 def VCVTPD2PSYrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2235 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2236 [(set VR128:$dst, (fpround VR256:$src))],
2237 IIC_SSE_CVT_PD_RR>, VEX, VEX_L, Sched<[WriteCvtF2F]>, VEX_WIG;
2238 def VCVTPD2PSYrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2239 "cvtpd2ps{y}\t{$src, $dst|$dst, $src}",
2240 [(set VR128:$dst, (fpround (loadv4f64 addr:$src)))],
2241 IIC_SSE_CVT_PD_RM>, VEX, VEX_L, Sched<[WriteCvtF2FLd]>, VEX_WIG;
2243 def : InstAlias<"vcvtpd2psy\t{$src, $dst|$dst, $src}",
2244 (VCVTPD2PSYrr VR128:$dst, VR256:$src), 0>;
2245 def : InstAlias<"vcvtpd2psy\t{$src, $dst|$dst, $src}",
2246 (VCVTPD2PSYrm VR128:$dst, f256mem:$src), 0>;
2248 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2249 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2250 [(set VR128:$dst, (X86vfpround (v2f64 VR128:$src)))],
2251 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2F]>;
2252 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2253 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2254 [(set VR128:$dst, (X86vfpround (memopv2f64 addr:$src)))],
2255 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtF2FLd]>;
2257 // AVX 256-bit register conversion intrinsics
2258 // FIXME: Migrate SSE conversion intrinsics matching to use patterns as below
2259 // whenever possible to avoid declaring two versions of each one.
2261 let Predicates = [HasAVX, NoVLX] in {
2262 // Match fpround and fpextend for 128/256-bit conversions
2263 let AddedComplexity = 15 in
2264 def : Pat<(X86vzmovl (v2f64 (bitconvert
2265 (v4f32 (X86vfpround (v2f64 VR128:$src)))))),
2266 (VCVTPD2PSrr VR128:$src)>;
2269 let Predicates = [UseSSE2] in {
2270 // Match fpround and fpextend for 128 conversions
2271 let AddedComplexity = 15 in
2272 def : Pat<(X86vzmovl (v2f64 (bitconvert
2273 (v4f32 (X86vfpround (v2f64 VR128:$src)))))),
2274 (CVTPD2PSrr VR128:$src)>;
2277 //===----------------------------------------------------------------------===//
2278 // SSE 1 & 2 - Compare Instructions
2279 //===----------------------------------------------------------------------===//
2281 // sse12_cmp_scalar - sse 1 & 2 compare scalar instructions
2282 multiclass sse12_cmp_scalar<RegisterClass RC, X86MemOperand x86memop,
2283 Operand CC, SDNode OpNode, ValueType VT,
2284 PatFrag ld_frag, string asm, string asm_alt,
2285 OpndItins itins, ImmLeaf immLeaf> {
2286 let isCommutable = 1 in
2287 def rr : SIi8<0xC2, MRMSrcReg,
2288 (outs RC:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
2289 [(set RC:$dst, (OpNode (VT RC:$src1), RC:$src2, immLeaf:$cc))],
2290 itins.rr>, Sched<[itins.Sched]>;
2291 def rm : SIi8<0xC2, MRMSrcMem,
2292 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
2293 [(set RC:$dst, (OpNode (VT RC:$src1),
2294 (ld_frag addr:$src2), immLeaf:$cc))],
2296 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2298 // Accept explicit immediate argument form instead of comparison code.
2299 let isAsmParserOnly = 1, hasSideEffects = 0 in {
2300 def rr_alt : SIi8<0xC2, MRMSrcReg, (outs RC:$dst),
2301 (ins RC:$src1, RC:$src2, u8imm:$cc), asm_alt, [],
2302 IIC_SSE_ALU_F32S_RR>, Sched<[itins.Sched]>;
2304 def rm_alt : SIi8<0xC2, MRMSrcMem, (outs RC:$dst),
2305 (ins RC:$src1, x86memop:$src2, u8imm:$cc), asm_alt, [],
2306 IIC_SSE_ALU_F32S_RM>,
2307 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2311 let ExeDomain = SSEPackedSingle in
2312 defm VCMPSS : sse12_cmp_scalar<FR32, f32mem, AVXCC, X86cmps, f32, loadf32,
2313 "cmp${cc}ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2314 "cmpss\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2315 SSE_ALU_F32S, i8immZExt5>, XS, VEX_4V, VEX_LIG, VEX_WIG;
2316 let ExeDomain = SSEPackedDouble in
2317 defm VCMPSD : sse12_cmp_scalar<FR64, f64mem, AVXCC, X86cmps, f64, loadf64,
2318 "cmp${cc}sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2319 "cmpsd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2320 SSE_ALU_F32S, i8immZExt5>, // same latency as 32 bit compare
2321 XD, VEX_4V, VEX_LIG, VEX_WIG;
2323 let Constraints = "$src1 = $dst" in {
2324 let ExeDomain = SSEPackedSingle in
2325 defm CMPSS : sse12_cmp_scalar<FR32, f32mem, SSECC, X86cmps, f32, loadf32,
2326 "cmp${cc}ss\t{$src2, $dst|$dst, $src2}",
2327 "cmpss\t{$cc, $src2, $dst|$dst, $src2, $cc}", SSE_ALU_F32S,
2329 let ExeDomain = SSEPackedDouble in
2330 defm CMPSD : sse12_cmp_scalar<FR64, f64mem, SSECC, X86cmps, f64, loadf64,
2331 "cmp${cc}sd\t{$src2, $dst|$dst, $src2}",
2332 "cmpsd\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2333 SSE_ALU_F64S, i8immZExt3>, XD;
2336 multiclass sse12_cmp_scalar_int<Operand memop, Operand CC,
2337 Intrinsic Int, string asm, OpndItins itins,
2338 ImmLeaf immLeaf, ComplexPattern mem_cpat> {
2339 def rr : SIi8<0xC2, MRMSrcReg, (outs VR128:$dst),
2340 (ins VR128:$src1, VR128:$src, CC:$cc), asm,
2341 [(set VR128:$dst, (Int VR128:$src1,
2342 VR128:$src, immLeaf:$cc))],
2344 Sched<[itins.Sched]>;
2346 def rm : SIi8<0xC2, MRMSrcMem, (outs VR128:$dst),
2347 (ins VR128:$src1, memop:$src, CC:$cc), asm,
2348 [(set VR128:$dst, (Int VR128:$src1,
2349 mem_cpat:$src, immLeaf:$cc))],
2351 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2354 let isCodeGenOnly = 1 in {
2355 // Aliases to match intrinsics which expect XMM operand(s).
2356 let ExeDomain = SSEPackedSingle in
2357 defm Int_VCMPSS : sse12_cmp_scalar_int<ssmem, AVXCC, int_x86_sse_cmp_ss,
2358 "cmp${cc}ss\t{$src, $src1, $dst|$dst, $src1, $src}",
2359 SSE_ALU_F32S, i8immZExt5, sse_load_f32>,
2361 let ExeDomain = SSEPackedDouble in
2362 defm Int_VCMPSD : sse12_cmp_scalar_int<sdmem, AVXCC, int_x86_sse2_cmp_sd,
2363 "cmp${cc}sd\t{$src, $src1, $dst|$dst, $src1, $src}",
2364 SSE_ALU_F32S, i8immZExt5, sse_load_f64>, // same latency as f32
2366 let Constraints = "$src1 = $dst" in {
2367 let ExeDomain = SSEPackedSingle in
2368 defm Int_CMPSS : sse12_cmp_scalar_int<ssmem, SSECC, int_x86_sse_cmp_ss,
2369 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
2370 SSE_ALU_F32S, i8immZExt3, sse_load_f32>, XS;
2371 let ExeDomain = SSEPackedDouble in
2372 defm Int_CMPSD : sse12_cmp_scalar_int<sdmem, SSECC, int_x86_sse2_cmp_sd,
2373 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
2374 SSE_ALU_F64S, i8immZExt3, sse_load_f64>,
2380 // sse12_ord_cmp - Unordered/Ordered scalar fp compare and set EFLAGS
2381 multiclass sse12_ord_cmp<bits<8> opc, RegisterClass RC, SDNode OpNode,
2382 ValueType vt, X86MemOperand x86memop,
2383 PatFrag ld_frag, string OpcodeStr> {
2384 def rr: SI<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
2385 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2386 [(set EFLAGS, (OpNode (vt RC:$src1), RC:$src2))],
2390 def rm: SI<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
2391 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2392 [(set EFLAGS, (OpNode (vt RC:$src1),
2393 (ld_frag addr:$src2)))],
2395 Sched<[WriteFAddLd, ReadAfterLd]>;
2398 // sse12_ord_cmp_int - Intrinsic version of sse12_ord_cmp
2399 multiclass sse12_ord_cmp_int<bits<8> opc, RegisterClass RC, SDNode OpNode,
2400 ValueType vt, Operand memop,
2401 ComplexPattern mem_cpat, string OpcodeStr> {
2402 def rr: SI<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
2403 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2404 [(set EFLAGS, (OpNode (vt RC:$src1), RC:$src2))],
2408 def rm: SI<opc, MRMSrcMem, (outs), (ins RC:$src1, memop:$src2),
2409 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2410 [(set EFLAGS, (OpNode (vt RC:$src1),
2413 Sched<[WriteFAddLd, ReadAfterLd]>;
2416 let Defs = [EFLAGS] in {
2417 defm VUCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
2418 "ucomiss">, PS, VEX, VEX_LIG, VEX_WIG;
2419 defm VUCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
2420 "ucomisd">, PD, VEX, VEX_LIG, VEX_WIG;
2421 let Pattern = []<dag> in {
2422 defm VCOMISS : sse12_ord_cmp<0x2F, FR32, undef, f32, f32mem, loadf32,
2423 "comiss">, PS, VEX, VEX_LIG, VEX_WIG;
2424 defm VCOMISD : sse12_ord_cmp<0x2F, FR64, undef, f64, f64mem, loadf64,
2425 "comisd">, PD, VEX, VEX_LIG, VEX_WIG;
2428 let isCodeGenOnly = 1 in {
2429 defm Int_VUCOMISS : sse12_ord_cmp_int<0x2E, VR128, X86ucomi, v4f32, ssmem,
2430 sse_load_f32, "ucomiss">, PS, VEX, VEX_WIG;
2431 defm Int_VUCOMISD : sse12_ord_cmp_int<0x2E, VR128, X86ucomi, v2f64, sdmem,
2432 sse_load_f64, "ucomisd">, PD, VEX, VEX_WIG;
2434 defm Int_VCOMISS : sse12_ord_cmp_int<0x2F, VR128, X86comi, v4f32, ssmem,
2435 sse_load_f32, "comiss">, PS, VEX, VEX_WIG;
2436 defm Int_VCOMISD : sse12_ord_cmp_int<0x2F, VR128, X86comi, v2f64, sdmem,
2437 sse_load_f64, "comisd">, PD, VEX, VEX_WIG;
2439 defm UCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
2441 defm UCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
2444 let Pattern = []<dag> in {
2445 defm COMISS : sse12_ord_cmp<0x2F, FR32, undef, f32, f32mem, loadf32,
2447 defm COMISD : sse12_ord_cmp<0x2F, FR64, undef, f64, f64mem, loadf64,
2451 let isCodeGenOnly = 1 in {
2452 defm Int_UCOMISS : sse12_ord_cmp_int<0x2E, VR128, X86ucomi, v4f32, ssmem,
2453 sse_load_f32, "ucomiss">, PS;
2454 defm Int_UCOMISD : sse12_ord_cmp_int<0x2E, VR128, X86ucomi, v2f64, sdmem,
2455 sse_load_f64, "ucomisd">, PD;
2457 defm Int_COMISS : sse12_ord_cmp_int<0x2F, VR128, X86comi, v4f32, ssmem,
2458 sse_load_f32, "comiss">, PS;
2459 defm Int_COMISD : sse12_ord_cmp_int<0x2F, VR128, X86comi, v2f64, sdmem,
2460 sse_load_f64, "comisd">, PD;
2462 } // Defs = [EFLAGS]
2464 // sse12_cmp_packed - sse 1 & 2 compare packed instructions
2465 multiclass sse12_cmp_packed<RegisterClass RC, X86MemOperand x86memop,
2466 Operand CC, ValueType VT, string asm,
2467 string asm_alt, Domain d, ImmLeaf immLeaf,
2468 PatFrag ld_frag, OpndItins itins = SSE_ALU_F32P> {
2469 let isCommutable = 1 in
2470 def rri : PIi8<0xC2, MRMSrcReg,
2471 (outs RC:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
2472 [(set RC:$dst, (VT (X86cmpp RC:$src1, RC:$src2, immLeaf:$cc)))],
2475 def rmi : PIi8<0xC2, MRMSrcMem,
2476 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
2478 (VT (X86cmpp RC:$src1, (ld_frag addr:$src2), immLeaf:$cc)))],
2480 Sched<[WriteFAddLd, ReadAfterLd]>;
2482 // Accept explicit immediate argument form instead of comparison code.
2483 let isAsmParserOnly = 1, hasSideEffects = 0 in {
2484 def rri_alt : PIi8<0xC2, MRMSrcReg,
2485 (outs RC:$dst), (ins RC:$src1, RC:$src2, u8imm:$cc),
2486 asm_alt, [], itins.rr, d>, Sched<[WriteFAdd]>;
2488 def rmi_alt : PIi8<0xC2, MRMSrcMem,
2489 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, u8imm:$cc),
2490 asm_alt, [], itins.rm, d>,
2491 Sched<[WriteFAddLd, ReadAfterLd]>;
2495 defm VCMPPS : sse12_cmp_packed<VR128, f128mem, AVXCC, v4f32,
2496 "cmp${cc}ps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2497 "cmpps\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2498 SSEPackedSingle, i8immZExt5, loadv4f32>, PS, VEX_4V, VEX_WIG;
2499 defm VCMPPD : sse12_cmp_packed<VR128, f128mem, AVXCC, v2f64,
2500 "cmp${cc}pd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2501 "cmppd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2502 SSEPackedDouble, i8immZExt5, loadv2f64>, PD, VEX_4V, VEX_WIG;
2503 defm VCMPPSY : sse12_cmp_packed<VR256, f256mem, AVXCC, v8f32,
2504 "cmp${cc}ps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2505 "cmpps\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2506 SSEPackedSingle, i8immZExt5, loadv8f32>, PS, VEX_4V, VEX_L;
2507 defm VCMPPDY : sse12_cmp_packed<VR256, f256mem, AVXCC, v4f64,
2508 "cmp${cc}pd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2509 "cmppd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2510 SSEPackedDouble, i8immZExt5, loadv4f64>, PD, VEX_4V, VEX_L;
2511 let Constraints = "$src1 = $dst" in {
2512 defm CMPPS : sse12_cmp_packed<VR128, f128mem, SSECC, v4f32,
2513 "cmp${cc}ps\t{$src2, $dst|$dst, $src2}",
2514 "cmpps\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2515 SSEPackedSingle, i8immZExt5, memopv4f32, SSE_ALU_F32P>, PS;
2516 defm CMPPD : sse12_cmp_packed<VR128, f128mem, SSECC, v2f64,
2517 "cmp${cc}pd\t{$src2, $dst|$dst, $src2}",
2518 "cmppd\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2519 SSEPackedDouble, i8immZExt5, memopv2f64, SSE_ALU_F64P>, PD;
2522 //===----------------------------------------------------------------------===//
2523 // SSE 1 & 2 - Shuffle Instructions
2524 //===----------------------------------------------------------------------===//
2526 /// sse12_shuffle - sse 1 & 2 fp shuffle instructions
2527 multiclass sse12_shuffle<RegisterClass RC, X86MemOperand x86memop,
2528 ValueType vt, string asm, PatFrag mem_frag,
2530 def rmi : PIi8<0xC6, MRMSrcMem, (outs RC:$dst),
2531 (ins RC:$src1, x86memop:$src2, u8imm:$src3), asm,
2532 [(set RC:$dst, (vt (X86Shufp RC:$src1, (mem_frag addr:$src2),
2533 (i8 imm:$src3))))], IIC_SSE_SHUFP, d>,
2534 Sched<[WriteFShuffleLd, ReadAfterLd]>;
2535 def rri : PIi8<0xC6, MRMSrcReg, (outs RC:$dst),
2536 (ins RC:$src1, RC:$src2, u8imm:$src3), asm,
2537 [(set RC:$dst, (vt (X86Shufp RC:$src1, RC:$src2,
2538 (i8 imm:$src3))))], IIC_SSE_SHUFP, d>,
2539 Sched<[WriteFShuffle]>;
2542 let Predicates = [HasAVX, NoVLX] in {
2543 defm VSHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
2544 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2545 loadv4f32, SSEPackedSingle>, PS, VEX_4V, VEX_WIG;
2546 defm VSHUFPSY : sse12_shuffle<VR256, f256mem, v8f32,
2547 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2548 loadv8f32, SSEPackedSingle>, PS, VEX_4V, VEX_L, VEX_WIG;
2549 defm VSHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
2550 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2551 loadv2f64, SSEPackedDouble>, PD, VEX_4V, VEX_WIG;
2552 defm VSHUFPDY : sse12_shuffle<VR256, f256mem, v4f64,
2553 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2554 loadv4f64, SSEPackedDouble>, PD, VEX_4V, VEX_L, VEX_WIG;
2556 let Constraints = "$src1 = $dst" in {
2557 defm SHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
2558 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2559 memopv4f32, SSEPackedSingle>, PS;
2560 defm SHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
2561 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2562 memopv2f64, SSEPackedDouble>, PD;
2565 let Predicates = [HasAVX, NoVLX] in {
2566 def : Pat<(v4i32 (X86Shufp VR128:$src1,
2567 (bc_v4i32 (loadv2i64 addr:$src2)), (i8 imm:$imm))),
2568 (VSHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>;
2569 def : Pat<(v4i32 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2570 (VSHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>;
2572 def : Pat<(v2i64 (X86Shufp VR128:$src1,
2573 (loadv2i64 addr:$src2), (i8 imm:$imm))),
2574 (VSHUFPDrmi VR128:$src1, addr:$src2, imm:$imm)>;
2575 def : Pat<(v2i64 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2576 (VSHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>;
2579 def : Pat<(v8i32 (X86Shufp VR256:$src1, VR256:$src2, (i8 imm:$imm))),
2580 (VSHUFPSYrri VR256:$src1, VR256:$src2, imm:$imm)>;
2581 def : Pat<(v8i32 (X86Shufp VR256:$src1,
2582 (bc_v8i32 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
2583 (VSHUFPSYrmi VR256:$src1, addr:$src2, imm:$imm)>;
2585 def : Pat<(v4i64 (X86Shufp VR256:$src1, VR256:$src2, (i8 imm:$imm))),
2586 (VSHUFPDYrri VR256:$src1, VR256:$src2, imm:$imm)>;
2587 def : Pat<(v4i64 (X86Shufp VR256:$src1,
2588 (loadv4i64 addr:$src2), (i8 imm:$imm))),
2589 (VSHUFPDYrmi VR256:$src1, addr:$src2, imm:$imm)>;
2592 let Predicates = [UseSSE1] in {
2593 def : Pat<(v4i32 (X86Shufp VR128:$src1,
2594 (bc_v4i32 (memopv2i64 addr:$src2)), (i8 imm:$imm))),
2595 (SHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>;
2596 def : Pat<(v4i32 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2597 (SHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>;
2600 let Predicates = [UseSSE2] in {
2601 // Generic SHUFPD patterns
2602 def : Pat<(v2i64 (X86Shufp VR128:$src1,
2603 (memopv2i64 addr:$src2), (i8 imm:$imm))),
2604 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$imm)>;
2605 def : Pat<(v2i64 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2606 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>;
2609 //===----------------------------------------------------------------------===//
2610 // SSE 1 & 2 - Unpack FP Instructions
2611 //===----------------------------------------------------------------------===//
2613 /// sse12_unpack_interleave - sse 1 & 2 fp unpack and interleave
2614 multiclass sse12_unpack_interleave<bits<8> opc, SDNode OpNode, ValueType vt,
2615 PatFrag mem_frag, RegisterClass RC,
2616 X86MemOperand x86memop, string asm,
2617 Domain d, bit IsCommutable = 0> {
2618 let isCommutable = IsCommutable in
2619 def rr : PI<opc, MRMSrcReg,
2620 (outs RC:$dst), (ins RC:$src1, RC:$src2),
2622 (vt (OpNode RC:$src1, RC:$src2)))],
2623 IIC_SSE_UNPCK, d>, Sched<[WriteFShuffle]>;
2624 def rm : PI<opc, MRMSrcMem,
2625 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
2627 (vt (OpNode RC:$src1,
2628 (mem_frag addr:$src2))))],
2630 Sched<[WriteFShuffleLd, ReadAfterLd]>;
2633 let Predicates = [HasAVX, NoVLX] in {
2634 defm VUNPCKHPS: sse12_unpack_interleave<0x15, X86Unpckh, v4f32, loadv4f32,
2635 VR128, f128mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2636 SSEPackedSingle>, PS, VEX_4V, VEX_WIG;
2637 defm VUNPCKHPD: sse12_unpack_interleave<0x15, X86Unpckh, v2f64, loadv2f64,
2638 VR128, f128mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2639 SSEPackedDouble>, PD, VEX_4V, VEX_WIG;
2640 defm VUNPCKLPS: sse12_unpack_interleave<0x14, X86Unpckl, v4f32, loadv4f32,
2641 VR128, f128mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2642 SSEPackedSingle>, PS, VEX_4V, VEX_WIG;
2643 defm VUNPCKLPD: sse12_unpack_interleave<0x14, X86Unpckl, v2f64, loadv2f64,
2644 VR128, f128mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2645 SSEPackedDouble>, PD, VEX_4V, VEX_WIG;
2647 defm VUNPCKHPSY: sse12_unpack_interleave<0x15, X86Unpckh, v8f32, loadv8f32,
2648 VR256, f256mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2649 SSEPackedSingle>, PS, VEX_4V, VEX_L, VEX_WIG;
2650 defm VUNPCKHPDY: sse12_unpack_interleave<0x15, X86Unpckh, v4f64, loadv4f64,
2651 VR256, f256mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2652 SSEPackedDouble>, PD, VEX_4V, VEX_L, VEX_WIG;
2653 defm VUNPCKLPSY: sse12_unpack_interleave<0x14, X86Unpckl, v8f32, loadv8f32,
2654 VR256, f256mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2655 SSEPackedSingle>, PS, VEX_4V, VEX_L, VEX_WIG;
2656 defm VUNPCKLPDY: sse12_unpack_interleave<0x14, X86Unpckl, v4f64, loadv4f64,
2657 VR256, f256mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2658 SSEPackedDouble>, PD, VEX_4V, VEX_L, VEX_WIG;
2659 }// Predicates = [HasAVX, NoVLX]
2660 let Constraints = "$src1 = $dst" in {
2661 defm UNPCKHPS: sse12_unpack_interleave<0x15, X86Unpckh, v4f32, memopv4f32,
2662 VR128, f128mem, "unpckhps\t{$src2, $dst|$dst, $src2}",
2663 SSEPackedSingle>, PS;
2664 defm UNPCKHPD: sse12_unpack_interleave<0x15, X86Unpckh, v2f64, memopv2f64,
2665 VR128, f128mem, "unpckhpd\t{$src2, $dst|$dst, $src2}",
2666 SSEPackedDouble, 1>, PD;
2667 defm UNPCKLPS: sse12_unpack_interleave<0x14, X86Unpckl, v4f32, memopv4f32,
2668 VR128, f128mem, "unpcklps\t{$src2, $dst|$dst, $src2}",
2669 SSEPackedSingle>, PS;
2670 defm UNPCKLPD: sse12_unpack_interleave<0x14, X86Unpckl, v2f64, memopv2f64,
2671 VR128, f128mem, "unpcklpd\t{$src2, $dst|$dst, $src2}",
2672 SSEPackedDouble>, PD;
2673 } // Constraints = "$src1 = $dst"
2675 let Predicates = [HasAVX1Only] in {
2676 def : Pat<(v8i32 (X86Unpckl VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
2677 (VUNPCKLPSYrm VR256:$src1, addr:$src2)>;
2678 def : Pat<(v8i32 (X86Unpckl VR256:$src1, VR256:$src2)),
2679 (VUNPCKLPSYrr VR256:$src1, VR256:$src2)>;
2680 def : Pat<(v8i32 (X86Unpckh VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
2681 (VUNPCKHPSYrm VR256:$src1, addr:$src2)>;
2682 def : Pat<(v8i32 (X86Unpckh VR256:$src1, VR256:$src2)),
2683 (VUNPCKHPSYrr VR256:$src1, VR256:$src2)>;
2685 def : Pat<(v4i64 (X86Unpckl VR256:$src1, (loadv4i64 addr:$src2))),
2686 (VUNPCKLPDYrm VR256:$src1, addr:$src2)>;
2687 def : Pat<(v4i64 (X86Unpckl VR256:$src1, VR256:$src2)),
2688 (VUNPCKLPDYrr VR256:$src1, VR256:$src2)>;
2689 def : Pat<(v4i64 (X86Unpckh VR256:$src1, (loadv4i64 addr:$src2))),
2690 (VUNPCKHPDYrm VR256:$src1, addr:$src2)>;
2691 def : Pat<(v4i64 (X86Unpckh VR256:$src1, VR256:$src2)),
2692 (VUNPCKHPDYrr VR256:$src1, VR256:$src2)>;
2695 //===----------------------------------------------------------------------===//
2696 // SSE 1 & 2 - Extract Floating-Point Sign mask
2697 //===----------------------------------------------------------------------===//
2699 /// sse12_extr_sign_mask - sse 1 & 2 unpack and interleave
2700 multiclass sse12_extr_sign_mask<RegisterClass RC, ValueType vt,
2701 string asm, Domain d> {
2702 def rr : PI<0x50, MRMSrcReg, (outs GR32orGR64:$dst), (ins RC:$src),
2703 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
2704 [(set GR32orGR64:$dst, (X86movmsk (vt RC:$src)))], IIC_SSE_MOVMSK, d>,
2705 Sched<[WriteVecLogic]>;
2708 let Predicates = [HasAVX] in {
2709 defm VMOVMSKPS : sse12_extr_sign_mask<VR128, v4f32, "movmskps",
2710 SSEPackedSingle>, PS, VEX, VEX_WIG;
2711 defm VMOVMSKPD : sse12_extr_sign_mask<VR128, v2f64, "movmskpd",
2712 SSEPackedDouble>, PD, VEX, VEX_WIG;
2713 defm VMOVMSKPSY : sse12_extr_sign_mask<VR256, v8f32, "movmskps",
2714 SSEPackedSingle>, PS, VEX, VEX_L, VEX_WIG;
2715 defm VMOVMSKPDY : sse12_extr_sign_mask<VR256, v4f64, "movmskpd",
2716 SSEPackedDouble>, PD, VEX, VEX_L, VEX_WIG;
2719 defm MOVMSKPS : sse12_extr_sign_mask<VR128, v4f32, "movmskps",
2720 SSEPackedSingle>, PS;
2721 defm MOVMSKPD : sse12_extr_sign_mask<VR128, v2f64, "movmskpd",
2722 SSEPackedDouble>, PD;
2724 //===---------------------------------------------------------------------===//
2725 // SSE2 - Packed Integer Logical Instructions
2726 //===---------------------------------------------------------------------===//
2728 let ExeDomain = SSEPackedInt in { // SSE integer instructions
2730 /// PDI_binop_rm - Simple SSE2 binary operator.
2731 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
2732 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
2733 X86MemOperand x86memop, OpndItins itins,
2734 bit IsCommutable, bit Is2Addr> {
2735 let isCommutable = IsCommutable in
2736 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
2737 (ins RC:$src1, RC:$src2),
2739 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2740 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2741 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))], itins.rr>,
2742 Sched<[itins.Sched]>;
2743 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
2744 (ins RC:$src1, x86memop:$src2),
2746 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2747 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2748 [(set RC:$dst, (OpVT (OpNode RC:$src1,
2749 (bitconvert (memop_frag addr:$src2)))))],
2751 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2753 } // ExeDomain = SSEPackedInt
2755 multiclass PDI_binop_all<bits<8> opc, string OpcodeStr, SDNode Opcode,
2756 ValueType OpVT128, ValueType OpVT256,
2757 OpndItins itins, bit IsCommutable = 0, Predicate prd> {
2758 let Predicates = [HasAVX, prd] in
2759 defm V#NAME : PDI_binop_rm<opc, !strconcat("v", OpcodeStr), Opcode, OpVT128,
2760 VR128, loadv2i64, i128mem, itins, IsCommutable, 0>, VEX_4V, VEX_WIG;
2762 let Constraints = "$src1 = $dst" in
2763 defm NAME : PDI_binop_rm<opc, OpcodeStr, Opcode, OpVT128, VR128,
2764 memopv2i64, i128mem, itins, IsCommutable, 1>;
2766 let Predicates = [HasAVX2, prd] in
2767 defm V#NAME#Y : PDI_binop_rm<opc, !strconcat("v", OpcodeStr), Opcode,
2768 OpVT256, VR256, loadv4i64, i256mem, itins,
2769 IsCommutable, 0>, VEX_4V, VEX_L, VEX_WIG;
2772 // These are ordered here for pattern ordering requirements with the fp versions
2774 defm PAND : PDI_binop_all<0xDB, "pand", and, v2i64, v4i64,
2775 SSE_VEC_BIT_ITINS_P, 1, NoVLX>;
2776 defm POR : PDI_binop_all<0xEB, "por", or, v2i64, v4i64,
2777 SSE_VEC_BIT_ITINS_P, 1, NoVLX>;
2778 defm PXOR : PDI_binop_all<0xEF, "pxor", xor, v2i64, v4i64,
2779 SSE_VEC_BIT_ITINS_P, 1, NoVLX>;
2780 defm PANDN : PDI_binop_all<0xDF, "pandn", X86andnp, v2i64, v4i64,
2781 SSE_VEC_BIT_ITINS_P, 0, NoVLX>;
2783 //===----------------------------------------------------------------------===//
2784 // SSE 1 & 2 - Logical Instructions
2785 //===----------------------------------------------------------------------===//
2787 /// sse12_fp_packed_logical - SSE 1 & 2 packed FP logical ops
2789 multiclass sse12_fp_packed_logical<bits<8> opc, string OpcodeStr,
2791 let Predicates = [HasAVX, NoVLX] in {
2792 defm V#NAME#PSY : sse12_fp_packed_logical_rm<opc, VR256, SSEPackedSingle,
2793 !strconcat(OpcodeStr, "ps"), f256mem,
2794 [(set VR256:$dst, (OpNode (bc_v4i64 (v8f32 VR256:$src1)),
2795 (bc_v4i64 (v8f32 VR256:$src2))))],
2796 [(set VR256:$dst, (OpNode (bc_v4i64 (v8f32 VR256:$src1)),
2797 (loadv4i64 addr:$src2)))], 0>, PS, VEX_4V, VEX_L, VEX_WIG;
2799 defm V#NAME#PDY : sse12_fp_packed_logical_rm<opc, VR256, SSEPackedDouble,
2800 !strconcat(OpcodeStr, "pd"), f256mem,
2801 [(set VR256:$dst, (OpNode (bc_v4i64 (v4f64 VR256:$src1)),
2802 (bc_v4i64 (v4f64 VR256:$src2))))],
2803 [(set VR256:$dst, (OpNode (bc_v4i64 (v4f64 VR256:$src1)),
2804 (loadv4i64 addr:$src2)))], 0>,
2805 PD, VEX_4V, VEX_L, VEX_WIG;
2807 defm V#NAME#PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
2808 !strconcat(OpcodeStr, "ps"), f128mem,
2809 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2810 (bc_v2i64 (v4f32 VR128:$src2))))],
2811 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2812 (loadv2i64 addr:$src2)))], 0>, PS, VEX_4V, VEX_WIG;
2814 defm V#NAME#PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
2815 !strconcat(OpcodeStr, "pd"), f128mem,
2816 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2817 (bc_v2i64 (v2f64 VR128:$src2))))],
2818 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2819 (loadv2i64 addr:$src2)))], 0>,
2820 PD, VEX_4V, VEX_WIG;
2823 let Constraints = "$src1 = $dst" in {
2824 defm PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
2825 !strconcat(OpcodeStr, "ps"), f128mem,
2826 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2827 (bc_v2i64 (v4f32 VR128:$src2))))],
2828 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2829 (memopv2i64 addr:$src2)))]>, PS;
2831 defm PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
2832 !strconcat(OpcodeStr, "pd"), f128mem,
2833 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2834 (bc_v2i64 (v2f64 VR128:$src2))))],
2835 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2836 (memopv2i64 addr:$src2)))]>, PD;
2840 defm AND : sse12_fp_packed_logical<0x54, "and", and>;
2841 defm OR : sse12_fp_packed_logical<0x56, "or", or>;
2842 defm XOR : sse12_fp_packed_logical<0x57, "xor", xor>;
2843 let isCommutable = 0 in
2844 defm ANDN : sse12_fp_packed_logical<0x55, "andn", X86andnp>;
2846 // If only AVX1 is supported, we need to handle integer operations with
2847 // floating point instructions since the integer versions aren't available.
2848 let Predicates = [HasAVX1Only] in {
2849 def : Pat<(v4i64 (and VR256:$src1, VR256:$src2)),
2850 (VANDPSYrr VR256:$src1, VR256:$src2)>;
2851 def : Pat<(v4i64 (or VR256:$src1, VR256:$src2)),
2852 (VORPSYrr VR256:$src1, VR256:$src2)>;
2853 def : Pat<(v4i64 (xor VR256:$src1, VR256:$src2)),
2854 (VXORPSYrr VR256:$src1, VR256:$src2)>;
2855 def : Pat<(v4i64 (X86andnp VR256:$src1, VR256:$src2)),
2856 (VANDNPSYrr VR256:$src1, VR256:$src2)>;
2858 def : Pat<(and VR256:$src1, (loadv4i64 addr:$src2)),
2859 (VANDPSYrm VR256:$src1, addr:$src2)>;
2860 def : Pat<(or VR256:$src1, (loadv4i64 addr:$src2)),
2861 (VORPSYrm VR256:$src1, addr:$src2)>;
2862 def : Pat<(xor VR256:$src1, (loadv4i64 addr:$src2)),
2863 (VXORPSYrm VR256:$src1, addr:$src2)>;
2864 def : Pat<(X86andnp VR256:$src1, (loadv4i64 addr:$src2)),
2865 (VANDNPSYrm VR256:$src1, addr:$src2)>;
2868 let Predicates = [HasAVX, NoVLX_Or_NoDQI] in {
2869 // Use packed logical operations for scalar ops.
2870 def : Pat<(f64 (X86fand FR64:$src1, FR64:$src2)),
2871 (COPY_TO_REGCLASS (VANDPDrr
2872 (COPY_TO_REGCLASS FR64:$src1, VR128),
2873 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2874 def : Pat<(f64 (X86for FR64:$src1, FR64:$src2)),
2875 (COPY_TO_REGCLASS (VORPDrr
2876 (COPY_TO_REGCLASS FR64:$src1, VR128),
2877 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2878 def : Pat<(f64 (X86fxor FR64:$src1, FR64:$src2)),
2879 (COPY_TO_REGCLASS (VXORPDrr
2880 (COPY_TO_REGCLASS FR64:$src1, VR128),
2881 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2882 def : Pat<(f64 (X86fandn FR64:$src1, FR64:$src2)),
2883 (COPY_TO_REGCLASS (VANDNPDrr
2884 (COPY_TO_REGCLASS FR64:$src1, VR128),
2885 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2887 def : Pat<(f32 (X86fand FR32:$src1, FR32:$src2)),
2888 (COPY_TO_REGCLASS (VANDPSrr
2889 (COPY_TO_REGCLASS FR32:$src1, VR128),
2890 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2891 def : Pat<(f32 (X86for FR32:$src1, FR32:$src2)),
2892 (COPY_TO_REGCLASS (VORPSrr
2893 (COPY_TO_REGCLASS FR32:$src1, VR128),
2894 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2895 def : Pat<(f32 (X86fxor FR32:$src1, FR32:$src2)),
2896 (COPY_TO_REGCLASS (VXORPSrr
2897 (COPY_TO_REGCLASS FR32:$src1, VR128),
2898 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2899 def : Pat<(f32 (X86fandn FR32:$src1, FR32:$src2)),
2900 (COPY_TO_REGCLASS (VANDNPSrr
2901 (COPY_TO_REGCLASS FR32:$src1, VR128),
2902 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2905 let Predicates = [UseSSE1] in {
2906 // Use packed logical operations for scalar ops.
2907 def : Pat<(f32 (X86fand FR32:$src1, FR32:$src2)),
2908 (COPY_TO_REGCLASS (ANDPSrr
2909 (COPY_TO_REGCLASS FR32:$src1, VR128),
2910 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2911 def : Pat<(f32 (X86for FR32:$src1, FR32:$src2)),
2912 (COPY_TO_REGCLASS (ORPSrr
2913 (COPY_TO_REGCLASS FR32:$src1, VR128),
2914 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2915 def : Pat<(f32 (X86fxor FR32:$src1, FR32:$src2)),
2916 (COPY_TO_REGCLASS (XORPSrr
2917 (COPY_TO_REGCLASS FR32:$src1, VR128),
2918 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2919 def : Pat<(f32 (X86fandn FR32:$src1, FR32:$src2)),
2920 (COPY_TO_REGCLASS (ANDNPSrr
2921 (COPY_TO_REGCLASS FR32:$src1, VR128),
2922 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2925 let Predicates = [UseSSE2] in {
2926 // Use packed logical operations for scalar ops.
2927 def : Pat<(f64 (X86fand FR64:$src1, FR64:$src2)),
2928 (COPY_TO_REGCLASS (ANDPDrr
2929 (COPY_TO_REGCLASS FR64:$src1, VR128),
2930 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2931 def : Pat<(f64 (X86for FR64:$src1, FR64:$src2)),
2932 (COPY_TO_REGCLASS (ORPDrr
2933 (COPY_TO_REGCLASS FR64:$src1, VR128),
2934 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2935 def : Pat<(f64 (X86fxor FR64:$src1, FR64:$src2)),
2936 (COPY_TO_REGCLASS (XORPDrr
2937 (COPY_TO_REGCLASS FR64:$src1, VR128),
2938 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2939 def : Pat<(f64 (X86fandn FR64:$src1, FR64:$src2)),
2940 (COPY_TO_REGCLASS (ANDNPDrr
2941 (COPY_TO_REGCLASS FR64:$src1, VR128),
2942 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2945 // Patterns for packed operations when we don't have integer type available.
2946 def : Pat<(v4f32 (X86fand VR128:$src1, VR128:$src2)),
2947 (ANDPSrr VR128:$src1, VR128:$src2)>;
2948 def : Pat<(v4f32 (X86for VR128:$src1, VR128:$src2)),
2949 (ORPSrr VR128:$src1, VR128:$src2)>;
2950 def : Pat<(v4f32 (X86fxor VR128:$src1, VR128:$src2)),
2951 (XORPSrr VR128:$src1, VR128:$src2)>;
2952 def : Pat<(v4f32 (X86fandn VR128:$src1, VR128:$src2)),
2953 (ANDNPSrr VR128:$src1, VR128:$src2)>;
2955 def : Pat<(X86fand VR128:$src1, (memopv4f32 addr:$src2)),
2956 (ANDPSrm VR128:$src1, addr:$src2)>;
2957 def : Pat<(X86for VR128:$src1, (memopv4f32 addr:$src2)),
2958 (ORPSrm VR128:$src1, addr:$src2)>;
2959 def : Pat<(X86fxor VR128:$src1, (memopv4f32 addr:$src2)),
2960 (XORPSrm VR128:$src1, addr:$src2)>;
2961 def : Pat<(X86fandn VR128:$src1, (memopv4f32 addr:$src2)),
2962 (ANDNPSrm VR128:$src1, addr:$src2)>;
2964 //===----------------------------------------------------------------------===//
2965 // SSE 1 & 2 - Arithmetic Instructions
2966 //===----------------------------------------------------------------------===//
2968 /// basic_sse12_fp_binop_xxx - SSE 1 & 2 binops come in both scalar and
2971 /// In addition, we also have a special variant of the scalar form here to
2972 /// represent the associated intrinsic operation. This form is unlike the
2973 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
2974 /// and leaves the top elements unmodified (therefore these cannot be commuted).
2976 /// These three forms can each be reg+reg or reg+mem.
