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:$dst, (v4i32 immAllOnesV))]>;
489 let Predicates = [HasAVX1Only, OptForMinSize] in {
490 def AVX1_SETALLONES: I<0, Pseudo, (outs VR256:$dst), (ins), "",
491 [(set VR256:$dst, (v8i32 immAllOnesV))]>;
493 let Predicates = [HasAVX2] in
494 def AVX2_SETALLONES : I<0, Pseudo, (outs VR256:$dst), (ins), "",
495 [(set VR256:$dst, (v8i32 immAllOnesV))]>;
498 //===----------------------------------------------------------------------===//
499 // SSE 1 & 2 - Move FP Scalar Instructions
501 // Move Instructions. Register-to-register movss/movsd is not used for FR32/64
502 // register copies because it's a partial register update; Register-to-register
503 // movss/movsd is not modeled as an INSERT_SUBREG because INSERT_SUBREG requires
504 // that the insert be implementable in terms of a copy, and just mentioned, we
505 // don't use movss/movsd for copies.
506 //===----------------------------------------------------------------------===//
508 multiclass sse12_move_rr<RegisterClass RC, SDNode OpNode, ValueType vt,
509 X86MemOperand x86memop, string base_opc,
510 string asm_opr, Domain d = GenericDomain,
512 let isCommutable = 1 in
513 def rr : SI<0x10, MRMSrcReg, (outs VR128:$dst),
514 (ins VR128:$src1, RC:$src2),
515 !strconcat(base_opc, asm_opr),
516 [(set VR128:$dst, (vt (OpNode VR128:$src1,
517 (scalar_to_vector RC:$src2))))],
518 IIC_SSE_MOV_S_RR, d>, Sched<[WriteFShuffle]>;
520 // For the disassembler
521 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in
522 def rr_REV : SI<0x11, MRMDestReg, (outs VR128:$dst),
523 (ins VR128:$src1, RC:$src2),
524 !strconcat(base_opc, asm_opr),
525 [], IIC_SSE_MOV_S_RR>, Sched<[WriteFShuffle]>,
526 FoldGenData<Name#rr>;
529 multiclass sse12_move<RegisterClass RC, SDNode OpNode, ValueType vt,
530 X86MemOperand x86memop, string OpcodeStr,
531 Domain d = GenericDomain, string Name> {
533 defm V#NAME : sse12_move_rr<RC, OpNode, vt, x86memop, OpcodeStr,
534 "\t{$src2, $src1, $dst|$dst, $src1, $src2}", d,
536 VEX_4V, VEX_LIG, VEX_WIG;
538 def V#NAME#mr : SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, RC:$src),
539 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
540 [(store RC:$src, addr:$dst)], IIC_SSE_MOV_S_MR, d>,
541 VEX, VEX_LIG, Sched<[WriteStore]>, VEX_WIG;
543 let Constraints = "$src1 = $dst" in {
544 defm NAME : sse12_move_rr<RC, OpNode, vt, x86memop, OpcodeStr,
545 "\t{$src2, $dst|$dst, $src2}", d, Name>;
548 def NAME#mr : SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, RC:$src),
549 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
550 [(store RC:$src, addr:$dst)], IIC_SSE_MOV_S_MR, d>,
554 // Loading from memory automatically zeroing upper bits.
555 multiclass sse12_move_rm<RegisterClass RC, X86MemOperand x86memop,
556 PatFrag mem_pat, string OpcodeStr,
557 Domain d = GenericDomain> {
558 def V#NAME#rm : SI<0x10, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
559 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
560 [(set RC:$dst, (mem_pat addr:$src))],
561 IIC_SSE_MOV_S_RM, d>, VEX, VEX_LIG, Sched<[WriteLoad]>, VEX_WIG;
562 def NAME#rm : SI<0x10, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
563 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
564 [(set RC:$dst, (mem_pat addr:$src))],
565 IIC_SSE_MOV_S_RM, d>, Sched<[WriteLoad]>;
568 defm MOVSS : sse12_move<FR32, X86Movss, v4f32, f32mem, "movss",
569 SSEPackedSingle, "MOVSS">, XS;
570 defm MOVSD : sse12_move<FR64, X86Movsd, v2f64, f64mem, "movsd",
571 SSEPackedDouble, "MOVSD">, XD;
573 let canFoldAsLoad = 1, isReMaterializable = 1 in {
574 defm MOVSS : sse12_move_rm<FR32, f32mem, loadf32, "movss",
575 SSEPackedSingle>, XS;
577 let AddedComplexity = 20 in
578 defm MOVSD : sse12_move_rm<FR64, f64mem, loadf64, "movsd",
579 SSEPackedDouble>, XD;
583 let Predicates = [UseAVX] in {
584 let AddedComplexity = 20 in {
585 // MOVSSrm zeros the high parts of the register; represent this
586 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
587 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
588 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
589 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
590 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
591 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
592 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
593 def : Pat<(v4f32 (X86vzload addr:$src)),
594 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
596 // MOVSDrm zeros the high parts of the register; represent this
597 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
598 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
599 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
600 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
601 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
602 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
603 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
604 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
605 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
606 def : Pat<(v2f64 (X86vzload addr:$src)),
607 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
609 // Represent the same patterns above but in the form they appear for
611 def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
612 (v4f32 (scalar_to_vector (loadf32 addr:$src))), (iPTR 0)))),
613 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_xmm)>;
614 def : Pat<(v8f32 (X86vzload addr:$src)),
615 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_xmm)>;
616 def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
617 (v2f64 (scalar_to_vector (loadf64 addr:$src))), (iPTR 0)))),
618 (SUBREG_TO_REG (i32 0), (VMOVSDrm addr:$src), sub_xmm)>;
619 def : Pat<(v4f64 (X86vzload addr:$src)),
620 (SUBREG_TO_REG (i32 0), (VMOVSDrm addr:$src), sub_xmm)>;
623 // Extract and store.
624 def : Pat<(store (f32 (extractelt (v4f32 VR128:$src), (iPTR 0))),
626 (VMOVSSmr addr:$dst, (COPY_TO_REGCLASS (v4f32 VR128:$src), FR32))>;
628 // Shuffle with VMOVSS
629 def : Pat<(v4i32 (X86Movss VR128:$src1, VR128:$src2)),
630 (VMOVSSrr (v4i32 VR128:$src1),
631 (COPY_TO_REGCLASS (v4i32 VR128:$src2), FR32))>;
632 def : Pat<(v4f32 (X86Movss VR128:$src1, VR128:$src2)),
633 (VMOVSSrr (v4f32 VR128:$src1),
634 (COPY_TO_REGCLASS (v4f32 VR128:$src2), FR32))>;
637 def : Pat<(v8i32 (X86Movss VR256:$src1, VR256:$src2)),
638 (SUBREG_TO_REG (i32 0),
639 (VMOVSSrr (EXTRACT_SUBREG (v8i32 VR256:$src1), sub_xmm),
640 (EXTRACT_SUBREG (v8i32 VR256:$src2), sub_xmm)),
642 def : Pat<(v8f32 (X86Movss VR256:$src1, VR256:$src2)),
643 (SUBREG_TO_REG (i32 0),
644 (VMOVSSrr (EXTRACT_SUBREG (v8f32 VR256:$src1), sub_xmm),
645 (EXTRACT_SUBREG (v8f32 VR256:$src2), sub_xmm)),
648 // Shuffle with VMOVSD
649 def : Pat<(v2i64 (X86Movsd VR128:$src1, VR128:$src2)),
650 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
651 def : Pat<(v2f64 (X86Movsd VR128:$src1, VR128:$src2)),
652 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
655 def : Pat<(v4i64 (X86Movsd VR256:$src1, VR256:$src2)),
656 (SUBREG_TO_REG (i32 0),
657 (VMOVSDrr (EXTRACT_SUBREG (v4i64 VR256:$src1), sub_xmm),
658 (EXTRACT_SUBREG (v4i64 VR256:$src2), sub_xmm)),
660 def : Pat<(v4f64 (X86Movsd VR256:$src1, VR256:$src2)),
661 (SUBREG_TO_REG (i32 0),
662 (VMOVSDrr (EXTRACT_SUBREG (v4f64 VR256:$src1), sub_xmm),
663 (EXTRACT_SUBREG (v4f64 VR256:$src2), sub_xmm)),
666 // FIXME: Instead of a X86Movlps there should be a X86Movsd here, the problem
667 // is during lowering, where it's not possible to recognize the fold cause
668 // it has two uses through a bitcast. One use disappears at isel time and the
669 // fold opportunity reappears.
670 def : Pat<(v2f64 (X86Movlpd VR128:$src1, VR128:$src2)),
671 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
672 def : Pat<(v2i64 (X86Movlpd VR128:$src1, VR128:$src2)),
673 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
674 def : Pat<(v4f32 (X86Movlps VR128:$src1, VR128:$src2)),
675 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
676 def : Pat<(v4i32 (X86Movlps VR128:$src1, VR128:$src2)),
677 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
680 let Predicates = [UseSSE1] in {
681 let Predicates = [NoSSE41], AddedComplexity = 15 in {
682 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
683 // MOVSS to the lower bits.
684 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
685 (MOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
686 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
687 (MOVSSrr (v4f32 (V_SET0)), (COPY_TO_REGCLASS VR128:$src, FR32))>;
688 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
689 (MOVSSrr (v4i32 (V_SET0)), (COPY_TO_REGCLASS VR128:$src, FR32))>;
692 let AddedComplexity = 20 in {
693 // MOVSSrm already zeros the high parts of the register.
694 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
695 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
696 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
697 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
698 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
699 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
700 def : Pat<(v4f32 (X86vzload addr:$src)),
701 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
704 // Extract and store.
705 def : Pat<(store (f32 (extractelt (v4f32 VR128:$src), (iPTR 0))),
707 (MOVSSmr addr:$dst, (COPY_TO_REGCLASS VR128:$src, FR32))>;
709 // Shuffle with MOVSS
710 def : Pat<(v4i32 (X86Movss VR128:$src1, VR128:$src2)),
711 (MOVSSrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR32))>;
712 def : Pat<(v4f32 (X86Movss VR128:$src1, VR128:$src2)),
713 (MOVSSrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR32))>;
716 let Predicates = [UseSSE2] in {
717 let Predicates = [NoSSE41], AddedComplexity = 15 in {
718 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
719 // MOVSD to the lower bits.
720 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
721 (MOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
724 let AddedComplexity = 20 in {
725 // MOVSDrm already zeros the high parts of the register.
726 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
727 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
728 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
729 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
730 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
731 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
732 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
733 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
734 def : Pat<(v2f64 (X86vzload addr:$src)),
735 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
738 // Shuffle with MOVSD
739 def : Pat<(v2i64 (X86Movsd VR128:$src1, VR128:$src2)),
740 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
741 def : Pat<(v2f64 (X86Movsd VR128:$src1, VR128:$src2)),
742 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
744 // FIXME: Instead of a X86Movlps there should be a X86Movsd here, the problem
745 // is during lowering, where it's not possible to recognize the fold because
746 // it has two uses through a bitcast. One use disappears at isel time and the
747 // fold opportunity reappears.
748 def : Pat<(v2f64 (X86Movlpd VR128:$src1, VR128:$src2)),
749 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
750 def : Pat<(v2i64 (X86Movlpd VR128:$src1, VR128:$src2)),
751 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
752 def : Pat<(v4f32 (X86Movlps VR128:$src1, VR128:$src2)),
753 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
754 def : Pat<(v4i32 (X86Movlps VR128:$src1, VR128:$src2)),
755 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
758 // Aliases to help the assembler pick two byte VEX encodings by swapping the
759 // operands relative to the normal instructions to use VEX.R instead of VEX.B.
760 def : InstAlias<"vmovss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
761 (VMOVSSrr_REV VR128L:$dst, VR128:$src1, VR128H:$src2), 0>;
762 def : InstAlias<"vmovsd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
763 (VMOVSDrr_REV VR128L:$dst, VR128:$src1, VR128H:$src2), 0>;
765 //===----------------------------------------------------------------------===//
766 // SSE 1 & 2 - Move Aligned/Unaligned FP Instructions
767 //===----------------------------------------------------------------------===//
769 multiclass sse12_mov_packed<bits<8> opc, RegisterClass RC,
770 X86MemOperand x86memop, PatFrag ld_frag,
771 string asm, Domain d,
773 let hasSideEffects = 0 in
774 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
775 !strconcat(asm, "\t{$src, $dst|$dst, $src}"), [], itins.rr, d>,
776 Sched<[WriteFShuffle]>;
777 let canFoldAsLoad = 1, isReMaterializable = 1 in
778 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
779 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
780 [(set RC:$dst, (ld_frag addr:$src))], itins.rm, d>,
784 let Predicates = [HasAVX, NoVLX] in {
785 defm VMOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
786 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
788 defm VMOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
789 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
791 defm VMOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
792 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
794 defm VMOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
795 "movupd", SSEPackedDouble, SSE_MOVU_ITINS>,
798 defm VMOVAPSY : sse12_mov_packed<0x28, VR256, f256mem, alignedloadv8f32,
799 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
800 PS, VEX, VEX_L, VEX_WIG;
801 defm VMOVAPDY : sse12_mov_packed<0x28, VR256, f256mem, alignedloadv4f64,
802 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
803 PD, VEX, VEX_L, VEX_WIG;
804 defm VMOVUPSY : sse12_mov_packed<0x10, VR256, f256mem, loadv8f32,
805 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
806 PS, VEX, VEX_L, VEX_WIG;
807 defm VMOVUPDY : sse12_mov_packed<0x10, VR256, f256mem, loadv4f64,
808 "movupd", SSEPackedDouble, SSE_MOVU_ITINS>,
809 PD, VEX, VEX_L, VEX_WIG;
812 let Predicates = [UseSSE1] in {
813 defm MOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
814 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
816 defm MOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
817 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
820 let Predicates = [UseSSE2] in {
821 defm MOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
822 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
824 defm MOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
825 "movupd", SSEPackedDouble, SSE_MOVU_ITINS>,
829 let SchedRW = [WriteStore], Predicates = [HasAVX, NoVLX] in {
830 def VMOVAPSmr : VPSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
831 "movaps\t{$src, $dst|$dst, $src}",
832 [(alignedstore (v4f32 VR128:$src), addr:$dst)],
833 IIC_SSE_MOVA_P_MR>, VEX, VEX_WIG;
834 def VMOVAPDmr : VPDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
835 "movapd\t{$src, $dst|$dst, $src}",
836 [(alignedstore (v2f64 VR128:$src), addr:$dst)],
837 IIC_SSE_MOVA_P_MR>, VEX, VEX_WIG;
838 def VMOVUPSmr : VPSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
839 "movups\t{$src, $dst|$dst, $src}",
840 [(store (v4f32 VR128:$src), addr:$dst)],
841 IIC_SSE_MOVU_P_MR>, VEX, VEX_WIG;
842 def VMOVUPDmr : VPDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
843 "movupd\t{$src, $dst|$dst, $src}",
844 [(store (v2f64 VR128:$src), addr:$dst)],
845 IIC_SSE_MOVU_P_MR>, VEX, VEX_WIG;
846 def VMOVAPSYmr : VPSI<0x29, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
847 "movaps\t{$src, $dst|$dst, $src}",
848 [(alignedstore256 (v8f32 VR256:$src), addr:$dst)],
849 IIC_SSE_MOVA_P_MR>, VEX, VEX_L, VEX_WIG;
850 def VMOVAPDYmr : VPDI<0x29, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
851 "movapd\t{$src, $dst|$dst, $src}",
852 [(alignedstore256 (v4f64 VR256:$src), addr:$dst)],
853 IIC_SSE_MOVA_P_MR>, VEX, VEX_L, VEX_WIG;
854 def VMOVUPSYmr : VPSI<0x11, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
855 "movups\t{$src, $dst|$dst, $src}",
856 [(store (v8f32 VR256:$src), addr:$dst)],
857 IIC_SSE_MOVU_P_MR>, VEX, VEX_L, VEX_WIG;
858 def VMOVUPDYmr : VPDI<0x11, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
859 "movupd\t{$src, $dst|$dst, $src}",
860 [(store (v4f64 VR256:$src), addr:$dst)],
861 IIC_SSE_MOVU_P_MR>, VEX, VEX_L, VEX_WIG;
865 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
866 SchedRW = [WriteFShuffle] in {
867 def VMOVAPSrr_REV : VPSI<0x29, MRMDestReg, (outs VR128:$dst),
869 "movaps\t{$src, $dst|$dst, $src}", [],
870 IIC_SSE_MOVA_P_RR>, VEX, VEX_WIG,
871 FoldGenData<"VMOVAPSrr">;
872 def VMOVAPDrr_REV : VPDI<0x29, MRMDestReg, (outs VR128:$dst),
874 "movapd\t{$src, $dst|$dst, $src}", [],
875 IIC_SSE_MOVA_P_RR>, VEX, VEX_WIG,
876 FoldGenData<"VMOVAPDrr">;
877 def VMOVUPSrr_REV : VPSI<0x11, MRMDestReg, (outs VR128:$dst),
879 "movups\t{$src, $dst|$dst, $src}", [],
880 IIC_SSE_MOVU_P_RR>, VEX, VEX_WIG,
881 FoldGenData<"VMOVUPSrr">;
882 def VMOVUPDrr_REV : VPDI<0x11, MRMDestReg, (outs VR128:$dst),
884 "movupd\t{$src, $dst|$dst, $src}", [],
885 IIC_SSE_MOVU_P_RR>, VEX, VEX_WIG,
886 FoldGenData<"VMOVUPDrr">;
887 def VMOVAPSYrr_REV : VPSI<0x29, MRMDestReg, (outs VR256:$dst),
889 "movaps\t{$src, $dst|$dst, $src}", [],
890 IIC_SSE_MOVA_P_RR>, VEX, VEX_L, VEX_WIG,
891 FoldGenData<"VMOVAPSYrr">;
892 def VMOVAPDYrr_REV : VPDI<0x29, MRMDestReg, (outs VR256:$dst),
894 "movapd\t{$src, $dst|$dst, $src}", [],
895 IIC_SSE_MOVA_P_RR>, VEX, VEX_L, VEX_WIG,
896 FoldGenData<"VMOVAPDYrr">;
897 def VMOVUPSYrr_REV : VPSI<0x11, MRMDestReg, (outs VR256:$dst),
899 "movups\t{$src, $dst|$dst, $src}", [],
900 IIC_SSE_MOVU_P_RR>, VEX, VEX_L, VEX_WIG,
901 FoldGenData<"VMOVUPSYrr">;
902 def VMOVUPDYrr_REV : VPDI<0x11, MRMDestReg, (outs VR256:$dst),
904 "movupd\t{$src, $dst|$dst, $src}", [],
905 IIC_SSE_MOVU_P_RR>, VEX, VEX_L, VEX_WIG,
906 FoldGenData<"VMOVUPDYrr">;
909 // Aliases to help the assembler pick two byte VEX encodings by swapping the
910 // operands relative to the normal instructions to use VEX.R instead of VEX.B.
911 def : InstAlias<"vmovaps\t{$src, $dst|$dst, $src}",
912 (VMOVAPSrr_REV VR128L:$dst, VR128H:$src), 0>;
913 def : InstAlias<"vmovapd\t{$src, $dst|$dst, $src}",
914 (VMOVAPDrr_REV VR128L:$dst, VR128H:$src), 0>;
915 def : InstAlias<"vmovups\t{$src, $dst|$dst, $src}",
916 (VMOVUPSrr_REV VR128L:$dst, VR128H:$src), 0>;
917 def : InstAlias<"vmovupd\t{$src, $dst|$dst, $src}",
918 (VMOVUPDrr_REV VR128L:$dst, VR128H:$src), 0>;
919 def : InstAlias<"vmovaps\t{$src, $dst|$dst, $src}",
920 (VMOVAPSYrr_REV VR256L:$dst, VR256H:$src), 0>;
921 def : InstAlias<"vmovapd\t{$src, $dst|$dst, $src}",
922 (VMOVAPDYrr_REV VR256L:$dst, VR256H:$src), 0>;
923 def : InstAlias<"vmovups\t{$src, $dst|$dst, $src}",
924 (VMOVUPSYrr_REV VR256L:$dst, VR256H:$src), 0>;
925 def : InstAlias<"vmovupd\t{$src, $dst|$dst, $src}",
926 (VMOVUPDYrr_REV VR256L:$dst, VR256H:$src), 0>;
928 let SchedRW = [WriteStore] in {
929 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
930 "movaps\t{$src, $dst|$dst, $src}",
931 [(alignedstore (v4f32 VR128:$src), addr:$dst)],
933 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
934 "movapd\t{$src, $dst|$dst, $src}",
935 [(alignedstore (v2f64 VR128:$src), addr:$dst)],
937 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
938 "movups\t{$src, $dst|$dst, $src}",
939 [(store (v4f32 VR128:$src), addr:$dst)],
941 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
942 "movupd\t{$src, $dst|$dst, $src}",
943 [(store (v2f64 VR128:$src), addr:$dst)],
948 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
949 SchedRW = [WriteFShuffle] in {
950 def MOVAPSrr_REV : PSI<0x29, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
951 "movaps\t{$src, $dst|$dst, $src}", [],
952 IIC_SSE_MOVA_P_RR>, FoldGenData<"MOVAPSrr">;
953 def MOVAPDrr_REV : PDI<0x29, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
954 "movapd\t{$src, $dst|$dst, $src}", [],
955 IIC_SSE_MOVA_P_RR>, FoldGenData<"MOVAPDrr">;
956 def MOVUPSrr_REV : PSI<0x11, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
957 "movups\t{$src, $dst|$dst, $src}", [],
958 IIC_SSE_MOVU_P_RR>, FoldGenData<"MOVUPSrr">;
959 def MOVUPDrr_REV : PDI<0x11, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
960 "movupd\t{$src, $dst|$dst, $src}", [],
961 IIC_SSE_MOVU_P_RR>, FoldGenData<"MOVUPDrr">;
964 let Predicates = [HasAVX, NoVLX] in {
965 // 256-bit load/store need to use floating point load/store in case we don't
966 // have AVX2. Execution domain fixing will convert to integer if AVX2 is
967 // available and changing the domain is beneficial.
968 def : Pat<(alignedloadv4i64 addr:$src),
969 (VMOVAPSYrm addr:$src)>;
970 def : Pat<(loadv4i64 addr:$src),
971 (VMOVUPSYrm addr:$src)>;
972 def : Pat<(alignedstore256 (v4i64 VR256:$src), addr:$dst),
973 (VMOVAPSYmr addr:$dst, VR256:$src)>;
974 def : Pat<(alignedstore256 (v8i32 VR256:$src), addr:$dst),
975 (VMOVAPSYmr addr:$dst, VR256:$src)>;
976 def : Pat<(alignedstore256 (v16i16 VR256:$src), addr:$dst),
977 (VMOVAPSYmr addr:$dst, VR256:$src)>;
978 def : Pat<(alignedstore256 (v32i8 VR256:$src), addr:$dst),
979 (VMOVAPSYmr addr:$dst, VR256:$src)>;
980 def : Pat<(store (v4i64 VR256:$src), addr:$dst),
981 (VMOVUPSYmr addr:$dst, VR256:$src)>;
982 def : Pat<(store (v8i32 VR256:$src), addr:$dst),
983 (VMOVUPSYmr addr:$dst, VR256:$src)>;
984 def : Pat<(store (v16i16 VR256:$src), addr:$dst),
985 (VMOVUPSYmr addr:$dst, VR256:$src)>;
986 def : Pat<(store (v32i8 VR256:$src), addr:$dst),
987 (VMOVUPSYmr addr:$dst, VR256:$src)>;
989 // Special patterns for storing subvector extracts of lower 128-bits
990 // Its cheaper to just use VMOVAPS/VMOVUPS instead of VEXTRACTF128mr
991 def : Pat<(alignedstore (v2f64 (extract_subvector
992 (v4f64 VR256:$src), (iPTR 0))), addr:$dst),
993 (VMOVAPDmr addr:$dst, (v2f64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
994 def : Pat<(alignedstore (v4f32 (extract_subvector
995 (v8f32 VR256:$src), (iPTR 0))), addr:$dst),
996 (VMOVAPSmr addr:$dst, (v4f32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
998 def : Pat<(store (v2f64 (extract_subvector
999 (v4f64 VR256:$src), (iPTR 0))), addr:$dst),
1000 (VMOVUPDmr addr:$dst, (v2f64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1001 def : Pat<(store (v4f32 (extract_subvector
1002 (v8f32 VR256:$src), (iPTR 0))), addr:$dst),
1003 (VMOVUPSmr addr:$dst, (v4f32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1006 // Use movaps / movups for SSE integer load / store (one byte shorter).
1007 // The instructions selected below are then converted to MOVDQA/MOVDQU
1008 // during the SSE domain pass.
1009 let Predicates = [UseSSE1] in {
1010 def : Pat<(alignedloadv2i64 addr:$src),
1011 (MOVAPSrm addr:$src)>;
1012 def : Pat<(loadv2i64 addr:$src),
1013 (MOVUPSrm addr:$src)>;
1015 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
1016 (MOVAPSmr addr:$dst, VR128:$src)>;
1017 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
1018 (MOVAPSmr addr:$dst, VR128:$src)>;
1019 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
1020 (MOVAPSmr addr:$dst, VR128:$src)>;
1021 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
1022 (MOVAPSmr addr:$dst, VR128:$src)>;
1023 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
1024 (MOVUPSmr addr:$dst, VR128:$src)>;
1025 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
1026 (MOVUPSmr addr:$dst, VR128:$src)>;
1027 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
1028 (MOVUPSmr addr:$dst, VR128:$src)>;
1029 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
1030 (MOVUPSmr addr:$dst, VR128:$src)>;
1033 //===----------------------------------------------------------------------===//
1034 // SSE 1 & 2 - Move Low packed FP Instructions
1035 //===----------------------------------------------------------------------===//
1037 multiclass sse12_mov_hilo_packed_base<bits<8>opc, SDNode psnode, SDNode pdnode,
1038 string base_opc, string asm_opr,
1039 InstrItinClass itin> {
1040 def PSrm : PI<opc, MRMSrcMem,
1041 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1042 !strconcat(base_opc, "s", asm_opr),
1044 (psnode VR128:$src1,
1045 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))],
1046 itin, SSEPackedSingle>, PS,
1047 Sched<[WriteFShuffleLd, ReadAfterLd]>;
1049 def PDrm : PI<opc, MRMSrcMem,
1050 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1051 !strconcat(base_opc, "d", asm_opr),
1052 [(set VR128:$dst, (v2f64 (pdnode VR128:$src1,
1053 (scalar_to_vector (loadf64 addr:$src2)))))],
1054 itin, SSEPackedDouble>, PD,
1055 Sched<[WriteFShuffleLd, ReadAfterLd]>;
1059 multiclass sse12_mov_hilo_packed<bits<8>opc, SDNode psnode, SDNode pdnode,
1060 string base_opc, InstrItinClass itin> {
1061 let Predicates = [UseAVX] in
1062 defm V#NAME : sse12_mov_hilo_packed_base<opc, psnode, pdnode, base_opc,
1063 "\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1064 itin>, VEX_4V, VEX_WIG;
1066 let Constraints = "$src1 = $dst" in
1067 defm NAME : sse12_mov_hilo_packed_base<opc, psnode, pdnode, base_opc,
1068 "\t{$src2, $dst|$dst, $src2}",
1072 let AddedComplexity = 20 in {
1073 defm MOVL : sse12_mov_hilo_packed<0x12, X86Movlps, X86Movlpd, "movlp",
1077 let SchedRW = [WriteStore] in {
1078 let Predicates = [UseAVX] in {
1079 def VMOVLPSmr : VPSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1080 "movlps\t{$src, $dst|$dst, $src}",
1081 [(store (f64 (extractelt (bc_v2f64 (v4f32 VR128:$src)),
1082 (iPTR 0))), addr:$dst)],
1083 IIC_SSE_MOV_LH>, VEX, VEX_WIG;
1084 def VMOVLPDmr : VPDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1085 "movlpd\t{$src, $dst|$dst, $src}",
1086 [(store (f64 (extractelt (v2f64 VR128:$src),
1087 (iPTR 0))), addr:$dst)],
1088 IIC_SSE_MOV_LH>, VEX, VEX_WIG;
1090 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1091 "movlps\t{$src, $dst|$dst, $src}",
1092 [(store (f64 (extractelt (bc_v2f64 (v4f32 VR128:$src)),
1093 (iPTR 0))), addr:$dst)],
1095 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1096 "movlpd\t{$src, $dst|$dst, $src}",
1097 [(store (f64 (extractelt (v2f64 VR128:$src),
1098 (iPTR 0))), addr:$dst)],
1102 let Predicates = [UseAVX] in {
1103 // Shuffle with VMOVLPS
1104 def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))),
1105 (VMOVLPSrm VR128:$src1, addr:$src2)>;
1106 def : Pat<(v4i32 (X86Movlps VR128:$src1, (load addr:$src2))),
1107 (VMOVLPSrm VR128:$src1, addr:$src2)>;
1109 // Shuffle with VMOVLPD
1110 def : Pat<(v2f64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1111 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1112 def : Pat<(v2i64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1113 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1114 def : Pat<(v2f64 (X86Movsd VR128:$src1,
1115 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))),
1116 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1119 def : Pat<(store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)),
1121 (VMOVLPSmr addr:$src1, VR128:$src2)>;
1122 def : Pat<(store (v4i32 (X86Movlps
1123 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)), addr:$src1),
1124 (VMOVLPSmr addr:$src1, VR128:$src2)>;
1125 def : Pat<(store (v2f64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1127 (VMOVLPDmr addr:$src1, VR128:$src2)>;
1128 def : Pat<(store (v2i64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1130 (VMOVLPDmr addr:$src1, VR128:$src2)>;
1133 let Predicates = [UseSSE1] in {
1134 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
1135 def : Pat<(store (i64 (extractelt (bc_v2i64 (v4f32 VR128:$src2)),
1136 (iPTR 0))), addr:$src1),
1137 (MOVLPSmr addr:$src1, VR128:$src2)>;
1139 // Shuffle with MOVLPS
1140 def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))),
1141 (MOVLPSrm VR128:$src1, addr:$src2)>;
1142 def : Pat<(v4i32 (X86Movlps VR128:$src1, (load addr:$src2))),
1143 (MOVLPSrm VR128:$src1, addr:$src2)>;
1144 def : Pat<(X86Movlps VR128:$src1,
1145 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1146 (MOVLPSrm VR128:$src1, addr:$src2)>;
1149 def : Pat<(store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)),
1151 (MOVLPSmr addr:$src1, VR128:$src2)>;
1152 def : Pat<(store (v4i32 (X86Movlps
1153 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
1155 (MOVLPSmr addr:$src1, VR128:$src2)>;
1158 let Predicates = [UseSSE2] in {
1159 // Shuffle with MOVLPD
1160 def : Pat<(v2f64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1161 (MOVLPDrm VR128:$src1, addr:$src2)>;
1162 def : Pat<(v2i64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1163 (MOVLPDrm VR128:$src1, addr:$src2)>;
1164 def : Pat<(v2f64 (X86Movsd VR128:$src1,
1165 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))),
1166 (MOVLPDrm VR128:$src1, addr:$src2)>;
1169 def : Pat<(store (v2f64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1171 (MOVLPDmr addr:$src1, VR128:$src2)>;
1172 def : Pat<(store (v2i64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1174 (MOVLPDmr addr:$src1, VR128:$src2)>;
1177 //===----------------------------------------------------------------------===//
1178 // SSE 1 & 2 - Move Hi packed FP Instructions
1179 //===----------------------------------------------------------------------===//
1181 let AddedComplexity = 20 in {
1182 defm MOVH : sse12_mov_hilo_packed<0x16, X86Movlhps, X86Movlhpd, "movhp",
1186 let SchedRW = [WriteStore] in {
1187 // v2f64 extract element 1 is always custom lowered to unpack high to low
1188 // and extract element 0 so the non-store version isn't too horrible.
1189 let Predicates = [UseAVX] in {
1190 def VMOVHPSmr : VPSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1191 "movhps\t{$src, $dst|$dst, $src}",
1192 [(store (f64 (extractelt
1193 (X86Unpckh (bc_v2f64 (v4f32 VR128:$src)),
1194 (bc_v2f64 (v4f32 VR128:$src))),
1195 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>, VEX, VEX_WIG;
1196 def VMOVHPDmr : VPDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1197 "movhpd\t{$src, $dst|$dst, $src}",
1198 [(store (f64 (extractelt
1199 (v2f64 (X86Unpckh VR128:$src, VR128:$src)),
1200 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>, VEX, VEX_WIG;
1202 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1203 "movhps\t{$src, $dst|$dst, $src}",
1204 [(store (f64 (extractelt
1205 (X86Unpckh (bc_v2f64 (v4f32 VR128:$src)),
1206 (bc_v2f64 (v4f32 VR128:$src))),
1207 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>;
1208 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1209 "movhpd\t{$src, $dst|$dst, $src}",
1210 [(store (f64 (extractelt
1211 (v2f64 (X86Unpckh VR128:$src, VR128:$src)),
1212 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>;
1215 let Predicates = [UseAVX] in {
1217 def : Pat<(X86Movlhps VR128:$src1,
1218 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1219 (VMOVHPSrm VR128:$src1, addr:$src2)>;
1220 def : Pat<(X86Movlhps VR128:$src1,
1221 (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
1222 (VMOVHPSrm VR128:$src1, addr:$src2)>;
1226 // FIXME: Instead of X86Unpckl, there should be a X86Movlhpd here, the problem
1227 // is during lowering, where it's not possible to recognize the load fold
1228 // cause it has two uses through a bitcast. One use disappears at isel time
1229 // and the fold opportunity reappears.
1230 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1231 (scalar_to_vector (loadf64 addr:$src2)))),
1232 (VMOVHPDrm VR128:$src1, addr:$src2)>;
1234 // Also handle an i64 load because that may get selected as a faster way to
1236 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1237 (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src2)))))),
1238 (VMOVHPDrm VR128:$src1, addr:$src2)>;
1240 def : Pat<(store (f64 (extractelt
1241 (bc_v2f64 (v4f32 (X86Movhlps VR128:$src, VR128:$src))),
1242 (iPTR 0))), addr:$dst),
1243 (VMOVHPDmr addr:$dst, VR128:$src)>;
1245 def : Pat<(store (f64 (extractelt
1246 (v2f64 (X86VPermilpi VR128:$src, (i8 1))),
1247 (iPTR 0))), addr:$dst),
1248 (VMOVHPDmr addr:$dst, VR128:$src)>;
1251 let Predicates = [UseSSE1] in {
1253 def : Pat<(X86Movlhps VR128:$src1,
1254 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1255 (MOVHPSrm VR128:$src1, addr:$src2)>;
1256 def : Pat<(X86Movlhps VR128:$src1,
1257 (bc_v4f32 (v2i64 (X86vzload addr:$src2)))),
1258 (MOVHPSrm VR128:$src1, addr:$src2)>;
1261 let Predicates = [UseSSE2] in {
1264 // FIXME: Instead of X86Unpckl, there should be a X86Movlhpd here, the problem
1265 // is during lowering, where it's not possible to recognize the load fold
1266 // cause it has two uses through a bitcast. One use disappears at isel time
1267 // and the fold opportunity reappears.
1268 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1269 (scalar_to_vector (loadf64 addr:$src2)))),
1270 (MOVHPDrm VR128:$src1, addr:$src2)>;
1272 // Also handle an i64 load because that may get selected as a faster way to
1274 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1275 (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src2)))))),
1276 (MOVHPDrm VR128:$src1, addr:$src2)>;
1278 def : Pat<(store (f64 (extractelt
1279 (bc_v2f64 (v4f32 (X86Movhlps VR128:$src, VR128:$src))),
1280 (iPTR 0))), addr:$dst),
1281 (MOVHPDmr addr:$dst, VR128:$src)>;
1283 def : Pat<(store (f64 (extractelt
1284 (v2f64 (X86Shufp VR128:$src, VR128:$src, (i8 1))),
1285 (iPTR 0))), addr:$dst),
1286 (MOVHPDmr addr:$dst, VR128:$src)>;
1289 //===----------------------------------------------------------------------===//
1290 // SSE 1 & 2 - Move Low to High and High to Low packed FP Instructions
1291 //===----------------------------------------------------------------------===//
1293 let AddedComplexity = 20, Predicates = [UseAVX] in {
1294 def VMOVLHPSrr : VPSI<0x16, MRMSrcReg, (outs VR128:$dst),
1295 (ins VR128:$src1, VR128:$src2),
1296 "movlhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1298 (v4f32 (X86Movlhps VR128:$src1, VR128:$src2)))],
1300 VEX_4V, Sched<[WriteFShuffle]>, VEX_WIG;
1301 def VMOVHLPSrr : VPSI<0x12, MRMSrcReg, (outs VR128:$dst),
1302 (ins VR128:$src1, VR128:$src2),
1303 "movhlps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1305 (v4f32 (X86Movhlps VR128:$src1, VR128:$src2)))],
1307 VEX_4V, Sched<[WriteFShuffle]>, VEX_WIG;
1309 let Constraints = "$src1 = $dst", AddedComplexity = 20 in {
1310 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
1311 (ins VR128:$src1, VR128:$src2),
1312 "movlhps\t{$src2, $dst|$dst, $src2}",
1314 (v4f32 (X86Movlhps VR128:$src1, VR128:$src2)))],
1315 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
1316 let isCommutable = 1 in
1317 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
1318 (ins VR128:$src1, VR128:$src2),
1319 "movhlps\t{$src2, $dst|$dst, $src2}",
1321 (v4f32 (X86Movhlps VR128:$src1, VR128:$src2)))],
1322 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
1325 let Predicates = [UseAVX] in {
1327 def : Pat<(v4i32 (X86Movlhps VR128:$src1, VR128:$src2)),
1328 (VMOVLHPSrr VR128:$src1, VR128:$src2)>;
1329 def : Pat<(v2i64 (X86Movlhps VR128:$src1, VR128:$src2)),
1330 (VMOVLHPSrr (v2i64 VR128:$src1), VR128:$src2)>;
1333 def : Pat<(v4i32 (X86Movhlps VR128:$src1, VR128:$src2)),
1334 (VMOVHLPSrr VR128:$src1, VR128:$src2)>;
1337 let Predicates = [UseSSE1] in {
1339 def : Pat<(v4i32 (X86Movlhps VR128:$src1, VR128:$src2)),
1340 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
1341 def : Pat<(v2i64 (X86Movlhps VR128:$src1, VR128:$src2)),
1342 (MOVLHPSrr (v2i64 VR128:$src1), VR128:$src2)>;
1345 def : Pat<(v4i32 (X86Movhlps VR128:$src1, VR128:$src2)),
1346 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
1349 //===----------------------------------------------------------------------===//
1350 // SSE 1 & 2 - Conversion Instructions
1351 //===----------------------------------------------------------------------===//
1353 def SSE_CVT_PD : OpndItins<
1354 IIC_SSE_CVT_PD_RR, IIC_SSE_CVT_PD_RM
1357 let Sched = WriteCvtI2F in
1358 def SSE_CVT_PS : OpndItins<
1359 IIC_SSE_CVT_PS_RR, IIC_SSE_CVT_PS_RM
1362 let Sched = WriteCvtI2F in
1363 def SSE_CVT_Scalar : OpndItins<
1364 IIC_SSE_CVT_Scalar_RR, IIC_SSE_CVT_Scalar_RM
1367 let Sched = WriteCvtF2I in
1368 def SSE_CVT_SS2SI_32 : OpndItins<
1369 IIC_SSE_CVT_SS2SI32_RR, IIC_SSE_CVT_SS2SI32_RM
1372 let Sched = WriteCvtF2I in
1373 def SSE_CVT_SS2SI_64 : OpndItins<
1374 IIC_SSE_CVT_SS2SI64_RR, IIC_SSE_CVT_SS2SI64_RM
1377 let Sched = WriteCvtF2I in
1378 def SSE_CVT_SD2SI : OpndItins<
1379 IIC_SSE_CVT_SD2SI_RR, IIC_SSE_CVT_SD2SI_RM
1382 // FIXME: We probably want to match the rm form only when optimizing for
1383 // size, to avoid false depenendecies (see sse_fp_unop_s for details)
1384 multiclass sse12_cvt_s<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1385 SDNode OpNode, X86MemOperand x86memop, PatFrag ld_frag,
1386 string asm, OpndItins itins> {
1387 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
1388 [(set DstRC:$dst, (OpNode SrcRC:$src))],
1389 itins.rr>, Sched<[itins.Sched]>;
1390 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
1391 [(set DstRC:$dst, (OpNode (ld_frag addr:$src)))],
1392 itins.rm>, Sched<[itins.Sched.Folded]>;
1395 multiclass sse12_cvt_p<bits<8> opc, RegisterClass RC, X86MemOperand x86memop,
1396 ValueType DstTy, ValueType SrcTy, PatFrag ld_frag,
1397 string asm, Domain d, OpndItins itins> {
1398 let hasSideEffects = 0 in {
1399 def rr : I<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src), asm,
1400 [(set RC:$dst, (DstTy (sint_to_fp (SrcTy RC:$src))))],
1401 itins.rr, d>, Sched<[itins.Sched]>;
1403 def rm : I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src), asm,
1404 [(set RC:$dst, (DstTy (sint_to_fp
1405 (SrcTy (bitconvert (ld_frag addr:$src))))))],
1406 itins.rm, d>, Sched<[itins.Sched.Folded]>;
1410 // FIXME: We probably want to match the rm form only when optimizing for
1411 // size, to avoid false depenendecies (see sse_fp_unop_s for details)
1412 multiclass sse12_vcvt_avx<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1413 X86MemOperand x86memop, string asm> {
1414 let hasSideEffects = 0, Predicates = [UseAVX] in {
1415 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src),
1416 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>,
1417 Sched<[WriteCvtI2F]>;
1419 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
1420 (ins DstRC:$src1, x86memop:$src),
1421 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>,
1422 Sched<[WriteCvtI2FLd, ReadAfterLd]>;
1423 } // hasSideEffects = 0
1426 let Predicates = [UseAVX] in {
1427 defm VCVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
1428 "cvttss2si\t{$src, $dst|$dst, $src}",
1431 defm VCVTTSS2SI64 : sse12_cvt_s<0x2C, FR32, GR64, fp_to_sint, f32mem, loadf32,
1432 "cvttss2si\t{$src, $dst|$dst, $src}",
1434 XS, VEX, VEX_W, VEX_LIG;
1435 defm VCVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
1436 "cvttsd2si\t{$src, $dst|$dst, $src}",
1439 defm VCVTTSD2SI64 : sse12_cvt_s<0x2C, FR64, GR64, fp_to_sint, f64mem, loadf64,
1440 "cvttsd2si\t{$src, $dst|$dst, $src}",
1442 XD, VEX, VEX_W, VEX_LIG;
1444 def : InstAlias<"vcvttss2si{l}\t{$src, $dst|$dst, $src}",
1445 (VCVTTSS2SIrr GR32:$dst, FR32:$src), 0>;
1446 def : InstAlias<"vcvttss2si{l}\t{$src, $dst|$dst, $src}",
1447 (VCVTTSS2SIrm GR32:$dst, f32mem:$src), 0>;
1448 def : InstAlias<"vcvttsd2si{l}\t{$src, $dst|$dst, $src}",
1449 (VCVTTSD2SIrr GR32:$dst, FR64:$src), 0>;
1450 def : InstAlias<"vcvttsd2si{l}\t{$src, $dst|$dst, $src}",
1451 (VCVTTSD2SIrm GR32:$dst, f64mem:$src), 0>;
1452 def : InstAlias<"vcvttss2si{q}\t{$src, $dst|$dst, $src}",
1453 (VCVTTSS2SI64rr GR64:$dst, FR32:$src), 0>;
1454 def : InstAlias<"vcvttss2si{q}\t{$src, $dst|$dst, $src}",
1455 (VCVTTSS2SI64rm GR64:$dst, f32mem:$src), 0>;
1456 def : InstAlias<"vcvttsd2si{q}\t{$src, $dst|$dst, $src}",
1457 (VCVTTSD2SI64rr GR64:$dst, FR64:$src), 0>;
1458 def : InstAlias<"vcvttsd2si{q}\t{$src, $dst|$dst, $src}",
1459 (VCVTTSD2SI64rm GR64:$dst, f64mem:$src), 0>;
1461 // The assembler can recognize rr 64-bit instructions by seeing a rxx
1462 // register, but the same isn't true when only using memory operands,
1463 // provide other assembly "l" and "q" forms to address this explicitly
1464 // where appropriate to do so.
1465 defm VCVTSI2SS : sse12_vcvt_avx<0x2A, GR32, FR32, i32mem, "cvtsi2ss{l}">,
1466 XS, VEX_4V, VEX_LIG;
1467 defm VCVTSI2SS64 : sse12_vcvt_avx<0x2A, GR64, FR32, i64mem, "cvtsi2ss{q}">,
1468 XS, VEX_4V, VEX_W, VEX_LIG;
1469 defm VCVTSI2SD : sse12_vcvt_avx<0x2A, GR32, FR64, i32mem, "cvtsi2sd{l}">,
1470 XD, VEX_4V, VEX_LIG;
1471 defm VCVTSI2SD64 : sse12_vcvt_avx<0x2A, GR64, FR64, i64mem, "cvtsi2sd{q}">,
1472 XD, VEX_4V, VEX_W, VEX_LIG;
1474 let Predicates = [UseAVX] in {
1475 def : InstAlias<"vcvtsi2ss\t{$src, $src1, $dst|$dst, $src1, $src}",
1476 (VCVTSI2SSrm FR64:$dst, FR64:$src1, i32mem:$src), 0>;
1477 def : InstAlias<"vcvtsi2sd\t{$src, $src1, $dst|$dst, $src1, $src}",
1478 (VCVTSI2SDrm FR64:$dst, FR64:$src1, i32mem:$src), 0>;
1480 def : Pat<(f32 (sint_to_fp (loadi32 addr:$src))),
1481 (VCVTSI2SSrm (f32 (IMPLICIT_DEF)), addr:$src)>;
1482 def : Pat<(f32 (sint_to_fp (loadi64 addr:$src))),
1483 (VCVTSI2SS64rm (f32 (IMPLICIT_DEF)), addr:$src)>;
1484 def : Pat<(f64 (sint_to_fp (loadi32 addr:$src))),
1485 (VCVTSI2SDrm (f64 (IMPLICIT_DEF)), addr:$src)>;
1486 def : Pat<(f64 (sint_to_fp (loadi64 addr:$src))),
1487 (VCVTSI2SD64rm (f64 (IMPLICIT_DEF)), addr:$src)>;
1489 def : Pat<(f32 (sint_to_fp GR32:$src)),
1490 (VCVTSI2SSrr (f32 (IMPLICIT_DEF)), GR32:$src)>;
1491 def : Pat<(f32 (sint_to_fp GR64:$src)),
1492 (VCVTSI2SS64rr (f32 (IMPLICIT_DEF)), GR64:$src)>;
1493 def : Pat<(f64 (sint_to_fp GR32:$src)),
1494 (VCVTSI2SDrr (f64 (IMPLICIT_DEF)), GR32:$src)>;
1495 def : Pat<(f64 (sint_to_fp GR64:$src)),
1496 (VCVTSI2SD64rr (f64 (IMPLICIT_DEF)), GR64:$src)>;
1499 defm CVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
1500 "cvttss2si\t{$src, $dst|$dst, $src}",
1501 SSE_CVT_SS2SI_32>, XS;
1502 defm CVTTSS2SI64 : sse12_cvt_s<0x2C, FR32, GR64, fp_to_sint, f32mem, loadf32,
1503 "cvttss2si\t{$src, $dst|$dst, $src}",
1504 SSE_CVT_SS2SI_64>, XS, REX_W;
1505 defm CVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
1506 "cvttsd2si\t{$src, $dst|$dst, $src}",
1508 defm CVTTSD2SI64 : sse12_cvt_s<0x2C, FR64, GR64, fp_to_sint, f64mem, loadf64,
1509 "cvttsd2si\t{$src, $dst|$dst, $src}",
1510 SSE_CVT_SD2SI>, XD, REX_W;
1511 defm CVTSI2SS : sse12_cvt_s<0x2A, GR32, FR32, sint_to_fp, i32mem, loadi32,
1512 "cvtsi2ss{l}\t{$src, $dst|$dst, $src}",
1513 SSE_CVT_Scalar>, XS;
1514 defm CVTSI2SS64 : sse12_cvt_s<0x2A, GR64, FR32, sint_to_fp, i64mem, loadi64,
1515 "cvtsi2ss{q}\t{$src, $dst|$dst, $src}",
1516 SSE_CVT_Scalar>, XS, REX_W;
1517 defm CVTSI2SD : sse12_cvt_s<0x2A, GR32, FR64, sint_to_fp, i32mem, loadi32,
1518 "cvtsi2sd{l}\t{$src, $dst|$dst, $src}",
1519 SSE_CVT_Scalar>, XD;
1520 defm CVTSI2SD64 : sse12_cvt_s<0x2A, GR64, FR64, sint_to_fp, i64mem, loadi64,
1521 "cvtsi2sd{q}\t{$src, $dst|$dst, $src}",
1522 SSE_CVT_Scalar>, XD, REX_W;
1524 def : InstAlias<"cvttss2si{l}\t{$src, $dst|$dst, $src}",
1525 (CVTTSS2SIrr GR32:$dst, FR32:$src), 0>;
1526 def : InstAlias<"cvttss2si{l}\t{$src, $dst|$dst, $src}",
1527 (CVTTSS2SIrm GR32:$dst, f32mem:$src), 0>;
1528 def : InstAlias<"cvttsd2si{l}\t{$src, $dst|$dst, $src}",
1529 (CVTTSD2SIrr GR32:$dst, FR64:$src), 0>;
1530 def : InstAlias<"cvttsd2si{l}\t{$src, $dst|$dst, $src}",
1531 (CVTTSD2SIrm GR32:$dst, f64mem:$src), 0>;
1532 def : InstAlias<"cvttss2si{q}\t{$src, $dst|$dst, $src}",
1533 (CVTTSS2SI64rr GR64:$dst, FR32:$src), 0>;
1534 def : InstAlias<"cvttss2si{q}\t{$src, $dst|$dst, $src}",
1535 (CVTTSS2SI64rm GR64:$dst, f32mem:$src), 0>;
1536 def : InstAlias<"cvttsd2si{q}\t{$src, $dst|$dst, $src}",
1537 (CVTTSD2SI64rr GR64:$dst, FR64:$src), 0>;
1538 def : InstAlias<"cvttsd2si{q}\t{$src, $dst|$dst, $src}",
1539 (CVTTSD2SI64rm GR64:$dst, f64mem:$src), 0>;
1541 def : InstAlias<"cvtsi2ss\t{$src, $dst|$dst, $src}",
1542 (CVTSI2SSrm FR64:$dst, i32mem:$src), 0>;
1543 def : InstAlias<"cvtsi2sd\t{$src, $dst|$dst, $src}",
1544 (CVTSI2SDrm FR64:$dst, i32mem:$src), 0>;
1546 // Conversion Instructions Intrinsics - Match intrinsics which expect MM
1547 // and/or XMM operand(s).
