1 //===-- X86InstrSSE.td - SSE Instruction Set ---------------*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the X86 SSE instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
16 class OpndItins<InstrItinClass arg_rr, InstrItinClass arg_rm> {
17 InstrItinClass rr = arg_rr;
18 InstrItinClass rm = arg_rm;
19 // InstrSchedModel info.
20 X86FoldableSchedWrite Sched = WriteFAdd;
23 class SizeItins<OpndItins arg_s, OpndItins arg_d> {
29 class ShiftOpndItins<InstrItinClass arg_rr, InstrItinClass arg_rm,
30 InstrItinClass arg_ri> {
31 InstrItinClass rr = arg_rr;
32 InstrItinClass rm = arg_rm;
33 InstrItinClass ri = arg_ri;
38 let Sched = WriteFAdd in {
39 def SSE_ALU_F32S : OpndItins<
40 IIC_SSE_ALU_F32S_RR, IIC_SSE_ALU_F32S_RM
43 def SSE_ALU_F64S : OpndItins<
44 IIC_SSE_ALU_F64S_RR, IIC_SSE_ALU_F64S_RM
48 def SSE_ALU_ITINS_S : SizeItins<
49 SSE_ALU_F32S, SSE_ALU_F64S
52 let Sched = WriteFMul in {
53 def SSE_MUL_F32S : OpndItins<
54 IIC_SSE_MUL_F32S_RR, IIC_SSE_MUL_F64S_RM
57 def SSE_MUL_F64S : OpndItins<
58 IIC_SSE_MUL_F64S_RR, IIC_SSE_MUL_F64S_RM
62 def SSE_MUL_ITINS_S : SizeItins<
63 SSE_MUL_F32S, SSE_MUL_F64S
66 let Sched = WriteFDiv in {
67 def SSE_DIV_F32S : OpndItins<
68 IIC_SSE_DIV_F32S_RR, IIC_SSE_DIV_F64S_RM
71 def SSE_DIV_F64S : OpndItins<
72 IIC_SSE_DIV_F64S_RR, IIC_SSE_DIV_F64S_RM
76 def SSE_DIV_ITINS_S : SizeItins<
77 SSE_DIV_F32S, SSE_DIV_F64S
81 let Sched = WriteFAdd in {
82 def SSE_ALU_F32P : OpndItins<
83 IIC_SSE_ALU_F32P_RR, IIC_SSE_ALU_F32P_RM
86 def SSE_ALU_F64P : OpndItins<
87 IIC_SSE_ALU_F64P_RR, IIC_SSE_ALU_F64P_RM
91 def SSE_ALU_ITINS_P : SizeItins<
92 SSE_ALU_F32P, SSE_ALU_F64P
95 let Sched = WriteFMul in {
96 def SSE_MUL_F32P : OpndItins<
97 IIC_SSE_MUL_F32P_RR, IIC_SSE_MUL_F64P_RM
100 def SSE_MUL_F64P : OpndItins<
101 IIC_SSE_MUL_F64P_RR, IIC_SSE_MUL_F64P_RM
105 def SSE_MUL_ITINS_P : SizeItins<
106 SSE_MUL_F32P, SSE_MUL_F64P
109 let Sched = WriteFDiv in {
110 def SSE_DIV_F32P : OpndItins<
111 IIC_SSE_DIV_F32P_RR, IIC_SSE_DIV_F64P_RM
114 def SSE_DIV_F64P : OpndItins<
115 IIC_SSE_DIV_F64P_RR, IIC_SSE_DIV_F64P_RM
119 def SSE_DIV_ITINS_P : SizeItins<
120 SSE_DIV_F32P, SSE_DIV_F64P
123 let Sched = WriteVecLogic in
124 def SSE_VEC_BIT_ITINS_P : OpndItins<
125 IIC_SSE_BIT_P_RR, IIC_SSE_BIT_P_RM
128 def SSE_BIT_ITINS_P : OpndItins<
129 IIC_SSE_BIT_P_RR, IIC_SSE_BIT_P_RM
132 let Sched = WriteVecALU in {
133 def SSE_INTALU_ITINS_P : OpndItins<
134 IIC_SSE_INTALU_P_RR, IIC_SSE_INTALU_P_RM
137 def SSE_INTALUQ_ITINS_P : OpndItins<
138 IIC_SSE_INTALUQ_P_RR, IIC_SSE_INTALUQ_P_RM
142 let Sched = WriteVecIMul in
143 def SSE_INTMUL_ITINS_P : OpndItins<
144 IIC_SSE_INTMUL_P_RR, IIC_SSE_INTMUL_P_RM
147 def SSE_INTSHIFT_ITINS_P : ShiftOpndItins<
148 IIC_SSE_INTSH_P_RR, IIC_SSE_INTSH_P_RM, IIC_SSE_INTSH_P_RI
151 def SSE_MOVA_ITINS : OpndItins<
152 IIC_SSE_MOVA_P_RR, IIC_SSE_MOVA_P_RM
155 def SSE_MOVU_ITINS : OpndItins<
156 IIC_SSE_MOVU_P_RR, IIC_SSE_MOVU_P_RM
159 def SSE_DPPD_ITINS : OpndItins<
160 IIC_SSE_DPPD_RR, IIC_SSE_DPPD_RM