2979 /// FIXME: once all 256-bit intrinsics are matched, cleanup and refactor those
2981 multiclass basic_sse12_fp_binop_p<bits<8> opc, string OpcodeStr,
2982 SDNode OpNode, SizeItins itins> {
2983 let Predicates = [HasAVX, NoVLX] in {
2984 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode,
2985 VR128, v4f32, f128mem, loadv4f32,
2986 SSEPackedSingle, itins.s, 0>, PS, VEX_4V, VEX_WIG;
2987 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode,
2988 VR128, v2f64, f128mem, loadv2f64,
2989 SSEPackedDouble, itins.d, 0>, PD, VEX_4V, VEX_WIG;
2991 defm V#NAME#PSY : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"),
2992 OpNode, VR256, v8f32, f256mem, loadv8f32,
2993 SSEPackedSingle, itins.s, 0>, PS, VEX_4V, VEX_L, VEX_WIG;
2994 defm V#NAME#PDY : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"),
2995 OpNode, VR256, v4f64, f256mem, loadv4f64,
2996 SSEPackedDouble, itins.d, 0>, PD, VEX_4V, VEX_L, VEX_WIG;
2999 let Constraints = "$src1 = $dst" in {
3000 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode, VR128,
3001 v4f32, f128mem, memopv4f32, SSEPackedSingle,
3003 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode, VR128,
3004 v2f64, f128mem, memopv2f64, SSEPackedDouble,
3009 multiclass basic_sse12_fp_binop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
3011 defm V#NAME#SS : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "ss"),
3012 OpNode, FR32, f32mem, SSEPackedSingle, itins.s, 0>,
3013 XS, VEX_4V, VEX_LIG, VEX_WIG;
3014 defm V#NAME#SD : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "sd"),
3015 OpNode, FR64, f64mem, SSEPackedDouble, itins.d, 0>,
3016 XD, VEX_4V, VEX_LIG, VEX_WIG;
3018 let Constraints = "$src1 = $dst" in {
3019 defm SS : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "ss"),
3020 OpNode, FR32, f32mem, SSEPackedSingle,
3022 defm SD : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "sd"),
3023 OpNode, FR64, f64mem, SSEPackedDouble,
3028 multiclass basic_sse12_fp_binop_s_int<bits<8> opc, string OpcodeStr,
3029 SDPatternOperator OpNode,
3031 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, OpNode, VR128, v4f32,
3032 !strconcat(OpcodeStr, "ss"), ssmem, sse_load_f32,
3033 SSEPackedSingle, itins.s, 0>, XS, VEX_4V, VEX_LIG, VEX_WIG;
3034 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, OpNode, VR128, v2f64,
3035 !strconcat(OpcodeStr, "sd"), sdmem, sse_load_f64,
3036 SSEPackedDouble, itins.d, 0>, XD, VEX_4V, VEX_LIG, VEX_WIG;
3038 let Constraints = "$src1 = $dst" in {
3039 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, OpNode, VR128, v4f32,
3040 !strconcat(OpcodeStr, "ss"), ssmem, sse_load_f32,
3041 SSEPackedSingle, itins.s>, XS;
3042 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, OpNode, VR128, v2f64,
3043 !strconcat(OpcodeStr, "sd"), sdmem, sse_load_f64,
3044 SSEPackedDouble, itins.d>, XD;
3048 // Binary Arithmetic instructions
3049 defm ADD : basic_sse12_fp_binop_p<0x58, "add", fadd, SSE_ALU_ITINS_P>,
3050 basic_sse12_fp_binop_s<0x58, "add", fadd, SSE_ALU_ITINS_S>,
3051 basic_sse12_fp_binop_s_int<0x58, "add", null_frag, SSE_ALU_ITINS_S>;
3052 defm MUL : basic_sse12_fp_binop_p<0x59, "mul", fmul, SSE_MUL_ITINS_P>,
3053 basic_sse12_fp_binop_s<0x59, "mul", fmul, SSE_MUL_ITINS_S>,
3054 basic_sse12_fp_binop_s_int<0x59, "mul", null_frag, SSE_MUL_ITINS_S>;
3055 let isCommutable = 0 in {
3056 defm SUB : basic_sse12_fp_binop_p<0x5C, "sub", fsub, SSE_ALU_ITINS_P>,
3057 basic_sse12_fp_binop_s<0x5C, "sub", fsub, SSE_ALU_ITINS_S>,
3058 basic_sse12_fp_binop_s_int<0x5C, "sub", null_frag,SSE_ALU_ITINS_S>;
3059 defm DIV : basic_sse12_fp_binop_p<0x5E, "div", fdiv, SSE_DIV_ITINS_P>,
3060 basic_sse12_fp_binop_s<0x5E, "div", fdiv, SSE_DIV_ITINS_S>,
3061 basic_sse12_fp_binop_s_int<0x5E, "div", null_frag,SSE_DIV_ITINS_S>;
3062 defm MAX : basic_sse12_fp_binop_p<0x5F, "max", X86fmax, SSE_ALU_ITINS_P>,
3063 basic_sse12_fp_binop_s<0x5F, "max", X86fmax, SSE_ALU_ITINS_S>,
3064 basic_sse12_fp_binop_s_int<0x5F, "max", X86fmaxs, SSE_ALU_ITINS_S>;
3065 defm MIN : basic_sse12_fp_binop_p<0x5D, "min", X86fmin, SSE_ALU_ITINS_P>,
3066 basic_sse12_fp_binop_s<0x5D, "min", X86fmin, SSE_ALU_ITINS_S>,
3067 basic_sse12_fp_binop_s_int<0x5D, "min", X86fmins, SSE_ALU_ITINS_S>;
3070 let isCodeGenOnly = 1 in {
3071 defm MAXC: basic_sse12_fp_binop_p<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_P>,
3072 basic_sse12_fp_binop_s<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_S>;
3073 defm MINC: basic_sse12_fp_binop_p<0x5D, "min", X86fminc, SSE_ALU_ITINS_P>,
3074 basic_sse12_fp_binop_s<0x5D, "min", X86fminc, SSE_ALU_ITINS_S>;
3077 // Patterns used to select SSE scalar fp arithmetic instructions from
3080 // (1) a scalar fp operation followed by a blend
3082 // The effect is that the backend no longer emits unnecessary vector
3083 // insert instructions immediately after SSE scalar fp instructions
3084 // like addss or mulss.
3086 // For example, given the following code:
3087 // __m128 foo(__m128 A, __m128 B) {
3092 // Previously we generated:
3093 // addss %xmm0, %xmm1
3094 // movss %xmm1, %xmm0
3097 // addss %xmm1, %xmm0
3099 // (2) a vector packed single/double fp operation followed by a vector insert
3101 // The effect is that the backend converts the packed fp instruction
3102 // followed by a vector insert into a single SSE scalar fp instruction.
3104 // For example, given the following code:
3105 // __m128 foo(__m128 A, __m128 B) {
3106 // __m128 C = A + B;
3107 // return (__m128) {c[0], a[1], a[2], a[3]};
3110 // Previously we generated:
3111 // addps %xmm0, %xmm1
3112 // movss %xmm1, %xmm0
3115 // addss %xmm1, %xmm0
3117 // TODO: Some canonicalization in lowering would simplify the number of
3118 // patterns we have to try to match.
3119 multiclass scalar_math_f32_patterns<SDNode Op, string OpcPrefix> {
3120 let Predicates = [UseSSE1] in {
3121 // extracted scalar math op with insert via movss
3122 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3123 (Op (f32 (extractelt (v4f32 VR128:$dst), (iPTR 0))),
3125 (!cast<I>(OpcPrefix#SSrr_Int) v4f32:$dst,
3126 (COPY_TO_REGCLASS FR32:$src, VR128))>;
3128 // vector math op with insert via movss
3129 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3130 (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3131 (!cast<I>(OpcPrefix#SSrr_Int) v4f32:$dst, v4f32:$src)>;
3134 // With SSE 4.1, blendi is preferred to movsd, so match that too.
3135 let Predicates = [UseSSE41] in {
3136 // extracted scalar math op with insert via blend
3137 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3138 (Op (f32 (extractelt (v4f32 VR128:$dst), (iPTR 0))),
3139 FR32:$src))), (i8 1))),
3140 (!cast<I>(OpcPrefix#SSrr_Int) v4f32:$dst,
3141 (COPY_TO_REGCLASS FR32:$src, VR128))>;
3143 // vector math op with insert via blend
3144 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$dst),
3145 (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)), (i8 1))),
3146 (!cast<I>(OpcPrefix#SSrr_Int)v4f32:$dst, v4f32:$src)>;
3150 // Repeat everything for AVX.
3151 let Predicates = [UseAVX] in {
3152 // extracted scalar math op with insert via movss
3153 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3154 (Op (f32 (extractelt (v4f32 VR128:$dst), (iPTR 0))),
3156 (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst,
3157 (COPY_TO_REGCLASS FR32:$src, VR128))>;
3159 // extracted scalar math op with insert via blend
3160 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3161 (Op (f32 (extractelt (v4f32 VR128:$dst), (iPTR 0))),
3162 FR32:$src))), (i8 1))),
3163 (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst,
3164 (COPY_TO_REGCLASS FR32:$src, VR128))>;
3166 // vector math op with insert via movss
3167 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3168 (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3169 (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst, v4f32:$src)>;
3171 // vector math op with insert via blend
3172 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$dst),
3173 (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)), (i8 1))),
3174 (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst, v4f32:$src)>;
3178 defm : scalar_math_f32_patterns<fadd, "ADD">;
3179 defm : scalar_math_f32_patterns<fsub, "SUB">;
3180 defm : scalar_math_f32_patterns<fmul, "MUL">;
3181 defm : scalar_math_f32_patterns<fdiv, "DIV">;
3183 multiclass scalar_math_f64_patterns<SDNode Op, string OpcPrefix> {
3184 let Predicates = [UseSSE2] in {
3185 // extracted scalar math op with insert via movsd
3186 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
3187 (Op (f64 (extractelt (v2f64 VR128:$dst), (iPTR 0))),
3189 (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst,
3190 (COPY_TO_REGCLASS FR64:$src, VR128))>;
3192 // vector math op with insert via movsd
3193 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3194 (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3195 (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
3198 // With SSE 4.1, blendi is preferred to movsd, so match those too.
3199 let Predicates = [UseSSE41] in {
3200 // extracted scalar math op with insert via blend
3201 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
3202 (Op (f64 (extractelt (v2f64 VR128:$dst), (iPTR 0))),
3203 FR64:$src))), (i8 1))),
3204 (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst,
3205 (COPY_TO_REGCLASS FR64:$src, VR128))>;
3207 // vector math op with insert via blend
3208 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst),
3209 (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)), (i8 1))),
3210 (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
3213 // Repeat everything for AVX.
3214 let Predicates = [UseAVX] in {
3215 // extracted scalar math op with insert via movsd
3216 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
3217 (Op (f64 (extractelt (v2f64 VR128:$dst), (iPTR 0))),
3219 (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst,
3220 (COPY_TO_REGCLASS FR64:$src, VR128))>;
3222 // extracted scalar math op with insert via blend
3223 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
3224 (Op (f64 (extractelt (v2f64 VR128:$dst), (iPTR 0))),
3225 FR64:$src))), (i8 1))),
3226 (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst,
3227 (COPY_TO_REGCLASS FR64:$src, VR128))>;
3229 // vector math op with insert via movsd
3230 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3231 (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3232 (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
3234 // vector math op with insert via blend
3235 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst),
3236 (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)), (i8 1))),
3237 (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
3241 defm : scalar_math_f64_patterns<fadd, "ADD">;
3242 defm : scalar_math_f64_patterns<fsub, "SUB">;
3243 defm : scalar_math_f64_patterns<fmul, "MUL">;
3244 defm : scalar_math_f64_patterns<fdiv, "DIV">;
3248 /// In addition, we also have a special variant of the scalar form here to
3249 /// represent the associated intrinsic operation. This form is unlike the
3250 /// plain scalar form, in that it takes an entire vector (instead of a
3251 /// scalar) and leaves the top elements undefined.
3253 /// And, we have a special variant form for a full-vector intrinsic form.
3255 let Sched = WriteFSqrt in {
3256 def SSE_SQRTPS : OpndItins<
3257 IIC_SSE_SQRTPS_RR, IIC_SSE_SQRTPS_RM
3260 def SSE_SQRTSS : OpndItins<
3261 IIC_SSE_SQRTSS_RR, IIC_SSE_SQRTSS_RM
3264 def SSE_SQRTPD : OpndItins<
3265 IIC_SSE_SQRTPD_RR, IIC_SSE_SQRTPD_RM
3268 def SSE_SQRTSD : OpndItins<
3269 IIC_SSE_SQRTSD_RR, IIC_SSE_SQRTSD_RM
3273 let Sched = WriteFRsqrt in {
3274 def SSE_RSQRTPS : OpndItins<
3275 IIC_SSE_RSQRTPS_RR, IIC_SSE_RSQRTPS_RM
3278 def SSE_RSQRTSS : OpndItins<
3279 IIC_SSE_RSQRTSS_RR, IIC_SSE_RSQRTSS_RM
3283 let Sched = WriteFRcp in {
3284 def SSE_RCPP : OpndItins<
3285 IIC_SSE_RCPP_RR, IIC_SSE_RCPP_RM
3288 def SSE_RCPS : OpndItins<
3289 IIC_SSE_RCPS_RR, IIC_SSE_RCPS_RM
3293 /// sse_fp_unop_s - SSE1 unops in scalar form
3294 /// For the non-AVX defs, we need $src1 to be tied to $dst because
3295 /// the HW instructions are 2 operand / destructive.
3296 multiclass sse_fp_unop_s<bits<8> opc, string OpcodeStr, RegisterClass RC,
3297 ValueType vt, ValueType ScalarVT,
3298 X86MemOperand x86memop,
3300 SDNode OpNode, Domain d, OpndItins itins,
3301 Predicate target, string Suffix> {
3302 let hasSideEffects = 0 in {
3303 def r : I<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1),
3304 !strconcat(OpcodeStr, "\t{$src1, $dst|$dst, $src1}"),
3305 [(set RC:$dst, (OpNode RC:$src1))], itins.rr, d>, Sched<[itins.Sched]>,
3308 def m : I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src1),
3309 !strconcat(OpcodeStr, "\t{$src1, $dst|$dst, $src1}"),
3310 [(set RC:$dst, (OpNode (load addr:$src1)))], itins.rm, d>,
3311 Sched<[itins.Sched.Folded, ReadAfterLd]>,
3312 Requires<[target, OptForSize]>;
3314 let isCodeGenOnly = 1, Constraints = "$src1 = $dst", ExeDomain = d in {
3315 def r_Int : I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
3316 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3317 []>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3319 def m_Int : I<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, x86memop:$src2),
3320 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3321 []>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3325 let Predicates = [target] in {
3326 // These are unary operations, but they are modeled as having 2 source operands
3327 // because the high elements of the destination are unchanged in SSE.
3328 def : Pat<(Intr VR128:$src),
3329 (!cast<Instruction>(NAME#Suffix##r_Int) VR128:$src, VR128:$src)>;
3331 // We don't want to fold scalar loads into these instructions unless
3332 // optimizing for size. This is because the folded instruction will have a
3333 // partial register update, while the unfolded sequence will not, e.g.
3335 // rcpss %xmm0, %xmm0
3336 // which has a clobber before the rcp, vs.
3338 let Predicates = [target, OptForSize] in {
3339 def : Pat<(Intr (scalar_to_vector (ScalarVT (load addr:$src2)))),
3340 (!cast<Instruction>(NAME#Suffix##m_Int)
3341 (vt (IMPLICIT_DEF)), addr:$src2)>;
3345 multiclass avx_fp_unop_s<bits<8> opc, string OpcodeStr, RegisterClass RC,
3346 ValueType vt, ValueType ScalarVT,
3347 X86MemOperand x86memop,
3348 Intrinsic Intr, SDNode OpNode, Domain d,
3349 OpndItins itins, string Suffix> {
3350 let hasSideEffects = 0 in {
3351 def r : I<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
3352 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3353 [], itins.rr, d>, Sched<[itins.Sched]>;
3355 def m : I<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
3356 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3357 [], itins.rm, d>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3358 let isCodeGenOnly = 1, ExeDomain = d in {
3359 def r_Int : I<opc, MRMSrcReg, (outs VR128:$dst),
3360 (ins VR128:$src1, VR128:$src2),
3361 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3362 []>, Sched<[itins.Sched.Folded]>;
3364 def m_Int : I<opc, MRMSrcMem, (outs VR128:$dst),
3365 (ins VR128:$src1, x86memop:$src2),
3366 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3367 []>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3371 // We don't want to fold scalar loads into these instructions unless
3372 // optimizing for size. This is because the folded instruction will have a
3373 // partial register update, while the unfolded sequence will not, e.g.
3374 // vmovss mem, %xmm0
3375 // vrcpss %xmm0, %xmm0, %xmm0
3376 // which has a clobber before the rcp, vs.
3377 // vrcpss mem, %xmm0, %xmm0
3378 // TODO: In theory, we could fold the load, and avoid the stall caused by
3379 // the partial register store, either in ExecutionDepsFix or with smarter RA.
3380 let Predicates = [UseAVX] in {
3381 def : Pat<(OpNode RC:$src), (!cast<Instruction>("V"#NAME#Suffix##r)
3382 (ScalarVT (IMPLICIT_DEF)), RC:$src)>;
3384 let Predicates = [HasAVX] in {
3385 def : Pat<(Intr VR128:$src),
3386 (!cast<Instruction>("V"#NAME#Suffix##r_Int) VR128:$src,
3389 let Predicates = [HasAVX, OptForSize] in {
3390 def : Pat<(Intr (scalar_to_vector (ScalarVT (load addr:$src2)))),
3391 (!cast<Instruction>("V"#NAME#Suffix##m_Int)
3392 (vt (IMPLICIT_DEF)), addr:$src2)>;
3394 let Predicates = [UseAVX, OptForSize] in {
3395 def : Pat<(ScalarVT (OpNode (load addr:$src))),
3396 (!cast<Instruction>("V"#NAME#Suffix##m) (ScalarVT (IMPLICIT_DEF)),
3401 /// sse1_fp_unop_p - SSE1 unops in packed form.
3402 multiclass sse1_fp_unop_p<bits<8> opc, string OpcodeStr, SDNode OpNode,
3403 OpndItins itins, list<Predicate> prds> {
3404 let Predicates = prds in {
3405 def V#NAME#PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3406 !strconcat("v", OpcodeStr,
3407 "ps\t{$src, $dst|$dst, $src}"),
3408 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))],
3409 itins.rr>, VEX, Sched<[itins.Sched]>, VEX_WIG;
3410 def V#NAME#PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3411 !strconcat("v", OpcodeStr,
3412 "ps\t{$src, $dst|$dst, $src}"),
3413 [(set VR128:$dst, (OpNode (loadv4f32 addr:$src)))],
3414 itins.rm>, VEX, Sched<[itins.Sched.Folded]>, VEX_WIG;
3415 def V#NAME#PSYr : PSI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3416 !strconcat("v", OpcodeStr,
3417 "ps\t{$src, $dst|$dst, $src}"),
3418 [(set VR256:$dst, (v8f32 (OpNode VR256:$src)))],
3419 itins.rr>, VEX, VEX_L, Sched<[itins.Sched]>, VEX_WIG;
3420 def V#NAME#PSYm : PSI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
3421 !strconcat("v", OpcodeStr,
3422 "ps\t{$src, $dst|$dst, $src}"),
3423 [(set VR256:$dst, (OpNode (loadv8f32 addr:$src)))],
3424 itins.rm>, VEX, VEX_L, Sched<[itins.Sched.Folded]>, VEX_WIG;
3427 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3428 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3429 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))], itins.rr>,
3430 Sched<[itins.Sched]>;
3431 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3432 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3433 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))], itins.rm>,
3434 Sched<[itins.Sched.Folded]>;
3437 /// sse2_fp_unop_p - SSE2 unops in vector forms.
3438 multiclass sse2_fp_unop_p<bits<8> opc, string OpcodeStr,
3439 SDNode OpNode, OpndItins itins> {
3440 let Predicates = [HasAVX] in {
3441 def V#NAME#PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3442 !strconcat("v", OpcodeStr,
3443 "pd\t{$src, $dst|$dst, $src}"),
3444 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))],
3445 itins.rr>, VEX, Sched<[itins.Sched]>, VEX_WIG;
3446 def V#NAME#PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3447 !strconcat("v", OpcodeStr,
3448 "pd\t{$src, $dst|$dst, $src}"),
3449 [(set VR128:$dst, (OpNode (loadv2f64 addr:$src)))],
3450 itins.rm>, VEX, Sched<[itins.Sched.Folded]>, VEX_WIG;
3451 def V#NAME#PDYr : PDI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3452 !strconcat("v", OpcodeStr,
3453 "pd\t{$src, $dst|$dst, $src}"),
3454 [(set VR256:$dst, (v4f64 (OpNode VR256:$src)))],
3455 itins.rr>, VEX, VEX_L, Sched<[itins.Sched]>, VEX_WIG;
3456 def V#NAME#PDYm : PDI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
3457 !strconcat("v", OpcodeStr,
3458 "pd\t{$src, $dst|$dst, $src}"),
3459 [(set VR256:$dst, (OpNode (loadv4f64 addr:$src)))],
3460 itins.rm>, VEX, VEX_L, Sched<[itins.Sched.Folded]>, VEX_WIG;
3463 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3464 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
3465 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))], itins.rr>,
3466 Sched<[itins.Sched]>;
3467 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3468 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
3469 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))], itins.rm>,
3470 Sched<[itins.Sched.Folded]>;
3473 multiclass sse1_fp_unop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
3475 defm SS : sse_fp_unop_s<opc, OpcodeStr##ss, FR32, v4f32, f32, f32mem,
3476 !cast<Intrinsic>("int_x86_sse_"##OpcodeStr##_ss), OpNode,
3477 SSEPackedSingle, itins, UseSSE1, "SS">, XS;
3478 defm V#NAME#SS : avx_fp_unop_s<opc, "v"#OpcodeStr##ss, FR32, v4f32, f32,
3480 !cast<Intrinsic>("int_x86_sse_"##OpcodeStr##_ss), OpNode,
3481 SSEPackedSingle, itins, "SS">, XS, VEX_4V, VEX_LIG, VEX_WIG;
3484 multiclass sse2_fp_unop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
3486 defm SD : sse_fp_unop_s<opc, OpcodeStr##sd, FR64, v2f64, f64, f64mem,
3487 !cast<Intrinsic>("int_x86_sse2_"##OpcodeStr##_sd),
3488 OpNode, SSEPackedDouble, itins, UseSSE2, "SD">, XD;
3489 defm V#NAME#SD : avx_fp_unop_s<opc, "v"#OpcodeStr##sd, FR64, v2f64, f64,
3491 !cast<Intrinsic>("int_x86_sse2_"##OpcodeStr##_sd),
3492 OpNode, SSEPackedDouble, itins, "SD">,
3493 XD, VEX_4V, VEX_LIG, VEX_WIG;
3497 defm SQRT : sse1_fp_unop_s<0x51, "sqrt", fsqrt, SSE_SQRTSS>,
3498 sse1_fp_unop_p<0x51, "sqrt", fsqrt, SSE_SQRTPS, [HasAVX]>,
3499 sse2_fp_unop_s<0x51, "sqrt", fsqrt, SSE_SQRTSD>,
3500 sse2_fp_unop_p<0x51, "sqrt", fsqrt, SSE_SQRTPD>;
3502 // Reciprocal approximations. Note that these typically require refinement
3503 // in order to obtain suitable precision.
3504 defm RSQRT : sse1_fp_unop_s<0x52, "rsqrt", X86frsqrt, SSE_RSQRTSS>,
3505 sse1_fp_unop_p<0x52, "rsqrt", X86frsqrt, SSE_RSQRTPS, [HasAVX, NoVLX] >;
3506 defm RCP : sse1_fp_unop_s<0x53, "rcp", X86frcp, SSE_RCPS>,
3507 sse1_fp_unop_p<0x53, "rcp", X86frcp, SSE_RCPP, [HasAVX, NoVLX]>;
3509 // There is no f64 version of the reciprocal approximation instructions.
3511 // TODO: We should add *scalar* op patterns for these just like we have for
3512 // the binops above. If the binop and unop patterns could all be unified
3513 // that would be even better.
3515 multiclass scalar_unary_math_patterns<Intrinsic Intr, string OpcPrefix,
3516 SDNode Move, ValueType VT,
3517 Predicate BasePredicate> {
3518 let Predicates = [BasePredicate] in {
3519 def : Pat<(VT (Move VT:$dst, (Intr VT:$src))),
3520 (!cast<I>(OpcPrefix#r_Int) VT:$dst, VT:$src)>;
3523 // With SSE 4.1, blendi is preferred to movs*, so match that too.
3524 let Predicates = [UseSSE41] in {
3525 def : Pat<(VT (X86Blendi VT:$dst, (Intr VT:$src), (i8 1))),
3526 (!cast<I>(OpcPrefix#r_Int) VT:$dst, VT:$src)>;
3529 // Repeat for AVX versions of the instructions.
3530 let Predicates = [HasAVX] in {
3531 def : Pat<(VT (Move VT:$dst, (Intr VT:$src))),
3532 (!cast<I>("V"#OpcPrefix#r_Int) VT:$dst, VT:$src)>;
3534 def : Pat<(VT (X86Blendi VT:$dst, (Intr VT:$src), (i8 1))),
3535 (!cast<I>("V"#OpcPrefix#r_Int) VT:$dst, VT:$src)>;
3539 defm : scalar_unary_math_patterns<int_x86_sse_rcp_ss, "RCPSS", X86Movss,
3541 defm : scalar_unary_math_patterns<int_x86_sse_rsqrt_ss, "RSQRTSS", X86Movss,
3543 defm : scalar_unary_math_patterns<int_x86_sse_sqrt_ss, "SQRTSS", X86Movss,
3545 defm : scalar_unary_math_patterns<int_x86_sse2_sqrt_sd, "SQRTSD", X86Movsd,
3549 //===----------------------------------------------------------------------===//
3550 // SSE 1 & 2 - Non-temporal stores
3551 //===----------------------------------------------------------------------===//
3553 let AddedComplexity = 400 in { // Prefer non-temporal versions
3554 let SchedRW = [WriteStore] in {
3555 let Predicates = [HasAVX, NoVLX] in {
3556 def VMOVNTPSmr : VPSI<0x2B, MRMDestMem, (outs),
3557 (ins f128mem:$dst, VR128:$src),
3558 "movntps\t{$src, $dst|$dst, $src}",
3559 [(alignednontemporalstore (v4f32 VR128:$src),
3561 IIC_SSE_MOVNT>, VEX, VEX_WIG;
3562 def VMOVNTPDmr : VPDI<0x2B, MRMDestMem, (outs),
3563 (ins f128mem:$dst, VR128:$src),
3564 "movntpd\t{$src, $dst|$dst, $src}",
3565 [(alignednontemporalstore (v2f64 VR128:$src),
3567 IIC_SSE_MOVNT>, VEX, VEX_WIG;
3569 let ExeDomain = SSEPackedInt in
3570 def VMOVNTDQmr : VPDI<0xE7, MRMDestMem, (outs),
3571 (ins i128mem:$dst, VR128:$src),
3572 "movntdq\t{$src, $dst|$dst, $src}",
3573 [(alignednontemporalstore (v2i64 VR128:$src),
3575 IIC_SSE_MOVNT>, VEX, VEX_WIG;
3577 def VMOVNTPSYmr : VPSI<0x2B, MRMDestMem, (outs),
3578 (ins f256mem:$dst, VR256:$src),
3579 "movntps\t{$src, $dst|$dst, $src}",
3580 [(alignednontemporalstore (v8f32 VR256:$src),
3582 IIC_SSE_MOVNT>, VEX, VEX_L, VEX_WIG;
3583 def VMOVNTPDYmr : VPDI<0x2B, MRMDestMem, (outs),
3584 (ins f256mem:$dst, VR256:$src),
3585 "movntpd\t{$src, $dst|$dst, $src}",
3586 [(alignednontemporalstore (v4f64 VR256:$src),
3588 IIC_SSE_MOVNT>, VEX, VEX_L, VEX_WIG;
3589 let ExeDomain = SSEPackedInt in
3590 def VMOVNTDQYmr : VPDI<0xE7, MRMDestMem, (outs),
3591 (ins i256mem:$dst, VR256:$src),
3592 "movntdq\t{$src, $dst|$dst, $src}",
3593 [(alignednontemporalstore (v4i64 VR256:$src),
3595 IIC_SSE_MOVNT>, VEX, VEX_L, VEX_WIG;
3598 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3599 "movntps\t{$src, $dst|$dst, $src}",
3600 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)],
3602 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3603 "movntpd\t{$src, $dst|$dst, $src}",
3604 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)],
3607 let ExeDomain = SSEPackedInt in
3608 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3609 "movntdq\t{$src, $dst|$dst, $src}",
3610 [(alignednontemporalstore (v2i64 VR128:$src), addr:$dst)],
3613 // There is no AVX form for instructions below this point
3614 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
3615 "movnti{l}\t{$src, $dst|$dst, $src}",
3616 [(nontemporalstore (i32 GR32:$src), addr:$dst)],
3618 PS, Requires<[HasSSE2]>;
3619 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
3620 "movnti{q}\t{$src, $dst|$dst, $src}",
3621 [(nontemporalstore (i64 GR64:$src), addr:$dst)],
3623 PS, Requires<[HasSSE2]>;
3624 } // SchedRW = [WriteStore]
3626 let Predicates = [HasAVX, NoVLX] in {
3627 def : Pat<(alignednontemporalstore (v8i32 VR256:$src), addr:$dst),
3628 (VMOVNTDQYmr addr:$dst, VR256:$src)>;
3629 def : Pat<(alignednontemporalstore (v16i16 VR256:$src), addr:$dst),
3630 (VMOVNTDQYmr addr:$dst, VR256:$src)>;
3631 def : Pat<(alignednontemporalstore (v32i8 VR256:$src), addr:$dst),
3632 (VMOVNTDQYmr addr:$dst, VR256:$src)>;
3634 def : Pat<(alignednontemporalstore (v4i32 VR128:$src), addr:$dst),
3635 (VMOVNTDQmr addr:$dst, VR128:$src)>;
3636 def : Pat<(alignednontemporalstore (v8i16 VR128:$src), addr:$dst),
3637 (VMOVNTDQmr addr:$dst, VR128:$src)>;
3638 def : Pat<(alignednontemporalstore (v16i8 VR128:$src), addr:$dst),
3639 (VMOVNTDQmr addr:$dst, VR128:$src)>;
3642 let Predicates = [UseSSE2] in {
3643 def : Pat<(alignednontemporalstore (v4i32 VR128:$src), addr:$dst),
3644 (MOVNTDQmr addr:$dst, VR128:$src)>;
3645 def : Pat<(alignednontemporalstore (v8i16 VR128:$src), addr:$dst),
3646 (MOVNTDQmr addr:$dst, VR128:$src)>;
3647 def : Pat<(alignednontemporalstore (v16i8 VR128:$src), addr:$dst),
3648 (MOVNTDQmr addr:$dst, VR128:$src)>;
3651 } // AddedComplexity
3653 //===----------------------------------------------------------------------===//
3654 // SSE 1 & 2 - Prefetch and memory fence
3655 //===----------------------------------------------------------------------===//
3657 // Prefetch intrinsic.
3658 let Predicates = [HasSSE1], SchedRW = [WriteLoad] in {
3659 def PREFETCHT0 : I<0x18, MRM1m, (outs), (ins i8mem:$src),
3660 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3), (i32 1))],
3661 IIC_SSE_PREFETCH>, TB;
3662 def PREFETCHT1 : I<0x18, MRM2m, (outs), (ins i8mem:$src),
3663 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2), (i32 1))],
3664 IIC_SSE_PREFETCH>, TB;
3665 def PREFETCHT2 : I<0x18, MRM3m, (outs), (ins i8mem:$src),
3666 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1), (i32 1))],
3667 IIC_SSE_PREFETCH>, TB;
3668 def PREFETCHNTA : I<0x18, MRM0m, (outs), (ins i8mem:$src),
3669 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0), (i32 1))],
3670 IIC_SSE_PREFETCH>, TB;
3673 // FIXME: How should flush instruction be modeled?
3674 let SchedRW = [WriteLoad] in {
3676 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
3677 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)],
3678 IIC_SSE_PREFETCH>, PS, Requires<[HasSSE2]>;
3681 let SchedRW = [WriteNop] in {
3682 // Pause. This "instruction" is encoded as "rep; nop", so even though it
3683 // was introduced with SSE2, it's backward compatible.
3684 def PAUSE : I<0x90, RawFrm, (outs), (ins),
3685 "pause", [(int_x86_sse2_pause)], IIC_SSE_PAUSE>,
3686 OBXS, Requires<[HasSSE2]>;
3689 let SchedRW = [WriteFence] in {
3690 // Load, store, and memory fence
3691 // TODO: As with mfence, we may want to ease the availablity of sfence/lfence
3692 // to include any 64-bit target.
3693 def SFENCE : I<0xAE, MRM_F8, (outs), (ins),
3694 "sfence", [(int_x86_sse_sfence)], IIC_SSE_SFENCE>,
3695 PS, Requires<[HasSSE1]>;
3696 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
3697 "lfence", [(int_x86_sse2_lfence)], IIC_SSE_LFENCE>,
3698 TB, Requires<[HasSSE2]>;
3699 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
3700 "mfence", [(int_x86_sse2_mfence)], IIC_SSE_MFENCE>,
3701 TB, Requires<[HasMFence]>;
3704 def : Pat<(X86MFence), (MFENCE)>;
3706 //===----------------------------------------------------------------------===//
3707 // SSE 1 & 2 - Load/Store XCSR register
3708 //===----------------------------------------------------------------------===//
3710 def VLDMXCSR : VPSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
3711 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)],
3712 IIC_SSE_LDMXCSR>, VEX, Sched<[WriteLoad]>, VEX_WIG;
3713 def VSTMXCSR : VPSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
3714 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)],
3715 IIC_SSE_STMXCSR>, VEX, Sched<[WriteStore]>, VEX_WIG;
3717 def LDMXCSR : I<0xAE, MRM2m, (outs), (ins i32mem:$src),
3718 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)],
3719 IIC_SSE_LDMXCSR>, TB, Sched<[WriteLoad]>;
3720 def STMXCSR : I<0xAE, MRM3m, (outs), (ins i32mem:$dst),
3721 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)],
3722 IIC_SSE_STMXCSR>, TB, Sched<[WriteStore]>;
3724 //===---------------------------------------------------------------------===//
3725 // SSE2 - Move Aligned/Unaligned Packed Integer Instructions
3726 //===---------------------------------------------------------------------===//
3728 let ExeDomain = SSEPackedInt in { // SSE integer instructions
3730 let hasSideEffects = 0, SchedRW = [WriteMove] in {
3731 def VMOVDQArr : VPDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3732 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>,
3734 def VMOVDQAYrr : VPDI<0x6F, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3735 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>,
3736 VEX, VEX_L, VEX_WIG;
3737 def VMOVDQUrr : VSSI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3738 "movdqu\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVU_P_RR>,
3740 def VMOVDQUYrr : VSSI<0x6F, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3741 "movdqu\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVU_P_RR>,
3742 VEX, VEX_L, VEX_WIG;
3746 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
3747 SchedRW = [WriteMove] in {
3748 def VMOVDQArr_REV : VPDI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3749 "movdqa\t{$src, $dst|$dst, $src}", [],
3752 def VMOVDQAYrr_REV : VPDI<0x7F, MRMDestReg, (outs VR256:$dst), (ins VR256:$src),
3753 "movdqa\t{$src, $dst|$dst, $src}", [],
3754 IIC_SSE_MOVA_P_RR>, VEX, VEX_L, VEX_WIG;
3755 def VMOVDQUrr_REV : VSSI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3756 "movdqu\t{$src, $dst|$dst, $src}", [],
3759 def VMOVDQUYrr_REV : VSSI<0x7F, MRMDestReg, (outs VR256:$dst), (ins VR256:$src),
3760 "movdqu\t{$src, $dst|$dst, $src}", [],
3761 IIC_SSE_MOVU_P_RR>, VEX, VEX_L, VEX_WIG;
3764 let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1,
3765 hasSideEffects = 0, SchedRW = [WriteLoad] in {
3766 let Predicates = [HasAVX,NoVLX] in
3767 def VMOVDQArm : VPDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3768 "movdqa\t{$src, $dst|$dst, $src}",
3769 [(set VR128:$dst, (alignedloadv2i64 addr:$src))],
3770 IIC_SSE_MOVA_P_RM>, VEX, VEX_WIG;
3771 def VMOVDQAYrm : VPDI<0x6F, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
3772 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RM>,
3773 VEX, VEX_L, VEX_WIG;
3774 let Predicates = [HasAVX,NoVLX] in
3775 def VMOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3776 "vmovdqu\t{$src, $dst|$dst, $src}",
3777 [(set VR128:$dst, (loadv2i64 addr:$src))],
3778 IIC_SSE_MOVU_P_RM>, XS, VEX, VEX_WIG;
3779 def VMOVDQUYrm : I<0x6F, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
3780 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_RM>,
3781 XS, VEX, VEX_L, VEX_WIG;
3784 let mayStore = 1, hasSideEffects = 0, SchedRW = [WriteStore] in {
3785 let Predicates = [HasAVX,NoVLX] in
3786 def VMOVDQAmr : VPDI<0x7F, MRMDestMem, (outs),
3787 (ins i128mem:$dst, VR128:$src),
3788 "movdqa\t{$src, $dst|$dst, $src}",
3789 [(alignedstore (v2i64 VR128:$src), addr:$dst)],
3790 IIC_SSE_MOVA_P_MR>, VEX, VEX_WIG;
3791 def VMOVDQAYmr : VPDI<0x7F, MRMDestMem, (outs),
3792 (ins i256mem:$dst, VR256:$src),
3793 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_MR>,
3794 VEX, VEX_L, VEX_WIG;
3795 let Predicates = [HasAVX,NoVLX] in
3796 def VMOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3797 "vmovdqu\t{$src, $dst|$dst, $src}",
3798 [(store (v2i64 VR128:$src), addr:$dst)], IIC_SSE_MOVU_P_MR>,
3800 def VMOVDQUYmr : I<0x7F, MRMDestMem, (outs), (ins i256mem:$dst, VR256:$src),
3801 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_MR>,
3802 XS, VEX, VEX_L, VEX_WIG;
3805 let SchedRW = [WriteMove] in {
3806 let hasSideEffects = 0 in {
3807 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3808 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>;
3810 def MOVDQUrr : I<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3811 "movdqu\t{$src, $dst|$dst, $src}",
3812 [], IIC_SSE_MOVU_P_RR>, XS, Requires<[UseSSE2]>;
3816 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in {
3817 def MOVDQArr_REV : PDI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3818 "movdqa\t{$src, $dst|$dst, $src}", [],
3821 def MOVDQUrr_REV : I<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3822 "movdqu\t{$src, $dst|$dst, $src}",
3823 [], IIC_SSE_MOVU_P_RR>, XS, Requires<[UseSSE2]>;
3827 let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1,
3828 hasSideEffects = 0, SchedRW = [WriteLoad] in {
3829 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3830 "movdqa\t{$src, $dst|$dst, $src}",
3831 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/],
3833 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3834 "movdqu\t{$src, $dst|$dst, $src}",
3835 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/],
3837 XS, Requires<[UseSSE2]>;
3840 let mayStore = 1, hasSideEffects = 0, SchedRW = [WriteStore] in {
3841 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3842 "movdqa\t{$src, $dst|$dst, $src}",
3843 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/],
3845 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3846 "movdqu\t{$src, $dst|$dst, $src}",
3847 [/*(store (v2i64 VR128:$src), addr:$dst)*/],
3849 XS, Requires<[UseSSE2]>;
3852 } // ExeDomain = SSEPackedInt
3854 // Aliases to help the assembler pick two byte VEX encodings by swapping the
3855 // operands relative to the normal instructions to use VEX.R instead of VEX.B.