1549 // FIXME: We probably want to match the rm form only when optimizing for
1550 // size, to avoid false depenendecies (see sse_fp_unop_s for details)
1551 multiclass sse12_cvt_sint<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1552 Intrinsic Int, Operand memop, ComplexPattern mem_cpat,
1553 string asm, OpndItins itins> {
1554 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src),
1555 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1556 [(set DstRC:$dst, (Int SrcRC:$src))], itins.rr>,
1557 Sched<[itins.Sched]>;
1558 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins memop:$src),
1559 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1560 [(set DstRC:$dst, (Int mem_cpat:$src))], itins.rm>,
1561 Sched<[itins.Sched.Folded]>;
1564 multiclass sse12_cvt_sint_3addr<bits<8> opc, RegisterClass SrcRC,
1565 RegisterClass DstRC, Intrinsic Int, X86MemOperand x86memop,
1566 PatFrag ld_frag, string asm, OpndItins itins,
1568 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src2),
1570 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
1571 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
1572 [(set DstRC:$dst, (Int DstRC:$src1, SrcRC:$src2))],
1573 itins.rr>, Sched<[itins.Sched]>;
1574 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
1575 (ins DstRC:$src1, x86memop:$src2),
1577 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
1578 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
1579 [(set DstRC:$dst, (Int DstRC:$src1, (ld_frag addr:$src2)))],
1580 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
1583 let Predicates = [UseAVX] in {
1584 defm VCVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32,
1585 int_x86_sse2_cvtsd2si, sdmem, sse_load_f64, "cvtsd2si",
1586 SSE_CVT_SD2SI>, XD, VEX, VEX_LIG;
1587 defm VCVTSD2SI64 : sse12_cvt_sint<0x2D, VR128, GR64,
1588 int_x86_sse2_cvtsd2si64, sdmem, sse_load_f64, "cvtsd2si",
1589 SSE_CVT_SD2SI>, XD, VEX, VEX_W, VEX_LIG;
1591 defm CVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse2_cvtsd2si,
1592 sdmem, sse_load_f64, "cvtsd2si", SSE_CVT_SD2SI>, XD;
1593 defm CVTSD2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse2_cvtsd2si64,
1594 sdmem, sse_load_f64, "cvtsd2si", SSE_CVT_SD2SI>, XD, REX_W;
1597 let isCodeGenOnly = 1 in {
1598 let Predicates = [UseAVX] in {
1599 defm Int_VCVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1600 int_x86_sse_cvtsi2ss, i32mem, loadi32, "cvtsi2ss{l}",
1601 SSE_CVT_Scalar, 0>, XS, VEX_4V;
1602 defm Int_VCVTSI2SS64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1603 int_x86_sse_cvtsi642ss, i64mem, loadi64, "cvtsi2ss{q}",
1604 SSE_CVT_Scalar, 0>, XS, VEX_4V,
1606 defm Int_VCVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1607 int_x86_sse2_cvtsi2sd, i32mem, loadi32, "cvtsi2sd{l}",
1608 SSE_CVT_Scalar, 0>, XD, VEX_4V;
1609 defm Int_VCVTSI2SD64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1610 int_x86_sse2_cvtsi642sd, i64mem, loadi64, "cvtsi2sd{q}",
1611 SSE_CVT_Scalar, 0>, XD,
1614 let Constraints = "$src1 = $dst" in {
1615 defm Int_CVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1616 int_x86_sse_cvtsi2ss, i32mem, loadi32,
1617 "cvtsi2ss{l}", SSE_CVT_Scalar>, XS;
1618 defm Int_CVTSI2SS64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1619 int_x86_sse_cvtsi642ss, i64mem, loadi64,
1620 "cvtsi2ss{q}", SSE_CVT_Scalar>, XS, REX_W;
1621 defm Int_CVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1622 int_x86_sse2_cvtsi2sd, i32mem, loadi32,
1623 "cvtsi2sd{l}", SSE_CVT_Scalar>, XD;
1624 defm Int_CVTSI2SD64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1625 int_x86_sse2_cvtsi642sd, i64mem, loadi64,
1626 "cvtsi2sd{q}", SSE_CVT_Scalar>, XD, REX_W;
1628 } // isCodeGenOnly = 1
1632 // Aliases for intrinsics
1633 let isCodeGenOnly = 1 in {
1634 let Predicates = [UseAVX] in {
1635 defm Int_VCVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
1636 ssmem, sse_load_f32, "cvttss2si",
1637 SSE_CVT_SS2SI_32>, XS, VEX;
1638 defm Int_VCVTTSS2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1639 int_x86_sse_cvttss2si64, ssmem, sse_load_f32,
1640 "cvttss2si", SSE_CVT_SS2SI_64>,
1642 defm Int_VCVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
1643 sdmem, sse_load_f64, "cvttsd2si",
1644 SSE_CVT_SD2SI>, XD, VEX;
1645 defm Int_VCVTTSD2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1646 int_x86_sse2_cvttsd2si64, sdmem, sse_load_f64,
1647 "cvttsd2si", SSE_CVT_SD2SI>,
1650 defm Int_CVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
1651 ssmem, sse_load_f32, "cvttss2si",
1652 SSE_CVT_SS2SI_32>, XS;
1653 defm Int_CVTTSS2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1654 int_x86_sse_cvttss2si64, ssmem, sse_load_f32,
1655 "cvttss2si", SSE_CVT_SS2SI_64>, XS, REX_W;
1656 defm Int_CVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
1657 sdmem, sse_load_f64, "cvttsd2si",
1659 defm Int_CVTTSD2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1660 int_x86_sse2_cvttsd2si64, sdmem, sse_load_f64,
1661 "cvttsd2si", SSE_CVT_SD2SI>, XD, REX_W;
1662 } // isCodeGenOnly = 1
1664 let Predicates = [UseAVX] in {
1665 defm VCVTSS2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse_cvtss2si,
1666 ssmem, sse_load_f32, "cvtss2si",
1667 SSE_CVT_SS2SI_32>, XS, VEX, VEX_LIG;
1668 defm VCVTSS2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse_cvtss2si64,
1669 ssmem, sse_load_f32, "cvtss2si",
1670 SSE_CVT_SS2SI_64>, XS, VEX, VEX_W, VEX_LIG;
1672 defm CVTSS2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse_cvtss2si,
1673 ssmem, sse_load_f32, "cvtss2si",
1674 SSE_CVT_SS2SI_32>, XS;
1675 defm CVTSS2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse_cvtss2si64,
1676 ssmem, sse_load_f32, "cvtss2si",
1677 SSE_CVT_SS2SI_64>, XS, REX_W;
1679 defm VCVTDQ2PS : sse12_cvt_p<0x5B, VR128, i128mem, v4f32, v4i32, loadv2i64,
1680 "vcvtdq2ps\t{$src, $dst|$dst, $src}",
1681 SSEPackedSingle, SSE_CVT_PS>,
1682 PS, VEX, Requires<[HasAVX, NoVLX]>, VEX_WIG;
1683 defm VCVTDQ2PSY : sse12_cvt_p<0x5B, VR256, i256mem, v8f32, v8i32, loadv4i64,
1684 "vcvtdq2ps\t{$src, $dst|$dst, $src}",
1685 SSEPackedSingle, SSE_CVT_PS>,
1686 PS, VEX, VEX_L, Requires<[HasAVX, NoVLX]>, VEX_WIG;
1688 defm CVTDQ2PS : sse12_cvt_p<0x5B, VR128, i128mem, v4f32, v4i32, memopv2i64,
1689 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1690 SSEPackedSingle, SSE_CVT_PS>,
1691 PS, Requires<[UseSSE2]>;
1693 let Predicates = [UseAVX] in {
1694 def : InstAlias<"vcvtss2si{l}\t{$src, $dst|$dst, $src}",
1695 (VCVTSS2SIrr GR32:$dst, VR128:$src), 0>;
1696 def : InstAlias<"vcvtss2si{l}\t{$src, $dst|$dst, $src}",
1697 (VCVTSS2SIrm GR32:$dst, ssmem:$src), 0>;
1698 def : InstAlias<"vcvtsd2si{l}\t{$src, $dst|$dst, $src}",
1699 (VCVTSD2SIrr GR32:$dst, VR128:$src), 0>;
1700 def : InstAlias<"vcvtsd2si{l}\t{$src, $dst|$dst, $src}",
1701 (VCVTSD2SIrm GR32:$dst, sdmem:$src), 0>;
1702 def : InstAlias<"vcvtss2si{q}\t{$src, $dst|$dst, $src}",
1703 (VCVTSS2SI64rr GR64:$dst, VR128:$src), 0>;
1704 def : InstAlias<"vcvtss2si{q}\t{$src, $dst|$dst, $src}",
1705 (VCVTSS2SI64rm GR64:$dst, ssmem:$src), 0>;
1706 def : InstAlias<"vcvtsd2si{q}\t{$src, $dst|$dst, $src}",
1707 (VCVTSD2SI64rr GR64:$dst, VR128:$src), 0>;
1708 def : InstAlias<"vcvtsd2si{q}\t{$src, $dst|$dst, $src}",
1709 (VCVTSD2SI64rm GR64:$dst, sdmem:$src), 0>;
1712 def : InstAlias<"cvtss2si{l}\t{$src, $dst|$dst, $src}",
1713 (CVTSS2SIrr GR32:$dst, VR128:$src), 0>;
1714 def : InstAlias<"cvtss2si{l}\t{$src, $dst|$dst, $src}",
1715 (CVTSS2SIrm GR32:$dst, ssmem:$src), 0>;
1716 def : InstAlias<"cvtsd2si{l}\t{$src, $dst|$dst, $src}",
1717 (CVTSD2SIrr GR32:$dst, VR128:$src), 0>;
1718 def : InstAlias<"cvtsd2si{l}\t{$src, $dst|$dst, $src}",
1719 (CVTSD2SIrm GR32:$dst, sdmem:$src), 0>;
1720 def : InstAlias<"cvtss2si{q}\t{$src, $dst|$dst, $src}",
1721 (CVTSS2SI64rr GR64:$dst, VR128:$src), 0>;
1722 def : InstAlias<"cvtss2si{q}\t{$src, $dst|$dst, $src}",
1723 (CVTSS2SI64rm GR64:$dst, ssmem:$src), 0>;
1724 def : InstAlias<"cvtsd2si{q}\t{$src, $dst|$dst, $src}",
1725 (CVTSD2SI64rr GR64:$dst, VR128:$src), 0>;
1726 def : InstAlias<"cvtsd2si{q}\t{$src, $dst|$dst, $src}",
1727 (CVTSD2SI64rm GR64:$dst, sdmem:$src), 0>;
1731 // Convert scalar double to scalar single
1732 let hasSideEffects = 0, Predicates = [UseAVX] in {
1733 def VCVTSD2SSrr : VSDI<0x5A, MRMSrcReg, (outs FR32:$dst),
1734 (ins FR32:$src1, FR64:$src2),
1735 "cvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}", [],
1736 IIC_SSE_CVT_Scalar_RR>, VEX_4V, VEX_LIG,
1737 Sched<[WriteCvtF2F]>, VEX_WIG;
1739 def VCVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst),
1740 (ins FR32:$src1, f64mem:$src2),
1741 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1742 [], IIC_SSE_CVT_Scalar_RM>,
1743 XD, Requires<[HasAVX, OptForSize]>, VEX_4V, VEX_LIG,
1744 Sched<[WriteCvtF2FLd, ReadAfterLd]>, VEX_WIG;
1747 def : Pat<(f32 (fpround FR64:$src)),
1748 (VCVTSD2SSrr (COPY_TO_REGCLASS FR64:$src, FR32), FR64:$src)>,
1751 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1752 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1753 [(set FR32:$dst, (fpround FR64:$src))],
1754 IIC_SSE_CVT_Scalar_RR>, Sched<[WriteCvtF2F]>;
1755 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1756 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1757 [(set FR32:$dst, (fpround (loadf64 addr:$src)))],
1758 IIC_SSE_CVT_Scalar_RM>,
1760 Requires<[UseSSE2, OptForSize]>, Sched<[WriteCvtF2FLd]>;
1762 let isCodeGenOnly = 1 in {
1763 def Int_VCVTSD2SSrr: I<0x5A, MRMSrcReg,
1764 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1765 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1767 (int_x86_sse2_cvtsd2ss VR128:$src1, VR128:$src2))],
1768 IIC_SSE_CVT_Scalar_RR>, XD, VEX_4V, VEX_WIG,
1769 Requires<[HasAVX]>, Sched<[WriteCvtF2F]>;
1770 def Int_VCVTSD2SSrm: I<0x5A, MRMSrcMem,
1771 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1772 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1773 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss
1774 VR128:$src1, sse_load_f64:$src2))],
1775 IIC_SSE_CVT_Scalar_RM>, XD, VEX_4V, VEX_WIG,
1776 Requires<[HasAVX]>, Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1778 let Constraints = "$src1 = $dst" in {
1779 def Int_CVTSD2SSrr: I<0x5A, MRMSrcReg,
1780 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1781 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1783 (int_x86_sse2_cvtsd2ss VR128:$src1, VR128:$src2))],
1784 IIC_SSE_CVT_Scalar_RR>, XD, Requires<[UseSSE2]>,
1785 Sched<[WriteCvtF2F]>;
1786 def Int_CVTSD2SSrm: I<0x5A, MRMSrcMem,
1787 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1788 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1789 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss
1790 VR128:$src1, sse_load_f64:$src2))],
1791 IIC_SSE_CVT_Scalar_RM>, XD, Requires<[UseSSE2]>,
1792 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1794 } // isCodeGenOnly = 1
1796 // Convert scalar single to scalar double
1797 // SSE2 instructions with XS prefix
1798 let hasSideEffects = 0, Predicates = [UseAVX] in {
1799 def VCVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst),
1800 (ins FR64:$src1, FR32:$src2),
1801 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1802 [], IIC_SSE_CVT_Scalar_RR>,
1803 XS, Requires<[HasAVX]>, VEX_4V, VEX_LIG,
1804 Sched<[WriteCvtF2F]>, VEX_WIG;
1806 def VCVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst),
1807 (ins FR64:$src1, f32mem:$src2),
1808 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1809 [], IIC_SSE_CVT_Scalar_RM>,
1810 XS, VEX_4V, VEX_LIG, Requires<[HasAVX, OptForSize]>,
1811 Sched<[WriteCvtF2FLd, ReadAfterLd]>, VEX_WIG;
1814 def : Pat<(f64 (fpextend FR32:$src)),
1815 (VCVTSS2SDrr (COPY_TO_REGCLASS FR32:$src, FR64), FR32:$src)>, Requires<[UseAVX]>;
1816 def : Pat<(fpextend (loadf32 addr:$src)),
1817 (VCVTSS2SDrm (f64 (IMPLICIT_DEF)), addr:$src)>, Requires<[UseAVX]>;
1819 def : Pat<(extloadf32 addr:$src),
1820 (VCVTSS2SDrm (f64 (IMPLICIT_DEF)), addr:$src)>,
1821 Requires<[UseAVX, OptForSize]>;
1822 def : Pat<(extloadf32 addr:$src),
1823 (VCVTSS2SDrr (f64 (IMPLICIT_DEF)), (VMOVSSrm addr:$src))>,
1824 Requires<[UseAVX, OptForSpeed]>;
1826 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1827 "cvtss2sd\t{$src, $dst|$dst, $src}",
1828 [(set FR64:$dst, (fpextend FR32:$src))],
1829 IIC_SSE_CVT_Scalar_RR>, XS,
1830 Requires<[UseSSE2]>, Sched<[WriteCvtF2F]>;
1831 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1832 "cvtss2sd\t{$src, $dst|$dst, $src}",
1833 [(set FR64:$dst, (extloadf32 addr:$src))],
1834 IIC_SSE_CVT_Scalar_RM>, XS,
1835 Requires<[UseSSE2, OptForSize]>, Sched<[WriteCvtF2FLd]>;
1837 // extload f32 -> f64. This matches load+fpextend because we have a hack in
1838 // the isel (PreprocessForFPConvert) that can introduce loads after dag
1840 // Since these loads aren't folded into the fpextend, we have to match it
1842 def : Pat<(fpextend (loadf32 addr:$src)),
1843 (CVTSS2SDrm addr:$src)>, Requires<[UseSSE2]>;
1844 def : Pat<(extloadf32 addr:$src),
1845 (CVTSS2SDrr (MOVSSrm addr:$src))>, Requires<[UseSSE2, OptForSpeed]>;
1847 let isCodeGenOnly = 1 in {
1848 def Int_VCVTSS2SDrr: I<0x5A, MRMSrcReg,
1849 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1850 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1852 (int_x86_sse2_cvtss2sd VR128:$src1, VR128:$src2))],
1853 IIC_SSE_CVT_Scalar_RR>, XS, VEX_4V, VEX_WIG,
1854 Requires<[HasAVX]>, Sched<[WriteCvtF2F]>;
1855 def Int_VCVTSS2SDrm: I<0x5A, MRMSrcMem,
1856 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
1857 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1859 (int_x86_sse2_cvtss2sd VR128:$src1, sse_load_f32:$src2))],
1860 IIC_SSE_CVT_Scalar_RM>, XS, VEX_4V, VEX_WIG,
1861 Requires<[HasAVX]>, Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1862 let Constraints = "$src1 = $dst" in { // SSE2 instructions with XS prefix
1863 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1864 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1865 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1867 (int_x86_sse2_cvtss2sd VR128:$src1, VR128:$src2))],
1868 IIC_SSE_CVT_Scalar_RR>, XS, Requires<[UseSSE2]>,
1869 Sched<[WriteCvtF2F]>;
1870 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1871 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
1872 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1874 (int_x86_sse2_cvtss2sd VR128:$src1, sse_load_f32:$src2))],
1875 IIC_SSE_CVT_Scalar_RM>, XS, Requires<[UseSSE2]>,
1876 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1878 } // isCodeGenOnly = 1
1880 // Patterns used for matching (v)cvtsi2ss, (v)cvtsi2sd, (v)cvtsd2ss and
1881 // (v)cvtss2sd intrinsic sequences from clang which produce unnecessary
1882 // vmovs{s,d} instructions
1883 let Predicates = [UseAVX] in {
1884 def : Pat<(v4f32 (X86Movss
1886 (v4f32 (scalar_to_vector
1887 (f32 (fpround (f64 (extractelt VR128:$src, (iPTR 0))))))))),
1888 (Int_VCVTSD2SSrr VR128:$dst, VR128:$src)>;
1890 def : Pat<(v2f64 (X86Movsd
1892 (v2f64 (scalar_to_vector
1893 (f64 (fpextend (f32 (extractelt VR128:$src, (iPTR 0))))))))),
1894 (Int_VCVTSS2SDrr VR128:$dst, VR128:$src)>;
1896 def : Pat<(v4f32 (X86Movss
1898 (v4f32 (scalar_to_vector (f32 (sint_to_fp GR64:$src)))))),
1899 (Int_VCVTSI2SS64rr VR128:$dst, GR64:$src)>;
1901 def : Pat<(v4f32 (X86Movss
1903 (v4f32 (scalar_to_vector (f32 (sint_to_fp GR32:$src)))))),
1904 (Int_VCVTSI2SSrr VR128:$dst, GR32:$src)>;
1906 def : Pat<(v2f64 (X86Movsd
1908 (v2f64 (scalar_to_vector (f64 (sint_to_fp GR64:$src)))))),
1909 (Int_VCVTSI2SD64rr VR128:$dst, GR64:$src)>;
1911 def : Pat<(v2f64 (X86Movsd
1913 (v2f64 (scalar_to_vector (f64 (sint_to_fp GR32:$src)))))),
1914 (Int_VCVTSI2SDrr VR128:$dst, GR32:$src)>;
1915 } // Predicates = [UseAVX]
1917 let Predicates = [UseSSE2] in {
1918 def : Pat<(v4f32 (X86Movss
1920 (v4f32 (scalar_to_vector
1921 (f32 (fpround (f64 (extractelt VR128:$src, (iPTR 0))))))))),
1922 (Int_CVTSD2SSrr VR128:$dst, VR128:$src)>;
1924 def : Pat<(v2f64 (X86Movsd
1926 (v2f64 (scalar_to_vector
1927 (f64 (fpextend (f32 (extractelt VR128:$src, (iPTR 0))))))))),
1928 (Int_CVTSS2SDrr VR128:$dst, VR128:$src)>;
1930 def : Pat<(v2f64 (X86Movsd
1932 (v2f64 (scalar_to_vector (f64 (sint_to_fp GR64:$src)))))),
1933 (Int_CVTSI2SD64rr VR128:$dst, GR64:$src)>;
1935 def : Pat<(v2f64 (X86Movsd
1937 (v2f64 (scalar_to_vector (f64 (sint_to_fp GR32:$src)))))),
1938 (Int_CVTSI2SDrr VR128:$dst, GR32:$src)>;
1939 } // Predicates = [UseSSE2]
1941 let Predicates = [UseSSE1] in {
1942 def : Pat<(v4f32 (X86Movss
1944 (v4f32 (scalar_to_vector (f32 (sint_to_fp GR64:$src)))))),
1945 (Int_CVTSI2SS64rr VR128:$dst, GR64:$src)>;
1947 def : Pat<(v4f32 (X86Movss
1949 (v4f32 (scalar_to_vector (f32 (sint_to_fp GR32:$src)))))),
1950 (Int_CVTSI2SSrr VR128:$dst, GR32:$src)>;
1951 } // Predicates = [UseSSE1]
1953 // Convert packed single/double fp to doubleword
1954 def VCVTPS2DQrr : VPDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1955 "cvtps2dq\t{$src, $dst|$dst, $src}",
1956 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))],
1957 IIC_SSE_CVT_PS_RR>, VEX, Sched<[WriteCvtF2I]>, VEX_WIG;
1958 def VCVTPS2DQrm : VPDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1959 "cvtps2dq\t{$src, $dst|$dst, $src}",
1961 (int_x86_sse2_cvtps2dq (loadv4f32 addr:$src)))],
1962 IIC_SSE_CVT_PS_RM>, VEX, Sched<[WriteCvtF2ILd]>, VEX_WIG;
1963 def VCVTPS2DQYrr : VPDI<0x5B, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
1964 "cvtps2dq\t{$src, $dst|$dst, $src}",
1966 (int_x86_avx_cvt_ps2dq_256 VR256:$src))],
1967 IIC_SSE_CVT_PS_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>, VEX_WIG;
1968 def VCVTPS2DQYrm : VPDI<0x5B, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
1969 "cvtps2dq\t{$src, $dst|$dst, $src}",
1971 (int_x86_avx_cvt_ps2dq_256 (loadv8f32 addr:$src)))],
1972 IIC_SSE_CVT_PS_RM>, VEX, VEX_L, Sched<[WriteCvtF2ILd]>, VEX_WIG;
1973 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1974 "cvtps2dq\t{$src, $dst|$dst, $src}",
1975 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))],
1976 IIC_SSE_CVT_PS_RR>, Sched<[WriteCvtF2I]>;
1977 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1978 "cvtps2dq\t{$src, $dst|$dst, $src}",
1980 (int_x86_sse2_cvtps2dq (memopv4f32 addr:$src)))],
1981 IIC_SSE_CVT_PS_RM>, Sched<[WriteCvtF2ILd]>;
1984 // Convert Packed Double FP to Packed DW Integers
1985 let Predicates = [HasAVX, NoVLX] in {
1986 // The assembler can recognize rr 256-bit instructions by seeing a ymm
1987 // register, but the same isn't true when using memory operands instead.
1988 // Provide other assembly rr and rm forms to address this explicitly.
1989 def VCVTPD2DQrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1990 "vcvtpd2dq\t{$src, $dst|$dst, $src}",
1992 (v4i32 (X86cvtp2Int (v2f64 VR128:$src))))]>,
1993 VEX, Sched<[WriteCvtF2I]>, VEX_WIG;
1996 def : InstAlias<"vcvtpd2dqx\t{$src, $dst|$dst, $src}",
1997 (VCVTPD2DQrr VR128:$dst, VR128:$src), 0>;
1998 def VCVTPD2DQrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1999 "vcvtpd2dq{x}\t{$src, $dst|$dst, $src}",
2001 (v4i32 (X86cvtp2Int (loadv2f64 addr:$src))))]>, VEX,
2002 Sched<[WriteCvtF2ILd]>, VEX_WIG;
2003 def : InstAlias<"vcvtpd2dqx\t{$src, $dst|$dst, $src}",
2004 (VCVTPD2DQrm VR128:$dst, f128mem:$src), 0>;
2007 def VCVTPD2DQYrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2008 "vcvtpd2dq\t{$src, $dst|$dst, $src}",
2010 (v4i32 (X86cvtp2Int (v4f64 VR256:$src))))]>,
2011 VEX, VEX_L, Sched<[WriteCvtF2I]>, VEX_WIG;
2012 def VCVTPD2DQYrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2013 "vcvtpd2dq{y}\t{$src, $dst|$dst, $src}",
2015 (v4i32 (X86cvtp2Int (loadv4f64 addr:$src))))]>,
2016 VEX, VEX_L, Sched<[WriteCvtF2ILd]>, VEX_WIG;
2017 def : InstAlias<"vcvtpd2dqy\t{$src, $dst|$dst, $src}",
2018 (VCVTPD2DQYrr VR128:$dst, VR256:$src), 0>;
2019 def : InstAlias<"vcvtpd2dqy\t{$src, $dst|$dst, $src}",
2020 (VCVTPD2DQYrm VR128:$dst, f256mem:$src), 0>;
2023 def CVTPD2DQrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2024 "cvtpd2dq\t{$src, $dst|$dst, $src}",
2026 (v4i32 (X86cvtp2Int (memopv2f64 addr:$src))))],
2027 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtF2ILd]>;
2028 def CVTPD2DQrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2029 "cvtpd2dq\t{$src, $dst|$dst, $src}",
2031 (v4i32 (X86cvtp2Int (v2f64 VR128:$src))))],
2032 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2I]>;
2034 // Convert with truncation packed single/double fp to doubleword
2035 // SSE2 packed instructions with XS prefix
2036 let Predicates = [HasAVX, NoVLX] in {
2037 def VCVTTPS2DQrr : VS2SI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2038 "cvttps2dq\t{$src, $dst|$dst, $src}",
2040 (v4i32 (fp_to_sint (v4f32 VR128:$src))))],
2041 IIC_SSE_CVT_PS_RR>, VEX, Sched<[WriteCvtF2I]>, VEX_WIG;
2042 def VCVTTPS2DQrm : VS2SI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2043 "cvttps2dq\t{$src, $dst|$dst, $src}",
2045 (v4i32 (fp_to_sint (loadv4f32 addr:$src))))],
2046 IIC_SSE_CVT_PS_RM>, VEX, Sched<[WriteCvtF2ILd]>, VEX_WIG;
2047 def VCVTTPS2DQYrr : VS2SI<0x5B, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
2048 "cvttps2dq\t{$src, $dst|$dst, $src}",
2050 (v8i32 (fp_to_sint (v8f32 VR256:$src))))],
2051 IIC_SSE_CVT_PS_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>, VEX_WIG;
2052 def VCVTTPS2DQYrm : VS2SI<0x5B, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
2053 "cvttps2dq\t{$src, $dst|$dst, $src}",
2055 (v8i32 (fp_to_sint (loadv8f32 addr:$src))))],
2056 IIC_SSE_CVT_PS_RM>, VEX, VEX_L,
2057 Sched<[WriteCvtF2ILd]>, VEX_WIG;
2060 def CVTTPS2DQrr : S2SI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2061 "cvttps2dq\t{$src, $dst|$dst, $src}",
2063 (v4i32 (fp_to_sint (v4f32 VR128:$src))))],
2064 IIC_SSE_CVT_PS_RR>, Sched<[WriteCvtF2I]>;
2065 def CVTTPS2DQrm : S2SI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2066 "cvttps2dq\t{$src, $dst|$dst, $src}",
2068 (v4i32 (fp_to_sint (memopv4f32 addr:$src))))],
2069 IIC_SSE_CVT_PS_RM>, Sched<[WriteCvtF2ILd]>;
2071 let Predicates = [HasAVX, NoVLX] in
2072 def VCVTTPD2DQrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2073 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2075 (v4i32 (X86cvttp2si (v2f64 VR128:$src))))],
2076 IIC_SSE_CVT_PD_RR>, VEX, Sched<[WriteCvtF2I]>, VEX_WIG;
2078 // The assembler can recognize rr 256-bit instructions by seeing a ymm
2079 // register, but the same isn't true when using memory operands instead.
2080 // Provide other assembly rr and rm forms to address this explicitly.
2083 def : InstAlias<"vcvttpd2dqx\t{$src, $dst|$dst, $src}",
2084 (VCVTTPD2DQrr VR128:$dst, VR128:$src), 0>;
2085 let Predicates = [HasAVX, NoVLX] in
2086 def VCVTTPD2DQrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2087 "cvttpd2dq{x}\t{$src, $dst|$dst, $src}",
2089 (v4i32 (X86cvttp2si (loadv2f64 addr:$src))))],
2090 IIC_SSE_CVT_PD_RM>, VEX, Sched<[WriteCvtF2ILd]>, VEX_WIG;
2091 def : InstAlias<"vcvttpd2dqx\t{$src, $dst|$dst, $src}",
2092 (VCVTTPD2DQrm VR128:$dst, f128mem:$src), 0>;
2095 let Predicates = [HasAVX, NoVLX] in {
2096 def VCVTTPD2DQYrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2097 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2099 (v4i32 (fp_to_sint (v4f64 VR256:$src))))],
2100 IIC_SSE_CVT_PD_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>, VEX_WIG;
2101 def VCVTTPD2DQYrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2102 "cvttpd2dq{y}\t{$src, $dst|$dst, $src}",
2104 (v4i32 (fp_to_sint (loadv4f64 addr:$src))))],
2105 IIC_SSE_CVT_PD_RM>, VEX, VEX_L, Sched<[WriteCvtF2ILd]>, VEX_WIG;
2107 def : InstAlias<"vcvttpd2dqy\t{$src, $dst|$dst, $src}",
2108 (VCVTTPD2DQYrr VR128:$dst, VR256:$src), 0>;
2109 def : InstAlias<"vcvttpd2dqy\t{$src, $dst|$dst, $src}",
2110 (VCVTTPD2DQYrm VR128:$dst, f256mem:$src), 0>;
2112 let Predicates = [HasAVX, NoVLX] in {
2113 let AddedComplexity = 15 in {
2114 def : Pat<(X86vzmovl (v2i64 (bitconvert
2115 (v4i32 (X86cvtp2Int (v2f64 VR128:$src)))))),
2116 (VCVTPD2DQrr VR128:$src)>;
2117 def : Pat<(X86vzmovl (v2i64 (bitconvert
2118 (v4i32 (X86cvttp2si (v2f64 VR128:$src)))))),
2119 (VCVTTPD2DQrr VR128:$src)>;
2121 } // Predicates = [HasAVX]
2123 def CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2124 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2126 (v4i32 (X86cvttp2si (v2f64 VR128:$src))))],
2127 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2I]>;
2128 def CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
2129 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2131 (v4i32 (X86cvttp2si (memopv2f64 addr:$src))))],
2132 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtF2ILd]>;
2134 let Predicates = [UseSSE2] in {
2135 let AddedComplexity = 15 in {
2136 def : Pat<(X86vzmovl (v2i64 (bitconvert
2137 (v4i32 (X86cvtp2Int (v2f64 VR128:$src)))))),
2138 (CVTPD2DQrr VR128:$src)>;
2139 def : Pat<(X86vzmovl (v2i64 (bitconvert
2140 (v4i32 (X86cvttp2si (v2f64 VR128:$src)))))),
2141 (CVTTPD2DQrr VR128:$src)>;
2143 } // Predicates = [UseSSE2]
2145 // Convert packed single to packed double
2146 let Predicates = [HasAVX, NoVLX] in {
2147 // SSE2 instructions without OpSize prefix
2148 def VCVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2149 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2150 [(set VR128:$dst, (v2f64 (X86vfpext (v4f32 VR128:$src))))],
2151 IIC_SSE_CVT_PD_RR>, PS, VEX, Sched<[WriteCvtF2F]>, VEX_WIG;
2152 def VCVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2153 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2154 [(set VR128:$dst, (v2f64 (extloadv2f32 addr:$src)))],
2155 IIC_SSE_CVT_PD_RM>, PS, VEX, Sched<[WriteCvtF2FLd]>, VEX_WIG;
2156 def VCVTPS2PDYrr : I<0x5A, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
2157 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2158 [(set VR256:$dst, (v4f64 (fpextend (v4f32 VR128:$src))))],
2159 IIC_SSE_CVT_PD_RR>, PS, VEX, VEX_L, Sched<[WriteCvtF2F]>, VEX_WIG;
2160 def VCVTPS2PDYrm : I<0x5A, MRMSrcMem, (outs VR256:$dst), (ins f128mem:$src),
2161 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2162 [(set VR256:$dst, (v4f64 (extloadv4f32 addr:$src)))],
2163 IIC_SSE_CVT_PD_RM>, PS, VEX, VEX_L, Sched<[WriteCvtF2FLd]>, VEX_WIG;
2166 let Predicates = [UseSSE2] in {
2167 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2168 "cvtps2pd\t{$src, $dst|$dst, $src}",
2169 [(set VR128:$dst, (v2f64 (X86vfpext (v4f32 VR128:$src))))],
2170 IIC_SSE_CVT_PD_RR>, PS, Sched<[WriteCvtF2F]>;
2171 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2172 "cvtps2pd\t{$src, $dst|$dst, $src}",
2173 [(set VR128:$dst, (v2f64 (extloadv2f32 addr:$src)))],
2174 IIC_SSE_CVT_PD_RM>, PS, Sched<[WriteCvtF2FLd]>;
2177 // Convert Packed DW Integers to Packed Double FP
2178 let Predicates = [HasAVX, NoVLX] in {
2179 let hasSideEffects = 0, mayLoad = 1 in
2180 def VCVTDQ2PDrm : S2SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2181 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2183 (v2f64 (X86VSintToFP (bc_v4i32 (v2i64 (X86vzload addr:$src))))))]>,
2184 VEX, Sched<[WriteCvtI2FLd]>, VEX_WIG;
2185 def VCVTDQ2PDrr : S2SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2186 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2188 (v2f64 (X86VSintToFP (v4i32 VR128:$src))))]>,
2189 VEX, Sched<[WriteCvtI2F]>, VEX_WIG;
2190 def VCVTDQ2PDYrm : S2SI<0xE6, MRMSrcMem, (outs VR256:$dst), (ins i128mem:$src),
2191 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2193 (v4f64 (sint_to_fp (bc_v4i32 (loadv2i64 addr:$src)))))]>,
2194 VEX, VEX_L, Sched<[WriteCvtI2FLd]>, VEX_WIG;
2195 def VCVTDQ2PDYrr : S2SI<0xE6, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
2196 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2198 (v4f64 (sint_to_fp (v4i32 VR128:$src))))]>,
2199 VEX, VEX_L, Sched<[WriteCvtI2F]>, VEX_WIG;
2202 let hasSideEffects = 0, mayLoad = 1 in
2203 def CVTDQ2PDrm : S2SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2204 "cvtdq2pd\t{$src, $dst|$dst, $src}",
2206 (v2f64 (X86VSintToFP (bc_v4i32 (v2i64 (X86vzload addr:$src))))))],
2207 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtI2FLd]>;
2208 def CVTDQ2PDrr : S2SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2209 "cvtdq2pd\t{$src, $dst|$dst, $src}",
2211 (v2f64 (X86VSintToFP (v4i32 VR128:$src))))],
2212 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtI2F]>;
2214 // AVX register conversion intrinsics
2215 let Predicates = [HasAVX, NoVLX] in {
2216 def : Pat<(v2f64 (X86VSintToFP (bc_v4i32 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
2217 (VCVTDQ2PDrm addr:$src)>;
2218 } // Predicates = [HasAVX, NoVLX]
2220 // SSE2 register conversion intrinsics
2221 let Predicates = [UseSSE2] in {
2222 def : Pat<(v2f64 (X86VSintToFP (bc_v4i32 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
2223 (CVTDQ2PDrm addr:$src)>;
2224 } // Predicates = [UseSSE2]
2226 // Convert packed double to packed single
2227 // The assembler can recognize rr 256-bit instructions by seeing a ymm
2228 // register, but the same isn't true when using memory operands instead.
2229 // Provide other assembly rr and rm forms to address this explicitly.
2230 let Predicates = [HasAVX, NoVLX] in
2231 def VCVTPD2PSrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2232 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2233 [(set VR128:$dst, (X86vfpround (v2f64 VR128:$src)))],
2234 IIC_SSE_CVT_PD_RR>, VEX, Sched<[WriteCvtF2F]>, VEX_WIG;
2237 def : InstAlias<"vcvtpd2psx\t{$src, $dst|$dst, $src}",
2238 (VCVTPD2PSrr VR128:$dst, VR128:$src), 0>;
2239 let Predicates = [HasAVX, NoVLX] in
2240 def VCVTPD2PSrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2241 "cvtpd2ps{x}\t{$src, $dst|$dst, $src}",
2242 [(set VR128:$dst, (X86vfpround (loadv2f64 addr:$src)))],
2243 IIC_SSE_CVT_PD_RM>, VEX, Sched<[WriteCvtF2FLd]>, VEX_WIG;
2244 def : InstAlias<"vcvtpd2psx\t{$src, $dst|$dst, $src}",
2245 (VCVTPD2PSrm VR128:$dst, f128mem:$src), 0>;
2248 let Predicates = [HasAVX, NoVLX] in {
2249 def VCVTPD2PSYrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2250 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2251 [(set VR128:$dst, (fpround VR256:$src))],
2252 IIC_SSE_CVT_PD_RR>, VEX, VEX_L, Sched<[WriteCvtF2F]>, VEX_WIG;
2253 def VCVTPD2PSYrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2254 "cvtpd2ps{y}\t{$src, $dst|$dst, $src}",
2255 [(set VR128:$dst, (fpround (loadv4f64 addr:$src)))],
2256 IIC_SSE_CVT_PD_RM>, VEX, VEX_L, Sched<[WriteCvtF2FLd]>, VEX_WIG;
2258 def : InstAlias<"vcvtpd2psy\t{$src, $dst|$dst, $src}",
2259 (VCVTPD2PSYrr VR128:$dst, VR256:$src), 0>;
2260 def : InstAlias<"vcvtpd2psy\t{$src, $dst|$dst, $src}",
2261 (VCVTPD2PSYrm VR128:$dst, f256mem:$src), 0>;
2263 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2264 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2265 [(set VR128:$dst, (X86vfpround (v2f64 VR128:$src)))],
2266 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2F]>;
2267 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2268 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2269 [(set VR128:$dst, (X86vfpround (memopv2f64 addr:$src)))],
2270 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtF2FLd]>;
2272 // AVX 256-bit register conversion intrinsics
2273 // FIXME: Migrate SSE conversion intrinsics matching to use patterns as below
2274 // whenever possible to avoid declaring two versions of each one.
2276 let Predicates = [HasAVX, NoVLX] in {
2277 // Match fpround and fpextend for 128/256-bit conversions
2278 let AddedComplexity = 15 in
2279 def : Pat<(X86vzmovl (v2f64 (bitconvert
2280 (v4f32 (X86vfpround (v2f64 VR128:$src)))))),
2281 (VCVTPD2PSrr VR128:$src)>;
2284 let Predicates = [UseSSE2] in {
2285 // Match fpround and fpextend for 128 conversions
2286 let AddedComplexity = 15 in
2287 def : Pat<(X86vzmovl (v2f64 (bitconvert
2288 (v4f32 (X86vfpround (v2f64 VR128:$src)))))),
2289 (CVTPD2PSrr VR128:$src)>;
2292 //===----------------------------------------------------------------------===//
2293 // SSE 1 & 2 - Compare Instructions
2294 //===----------------------------------------------------------------------===//
2296 // sse12_cmp_scalar - sse 1 & 2 compare scalar instructions
2297 multiclass sse12_cmp_scalar<RegisterClass RC, X86MemOperand x86memop,
2298 Operand CC, SDNode OpNode, ValueType VT,
2299 PatFrag ld_frag, string asm, string asm_alt,
2300 OpndItins itins, ImmLeaf immLeaf> {
2301 let isCommutable = 1 in
2302 def rr : SIi8<0xC2, MRMSrcReg,
2303 (outs RC:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
2304 [(set RC:$dst, (OpNode (VT RC:$src1), RC:$src2, immLeaf:$cc))],
2305 itins.rr>, Sched<[itins.Sched]>;
2306 def rm : SIi8<0xC2, MRMSrcMem,
2307 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
2308 [(set RC:$dst, (OpNode (VT RC:$src1),
2309 (ld_frag addr:$src2), immLeaf:$cc))],
2311 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2313 // Accept explicit immediate argument form instead of comparison code.
2314 let isAsmParserOnly = 1, hasSideEffects = 0 in {
2315 def rr_alt : SIi8<0xC2, MRMSrcReg, (outs RC:$dst),
2316 (ins RC:$src1, RC:$src2, u8imm:$cc), asm_alt, [],
2317 IIC_SSE_ALU_F32S_RR>, Sched<[itins.Sched]>;
2319 def rm_alt : SIi8<0xC2, MRMSrcMem, (outs RC:$dst),
2320 (ins RC:$src1, x86memop:$src2, u8imm:$cc), asm_alt, [],
2321 IIC_SSE_ALU_F32S_RM>,
2322 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2326 let ExeDomain = SSEPackedSingle in
2327 defm VCMPSS : sse12_cmp_scalar<FR32, f32mem, AVXCC, X86cmps, f32, loadf32,
2328 "cmp${cc}ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2329 "cmpss\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2330 SSE_ALU_F32S, i8immZExt5>, XS, VEX_4V, VEX_LIG, VEX_WIG;
2331 let ExeDomain = SSEPackedDouble in
2332 defm VCMPSD : sse12_cmp_scalar<FR64, f64mem, AVXCC, X86cmps, f64, loadf64,
2333 "cmp${cc}sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2334 "cmpsd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2335 SSE_ALU_F32S, i8immZExt5>, // same latency as 32 bit compare
2336 XD, VEX_4V, VEX_LIG, VEX_WIG;
2338 let Constraints = "$src1 = $dst" in {
2339 let ExeDomain = SSEPackedSingle in
2340 defm CMPSS : sse12_cmp_scalar<FR32, f32mem, SSECC, X86cmps, f32, loadf32,
2341 "cmp${cc}ss\t{$src2, $dst|$dst, $src2}",
2342 "cmpss\t{$cc, $src2, $dst|$dst, $src2, $cc}", SSE_ALU_F32S,
2344 let ExeDomain = SSEPackedDouble in
2345 defm CMPSD : sse12_cmp_scalar<FR64, f64mem, SSECC, X86cmps, f64, loadf64,
2346 "cmp${cc}sd\t{$src2, $dst|$dst, $src2}",
2347 "cmpsd\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2348 SSE_ALU_F64S, i8immZExt3>, XD;
2351 multiclass sse12_cmp_scalar_int<Operand memop, Operand CC,
2352 Intrinsic Int, string asm, OpndItins itins,
2353 ImmLeaf immLeaf, ComplexPattern mem_cpat> {
2354 def rr : SIi8<0xC2, MRMSrcReg, (outs VR128:$dst),
2355 (ins VR128:$src1, VR128:$src, CC:$cc), asm,
2356 [(set VR128:$dst, (Int VR128:$src1,
2357 VR128:$src, immLeaf:$cc))],
2359 Sched<[itins.Sched]>;
2361 def rm : SIi8<0xC2, MRMSrcMem, (outs VR128:$dst),
2362 (ins VR128:$src1, memop:$src, CC:$cc), asm,
2363 [(set VR128:$dst, (Int VR128:$src1,
2364 mem_cpat:$src, immLeaf:$cc))],
2366 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2369 let isCodeGenOnly = 1 in {
2370 // Aliases to match intrinsics which expect XMM operand(s).
2371 let ExeDomain = SSEPackedSingle in
2372 defm Int_VCMPSS : sse12_cmp_scalar_int<ssmem, AVXCC, int_x86_sse_cmp_ss,
2373 "cmp${cc}ss\t{$src, $src1, $dst|$dst, $src1, $src}",
2374 SSE_ALU_F32S, i8immZExt5, sse_load_f32>,
2376 let ExeDomain = SSEPackedDouble in
2377 defm Int_VCMPSD : sse12_cmp_scalar_int<sdmem, AVXCC, int_x86_sse2_cmp_sd,
2378 "cmp${cc}sd\t{$src, $src1, $dst|$dst, $src1, $src}",
2379 SSE_ALU_F32S, i8immZExt5, sse_load_f64>, // same latency as f32
2381 let Constraints = "$src1 = $dst" in {
2382 let ExeDomain = SSEPackedSingle in
2383 defm Int_CMPSS : sse12_cmp_scalar_int<ssmem, SSECC, int_x86_sse_cmp_ss,
2384 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
2385 SSE_ALU_F32S, i8immZExt3, sse_load_f32>, XS;
2386 let ExeDomain = SSEPackedDouble in
2387 defm Int_CMPSD : sse12_cmp_scalar_int<sdmem, SSECC, int_x86_sse2_cmp_sd,
2388 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
2389 SSE_ALU_F64S, i8immZExt3, sse_load_f64>,
2395 // sse12_ord_cmp - Unordered/Ordered scalar fp compare and set EFLAGS
2396 multiclass sse12_ord_cmp<bits<8> opc, RegisterClass RC, SDNode OpNode,
2397 ValueType vt, X86MemOperand x86memop,
2398 PatFrag ld_frag, string OpcodeStr> {
2399 def rr: SI<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
2400 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2401 [(set EFLAGS, (OpNode (vt RC:$src1), RC:$src2))],
2405 def rm: SI<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
2406 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2407 [(set EFLAGS, (OpNode (vt RC:$src1),
2408 (ld_frag addr:$src2)))],
2410 Sched<[WriteFAddLd, ReadAfterLd]>;
2413 // sse12_ord_cmp_int - Intrinsic version of sse12_ord_cmp
2414 multiclass sse12_ord_cmp_int<bits<8> opc, RegisterClass RC, SDNode OpNode,
2415 ValueType vt, Operand memop,
2416 ComplexPattern mem_cpat, string OpcodeStr> {
2417 def rr: SI<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
2418 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2419 [(set EFLAGS, (OpNode (vt RC:$src1), RC:$src2))],
2423 def rm: SI<opc, MRMSrcMem, (outs), (ins RC:$src1, memop:$src2),
2424 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2425 [(set EFLAGS, (OpNode (vt RC:$src1),
2428 Sched<[WriteFAddLd, ReadAfterLd]>;
2431 let Defs = [EFLAGS] in {
2432 defm VUCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
2433 "ucomiss">, PS, VEX, VEX_LIG, VEX_WIG;
2434 defm VUCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
2435 "ucomisd">, PD, VEX, VEX_LIG, VEX_WIG;
2436 let Pattern = []<dag> in {
2437 defm VCOMISS : sse12_ord_cmp<0x2F, FR32, undef, f32, f32mem, loadf32,
2438 "comiss">, PS, VEX, VEX_LIG, VEX_WIG;
2439 defm VCOMISD : sse12_ord_cmp<0x2F, FR64, undef, f64, f64mem, loadf64,
2440 "comisd">, PD, VEX, VEX_LIG, VEX_WIG;
2443 let isCodeGenOnly = 1 in {
2444 defm Int_VUCOMISS : sse12_ord_cmp_int<0x2E, VR128, X86ucomi, v4f32, ssmem,
2445 sse_load_f32, "ucomiss">, PS, VEX, VEX_WIG;
2446 defm Int_VUCOMISD : sse12_ord_cmp_int<0x2E, VR128, X86ucomi, v2f64, sdmem,
2447 sse_load_f64, "ucomisd">, PD, VEX, VEX_WIG;
2449 defm Int_VCOMISS : sse12_ord_cmp_int<0x2F, VR128, X86comi, v4f32, ssmem,
2450 sse_load_f32, "comiss">, PS, VEX, VEX_WIG;
2451 defm Int_VCOMISD : sse12_ord_cmp_int<0x2F, VR128, X86comi, v2f64, sdmem,
2452 sse_load_f64, "comisd">, PD, VEX, VEX_WIG;
2454 defm UCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
2456 defm UCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
2459 let Pattern = []<dag> in {
2460 defm COMISS : sse12_ord_cmp<0x2F, FR32, undef, f32, f32mem, loadf32,
2462 defm COMISD : sse12_ord_cmp<0x2F, FR64, undef, f64, f64mem, loadf64,
2466 let isCodeGenOnly = 1 in {
2467 defm Int_UCOMISS : sse12_ord_cmp_int<0x2E, VR128, X86ucomi, v4f32, ssmem,
2468 sse_load_f32, "ucomiss">, PS;
2469 defm Int_UCOMISD : sse12_ord_cmp_int<0x2E, VR128, X86ucomi, v2f64, sdmem,
2470 sse_load_f64, "ucomisd">, PD;
2472 defm Int_COMISS : sse12_ord_cmp_int<0x2F, VR128, X86comi, v4f32, ssmem,
2473 sse_load_f32, "comiss">, PS;
2474 defm Int_COMISD : sse12_ord_cmp_int<0x2F, VR128, X86comi, v2f64, sdmem,
2475 sse_load_f64, "comisd">, PD;
2477 } // Defs = [EFLAGS]
2479 // sse12_cmp_packed - sse 1 & 2 compare packed instructions
2480 multiclass sse12_cmp_packed<RegisterClass RC, X86MemOperand x86memop,
2481 Operand CC, ValueType VT, string asm,
2482 string asm_alt, Domain d, ImmLeaf immLeaf,
2483 PatFrag ld_frag, OpndItins itins = SSE_ALU_F32P> {
2484 let isCommutable = 1 in
2485 def rri : PIi8<0xC2, MRMSrcReg,
2486 (outs RC:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
2487 [(set RC:$dst, (VT (X86cmpp RC:$src1, RC:$src2, immLeaf:$cc)))],
2490 def rmi : PIi8<0xC2, MRMSrcMem,
2491 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
2493 (VT (X86cmpp RC:$src1, (ld_frag addr:$src2), immLeaf:$cc)))],
2495 Sched<[WriteFAddLd, ReadAfterLd]>;
2497 // Accept explicit immediate argument form instead of comparison code.