163 def SSE_DPPS_ITINS : OpndItins<
164 IIC_SSE_DPPS_RR, IIC_SSE_DPPD_RM
167 def DEFAULT_ITINS : OpndItins<
168 IIC_ALU_NONMEM, IIC_ALU_MEM
171 def SSE_EXTRACT_ITINS : OpndItins<
172 IIC_SSE_EXTRACTPS_RR, IIC_SSE_EXTRACTPS_RM
175 def SSE_INSERT_ITINS : OpndItins<
176 IIC_SSE_INSERTPS_RR, IIC_SSE_INSERTPS_RM
179 let Sched = WriteMPSAD in
180 def SSE_MPSADBW_ITINS : OpndItins<
181 IIC_SSE_MPSADBW_RR, IIC_SSE_MPSADBW_RM
184 let Sched = WriteVecIMul in
185 def SSE_PMULLD_ITINS : OpndItins<
186 IIC_SSE_PMULLD_RR, IIC_SSE_PMULLD_RM
189 // Definitions for backward compatibility.
190 // The instructions mapped on these definitions uses a different itinerary
191 // than the actual scheduling model.
192 let Sched = WriteShuffle in
193 def DEFAULT_ITINS_SHUFFLESCHED : OpndItins<
194 IIC_ALU_NONMEM, IIC_ALU_MEM
197 let Sched = WriteVecIMul in
198 def DEFAULT_ITINS_VECIMULSCHED : OpndItins<
199 IIC_ALU_NONMEM, IIC_ALU_MEM
202 let Sched = WriteShuffle in
203 def SSE_INTALU_ITINS_SHUFF_P : OpndItins<
204 IIC_SSE_INTALU_P_RR, IIC_SSE_INTALU_P_RM
207 let Sched = WriteMPSAD in
208 def DEFAULT_ITINS_MPSADSCHED : OpndItins<
209 IIC_ALU_NONMEM, IIC_ALU_MEM
212 let Sched = WriteFBlend in
213 def DEFAULT_ITINS_FBLENDSCHED : OpndItins<
214 IIC_ALU_NONMEM, IIC_ALU_MEM
217 let Sched = WriteBlend in
218 def DEFAULT_ITINS_BLENDSCHED : OpndItins<
219 IIC_ALU_NONMEM, IIC_ALU_MEM
222 let Sched = WriteVarBlend in
223 def DEFAULT_ITINS_VARBLENDSCHED : OpndItins<
224 IIC_ALU_NONMEM, IIC_ALU_MEM
227 let Sched = WriteFBlend in
228 def SSE_INTALU_ITINS_FBLEND_P : OpndItins<
229 IIC_SSE_INTALU_P_RR, IIC_SSE_INTALU_P_RM
232 let Sched = WriteBlend in
233 def SSE_INTALU_ITINS_BLEND_P : OpndItins<
234 IIC_SSE_INTALU_P_RR, IIC_SSE_INTALU_P_RM
237 //===----------------------------------------------------------------------===//
238 // SSE 1 & 2 Instructions Classes
239 //===----------------------------------------------------------------------===//
241 /// sse12_fp_scalar - SSE 1 & 2 scalar instructions class
242 multiclass sse12_fp_scalar<bits<8> opc, string OpcodeStr, SDNode OpNode,
243 RegisterClass RC, X86MemOperand x86memop,
244 Domain d, OpndItins itins, bit Is2Addr = 1> {
245 let isCommutable = 1 in {
246 def rr : SI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
248 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
249 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
250 [(set RC:$dst, (OpNode RC:$src1, RC:$src2))], itins.rr, d>,
251 Sched<[itins.Sched]>;
253 def rm : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
255 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
256 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
257 [(set RC:$dst, (OpNode RC:$src1, (load addr:$src2)))], itins.rm, d>,
258 Sched<[itins.Sched.Folded, ReadAfterLd]>;
261 /// sse12_fp_scalar_int - SSE 1 & 2 scalar instructions intrinsics class
262 multiclass sse12_fp_scalar_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
263 string asm, string SSEVer, string FPSizeStr,