3856 def : InstAlias<"vmovdqa\t{$src, $dst|$dst, $src}",
3857 (VMOVDQArr_REV VR128L:$dst, VR128H:$src), 0>;
3858 def : InstAlias<"vmovdqa\t{$src, $dst|$dst, $src}",
3859 (VMOVDQAYrr_REV VR256L:$dst, VR256H:$src), 0>;
3860 def : InstAlias<"vmovdqu\t{$src, $dst|$dst, $src}",
3861 (VMOVDQUrr_REV VR128L:$dst, VR128H:$src), 0>;
3862 def : InstAlias<"vmovdqu\t{$src, $dst|$dst, $src}",
3863 (VMOVDQUYrr_REV VR256L:$dst, VR256H:$src), 0>;
3865 let Predicates = [HasAVX, NoVLX] in {
3866 // Additional patterns for other integer sizes.
3867 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3868 (VMOVDQAmr addr:$dst, VR128:$src)>;
3869 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3870 (VMOVDQAmr addr:$dst, VR128:$src)>;
3871 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3872 (VMOVDQAmr addr:$dst, VR128:$src)>;
3873 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3874 (VMOVDQUmr addr:$dst, VR128:$src)>;
3875 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3876 (VMOVDQUmr addr:$dst, VR128:$src)>;
3877 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3878 (VMOVDQUmr addr:$dst, VR128:$src)>;
3880 // Special patterns for storing subvector extracts of lower 128-bits
3881 // Its cheaper to just use VMOVDQA/VMOVDQU instead of VEXTRACTF128mr
3882 def : Pat<(alignedstore (v2i64 (extract_subvector
3883 (v4i64 VR256:$src), (iPTR 0))), addr:$dst),
3884 (VMOVDQAmr addr:$dst, (v2i64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3885 def : Pat<(alignedstore (v4i32 (extract_subvector
3886 (v8i32 VR256:$src), (iPTR 0))), addr:$dst),
3887 (VMOVDQAmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3888 def : Pat<(alignedstore (v8i16 (extract_subvector
3889 (v16i16 VR256:$src), (iPTR 0))), addr:$dst),
3890 (VMOVDQAmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3891 def : Pat<(alignedstore (v16i8 (extract_subvector
3892 (v32i8 VR256:$src), (iPTR 0))), addr:$dst),
3893 (VMOVDQAmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3895 def : Pat<(store (v2i64 (extract_subvector
3896 (v4i64 VR256:$src), (iPTR 0))), addr:$dst),
3897 (VMOVDQUmr addr:$dst, (v2i64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3898 def : Pat<(store (v4i32 (extract_subvector
3899 (v8i32 VR256:$src), (iPTR 0))), addr:$dst),
3900 (VMOVDQUmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3901 def : Pat<(store (v8i16 (extract_subvector
3902 (v16i16 VR256:$src), (iPTR 0))), addr:$dst),
3903 (VMOVDQUmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3904 def : Pat<(store (v16i8 (extract_subvector
3905 (v32i8 VR256:$src), (iPTR 0))), addr:$dst),
3906 (VMOVDQUmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3909 //===---------------------------------------------------------------------===//
3910 // SSE2 - Packed Integer Arithmetic Instructions
3911 //===---------------------------------------------------------------------===//
3913 let Sched = WriteVecIMul in
3914 def SSE_PMADD : OpndItins<
3915 IIC_SSE_PMADD, IIC_SSE_PMADD
3918 let ExeDomain = SSEPackedInt in { // SSE integer instructions
3920 /// PDI_binop_rm2 - Simple SSE2 binary operator with different src and dst types
3921 multiclass PDI_binop_rm2<bits<8> opc, string OpcodeStr, SDNode OpNode,
3922 ValueType DstVT, ValueType SrcVT, RegisterClass RC,
3923 PatFrag memop_frag, X86MemOperand x86memop,
3924 OpndItins itins, bit Is2Addr = 1> {
3925 let isCommutable = 1 in
3926 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
3927 (ins RC:$src1, RC:$src2),
3929 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3930 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3931 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1), RC:$src2)))]>,
3932 Sched<[itins.Sched]>;
3933 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
3934 (ins RC:$src1, x86memop:$src2),
3936 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3937 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3938 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1),
3939 (bitconvert (memop_frag addr:$src2)))))]>,
3940 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3942 } // ExeDomain = SSEPackedInt
3944 defm PADDB : PDI_binop_all<0xFC, "paddb", add, v16i8, v32i8,
3945 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3946 defm PADDW : PDI_binop_all<0xFD, "paddw", add, v8i16, v16i16,
3947 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3948 defm PADDD : PDI_binop_all<0xFE, "paddd", add, v4i32, v8i32,
3949 SSE_INTALU_ITINS_P, 1, NoVLX>;
3950 defm PADDQ : PDI_binop_all<0xD4, "paddq", add, v2i64, v4i64,
3951 SSE_INTALUQ_ITINS_P, 1, NoVLX>;
3952 defm PADDSB : PDI_binop_all<0xEC, "paddsb", X86adds, v16i8, v32i8,
3953 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3954 defm PADDSW : PDI_binop_all<0xED, "paddsw", X86adds, v8i16, v16i16,
3955 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3956 defm PADDUSB : PDI_binop_all<0xDC, "paddusb", X86addus, v16i8, v32i8,
3957 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3958 defm PADDUSW : PDI_binop_all<0xDD, "paddusw", X86addus, v8i16, v16i16,
3959 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3960 defm PMULLW : PDI_binop_all<0xD5, "pmullw", mul, v8i16, v16i16,
3961 SSE_INTMUL_ITINS_P, 1, NoVLX_Or_NoBWI>;
3962 defm PMULHUW : PDI_binop_all<0xE4, "pmulhuw", mulhu, v8i16, v16i16,
3963 SSE_INTMUL_ITINS_P, 1, NoVLX_Or_NoBWI>;
3964 defm PMULHW : PDI_binop_all<0xE5, "pmulhw", mulhs, v8i16, v16i16,
3965 SSE_INTMUL_ITINS_P, 1, NoVLX_Or_NoBWI>;
3966 defm PSUBB : PDI_binop_all<0xF8, "psubb", sub, v16i8, v32i8,
3967 SSE_INTALU_ITINS_P, 0, NoVLX_Or_NoBWI>;
3968 defm PSUBW : PDI_binop_all<0xF9, "psubw", sub, v8i16, v16i16,
3969 SSE_INTALU_ITINS_P, 0, NoVLX_Or_NoBWI>;
3970 defm PSUBD : PDI_binop_all<0xFA, "psubd", sub, v4i32, v8i32,
3971 SSE_INTALU_ITINS_P, 0, NoVLX>;
3972 defm PSUBQ : PDI_binop_all<0xFB, "psubq", sub, v2i64, v4i64,
3973 SSE_INTALUQ_ITINS_P, 0, NoVLX>;
3974 defm PSUBSB : PDI_binop_all<0xE8, "psubsb", X86subs, v16i8, v32i8,
3975 SSE_INTALU_ITINS_P, 0, NoVLX_Or_NoBWI>;
3976 defm PSUBSW : PDI_binop_all<0xE9, "psubsw", X86subs, v8i16, v16i16,
3977 SSE_INTALU_ITINS_P, 0, NoVLX_Or_NoBWI>;
3978 defm PSUBUSB : PDI_binop_all<0xD8, "psubusb", X86subus, v16i8, v32i8,
3979 SSE_INTALU_ITINS_P, 0, NoVLX_Or_NoBWI>;
3980 defm PSUBUSW : PDI_binop_all<0xD9, "psubusw", X86subus, v8i16, v16i16,
3981 SSE_INTALU_ITINS_P, 0, NoVLX_Or_NoBWI>;
3982 defm PMINUB : PDI_binop_all<0xDA, "pminub", umin, v16i8, v32i8,
3983 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3984 defm PMINSW : PDI_binop_all<0xEA, "pminsw", smin, v8i16, v16i16,
3985 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3986 defm PMAXUB : PDI_binop_all<0xDE, "pmaxub", umax, v16i8, v32i8,
3987 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3988 defm PMAXSW : PDI_binop_all<0xEE, "pmaxsw", smax, v8i16, v16i16,
3989 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3990 defm PAVGB : PDI_binop_all<0xE0, "pavgb", X86avg, v16i8, v32i8,
3991 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3992 defm PAVGW : PDI_binop_all<0xE3, "pavgw", X86avg, v8i16, v16i16,
3993 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3995 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in
3996 defm VPMADDWD : PDI_binop_rm2<0xF5, "vpmaddwd", X86vpmaddwd, v4i32, v8i16, VR128,
3997 loadv2i64, i128mem, SSE_PMADD, 0>, VEX_4V, VEX_WIG;
3999 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in
4000 defm VPMADDWDY : PDI_binop_rm2<0xF5, "vpmaddwd", X86vpmaddwd, v8i32, v16i16,
4001 VR256, loadv4i64, i256mem, SSE_PMADD,
4002 0>, VEX_4V, VEX_L, VEX_WIG;
4003 let Constraints = "$src1 = $dst" in
4004 defm PMADDWD : PDI_binop_rm2<0xF5, "pmaddwd", X86vpmaddwd, v4i32, v8i16, VR128,
4005 memopv2i64, i128mem, SSE_PMADD>;
4007 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in
4008 defm VPSADBW : PDI_binop_rm2<0xF6, "vpsadbw", X86psadbw, v2i64, v16i8, VR128,
4009 loadv2i64, i128mem, SSE_INTMUL_ITINS_P, 0>,
4011 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in
4012 defm VPSADBWY : PDI_binop_rm2<0xF6, "vpsadbw", X86psadbw, v4i64, v32i8, VR256,
4013 loadv4i64, i256mem, SSE_INTMUL_ITINS_P, 0>,
4014 VEX_4V, VEX_L, VEX_WIG;
4015 let Constraints = "$src1 = $dst" in
4016 defm PSADBW : PDI_binop_rm2<0xF6, "psadbw", X86psadbw, v2i64, v16i8, VR128,
4017 memopv2i64, i128mem, SSE_INTALU_ITINS_P>;
4019 let Predicates = [HasAVX, NoVLX] in
4020 defm VPMULUDQ : PDI_binop_rm2<0xF4, "vpmuludq", X86pmuludq, v2i64, v4i32, VR128,
4021 loadv2i64, i128mem, SSE_INTMUL_ITINS_P, 0>,
4023 let Predicates = [HasAVX2, NoVLX] in
4024 defm VPMULUDQY : PDI_binop_rm2<0xF4, "vpmuludq", X86pmuludq, v4i64, v8i32,
4025 VR256, loadv4i64, i256mem,
4026 SSE_INTMUL_ITINS_P, 0>, VEX_4V, VEX_L, VEX_WIG;
4027 let Constraints = "$src1 = $dst" in
4028 defm PMULUDQ : PDI_binop_rm2<0xF4, "pmuludq", X86pmuludq, v2i64, v4i32, VR128,
4029 memopv2i64, i128mem, SSE_INTMUL_ITINS_P>;
4031 //===---------------------------------------------------------------------===//
4032 // SSE2 - Packed Integer Logical Instructions
4033 //===---------------------------------------------------------------------===//
4035 multiclass PDI_binop_rmi<bits<8> opc, bits<8> opc2, Format ImmForm,
4036 string OpcodeStr, SDNode OpNode,
4037 SDNode OpNode2, RegisterClass RC,
4038 ValueType DstVT, ValueType SrcVT,
4039 PatFrag ld_frag, bit Is2Addr = 1> {
4040 // src2 is always 128-bit
4041 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
4042 (ins RC:$src1, VR128:$src2),
4044 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4045 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4046 [(set RC:$dst, (DstVT (OpNode RC:$src1, (SrcVT VR128:$src2))))],
4047 SSE_INTSHIFT_ITINS_P.rr>, Sched<[WriteVecShift]>;
4048 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
4049 (ins RC:$src1, i128mem:$src2),
4051 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4052 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4053 [(set RC:$dst, (DstVT (OpNode RC:$src1,
4054 (SrcVT (bitconvert (ld_frag addr:$src2))))))],
4055 SSE_INTSHIFT_ITINS_P.rm>, Sched<[WriteVecShiftLd, ReadAfterLd]>;
4056 def ri : PDIi8<opc2, ImmForm, (outs RC:$dst),
4057 (ins RC:$src1, u8imm:$src2),
4059 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4060 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4061 [(set RC:$dst, (DstVT (OpNode2 RC:$src1, (i8 imm:$src2))))],
4062 SSE_INTSHIFT_ITINS_P.ri>, Sched<[WriteVecShift]>;
4065 multiclass PDI_binop_rmi_all<bits<8> opc, bits<8> opc2, Format ImmForm,
4066 string OpcodeStr, SDNode OpNode,
4067 SDNode OpNode2, ValueType DstVT128,
4068 ValueType DstVT256, ValueType SrcVT,
4070 let Predicates = [HasAVX, prd] in
4071 defm V#NAME : PDI_binop_rmi<opc, opc2, ImmForm, !strconcat("v", OpcodeStr),
4072 OpNode, OpNode2, VR128, DstVT128, SrcVT,
4073 loadv2i64, 0>, VEX_4V, VEX_WIG;
4074 let Predicates = [HasAVX2, prd] in
4075 defm V#NAME#Y : PDI_binop_rmi<opc, opc2, ImmForm, !strconcat("v", OpcodeStr),
4076 OpNode, OpNode2, VR256, DstVT256, SrcVT,
4077 loadv2i64, 0>, VEX_4V, VEX_L, VEX_WIG;
4078 let Constraints = "$src1 = $dst" in
4079 defm NAME : PDI_binop_rmi<opc, opc2, ImmForm, OpcodeStr, OpNode, OpNode2,
4080 VR128, DstVT128, SrcVT, memopv2i64>;
4083 multiclass PDI_binop_ri<bits<8> opc, Format ImmForm, string OpcodeStr,
4084 SDNode OpNode, RegisterClass RC, ValueType VT,
4086 def ri : PDIi8<opc, ImmForm, (outs RC:$dst), (ins RC:$src1, u8imm:$src2),
4088 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4089 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4090 [(set RC:$dst, (VT (OpNode RC:$src1, (i8 imm:$src2))))],
4091 IIC_SSE_INTSHDQ_P_RI>, Sched<[WriteVecShift]>;
4094 multiclass PDI_binop_ri_all<bits<8> opc, Format ImmForm, string OpcodeStr,
4096 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in
4097 defm V#NAME : PDI_binop_ri<opc, ImmForm, !strconcat("v", OpcodeStr), OpNode,
4098 VR128, v16i8, 0>, VEX_4V, VEX_WIG;
4099 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in
4100 defm V#NAME#Y : PDI_binop_ri<opc, ImmForm, !strconcat("v", OpcodeStr), OpNode,
4101 VR256, v32i8, 0>, VEX_4V, VEX_L, VEX_WIG;
4102 let Constraints = "$src1 = $dst" in
4103 defm NAME : PDI_binop_ri<opc, ImmForm, OpcodeStr, OpNode, VR128, v16i8>;
4106 let ExeDomain = SSEPackedInt in {
4107 defm PSLLW : PDI_binop_rmi_all<0xF1, 0x71, MRM6r, "psllw", X86vshl, X86vshli,
4108 v8i16, v16i16, v8i16, NoVLX_Or_NoBWI>;
4109 defm PSLLD : PDI_binop_rmi_all<0xF2, 0x72, MRM6r, "pslld", X86vshl, X86vshli,
4110 v4i32, v8i32, v4i32, NoVLX>;
4111 defm PSLLQ : PDI_binop_rmi_all<0xF3, 0x73, MRM6r, "psllq", X86vshl, X86vshli,
4112 v2i64, v4i64, v2i64, NoVLX>;
4114 defm PSRLW : PDI_binop_rmi_all<0xD1, 0x71, MRM2r, "psrlw", X86vsrl, X86vsrli,
4115 v8i16, v16i16, v8i16, NoVLX_Or_NoBWI>;
4116 defm PSRLD : PDI_binop_rmi_all<0xD2, 0x72, MRM2r, "psrld", X86vsrl, X86vsrli,
4117 v4i32, v8i32, v4i32, NoVLX>;
4118 defm PSRLQ : PDI_binop_rmi_all<0xD3, 0x73, MRM2r, "psrlq", X86vsrl, X86vsrli,
4119 v2i64, v4i64, v2i64, NoVLX>;
4121 defm PSRAW : PDI_binop_rmi_all<0xE1, 0x71, MRM4r, "psraw", X86vsra, X86vsrai,
4122 v8i16, v16i16, v8i16, NoVLX_Or_NoBWI>;
4123 defm PSRAD : PDI_binop_rmi_all<0xE2, 0x72, MRM4r, "psrad", X86vsra, X86vsrai,
4124 v4i32, v8i32, v4i32, NoVLX>;
4126 defm PSLLDQ : PDI_binop_ri_all<0x73, MRM7r, "pslldq", X86vshldq>;
4127 defm PSRLDQ : PDI_binop_ri_all<0x73, MRM3r, "psrldq", X86vshrdq>;
4128 // PSRADQri doesn't exist in SSE[1-3].
4129 } // ExeDomain = SSEPackedInt
4131 //===---------------------------------------------------------------------===//
4132 // SSE2 - Packed Integer Comparison Instructions
4133 //===---------------------------------------------------------------------===//
4135 defm PCMPEQB : PDI_binop_all<0x74, "pcmpeqb", X86pcmpeq, v16i8, v32i8,
4136 SSE_INTALU_ITINS_P, 1, TruePredicate>;
4137 defm PCMPEQW : PDI_binop_all<0x75, "pcmpeqw", X86pcmpeq, v8i16, v16i16,
4138 SSE_INTALU_ITINS_P, 1, TruePredicate>;
4139 defm PCMPEQD : PDI_binop_all<0x76, "pcmpeqd", X86pcmpeq, v4i32, v8i32,
4140 SSE_INTALU_ITINS_P, 1, TruePredicate>;
4141 defm PCMPGTB : PDI_binop_all<0x64, "pcmpgtb", X86pcmpgt, v16i8, v32i8,
4142 SSE_INTALU_ITINS_P, 0, TruePredicate>;
4143 defm PCMPGTW : PDI_binop_all<0x65, "pcmpgtw", X86pcmpgt, v8i16, v16i16,
4144 SSE_INTALU_ITINS_P, 0, TruePredicate>;
4145 defm PCMPGTD : PDI_binop_all<0x66, "pcmpgtd", X86pcmpgt, v4i32, v8i32,
4146 SSE_INTALU_ITINS_P, 0, TruePredicate>;
4148 //===---------------------------------------------------------------------===//
4149 // SSE2 - Packed Integer Shuffle Instructions
4150 //===---------------------------------------------------------------------===//
4152 let ExeDomain = SSEPackedInt in {
4153 multiclass sse2_pshuffle<string OpcodeStr, ValueType vt128, ValueType vt256,
4154 SDNode OpNode, Predicate prd> {
4155 let Predicates = [HasAVX, prd] in {
4156 def V#NAME#ri : Ii8<0x70, MRMSrcReg, (outs VR128:$dst),
4157 (ins VR128:$src1, u8imm:$src2),
4158 !strconcat("v", OpcodeStr,
4159 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4161 (vt128 (OpNode VR128:$src1, (i8 imm:$src2))))],
4162 IIC_SSE_PSHUF_RI>, VEX, Sched<[WriteShuffle]>, VEX_WIG;
4163 def V#NAME#mi : Ii8<0x70, MRMSrcMem, (outs VR128:$dst),
4164 (ins i128mem:$src1, u8imm:$src2),
4165 !strconcat("v", OpcodeStr,
4166 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4168 (vt128 (OpNode (bitconvert (loadv2i64 addr:$src1)),
4169 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>, VEX,
4170 Sched<[WriteShuffleLd]>, VEX_WIG;
4173 let Predicates = [HasAVX2, prd] in {
4174 def V#NAME#Yri : Ii8<0x70, MRMSrcReg, (outs VR256:$dst),
4175 (ins VR256:$src1, u8imm:$src2),
4176 !strconcat("v", OpcodeStr,
4177 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4179 (vt256 (OpNode VR256:$src1, (i8 imm:$src2))))],
4180 IIC_SSE_PSHUF_RI>, VEX, VEX_L, Sched<[WriteShuffle]>, VEX_WIG;
4181 def V#NAME#Ymi : Ii8<0x70, MRMSrcMem, (outs VR256:$dst),
4182 (ins i256mem:$src1, u8imm:$src2),
4183 !strconcat("v", OpcodeStr,
4184 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4186 (vt256 (OpNode (bitconvert (loadv4i64 addr:$src1)),
4187 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>, VEX, VEX_L,
4188 Sched<[WriteShuffleLd]>, VEX_WIG;
4191 let Predicates = [UseSSE2] in {
4192 def ri : Ii8<0x70, MRMSrcReg,
4193 (outs VR128:$dst), (ins VR128:$src1, u8imm:$src2),
4194 !strconcat(OpcodeStr,
4195 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4197 (vt128 (OpNode VR128:$src1, (i8 imm:$src2))))],
4198 IIC_SSE_PSHUF_RI>, Sched<[WriteShuffle]>;
4199 def mi : Ii8<0x70, MRMSrcMem,
4200 (outs VR128:$dst), (ins i128mem:$src1, u8imm:$src2),
4201 !strconcat(OpcodeStr,
4202 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4204 (vt128 (OpNode (bitconvert (memopv2i64 addr:$src1)),
4205 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>,
4206 Sched<[WriteShuffleLd, ReadAfterLd]>;
4209 } // ExeDomain = SSEPackedInt
4211 defm PSHUFD : sse2_pshuffle<"pshufd", v4i32, v8i32, X86PShufd, NoVLX>, PD;
4212 defm PSHUFHW : sse2_pshuffle<"pshufhw", v8i16, v16i16, X86PShufhw,
4213 NoVLX_Or_NoBWI>, XS;
4214 defm PSHUFLW : sse2_pshuffle<"pshuflw", v8i16, v16i16, X86PShuflw,
4215 NoVLX_Or_NoBWI>, XD;
4217 //===---------------------------------------------------------------------===//
4218 // Packed Integer Pack Instructions (SSE & AVX)
4219 //===---------------------------------------------------------------------===//
4221 let ExeDomain = SSEPackedInt in {
4222 multiclass sse2_pack<bits<8> opc, string OpcodeStr, ValueType OutVT,
4223 ValueType ArgVT, SDNode OpNode, PatFrag ld_frag,
4225 def rr : PDI<opc, MRMSrcReg,
4226 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4228 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4229 !strconcat(OpcodeStr,
4230 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4232 (OutVT (OpNode (ArgVT VR128:$src1), VR128:$src2)))]>,
4233 Sched<[WriteShuffle]>;
4234 def rm : PDI<opc, MRMSrcMem,
4235 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4237 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4238 !strconcat(OpcodeStr,
4239 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4241 (OutVT (OpNode (ArgVT VR128:$src1),
4242 (bitconvert (ld_frag addr:$src2)))))]>,
4243 Sched<[WriteShuffleLd, ReadAfterLd]>;
4246 multiclass sse2_pack_y<bits<8> opc, string OpcodeStr, ValueType OutVT,
4247 ValueType ArgVT, SDNode OpNode> {
4248 def Yrr : PDI<opc, MRMSrcReg,
4249 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4250 !strconcat(OpcodeStr,
4251 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4253 (OutVT (OpNode (ArgVT VR256:$src1), VR256:$src2)))]>,
4254 Sched<[WriteShuffle]>;
4255 def Yrm : PDI<opc, MRMSrcMem,
4256 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4257 !strconcat(OpcodeStr,
4258 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4260 (OutVT (OpNode (ArgVT VR256:$src1),
4261 (bitconvert (loadv4i64 addr:$src2)))))]>,
4262 Sched<[WriteShuffleLd, ReadAfterLd]>;
4265 multiclass sse4_pack<bits<8> opc, string OpcodeStr, ValueType OutVT,
4266 ValueType ArgVT, SDNode OpNode, PatFrag ld_frag,
4268 def rr : SS48I<opc, MRMSrcReg,
4269 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4271 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4272 !strconcat(OpcodeStr,
4273 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4275 (OutVT (OpNode (ArgVT VR128:$src1), VR128:$src2)))]>,
4276 Sched<[WriteShuffle]>;
4277 def rm : SS48I<opc, MRMSrcMem,
4278 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4280 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4281 !strconcat(OpcodeStr,
4282 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4284 (OutVT (OpNode (ArgVT VR128:$src1),
4285 (bitconvert (ld_frag addr:$src2)))))]>,
4286 Sched<[WriteShuffleLd, ReadAfterLd]>;
4289 multiclass sse4_pack_y<bits<8> opc, string OpcodeStr, ValueType OutVT,
4290 ValueType ArgVT, SDNode OpNode> {
4291 def Yrr : SS48I<opc, MRMSrcReg,
4292 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4293 !strconcat(OpcodeStr,
4294 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4296 (OutVT (OpNode (ArgVT VR256:$src1), VR256:$src2)))]>,
4297 Sched<[WriteShuffle]>;
4298 def Yrm : SS48I<opc, MRMSrcMem,
4299 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4300 !strconcat(OpcodeStr,
4301 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4303 (OutVT (OpNode (ArgVT VR256:$src1),
4304 (bitconvert (loadv4i64 addr:$src2)))))]>,
4305 Sched<[WriteShuffleLd, ReadAfterLd]>;
4308 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
4309 defm VPACKSSWB : sse2_pack<0x63, "vpacksswb", v16i8, v8i16, X86Packss,
4310 loadv2i64, 0>, VEX_4V, VEX_WIG;
4311 defm VPACKSSDW : sse2_pack<0x6B, "vpackssdw", v8i16, v4i32, X86Packss,
4312 loadv2i64, 0>, VEX_4V, VEX_WIG;
4314 defm VPACKUSWB : sse2_pack<0x67, "vpackuswb", v16i8, v8i16, X86Packus,
4315 loadv2i64, 0>, VEX_4V, VEX_WIG;
4316 defm VPACKUSDW : sse4_pack<0x2B, "vpackusdw", v8i16, v4i32, X86Packus,
4317 loadv2i64, 0>, VEX_4V;
4320 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
4321 defm VPACKSSWB : sse2_pack_y<0x63, "vpacksswb", v32i8, v16i16, X86Packss>,
4322 VEX_4V, VEX_L, VEX_WIG;
4323 defm VPACKSSDW : sse2_pack_y<0x6B, "vpackssdw", v16i16, v8i32, X86Packss>,
4324 VEX_4V, VEX_L, VEX_WIG;
4326 defm VPACKUSWB : sse2_pack_y<0x67, "vpackuswb", v32i8, v16i16, X86Packus>,
4327 VEX_4V, VEX_L, VEX_WIG;
4328 defm VPACKUSDW : sse4_pack_y<0x2B, "vpackusdw", v16i16, v8i32, X86Packus>,
4332 let Constraints = "$src1 = $dst" in {
4333 defm PACKSSWB : sse2_pack<0x63, "packsswb", v16i8, v8i16, X86Packss,
4335 defm PACKSSDW : sse2_pack<0x6B, "packssdw", v8i16, v4i32, X86Packss,
4338 defm PACKUSWB : sse2_pack<0x67, "packuswb", v16i8, v8i16, X86Packus,
4341 defm PACKUSDW : sse4_pack<0x2B, "packusdw", v8i16, v4i32, X86Packus,
4344 } // ExeDomain = SSEPackedInt
4346 //===---------------------------------------------------------------------===//
4347 // SSE2 - Packed Integer Unpack Instructions
4348 //===---------------------------------------------------------------------===//
4350 let ExeDomain = SSEPackedInt in {
4351 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
4352 SDNode OpNode, PatFrag ld_frag, bit Is2Addr = 1> {
4353 def rr : PDI<opc, MRMSrcReg,
4354 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4356 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
4357 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4358 [(set VR128:$dst, (vt (OpNode VR128:$src1, VR128:$src2)))],
4359 IIC_SSE_UNPCK>, Sched<[WriteShuffle]>;
4360 def rm : PDI<opc, MRMSrcMem,
4361 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4363 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
4364 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4365 [(set VR128:$dst, (vt (OpNode VR128:$src1,
4366 (bitconvert (ld_frag addr:$src2)))))],
4368 Sched<[WriteShuffleLd, ReadAfterLd]>;
4371 multiclass sse2_unpack_y<bits<8> opc, string OpcodeStr, ValueType vt,
4373 def Yrr : PDI<opc, MRMSrcReg,
4374 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4375 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4376 [(set VR256:$dst, (vt (OpNode VR256:$src1, VR256:$src2)))]>,
4377 Sched<[WriteShuffle]>;
4378 def Yrm : PDI<opc, MRMSrcMem,
4379 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4380 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4381 [(set VR256:$dst, (vt (OpNode VR256:$src1,
4382 (bitconvert (loadv4i64 addr:$src2)))))]>,
4383 Sched<[WriteShuffleLd, ReadAfterLd]>;
4387 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
4388 defm VPUNPCKLBW : sse2_unpack<0x60, "vpunpcklbw", v16i8, X86Unpckl,
4389 loadv2i64, 0>, VEX_4V, VEX_WIG;
4390 defm VPUNPCKLWD : sse2_unpack<0x61, "vpunpcklwd", v8i16, X86Unpckl,
4391 loadv2i64, 0>, VEX_4V, VEX_WIG;
4392 defm VPUNPCKHBW : sse2_unpack<0x68, "vpunpckhbw", v16i8, X86Unpckh,
4393 loadv2i64, 0>, VEX_4V, VEX_WIG;
4394 defm VPUNPCKHWD : sse2_unpack<0x69, "vpunpckhwd", v8i16, X86Unpckh,
4395 loadv2i64, 0>, VEX_4V, VEX_WIG;
4397 let Predicates = [HasAVX, NoVLX] in {
4398 defm VPUNPCKLDQ : sse2_unpack<0x62, "vpunpckldq", v4i32, X86Unpckl,
4399 loadv2i64, 0>, VEX_4V, VEX_WIG;
4400 defm VPUNPCKLQDQ : sse2_unpack<0x6C, "vpunpcklqdq", v2i64, X86Unpckl,
4401 loadv2i64, 0>, VEX_4V, VEX_WIG;
4402 defm VPUNPCKHDQ : sse2_unpack<0x6A, "vpunpckhdq", v4i32, X86Unpckh,
4403 loadv2i64, 0>, VEX_4V, VEX_WIG;
4404 defm VPUNPCKHQDQ : sse2_unpack<0x6D, "vpunpckhqdq", v2i64, X86Unpckh,
4405 loadv2i64, 0>, VEX_4V, VEX_WIG;
4408 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
4409 defm VPUNPCKLBW : sse2_unpack_y<0x60, "vpunpcklbw", v32i8, X86Unpckl>,
4410 VEX_4V, VEX_L, VEX_WIG;
4411 defm VPUNPCKLWD : sse2_unpack_y<0x61, "vpunpcklwd", v16i16, X86Unpckl>,
4412 VEX_4V, VEX_L, VEX_WIG;
4413 defm VPUNPCKHBW : sse2_unpack_y<0x68, "vpunpckhbw", v32i8, X86Unpckh>,
4414 VEX_4V, VEX_L, VEX_WIG;
4415 defm VPUNPCKHWD : sse2_unpack_y<0x69, "vpunpckhwd", v16i16, X86Unpckh>,
4416 VEX_4V, VEX_L, VEX_WIG;
4418 let Predicates = [HasAVX2, NoVLX] in {
4419 defm VPUNPCKLDQ : sse2_unpack_y<0x62, "vpunpckldq", v8i32, X86Unpckl>,
4420 VEX_4V, VEX_L, VEX_WIG;
4421 defm VPUNPCKLQDQ : sse2_unpack_y<0x6C, "vpunpcklqdq", v4i64, X86Unpckl>,
4422 VEX_4V, VEX_L, VEX_WIG;
4423 defm VPUNPCKHDQ : sse2_unpack_y<0x6A, "vpunpckhdq", v8i32, X86Unpckh>,
4424 VEX_4V, VEX_L, VEX_WIG;
4425 defm VPUNPCKHQDQ : sse2_unpack_y<0x6D, "vpunpckhqdq", v4i64, X86Unpckh>,
4426 VEX_4V, VEX_L, VEX_WIG;
4429 let Constraints = "$src1 = $dst" in {
4430 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, X86Unpckl,
4432 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, X86Unpckl,
4434 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, X86Unpckl,
4436 defm PUNPCKLQDQ : sse2_unpack<0x6C, "punpcklqdq", v2i64, X86Unpckl,
4439 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, X86Unpckh,
4441 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, X86Unpckh,
4443 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, X86Unpckh,
4445 defm PUNPCKHQDQ : sse2_unpack<0x6D, "punpckhqdq", v2i64, X86Unpckh,
4448 } // ExeDomain = SSEPackedInt
4450 //===---------------------------------------------------------------------===//
4451 // SSE2 - Packed Integer Extract and Insert
4452 //===---------------------------------------------------------------------===//
4454 let ExeDomain = SSEPackedInt in {
4455 multiclass sse2_pinsrw<bit Is2Addr = 1> {
4456 def rri : Ii8<0xC4, MRMSrcReg,
4457 (outs VR128:$dst), (ins VR128:$src1,
4458 GR32orGR64:$src2, u8imm:$src3),
4460 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
4461 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4463 (X86pinsrw VR128:$src1, GR32orGR64:$src2, imm:$src3))],
4464 IIC_SSE_PINSRW>, Sched<[WriteShuffle]>;
4465 def rmi : Ii8<0xC4, MRMSrcMem,
4466 (outs VR128:$dst), (ins VR128:$src1,
4467 i16mem:$src2, u8imm:$src3),
4469 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
4470 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4472 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
4473 imm:$src3))], IIC_SSE_PINSRW>,
4474 Sched<[WriteShuffleLd, ReadAfterLd]>;
4478 let Predicates = [HasAVX, NoBWI] in
4479 def VPEXTRWri : Ii8<0xC5, MRMSrcReg,
4480 (outs GR32orGR64:$dst), (ins VR128:$src1, u8imm:$src2),
4481 "vpextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4482 [(set GR32orGR64:$dst, (X86pextrw (v8i16 VR128:$src1),
4483 imm:$src2))]>, PD, VEX,
4484 Sched<[WriteShuffle]>;
4485 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
4486 (outs GR32orGR64:$dst), (ins VR128:$src1, u8imm:$src2),
4487 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4488 [(set GR32orGR64:$dst, (X86pextrw (v8i16 VR128:$src1),
4489 imm:$src2))], IIC_SSE_PEXTRW>,
4490 Sched<[WriteShuffleLd, ReadAfterLd]>;
4493 let Predicates = [HasAVX, NoBWI] in
4494 defm VPINSRW : sse2_pinsrw<0>, PD, VEX_4V;
4496 let Predicates = [UseSSE2], Constraints = "$src1 = $dst" in
4497 defm PINSRW : sse2_pinsrw, PD;
4499 } // ExeDomain = SSEPackedInt
4501 //===---------------------------------------------------------------------===//
4502 // SSE2 - Packed Mask Creation
4503 //===---------------------------------------------------------------------===//
4505 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecLogic] in {
4507 def VPMOVMSKBrr : VPDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst),
4509 "pmovmskb\t{$src, $dst|$dst, $src}",
4510 [(set GR32orGR64:$dst, (X86movmsk (v16i8 VR128:$src)))],
4511 IIC_SSE_MOVMSK>, VEX, VEX_WIG;
4513 let Predicates = [HasAVX2] in {
4514 def VPMOVMSKBYrr : VPDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst),
4516 "pmovmskb\t{$src, $dst|$dst, $src}",
4517 [(set GR32orGR64:$dst, (X86movmsk (v32i8 VR256:$src)))]>,
4518 VEX, VEX_L, VEX_WIG;
4521 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst), (ins VR128:$src),
4522 "pmovmskb\t{$src, $dst|$dst, $src}",
4523 [(set GR32orGR64:$dst, (X86movmsk (v16i8 VR128:$src)))],
4526 } // ExeDomain = SSEPackedInt
4528 //===---------------------------------------------------------------------===//
4529 // SSE2 - Conditional Store
4530 //===---------------------------------------------------------------------===//
4532 let ExeDomain = SSEPackedInt, SchedRW = [WriteStore] in {
4534 let Uses = [EDI], Predicates = [HasAVX,Not64BitMode] in
4535 def VMASKMOVDQU : VPDI<0xF7, MRMSrcReg, (outs),
4536 (ins VR128:$src, VR128:$mask),
4537 "maskmovdqu\t{$mask, $src|$src, $mask}",
4538 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)],
4539 IIC_SSE_MASKMOV>, VEX, VEX_WIG;
4540 let Uses = [RDI], Predicates = [HasAVX,In64BitMode] in
4541 def VMASKMOVDQU64 : VPDI<0xF7, MRMSrcReg, (outs),
4542 (ins VR128:$src, VR128:$mask),
4543 "maskmovdqu\t{$mask, $src|$src, $mask}",
4544 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)],
4545 IIC_SSE_MASKMOV>, VEX, VEX_WIG;
4547 let Uses = [EDI], Predicates = [UseSSE2,Not64BitMode] in
4548 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
4549 "maskmovdqu\t{$mask, $src|$src, $mask}",
4550 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)],
4552 let Uses = [RDI], Predicates = [UseSSE2,In64BitMode] in
4553 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
4554 "maskmovdqu\t{$mask, $src|$src, $mask}",
4555 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)],
4558 } // ExeDomain = SSEPackedInt
4560 //===---------------------------------------------------------------------===//
4561 // SSE2 - Move Doubleword/Quadword
4562 //===---------------------------------------------------------------------===//
4564 //===---------------------------------------------------------------------===//
4565 // Move Int Doubleword to Packed Double Int
4567 let ExeDomain = SSEPackedInt in {
4568 def VMOVDI2PDIrr : VS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
4569 "movd\t{$src, $dst|$dst, $src}",
4571 (v4i32 (scalar_to_vector GR32:$src)))], IIC_SSE_MOVDQ>,
4572 VEX, Sched<[WriteMove]>;
4573 def VMOVDI2PDIrm : VS2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
4574 "movd\t{$src, $dst|$dst, $src}",
4576 (v4i32 (scalar_to_vector (loadi32 addr:$src))))],
4578 VEX, Sched<[WriteLoad]>;
4579 def VMOV64toPQIrr : VRS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4580 "movq\t{$src, $dst|$dst, $src}",
4582 (v2i64 (scalar_to_vector GR64:$src)))],
4583 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4584 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0, mayLoad = 1 in
4585 def VMOV64toPQIrm : VRS2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4586 "movq\t{$src, $dst|$dst, $src}",
4587 [], IIC_SSE_MOVDQ>, VEX, Sched<[WriteLoad]>;
4588 let isCodeGenOnly = 1 in
4589 def VMOV64toSDrr : VRS2I<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
4590 "movq\t{$src, $dst|$dst, $src}",
4591 [(set FR64:$dst, (bitconvert GR64:$src))],
4592 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4594 def MOVDI2PDIrr : S2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
4595 "movd\t{$src, $dst|$dst, $src}",
4597 (v4i32 (scalar_to_vector GR32:$src)))], IIC_SSE_MOVDQ>,
4599 def MOVDI2PDIrm : S2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
4600 "movd\t{$src, $dst|$dst, $src}",
4602 (v4i32 (scalar_to_vector (loadi32 addr:$src))))],
4603 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4604 def MOV64toPQIrr : RS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4605 "movq\t{$src, $dst|$dst, $src}",
4607 (v2i64 (scalar_to_vector GR64:$src)))],
4608 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4609 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0, mayLoad = 1 in
4610 def MOV64toPQIrm : RS2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4611 "movq\t{$src, $dst|$dst, $src}",
4612 [], IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4613 let isCodeGenOnly = 1 in
4614 def MOV64toSDrr : RS2I<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
4615 "movq\t{$src, $dst|$dst, $src}",
4616 [(set FR64:$dst, (bitconvert GR64:$src))],
4617 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4618 } // ExeDomain = SSEPackedInt
4620 //===---------------------------------------------------------------------===//
4621 // Move Int Doubleword to Single Scalar
4623 let ExeDomain = SSEPackedInt, isCodeGenOnly = 1 in {
4624 def VMOVDI2SSrr : VS2I<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
4625 "movd\t{$src, $dst|$dst, $src}",
4626 [(set FR32:$dst, (bitconvert GR32:$src))],
4627 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4629 def VMOVDI2SSrm : VS2I<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
4630 "movd\t{$src, $dst|$dst, $src}",
4631 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))],
4633 VEX, Sched<[WriteLoad]>;
4634 def MOVDI2SSrr : S2I<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
4635 "movd\t{$src, $dst|$dst, $src}",
4636 [(set FR32:$dst, (bitconvert GR32:$src))],
4637 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4639 def MOVDI2SSrm : S2I<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
4640 "movd\t{$src, $dst|$dst, $src}",
4641 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))],
4642 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4643 } // ExeDomain = SSEPackedInt, isCodeGenOnly = 1
4645 //===---------------------------------------------------------------------===//
4646 // Move Packed Doubleword Int to Packed Double Int
4648 let ExeDomain = SSEPackedInt in {
4649 def VMOVPDI2DIrr : VS2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
4650 "movd\t{$src, $dst|$dst, $src}",
4651 [(set GR32:$dst, (extractelt (v4i32 VR128:$src),
4652 (iPTR 0)))], IIC_SSE_MOVD_ToGP>, VEX,
4654 def VMOVPDI2DImr : VS2I<0x7E, MRMDestMem, (outs),
4655 (ins i32mem:$dst, VR128:$src),
4656 "movd\t{$src, $dst|$dst, $src}",
4657 [(store (i32 (extractelt (v4i32 VR128:$src),
4658 (iPTR 0))), addr:$dst)], IIC_SSE_MOVDQ>,
4659 VEX, Sched<[WriteStore]>;
4660 def MOVPDI2DIrr : S2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
4661 "movd\t{$src, $dst|$dst, $src}",
4662 [(set GR32:$dst, (extractelt (v4i32 VR128:$src),
4663 (iPTR 0)))], IIC_SSE_MOVD_ToGP>,
4665 def MOVPDI2DImr : S2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
4666 "movd\t{$src, $dst|$dst, $src}",
4667 [(store (i32 (extractelt (v4i32 VR128:$src),
4668 (iPTR 0))), addr:$dst)],
4669 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4670 } // ExeDomain = SSEPackedInt
4671 //===---------------------------------------------------------------------===//
4672 // Move Packed Doubleword Int first element to Doubleword Int
4674 let ExeDomain = SSEPackedInt in {
4675 let SchedRW = [WriteMove] in {
4676 def VMOVPQIto64rr : VRS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
4677 "movq\t{$src, $dst|$dst, $src}",
4678 [(set GR64:$dst, (extractelt (v2i64 VR128:$src),
4683 def MOVPQIto64rr : RS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
4684 "movq\t{$src, $dst|$dst, $src}",
4685 [(set GR64:$dst, (extractelt (v2i64 VR128:$src),
4690 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0, mayStore = 1 in
4691 def VMOVPQIto64mr : VRS2I<0x7E, MRMDestMem, (outs),
4692 (ins i64mem:$dst, VR128:$src),
4693 "movq\t{$src, $dst|$dst, $src}",
4694 [], IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4695 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0, mayStore = 1 in
4696 def MOVPQIto64mr : RS2I<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
4697 "movq\t{$src, $dst|$dst, $src}",
4698 [], IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4699 } // ExeDomain = SSEPackedInt
4701 //===---------------------------------------------------------------------===//
4702 // Bitcast FR64 <-> GR64
4704 let ExeDomain = SSEPackedInt, isCodeGenOnly = 1 in {
4705 let Predicates = [UseAVX] in
4706 def VMOV64toSDrm : VS2SI<0x7E, MRMSrcMem, (outs FR64:$dst), (ins i64mem:$src),
4707 "movq\t{$src, $dst|$dst, $src}",
4708 [(set FR64:$dst, (bitconvert (loadi64 addr:$src)))]>,
4709 VEX, Sched<[WriteLoad]>;
4710 def VMOVSDto64rr : VRS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins FR64:$src),
4711 "movq\t{$src, $dst|$dst, $src}",
4712 [(set GR64:$dst, (bitconvert FR64:$src))],
4713 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4714 def VMOVSDto64mr : VRS2I<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, FR64:$src),
4715 "movq\t{$src, $dst|$dst, $src}",
4716 [(store (i64 (bitconvert FR64:$src)), addr:$dst)],
4717 IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4719 def MOV64toSDrm : S2SI<0x7E, MRMSrcMem, (outs FR64:$dst), (ins i64mem:$src),
4720 "movq\t{$src, $dst|$dst, $src}",
4721 [(set FR64:$dst, (bitconvert (loadi64 addr:$src)))],
4722 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4723 def MOVSDto64rr : RS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins FR64:$src),
4724 "movq\t{$src, $dst|$dst, $src}",
4725 [(set GR64:$dst, (bitconvert FR64:$src))],
4726 IIC_SSE_MOVD_ToGP>, Sched<[WriteMove]>;
4727 def MOVSDto64mr : RS2I<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, FR64:$src),
4728 "movq\t{$src, $dst|$dst, $src}",
4729 [(store (i64 (bitconvert FR64:$src)), addr:$dst)],
4730 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4731 } // ExeDomain = SSEPackedInt, isCodeGenOnly = 1
4733 //===---------------------------------------------------------------------===//
4734 // Move Scalar Single to Double Int
4736 let ExeDomain = SSEPackedInt, isCodeGenOnly = 1 in {
4737 def VMOVSS2DIrr : VS2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
4738 "movd\t{$src, $dst|$dst, $src}",
4739 [(set GR32:$dst, (bitconvert FR32:$src))],
4740 IIC_SSE_MOVD_ToGP>, VEX, Sched<[WriteMove]>;
4741 def VMOVSS2DImr : VS2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
4742 "movd\t{$src, $dst|$dst, $src}",
4743 [(store (i32 (bitconvert FR32:$src)), addr:$dst)],
4744 IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4745 def MOVSS2DIrr : S2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
4746 "movd\t{$src, $dst|$dst, $src}",
4747 [(set GR32:$dst, (bitconvert FR32:$src))],
4748 IIC_SSE_MOVD_ToGP>, Sched<[WriteMove]>;
4749 def MOVSS2DImr : S2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
4750 "movd\t{$src, $dst|$dst, $src}",
4751 [(store (i32 (bitconvert FR32:$src)), addr:$dst)],
4752 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4753 } // ExeDomain = SSEPackedInt, isCodeGenOnly = 1
4755 let Predicates = [UseAVX] in {
4756 let AddedComplexity = 15 in {
4757 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector GR32:$src)))),
4758 (VMOVDI2PDIrr GR32:$src)>;
4760 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector GR64:$src)))),
4761 (VMOV64toPQIrr GR64:$src)>;
4763 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
4764 (v2i64 (scalar_to_vector GR64:$src)),(iPTR 0)))),
4765 (SUBREG_TO_REG (i64 0), (VMOV64toPQIrr GR64:$src), sub_xmm)>;
4767 // AVX 128-bit movd/movq instructions write zeros in the high 128-bit part.