2498 let isAsmParserOnly = 1, hasSideEffects = 0 in {
2499 def rri_alt : PIi8<0xC2, MRMSrcReg,
2500 (outs RC:$dst), (ins RC:$src1, RC:$src2, u8imm:$cc),
2501 asm_alt, [], itins.rr, d>, Sched<[WriteFAdd]>;
2503 def rmi_alt : PIi8<0xC2, MRMSrcMem,
2504 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, u8imm:$cc),
2505 asm_alt, [], itins.rm, d>,
2506 Sched<[WriteFAddLd, ReadAfterLd]>;
2510 defm VCMPPS : sse12_cmp_packed<VR128, f128mem, AVXCC, v4f32,
2511 "cmp${cc}ps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2512 "cmpps\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2513 SSEPackedSingle, i8immZExt5, loadv4f32>, PS, VEX_4V, VEX_WIG;
2514 defm VCMPPD : sse12_cmp_packed<VR128, f128mem, AVXCC, v2f64,
2515 "cmp${cc}pd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2516 "cmppd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2517 SSEPackedDouble, i8immZExt5, loadv2f64>, PD, VEX_4V, VEX_WIG;
2518 defm VCMPPSY : sse12_cmp_packed<VR256, f256mem, AVXCC, v8f32,
2519 "cmp${cc}ps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2520 "cmpps\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2521 SSEPackedSingle, i8immZExt5, loadv8f32>, PS, VEX_4V, VEX_L;
2522 defm VCMPPDY : sse12_cmp_packed<VR256, f256mem, AVXCC, v4f64,
2523 "cmp${cc}pd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2524 "cmppd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2525 SSEPackedDouble, i8immZExt5, loadv4f64>, PD, VEX_4V, VEX_L;
2526 let Constraints = "$src1 = $dst" in {
2527 defm CMPPS : sse12_cmp_packed<VR128, f128mem, SSECC, v4f32,
2528 "cmp${cc}ps\t{$src2, $dst|$dst, $src2}",
2529 "cmpps\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2530 SSEPackedSingle, i8immZExt5, memopv4f32, SSE_ALU_F32P>, PS;
2531 defm CMPPD : sse12_cmp_packed<VR128, f128mem, SSECC, v2f64,
2532 "cmp${cc}pd\t{$src2, $dst|$dst, $src2}",
2533 "cmppd\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2534 SSEPackedDouble, i8immZExt5, memopv2f64, SSE_ALU_F64P>, PD;
2537 //===----------------------------------------------------------------------===//
2538 // SSE 1 & 2 - Shuffle Instructions
2539 //===----------------------------------------------------------------------===//
2541 /// sse12_shuffle - sse 1 & 2 fp shuffle instructions
2542 multiclass sse12_shuffle<RegisterClass RC, X86MemOperand x86memop,
2543 ValueType vt, string asm, PatFrag mem_frag,
2545 def rmi : PIi8<0xC6, MRMSrcMem, (outs RC:$dst),
2546 (ins RC:$src1, x86memop:$src2, u8imm:$src3), asm,
2547 [(set RC:$dst, (vt (X86Shufp RC:$src1, (mem_frag addr:$src2),
2548 (i8 imm:$src3))))], IIC_SSE_SHUFP, d>,
2549 Sched<[WriteFShuffleLd, ReadAfterLd]>;
2550 def rri : PIi8<0xC6, MRMSrcReg, (outs RC:$dst),
2551 (ins RC:$src1, RC:$src2, u8imm:$src3), asm,
2552 [(set RC:$dst, (vt (X86Shufp RC:$src1, RC:$src2,
2553 (i8 imm:$src3))))], IIC_SSE_SHUFP, d>,
2554 Sched<[WriteFShuffle]>;
2557 let Predicates = [HasAVX, NoVLX] in {
2558 defm VSHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
2559 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2560 loadv4f32, SSEPackedSingle>, PS, VEX_4V, VEX_WIG;
2561 defm VSHUFPSY : sse12_shuffle<VR256, f256mem, v8f32,
2562 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2563 loadv8f32, SSEPackedSingle>, PS, VEX_4V, VEX_L, VEX_WIG;
2564 defm VSHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
2565 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2566 loadv2f64, SSEPackedDouble>, PD, VEX_4V, VEX_WIG;
2567 defm VSHUFPDY : sse12_shuffle<VR256, f256mem, v4f64,
2568 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2569 loadv4f64, SSEPackedDouble>, PD, VEX_4V, VEX_L, VEX_WIG;
2571 let Constraints = "$src1 = $dst" in {
2572 defm SHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
2573 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2574 memopv4f32, SSEPackedSingle>, PS;
2575 defm SHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
2576 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2577 memopv2f64, SSEPackedDouble>, PD;
2580 let Predicates = [HasAVX, NoVLX] in {
2581 def : Pat<(v4i32 (X86Shufp VR128:$src1,
2582 (bc_v4i32 (loadv2i64 addr:$src2)), (i8 imm:$imm))),
2583 (VSHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>;
2584 def : Pat<(v4i32 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2585 (VSHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>;
2587 def : Pat<(v2i64 (X86Shufp VR128:$src1,
2588 (loadv2i64 addr:$src2), (i8 imm:$imm))),
2589 (VSHUFPDrmi VR128:$src1, addr:$src2, imm:$imm)>;
2590 def : Pat<(v2i64 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2591 (VSHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>;
2594 def : Pat<(v8i32 (X86Shufp VR256:$src1, VR256:$src2, (i8 imm:$imm))),
2595 (VSHUFPSYrri VR256:$src1, VR256:$src2, imm:$imm)>;
2596 def : Pat<(v8i32 (X86Shufp VR256:$src1,
2597 (bc_v8i32 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
2598 (VSHUFPSYrmi VR256:$src1, addr:$src2, imm:$imm)>;
2600 def : Pat<(v4i64 (X86Shufp VR256:$src1, VR256:$src2, (i8 imm:$imm))),
2601 (VSHUFPDYrri VR256:$src1, VR256:$src2, imm:$imm)>;
2602 def : Pat<(v4i64 (X86Shufp VR256:$src1,
2603 (loadv4i64 addr:$src2), (i8 imm:$imm))),
2604 (VSHUFPDYrmi VR256:$src1, addr:$src2, imm:$imm)>;
2607 let Predicates = [UseSSE1] in {
2608 def : Pat<(v4i32 (X86Shufp VR128:$src1,
2609 (bc_v4i32 (memopv2i64 addr:$src2)), (i8 imm:$imm))),
2610 (SHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>;
2611 def : Pat<(v4i32 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2612 (SHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>;
2615 let Predicates = [UseSSE2] in {
2616 // Generic SHUFPD patterns
2617 def : Pat<(v2i64 (X86Shufp VR128:$src1,
2618 (memopv2i64 addr:$src2), (i8 imm:$imm))),
2619 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$imm)>;
2620 def : Pat<(v2i64 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2621 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>;
2624 //===----------------------------------------------------------------------===//
2625 // SSE 1 & 2 - Unpack FP Instructions
2626 //===----------------------------------------------------------------------===//
2628 /// sse12_unpack_interleave - sse 1 & 2 fp unpack and interleave
2629 multiclass sse12_unpack_interleave<bits<8> opc, SDNode OpNode, ValueType vt,
2630 PatFrag mem_frag, RegisterClass RC,
2631 X86MemOperand x86memop, string asm,
2632 Domain d, bit IsCommutable = 0> {
2633 let isCommutable = IsCommutable in
2634 def rr : PI<opc, MRMSrcReg,
2635 (outs RC:$dst), (ins RC:$src1, RC:$src2),
2637 (vt (OpNode RC:$src1, RC:$src2)))],
2638 IIC_SSE_UNPCK, d>, Sched<[WriteFShuffle]>;
2639 def rm : PI<opc, MRMSrcMem,
2640 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
2642 (vt (OpNode RC:$src1,
2643 (mem_frag addr:$src2))))],
2645 Sched<[WriteFShuffleLd, ReadAfterLd]>;
2648 let Predicates = [HasAVX, NoVLX] in {
2649 defm VUNPCKHPS: sse12_unpack_interleave<0x15, X86Unpckh, v4f32, loadv4f32,
2650 VR128, f128mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2651 SSEPackedSingle>, PS, VEX_4V, VEX_WIG;
2652 defm VUNPCKHPD: sse12_unpack_interleave<0x15, X86Unpckh, v2f64, loadv2f64,
2653 VR128, f128mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2654 SSEPackedDouble>, PD, VEX_4V, VEX_WIG;
2655 defm VUNPCKLPS: sse12_unpack_interleave<0x14, X86Unpckl, v4f32, loadv4f32,
2656 VR128, f128mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2657 SSEPackedSingle>, PS, VEX_4V, VEX_WIG;
2658 defm VUNPCKLPD: sse12_unpack_interleave<0x14, X86Unpckl, v2f64, loadv2f64,
2659 VR128, f128mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2660 SSEPackedDouble>, PD, VEX_4V, VEX_WIG;
2662 defm VUNPCKHPSY: sse12_unpack_interleave<0x15, X86Unpckh, v8f32, loadv8f32,
2663 VR256, f256mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2664 SSEPackedSingle>, PS, VEX_4V, VEX_L, VEX_WIG;
2665 defm VUNPCKHPDY: sse12_unpack_interleave<0x15, X86Unpckh, v4f64, loadv4f64,
2666 VR256, f256mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2667 SSEPackedDouble>, PD, VEX_4V, VEX_L, VEX_WIG;
2668 defm VUNPCKLPSY: sse12_unpack_interleave<0x14, X86Unpckl, v8f32, loadv8f32,
2669 VR256, f256mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2670 SSEPackedSingle>, PS, VEX_4V, VEX_L, VEX_WIG;
2671 defm VUNPCKLPDY: sse12_unpack_interleave<0x14, X86Unpckl, v4f64, loadv4f64,
2672 VR256, f256mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2673 SSEPackedDouble>, PD, VEX_4V, VEX_L, VEX_WIG;
2674 }// Predicates = [HasAVX, NoVLX]
2675 let Constraints = "$src1 = $dst" in {
2676 defm UNPCKHPS: sse12_unpack_interleave<0x15, X86Unpckh, v4f32, memopv4f32,
2677 VR128, f128mem, "unpckhps\t{$src2, $dst|$dst, $src2}",
2678 SSEPackedSingle>, PS;
2679 defm UNPCKHPD: sse12_unpack_interleave<0x15, X86Unpckh, v2f64, memopv2f64,
2680 VR128, f128mem, "unpckhpd\t{$src2, $dst|$dst, $src2}",
2681 SSEPackedDouble, 1>, PD;
2682 defm UNPCKLPS: sse12_unpack_interleave<0x14, X86Unpckl, v4f32, memopv4f32,
2683 VR128, f128mem, "unpcklps\t{$src2, $dst|$dst, $src2}",
2684 SSEPackedSingle>, PS;
2685 defm UNPCKLPD: sse12_unpack_interleave<0x14, X86Unpckl, v2f64, memopv2f64,
2686 VR128, f128mem, "unpcklpd\t{$src2, $dst|$dst, $src2}",
2687 SSEPackedDouble>, PD;
2688 } // Constraints = "$src1 = $dst"
2690 let Predicates = [HasAVX1Only] in {
2691 def : Pat<(v8i32 (X86Unpckl VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
2692 (VUNPCKLPSYrm VR256:$src1, addr:$src2)>;
2693 def : Pat<(v8i32 (X86Unpckl VR256:$src1, VR256:$src2)),
2694 (VUNPCKLPSYrr VR256:$src1, VR256:$src2)>;
2695 def : Pat<(v8i32 (X86Unpckh VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
2696 (VUNPCKHPSYrm VR256:$src1, addr:$src2)>;
2697 def : Pat<(v8i32 (X86Unpckh VR256:$src1, VR256:$src2)),
2698 (VUNPCKHPSYrr VR256:$src1, VR256:$src2)>;
2700 def : Pat<(v4i64 (X86Unpckl VR256:$src1, (loadv4i64 addr:$src2))),
2701 (VUNPCKLPDYrm VR256:$src1, addr:$src2)>;
2702 def : Pat<(v4i64 (X86Unpckl VR256:$src1, VR256:$src2)),
2703 (VUNPCKLPDYrr VR256:$src1, VR256:$src2)>;
2704 def : Pat<(v4i64 (X86Unpckh VR256:$src1, (loadv4i64 addr:$src2))),
2705 (VUNPCKHPDYrm VR256:$src1, addr:$src2)>;
2706 def : Pat<(v4i64 (X86Unpckh VR256:$src1, VR256:$src2)),
2707 (VUNPCKHPDYrr VR256:$src1, VR256:$src2)>;
2710 //===----------------------------------------------------------------------===//
2711 // SSE 1 & 2 - Extract Floating-Point Sign mask
2712 //===----------------------------------------------------------------------===//
2714 /// sse12_extr_sign_mask - sse 1 & 2 unpack and interleave
2715 multiclass sse12_extr_sign_mask<RegisterClass RC, ValueType vt,
2716 string asm, Domain d> {
2717 def rr : PI<0x50, MRMSrcReg, (outs GR32orGR64:$dst), (ins RC:$src),
2718 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
2719 [(set GR32orGR64:$dst, (X86movmsk (vt RC:$src)))], IIC_SSE_MOVMSK, d>,
2720 Sched<[WriteVecLogic]>;
2723 let Predicates = [HasAVX] in {
2724 defm VMOVMSKPS : sse12_extr_sign_mask<VR128, v4f32, "movmskps",
2725 SSEPackedSingle>, PS, VEX, VEX_WIG;
2726 defm VMOVMSKPD : sse12_extr_sign_mask<VR128, v2f64, "movmskpd",
2727 SSEPackedDouble>, PD, VEX, VEX_WIG;
2728 defm VMOVMSKPSY : sse12_extr_sign_mask<VR256, v8f32, "movmskps",
2729 SSEPackedSingle>, PS, VEX, VEX_L, VEX_WIG;
2730 defm VMOVMSKPDY : sse12_extr_sign_mask<VR256, v4f64, "movmskpd",
2731 SSEPackedDouble>, PD, VEX, VEX_L, VEX_WIG;
2734 defm MOVMSKPS : sse12_extr_sign_mask<VR128, v4f32, "movmskps",
2735 SSEPackedSingle>, PS;
2736 defm MOVMSKPD : sse12_extr_sign_mask<VR128, v2f64, "movmskpd",
2737 SSEPackedDouble>, PD;
2739 //===---------------------------------------------------------------------===//
2740 // SSE2 - Packed Integer Logical Instructions
2741 //===---------------------------------------------------------------------===//
2743 let ExeDomain = SSEPackedInt in { // SSE integer instructions
2745 /// PDI_binop_rm - Simple SSE2 binary operator.
2746 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
2747 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
2748 X86MemOperand x86memop, OpndItins itins,
2749 bit IsCommutable, bit Is2Addr> {
2750 let isCommutable = IsCommutable in
2751 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
2752 (ins RC:$src1, RC:$src2),
2754 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2755 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2756 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))], itins.rr>,
2757 Sched<[itins.Sched]>;
2758 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
2759 (ins RC:$src1, x86memop:$src2),
2761 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2762 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2763 [(set RC:$dst, (OpVT (OpNode RC:$src1,
2764 (bitconvert (memop_frag addr:$src2)))))],
2766 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2768 } // ExeDomain = SSEPackedInt
2770 multiclass PDI_binop_all<bits<8> opc, string OpcodeStr, SDNode Opcode,
2771 ValueType OpVT128, ValueType OpVT256,
2772 OpndItins itins, bit IsCommutable = 0, Predicate prd> {
2773 let Predicates = [HasAVX, prd] in
2774 defm V#NAME : PDI_binop_rm<opc, !strconcat("v", OpcodeStr), Opcode, OpVT128,
2775 VR128, loadv2i64, i128mem, itins, IsCommutable, 0>, VEX_4V, VEX_WIG;
2777 let Constraints = "$src1 = $dst" in
2778 defm NAME : PDI_binop_rm<opc, OpcodeStr, Opcode, OpVT128, VR128,
2779 memopv2i64, i128mem, itins, IsCommutable, 1>;
2781 let Predicates = [HasAVX2, prd] in
2782 defm V#NAME#Y : PDI_binop_rm<opc, !strconcat("v", OpcodeStr), Opcode,
2783 OpVT256, VR256, loadv4i64, i256mem, itins,
2784 IsCommutable, 0>, VEX_4V, VEX_L, VEX_WIG;
2787 // These are ordered here for pattern ordering requirements with the fp versions
2789 defm PAND : PDI_binop_all<0xDB, "pand", and, v2i64, v4i64,
2790 SSE_VEC_BIT_ITINS_P, 1, NoVLX>;
2791 defm POR : PDI_binop_all<0xEB, "por", or, v2i64, v4i64,
2792 SSE_VEC_BIT_ITINS_P, 1, NoVLX>;
2793 defm PXOR : PDI_binop_all<0xEF, "pxor", xor, v2i64, v4i64,
2794 SSE_VEC_BIT_ITINS_P, 1, NoVLX>;
2795 defm PANDN : PDI_binop_all<0xDF, "pandn", X86andnp, v2i64, v4i64,
2796 SSE_VEC_BIT_ITINS_P, 0, NoVLX>;
2798 //===----------------------------------------------------------------------===//
2799 // SSE 1 & 2 - Logical Instructions
2800 //===----------------------------------------------------------------------===//
2802 /// sse12_fp_packed_logical - SSE 1 & 2 packed FP logical ops
2804 multiclass sse12_fp_packed_logical<bits<8> opc, string OpcodeStr,
2806 let Predicates = [HasAVX, NoVLX] in {
2807 defm V#NAME#PSY : sse12_fp_packed_logical_rm<opc, VR256, SSEPackedSingle,
2808 !strconcat(OpcodeStr, "ps"), f256mem,
2809 [(set VR256:$dst, (OpNode (bc_v4i64 (v8f32 VR256:$src1)),
2810 (bc_v4i64 (v8f32 VR256:$src2))))],
2811 [(set VR256:$dst, (OpNode (bc_v4i64 (v8f32 VR256:$src1)),
2812 (loadv4i64 addr:$src2)))], 0>, PS, VEX_4V, VEX_L, VEX_WIG;
2814 defm V#NAME#PDY : sse12_fp_packed_logical_rm<opc, VR256, SSEPackedDouble,
2815 !strconcat(OpcodeStr, "pd"), f256mem,
2816 [(set VR256:$dst, (OpNode (bc_v4i64 (v4f64 VR256:$src1)),
2817 (bc_v4i64 (v4f64 VR256:$src2))))],
2818 [(set VR256:$dst, (OpNode (bc_v4i64 (v4f64 VR256:$src1)),
2819 (loadv4i64 addr:$src2)))], 0>,
2820 PD, VEX_4V, VEX_L, VEX_WIG;
2822 defm V#NAME#PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
2823 !strconcat(OpcodeStr, "ps"), f128mem,
2824 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2825 (bc_v2i64 (v4f32 VR128:$src2))))],
2826 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2827 (loadv2i64 addr:$src2)))], 0>, PS, VEX_4V, VEX_WIG;
2829 defm V#NAME#PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
2830 !strconcat(OpcodeStr, "pd"), f128mem,
2831 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2832 (bc_v2i64 (v2f64 VR128:$src2))))],
2833 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2834 (loadv2i64 addr:$src2)))], 0>,
2835 PD, VEX_4V, VEX_WIG;
2838 let Constraints = "$src1 = $dst" in {
2839 defm PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
2840 !strconcat(OpcodeStr, "ps"), f128mem,
2841 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2842 (bc_v2i64 (v4f32 VR128:$src2))))],
2843 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2844 (memopv2i64 addr:$src2)))]>, PS;
2846 defm PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
2847 !strconcat(OpcodeStr, "pd"), f128mem,
2848 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2849 (bc_v2i64 (v2f64 VR128:$src2))))],
2850 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2851 (memopv2i64 addr:$src2)))]>, PD;
2855 defm AND : sse12_fp_packed_logical<0x54, "and", and>;
2856 defm OR : sse12_fp_packed_logical<0x56, "or", or>;
2857 defm XOR : sse12_fp_packed_logical<0x57, "xor", xor>;
2858 let isCommutable = 0 in
2859 defm ANDN : sse12_fp_packed_logical<0x55, "andn", X86andnp>;
2861 // If only AVX1 is supported, we need to handle integer operations with
2862 // floating point instructions since the integer versions aren't available.
2863 let Predicates = [HasAVX1Only] in {
2864 def : Pat<(v4i64 (and VR256:$src1, VR256:$src2)),
2865 (VANDPSYrr VR256:$src1, VR256:$src2)>;
2866 def : Pat<(v4i64 (or VR256:$src1, VR256:$src2)),
2867 (VORPSYrr VR256:$src1, VR256:$src2)>;
2868 def : Pat<(v4i64 (xor VR256:$src1, VR256:$src2)),
2869 (VXORPSYrr VR256:$src1, VR256:$src2)>;
2870 def : Pat<(v4i64 (X86andnp VR256:$src1, VR256:$src2)),
2871 (VANDNPSYrr VR256:$src1, VR256:$src2)>;
2873 def : Pat<(and VR256:$src1, (loadv4i64 addr:$src2)),
2874 (VANDPSYrm VR256:$src1, addr:$src2)>;
2875 def : Pat<(or VR256:$src1, (loadv4i64 addr:$src2)),
2876 (VORPSYrm VR256:$src1, addr:$src2)>;
2877 def : Pat<(xor VR256:$src1, (loadv4i64 addr:$src2)),
2878 (VXORPSYrm VR256:$src1, addr:$src2)>;
2879 def : Pat<(X86andnp VR256:$src1, (loadv4i64 addr:$src2)),
2880 (VANDNPSYrm VR256:$src1, addr:$src2)>;
2883 let Predicates = [HasAVX, NoVLX_Or_NoDQI] in {
2884 // Use packed logical operations for scalar ops.
2885 def : Pat<(f64 (X86fand FR64:$src1, FR64:$src2)),
2886 (COPY_TO_REGCLASS (VANDPDrr
2887 (COPY_TO_REGCLASS FR64:$src1, VR128),
2888 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2889 def : Pat<(f64 (X86for FR64:$src1, FR64:$src2)),
2890 (COPY_TO_REGCLASS (VORPDrr
2891 (COPY_TO_REGCLASS FR64:$src1, VR128),
2892 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2893 def : Pat<(f64 (X86fxor FR64:$src1, FR64:$src2)),
2894 (COPY_TO_REGCLASS (VXORPDrr
2895 (COPY_TO_REGCLASS FR64:$src1, VR128),
2896 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2897 def : Pat<(f64 (X86fandn FR64:$src1, FR64:$src2)),
2898 (COPY_TO_REGCLASS (VANDNPDrr
2899 (COPY_TO_REGCLASS FR64:$src1, VR128),
2900 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2902 def : Pat<(f32 (X86fand FR32:$src1, FR32:$src2)),
2903 (COPY_TO_REGCLASS (VANDPSrr
2904 (COPY_TO_REGCLASS FR32:$src1, VR128),
2905 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2906 def : Pat<(f32 (X86for FR32:$src1, FR32:$src2)),
2907 (COPY_TO_REGCLASS (VORPSrr
2908 (COPY_TO_REGCLASS FR32:$src1, VR128),
2909 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2910 def : Pat<(f32 (X86fxor FR32:$src1, FR32:$src2)),
2911 (COPY_TO_REGCLASS (VXORPSrr
2912 (COPY_TO_REGCLASS FR32:$src1, VR128),
2913 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2914 def : Pat<(f32 (X86fandn FR32:$src1, FR32:$src2)),
2915 (COPY_TO_REGCLASS (VANDNPSrr
2916 (COPY_TO_REGCLASS FR32:$src1, VR128),
2917 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2920 let Predicates = [UseSSE1] in {
2921 // Use packed logical operations for scalar ops.
2922 def : Pat<(f32 (X86fand FR32:$src1, FR32:$src2)),
2923 (COPY_TO_REGCLASS (ANDPSrr
2924 (COPY_TO_REGCLASS FR32:$src1, VR128),
2925 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2926 def : Pat<(f32 (X86for FR32:$src1, FR32:$src2)),
2927 (COPY_TO_REGCLASS (ORPSrr
2928 (COPY_TO_REGCLASS FR32:$src1, VR128),
2929 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2930 def : Pat<(f32 (X86fxor FR32:$src1, FR32:$src2)),
2931 (COPY_TO_REGCLASS (XORPSrr
2932 (COPY_TO_REGCLASS FR32:$src1, VR128),
2933 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2934 def : Pat<(f32 (X86fandn FR32:$src1, FR32:$src2)),
2935 (COPY_TO_REGCLASS (ANDNPSrr
2936 (COPY_TO_REGCLASS FR32:$src1, VR128),
2937 (COPY_TO_REGCLASS FR32:$src2, VR128)), FR32)>;
2940 let Predicates = [UseSSE2] in {
2941 // Use packed logical operations for scalar ops.
2942 def : Pat<(f64 (X86fand FR64:$src1, FR64:$src2)),
2943 (COPY_TO_REGCLASS (ANDPDrr
2944 (COPY_TO_REGCLASS FR64:$src1, VR128),
2945 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2946 def : Pat<(f64 (X86for FR64:$src1, FR64:$src2)),
2947 (COPY_TO_REGCLASS (ORPDrr
2948 (COPY_TO_REGCLASS FR64:$src1, VR128),
2949 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2950 def : Pat<(f64 (X86fxor FR64:$src1, FR64:$src2)),
2951 (COPY_TO_REGCLASS (XORPDrr
2952 (COPY_TO_REGCLASS FR64:$src1, VR128),
2953 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2954 def : Pat<(f64 (X86fandn FR64:$src1, FR64:$src2)),
2955 (COPY_TO_REGCLASS (ANDNPDrr
2956 (COPY_TO_REGCLASS FR64:$src1, VR128),
2957 (COPY_TO_REGCLASS FR64:$src2, VR128)), FR64)>;
2960 // Patterns for packed operations when we don't have integer type available.
2961 def : Pat<(v4f32 (X86fand VR128:$src1, VR128:$src2)),
2962 (ANDPSrr VR128:$src1, VR128:$src2)>;
2963 def : Pat<(v4f32 (X86for VR128:$src1, VR128:$src2)),
2964 (ORPSrr VR128:$src1, VR128:$src2)>;
2965 def : Pat<(v4f32 (X86fxor VR128:$src1, VR128:$src2)),
2966 (XORPSrr VR128:$src1, VR128:$src2)>;
2967 def : Pat<(v4f32 (X86fandn VR128:$src1, VR128:$src2)),
2968 (ANDNPSrr VR128:$src1, VR128:$src2)>;
2970 def : Pat<(X86fand VR128:$src1, (memopv4f32 addr:$src2)),
2971 (ANDPSrm VR128:$src1, addr:$src2)>;
2972 def : Pat<(X86for VR128:$src1, (memopv4f32 addr:$src2)),
2973 (ORPSrm VR128:$src1, addr:$src2)>;
2974 def : Pat<(X86fxor VR128:$src1, (memopv4f32 addr:$src2)),
2975 (XORPSrm VR128:$src1, addr:$src2)>;
2976 def : Pat<(X86fandn VR128:$src1, (memopv4f32 addr:$src2)),
2977 (ANDNPSrm VR128:$src1, addr:$src2)>;
2979 //===----------------------------------------------------------------------===//
2980 // SSE 1 & 2 - Arithmetic Instructions
2981 //===----------------------------------------------------------------------===//
2983 /// basic_sse12_fp_binop_xxx - SSE 1 & 2 binops come in both scalar and
2986 /// In addition, we also have a special variant of the scalar form here to
2987 /// represent the associated intrinsic operation. This form is unlike the
2988 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
2989 /// and leaves the top elements unmodified (therefore these cannot be commuted).
2991 /// These three forms can each be reg+reg or reg+mem.
2994 /// FIXME: once all 256-bit intrinsics are matched, cleanup and refactor those
2996 multiclass basic_sse12_fp_binop_p<bits<8> opc, string OpcodeStr,
2997 SDNode OpNode, SizeItins itins> {
2998 let Predicates = [HasAVX, NoVLX] in {
2999 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode,
3000 VR128, v4f32, f128mem, loadv4f32,
3001 SSEPackedSingle, itins.s, 0>, PS, VEX_4V, VEX_WIG;
3002 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode,
3003 VR128, v2f64, f128mem, loadv2f64,
3004 SSEPackedDouble, itins.d, 0>, PD, VEX_4V, VEX_WIG;
3006 defm V#NAME#PSY : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"),
3007 OpNode, VR256, v8f32, f256mem, loadv8f32,
3008 SSEPackedSingle, itins.s, 0>, PS, VEX_4V, VEX_L, VEX_WIG;
3009 defm V#NAME#PDY : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"),
3010 OpNode, VR256, v4f64, f256mem, loadv4f64,
3011 SSEPackedDouble, itins.d, 0>, PD, VEX_4V, VEX_L, VEX_WIG;
3014 let Constraints = "$src1 = $dst" in {
3015 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode, VR128,
3016 v4f32, f128mem, memopv4f32, SSEPackedSingle,
3018 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode, VR128,
3019 v2f64, f128mem, memopv2f64, SSEPackedDouble,
3024 multiclass basic_sse12_fp_binop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
3026 defm V#NAME#SS : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "ss"),
3027 OpNode, FR32, f32mem, SSEPackedSingle, itins.s, 0>,
3028 XS, VEX_4V, VEX_LIG, VEX_WIG;
3029 defm V#NAME#SD : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "sd"),
3030 OpNode, FR64, f64mem, SSEPackedDouble, itins.d, 0>,
3031 XD, VEX_4V, VEX_LIG, VEX_WIG;
3033 let Constraints = "$src1 = $dst" in {
3034 defm SS : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "ss"),
3035 OpNode, FR32, f32mem, SSEPackedSingle,
3037 defm SD : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "sd"),
3038 OpNode, FR64, f64mem, SSEPackedDouble,
3043 multiclass basic_sse12_fp_binop_s_int<bits<8> opc, string OpcodeStr,
3044 SDPatternOperator OpNode,
3046 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, OpNode, VR128, v4f32,
3047 !strconcat(OpcodeStr, "ss"), ssmem, sse_load_f32,
3048 SSEPackedSingle, itins.s, 0>, XS, VEX_4V, VEX_LIG, VEX_WIG;
3049 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, OpNode, VR128, v2f64,
3050 !strconcat(OpcodeStr, "sd"), sdmem, sse_load_f64,
3051 SSEPackedDouble, itins.d, 0>, XD, VEX_4V, VEX_LIG, VEX_WIG;
3053 let Constraints = "$src1 = $dst" in {
3054 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, OpNode, VR128, v4f32,
3055 !strconcat(OpcodeStr, "ss"), ssmem, sse_load_f32,
3056 SSEPackedSingle, itins.s>, XS;
3057 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, OpNode, VR128, v2f64,
3058 !strconcat(OpcodeStr, "sd"), sdmem, sse_load_f64,
3059 SSEPackedDouble, itins.d>, XD;
3063 // Binary Arithmetic instructions
3064 defm ADD : basic_sse12_fp_binop_p<0x58, "add", fadd, SSE_ALU_ITINS_P>,
3065 basic_sse12_fp_binop_s<0x58, "add", fadd, SSE_ALU_ITINS_S>,
3066 basic_sse12_fp_binop_s_int<0x58, "add", null_frag, SSE_ALU_ITINS_S>;
3067 defm MUL : basic_sse12_fp_binop_p<0x59, "mul", fmul, SSE_MUL_ITINS_P>,
3068 basic_sse12_fp_binop_s<0x59, "mul", fmul, SSE_MUL_ITINS_S>,
3069 basic_sse12_fp_binop_s_int<0x59, "mul", null_frag, SSE_MUL_ITINS_S>;
3070 let isCommutable = 0 in {
3071 defm SUB : basic_sse12_fp_binop_p<0x5C, "sub", fsub, SSE_ALU_ITINS_P>,
3072 basic_sse12_fp_binop_s<0x5C, "sub", fsub, SSE_ALU_ITINS_S>,
3073 basic_sse12_fp_binop_s_int<0x5C, "sub", null_frag,SSE_ALU_ITINS_S>;
3074 defm DIV : basic_sse12_fp_binop_p<0x5E, "div", fdiv, SSE_DIV_ITINS_P>,
3075 basic_sse12_fp_binop_s<0x5E, "div", fdiv, SSE_DIV_ITINS_S>,
3076 basic_sse12_fp_binop_s_int<0x5E, "div", null_frag,SSE_DIV_ITINS_S>;
3077 defm MAX : basic_sse12_fp_binop_p<0x5F, "max", X86fmax, SSE_ALU_ITINS_P>,
3078 basic_sse12_fp_binop_s<0x5F, "max", X86fmax, SSE_ALU_ITINS_S>,
3079 basic_sse12_fp_binop_s_int<0x5F, "max", X86fmaxs, SSE_ALU_ITINS_S>;
3080 defm MIN : basic_sse12_fp_binop_p<0x5D, "min", X86fmin, SSE_ALU_ITINS_P>,
3081 basic_sse12_fp_binop_s<0x5D, "min", X86fmin, SSE_ALU_ITINS_S>,
3082 basic_sse12_fp_binop_s_int<0x5D, "min", X86fmins, SSE_ALU_ITINS_S>;
3085 let isCodeGenOnly = 1 in {
3086 defm MAXC: basic_sse12_fp_binop_p<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_P>,
3087 basic_sse12_fp_binop_s<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_S>;
3088 defm MINC: basic_sse12_fp_binop_p<0x5D, "min", X86fminc, SSE_ALU_ITINS_P>,
3089 basic_sse12_fp_binop_s<0x5D, "min", X86fminc, SSE_ALU_ITINS_S>;
3092 // Patterns used to select SSE scalar fp arithmetic instructions from
3095 // (1) a scalar fp operation followed by a blend
3097 // The effect is that the backend no longer emits unnecessary vector
3098 // insert instructions immediately after SSE scalar fp instructions
3099 // like addss or mulss.
3101 // For example, given the following code:
3102 // __m128 foo(__m128 A, __m128 B) {
3107 // Previously we generated:
3108 // addss %xmm0, %xmm1
3109 // movss %xmm1, %xmm0
3112 // addss %xmm1, %xmm0
3114 // (2) a vector packed single/double fp operation followed by a vector insert
3116 // The effect is that the backend converts the packed fp instruction
3117 // followed by a vector insert into a single SSE scalar fp instruction.
3119 // For example, given the following code:
3120 // __m128 foo(__m128 A, __m128 B) {
3121 // __m128 C = A + B;
3122 // return (__m128) {c[0], a[1], a[2], a[3]};
3125 // Previously we generated:
3126 // addps %xmm0, %xmm1
3127 // movss %xmm1, %xmm0
3130 // addss %xmm1, %xmm0
3132 // TODO: Some canonicalization in lowering would simplify the number of
3133 // patterns we have to try to match.
3134 multiclass scalar_math_f32_patterns<SDNode Op, string OpcPrefix> {
3135 let Predicates = [UseSSE1] in {
3136 // extracted scalar math op with insert via movss
3137 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3138 (Op (f32 (extractelt (v4f32 VR128:$dst), (iPTR 0))),
3140 (!cast<I>(OpcPrefix#SSrr_Int) v4f32:$dst,
3141 (COPY_TO_REGCLASS FR32:$src, VR128))>;
3143 // vector math op with insert via movss
3144 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3145 (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3146 (!cast<I>(OpcPrefix#SSrr_Int) v4f32:$dst, v4f32:$src)>;
3149 // With SSE 4.1, blendi is preferred to movsd, so match that too.
3150 let Predicates = [UseSSE41] in {
3151 // extracted scalar math op with insert via blend
3152 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3153 (Op (f32 (extractelt (v4f32 VR128:$dst), (iPTR 0))),
3154 FR32:$src))), (i8 1))),
3155 (!cast<I>(OpcPrefix#SSrr_Int) v4f32:$dst,
3156 (COPY_TO_REGCLASS FR32:$src, VR128))>;
3158 // vector math op with insert via blend
3159 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$dst),
3160 (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)), (i8 1))),
3161 (!cast<I>(OpcPrefix#SSrr_Int)v4f32:$dst, v4f32:$src)>;
3165 // Repeat everything for AVX.
3166 let Predicates = [UseAVX] in {
3167 // extracted scalar math op with insert via movss
3168 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3169 (Op (f32 (extractelt (v4f32 VR128:$dst), (iPTR 0))),
3171 (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst,
3172 (COPY_TO_REGCLASS FR32:$src, VR128))>;
3174 // extracted scalar math op with insert via blend
3175 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3176 (Op (f32 (extractelt (v4f32 VR128:$dst), (iPTR 0))),
3177 FR32:$src))), (i8 1))),
3178 (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst,
3179 (COPY_TO_REGCLASS FR32:$src, VR128))>;
3181 // vector math op with insert via movss
3182 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3183 (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3184 (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst, v4f32:$src)>;
3186 // vector math op with insert via blend
3187 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$dst),
3188 (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)), (i8 1))),
3189 (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst, v4f32:$src)>;
3193 defm : scalar_math_f32_patterns<fadd, "ADD">;
3194 defm : scalar_math_f32_patterns<fsub, "SUB">;
3195 defm : scalar_math_f32_patterns<fmul, "MUL">;
3196 defm : scalar_math_f32_patterns<fdiv, "DIV">;
3198 multiclass scalar_math_f64_patterns<SDNode Op, string OpcPrefix> {
3199 let Predicates = [UseSSE2] in {
3200 // extracted scalar math op with insert via movsd
3201 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
3202 (Op (f64 (extractelt (v2f64 VR128:$dst), (iPTR 0))),
3204 (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst,
3205 (COPY_TO_REGCLASS FR64:$src, VR128))>;
3207 // vector math op with insert via movsd
3208 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3209 (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3210 (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
3213 // With SSE 4.1, blendi is preferred to movsd, so match those too.
3214 let Predicates = [UseSSE41] in {
3215 // extracted scalar math op with insert via blend
3216 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
3217 (Op (f64 (extractelt (v2f64 VR128:$dst), (iPTR 0))),
3218 FR64:$src))), (i8 1))),
3219 (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst,
3220 (COPY_TO_REGCLASS FR64:$src, VR128))>;
3222 // vector math op with insert via blend
3223 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst),
3224 (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)), (i8 1))),
3225 (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
3228 // Repeat everything for AVX.
3229 let Predicates = [UseAVX] in {
3230 // extracted scalar math op with insert via movsd
3231 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
3232 (Op (f64 (extractelt (v2f64 VR128:$dst), (iPTR 0))),
3234 (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst,
3235 (COPY_TO_REGCLASS FR64:$src, VR128))>;
3237 // extracted scalar math op with insert via blend
3238 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
3239 (Op (f64 (extractelt (v2f64 VR128:$dst), (iPTR 0))),
3240 FR64:$src))), (i8 1))),
3241 (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst,
3242 (COPY_TO_REGCLASS FR64:$src, VR128))>;
3244 // vector math op with insert via movsd
3245 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3246 (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3247 (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
3249 // vector math op with insert via blend
3250 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst),
3251 (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)), (i8 1))),
3252 (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
3256 defm : scalar_math_f64_patterns<fadd, "ADD">;
3257 defm : scalar_math_f64_patterns<fsub, "SUB">;
3258 defm : scalar_math_f64_patterns<fmul, "MUL">;
3259 defm : scalar_math_f64_patterns<fdiv, "DIV">;
3263 /// In addition, we also have a special variant of the scalar form here to
3264 /// represent the associated intrinsic operation. This form is unlike the
3265 /// plain scalar form, in that it takes an entire vector (instead of a
3266 /// scalar) and leaves the top elements undefined.
3268 /// And, we have a special variant form for a full-vector intrinsic form.
3270 let Sched = WriteFSqrt in {
3271 def SSE_SQRTPS : OpndItins<
3272 IIC_SSE_SQRTPS_RR, IIC_SSE_SQRTPS_RM
3275 def SSE_SQRTSS : OpndItins<
3276 IIC_SSE_SQRTSS_RR, IIC_SSE_SQRTSS_RM
3279 def SSE_SQRTPD : OpndItins<
3280 IIC_SSE_SQRTPD_RR, IIC_SSE_SQRTPD_RM
3283 def SSE_SQRTSD : OpndItins<
3284 IIC_SSE_SQRTSD_RR, IIC_SSE_SQRTSD_RM
3288 let Sched = WriteFRsqrt in {
3289 def SSE_RSQRTPS : OpndItins<
3290 IIC_SSE_RSQRTPS_RR, IIC_SSE_RSQRTPS_RM
3293 def SSE_RSQRTSS : OpndItins<
3294 IIC_SSE_RSQRTSS_RR, IIC_SSE_RSQRTSS_RM
3298 let Sched = WriteFRcp in {
3299 def SSE_RCPP : OpndItins<
3300 IIC_SSE_RCPP_RR, IIC_SSE_RCPP_RM
3303 def SSE_RCPS : OpndItins<
3304 IIC_SSE_RCPS_RR, IIC_SSE_RCPS_RM
3308 /// sse_fp_unop_s - SSE1 unops in scalar form
3309 /// For the non-AVX defs, we need $src1 to be tied to $dst because
3310 /// the HW instructions are 2 operand / destructive.
3311 multiclass sse_fp_unop_s<bits<8> opc, string OpcodeStr, RegisterClass RC,
3312 ValueType vt, ValueType ScalarVT,
3313 X86MemOperand x86memop,
3315 SDNode OpNode, Domain d, OpndItins itins,
3316 Predicate target, string Suffix> {
3317 let hasSideEffects = 0 in {
3318 def r : I<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1),
3319 !strconcat(OpcodeStr, "\t{$src1, $dst|$dst, $src1}"),
3320 [(set RC:$dst, (OpNode RC:$src1))], itins.rr, d>, Sched<[itins.Sched]>,
3323 def m : I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src1),
3324 !strconcat(OpcodeStr, "\t{$src1, $dst|$dst, $src1}"),
3325 [(set RC:$dst, (OpNode (load addr:$src1)))], itins.rm, d>,
3326 Sched<[itins.Sched.Folded, ReadAfterLd]>,
3327 Requires<[target, OptForSize]>;
3329 let isCodeGenOnly = 1, Constraints = "$src1 = $dst", ExeDomain = d in {
3330 def r_Int : I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
3331 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3332 []>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3334 def m_Int : I<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, x86memop:$src2),
3335 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3336 []>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3340 let Predicates = [target] in {
3341 // These are unary operations, but they are modeled as having 2 source operands
3342 // because the high elements of the destination are unchanged in SSE.
3343 def : Pat<(Intr VR128:$src),
3344 (!cast<Instruction>(NAME#Suffix##r_Int) VR128:$src, VR128:$src)>;
3346 // We don't want to fold scalar loads into these instructions unless
3347 // optimizing for size. This is because the folded instruction will have a
3348 // partial register update, while the unfolded sequence will not, e.g.
3350 // rcpss %xmm0, %xmm0
3351 // which has a clobber before the rcp, vs.
3353 let Predicates = [target, OptForSize] in {
3354 def : Pat<(Intr (scalar_to_vector (ScalarVT (load addr:$src2)))),
3355 (!cast<Instruction>(NAME#Suffix##m_Int)
3356 (vt (IMPLICIT_DEF)), addr:$src2)>;
3360 multiclass avx_fp_unop_s<bits<8> opc, string OpcodeStr, RegisterClass RC,
3361 ValueType vt, ValueType ScalarVT,
3362 X86MemOperand x86memop,
3363 Intrinsic Intr, SDNode OpNode, Domain d,
3364 OpndItins itins, string Suffix> {
3365 let hasSideEffects = 0 in {
3366 def r : I<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
3367 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3368 [], itins.rr, d>, Sched<[itins.Sched]>;
3370 def m : I<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
3371 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3372 [], itins.rm, d>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3373 let isCodeGenOnly = 1, ExeDomain = d in {
3374 def r_Int : I<opc, MRMSrcReg, (outs VR128:$dst),
3375 (ins VR128:$src1, VR128:$src2),
3376 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3377 []>, Sched<[itins.Sched.Folded]>;
3379 def m_Int : I<opc, MRMSrcMem, (outs VR128:$dst),
3380 (ins VR128:$src1, x86memop:$src2),
3381 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3382 []>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3386 // We don't want to fold scalar loads into these instructions unless
3387 // optimizing for size. This is because the folded instruction will have a
3388 // partial register update, while the unfolded sequence will not, e.g.
3389 // vmovss mem, %xmm0
3390 // vrcpss %xmm0, %xmm0, %xmm0
3391 // which has a clobber before the rcp, vs.
3392 // vrcpss mem, %xmm0, %xmm0
3393 // TODO: In theory, we could fold the load, and avoid the stall caused by
3394 // the partial register store, either in ExecutionDepsFix or with smarter RA.
3395 let Predicates = [UseAVX] in {
3396 def : Pat<(OpNode RC:$src), (!cast<Instruction>("V"#NAME#Suffix##r)
3397 (ScalarVT (IMPLICIT_DEF)), RC:$src)>;
3399 let Predicates = [HasAVX] in {
3400 def : Pat<(Intr VR128:$src),
3401 (!cast<Instruction>("V"#NAME#Suffix##r_Int) VR128:$src,
3404 let Predicates = [HasAVX, OptForSize] in {
3405 def : Pat<(Intr (scalar_to_vector (ScalarVT (load addr:$src2)))),
3406 (!cast<Instruction>("V"#NAME#Suffix##m_Int)
3407 (vt (IMPLICIT_DEF)), addr:$src2)>;
3409 let Predicates = [UseAVX, OptForSize] in {
3410 def : Pat<(ScalarVT (OpNode (load addr:$src))),
3411 (!cast<Instruction>("V"#NAME#Suffix##m) (ScalarVT (IMPLICIT_DEF)),
3416 /// sse1_fp_unop_p - SSE1 unops in packed form.
3417 multiclass sse1_fp_unop_p<bits<8> opc, string OpcodeStr, SDNode OpNode,
3418 OpndItins itins, list<Predicate> prds> {
3419 let Predicates = prds in {
3420 def V#NAME#PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3421 !strconcat("v", OpcodeStr,
3422 "ps\t{$src, $dst|$dst, $src}"),
3423 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))],
3424 itins.rr>, VEX, Sched<[itins.Sched]>, VEX_WIG;
3425 def V#NAME#PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3426 !strconcat("v", OpcodeStr,
3427 "ps\t{$src, $dst|$dst, $src}"),
3428 [(set VR128:$dst, (OpNode (loadv4f32 addr:$src)))],
3429 itins.rm>, VEX, Sched<[itins.Sched.Folded]>, VEX_WIG;
3430 def V#NAME#PSYr : PSI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3431 !strconcat("v", OpcodeStr,
3432 "ps\t{$src, $dst|$dst, $src}"),
3433 [(set VR256:$dst, (v8f32 (OpNode VR256:$src)))],
3434 itins.rr>, VEX, VEX_L, Sched<[itins.Sched]>, VEX_WIG;
3435 def V#NAME#PSYm : PSI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
3436 !strconcat("v", OpcodeStr,
3437 "ps\t{$src, $dst|$dst, $src}"),
3438 [(set VR256:$dst, (OpNode (loadv8f32 addr:$src)))],
3439 itins.rm>, VEX, VEX_L, Sched<[itins.Sched.Folded]>, VEX_WIG;
3442 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3443 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3444 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))], itins.rr>,
3445 Sched<[itins.Sched]>;
3446 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3447 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3448 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))], itins.rm>,
3449 Sched<[itins.Sched.Folded]>;
3452 /// sse2_fp_unop_p - SSE2 unops in vector forms.
3453 multiclass sse2_fp_unop_p<bits<8> opc, string OpcodeStr,
3454 SDNode OpNode, OpndItins itins> {
3455 let Predicates = [HasAVX] in {
3456 def V#NAME#PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3457 !strconcat("v", OpcodeStr,
3458 "pd\t{$src, $dst|$dst, $src}"),
3459 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))],
3460 itins.rr>, VEX, Sched<[itins.Sched]>, VEX_WIG;
3461 def V#NAME#PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3462 !strconcat("v", OpcodeStr,
3463 "pd\t{$src, $dst|$dst, $src}"),
3464 [(set VR128:$dst, (OpNode (loadv2f64 addr:$src)))],
3465 itins.rm>, VEX, Sched<[itins.Sched.Folded]>, VEX_WIG;
3466 def V#NAME#PDYr : PDI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3467 !strconcat("v", OpcodeStr,
3468 "pd\t{$src, $dst|$dst, $src}"),
3469 [(set VR256:$dst, (v4f64 (OpNode VR256:$src)))],
3470 itins.rr>, VEX, VEX_L, Sched<[itins.Sched]>, VEX_WIG;
3471 def V#NAME#PDYm : PDI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
3472 !strconcat("v", OpcodeStr,
3473 "pd\t{$src, $dst|$dst, $src}"),
3474 [(set VR256:$dst, (OpNode (loadv4f64 addr:$src)))],
3475 itins.rm>, VEX, VEX_L, Sched<[itins.Sched.Folded]>, VEX_WIG;
3478 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3479 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
3480 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))], itins.rr>,
3481 Sched<[itins.Sched]>;
3482 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3483 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
3484 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))], itins.rm>,
3485 Sched<[itins.Sched.Folded]>;
3488 multiclass sse1_fp_unop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
3490 defm SS : sse_fp_unop_s<opc, OpcodeStr##ss, FR32, v4f32, f32, f32mem,
3491 !cast<Intrinsic>("int_x86_sse_"##OpcodeStr##_ss), OpNode,
3492 SSEPackedSingle, itins, UseSSE1, "SS">, XS;
3493 defm V#NAME#SS : avx_fp_unop_s<opc, "v"#OpcodeStr##ss, FR32, v4f32, f32,
3495 !cast<Intrinsic>("int_x86_sse_"##OpcodeStr##_ss), OpNode,
3496 SSEPackedSingle, itins, "SS">, XS, VEX_4V, VEX_LIG, VEX_WIG;
3499 multiclass sse2_fp_unop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
3501 defm SD : sse_fp_unop_s<opc, OpcodeStr##sd, FR64, v2f64, f64, f64mem,
3502 !cast<Intrinsic>("int_x86_sse2_"##OpcodeStr##_sd),
3503 OpNode, SSEPackedDouble, itins, UseSSE2, "SD">, XD;
3504 defm V#NAME#SD : avx_fp_unop_s<opc, "v"#OpcodeStr##sd, FR64, v2f64, f64,
3506 !cast<Intrinsic>("int_x86_sse2_"##OpcodeStr##_sd),
3507 OpNode, SSEPackedDouble, itins, "SD">,
3508 XD, VEX_4V, VEX_LIG, VEX_WIG;
3512 defm SQRT : sse1_fp_unop_s<0x51, "sqrt", fsqrt, SSE_SQRTSS>,
3513 sse1_fp_unop_p<0x51, "sqrt", fsqrt, SSE_SQRTPS, [HasAVX]>,
3514 sse2_fp_unop_s<0x51, "sqrt", fsqrt, SSE_SQRTSD>,
3515 sse2_fp_unop_p<0x51, "sqrt", fsqrt, SSE_SQRTPD>;
3517 // Reciprocal approximations. Note that these typically require refinement
3518 // in order to obtain suitable precision.
3519 defm RSQRT : sse1_fp_unop_s<0x52, "rsqrt", X86frsqrt, SSE_RSQRTSS>,
3520 sse1_fp_unop_p<0x52, "rsqrt", X86frsqrt, SSE_RSQRTPS, [HasAVX, NoVLX] >;
3521 defm RCP : sse1_fp_unop_s<0x53, "rcp", X86frcp, SSE_RCPS>,
3522 sse1_fp_unop_p<0x53, "rcp", X86frcp, SSE_RCPP, [HasAVX, NoVLX]>;
3524 // There is no f64 version of the reciprocal approximation instructions.
3526 // TODO: We should add *scalar* op patterns for these just like we have for
3527 // the binops above. If the binop and unop patterns could all be unified
3528 // that would be even better.
3530 multiclass scalar_unary_math_patterns<Intrinsic Intr, string OpcPrefix,
3531 SDNode Move, ValueType VT,
3532 Predicate BasePredicate> {
3533 let Predicates = [BasePredicate] in {
3534 def : Pat<(VT (Move VT:$dst, (Intr VT:$src))),
3535 (!cast<I>(OpcPrefix#r_Int) VT:$dst, VT:$src)>;
3538 // With SSE 4.1, blendi is preferred to movs*, so match that too.
3539 let Predicates = [UseSSE41] in {
3540 def : Pat<(VT (X86Blendi VT:$dst, (Intr VT:$src), (i8 1))),
3541 (!cast<I>(OpcPrefix#r_Int) VT:$dst, VT:$src)>;
3544 // Repeat for AVX versions of the instructions.