264 Operand memopr, ComplexPattern mem_cpat,
265 Domain d, OpndItins itins, bit Is2Addr = 1> {
266 let isCodeGenOnly = 1 in {
267 def rr_Int : SI<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, (!cast<Intrinsic>(
272 !strconcat("int_x86_sse", SSEVer, "_", OpcodeStr, FPSizeStr))
273 RC:$src1, RC:$src2))], itins.rr, d>,
274 Sched<[itins.Sched]>;
275 def rm_Int : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memopr:$src2),
277 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
278 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
279 [(set RC:$dst, (!cast<Intrinsic>(!strconcat("int_x86_sse",
280 SSEVer, "_", OpcodeStr, FPSizeStr))
281 RC:$src1, mem_cpat:$src2))], itins.rm, d>,
282 Sched<[itins.Sched.Folded, ReadAfterLd]>;
286 /// sse12_fp_packed - SSE 1 & 2 packed instructions class
287 multiclass sse12_fp_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
288 RegisterClass RC, ValueType vt,
289 X86MemOperand x86memop, PatFrag mem_frag,
290 Domain d, OpndItins itins, bit Is2Addr = 1> {
291 let isCommutable = 1 in
292 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
294 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
295 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
296 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], itins.rr, d>,
297 Sched<[itins.Sched]>;
299 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
301 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
302 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
303 [(set RC:$dst, (OpNode RC:$src1, (mem_frag addr:$src2)))],
305 Sched<[itins.Sched.Folded, ReadAfterLd]>;
308 /// sse12_fp_packed_logical_rm - SSE 1 & 2 packed instructions class
309 multiclass sse12_fp_packed_logical_rm<bits<8> opc, RegisterClass RC, Domain d,
310 string OpcodeStr, X86MemOperand x86memop,
311 list<dag> pat_rr, list<dag> pat_rm,
313 let isCommutable = 1, hasSideEffects = 0 in
314 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
316 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
317 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
318 pat_rr, NoItinerary, d>,
319 Sched<[WriteVecLogic]>;
320 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
322 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
323 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
324 pat_rm, NoItinerary, d>,
325 Sched<[WriteVecLogicLd, ReadAfterLd]>;
328 //===----------------------------------------------------------------------===//
329 // Non-instruction patterns
330 //===----------------------------------------------------------------------===//
332 // A vector extract of the first f32/f64 position is a subregister copy
333 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
334 (COPY_TO_REGCLASS (v4f32 VR128:$src), FR32)>;
335 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
336 (COPY_TO_REGCLASS (v2f64 VR128:$src), FR64)>;
338 // A 128-bit subvector extract from the first 256-bit vector position
339 // is a subregister copy that needs no instruction.