4768 // These instructions also write zeros in the high part of a 256-bit register.
4769 let AddedComplexity = 20 in {
4770 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector (zextloadi64i32 addr:$src))))),
4771 (VMOVDI2PDIrm addr:$src)>;
4772 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
4773 (VMOVDI2PDIrm addr:$src)>;
4774 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
4775 (VMOVDI2PDIrm addr:$src)>;
4776 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
4777 (VMOVDI2PDIrm addr:$src)>;
4778 def : Pat<(v4i32 (X86vzload addr:$src)),
4779 (VMOVDI2PDIrm addr:$src)>;
4780 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
4781 (v4i32 (scalar_to_vector (loadi32 addr:$src))), (iPTR 0)))),
4782 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrm addr:$src), sub_xmm)>;
4783 def : Pat<(v8i32 (X86vzload addr:$src)),
4784 (SUBREG_TO_REG (i64 0), (VMOVDI2PDIrm addr:$src), sub_xmm)>;
4786 // Use regular 128-bit instructions to match 256-bit scalar_to_vec+zext.
4787 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
4788 (v4i32 (scalar_to_vector GR32:$src)),(iPTR 0)))),
4789 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrr GR32:$src), sub_xmm)>;
4792 let Predicates = [UseSSE2] in {
4793 let AddedComplexity = 15 in {
4794 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector GR32:$src)))),
4795 (MOVDI2PDIrr GR32:$src)>;
4797 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector GR64:$src)))),
4798 (MOV64toPQIrr GR64:$src)>;
4800 let AddedComplexity = 20 in {
4801 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector (zextloadi64i32 addr:$src))))),
4802 (MOVDI2PDIrm addr:$src)>;
4803 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
4804 (MOVDI2PDIrm addr:$src)>;
4805 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
4806 (MOVDI2PDIrm addr:$src)>;
4807 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
4808 (MOVDI2PDIrm addr:$src)>;
4809 def : Pat<(v4i32 (X86vzload addr:$src)),
4810 (MOVDI2PDIrm addr:$src)>;
4814 // Before the MC layer of LLVM existed, clang emitted "movd" assembly instead of
4815 // "movq" due to MacOS parsing limitation. In order to parse old assembly, we add
4817 def : InstAlias<"movd\t{$src, $dst|$dst, $src}",
4818 (MOV64toPQIrr VR128:$dst, GR64:$src), 0>;
4819 def : InstAlias<"movd\t{$src, $dst|$dst, $src}",
4820 (MOVPQIto64rr GR64:$dst, VR128:$src), 0>;
4821 // Allow "vmovd" but print "vmovq" since we don't need compatibility for AVX.
4822 def : InstAlias<"vmovd\t{$src, $dst|$dst, $src}",
4823 (VMOV64toPQIrr VR128:$dst, GR64:$src), 0>;
4824 def : InstAlias<"vmovd\t{$src, $dst|$dst, $src}",
4825 (VMOVPQIto64rr GR64:$dst, VR128:$src), 0>;
4827 //===---------------------------------------------------------------------===//
4828 // SSE2 - Move Quadword
4829 //===---------------------------------------------------------------------===//
4831 //===---------------------------------------------------------------------===//
4832 // Move Quadword Int to Packed Quadword Int
4835 let ExeDomain = SSEPackedInt, SchedRW = [WriteLoad] in {
4836 def VMOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4837 "vmovq\t{$src, $dst|$dst, $src}",
4839 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
4840 VEX, Requires<[UseAVX]>, VEX_WIG;
4841 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4842 "movq\t{$src, $dst|$dst, $src}",
4844 (v2i64 (scalar_to_vector (loadi64 addr:$src))))],
4846 Requires<[UseSSE2]>; // SSE2 instruction with XS Prefix
4847 } // ExeDomain, SchedRW
4849 //===---------------------------------------------------------------------===//
4850 // Move Packed Quadword Int to Quadword Int
4852 let ExeDomain = SSEPackedInt, SchedRW = [WriteStore] in {
4853 def VMOVPQI2QImr : VS2I<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
4854 "movq\t{$src, $dst|$dst, $src}",
4855 [(store (i64 (extractelt (v2i64 VR128:$src),
4856 (iPTR 0))), addr:$dst)],
4857 IIC_SSE_MOVDQ>, VEX, VEX_WIG;
4858 def MOVPQI2QImr : S2I<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
4859 "movq\t{$src, $dst|$dst, $src}",
4860 [(store (i64 (extractelt (v2i64 VR128:$src),
4861 (iPTR 0))), addr:$dst)],
4863 } // ExeDomain, SchedRW
4865 // For disassembler only
4866 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
4867 SchedRW = [WriteVecLogic] in {
4868 def VMOVPQI2QIrr : VS2I<0xD6, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
4869 "movq\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVQ_RR>, VEX, VEX_WIG;
4870 def MOVPQI2QIrr : S2I<0xD6, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
4871 "movq\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVQ_RR>;
4874 // Aliases to help the assembler pick two byte VEX encodings by swapping the
4875 // operands relative to the normal instructions to use VEX.R instead of VEX.B.
4876 def : InstAlias<"vmovq\t{$src, $dst|$dst, $src}",
4877 (VMOVPQI2QIrr VR128L:$dst, VR128H:$src), 0>;
4879 let Predicates = [UseAVX], AddedComplexity = 20 in {
4880 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
4881 (VMOVQI2PQIrm addr:$src)>;
4882 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
4883 (VMOVQI2PQIrm addr:$src)>;
4884 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
4885 (VMOVQI2PQIrm addr:$src)>;
4886 def : Pat<(v2i64 (X86vzload addr:$src)),
4887 (VMOVQI2PQIrm addr:$src)>;
4888 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
4889 (v2i64 (scalar_to_vector (loadi64 addr:$src))), (iPTR 0)))),
4890 (SUBREG_TO_REG (i64 0), (VMOVQI2PQIrm addr:$src), sub_xmm)>;
4891 def : Pat<(v4i64 (X86vzload addr:$src)),
4892 (SUBREG_TO_REG (i64 0), (VMOVQI2PQIrm addr:$src), sub_xmm)>;
4895 let Predicates = [UseSSE2], AddedComplexity = 20 in {
4896 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
4897 (MOVQI2PQIrm addr:$src)>;
4898 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
4899 (MOVQI2PQIrm addr:$src)>;
4900 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
4901 (MOVQI2PQIrm addr:$src)>;
4902 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVQI2PQIrm addr:$src)>;
4905 //===---------------------------------------------------------------------===//
4906 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
4907 // IA32 document. movq xmm1, xmm2 does clear the high bits.
4909 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecLogic] in {
4910 let AddedComplexity = 15 in
4911 def VMOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
4912 "vmovq\t{$src, $dst|$dst, $src}",
4913 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))],
4915 XS, VEX, Requires<[UseAVX]>, VEX_WIG;
4916 let AddedComplexity = 15 in
4917 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
4918 "movq\t{$src, $dst|$dst, $src}",
4919 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))],
4921 XS, Requires<[UseSSE2]>;
4922 } // ExeDomain, SchedRW
4924 let AddedComplexity = 20 in {
4925 let Predicates = [UseAVX] in {
4926 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
4927 (VMOVZPQILo2PQIrr VR128:$src)>;
4929 let Predicates = [UseSSE2] in {
4930 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
4931 (MOVZPQILo2PQIrr VR128:$src)>;
4935 //===---------------------------------------------------------------------===//
4936 // SSE3 - Replicate Single FP - MOVSHDUP and MOVSLDUP
4937 //===---------------------------------------------------------------------===//
4938 multiclass sse3_replicate_sfp<bits<8> op, SDNode OpNode, string OpcodeStr,
4939 ValueType vt, RegisterClass RC, PatFrag mem_frag,
4940 X86MemOperand x86memop> {
4941 def rr : S3SI<op, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
4942 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
4943 [(set RC:$dst, (vt (OpNode RC:$src)))],
4944 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
4945 def rm : S3SI<op, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
4946 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
4947 [(set RC:$dst, (OpNode (mem_frag addr:$src)))],
4948 IIC_SSE_MOV_LH>, Sched<[WriteLoad]>;
4951 let Predicates = [HasAVX, NoVLX] in {
4952 defm VMOVSHDUP : sse3_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
4953 v4f32, VR128, loadv4f32, f128mem>, VEX, VEX_WIG;
4954 defm VMOVSLDUP : sse3_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
4955 v4f32, VR128, loadv4f32, f128mem>, VEX, VEX_WIG;
4956 defm VMOVSHDUPY : sse3_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
4957 v8f32, VR256, loadv8f32, f256mem>, VEX, VEX_L, VEX_WIG;
4958 defm VMOVSLDUPY : sse3_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
4959 v8f32, VR256, loadv8f32, f256mem>, VEX, VEX_L, VEX_WIG;
4961 defm MOVSHDUP : sse3_replicate_sfp<0x16, X86Movshdup, "movshdup", v4f32, VR128,
4962 memopv4f32, f128mem>;
4963 defm MOVSLDUP : sse3_replicate_sfp<0x12, X86Movsldup, "movsldup", v4f32, VR128,
4964 memopv4f32, f128mem>;
4966 let Predicates = [HasAVX, NoVLX] in {
4967 def : Pat<(v4i32 (X86Movshdup VR128:$src)),
4968 (VMOVSHDUPrr VR128:$src)>;
4969 def : Pat<(v4i32 (X86Movshdup (bc_v4i32 (loadv2i64 addr:$src)))),
4970 (VMOVSHDUPrm addr:$src)>;
4971 def : Pat<(v4i32 (X86Movsldup VR128:$src)),
4972 (VMOVSLDUPrr VR128:$src)>;
4973 def : Pat<(v4i32 (X86Movsldup (bc_v4i32 (loadv2i64 addr:$src)))),
4974 (VMOVSLDUPrm addr:$src)>;
4975 def : Pat<(v8i32 (X86Movshdup VR256:$src)),
4976 (VMOVSHDUPYrr VR256:$src)>;
4977 def : Pat<(v8i32 (X86Movshdup (bc_v8i32 (loadv4i64 addr:$src)))),
4978 (VMOVSHDUPYrm addr:$src)>;
4979 def : Pat<(v8i32 (X86Movsldup VR256:$src)),
4980 (VMOVSLDUPYrr VR256:$src)>;
4981 def : Pat<(v8i32 (X86Movsldup (bc_v8i32 (loadv4i64 addr:$src)))),
4982 (VMOVSLDUPYrm addr:$src)>;
4985 let Predicates = [UseSSE3] in {
4986 def : Pat<(v4i32 (X86Movshdup VR128:$src)),
4987 (MOVSHDUPrr VR128:$src)>;
4988 def : Pat<(v4i32 (X86Movshdup (bc_v4i32 (memopv2i64 addr:$src)))),
4989 (MOVSHDUPrm addr:$src)>;
4990 def : Pat<(v4i32 (X86Movsldup VR128:$src)),
4991 (MOVSLDUPrr VR128:$src)>;
4992 def : Pat<(v4i32 (X86Movsldup (bc_v4i32 (memopv2i64 addr:$src)))),
4993 (MOVSLDUPrm addr:$src)>;
4996 //===---------------------------------------------------------------------===//
4997 // SSE3 - Replicate Double FP - MOVDDUP
4998 //===---------------------------------------------------------------------===//
5000 multiclass sse3_replicate_dfp<string OpcodeStr> {
5001 def rr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5002 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5003 [(set VR128:$dst, (v2f64 (X86Movddup VR128:$src)))],
5004 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
5005 def rm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
5006 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5009 (scalar_to_vector (loadf64 addr:$src)))))],
5010 IIC_SSE_MOV_LH>, Sched<[WriteLoad]>;
5013 // FIXME: Merge with above classe when there're patterns for the ymm version
5014 multiclass sse3_replicate_dfp_y<string OpcodeStr> {
5015 def rr : S3DI<0x12, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
5016 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5017 [(set VR256:$dst, (v4f64 (X86Movddup VR256:$src)))]>,
5018 Sched<[WriteFShuffle]>;
5019 def rm : S3DI<0x12, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
5020 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5022 (v4f64 (X86Movddup (loadv4f64 addr:$src))))]>,
5026 let Predicates = [HasAVX, NoVLX] in {
5027 defm VMOVDDUP : sse3_replicate_dfp<"vmovddup">, VEX, VEX_WIG;
5028 defm VMOVDDUPY : sse3_replicate_dfp_y<"vmovddup">, VEX, VEX_L, VEX_WIG;
5031 defm MOVDDUP : sse3_replicate_dfp<"movddup">;
5034 let Predicates = [HasAVX, NoVLX] in {
5035 def : Pat<(X86Movddup (loadv2f64 addr:$src)),
5036 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5039 def : Pat<(X86Movddup (loadv4i64 addr:$src)),
5040 (VMOVDDUPYrm addr:$src)>;
5041 def : Pat<(X86Movddup (v4i64 VR256:$src)),
5042 (VMOVDDUPYrr VR256:$src)>;
5045 let Predicates = [HasAVX, NoVLX] in
5046 def : Pat<(v2f64 (X86VBroadcast (loadf64 addr:$src))),
5047 (VMOVDDUPrm addr:$src)>;
5048 let Predicates = [HasAVX1Only] in
5049 def : Pat<(v2i64 (X86VBroadcast (loadi64 addr:$src))),
5050 (VMOVDDUPrm addr:$src)>;
5052 let Predicates = [UseSSE3] in {
5053 def : Pat<(X86Movddup (memopv2f64 addr:$src)),
5054 (MOVDDUPrm addr:$src)>;
5057 //===---------------------------------------------------------------------===//
5058 // SSE3 - Move Unaligned Integer
5059 //===---------------------------------------------------------------------===//
5061 let SchedRW = [WriteLoad] in {
5062 let Predicates = [HasAVX] in {
5063 def VLDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5064 "vlddqu\t{$src, $dst|$dst, $src}",
5065 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>, VEX, VEX_WIG;
5066 def VLDDQUYrm : S3DI<0xF0, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
5067 "vlddqu\t{$src, $dst|$dst, $src}",
5068 [(set VR256:$dst, (int_x86_avx_ldu_dq_256 addr:$src))]>,
5069 VEX, VEX_L, VEX_WIG;
5071 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5072 "lddqu\t{$src, $dst|$dst, $src}",
5073 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))],
5077 //===---------------------------------------------------------------------===//
5078 // SSE3 - Arithmetic
5079 //===---------------------------------------------------------------------===//
5081 multiclass sse3_addsub<Intrinsic Int, string OpcodeStr, RegisterClass RC,
5082 X86MemOperand x86memop, OpndItins itins,
5083 PatFrag ld_frag, bit Is2Addr = 1> {
5084 def rr : I<0xD0, MRMSrcReg,
5085 (outs RC:$dst), (ins RC:$src1, RC:$src2),
5087 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5088 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5089 [(set RC:$dst, (Int RC:$src1, RC:$src2))], itins.rr>,
5090 Sched<[itins.Sched]>;
5091 def rm : I<0xD0, MRMSrcMem,
5092 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5094 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5095 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5096 [(set RC:$dst, (Int RC:$src1, (ld_frag addr:$src2)))], itins.rr>,
5097 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5100 let Predicates = [HasAVX] in {
5101 let ExeDomain = SSEPackedSingle in {
5102 defm VADDSUBPS : sse3_addsub<int_x86_sse3_addsub_ps, "vaddsubps", VR128,
5103 f128mem, SSE_ALU_F32P, loadv4f32, 0>, XD, VEX_4V, VEX_WIG;
5104 defm VADDSUBPSY : sse3_addsub<int_x86_avx_addsub_ps_256, "vaddsubps", VR256,
5105 f256mem, SSE_ALU_F32P, loadv8f32, 0>, XD, VEX_4V, VEX_L, VEX_WIG;
5107 let ExeDomain = SSEPackedDouble in {
5108 defm VADDSUBPD : sse3_addsub<int_x86_sse3_addsub_pd, "vaddsubpd", VR128,
5109 f128mem, SSE_ALU_F64P, loadv2f64, 0>, PD, VEX_4V, VEX_WIG;
5110 defm VADDSUBPDY : sse3_addsub<int_x86_avx_addsub_pd_256, "vaddsubpd", VR256,
5111 f256mem, SSE_ALU_F64P, loadv4f64, 0>, PD, VEX_4V, VEX_L, VEX_WIG;
5114 let Constraints = "$src1 = $dst", Predicates = [UseSSE3] in {
5115 let ExeDomain = SSEPackedSingle in
5116 defm ADDSUBPS : sse3_addsub<int_x86_sse3_addsub_ps, "addsubps", VR128,
5117 f128mem, SSE_ALU_F32P, memopv4f32>, XD;
5118 let ExeDomain = SSEPackedDouble in
5119 defm ADDSUBPD : sse3_addsub<int_x86_sse3_addsub_pd, "addsubpd", VR128,
5120 f128mem, SSE_ALU_F64P, memopv2f64>, PD;
5123 // Patterns used to select 'addsub' instructions.
5124 let Predicates = [HasAVX] in {
5125 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 VR128:$rhs))),
5126 (VADDSUBPSrr VR128:$lhs, VR128:$rhs)>;
5127 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (loadv4f32 addr:$rhs))),
5128 (VADDSUBPSrm VR128:$lhs, f128mem:$rhs)>;
5129 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 VR128:$rhs))),
5130 (VADDSUBPDrr VR128:$lhs, VR128:$rhs)>;
5131 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (loadv2f64 addr:$rhs))),
5132 (VADDSUBPDrm VR128:$lhs, f128mem:$rhs)>;
5134 def : Pat<(v8f32 (X86Addsub (v8f32 VR256:$lhs), (v8f32 VR256:$rhs))),
5135 (VADDSUBPSYrr VR256:$lhs, VR256:$rhs)>;
5136 def : Pat<(v8f32 (X86Addsub (v8f32 VR256:$lhs), (loadv8f32 addr:$rhs))),
5137 (VADDSUBPSYrm VR256:$lhs, f256mem:$rhs)>;
5138 def : Pat<(v4f64 (X86Addsub (v4f64 VR256:$lhs), (v4f64 VR256:$rhs))),
5139 (VADDSUBPDYrr VR256:$lhs, VR256:$rhs)>;
5140 def : Pat<(v4f64 (X86Addsub (v4f64 VR256:$lhs), (loadv4f64 addr:$rhs))),
5141 (VADDSUBPDYrm VR256:$lhs, f256mem:$rhs)>;
5144 let Predicates = [UseSSE3] in {
5145 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 VR128:$rhs))),
5146 (ADDSUBPSrr VR128:$lhs, VR128:$rhs)>;
5147 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (memopv4f32 addr:$rhs))),
5148 (ADDSUBPSrm VR128:$lhs, f128mem:$rhs)>;
5149 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 VR128:$rhs))),
5150 (ADDSUBPDrr VR128:$lhs, VR128:$rhs)>;
5151 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (memopv2f64 addr:$rhs))),
5152 (ADDSUBPDrm VR128:$lhs, f128mem:$rhs)>;
5155 //===---------------------------------------------------------------------===//
5156 // SSE3 Instructions
5157 //===---------------------------------------------------------------------===//
5160 multiclass S3D_Int<bits<8> o, string OpcodeStr, ValueType vt, RegisterClass RC,
5161 X86MemOperand x86memop, SDNode OpNode, PatFrag ld_frag,
5163 def rr : S3DI<o, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
5165 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5166 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5167 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], IIC_SSE_HADDSUB_RR>,
5170 def rm : S3DI<o, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5172 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5173 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5174 [(set RC:$dst, (vt (OpNode RC:$src1, (ld_frag addr:$src2))))],
5175 IIC_SSE_HADDSUB_RM>, Sched<[WriteFAddLd, ReadAfterLd]>;
5177 multiclass S3_Int<bits<8> o, string OpcodeStr, ValueType vt, RegisterClass RC,
5178 X86MemOperand x86memop, SDNode OpNode, PatFrag ld_frag,
5180 def rr : S3I<o, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
5182 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5183 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5184 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], IIC_SSE_HADDSUB_RR>,
5187 def rm : S3I<o, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5189 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5190 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5191 [(set RC:$dst, (vt (OpNode RC:$src1, (ld_frag addr:$src2))))],
5192 IIC_SSE_HADDSUB_RM>, Sched<[WriteFAddLd, ReadAfterLd]>;
5195 let Predicates = [HasAVX] in {
5196 let ExeDomain = SSEPackedSingle in {
5197 defm VHADDPS : S3D_Int<0x7C, "vhaddps", v4f32, VR128, f128mem,
5198 X86fhadd, loadv4f32, 0>, VEX_4V, VEX_WIG;
5199 defm VHSUBPS : S3D_Int<0x7D, "vhsubps", v4f32, VR128, f128mem,
5200 X86fhsub, loadv4f32, 0>, VEX_4V, VEX_WIG;
5201 defm VHADDPSY : S3D_Int<0x7C, "vhaddps", v8f32, VR256, f256mem,
5202 X86fhadd, loadv8f32, 0>, VEX_4V, VEX_L, VEX_WIG;
5203 defm VHSUBPSY : S3D_Int<0x7D, "vhsubps", v8f32, VR256, f256mem,
5204 X86fhsub, loadv8f32, 0>, VEX_4V, VEX_L, VEX_WIG;
5206 let ExeDomain = SSEPackedDouble in {
5207 defm VHADDPD : S3_Int <0x7C, "vhaddpd", v2f64, VR128, f128mem,
5208 X86fhadd, loadv2f64, 0>, VEX_4V, VEX_WIG;
5209 defm VHSUBPD : S3_Int <0x7D, "vhsubpd", v2f64, VR128, f128mem,
5210 X86fhsub, loadv2f64, 0>, VEX_4V, VEX_WIG;
5211 defm VHADDPDY : S3_Int <0x7C, "vhaddpd", v4f64, VR256, f256mem,
5212 X86fhadd, loadv4f64, 0>, VEX_4V, VEX_L, VEX_WIG;
5213 defm VHSUBPDY : S3_Int <0x7D, "vhsubpd", v4f64, VR256, f256mem,
5214 X86fhsub, loadv4f64, 0>, VEX_4V, VEX_L, VEX_WIG;
5218 let Constraints = "$src1 = $dst" in {
5219 let ExeDomain = SSEPackedSingle in {
5220 defm HADDPS : S3D_Int<0x7C, "haddps", v4f32, VR128, f128mem, X86fhadd,
5222 defm HSUBPS : S3D_Int<0x7D, "hsubps", v4f32, VR128, f128mem, X86fhsub,
5225 let ExeDomain = SSEPackedDouble in {
5226 defm HADDPD : S3_Int<0x7C, "haddpd", v2f64, VR128, f128mem, X86fhadd,
5228 defm HSUBPD : S3_Int<0x7D, "hsubpd", v2f64, VR128, f128mem, X86fhsub,
5233 //===---------------------------------------------------------------------===//
5234 // SSSE3 - Packed Absolute Instructions
5235 //===---------------------------------------------------------------------===//
5238 /// SS3I_unop_rm_int - Simple SSSE3 unary op whose type can be v*{i8,i16,i32}.
5239 multiclass SS3I_unop_rm<bits<8> opc, string OpcodeStr, ValueType vt,
5240 SDNode OpNode, PatFrag ld_frag> {
5241 def rr : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
5243 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5244 [(set VR128:$dst, (vt (OpNode VR128:$src)))],
5245 IIC_SSE_PABS_RR>, Sched<[WriteVecALU]>;
5247 def rm : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
5249 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5251 (vt (OpNode (bitconvert (ld_frag addr:$src)))))],
5252 IIC_SSE_PABS_RM>, Sched<[WriteVecALULd]>;
5255 /// SS3I_unop_rm_int_y - Simple SSSE3 unary op whose type can be v*{i8,i16,i32}.
5256 multiclass SS3I_unop_rm_y<bits<8> opc, string OpcodeStr, ValueType vt,
5258 def Yrr : SS38I<opc, MRMSrcReg, (outs VR256:$dst),
5260 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5261 [(set VR256:$dst, (vt (OpNode VR256:$src)))]>,
5262 Sched<[WriteVecALU]>;
5264 def Yrm : SS38I<opc, MRMSrcMem, (outs VR256:$dst),
5266 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5268 (vt (OpNode (bitconvert (loadv4i64 addr:$src)))))]>,
5269 Sched<[WriteVecALULd]>;
5272 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5273 defm VPABSB : SS3I_unop_rm<0x1C, "vpabsb", v16i8, abs, loadv2i64>, VEX, VEX_WIG;
5274 defm VPABSW : SS3I_unop_rm<0x1D, "vpabsw", v8i16, abs, loadv2i64>, VEX, VEX_WIG;
5276 let Predicates = [HasAVX, NoVLX] in {
5277 defm VPABSD : SS3I_unop_rm<0x1E, "vpabsd", v4i32, abs, loadv2i64>, VEX, VEX_WIG;
5279 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
5280 defm VPABSB : SS3I_unop_rm_y<0x1C, "vpabsb", v32i8, abs>, VEX, VEX_L, VEX_WIG;
5281 defm VPABSW : SS3I_unop_rm_y<0x1D, "vpabsw", v16i16, abs>, VEX, VEX_L, VEX_WIG;
5283 let Predicates = [HasAVX2, NoVLX] in {
5284 defm VPABSD : SS3I_unop_rm_y<0x1E, "vpabsd", v8i32, abs>, VEX, VEX_L, VEX_WIG;
5287 defm PABSB : SS3I_unop_rm<0x1C, "pabsb", v16i8, abs, memopv2i64>;
5288 defm PABSW : SS3I_unop_rm<0x1D, "pabsw", v8i16, abs, memopv2i64>;
5289 defm PABSD : SS3I_unop_rm<0x1E, "pabsd", v4i32, abs, memopv2i64>;
5291 //===---------------------------------------------------------------------===//
5292 // SSSE3 - Packed Binary Operator Instructions
5293 //===---------------------------------------------------------------------===//
5295 let Sched = WriteVecALU in {
5296 def SSE_PHADDSUBD : OpndItins<
5297 IIC_SSE_PHADDSUBD_RR, IIC_SSE_PHADDSUBD_RM
5299 def SSE_PHADDSUBSW : OpndItins<
5300 IIC_SSE_PHADDSUBSW_RR, IIC_SSE_PHADDSUBSW_RM
5302 def SSE_PHADDSUBW : OpndItins<
5303 IIC_SSE_PHADDSUBW_RR, IIC_SSE_PHADDSUBW_RM
5306 let Sched = WriteShuffle in
5307 def SSE_PSHUFB : OpndItins<
5308 IIC_SSE_PSHUFB_RR, IIC_SSE_PSHUFB_RM
5310 let Sched = WriteVecALU in
5311 def SSE_PSIGN : OpndItins<
5312 IIC_SSE_PSIGN_RR, IIC_SSE_PSIGN_RM
5314 let Sched = WriteVecIMul in
5315 def SSE_PMULHRSW : OpndItins<
5316 IIC_SSE_PMULHRSW, IIC_SSE_PMULHRSW
5319 /// SS3I_binop_rm - Simple SSSE3 bin op
5320 multiclass SS3I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
5321 ValueType DstVT, ValueType OpVT, RegisterClass RC,
5322 PatFrag memop_frag, X86MemOperand x86memop,
5323 OpndItins itins, bit Is2Addr = 1> {
5324 let isCommutable = 1 in
5325 def rr : SS38I<opc, MRMSrcReg, (outs RC:$dst),
5326 (ins RC:$src1, RC:$src2),
5328 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5329 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5330 [(set RC:$dst, (DstVT (OpNode (OpVT RC:$src1), RC:$src2)))], itins.rr>,
5331 Sched<[itins.Sched]>;
5332 def rm : SS38I<opc, MRMSrcMem, (outs RC:$dst),
5333 (ins RC:$src1, x86memop:$src2),
5335 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5336 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5338 (DstVT (OpNode (OpVT RC:$src1),
5339 (bitconvert (memop_frag addr:$src2)))))], itins.rm>,
5340 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5343 /// SS3I_binop_rm_int - Simple SSSE3 bin op whose type can be v*{i8,i16,i32}.
5344 multiclass SS3I_binop_rm_int<bits<8> opc, string OpcodeStr,
5345 Intrinsic IntId128, OpndItins itins,
5346 PatFrag ld_frag, bit Is2Addr = 1> {
5347 let isCommutable = 1 in
5348 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
5349 (ins VR128:$src1, VR128:$src2),
5351 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5352 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5353 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
5354 Sched<[itins.Sched]>;
5355 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
5356 (ins VR128:$src1, i128mem:$src2),
5358 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5359 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5361 (IntId128 VR128:$src1,
5362 (bitconvert (ld_frag addr:$src2))))]>,
5363 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5366 multiclass SS3I_binop_rm_int_y<bits<8> opc, string OpcodeStr,
5368 X86FoldableSchedWrite Sched> {
5369 let isCommutable = 1 in
5370 def rr256 : SS38I<opc, MRMSrcReg, (outs VR256:$dst),
5371 (ins VR256:$src1, VR256:$src2),
5372 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5373 [(set VR256:$dst, (IntId256 VR256:$src1, VR256:$src2))]>,
5375 def rm256 : SS38I<opc, MRMSrcMem, (outs VR256:$dst),
5376 (ins VR256:$src1, i256mem:$src2),
5377 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5379 (IntId256 VR256:$src1, (bitconvert (loadv4i64 addr:$src2))))]>,
5380 Sched<[Sched.Folded, ReadAfterLd]>;
5383 let ImmT = NoImm, Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5384 let isCommutable = 0 in {
5385 defm VPSHUFB : SS3I_binop_rm<0x00, "vpshufb", X86pshufb, v16i8, v16i8,
5386 VR128, loadv2i64, i128mem,
5387 SSE_PSHUFB, 0>, VEX_4V, VEX_WIG;
5388 defm VPMADDUBSW : SS3I_binop_rm<0x04, "vpmaddubsw", X86vpmaddubsw, v8i16,
5389 v16i8, VR128, loadv2i64, i128mem,
5390 SSE_PMADD, 0>, VEX_4V, VEX_WIG;
5392 defm VPMULHRSW : SS3I_binop_rm<0x0B, "vpmulhrsw", X86mulhrs, v8i16, v8i16,
5393 VR128, loadv2i64, i128mem,
5394 SSE_PMULHRSW, 0>, VEX_4V, VEX_WIG;
5397 let ImmT = NoImm, Predicates = [HasAVX] in {
5398 let isCommutable = 0 in {
5399 defm VPHADDW : SS3I_binop_rm<0x01, "vphaddw", X86hadd, v8i16, v8i16, VR128,
5401 SSE_PHADDSUBW, 0>, VEX_4V, VEX_WIG;
5402 defm VPHADDD : SS3I_binop_rm<0x02, "vphaddd", X86hadd, v4i32, v4i32, VR128,
5404 SSE_PHADDSUBD, 0>, VEX_4V, VEX_WIG;
5405 defm VPHSUBW : SS3I_binop_rm<0x05, "vphsubw", X86hsub, v8i16, v8i16, VR128,
5407 SSE_PHADDSUBW, 0>, VEX_4V, VEX_WIG;
5408 defm VPHSUBD : SS3I_binop_rm<0x06, "vphsubd", X86hsub, v4i32, v4i32, VR128,
5410 SSE_PHADDSUBD, 0>, VEX_4V;
5411 defm VPSIGNB : SS3I_binop_rm_int<0x08, "vpsignb",
5412 int_x86_ssse3_psign_b_128,
5413 SSE_PSIGN, loadv2i64, 0>, VEX_4V, VEX_WIG;
5414 defm VPSIGNW : SS3I_binop_rm_int<0x09, "vpsignw",
5415 int_x86_ssse3_psign_w_128,
5416 SSE_PSIGN, loadv2i64, 0>, VEX_4V, VEX_WIG;
5417 defm VPSIGND : SS3I_binop_rm_int<0x0A, "vpsignd",
5418 int_x86_ssse3_psign_d_128,
5419 SSE_PSIGN, loadv2i64, 0>, VEX_4V, VEX_WIG;
5420 defm VPHADDSW : SS3I_binop_rm_int<0x03, "vphaddsw",
5421 int_x86_ssse3_phadd_sw_128,
5422 SSE_PHADDSUBSW, loadv2i64, 0>, VEX_4V, VEX_WIG;
5423 defm VPHSUBSW : SS3I_binop_rm_int<0x07, "vphsubsw",
5424 int_x86_ssse3_phsub_sw_128,
5425 SSE_PHADDSUBSW, loadv2i64, 0>, VEX_4V, VEX_WIG;
5429 let ImmT = NoImm, Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
5430 let isCommutable = 0 in {
5431 defm VPSHUFBY : SS3I_binop_rm<0x00, "vpshufb", X86pshufb, v32i8, v32i8,
5432 VR256, loadv4i64, i256mem,
5433 SSE_PSHUFB, 0>, VEX_4V, VEX_L, VEX_WIG;
5434 defm VPMADDUBSWY : SS3I_binop_rm<0x04, "vpmaddubsw", X86vpmaddubsw, v16i16,
5435 v32i8, VR256, loadv4i64, i256mem,
5436 SSE_PMADD, 0>, VEX_4V, VEX_L, VEX_WIG;
5438 defm VPMULHRSWY : SS3I_binop_rm<0x0B, "vpmulhrsw", X86mulhrs, v16i16, v16i16,
5439 VR256, loadv4i64, i256mem,
5440 SSE_PMULHRSW, 0>, VEX_4V, VEX_L, VEX_WIG;
5443 let ImmT = NoImm, Predicates = [HasAVX2] in {
5444 let isCommutable = 0 in {
5445 defm VPHADDWY : SS3I_binop_rm<0x01, "vphaddw", X86hadd, v16i16, v16i16,
5446 VR256, loadv4i64, i256mem,
5447 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L, VEX_WIG;
5448 defm VPHADDDY : SS3I_binop_rm<0x02, "vphaddd", X86hadd, v8i32, v8i32, VR256,
5450 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L, VEX_WIG;
5451 defm VPHSUBWY : SS3I_binop_rm<0x05, "vphsubw", X86hsub, v16i16, v16i16,
5452 VR256, loadv4i64, i256mem,
5453 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L, VEX_WIG;
5454 defm VPHSUBDY : SS3I_binop_rm<0x06, "vphsubd", X86hsub, v8i32, v8i32, VR256,
5456 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5457 defm VPSIGNBY : SS3I_binop_rm_int_y<0x08, "vpsignb", int_x86_avx2_psign_b,
5458 WriteVecALU>, VEX_4V, VEX_L, VEX_WIG;
5459 defm VPSIGNWY : SS3I_binop_rm_int_y<0x09, "vpsignw", int_x86_avx2_psign_w,
5460 WriteVecALU>, VEX_4V, VEX_L, VEX_WIG;
5461 defm VPSIGNDY : SS3I_binop_rm_int_y<0x0A, "vpsignd", int_x86_avx2_psign_d,
5462 WriteVecALU>, VEX_4V, VEX_L, VEX_WIG;
5463 defm VPHADDSW : SS3I_binop_rm_int_y<0x03, "vphaddsw",
5464 int_x86_avx2_phadd_sw,
5465 WriteVecALU>, VEX_4V, VEX_L, VEX_WIG;
5466 defm VPHSUBSW : SS3I_binop_rm_int_y<0x07, "vphsubsw",
5467 int_x86_avx2_phsub_sw,
5468 WriteVecALU>, VEX_4V, VEX_L, VEX_WIG;
5472 // None of these have i8 immediate fields.