3545 let Predicates = [HasAVX] in {
3546 def : Pat<(VT (Move VT:$dst, (Intr VT:$src))),
3547 (!cast<I>("V"#OpcPrefix#r_Int) VT:$dst, VT:$src)>;
3549 def : Pat<(VT (X86Blendi VT:$dst, (Intr VT:$src), (i8 1))),
3550 (!cast<I>("V"#OpcPrefix#r_Int) VT:$dst, VT:$src)>;
3554 defm : scalar_unary_math_patterns<int_x86_sse_rcp_ss, "RCPSS", X86Movss,
3556 defm : scalar_unary_math_patterns<int_x86_sse_rsqrt_ss, "RSQRTSS", X86Movss,
3558 defm : scalar_unary_math_patterns<int_x86_sse_sqrt_ss, "SQRTSS", X86Movss,
3560 defm : scalar_unary_math_patterns<int_x86_sse2_sqrt_sd, "SQRTSD", X86Movsd,
3564 //===----------------------------------------------------------------------===//
3565 // SSE 1 & 2 - Non-temporal stores
3566 //===----------------------------------------------------------------------===//
3568 let AddedComplexity = 400 in { // Prefer non-temporal versions
3569 let SchedRW = [WriteStore] in {
3570 let Predicates = [HasAVX, NoVLX] in {
3571 def VMOVNTPSmr : VPSI<0x2B, MRMDestMem, (outs),
3572 (ins f128mem:$dst, VR128:$src),
3573 "movntps\t{$src, $dst|$dst, $src}",
3574 [(alignednontemporalstore (v4f32 VR128:$src),
3576 IIC_SSE_MOVNT>, VEX, VEX_WIG;
3577 def VMOVNTPDmr : VPDI<0x2B, MRMDestMem, (outs),
3578 (ins f128mem:$dst, VR128:$src),
3579 "movntpd\t{$src, $dst|$dst, $src}",
3580 [(alignednontemporalstore (v2f64 VR128:$src),
3582 IIC_SSE_MOVNT>, VEX, VEX_WIG;
3584 let ExeDomain = SSEPackedInt in
3585 def VMOVNTDQmr : VPDI<0xE7, MRMDestMem, (outs),
3586 (ins i128mem:$dst, VR128:$src),
3587 "movntdq\t{$src, $dst|$dst, $src}",
3588 [(alignednontemporalstore (v2i64 VR128:$src),
3590 IIC_SSE_MOVNT>, VEX, VEX_WIG;
3592 def VMOVNTPSYmr : VPSI<0x2B, MRMDestMem, (outs),
3593 (ins f256mem:$dst, VR256:$src),
3594 "movntps\t{$src, $dst|$dst, $src}",
3595 [(alignednontemporalstore (v8f32 VR256:$src),
3597 IIC_SSE_MOVNT>, VEX, VEX_L, VEX_WIG;
3598 def VMOVNTPDYmr : VPDI<0x2B, MRMDestMem, (outs),
3599 (ins f256mem:$dst, VR256:$src),
3600 "movntpd\t{$src, $dst|$dst, $src}",
3601 [(alignednontemporalstore (v4f64 VR256:$src),
3603 IIC_SSE_MOVNT>, VEX, VEX_L, VEX_WIG;
3604 let ExeDomain = SSEPackedInt in
3605 def VMOVNTDQYmr : VPDI<0xE7, MRMDestMem, (outs),
3606 (ins i256mem:$dst, VR256:$src),
3607 "movntdq\t{$src, $dst|$dst, $src}",
3608 [(alignednontemporalstore (v4i64 VR256:$src),
3610 IIC_SSE_MOVNT>, VEX, VEX_L, VEX_WIG;
3613 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3614 "movntps\t{$src, $dst|$dst, $src}",
3615 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)],
3617 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3618 "movntpd\t{$src, $dst|$dst, $src}",
3619 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)],
3622 let ExeDomain = SSEPackedInt in
3623 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3624 "movntdq\t{$src, $dst|$dst, $src}",
3625 [(alignednontemporalstore (v2i64 VR128:$src), addr:$dst)],
3628 // There is no AVX form for instructions below this point
3629 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
3630 "movnti{l}\t{$src, $dst|$dst, $src}",
3631 [(nontemporalstore (i32 GR32:$src), addr:$dst)],
3633 PS, Requires<[HasSSE2]>;
3634 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
3635 "movnti{q}\t{$src, $dst|$dst, $src}",
3636 [(nontemporalstore (i64 GR64:$src), addr:$dst)],
3638 PS, Requires<[HasSSE2]>;
3639 } // SchedRW = [WriteStore]
3641 let Predicates = [HasAVX, NoVLX] in {
3642 def : Pat<(alignednontemporalstore (v8i32 VR256:$src), addr:$dst),
3643 (VMOVNTDQYmr addr:$dst, VR256:$src)>;
3644 def : Pat<(alignednontemporalstore (v16i16 VR256:$src), addr:$dst),
3645 (VMOVNTDQYmr addr:$dst, VR256:$src)>;
3646 def : Pat<(alignednontemporalstore (v32i8 VR256:$src), addr:$dst),
3647 (VMOVNTDQYmr addr:$dst, VR256:$src)>;
3649 def : Pat<(alignednontemporalstore (v4i32 VR128:$src), addr:$dst),
3650 (VMOVNTDQmr addr:$dst, VR128:$src)>;
3651 def : Pat<(alignednontemporalstore (v8i16 VR128:$src), addr:$dst),
3652 (VMOVNTDQmr addr:$dst, VR128:$src)>;
3653 def : Pat<(alignednontemporalstore (v16i8 VR128:$src), addr:$dst),
3654 (VMOVNTDQmr addr:$dst, VR128:$src)>;
3657 let Predicates = [UseSSE2] in {
3658 def : Pat<(alignednontemporalstore (v4i32 VR128:$src), addr:$dst),
3659 (MOVNTDQmr addr:$dst, VR128:$src)>;
3660 def : Pat<(alignednontemporalstore (v8i16 VR128:$src), addr:$dst),
3661 (MOVNTDQmr addr:$dst, VR128:$src)>;
3662 def : Pat<(alignednontemporalstore (v16i8 VR128:$src), addr:$dst),
3663 (MOVNTDQmr addr:$dst, VR128:$src)>;
3666 } // AddedComplexity
3668 //===----------------------------------------------------------------------===//
3669 // SSE 1 & 2 - Prefetch and memory fence
3670 //===----------------------------------------------------------------------===//
3672 // Prefetch intrinsic.
3673 let Predicates = [HasSSE1], SchedRW = [WriteLoad] in {
3674 def PREFETCHT0 : I<0x18, MRM1m, (outs), (ins i8mem:$src),
3675 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3), (i32 1))],
3676 IIC_SSE_PREFETCH>, TB;
3677 def PREFETCHT1 : I<0x18, MRM2m, (outs), (ins i8mem:$src),
3678 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2), (i32 1))],
3679 IIC_SSE_PREFETCH>, TB;
3680 def PREFETCHT2 : I<0x18, MRM3m, (outs), (ins i8mem:$src),
3681 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1), (i32 1))],
3682 IIC_SSE_PREFETCH>, TB;
3683 def PREFETCHNTA : I<0x18, MRM0m, (outs), (ins i8mem:$src),
3684 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0), (i32 1))],
3685 IIC_SSE_PREFETCH>, TB;
3688 // FIXME: How should flush instruction be modeled?
3689 let SchedRW = [WriteLoad] in {
3691 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
3692 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)],
3693 IIC_SSE_PREFETCH>, PS, Requires<[HasSSE2]>;
3696 let SchedRW = [WriteNop] in {
3697 // Pause. This "instruction" is encoded as "rep; nop", so even though it
3698 // was introduced with SSE2, it's backward compatible.
3699 def PAUSE : I<0x90, RawFrm, (outs), (ins),
3700 "pause", [(int_x86_sse2_pause)], IIC_SSE_PAUSE>, OBXS;
3703 let SchedRW = [WriteFence] in {
3704 // Load, store, and memory fence
3705 // TODO: As with mfence, we may want to ease the availablity of sfence/lfence
3706 // to include any 64-bit target.
3707 def SFENCE : I<0xAE, MRM_F8, (outs), (ins),
3708 "sfence", [(int_x86_sse_sfence)], IIC_SSE_SFENCE>,
3709 PS, Requires<[HasSSE1]>;
3710 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
3711 "lfence", [(int_x86_sse2_lfence)], IIC_SSE_LFENCE>,
3712 TB, Requires<[HasSSE2]>;
3713 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
3714 "mfence", [(int_x86_sse2_mfence)], IIC_SSE_MFENCE>,
3715 TB, Requires<[HasMFence]>;
3718 def : Pat<(X86MFence), (MFENCE)>;
3720 //===----------------------------------------------------------------------===//
3721 // SSE 1 & 2 - Load/Store XCSR register
3722 //===----------------------------------------------------------------------===//
3724 def VLDMXCSR : VPSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
3725 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)],
3726 IIC_SSE_LDMXCSR>, VEX, Sched<[WriteLoad]>, VEX_WIG;
3727 def VSTMXCSR : VPSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
3728 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)],
3729 IIC_SSE_STMXCSR>, VEX, Sched<[WriteStore]>, VEX_WIG;
3731 def LDMXCSR : I<0xAE, MRM2m, (outs), (ins i32mem:$src),
3732 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)],
3733 IIC_SSE_LDMXCSR>, TB, Sched<[WriteLoad]>;
3734 def STMXCSR : I<0xAE, MRM3m, (outs), (ins i32mem:$dst),
3735 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)],
3736 IIC_SSE_STMXCSR>, TB, Sched<[WriteStore]>;
3738 //===---------------------------------------------------------------------===//
3739 // SSE2 - Move Aligned/Unaligned Packed Integer Instructions
3740 //===---------------------------------------------------------------------===//
3742 let ExeDomain = SSEPackedInt in { // SSE integer instructions
3744 let hasSideEffects = 0, SchedRW = [WriteMove] in {
3745 def VMOVDQArr : VPDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3746 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>,
3748 def VMOVDQAYrr : VPDI<0x6F, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3749 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>,
3750 VEX, VEX_L, VEX_WIG;
3751 def VMOVDQUrr : VSSI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3752 "movdqu\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVU_P_RR>,
3754 def VMOVDQUYrr : VSSI<0x6F, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3755 "movdqu\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVU_P_RR>,
3756 VEX, VEX_L, VEX_WIG;
3760 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
3761 SchedRW = [WriteMove] in {
3762 def VMOVDQArr_REV : VPDI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3763 "movdqa\t{$src, $dst|$dst, $src}", [],
3765 VEX, VEX_WIG, FoldGenData<"VMOVDQArr">;
3766 def VMOVDQAYrr_REV : VPDI<0x7F, MRMDestReg, (outs VR256:$dst), (ins VR256:$src),
3767 "movdqa\t{$src, $dst|$dst, $src}", [],
3768 IIC_SSE_MOVA_P_RR>, VEX, VEX_L, VEX_WIG,
3769 FoldGenData<"VMOVDQAYrr">;
3770 def VMOVDQUrr_REV : VSSI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3771 "movdqu\t{$src, $dst|$dst, $src}", [],
3773 VEX, VEX_WIG, FoldGenData<"VMOVDQUrr">;
3774 def VMOVDQUYrr_REV : VSSI<0x7F, MRMDestReg, (outs VR256:$dst), (ins VR256:$src),
3775 "movdqu\t{$src, $dst|$dst, $src}", [],
3776 IIC_SSE_MOVU_P_RR>, VEX, VEX_L, VEX_WIG,
3777 FoldGenData<"VMOVDQUYrr">;
3780 let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1,
3781 hasSideEffects = 0, SchedRW = [WriteLoad] in {
3782 let Predicates = [HasAVX,NoVLX] in
3783 def VMOVDQArm : VPDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3784 "movdqa\t{$src, $dst|$dst, $src}",
3785 [(set VR128:$dst, (alignedloadv2i64 addr:$src))],
3786 IIC_SSE_MOVA_P_RM>, VEX, VEX_WIG;
3787 def VMOVDQAYrm : VPDI<0x6F, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
3788 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RM>,
3789 VEX, VEX_L, VEX_WIG;
3790 let Predicates = [HasAVX,NoVLX] in
3791 def VMOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3792 "vmovdqu\t{$src, $dst|$dst, $src}",
3793 [(set VR128:$dst, (loadv2i64 addr:$src))],
3794 IIC_SSE_MOVU_P_RM>, XS, VEX, VEX_WIG;
3795 def VMOVDQUYrm : I<0x6F, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
3796 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_RM>,
3797 XS, VEX, VEX_L, VEX_WIG;
3800 let mayStore = 1, hasSideEffects = 0, SchedRW = [WriteStore] in {
3801 let Predicates = [HasAVX,NoVLX] in
3802 def VMOVDQAmr : VPDI<0x7F, MRMDestMem, (outs),
3803 (ins i128mem:$dst, VR128:$src),
3804 "movdqa\t{$src, $dst|$dst, $src}",
3805 [(alignedstore (v2i64 VR128:$src), addr:$dst)],
3806 IIC_SSE_MOVA_P_MR>, VEX, VEX_WIG;
3807 def VMOVDQAYmr : VPDI<0x7F, MRMDestMem, (outs),
3808 (ins i256mem:$dst, VR256:$src),
3809 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_MR>,
3810 VEX, VEX_L, VEX_WIG;
3811 let Predicates = [HasAVX,NoVLX] in
3812 def VMOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3813 "vmovdqu\t{$src, $dst|$dst, $src}",
3814 [(store (v2i64 VR128:$src), addr:$dst)], IIC_SSE_MOVU_P_MR>,
3816 def VMOVDQUYmr : I<0x7F, MRMDestMem, (outs), (ins i256mem:$dst, VR256:$src),
3817 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_MR>,
3818 XS, VEX, VEX_L, VEX_WIG;
3821 let SchedRW = [WriteMove] in {
3822 let hasSideEffects = 0 in {
3823 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3824 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>;
3826 def MOVDQUrr : I<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3827 "movdqu\t{$src, $dst|$dst, $src}",
3828 [], IIC_SSE_MOVU_P_RR>, XS, Requires<[UseSSE2]>;
3832 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in {
3833 def MOVDQArr_REV : PDI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3834 "movdqa\t{$src, $dst|$dst, $src}", [],
3835 IIC_SSE_MOVA_P_RR>, FoldGenData<"MOVDQArr">;
3837 def MOVDQUrr_REV : I<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3838 "movdqu\t{$src, $dst|$dst, $src}",
3839 [], IIC_SSE_MOVU_P_RR>, XS, Requires<[UseSSE2]>,
3840 FoldGenData<"MOVDQUrr">;
3844 let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1,
3845 hasSideEffects = 0, SchedRW = [WriteLoad] in {
3846 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3847 "movdqa\t{$src, $dst|$dst, $src}",
3848 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/],
3850 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3851 "movdqu\t{$src, $dst|$dst, $src}",
3852 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/],
3854 XS, Requires<[UseSSE2]>;
3857 let mayStore = 1, hasSideEffects = 0, SchedRW = [WriteStore] in {
3858 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3859 "movdqa\t{$src, $dst|$dst, $src}",
3860 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/],
3862 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3863 "movdqu\t{$src, $dst|$dst, $src}",
3864 [/*(store (v2i64 VR128:$src), addr:$dst)*/],
3866 XS, Requires<[UseSSE2]>;
3869 } // ExeDomain = SSEPackedInt
3871 // Aliases to help the assembler pick two byte VEX encodings by swapping the
3872 // operands relative to the normal instructions to use VEX.R instead of VEX.B.
3873 def : InstAlias<"vmovdqa\t{$src, $dst|$dst, $src}",
3874 (VMOVDQArr_REV VR128L:$dst, VR128H:$src), 0>;
3875 def : InstAlias<"vmovdqa\t{$src, $dst|$dst, $src}",
3876 (VMOVDQAYrr_REV VR256L:$dst, VR256H:$src), 0>;
3877 def : InstAlias<"vmovdqu\t{$src, $dst|$dst, $src}",
3878 (VMOVDQUrr_REV VR128L:$dst, VR128H:$src), 0>;
3879 def : InstAlias<"vmovdqu\t{$src, $dst|$dst, $src}",
3880 (VMOVDQUYrr_REV VR256L:$dst, VR256H:$src), 0>;
3882 let Predicates = [HasAVX, NoVLX] in {
3883 // Additional patterns for other integer sizes.
3884 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3885 (VMOVDQAmr addr:$dst, VR128:$src)>;
3886 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3887 (VMOVDQAmr addr:$dst, VR128:$src)>;
3888 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3889 (VMOVDQAmr addr:$dst, VR128:$src)>;
3890 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3891 (VMOVDQUmr addr:$dst, VR128:$src)>;
3892 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3893 (VMOVDQUmr addr:$dst, VR128:$src)>;
3894 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3895 (VMOVDQUmr addr:$dst, VR128:$src)>;
3897 // Special patterns for storing subvector extracts of lower 128-bits
3898 // Its cheaper to just use VMOVDQA/VMOVDQU instead of VEXTRACTF128mr
3899 def : Pat<(alignedstore (v2i64 (extract_subvector
3900 (v4i64 VR256:$src), (iPTR 0))), addr:$dst),
3901 (VMOVDQAmr addr:$dst, (v2i64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3902 def : Pat<(alignedstore (v4i32 (extract_subvector
3903 (v8i32 VR256:$src), (iPTR 0))), addr:$dst),
3904 (VMOVDQAmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3905 def : Pat<(alignedstore (v8i16 (extract_subvector
3906 (v16i16 VR256:$src), (iPTR 0))), addr:$dst),
3907 (VMOVDQAmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3908 def : Pat<(alignedstore (v16i8 (extract_subvector
3909 (v32i8 VR256:$src), (iPTR 0))), addr:$dst),
3910 (VMOVDQAmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3912 def : Pat<(store (v2i64 (extract_subvector
3913 (v4i64 VR256:$src), (iPTR 0))), addr:$dst),
3914 (VMOVDQUmr addr:$dst, (v2i64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3915 def : Pat<(store (v4i32 (extract_subvector
3916 (v8i32 VR256:$src), (iPTR 0))), addr:$dst),
3917 (VMOVDQUmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3918 def : Pat<(store (v8i16 (extract_subvector
3919 (v16i16 VR256:$src), (iPTR 0))), addr:$dst),
3920 (VMOVDQUmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3921 def : Pat<(store (v16i8 (extract_subvector
3922 (v32i8 VR256:$src), (iPTR 0))), addr:$dst),
3923 (VMOVDQUmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
3926 //===---------------------------------------------------------------------===//
3927 // SSE2 - Packed Integer Arithmetic Instructions
3928 //===---------------------------------------------------------------------===//
3930 let Sched = WriteVecIMul in
3931 def SSE_PMADD : OpndItins<
3932 IIC_SSE_PMADD, IIC_SSE_PMADD
3935 let ExeDomain = SSEPackedInt in { // SSE integer instructions
3937 /// PDI_binop_rm2 - Simple SSE2 binary operator with different src and dst types
3938 multiclass PDI_binop_rm2<bits<8> opc, string OpcodeStr, SDNode OpNode,
3939 ValueType DstVT, ValueType SrcVT, RegisterClass RC,
3940 PatFrag memop_frag, X86MemOperand x86memop,
3941 OpndItins itins, bit Is2Addr = 1> {
3942 let isCommutable = 1 in
3943 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
3944 (ins RC:$src1, RC:$src2),
3946 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3947 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3948 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1), RC:$src2)))]>,
3949 Sched<[itins.Sched]>;
3950 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
3951 (ins RC:$src1, x86memop:$src2),
3953 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3954 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3955 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1),
3956 (bitconvert (memop_frag addr:$src2)))))]>,
3957 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3959 } // ExeDomain = SSEPackedInt
3961 defm PADDB : PDI_binop_all<0xFC, "paddb", add, v16i8, v32i8,
3962 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3963 defm PADDW : PDI_binop_all<0xFD, "paddw", add, v8i16, v16i16,
3964 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3965 defm PADDD : PDI_binop_all<0xFE, "paddd", add, v4i32, v8i32,
3966 SSE_INTALU_ITINS_P, 1, NoVLX>;
3967 defm PADDQ : PDI_binop_all<0xD4, "paddq", add, v2i64, v4i64,
3968 SSE_INTALUQ_ITINS_P, 1, NoVLX>;
3969 defm PADDSB : PDI_binop_all<0xEC, "paddsb", X86adds, v16i8, v32i8,
3970 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3971 defm PADDSW : PDI_binop_all<0xED, "paddsw", X86adds, v8i16, v16i16,
3972 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3973 defm PADDUSB : PDI_binop_all<0xDC, "paddusb", X86addus, v16i8, v32i8,
3974 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3975 defm PADDUSW : PDI_binop_all<0xDD, "paddusw", X86addus, v8i16, v16i16,
3976 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
3977 defm PMULLW : PDI_binop_all<0xD5, "pmullw", mul, v8i16, v16i16,
3978 SSE_INTMUL_ITINS_P, 1, NoVLX_Or_NoBWI>;
3979 defm PMULHUW : PDI_binop_all<0xE4, "pmulhuw", mulhu, v8i16, v16i16,
3980 SSE_INTMUL_ITINS_P, 1, NoVLX_Or_NoBWI>;
3981 defm PMULHW : PDI_binop_all<0xE5, "pmulhw", mulhs, v8i16, v16i16,
3982 SSE_INTMUL_ITINS_P, 1, NoVLX_Or_NoBWI>;
3983 defm PSUBB : PDI_binop_all<0xF8, "psubb", sub, v16i8, v32i8,
3984 SSE_INTALU_ITINS_P, 0, NoVLX_Or_NoBWI>;
3985 defm PSUBW : PDI_binop_all<0xF9, "psubw", sub, v8i16, v16i16,
3986 SSE_INTALU_ITINS_P, 0, NoVLX_Or_NoBWI>;
3987 defm PSUBD : PDI_binop_all<0xFA, "psubd", sub, v4i32, v8i32,
3988 SSE_INTALU_ITINS_P, 0, NoVLX>;
3989 defm PSUBQ : PDI_binop_all<0xFB, "psubq", sub, v2i64, v4i64,
3990 SSE_INTALUQ_ITINS_P, 0, NoVLX>;
3991 defm PSUBSB : PDI_binop_all<0xE8, "psubsb", X86subs, v16i8, v32i8,
3992 SSE_INTALU_ITINS_P, 0, NoVLX_Or_NoBWI>;
3993 defm PSUBSW : PDI_binop_all<0xE9, "psubsw", X86subs, v8i16, v16i16,
3994 SSE_INTALU_ITINS_P, 0, NoVLX_Or_NoBWI>;
3995 defm PSUBUSB : PDI_binop_all<0xD8, "psubusb", X86subus, v16i8, v32i8,
3996 SSE_INTALU_ITINS_P, 0, NoVLX_Or_NoBWI>;
3997 defm PSUBUSW : PDI_binop_all<0xD9, "psubusw", X86subus, v8i16, v16i16,
3998 SSE_INTALU_ITINS_P, 0, NoVLX_Or_NoBWI>;
3999 defm PMINUB : PDI_binop_all<0xDA, "pminub", umin, v16i8, v32i8,
4000 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
4001 defm PMINSW : PDI_binop_all<0xEA, "pminsw", smin, v8i16, v16i16,
4002 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
4003 defm PMAXUB : PDI_binop_all<0xDE, "pmaxub", umax, v16i8, v32i8,
4004 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
4005 defm PMAXSW : PDI_binop_all<0xEE, "pmaxsw", smax, v8i16, v16i16,
4006 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
4007 defm PAVGB : PDI_binop_all<0xE0, "pavgb", X86avg, v16i8, v32i8,
4008 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
4009 defm PAVGW : PDI_binop_all<0xE3, "pavgw", X86avg, v8i16, v16i16,
4010 SSE_INTALU_ITINS_P, 1, NoVLX_Or_NoBWI>;
4012 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in
4013 defm VPMADDWD : PDI_binop_rm2<0xF5, "vpmaddwd", X86vpmaddwd, v4i32, v8i16, VR128,
4014 loadv2i64, i128mem, SSE_PMADD, 0>, VEX_4V, VEX_WIG;
4016 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in
4017 defm VPMADDWDY : PDI_binop_rm2<0xF5, "vpmaddwd", X86vpmaddwd, v8i32, v16i16,
4018 VR256, loadv4i64, i256mem, SSE_PMADD,
4019 0>, VEX_4V, VEX_L, VEX_WIG;
4020 let Constraints = "$src1 = $dst" in
4021 defm PMADDWD : PDI_binop_rm2<0xF5, "pmaddwd", X86vpmaddwd, v4i32, v8i16, VR128,
4022 memopv2i64, i128mem, SSE_PMADD>;
4024 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in
4025 defm VPSADBW : PDI_binop_rm2<0xF6, "vpsadbw", X86psadbw, v2i64, v16i8, VR128,
4026 loadv2i64, i128mem, SSE_INTMUL_ITINS_P, 0>,
4028 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in
4029 defm VPSADBWY : PDI_binop_rm2<0xF6, "vpsadbw", X86psadbw, v4i64, v32i8, VR256,
4030 loadv4i64, i256mem, SSE_INTMUL_ITINS_P, 0>,
4031 VEX_4V, VEX_L, VEX_WIG;
4032 let Constraints = "$src1 = $dst" in
4033 defm PSADBW : PDI_binop_rm2<0xF6, "psadbw", X86psadbw, v2i64, v16i8, VR128,
4034 memopv2i64, i128mem, SSE_INTALU_ITINS_P>;
4036 let Predicates = [HasAVX, NoVLX] in
4037 defm VPMULUDQ : PDI_binop_rm2<0xF4, "vpmuludq", X86pmuludq, v2i64, v4i32, VR128,
4038 loadv2i64, i128mem, SSE_INTMUL_ITINS_P, 0>,
4040 let Predicates = [HasAVX2, NoVLX] in
4041 defm VPMULUDQY : PDI_binop_rm2<0xF4, "vpmuludq", X86pmuludq, v4i64, v8i32,
4042 VR256, loadv4i64, i256mem,
4043 SSE_INTMUL_ITINS_P, 0>, VEX_4V, VEX_L, VEX_WIG;
4044 let Constraints = "$src1 = $dst" in
4045 defm PMULUDQ : PDI_binop_rm2<0xF4, "pmuludq", X86pmuludq, v2i64, v4i32, VR128,
4046 memopv2i64, i128mem, SSE_INTMUL_ITINS_P>;
4048 //===---------------------------------------------------------------------===//
4049 // SSE2 - Packed Integer Logical Instructions
4050 //===---------------------------------------------------------------------===//
4052 multiclass PDI_binop_rmi<bits<8> opc, bits<8> opc2, Format ImmForm,
4053 string OpcodeStr, SDNode OpNode,
4054 SDNode OpNode2, RegisterClass RC,
4055 ValueType DstVT, ValueType SrcVT,
4056 PatFrag ld_frag, bit Is2Addr = 1> {
4057 // src2 is always 128-bit
4058 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
4059 (ins RC:$src1, VR128:$src2),
4061 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4062 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4063 [(set RC:$dst, (DstVT (OpNode RC:$src1, (SrcVT VR128:$src2))))],
4064 SSE_INTSHIFT_ITINS_P.rr>, Sched<[WriteVecShift]>;
4065 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
4066 (ins RC:$src1, i128mem:$src2),
4068 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4069 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4070 [(set RC:$dst, (DstVT (OpNode RC:$src1,
4071 (SrcVT (bitconvert (ld_frag addr:$src2))))))],
4072 SSE_INTSHIFT_ITINS_P.rm>, Sched<[WriteVecShiftLd, ReadAfterLd]>;
4073 def ri : PDIi8<opc2, ImmForm, (outs RC:$dst),
4074 (ins RC:$src1, u8imm:$src2),
4076 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4077 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4078 [(set RC:$dst, (DstVT (OpNode2 RC:$src1, (i8 imm:$src2))))],
4079 SSE_INTSHIFT_ITINS_P.ri>, Sched<[WriteVecShift]>;
4082 multiclass PDI_binop_rmi_all<bits<8> opc, bits<8> opc2, Format ImmForm,
4083 string OpcodeStr, SDNode OpNode,
4084 SDNode OpNode2, ValueType DstVT128,
4085 ValueType DstVT256, ValueType SrcVT,
4087 let Predicates = [HasAVX, prd] in
4088 defm V#NAME : PDI_binop_rmi<opc, opc2, ImmForm, !strconcat("v", OpcodeStr),
4089 OpNode, OpNode2, VR128, DstVT128, SrcVT,
4090 loadv2i64, 0>, VEX_4V, VEX_WIG;
4091 let Predicates = [HasAVX2, prd] in
4092 defm V#NAME#Y : PDI_binop_rmi<opc, opc2, ImmForm, !strconcat("v", OpcodeStr),
4093 OpNode, OpNode2, VR256, DstVT256, SrcVT,
4094 loadv2i64, 0>, VEX_4V, VEX_L, VEX_WIG;
4095 let Constraints = "$src1 = $dst" in
4096 defm NAME : PDI_binop_rmi<opc, opc2, ImmForm, OpcodeStr, OpNode, OpNode2,
4097 VR128, DstVT128, SrcVT, memopv2i64>;
4100 multiclass PDI_binop_ri<bits<8> opc, Format ImmForm, string OpcodeStr,
4101 SDNode OpNode, RegisterClass RC, ValueType VT,
4103 def ri : PDIi8<opc, ImmForm, (outs RC:$dst), (ins RC:$src1, u8imm:$src2),
4105 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4106 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4107 [(set RC:$dst, (VT (OpNode RC:$src1, (i8 imm:$src2))))],
4108 IIC_SSE_INTSHDQ_P_RI>, Sched<[WriteVecShift]>;
4111 multiclass PDI_binop_ri_all<bits<8> opc, Format ImmForm, string OpcodeStr,
4113 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in
4114 defm V#NAME : PDI_binop_ri<opc, ImmForm, !strconcat("v", OpcodeStr), OpNode,
4115 VR128, v16i8, 0>, VEX_4V, VEX_WIG;
4116 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in
4117 defm V#NAME#Y : PDI_binop_ri<opc, ImmForm, !strconcat("v", OpcodeStr), OpNode,
4118 VR256, v32i8, 0>, VEX_4V, VEX_L, VEX_WIG;
4119 let Constraints = "$src1 = $dst" in
4120 defm NAME : PDI_binop_ri<opc, ImmForm, OpcodeStr, OpNode, VR128, v16i8>;
4123 let ExeDomain = SSEPackedInt in {
4124 defm PSLLW : PDI_binop_rmi_all<0xF1, 0x71, MRM6r, "psllw", X86vshl, X86vshli,
4125 v8i16, v16i16, v8i16, NoVLX_Or_NoBWI>;
4126 defm PSLLD : PDI_binop_rmi_all<0xF2, 0x72, MRM6r, "pslld", X86vshl, X86vshli,
4127 v4i32, v8i32, v4i32, NoVLX>;
4128 defm PSLLQ : PDI_binop_rmi_all<0xF3, 0x73, MRM6r, "psllq", X86vshl, X86vshli,
4129 v2i64, v4i64, v2i64, NoVLX>;
4131 defm PSRLW : PDI_binop_rmi_all<0xD1, 0x71, MRM2r, "psrlw", X86vsrl, X86vsrli,
4132 v8i16, v16i16, v8i16, NoVLX_Or_NoBWI>;
4133 defm PSRLD : PDI_binop_rmi_all<0xD2, 0x72, MRM2r, "psrld", X86vsrl, X86vsrli,
4134 v4i32, v8i32, v4i32, NoVLX>;
4135 defm PSRLQ : PDI_binop_rmi_all<0xD3, 0x73, MRM2r, "psrlq", X86vsrl, X86vsrli,
4136 v2i64, v4i64, v2i64, NoVLX>;
4138 defm PSRAW : PDI_binop_rmi_all<0xE1, 0x71, MRM4r, "psraw", X86vsra, X86vsrai,
4139 v8i16, v16i16, v8i16, NoVLX_Or_NoBWI>;
4140 defm PSRAD : PDI_binop_rmi_all<0xE2, 0x72, MRM4r, "psrad", X86vsra, X86vsrai,
4141 v4i32, v8i32, v4i32, NoVLX>;
4143 defm PSLLDQ : PDI_binop_ri_all<0x73, MRM7r, "pslldq", X86vshldq>;
4144 defm PSRLDQ : PDI_binop_ri_all<0x73, MRM3r, "psrldq", X86vshrdq>;
4145 // PSRADQri doesn't exist in SSE[1-3].
4146 } // ExeDomain = SSEPackedInt
4148 //===---------------------------------------------------------------------===//
4149 // SSE2 - Packed Integer Comparison Instructions
4150 //===---------------------------------------------------------------------===//
4152 defm PCMPEQB : PDI_binop_all<0x74, "pcmpeqb", X86pcmpeq, v16i8, v32i8,
4153 SSE_INTALU_ITINS_P, 1, TruePredicate>;
4154 defm PCMPEQW : PDI_binop_all<0x75, "pcmpeqw", X86pcmpeq, v8i16, v16i16,
4155 SSE_INTALU_ITINS_P, 1, TruePredicate>;
4156 defm PCMPEQD : PDI_binop_all<0x76, "pcmpeqd", X86pcmpeq, v4i32, v8i32,
4157 SSE_INTALU_ITINS_P, 1, TruePredicate>;
4158 defm PCMPGTB : PDI_binop_all<0x64, "pcmpgtb", X86pcmpgt, v16i8, v32i8,
4159 SSE_INTALU_ITINS_P, 0, TruePredicate>;
4160 defm PCMPGTW : PDI_binop_all<0x65, "pcmpgtw", X86pcmpgt, v8i16, v16i16,
4161 SSE_INTALU_ITINS_P, 0, TruePredicate>;
4162 defm PCMPGTD : PDI_binop_all<0x66, "pcmpgtd", X86pcmpgt, v4i32, v8i32,
4163 SSE_INTALU_ITINS_P, 0, TruePredicate>;
4165 //===---------------------------------------------------------------------===//
4166 // SSE2 - Packed Integer Shuffle Instructions
4167 //===---------------------------------------------------------------------===//
4169 let ExeDomain = SSEPackedInt in {
4170 multiclass sse2_pshuffle<string OpcodeStr, ValueType vt128, ValueType vt256,
4171 SDNode OpNode, Predicate prd> {
4172 let Predicates = [HasAVX, prd] in {
4173 def V#NAME#ri : Ii8<0x70, MRMSrcReg, (outs VR128:$dst),
4174 (ins VR128:$src1, u8imm:$src2),
4175 !strconcat("v", OpcodeStr,
4176 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4178 (vt128 (OpNode VR128:$src1, (i8 imm:$src2))))],
4179 IIC_SSE_PSHUF_RI>, VEX, Sched<[WriteShuffle]>, VEX_WIG;
4180 def V#NAME#mi : Ii8<0x70, MRMSrcMem, (outs VR128:$dst),
4181 (ins i128mem:$src1, u8imm:$src2),
4182 !strconcat("v", OpcodeStr,
4183 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4185 (vt128 (OpNode (bitconvert (loadv2i64 addr:$src1)),
4186 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>, VEX,
4187 Sched<[WriteShuffleLd]>, VEX_WIG;
4190 let Predicates = [HasAVX2, prd] in {
4191 def V#NAME#Yri : Ii8<0x70, MRMSrcReg, (outs VR256:$dst),
4192 (ins VR256:$src1, u8imm:$src2),
4193 !strconcat("v", OpcodeStr,
4194 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4196 (vt256 (OpNode VR256:$src1, (i8 imm:$src2))))],
4197 IIC_SSE_PSHUF_RI>, VEX, VEX_L, Sched<[WriteShuffle]>, VEX_WIG;
4198 def V#NAME#Ymi : Ii8<0x70, MRMSrcMem, (outs VR256:$dst),
4199 (ins i256mem:$src1, u8imm:$src2),
4200 !strconcat("v", OpcodeStr,
4201 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4203 (vt256 (OpNode (bitconvert (loadv4i64 addr:$src1)),
4204 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>, VEX, VEX_L,
4205 Sched<[WriteShuffleLd]>, VEX_WIG;
4208 let Predicates = [UseSSE2] in {
4209 def ri : Ii8<0x70, MRMSrcReg,
4210 (outs VR128:$dst), (ins VR128:$src1, u8imm:$src2),
4211 !strconcat(OpcodeStr,
4212 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4214 (vt128 (OpNode VR128:$src1, (i8 imm:$src2))))],
4215 IIC_SSE_PSHUF_RI>, Sched<[WriteShuffle]>;
4216 def mi : Ii8<0x70, MRMSrcMem,
4217 (outs VR128:$dst), (ins i128mem:$src1, u8imm:$src2),
4218 !strconcat(OpcodeStr,
4219 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4221 (vt128 (OpNode (bitconvert (memopv2i64 addr:$src1)),
4222 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>,
4223 Sched<[WriteShuffleLd, ReadAfterLd]>;
4226 } // ExeDomain = SSEPackedInt
4228 defm PSHUFD : sse2_pshuffle<"pshufd", v4i32, v8i32, X86PShufd, NoVLX>, PD;
4229 defm PSHUFHW : sse2_pshuffle<"pshufhw", v8i16, v16i16, X86PShufhw,
4230 NoVLX_Or_NoBWI>, XS;
4231 defm PSHUFLW : sse2_pshuffle<"pshuflw", v8i16, v16i16, X86PShuflw,
4232 NoVLX_Or_NoBWI>, XD;
4234 //===---------------------------------------------------------------------===//
4235 // Packed Integer Pack Instructions (SSE & AVX)
4236 //===---------------------------------------------------------------------===//
4238 let ExeDomain = SSEPackedInt in {
4239 multiclass sse2_pack<bits<8> opc, string OpcodeStr, ValueType OutVT,
4240 ValueType ArgVT, SDNode OpNode, PatFrag ld_frag,
4242 def rr : PDI<opc, MRMSrcReg,
4243 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4245 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4246 !strconcat(OpcodeStr,
4247 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4249 (OutVT (OpNode (ArgVT VR128:$src1), VR128:$src2)))]>,
4250 Sched<[WriteShuffle]>;
4251 def rm : PDI<opc, MRMSrcMem,
4252 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4254 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4255 !strconcat(OpcodeStr,
4256 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4258 (OutVT (OpNode (ArgVT VR128:$src1),
4259 (bitconvert (ld_frag addr:$src2)))))]>,
4260 Sched<[WriteShuffleLd, ReadAfterLd]>;
4263 multiclass sse2_pack_y<bits<8> opc, string OpcodeStr, ValueType OutVT,
4264 ValueType ArgVT, SDNode OpNode> {
4265 def Yrr : PDI<opc, MRMSrcReg,
4266 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4267 !strconcat(OpcodeStr,
4268 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4270 (OutVT (OpNode (ArgVT VR256:$src1), VR256:$src2)))]>,
4271 Sched<[WriteShuffle]>;
4272 def Yrm : PDI<opc, MRMSrcMem,
4273 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4274 !strconcat(OpcodeStr,
4275 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4277 (OutVT (OpNode (ArgVT VR256:$src1),
4278 (bitconvert (loadv4i64 addr:$src2)))))]>,
4279 Sched<[WriteShuffleLd, ReadAfterLd]>;
4282 multiclass sse4_pack<bits<8> opc, string OpcodeStr, ValueType OutVT,
4283 ValueType ArgVT, SDNode OpNode, PatFrag ld_frag,
4285 def rr : SS48I<opc, MRMSrcReg,
4286 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4288 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4289 !strconcat(OpcodeStr,
4290 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4292 (OutVT (OpNode (ArgVT VR128:$src1), VR128:$src2)))]>,
4293 Sched<[WriteShuffle]>;
4294 def rm : SS48I<opc, MRMSrcMem,
4295 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4297 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4298 !strconcat(OpcodeStr,
4299 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4301 (OutVT (OpNode (ArgVT VR128:$src1),
4302 (bitconvert (ld_frag addr:$src2)))))]>,
4303 Sched<[WriteShuffleLd, ReadAfterLd]>;
4306 multiclass sse4_pack_y<bits<8> opc, string OpcodeStr, ValueType OutVT,
4307 ValueType ArgVT, SDNode OpNode> {
4308 def Yrr : SS48I<opc, MRMSrcReg,
4309 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4310 !strconcat(OpcodeStr,
4311 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4313 (OutVT (OpNode (ArgVT VR256:$src1), VR256:$src2)))]>,
4314 Sched<[WriteShuffle]>;
4315 def Yrm : SS48I<opc, MRMSrcMem,
4316 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4317 !strconcat(OpcodeStr,
4318 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4320 (OutVT (OpNode (ArgVT VR256:$src1),
4321 (bitconvert (loadv4i64 addr:$src2)))))]>,
4322 Sched<[WriteShuffleLd, ReadAfterLd]>;
4325 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
4326 defm VPACKSSWB : sse2_pack<0x63, "vpacksswb", v16i8, v8i16, X86Packss,
4327 loadv2i64, 0>, VEX_4V, VEX_WIG;
4328 defm VPACKSSDW : sse2_pack<0x6B, "vpackssdw", v8i16, v4i32, X86Packss,
4329 loadv2i64, 0>, VEX_4V, VEX_WIG;
4331 defm VPACKUSWB : sse2_pack<0x67, "vpackuswb", v16i8, v8i16, X86Packus,
4332 loadv2i64, 0>, VEX_4V, VEX_WIG;
4333 defm VPACKUSDW : sse4_pack<0x2B, "vpackusdw", v8i16, v4i32, X86Packus,
4334 loadv2i64, 0>, VEX_4V;
4337 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
4338 defm VPACKSSWB : sse2_pack_y<0x63, "vpacksswb", v32i8, v16i16, X86Packss>,
4339 VEX_4V, VEX_L, VEX_WIG;
4340 defm VPACKSSDW : sse2_pack_y<0x6B, "vpackssdw", v16i16, v8i32, X86Packss>,
4341 VEX_4V, VEX_L, VEX_WIG;
4343 defm VPACKUSWB : sse2_pack_y<0x67, "vpackuswb", v32i8, v16i16, X86Packus>,
4344 VEX_4V, VEX_L, VEX_WIG;
4345 defm VPACKUSDW : sse4_pack_y<0x2B, "vpackusdw", v16i16, v8i32, X86Packus>,
4349 let Constraints = "$src1 = $dst" in {
4350 defm PACKSSWB : sse2_pack<0x63, "packsswb", v16i8, v8i16, X86Packss,
4352 defm PACKSSDW : sse2_pack<0x6B, "packssdw", v8i16, v4i32, X86Packss,
4355 defm PACKUSWB : sse2_pack<0x67, "packuswb", v16i8, v8i16, X86Packus,
4358 defm PACKUSDW : sse4_pack<0x2B, "packusdw", v8i16, v4i32, X86Packus,
4361 } // ExeDomain = SSEPackedInt
4363 //===---------------------------------------------------------------------===//
4364 // SSE2 - Packed Integer Unpack Instructions
4365 //===---------------------------------------------------------------------===//
4367 let ExeDomain = SSEPackedInt in {
4368 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
4369 SDNode OpNode, PatFrag ld_frag, bit Is2Addr = 1> {
4370 def rr : PDI<opc, MRMSrcReg,
4371 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4373 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
4374 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4375 [(set VR128:$dst, (vt (OpNode VR128:$src1, VR128:$src2)))],
4376 IIC_SSE_UNPCK>, Sched<[WriteShuffle]>;
4377 def rm : PDI<opc, MRMSrcMem,
4378 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4380 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
4381 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4382 [(set VR128:$dst, (vt (OpNode VR128:$src1,
4383 (bitconvert (ld_frag addr:$src2)))))],
4385 Sched<[WriteShuffleLd, ReadAfterLd]>;
4388 multiclass sse2_unpack_y<bits<8> opc, string OpcodeStr, ValueType vt,
4390 def Yrr : PDI<opc, MRMSrcReg,
4391 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4392 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4393 [(set VR256:$dst, (vt (OpNode VR256:$src1, VR256:$src2)))]>,
4394 Sched<[WriteShuffle]>;
4395 def Yrm : PDI<opc, MRMSrcMem,
4396 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4397 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4398 [(set VR256:$dst, (vt (OpNode VR256:$src1,
4399 (bitconvert (loadv4i64 addr:$src2)))))]>,
4400 Sched<[WriteShuffleLd, ReadAfterLd]>;
4404 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
4405 defm VPUNPCKLBW : sse2_unpack<0x60, "vpunpcklbw", v16i8, X86Unpckl,
4406 loadv2i64, 0>, VEX_4V, VEX_WIG;
4407 defm VPUNPCKLWD : sse2_unpack<0x61, "vpunpcklwd", v8i16, X86Unpckl,
4408 loadv2i64, 0>, VEX_4V, VEX_WIG;
4409 defm VPUNPCKHBW : sse2_unpack<0x68, "vpunpckhbw", v16i8, X86Unpckh,
4410 loadv2i64, 0>, VEX_4V, VEX_WIG;
4411 defm VPUNPCKHWD : sse2_unpack<0x69, "vpunpckhwd", v8i16, X86Unpckh,
4412 loadv2i64, 0>, VEX_4V, VEX_WIG;
4414 let Predicates = [HasAVX, NoVLX] in {
4415 defm VPUNPCKLDQ : sse2_unpack<0x62, "vpunpckldq", v4i32, X86Unpckl,
4416 loadv2i64, 0>, VEX_4V, VEX_WIG;
4417 defm VPUNPCKLQDQ : sse2_unpack<0x6C, "vpunpcklqdq", v2i64, X86Unpckl,
4418 loadv2i64, 0>, VEX_4V, VEX_WIG;
4419 defm VPUNPCKHDQ : sse2_unpack<0x6A, "vpunpckhdq", v4i32, X86Unpckh,
4420 loadv2i64, 0>, VEX_4V, VEX_WIG;
4421 defm VPUNPCKHQDQ : sse2_unpack<0x6D, "vpunpckhqdq", v2i64, X86Unpckh,
4422 loadv2i64, 0>, VEX_4V, VEX_WIG;
4425 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
4426 defm VPUNPCKLBW : sse2_unpack_y<0x60, "vpunpcklbw", v32i8, X86Unpckl>,
4427 VEX_4V, VEX_L, VEX_WIG;
4428 defm VPUNPCKLWD : sse2_unpack_y<0x61, "vpunpcklwd", v16i16, X86Unpckl>,
4429 VEX_4V, VEX_L, VEX_WIG;
4430 defm VPUNPCKHBW : sse2_unpack_y<0x68, "vpunpckhbw", v32i8, X86Unpckh>,
4431 VEX_4V, VEX_L, VEX_WIG;
4432 defm VPUNPCKHWD : sse2_unpack_y<0x69, "vpunpckhwd", v16i16, X86Unpckh>,
4433 VEX_4V, VEX_L, VEX_WIG;
4435 let Predicates = [HasAVX2, NoVLX] in {
4436 defm VPUNPCKLDQ : sse2_unpack_y<0x62, "vpunpckldq", v8i32, X86Unpckl>,
4437 VEX_4V, VEX_L, VEX_WIG;
4438 defm VPUNPCKLQDQ : sse2_unpack_y<0x6C, "vpunpcklqdq", v4i64, X86Unpckl>,
4439 VEX_4V, VEX_L, VEX_WIG;
4440 defm VPUNPCKHDQ : sse2_unpack_y<0x6A, "vpunpckhdq", v8i32, X86Unpckh>,
4441 VEX_4V, VEX_L, VEX_WIG;
4442 defm VPUNPCKHQDQ : sse2_unpack_y<0x6D, "vpunpckhqdq", v4i64, X86Unpckh>,
4443 VEX_4V, VEX_L, VEX_WIG;
4446 let Constraints = "$src1 = $dst" in {
4447 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, X86Unpckl,
4449 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, X86Unpckl,
4451 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, X86Unpckl,
4453 defm PUNPCKLQDQ : sse2_unpack<0x6C, "punpcklqdq", v2i64, X86Unpckl,
4456 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, X86Unpckh,
4458 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, X86Unpckh,
4460 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, X86Unpckh,
4462 defm PUNPCKHQDQ : sse2_unpack<0x6D, "punpckhqdq", v2i64, X86Unpckh,
4465 } // ExeDomain = SSEPackedInt
4467 //===---------------------------------------------------------------------===//
4468 // SSE2 - Packed Integer Extract and Insert
4469 //===---------------------------------------------------------------------===//
4471 let ExeDomain = SSEPackedInt in {
4472 multiclass sse2_pinsrw<bit Is2Addr = 1> {
4473 def rri : Ii8<0xC4, MRMSrcReg,
4474 (outs VR128:$dst), (ins VR128:$src1,
4475 GR32orGR64:$src2, u8imm:$src3),
4477 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
4478 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4480 (X86pinsrw VR128:$src1, GR32orGR64:$src2, imm:$src3))],
4481 IIC_SSE_PINSRW>, Sched<[WriteShuffle]>;
4482 def rmi : Ii8<0xC4, MRMSrcMem,
4483 (outs VR128:$dst), (ins VR128:$src1,
4484 i16mem:$src2, u8imm:$src3),
4486 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
4487 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4489 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
4490 imm:$src3))], IIC_SSE_PINSRW>,
4491 Sched<[WriteShuffleLd, ReadAfterLd]>;
4495 let Predicates = [HasAVX, NoBWI] in
4496 def VPEXTRWri : Ii8<0xC5, MRMSrcReg,
4497 (outs GR32orGR64:$dst), (ins VR128:$src1, u8imm:$src2),
4498 "vpextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4499 [(set GR32orGR64:$dst, (X86pextrw (v8i16 VR128:$src1),
4500 imm:$src2))]>, PD, VEX,
4501 Sched<[WriteShuffle]>;
4502 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
4503 (outs GR32orGR64:$dst), (ins VR128:$src1, u8imm:$src2),
4504 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4505 [(set GR32orGR64:$dst, (X86pextrw (v8i16 VR128:$src1),
4506 imm:$src2))], IIC_SSE_PEXTRW>,
4507 Sched<[WriteShuffleLd, ReadAfterLd]>;
4510 let Predicates = [HasAVX, NoBWI] in
4511 defm VPINSRW : sse2_pinsrw<0>, PD, VEX_4V;
4513 let Predicates = [UseSSE2], Constraints = "$src1 = $dst" in
4514 defm PINSRW : sse2_pinsrw, PD;
4516 } // ExeDomain = SSEPackedInt
4518 //===---------------------------------------------------------------------===//
4519 // SSE2 - Packed Mask Creation
4520 //===---------------------------------------------------------------------===//
4522 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecLogic] in {
4524 def VPMOVMSKBrr : VPDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst),
4526 "pmovmskb\t{$src, $dst|$dst, $src}",
4527 [(set GR32orGR64:$dst, (X86movmsk (v16i8 VR128:$src)))],
4528 IIC_SSE_MOVMSK>, VEX, VEX_WIG;
4530 let Predicates = [HasAVX2] in {
4531 def VPMOVMSKBYrr : VPDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst),
4533 "pmovmskb\t{$src, $dst|$dst, $src}",
4534 [(set GR32orGR64:$dst, (X86movmsk (v32i8 VR256:$src)))]>,
4535 VEX, VEX_L, VEX_WIG;
4538 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst), (ins VR128:$src),
4539 "pmovmskb\t{$src, $dst|$dst, $src}",
4540 [(set GR32orGR64:$dst, (X86movmsk (v16i8 VR128:$src)))],
4543 } // ExeDomain = SSEPackedInt
4545 //===---------------------------------------------------------------------===//
4546 // SSE2 - Conditional Store
4547 //===---------------------------------------------------------------------===//
4549 let ExeDomain = SSEPackedInt, SchedRW = [WriteStore] in {
4551 let Uses = [EDI], Predicates = [HasAVX,Not64BitMode] in
4552 def VMASKMOVDQU : VPDI<0xF7, MRMSrcReg, (outs),
4553 (ins VR128:$src, VR128:$mask),
4554 "maskmovdqu\t{$mask, $src|$src, $mask}",
4555 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)],
4556 IIC_SSE_MASKMOV>, VEX, VEX_WIG;
4557 let Uses = [RDI], Predicates = [HasAVX,In64BitMode] in
4558 def VMASKMOVDQU64 : VPDI<0xF7, MRMSrcReg, (outs),
4559 (ins VR128:$src, VR128:$mask),
4560 "maskmovdqu\t{$mask, $src|$src, $mask}",
4561 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)],
4562 IIC_SSE_MASKMOV>, VEX, VEX_WIG;
4564 let Uses = [EDI], Predicates = [UseSSE2,Not64BitMode] in
4565 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
4566 "maskmovdqu\t{$mask, $src|$src, $mask}",
4567 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)],
4569 let Uses = [RDI], Predicates = [UseSSE2,In64BitMode] in
4570 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
4571 "maskmovdqu\t{$mask, $src|$src, $mask}",
4572 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)],
4575 } // ExeDomain = SSEPackedInt
4577 //===---------------------------------------------------------------------===//
4578 // SSE2 - Move Doubleword/Quadword
4579 //===---------------------------------------------------------------------===//
4581 //===---------------------------------------------------------------------===//
4582 // Move Int Doubleword to Packed Double Int
4584 let ExeDomain = SSEPackedInt in {
4585 def VMOVDI2PDIrr : VS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
4586 "movd\t{$src, $dst|$dst, $src}",
4588 (v4i32 (scalar_to_vector GR32:$src)))], IIC_SSE_MOVDQ>,
4589 VEX, Sched<[WriteMove]>;
4590 def VMOVDI2PDIrm : VS2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
4591 "movd\t{$src, $dst|$dst, $src}",
4593 (v4i32 (scalar_to_vector (loadi32 addr:$src))))],
4595 VEX, Sched<[WriteLoad]>;
4596 def VMOV64toPQIrr : VRS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4597 "movq\t{$src, $dst|$dst, $src}",
4599 (v2i64 (scalar_to_vector GR64:$src)))],
4600 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4601 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0, mayLoad = 1 in
4602 def VMOV64toPQIrm : VRS2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4603 "movq\t{$src, $dst|$dst, $src}",
4604 [], IIC_SSE_MOVDQ>, VEX, Sched<[WriteLoad]>;
4605 let isCodeGenOnly = 1 in
4606 def VMOV64toSDrr : VRS2I<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
4607 "movq\t{$src, $dst|$dst, $src}",
4608 [(set FR64:$dst, (bitconvert GR64:$src))],
4609 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4611 def MOVDI2PDIrr : S2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
4612 "movd\t{$src, $dst|$dst, $src}",
4614 (v4i32 (scalar_to_vector GR32:$src)))], IIC_SSE_MOVDQ>,
4616 def MOVDI2PDIrm : S2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
4617 "movd\t{$src, $dst|$dst, $src}",
4619 (v4i32 (scalar_to_vector (loadi32 addr:$src))))],
4620 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4621 def MOV64toPQIrr : RS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4622 "movq\t{$src, $dst|$dst, $src}",
4624 (v2i64 (scalar_to_vector GR64:$src)))],
4625 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4626 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0, mayLoad = 1 in
4627 def MOV64toPQIrm : RS2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4628 "movq\t{$src, $dst|$dst, $src}",
4629 [], IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4630 let isCodeGenOnly = 1 in
4631 def MOV64toSDrr : RS2I<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
4632 "movq\t{$src, $dst|$dst, $src}",
4633 [(set FR64:$dst, (bitconvert GR64:$src))],
4634 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4635 } // ExeDomain = SSEPackedInt
4637 //===---------------------------------------------------------------------===//
4638 // Move Int Doubleword to Single Scalar
4640 let ExeDomain = SSEPackedInt, isCodeGenOnly = 1 in {
4641 def VMOVDI2SSrr : VS2I<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
4642 "movd\t{$src, $dst|$dst, $src}",
4643 [(set FR32:$dst, (bitconvert GR32:$src))],
4644 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4646 def VMOVDI2SSrm : VS2I<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
4647 "movd\t{$src, $dst|$dst, $src}",
4648 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))],
4650 VEX, Sched<[WriteLoad]>;
4651 def MOVDI2SSrr : S2I<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
4652 "movd\t{$src, $dst|$dst, $src}",
4653 [(set FR32:$dst, (bitconvert GR32:$src))],
4654 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4656 def MOVDI2SSrm : S2I<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
4657 "movd\t{$src, $dst|$dst, $src}",
4658 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))],
4659 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4660 } // ExeDomain = SSEPackedInt, isCodeGenOnly = 1
4662 //===---------------------------------------------------------------------===//
4663 // Move Packed Doubleword Int to Packed Double Int
4665 let ExeDomain = SSEPackedInt in {
4666 def VMOVPDI2DIrr : VS2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
4667 "movd\t{$src, $dst|$dst, $src}",
4668 [(set GR32:$dst, (extractelt (v4i32 VR128:$src),
4669 (iPTR 0)))], IIC_SSE_MOVD_ToGP>, VEX,
4671 def VMOVPDI2DImr : VS2I<0x7E, MRMDestMem, (outs),
4672 (ins i32mem:$dst, VR128:$src),
4673 "movd\t{$src, $dst|$dst, $src}",
4674 [(store (i32 (extractelt (v4i32 VR128:$src),
4675 (iPTR 0))), addr:$dst)], IIC_SSE_MOVDQ>,
4676 VEX, Sched<[WriteStore]>;
4677 def MOVPDI2DIrr : S2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
4678 "movd\t{$src, $dst|$dst, $src}",
4679 [(set GR32:$dst, (extractelt (v4i32 VR128:$src),
4680 (iPTR 0)))], IIC_SSE_MOVD_ToGP>,
4682 def MOVPDI2DImr : S2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
4683 "movd\t{$src, $dst|$dst, $src}",
4684 [(store (i32 (extractelt (v4i32 VR128:$src),
4685 (iPTR 0))), addr:$dst)],
4686 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4687 } // ExeDomain = SSEPackedInt
4688 //===---------------------------------------------------------------------===//
4689 // Move Packed Doubleword Int first element to Doubleword Int
4691 let ExeDomain = SSEPackedInt in {
4692 let SchedRW = [WriteMove] in {
4693 def VMOVPQIto64rr : VRS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
4694 "movq\t{$src, $dst|$dst, $src}",
4695 [(set GR64:$dst, (extractelt (v2i64 VR128:$src),
4700 def MOVPQIto64rr : RS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
4701 "movq\t{$src, $dst|$dst, $src}",
4702 [(set GR64:$dst, (extractelt (v2i64 VR128:$src),
4707 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0, mayStore = 1 in
4708 def VMOVPQIto64mr : VRS2I<0x7E, MRMDestMem, (outs),
4709 (ins i64mem:$dst, VR128:$src),
4710 "movq\t{$src, $dst|$dst, $src}",
4711 [], IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4712 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0, mayStore = 1 in
4713 def MOVPQIto64mr : RS2I<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
4714 "movq\t{$src, $dst|$dst, $src}",
4715 [], IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4716 } // ExeDomain = SSEPackedInt
4718 //===---------------------------------------------------------------------===//
4719 // Bitcast FR64 <-> GR64
4721 let ExeDomain = SSEPackedInt, isCodeGenOnly = 1 in {
4722 let Predicates = [UseAVX] in
4723 def VMOV64toSDrm : VS2SI<0x7E, MRMSrcMem, (outs FR64:$dst), (ins i64mem:$src),
4724 "movq\t{$src, $dst|$dst, $src}",
4725 [(set FR64:$dst, (bitconvert (loadi64 addr:$src)))]>,
4726 VEX, Sched<[WriteLoad]>;
4727 def VMOVSDto64rr : VRS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins FR64:$src),
4728 "movq\t{$src, $dst|$dst, $src}",
4729 [(set GR64:$dst, (bitconvert FR64:$src))],
4730 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4731 def VMOVSDto64mr : VRS2I<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, FR64:$src),
4732 "movq\t{$src, $dst|$dst, $src}",
4733 [(store (i64 (bitconvert FR64:$src)), addr:$dst)],
4734 IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4736 def MOV64toSDrm : S2SI<0x7E, MRMSrcMem, (outs FR64:$dst), (ins i64mem:$src),
4737 "movq\t{$src, $dst|$dst, $src}",
4738 [(set FR64:$dst, (bitconvert (loadi64 addr:$src)))],
4739 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4740 def MOVSDto64rr : RS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins FR64:$src),
4741 "movq\t{$src, $dst|$dst, $src}",
4742 [(set GR64:$dst, (bitconvert FR64:$src))],
4743 IIC_SSE_MOVD_ToGP>, Sched<[WriteMove]>;
4744 def MOVSDto64mr : RS2I<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, FR64:$src),
4745 "movq\t{$src, $dst|$dst, $src}",
4746 [(store (i64 (bitconvert FR64:$src)), addr:$dst)],
4747 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4748 } // ExeDomain = SSEPackedInt, isCodeGenOnly = 1
4750 //===---------------------------------------------------------------------===//
4751 // Move Scalar Single to Double Int
4753 let ExeDomain = SSEPackedInt, isCodeGenOnly = 1 in {
4754 def VMOVSS2DIrr : VS2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
4755 "movd\t{$src, $dst|$dst, $src}",
4756 [(set GR32:$dst, (bitconvert FR32:$src))],
4757 IIC_SSE_MOVD_ToGP>, VEX, Sched<[WriteMove]>;
4758 def VMOVSS2DImr : VS2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
4759 "movd\t{$src, $dst|$dst, $src}",
4760 [(store (i32 (bitconvert FR32:$src)), addr:$dst)],
4761 IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4762 def MOVSS2DIrr : S2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
4763 "movd\t{$src, $dst|$dst, $src}",
4764 [(set GR32:$dst, (bitconvert FR32:$src))],
4765 IIC_SSE_MOVD_ToGP>, Sched<[WriteMove]>;
4766 def MOVSS2DImr : S2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
4767 "movd\t{$src, $dst|$dst, $src}",
4768 [(store (i32 (bitconvert FR32:$src)), addr:$dst)],
4769 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4770 } // ExeDomain = SSEPackedInt, isCodeGenOnly = 1
4772 let Predicates = [UseAVX] in {
4773 let AddedComplexity = 15 in {
4774 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector GR32:$src)))),
4775 (VMOVDI2PDIrr GR32:$src)>;
4777 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector GR64:$src)))),
4778 (VMOV64toPQIrr GR64:$src)>;
4780 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
4781 (v2i64 (scalar_to_vector GR64:$src)),(iPTR 0)))),
4782 (SUBREG_TO_REG (i64 0), (VMOV64toPQIrr GR64:$src), sub_xmm)>;
4784 // AVX 128-bit movd/movq instructions write zeros in the high 128-bit part.