340 def : Pat<(v4i32 (extract_subvector (v8i32 VR256:$src), (iPTR 0))),
341 (v4i32 (EXTRACT_SUBREG (v8i32 VR256:$src), sub_xmm))>;
342 def : Pat<(v4f32 (extract_subvector (v8f32 VR256:$src), (iPTR 0))),
343 (v4f32 (EXTRACT_SUBREG (v8f32 VR256:$src), sub_xmm))>;
345 def : Pat<(v2i64 (extract_subvector (v4i64 VR256:$src), (iPTR 0))),
346 (v2i64 (EXTRACT_SUBREG (v4i64 VR256:$src), sub_xmm))>;
347 def : Pat<(v2f64 (extract_subvector (v4f64 VR256:$src), (iPTR 0))),
348 (v2f64 (EXTRACT_SUBREG (v4f64 VR256:$src), sub_xmm))>;
350 def : Pat<(v8i16 (extract_subvector (v16i16 VR256:$src), (iPTR 0))),
351 (v8i16 (EXTRACT_SUBREG (v16i16 VR256:$src), sub_xmm))>;
352 def : Pat<(v16i8 (extract_subvector (v32i8 VR256:$src), (iPTR 0))),
353 (v16i8 (EXTRACT_SUBREG (v32i8 VR256:$src), sub_xmm))>;
355 // A 128-bit subvector insert to the first 256-bit vector position
356 // is a subregister copy that needs no instruction.
357 let AddedComplexity = 25 in { // to give priority over vinsertf128rm
358 def : Pat<(insert_subvector undef, (v2i64 VR128:$src), (iPTR 0)),
359 (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
360 def : Pat<(insert_subvector undef, (v2f64 VR128:$src), (iPTR 0)),
361 (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
362 def : Pat<(insert_subvector undef, (v4i32 VR128:$src), (iPTR 0)),
363 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
364 def : Pat<(insert_subvector undef, (v4f32 VR128:$src), (iPTR 0)),
365 (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
366 def : Pat<(insert_subvector undef, (v8i16 VR128:$src), (iPTR 0)),
367 (INSERT_SUBREG (v16i16 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
368 def : Pat<(insert_subvector undef, (v16i8 VR128:$src), (iPTR 0)),
369 (INSERT_SUBREG (v32i8 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
372 // Implicitly promote a 32-bit scalar to a vector.
373 def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
374 (COPY_TO_REGCLASS FR32:$src, VR128)>;
375 def : Pat<(v8f32 (scalar_to_vector FR32:$src)),
376 (COPY_TO_REGCLASS FR32:$src, VR128)>;
377 // Implicitly promote a 64-bit scalar to a vector.