5473 let ImmT = NoImm, Constraints = "$src1 = $dst" in {
5474 let isCommutable = 0 in {
5475 defm PHADDW : SS3I_binop_rm<0x01, "phaddw", X86hadd, v8i16, v8i16, VR128,
5476 memopv2i64, i128mem, SSE_PHADDSUBW>;
5477 defm PHADDD : SS3I_binop_rm<0x02, "phaddd", X86hadd, v4i32, v4i32, VR128,
5478 memopv2i64, i128mem, SSE_PHADDSUBD>;
5479 defm PHSUBW : SS3I_binop_rm<0x05, "phsubw", X86hsub, v8i16, v8i16, VR128,
5480 memopv2i64, i128mem, SSE_PHADDSUBW>;
5481 defm PHSUBD : SS3I_binop_rm<0x06, "phsubd", X86hsub, v4i32, v4i32, VR128,
5482 memopv2i64, i128mem, SSE_PHADDSUBD>;
5483 defm PSIGNB : SS3I_binop_rm_int<0x08, "psignb", int_x86_ssse3_psign_b_128,
5484 SSE_PSIGN, memopv2i64>;
5485 defm PSIGNW : SS3I_binop_rm_int<0x09, "psignw", int_x86_ssse3_psign_w_128,
5486 SSE_PSIGN, memopv2i64>;
5487 defm PSIGND : SS3I_binop_rm_int<0x0A, "psignd", int_x86_ssse3_psign_d_128,
5488 SSE_PSIGN, memopv2i64>;
5489 defm PSHUFB : SS3I_binop_rm<0x00, "pshufb", X86pshufb, v16i8, v16i8, VR128,
5490 memopv2i64, i128mem, SSE_PSHUFB>;
5491 defm PHADDSW : SS3I_binop_rm_int<0x03, "phaddsw",
5492 int_x86_ssse3_phadd_sw_128,
5493 SSE_PHADDSUBSW, memopv2i64>;
5494 defm PHSUBSW : SS3I_binop_rm_int<0x07, "phsubsw",
5495 int_x86_ssse3_phsub_sw_128,
5496 SSE_PHADDSUBSW, memopv2i64>;
5497 defm PMADDUBSW : SS3I_binop_rm<0x04, "pmaddubsw", X86vpmaddubsw, v8i16,
5498 v16i8, VR128, memopv2i64, i128mem,
5501 defm PMULHRSW : SS3I_binop_rm<0x0B, "pmulhrsw", X86mulhrs, v8i16, v8i16,
5502 VR128, memopv2i64, i128mem, SSE_PMULHRSW>;
5505 //===---------------------------------------------------------------------===//
5506 // SSSE3 - Packed Align Instruction Patterns
5507 //===---------------------------------------------------------------------===//
5509 multiclass ssse3_palignr<string asm, bit Is2Addr = 1> {
5510 let hasSideEffects = 0 in {
5511 def rri : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
5512 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
5514 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
5516 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
5517 [], IIC_SSE_PALIGNRR>, Sched<[WriteShuffle]>;
5519 def rmi : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
5520 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
5522 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
5524 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
5525 [], IIC_SSE_PALIGNRM>, Sched<[WriteShuffleLd, ReadAfterLd]>;
5529 multiclass ssse3_palignr_y<string asm, bit Is2Addr = 1> {
5530 let hasSideEffects = 0 in {
5531 def Yrri : SS3AI<0x0F, MRMSrcReg, (outs VR256:$dst),
5532 (ins VR256:$src1, VR256:$src2, u8imm:$src3),
5534 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
5535 []>, Sched<[WriteShuffle]>;
5537 def Yrmi : SS3AI<0x0F, MRMSrcMem, (outs VR256:$dst),
5538 (ins VR256:$src1, i256mem:$src2, u8imm:$src3),
5540 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
5541 []>, Sched<[WriteShuffleLd, ReadAfterLd]>;
5545 let Predicates = [HasAVX] in
5546 defm VPALIGNR : ssse3_palignr<"vpalignr", 0>, VEX_4V, VEX_WIG;
5547 let Predicates = [HasAVX2] in
5548 defm VPALIGNR : ssse3_palignr_y<"vpalignr", 0>, VEX_4V, VEX_L, VEX_WIG;
5549 let Constraints = "$src1 = $dst", Predicates = [UseSSSE3] in
5550 defm PALIGNR : ssse3_palignr<"palignr">;
5552 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
5553 def : Pat<(v8i32 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5554 (VPALIGNRYrri VR256:$src1, VR256:$src2, imm:$imm)>;
5555 def : Pat<(v8f32 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5556 (VPALIGNRYrri VR256:$src1, VR256:$src2, imm:$imm)>;
5557 def : Pat<(v16i16 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5558 (VPALIGNRYrri VR256:$src1, VR256:$src2, imm:$imm)>;
5559 def : Pat<(v32i8 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5560 (VPALIGNRYrri VR256:$src1, VR256:$src2, imm:$imm)>;
5563 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5564 def : Pat<(v4i32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5565 (VPALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5566 def : Pat<(v4f32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5567 (VPALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5568 def : Pat<(v8i16 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5569 (VPALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5570 def : Pat<(v16i8 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5571 (VPALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5574 let Predicates = [UseSSSE3] in {
5575 def : Pat<(v4i32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5576 (PALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5577 def : Pat<(v4f32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5578 (PALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5579 def : Pat<(v8i16 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5580 (PALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5581 def : Pat<(v16i8 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5582 (PALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5585 //===---------------------------------------------------------------------===//
5586 // SSSE3 - Thread synchronization
5587 //===---------------------------------------------------------------------===//
5589 let SchedRW = [WriteSystem] in {
5590 let usesCustomInserter = 1 in {
5591 def MONITOR : PseudoI<(outs), (ins i32mem:$src1, GR32:$src2, GR32:$src3),
5592 [(int_x86_sse3_monitor addr:$src1, GR32:$src2, GR32:$src3)]>,
5593 Requires<[HasSSE3]>;
5596 let Uses = [EAX, ECX, EDX] in
5597 def MONITORrrr : I<0x01, MRM_C8, (outs), (ins), "monitor", [], IIC_SSE_MONITOR>,
5598 TB, Requires<[HasSSE3]>;
5600 let Uses = [ECX, EAX] in
5601 def MWAITrr : I<0x01, MRM_C9, (outs), (ins), "mwait",
5602 [(int_x86_sse3_mwait ECX, EAX)], IIC_SSE_MWAIT>,
5603 TB, Requires<[HasSSE3]>;
5606 def : InstAlias<"mwait\t{%eax, %ecx|ecx, eax}", (MWAITrr)>, Requires<[Not64BitMode]>;
5607 def : InstAlias<"mwait\t{%rax, %rcx|rcx, rax}", (MWAITrr)>, Requires<[In64BitMode]>;
5609 def : InstAlias<"monitor\t{%eax, %ecx, %edx|edx, ecx, eax}", (MONITORrrr)>,
5610 Requires<[Not64BitMode]>;
5611 def : InstAlias<"monitor\t{%rax, %rcx, %rdx|rdx, rcx, rax}", (MONITORrrr)>,
5612 Requires<[In64BitMode]>;
5614 //===----------------------------------------------------------------------===//
5615 // SSE4.1 - Packed Move with Sign/Zero Extend
5616 //===----------------------------------------------------------------------===//
5618 multiclass SS41I_pmovx_rrrm<bits<8> opc, string OpcodeStr, X86MemOperand MemOp,
5619 RegisterClass OutRC, RegisterClass InRC,
5621 def rr : SS48I<opc, MRMSrcReg, (outs OutRC:$dst), (ins InRC:$src),
5622 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5624 Sched<[itins.Sched]>;
5626 def rm : SS48I<opc, MRMSrcMem, (outs OutRC:$dst), (ins MemOp:$src),
5627 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5629 itins.rm>, Sched<[itins.Sched.Folded]>;
5632 multiclass SS41I_pmovx_rm_all<bits<8> opc, string OpcodeStr,
5633 X86MemOperand MemOp, X86MemOperand MemYOp,
5634 OpndItins SSEItins, OpndItins AVXItins,
5635 OpndItins AVX2Itins, Predicate prd> {
5636 defm NAME : SS41I_pmovx_rrrm<opc, OpcodeStr, MemOp, VR128, VR128, SSEItins>;
5637 let Predicates = [HasAVX, prd] in
5638 defm V#NAME : SS41I_pmovx_rrrm<opc, !strconcat("v", OpcodeStr), MemOp,
5639 VR128, VR128, AVXItins>, VEX, VEX_WIG;
5640 let Predicates = [HasAVX2, prd] in
5641 defm V#NAME#Y : SS41I_pmovx_rrrm<opc, !strconcat("v", OpcodeStr), MemYOp,
5642 VR256, VR128, AVX2Itins>, VEX, VEX_L, VEX_WIG;
5645 multiclass SS41I_pmovx_rm<bits<8> opc, string OpcodeStr, X86MemOperand MemOp,
5646 X86MemOperand MemYOp, Predicate prd> {
5647 defm PMOVSX#NAME : SS41I_pmovx_rm_all<opc, !strconcat("pmovsx", OpcodeStr),
5649 SSE_INTALU_ITINS_SHUFF_P,
5650 DEFAULT_ITINS_SHUFFLESCHED,
5651 DEFAULT_ITINS_SHUFFLESCHED, prd>;
5652 defm PMOVZX#NAME : SS41I_pmovx_rm_all<!add(opc, 0x10),
5653 !strconcat("pmovzx", OpcodeStr),
5655 SSE_INTALU_ITINS_SHUFF_P,
5656 DEFAULT_ITINS_SHUFFLESCHED,
5657 DEFAULT_ITINS_SHUFFLESCHED, prd>;
5660 defm BW : SS41I_pmovx_rm<0x20, "bw", i64mem, i128mem, NoVLX_Or_NoBWI>;
5661 defm WD : SS41I_pmovx_rm<0x23, "wd", i64mem, i128mem, NoVLX>;
5662 defm DQ : SS41I_pmovx_rm<0x25, "dq", i64mem, i128mem, NoVLX>;
5664 defm BD : SS41I_pmovx_rm<0x21, "bd", i32mem, i64mem, NoVLX>;
5665 defm WQ : SS41I_pmovx_rm<0x24, "wq", i32mem, i64mem, NoVLX>;
5667 defm BQ : SS41I_pmovx_rm<0x22, "bq", i16mem, i32mem, NoVLX>;
5670 multiclass SS41I_pmovx_avx2_patterns<string OpcPrefix, string ExtTy, SDNode ExtOp> {
5671 // Register-Register patterns
5672 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5673 def : Pat<(v16i16 (ExtOp (v16i8 VR128:$src))),
5674 (!cast<I>(OpcPrefix#BWYrr) VR128:$src)>;
5676 let Predicates = [HasAVX, NoVLX] in {
5677 def : Pat<(v8i32 (ExtOp (v16i8 VR128:$src))),
5678 (!cast<I>(OpcPrefix#BDYrr) VR128:$src)>;
5679 def : Pat<(v4i64 (ExtOp (v16i8 VR128:$src))),
5680 (!cast<I>(OpcPrefix#BQYrr) VR128:$src)>;
5682 def : Pat<(v8i32 (ExtOp (v8i16 VR128:$src))),
5683 (!cast<I>(OpcPrefix#WDYrr) VR128:$src)>;
5684 def : Pat<(v4i64 (ExtOp (v8i16 VR128:$src))),
5685 (!cast<I>(OpcPrefix#WQYrr) VR128:$src)>;
5687 def : Pat<(v4i64 (ExtOp (v4i32 VR128:$src))),
5688 (!cast<I>(OpcPrefix#DQYrr) VR128:$src)>;
5691 // Simple Register-Memory patterns
5692 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5693 def : Pat<(v16i16 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5694 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5696 let Predicates = [HasAVX, NoVLX] in {
5697 def : Pat<(v8i32 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5698 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5699 def : Pat<(v4i64 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5700 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5702 def : Pat<(v8i32 (!cast<PatFrag>(ExtTy#"extloadvi16") addr:$src)),
5703 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5704 def : Pat<(v4i64 (!cast<PatFrag>(ExtTy#"extloadvi16") addr:$src)),
5705 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5707 def : Pat<(v4i64 (!cast<PatFrag>(ExtTy#"extloadvi32") addr:$src)),
5708 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5711 // AVX2 Register-Memory patterns
5712 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5713 def : Pat<(v16i16 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5714 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5715 def : Pat<(v16i16 (ExtOp (v16i8 (vzmovl_v2i64 addr:$src)))),
5716 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5717 def : Pat<(v16i16 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5718 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5720 let Predicates = [HasAVX, NoVLX] in {
5721 def : Pat<(v8i32 (ExtOp (bc_v16i8 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
5722 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5723 def : Pat<(v8i32 (ExtOp (v16i8 (vzmovl_v2i64 addr:$src)))),
5724 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5725 def : Pat<(v8i32 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5726 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5727 def : Pat<(v8i32 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5728 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5730 def : Pat<(v4i64 (ExtOp (bc_v16i8 (v4i32 (scalar_to_vector (loadi32 addr:$src)))))),
5731 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5732 def : Pat<(v4i64 (ExtOp (v16i8 (vzmovl_v4i32 addr:$src)))),
5733 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5734 def : Pat<(v4i64 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5735 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5736 def : Pat<(v4i64 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5737 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5739 def : Pat<(v8i32 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
5740 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5741 def : Pat<(v8i32 (ExtOp (v8i16 (vzmovl_v2i64 addr:$src)))),
5742 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5743 def : Pat<(v8i32 (ExtOp (v8i16 (vzload_v2i64 addr:$src)))),
5744 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5746 def : Pat<(v4i64 (ExtOp (bc_v8i16 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
5747 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5748 def : Pat<(v4i64 (ExtOp (v8i16 (vzmovl_v2i64 addr:$src)))),
5749 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5750 def : Pat<(v4i64 (ExtOp (v8i16 (vzload_v2i64 addr:$src)))),
5751 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5752 def : Pat<(v4i64 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
5753 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5755 def : Pat<(v4i64 (ExtOp (bc_v4i32 (loadv2i64 addr:$src)))),
5756 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5757 def : Pat<(v4i64 (ExtOp (v4i32 (vzmovl_v2i64 addr:$src)))),
5758 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5759 def : Pat<(v4i64 (ExtOp (v4i32 (vzload_v2i64 addr:$src)))),
5760 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5764 defm : SS41I_pmovx_avx2_patterns<"VPMOVSX", "s", X86vsext>;
5765 defm : SS41I_pmovx_avx2_patterns<"VPMOVZX", "z", X86vzext>;
5767 // SSE4.1/AVX patterns.
5768 multiclass SS41I_pmovx_patterns<string OpcPrefix, string ExtTy,
5769 SDNode ExtOp, PatFrag ExtLoad16> {
5770 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5771 def : Pat<(v8i16 (ExtOp (v16i8 VR128:$src))),
5772 (!cast<I>(OpcPrefix#BWrr) VR128:$src)>;
5774 let Predicates = [HasAVX, NoVLX] in {
5775 def : Pat<(v4i32 (ExtOp (v16i8 VR128:$src))),
5776 (!cast<I>(OpcPrefix#BDrr) VR128:$src)>;
5777 def : Pat<(v2i64 (ExtOp (v16i8 VR128:$src))),
5778 (!cast<I>(OpcPrefix#BQrr) VR128:$src)>;
5780 def : Pat<(v4i32 (ExtOp (v8i16 VR128:$src))),
5781 (!cast<I>(OpcPrefix#WDrr) VR128:$src)>;
5782 def : Pat<(v2i64 (ExtOp (v8i16 VR128:$src))),
5783 (!cast<I>(OpcPrefix#WQrr) VR128:$src)>;
5785 def : Pat<(v2i64 (ExtOp (v4i32 VR128:$src))),
5786 (!cast<I>(OpcPrefix#DQrr) VR128:$src)>;
5788 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5789 def : Pat<(v8i16 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5790 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
5792 let Predicates = [HasAVX, NoVLX] in {
5793 def : Pat<(v4i32 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5794 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
5795 def : Pat<(v2i64 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5796 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
5798 def : Pat<(v4i32 (!cast<PatFrag>(ExtTy#"extloadvi16") addr:$src)),
5799 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
5800 def : Pat<(v2i64 (!cast<PatFrag>(ExtTy#"extloadvi16") addr:$src)),
5801 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
5803 def : Pat<(v2i64 (!cast<PatFrag>(ExtTy#"extloadvi32") addr:$src)),
5804 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
5806 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5807 def : Pat<(v8i16 (ExtOp (bc_v16i8 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
5808 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
5809 def : Pat<(v8i16 (ExtOp (bc_v16i8 (v2f64 (scalar_to_vector (loadf64 addr:$src)))))),
5810 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
5811 def : Pat<(v8i16 (ExtOp (v16i8 (vzmovl_v2i64 addr:$src)))),
5812 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
5813 def : Pat<(v8i16 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5814 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
5815 def : Pat<(v8i16 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5816 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
5818 let Predicates = [HasAVX, NoVLX] in {
5819 def : Pat<(v4i32 (ExtOp (bc_v16i8 (v4i32 (scalar_to_vector (loadi32 addr:$src)))))),
5820 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
5821 def : Pat<(v4i32 (ExtOp (v16i8 (vzmovl_v4i32 addr:$src)))),
5822 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
5823 def : Pat<(v4i32 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5824 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
5825 def : Pat<(v4i32 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5826 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
5828 def : Pat<(v2i64 (ExtOp (bc_v16i8 (v4i32 (scalar_to_vector (ExtLoad16 addr:$src)))))),
5829 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
5830 def : Pat<(v2i64 (ExtOp (v16i8 (vzmovl_v4i32 addr:$src)))),
5831 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
5832 def : Pat<(v2i64 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5833 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
5834 def : Pat<(v2i64 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5835 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
5837 def : Pat<(v4i32 (ExtOp (bc_v8i16 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
5838 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
5839 def : Pat<(v4i32 (ExtOp (bc_v8i16 (v2f64 (scalar_to_vector (loadf64 addr:$src)))))),
5840 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
5841 def : Pat<(v4i32 (ExtOp (v8i16 (vzmovl_v2i64 addr:$src)))),
5842 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
5843 def : Pat<(v4i32 (ExtOp (v8i16 (vzload_v2i64 addr:$src)))),
5844 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
5845 def : Pat<(v4i32 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
5846 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
5848 def : Pat<(v2i64 (ExtOp (bc_v8i16 (v4i32 (scalar_to_vector (loadi32 addr:$src)))))),
5849 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
5850 def : Pat<(v2i64 (ExtOp (v8i16 (vzmovl_v4i32 addr:$src)))),
5851 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
5852 def : Pat<(v2i64 (ExtOp (v8i16 (vzload_v2i64 addr:$src)))),
5853 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
5854 def : Pat<(v2i64 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
5855 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
5857 def : Pat<(v2i64 (ExtOp (bc_v4i32 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
5858 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
5859 def : Pat<(v2i64 (ExtOp (bc_v4i32 (v2f64 (scalar_to_vector (loadf64 addr:$src)))))),
5860 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
5861 def : Pat<(v2i64 (ExtOp (v4i32 (vzmovl_v2i64 addr:$src)))),
5862 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
5863 def : Pat<(v2i64 (ExtOp (v4i32 (vzload_v2i64 addr:$src)))),
5864 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
5865 def : Pat<(v2i64 (ExtOp (bc_v4i32 (loadv2i64 addr:$src)))),
5866 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
5870 defm : SS41I_pmovx_patterns<"VPMOVSX", "s", sext_invec, extloadi32i16>;
5871 defm : SS41I_pmovx_patterns<"VPMOVZX", "z", zext_invec, loadi16_anyext>;
5873 let Predicates = [UseSSE41] in {
5874 defm : SS41I_pmovx_patterns<"PMOVSX", "s", sext_invec, extloadi32i16>;
5875 defm : SS41I_pmovx_patterns<"PMOVZX", "z", zext_invec, loadi16_anyext>;
5878 //===----------------------------------------------------------------------===//
5879 // SSE4.1 - Extract Instructions
5880 //===----------------------------------------------------------------------===//
5882 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
5883 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
5884 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
5885 (ins VR128:$src1, u8imm:$src2),
5886 !strconcat(OpcodeStr,
5887 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5888 [(set GR32orGR64:$dst, (X86pextrb (v16i8 VR128:$src1),
5890 Sched<[WriteShuffle]>;
5891 let hasSideEffects = 0, mayStore = 1,
5892 SchedRW = [WriteShuffleLd, WriteRMW] in
5893 def mr : SS4AIi8<opc, MRMDestMem, (outs),
5894 (ins i8mem:$dst, VR128:$src1, u8imm:$src2),
5895 !strconcat(OpcodeStr,
5896 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5897 [(store (i8 (trunc (assertzext (X86pextrb (v16i8 VR128:$src1),
5898 imm:$src2)))), addr:$dst)]>;
5901 let Predicates = [HasAVX, NoBWI] in
5902 defm VPEXTRB : SS41I_extract8<0x14, "vpextrb">, VEX;
5904 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
5907 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
5908 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
5909 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in
5910 def rr_REV : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
5911 (ins VR128:$src1, u8imm:$src2),
5912 !strconcat(OpcodeStr,
5913 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5914 []>, Sched<[WriteShuffle]>;
5916 let hasSideEffects = 0, mayStore = 1,
5917 SchedRW = [WriteShuffleLd, WriteRMW] in
5918 def mr : SS4AIi8<opc, MRMDestMem, (outs),
5919 (ins i16mem:$dst, VR128:$src1, u8imm:$src2),
5920 !strconcat(OpcodeStr,
5921 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5922 [(store (i16 (trunc (assertzext (X86pextrw (v8i16 VR128:$src1),
5923 imm:$src2)))), addr:$dst)]>;
5926 let Predicates = [HasAVX, NoBWI] in
5927 defm VPEXTRW : SS41I_extract16<0x15, "vpextrw">, VEX;
5929 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
5932 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
5933 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
5934 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
5935 (ins VR128:$src1, u8imm:$src2),
5936 !strconcat(OpcodeStr,
5937 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5939 (extractelt (v4i32 VR128:$src1), imm:$src2))]>,
5940 Sched<[WriteShuffle]>;
5941 let SchedRW = [WriteShuffleLd, WriteRMW] in
5942 def mr : SS4AIi8<opc, MRMDestMem, (outs),
5943 (ins i32mem:$dst, VR128:$src1, u8imm:$src2),
5944 !strconcat(OpcodeStr,
5945 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5946 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
5950 let Predicates = [HasAVX, NoDQI] in
5951 defm VPEXTRD : SS41I_extract32<0x16, "vpextrd">, VEX;
5953 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
5955 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
5956 multiclass SS41I_extract64<bits<8> opc, string OpcodeStr> {
5957 def rr : SS4AIi8<opc, MRMDestReg, (outs GR64:$dst),
5958 (ins VR128:$src1, u8imm:$src2),
5959 !strconcat(OpcodeStr,
5960 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5962 (extractelt (v2i64 VR128:$src1), imm:$src2))]>,
5963 Sched<[WriteShuffle]>;
5964 let SchedRW = [WriteShuffleLd, WriteRMW] in
5965 def mr : SS4AIi8<opc, MRMDestMem, (outs),
5966 (ins i64mem:$dst, VR128:$src1, u8imm:$src2),
5967 !strconcat(OpcodeStr,
5968 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5969 [(store (extractelt (v2i64 VR128:$src1), imm:$src2),
5973 let Predicates = [HasAVX, NoDQI] in
5974 defm VPEXTRQ : SS41I_extract64<0x16, "vpextrq">, VEX, VEX_W;
5976 defm PEXTRQ : SS41I_extract64<0x16, "pextrq">, REX_W;
5978 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
5980 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr,
5981 OpndItins itins = DEFAULT_ITINS> {
5982 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
5983 (ins VR128:$src1, u8imm:$src2),
5984 !strconcat(OpcodeStr,
5985 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5986 [(set GR32orGR64:$dst,
5987 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))],
5988 itins.rr>, Sched<[WriteFBlend]>;
5989 let SchedRW = [WriteFBlendLd, WriteRMW] in
5990 def mr : SS4AIi8<opc, MRMDestMem, (outs),
5991 (ins f32mem:$dst, VR128:$src1, u8imm:$src2),
5992 !strconcat(OpcodeStr,
5993 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5994 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
5995 addr:$dst)], itins.rm>;
5998 let ExeDomain = SSEPackedSingle in {
5999 let Predicates = [UseAVX] in
6000 defm VEXTRACTPS : SS41I_extractf32<0x17, "vextractps">, VEX, VEX_WIG;
6001 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps", SSE_EXTRACT_ITINS>;
6004 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
6005 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
6008 (VEXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
6010 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
6013 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
6014 Requires<[UseSSE41]>;
6016 //===----------------------------------------------------------------------===//
6017 // SSE4.1 - Insert Instructions
6018 //===----------------------------------------------------------------------===//
6020 multiclass SS41I_insert8<bits<8> opc, string asm, bit Is2Addr = 1> {
6021 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6022 (ins VR128:$src1, GR32orGR64:$src2, u8imm:$src3),
6024 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6026 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6028 (X86pinsrb VR128:$src1, GR32orGR64:$src2, imm:$src3))]>,
6029 Sched<[WriteShuffle]>;
6030 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6031 (ins VR128:$src1, i8mem:$src2, u8imm:$src3),
6033 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6035 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6037 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
6038 imm:$src3))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6041 let Predicates = [HasAVX, NoBWI] in
6042 defm VPINSRB : SS41I_insert8<0x20, "vpinsrb", 0>, VEX_4V;
6043 let Constraints = "$src1 = $dst" in
6044 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
6046 multiclass SS41I_insert32<bits<8> opc, string asm, bit Is2Addr = 1> {
6047 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6048 (ins VR128:$src1, GR32:$src2, u8imm:$src3),
6050 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6052 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6054 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
6055 Sched<[WriteShuffle]>;
6056 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6057 (ins VR128:$src1, i32mem:$src2, u8imm:$src3),
6059 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6061 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6063 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
6064 imm:$src3)))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6067 let Predicates = [HasAVX, NoDQI] in
6068 defm VPINSRD : SS41I_insert32<0x22, "vpinsrd", 0>, VEX_4V;
6069 let Constraints = "$src1 = $dst" in
6070 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
6072 multiclass SS41I_insert64<bits<8> opc, string asm, bit Is2Addr = 1> {
6073 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6074 (ins VR128:$src1, GR64:$src2, u8imm:$src3),
6076 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6078 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6080 (v2i64 (insertelt VR128:$src1, GR64:$src2, imm:$src3)))]>,
6081 Sched<[WriteShuffle]>;
6082 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6083 (ins VR128:$src1, i64mem:$src2, u8imm:$src3),
6085 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6087 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6089 (v2i64 (insertelt VR128:$src1, (loadi64 addr:$src2),
6090 imm:$src3)))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6093 let Predicates = [HasAVX, NoDQI] in
6094 defm VPINSRQ : SS41I_insert64<0x22, "vpinsrq", 0>, VEX_4V, VEX_W;
6095 let Constraints = "$src1 = $dst" in
6096 defm PINSRQ : SS41I_insert64<0x22, "pinsrq">, REX_W;
6098 // insertps has a few different modes, there's the first two here below which
6099 // are optimized inserts that won't zero arbitrary elements in the destination
6100 // vector. The next one matches the intrinsic and could zero arbitrary elements
6101 // in the target vector.
6102 multiclass SS41I_insertf32<bits<8> opc, string asm, bit Is2Addr = 1,
6103 OpndItins itins = DEFAULT_ITINS> {
6104 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6105 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
6107 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6109 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6111 (X86insertps VR128:$src1, VR128:$src2, imm:$src3))], itins.rr>,
6112 Sched<[WriteFShuffle]>;
6113 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6114 (ins VR128:$src1, f32mem:$src2, u8imm:$src3),
6116 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6118 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6120 (X86insertps VR128:$src1,
6121 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
6122 imm:$src3))], itins.rm>,
6123 Sched<[WriteFShuffleLd, ReadAfterLd]>;
6126 let ExeDomain = SSEPackedSingle in {
6127 let Predicates = [UseAVX] in
6128 defm VINSERTPS : SS41I_insertf32<0x21, "vinsertps", 0>, VEX_4V, VEX_WIG;
6129 let Constraints = "$src1 = $dst" in
6130 defm INSERTPS : SS41I_insertf32<0x21, "insertps", 1, SSE_INSERT_ITINS>;
6133 let Predicates = [UseSSE41] in {
6134 // If we're inserting an element from a load or a null pshuf of a load,
6135 // fold the load into the insertps instruction.
6136 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd (v4f32
6137 (scalar_to_vector (loadf32 addr:$src2))), (i8 0)),
6139 (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6140 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd
6141 (loadv4f32 addr:$src2), (i8 0)), imm:$src3)),
6142 (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6145 let Predicates = [UseAVX] in {
6146 // If we're inserting an element from a vbroadcast of a load, fold the
6147 // load into the X86insertps instruction.
6148 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
6149 (X86VBroadcast (loadf32 addr:$src2)), imm:$src3)),
6150 (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6151 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
6152 (X86VBroadcast (loadv4f32 addr:$src2)), imm:$src3)),
6153 (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6156 //===----------------------------------------------------------------------===//
6157 // SSE4.1 - Round Instructions
6158 //===----------------------------------------------------------------------===//
6160 multiclass sse41_fp_unop_p<bits<8> opcps, bits<8> opcpd, string OpcodeStr,
6161 X86MemOperand x86memop, RegisterClass RC,
6162 PatFrag mem_frag32, PatFrag mem_frag64,
6163 Intrinsic V4F32Int, Intrinsic V2F64Int> {
6164 let ExeDomain = SSEPackedSingle in {
6165 // Intrinsic operation, reg.
6166 // Vector intrinsic operation, reg
6167 def PSr : SS4AIi8<opcps, MRMSrcReg,
6168 (outs RC:$dst), (ins RC:$src1, i32u8imm:$src2),
6169 !strconcat(OpcodeStr,
6170 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6171 [(set RC:$dst, (V4F32Int RC:$src1, imm:$src2))],
6172 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAdd]>;
6174 // Vector intrinsic operation, mem
6175 def PSm : SS4AIi8<opcps, MRMSrcMem,
6176 (outs RC:$dst), (ins x86memop:$src1, i32u8imm:$src2),
6177 !strconcat(OpcodeStr,
6178 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6180 (V4F32Int (mem_frag32 addr:$src1),imm:$src2))],
6181 IIC_SSE_ROUNDPS_MEM>, Sched<[WriteFAddLd]>;
6182 } // ExeDomain = SSEPackedSingle
6184 let ExeDomain = SSEPackedDouble in {
6185 // Vector intrinsic operation, reg
6186 def PDr : SS4AIi8<opcpd, MRMSrcReg,
6187 (outs RC:$dst), (ins RC:$src1, i32u8imm:$src2),
6188 !strconcat(OpcodeStr,
6189 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6190 [(set RC:$dst, (V2F64Int RC:$src1, imm:$src2))],
6191 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAdd]>;
6193 // Vector intrinsic operation, mem
6194 def PDm : SS4AIi8<opcpd, MRMSrcMem,
6195 (outs RC:$dst), (ins x86memop:$src1, i32u8imm:$src2),
6196 !strconcat(OpcodeStr,
6197 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6199 (V2F64Int (mem_frag64 addr:$src1),imm:$src2))],
6200 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAddLd]>;
6201 } // ExeDomain = SSEPackedDouble
6204 multiclass avx_fp_unop_rm<bits<8> opcss, bits<8> opcsd,
6206 let ExeDomain = GenericDomain, hasSideEffects = 0 in {
6207 def SSr : SS4AIi8<opcss, MRMSrcReg,
6208 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2, i32u8imm:$src3),
6209 !strconcat(OpcodeStr,
6210 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
6211 []>, Sched<[WriteFAdd]>;
6214 def SSm : SS4AIi8<opcss, MRMSrcMem,
6215 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2, i32u8imm:$src3),
6216 !strconcat(OpcodeStr,
6217 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
6218 []>, Sched<[WriteFAddLd, ReadAfterLd]>;
6220 def SDr : SS4AIi8<opcsd, MRMSrcReg,
6221 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2, i32u8imm:$src3),
6222 !strconcat(OpcodeStr,
6223 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
6224 []>, Sched<[WriteFAdd]>;
6227 def SDm : SS4AIi8<opcsd, MRMSrcMem,
6228 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2, i32u8imm:$src3),
6229 !strconcat(OpcodeStr,
6230 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
6231 []>, Sched<[WriteFAddLd, ReadAfterLd]>;
6232 } // ExeDomain = GenericDomain, hasSideEffects = 0
6235 multiclass sse41_fp_unop_s<bits<8> opcss, bits<8> opcsd,
6237 let ExeDomain = GenericDomain, hasSideEffects = 0 in {
6238 def SSr : SS4AIi8<opcss, MRMSrcReg,
6239 (outs FR32:$dst), (ins FR32:$src1, i32u8imm:$src2),
6240 !strconcat(OpcodeStr,
6241 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6242 []>, Sched<[WriteFAdd]>;
6245 def SSm : SS4AIi8<opcss, MRMSrcMem,
6246 (outs FR32:$dst), (ins f32mem:$src1, i32u8imm:$src2),
6247 !strconcat(OpcodeStr,
6248 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6249 []>, Sched<[WriteFAddLd, ReadAfterLd]>;
6251 def SDr : SS4AIi8<opcsd, MRMSrcReg,
6252 (outs FR64:$dst), (ins FR64:$src1, i32u8imm:$src2),
6253 !strconcat(OpcodeStr,
6254 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6255 []>, Sched<[WriteFAdd]>;
6258 def SDm : SS4AIi8<opcsd, MRMSrcMem,
6259 (outs FR64:$dst), (ins f64mem:$src1, i32u8imm:$src2),
6260 !strconcat(OpcodeStr,
6261 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6262 []>, Sched<[WriteFAddLd, ReadAfterLd]>;
6263 } // ExeDomain = GenericDomain, hasSideEffects = 0
6266 multiclass sse41_fp_binop_s<bits<8> opcss, bits<8> opcsd,
6269 Intrinsic F64Int, bit Is2Addr = 1> {
6270 let ExeDomain = GenericDomain, isCodeGenOnly = 1 in {
6271 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
6272 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32u8imm:$src3),
6274 !strconcat(OpcodeStr,
6275 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6276 !strconcat(OpcodeStr,
6277 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6278 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
6281 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
6282 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2, i32u8imm:$src3),
6284 !strconcat(OpcodeStr,
6285 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6286 !strconcat(OpcodeStr,
6287 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6289 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
6290 Sched<[WriteFAddLd, ReadAfterLd]>;
6292 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
6293 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32u8imm:$src3),
6295 !strconcat(OpcodeStr,
6296 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6297 !strconcat(OpcodeStr,
6298 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6299 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
6302 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
6303 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2, i32u8imm:$src3),
6305 !strconcat(OpcodeStr,
6306 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6307 !strconcat(OpcodeStr,
6308 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6310 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
6311 Sched<[WriteFAddLd, ReadAfterLd]>;
6312 } // ExeDomain = GenericDomain, isCodeGenOnly = 1
6315 // FP round - roundss, roundps, roundsd, roundpd
6316 let Predicates = [HasAVX] in {
6318 defm VROUND : sse41_fp_unop_p<0x08, 0x09, "vround", f128mem, VR128,
6319 loadv4f32, loadv2f64,
6320 int_x86_sse41_round_ps,
6321 int_x86_sse41_round_pd>, VEX, VEX_WIG;
6322 defm VROUNDY : sse41_fp_unop_p<0x08, 0x09, "vround", f256mem, VR256,
6323 loadv8f32, loadv4f64,
6324 int_x86_avx_round_ps_256,
6325 int_x86_avx_round_pd_256>, VEX, VEX_L, VEX_WIG;
6326 defm VROUND : sse41_fp_binop_s<0x0A, 0x0B, "vround",
6327 int_x86_sse41_round_ss,
6328 int_x86_sse41_round_sd, 0>, VEX_4V, VEX_LIG, VEX_WIG;
6329 defm VROUND : avx_fp_unop_rm<0x0A, 0x0B, "vround">, VEX_4V, VEX_LIG;
6332 let Predicates = [UseAVX] in {
6333 def : Pat<(ffloor FR32:$src),
6334 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x9))>;
6335 def : Pat<(f64 (ffloor FR64:$src)),
6336 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x9))>;
6337 def : Pat<(f32 (fnearbyint FR32:$src)),
6338 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0xC))>;
6339 def : Pat<(f64 (fnearbyint FR64:$src)),
6340 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0xC))>;
6341 def : Pat<(f32 (fceil FR32:$src)),
6342 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0xA))>;
6343 def : Pat<(f64 (fceil FR64:$src)),
6344 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0xA))>;
6345 def : Pat<(f32 (frint FR32:$src)),
6346 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x4))>;
6347 def : Pat<(f64 (frint FR64:$src)),
6348 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x4))>;
6349 def : Pat<(f32 (ftrunc FR32:$src)),
6350 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0xB))>;
6351 def : Pat<(f64 (ftrunc FR64:$src)),
6352 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0xB))>;
6355 let Predicates = [HasAVX] in {
6356 def : Pat<(v4f32 (ffloor VR128:$src)),
6357 (VROUNDPSr VR128:$src, (i32 0x9))>;
6358 def : Pat<(v4f32 (fnearbyint VR128:$src)),
6359 (VROUNDPSr VR128:$src, (i32 0xC))>;
6360 def : Pat<(v4f32 (fceil VR128:$src)),
6361 (VROUNDPSr VR128:$src, (i32 0xA))>;
6362 def : Pat<(v4f32 (frint VR128:$src)),
6363 (VROUNDPSr VR128:$src, (i32 0x4))>;
6364 def : Pat<(v4f32 (ftrunc VR128:$src)),
6365 (VROUNDPSr VR128:$src, (i32 0xB))>;
6367 def : Pat<(v2f64 (ffloor VR128:$src)),
6368 (VROUNDPDr VR128:$src, (i32 0x9))>;
6369 def : Pat<(v2f64 (fnearbyint VR128:$src)),
6370 (VROUNDPDr VR128:$src, (i32 0xC))>;
6371 def : Pat<(v2f64 (fceil VR128:$src)),
6372 (VROUNDPDr VR128:$src, (i32 0xA))>;
6373 def : Pat<(v2f64 (frint VR128:$src)),
6374 (VROUNDPDr VR128:$src, (i32 0x4))>;
6375 def : Pat<(v2f64 (ftrunc VR128:$src)),
6376 (VROUNDPDr VR128:$src, (i32 0xB))>;
6378 def : Pat<(v8f32 (ffloor VR256:$src)),
6379 (VROUNDYPSr VR256:$src, (i32 0x9))>;
6380 def : Pat<(v8f32 (fnearbyint VR256:$src)),
6381 (VROUNDYPSr VR256:$src, (i32 0xC))>;
6382 def : Pat<(v8f32 (fceil VR256:$src)),
6383 (VROUNDYPSr VR256:$src, (i32 0xA))>;
6384 def : Pat<(v8f32 (frint VR256:$src)),
6385 (VROUNDYPSr VR256:$src, (i32 0x4))>;
6386 def : Pat<(v8f32 (ftrunc VR256:$src)),
6387 (VROUNDYPSr VR256:$src, (i32 0xB))>;
6389 def : Pat<(v4f64 (ffloor VR256:$src)),
6390 (VROUNDYPDr VR256:$src, (i32 0x9))>;
6391 def : Pat<(v4f64 (fnearbyint VR256:$src)),
6392 (VROUNDYPDr VR256:$src, (i32 0xC))>;
6393 def : Pat<(v4f64 (fceil VR256:$src)),
6394 (VROUNDYPDr VR256:$src, (i32 0xA))>;
6395 def : Pat<(v4f64 (frint VR256:$src)),
6396 (VROUNDYPDr VR256:$src, (i32 0x4))>;
6397 def : Pat<(v4f64 (ftrunc VR256:$src)),
6398 (VROUNDYPDr VR256:$src, (i32 0xB))>;
6401 defm ROUND : sse41_fp_unop_p<0x08, 0x09, "round", f128mem, VR128,
6402 memopv4f32, memopv2f64, int_x86_sse41_round_ps,
6403 int_x86_sse41_round_pd>;
6405 defm ROUND : sse41_fp_unop_s<0x0A, 0x0B, "round">;
6407 let Constraints = "$src1 = $dst" in
6408 defm ROUND : sse41_fp_binop_s<0x0A, 0x0B, "round",
6409 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
6411 let Predicates = [UseSSE41] in {
6412 def : Pat<(ffloor FR32:$src),
6413 (ROUNDSSr FR32:$src, (i32 0x9))>;
6414 def : Pat<(f64 (ffloor FR64:$src)),
6415 (ROUNDSDr FR64:$src, (i32 0x9))>;
6416 def : Pat<(f32 (fnearbyint FR32:$src)),
6417 (ROUNDSSr FR32:$src, (i32 0xC))>;
6418 def : Pat<(f64 (fnearbyint FR64:$src)),
6419 (ROUNDSDr FR64:$src, (i32 0xC))>;
6420 def : Pat<(f32 (fceil FR32:$src)),
6421 (ROUNDSSr FR32:$src, (i32 0xA))>;
6422 def : Pat<(f64 (fceil FR64:$src)),
6423 (ROUNDSDr FR64:$src, (i32 0xA))>;
6424 def : Pat<(f32 (frint FR32:$src)),
6425 (ROUNDSSr FR32:$src, (i32 0x4))>;
6426 def : Pat<(f64 (frint FR64:$src)),
6427 (ROUNDSDr FR64:$src, (i32 0x4))>;
6428 def : Pat<(f32 (ftrunc FR32:$src)),
6429 (ROUNDSSr FR32:$src, (i32 0xB))>;
6430 def : Pat<(f64 (ftrunc FR64:$src)),
6431 (ROUNDSDr FR64:$src, (i32 0xB))>;
6433 def : Pat<(v4f32 (ffloor VR128:$src)),
6434 (ROUNDPSr VR128:$src, (i32 0x9))>;
6435 def : Pat<(v4f32 (fnearbyint VR128:$src)),
6436 (ROUNDPSr VR128:$src, (i32 0xC))>;
6437 def : Pat<(v4f32 (fceil VR128:$src)),
6438 (ROUNDPSr VR128:$src, (i32 0xA))>;
6439 def : Pat<(v4f32 (frint VR128:$src)),
6440 (ROUNDPSr VR128:$src, (i32 0x4))>;
6441 def : Pat<(v4f32 (ftrunc VR128:$src)),
6442 (ROUNDPSr VR128:$src, (i32 0xB))>;
6444 def : Pat<(v2f64 (ffloor VR128:$src)),
6445 (ROUNDPDr VR128:$src, (i32 0x9))>;
6446 def : Pat<(v2f64 (fnearbyint VR128:$src)),
6447 (ROUNDPDr VR128:$src, (i32 0xC))>;
6448 def : Pat<(v2f64 (fceil VR128:$src)),
6449 (ROUNDPDr VR128:$src, (i32 0xA))>;
6450 def : Pat<(v2f64 (frint VR128:$src)),
6451 (ROUNDPDr VR128:$src, (i32 0x4))>;
6452 def : Pat<(v2f64 (ftrunc VR128:$src)),
6453 (ROUNDPDr VR128:$src, (i32 0xB))>;
6456 //===----------------------------------------------------------------------===//
6457 // SSE4.1 - Packed Bit Test
6458 //===----------------------------------------------------------------------===//
6460 // ptest instruction we'll lower to this in X86ISelLowering primarily from
6461 // the intel intrinsic that corresponds to this.