4785 // These instructions also write zeros in the high part of a 256-bit register.
4786 let AddedComplexity = 20 in {
4787 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector (zextloadi64i32 addr:$src))))),
4788 (VMOVDI2PDIrm addr:$src)>;
4789 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
4790 (VMOVDI2PDIrm addr:$src)>;
4791 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
4792 (VMOVDI2PDIrm addr:$src)>;
4793 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
4794 (VMOVDI2PDIrm addr:$src)>;
4795 def : Pat<(v4i32 (X86vzload addr:$src)),
4796 (VMOVDI2PDIrm addr:$src)>;
4797 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
4798 (v4i32 (scalar_to_vector (loadi32 addr:$src))), (iPTR 0)))),
4799 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrm addr:$src), sub_xmm)>;
4800 def : Pat<(v8i32 (X86vzload addr:$src)),
4801 (SUBREG_TO_REG (i64 0), (VMOVDI2PDIrm addr:$src), sub_xmm)>;
4803 // Use regular 128-bit instructions to match 256-bit scalar_to_vec+zext.
4804 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
4805 (v4i32 (scalar_to_vector GR32:$src)),(iPTR 0)))),
4806 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrr GR32:$src), sub_xmm)>;
4809 let Predicates = [UseSSE2] in {
4810 let AddedComplexity = 15 in {
4811 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector GR32:$src)))),
4812 (MOVDI2PDIrr GR32:$src)>;
4814 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector GR64:$src)))),
4815 (MOV64toPQIrr GR64:$src)>;
4817 let AddedComplexity = 20 in {
4818 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector (zextloadi64i32 addr:$src))))),
4819 (MOVDI2PDIrm addr:$src)>;
4820 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
4821 (MOVDI2PDIrm addr:$src)>;
4822 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
4823 (MOVDI2PDIrm addr:$src)>;
4824 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
4825 (MOVDI2PDIrm addr:$src)>;
4826 def : Pat<(v4i32 (X86vzload addr:$src)),
4827 (MOVDI2PDIrm addr:$src)>;
4831 // Before the MC layer of LLVM existed, clang emitted "movd" assembly instead of
4832 // "movq" due to MacOS parsing limitation. In order to parse old assembly, we add
4834 def : InstAlias<"movd\t{$src, $dst|$dst, $src}",
4835 (MOV64toPQIrr VR128:$dst, GR64:$src), 0>;
4836 def : InstAlias<"movd\t{$src, $dst|$dst, $src}",
4837 (MOVPQIto64rr GR64:$dst, VR128:$src), 0>;
4838 // Allow "vmovd" but print "vmovq" since we don't need compatibility for AVX.
4839 def : InstAlias<"vmovd\t{$src, $dst|$dst, $src}",
4840 (VMOV64toPQIrr VR128:$dst, GR64:$src), 0>;
4841 def : InstAlias<"vmovd\t{$src, $dst|$dst, $src}",
4842 (VMOVPQIto64rr GR64:$dst, VR128:$src), 0>;
4844 //===---------------------------------------------------------------------===//
4845 // SSE2 - Move Quadword
4846 //===---------------------------------------------------------------------===//
4848 //===---------------------------------------------------------------------===//
4849 // Move Quadword Int to Packed Quadword Int
4852 let ExeDomain = SSEPackedInt, SchedRW = [WriteLoad] in {
4853 def VMOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4854 "vmovq\t{$src, $dst|$dst, $src}",
4856 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
4857 VEX, Requires<[UseAVX]>, VEX_WIG;
4858 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4859 "movq\t{$src, $dst|$dst, $src}",
4861 (v2i64 (scalar_to_vector (loadi64 addr:$src))))],
4863 Requires<[UseSSE2]>; // SSE2 instruction with XS Prefix
4864 } // ExeDomain, SchedRW
4866 //===---------------------------------------------------------------------===//
4867 // Move Packed Quadword Int to Quadword Int
4869 let ExeDomain = SSEPackedInt, SchedRW = [WriteStore] in {
4870 def VMOVPQI2QImr : VS2I<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
4871 "movq\t{$src, $dst|$dst, $src}",
4872 [(store (i64 (extractelt (v2i64 VR128:$src),
4873 (iPTR 0))), addr:$dst)],
4874 IIC_SSE_MOVDQ>, VEX, VEX_WIG;
4875 def MOVPQI2QImr : S2I<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
4876 "movq\t{$src, $dst|$dst, $src}",
4877 [(store (i64 (extractelt (v2i64 VR128:$src),
4878 (iPTR 0))), addr:$dst)],
4880 } // ExeDomain, SchedRW
4882 // For disassembler only
4883 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
4884 SchedRW = [WriteVecLogic] in {
4885 def VMOVPQI2QIrr : VS2I<0xD6, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
4886 "movq\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVQ_RR>, VEX, VEX_WIG;
4887 def MOVPQI2QIrr : S2I<0xD6, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
4888 "movq\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVQ_RR>;
4891 // Aliases to help the assembler pick two byte VEX encodings by swapping the
4892 // operands relative to the normal instructions to use VEX.R instead of VEX.B.
4893 def : InstAlias<"vmovq\t{$src, $dst|$dst, $src}",
4894 (VMOVPQI2QIrr VR128L:$dst, VR128H:$src), 0>;
4896 let Predicates = [UseAVX], AddedComplexity = 20 in {
4897 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
4898 (VMOVQI2PQIrm addr:$src)>;
4899 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
4900 (VMOVQI2PQIrm addr:$src)>;
4901 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
4902 (VMOVQI2PQIrm addr:$src)>;
4903 def : Pat<(v2i64 (X86vzload addr:$src)),
4904 (VMOVQI2PQIrm addr:$src)>;
4905 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
4906 (v2i64 (scalar_to_vector (loadi64 addr:$src))), (iPTR 0)))),
4907 (SUBREG_TO_REG (i64 0), (VMOVQI2PQIrm addr:$src), sub_xmm)>;
4908 def : Pat<(v4i64 (X86vzload addr:$src)),
4909 (SUBREG_TO_REG (i64 0), (VMOVQI2PQIrm addr:$src), sub_xmm)>;
4912 let Predicates = [UseSSE2], AddedComplexity = 20 in {
4913 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
4914 (MOVQI2PQIrm addr:$src)>;
4915 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
4916 (MOVQI2PQIrm addr:$src)>;
4917 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
4918 (MOVQI2PQIrm addr:$src)>;
4919 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVQI2PQIrm addr:$src)>;
4922 //===---------------------------------------------------------------------===//
4923 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
4924 // IA32 document. movq xmm1, xmm2 does clear the high bits.
4926 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecLogic] in {
4927 let AddedComplexity = 15 in
4928 def VMOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
4929 "vmovq\t{$src, $dst|$dst, $src}",
4930 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))],
4932 XS, VEX, Requires<[UseAVX]>, VEX_WIG;
4933 let AddedComplexity = 15 in
4934 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
4935 "movq\t{$src, $dst|$dst, $src}",
4936 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))],
4938 XS, Requires<[UseSSE2]>;
4939 } // ExeDomain, SchedRW
4941 let AddedComplexity = 20 in {
4942 let Predicates = [UseAVX] in {
4943 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
4944 (VMOVZPQILo2PQIrr VR128:$src)>;
4946 let Predicates = [UseSSE2] in {
4947 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
4948 (MOVZPQILo2PQIrr VR128:$src)>;
4952 //===---------------------------------------------------------------------===//
4953 // SSE3 - Replicate Single FP - MOVSHDUP and MOVSLDUP
4954 //===---------------------------------------------------------------------===//
4955 multiclass sse3_replicate_sfp<bits<8> op, SDNode OpNode, string OpcodeStr,
4956 ValueType vt, RegisterClass RC, PatFrag mem_frag,
4957 X86MemOperand x86memop> {
4958 def rr : S3SI<op, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
4959 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
4960 [(set RC:$dst, (vt (OpNode RC:$src)))],
4961 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
4962 def rm : S3SI<op, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
4963 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
4964 [(set RC:$dst, (OpNode (mem_frag addr:$src)))],
4965 IIC_SSE_MOV_LH>, Sched<[WriteLoad]>;
4968 let Predicates = [HasAVX, NoVLX] in {
4969 defm VMOVSHDUP : sse3_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
4970 v4f32, VR128, loadv4f32, f128mem>, VEX, VEX_WIG;
4971 defm VMOVSLDUP : sse3_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
4972 v4f32, VR128, loadv4f32, f128mem>, VEX, VEX_WIG;
4973 defm VMOVSHDUPY : sse3_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
4974 v8f32, VR256, loadv8f32, f256mem>, VEX, VEX_L, VEX_WIG;
4975 defm VMOVSLDUPY : sse3_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
4976 v8f32, VR256, loadv8f32, f256mem>, VEX, VEX_L, VEX_WIG;
4978 defm MOVSHDUP : sse3_replicate_sfp<0x16, X86Movshdup, "movshdup", v4f32, VR128,
4979 memopv4f32, f128mem>;
4980 defm MOVSLDUP : sse3_replicate_sfp<0x12, X86Movsldup, "movsldup", v4f32, VR128,
4981 memopv4f32, f128mem>;
4983 let Predicates = [HasAVX, NoVLX] in {
4984 def : Pat<(v4i32 (X86Movshdup VR128:$src)),
4985 (VMOVSHDUPrr VR128:$src)>;
4986 def : Pat<(v4i32 (X86Movshdup (bc_v4i32 (loadv2i64 addr:$src)))),
4987 (VMOVSHDUPrm addr:$src)>;
4988 def : Pat<(v4i32 (X86Movsldup VR128:$src)),
4989 (VMOVSLDUPrr VR128:$src)>;
4990 def : Pat<(v4i32 (X86Movsldup (bc_v4i32 (loadv2i64 addr:$src)))),
4991 (VMOVSLDUPrm addr:$src)>;
4992 def : Pat<(v8i32 (X86Movshdup VR256:$src)),
4993 (VMOVSHDUPYrr VR256:$src)>;
4994 def : Pat<(v8i32 (X86Movshdup (bc_v8i32 (loadv4i64 addr:$src)))),
4995 (VMOVSHDUPYrm addr:$src)>;
4996 def : Pat<(v8i32 (X86Movsldup VR256:$src)),
4997 (VMOVSLDUPYrr VR256:$src)>;
4998 def : Pat<(v8i32 (X86Movsldup (bc_v8i32 (loadv4i64 addr:$src)))),
4999 (VMOVSLDUPYrm addr:$src)>;
5002 let Predicates = [UseSSE3] in {
5003 def : Pat<(v4i32 (X86Movshdup VR128:$src)),
5004 (MOVSHDUPrr VR128:$src)>;
5005 def : Pat<(v4i32 (X86Movshdup (bc_v4i32 (memopv2i64 addr:$src)))),
5006 (MOVSHDUPrm addr:$src)>;
5007 def : Pat<(v4i32 (X86Movsldup VR128:$src)),
5008 (MOVSLDUPrr VR128:$src)>;
5009 def : Pat<(v4i32 (X86Movsldup (bc_v4i32 (memopv2i64 addr:$src)))),
5010 (MOVSLDUPrm addr:$src)>;
5013 //===---------------------------------------------------------------------===//
5014 // SSE3 - Replicate Double FP - MOVDDUP
5015 //===---------------------------------------------------------------------===//
5017 multiclass sse3_replicate_dfp<string OpcodeStr> {
5018 def rr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5019 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5020 [(set VR128:$dst, (v2f64 (X86Movddup VR128:$src)))],
5021 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
5022 def rm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
5023 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5026 (scalar_to_vector (loadf64 addr:$src)))))],
5027 IIC_SSE_MOV_LH>, Sched<[WriteLoad]>;
5030 // FIXME: Merge with above classe when there're patterns for the ymm version
5031 multiclass sse3_replicate_dfp_y<string OpcodeStr> {
5032 def rr : S3DI<0x12, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
5033 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5034 [(set VR256:$dst, (v4f64 (X86Movddup VR256:$src)))]>,
5035 Sched<[WriteFShuffle]>;
5036 def rm : S3DI<0x12, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
5037 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5039 (v4f64 (X86Movddup (loadv4f64 addr:$src))))]>,
5043 let Predicates = [HasAVX, NoVLX] in {
5044 defm VMOVDDUP : sse3_replicate_dfp<"vmovddup">, VEX, VEX_WIG;
5045 defm VMOVDDUPY : sse3_replicate_dfp_y<"vmovddup">, VEX, VEX_L, VEX_WIG;
5048 defm MOVDDUP : sse3_replicate_dfp<"movddup">;
5051 let Predicates = [HasAVX, NoVLX] in {
5052 def : Pat<(X86Movddup (loadv2f64 addr:$src)),
5053 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5056 def : Pat<(X86Movddup (loadv4i64 addr:$src)),
5057 (VMOVDDUPYrm addr:$src)>;
5058 def : Pat<(X86Movddup (v4i64 VR256:$src)),
5059 (VMOVDDUPYrr VR256:$src)>;
5062 let Predicates = [HasAVX, NoVLX] in
5063 def : Pat<(v2f64 (X86VBroadcast (loadf64 addr:$src))),
5064 (VMOVDDUPrm addr:$src)>;
5065 let Predicates = [HasAVX1Only] in
5066 def : Pat<(v2i64 (X86VBroadcast (loadi64 addr:$src))),
5067 (VMOVDDUPrm addr:$src)>;
5069 let Predicates = [UseSSE3] in {
5070 def : Pat<(X86Movddup (memopv2f64 addr:$src)),
5071 (MOVDDUPrm addr:$src)>;
5074 //===---------------------------------------------------------------------===//
5075 // SSE3 - Move Unaligned Integer
5076 //===---------------------------------------------------------------------===//
5078 let SchedRW = [WriteLoad] in {
5079 let Predicates = [HasAVX] in {
5080 def VLDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5081 "vlddqu\t{$src, $dst|$dst, $src}",
5082 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>, VEX, VEX_WIG;
5083 def VLDDQUYrm : S3DI<0xF0, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
5084 "vlddqu\t{$src, $dst|$dst, $src}",
5085 [(set VR256:$dst, (int_x86_avx_ldu_dq_256 addr:$src))]>,
5086 VEX, VEX_L, VEX_WIG;
5088 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5089 "lddqu\t{$src, $dst|$dst, $src}",
5090 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))],
5094 //===---------------------------------------------------------------------===//
5095 // SSE3 - Arithmetic
5096 //===---------------------------------------------------------------------===//
5098 multiclass sse3_addsub<Intrinsic Int, string OpcodeStr, RegisterClass RC,
5099 X86MemOperand x86memop, OpndItins itins,
5100 PatFrag ld_frag, bit Is2Addr = 1> {
5101 def rr : I<0xD0, MRMSrcReg,
5102 (outs RC:$dst), (ins RC:$src1, RC:$src2),
5104 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5105 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5106 [(set RC:$dst, (Int RC:$src1, RC:$src2))], itins.rr>,
5107 Sched<[itins.Sched]>;
5108 def rm : I<0xD0, MRMSrcMem,
5109 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5111 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5112 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5113 [(set RC:$dst, (Int RC:$src1, (ld_frag addr:$src2)))], itins.rr>,
5114 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5117 let Predicates = [HasAVX] in {
5118 let ExeDomain = SSEPackedSingle in {
5119 defm VADDSUBPS : sse3_addsub<int_x86_sse3_addsub_ps, "vaddsubps", VR128,
5120 f128mem, SSE_ALU_F32P, loadv4f32, 0>, XD, VEX_4V, VEX_WIG;
5121 defm VADDSUBPSY : sse3_addsub<int_x86_avx_addsub_ps_256, "vaddsubps", VR256,
5122 f256mem, SSE_ALU_F32P, loadv8f32, 0>, XD, VEX_4V, VEX_L, VEX_WIG;
5124 let ExeDomain = SSEPackedDouble in {
5125 defm VADDSUBPD : sse3_addsub<int_x86_sse3_addsub_pd, "vaddsubpd", VR128,
5126 f128mem, SSE_ALU_F64P, loadv2f64, 0>, PD, VEX_4V, VEX_WIG;
5127 defm VADDSUBPDY : sse3_addsub<int_x86_avx_addsub_pd_256, "vaddsubpd", VR256,
5128 f256mem, SSE_ALU_F64P, loadv4f64, 0>, PD, VEX_4V, VEX_L, VEX_WIG;
5131 let Constraints = "$src1 = $dst", Predicates = [UseSSE3] in {
5132 let ExeDomain = SSEPackedSingle in
5133 defm ADDSUBPS : sse3_addsub<int_x86_sse3_addsub_ps, "addsubps", VR128,
5134 f128mem, SSE_ALU_F32P, memopv4f32>, XD;
5135 let ExeDomain = SSEPackedDouble in
5136 defm ADDSUBPD : sse3_addsub<int_x86_sse3_addsub_pd, "addsubpd", VR128,
5137 f128mem, SSE_ALU_F64P, memopv2f64>, PD;
5140 // Patterns used to select 'addsub' instructions.
5141 let Predicates = [HasAVX] in {
5142 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 VR128:$rhs))),
5143 (VADDSUBPSrr VR128:$lhs, VR128:$rhs)>;
5144 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (loadv4f32 addr:$rhs))),
5145 (VADDSUBPSrm VR128:$lhs, f128mem:$rhs)>;
5146 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 VR128:$rhs))),
5147 (VADDSUBPDrr VR128:$lhs, VR128:$rhs)>;
5148 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (loadv2f64 addr:$rhs))),
5149 (VADDSUBPDrm VR128:$lhs, f128mem:$rhs)>;
5151 def : Pat<(v8f32 (X86Addsub (v8f32 VR256:$lhs), (v8f32 VR256:$rhs))),
5152 (VADDSUBPSYrr VR256:$lhs, VR256:$rhs)>;
5153 def : Pat<(v8f32 (X86Addsub (v8f32 VR256:$lhs), (loadv8f32 addr:$rhs))),
5154 (VADDSUBPSYrm VR256:$lhs, f256mem:$rhs)>;
5155 def : Pat<(v4f64 (X86Addsub (v4f64 VR256:$lhs), (v4f64 VR256:$rhs))),
5156 (VADDSUBPDYrr VR256:$lhs, VR256:$rhs)>;
5157 def : Pat<(v4f64 (X86Addsub (v4f64 VR256:$lhs), (loadv4f64 addr:$rhs))),
5158 (VADDSUBPDYrm VR256:$lhs, f256mem:$rhs)>;
5161 let Predicates = [UseSSE3] in {
5162 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 VR128:$rhs))),
5163 (ADDSUBPSrr VR128:$lhs, VR128:$rhs)>;
5164 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (memopv4f32 addr:$rhs))),
5165 (ADDSUBPSrm VR128:$lhs, f128mem:$rhs)>;
5166 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 VR128:$rhs))),
5167 (ADDSUBPDrr VR128:$lhs, VR128:$rhs)>;
5168 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (memopv2f64 addr:$rhs))),
5169 (ADDSUBPDrm VR128:$lhs, f128mem:$rhs)>;
5172 //===---------------------------------------------------------------------===//
5173 // SSE3 Instructions
5174 //===---------------------------------------------------------------------===//
5177 multiclass S3D_Int<bits<8> o, string OpcodeStr, ValueType vt, RegisterClass RC,
5178 X86MemOperand x86memop, SDNode OpNode, PatFrag ld_frag,
5180 def rr : S3DI<o, MRMSrcReg, (outs RC:$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>,
5185 Sched<[WriteFHAdd]>;
5187 def rm : S3DI<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<[WriteFHAddLd, ReadAfterLd]>;
5194 multiclass S3_Int<bits<8> o, string OpcodeStr, ValueType vt, RegisterClass RC,
5195 X86MemOperand x86memop, SDNode OpNode, PatFrag ld_frag,
5197 def rr : S3I<o, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
5199 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5200 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5201 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], IIC_SSE_HADDSUB_RR>,
5202 Sched<[WriteFHAdd]>;
5204 def rm : S3I<o, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5206 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5207 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5208 [(set RC:$dst, (vt (OpNode RC:$src1, (ld_frag addr:$src2))))],
5209 IIC_SSE_HADDSUB_RM>, Sched<[WriteFHAddLd, ReadAfterLd]>;
5212 let Predicates = [HasAVX] in {
5213 let ExeDomain = SSEPackedSingle in {
5214 defm VHADDPS : S3D_Int<0x7C, "vhaddps", v4f32, VR128, f128mem,
5215 X86fhadd, loadv4f32, 0>, VEX_4V, VEX_WIG;
5216 defm VHSUBPS : S3D_Int<0x7D, "vhsubps", v4f32, VR128, f128mem,
5217 X86fhsub, loadv4f32, 0>, VEX_4V, VEX_WIG;
5218 defm VHADDPSY : S3D_Int<0x7C, "vhaddps", v8f32, VR256, f256mem,
5219 X86fhadd, loadv8f32, 0>, VEX_4V, VEX_L, VEX_WIG;
5220 defm VHSUBPSY : S3D_Int<0x7D, "vhsubps", v8f32, VR256, f256mem,
5221 X86fhsub, loadv8f32, 0>, VEX_4V, VEX_L, VEX_WIG;
5223 let ExeDomain = SSEPackedDouble in {
5224 defm VHADDPD : S3_Int <0x7C, "vhaddpd", v2f64, VR128, f128mem,
5225 X86fhadd, loadv2f64, 0>, VEX_4V, VEX_WIG;
5226 defm VHSUBPD : S3_Int <0x7D, "vhsubpd", v2f64, VR128, f128mem,
5227 X86fhsub, loadv2f64, 0>, VEX_4V, VEX_WIG;
5228 defm VHADDPDY : S3_Int <0x7C, "vhaddpd", v4f64, VR256, f256mem,
5229 X86fhadd, loadv4f64, 0>, VEX_4V, VEX_L, VEX_WIG;
5230 defm VHSUBPDY : S3_Int <0x7D, "vhsubpd", v4f64, VR256, f256mem,
5231 X86fhsub, loadv4f64, 0>, VEX_4V, VEX_L, VEX_WIG;
5235 let Constraints = "$src1 = $dst" in {
5236 let ExeDomain = SSEPackedSingle in {
5237 defm HADDPS : S3D_Int<0x7C, "haddps", v4f32, VR128, f128mem, X86fhadd,
5239 defm HSUBPS : S3D_Int<0x7D, "hsubps", v4f32, VR128, f128mem, X86fhsub,
5242 let ExeDomain = SSEPackedDouble in {
5243 defm HADDPD : S3_Int<0x7C, "haddpd", v2f64, VR128, f128mem, X86fhadd,
5245 defm HSUBPD : S3_Int<0x7D, "hsubpd", v2f64, VR128, f128mem, X86fhsub,
5250 //===---------------------------------------------------------------------===//
5251 // SSSE3 - Packed Absolute Instructions
5252 //===---------------------------------------------------------------------===//
5255 /// SS3I_unop_rm_int - Simple SSSE3 unary op whose type can be v*{i8,i16,i32}.
5256 multiclass SS3I_unop_rm<bits<8> opc, string OpcodeStr, ValueType vt,
5257 SDNode OpNode, PatFrag ld_frag> {
5258 def rr : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
5260 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5261 [(set VR128:$dst, (vt (OpNode VR128:$src)))],
5262 IIC_SSE_PABS_RR>, Sched<[WriteVecALU]>;
5264 def rm : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
5266 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5268 (vt (OpNode (bitconvert (ld_frag addr:$src)))))],
5269 IIC_SSE_PABS_RM>, Sched<[WriteVecALULd]>;
5272 /// SS3I_unop_rm_int_y - Simple SSSE3 unary op whose type can be v*{i8,i16,i32}.
5273 multiclass SS3I_unop_rm_y<bits<8> opc, string OpcodeStr, ValueType vt,
5275 def Yrr : SS38I<opc, MRMSrcReg, (outs VR256:$dst),
5277 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5278 [(set VR256:$dst, (vt (OpNode VR256:$src)))]>,
5279 Sched<[WriteVecALU]>;
5281 def Yrm : SS38I<opc, MRMSrcMem, (outs VR256:$dst),
5283 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5285 (vt (OpNode (bitconvert (loadv4i64 addr:$src)))))]>,
5286 Sched<[WriteVecALULd]>;
5289 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5290 defm VPABSB : SS3I_unop_rm<0x1C, "vpabsb", v16i8, abs, loadv2i64>, VEX, VEX_WIG;
5291 defm VPABSW : SS3I_unop_rm<0x1D, "vpabsw", v8i16, abs, loadv2i64>, VEX, VEX_WIG;
5293 let Predicates = [HasAVX, NoVLX] in {
5294 defm VPABSD : SS3I_unop_rm<0x1E, "vpabsd", v4i32, abs, loadv2i64>, VEX, VEX_WIG;
5296 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
5297 defm VPABSB : SS3I_unop_rm_y<0x1C, "vpabsb", v32i8, abs>, VEX, VEX_L, VEX_WIG;
5298 defm VPABSW : SS3I_unop_rm_y<0x1D, "vpabsw", v16i16, abs>, VEX, VEX_L, VEX_WIG;
5300 let Predicates = [HasAVX2, NoVLX] in {
5301 defm VPABSD : SS3I_unop_rm_y<0x1E, "vpabsd", v8i32, abs>, VEX, VEX_L, VEX_WIG;
5304 defm PABSB : SS3I_unop_rm<0x1C, "pabsb", v16i8, abs, memopv2i64>;
5305 defm PABSW : SS3I_unop_rm<0x1D, "pabsw", v8i16, abs, memopv2i64>;
5306 defm PABSD : SS3I_unop_rm<0x1E, "pabsd", v4i32, abs, memopv2i64>;
5308 //===---------------------------------------------------------------------===//
5309 // SSSE3 - Packed Binary Operator Instructions
5310 //===---------------------------------------------------------------------===//
5312 let Sched = WritePHAdd in {
5313 def SSE_PHADDSUBD : OpndItins<
5314 IIC_SSE_PHADDSUBD_RR, IIC_SSE_PHADDSUBD_RM
5316 def SSE_PHADDSUBSW : OpndItins<
5317 IIC_SSE_PHADDSUBSW_RR, IIC_SSE_PHADDSUBSW_RM
5319 def SSE_PHADDSUBW : OpndItins<
5320 IIC_SSE_PHADDSUBW_RR, IIC_SSE_PHADDSUBW_RM
5323 let Sched = WriteShuffle in
5324 def SSE_PSHUFB : OpndItins<
5325 IIC_SSE_PSHUFB_RR, IIC_SSE_PSHUFB_RM
5327 let Sched = WriteVecALU in
5328 def SSE_PSIGN : OpndItins<
5329 IIC_SSE_PSIGN_RR, IIC_SSE_PSIGN_RM
5331 let Sched = WriteVecIMul in
5332 def SSE_PMULHRSW : OpndItins<
5333 IIC_SSE_PMULHRSW, IIC_SSE_PMULHRSW
5336 /// SS3I_binop_rm - Simple SSSE3 bin op
5337 multiclass SS3I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
5338 ValueType DstVT, ValueType OpVT, RegisterClass RC,
5339 PatFrag memop_frag, X86MemOperand x86memop,
5340 OpndItins itins, bit Is2Addr = 1> {
5341 let isCommutable = 1 in
5342 def rr : SS38I<opc, MRMSrcReg, (outs RC:$dst),
5343 (ins RC:$src1, RC:$src2),
5345 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5346 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5347 [(set RC:$dst, (DstVT (OpNode (OpVT RC:$src1), RC:$src2)))], itins.rr>,
5348 Sched<[itins.Sched]>;
5349 def rm : SS38I<opc, MRMSrcMem, (outs RC:$dst),
5350 (ins RC:$src1, x86memop:$src2),
5352 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5353 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5355 (DstVT (OpNode (OpVT RC:$src1),
5356 (bitconvert (memop_frag addr:$src2)))))], itins.rm>,
5357 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5360 /// SS3I_binop_rm_int - Simple SSSE3 bin op whose type can be v*{i8,i16,i32}.
5361 multiclass SS3I_binop_rm_int<bits<8> opc, string OpcodeStr,
5362 Intrinsic IntId128, OpndItins itins,
5363 PatFrag ld_frag, bit Is2Addr = 1> {
5364 let isCommutable = 1 in
5365 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
5366 (ins VR128:$src1, VR128:$src2),
5368 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5369 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5370 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
5371 Sched<[itins.Sched]>;
5372 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
5373 (ins VR128:$src1, i128mem:$src2),
5375 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5376 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5378 (IntId128 VR128:$src1,
5379 (bitconvert (ld_frag addr:$src2))))]>,
5380 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5383 multiclass SS3I_binop_rm_int_y<bits<8> opc, string OpcodeStr,
5385 X86FoldableSchedWrite Sched> {
5386 let isCommutable = 1 in
5387 def rr256 : SS38I<opc, MRMSrcReg, (outs VR256:$dst),
5388 (ins VR256:$src1, VR256:$src2),
5389 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5390 [(set VR256:$dst, (IntId256 VR256:$src1, VR256:$src2))]>,
5392 def rm256 : SS38I<opc, MRMSrcMem, (outs VR256:$dst),
5393 (ins VR256:$src1, i256mem:$src2),
5394 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5396 (IntId256 VR256:$src1, (bitconvert (loadv4i64 addr:$src2))))]>,
5397 Sched<[Sched.Folded, ReadAfterLd]>;
5400 let ImmT = NoImm, Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5401 let isCommutable = 0 in {
5402 defm VPSHUFB : SS3I_binop_rm<0x00, "vpshufb", X86pshufb, v16i8, v16i8,
5403 VR128, loadv2i64, i128mem,
5404 SSE_PSHUFB, 0>, VEX_4V, VEX_WIG;
5405 defm VPMADDUBSW : SS3I_binop_rm<0x04, "vpmaddubsw", X86vpmaddubsw, v8i16,
5406 v16i8, VR128, loadv2i64, i128mem,
5407 SSE_PMADD, 0>, VEX_4V, VEX_WIG;
5409 defm VPMULHRSW : SS3I_binop_rm<0x0B, "vpmulhrsw", X86mulhrs, v8i16, v8i16,
5410 VR128, loadv2i64, i128mem,
5411 SSE_PMULHRSW, 0>, VEX_4V, VEX_WIG;
5414 let ImmT = NoImm, Predicates = [HasAVX] in {
5415 let isCommutable = 0 in {
5416 defm VPHADDW : SS3I_binop_rm<0x01, "vphaddw", X86hadd, v8i16, v8i16, VR128,
5418 SSE_PHADDSUBW, 0>, VEX_4V, VEX_WIG;
5419 defm VPHADDD : SS3I_binop_rm<0x02, "vphaddd", X86hadd, v4i32, v4i32, VR128,
5421 SSE_PHADDSUBD, 0>, VEX_4V, VEX_WIG;
5422 defm VPHSUBW : SS3I_binop_rm<0x05, "vphsubw", X86hsub, v8i16, v8i16, VR128,
5424 SSE_PHADDSUBW, 0>, VEX_4V, VEX_WIG;
5425 defm VPHSUBD : SS3I_binop_rm<0x06, "vphsubd", X86hsub, v4i32, v4i32, VR128,
5427 SSE_PHADDSUBD, 0>, VEX_4V;
5428 defm VPSIGNB : SS3I_binop_rm_int<0x08, "vpsignb",
5429 int_x86_ssse3_psign_b_128,
5430 SSE_PSIGN, loadv2i64, 0>, VEX_4V, VEX_WIG;
5431 defm VPSIGNW : SS3I_binop_rm_int<0x09, "vpsignw",
5432 int_x86_ssse3_psign_w_128,
5433 SSE_PSIGN, loadv2i64, 0>, VEX_4V, VEX_WIG;
5434 defm VPSIGND : SS3I_binop_rm_int<0x0A, "vpsignd",
5435 int_x86_ssse3_psign_d_128,
5436 SSE_PSIGN, loadv2i64, 0>, VEX_4V, VEX_WIG;
5437 defm VPHADDSW : SS3I_binop_rm_int<0x03, "vphaddsw",
5438 int_x86_ssse3_phadd_sw_128,
5439 SSE_PHADDSUBSW, loadv2i64, 0>, VEX_4V, VEX_WIG;
5440 defm VPHSUBSW : SS3I_binop_rm_int<0x07, "vphsubsw",
5441 int_x86_ssse3_phsub_sw_128,
5442 SSE_PHADDSUBSW, loadv2i64, 0>, VEX_4V, VEX_WIG;
5446 let ImmT = NoImm, Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
5447 let isCommutable = 0 in {
5448 defm VPSHUFBY : SS3I_binop_rm<0x00, "vpshufb", X86pshufb, v32i8, v32i8,
5449 VR256, loadv4i64, i256mem,
5450 SSE_PSHUFB, 0>, VEX_4V, VEX_L, VEX_WIG;
5451 defm VPMADDUBSWY : SS3I_binop_rm<0x04, "vpmaddubsw", X86vpmaddubsw, v16i16,
5452 v32i8, VR256, loadv4i64, i256mem,
5453 SSE_PMADD, 0>, VEX_4V, VEX_L, VEX_WIG;
5455 defm VPMULHRSWY : SS3I_binop_rm<0x0B, "vpmulhrsw", X86mulhrs, v16i16, v16i16,
5456 VR256, loadv4i64, i256mem,
5457 SSE_PMULHRSW, 0>, VEX_4V, VEX_L, VEX_WIG;
5460 let ImmT = NoImm, Predicates = [HasAVX2] in {
5461 let isCommutable = 0 in {
5462 defm VPHADDWY : SS3I_binop_rm<0x01, "vphaddw", X86hadd, v16i16, v16i16,
5463 VR256, loadv4i64, i256mem,
5464 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L, VEX_WIG;
5465 defm VPHADDDY : SS3I_binop_rm<0x02, "vphaddd", X86hadd, v8i32, v8i32, VR256,
5467 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L, VEX_WIG;
5468 defm VPHSUBWY : SS3I_binop_rm<0x05, "vphsubw", X86hsub, v16i16, v16i16,
5469 VR256, loadv4i64, i256mem,
5470 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L, VEX_WIG;
5471 defm VPHSUBDY : SS3I_binop_rm<0x06, "vphsubd", X86hsub, v8i32, v8i32, VR256,
5473 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5474 defm VPSIGNBY : SS3I_binop_rm_int_y<0x08, "vpsignb", int_x86_avx2_psign_b,
5475 WriteVecALU>, VEX_4V, VEX_L, VEX_WIG;
5476 defm VPSIGNWY : SS3I_binop_rm_int_y<0x09, "vpsignw", int_x86_avx2_psign_w,
5477 WriteVecALU>, VEX_4V, VEX_L, VEX_WIG;
5478 defm VPSIGNDY : SS3I_binop_rm_int_y<0x0A, "vpsignd", int_x86_avx2_psign_d,
5479 WriteVecALU>, VEX_4V, VEX_L, VEX_WIG;
5480 defm VPHADDSW : SS3I_binop_rm_int_y<0x03, "vphaddsw",
5481 int_x86_avx2_phadd_sw,
5482 WriteVecALU>, VEX_4V, VEX_L, VEX_WIG;
5483 defm VPHSUBSW : SS3I_binop_rm_int_y<0x07, "vphsubsw",
5484 int_x86_avx2_phsub_sw,
5485 WriteVecALU>, VEX_4V, VEX_L, VEX_WIG;
5489 // None of these have i8 immediate fields.