378 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
379 (COPY_TO_REGCLASS FR64:$src, VR128)>;
380 def : Pat<(v4f64 (scalar_to_vector FR64:$src)),
381 (COPY_TO_REGCLASS FR64:$src, VR128)>;
383 // Bitcasts between 128-bit vector types. Return the original type since
384 // no instruction is needed for the conversion
385 let Predicates = [HasSSE2] in {
386 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
387 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
388 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
389 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
390 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
391 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
392 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
393 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
394 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
395 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
396 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
397 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
398 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
399 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
400 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
401 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
402 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
403 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
404 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
405 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
406 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
407 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
408 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
409 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
410 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
411 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
412 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
413 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
414 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
415 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
418 // Bitcasts between 256-bit vector types. Return the original type since
419 // no instruction is needed for the conversion
420 let Predicates = [HasAVX] in {
421 def : Pat<(v4f64 (bitconvert (v8f32 VR256:$src))), (v4f64 VR256:$src)>;
422 def : Pat<(v4f64 (bitconvert (v8i32 VR256:$src))), (v4f64 VR256:$src)>;
423 def : Pat<(v4f64 (bitconvert (v4i64 VR256:$src))), (v4f64 VR256:$src)>;
424 def : Pat<(v4f64 (bitconvert (v16i16 VR256:$src))), (v4f64 VR256:$src)>;
425 def : Pat<(v4f64 (bitconvert (v32i8 VR256:$src))), (v4f64 VR256:$src)>;
426 def : Pat<(v8f32 (bitconvert (v8i32 VR256:$src))), (v8f32 VR256:$src)>;
427 def : Pat<(v8f32 (bitconvert (v4i64 VR256:$src))), (v8f32 VR256:$src)>;
428 def : Pat<(v8f32 (bitconvert (v4f64 VR256:$src))), (v8f32 VR256:$src)>;
429 def : Pat<(v8f32 (bitconvert (v32i8 VR256:$src))), (v8f32 VR256:$src)>;
430 def : Pat<(v8f32 (bitconvert (v16i16 VR256:$src))), (v8f32 VR256:$src)>;
431 def : Pat<(v4i64 (bitconvert (v8f32 VR256:$src))), (v4i64 VR256:$src)>;
432 def : Pat<(v4i64 (bitconvert (v8i32 VR256:$src))), (v4i64 VR256:$src)>;
433 def : Pat<(v4i64 (bitconvert (v4f64 VR256:$src))), (v4i64 VR256:$src)>;
434 def : Pat<(v4i64 (bitconvert (v32i8 VR256:$src))), (v4i64 VR256:$src)>;
435 def : Pat<(v4i64 (bitconvert (v16i16 VR256:$src))), (v4i64 VR256:$src)>;
436 def : Pat<(v32i8 (bitconvert (v4f64 VR256:$src))), (v32i8 VR256:$src)>;
437 def : Pat<(v32i8 (bitconvert (v4i64 VR256:$src))), (v32i8 VR256:$src)>;
438 def : Pat<(v32i8 (bitconvert (v8f32 VR256:$src))), (v32i8 VR256:$src)>;
439 def : Pat<(v32i8 (bitconvert (v8i32 VR256:$src))), (v32i8 VR256:$src)>;
440 def : Pat<(v32i8 (bitconvert (v16i16 VR256:$src))), (v32i8 VR256:$src)>;
441 def : Pat<(v8i32 (bitconvert (v32i8 VR256:$src))), (v8i32 VR256:$src)>;
442 def : Pat<(v8i32 (bitconvert (v16i16 VR256:$src))), (v8i32 VR256:$src)>;
443 def : Pat<(v8i32 (bitconvert (v8f32 VR256:$src))), (v8i32 VR256:$src)>;
444 def : Pat<(v8i32 (bitconvert (v4i64 VR256:$src))), (v8i32 VR256:$src)>;
445 def : Pat<(v8i32 (bitconvert (v4f64 VR256:$src))), (v8i32 VR256:$src)>;
446 def : Pat<(v16i16 (bitconvert (v8f32 VR256:$src))), (v16i16 VR256:$src)>;
447 def : Pat<(v16i16 (bitconvert (v8i32 VR256:$src))), (v16i16 VR256:$src)>;
448 def : Pat<(v16i16 (bitconvert (v4i64 VR256:$src))), (v16i16 VR256:$src)>;
449 def : Pat<(v16i16 (bitconvert (v4f64 VR256:$src))), (v16i16 VR256:$src)>;
450 def : Pat<(v16i16 (bitconvert (v32i8 VR256:$src))), (v16i16 VR256:$src)>;
453 // Alias instructions that map fld0 to xorps for sse or vxorps for avx.