6462 let Defs = [EFLAGS], Predicates = [HasAVX] in {
6463 def VPTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
6464 "vptest\t{$src2, $src1|$src1, $src2}",
6465 [(set EFLAGS, (X86ptest VR128:$src1, (v2i64 VR128:$src2)))]>,
6466 Sched<[WriteVecLogic]>, VEX, VEX_WIG;
6467 def VPTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
6468 "vptest\t{$src2, $src1|$src1, $src2}",
6469 [(set EFLAGS,(X86ptest VR128:$src1, (loadv2i64 addr:$src2)))]>,
6470 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX, VEX_WIG;
6472 def VPTESTYrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR256:$src1, VR256:$src2),
6473 "vptest\t{$src2, $src1|$src1, $src2}",
6474 [(set EFLAGS, (X86ptest VR256:$src1, (v4i64 VR256:$src2)))]>,
6475 Sched<[WriteVecLogic]>, VEX, VEX_L, VEX_WIG;
6476 def VPTESTYrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR256:$src1, i256mem:$src2),
6477 "vptest\t{$src2, $src1|$src1, $src2}",
6478 [(set EFLAGS,(X86ptest VR256:$src1, (loadv4i64 addr:$src2)))]>,
6479 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX, VEX_L, VEX_WIG;
6482 let Defs = [EFLAGS] in {
6483 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
6484 "ptest\t{$src2, $src1|$src1, $src2}",
6485 [(set EFLAGS, (X86ptest VR128:$src1, (v2i64 VR128:$src2)))]>,
6486 Sched<[WriteVecLogic]>;
6487 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
6488 "ptest\t{$src2, $src1|$src1, $src2}",
6489 [(set EFLAGS, (X86ptest VR128:$src1, (memopv2i64 addr:$src2)))]>,
6490 Sched<[WriteVecLogicLd, ReadAfterLd]>;
6493 // The bit test instructions below are AVX only
6494 multiclass avx_bittest<bits<8> opc, string OpcodeStr, RegisterClass RC,
6495 X86MemOperand x86memop, PatFrag mem_frag, ValueType vt> {
6496 def rr : SS48I<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
6497 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
6498 [(set EFLAGS, (X86testp RC:$src1, (vt RC:$src2)))]>,
6499 Sched<[WriteVecLogic]>, VEX;
6500 def rm : SS48I<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
6501 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
6502 [(set EFLAGS, (X86testp RC:$src1, (mem_frag addr:$src2)))]>,
6503 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX;
6506 let Defs = [EFLAGS], Predicates = [HasAVX] in {
6507 let ExeDomain = SSEPackedSingle in {
6508 defm VTESTPS : avx_bittest<0x0E, "vtestps", VR128, f128mem, loadv4f32, v4f32>;
6509 defm VTESTPSY : avx_bittest<0x0E, "vtestps", VR256, f256mem, loadv8f32, v8f32>,
6512 let ExeDomain = SSEPackedDouble in {
6513 defm VTESTPD : avx_bittest<0x0F, "vtestpd", VR128, f128mem, loadv2f64, v2f64>;
6514 defm VTESTPDY : avx_bittest<0x0F, "vtestpd", VR256, f256mem, loadv4f64, v4f64>,
6519 //===----------------------------------------------------------------------===//
6520 // SSE4.1 - Misc Instructions
6521 //===----------------------------------------------------------------------===//
6523 let Defs = [EFLAGS], Predicates = [HasPOPCNT] in {
6524 def POPCNT16rr : I<0xB8, MRMSrcReg, (outs GR16:$dst), (ins GR16:$src),
6525 "popcnt{w}\t{$src, $dst|$dst, $src}",
6526 [(set GR16:$dst, (ctpop GR16:$src)), (implicit EFLAGS)],
6527 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>,
6529 def POPCNT16rm : I<0xB8, MRMSrcMem, (outs GR16:$dst), (ins i16mem:$src),
6530 "popcnt{w}\t{$src, $dst|$dst, $src}",
6531 [(set GR16:$dst, (ctpop (loadi16 addr:$src))),
6532 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
6533 Sched<[WriteFAddLd]>, OpSize16, XS;
6535 def POPCNT32rr : I<0xB8, MRMSrcReg, (outs GR32:$dst), (ins GR32:$src),
6536 "popcnt{l}\t{$src, $dst|$dst, $src}",
6537 [(set GR32:$dst, (ctpop GR32:$src)), (implicit EFLAGS)],
6538 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>,
6541 def POPCNT32rm : I<0xB8, MRMSrcMem, (outs GR32:$dst), (ins i32mem:$src),
6542 "popcnt{l}\t{$src, $dst|$dst, $src}",
6543 [(set GR32:$dst, (ctpop (loadi32 addr:$src))),
6544 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
6545 Sched<[WriteFAddLd]>, OpSize32, XS;
6547 def POPCNT64rr : RI<0xB8, MRMSrcReg, (outs GR64:$dst), (ins GR64:$src),
6548 "popcnt{q}\t{$src, $dst|$dst, $src}",
6549 [(set GR64:$dst, (ctpop GR64:$src)), (implicit EFLAGS)],
6550 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>, XS;
6551 def POPCNT64rm : RI<0xB8, MRMSrcMem, (outs GR64:$dst), (ins i64mem:$src),
6552 "popcnt{q}\t{$src, $dst|$dst, $src}",
6553 [(set GR64:$dst, (ctpop (loadi64 addr:$src))),
6554 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
6555 Sched<[WriteFAddLd]>, XS;
6560 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
6561 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
6562 Intrinsic IntId128, PatFrag ld_frag,
6563 X86FoldableSchedWrite Sched> {
6564 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
6566 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6567 [(set VR128:$dst, (IntId128 VR128:$src))]>,
6569 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
6571 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6573 (IntId128 (bitconvert (ld_frag addr:$src))))]>,
6574 Sched<[Sched.Folded]>;
6577 // PHMIN has the same profile as PSAD, thus we use the same scheduling
6578 // model, although the naming is misleading.
6579 let Predicates = [HasAVX] in
6580 defm VPHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "vphminposuw",
6581 int_x86_sse41_phminposuw, loadv2i64,
6582 WriteVecIMul>, VEX, VEX_WIG;
6583 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
6584 int_x86_sse41_phminposuw, memopv2i64,
6587 /// SS48I_binop_rm - Simple SSE41 binary operator.
6588 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
6589 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
6590 X86MemOperand x86memop, bit Is2Addr = 1,
6591 OpndItins itins = SSE_INTALU_ITINS_P> {
6592 let isCommutable = 1 in
6593 def rr : SS48I<opc, MRMSrcReg, (outs RC:$dst),
6594 (ins RC:$src1, RC:$src2),
6596 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
6597 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
6598 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))]>,
6599 Sched<[itins.Sched]>;
6600 def rm : SS48I<opc, MRMSrcMem, (outs RC:$dst),
6601 (ins RC:$src1, x86memop:$src2),
6603 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
6604 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
6606 (OpVT (OpNode RC:$src1, (bitconvert (memop_frag addr:$src2)))))]>,
6607 Sched<[itins.Sched.Folded, ReadAfterLd]>;
6610 /// SS48I_binop_rm2 - Simple SSE41 binary operator with different src and dst
6612 multiclass SS48I_binop_rm2<bits<8> opc, string OpcodeStr, SDNode OpNode,
6613 ValueType DstVT, ValueType SrcVT, RegisterClass RC,
6614 PatFrag memop_frag, X86MemOperand x86memop,
6616 bit IsCommutable = 0, bit Is2Addr = 1> {
6617 let isCommutable = IsCommutable in
6618 def rr : SS48I<opc, MRMSrcReg, (outs RC:$dst),
6619 (ins RC:$src1, RC:$src2),
6621 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
6622 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
6623 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1), RC:$src2)))]>,
6624 Sched<[itins.Sched]>;
6625 def rm : SS48I<opc, MRMSrcMem, (outs RC:$dst),
6626 (ins RC:$src1, x86memop:$src2),
6628 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
6629 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
6630 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1),
6631 (bitconvert (memop_frag addr:$src2)))))]>,
6632 Sched<[itins.Sched.Folded, ReadAfterLd]>;
6635 let Predicates = [HasAVX, NoVLX] in {
6636 defm VPMINSD : SS48I_binop_rm<0x39, "vpminsd", smin, v4i32, VR128,
6637 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6639 defm VPMINUD : SS48I_binop_rm<0x3B, "vpminud", umin, v4i32, VR128,
6640 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6642 defm VPMAXSD : SS48I_binop_rm<0x3D, "vpmaxsd", smax, v4i32, VR128,
6643 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6645 defm VPMAXUD : SS48I_binop_rm<0x3F, "vpmaxud", umax, v4i32, VR128,
6646 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6648 defm VPMULDQ : SS48I_binop_rm2<0x28, "vpmuldq", X86pmuldq, v2i64, v4i32,
6649 VR128, loadv2i64, i128mem,
6650 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V, VEX_WIG;
6652 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
6653 defm VPMINSB : SS48I_binop_rm<0x38, "vpminsb", smin, v16i8, VR128,
6654 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6656 defm VPMINUW : SS48I_binop_rm<0x3A, "vpminuw", umin, v8i16, VR128,
6657 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6659 defm VPMAXSB : SS48I_binop_rm<0x3C, "vpmaxsb", smax, v16i8, VR128,
6660 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6662 defm VPMAXUW : SS48I_binop_rm<0x3E, "vpmaxuw", umax, v8i16, VR128,
6663 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6667 let Predicates = [HasAVX2, NoVLX] in {
6668 defm VPMINSDY : SS48I_binop_rm<0x39, "vpminsd", smin, v8i32, VR256,
6669 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6670 VEX_4V, VEX_L, VEX_WIG;
6671 defm VPMINUDY : SS48I_binop_rm<0x3B, "vpminud", umin, v8i32, VR256,
6672 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6673 VEX_4V, VEX_L, VEX_WIG;
6674 defm VPMAXSDY : SS48I_binop_rm<0x3D, "vpmaxsd", smax, v8i32, VR256,
6675 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6676 VEX_4V, VEX_L, VEX_WIG;
6677 defm VPMAXUDY : SS48I_binop_rm<0x3F, "vpmaxud", umax, v8i32, VR256,
6678 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6679 VEX_4V, VEX_L, VEX_WIG;
6680 defm VPMULDQY : SS48I_binop_rm2<0x28, "vpmuldq", X86pmuldq, v4i64, v8i32,
6681 VR256, loadv4i64, i256mem,
6682 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V, VEX_L, VEX_WIG;
6684 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
6685 defm VPMINSBY : SS48I_binop_rm<0x38, "vpminsb", smin, v32i8, VR256,
6686 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6687 VEX_4V, VEX_L, VEX_WIG;
6688 defm VPMINUWY : SS48I_binop_rm<0x3A, "vpminuw", umin, v16i16, VR256,
6689 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6690 VEX_4V, VEX_L, VEX_WIG;
6691 defm VPMAXSBY : SS48I_binop_rm<0x3C, "vpmaxsb", smax, v32i8, VR256,
6692 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6693 VEX_4V, VEX_L, VEX_WIG;
6694 defm VPMAXUWY : SS48I_binop_rm<0x3E, "vpmaxuw", umax, v16i16, VR256,
6695 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6696 VEX_4V, VEX_L, VEX_WIG;
6699 let Constraints = "$src1 = $dst" in {
6700 defm PMINSB : SS48I_binop_rm<0x38, "pminsb", smin, v16i8, VR128,
6701 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6702 defm PMINSD : SS48I_binop_rm<0x39, "pminsd", smin, v4i32, VR128,
6703 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6704 defm PMINUD : SS48I_binop_rm<0x3B, "pminud", umin, v4i32, VR128,
6705 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6706 defm PMINUW : SS48I_binop_rm<0x3A, "pminuw", umin, v8i16, VR128,
6707 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6708 defm PMAXSB : SS48I_binop_rm<0x3C, "pmaxsb", smax, v16i8, VR128,
6709 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6710 defm PMAXSD : SS48I_binop_rm<0x3D, "pmaxsd", smax, v4i32, VR128,
6711 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6712 defm PMAXUD : SS48I_binop_rm<0x3F, "pmaxud", umax, v4i32, VR128,
6713 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6714 defm PMAXUW : SS48I_binop_rm<0x3E, "pmaxuw", umax, v8i16, VR128,
6715 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6716 defm PMULDQ : SS48I_binop_rm2<0x28, "pmuldq", X86pmuldq, v2i64, v4i32,
6717 VR128, memopv2i64, i128mem,
6718 SSE_INTMUL_ITINS_P, 1>;
6721 let Predicates = [HasAVX, NoVLX] in {
6722 defm VPMULLD : SS48I_binop_rm<0x40, "vpmulld", mul, v4i32, VR128,
6723 loadv2i64, i128mem, 0, SSE_PMULLD_ITINS>,
6725 defm VPCMPEQQ : SS48I_binop_rm<0x29, "vpcmpeqq", X86pcmpeq, v2i64, VR128,
6726 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6729 let Predicates = [HasAVX2] in {
6730 defm VPMULLDY : SS48I_binop_rm<0x40, "vpmulld", mul, v8i32, VR256,
6731 loadv4i64, i256mem, 0, SSE_PMULLD_ITINS>,
6732 VEX_4V, VEX_L, VEX_WIG;
6733 defm VPCMPEQQY : SS48I_binop_rm<0x29, "vpcmpeqq", X86pcmpeq, v4i64, VR256,
6734 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6735 VEX_4V, VEX_L, VEX_WIG;
6738 let Constraints = "$src1 = $dst" in {
6739 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, VR128,
6740 memopv2i64, i128mem, 1, SSE_PMULLD_ITINS>;
6741 defm PCMPEQQ : SS48I_binop_rm<0x29, "pcmpeqq", X86pcmpeq, v2i64, VR128,
6742 memopv2i64, i128mem, 1, SSE_INTALUQ_ITINS_P>;
6745 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
6746 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
6747 Intrinsic IntId, RegisterClass RC, PatFrag memop_frag,
6748 X86MemOperand x86memop, bit Is2Addr = 1,
6749 OpndItins itins = DEFAULT_ITINS> {
6750 let isCommutable = 1 in
6751 def rri : SS4AIi8<opc, MRMSrcReg, (outs RC:$dst),
6752 (ins RC:$src1, RC:$src2, u8imm:$src3),
6754 !strconcat(OpcodeStr,
6755 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6756 !strconcat(OpcodeStr,
6757 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6758 [(set RC:$dst, (IntId RC:$src1, RC:$src2, imm:$src3))], itins.rr>,
6759 Sched<[itins.Sched]>;
6760 def rmi : SS4AIi8<opc, MRMSrcMem, (outs RC:$dst),
6761 (ins RC:$src1, x86memop:$src2, u8imm:$src3),
6763 !strconcat(OpcodeStr,
6764 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6765 !strconcat(OpcodeStr,
6766 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6769 (bitconvert (memop_frag addr:$src2)), imm:$src3))], itins.rm>,
6770 Sched<[itins.Sched.Folded, ReadAfterLd]>;
6773 /// SS41I_binop_rmi - SSE 4.1 binary operator with 8-bit immediate
6774 multiclass SS41I_binop_rmi<bits<8> opc, string OpcodeStr, SDNode OpNode,
6775 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
6776 X86MemOperand x86memop, bit Is2Addr = 1,
6777 OpndItins itins = DEFAULT_ITINS> {
6778 let isCommutable = 1 in
6779 def rri : SS4AIi8<opc, MRMSrcReg, (outs RC:$dst),
6780 (ins RC:$src1, RC:$src2, u8imm:$src3),
6782 !strconcat(OpcodeStr,
6783 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6784 !strconcat(OpcodeStr,
6785 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6786 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2, imm:$src3)))],
6787 itins.rr>, Sched<[itins.Sched]>;
6788 def rmi : SS4AIi8<opc, MRMSrcMem, (outs RC:$dst),
6789 (ins RC:$src1, x86memop:$src2, u8imm:$src3),
6791 !strconcat(OpcodeStr,
6792 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6793 !strconcat(OpcodeStr,
6794 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6796 (OpVT (OpNode RC:$src1,
6797 (bitconvert (memop_frag addr:$src2)), imm:$src3)))], itins.rm>,
6798 Sched<[itins.Sched.Folded, ReadAfterLd]>;
6801 let Predicates = [HasAVX] in {
6802 let isCommutable = 0 in {
6803 defm VMPSADBW : SS41I_binop_rmi_int<0x42, "vmpsadbw", int_x86_sse41_mpsadbw,
6804 VR128, loadv2i64, i128mem, 0,
6805 DEFAULT_ITINS_MPSADSCHED>, VEX_4V, VEX_WIG;
6808 let ExeDomain = SSEPackedSingle in {
6809 defm VBLENDPS : SS41I_binop_rmi<0x0C, "vblendps", X86Blendi, v4f32,
6810 VR128, loadv4f32, f128mem, 0,
6811 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V, VEX_WIG;
6812 defm VBLENDPSY : SS41I_binop_rmi<0x0C, "vblendps", X86Blendi, v8f32,
6813 VR256, loadv8f32, f256mem, 0,
6814 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V, VEX_L, VEX_WIG;
6816 let ExeDomain = SSEPackedDouble in {
6817 defm VBLENDPD : SS41I_binop_rmi<0x0D, "vblendpd", X86Blendi, v2f64,
6818 VR128, loadv2f64, f128mem, 0,
6819 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V, VEX_WIG;
6820 defm VBLENDPDY : SS41I_binop_rmi<0x0D, "vblendpd", X86Blendi, v4f64,
6821 VR256, loadv4f64, f256mem, 0,
6822 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V, VEX_L, VEX_WIG;
6824 defm VPBLENDW : SS41I_binop_rmi<0x0E, "vpblendw", X86Blendi, v8i16,
6825 VR128, loadv2i64, i128mem, 0,
6826 DEFAULT_ITINS_BLENDSCHED>, VEX_4V, VEX_WIG;
6828 let ExeDomain = SSEPackedSingle in
6829 defm VDPPS : SS41I_binop_rmi_int<0x40, "vdpps", int_x86_sse41_dpps,
6830 VR128, loadv4f32, f128mem, 0,
6831 SSE_DPPS_ITINS>, VEX_4V, VEX_WIG;
6832 let ExeDomain = SSEPackedDouble in
6833 defm VDPPD : SS41I_binop_rmi_int<0x41, "vdppd", int_x86_sse41_dppd,
6834 VR128, loadv2f64, f128mem, 0,
6835 SSE_DPPS_ITINS>, VEX_4V, VEX_WIG;
6836 let ExeDomain = SSEPackedSingle in
6837 defm VDPPSY : SS41I_binop_rmi_int<0x40, "vdpps", int_x86_avx_dp_ps_256,
6838 VR256, loadv8f32, i256mem, 0,
6839 SSE_DPPS_ITINS>, VEX_4V, VEX_L, VEX_WIG;
6842 let Predicates = [HasAVX2] in {
6843 let isCommutable = 0 in {
6844 defm VMPSADBWY : SS41I_binop_rmi_int<0x42, "vmpsadbw", int_x86_avx2_mpsadbw,
6845 VR256, loadv4i64, i256mem, 0,
6846 DEFAULT_ITINS_MPSADSCHED>, VEX_4V, VEX_L, VEX_WIG;
6848 defm VPBLENDWY : SS41I_binop_rmi<0x0E, "vpblendw", X86Blendi, v16i16,
6849 VR256, loadv4i64, i256mem, 0,
6850 DEFAULT_ITINS_BLENDSCHED>, VEX_4V, VEX_L, VEX_WIG;
6853 let Constraints = "$src1 = $dst" in {
6854 let isCommutable = 0 in {
6855 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw", int_x86_sse41_mpsadbw,
6856 VR128, memopv2i64, i128mem,
6857 1, SSE_MPSADBW_ITINS>;
6859 let ExeDomain = SSEPackedSingle in
6860 defm BLENDPS : SS41I_binop_rmi<0x0C, "blendps", X86Blendi, v4f32,
6861 VR128, memopv4f32, f128mem,
6862 1, SSE_INTALU_ITINS_FBLEND_P>;
6863 let ExeDomain = SSEPackedDouble in
6864 defm BLENDPD : SS41I_binop_rmi<0x0D, "blendpd", X86Blendi, v2f64,
6865 VR128, memopv2f64, f128mem,
6866 1, SSE_INTALU_ITINS_FBLEND_P>;
6867 defm PBLENDW : SS41I_binop_rmi<0x0E, "pblendw", X86Blendi, v8i16,
6868 VR128, memopv2i64, i128mem,
6869 1, SSE_INTALU_ITINS_BLEND_P>;
6870 let ExeDomain = SSEPackedSingle in
6871 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps", int_x86_sse41_dpps,
6872 VR128, memopv4f32, f128mem, 1,
6874 let ExeDomain = SSEPackedDouble in
6875 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd", int_x86_sse41_dppd,
6876 VR128, memopv2f64, f128mem, 1,
6880 // For insertion into the zero index (low half) of a 256-bit vector, it is
6881 // more efficient to generate a blend with immediate instead of an insert*128.
6882 let Predicates = [HasAVX] in {
6883 def : Pat<(insert_subvector (v4f64 VR256:$src1), (v2f64 VR128:$src2), (iPTR 0)),
6884 (VBLENDPDYrri VR256:$src1,
6885 (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)),
6886 VR128:$src2, sub_xmm), 0x3)>;
6887 def : Pat<(insert_subvector (v8f32 VR256:$src1), (v4f32 VR128:$src2), (iPTR 0)),
6888 (VBLENDPSYrri VR256:$src1,
6889 (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)),
6890 VR128:$src2, sub_xmm), 0xf)>;
6893 /// SS41I_quaternary_int_avx - AVX SSE 4.1 with 4 operators
6894 multiclass SS41I_quaternary_int_avx<bits<8> opc, string OpcodeStr,
6895 RegisterClass RC, X86MemOperand x86memop,
6896 PatFrag mem_frag, Intrinsic IntId,
6897 X86FoldableSchedWrite Sched> {
6898 def rr : Ii8Reg<opc, MRMSrcReg, (outs RC:$dst),
6899 (ins RC:$src1, RC:$src2, RC:$src3),
6900 !strconcat(OpcodeStr,
6901 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
6902 [(set RC:$dst, (IntId RC:$src1, RC:$src2, RC:$src3))],
6903 NoItinerary, SSEPackedInt>, TAPD, VEX_4V,
6906 def rm : Ii8Reg<opc, MRMSrcMem, (outs RC:$dst),
6907 (ins RC:$src1, x86memop:$src2, RC:$src3),
6908 !strconcat(OpcodeStr,
6909 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
6911 (IntId RC:$src1, (bitconvert (mem_frag addr:$src2)),
6913 NoItinerary, SSEPackedInt>, TAPD, VEX_4V,
6914 Sched<[Sched.Folded, ReadAfterLd]>;
6917 let Predicates = [HasAVX] in {
6918 let ExeDomain = SSEPackedDouble in {
6919 defm VBLENDVPD : SS41I_quaternary_int_avx<0x4B, "vblendvpd", VR128, f128mem,
6920 loadv2f64, int_x86_sse41_blendvpd,
6922 defm VBLENDVPDY : SS41I_quaternary_int_avx<0x4B, "vblendvpd", VR256, f256mem,
6923 loadv4f64, int_x86_avx_blendv_pd_256,
6924 WriteFVarBlend>, VEX_L;
6925 } // ExeDomain = SSEPackedDouble
6926 let ExeDomain = SSEPackedSingle in {
6927 defm VBLENDVPS : SS41I_quaternary_int_avx<0x4A, "vblendvps", VR128, f128mem,
6928 loadv4f32, int_x86_sse41_blendvps,
6930 defm VBLENDVPSY : SS41I_quaternary_int_avx<0x4A, "vblendvps", VR256, f256mem,
6931 loadv8f32, int_x86_avx_blendv_ps_256,
6932 WriteFVarBlend>, VEX_L;
6933 } // ExeDomain = SSEPackedSingle
6934 defm VPBLENDVB : SS41I_quaternary_int_avx<0x4C, "vpblendvb", VR128, i128mem,
6935 loadv2i64, int_x86_sse41_pblendvb,
6939 let Predicates = [HasAVX2] in {
6940 defm VPBLENDVBY : SS41I_quaternary_int_avx<0x4C, "vpblendvb", VR256, i256mem,
6941 loadv4i64, int_x86_avx2_pblendvb,
6942 WriteVarBlend>, VEX_L;
6945 let Predicates = [HasAVX] in {
6946 def : Pat<(v16i8 (vselect (v16i8 VR128:$mask), (v16i8 VR128:$src1),
6947 (v16i8 VR128:$src2))),
6948 (VPBLENDVBrr VR128:$src2, VR128:$src1, VR128:$mask)>;
6949 def : Pat<(v4i32 (vselect (v4i32 VR128:$mask), (v4i32 VR128:$src1),
6950 (v4i32 VR128:$src2))),
6951 (VBLENDVPSrr VR128:$src2, VR128:$src1, VR128:$mask)>;
6952 def : Pat<(v4f32 (vselect (v4i32 VR128:$mask), (v4f32 VR128:$src1),
6953 (v4f32 VR128:$src2))),
6954 (VBLENDVPSrr VR128:$src2, VR128:$src1, VR128:$mask)>;
6955 def : Pat<(v2i64 (vselect (v2i64 VR128:$mask), (v2i64 VR128:$src1),
6956 (v2i64 VR128:$src2))),
6957 (VBLENDVPDrr VR128:$src2, VR128:$src1, VR128:$mask)>;
6958 def : Pat<(v2f64 (vselect (v2i64 VR128:$mask), (v2f64 VR128:$src1),
6959 (v2f64 VR128:$src2))),
6960 (VBLENDVPDrr VR128:$src2, VR128:$src1, VR128:$mask)>;
6961 def : Pat<(v8i32 (vselect (v8i32 VR256:$mask), (v8i32 VR256:$src1),
6962 (v8i32 VR256:$src2))),
6963 (VBLENDVPSYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
6964 def : Pat<(v8f32 (vselect (v8i32 VR256:$mask), (v8f32 VR256:$src1),
6965 (v8f32 VR256:$src2))),
6966 (VBLENDVPSYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
6967 def : Pat<(v4i64 (vselect (v4i64 VR256:$mask), (v4i64 VR256:$src1),
6968 (v4i64 VR256:$src2))),
6969 (VBLENDVPDYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
6970 def : Pat<(v4f64 (vselect (v4i64 VR256:$mask), (v4f64 VR256:$src1),
6971 (v4f64 VR256:$src2))),
6972 (VBLENDVPDYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
6975 let Predicates = [HasAVX2] in {
6976 def : Pat<(v32i8 (vselect (v32i8 VR256:$mask), (v32i8 VR256:$src1),
6977 (v32i8 VR256:$src2))),
6978 (VPBLENDVBYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
6982 // FIXME: Prefer a movss or movsd over a blendps when optimizing for size or
6983 // on targets where they have equal performance. These were changed to use
6984 // blends because blends have better throughput on SandyBridge and Haswell, but
6985 // movs[s/d] are 1-2 byte shorter instructions.
6986 let Predicates = [UseAVX] in {
6987 let AddedComplexity = 15 in {
6988 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
6989 // MOVS{S,D} to the lower bits.
6990 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
6991 (VMOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
6992 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
6993 (VBLENDPSrri (v4f32 (V_SET0)), VR128:$src, (i8 1))>;
6994 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
6995 (VPBLENDWrri (v4i32 (V_SET0)), VR128:$src, (i8 3))>;
6996 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
6997 (VMOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
6999 // Move low f32 and clear high bits.
7000 def : Pat<(v8f32 (X86vzmovl (v8f32 VR256:$src))),
7001 (VBLENDPSYrri (v8f32 (AVX_SET0)), VR256:$src, (i8 1))>;
7003 // Move low f64 and clear high bits.
7004 def : Pat<(v4f64 (X86vzmovl (v4f64 VR256:$src))),
7005 (VBLENDPDYrri (v4f64 (AVX_SET0)), VR256:$src, (i8 1))>;
7008 // These will incur an FP/int domain crossing penalty, but it may be the only
7009 // way without AVX2. Do not add any complexity because we may be able to match
7010 // more optimal patterns defined earlier in this file.
7011 def : Pat<(v8i32 (X86vzmovl (v8i32 VR256:$src))),
7012 (VBLENDPSYrri (v8i32 (AVX_SET0)), VR256:$src, (i8 1))>;
7013 def : Pat<(v4i64 (X86vzmovl (v4i64 VR256:$src))),
7014 (VBLENDPDYrri (v4i64 (AVX_SET0)), VR256:$src, (i8 1))>;
7017 // FIXME: Prefer a movss or movsd over a blendps when optimizing for size or
7018 // on targets where they have equal performance. These were changed to use
7019 // blends because blends have better throughput on SandyBridge and Haswell, but
7020 // movs[s/d] are 1-2 byte shorter instructions.
7021 let Predicates = [UseSSE41], AddedComplexity = 15 in {
7022 // With SSE41 we can use blends for these patterns.