5490 let ImmT = NoImm, Constraints = "$src1 = $dst" in {
5491 let isCommutable = 0 in {
5492 defm PHADDW : SS3I_binop_rm<0x01, "phaddw", X86hadd, v8i16, v8i16, VR128,
5493 memopv2i64, i128mem, SSE_PHADDSUBW>;
5494 defm PHADDD : SS3I_binop_rm<0x02, "phaddd", X86hadd, v4i32, v4i32, VR128,
5495 memopv2i64, i128mem, SSE_PHADDSUBD>;
5496 defm PHSUBW : SS3I_binop_rm<0x05, "phsubw", X86hsub, v8i16, v8i16, VR128,
5497 memopv2i64, i128mem, SSE_PHADDSUBW>;
5498 defm PHSUBD : SS3I_binop_rm<0x06, "phsubd", X86hsub, v4i32, v4i32, VR128,
5499 memopv2i64, i128mem, SSE_PHADDSUBD>;
5500 defm PSIGNB : SS3I_binop_rm_int<0x08, "psignb", int_x86_ssse3_psign_b_128,
5501 SSE_PSIGN, memopv2i64>;
5502 defm PSIGNW : SS3I_binop_rm_int<0x09, "psignw", int_x86_ssse3_psign_w_128,
5503 SSE_PSIGN, memopv2i64>;
5504 defm PSIGND : SS3I_binop_rm_int<0x0A, "psignd", int_x86_ssse3_psign_d_128,
5505 SSE_PSIGN, memopv2i64>;
5506 defm PSHUFB : SS3I_binop_rm<0x00, "pshufb", X86pshufb, v16i8, v16i8, VR128,
5507 memopv2i64, i128mem, SSE_PSHUFB>;
5508 defm PHADDSW : SS3I_binop_rm_int<0x03, "phaddsw",
5509 int_x86_ssse3_phadd_sw_128,
5510 SSE_PHADDSUBSW, memopv2i64>;
5511 defm PHSUBSW : SS3I_binop_rm_int<0x07, "phsubsw",
5512 int_x86_ssse3_phsub_sw_128,
5513 SSE_PHADDSUBSW, memopv2i64>;
5514 defm PMADDUBSW : SS3I_binop_rm<0x04, "pmaddubsw", X86vpmaddubsw, v8i16,
5515 v16i8, VR128, memopv2i64, i128mem,
5518 defm PMULHRSW : SS3I_binop_rm<0x0B, "pmulhrsw", X86mulhrs, v8i16, v8i16,
5519 VR128, memopv2i64, i128mem, SSE_PMULHRSW>;
5522 //===---------------------------------------------------------------------===//
5523 // SSSE3 - Packed Align Instruction Patterns
5524 //===---------------------------------------------------------------------===//
5526 multiclass ssse3_palignr<string asm, bit Is2Addr = 1> {
5527 let hasSideEffects = 0 in {
5528 def rri : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
5529 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
5531 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
5533 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
5534 [], IIC_SSE_PALIGNRR>, Sched<[WriteShuffle]>;
5536 def rmi : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
5537 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
5539 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
5541 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
5542 [], IIC_SSE_PALIGNRM>, Sched<[WriteShuffleLd, ReadAfterLd]>;
5546 multiclass ssse3_palignr_y<string asm, bit Is2Addr = 1> {
5547 let hasSideEffects = 0 in {
5548 def Yrri : SS3AI<0x0F, MRMSrcReg, (outs VR256:$dst),
5549 (ins VR256:$src1, VR256:$src2, u8imm:$src3),
5551 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
5552 []>, Sched<[WriteShuffle]>;
5554 def Yrmi : SS3AI<0x0F, MRMSrcMem, (outs VR256:$dst),
5555 (ins VR256:$src1, i256mem:$src2, u8imm:$src3),
5557 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
5558 []>, Sched<[WriteShuffleLd, ReadAfterLd]>;
5562 let Predicates = [HasAVX] in
5563 defm VPALIGNR : ssse3_palignr<"vpalignr", 0>, VEX_4V, VEX_WIG;
5564 let Predicates = [HasAVX2] in
5565 defm VPALIGNR : ssse3_palignr_y<"vpalignr", 0>, VEX_4V, VEX_L, VEX_WIG;
5566 let Constraints = "$src1 = $dst", Predicates = [UseSSSE3] in
5567 defm PALIGNR : ssse3_palignr<"palignr">;
5569 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
5570 def : Pat<(v8i32 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5571 (VPALIGNRYrri VR256:$src1, VR256:$src2, imm:$imm)>;
5572 def : Pat<(v8f32 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5573 (VPALIGNRYrri VR256:$src1, VR256:$src2, imm:$imm)>;
5574 def : Pat<(v16i16 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5575 (VPALIGNRYrri VR256:$src1, VR256:$src2, imm:$imm)>;
5576 def : Pat<(v32i8 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5577 (VPALIGNRYrri VR256:$src1, VR256:$src2, imm:$imm)>;
5580 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5581 def : Pat<(v4i32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5582 (VPALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5583 def : Pat<(v4f32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5584 (VPALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5585 def : Pat<(v8i16 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5586 (VPALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5587 def : Pat<(v16i8 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5588 (VPALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5591 let Predicates = [UseSSSE3] in {
5592 def : Pat<(v4i32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5593 (PALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5594 def : Pat<(v4f32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5595 (PALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5596 def : Pat<(v8i16 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5597 (PALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5598 def : Pat<(v16i8 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5599 (PALIGNRrri VR128:$src1, VR128:$src2, imm:$imm)>;
5602 //===---------------------------------------------------------------------===//
5603 // SSSE3 - Thread synchronization
5604 //===---------------------------------------------------------------------===//
5606 let SchedRW = [WriteSystem] in {
5607 let usesCustomInserter = 1 in {
5608 def MONITOR : PseudoI<(outs), (ins i32mem:$src1, GR32:$src2, GR32:$src3),
5609 [(int_x86_sse3_monitor addr:$src1, GR32:$src2, GR32:$src3)]>,
5610 Requires<[HasSSE3]>;
5613 let Uses = [EAX, ECX, EDX] in
5614 def MONITORrrr : I<0x01, MRM_C8, (outs), (ins), "monitor", [], IIC_SSE_MONITOR>,
5615 TB, Requires<[HasSSE3]>;
5617 let Uses = [ECX, EAX] in
5618 def MWAITrr : I<0x01, MRM_C9, (outs), (ins), "mwait",
5619 [(int_x86_sse3_mwait ECX, EAX)], IIC_SSE_MWAIT>,
5620 TB, Requires<[HasSSE3]>;
5623 def : InstAlias<"mwait\t{%eax, %ecx|ecx, eax}", (MWAITrr)>, Requires<[Not64BitMode]>;
5624 def : InstAlias<"mwait\t{%rax, %rcx|rcx, rax}", (MWAITrr)>, Requires<[In64BitMode]>;
5626 def : InstAlias<"monitor\t{%eax, %ecx, %edx|edx, ecx, eax}", (MONITORrrr)>,
5627 Requires<[Not64BitMode]>;
5628 def : InstAlias<"monitor\t{%rax, %rcx, %rdx|rdx, rcx, rax}", (MONITORrrr)>,
5629 Requires<[In64BitMode]>;
5631 //===----------------------------------------------------------------------===//
5632 // SSE4.1 - Packed Move with Sign/Zero Extend
5633 //===----------------------------------------------------------------------===//
5635 multiclass SS41I_pmovx_rrrm<bits<8> opc, string OpcodeStr, X86MemOperand MemOp,
5636 RegisterClass OutRC, RegisterClass InRC,
5638 def rr : SS48I<opc, MRMSrcReg, (outs OutRC:$dst), (ins InRC:$src),
5639 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5641 Sched<[itins.Sched]>;
5643 def rm : SS48I<opc, MRMSrcMem, (outs OutRC:$dst), (ins MemOp:$src),
5644 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5646 itins.rm>, Sched<[itins.Sched.Folded]>;
5649 multiclass SS41I_pmovx_rm_all<bits<8> opc, string OpcodeStr,
5650 X86MemOperand MemOp, X86MemOperand MemYOp,
5651 OpndItins SSEItins, OpndItins AVXItins,
5652 OpndItins AVX2Itins, Predicate prd> {
5653 defm NAME : SS41I_pmovx_rrrm<opc, OpcodeStr, MemOp, VR128, VR128, SSEItins>;
5654 let Predicates = [HasAVX, prd] in
5655 defm V#NAME : SS41I_pmovx_rrrm<opc, !strconcat("v", OpcodeStr), MemOp,
5656 VR128, VR128, AVXItins>, VEX, VEX_WIG;
5657 let Predicates = [HasAVX2, prd] in
5658 defm V#NAME#Y : SS41I_pmovx_rrrm<opc, !strconcat("v", OpcodeStr), MemYOp,
5659 VR256, VR128, AVX2Itins>, VEX, VEX_L, VEX_WIG;
5662 multiclass SS41I_pmovx_rm<bits<8> opc, string OpcodeStr, X86MemOperand MemOp,
5663 X86MemOperand MemYOp, Predicate prd> {
5664 defm PMOVSX#NAME : SS41I_pmovx_rm_all<opc, !strconcat("pmovsx", OpcodeStr),
5666 SSE_INTALU_ITINS_SHUFF_P,
5667 DEFAULT_ITINS_SHUFFLESCHED,
5668 DEFAULT_ITINS_SHUFFLESCHED, prd>;
5669 defm PMOVZX#NAME : SS41I_pmovx_rm_all<!add(opc, 0x10),
5670 !strconcat("pmovzx", OpcodeStr),
5672 SSE_INTALU_ITINS_SHUFF_P,
5673 DEFAULT_ITINS_SHUFFLESCHED,
5674 DEFAULT_ITINS_SHUFFLESCHED, prd>;
5677 defm BW : SS41I_pmovx_rm<0x20, "bw", i64mem, i128mem, NoVLX_Or_NoBWI>;
5678 defm WD : SS41I_pmovx_rm<0x23, "wd", i64mem, i128mem, NoVLX>;
5679 defm DQ : SS41I_pmovx_rm<0x25, "dq", i64mem, i128mem, NoVLX>;
5681 defm BD : SS41I_pmovx_rm<0x21, "bd", i32mem, i64mem, NoVLX>;
5682 defm WQ : SS41I_pmovx_rm<0x24, "wq", i32mem, i64mem, NoVLX>;
5684 defm BQ : SS41I_pmovx_rm<0x22, "bq", i16mem, i32mem, NoVLX>;
5687 multiclass SS41I_pmovx_avx2_patterns<string OpcPrefix, string ExtTy, SDNode ExtOp> {
5688 // Register-Register patterns
5689 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5690 def : Pat<(v16i16 (ExtOp (v16i8 VR128:$src))),
5691 (!cast<I>(OpcPrefix#BWYrr) VR128:$src)>;
5693 let Predicates = [HasAVX, NoVLX] in {
5694 def : Pat<(v8i32 (ExtOp (v16i8 VR128:$src))),
5695 (!cast<I>(OpcPrefix#BDYrr) VR128:$src)>;
5696 def : Pat<(v4i64 (ExtOp (v16i8 VR128:$src))),
5697 (!cast<I>(OpcPrefix#BQYrr) VR128:$src)>;
5699 def : Pat<(v8i32 (ExtOp (v8i16 VR128:$src))),
5700 (!cast<I>(OpcPrefix#WDYrr) VR128:$src)>;
5701 def : Pat<(v4i64 (ExtOp (v8i16 VR128:$src))),
5702 (!cast<I>(OpcPrefix#WQYrr) VR128:$src)>;
5704 def : Pat<(v4i64 (ExtOp (v4i32 VR128:$src))),
5705 (!cast<I>(OpcPrefix#DQYrr) VR128:$src)>;
5708 // Simple Register-Memory patterns
5709 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5710 def : Pat<(v16i16 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5711 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5713 let Predicates = [HasAVX, NoVLX] in {
5714 def : Pat<(v8i32 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5715 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5716 def : Pat<(v4i64 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5717 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5719 def : Pat<(v8i32 (!cast<PatFrag>(ExtTy#"extloadvi16") addr:$src)),
5720 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5721 def : Pat<(v4i64 (!cast<PatFrag>(ExtTy#"extloadvi16") addr:$src)),
5722 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5724 def : Pat<(v4i64 (!cast<PatFrag>(ExtTy#"extloadvi32") addr:$src)),
5725 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5728 // AVX2 Register-Memory patterns
5729 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5730 def : Pat<(v16i16 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5731 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5732 def : Pat<(v16i16 (ExtOp (v16i8 (vzmovl_v2i64 addr:$src)))),
5733 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5734 def : Pat<(v16i16 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5735 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5737 let Predicates = [HasAVX, NoVLX] in {
5738 def : Pat<(v8i32 (ExtOp (bc_v16i8 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
5739 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5740 def : Pat<(v8i32 (ExtOp (v16i8 (vzmovl_v2i64 addr:$src)))),
5741 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5742 def : Pat<(v8i32 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5743 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5744 def : Pat<(v8i32 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5745 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5747 def : Pat<(v4i64 (ExtOp (bc_v16i8 (v4i32 (scalar_to_vector (loadi32 addr:$src)))))),
5748 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5749 def : Pat<(v4i64 (ExtOp (v16i8 (vzmovl_v4i32 addr:$src)))),
5750 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5751 def : Pat<(v4i64 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5752 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5753 def : Pat<(v4i64 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5754 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5756 def : Pat<(v8i32 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
5757 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5758 def : Pat<(v8i32 (ExtOp (v8i16 (vzmovl_v2i64 addr:$src)))),
5759 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5760 def : Pat<(v8i32 (ExtOp (v8i16 (vzload_v2i64 addr:$src)))),
5761 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5763 def : Pat<(v4i64 (ExtOp (bc_v8i16 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
5764 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5765 def : Pat<(v4i64 (ExtOp (v8i16 (vzmovl_v2i64 addr:$src)))),
5766 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5767 def : Pat<(v4i64 (ExtOp (v8i16 (vzload_v2i64 addr:$src)))),
5768 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5769 def : Pat<(v4i64 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
5770 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5772 def : Pat<(v4i64 (ExtOp (bc_v4i32 (loadv2i64 addr:$src)))),
5773 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5774 def : Pat<(v4i64 (ExtOp (v4i32 (vzmovl_v2i64 addr:$src)))),
5775 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5776 def : Pat<(v4i64 (ExtOp (v4i32 (vzload_v2i64 addr:$src)))),
5777 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5781 defm : SS41I_pmovx_avx2_patterns<"VPMOVSX", "s", X86vsext>;
5782 defm : SS41I_pmovx_avx2_patterns<"VPMOVZX", "z", X86vzext>;
5784 // SSE4.1/AVX patterns.
5785 multiclass SS41I_pmovx_patterns<string OpcPrefix, string ExtTy,
5786 SDNode ExtOp, PatFrag ExtLoad16> {
5787 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5788 def : Pat<(v8i16 (ExtOp (v16i8 VR128:$src))),
5789 (!cast<I>(OpcPrefix#BWrr) VR128:$src)>;
5791 let Predicates = [HasAVX, NoVLX] in {
5792 def : Pat<(v4i32 (ExtOp (v16i8 VR128:$src))),
5793 (!cast<I>(OpcPrefix#BDrr) VR128:$src)>;
5794 def : Pat<(v2i64 (ExtOp (v16i8 VR128:$src))),
5795 (!cast<I>(OpcPrefix#BQrr) VR128:$src)>;
5797 def : Pat<(v4i32 (ExtOp (v8i16 VR128:$src))),
5798 (!cast<I>(OpcPrefix#WDrr) VR128:$src)>;
5799 def : Pat<(v2i64 (ExtOp (v8i16 VR128:$src))),
5800 (!cast<I>(OpcPrefix#WQrr) VR128:$src)>;
5802 def : Pat<(v2i64 (ExtOp (v4i32 VR128:$src))),
5803 (!cast<I>(OpcPrefix#DQrr) VR128:$src)>;
5805 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5806 def : Pat<(v8i16 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5807 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
5809 let Predicates = [HasAVX, NoVLX] in {
5810 def : Pat<(v4i32 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5811 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
5812 def : Pat<(v2i64 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5813 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
5815 def : Pat<(v4i32 (!cast<PatFrag>(ExtTy#"extloadvi16") addr:$src)),
5816 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
5817 def : Pat<(v2i64 (!cast<PatFrag>(ExtTy#"extloadvi16") addr:$src)),
5818 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
5820 def : Pat<(v2i64 (!cast<PatFrag>(ExtTy#"extloadvi32") addr:$src)),
5821 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
5823 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
5824 def : Pat<(v8i16 (ExtOp (bc_v16i8 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
5825 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
5826 def : Pat<(v8i16 (ExtOp (bc_v16i8 (v2f64 (scalar_to_vector (loadf64 addr:$src)))))),
5827 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
5828 def : Pat<(v8i16 (ExtOp (v16i8 (vzmovl_v2i64 addr:$src)))),
5829 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
5830 def : Pat<(v8i16 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5831 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
5832 def : Pat<(v8i16 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5833 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
5835 let Predicates = [HasAVX, NoVLX] in {
5836 def : Pat<(v4i32 (ExtOp (bc_v16i8 (v4i32 (scalar_to_vector (loadi32 addr:$src)))))),
5837 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
5838 def : Pat<(v4i32 (ExtOp (v16i8 (vzmovl_v4i32 addr:$src)))),
5839 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
5840 def : Pat<(v4i32 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5841 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
5842 def : Pat<(v4i32 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5843 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
5845 def : Pat<(v2i64 (ExtOp (bc_v16i8 (v4i32 (scalar_to_vector (ExtLoad16 addr:$src)))))),
5846 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
5847 def : Pat<(v2i64 (ExtOp (v16i8 (vzmovl_v4i32 addr:$src)))),
5848 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
5849 def : Pat<(v2i64 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5850 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
5851 def : Pat<(v2i64 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5852 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
5854 def : Pat<(v4i32 (ExtOp (bc_v8i16 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
5855 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
5856 def : Pat<(v4i32 (ExtOp (bc_v8i16 (v2f64 (scalar_to_vector (loadf64 addr:$src)))))),
5857 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
5858 def : Pat<(v4i32 (ExtOp (v8i16 (vzmovl_v2i64 addr:$src)))),
5859 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
5860 def : Pat<(v4i32 (ExtOp (v8i16 (vzload_v2i64 addr:$src)))),
5861 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
5862 def : Pat<(v4i32 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
5863 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
5865 def : Pat<(v2i64 (ExtOp (bc_v8i16 (v4i32 (scalar_to_vector (loadi32 addr:$src)))))),
5866 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
5867 def : Pat<(v2i64 (ExtOp (v8i16 (vzmovl_v4i32 addr:$src)))),
5868 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
5869 def : Pat<(v2i64 (ExtOp (v8i16 (vzload_v2i64 addr:$src)))),
5870 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
5871 def : Pat<(v2i64 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
5872 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
5874 def : Pat<(v2i64 (ExtOp (bc_v4i32 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
5875 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
5876 def : Pat<(v2i64 (ExtOp (bc_v4i32 (v2f64 (scalar_to_vector (loadf64 addr:$src)))))),
5877 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
5878 def : Pat<(v2i64 (ExtOp (v4i32 (vzmovl_v2i64 addr:$src)))),
5879 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
5880 def : Pat<(v2i64 (ExtOp (v4i32 (vzload_v2i64 addr:$src)))),
5881 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
5882 def : Pat<(v2i64 (ExtOp (bc_v4i32 (loadv2i64 addr:$src)))),
5883 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
5887 defm : SS41I_pmovx_patterns<"VPMOVSX", "s", sext_invec, extloadi32i16>;
5888 defm : SS41I_pmovx_patterns<"VPMOVZX", "z", zext_invec, loadi16_anyext>;
5890 let Predicates = [UseSSE41] in {
5891 defm : SS41I_pmovx_patterns<"PMOVSX", "s", sext_invec, extloadi32i16>;
5892 defm : SS41I_pmovx_patterns<"PMOVZX", "z", zext_invec, loadi16_anyext>;
5895 //===----------------------------------------------------------------------===//
5896 // SSE4.1 - Extract Instructions
5897 //===----------------------------------------------------------------------===//
5899 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
5900 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
5901 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
5902 (ins VR128:$src1, u8imm:$src2),
5903 !strconcat(OpcodeStr,
5904 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5905 [(set GR32orGR64:$dst, (X86pextrb (v16i8 VR128:$src1),
5907 Sched<[WriteShuffle]>;
5908 let hasSideEffects = 0, mayStore = 1,
5909 SchedRW = [WriteShuffleLd, WriteRMW] in
5910 def mr : SS4AIi8<opc, MRMDestMem, (outs),
5911 (ins i8mem:$dst, VR128:$src1, u8imm:$src2),
5912 !strconcat(OpcodeStr,
5913 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5914 [(store (i8 (trunc (assertzext (X86pextrb (v16i8 VR128:$src1),
5915 imm:$src2)))), addr:$dst)]>;
5918 let Predicates = [HasAVX, NoBWI] in
5919 defm VPEXTRB : SS41I_extract8<0x14, "vpextrb">, VEX;
5921 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
5924 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
5925 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
5926 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in
5927 def rr_REV : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
5928 (ins VR128:$src1, u8imm:$src2),
5929 !strconcat(OpcodeStr,
5930 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5931 []>, Sched<[WriteShuffle]>, FoldGenData<NAME#ri>;
5933 let hasSideEffects = 0, mayStore = 1,
5934 SchedRW = [WriteShuffleLd, WriteRMW] in
5935 def mr : SS4AIi8<opc, MRMDestMem, (outs),
5936 (ins i16mem:$dst, VR128:$src1, u8imm:$src2),
5937 !strconcat(OpcodeStr,
5938 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5939 [(store (i16 (trunc (assertzext (X86pextrw (v8i16 VR128:$src1),
5940 imm:$src2)))), addr:$dst)]>;
5943 let Predicates = [HasAVX, NoBWI] in
5944 defm VPEXTRW : SS41I_extract16<0x15, "vpextrw">, VEX;
5946 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
5949 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
5950 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
5951 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
5952 (ins VR128:$src1, u8imm:$src2),
5953 !strconcat(OpcodeStr,
5954 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5956 (extractelt (v4i32 VR128:$src1), imm:$src2))]>,
5957 Sched<[WriteShuffle]>;
5958 let SchedRW = [WriteShuffleLd, WriteRMW] in
5959 def mr : SS4AIi8<opc, MRMDestMem, (outs),
5960 (ins i32mem:$dst, VR128:$src1, u8imm:$src2),
5961 !strconcat(OpcodeStr,
5962 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5963 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
5967 let Predicates = [HasAVX, NoDQI] in
5968 defm VPEXTRD : SS41I_extract32<0x16, "vpextrd">, VEX;
5970 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
5972 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
5973 multiclass SS41I_extract64<bits<8> opc, string OpcodeStr> {
5974 def rr : SS4AIi8<opc, MRMDestReg, (outs GR64:$dst),
5975 (ins VR128:$src1, u8imm:$src2),
5976 !strconcat(OpcodeStr,
5977 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5979 (extractelt (v2i64 VR128:$src1), imm:$src2))]>,
5980 Sched<[WriteShuffle]>;
5981 let SchedRW = [WriteShuffleLd, WriteRMW] in
5982 def mr : SS4AIi8<opc, MRMDestMem, (outs),
5983 (ins i64mem:$dst, VR128:$src1, u8imm:$src2),
5984 !strconcat(OpcodeStr,
5985 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5986 [(store (extractelt (v2i64 VR128:$src1), imm:$src2),
5990 let Predicates = [HasAVX, NoDQI] in
5991 defm VPEXTRQ : SS41I_extract64<0x16, "vpextrq">, VEX, VEX_W;
5993 defm PEXTRQ : SS41I_extract64<0x16, "pextrq">, REX_W;
5995 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
5997 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr,
5998 OpndItins itins = DEFAULT_ITINS> {
5999 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
6000 (ins VR128:$src1, u8imm:$src2),
6001 !strconcat(OpcodeStr,
6002 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6003 [(set GR32orGR64:$dst,
6004 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))],
6005 itins.rr>, Sched<[WriteFBlend]>;
6006 let SchedRW = [WriteFBlendLd, WriteRMW] in
6007 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6008 (ins f32mem:$dst, VR128:$src1, u8imm:$src2),
6009 !strconcat(OpcodeStr,
6010 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6011 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
6012 addr:$dst)], itins.rm>;
6015 let ExeDomain = SSEPackedSingle in {
6016 let Predicates = [UseAVX] in
6017 defm VEXTRACTPS : SS41I_extractf32<0x17, "vextractps">, VEX, VEX_WIG;
6018 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps", SSE_EXTRACT_ITINS>;
6021 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
6022 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
6025 (VEXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
6027 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
6030 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
6031 Requires<[UseSSE41]>;
6033 //===----------------------------------------------------------------------===//
6034 // SSE4.1 - Insert Instructions
6035 //===----------------------------------------------------------------------===//
6037 multiclass SS41I_insert8<bits<8> opc, string asm, bit Is2Addr = 1> {
6038 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6039 (ins VR128:$src1, GR32orGR64:$src2, u8imm:$src3),
6041 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6043 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6045 (X86pinsrb VR128:$src1, GR32orGR64:$src2, imm:$src3))]>,
6046 Sched<[WriteShuffle]>;
6047 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6048 (ins VR128:$src1, i8mem:$src2, u8imm:$src3),
6050 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6052 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6054 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
6055 imm:$src3))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6058 let Predicates = [HasAVX, NoBWI] in
6059 defm VPINSRB : SS41I_insert8<0x20, "vpinsrb", 0>, VEX_4V;
6060 let Constraints = "$src1 = $dst" in
6061 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
6063 multiclass SS41I_insert32<bits<8> opc, string asm, bit Is2Addr = 1> {
6064 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6065 (ins VR128:$src1, GR32:$src2, u8imm:$src3),
6067 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6069 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6071 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
6072 Sched<[WriteShuffle]>;
6073 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6074 (ins VR128:$src1, i32mem:$src2, u8imm:$src3),
6076 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6078 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6080 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
6081 imm:$src3)))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6084 let Predicates = [HasAVX, NoDQI] in
6085 defm VPINSRD : SS41I_insert32<0x22, "vpinsrd", 0>, VEX_4V;
6086 let Constraints = "$src1 = $dst" in
6087 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
6089 multiclass SS41I_insert64<bits<8> opc, string asm, bit Is2Addr = 1> {
6090 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6091 (ins VR128:$src1, GR64:$src2, u8imm:$src3),
6093 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6095 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6097 (v2i64 (insertelt VR128:$src1, GR64:$src2, imm:$src3)))]>,
6098 Sched<[WriteShuffle]>;
6099 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6100 (ins VR128:$src1, i64mem:$src2, u8imm:$src3),
6102 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6104 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6106 (v2i64 (insertelt VR128:$src1, (loadi64 addr:$src2),
6107 imm:$src3)))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6110 let Predicates = [HasAVX, NoDQI] in
6111 defm VPINSRQ : SS41I_insert64<0x22, "vpinsrq", 0>, VEX_4V, VEX_W;
6112 let Constraints = "$src1 = $dst" in
6113 defm PINSRQ : SS41I_insert64<0x22, "pinsrq">, REX_W;
6115 // insertps has a few different modes, there's the first two here below which
6116 // are optimized inserts that won't zero arbitrary elements in the destination
6117 // vector. The next one matches the intrinsic and could zero arbitrary elements
6118 // in the target vector.
6119 multiclass SS41I_insertf32<bits<8> opc, string asm, bit Is2Addr = 1,
6120 OpndItins itins = DEFAULT_ITINS> {
6121 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6122 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
6124 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6126 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6128 (X86insertps VR128:$src1, VR128:$src2, imm:$src3))], itins.rr>,
6129 Sched<[WriteFShuffle]>;
6130 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6131 (ins VR128:$src1, f32mem:$src2, u8imm:$src3),
6133 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6135 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6137 (X86insertps VR128:$src1,
6138 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
6139 imm:$src3))], itins.rm>,
6140 Sched<[WriteFShuffleLd, ReadAfterLd]>;
6143 let ExeDomain = SSEPackedSingle in {
6144 let Predicates = [UseAVX] in
6145 defm VINSERTPS : SS41I_insertf32<0x21, "vinsertps", 0>, VEX_4V, VEX_WIG;
6146 let Constraints = "$src1 = $dst" in
6147 defm INSERTPS : SS41I_insertf32<0x21, "insertps", 1, SSE_INSERT_ITINS>;
6150 let Predicates = [UseSSE41] in {
6151 // If we're inserting an element from a load or a null pshuf of a load,
6152 // fold the load into the insertps instruction.
6153 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd (v4f32
6154 (scalar_to_vector (loadf32 addr:$src2))), (i8 0)),
6156 (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6157 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd
6158 (loadv4f32 addr:$src2), (i8 0)), imm:$src3)),
6159 (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6162 let Predicates = [UseAVX] in {
6163 // If we're inserting an element from a vbroadcast of a load, fold the
6164 // load into the X86insertps instruction.
6165 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
6166 (X86VBroadcast (loadf32 addr:$src2)), imm:$src3)),
6167 (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6168 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
6169 (X86VBroadcast (loadv4f32 addr:$src2)), imm:$src3)),
6170 (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6173 //===----------------------------------------------------------------------===//
6174 // SSE4.1 - Round Instructions
6175 //===----------------------------------------------------------------------===//
6177 multiclass sse41_fp_unop_p<bits<8> opcps, bits<8> opcpd, string OpcodeStr,
6178 X86MemOperand x86memop, RegisterClass RC,
6179 PatFrag mem_frag32, PatFrag mem_frag64,
6180 Intrinsic V4F32Int, Intrinsic V2F64Int> {
6181 let ExeDomain = SSEPackedSingle in {
6182 // Intrinsic operation, reg.
6183 // Vector intrinsic operation, reg
6184 def PSr : SS4AIi8<opcps, MRMSrcReg,
6185 (outs RC:$dst), (ins RC:$src1, i32u8imm:$src2),
6186 !strconcat(OpcodeStr,
6187 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6188 [(set RC:$dst, (V4F32Int RC:$src1, imm:$src2))],
6189 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAdd]>;
6191 // Vector intrinsic operation, mem
6192 def PSm : SS4AIi8<opcps, MRMSrcMem,
6193 (outs RC:$dst), (ins x86memop:$src1, i32u8imm:$src2),
6194 !strconcat(OpcodeStr,
6195 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6197 (V4F32Int (mem_frag32 addr:$src1),imm:$src2))],
6198 IIC_SSE_ROUNDPS_MEM>, Sched<[WriteFAddLd]>;
6199 } // ExeDomain = SSEPackedSingle
6201 let ExeDomain = SSEPackedDouble in {
6202 // Vector intrinsic operation, reg
6203 def PDr : SS4AIi8<opcpd, MRMSrcReg,
6204 (outs RC:$dst), (ins RC:$src1, i32u8imm:$src2),
6205 !strconcat(OpcodeStr,
6206 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6207 [(set RC:$dst, (V2F64Int RC:$src1, imm:$src2))],
6208 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAdd]>;
6210 // Vector intrinsic operation, mem
6211 def PDm : SS4AIi8<opcpd, MRMSrcMem,
6212 (outs RC:$dst), (ins x86memop:$src1, i32u8imm:$src2),
6213 !strconcat(OpcodeStr,
6214 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6216 (V2F64Int (mem_frag64 addr:$src1),imm:$src2))],
6217 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAddLd]>;
6218 } // ExeDomain = SSEPackedDouble
6221 multiclass avx_fp_unop_rm<bits<8> opcss, bits<8> opcsd,
6223 let ExeDomain = GenericDomain, hasSideEffects = 0 in {
6224 def SSr : SS4AIi8<opcss, MRMSrcReg,
6225 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2, i32u8imm:$src3),
6226 !strconcat(OpcodeStr,
6227 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
6228 []>, Sched<[WriteFAdd]>;
6231 def SSm : SS4AIi8<opcss, MRMSrcMem,
6232 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2, i32u8imm:$src3),
6233 !strconcat(OpcodeStr,
6234 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
6235 []>, Sched<[WriteFAddLd, ReadAfterLd]>;
6237 def SDr : SS4AIi8<opcsd, MRMSrcReg,
6238 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2, i32u8imm:$src3),
6239 !strconcat(OpcodeStr,
6240 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
6241 []>, Sched<[WriteFAdd]>;
6244 def SDm : SS4AIi8<opcsd, MRMSrcMem,
6245 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2, i32u8imm:$src3),
6246 !strconcat(OpcodeStr,
6247 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
6248 []>, Sched<[WriteFAddLd, ReadAfterLd]>;
6249 } // ExeDomain = GenericDomain, hasSideEffects = 0
6252 multiclass sse41_fp_unop_s<bits<8> opcss, bits<8> opcsd,
6254 let ExeDomain = GenericDomain, hasSideEffects = 0 in {
6255 def SSr : SS4AIi8<opcss, MRMSrcReg,
6256 (outs FR32:$dst), (ins FR32:$src1, i32u8imm:$src2),
6257 !strconcat(OpcodeStr,
6258 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6259 []>, Sched<[WriteFAdd]>;
6262 def SSm : SS4AIi8<opcss, MRMSrcMem,
6263 (outs FR32:$dst), (ins f32mem:$src1, i32u8imm:$src2),
6264 !strconcat(OpcodeStr,
6265 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6266 []>, Sched<[WriteFAddLd, ReadAfterLd]>;
6268 def SDr : SS4AIi8<opcsd, MRMSrcReg,
6269 (outs FR64:$dst), (ins FR64:$src1, i32u8imm:$src2),
6270 !strconcat(OpcodeStr,
6271 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6272 []>, Sched<[WriteFAdd]>;
6275 def SDm : SS4AIi8<opcsd, MRMSrcMem,
6276 (outs FR64:$dst), (ins f64mem:$src1, i32u8imm:$src2),
6277 !strconcat(OpcodeStr,
6278 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6279 []>, Sched<[WriteFAddLd, ReadAfterLd]>;
6280 } // ExeDomain = GenericDomain, hasSideEffects = 0
6283 multiclass sse41_fp_binop_s<bits<8> opcss, bits<8> opcsd,
6286 Intrinsic F64Int, bit Is2Addr = 1> {
6287 let ExeDomain = GenericDomain, isCodeGenOnly = 1 in {
6288 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
6289 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32u8imm:$src3),
6291 !strconcat(OpcodeStr,
6292 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6293 !strconcat(OpcodeStr,
6294 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6295 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
6298 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
6299 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2, i32u8imm:$src3),
6301 !strconcat(OpcodeStr,
6302 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6303 !strconcat(OpcodeStr,
6304 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6306 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
6307 Sched<[WriteFAddLd, ReadAfterLd]>;
6309 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
6310 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32u8imm:$src3),
6312 !strconcat(OpcodeStr,
6313 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6314 !strconcat(OpcodeStr,
6315 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6316 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
6319 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
6320 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2, i32u8imm:$src3),
6322 !strconcat(OpcodeStr,
6323 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6324 !strconcat(OpcodeStr,
6325 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6327 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
6328 Sched<[WriteFAddLd, ReadAfterLd]>;
6329 } // ExeDomain = GenericDomain, isCodeGenOnly = 1
6332 // FP round - roundss, roundps, roundsd, roundpd
6333 let Predicates = [HasAVX] in {
6335 defm VROUND : sse41_fp_unop_p<0x08, 0x09, "vround", f128mem, VR128,
6336 loadv4f32, loadv2f64,
6337 int_x86_sse41_round_ps,
6338 int_x86_sse41_round_pd>, VEX, VEX_WIG;
6339 defm VROUNDY : sse41_fp_unop_p<0x08, 0x09, "vround", f256mem, VR256,
6340 loadv8f32, loadv4f64,
6341 int_x86_avx_round_ps_256,
6342 int_x86_avx_round_pd_256>, VEX, VEX_L, VEX_WIG;
6343 defm VROUND : sse41_fp_binop_s<0x0A, 0x0B, "vround",
6344 int_x86_sse41_round_ss,
6345 int_x86_sse41_round_sd, 0>, VEX_4V, VEX_LIG, VEX_WIG;
6346 defm VROUND : avx_fp_unop_rm<0x0A, 0x0B, "vround">, VEX_4V, VEX_LIG;
6349 let Predicates = [UseAVX] in {
6350 def : Pat<(ffloor FR32:$src),
6351 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x9))>;
6352 def : Pat<(f64 (ffloor FR64:$src)),
6353 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x9))>;
6354 def : Pat<(f32 (fnearbyint FR32:$src)),
6355 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0xC))>;
6356 def : Pat<(f64 (fnearbyint FR64:$src)),
6357 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0xC))>;
6358 def : Pat<(f32 (fceil FR32:$src)),
6359 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0xA))>;
6360 def : Pat<(f64 (fceil FR64:$src)),
6361 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0xA))>;
6362 def : Pat<(f32 (frint FR32:$src)),
6363 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x4))>;
6364 def : Pat<(f64 (frint FR64:$src)),
6365 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x4))>;
6366 def : Pat<(f32 (ftrunc FR32:$src)),
6367 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0xB))>;
6368 def : Pat<(f64 (ftrunc FR64:$src)),
6369 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0xB))>;
6372 let Predicates = [HasAVX] in {
6373 def : Pat<(v4f32 (ffloor VR128:$src)),
6374 (VROUNDPSr VR128:$src, (i32 0x9))>;
6375 def : Pat<(v4f32 (fnearbyint VR128:$src)),
6376 (VROUNDPSr VR128:$src, (i32 0xC))>;
6377 def : Pat<(v4f32 (fceil VR128:$src)),
6378 (VROUNDPSr VR128:$src, (i32 0xA))>;
6379 def : Pat<(v4f32 (frint VR128:$src)),
6380 (VROUNDPSr VR128:$src, (i32 0x4))>;
6381 def : Pat<(v4f32 (ftrunc VR128:$src)),
6382 (VROUNDPSr VR128:$src, (i32 0xB))>;
6384 def : Pat<(v2f64 (ffloor VR128:$src)),
6385 (VROUNDPDr VR128:$src, (i32 0x9))>;
6386 def : Pat<(v2f64 (fnearbyint VR128:$src)),
6387 (VROUNDPDr VR128:$src, (i32 0xC))>;
6388 def : Pat<(v2f64 (fceil VR128:$src)),
6389 (VROUNDPDr VR128:$src, (i32 0xA))>;
6390 def : Pat<(v2f64 (frint VR128:$src)),
6391 (VROUNDPDr VR128:$src, (i32 0x4))>;
6392 def : Pat<(v2f64 (ftrunc VR128:$src)),
6393 (VROUNDPDr VR128:$src, (i32 0xB))>;
6395 def : Pat<(v8f32 (ffloor VR256:$src)),
6396 (VROUNDYPSr VR256:$src, (i32 0x9))>;
6397 def : Pat<(v8f32 (fnearbyint VR256:$src)),
6398 (VROUNDYPSr VR256:$src, (i32 0xC))>;
6399 def : Pat<(v8f32 (fceil VR256:$src)),
6400 (VROUNDYPSr VR256:$src, (i32 0xA))>;
6401 def : Pat<(v8f32 (frint VR256:$src)),
6402 (VROUNDYPSr VR256:$src, (i32 0x4))>;
6403 def : Pat<(v8f32 (ftrunc VR256:$src)),
6404 (VROUNDYPSr VR256:$src, (i32 0xB))>;
6406 def : Pat<(v4f64 (ffloor VR256:$src)),
6407 (VROUNDYPDr VR256:$src, (i32 0x9))>;
6408 def : Pat<(v4f64 (fnearbyint VR256:$src)),
6409 (VROUNDYPDr VR256:$src, (i32 0xC))>;
6410 def : Pat<(v4f64 (fceil VR256:$src)),
6411 (VROUNDYPDr VR256:$src, (i32 0xA))>;
6412 def : Pat<(v4f64 (frint VR256:$src)),
6413 (VROUNDYPDr VR256:$src, (i32 0x4))>;
6414 def : Pat<(v4f64 (ftrunc VR256:$src)),
6415 (VROUNDYPDr VR256:$src, (i32 0xB))>;
6418 defm ROUND : sse41_fp_unop_p<0x08, 0x09, "round", f128mem, VR128,
6419 memopv4f32, memopv2f64, int_x86_sse41_round_ps,
6420 int_x86_sse41_round_pd>;
6422 defm ROUND : sse41_fp_unop_s<0x0A, 0x0B, "round">;
6424 let Constraints = "$src1 = $dst" in
6425 defm ROUND : sse41_fp_binop_s<0x0A, 0x0B, "round",
6426 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
6428 let Predicates = [UseSSE41] in {
6429 def : Pat<(ffloor FR32:$src),
6430 (ROUNDSSr FR32:$src, (i32 0x9))>;
6431 def : Pat<(f64 (ffloor FR64:$src)),
6432 (ROUNDSDr FR64:$src, (i32 0x9))>;
6433 def : Pat<(f32 (fnearbyint FR32:$src)),
6434 (ROUNDSSr FR32:$src, (i32 0xC))>;
6435 def : Pat<(f64 (fnearbyint FR64:$src)),
6436 (ROUNDSDr FR64:$src, (i32 0xC))>;
6437 def : Pat<(f32 (fceil FR32:$src)),
6438 (ROUNDSSr FR32:$src, (i32 0xA))>;
6439 def : Pat<(f64 (fceil FR64:$src)),
6440 (ROUNDSDr FR64:$src, (i32 0xA))>;
6441 def : Pat<(f32 (frint FR32:$src)),
6442 (ROUNDSSr FR32:$src, (i32 0x4))>;
6443 def : Pat<(f64 (frint FR64:$src)),
6444 (ROUNDSDr FR64:$src, (i32 0x4))>;
6445 def : Pat<(f32 (ftrunc FR32:$src)),
6446 (ROUNDSSr FR32:$src, (i32 0xB))>;
6447 def : Pat<(f64 (ftrunc FR64:$src)),
6448 (ROUNDSDr FR64:$src, (i32 0xB))>;
6450 def : Pat<(v4f32 (ffloor VR128:$src)),
6451 (ROUNDPSr VR128:$src, (i32 0x9))>;
6452 def : Pat<(v4f32 (fnearbyint VR128:$src)),
6453 (ROUNDPSr VR128:$src, (i32 0xC))>;
6454 def : Pat<(v4f32 (fceil VR128:$src)),
6455 (ROUNDPSr VR128:$src, (i32 0xA))>;
6456 def : Pat<(v4f32 (frint VR128:$src)),
6457 (ROUNDPSr VR128:$src, (i32 0x4))>;
6458 def : Pat<(v4f32 (ftrunc VR128:$src)),
6459 (ROUNDPSr VR128:$src, (i32 0xB))>;
6461 def : Pat<(v2f64 (ffloor VR128:$src)),
6462 (ROUNDPDr VR128:$src, (i32 0x9))>;
6463 def : Pat<(v2f64 (fnearbyint VR128:$src)),
6464 (ROUNDPDr VR128:$src, (i32 0xC))>;
6465 def : Pat<(v2f64 (fceil VR128:$src)),
6466 (ROUNDPDr VR128:$src, (i32 0xA))>;
6467 def : Pat<(v2f64 (frint VR128:$src)),
6468 (ROUNDPDr VR128:$src, (i32 0x4))>;
6469 def : Pat<(v2f64 (ftrunc VR128:$src)),
6470 (ROUNDPDr VR128:$src, (i32 0xB))>;
6473 //===----------------------------------------------------------------------===//
6474 // SSE4.1 - Packed Bit Test
6475 //===----------------------------------------------------------------------===//
6477 // ptest instruction we'll lower to this in X86ISelLowering primarily from
6478 // the intel intrinsic that corresponds to this.
6479 let Defs = [EFLAGS], Predicates = [HasAVX] in {
6480 def VPTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
6481 "vptest\t{$src2, $src1|$src1, $src2}",
6482 [(set EFLAGS, (X86ptest VR128:$src1, (v2i64 VR128:$src2)))]>,
6483 Sched<[WriteVecLogic]>, VEX, VEX_WIG;
6484 def VPTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
6485 "vptest\t{$src2, $src1|$src1, $src2}",
6486 [(set EFLAGS,(X86ptest VR128:$src1, (loadv2i64 addr:$src2)))]>,
6487 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX, VEX_WIG;
6489 def VPTESTYrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR256:$src1, VR256:$src2),
6490 "vptest\t{$src2, $src1|$src1, $src2}",
6491 [(set EFLAGS, (X86ptest VR256:$src1, (v4i64 VR256:$src2)))]>,
6492 Sched<[WriteVecLogic]>, VEX, VEX_L, VEX_WIG;
6493 def VPTESTYrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR256:$src1, i256mem:$src2),
6494 "vptest\t{$src2, $src1|$src1, $src2}",
6495 [(set EFLAGS,(X86ptest VR256:$src1, (loadv4i64 addr:$src2)))]>,
6496 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX, VEX_L, VEX_WIG;
6499 let Defs = [EFLAGS] in {
6500 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
6501 "ptest\t{$src2, $src1|$src1, $src2}",
6502 [(set EFLAGS, (X86ptest VR128:$src1, (v2i64 VR128:$src2)))]>,
6503 Sched<[WriteVecLogic]>;
6504 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
6505 "ptest\t{$src2, $src1|$src1, $src2}",
6506 [(set EFLAGS, (X86ptest VR128:$src1, (memopv2i64 addr:$src2)))]>,
6507 Sched<[WriteVecLogicLd, ReadAfterLd]>;
6510 // The bit test instructions below are AVX only
6511 multiclass avx_bittest<bits<8> opc, string OpcodeStr, RegisterClass RC,
6512 X86MemOperand x86memop, PatFrag mem_frag, ValueType vt> {
6513 def rr : SS48I<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
6514 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
6515 [(set EFLAGS, (X86testp RC:$src1, (vt RC:$src2)))]>,
6516 Sched<[WriteVecLogic]>, VEX;
6517 def rm : SS48I<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
6518 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
6519 [(set EFLAGS, (X86testp RC:$src1, (mem_frag addr:$src2)))]>,
6520 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX;
6523 let Defs = [EFLAGS], Predicates = [HasAVX] in {
6524 let ExeDomain = SSEPackedSingle in {
6525 defm VTESTPS : avx_bittest<0x0E, "vtestps", VR128, f128mem, loadv4f32, v4f32>;
6526 defm VTESTPSY : avx_bittest<0x0E, "vtestps", VR256, f256mem, loadv8f32, v8f32>,
6529 let ExeDomain = SSEPackedDouble in {
6530 defm VTESTPD : avx_bittest<0x0F, "vtestpd", VR128, f128mem, loadv2f64, v2f64>;
6531 defm VTESTPDY : avx_bittest<0x0F, "vtestpd", VR256, f256mem, loadv4f64, v4f64>,
6536 //===----------------------------------------------------------------------===//
6537 // SSE4.1 - Misc Instructions
6538 //===----------------------------------------------------------------------===//
6540 let Defs = [EFLAGS], Predicates = [HasPOPCNT] in {
6541 def POPCNT16rr : I<0xB8, MRMSrcReg, (outs GR16:$dst), (ins GR16:$src),
6542 "popcnt{w}\t{$src, $dst|$dst, $src}",
6543 [(set GR16:$dst, (ctpop GR16:$src)), (implicit EFLAGS)],
6544 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>,
6546 def POPCNT16rm : I<0xB8, MRMSrcMem, (outs GR16:$dst), (ins i16mem:$src),
6547 "popcnt{w}\t{$src, $dst|$dst, $src}",
6548 [(set GR16:$dst, (ctpop (loadi16 addr:$src))),
6549 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
6550 Sched<[WriteFAddLd]>, OpSize16, XS;
6552 def POPCNT32rr : I<0xB8, MRMSrcReg, (outs GR32:$dst), (ins GR32:$src),
6553 "popcnt{l}\t{$src, $dst|$dst, $src}",
6554 [(set GR32:$dst, (ctpop GR32:$src)), (implicit EFLAGS)],
6555 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>,
6558 def POPCNT32rm : I<0xB8, MRMSrcMem, (outs GR32:$dst), (ins i32mem:$src),
6559 "popcnt{l}\t{$src, $dst|$dst, $src}",
6560 [(set GR32:$dst, (ctpop (loadi32 addr:$src))),
6561 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
6562 Sched<[WriteFAddLd]>, OpSize32, XS;
6564 def POPCNT64rr : RI<0xB8, MRMSrcReg, (outs GR64:$dst), (ins GR64:$src),
6565 "popcnt{q}\t{$src, $dst|$dst, $src}",
6566 [(set GR64:$dst, (ctpop GR64:$src)), (implicit EFLAGS)],
6567 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>, XS;
6568 def POPCNT64rm : RI<0xB8, MRMSrcMem, (outs GR64:$dst), (ins i64mem:$src),
6569 "popcnt{q}\t{$src, $dst|$dst, $src}",
6570 [(set GR64:$dst, (ctpop (loadi64 addr:$src))),
6571 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
6572 Sched<[WriteFAddLd]>, XS;
6577 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
6578 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
6579 Intrinsic IntId128, PatFrag ld_frag,
6580 X86FoldableSchedWrite Sched> {
6581 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
6583 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6584 [(set VR128:$dst, (IntId128 VR128:$src))]>,
6586 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
6588 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6590 (IntId128 (bitconvert (ld_frag addr:$src))))]>,
6591 Sched<[Sched.Folded]>;
6594 // PHMIN has the same profile as PSAD, thus we use the same scheduling
6595 // model, although the naming is misleading.
6596 let Predicates = [HasAVX] in
6597 defm VPHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "vphminposuw",
6598 int_x86_sse41_phminposuw, loadv2i64,
6599 WriteVecIMul>, VEX, VEX_WIG;
6600 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
6601 int_x86_sse41_phminposuw, memopv2i64,
6604 /// SS48I_binop_rm - Simple SSE41 binary operator.
6605 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
6606 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
6607 X86MemOperand x86memop, bit Is2Addr = 1,
6608 OpndItins itins = SSE_INTALU_ITINS_P> {
6609 let isCommutable = 1 in
6610 def rr : SS48I<opc, MRMSrcReg, (outs RC:$dst),
6611 (ins RC:$src1, RC:$src2),
6613 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
6614 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
6615 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))]>,
6616 Sched<[itins.Sched]>;
6617 def rm : SS48I<opc, MRMSrcMem, (outs RC:$dst),
6618 (ins RC:$src1, x86memop:$src2),
6620 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
6621 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
6623 (OpVT (OpNode RC:$src1, (bitconvert (memop_frag addr:$src2)))))]>,
6624 Sched<[itins.Sched.Folded, ReadAfterLd]>;
6627 /// SS48I_binop_rm2 - Simple SSE41 binary operator with different src and dst
6629 multiclass SS48I_binop_rm2<bits<8> opc, string OpcodeStr, SDNode OpNode,
6630 ValueType DstVT, ValueType SrcVT, RegisterClass RC,
6631 PatFrag memop_frag, X86MemOperand x86memop,
6633 bit IsCommutable = 0, bit Is2Addr = 1> {
6634 let isCommutable = IsCommutable in
6635 def rr : SS48I<opc, MRMSrcReg, (outs RC:$dst),
6636 (ins RC:$src1, RC:$src2),
6638 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
6639 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
6640 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1), RC:$src2)))]>,
6641 Sched<[itins.Sched]>;
6642 def rm : SS48I<opc, MRMSrcMem, (outs RC:$dst),
6643 (ins RC:$src1, x86memop:$src2),
6645 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
6646 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
6647 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1),
6648 (bitconvert (memop_frag addr:$src2)))))]>,
6649 Sched<[itins.Sched.Folded, ReadAfterLd]>;
6652 let Predicates = [HasAVX, NoVLX] in {
6653 defm VPMINSD : SS48I_binop_rm<0x39, "vpminsd", smin, v4i32, VR128,
6654 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6656 defm VPMINUD : SS48I_binop_rm<0x3B, "vpminud", umin, v4i32, VR128,
6657 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6659 defm VPMAXSD : SS48I_binop_rm<0x3D, "vpmaxsd", smax, v4i32, VR128,
6660 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6662 defm VPMAXUD : SS48I_binop_rm<0x3F, "vpmaxud", umax, v4i32, VR128,
6663 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6665 defm VPMULDQ : SS48I_binop_rm2<0x28, "vpmuldq", X86pmuldq, v2i64, v4i32,
6666 VR128, loadv2i64, i128mem,
6667 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V, VEX_WIG;
6669 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
6670 defm VPMINSB : SS48I_binop_rm<0x38, "vpminsb", smin, v16i8, VR128,
6671 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6673 defm VPMINUW : SS48I_binop_rm<0x3A, "vpminuw", umin, v8i16, VR128,
6674 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6676 defm VPMAXSB : SS48I_binop_rm<0x3C, "vpmaxsb", smax, v16i8, VR128,
6677 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6679 defm VPMAXUW : SS48I_binop_rm<0x3E, "vpmaxuw", umax, v8i16, VR128,
6680 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6684 let Predicates = [HasAVX2, NoVLX] in {
6685 defm VPMINSDY : SS48I_binop_rm<0x39, "vpminsd", smin, v8i32, VR256,
6686 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6687 VEX_4V, VEX_L, VEX_WIG;
6688 defm VPMINUDY : SS48I_binop_rm<0x3B, "vpminud", umin, v8i32, VR256,
6689 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6690 VEX_4V, VEX_L, VEX_WIG;
6691 defm VPMAXSDY : SS48I_binop_rm<0x3D, "vpmaxsd", smax, v8i32, VR256,
6692 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6693 VEX_4V, VEX_L, VEX_WIG;
6694 defm VPMAXUDY : SS48I_binop_rm<0x3F, "vpmaxud", umax, v8i32, VR256,
6695 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6696 VEX_4V, VEX_L, VEX_WIG;
6697 defm VPMULDQY : SS48I_binop_rm2<0x28, "vpmuldq", X86pmuldq, v4i64, v8i32,
6698 VR256, loadv4i64, i256mem,
6699 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V, VEX_L, VEX_WIG;
6701 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
6702 defm VPMINSBY : SS48I_binop_rm<0x38, "vpminsb", smin, v32i8, VR256,
6703 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6704 VEX_4V, VEX_L, VEX_WIG;
6705 defm VPMINUWY : SS48I_binop_rm<0x3A, "vpminuw", umin, v16i16, VR256,
6706 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6707 VEX_4V, VEX_L, VEX_WIG;
6708 defm VPMAXSBY : SS48I_binop_rm<0x3C, "vpmaxsb", smax, v32i8, VR256,
6709 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6710 VEX_4V, VEX_L, VEX_WIG;
6711 defm VPMAXUWY : SS48I_binop_rm<0x3E, "vpmaxuw", umax, v16i16, VR256,
6712 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6713 VEX_4V, VEX_L, VEX_WIG;
6716 let Constraints = "$src1 = $dst" in {
6717 defm PMINSB : SS48I_binop_rm<0x38, "pminsb", smin, v16i8, VR128,
6718 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6719 defm PMINSD : SS48I_binop_rm<0x39, "pminsd", smin, v4i32, VR128,
6720 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6721 defm PMINUD : SS48I_binop_rm<0x3B, "pminud", umin, v4i32, VR128,
6722 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6723 defm PMINUW : SS48I_binop_rm<0x3A, "pminuw", umin, v8i16, VR128,
6724 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6725 defm PMAXSB : SS48I_binop_rm<0x3C, "pmaxsb", smax, v16i8, VR128,
6726 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6727 defm PMAXSD : SS48I_binop_rm<0x3D, "pmaxsd", smax, v4i32, VR128,
6728 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6729 defm PMAXUD : SS48I_binop_rm<0x3F, "pmaxud", umax, v4i32, VR128,
6730 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6731 defm PMAXUW : SS48I_binop_rm<0x3E, "pmaxuw", umax, v8i16, VR128,
6732 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6733 defm PMULDQ : SS48I_binop_rm2<0x28, "pmuldq", X86pmuldq, v2i64, v4i32,
6734 VR128, memopv2i64, i128mem,
6735 SSE_INTMUL_ITINS_P, 1>;
6738 let Predicates = [HasAVX, NoVLX] in
6739 defm VPMULLD : SS48I_binop_rm<0x40, "vpmulld", mul, v4i32, VR128,
6740 loadv2i64, i128mem, 0, SSE_PMULLD_ITINS>,
6742 let Predicates = [HasAVX] in
6743 defm VPCMPEQQ : SS48I_binop_rm<0x29, "vpcmpeqq", X86pcmpeq, v2i64, VR128,
6744 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6747 let Predicates = [HasAVX2, NoVLX] in
6748 defm VPMULLDY : SS48I_binop_rm<0x40, "vpmulld", mul, v8i32, VR256,
6749 loadv4i64, i256mem, 0, SSE_PMULLD_ITINS>,
6750 VEX_4V, VEX_L, VEX_WIG;
6751 let Predicates = [HasAVX2] in
6752 defm VPCMPEQQY : SS48I_binop_rm<0x29, "vpcmpeqq", X86pcmpeq, v4i64, VR256,
6753 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6754 VEX_4V, VEX_L, VEX_WIG;
6756 let Constraints = "$src1 = $dst" in {
6757 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, VR128,
6758 memopv2i64, i128mem, 1, SSE_PMULLD_ITINS>;
6759 defm PCMPEQQ : SS48I_binop_rm<0x29, "pcmpeqq", X86pcmpeq, v2i64, VR128,
6760 memopv2i64, i128mem, 1, SSE_INTALUQ_ITINS_P>;
6763 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
6764 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
6765 Intrinsic IntId, RegisterClass RC, PatFrag memop_frag,
6766 X86MemOperand x86memop, bit Is2Addr = 1,
6767 OpndItins itins = DEFAULT_ITINS> {
6768 let isCommutable = 1 in
6769 def rri : SS4AIi8<opc, MRMSrcReg, (outs RC:$dst),
6770 (ins RC:$src1, RC:$src2, u8imm:$src3),
6772 !strconcat(OpcodeStr,
6773 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6774 !strconcat(OpcodeStr,
6775 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6776 [(set RC:$dst, (IntId RC:$src1, RC:$src2, imm:$src3))], itins.rr>,
6777 Sched<[itins.Sched]>;
6778 def rmi : SS4AIi8<opc, MRMSrcMem, (outs RC:$dst),
6779 (ins RC:$src1, x86memop:$src2, u8imm:$src3),
6781 !strconcat(OpcodeStr,
6782 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6783 !strconcat(OpcodeStr,
6784 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6787 (bitconvert (memop_frag addr:$src2)), imm:$src3))], itins.rm>,
6788 Sched<[itins.Sched.Folded, ReadAfterLd]>;
6791 /// SS41I_binop_rmi - SSE 4.1 binary operator with 8-bit immediate
6792 multiclass SS41I_binop_rmi<bits<8> opc, string OpcodeStr, SDNode OpNode,
6793 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
6794 X86MemOperand x86memop, bit Is2Addr = 1,
6795 OpndItins itins = DEFAULT_ITINS> {
6796 let isCommutable = 1 in
6797 def rri : SS4AIi8<opc, MRMSrcReg, (outs RC:$dst),
6798 (ins RC:$src1, RC:$src2, u8imm:$src3),
6800 !strconcat(OpcodeStr,
6801 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6802 !strconcat(OpcodeStr,
6803 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6804 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2, imm:$src3)))],
6805 itins.rr>, Sched<[itins.Sched]>;
6806 def rmi : SS4AIi8<opc, MRMSrcMem, (outs RC:$dst),
6807 (ins RC:$src1, x86memop:$src2, u8imm:$src3),
6809 !strconcat(OpcodeStr,
6810 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6811 !strconcat(OpcodeStr,
6812 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6814 (OpVT (OpNode RC:$src1,
6815 (bitconvert (memop_frag addr:$src2)), imm:$src3)))], itins.rm>,
6816 Sched<[itins.Sched.Folded, ReadAfterLd]>;
6819 let Predicates = [HasAVX] in {
6820 let isCommutable = 0 in {
6821 defm VMPSADBW : SS41I_binop_rmi_int<0x42, "vmpsadbw", int_x86_sse41_mpsadbw,
6822 VR128, loadv2i64, i128mem, 0,
6823 DEFAULT_ITINS_MPSADSCHED>, VEX_4V, VEX_WIG;
6826 let ExeDomain = SSEPackedSingle in {
6827 defm VBLENDPS : SS41I_binop_rmi<0x0C, "vblendps", X86Blendi, v4f32,
6828 VR128, loadv4f32, f128mem, 0,
6829 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V, VEX_WIG;
6830 defm VBLENDPSY : SS41I_binop_rmi<0x0C, "vblendps", X86Blendi, v8f32,
6831 VR256, loadv8f32, f256mem, 0,
6832 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V, VEX_L, VEX_WIG;
6834 let ExeDomain = SSEPackedDouble in {
6835 defm VBLENDPD : SS41I_binop_rmi<0x0D, "vblendpd", X86Blendi, v2f64,
6836 VR128, loadv2f64, f128mem, 0,
6837 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V, VEX_WIG;
6838 defm VBLENDPDY : SS41I_binop_rmi<0x0D, "vblendpd", X86Blendi, v4f64,
6839 VR256, loadv4f64, f256mem, 0,
6840 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V, VEX_L, VEX_WIG;
6842 defm VPBLENDW : SS41I_binop_rmi<0x0E, "vpblendw", X86Blendi, v8i16,
6843 VR128, loadv2i64, i128mem, 0,
6844 DEFAULT_ITINS_BLENDSCHED>, VEX_4V, VEX_WIG;
6846 let ExeDomain = SSEPackedSingle in
6847 defm VDPPS : SS41I_binop_rmi_int<0x40, "vdpps", int_x86_sse41_dpps,
6848 VR128, loadv4f32, f128mem, 0,
6849 SSE_DPPS_ITINS>, VEX_4V, VEX_WIG;
6850 let ExeDomain = SSEPackedDouble in
6851 defm VDPPD : SS41I_binop_rmi_int<0x41, "vdppd", int_x86_sse41_dppd,
6852 VR128, loadv2f64, f128mem, 0,
6853 SSE_DPPS_ITINS>, VEX_4V, VEX_WIG;
6854 let ExeDomain = SSEPackedSingle in
6855 defm VDPPSY : SS41I_binop_rmi_int<0x40, "vdpps", int_x86_avx_dp_ps_256,
6856 VR256, loadv8f32, i256mem, 0,
6857 SSE_DPPS_ITINS>, VEX_4V, VEX_L, VEX_WIG;
6860 let Predicates = [HasAVX2] in {
6861 let isCommutable = 0 in {
6862 defm VMPSADBWY : SS41I_binop_rmi_int<0x42, "vmpsadbw", int_x86_avx2_mpsadbw,
6863 VR256, loadv4i64, i256mem, 0,
6864 DEFAULT_ITINS_MPSADSCHED>, VEX_4V, VEX_L, VEX_WIG;
6866 defm VPBLENDWY : SS41I_binop_rmi<0x0E, "vpblendw", X86Blendi, v16i16,
6867 VR256, loadv4i64, i256mem, 0,
6868 DEFAULT_ITINS_BLENDSCHED>, VEX_4V, VEX_L, VEX_WIG;
6871 let Constraints = "$src1 = $dst" in {
6872 let isCommutable = 0 in {
6873 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw", int_x86_sse41_mpsadbw,
6874 VR128, memopv2i64, i128mem,
6875 1, SSE_MPSADBW_ITINS>;
6877 let ExeDomain = SSEPackedSingle in
6878 defm BLENDPS : SS41I_binop_rmi<0x0C, "blendps", X86Blendi, v4f32,
6879 VR128, memopv4f32, f128mem,
6880 1, SSE_INTALU_ITINS_FBLEND_P>;
6881 let ExeDomain = SSEPackedDouble in
6882 defm BLENDPD : SS41I_binop_rmi<0x0D, "blendpd", X86Blendi, v2f64,
6883 VR128, memopv2f64, f128mem,
6884 1, SSE_INTALU_ITINS_FBLEND_P>;
6885 defm PBLENDW : SS41I_binop_rmi<0x0E, "pblendw", X86Blendi, v8i16,
6886 VR128, memopv2i64, i128mem,
6887 1, SSE_INTALU_ITINS_BLEND_P>;
6888 let ExeDomain = SSEPackedSingle in
6889 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps", int_x86_sse41_dpps,
6890 VR128, memopv4f32, f128mem, 1,
6892 let ExeDomain = SSEPackedDouble in
6893 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd", int_x86_sse41_dppd,
6894 VR128, memopv2f64, f128mem, 1,
6898 // For insertion into the zero index (low half) of a 256-bit vector, it is
6899 // more efficient to generate a blend with immediate instead of an insert*128.