454 // This is expanded by ExpandPostRAPseudos.
455 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
456 isPseudo = 1, SchedRW = [WriteZero] in {
457 def FsFLD0SS : I<0, Pseudo, (outs FR32:$dst), (ins), "",
458 [(set FR32:$dst, fp32imm0)]>, Requires<[HasSSE1]>;
459 def FsFLD0SD : I<0, Pseudo, (outs FR64:$dst), (ins), "",
460 [(set FR64:$dst, fpimm0)]>, Requires<[HasSSE2]>;
463 //===----------------------------------------------------------------------===//
464 // AVX & SSE - Zero/One Vectors
465 //===----------------------------------------------------------------------===//
467 // Alias instruction that maps zero vector to pxor / xorp* for sse.
468 // This is expanded by ExpandPostRAPseudos to an xorps / vxorps, and then
469 // swizzled by ExecutionDepsFix to pxor.
470 // We set canFoldAsLoad because this can be converted to a constant-pool
471 // load of an all-zeros value if folding it would be beneficial.
472 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
473 isPseudo = 1, SchedRW = [WriteZero] in {
474 def V_SET0 : I<0, Pseudo, (outs VR128:$dst), (ins), "",
475 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
478 def : Pat<(v2f64 immAllZerosV), (V_SET0)>;
479 def : Pat<(v4i32 immAllZerosV), (V_SET0)>;
480 def : Pat<(v2i64 immAllZerosV), (V_SET0)>;
481 def : Pat<(v8i16 immAllZerosV), (V_SET0)>;
482 def : Pat<(v16i8 immAllZerosV), (V_SET0)>;
485 // The same as done above but for AVX. The 256-bit AVX1 ISA doesn't support PI,
486 // and doesn't need it because on sandy bridge the register is set to zero
487 // at the rename stage without using any execution unit, so SET0PSY
488 // and SET0PDY can be used for vector int instructions without penalty
489 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
490 isPseudo = 1, Predicates = [HasAVX], SchedRW = [WriteZero] in {
491 def AVX_SET0 : I<0, Pseudo, (outs VR256:$dst), (ins), "",
492 [(set VR256:$dst, (v8f32 immAllZerosV))]>;
495 let Predicates = [HasAVX] in
496 def : Pat<(v4f64 immAllZerosV), (AVX_SET0)>;
498 let Predicates = [HasAVX2] in {
499 def : Pat<(v4i64 immAllZerosV), (AVX_SET0)>;
500 def : Pat<(v8i32 immAllZerosV), (AVX_SET0)>;
501 def : Pat<(v16i16 immAllZerosV), (AVX_SET0)>;
502 def : Pat<(v32i8 immAllZerosV), (AVX_SET0)>;
505 // AVX1 has no support for 256-bit integer instructions, but since the 128-bit
506 // VPXOR instruction writes zero to its upper part, it's safe build zeros.
507 let Predicates = [HasAVX1Only] in {
508 def : Pat<(v32i8 immAllZerosV), (SUBREG_TO_REG (i8 0), (V_SET0), sub_xmm)>;
509 def : Pat<(bc_v32i8 (v8f32 immAllZerosV)),
510 (SUBREG_TO_REG (i8 0), (V_SET0), sub_xmm)>;
512 def : Pat<(v16i16 immAllZerosV), (SUBREG_TO_REG (i16 0), (V_SET0), sub_xmm)>;
513 def : Pat<(bc_v16i16 (v8f32 immAllZerosV)),
514 (SUBREG_TO_REG (i16 0), (V_SET0), sub_xmm)>;
516 def : Pat<(v8i32 immAllZerosV), (SUBREG_TO_REG (i32 0), (V_SET0), sub_xmm)>;
517 def : Pat<(bc_v8i32 (v8f32 immAllZerosV)),
518 (SUBREG_TO_REG (i32 0), (V_SET0), sub_xmm)>;
520 def : Pat<(v4i64 immAllZerosV), (SUBREG_TO_REG (i64 0), (V_SET0), sub_xmm)>;
521 def : Pat<(bc_v4i64 (v8f32 immAllZerosV)),
522 (SUBREG_TO_REG (i64 0), (V_SET0), sub_xmm)>;
525 // We set canFoldAsLoad because this can be converted to a constant-pool
526 // load of an all-ones value if folding it would be beneficial.