7023 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
7024 (BLENDPSrri (v4f32 (V_SET0)), VR128:$src, (i8 1))>;
7025 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
7026 (PBLENDWrri (v4i32 (V_SET0)), VR128:$src, (i8 3))>;
7030 /// SS41I_ternary_int - SSE 4.1 ternary operator
7031 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
7032 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
7033 X86MemOperand x86memop, Intrinsic IntId,
7034 OpndItins itins = DEFAULT_ITINS> {
7035 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
7036 (ins VR128:$src1, VR128:$src2),
7037 !strconcat(OpcodeStr,
7038 "\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}"),
7039 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))],
7040 itins.rr>, Sched<[itins.Sched]>;
7042 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
7043 (ins VR128:$src1, x86memop:$src2),
7044 !strconcat(OpcodeStr,
7045 "\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}"),
7048 (bitconvert (mem_frag addr:$src2)), XMM0))],
7049 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
7053 let ExeDomain = SSEPackedDouble in
7054 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", memopv2f64, f128mem,
7055 int_x86_sse41_blendvpd,
7056 DEFAULT_ITINS_FBLENDSCHED>;
7057 let ExeDomain = SSEPackedSingle in
7058 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", memopv4f32, f128mem,
7059 int_x86_sse41_blendvps,
7060 DEFAULT_ITINS_FBLENDSCHED>;
7061 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", memopv2i64, i128mem,
7062 int_x86_sse41_pblendvb,
7063 DEFAULT_ITINS_VARBLENDSCHED>;
7065 // Aliases with the implicit xmm0 argument
7066 def : InstAlias<"blendvpd\t{$src2, $dst|$dst, $src2}",
7067 (BLENDVPDrr0 VR128:$dst, VR128:$src2), 0>;
7068 def : InstAlias<"blendvpd\t{$src2, $dst|$dst, $src2}",
7069 (BLENDVPDrm0 VR128:$dst, f128mem:$src2), 0>;
7070 def : InstAlias<"blendvps\t{$src2, $dst|$dst, $src2}",
7071 (BLENDVPSrr0 VR128:$dst, VR128:$src2), 0>;
7072 def : InstAlias<"blendvps\t{$src2, $dst|$dst, $src2}",
7073 (BLENDVPSrm0 VR128:$dst, f128mem:$src2), 0>;
7074 def : InstAlias<"pblendvb\t{$src2, $dst|$dst, $src2}",
7075 (PBLENDVBrr0 VR128:$dst, VR128:$src2), 0>;
7076 def : InstAlias<"pblendvb\t{$src2, $dst|$dst, $src2}",
7077 (PBLENDVBrm0 VR128:$dst, i128mem:$src2), 0>;
7079 let Predicates = [UseSSE41] in {
7080 def : Pat<(v16i8 (vselect (v16i8 XMM0), (v16i8 VR128:$src1),
7081 (v16i8 VR128:$src2))),
7082 (PBLENDVBrr0 VR128:$src2, VR128:$src1)>;
7083 def : Pat<(v4i32 (vselect (v4i32 XMM0), (v4i32 VR128:$src1),
7084 (v4i32 VR128:$src2))),
7085 (BLENDVPSrr0 VR128:$src2, VR128:$src1)>;
7086 def : Pat<(v4f32 (vselect (v4i32 XMM0), (v4f32 VR128:$src1),
7087 (v4f32 VR128:$src2))),
7088 (BLENDVPSrr0 VR128:$src2, VR128:$src1)>;
7089 def : Pat<(v2i64 (vselect (v2i64 XMM0), (v2i64 VR128:$src1),
7090 (v2i64 VR128:$src2))),
7091 (BLENDVPDrr0 VR128:$src2, VR128:$src1)>;
7092 def : Pat<(v2f64 (vselect (v2i64 XMM0), (v2f64 VR128:$src1),
7093 (v2f64 VR128:$src2))),
7094 (BLENDVPDrr0 VR128:$src2, VR128:$src1)>;
7097 let AddedComplexity = 400 in { // Prefer non-temporal versions
7098 let SchedRW = [WriteLoad] in {
7099 let Predicates = [HasAVX, NoVLX] in
7100 def VMOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
7101 "vmovntdqa\t{$src, $dst|$dst, $src}", []>,
7103 let Predicates = [HasAVX2, NoVLX] in
7104 def VMOVNTDQAYrm : SS48I<0x2A, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
7105 "vmovntdqa\t{$src, $dst|$dst, $src}", []>,
7106 VEX, VEX_L, VEX_WIG;
7107 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
7108 "movntdqa\t{$src, $dst|$dst, $src}", []>;
7111 let Predicates = [HasAVX2, NoVLX] in {
7112 def : Pat<(v8f32 (alignednontemporalload addr:$src)),
7113 (VMOVNTDQAYrm addr:$src)>;
7114 def : Pat<(v4f64 (alignednontemporalload addr:$src)),
7115 (VMOVNTDQAYrm addr:$src)>;
7116 def : Pat<(v4i64 (alignednontemporalload addr:$src)),
7117 (VMOVNTDQAYrm addr:$src)>;
7120 let Predicates = [HasAVX, NoVLX] in {
7121 def : Pat<(v4f32 (alignednontemporalload addr:$src)),
7122 (VMOVNTDQArm addr:$src)>;
7123 def : Pat<(v2f64 (alignednontemporalload addr:$src)),
7124 (VMOVNTDQArm addr:$src)>;
7125 def : Pat<(v2i64 (alignednontemporalload addr:$src)),
7126 (VMOVNTDQArm addr:$src)>;
7129 let Predicates = [UseSSE41] in {
7130 def : Pat<(v4f32 (alignednontemporalload addr:$src)),
7131 (MOVNTDQArm addr:$src)>;
7132 def : Pat<(v2f64 (alignednontemporalload addr:$src)),
7133 (MOVNTDQArm addr:$src)>;
7134 def : Pat<(v2i64 (alignednontemporalload addr:$src)),
7135 (MOVNTDQArm addr:$src)>;
7138 } // AddedComplexity
7140 //===----------------------------------------------------------------------===//
7141 // SSE4.2 - Compare Instructions
7142 //===----------------------------------------------------------------------===//
7144 /// SS42I_binop_rm - Simple SSE 4.2 binary operator
7145 multiclass SS42I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
7146 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
7147 X86MemOperand x86memop, OpndItins itins,
7149 def rr : SS428I<opc, MRMSrcReg, (outs RC:$dst),
7150 (ins RC:$src1, RC:$src2),
7152 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7153 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7154 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))]>, Sched<[itins.Sched]>;
7155 def rm : SS428I<opc, MRMSrcMem, (outs RC:$dst),
7156 (ins RC:$src1, x86memop:$src2),
7158 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7159 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7161 (OpVT (OpNode RC:$src1, (memop_frag addr:$src2))))]>,
7162 Sched<[itins.Sched.Folded, ReadAfterLd]>;
7165 let Predicates = [HasAVX] in
7166 defm VPCMPGTQ : SS42I_binop_rm<0x37, "vpcmpgtq", X86pcmpgt, v2i64, VR128,
7167 loadv2i64, i128mem, SSE_INTALU_ITINS_P, 0>,
7170 let Predicates = [HasAVX2] in
7171 defm VPCMPGTQY : SS42I_binop_rm<0x37, "vpcmpgtq", X86pcmpgt, v4i64, VR256,
7172 loadv4i64, i256mem, SSE_INTALU_ITINS_P, 0>,
7173 VEX_4V, VEX_L, VEX_WIG;
7175 let Constraints = "$src1 = $dst" in
7176 defm PCMPGTQ : SS42I_binop_rm<0x37, "pcmpgtq", X86pcmpgt, v2i64, VR128,
7177 memopv2i64, i128mem, SSE_INTALU_ITINS_P>;
7179 //===----------------------------------------------------------------------===//
7180 // SSE4.2 - String/text Processing Instructions
7181 //===----------------------------------------------------------------------===//
7183 // Packed Compare Implicit Length Strings, Return Mask
7184 multiclass pseudo_pcmpistrm<string asm, PatFrag ld_frag> {
7185 def REG : PseudoI<(outs VR128:$dst),
7186 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7187 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
7189 def MEM : PseudoI<(outs VR128:$dst),
7190 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7191 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1,
7192 (bc_v16i8 (ld_frag addr:$src2)), imm:$src3))]>;
7195 let Defs = [EFLAGS], usesCustomInserter = 1 in {
7196 defm VPCMPISTRM128 : pseudo_pcmpistrm<"#VPCMPISTRM128", loadv2i64>,
7197 Requires<[HasAVX]>, VEX_WIG;
7198 defm PCMPISTRM128 : pseudo_pcmpistrm<"#PCMPISTRM128", memopv2i64>,
7199 Requires<[UseSSE42]>;
7202 multiclass pcmpistrm_SS42AI<string asm> {
7203 def rr : SS42AI<0x62, MRMSrcReg, (outs),
7204 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7205 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7206 []>, Sched<[WritePCmpIStrM]>;
7208 def rm :SS42AI<0x62, MRMSrcMem, (outs),
7209 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7210 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7211 []>, Sched<[WritePCmpIStrMLd, ReadAfterLd]>;
7214 let Defs = [XMM0, EFLAGS], hasSideEffects = 0 in {
7215 let Predicates = [HasAVX] in
7216 defm VPCMPISTRM128 : pcmpistrm_SS42AI<"vpcmpistrm">, VEX;
7217 defm PCMPISTRM128 : pcmpistrm_SS42AI<"pcmpistrm"> ;
7220 // Packed Compare Explicit Length Strings, Return Mask
7221 multiclass pseudo_pcmpestrm<string asm, PatFrag ld_frag> {
7222 def REG : PseudoI<(outs VR128:$dst),
7223 (ins VR128:$src1, VR128:$src3, u8imm:$src5),
7224 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
7225 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>;
7226 def MEM : PseudoI<(outs VR128:$dst),
7227 (ins VR128:$src1, i128mem:$src3, u8imm:$src5),
7228 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128 VR128:$src1, EAX,
7229 (bc_v16i8 (ld_frag addr:$src3)), EDX, imm:$src5))]>;
7232 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
7233 defm VPCMPESTRM128 : pseudo_pcmpestrm<"#VPCMPESTRM128", loadv2i64>,
7235 defm PCMPESTRM128 : pseudo_pcmpestrm<"#PCMPESTRM128", memopv2i64>,
7236 Requires<[UseSSE42]>;
7239 multiclass SS42AI_pcmpestrm<string asm> {
7240 def rr : SS42AI<0x60, MRMSrcReg, (outs),
7241 (ins VR128:$src1, VR128:$src3, u8imm:$src5),
7242 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7243 []>, Sched<[WritePCmpEStrM]>;
7245 def rm : SS42AI<0x60, MRMSrcMem, (outs),
7246 (ins VR128:$src1, i128mem:$src3, u8imm:$src5),
7247 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7248 []>, Sched<[WritePCmpEStrMLd, ReadAfterLd]>;
7251 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX], hasSideEffects = 0 in {
7252 let Predicates = [HasAVX] in
7253 defm VPCMPESTRM128 : SS42AI_pcmpestrm<"vpcmpestrm">, VEX;
7254 defm PCMPESTRM128 : SS42AI_pcmpestrm<"pcmpestrm">;
7257 // Packed Compare Implicit Length Strings, Return Index
7258 multiclass pseudo_pcmpistri<string asm, PatFrag ld_frag> {
7259 def REG : PseudoI<(outs GR32:$dst),
7260 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7261 [(set GR32:$dst, EFLAGS,
7262 (X86pcmpistri VR128:$src1, VR128:$src2, imm:$src3))]>;
7263 def MEM : PseudoI<(outs GR32:$dst),
7264 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7265 [(set GR32:$dst, EFLAGS, (X86pcmpistri VR128:$src1,
7266 (bc_v16i8 (ld_frag addr:$src2)), imm:$src3))]>;
7269 let Defs = [EFLAGS], usesCustomInserter = 1 in {
7270 defm VPCMPISTRI : pseudo_pcmpistri<"#VPCMPISTRI", loadv2i64>,
7271 Requires<[HasAVX]>, VEX_WIG;
7272 defm PCMPISTRI : pseudo_pcmpistri<"#PCMPISTRI", memopv2i64>,
7273 Requires<[UseSSE42]>;
7276 multiclass SS42AI_pcmpistri<string asm> {
7277 def rr : SS42AI<0x63, MRMSrcReg, (outs),
7278 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7279 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7280 []>, Sched<[WritePCmpIStrI]>;
7282 def rm : SS42AI<0x63, MRMSrcMem, (outs),
7283 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7284 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7285 []>, Sched<[WritePCmpIStrILd, ReadAfterLd]>;
7288 let Defs = [ECX, EFLAGS], hasSideEffects = 0 in {
7289 let Predicates = [HasAVX] in
7290 defm VPCMPISTRI : SS42AI_pcmpistri<"vpcmpistri">, VEX;
7291 defm PCMPISTRI : SS42AI_pcmpistri<"pcmpistri">;
7294 // Packed Compare Explicit Length Strings, Return Index
7295 multiclass pseudo_pcmpestri<string asm, PatFrag ld_frag> {
7296 def REG : PseudoI<(outs GR32:$dst),
7297 (ins VR128:$src1, VR128:$src3, u8imm:$src5),
7298 [(set GR32:$dst, EFLAGS,
7299 (X86pcmpestri VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>;
7300 def MEM : PseudoI<(outs GR32:$dst),
7301 (ins VR128:$src1, i128mem:$src3, u8imm:$src5),
7302 [(set GR32:$dst, EFLAGS,
7303 (X86pcmpestri VR128:$src1, EAX, (bc_v16i8 (ld_frag addr:$src3)), EDX,
7307 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
7308 defm VPCMPESTRI : pseudo_pcmpestri<"#VPCMPESTRI", loadv2i64>,
7310 defm PCMPESTRI : pseudo_pcmpestri<"#PCMPESTRI", memopv2i64>,
7311 Requires<[UseSSE42]>;
7314 multiclass SS42AI_pcmpestri<string asm> {
7315 def rr : SS42AI<0x61, MRMSrcReg, (outs),
7316 (ins VR128:$src1, VR128:$src3, u8imm:$src5),
7317 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7318 []>, Sched<[WritePCmpEStrI]>;
7320 def rm : SS42AI<0x61, MRMSrcMem, (outs),
7321 (ins VR128:$src1, i128mem:$src3, u8imm:$src5),
7322 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7323 []>, Sched<[WritePCmpEStrILd, ReadAfterLd]>;
7326 let Defs = [ECX, EFLAGS], Uses = [EAX, EDX], hasSideEffects = 0 in {
7327 let Predicates = [HasAVX] in
7328 defm VPCMPESTRI : SS42AI_pcmpestri<"vpcmpestri">, VEX;
7329 defm PCMPESTRI : SS42AI_pcmpestri<"pcmpestri">;
7332 //===----------------------------------------------------------------------===//
7333 // SSE4.2 - CRC Instructions
7334 //===----------------------------------------------------------------------===//
7336 // No CRC instructions have AVX equivalents
7338 // crc intrinsic instruction
7339 // This set of instructions are only rm, the only difference is the size
7341 class SS42I_crc32r<bits<8> opc, string asm, RegisterClass RCOut,
7342 RegisterClass RCIn, SDPatternOperator Int> :
7343 SS42FI<opc, MRMSrcReg, (outs RCOut:$dst), (ins RCOut:$src1, RCIn:$src2),
7344 !strconcat(asm, "\t{$src2, $src1|$src1, $src2}"),
7345 [(set RCOut:$dst, (Int RCOut:$src1, RCIn:$src2))], IIC_CRC32_REG>,
7348 class SS42I_crc32m<bits<8> opc, string asm, RegisterClass RCOut,
7349 X86MemOperand x86memop, SDPatternOperator Int> :
7350 SS42FI<opc, MRMSrcMem, (outs RCOut:$dst), (ins RCOut:$src1, x86memop:$src2),
7351 !strconcat(asm, "\t{$src2, $src1|$src1, $src2}"),
7352 [(set RCOut:$dst, (Int RCOut:$src1, (load addr:$src2)))],
7353 IIC_CRC32_MEM>, Sched<[WriteFAddLd, ReadAfterLd]>;
7355 let Constraints = "$src1 = $dst" in {
7356 def CRC32r32m8 : SS42I_crc32m<0xF0, "crc32{b}", GR32, i8mem,
7357 int_x86_sse42_crc32_32_8>;
7358 def CRC32r32r8 : SS42I_crc32r<0xF0, "crc32{b}", GR32, GR8,
7359 int_x86_sse42_crc32_32_8>;
7360 def CRC32r32m16 : SS42I_crc32m<0xF1, "crc32{w}", GR32, i16mem,
7361 int_x86_sse42_crc32_32_16>, OpSize16;
7362 def CRC32r32r16 : SS42I_crc32r<0xF1, "crc32{w}", GR32, GR16,
7363 int_x86_sse42_crc32_32_16>, OpSize16;
7364 def CRC32r32m32 : SS42I_crc32m<0xF1, "crc32{l}", GR32, i32mem,
7365 int_x86_sse42_crc32_32_32>, OpSize32;
7366 def CRC32r32r32 : SS42I_crc32r<0xF1, "crc32{l}", GR32, GR32,
7367 int_x86_sse42_crc32_32_32>, OpSize32;
7368 def CRC32r64m64 : SS42I_crc32m<0xF1, "crc32{q}", GR64, i64mem,
7369 int_x86_sse42_crc32_64_64>, REX_W;
7370 def CRC32r64r64 : SS42I_crc32r<0xF1, "crc32{q}", GR64, GR64,
7371 int_x86_sse42_crc32_64_64>, REX_W;
7372 let hasSideEffects = 0 in {
7374 def CRC32r64m8 : SS42I_crc32m<0xF0, "crc32{b}", GR64, i8mem,
7376 def CRC32r64r8 : SS42I_crc32r<0xF0, "crc32{b}", GR64, GR8,
7381 //===----------------------------------------------------------------------===//
7382 // SHA-NI Instructions
7383 //===----------------------------------------------------------------------===//
7385 multiclass SHAI_binop<bits<8> Opc, string OpcodeStr, Intrinsic IntId,
7387 def rr : I<Opc, MRMSrcReg, (outs VR128:$dst),
7388 (ins VR128:$src1, VR128:$src2),
7390 !strconcat(OpcodeStr, "\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}"),
7391 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}")),
7393 (set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0)),
7394 (set VR128:$dst, (IntId VR128:$src1, VR128:$src2)))]>, T8;
7396 def rm : I<Opc, MRMSrcMem, (outs VR128:$dst),
7397 (ins VR128:$src1, i128mem:$src2),
7399 !strconcat(OpcodeStr, "\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}"),
7400 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}")),
7402 (set VR128:$dst, (IntId VR128:$src1,
7403 (bc_v4i32 (memopv2i64 addr:$src2)), XMM0)),
7404 (set VR128:$dst, (IntId VR128:$src1,
7405 (bc_v4i32 (memopv2i64 addr:$src2)))))]>, T8;
7408 let Constraints = "$src1 = $dst", Predicates = [HasSHA] in {
7409 def SHA1RNDS4rri : Ii8<0xCC, MRMSrcReg, (outs VR128:$dst),
7410 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7411 "sha1rnds4\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7413 (int_x86_sha1rnds4 VR128:$src1, VR128:$src2,
7414 (i8 imm:$src3)))]>, TA;
7415 def SHA1RNDS4rmi : Ii8<0xCC, MRMSrcMem, (outs VR128:$dst),
7416 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7417 "sha1rnds4\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7419 (int_x86_sha1rnds4 VR128:$src1,
7420 (bc_v4i32 (memopv2i64 addr:$src2)),
7421 (i8 imm:$src3)))]>, TA;
7423 defm SHA1NEXTE : SHAI_binop<0xC8, "sha1nexte", int_x86_sha1nexte>;
7424 defm SHA1MSG1 : SHAI_binop<0xC9, "sha1msg1", int_x86_sha1msg1>;
7425 defm SHA1MSG2 : SHAI_binop<0xCA, "sha1msg2", int_x86_sha1msg2>;
7428 defm SHA256RNDS2 : SHAI_binop<0xCB, "sha256rnds2", int_x86_sha256rnds2, 1>;
7430 defm SHA256MSG1 : SHAI_binop<0xCC, "sha256msg1", int_x86_sha256msg1>;
7431 defm SHA256MSG2 : SHAI_binop<0xCD, "sha256msg2", int_x86_sha256msg2>;
7434 // Aliases with explicit %xmm0
7435 def : InstAlias<"sha256rnds2\t{$src2, $dst|$dst, $src2}",
7436 (SHA256RNDS2rr VR128:$dst, VR128:$src2), 0>;
7437 def : InstAlias<"sha256rnds2\t{$src2, $dst|$dst, $src2}",
7438 (SHA256RNDS2rm VR128:$dst, i128mem:$src2), 0>;
7440 //===----------------------------------------------------------------------===//
7441 // AES-NI Instructions
7442 //===----------------------------------------------------------------------===//
7444 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId128,
7445 PatFrag ld_frag, bit Is2Addr = 1> {
7446 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
7447 (ins VR128:$src1, VR128:$src2),
7449 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7450 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7451 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
7452 Sched<[WriteAESDecEnc]>;
7453 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
7454 (ins VR128:$src1, i128mem:$src2),
7456 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7457 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7459 (IntId128 VR128:$src1, (ld_frag addr:$src2)))]>,
7460 Sched<[WriteAESDecEncLd, ReadAfterLd]>;
7463 // Perform One Round of an AES Encryption/Decryption Flow
7464 let Predicates = [HasAVX, HasAES] in {
7465 defm VAESENC : AESI_binop_rm_int<0xDC, "vaesenc",
7466 int_x86_aesni_aesenc, loadv2i64, 0>, VEX_4V, VEX_WIG;
7467 defm VAESENCLAST : AESI_binop_rm_int<0xDD, "vaesenclast",
7468 int_x86_aesni_aesenclast, loadv2i64, 0>, VEX_4V, VEX_WIG;
7469 defm VAESDEC : AESI_binop_rm_int<0xDE, "vaesdec",
7470 int_x86_aesni_aesdec, loadv2i64, 0>, VEX_4V, VEX_WIG;
7471 defm VAESDECLAST : AESI_binop_rm_int<0xDF, "vaesdeclast",
7472 int_x86_aesni_aesdeclast, loadv2i64, 0>, VEX_4V, VEX_WIG;
7475 let Constraints = "$src1 = $dst" in {
7476 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
7477 int_x86_aesni_aesenc, memopv2i64>;
7478 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
7479 int_x86_aesni_aesenclast, memopv2i64>;
7480 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
7481 int_x86_aesni_aesdec, memopv2i64>;
7482 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
7483 int_x86_aesni_aesdeclast, memopv2i64>;
7486 // Perform the AES InvMixColumn Transformation
7487 let Predicates = [HasAVX, HasAES] in {
7488 def VAESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
7490 "vaesimc\t{$src1, $dst|$dst, $src1}",
7492 (int_x86_aesni_aesimc VR128:$src1))]>, Sched<[WriteAESIMC]>,
7494 def VAESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
7495 (ins i128mem:$src1),
7496 "vaesimc\t{$src1, $dst|$dst, $src1}",
7497 [(set VR128:$dst, (int_x86_aesni_aesimc (loadv2i64 addr:$src1)))]>,
7498 Sched<[WriteAESIMCLd]>, VEX, VEX_WIG;
7500 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
7502 "aesimc\t{$src1, $dst|$dst, $src1}",
7504 (int_x86_aesni_aesimc VR128:$src1))]>, Sched<[WriteAESIMC]>;
7505 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
7506 (ins i128mem:$src1),
7507 "aesimc\t{$src1, $dst|$dst, $src1}",
7508 [(set VR128:$dst, (int_x86_aesni_aesimc (memopv2i64 addr:$src1)))]>,
7509 Sched<[WriteAESIMCLd]>;
7511 // AES Round Key Generation Assist
7512 let Predicates = [HasAVX, HasAES] in {
7513 def VAESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
7514 (ins VR128:$src1, u8imm:$src2),
7515 "vaeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7517 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
7518 Sched<[WriteAESKeyGen]>, VEX, VEX_WIG;
7519 def VAESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
7520 (ins i128mem:$src1, u8imm:$src2),
7521 "vaeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7523 (int_x86_aesni_aeskeygenassist (loadv2i64 addr:$src1), imm:$src2))]>,
7524 Sched<[WriteAESKeyGenLd]>, VEX, VEX_WIG;
7526 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
7527 (ins VR128:$src1, u8imm:$src2),
7528 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7530 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
7531 Sched<[WriteAESKeyGen]>;
7532 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
7533 (ins i128mem:$src1, u8imm:$src2),
7534 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7536 (int_x86_aesni_aeskeygenassist (memopv2i64 addr:$src1), imm:$src2))]>,
7537 Sched<[WriteAESKeyGenLd]>;
7539 //===----------------------------------------------------------------------===//
7540 // PCLMUL Instructions
7541 //===----------------------------------------------------------------------===//
7543 // AVX carry-less Multiplication instructions
7544 let isCommutable = 1 in
7545 def VPCLMULQDQrr : AVXPCLMULIi8<0x44, MRMSrcReg, (outs VR128:$dst),
7546 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7547 "vpclmulqdq\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7549 (int_x86_pclmulqdq VR128:$src1, VR128:$src2, imm:$src3))]>,
7550 Sched<[WriteCLMul]>, VEX_WIG;
7552 def VPCLMULQDQrm : AVXPCLMULIi8<0x44, MRMSrcMem, (outs VR128:$dst),
7553 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7554 "vpclmulqdq\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7555 [(set VR128:$dst, (int_x86_pclmulqdq VR128:$src1,
7556 (loadv2i64 addr:$src2), imm:$src3))]>,
7557 Sched<[WriteCLMulLd, ReadAfterLd]>, VEX_WIG;
7559 // Carry-less Multiplication instructions
7560 let Constraints = "$src1 = $dst" in {
7561 let isCommutable = 1 in
7562 def PCLMULQDQrr : PCLMULIi8<0x44, MRMSrcReg, (outs VR128:$dst),
7563 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7564 "pclmulqdq\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7566 (int_x86_pclmulqdq VR128:$src1, VR128:$src2, imm:$src3))],
7567 IIC_SSE_PCLMULQDQ_RR>, Sched<[WriteCLMul]>;
7569 def PCLMULQDQrm : PCLMULIi8<0x44, MRMSrcMem, (outs VR128:$dst),
7570 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7571 "pclmulqdq\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7572 [(set VR128:$dst, (int_x86_pclmulqdq VR128:$src1,
7573 (memopv2i64 addr:$src2), imm:$src3))],
7574 IIC_SSE_PCLMULQDQ_RM>,
7575 Sched<[WriteCLMulLd, ReadAfterLd]>;
7576 } // Constraints = "$src1 = $dst"
7579 multiclass pclmul_alias<string asm, int immop> {
7580 def : InstAlias<!strconcat("pclmul", asm, "dq {$src, $dst|$dst, $src}"),
7581 (PCLMULQDQrr VR128:$dst, VR128:$src, immop), 0>;
7583 def : InstAlias<!strconcat("pclmul", asm, "dq {$src, $dst|$dst, $src}"),
7584 (PCLMULQDQrm VR128:$dst, i128mem:$src, immop), 0>;
7586 def : InstAlias<!strconcat("vpclmul", asm,
7587 "dq {$src2, $src1, $dst|$dst, $src1, $src2}"),
7588 (VPCLMULQDQrr VR128:$dst, VR128:$src1, VR128:$src2, immop),
7591 def : InstAlias<!strconcat("vpclmul", asm,
7592 "dq {$src2, $src1, $dst|$dst, $src1, $src2}"),
7593 (VPCLMULQDQrm VR128:$dst, VR128:$src1, i128mem:$src2, immop),
7596 defm : pclmul_alias<"hqhq", 0x11>;
7597 defm : pclmul_alias<"hqlq", 0x01>;
7598 defm : pclmul_alias<"lqhq", 0x10>;
7599 defm : pclmul_alias<"lqlq", 0x00>;
7601 //===----------------------------------------------------------------------===//
7602 // SSE4A Instructions
7603 //===----------------------------------------------------------------------===//
7605 let Predicates = [HasSSE4A] in {
7607 let ExeDomain = SSEPackedInt in {
7608 let Constraints = "$src = $dst" in {
7609 def EXTRQI : Ii8<0x78, MRMXr, (outs VR128:$dst),
7610 (ins VR128:$src, u8imm:$len, u8imm:$idx),
7611 "extrq\t{$idx, $len, $src|$src, $len, $idx}",
7612 [(set VR128:$dst, (X86extrqi VR128:$src, imm:$len,
7614 def EXTRQ : I<0x79, MRMSrcReg, (outs VR128:$dst),
7615 (ins VR128:$src, VR128:$mask),
7616 "extrq\t{$mask, $src|$src, $mask}",
7617 [(set VR128:$dst, (int_x86_sse4a_extrq VR128:$src,
7618 VR128:$mask))]>, PD;
7620 def INSERTQI : Ii8<0x78, MRMSrcReg, (outs VR128:$dst),
7621 (ins VR128:$src, VR128:$src2, u8imm:$len, u8imm:$idx),
7622 "insertq\t{$idx, $len, $src2, $src|$src, $src2, $len, $idx}",
7623 [(set VR128:$dst, (X86insertqi VR128:$src, VR128:$src2,
7624 imm:$len, imm:$idx))]>, XD;
7625 def INSERTQ : I<0x79, MRMSrcReg, (outs VR128:$dst),
7626 (ins VR128:$src, VR128:$mask),
7627 "insertq\t{$mask, $src|$src, $mask}",
7628 [(set VR128:$dst, (int_x86_sse4a_insertq VR128:$src,
7629 VR128:$mask))]>, XD;
7631 } // ExeDomain = SSEPackedInt
7633 // Non-temporal (unaligned) scalar stores.
7634 let AddedComplexity = 400 in { // Prefer non-temporal versions
7635 let mayStore = 1, SchedRW = [WriteStore] in {
7636 def MOVNTSS : I<0x2B, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
7637 "movntss\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVNT>, XS;
7639 def MOVNTSD : I<0x2B, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
7640 "movntsd\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVNT>, XD;
7643 def : Pat<(nontemporalstore FR32:$src, addr:$dst),
7644 (MOVNTSS addr:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
7646 def : Pat<(nontemporalstore FR64:$src, addr:$dst),
7647 (MOVNTSD addr:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
7649 } // AddedComplexity
7652 //===----------------------------------------------------------------------===//
7654 //===----------------------------------------------------------------------===//
7656 //===----------------------------------------------------------------------===//
7657 // VBROADCAST - Load from memory and broadcast to all elements of the
7658 // destination operand
7660 class avx_broadcast_rm<bits<8> opc, string OpcodeStr, RegisterClass RC,
7661 X86MemOperand x86memop, ValueType VT,
7662 PatFrag ld_frag, SchedWrite Sched> :
7663 AVX8I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
7664 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
7665 [(set RC:$dst, (VT (X86VBroadcast (ld_frag addr:$src))))]>,
7666 Sched<[Sched]>, VEX;
7668 // AVX2 adds register forms
7669 class avx2_broadcast_rr<bits<8> opc, string OpcodeStr, RegisterClass RC,
7670 ValueType ResVT, ValueType OpVT, SchedWrite Sched> :
7671 AVX28I<opc, MRMSrcReg, (outs RC:$dst), (ins VR128:$src),
7672 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
7673 [(set RC:$dst, (ResVT (X86VBroadcast (OpVT VR128:$src))))]>,
7674 Sched<[Sched]>, VEX;
7676 let ExeDomain = SSEPackedSingle, Predicates = [HasAVX, NoVLX] in {
7677 def VBROADCASTSSrm : avx_broadcast_rm<0x18, "vbroadcastss", VR128,
7678 f32mem, v4f32, loadf32, WriteLoad>;
7679 def VBROADCASTSSYrm : avx_broadcast_rm<0x18, "vbroadcastss", VR256,
7680 f32mem, v8f32, loadf32,
7681 WriteFShuffleLd>, VEX_L;
7683 let ExeDomain = SSEPackedDouble, Predicates = [HasAVX, NoVLX] in
7684 def VBROADCASTSDYrm : avx_broadcast_rm<0x19, "vbroadcastsd", VR256, f64mem,
7685 v4f64, loadf64, WriteFShuffleLd>, VEX_L;
7687 let ExeDomain = SSEPackedSingle, Predicates = [HasAVX2, NoVLX] in {
7688 def VBROADCASTSSrr : avx2_broadcast_rr<0x18, "vbroadcastss", VR128,
7689 v4f32, v4f32, WriteFShuffle>;
7690 def VBROADCASTSSYrr : avx2_broadcast_rr<0x18, "vbroadcastss", VR256,
7691 v8f32, v4f32, WriteFShuffle256>, VEX_L;
7693 let ExeDomain = SSEPackedDouble, Predicates = [HasAVX2, NoVLX] in
7694 def VBROADCASTSDYrr : avx2_broadcast_rr<0x19, "vbroadcastsd", VR256,
7695 v4f64, v2f64, WriteFShuffle256>, VEX_L;
7697 //===----------------------------------------------------------------------===//
7698 // VBROADCAST*128 - Load from memory and broadcast 128-bit vector to both
7699 // halves of a 256-bit vector.
7701 let mayLoad = 1, hasSideEffects = 0, Predicates = [HasAVX2] in
7702 def VBROADCASTI128 : AVX8I<0x5A, MRMSrcMem, (outs VR256:$dst),
7704 "vbroadcasti128\t{$src, $dst|$dst, $src}", []>,
7705 Sched<[WriteLoad]>, VEX, VEX_L;
7707 let mayLoad = 1, hasSideEffects = 0, Predicates = [HasAVX] in
7708 def VBROADCASTF128 : AVX8I<0x1A, MRMSrcMem, (outs VR256:$dst),
7710 "vbroadcastf128\t{$src, $dst|$dst, $src}", []>,
7711 Sched<[WriteFShuffleLd]>, VEX, VEX_L;
7713 let Predicates = [HasAVX2, NoVLX] in {
7714 def : Pat<(v4i64 (X86SubVBroadcast (loadv2i64 addr:$src))),
7715 (VBROADCASTI128 addr:$src)>;
7716 def : Pat<(v8i32 (X86SubVBroadcast (bc_v4i32 (loadv2i64 addr:$src)))),
7717 (VBROADCASTI128 addr:$src)>;
7718 def : Pat<(v16i16 (X86SubVBroadcast (bc_v8i16 (loadv2i64 addr:$src)))),
7719 (VBROADCASTI128 addr:$src)>;
7720 def : Pat<(v32i8 (X86SubVBroadcast (bc_v16i8 (loadv2i64 addr:$src)))),
7721 (VBROADCASTI128 addr:$src)>;
7724 let Predicates = [HasAVX, NoVLX] in {
7725 def : Pat<(v4f64 (X86SubVBroadcast (loadv2f64 addr:$src))),
7726 (VBROADCASTF128 addr:$src)>;
7727 def : Pat<(v8f32 (X86SubVBroadcast (loadv4f32 addr:$src))),
7728 (VBROADCASTF128 addr:$src)>;
7731 let Predicates = [HasAVX1Only] in {
7732 def : Pat<(v4i64 (X86SubVBroadcast (loadv2i64 addr:$src))),
7733 (VBROADCASTF128 addr:$src)>;
7734 def : Pat<(v8i32 (X86SubVBroadcast (bc_v4i32 (loadv2i64 addr:$src)))),
7735 (VBROADCASTF128 addr:$src)>;
7736 def : Pat<(v16i16 (X86SubVBroadcast (bc_v8i16 (loadv2i64 addr:$src)))),
7737 (VBROADCASTF128 addr:$src)>;
7738 def : Pat<(v32i8 (X86SubVBroadcast (bc_v16i8 (loadv2i64 addr:$src)))),
7739 (VBROADCASTF128 addr:$src)>;
7742 //===----------------------------------------------------------------------===//
7743 // VINSERTF128 - Insert packed floating-point values
7745 let hasSideEffects = 0, ExeDomain = SSEPackedSingle in {
7746 def VINSERTF128rr : AVXAIi8<0x18, MRMSrcReg, (outs VR256:$dst),
7747 (ins VR256:$src1, VR128:$src2, u8imm:$src3),
7748 "vinsertf128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7749 []>, Sched<[WriteFShuffle]>, VEX_4V, VEX_L;
7751 def VINSERTF128rm : AVXAIi8<0x18, MRMSrcMem, (outs VR256:$dst),
7752 (ins VR256:$src1, f128mem:$src2, u8imm:$src3),
7753 "vinsertf128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7754 []>, Sched<[WriteFShuffleLd, ReadAfterLd]>, VEX_4V, VEX_L;
7758 // Without AVX2 we need to concat two v4i32 V_SETALLONES to create a 256-bit
7760 let Predicates = [HasAVX1Only] in
7761 def : Pat<(v8i32 immAllOnesV),
7763 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), (V_SETALLONES), sub_xmm),
7764 (V_SETALLONES), 1)>;
7766 multiclass vinsert_lowering<string InstrStr, ValueType From, ValueType To,
7767 PatFrag memop_frag> {
7768 def : Pat<(vinsert128_insert:$ins (To VR256:$src1), (From VR128:$src2),
7770 (!cast<Instruction>(InstrStr#rr) VR256:$src1, VR128:$src2,
7771 (INSERT_get_vinsert128_imm VR256:$ins))>;
7772 def : Pat<(vinsert128_insert:$ins (To VR256:$src1),
7773 (From (bitconvert (memop_frag addr:$src2))),
7775 (!cast<Instruction>(InstrStr#rm) VR256:$src1, addr:$src2,
7776 (INSERT_get_vinsert128_imm VR256:$ins))>;
7779 let Predicates = [HasAVX, NoVLX] in {
7780 defm : vinsert_lowering<"VINSERTF128", v4f32, v8f32, loadv4f32>;
7781 defm : vinsert_lowering<"VINSERTF128", v2f64, v4f64, loadv2f64>;
7784 let Predicates = [HasAVX1Only] in {
7785 defm : vinsert_lowering<"VINSERTF128", v2i64, v4i64, loadv2i64>;
7786 defm : vinsert_lowering<"VINSERTF128", v4i32, v8i32, loadv2i64>;
7787 defm : vinsert_lowering<"VINSERTF128", v8i16, v16i16, loadv2i64>;
7788 defm : vinsert_lowering<"VINSERTF128", v16i8, v32i8, loadv2i64>;
7791 //===----------------------------------------------------------------------===//
7792 // VEXTRACTF128 - Extract packed floating-point values
7794 let hasSideEffects = 0, ExeDomain = SSEPackedSingle in {
7795 def VEXTRACTF128rr : AVXAIi8<0x19, MRMDestReg, (outs VR128:$dst),
7796 (ins VR256:$src1, u8imm:$src2),
7797 "vextractf128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7798 []>, Sched<[WriteFShuffle]>, VEX, VEX_L;
7800 def VEXTRACTF128mr : AVXAIi8<0x19, MRMDestMem, (outs),
7801 (ins f128mem:$dst, VR256:$src1, u8imm:$src2),
7802 "vextractf128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7803 []>, Sched<[WriteStore]>, VEX, VEX_L;
7806 multiclass vextract_lowering<string InstrStr, ValueType From, ValueType To> {
7807 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
7808 (To (!cast<Instruction>(InstrStr#rr)
7810 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
7811 def : Pat<(store (To (vextract128_extract:$ext (From VR256:$src1),
7812 (iPTR imm))), addr:$dst),
7813 (!cast<Instruction>(InstrStr#mr) addr:$dst, VR256:$src1,
7814 (EXTRACT_get_vextract128_imm VR128:$ext))>;
7818 let Predicates = [HasAVX, NoVLX] in {
7819 defm : vextract_lowering<"VEXTRACTF128", v8f32, v4f32>;
7820 defm : vextract_lowering<"VEXTRACTF128", v4f64, v2f64>;
7823 let Predicates = [HasAVX1Only] in {
7824 defm : vextract_lowering<"VEXTRACTF128", v4i64, v2i64>;
7825 defm : vextract_lowering<"VEXTRACTF128", v8i32, v4i32>;
7826 defm : vextract_lowering<"VEXTRACTF128", v16i16, v8i16>;
7827 defm : vextract_lowering<"VEXTRACTF128", v32i8, v16i8>;
7830 //===----------------------------------------------------------------------===//
7831 // VMASKMOV - Conditional SIMD Packed Loads and Stores
7833 multiclass avx_movmask_rm<bits<8> opc_rm, bits<8> opc_mr, string OpcodeStr,
7834 Intrinsic IntLd, Intrinsic IntLd256,
7835 Intrinsic IntSt, Intrinsic IntSt256> {
7836 def rm : AVX8I<opc_rm, MRMSrcMem, (outs VR128:$dst),
7837 (ins VR128:$src1, f128mem:$src2),
7838 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7839 [(set VR128:$dst, (IntLd addr:$src2, VR128:$src1))]>,
7841 def Yrm : AVX8I<opc_rm, MRMSrcMem, (outs VR256:$dst),
7842 (ins VR256:$src1, f256mem:$src2),
7843 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7844 [(set VR256:$dst, (IntLd256 addr:$src2, VR256:$src1))]>,
7846 def mr : AVX8I<opc_mr, MRMDestMem, (outs),
7847 (ins f128mem:$dst, VR128:$src1, VR128:$src2),
7848 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7849 [(IntSt addr:$dst, VR128:$src1, VR128:$src2)]>, VEX_4V;
7850 def Ymr : AVX8I<opc_mr, MRMDestMem, (outs),
7851 (ins f256mem:$dst, VR256:$src1, VR256:$src2),
7852 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7853 [(IntSt256 addr:$dst, VR256:$src1, VR256:$src2)]>, VEX_4V, VEX_L;
7856 let ExeDomain = SSEPackedSingle in
7857 defm VMASKMOVPS : avx_movmask_rm<0x2C, 0x2E, "vmaskmovps",
7858 int_x86_avx_maskload_ps,
7859 int_x86_avx_maskload_ps_256,
7860 int_x86_avx_maskstore_ps,
7861 int_x86_avx_maskstore_ps_256>;
7862 let ExeDomain = SSEPackedDouble in
7863 defm VMASKMOVPD : avx_movmask_rm<0x2D, 0x2F, "vmaskmovpd",
7864 int_x86_avx_maskload_pd,
7865 int_x86_avx_maskload_pd_256,
7866 int_x86_avx_maskstore_pd,
7867 int_x86_avx_maskstore_pd_256>;
7869 //===----------------------------------------------------------------------===//
7870 // VPERMIL - Permute Single and Double Floating-Point Values
7872 multiclass avx_permil<bits<8> opc_rm, bits<8> opc_rmi, string OpcodeStr,
7873 RegisterClass RC, X86MemOperand x86memop_f,
7874 X86MemOperand x86memop_i, PatFrag i_frag,
7875 ValueType f_vt, ValueType i_vt> {
7876 let Predicates = [HasAVX, NoVLX] in {
7877 def rr : AVX8I<opc_rm, MRMSrcReg, (outs RC:$dst),
7878 (ins RC:$src1, RC:$src2),
7879 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7880 [(set RC:$dst, (f_vt (X86VPermilpv RC:$src1, (i_vt RC:$src2))))]>, VEX_4V,
7881 Sched<[WriteFShuffle]>;
7882 def rm : AVX8I<opc_rm, MRMSrcMem, (outs RC:$dst),
7883 (ins RC:$src1, x86memop_i:$src2),
7884 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7885 [(set RC:$dst, (f_vt (X86VPermilpv RC:$src1,
7886 (i_vt (bitconvert (i_frag addr:$src2))))))]>, VEX_4V,
7887 Sched<[WriteFShuffleLd, ReadAfterLd]>;
7889 def ri : AVXAIi8<opc_rmi, MRMSrcReg, (outs RC:$dst),
7890 (ins RC:$src1, u8imm:$src2),
7891 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7892 [(set RC:$dst, (f_vt (X86VPermilpi RC:$src1, (i8 imm:$src2))))]>, VEX,
7893 Sched<[WriteFShuffle]>;
7894 def mi : AVXAIi8<opc_rmi, MRMSrcMem, (outs RC:$dst),
7895 (ins x86memop_f:$src1, u8imm:$src2),
7896 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7898 (f_vt (X86VPermilpi (load addr:$src1), (i8 imm:$src2))))]>, VEX,
7899 Sched<[WriteFShuffleLd]>;
7900 }// Predicates = [HasAVX, NoVLX]
7903 let ExeDomain = SSEPackedSingle in {
7904 defm VPERMILPS : avx_permil<0x0C, 0x04, "vpermilps", VR128, f128mem, i128mem,
7905 loadv2i64, v4f32, v4i32>;
7906 defm VPERMILPSY : avx_permil<0x0C, 0x04, "vpermilps", VR256, f256mem, i256mem,
7907 loadv4i64, v8f32, v8i32>, VEX_L;
7909 let ExeDomain = SSEPackedDouble in {
7910 defm VPERMILPD : avx_permil<0x0D, 0x05, "vpermilpd", VR128, f128mem, i128mem,
7911 loadv2i64, v2f64, v2i64>;
7912 defm VPERMILPDY : avx_permil<0x0D, 0x05, "vpermilpd", VR256, f256mem, i256mem,
7913 loadv4i64, v4f64, v4i64>, VEX_L;
7916 //===----------------------------------------------------------------------===//
7917 // VPERM2F128 - Permute Floating-Point Values in 128-bit chunks
7919 let ExeDomain = SSEPackedSingle in {
7920 let isCommutable = 1 in
7921 def VPERM2F128rr : AVXAIi8<0x06, MRMSrcReg, (outs VR256:$dst),
7922 (ins VR256:$src1, VR256:$src2, u8imm:$src3),
7923 "vperm2f128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7924 [(set VR256:$dst, (v8f32 (X86VPerm2x128 VR256:$src1, VR256:$src2,
7925 (i8 imm:$src3))))]>, VEX_4V, VEX_L,
7926 Sched<[WriteFShuffle]>;
7927 def VPERM2F128rm : AVXAIi8<0x06, MRMSrcMem, (outs VR256:$dst),
7928 (ins VR256:$src1, f256mem:$src2, u8imm:$src3),
7929 "vperm2f128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7930 [(set VR256:$dst, (X86VPerm2x128 VR256:$src1, (loadv8f32 addr:$src2),
7931 (i8 imm:$src3)))]>, VEX_4V, VEX_L,
7932 Sched<[WriteFShuffleLd, ReadAfterLd]>;
7935 let Predicates = [HasAVX] in {
7936 def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
7937 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
7938 def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1,
7939 (loadv4f64 addr:$src2), (i8 imm:$imm))),
7940 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
7943 let Predicates = [HasAVX1Only] in {
7944 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
7945 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
7946 def : Pat<(v4i64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
7947 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
7948 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
7949 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
7950 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
7951 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
7953 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1,
7954 (bc_v8i32 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
7955 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
7956 def : Pat<(v4i64 (X86VPerm2x128 VR256:$src1,
7957 (loadv4i64 addr:$src2), (i8 imm:$imm))),
7958 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
7959 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1,
7960 (bc_v32i8 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
7961 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
7962 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1,
7963 (bc_v16i16 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
7964 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
7967 //===----------------------------------------------------------------------===//
7968 // VZERO - Zero YMM registers
7970 // Note, these instruction do not affect the YMM16-YMM31.