6900 let Predicates = [HasAVX] in {
6901 def : Pat<(insert_subvector (v4f64 VR256:$src1), (v2f64 VR128:$src2), (iPTR 0)),
6902 (VBLENDPDYrri VR256:$src1,
6903 (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)),
6904 VR128:$src2, sub_xmm), 0x3)>;
6905 def : Pat<(insert_subvector (v8f32 VR256:$src1), (v4f32 VR128:$src2), (iPTR 0)),
6906 (VBLENDPSYrri VR256:$src1,
6907 (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)),
6908 VR128:$src2, sub_xmm), 0xf)>;
6911 /// SS41I_quaternary_int_avx - AVX SSE 4.1 with 4 operators
6912 multiclass SS41I_quaternary_int_avx<bits<8> opc, string OpcodeStr,
6913 RegisterClass RC, X86MemOperand x86memop,
6914 PatFrag mem_frag, Intrinsic IntId,
6915 X86FoldableSchedWrite Sched> {
6916 def rr : Ii8Reg<opc, MRMSrcReg, (outs RC:$dst),
6917 (ins RC:$src1, RC:$src2, RC:$src3),
6918 !strconcat(OpcodeStr,
6919 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
6920 [(set RC:$dst, (IntId RC:$src1, RC:$src2, RC:$src3))],
6921 NoItinerary, SSEPackedInt>, TAPD, VEX_4V,
6924 def rm : Ii8Reg<opc, MRMSrcMem, (outs RC:$dst),
6925 (ins RC:$src1, x86memop:$src2, RC:$src3),
6926 !strconcat(OpcodeStr,
6927 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
6929 (IntId RC:$src1, (bitconvert (mem_frag addr:$src2)),
6931 NoItinerary, SSEPackedInt>, TAPD, VEX_4V,
6932 Sched<[Sched.Folded, ReadAfterLd]>;
6935 let Predicates = [HasAVX] in {
6936 let ExeDomain = SSEPackedDouble in {
6937 defm VBLENDVPD : SS41I_quaternary_int_avx<0x4B, "vblendvpd", VR128, f128mem,
6938 loadv2f64, int_x86_sse41_blendvpd,
6940 defm VBLENDVPDY : SS41I_quaternary_int_avx<0x4B, "vblendvpd", VR256, f256mem,
6941 loadv4f64, int_x86_avx_blendv_pd_256,
6942 WriteFVarBlend>, VEX_L;
6943 } // ExeDomain = SSEPackedDouble
6944 let ExeDomain = SSEPackedSingle in {
6945 defm VBLENDVPS : SS41I_quaternary_int_avx<0x4A, "vblendvps", VR128, f128mem,
6946 loadv4f32, int_x86_sse41_blendvps,
6948 defm VBLENDVPSY : SS41I_quaternary_int_avx<0x4A, "vblendvps", VR256, f256mem,
6949 loadv8f32, int_x86_avx_blendv_ps_256,
6950 WriteFVarBlend>, VEX_L;
6951 } // ExeDomain = SSEPackedSingle
6952 defm VPBLENDVB : SS41I_quaternary_int_avx<0x4C, "vpblendvb", VR128, i128mem,
6953 loadv2i64, int_x86_sse41_pblendvb,
6957 let Predicates = [HasAVX2] in {
6958 defm VPBLENDVBY : SS41I_quaternary_int_avx<0x4C, "vpblendvb", VR256, i256mem,
6959 loadv4i64, int_x86_avx2_pblendvb,
6960 WriteVarBlend>, VEX_L;
6963 let Predicates = [HasAVX] in {
6964 def : Pat<(v16i8 (vselect (v16i8 VR128:$mask), (v16i8 VR128:$src1),
6965 (v16i8 VR128:$src2))),
6966 (VPBLENDVBrr VR128:$src2, VR128:$src1, VR128:$mask)>;
6967 def : Pat<(v4i32 (vselect (v4i32 VR128:$mask), (v4i32 VR128:$src1),
6968 (v4i32 VR128:$src2))),
6969 (VBLENDVPSrr VR128:$src2, VR128:$src1, VR128:$mask)>;
6970 def : Pat<(v4f32 (vselect (v4i32 VR128:$mask), (v4f32 VR128:$src1),
6971 (v4f32 VR128:$src2))),
6972 (VBLENDVPSrr VR128:$src2, VR128:$src1, VR128:$mask)>;
6973 def : Pat<(v2i64 (vselect (v2i64 VR128:$mask), (v2i64 VR128:$src1),
6974 (v2i64 VR128:$src2))),
6975 (VBLENDVPDrr VR128:$src2, VR128:$src1, VR128:$mask)>;
6976 def : Pat<(v2f64 (vselect (v2i64 VR128:$mask), (v2f64 VR128:$src1),
6977 (v2f64 VR128:$src2))),
6978 (VBLENDVPDrr VR128:$src2, VR128:$src1, VR128:$mask)>;
6979 def : Pat<(v8i32 (vselect (v8i32 VR256:$mask), (v8i32 VR256:$src1),
6980 (v8i32 VR256:$src2))),
6981 (VBLENDVPSYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
6982 def : Pat<(v8f32 (vselect (v8i32 VR256:$mask), (v8f32 VR256:$src1),
6983 (v8f32 VR256:$src2))),
6984 (VBLENDVPSYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
6985 def : Pat<(v4i64 (vselect (v4i64 VR256:$mask), (v4i64 VR256:$src1),
6986 (v4i64 VR256:$src2))),
6987 (VBLENDVPDYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
6988 def : Pat<(v4f64 (vselect (v4i64 VR256:$mask), (v4f64 VR256:$src1),
6989 (v4f64 VR256:$src2))),
6990 (VBLENDVPDYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
6993 let Predicates = [HasAVX2] in {
6994 def : Pat<(v32i8 (vselect (v32i8 VR256:$mask), (v32i8 VR256:$src1),
6995 (v32i8 VR256:$src2))),
6996 (VPBLENDVBYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7000 // FIXME: Prefer a movss or movsd over a blendps when optimizing for size or
7001 // on targets where they have equal performance. These were changed to use
7002 // blends because blends have better throughput on SandyBridge and Haswell, but
7003 // movs[s/d] are 1-2 byte shorter instructions.
7004 let Predicates = [UseAVX] in {
7005 let AddedComplexity = 15 in {
7006 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
7007 // MOVS{S,D} to the lower bits.
7008 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
7009 (VMOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
7010 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
7011 (VBLENDPSrri (v4f32 (V_SET0)), VR128:$src, (i8 1))>;
7012 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
7013 (VPBLENDWrri (v4i32 (V_SET0)), VR128:$src, (i8 3))>;
7014 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
7015 (VMOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
7017 // Move low f32 and clear high bits.
7018 def : Pat<(v8f32 (X86vzmovl (v8f32 VR256:$src))),
7019 (VBLENDPSYrri (v8f32 (AVX_SET0)), VR256:$src, (i8 1))>;
7021 // Move low f64 and clear high bits.
7022 def : Pat<(v4f64 (X86vzmovl (v4f64 VR256:$src))),
7023 (VBLENDPDYrri (v4f64 (AVX_SET0)), VR256:$src, (i8 1))>;
7026 // These will incur an FP/int domain crossing penalty, but it may be the only
7027 // way without AVX2. Do not add any complexity because we may be able to match
7028 // more optimal patterns defined earlier in this file.
7029 def : Pat<(v8i32 (X86vzmovl (v8i32 VR256:$src))),
7030 (VBLENDPSYrri (v8i32 (AVX_SET0)), VR256:$src, (i8 1))>;
7031 def : Pat<(v4i64 (X86vzmovl (v4i64 VR256:$src))),
7032 (VBLENDPDYrri (v4i64 (AVX_SET0)), VR256:$src, (i8 1))>;
7035 // FIXME: Prefer a movss or movsd over a blendps when optimizing for size or
7036 // on targets where they have equal performance. These were changed to use
7037 // blends because blends have better throughput on SandyBridge and Haswell, but
7038 // movs[s/d] are 1-2 byte shorter instructions.
7039 let Predicates = [UseSSE41], AddedComplexity = 15 in {
7040 // With SSE41 we can use blends for these patterns.
7041 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
7042 (BLENDPSrri (v4f32 (V_SET0)), VR128:$src, (i8 1))>;
7043 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
7044 (PBLENDWrri (v4i32 (V_SET0)), VR128:$src, (i8 3))>;
7048 /// SS41I_ternary_int - SSE 4.1 ternary operator
7049 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
7050 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
7051 X86MemOperand x86memop, Intrinsic IntId,
7052 OpndItins itins = DEFAULT_ITINS> {
7053 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
7054 (ins VR128:$src1, VR128:$src2),
7055 !strconcat(OpcodeStr,
7056 "\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}"),
7057 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))],
7058 itins.rr>, Sched<[itins.Sched]>;
7060 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
7061 (ins VR128:$src1, x86memop:$src2),
7062 !strconcat(OpcodeStr,
7063 "\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}"),
7066 (bitconvert (mem_frag addr:$src2)), XMM0))],
7067 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
7071 let ExeDomain = SSEPackedDouble in
7072 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", memopv2f64, f128mem,
7073 int_x86_sse41_blendvpd,
7074 DEFAULT_ITINS_FBLENDSCHED>;
7075 let ExeDomain = SSEPackedSingle in
7076 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", memopv4f32, f128mem,
7077 int_x86_sse41_blendvps,
7078 DEFAULT_ITINS_FBLENDSCHED>;
7079 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", memopv2i64, i128mem,
7080 int_x86_sse41_pblendvb,
7081 DEFAULT_ITINS_VARBLENDSCHED>;
7083 // Aliases with the implicit xmm0 argument
7084 def : InstAlias<"blendvpd\t{$src2, $dst|$dst, $src2}",
7085 (BLENDVPDrr0 VR128:$dst, VR128:$src2), 0>;
7086 def : InstAlias<"blendvpd\t{$src2, $dst|$dst, $src2}",
7087 (BLENDVPDrm0 VR128:$dst, f128mem:$src2), 0>;
7088 def : InstAlias<"blendvps\t{$src2, $dst|$dst, $src2}",
7089 (BLENDVPSrr0 VR128:$dst, VR128:$src2), 0>;
7090 def : InstAlias<"blendvps\t{$src2, $dst|$dst, $src2}",
7091 (BLENDVPSrm0 VR128:$dst, f128mem:$src2), 0>;
7092 def : InstAlias<"pblendvb\t{$src2, $dst|$dst, $src2}",
7093 (PBLENDVBrr0 VR128:$dst, VR128:$src2), 0>;
7094 def : InstAlias<"pblendvb\t{$src2, $dst|$dst, $src2}",
7095 (PBLENDVBrm0 VR128:$dst, i128mem:$src2), 0>;
7097 let Predicates = [UseSSE41] in {
7098 def : Pat<(v16i8 (vselect (v16i8 XMM0), (v16i8 VR128:$src1),
7099 (v16i8 VR128:$src2))),
7100 (PBLENDVBrr0 VR128:$src2, VR128:$src1)>;
7101 def : Pat<(v4i32 (vselect (v4i32 XMM0), (v4i32 VR128:$src1),
7102 (v4i32 VR128:$src2))),
7103 (BLENDVPSrr0 VR128:$src2, VR128:$src1)>;
7104 def : Pat<(v4f32 (vselect (v4i32 XMM0), (v4f32 VR128:$src1),
7105 (v4f32 VR128:$src2))),
7106 (BLENDVPSrr0 VR128:$src2, VR128:$src1)>;
7107 def : Pat<(v2i64 (vselect (v2i64 XMM0), (v2i64 VR128:$src1),
7108 (v2i64 VR128:$src2))),
7109 (BLENDVPDrr0 VR128:$src2, VR128:$src1)>;
7110 def : Pat<(v2f64 (vselect (v2i64 XMM0), (v2f64 VR128:$src1),
7111 (v2f64 VR128:$src2))),
7112 (BLENDVPDrr0 VR128:$src2, VR128:$src1)>;
7115 let AddedComplexity = 400 in { // Prefer non-temporal versions
7116 let SchedRW = [WriteLoad] in {
7117 let Predicates = [HasAVX, NoVLX] in
7118 def VMOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
7119 "vmovntdqa\t{$src, $dst|$dst, $src}", []>,
7121 let Predicates = [HasAVX2, NoVLX] in
7122 def VMOVNTDQAYrm : SS48I<0x2A, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
7123 "vmovntdqa\t{$src, $dst|$dst, $src}", []>,
7124 VEX, VEX_L, VEX_WIG;
7125 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
7126 "movntdqa\t{$src, $dst|$dst, $src}", []>;
7129 let Predicates = [HasAVX2, NoVLX] in {
7130 def : Pat<(v8f32 (alignednontemporalload addr:$src)),
7131 (VMOVNTDQAYrm addr:$src)>;
7132 def : Pat<(v4f64 (alignednontemporalload addr:$src)),
7133 (VMOVNTDQAYrm addr:$src)>;
7134 def : Pat<(v4i64 (alignednontemporalload addr:$src)),
7135 (VMOVNTDQAYrm addr:$src)>;
7138 let Predicates = [HasAVX, NoVLX] in {
7139 def : Pat<(v4f32 (alignednontemporalload addr:$src)),
7140 (VMOVNTDQArm addr:$src)>;
7141 def : Pat<(v2f64 (alignednontemporalload addr:$src)),
7142 (VMOVNTDQArm addr:$src)>;
7143 def : Pat<(v2i64 (alignednontemporalload addr:$src)),
7144 (VMOVNTDQArm addr:$src)>;
7147 let Predicates = [UseSSE41] in {
7148 def : Pat<(v4f32 (alignednontemporalload addr:$src)),
7149 (MOVNTDQArm addr:$src)>;
7150 def : Pat<(v2f64 (alignednontemporalload addr:$src)),
7151 (MOVNTDQArm addr:$src)>;
7152 def : Pat<(v2i64 (alignednontemporalload addr:$src)),
7153 (MOVNTDQArm addr:$src)>;
7156 } // AddedComplexity
7158 //===----------------------------------------------------------------------===//
7159 // SSE4.2 - Compare Instructions
7160 //===----------------------------------------------------------------------===//
7162 /// SS42I_binop_rm - Simple SSE 4.2 binary operator
7163 multiclass SS42I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
7164 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
7165 X86MemOperand x86memop, OpndItins itins,
7167 def rr : SS428I<opc, MRMSrcReg, (outs RC:$dst),
7168 (ins RC:$src1, RC:$src2),
7170 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7171 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7172 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))]>, Sched<[itins.Sched]>;
7173 def rm : SS428I<opc, MRMSrcMem, (outs RC:$dst),
7174 (ins RC:$src1, x86memop:$src2),
7176 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7177 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7179 (OpVT (OpNode RC:$src1, (memop_frag addr:$src2))))]>,
7180 Sched<[itins.Sched.Folded, ReadAfterLd]>;
7183 let Predicates = [HasAVX] in
7184 defm VPCMPGTQ : SS42I_binop_rm<0x37, "vpcmpgtq", X86pcmpgt, v2i64, VR128,
7185 loadv2i64, i128mem, SSE_INTALU_ITINS_P, 0>,
7188 let Predicates = [HasAVX2] in
7189 defm VPCMPGTQY : SS42I_binop_rm<0x37, "vpcmpgtq", X86pcmpgt, v4i64, VR256,
7190 loadv4i64, i256mem, SSE_INTALU_ITINS_P, 0>,
7191 VEX_4V, VEX_L, VEX_WIG;
7193 let Constraints = "$src1 = $dst" in
7194 defm PCMPGTQ : SS42I_binop_rm<0x37, "pcmpgtq", X86pcmpgt, v2i64, VR128,
7195 memopv2i64, i128mem, SSE_INTALU_ITINS_P>;
7197 //===----------------------------------------------------------------------===//
7198 // SSE4.2 - String/text Processing Instructions
7199 //===----------------------------------------------------------------------===//
7201 // Packed Compare Implicit Length Strings, Return Mask
7202 multiclass pseudo_pcmpistrm<string asm, PatFrag ld_frag> {
7203 def REG : PseudoI<(outs VR128:$dst),
7204 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7205 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
7207 def MEM : PseudoI<(outs VR128:$dst),
7208 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7209 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1,
7210 (bc_v16i8 (ld_frag addr:$src2)), imm:$src3))]>;
7213 let Defs = [EFLAGS], usesCustomInserter = 1 in {
7214 defm VPCMPISTRM128 : pseudo_pcmpistrm<"#VPCMPISTRM128", loadv2i64>,
7215 Requires<[HasAVX]>, VEX_WIG;
7216 defm PCMPISTRM128 : pseudo_pcmpistrm<"#PCMPISTRM128", memopv2i64>,
7217 Requires<[UseSSE42]>;
7220 multiclass pcmpistrm_SS42AI<string asm> {
7221 def rr : SS42AI<0x62, MRMSrcReg, (outs),
7222 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7223 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7224 []>, Sched<[WritePCmpIStrM]>;
7226 def rm :SS42AI<0x62, MRMSrcMem, (outs),
7227 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7228 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7229 []>, Sched<[WritePCmpIStrMLd, ReadAfterLd]>;
7232 let Defs = [XMM0, EFLAGS], hasSideEffects = 0 in {
7233 let Predicates = [HasAVX] in
7234 defm VPCMPISTRM128 : pcmpistrm_SS42AI<"vpcmpistrm">, VEX;
7235 defm PCMPISTRM128 : pcmpistrm_SS42AI<"pcmpistrm"> ;
7238 // Packed Compare Explicit Length Strings, Return Mask
7239 multiclass pseudo_pcmpestrm<string asm, PatFrag ld_frag> {
7240 def REG : PseudoI<(outs VR128:$dst),
7241 (ins VR128:$src1, VR128:$src3, u8imm:$src5),
7242 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
7243 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>;
7244 def MEM : PseudoI<(outs VR128:$dst),
7245 (ins VR128:$src1, i128mem:$src3, u8imm:$src5),
7246 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128 VR128:$src1, EAX,
7247 (bc_v16i8 (ld_frag addr:$src3)), EDX, imm:$src5))]>;
7250 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
7251 defm VPCMPESTRM128 : pseudo_pcmpestrm<"#VPCMPESTRM128", loadv2i64>,
7253 defm PCMPESTRM128 : pseudo_pcmpestrm<"#PCMPESTRM128", memopv2i64>,
7254 Requires<[UseSSE42]>;
7257 multiclass SS42AI_pcmpestrm<string asm> {
7258 def rr : SS42AI<0x60, MRMSrcReg, (outs),
7259 (ins VR128:$src1, VR128:$src3, u8imm:$src5),
7260 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7261 []>, Sched<[WritePCmpEStrM]>;
7263 def rm : SS42AI<0x60, MRMSrcMem, (outs),
7264 (ins VR128:$src1, i128mem:$src3, u8imm:$src5),
7265 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7266 []>, Sched<[WritePCmpEStrMLd, ReadAfterLd]>;
7269 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX], hasSideEffects = 0 in {
7270 let Predicates = [HasAVX] in
7271 defm VPCMPESTRM128 : SS42AI_pcmpestrm<"vpcmpestrm">, VEX;
7272 defm PCMPESTRM128 : SS42AI_pcmpestrm<"pcmpestrm">;
7275 // Packed Compare Implicit Length Strings, Return Index
7276 multiclass pseudo_pcmpistri<string asm, PatFrag ld_frag> {
7277 def REG : PseudoI<(outs GR32:$dst),
7278 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7279 [(set GR32:$dst, EFLAGS,
7280 (X86pcmpistri VR128:$src1, VR128:$src2, imm:$src3))]>;
7281 def MEM : PseudoI<(outs GR32:$dst),
7282 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7283 [(set GR32:$dst, EFLAGS, (X86pcmpistri VR128:$src1,
7284 (bc_v16i8 (ld_frag addr:$src2)), imm:$src3))]>;
7287 let Defs = [EFLAGS], usesCustomInserter = 1 in {
7288 defm VPCMPISTRI : pseudo_pcmpistri<"#VPCMPISTRI", loadv2i64>,
7289 Requires<[HasAVX]>, VEX_WIG;
7290 defm PCMPISTRI : pseudo_pcmpistri<"#PCMPISTRI", memopv2i64>,
7291 Requires<[UseSSE42]>;
7294 multiclass SS42AI_pcmpistri<string asm> {
7295 def rr : SS42AI<0x63, MRMSrcReg, (outs),
7296 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7297 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7298 []>, Sched<[WritePCmpIStrI]>;
7300 def rm : SS42AI<0x63, MRMSrcMem, (outs),
7301 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7302 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7303 []>, Sched<[WritePCmpIStrILd, ReadAfterLd]>;
7306 let Defs = [ECX, EFLAGS], hasSideEffects = 0 in {
7307 let Predicates = [HasAVX] in
7308 defm VPCMPISTRI : SS42AI_pcmpistri<"vpcmpistri">, VEX;
7309 defm PCMPISTRI : SS42AI_pcmpistri<"pcmpistri">;
7312 // Packed Compare Explicit Length Strings, Return Index
7313 multiclass pseudo_pcmpestri<string asm, PatFrag ld_frag> {
7314 def REG : PseudoI<(outs GR32:$dst),
7315 (ins VR128:$src1, VR128:$src3, u8imm:$src5),
7316 [(set GR32:$dst, EFLAGS,
7317 (X86pcmpestri VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>;
7318 def MEM : PseudoI<(outs GR32:$dst),
7319 (ins VR128:$src1, i128mem:$src3, u8imm:$src5),
7320 [(set GR32:$dst, EFLAGS,
7321 (X86pcmpestri VR128:$src1, EAX, (bc_v16i8 (ld_frag addr:$src3)), EDX,
7325 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
7326 defm VPCMPESTRI : pseudo_pcmpestri<"#VPCMPESTRI", loadv2i64>,
7328 defm PCMPESTRI : pseudo_pcmpestri<"#PCMPESTRI", memopv2i64>,
7329 Requires<[UseSSE42]>;
7332 multiclass SS42AI_pcmpestri<string asm> {
7333 def rr : SS42AI<0x61, MRMSrcReg, (outs),
7334 (ins VR128:$src1, VR128:$src3, u8imm:$src5),
7335 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7336 []>, Sched<[WritePCmpEStrI]>;
7338 def rm : SS42AI<0x61, MRMSrcMem, (outs),
7339 (ins VR128:$src1, i128mem:$src3, u8imm:$src5),
7340 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7341 []>, Sched<[WritePCmpEStrILd, ReadAfterLd]>;
7344 let Defs = [ECX, EFLAGS], Uses = [EAX, EDX], hasSideEffects = 0 in {
7345 let Predicates = [HasAVX] in
7346 defm VPCMPESTRI : SS42AI_pcmpestri<"vpcmpestri">, VEX;
7347 defm PCMPESTRI : SS42AI_pcmpestri<"pcmpestri">;
7350 //===----------------------------------------------------------------------===//
7351 // SSE4.2 - CRC Instructions
7352 //===----------------------------------------------------------------------===//
7354 // No CRC instructions have AVX equivalents
7356 // crc intrinsic instruction
7357 // This set of instructions are only rm, the only difference is the size
7359 class SS42I_crc32r<bits<8> opc, string asm, RegisterClass RCOut,
7360 RegisterClass RCIn, SDPatternOperator Int> :
7361 SS42FI<opc, MRMSrcReg, (outs RCOut:$dst), (ins RCOut:$src1, RCIn:$src2),
7362 !strconcat(asm, "\t{$src2, $src1|$src1, $src2}"),
7363 [(set RCOut:$dst, (Int RCOut:$src1, RCIn:$src2))], IIC_CRC32_REG>,
7366 class SS42I_crc32m<bits<8> opc, string asm, RegisterClass RCOut,
7367 X86MemOperand x86memop, SDPatternOperator Int> :
7368 SS42FI<opc, MRMSrcMem, (outs RCOut:$dst), (ins RCOut:$src1, x86memop:$src2),
7369 !strconcat(asm, "\t{$src2, $src1|$src1, $src2}"),
7370 [(set RCOut:$dst, (Int RCOut:$src1, (load addr:$src2)))],
7371 IIC_CRC32_MEM>, Sched<[WriteFAddLd, ReadAfterLd]>;
7373 let Constraints = "$src1 = $dst" in {
7374 def CRC32r32m8 : SS42I_crc32m<0xF0, "crc32{b}", GR32, i8mem,
7375 int_x86_sse42_crc32_32_8>;
7376 def CRC32r32r8 : SS42I_crc32r<0xF0, "crc32{b}", GR32, GR8,
7377 int_x86_sse42_crc32_32_8>;
7378 def CRC32r32m16 : SS42I_crc32m<0xF1, "crc32{w}", GR32, i16mem,
7379 int_x86_sse42_crc32_32_16>, OpSize16;
7380 def CRC32r32r16 : SS42I_crc32r<0xF1, "crc32{w}", GR32, GR16,
7381 int_x86_sse42_crc32_32_16>, OpSize16;
7382 def CRC32r32m32 : SS42I_crc32m<0xF1, "crc32{l}", GR32, i32mem,
7383 int_x86_sse42_crc32_32_32>, OpSize32;
7384 def CRC32r32r32 : SS42I_crc32r<0xF1, "crc32{l}", GR32, GR32,
7385 int_x86_sse42_crc32_32_32>, OpSize32;
7386 def CRC32r64m64 : SS42I_crc32m<0xF1, "crc32{q}", GR64, i64mem,
7387 int_x86_sse42_crc32_64_64>, REX_W;
7388 def CRC32r64r64 : SS42I_crc32r<0xF1, "crc32{q}", GR64, GR64,
7389 int_x86_sse42_crc32_64_64>, REX_W;
7390 let hasSideEffects = 0 in {
7392 def CRC32r64m8 : SS42I_crc32m<0xF0, "crc32{b}", GR64, i8mem,
7394 def CRC32r64r8 : SS42I_crc32r<0xF0, "crc32{b}", GR64, GR8,
7399 //===----------------------------------------------------------------------===//
7400 // SHA-NI Instructions
7401 //===----------------------------------------------------------------------===//
7403 multiclass SHAI_binop<bits<8> Opc, string OpcodeStr, Intrinsic IntId,
7405 def rr : I<Opc, MRMSrcReg, (outs VR128:$dst),
7406 (ins VR128:$src1, VR128:$src2),
7408 !strconcat(OpcodeStr, "\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}"),
7409 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}")),
7411 (set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0)),
7412 (set VR128:$dst, (IntId VR128:$src1, VR128:$src2)))]>, T8;
7414 def rm : I<Opc, MRMSrcMem, (outs VR128:$dst),
7415 (ins VR128:$src1, i128mem:$src2),
7417 !strconcat(OpcodeStr, "\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}"),
7418 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}")),
7420 (set VR128:$dst, (IntId VR128:$src1,
7421 (bc_v4i32 (memopv2i64 addr:$src2)), XMM0)),
7422 (set VR128:$dst, (IntId VR128:$src1,
7423 (bc_v4i32 (memopv2i64 addr:$src2)))))]>, T8;
7426 let Constraints = "$src1 = $dst", Predicates = [HasSHA] in {
7427 def SHA1RNDS4rri : Ii8<0xCC, MRMSrcReg, (outs VR128:$dst),
7428 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7429 "sha1rnds4\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7431 (int_x86_sha1rnds4 VR128:$src1, VR128:$src2,
7432 (i8 imm:$src3)))]>, TA;
7433 def SHA1RNDS4rmi : Ii8<0xCC, MRMSrcMem, (outs VR128:$dst),
7434 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7435 "sha1rnds4\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7437 (int_x86_sha1rnds4 VR128:$src1,
7438 (bc_v4i32 (memopv2i64 addr:$src2)),
7439 (i8 imm:$src3)))]>, TA;
7441 defm SHA1NEXTE : SHAI_binop<0xC8, "sha1nexte", int_x86_sha1nexte>;
7442 defm SHA1MSG1 : SHAI_binop<0xC9, "sha1msg1", int_x86_sha1msg1>;
7443 defm SHA1MSG2 : SHAI_binop<0xCA, "sha1msg2", int_x86_sha1msg2>;
7446 defm SHA256RNDS2 : SHAI_binop<0xCB, "sha256rnds2", int_x86_sha256rnds2, 1>;
7448 defm SHA256MSG1 : SHAI_binop<0xCC, "sha256msg1", int_x86_sha256msg1>;
7449 defm SHA256MSG2 : SHAI_binop<0xCD, "sha256msg2", int_x86_sha256msg2>;
7452 // Aliases with explicit %xmm0
7453 def : InstAlias<"sha256rnds2\t{$src2, $dst|$dst, $src2}",
7454 (SHA256RNDS2rr VR128:$dst, VR128:$src2), 0>;
7455 def : InstAlias<"sha256rnds2\t{$src2, $dst|$dst, $src2}",
7456 (SHA256RNDS2rm VR128:$dst, i128mem:$src2), 0>;
7458 //===----------------------------------------------------------------------===//
7459 // AES-NI Instructions
7460 //===----------------------------------------------------------------------===//
7462 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId128,
7463 PatFrag ld_frag, bit Is2Addr = 1> {
7464 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
7465 (ins VR128:$src1, VR128:$src2),
7467 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7468 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7469 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
7470 Sched<[WriteAESDecEnc]>;
7471 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
7472 (ins VR128:$src1, i128mem:$src2),
7474 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7475 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7477 (IntId128 VR128:$src1, (ld_frag addr:$src2)))]>,
7478 Sched<[WriteAESDecEncLd, ReadAfterLd]>;
7481 // Perform One Round of an AES Encryption/Decryption Flow
7482 let Predicates = [HasAVX, HasAES] in {
7483 defm VAESENC : AESI_binop_rm_int<0xDC, "vaesenc",
7484 int_x86_aesni_aesenc, loadv2i64, 0>, VEX_4V, VEX_WIG;
7485 defm VAESENCLAST : AESI_binop_rm_int<0xDD, "vaesenclast",
7486 int_x86_aesni_aesenclast, loadv2i64, 0>, VEX_4V, VEX_WIG;
7487 defm VAESDEC : AESI_binop_rm_int<0xDE, "vaesdec",
7488 int_x86_aesni_aesdec, loadv2i64, 0>, VEX_4V, VEX_WIG;
7489 defm VAESDECLAST : AESI_binop_rm_int<0xDF, "vaesdeclast",
7490 int_x86_aesni_aesdeclast, loadv2i64, 0>, VEX_4V, VEX_WIG;
7493 let Constraints = "$src1 = $dst" in {
7494 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
7495 int_x86_aesni_aesenc, memopv2i64>;
7496 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
7497 int_x86_aesni_aesenclast, memopv2i64>;
7498 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
7499 int_x86_aesni_aesdec, memopv2i64>;
7500 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
7501 int_x86_aesni_aesdeclast, memopv2i64>;
7504 // Perform the AES InvMixColumn Transformation
7505 let Predicates = [HasAVX, HasAES] in {
7506 def VAESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
7508 "vaesimc\t{$src1, $dst|$dst, $src1}",
7510 (int_x86_aesni_aesimc VR128:$src1))]>, Sched<[WriteAESIMC]>,
7512 def VAESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
7513 (ins i128mem:$src1),
7514 "vaesimc\t{$src1, $dst|$dst, $src1}",
7515 [(set VR128:$dst, (int_x86_aesni_aesimc (loadv2i64 addr:$src1)))]>,
7516 Sched<[WriteAESIMCLd]>, VEX, VEX_WIG;
7518 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
7520 "aesimc\t{$src1, $dst|$dst, $src1}",
7522 (int_x86_aesni_aesimc VR128:$src1))]>, Sched<[WriteAESIMC]>;
7523 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
7524 (ins i128mem:$src1),
7525 "aesimc\t{$src1, $dst|$dst, $src1}",
7526 [(set VR128:$dst, (int_x86_aesni_aesimc (memopv2i64 addr:$src1)))]>,
7527 Sched<[WriteAESIMCLd]>;
7529 // AES Round Key Generation Assist
7530 let Predicates = [HasAVX, HasAES] in {
7531 def VAESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
7532 (ins VR128:$src1, u8imm:$src2),
7533 "vaeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7535 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
7536 Sched<[WriteAESKeyGen]>, VEX, VEX_WIG;
7537 def VAESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
7538 (ins i128mem:$src1, u8imm:$src2),
7539 "vaeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7541 (int_x86_aesni_aeskeygenassist (loadv2i64 addr:$src1), imm:$src2))]>,
7542 Sched<[WriteAESKeyGenLd]>, VEX, VEX_WIG;
7544 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
7545 (ins VR128:$src1, u8imm:$src2),
7546 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7548 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
7549 Sched<[WriteAESKeyGen]>;
7550 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
7551 (ins i128mem:$src1, u8imm:$src2),
7552 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7554 (int_x86_aesni_aeskeygenassist (memopv2i64 addr:$src1), imm:$src2))]>,
7555 Sched<[WriteAESKeyGenLd]>;
7557 //===----------------------------------------------------------------------===//
7558 // PCLMUL Instructions
7559 //===----------------------------------------------------------------------===//
7561 // AVX carry-less Multiplication instructions
7562 let isCommutable = 1 in
7563 def VPCLMULQDQrr : AVXPCLMULIi8<0x44, MRMSrcReg, (outs VR128:$dst),
7564 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7565 "vpclmulqdq\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7567 (int_x86_pclmulqdq VR128:$src1, VR128:$src2, imm:$src3))]>,
7568 Sched<[WriteCLMul]>, VEX_WIG;
7570 def VPCLMULQDQrm : AVXPCLMULIi8<0x44, MRMSrcMem, (outs VR128:$dst),
7571 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7572 "vpclmulqdq\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7573 [(set VR128:$dst, (int_x86_pclmulqdq VR128:$src1,
7574 (loadv2i64 addr:$src2), imm:$src3))]>,
7575 Sched<[WriteCLMulLd, ReadAfterLd]>, VEX_WIG;
7577 // Carry-less Multiplication instructions
7578 let Constraints = "$src1 = $dst" in {
7579 let isCommutable = 1 in
7580 def PCLMULQDQrr : PCLMULIi8<0x44, MRMSrcReg, (outs VR128:$dst),
7581 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7582 "pclmulqdq\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7584 (int_x86_pclmulqdq VR128:$src1, VR128:$src2, imm:$src3))],
7585 IIC_SSE_PCLMULQDQ_RR>, Sched<[WriteCLMul]>;
7587 def PCLMULQDQrm : PCLMULIi8<0x44, MRMSrcMem, (outs VR128:$dst),
7588 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7589 "pclmulqdq\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7590 [(set VR128:$dst, (int_x86_pclmulqdq VR128:$src1,
7591 (memopv2i64 addr:$src2), imm:$src3))],
7592 IIC_SSE_PCLMULQDQ_RM>,
7593 Sched<[WriteCLMulLd, ReadAfterLd]>;
7594 } // Constraints = "$src1 = $dst"
7597 multiclass pclmul_alias<string asm, int immop> {
7598 def : InstAlias<!strconcat("pclmul", asm, "dq {$src, $dst|$dst, $src}"),
7599 (PCLMULQDQrr VR128:$dst, VR128:$src, immop), 0>;
7601 def : InstAlias<!strconcat("pclmul", asm, "dq {$src, $dst|$dst, $src}"),
7602 (PCLMULQDQrm VR128:$dst, i128mem:$src, immop), 0>;
7604 def : InstAlias<!strconcat("vpclmul", asm,
7605 "dq {$src2, $src1, $dst|$dst, $src1, $src2}"),
7606 (VPCLMULQDQrr VR128:$dst, VR128:$src1, VR128:$src2, immop),
7609 def : InstAlias<!strconcat("vpclmul", asm,
7610 "dq {$src2, $src1, $dst|$dst, $src1, $src2}"),
7611 (VPCLMULQDQrm VR128:$dst, VR128:$src1, i128mem:$src2, immop),
7614 defm : pclmul_alias<"hqhq", 0x11>;
7615 defm : pclmul_alias<"hqlq", 0x01>;
7616 defm : pclmul_alias<"lqhq", 0x10>;
7617 defm : pclmul_alias<"lqlq", 0x00>;
7619 //===----------------------------------------------------------------------===//
7620 // SSE4A Instructions
7621 //===----------------------------------------------------------------------===//
7623 let Predicates = [HasSSE4A] in {
7625 let ExeDomain = SSEPackedInt in {
7626 let Constraints = "$src = $dst" in {
7627 def EXTRQI : Ii8<0x78, MRMXr, (outs VR128:$dst),
7628 (ins VR128:$src, u8imm:$len, u8imm:$idx),
7629 "extrq\t{$idx, $len, $src|$src, $len, $idx}",
7630 [(set VR128:$dst, (X86extrqi VR128:$src, imm:$len,
7632 def EXTRQ : I<0x79, MRMSrcReg, (outs VR128:$dst),
7633 (ins VR128:$src, VR128:$mask),
7634 "extrq\t{$mask, $src|$src, $mask}",
7635 [(set VR128:$dst, (int_x86_sse4a_extrq VR128:$src,
7636 VR128:$mask))]>, PD;
7638 def INSERTQI : Ii8<0x78, MRMSrcReg, (outs VR128:$dst),
7639 (ins VR128:$src, VR128:$src2, u8imm:$len, u8imm:$idx),
7640 "insertq\t{$idx, $len, $src2, $src|$src, $src2, $len, $idx}",
7641 [(set VR128:$dst, (X86insertqi VR128:$src, VR128:$src2,
7642 imm:$len, imm:$idx))]>, XD;
7643 def INSERTQ : I<0x79, MRMSrcReg, (outs VR128:$dst),
7644 (ins VR128:$src, VR128:$mask),
7645 "insertq\t{$mask, $src|$src, $mask}",
7646 [(set VR128:$dst, (int_x86_sse4a_insertq VR128:$src,
7647 VR128:$mask))]>, XD;
7649 } // ExeDomain = SSEPackedInt
7651 // Non-temporal (unaligned) scalar stores.
7652 let AddedComplexity = 400 in { // Prefer non-temporal versions
7653 let mayStore = 1, SchedRW = [WriteStore] in {
7654 def MOVNTSS : I<0x2B, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
7655 "movntss\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVNT>, XS;
7657 def MOVNTSD : I<0x2B, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
7658 "movntsd\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVNT>, XD;
7661 def : Pat<(nontemporalstore FR32:$src, addr:$dst),
7662 (MOVNTSS addr:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
7664 def : Pat<(nontemporalstore FR64:$src, addr:$dst),
7665 (MOVNTSD addr:$dst, (COPY_TO_REGCLASS FR64:$src, VR128))>;
7667 } // AddedComplexity
7670 //===----------------------------------------------------------------------===//
7672 //===----------------------------------------------------------------------===//
7674 //===----------------------------------------------------------------------===//
7675 // VBROADCAST - Load from memory and broadcast to all elements of the
7676 // destination operand
7678 class avx_broadcast_rm<bits<8> opc, string OpcodeStr, RegisterClass RC,
7679 X86MemOperand x86memop, ValueType VT,
7680 PatFrag ld_frag, SchedWrite Sched> :
7681 AVX8I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
7682 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
7683 [(set RC:$dst, (VT (X86VBroadcast (ld_frag addr:$src))))]>,
7684 Sched<[Sched]>, VEX;
7686 // AVX2 adds register forms
7687 class avx2_broadcast_rr<bits<8> opc, string OpcodeStr, RegisterClass RC,
7688 ValueType ResVT, ValueType OpVT, SchedWrite Sched> :
7689 AVX28I<opc, MRMSrcReg, (outs RC:$dst), (ins VR128:$src),
7690 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
7691 [(set RC:$dst, (ResVT (X86VBroadcast (OpVT VR128:$src))))]>,
7692 Sched<[Sched]>, VEX;
7694 let ExeDomain = SSEPackedSingle, Predicates = [HasAVX, NoVLX] in {
7695 def VBROADCASTSSrm : avx_broadcast_rm<0x18, "vbroadcastss", VR128,
7696 f32mem, v4f32, loadf32, WriteLoad>;
7697 def VBROADCASTSSYrm : avx_broadcast_rm<0x18, "vbroadcastss", VR256,
7698 f32mem, v8f32, loadf32,
7699 WriteFShuffleLd>, VEX_L;
7701 let ExeDomain = SSEPackedDouble, Predicates = [HasAVX, NoVLX] in
7702 def VBROADCASTSDYrm : avx_broadcast_rm<0x19, "vbroadcastsd", VR256, f64mem,
7703 v4f64, loadf64, WriteFShuffleLd>, VEX_L;
7705 let ExeDomain = SSEPackedSingle, Predicates = [HasAVX2, NoVLX] in {
7706 def VBROADCASTSSrr : avx2_broadcast_rr<0x18, "vbroadcastss", VR128,
7707 v4f32, v4f32, WriteFShuffle>;
7708 def VBROADCASTSSYrr : avx2_broadcast_rr<0x18, "vbroadcastss", VR256,
7709 v8f32, v4f32, WriteFShuffle256>, VEX_L;
7711 let ExeDomain = SSEPackedDouble, Predicates = [HasAVX2, NoVLX] in
7712 def VBROADCASTSDYrr : avx2_broadcast_rr<0x19, "vbroadcastsd", VR256,
7713 v4f64, v2f64, WriteFShuffle256>, VEX_L;
7715 //===----------------------------------------------------------------------===//
7716 // VBROADCAST*128 - Load from memory and broadcast 128-bit vector to both
7717 // halves of a 256-bit vector.
7719 let mayLoad = 1, hasSideEffects = 0, Predicates = [HasAVX2] in
7720 def VBROADCASTI128 : AVX8I<0x5A, MRMSrcMem, (outs VR256:$dst),
7722 "vbroadcasti128\t{$src, $dst|$dst, $src}", []>,
7723 Sched<[WriteLoad]>, VEX, VEX_L;
7725 let mayLoad = 1, hasSideEffects = 0, Predicates = [HasAVX] in
7726 def VBROADCASTF128 : AVX8I<0x1A, MRMSrcMem, (outs VR256:$dst),
7728 "vbroadcastf128\t{$src, $dst|$dst, $src}", []>,
7729 Sched<[WriteFShuffleLd]>, VEX, VEX_L;
7731 let Predicates = [HasAVX2, NoVLX] in {
7732 def : Pat<(v4i64 (X86SubVBroadcast (loadv2i64 addr:$src))),
7733 (VBROADCASTI128 addr:$src)>;
7734 def : Pat<(v8i32 (X86SubVBroadcast (bc_v4i32 (loadv2i64 addr:$src)))),
7735 (VBROADCASTI128 addr:$src)>;
7736 def : Pat<(v16i16 (X86SubVBroadcast (bc_v8i16 (loadv2i64 addr:$src)))),
7737 (VBROADCASTI128 addr:$src)>;
7738 def : Pat<(v32i8 (X86SubVBroadcast (bc_v16i8 (loadv2i64 addr:$src)))),
7739 (VBROADCASTI128 addr:$src)>;
7742 let Predicates = [HasAVX, NoVLX] in {
7743 def : Pat<(v4f64 (X86SubVBroadcast (loadv2f64 addr:$src))),
7744 (VBROADCASTF128 addr:$src)>;
7745 def : Pat<(v8f32 (X86SubVBroadcast (loadv4f32 addr:$src))),
7746 (VBROADCASTF128 addr:$src)>;
7749 let Predicates = [HasAVX1Only] in {
7750 def : Pat<(v4i64 (X86SubVBroadcast (loadv2i64 addr:$src))),
7751 (VBROADCASTF128 addr:$src)>;
7752 def : Pat<(v8i32 (X86SubVBroadcast (bc_v4i32 (loadv2i64 addr:$src)))),
7753 (VBROADCASTF128 addr:$src)>;
7754 def : Pat<(v16i16 (X86SubVBroadcast (bc_v8i16 (loadv2i64 addr:$src)))),
7755 (VBROADCASTF128 addr:$src)>;
7756 def : Pat<(v32i8 (X86SubVBroadcast (bc_v16i8 (loadv2i64 addr:$src)))),
7757 (VBROADCASTF128 addr:$src)>;
7760 //===----------------------------------------------------------------------===//
7761 // VINSERTF128 - Insert packed floating-point values
7763 let hasSideEffects = 0, ExeDomain = SSEPackedSingle in {
7764 def VINSERTF128rr : AVXAIi8<0x18, MRMSrcReg, (outs VR256:$dst),
7765 (ins VR256:$src1, VR128:$src2, u8imm:$src3),
7766 "vinsertf128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7767 []>, Sched<[WriteFShuffle]>, VEX_4V, VEX_L;
7769 def VINSERTF128rm : AVXAIi8<0x18, MRMSrcMem, (outs VR256:$dst),
7770 (ins VR256:$src1, f128mem:$src2, u8imm:$src3),
7771 "vinsertf128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7772 []>, Sched<[WriteFShuffleLd, ReadAfterLd]>, VEX_4V, VEX_L;
7775 // To create a 256-bit all ones value, we should produce VCMPTRUEPS
7776 // with YMM register containing zero.