527 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
528 isPseudo = 1, SchedRW = [WriteZero] in {
529 def V_SETALLONES : I<0, Pseudo, (outs VR128:$dst), (ins), "",
530 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
531 let Predicates = [HasAVX2] in
532 def AVX2_SETALLONES : I<0, Pseudo, (outs VR256:$dst), (ins), "",
533 [(set VR256:$dst, (v8i32 immAllOnesV))]>;
537 //===----------------------------------------------------------------------===//
538 // SSE 1 & 2 - Move FP Scalar Instructions
540 // Move Instructions. Register-to-register movss/movsd is not used for FR32/64
541 // register copies because it's a partial register update; Register-to-register
542 // movss/movsd is not modeled as an INSERT_SUBREG because INSERT_SUBREG requires
543 // that the insert be implementable in terms of a copy, and just mentioned, we
544 // don't use movss/movsd for copies.
545 //===----------------------------------------------------------------------===//
547 multiclass sse12_move_rr<RegisterClass RC, SDNode OpNode, ValueType vt,
548 X86MemOperand x86memop, string base_opc,
549 string asm_opr, Domain d = GenericDomain> {
550 def rr : SI<0x10, MRMSrcReg, (outs VR128:$dst),
551 (ins VR128:$src1, RC:$src2),
552 !strconcat(base_opc, asm_opr),
553 [(set VR128:$dst, (vt (OpNode VR128:$src1,
554 (scalar_to_vector RC:$src2))))],
555 IIC_SSE_MOV_S_RR, d>, Sched<[WriteFShuffle]>;
557 // For the disassembler
558 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in
559 def rr_REV : SI<0x11, MRMDestReg, (outs VR128:$dst),
560 (ins VR128:$src1, RC:$src2),
561 !strconcat(base_opc, asm_opr),
562 [], IIC_SSE_MOV_S_RR>, Sched<[WriteFShuffle]>;
565 multiclass sse12_move<RegisterClass RC, SDNode OpNode, ValueType vt,
566 X86MemOperand x86memop, string OpcodeStr,
567 Domain d = GenericDomain> {
569 defm V#NAME : sse12_move_rr<RC, OpNode, vt, x86memop, OpcodeStr,
570 "\t{$src2, $src1, $dst|$dst, $src1, $src2}", d>,
573 def V#NAME#mr : SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, RC:$src),
574 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
575 [(store RC:$src, addr:$dst)], IIC_SSE_MOV_S_MR, d>,
576 VEX, VEX_LIG, Sched<[WriteStore]>;
578 let Constraints = "$src1 = $dst" in {
579 defm NAME : sse12_move_rr<RC, OpNode, vt, x86memop, OpcodeStr,
580 "\t{$src2, $dst|$dst, $src2}", d>;
583 def NAME#mr : SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, RC:$src),
584 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
585 [(store RC:$src, addr:$dst)], IIC_SSE_MOV_S_MR, d>,
589 // Loading from memory automatically zeroing upper bits.