7971 let Defs = [YMM0, YMM1, YMM2, YMM3, YMM4, YMM5, YMM6, YMM7,
7972 YMM8, YMM9, YMM10, YMM11, YMM12, YMM13, YMM14, YMM15] in {
7973 // Zero All YMM registers
7974 def VZEROALL : I<0x77, RawFrm, (outs), (ins), "vzeroall",
7975 [(int_x86_avx_vzeroall)]>, PS, VEX, VEX_L, Requires<[HasAVX]>, VEX_WIG;
7977 // Zero Upper bits of YMM registers
7978 def VZEROUPPER : I<0x77, RawFrm, (outs), (ins), "vzeroupper",
7979 [(int_x86_avx_vzeroupper)]>, PS, VEX, Requires<[HasAVX]>, VEX_WIG;
7982 //===----------------------------------------------------------------------===//
7983 // Half precision conversion instructions
7984 //===----------------------------------------------------------------------===//
7985 multiclass f16c_ph2ps<RegisterClass RC, X86MemOperand x86memop, Intrinsic Int> {
7986 def rr : I<0x13, MRMSrcReg, (outs RC:$dst), (ins VR128:$src),
7987 "vcvtph2ps\t{$src, $dst|$dst, $src}",
7988 [(set RC:$dst, (Int VR128:$src))]>,
7989 T8PD, VEX, Sched<[WriteCvtF2F]>;
7990 let hasSideEffects = 0, mayLoad = 1 in
7991 def rm : I<0x13, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
7992 "vcvtph2ps\t{$src, $dst|$dst, $src}", []>, T8PD, VEX,
7993 Sched<[WriteCvtF2FLd]>;
7996 multiclass f16c_ps2ph<RegisterClass RC, X86MemOperand x86memop, Intrinsic Int> {
7997 def rr : Ii8<0x1D, MRMDestReg, (outs VR128:$dst),
7998 (ins RC:$src1, i32u8imm:$src2),
7999 "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8000 [(set VR128:$dst, (Int RC:$src1, imm:$src2))]>,
8001 TAPD, VEX, Sched<[WriteCvtF2F]>;
8002 let hasSideEffects = 0, mayStore = 1,
8003 SchedRW = [WriteCvtF2FLd, WriteRMW] in
8004 def mr : Ii8<0x1D, MRMDestMem, (outs),
8005 (ins x86memop:$dst, RC:$src1, i32u8imm:$src2),
8006 "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8010 let Predicates = [HasF16C] in {
8011 defm VCVTPH2PS : f16c_ph2ps<VR128, f64mem, int_x86_vcvtph2ps_128>;
8012 defm VCVTPH2PSY : f16c_ph2ps<VR256, f128mem, int_x86_vcvtph2ps_256>, VEX_L;
8013 defm VCVTPS2PH : f16c_ps2ph<VR128, f64mem, int_x86_vcvtps2ph_128>;
8014 defm VCVTPS2PHY : f16c_ps2ph<VR256, f128mem, int_x86_vcvtps2ph_256>, VEX_L;
8016 // Pattern match vcvtph2ps of a scalar i64 load.
8017 def : Pat<(int_x86_vcvtph2ps_128 (vzmovl_v2i64 addr:$src)),
8018 (VCVTPH2PSrm addr:$src)>;
8019 def : Pat<(int_x86_vcvtph2ps_128 (vzload_v2i64 addr:$src)),
8020 (VCVTPH2PSrm addr:$src)>;
8021 def : Pat<(int_x86_vcvtph2ps_128 (bitconvert
8022 (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
8023 (VCVTPH2PSrm addr:$src)>;
8025 def : Pat<(store (f64 (extractelt (bc_v2f64 (v8i16
8026 (int_x86_vcvtps2ph_128 VR128:$src1, i32:$src2))), (iPTR 0))),
8028 (VCVTPS2PHmr addr:$dst, VR128:$src1, imm:$src2)>;
8029 def : Pat<(store (i64 (extractelt (bc_v2i64 (v8i16
8030 (int_x86_vcvtps2ph_128 VR128:$src1, i32:$src2))), (iPTR 0))),
8032 (VCVTPS2PHmr addr:$dst, VR128:$src1, imm:$src2)>;
8033 def : Pat<(store (v8i16 (int_x86_vcvtps2ph_256 VR256:$src1, i32:$src2)),
8035 (VCVTPS2PHYmr addr:$dst, VR256:$src1, imm:$src2)>;
8038 // Patterns for matching conversions from float to half-float and vice versa.
8039 let Predicates = [HasF16C, NoVLX] in {
8040 // Use MXCSR.RC for rounding instead of explicitly specifying the default
8041 // rounding mode (Nearest-Even, encoded as 0). Both are equivalent in the
8042 // configurations we support (the default). However, falling back to MXCSR is
8043 // more consistent with other instructions, which are always controlled by it.
8044 // It's encoded as 0b100.
8045 def : Pat<(fp_to_f16 FR32:$src),
8046 (i16 (EXTRACT_SUBREG (VMOVPDI2DIrr (VCVTPS2PHrr
8047 (COPY_TO_REGCLASS FR32:$src, VR128), 4)), sub_16bit))>;
8049 def : Pat<(f16_to_fp GR16:$src),
8050 (f32 (COPY_TO_REGCLASS (VCVTPH2PSrr
8051 (COPY_TO_REGCLASS (MOVSX32rr16 GR16:$src), VR128)), FR32)) >;
8053 def : Pat<(f16_to_fp (i16 (fp_to_f16 FR32:$src))),
8054 (f32 (COPY_TO_REGCLASS (VCVTPH2PSrr
8055 (VCVTPS2PHrr (COPY_TO_REGCLASS FR32:$src, VR128), 4)), FR32)) >;
8058 //===----------------------------------------------------------------------===//
8059 // AVX2 Instructions
8060 //===----------------------------------------------------------------------===//
8062 /// AVX2_binop_rmi - AVX2 binary operator with 8-bit immediate
8063 multiclass AVX2_binop_rmi<bits<8> opc, string OpcodeStr, SDNode OpNode,
8064 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
8065 X86MemOperand x86memop> {
8066 let isCommutable = 1 in
8067 def rri : AVX2AIi8<opc, MRMSrcReg, (outs RC:$dst),
8068 (ins RC:$src1, RC:$src2, u8imm:$src3),
8069 !strconcat(OpcodeStr,
8070 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
8071 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2, imm:$src3)))]>,
8072 Sched<[WriteBlend]>, VEX_4V;
8073 def rmi : AVX2AIi8<opc, MRMSrcMem, (outs RC:$dst),
8074 (ins RC:$src1, x86memop:$src2, u8imm:$src3),
8075 !strconcat(OpcodeStr,
8076 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
8078 (OpVT (OpNode RC:$src1,
8079 (bitconvert (memop_frag addr:$src2)), imm:$src3)))]>,
8080 Sched<[WriteBlendLd, ReadAfterLd]>, VEX_4V;
8083 defm VPBLENDD : AVX2_binop_rmi<0x02, "vpblendd", X86Blendi, v4i32,
8084 VR128, loadv2i64, i128mem>;
8085 defm VPBLENDDY : AVX2_binop_rmi<0x02, "vpblendd", X86Blendi, v8i32,
8086 VR256, loadv4i64, i256mem>, VEX_L;
8088 // For insertion into the zero index (low half) of a 256-bit vector, it is
8089 // more efficient to generate a blend with immediate instead of an insert*128.
8090 let Predicates = [HasAVX2] in {
8091 def : Pat<(insert_subvector (v8i32 VR256:$src1), (v4i32 VR128:$src2), (iPTR 0)),
8092 (VPBLENDDYrri VR256:$src1,
8093 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8094 VR128:$src2, sub_xmm), 0xf)>;
8095 def : Pat<(insert_subvector (v4i64 VR256:$src1), (v2i64 VR128:$src2), (iPTR 0)),
8096 (VPBLENDDYrri VR256:$src1,
8097 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8098 VR128:$src2, sub_xmm), 0xf)>;
8099 def : Pat<(insert_subvector (v16i16 VR256:$src1), (v8i16 VR128:$src2), (iPTR 0)),
8100 (VPBLENDDYrri VR256:$src1,
8101 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8102 VR128:$src2, sub_xmm), 0xf)>;
8103 def : Pat<(insert_subvector (v32i8 VR256:$src1), (v16i8 VR128:$src2), (iPTR 0)),
8104 (VPBLENDDYrri VR256:$src1,
8105 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8106 VR128:$src2, sub_xmm), 0xf)>;
8109 let Predicates = [HasAVX1Only] in {
8110 def : Pat<(insert_subvector (v8i32 VR256:$src1), (v4i32 VR128:$src2), (iPTR 0)),
8111 (VBLENDPSYrri VR256:$src1,
8112 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8113 VR128:$src2, sub_xmm), 0xf)>;
8114 def : Pat<(insert_subvector (v4i64 VR256:$src1), (v2i64 VR128:$src2), (iPTR 0)),
8115 (VBLENDPSYrri VR256:$src1,
8116 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8117 VR128:$src2, sub_xmm), 0xf)>;
8118 def : Pat<(insert_subvector (v16i16 VR256:$src1), (v8i16 VR128:$src2), (iPTR 0)),
8119 (VBLENDPSYrri VR256:$src1,
8120 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8121 VR128:$src2, sub_xmm), 0xf)>;
8122 def : Pat<(insert_subvector (v32i8 VR256:$src1), (v16i8 VR128:$src2), (iPTR 0)),
8123 (VBLENDPSYrri VR256:$src1,
8124 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8125 VR128:$src2, sub_xmm), 0xf)>;
8128 //===----------------------------------------------------------------------===//
8129 // VPBROADCAST - Load from memory and broadcast to all elements of the
8130 // destination operand
8132 multiclass avx2_broadcast<bits<8> opc, string OpcodeStr,
8133 X86MemOperand x86memop, PatFrag ld_frag,
8134 ValueType OpVT128, ValueType OpVT256, Predicate prd> {
8135 let Predicates = [HasAVX2, prd] in {
8136 def rr : AVX28I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
8137 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8139 (OpVT128 (X86VBroadcast (OpVT128 VR128:$src))))]>,
8140 Sched<[WriteShuffle]>, VEX;
8141 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst), (ins x86memop:$src),
8142 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8144 (OpVT128 (X86VBroadcast (ld_frag addr:$src))))]>,
8145 Sched<[WriteLoad]>, VEX;
8146 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
8147 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8149 (OpVT256 (X86VBroadcast (OpVT128 VR128:$src))))]>,
8150 Sched<[WriteShuffle256]>, VEX, VEX_L;
8151 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst), (ins x86memop:$src),
8152 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8154 (OpVT256 (X86VBroadcast (ld_frag addr:$src))))]>,
8155 Sched<[WriteLoad]>, VEX, VEX_L;
8157 // Provide aliases for broadcast from the same register class that
8158 // automatically does the extract.
8159 def : Pat<(OpVT256 (X86VBroadcast (OpVT256 VR256:$src))),
8160 (!cast<Instruction>(NAME#"Yrr")
8161 (OpVT128 (EXTRACT_SUBREG (OpVT256 VR256:$src),sub_xmm)))>;
8165 defm VPBROADCASTB : avx2_broadcast<0x78, "vpbroadcastb", i8mem, loadi8,
8166 v16i8, v32i8, NoVLX_Or_NoBWI>;
8167 defm VPBROADCASTW : avx2_broadcast<0x79, "vpbroadcastw", i16mem, loadi16,
8168 v8i16, v16i16, NoVLX_Or_NoBWI>;
8169 defm VPBROADCASTD : avx2_broadcast<0x58, "vpbroadcastd", i32mem, loadi32,
8170 v4i32, v8i32, NoVLX>;
8171 defm VPBROADCASTQ : avx2_broadcast<0x59, "vpbroadcastq", i64mem, loadi64,
8172 v2i64, v4i64, NoVLX>;
8174 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
8175 // 32-bit targets will fail to load a i64 directly but can use ZEXT_LOAD.
8176 def : Pat<(v2i64 (X86VBroadcast (v2i64 (X86vzload addr:$src)))),
8177 (VPBROADCASTQrm addr:$src)>;
8178 def : Pat<(v4i64 (X86VBroadcast (v4i64 (X86vzload addr:$src)))),
8179 (VPBROADCASTQYrm addr:$src)>;
8180 // loadi16 is tricky to fold, because !isTypeDesirableForOp, justifiably.
8181 // This means we'll encounter truncated i32 loads; match that here.
8182 def : Pat<(v8i16 (X86VBroadcast (i16 (trunc (i32 (load addr:$src)))))),
8183 (VPBROADCASTWrm addr:$src)>;
8184 def : Pat<(v16i16 (X86VBroadcast (i16 (trunc (i32 (load addr:$src)))))),
8185 (VPBROADCASTWYrm addr:$src)>;
8186 def : Pat<(v8i16 (X86VBroadcast
8187 (i16 (trunc (i32 (zextloadi16 addr:$src)))))),
8188 (VPBROADCASTWrm addr:$src)>;
8189 def : Pat<(v16i16 (X86VBroadcast
8190 (i16 (trunc (i32 (zextloadi16 addr:$src)))))),
8191 (VPBROADCASTWYrm addr:$src)>;
8194 let Predicates = [HasAVX2, NoVLX] in {
8195 // Provide aliases for broadcast from the same register class that
8196 // automatically does the extract.
8197 def : Pat<(v8f32 (X86VBroadcast (v8f32 VR256:$src))),
8198 (VBROADCASTSSYrr (v4f32 (EXTRACT_SUBREG (v8f32 VR256:$src),
8200 def : Pat<(v4f64 (X86VBroadcast (v4f64 VR256:$src))),
8201 (VBROADCASTSDYrr (v2f64 (EXTRACT_SUBREG (v4f64 VR256:$src),
8205 let Predicates = [HasAVX2, NoVLX] in {
8206 // Provide fallback in case the load node that is used in the patterns above
8207 // is used by additional users, which prevents the pattern selection.
8208 def : Pat<(v4f32 (X86VBroadcast FR32:$src)),
8209 (VBROADCASTSSrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
8210 def : Pat<(v8f32 (X86VBroadcast FR32:$src)),
8211 (VBROADCASTSSYrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
8212 def : Pat<(v4f64 (X86VBroadcast FR64:$src)),
8213 (VBROADCASTSDYrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
8216 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
8217 def : Pat<(v16i8 (X86VBroadcast GR8:$src)),
8218 (VPBROADCASTBrr (COPY_TO_REGCLASS
8219 (i32 (INSERT_SUBREG (i32 (IMPLICIT_DEF)),
8220 GR8:$src, sub_8bit)),
8222 def : Pat<(v32i8 (X86VBroadcast GR8:$src)),
8223 (VPBROADCASTBYrr (COPY_TO_REGCLASS
8224 (i32 (INSERT_SUBREG (i32 (IMPLICIT_DEF)),
8225 GR8:$src, sub_8bit)),
8228 def : Pat<(v8i16 (X86VBroadcast GR16:$src)),
8229 (VPBROADCASTWrr (COPY_TO_REGCLASS
8230 (i32 (INSERT_SUBREG (i32 (IMPLICIT_DEF)),
8231 GR16:$src, sub_16bit)),
8233 def : Pat<(v16i16 (X86VBroadcast GR16:$src)),
8234 (VPBROADCASTWYrr (COPY_TO_REGCLASS
8235 (i32 (INSERT_SUBREG (i32 (IMPLICIT_DEF)),
8236 GR16:$src, sub_16bit)),
8239 let Predicates = [HasAVX2, NoVLX] in {
8240 def : Pat<(v4i32 (X86VBroadcast GR32:$src)),
8241 (VPBROADCASTDrr (COPY_TO_REGCLASS GR32:$src, VR128))>;
8242 def : Pat<(v8i32 (X86VBroadcast GR32:$src)),
8243 (VPBROADCASTDYrr (COPY_TO_REGCLASS GR32:$src, VR128))>;
8244 def : Pat<(v2i64 (X86VBroadcast GR64:$src)),
8245 (VPBROADCASTQrr (COPY_TO_REGCLASS GR64:$src, VR128))>;
8246 def : Pat<(v4i64 (X86VBroadcast GR64:$src)),
8247 (VPBROADCASTQYrr (COPY_TO_REGCLASS GR64:$src, VR128))>;
8250 // AVX1 broadcast patterns
8251 let Predicates = [HasAVX1Only] in {
8252 def : Pat<(v8i32 (X86VBroadcast (loadi32 addr:$src))),
8253 (VBROADCASTSSYrm addr:$src)>;
8254 def : Pat<(v4i64 (X86VBroadcast (loadi64 addr:$src))),
8255 (VBROADCASTSDYrm addr:$src)>;
8256 def : Pat<(v4i32 (X86VBroadcast (loadi32 addr:$src))),
8257 (VBROADCASTSSrm addr:$src)>;
8260 // Provide fallback in case the load node that is used in the patterns above
8261 // is used by additional users, which prevents the pattern selection.
8262 let Predicates = [HasAVX, NoVLX] in {
8263 // 128bit broadcasts:
8264 def : Pat<(v2f64 (X86VBroadcast f64:$src)),
8265 (VMOVDDUPrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
8268 let Predicates = [HasAVX1Only] in {
8269 def : Pat<(v4f32 (X86VBroadcast FR32:$src)),
8270 (VPERMILPSri (COPY_TO_REGCLASS FR32:$src, VR128), 0)>;
8271 def : Pat<(v8f32 (X86VBroadcast FR32:$src)),
8272 (VINSERTF128rr (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)),
8273 (VPERMILPSri (COPY_TO_REGCLASS FR32:$src, VR128), 0), sub_xmm),
8274 (VPERMILPSri (COPY_TO_REGCLASS FR32:$src, VR128), 0), 1)>;
8275 def : Pat<(v4f64 (X86VBroadcast FR64:$src)),
8276 (VINSERTF128rr (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)),
8277 (VMOVDDUPrr (COPY_TO_REGCLASS FR64:$src, VR128)), sub_xmm),
8278 (VMOVDDUPrr (COPY_TO_REGCLASS FR64:$src, VR128)), 1)>;
8280 def : Pat<(v4i32 (X86VBroadcast GR32:$src)),
8281 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0)>;
8282 def : Pat<(v8i32 (X86VBroadcast GR32:$src)),
8283 (VINSERTF128rr (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8284 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0), sub_xmm),
8285 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0), 1)>;
8286 def : Pat<(v4i64 (X86VBroadcast GR64:$src)),
8287 (VINSERTF128rr (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)),
8288 (VPSHUFDri (COPY_TO_REGCLASS GR64:$src, VR128), 0x44), sub_xmm),
8289 (VPSHUFDri (COPY_TO_REGCLASS GR64:$src, VR128), 0x44), 1)>;
8291 def : Pat<(v2i64 (X86VBroadcast i64:$src)),
8292 (VPSHUFDri (COPY_TO_REGCLASS GR64:$src, VR128), 0x44)>;
8295 //===----------------------------------------------------------------------===//
8296 // VPERM - Permute instructions
8299 multiclass avx2_perm<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
8300 ValueType OpVT, X86FoldableSchedWrite Sched,
8301 X86MemOperand memOp> {
8302 let Predicates = [HasAVX2, NoVLX] in {
8303 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst),
8304 (ins VR256:$src1, VR256:$src2),
8305 !strconcat(OpcodeStr,
8306 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8308 (OpVT (X86VPermv VR256:$src1, VR256:$src2)))]>,
8309 Sched<[Sched]>, VEX_4V, VEX_L;
8310 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst),
8311 (ins VR256:$src1, memOp:$src2),
8312 !strconcat(OpcodeStr,
8313 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8315 (OpVT (X86VPermv VR256:$src1,
8316 (bitconvert (mem_frag addr:$src2)))))]>,
8317 Sched<[Sched.Folded, ReadAfterLd]>, VEX_4V, VEX_L;
8321 defm VPERMD : avx2_perm<0x36, "vpermd", loadv4i64, v8i32, WriteShuffle256,
8323 let ExeDomain = SSEPackedSingle in
8324 defm VPERMPS : avx2_perm<0x16, "vpermps", loadv8f32, v8f32, WriteFShuffle256,
8327 multiclass avx2_perm_imm<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
8328 ValueType OpVT, X86FoldableSchedWrite Sched,
8329 X86MemOperand memOp> {
8330 let Predicates = [HasAVX2, NoVLX] in {
8331 def Yri : AVX2AIi8<opc, MRMSrcReg, (outs VR256:$dst),
8332 (ins VR256:$src1, u8imm:$src2),
8333 !strconcat(OpcodeStr,
8334 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8336 (OpVT (X86VPermi VR256:$src1, (i8 imm:$src2))))]>,
8337 Sched<[Sched]>, VEX, VEX_L;
8338 def Ymi : AVX2AIi8<opc, MRMSrcMem, (outs VR256:$dst),
8339 (ins memOp:$src1, u8imm:$src2),
8340 !strconcat(OpcodeStr,
8341 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8343 (OpVT (X86VPermi (mem_frag addr:$src1),
8344 (i8 imm:$src2))))]>,
8345 Sched<[Sched.Folded, ReadAfterLd]>, VEX, VEX_L;
8349 defm VPERMQ : avx2_perm_imm<0x00, "vpermq", loadv4i64, v4i64,
8350 WriteShuffle256, i256mem>, VEX_W;
8351 let ExeDomain = SSEPackedDouble in
8352 defm VPERMPD : avx2_perm_imm<0x01, "vpermpd", loadv4f64, v4f64,
8353 WriteFShuffle256, f256mem>, VEX_W;
8355 //===----------------------------------------------------------------------===//
8356 // VPERM2I128 - Permute Floating-Point Values in 128-bit chunks
8358 let isCommutable = 1 in
8359 def VPERM2I128rr : AVX2AIi8<0x46, MRMSrcReg, (outs VR256:$dst),
8360 (ins VR256:$src1, VR256:$src2, u8imm:$src3),
8361 "vperm2i128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8362 [(set VR256:$dst, (v4i64 (X86VPerm2x128 VR256:$src1, VR256:$src2,
8363 (i8 imm:$src3))))]>, Sched<[WriteShuffle256]>,
8365 def VPERM2I128rm : AVX2AIi8<0x46, MRMSrcMem, (outs VR256:$dst),
8366 (ins VR256:$src1, f256mem:$src2, u8imm:$src3),
8367 "vperm2i128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8368 [(set VR256:$dst, (X86VPerm2x128 VR256:$src1, (loadv4i64 addr:$src2),
8370 Sched<[WriteShuffle256Ld, ReadAfterLd]>, VEX_4V, VEX_L;
8372 let Predicates = [HasAVX2] in {
8373 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8374 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8375 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8376 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8377 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8378 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8380 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, (bc_v32i8 (loadv4i64 addr:$src2)),
8382 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8383 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1,
8384 (bc_v16i16 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8385 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8386 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)),
8388 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8392 //===----------------------------------------------------------------------===//
8393 // VINSERTI128 - Insert packed integer values
8395 let hasSideEffects = 0 in {
8396 def VINSERTI128rr : AVX2AIi8<0x38, MRMSrcReg, (outs VR256:$dst),
8397 (ins VR256:$src1, VR128:$src2, u8imm:$src3),
8398 "vinserti128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8399 []>, Sched<[WriteShuffle256]>, VEX_4V, VEX_L;
8401 def VINSERTI128rm : AVX2AIi8<0x38, MRMSrcMem, (outs VR256:$dst),
8402 (ins VR256:$src1, i128mem:$src2, u8imm:$src3),
8403 "vinserti128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8404 []>, Sched<[WriteShuffle256Ld, ReadAfterLd]>, VEX_4V, VEX_L;
8407 let Predicates = [HasAVX2, NoVLX] in {
8408 defm : vinsert_lowering<"VINSERTI128", v2i64, v4i64, loadv2i64>;
8409 defm : vinsert_lowering<"VINSERTI128", v4i32, v8i32, loadv2i64>;
8410 defm : vinsert_lowering<"VINSERTI128", v8i16, v16i16, loadv2i64>;
8411 defm : vinsert_lowering<"VINSERTI128", v16i8, v32i8, loadv2i64>;
8414 //===----------------------------------------------------------------------===//
8415 // VEXTRACTI128 - Extract packed integer values
8417 def VEXTRACTI128rr : AVX2AIi8<0x39, MRMDestReg, (outs VR128:$dst),
8418 (ins VR256:$src1, u8imm:$src2),
8419 "vextracti128\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8420 Sched<[WriteShuffle256]>, VEX, VEX_L;
8421 let hasSideEffects = 0, mayStore = 1 in
8422 def VEXTRACTI128mr : AVX2AIi8<0x39, MRMDestMem, (outs),
8423 (ins i128mem:$dst, VR256:$src1, u8imm:$src2),
8424 "vextracti128\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8425 Sched<[WriteStore]>, VEX, VEX_L;
8427 let Predicates = [HasAVX2, NoVLX] in {
8428 defm : vextract_lowering<"VEXTRACTI128", v4i64, v2i64>;
8429 defm : vextract_lowering<"VEXTRACTI128", v8i32, v4i32>;
8430 defm : vextract_lowering<"VEXTRACTI128", v16i16, v8i16>;
8431 defm : vextract_lowering<"VEXTRACTI128", v32i8, v16i8>;
8434 //===----------------------------------------------------------------------===//
8435 // VPMASKMOV - Conditional SIMD Integer Packed Loads and Stores
8437 multiclass avx2_pmovmask<string OpcodeStr,
8438 Intrinsic IntLd128, Intrinsic IntLd256,
8439 Intrinsic IntSt128, Intrinsic IntSt256> {
8440 def rm : AVX28I<0x8c, MRMSrcMem, (outs VR128:$dst),
8441 (ins VR128:$src1, i128mem:$src2),
8442 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8443 [(set VR128:$dst, (IntLd128 addr:$src2, VR128:$src1))]>, VEX_4V;
8444 def Yrm : AVX28I<0x8c, MRMSrcMem, (outs VR256:$dst),
8445 (ins VR256:$src1, i256mem:$src2),
8446 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8447 [(set VR256:$dst, (IntLd256 addr:$src2, VR256:$src1))]>,
8449 def mr : AVX28I<0x8e, MRMDestMem, (outs),
8450 (ins i128mem:$dst, VR128:$src1, VR128:$src2),
8451 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8452 [(IntSt128 addr:$dst, VR128:$src1, VR128:$src2)]>, VEX_4V;
8453 def Ymr : AVX28I<0x8e, MRMDestMem, (outs),
8454 (ins i256mem:$dst, VR256:$src1, VR256:$src2),
8455 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8456 [(IntSt256 addr:$dst, VR256:$src1, VR256:$src2)]>, VEX_4V, VEX_L;
8459 defm VPMASKMOVD : avx2_pmovmask<"vpmaskmovd",
8460 int_x86_avx2_maskload_d,
8461 int_x86_avx2_maskload_d_256,
8462 int_x86_avx2_maskstore_d,
8463 int_x86_avx2_maskstore_d_256>;
8464 defm VPMASKMOVQ : avx2_pmovmask<"vpmaskmovq",
8465 int_x86_avx2_maskload_q,
8466 int_x86_avx2_maskload_q_256,
8467 int_x86_avx2_maskstore_q,
8468 int_x86_avx2_maskstore_q_256>, VEX_W;
8470 multiclass maskmov_lowering<string InstrStr, RegisterClass RC, ValueType VT,
8471 ValueType MaskVT, string BlendStr, ValueType ZeroVT> {
8473 def: Pat<(X86mstore addr:$ptr, (MaskVT RC:$mask), (VT RC:$src)),
8474 (!cast<Instruction>(InstrStr#"mr") addr:$ptr, RC:$mask, RC:$src)>;
8476 def: Pat<(VT (X86mload addr:$ptr, (MaskVT RC:$mask), undef)),
8477 (!cast<Instruction>(InstrStr#"rm") RC:$mask, addr:$ptr)>;
8478 def: Pat<(VT (X86mload addr:$ptr, (MaskVT RC:$mask),
8479 (VT (bitconvert (ZeroVT immAllZerosV))))),
8480 (!cast<Instruction>(InstrStr#"rm") RC:$mask, addr:$ptr)>;
8481 def: Pat<(VT (X86mload addr:$ptr, (MaskVT RC:$mask), (VT RC:$src0))),
8482 (!cast<Instruction>(BlendStr#"rr")
8484 (!cast<Instruction>(InstrStr#"rm") RC:$mask, addr:$ptr),
8487 let Predicates = [HasAVX] in {
8488 defm : maskmov_lowering<"VMASKMOVPS", VR128, v4f32, v4i32, "VBLENDVPS", v4i32>;
8489 defm : maskmov_lowering<"VMASKMOVPD", VR128, v2f64, v2i64, "VBLENDVPD", v4i32>;
8490 defm : maskmov_lowering<"VMASKMOVPSY", VR256, v8f32, v8i32, "VBLENDVPSY", v8i32>;
8491 defm : maskmov_lowering<"VMASKMOVPDY", VR256, v4f64, v4i64, "VBLENDVPDY", v8i32>;
8493 let Predicates = [HasAVX1Only] in {
8494 // load/store i32/i64 not supported use ps/pd version
8495 defm : maskmov_lowering<"VMASKMOVPSY", VR256, v8i32, v8i32, "VBLENDVPSY", v8i32>;
8496 defm : maskmov_lowering<"VMASKMOVPDY", VR256, v4i64, v4i64, "VBLENDVPDY", v8i32>;
8497 defm : maskmov_lowering<"VMASKMOVPS", VR128, v4i32, v4i32, "VBLENDVPS", v4i32>;
8498 defm : maskmov_lowering<"VMASKMOVPD", VR128, v2i64, v2i64, "VBLENDVPD", v4i32>;
8500 let Predicates = [HasAVX2] in {
8501 defm : maskmov_lowering<"VPMASKMOVDY", VR256, v8i32, v8i32, "VBLENDVPSY", v8i32>;
8502 defm : maskmov_lowering<"VPMASKMOVQY", VR256, v4i64, v4i64, "VBLENDVPDY", v8i32>;
8503 defm : maskmov_lowering<"VPMASKMOVD", VR128, v4i32, v4i32, "VBLENDVPS", v4i32>;
8504 defm : maskmov_lowering<"VPMASKMOVQ", VR128, v2i64, v2i64, "VBLENDVPD", v4i32>;
8507 //===----------------------------------------------------------------------===//
8508 // SubVector Broadcasts
8509 // Provide fallback in case the load node that is used in the patterns above
8510 // is used by additional users, which prevents the pattern selection.
8512 let Predicates = [HasAVX2, NoVLX] in {
8513 def : Pat<(v4i64 (X86SubVBroadcast (v2i64 VR128:$src))),
8514 (VINSERTI128rr (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8515 (v2i64 VR128:$src), 1)>;
8516 def : Pat<(v8i32 (X86SubVBroadcast (v4i32 VR128:$src))),
8517 (VINSERTI128rr (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8518 (v4i32 VR128:$src), 1)>;
8519 def : Pat<(v16i16 (X86SubVBroadcast (v8i16 VR128:$src))),
8520 (VINSERTI128rr (INSERT_SUBREG (v16i16 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8521 (v8i16 VR128:$src), 1)>;
8522 def : Pat<(v32i8 (X86SubVBroadcast (v16i8 VR128:$src))),
8523 (VINSERTI128rr (INSERT_SUBREG (v32i8 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8524 (v16i8 VR128:$src), 1)>;
8527 let Predicates = [HasAVX, NoVLX] in {
8528 def : Pat<(v4f64 (X86SubVBroadcast (v2f64 VR128:$src))),
8529 (VINSERTF128rr (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8530 (v2f64 VR128:$src), 1)>;
8531 def : Pat<(v8f32 (X86SubVBroadcast (v4f32 VR128:$src))),
8532 (VINSERTF128rr (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8533 (v4f32 VR128:$src), 1)>;
8536 let Predicates = [HasAVX1Only] in {
8537 def : Pat<(v4i64 (X86SubVBroadcast (v2i64 VR128:$src))),
8538 (VINSERTF128rr (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8539 (v2i64 VR128:$src), 1)>;
8540 def : Pat<(v8i32 (X86SubVBroadcast (v4i32 VR128:$src))),
8541 (VINSERTF128rr (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8542 (v4i32 VR128:$src), 1)>;
8543 def : Pat<(v16i16 (X86SubVBroadcast (v8i16 VR128:$src))),
8544 (VINSERTF128rr (INSERT_SUBREG (v16i16 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8545 (v8i16 VR128:$src), 1)>;
8546 def : Pat<(v32i8 (X86SubVBroadcast (v16i8 VR128:$src))),
8547 (VINSERTF128rr (INSERT_SUBREG (v32i8 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8548 (v16i8 VR128:$src), 1)>;
8551 //===----------------------------------------------------------------------===//
8552 // Variable Bit Shifts
8554 multiclass avx2_var_shift<bits<8> opc, string OpcodeStr, SDNode OpNode,
8555 ValueType vt128, ValueType vt256> {
8556 def rr : AVX28I<opc, MRMSrcReg, (outs VR128:$dst),
8557 (ins VR128:$src1, VR128:$src2),
8558 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8560 (vt128 (OpNode VR128:$src1, (vt128 VR128:$src2))))]>,
8561 VEX_4V, Sched<[WriteVarVecShift]>;
8562 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst),
8563 (ins VR128:$src1, i128mem:$src2),
8564 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8566 (vt128 (OpNode VR128:$src1,
8567 (vt128 (bitconvert (loadv2i64 addr:$src2))))))]>,
8568 VEX_4V, Sched<[WriteVarVecShiftLd, ReadAfterLd]>;
8569 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst),
8570 (ins VR256:$src1, VR256:$src2),
8571 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8573 (vt256 (OpNode VR256:$src1, (vt256 VR256:$src2))))]>,
8574 VEX_4V, VEX_L, Sched<[WriteVarVecShift]>;
8575 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst),
8576 (ins VR256:$src1, i256mem:$src2),
8577 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8579 (vt256 (OpNode VR256:$src1,
8580 (vt256 (bitconvert (loadv4i64 addr:$src2))))))]>,
8581 VEX_4V, VEX_L, Sched<[WriteVarVecShiftLd, ReadAfterLd]>;
8584 let Predicates = [HasAVX2, NoVLX] in {
8585 defm VPSLLVD : avx2_var_shift<0x47, "vpsllvd", shl, v4i32, v8i32>;
8586 defm VPSLLVQ : avx2_var_shift<0x47, "vpsllvq", shl, v2i64, v4i64>, VEX_W;
8587 defm VPSRLVD : avx2_var_shift<0x45, "vpsrlvd", srl, v4i32, v8i32>;
8588 defm VPSRLVQ : avx2_var_shift<0x45, "vpsrlvq", srl, v2i64, v4i64>, VEX_W;
8589 defm VPSRAVD : avx2_var_shift<0x46, "vpsravd", sra, v4i32, v8i32>;
8591 def : Pat<(v4i32 (X86vsrav VR128:$src1, VR128:$src2)),
8592 (VPSRAVDrr VR128:$src1, VR128:$src2)>;
8593 def : Pat<(v4i32 (X86vsrav VR128:$src1,
8594 (bitconvert (loadv2i64 addr:$src2)))),
8595 (VPSRAVDrm VR128:$src1, addr:$src2)>;
8596 def : Pat<(v8i32 (X86vsrav VR256:$src1, VR256:$src2)),
8597 (VPSRAVDYrr VR256:$src1, VR256:$src2)>;
8598 def : Pat<(v8i32 (X86vsrav VR256:$src1,
8599 (bitconvert (loadv4i64 addr:$src2)))),
8600 (VPSRAVDYrm VR256:$src1, addr:$src2)>;
8605 //===----------------------------------------------------------------------===//
8606 // VGATHER - GATHER Operations
8607 multiclass avx2_gather<bits<8> opc, string OpcodeStr, RegisterClass RC256,
8608 X86MemOperand memop128, X86MemOperand memop256> {
8609 def rm : AVX28I<opc, MRMSrcMem4VOp3, (outs VR128:$dst, VR128:$mask_wb),
8610 (ins VR128:$src1, memop128:$src2, VR128:$mask),
8611 !strconcat(OpcodeStr,
8612 "\t{$mask, $src2, $dst|$dst, $src2, $mask}"),
8614 def Yrm : AVX28I<opc, MRMSrcMem4VOp3, (outs RC256:$dst, RC256:$mask_wb),
8615 (ins RC256:$src1, memop256:$src2, RC256:$mask),
8616 !strconcat(OpcodeStr,
8617 "\t{$mask, $src2, $dst|$dst, $src2, $mask}"),
8621 let mayLoad = 1, hasSideEffects = 0, Constraints
8622 = "@earlyclobber $dst,@earlyclobber $mask_wb, $src1 = $dst, $mask = $mask_wb"
8624 defm VPGATHERDQ : avx2_gather<0x90, "vpgatherdq", VR256, vx128mem, vx256mem>, VEX_W;
8625 defm VPGATHERQQ : avx2_gather<0x91, "vpgatherqq", VR256, vx128mem, vy256mem>, VEX_W;
8626 defm VPGATHERDD : avx2_gather<0x90, "vpgatherdd", VR256, vx128mem, vy256mem>;
8627 defm VPGATHERQD : avx2_gather<0x91, "vpgatherqd", VR128, vx64mem, vy128mem>;
8629 let ExeDomain = SSEPackedDouble in {
8630 defm VGATHERDPD : avx2_gather<0x92, "vgatherdpd", VR256, vx128mem, vx256mem>, VEX_W;
8631 defm VGATHERQPD : avx2_gather<0x93, "vgatherqpd", VR256, vx128mem, vy256mem>, VEX_W;
8634 let ExeDomain = SSEPackedSingle in {
8635 defm VGATHERDPS : avx2_gather<0x92, "vgatherdps", VR256, vx128mem, vy256mem>;
8636 defm VGATHERQPS : avx2_gather<0x93, "vgatherqps", VR128, vx64mem, vy128mem>;
8640 //===----------------------------------------------------------------------===//
8641 // Extra selection patterns for FR128, f128, f128mem
8643 // movaps is shorter than movdqa. movaps is in SSE and movdqa is in SSE2.
8644 def : Pat<(store (f128 FR128:$src), addr:$dst),
8645 (MOVAPSmr addr:$dst, (COPY_TO_REGCLASS (f128 FR128:$src), VR128))>;
8647 def : Pat<(loadf128 addr:$src),
8648 (COPY_TO_REGCLASS (MOVAPSrm addr:$src), FR128)>;
8650 // andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
8651 def : Pat<(X86fand FR128:$src1, (loadf128 addr:$src2)),
8653 (ANDPSrm (COPY_TO_REGCLASS FR128:$src1, VR128), f128mem:$src2),
8656 def : Pat<(X86fand FR128:$src1, FR128:$src2),
8658 (ANDPSrr (COPY_TO_REGCLASS FR128:$src1, VR128),
8659 (COPY_TO_REGCLASS FR128:$src2, VR128)), FR128)>;
8661 def : Pat<(and FR128:$src1, FR128:$src2),
8663 (ANDPSrr (COPY_TO_REGCLASS FR128:$src1, VR128),
8664 (COPY_TO_REGCLASS FR128:$src2, VR128)), FR128)>;
8666 def : Pat<(X86for FR128:$src1, (loadf128 addr:$src2)),
8668 (ORPSrm (COPY_TO_REGCLASS FR128:$src1, VR128), f128mem:$src2),
8671 def : Pat<(X86for FR128:$src1, FR128:$src2),
8673 (ORPSrr (COPY_TO_REGCLASS FR128:$src1, VR128),
8674 (COPY_TO_REGCLASS FR128:$src2, VR128)), FR128)>;
8676 def : Pat<(or FR128:$src1, FR128:$src2),
8678 (ORPSrr (COPY_TO_REGCLASS FR128:$src1, VR128),
8679 (COPY_TO_REGCLASS FR128:$src2, VR128)), FR128)>;
8681 def : Pat<(X86fxor FR128:$src1, (loadf128 addr:$src2)),
8683 (XORPSrm (COPY_TO_REGCLASS FR128:$src1, VR128), f128mem:$src2),
8686 def : Pat<(X86fxor FR128:$src1, FR128:$src2),
8688 (XORPSrr (COPY_TO_REGCLASS FR128:$src1, VR128),
8689 (COPY_TO_REGCLASS FR128:$src2, VR128)), FR128)>;
8691 def : Pat<(xor FR128:$src1, FR128:$src2),
8693 (XORPSrr (COPY_TO_REGCLASS FR128:$src1, VR128),
8694 (COPY_TO_REGCLASS FR128:$src2, VR128)), FR128)>;