7777 // FIXME: Avoid producing vxorps to clear the fake inputs.
7778 let Predicates = [HasAVX1Only] in {
7779 def : Pat<(v8i32 immAllOnesV), (VCMPPSYrri (AVX_SET0), (AVX_SET0), 0xf)>;
7782 multiclass vinsert_lowering<string InstrStr, ValueType From, ValueType To,
7783 PatFrag memop_frag> {
7784 def : Pat<(vinsert128_insert:$ins (To VR256:$src1), (From VR128:$src2),
7786 (!cast<Instruction>(InstrStr#rr) VR256:$src1, VR128:$src2,
7787 (INSERT_get_vinsert128_imm VR256:$ins))>;
7788 def : Pat<(vinsert128_insert:$ins (To VR256:$src1),
7789 (From (bitconvert (memop_frag addr:$src2))),
7791 (!cast<Instruction>(InstrStr#rm) VR256:$src1, addr:$src2,
7792 (INSERT_get_vinsert128_imm VR256:$ins))>;
7795 let Predicates = [HasAVX, NoVLX] in {
7796 defm : vinsert_lowering<"VINSERTF128", v4f32, v8f32, loadv4f32>;
7797 defm : vinsert_lowering<"VINSERTF128", v2f64, v4f64, loadv2f64>;
7800 let Predicates = [HasAVX1Only] in {
7801 defm : vinsert_lowering<"VINSERTF128", v2i64, v4i64, loadv2i64>;
7802 defm : vinsert_lowering<"VINSERTF128", v4i32, v8i32, loadv2i64>;
7803 defm : vinsert_lowering<"VINSERTF128", v8i16, v16i16, loadv2i64>;
7804 defm : vinsert_lowering<"VINSERTF128", v16i8, v32i8, loadv2i64>;
7807 //===----------------------------------------------------------------------===//
7808 // VEXTRACTF128 - Extract packed floating-point values
7810 let hasSideEffects = 0, ExeDomain = SSEPackedSingle in {
7811 def VEXTRACTF128rr : AVXAIi8<0x19, MRMDestReg, (outs VR128:$dst),
7812 (ins VR256:$src1, u8imm:$src2),
7813 "vextractf128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7814 []>, Sched<[WriteFShuffle]>, VEX, VEX_L;
7816 def VEXTRACTF128mr : AVXAIi8<0x19, MRMDestMem, (outs),
7817 (ins f128mem:$dst, VR256:$src1, u8imm:$src2),
7818 "vextractf128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7819 []>, Sched<[WriteStore]>, VEX, VEX_L;
7822 multiclass vextract_lowering<string InstrStr, ValueType From, ValueType To> {
7823 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
7824 (To (!cast<Instruction>(InstrStr#rr)
7826 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
7827 def : Pat<(store (To (vextract128_extract:$ext (From VR256:$src1),
7828 (iPTR imm))), addr:$dst),
7829 (!cast<Instruction>(InstrStr#mr) addr:$dst, VR256:$src1,
7830 (EXTRACT_get_vextract128_imm VR128:$ext))>;
7834 let Predicates = [HasAVX, NoVLX] in {
7835 defm : vextract_lowering<"VEXTRACTF128", v8f32, v4f32>;
7836 defm : vextract_lowering<"VEXTRACTF128", v4f64, v2f64>;
7839 let Predicates = [HasAVX1Only] in {
7840 defm : vextract_lowering<"VEXTRACTF128", v4i64, v2i64>;
7841 defm : vextract_lowering<"VEXTRACTF128", v8i32, v4i32>;
7842 defm : vextract_lowering<"VEXTRACTF128", v16i16, v8i16>;
7843 defm : vextract_lowering<"VEXTRACTF128", v32i8, v16i8>;
7846 //===----------------------------------------------------------------------===//
7847 // VMASKMOV - Conditional SIMD Packed Loads and Stores
7849 multiclass avx_movmask_rm<bits<8> opc_rm, bits<8> opc_mr, string OpcodeStr,
7850 Intrinsic IntLd, Intrinsic IntLd256,
7851 Intrinsic IntSt, Intrinsic IntSt256> {
7852 def rm : AVX8I<opc_rm, MRMSrcMem, (outs VR128:$dst),
7853 (ins VR128:$src1, f128mem:$src2),
7854 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7855 [(set VR128:$dst, (IntLd addr:$src2, VR128:$src1))]>,
7857 def Yrm : AVX8I<opc_rm, MRMSrcMem, (outs VR256:$dst),
7858 (ins VR256:$src1, f256mem:$src2),
7859 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7860 [(set VR256:$dst, (IntLd256 addr:$src2, VR256:$src1))]>,
7862 def mr : AVX8I<opc_mr, MRMDestMem, (outs),
7863 (ins f128mem:$dst, VR128:$src1, VR128:$src2),
7864 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7865 [(IntSt addr:$dst, VR128:$src1, VR128:$src2)]>, VEX_4V;
7866 def Ymr : AVX8I<opc_mr, MRMDestMem, (outs),
7867 (ins f256mem:$dst, VR256:$src1, VR256:$src2),
7868 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7869 [(IntSt256 addr:$dst, VR256:$src1, VR256:$src2)]>, VEX_4V, VEX_L;
7872 let ExeDomain = SSEPackedSingle in
7873 defm VMASKMOVPS : avx_movmask_rm<0x2C, 0x2E, "vmaskmovps",
7874 int_x86_avx_maskload_ps,
7875 int_x86_avx_maskload_ps_256,
7876 int_x86_avx_maskstore_ps,
7877 int_x86_avx_maskstore_ps_256>;
7878 let ExeDomain = SSEPackedDouble in
7879 defm VMASKMOVPD : avx_movmask_rm<0x2D, 0x2F, "vmaskmovpd",
7880 int_x86_avx_maskload_pd,
7881 int_x86_avx_maskload_pd_256,
7882 int_x86_avx_maskstore_pd,
7883 int_x86_avx_maskstore_pd_256>;
7885 //===----------------------------------------------------------------------===//
7886 // VPERMIL - Permute Single and Double Floating-Point Values
7888 multiclass avx_permil<bits<8> opc_rm, bits<8> opc_rmi, string OpcodeStr,
7889 RegisterClass RC, X86MemOperand x86memop_f,
7890 X86MemOperand x86memop_i, PatFrag i_frag,
7891 ValueType f_vt, ValueType i_vt> {
7892 let Predicates = [HasAVX, NoVLX] in {
7893 def rr : AVX8I<opc_rm, MRMSrcReg, (outs RC:$dst),
7894 (ins RC:$src1, RC:$src2),
7895 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7896 [(set RC:$dst, (f_vt (X86VPermilpv RC:$src1, (i_vt RC:$src2))))]>, VEX_4V,
7897 Sched<[WriteFShuffle]>;
7898 def rm : AVX8I<opc_rm, MRMSrcMem, (outs RC:$dst),
7899 (ins RC:$src1, x86memop_i:$src2),
7900 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7901 [(set RC:$dst, (f_vt (X86VPermilpv RC:$src1,
7902 (i_vt (bitconvert (i_frag addr:$src2))))))]>, VEX_4V,
7903 Sched<[WriteFShuffleLd, ReadAfterLd]>;
7905 def ri : AVXAIi8<opc_rmi, MRMSrcReg, (outs RC:$dst),
7906 (ins RC:$src1, u8imm:$src2),
7907 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7908 [(set RC:$dst, (f_vt (X86VPermilpi RC:$src1, (i8 imm:$src2))))]>, VEX,
7909 Sched<[WriteFShuffle]>;
7910 def mi : AVXAIi8<opc_rmi, MRMSrcMem, (outs RC:$dst),
7911 (ins x86memop_f:$src1, u8imm:$src2),
7912 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
7914 (f_vt (X86VPermilpi (load addr:$src1), (i8 imm:$src2))))]>, VEX,
7915 Sched<[WriteFShuffleLd]>;
7916 }// Predicates = [HasAVX, NoVLX]
7919 let ExeDomain = SSEPackedSingle in {
7920 defm VPERMILPS : avx_permil<0x0C, 0x04, "vpermilps", VR128, f128mem, i128mem,
7921 loadv2i64, v4f32, v4i32>;
7922 defm VPERMILPSY : avx_permil<0x0C, 0x04, "vpermilps", VR256, f256mem, i256mem,
7923 loadv4i64, v8f32, v8i32>, VEX_L;
7925 let ExeDomain = SSEPackedDouble in {
7926 defm VPERMILPD : avx_permil<0x0D, 0x05, "vpermilpd", VR128, f128mem, i128mem,
7927 loadv2i64, v2f64, v2i64>;
7928 defm VPERMILPDY : avx_permil<0x0D, 0x05, "vpermilpd", VR256, f256mem, i256mem,
7929 loadv4i64, v4f64, v4i64>, VEX_L;
7932 //===----------------------------------------------------------------------===//
7933 // VPERM2F128 - Permute Floating-Point Values in 128-bit chunks
7935 let ExeDomain = SSEPackedSingle in {
7936 let isCommutable = 1 in
7937 def VPERM2F128rr : AVXAIi8<0x06, MRMSrcReg, (outs VR256:$dst),
7938 (ins VR256:$src1, VR256:$src2, u8imm:$src3),
7939 "vperm2f128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7940 [(set VR256:$dst, (v8f32 (X86VPerm2x128 VR256:$src1, VR256:$src2,
7941 (i8 imm:$src3))))]>, VEX_4V, VEX_L,
7942 Sched<[WriteFShuffle]>;
7943 def VPERM2F128rm : AVXAIi8<0x06, MRMSrcMem, (outs VR256:$dst),
7944 (ins VR256:$src1, f256mem:$src2, u8imm:$src3),
7945 "vperm2f128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7946 [(set VR256:$dst, (X86VPerm2x128 VR256:$src1, (loadv8f32 addr:$src2),
7947 (i8 imm:$src3)))]>, VEX_4V, VEX_L,
7948 Sched<[WriteFShuffleLd, ReadAfterLd]>;
7951 let Predicates = [HasAVX] in {
7952 def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
7953 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
7954 def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1,
7955 (loadv4f64 addr:$src2), (i8 imm:$imm))),
7956 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
7959 let Predicates = [HasAVX1Only] in {
7960 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
7961 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
7962 def : Pat<(v4i64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
7963 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
7964 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
7965 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
7966 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
7967 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
7969 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1,
7970 (bc_v8i32 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
7971 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
7972 def : Pat<(v4i64 (X86VPerm2x128 VR256:$src1,
7973 (loadv4i64 addr:$src2), (i8 imm:$imm))),
7974 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
7975 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1,
7976 (bc_v32i8 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
7977 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
7978 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1,
7979 (bc_v16i16 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
7980 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
7983 //===----------------------------------------------------------------------===//
7984 // VZERO - Zero YMM registers
7986 // Note, these instruction do not affect the YMM16-YMM31.
7987 let Defs = [YMM0, YMM1, YMM2, YMM3, YMM4, YMM5, YMM6, YMM7,
7988 YMM8, YMM9, YMM10, YMM11, YMM12, YMM13, YMM14, YMM15] in {
7989 // Zero All YMM registers
7990 def VZEROALL : I<0x77, RawFrm, (outs), (ins), "vzeroall",
7991 [(int_x86_avx_vzeroall)]>, PS, VEX, VEX_L, Requires<[HasAVX]>, VEX_WIG;
7993 // Zero Upper bits of YMM registers
7994 def VZEROUPPER : I<0x77, RawFrm, (outs), (ins), "vzeroupper",
7995 [(int_x86_avx_vzeroupper)]>, PS, VEX, Requires<[HasAVX]>, VEX_WIG;
7998 //===----------------------------------------------------------------------===//
7999 // Half precision conversion instructions
8000 //===----------------------------------------------------------------------===//
8001 multiclass f16c_ph2ps<RegisterClass RC, X86MemOperand x86memop, Intrinsic Int> {
8002 def rr : I<0x13, MRMSrcReg, (outs RC:$dst), (ins VR128:$src),
8003 "vcvtph2ps\t{$src, $dst|$dst, $src}",
8004 [(set RC:$dst, (Int VR128:$src))]>,
8005 T8PD, VEX, Sched<[WriteCvtF2F]>;
8006 let hasSideEffects = 0, mayLoad = 1 in
8007 def rm : I<0x13, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
8008 "vcvtph2ps\t{$src, $dst|$dst, $src}", []>, T8PD, VEX,
8009 Sched<[WriteCvtF2FLd]>;
8012 multiclass f16c_ps2ph<RegisterClass RC, X86MemOperand x86memop, Intrinsic Int> {
8013 def rr : Ii8<0x1D, MRMDestReg, (outs VR128:$dst),
8014 (ins RC:$src1, i32u8imm:$src2),
8015 "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8016 [(set VR128:$dst, (Int RC:$src1, imm:$src2))]>,
8017 TAPD, VEX, Sched<[WriteCvtF2F]>;
8018 let hasSideEffects = 0, mayStore = 1,
8019 SchedRW = [WriteCvtF2FLd, WriteRMW] in
8020 def mr : Ii8<0x1D, MRMDestMem, (outs),
8021 (ins x86memop:$dst, RC:$src1, i32u8imm:$src2),
8022 "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8026 let Predicates = [HasF16C] in {
8027 defm VCVTPH2PS : f16c_ph2ps<VR128, f64mem, int_x86_vcvtph2ps_128>;
8028 defm VCVTPH2PSY : f16c_ph2ps<VR256, f128mem, int_x86_vcvtph2ps_256>, VEX_L;
8029 defm VCVTPS2PH : f16c_ps2ph<VR128, f64mem, int_x86_vcvtps2ph_128>;
8030 defm VCVTPS2PHY : f16c_ps2ph<VR256, f128mem, int_x86_vcvtps2ph_256>, VEX_L;
8032 // Pattern match vcvtph2ps of a scalar i64 load.
8033 def : Pat<(int_x86_vcvtph2ps_128 (vzmovl_v2i64 addr:$src)),
8034 (VCVTPH2PSrm addr:$src)>;
8035 def : Pat<(int_x86_vcvtph2ps_128 (vzload_v2i64 addr:$src)),
8036 (VCVTPH2PSrm addr:$src)>;
8037 def : Pat<(int_x86_vcvtph2ps_128 (bitconvert
8038 (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
8039 (VCVTPH2PSrm addr:$src)>;
8041 def : Pat<(store (f64 (extractelt (bc_v2f64 (v8i16
8042 (int_x86_vcvtps2ph_128 VR128:$src1, i32:$src2))), (iPTR 0))),
8044 (VCVTPS2PHmr addr:$dst, VR128:$src1, imm:$src2)>;
8045 def : Pat<(store (i64 (extractelt (bc_v2i64 (v8i16
8046 (int_x86_vcvtps2ph_128 VR128:$src1, i32:$src2))), (iPTR 0))),
8048 (VCVTPS2PHmr addr:$dst, VR128:$src1, imm:$src2)>;
8049 def : Pat<(store (v8i16 (int_x86_vcvtps2ph_256 VR256:$src1, i32:$src2)),
8051 (VCVTPS2PHYmr addr:$dst, VR256:$src1, imm:$src2)>;
8054 // Patterns for matching conversions from float to half-float and vice versa.
8055 let Predicates = [HasF16C, NoVLX] in {
8056 // Use MXCSR.RC for rounding instead of explicitly specifying the default
8057 // rounding mode (Nearest-Even, encoded as 0). Both are equivalent in the
8058 // configurations we support (the default). However, falling back to MXCSR is
8059 // more consistent with other instructions, which are always controlled by it.
8060 // It's encoded as 0b100.
8061 def : Pat<(fp_to_f16 FR32:$src),
8062 (i16 (EXTRACT_SUBREG (VMOVPDI2DIrr (VCVTPS2PHrr
8063 (COPY_TO_REGCLASS FR32:$src, VR128), 4)), sub_16bit))>;
8065 def : Pat<(f16_to_fp GR16:$src),
8066 (f32 (COPY_TO_REGCLASS (VCVTPH2PSrr
8067 (COPY_TO_REGCLASS (MOVSX32rr16 GR16:$src), VR128)), FR32)) >;
8069 def : Pat<(f16_to_fp (i16 (fp_to_f16 FR32:$src))),
8070 (f32 (COPY_TO_REGCLASS (VCVTPH2PSrr
8071 (VCVTPS2PHrr (COPY_TO_REGCLASS FR32:$src, VR128), 4)), FR32)) >;
8074 //===----------------------------------------------------------------------===//
8075 // AVX2 Instructions
8076 //===----------------------------------------------------------------------===//
8078 /// AVX2_binop_rmi - AVX2 binary operator with 8-bit immediate
8079 multiclass AVX2_binop_rmi<bits<8> opc, string OpcodeStr, SDNode OpNode,
8080 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
8081 X86MemOperand x86memop> {
8082 let isCommutable = 1 in
8083 def rri : AVX2AIi8<opc, MRMSrcReg, (outs RC:$dst),
8084 (ins RC:$src1, RC:$src2, u8imm:$src3),
8085 !strconcat(OpcodeStr,
8086 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
8087 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2, imm:$src3)))]>,
8088 Sched<[WriteBlend]>, VEX_4V;
8089 def rmi : AVX2AIi8<opc, MRMSrcMem, (outs RC:$dst),
8090 (ins RC:$src1, x86memop:$src2, u8imm:$src3),
8091 !strconcat(OpcodeStr,
8092 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
8094 (OpVT (OpNode RC:$src1,
8095 (bitconvert (memop_frag addr:$src2)), imm:$src3)))]>,
8096 Sched<[WriteBlendLd, ReadAfterLd]>, VEX_4V;
8099 defm VPBLENDD : AVX2_binop_rmi<0x02, "vpblendd", X86Blendi, v4i32,
8100 VR128, loadv2i64, i128mem>;
8101 defm VPBLENDDY : AVX2_binop_rmi<0x02, "vpblendd", X86Blendi, v8i32,
8102 VR256, loadv4i64, i256mem>, VEX_L;
8104 // For insertion into the zero index (low half) of a 256-bit vector, it is
8105 // more efficient to generate a blend with immediate instead of an insert*128.
8106 let Predicates = [HasAVX2] in {
8107 def : Pat<(insert_subvector (v8i32 VR256:$src1), (v4i32 VR128:$src2), (iPTR 0)),
8108 (VPBLENDDYrri VR256:$src1,
8109 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8110 VR128:$src2, sub_xmm), 0xf)>;
8111 def : Pat<(insert_subvector (v4i64 VR256:$src1), (v2i64 VR128:$src2), (iPTR 0)),
8112 (VPBLENDDYrri VR256:$src1,
8113 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8114 VR128:$src2, sub_xmm), 0xf)>;
8115 def : Pat<(insert_subvector (v16i16 VR256:$src1), (v8i16 VR128:$src2), (iPTR 0)),
8116 (VPBLENDDYrri VR256:$src1,
8117 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8118 VR128:$src2, sub_xmm), 0xf)>;
8119 def : Pat<(insert_subvector (v32i8 VR256:$src1), (v16i8 VR128:$src2), (iPTR 0)),
8120 (VPBLENDDYrri VR256:$src1,
8121 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8122 VR128:$src2, sub_xmm), 0xf)>;
8125 let Predicates = [HasAVX1Only] in {
8126 def : Pat<(insert_subvector (v8i32 VR256:$src1), (v4i32 VR128:$src2), (iPTR 0)),
8127 (VBLENDPSYrri VR256:$src1,
8128 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8129 VR128:$src2, sub_xmm), 0xf)>;
8130 def : Pat<(insert_subvector (v4i64 VR256:$src1), (v2i64 VR128:$src2), (iPTR 0)),
8131 (VBLENDPSYrri VR256:$src1,
8132 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8133 VR128:$src2, sub_xmm), 0xf)>;
8134 def : Pat<(insert_subvector (v16i16 VR256:$src1), (v8i16 VR128:$src2), (iPTR 0)),
8135 (VBLENDPSYrri VR256:$src1,
8136 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8137 VR128:$src2, sub_xmm), 0xf)>;
8138 def : Pat<(insert_subvector (v32i8 VR256:$src1), (v16i8 VR128:$src2), (iPTR 0)),
8139 (VBLENDPSYrri VR256:$src1,
8140 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8141 VR128:$src2, sub_xmm), 0xf)>;
8144 //===----------------------------------------------------------------------===//
8145 // VPBROADCAST - Load from memory and broadcast to all elements of the
8146 // destination operand
8148 multiclass avx2_broadcast<bits<8> opc, string OpcodeStr,
8149 X86MemOperand x86memop, PatFrag ld_frag,
8150 ValueType OpVT128, ValueType OpVT256, Predicate prd> {
8151 let Predicates = [HasAVX2, prd] in {
8152 def rr : AVX28I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
8153 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8155 (OpVT128 (X86VBroadcast (OpVT128 VR128:$src))))]>,
8156 Sched<[WriteShuffle]>, VEX;
8157 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst), (ins x86memop:$src),
8158 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8160 (OpVT128 (X86VBroadcast (ld_frag addr:$src))))]>,
8161 Sched<[WriteLoad]>, VEX;
8162 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
8163 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8165 (OpVT256 (X86VBroadcast (OpVT128 VR128:$src))))]>,
8166 Sched<[WriteShuffle256]>, VEX, VEX_L;
8167 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst), (ins x86memop:$src),
8168 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8170 (OpVT256 (X86VBroadcast (ld_frag addr:$src))))]>,
8171 Sched<[WriteLoad]>, VEX, VEX_L;
8173 // Provide aliases for broadcast from the same register class that
8174 // automatically does the extract.
8175 def : Pat<(OpVT256 (X86VBroadcast (OpVT256 VR256:$src))),
8176 (!cast<Instruction>(NAME#"Yrr")
8177 (OpVT128 (EXTRACT_SUBREG (OpVT256 VR256:$src),sub_xmm)))>;
8181 defm VPBROADCASTB : avx2_broadcast<0x78, "vpbroadcastb", i8mem, loadi8,
8182 v16i8, v32i8, NoVLX_Or_NoBWI>;
8183 defm VPBROADCASTW : avx2_broadcast<0x79, "vpbroadcastw", i16mem, loadi16,
8184 v8i16, v16i16, NoVLX_Or_NoBWI>;
8185 defm VPBROADCASTD : avx2_broadcast<0x58, "vpbroadcastd", i32mem, loadi32,
8186 v4i32, v8i32, NoVLX>;
8187 defm VPBROADCASTQ : avx2_broadcast<0x59, "vpbroadcastq", i64mem, loadi64,
8188 v2i64, v4i64, NoVLX>;
8190 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
8191 // 32-bit targets will fail to load a i64 directly but can use ZEXT_LOAD.
8192 def : Pat<(v2i64 (X86VBroadcast (v2i64 (X86vzload addr:$src)))),
8193 (VPBROADCASTQrm addr:$src)>;
8194 def : Pat<(v4i64 (X86VBroadcast (v4i64 (X86vzload addr:$src)))),
8195 (VPBROADCASTQYrm addr:$src)>;
8196 // loadi16 is tricky to fold, because !isTypeDesirableForOp, justifiably.
8197 // This means we'll encounter truncated i32 loads; match that here.
8198 def : Pat<(v8i16 (X86VBroadcast (i16 (trunc (i32 (load addr:$src)))))),
8199 (VPBROADCASTWrm addr:$src)>;
8200 def : Pat<(v16i16 (X86VBroadcast (i16 (trunc (i32 (load addr:$src)))))),
8201 (VPBROADCASTWYrm addr:$src)>;
8202 def : Pat<(v8i16 (X86VBroadcast
8203 (i16 (trunc (i32 (zextloadi16 addr:$src)))))),
8204 (VPBROADCASTWrm addr:$src)>;
8205 def : Pat<(v16i16 (X86VBroadcast
8206 (i16 (trunc (i32 (zextloadi16 addr:$src)))))),
8207 (VPBROADCASTWYrm addr:$src)>;
8210 let Predicates = [HasAVX2, NoVLX] in {
8211 // Provide aliases for broadcast from the same register class that
8212 // automatically does the extract.
8213 def : Pat<(v8f32 (X86VBroadcast (v8f32 VR256:$src))),
8214 (VBROADCASTSSYrr (v4f32 (EXTRACT_SUBREG (v8f32 VR256:$src),
8216 def : Pat<(v4f64 (X86VBroadcast (v4f64 VR256:$src))),
8217 (VBROADCASTSDYrr (v2f64 (EXTRACT_SUBREG (v4f64 VR256:$src),
8221 let Predicates = [HasAVX2, NoVLX] in {
8222 // Provide fallback in case the load node that is used in the patterns above
8223 // is used by additional users, which prevents the pattern selection.
8224 def : Pat<(v4f32 (X86VBroadcast FR32:$src)),
8225 (VBROADCASTSSrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
8226 def : Pat<(v8f32 (X86VBroadcast FR32:$src)),
8227 (VBROADCASTSSYrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
8228 def : Pat<(v4f64 (X86VBroadcast FR64:$src)),
8229 (VBROADCASTSDYrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
8232 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
8233 def : Pat<(v16i8 (X86VBroadcast GR8:$src)),
8234 (VPBROADCASTBrr (COPY_TO_REGCLASS
8235 (i32 (INSERT_SUBREG (i32 (IMPLICIT_DEF)),
8236 GR8:$src, sub_8bit)),
8238 def : Pat<(v32i8 (X86VBroadcast GR8:$src)),
8239 (VPBROADCASTBYrr (COPY_TO_REGCLASS
8240 (i32 (INSERT_SUBREG (i32 (IMPLICIT_DEF)),
8241 GR8:$src, sub_8bit)),
8244 def : Pat<(v8i16 (X86VBroadcast GR16:$src)),
8245 (VPBROADCASTWrr (COPY_TO_REGCLASS
8246 (i32 (INSERT_SUBREG (i32 (IMPLICIT_DEF)),
8247 GR16:$src, sub_16bit)),
8249 def : Pat<(v16i16 (X86VBroadcast GR16:$src)),
8250 (VPBROADCASTWYrr (COPY_TO_REGCLASS
8251 (i32 (INSERT_SUBREG (i32 (IMPLICIT_DEF)),
8252 GR16:$src, sub_16bit)),
8255 let Predicates = [HasAVX2, NoVLX] in {
8256 def : Pat<(v4i32 (X86VBroadcast GR32:$src)),
8257 (VPBROADCASTDrr (COPY_TO_REGCLASS GR32:$src, VR128))>;
8258 def : Pat<(v8i32 (X86VBroadcast GR32:$src)),
8259 (VPBROADCASTDYrr (COPY_TO_REGCLASS GR32:$src, VR128))>;
8260 def : Pat<(v2i64 (X86VBroadcast GR64:$src)),
8261 (VPBROADCASTQrr (COPY_TO_REGCLASS GR64:$src, VR128))>;
8262 def : Pat<(v4i64 (X86VBroadcast GR64:$src)),
8263 (VPBROADCASTQYrr (COPY_TO_REGCLASS GR64:$src, VR128))>;
8266 // AVX1 broadcast patterns
8267 let Predicates = [HasAVX1Only] in {
8268 def : Pat<(v8i32 (X86VBroadcast (loadi32 addr:$src))),
8269 (VBROADCASTSSYrm addr:$src)>;
8270 def : Pat<(v4i64 (X86VBroadcast (loadi64 addr:$src))),
8271 (VBROADCASTSDYrm addr:$src)>;
8272 def : Pat<(v4i32 (X86VBroadcast (loadi32 addr:$src))),
8273 (VBROADCASTSSrm addr:$src)>;
8276 // Provide fallback in case the load node that is used in the patterns above
8277 // is used by additional users, which prevents the pattern selection.
8278 let Predicates = [HasAVX, NoVLX] in {
8279 // 128bit broadcasts:
8280 def : Pat<(v2f64 (X86VBroadcast f64:$src)),
8281 (VMOVDDUPrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
8284 let Predicates = [HasAVX1Only] in {
8285 def : Pat<(v4f32 (X86VBroadcast FR32:$src)),
8286 (VPERMILPSri (COPY_TO_REGCLASS FR32:$src, VR128), 0)>;
8287 def : Pat<(v8f32 (X86VBroadcast FR32:$src)),
8288 (VINSERTF128rr (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)),
8289 (VPERMILPSri (COPY_TO_REGCLASS FR32:$src, VR128), 0), sub_xmm),
8290 (VPERMILPSri (COPY_TO_REGCLASS FR32:$src, VR128), 0), 1)>;
8291 def : Pat<(v4f64 (X86VBroadcast FR64:$src)),
8292 (VINSERTF128rr (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)),
8293 (VMOVDDUPrr (COPY_TO_REGCLASS FR64:$src, VR128)), sub_xmm),
8294 (VMOVDDUPrr (COPY_TO_REGCLASS FR64:$src, VR128)), 1)>;
8296 def : Pat<(v4i32 (X86VBroadcast GR32:$src)),
8297 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0)>;
8298 def : Pat<(v8i32 (X86VBroadcast GR32:$src)),
8299 (VINSERTF128rr (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8300 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0), sub_xmm),
8301 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0), 1)>;
8302 def : Pat<(v4i64 (X86VBroadcast GR64:$src)),
8303 (VINSERTF128rr (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)),
8304 (VPSHUFDri (COPY_TO_REGCLASS GR64:$src, VR128), 0x44), sub_xmm),
8305 (VPSHUFDri (COPY_TO_REGCLASS GR64:$src, VR128), 0x44), 1)>;
8307 def : Pat<(v2i64 (X86VBroadcast i64:$src)),
8308 (VPSHUFDri (COPY_TO_REGCLASS GR64:$src, VR128), 0x44)>;
8311 //===----------------------------------------------------------------------===//
8312 // VPERM - Permute instructions
8315 multiclass avx2_perm<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
8316 ValueType OpVT, X86FoldableSchedWrite Sched,
8317 X86MemOperand memOp> {
8318 let Predicates = [HasAVX2, NoVLX] in {
8319 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst),
8320 (ins VR256:$src1, VR256:$src2),
8321 !strconcat(OpcodeStr,
8322 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8324 (OpVT (X86VPermv VR256:$src1, VR256:$src2)))]>,
8325 Sched<[Sched]>, VEX_4V, VEX_L;
8326 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst),
8327 (ins VR256:$src1, memOp:$src2),
8328 !strconcat(OpcodeStr,
8329 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8331 (OpVT (X86VPermv VR256:$src1,
8332 (bitconvert (mem_frag addr:$src2)))))]>,
8333 Sched<[Sched.Folded, ReadAfterLd]>, VEX_4V, VEX_L;
8337 defm VPERMD : avx2_perm<0x36, "vpermd", loadv4i64, v8i32, WriteShuffle256,
8339 let ExeDomain = SSEPackedSingle in
8340 defm VPERMPS : avx2_perm<0x16, "vpermps", loadv8f32, v8f32, WriteFShuffle256,
8343 multiclass avx2_perm_imm<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
8344 ValueType OpVT, X86FoldableSchedWrite Sched,
8345 X86MemOperand memOp> {
8346 let Predicates = [HasAVX2, NoVLX] in {
8347 def Yri : AVX2AIi8<opc, MRMSrcReg, (outs VR256:$dst),
8348 (ins VR256:$src1, u8imm:$src2),
8349 !strconcat(OpcodeStr,
8350 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8352 (OpVT (X86VPermi VR256:$src1, (i8 imm:$src2))))]>,
8353 Sched<[Sched]>, VEX, VEX_L;
8354 def Ymi : AVX2AIi8<opc, MRMSrcMem, (outs VR256:$dst),
8355 (ins memOp:$src1, u8imm:$src2),
8356 !strconcat(OpcodeStr,
8357 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8359 (OpVT (X86VPermi (mem_frag addr:$src1),
8360 (i8 imm:$src2))))]>,
8361 Sched<[Sched.Folded, ReadAfterLd]>, VEX, VEX_L;
8365 defm VPERMQ : avx2_perm_imm<0x00, "vpermq", loadv4i64, v4i64,
8366 WriteShuffle256, i256mem>, VEX_W;
8367 let ExeDomain = SSEPackedDouble in
8368 defm VPERMPD : avx2_perm_imm<0x01, "vpermpd", loadv4f64, v4f64,
8369 WriteFShuffle256, f256mem>, VEX_W;
8371 //===----------------------------------------------------------------------===//
8372 // VPERM2I128 - Permute Floating-Point Values in 128-bit chunks
8374 let isCommutable = 1 in
8375 def VPERM2I128rr : AVX2AIi8<0x46, MRMSrcReg, (outs VR256:$dst),
8376 (ins VR256:$src1, VR256:$src2, u8imm:$src3),
8377 "vperm2i128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8378 [(set VR256:$dst, (v4i64 (X86VPerm2x128 VR256:$src1, VR256:$src2,
8379 (i8 imm:$src3))))]>, Sched<[WriteShuffle256]>,
8381 def VPERM2I128rm : AVX2AIi8<0x46, MRMSrcMem, (outs VR256:$dst),
8382 (ins VR256:$src1, f256mem:$src2, u8imm:$src3),
8383 "vperm2i128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8384 [(set VR256:$dst, (X86VPerm2x128 VR256:$src1, (loadv4i64 addr:$src2),
8386 Sched<[WriteShuffle256Ld, ReadAfterLd]>, VEX_4V, VEX_L;
8388 let Predicates = [HasAVX2] in {
8389 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8390 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8391 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8392 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8393 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8394 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8396 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, (bc_v32i8 (loadv4i64 addr:$src2)),
8398 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8399 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1,
8400 (bc_v16i16 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8401 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8402 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)),
8404 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8408 //===----------------------------------------------------------------------===//
8409 // VINSERTI128 - Insert packed integer values
8411 let hasSideEffects = 0 in {
8412 def VINSERTI128rr : AVX2AIi8<0x38, MRMSrcReg, (outs VR256:$dst),
8413 (ins VR256:$src1, VR128:$src2, u8imm:$src3),
8414 "vinserti128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8415 []>, Sched<[WriteShuffle256]>, VEX_4V, VEX_L;
8417 def VINSERTI128rm : AVX2AIi8<0x38, MRMSrcMem, (outs VR256:$dst),
8418 (ins VR256:$src1, i128mem:$src2, u8imm:$src3),
8419 "vinserti128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8420 []>, Sched<[WriteShuffle256Ld, ReadAfterLd]>, VEX_4V, VEX_L;
8423 let Predicates = [HasAVX2, NoVLX] in {
8424 defm : vinsert_lowering<"VINSERTI128", v2i64, v4i64, loadv2i64>;
8425 defm : vinsert_lowering<"VINSERTI128", v4i32, v8i32, loadv2i64>;
8426 defm : vinsert_lowering<"VINSERTI128", v8i16, v16i16, loadv2i64>;
8427 defm : vinsert_lowering<"VINSERTI128", v16i8, v32i8, loadv2i64>;
8430 //===----------------------------------------------------------------------===//
8431 // VEXTRACTI128 - Extract packed integer values
8433 def VEXTRACTI128rr : AVX2AIi8<0x39, MRMDestReg, (outs VR128:$dst),
8434 (ins VR256:$src1, u8imm:$src2),
8435 "vextracti128\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8436 Sched<[WriteShuffle256]>, VEX, VEX_L;
8437 let hasSideEffects = 0, mayStore = 1 in
8438 def VEXTRACTI128mr : AVX2AIi8<0x39, MRMDestMem, (outs),
8439 (ins i128mem:$dst, VR256:$src1, u8imm:$src2),
8440 "vextracti128\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8441 Sched<[WriteStore]>, VEX, VEX_L;
8443 let Predicates = [HasAVX2, NoVLX] in {
8444 defm : vextract_lowering<"VEXTRACTI128", v4i64, v2i64>;
8445 defm : vextract_lowering<"VEXTRACTI128", v8i32, v4i32>;
8446 defm : vextract_lowering<"VEXTRACTI128", v16i16, v8i16>;
8447 defm : vextract_lowering<"VEXTRACTI128", v32i8, v16i8>;
8450 //===----------------------------------------------------------------------===//
8451 // VPMASKMOV - Conditional SIMD Integer Packed Loads and Stores
8453 multiclass avx2_pmovmask<string OpcodeStr,
8454 Intrinsic IntLd128, Intrinsic IntLd256,
8455 Intrinsic IntSt128, Intrinsic IntSt256> {
8456 def rm : AVX28I<0x8c, MRMSrcMem, (outs VR128:$dst),
8457 (ins VR128:$src1, i128mem:$src2),
8458 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8459 [(set VR128:$dst, (IntLd128 addr:$src2, VR128:$src1))]>, VEX_4V;
8460 def Yrm : AVX28I<0x8c, MRMSrcMem, (outs VR256:$dst),
8461 (ins VR256:$src1, i256mem:$src2),
8462 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8463 [(set VR256:$dst, (IntLd256 addr:$src2, VR256:$src1))]>,
8465 def mr : AVX28I<0x8e, MRMDestMem, (outs),
8466 (ins i128mem:$dst, VR128:$src1, VR128:$src2),
8467 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8468 [(IntSt128 addr:$dst, VR128:$src1, VR128:$src2)]>, VEX_4V;
8469 def Ymr : AVX28I<0x8e, MRMDestMem, (outs),
8470 (ins i256mem:$dst, VR256:$src1, VR256:$src2),
8471 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8472 [(IntSt256 addr:$dst, VR256:$src1, VR256:$src2)]>, VEX_4V, VEX_L;
8475 defm VPMASKMOVD : avx2_pmovmask<"vpmaskmovd",
8476 int_x86_avx2_maskload_d,
8477 int_x86_avx2_maskload_d_256,
8478 int_x86_avx2_maskstore_d,
8479 int_x86_avx2_maskstore_d_256>;
8480 defm VPMASKMOVQ : avx2_pmovmask<"vpmaskmovq",
8481 int_x86_avx2_maskload_q,
8482 int_x86_avx2_maskload_q_256,
8483 int_x86_avx2_maskstore_q,
8484 int_x86_avx2_maskstore_q_256>, VEX_W;
8486 multiclass maskmov_lowering<string InstrStr, RegisterClass RC, ValueType VT,
8487 ValueType MaskVT, string BlendStr, ValueType ZeroVT> {
8489 def: Pat<(X86mstore addr:$ptr, (MaskVT RC:$mask), (VT RC:$src)),
8490 (!cast<Instruction>(InstrStr#"mr") addr:$ptr, RC:$mask, RC:$src)>;
8492 def: Pat<(VT (X86mload addr:$ptr, (MaskVT RC:$mask), undef)),
8493 (!cast<Instruction>(InstrStr#"rm") RC:$mask, addr:$ptr)>;
8494 def: Pat<(VT (X86mload addr:$ptr, (MaskVT RC:$mask),
8495 (VT (bitconvert (ZeroVT immAllZerosV))))),
8496 (!cast<Instruction>(InstrStr#"rm") RC:$mask, addr:$ptr)>;
8497 def: Pat<(VT (X86mload addr:$ptr, (MaskVT RC:$mask), (VT RC:$src0))),
8498 (!cast<Instruction>(BlendStr#"rr")
8500 (!cast<Instruction>(InstrStr#"rm") RC:$mask, addr:$ptr),
8503 let Predicates = [HasAVX] in {
8504 defm : maskmov_lowering<"VMASKMOVPS", VR128, v4f32, v4i32, "VBLENDVPS", v4i32>;
8505 defm : maskmov_lowering<"VMASKMOVPD", VR128, v2f64, v2i64, "VBLENDVPD", v4i32>;
8506 defm : maskmov_lowering<"VMASKMOVPSY", VR256, v8f32, v8i32, "VBLENDVPSY", v8i32>;
8507 defm : maskmov_lowering<"VMASKMOVPDY", VR256, v4f64, v4i64, "VBLENDVPDY", v8i32>;
8509 let Predicates = [HasAVX1Only] in {
8510 // load/store i32/i64 not supported use ps/pd version
8511 defm : maskmov_lowering<"VMASKMOVPSY", VR256, v8i32, v8i32, "VBLENDVPSY", v8i32>;
8512 defm : maskmov_lowering<"VMASKMOVPDY", VR256, v4i64, v4i64, "VBLENDVPDY", v8i32>;
8513 defm : maskmov_lowering<"VMASKMOVPS", VR128, v4i32, v4i32, "VBLENDVPS", v4i32>;
8514 defm : maskmov_lowering<"VMASKMOVPD", VR128, v2i64, v2i64, "VBLENDVPD", v4i32>;
8516 let Predicates = [HasAVX2] in {
8517 defm : maskmov_lowering<"VPMASKMOVDY", VR256, v8i32, v8i32, "VBLENDVPSY", v8i32>;
8518 defm : maskmov_lowering<"VPMASKMOVQY", VR256, v4i64, v4i64, "VBLENDVPDY", v8i32>;
8519 defm : maskmov_lowering<"VPMASKMOVD", VR128, v4i32, v4i32, "VBLENDVPS", v4i32>;
8520 defm : maskmov_lowering<"VPMASKMOVQ", VR128, v2i64, v2i64, "VBLENDVPD", v4i32>;
8523 //===----------------------------------------------------------------------===//
8524 // SubVector Broadcasts
8525 // Provide fallback in case the load node that is used in the patterns above
8526 // is used by additional users, which prevents the pattern selection.
8528 let Predicates = [HasAVX2, NoVLX] in {
8529 def : Pat<(v4i64 (X86SubVBroadcast (v2i64 VR128:$src))),
8530 (VINSERTI128rr (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8531 (v2i64 VR128:$src), 1)>;
8532 def : Pat<(v8i32 (X86SubVBroadcast (v4i32 VR128:$src))),
8533 (VINSERTI128rr (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8534 (v4i32 VR128:$src), 1)>;
8535 def : Pat<(v16i16 (X86SubVBroadcast (v8i16 VR128:$src))),
8536 (VINSERTI128rr (INSERT_SUBREG (v16i16 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8537 (v8i16 VR128:$src), 1)>;
8538 def : Pat<(v32i8 (X86SubVBroadcast (v16i8 VR128:$src))),
8539 (VINSERTI128rr (INSERT_SUBREG (v32i8 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8540 (v16i8 VR128:$src), 1)>;
8543 let Predicates = [HasAVX, NoVLX] in {
8544 def : Pat<(v4f64 (X86SubVBroadcast (v2f64 VR128:$src))),
8545 (VINSERTF128rr (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8546 (v2f64 VR128:$src), 1)>;
8547 def : Pat<(v8f32 (X86SubVBroadcast (v4f32 VR128:$src))),
8548 (VINSERTF128rr (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8549 (v4f32 VR128:$src), 1)>;
8552 let Predicates = [HasAVX1Only] in {
8553 def : Pat<(v4i64 (X86SubVBroadcast (v2i64 VR128:$src))),
8554 (VINSERTF128rr (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8555 (v2i64 VR128:$src), 1)>;
8556 def : Pat<(v8i32 (X86SubVBroadcast (v4i32 VR128:$src))),
8557 (VINSERTF128rr (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8558 (v4i32 VR128:$src), 1)>;
8559 def : Pat<(v16i16 (X86SubVBroadcast (v8i16 VR128:$src))),
8560 (VINSERTF128rr (INSERT_SUBREG (v16i16 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8561 (v8i16 VR128:$src), 1)>;
8562 def : Pat<(v32i8 (X86SubVBroadcast (v16i8 VR128:$src))),
8563 (VINSERTF128rr (INSERT_SUBREG (v32i8 (IMPLICIT_DEF)), VR128:$src, sub_xmm),
8564 (v16i8 VR128:$src), 1)>;
8567 //===----------------------------------------------------------------------===//
8568 // Variable Bit Shifts
8570 multiclass avx2_var_shift<bits<8> opc, string OpcodeStr, SDNode OpNode,
8571 ValueType vt128, ValueType vt256> {
8572 def rr : AVX28I<opc, MRMSrcReg, (outs VR128:$dst),
8573 (ins VR128:$src1, VR128:$src2),
8574 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8576 (vt128 (OpNode VR128:$src1, (vt128 VR128:$src2))))]>,
8577 VEX_4V, Sched<[WriteVarVecShift]>;
8578 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst),
8579 (ins VR128:$src1, i128mem:$src2),
8580 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8582 (vt128 (OpNode VR128:$src1,
8583 (vt128 (bitconvert (loadv2i64 addr:$src2))))))]>,
8584 VEX_4V, Sched<[WriteVarVecShiftLd, ReadAfterLd]>;
8585 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst),
8586 (ins VR256:$src1, VR256:$src2),
8587 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8589 (vt256 (OpNode VR256:$src1, (vt256 VR256:$src2))))]>,
8590 VEX_4V, VEX_L, Sched<[WriteVarVecShift]>;
8591 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst),
8592 (ins VR256:$src1, i256mem:$src2),
8593 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8595 (vt256 (OpNode VR256:$src1,
8596 (vt256 (bitconvert (loadv4i64 addr:$src2))))))]>,
8597 VEX_4V, VEX_L, Sched<[WriteVarVecShiftLd, ReadAfterLd]>;
8600 let Predicates = [HasAVX2, NoVLX] in {
8601 defm VPSLLVD : avx2_var_shift<0x47, "vpsllvd", shl, v4i32, v8i32>;
8602 defm VPSLLVQ : avx2_var_shift<0x47, "vpsllvq", shl, v2i64, v4i64>, VEX_W;
8603 defm VPSRLVD : avx2_var_shift<0x45, "vpsrlvd", srl, v4i32, v8i32>;
8604 defm VPSRLVQ : avx2_var_shift<0x45, "vpsrlvq", srl, v2i64, v4i64>, VEX_W;
8605 defm VPSRAVD : avx2_var_shift<0x46, "vpsravd", sra, v4i32, v8i32>;
8607 def : Pat<(v4i32 (X86vsrav VR128:$src1, VR128:$src2)),
8608 (VPSRAVDrr VR128:$src1, VR128:$src2)>;
8609 def : Pat<(v4i32 (X86vsrav VR128:$src1,
8610 (bitconvert (loadv2i64 addr:$src2)))),
8611 (VPSRAVDrm VR128:$src1, addr:$src2)>;
8612 def : Pat<(v8i32 (X86vsrav VR256:$src1, VR256:$src2)),
8613 (VPSRAVDYrr VR256:$src1, VR256:$src2)>;
8614 def : Pat<(v8i32 (X86vsrav VR256:$src1,
8615 (bitconvert (loadv4i64 addr:$src2)))),
8616 (VPSRAVDYrm VR256:$src1, addr:$src2)>;
8621 //===----------------------------------------------------------------------===//
8622 // VGATHER - GATHER Operations
8623 multiclass avx2_gather<bits<8> opc, string OpcodeStr, RegisterClass RC256,
8624 X86MemOperand memop128, X86MemOperand memop256> {
8625 def rm : AVX28I<opc, MRMSrcMem4VOp3, (outs VR128:$dst, VR128:$mask_wb),
8626 (ins VR128:$src1, memop128:$src2, VR128:$mask),
8627 !strconcat(OpcodeStr,
8628 "\t{$mask, $src2, $dst|$dst, $src2, $mask}"),
8630 def Yrm : AVX28I<opc, MRMSrcMem4VOp3, (outs RC256:$dst, RC256:$mask_wb),
8631 (ins RC256:$src1, memop256:$src2, RC256:$mask),
8632 !strconcat(OpcodeStr,
8633 "\t{$mask, $src2, $dst|$dst, $src2, $mask}"),
8637 let mayLoad = 1, hasSideEffects = 0, Constraints
8638 = "@earlyclobber $dst,@earlyclobber $mask_wb, $src1 = $dst, $mask = $mask_wb"
8640 defm VPGATHERDQ : avx2_gather<0x90, "vpgatherdq", VR256, vx128mem, vx256mem>, VEX_W;
8641 defm VPGATHERQQ : avx2_gather<0x91, "vpgatherqq", VR256, vx128mem, vy256mem>, VEX_W;
8642 defm VPGATHERDD : avx2_gather<0x90, "vpgatherdd", VR256, vx128mem, vy256mem>;
8643 defm VPGATHERQD : avx2_gather<0x91, "vpgatherqd", VR128, vx64mem, vy128mem>;
8645 let ExeDomain = SSEPackedDouble in {
8646 defm VGATHERDPD : avx2_gather<0x92, "vgatherdpd", VR256, vx128mem, vx256mem>, VEX_W;
8647 defm VGATHERQPD : avx2_gather<0x93, "vgatherqpd", VR256, vx128mem, vy256mem>, VEX_W;
8650 let ExeDomain = SSEPackedSingle in {
8651 defm VGATHERDPS : avx2_gather<0x92, "vgatherdps", VR256, vx128mem, vy256mem>;
8652 defm VGATHERQPS : avx2_gather<0x93, "vgatherqps", VR128, vx64mem, vy128mem>;
8656 //===----------------------------------------------------------------------===//
8657 // Extra selection patterns for FR128, f128, f128mem
8659 // movaps is shorter than movdqa. movaps is in SSE and movdqa is in SSE2.
8660 def : Pat<(store (f128 FR128:$src), addr:$dst),
8661 (MOVAPSmr addr:$dst, (COPY_TO_REGCLASS (f128 FR128:$src), VR128))>;
8663 def : Pat<(loadf128 addr:$src),
8664 (COPY_TO_REGCLASS (MOVAPSrm addr:$src), FR128)>;
8666 // andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
8667 def : Pat<(X86fand FR128:$src1, (loadf128 addr:$src2)),
8669 (ANDPSrm (COPY_TO_REGCLASS FR128:$src1, VR128), f128mem:$src2),
8672 def : Pat<(X86fand FR128:$src1, FR128:$src2),
8674 (ANDPSrr (COPY_TO_REGCLASS FR128:$src1, VR128),
8675 (COPY_TO_REGCLASS FR128:$src2, VR128)), FR128)>;
8677 def : Pat<(and FR128:$src1, FR128:$src2),
8679 (ANDPSrr (COPY_TO_REGCLASS FR128:$src1, VR128),
8680 (COPY_TO_REGCLASS FR128:$src2, VR128)), FR128)>;
8682 def : Pat<(X86for FR128:$src1, (loadf128 addr:$src2)),
8684 (ORPSrm (COPY_TO_REGCLASS FR128:$src1, VR128), f128mem:$src2),
8687 def : Pat<(X86for FR128:$src1, FR128:$src2),
8689 (ORPSrr (COPY_TO_REGCLASS FR128:$src1, VR128),
8690 (COPY_TO_REGCLASS FR128:$src2, VR128)), FR128)>;
8692 def : Pat<(or FR128:$src1, FR128:$src2),
8694 (ORPSrr (COPY_TO_REGCLASS FR128:$src1, VR128),
8695 (COPY_TO_REGCLASS FR128:$src2, VR128)), FR128)>;
8697 def : Pat<(X86fxor FR128:$src1, (loadf128 addr:$src2)),
8699 (XORPSrm (COPY_TO_REGCLASS FR128:$src1, VR128), f128mem:$src2),
8702 def : Pat<(X86fxor FR128:$src1, FR128:$src2),
8704 (XORPSrr (COPY_TO_REGCLASS FR128:$src1, VR128),
8705 (COPY_TO_REGCLASS FR128:$src2, VR128)), FR128)>;
8707 def : Pat<(xor FR128:$src1, FR128:$src2),
8709 (XORPSrr (COPY_TO_REGCLASS FR128:$src1, VR128),
8710 (COPY_TO_REGCLASS FR128:$src2, VR128)), FR128)>;