590 multiclass sse12_move_rm<RegisterClass RC, X86MemOperand x86memop,
591 PatFrag mem_pat, string OpcodeStr,
592 Domain d = GenericDomain> {
593 def V#NAME#rm : SI<0x10, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
594 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
595 [(set RC:$dst, (mem_pat addr:$src))],
596 IIC_SSE_MOV_S_RM, d>, VEX, VEX_LIG, Sched<[WriteLoad]>;
597 def NAME#rm : SI<0x10, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
598 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
599 [(set RC:$dst, (mem_pat addr:$src))],
600 IIC_SSE_MOV_S_RM, d>, Sched<[WriteLoad]>;
603 defm MOVSS : sse12_move<FR32, X86Movss, v4f32, f32mem, "movss",
604 SSEPackedSingle>, XS;
605 defm MOVSD : sse12_move<FR64, X86Movsd, v2f64, f64mem, "movsd",
606 SSEPackedDouble>, XD;
608 let canFoldAsLoad = 1, isReMaterializable = 1 in {
609 defm MOVSS : sse12_move_rm<FR32, f32mem, loadf32, "movss",
610 SSEPackedSingle>, XS;
612 let AddedComplexity = 20 in
613 defm MOVSD : sse12_move_rm<FR64, f64mem, loadf64, "movsd",
614 SSEPackedDouble>, XD;
618 let Predicates = [UseAVX] in {
619 let AddedComplexity = 20 in {
620 // MOVSSrm zeros the high parts of the register; represent this
621 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
622 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
623 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
624 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
625 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
626 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
627 (COPY_TO_REGCLASS (VMOVSSrm addr:$src), VR128)>;
629 // MOVSDrm zeros the high parts of the register; represent this
630 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
631 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
632 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
633 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
634 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
635 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
636 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
637 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
638 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
639 def : Pat<(v2f64 (X86vzload addr:$src)),
640 (COPY_TO_REGCLASS (VMOVSDrm addr:$src), VR128)>;
642 // Represent the same patterns above but in the form they appear for
644 def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
645 (v4f32 (scalar_to_vector (loadf32 addr:$src))), (iPTR 0)))),
646 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_xmm)>;
647 def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
648 (v2f64 (scalar_to_vector (loadf64 addr:$src))), (iPTR 0)))),
649 (SUBREG_TO_REG (i32 0), (VMOVSDrm addr:$src), sub_xmm)>;
652 // Extract and store.
653 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
655 (VMOVSSmr addr:$dst, (COPY_TO_REGCLASS (v4f32 VR128:$src), FR32))>;
656 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
658 (VMOVSDmr addr:$dst, (COPY_TO_REGCLASS (v2f64 VR128:$src), FR64))>;
660 // Shuffle with VMOVSS
661 def : Pat<(v4i32 (X86Movss VR128:$src1, VR128:$src2)),
662 (VMOVSSrr (v4i32 VR128:$src1),
663 (COPY_TO_REGCLASS (v4i32 VR128:$src2), FR32))>;
664 def : Pat<(v4f32 (X86Movss VR128:$src1, VR128:$src2)),
665 (VMOVSSrr (v4f32 VR128:$src1),
666 (COPY_TO_REGCLASS (v4f32 VR128:$src2), FR32))>;
669 def : Pat<(v8i32 (X86Movss VR256:$src1, VR256:$src2)),
670 (SUBREG_TO_REG (i32 0),
671 (VMOVSSrr (EXTRACT_SUBREG (v8i32 VR256:$src1), sub_xmm),
672 (EXTRACT_SUBREG (v8i32 VR256:$src2), sub_xmm)),
674 def : Pat<(v8f32 (X86Movss VR256:$src1, VR256:$src2)),
675 (SUBREG_TO_REG (i32 0),
676 (VMOVSSrr (EXTRACT_SUBREG (v8f32 VR256:$src1), sub_xmm),
677 (EXTRACT_SUBREG (v8f32 VR256:$src2), sub_xmm)),
680 // Shuffle with VMOVSD
681 def : Pat<(v2i64 (X86Movsd VR128:$src1, VR128:$src2)),
682 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
683 def : Pat<(v2f64 (X86Movsd VR128:$src1, VR128:$src2)),
684 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
685 def : Pat<(v4f32 (X86Movsd VR128:$src1, VR128:$src2)),
686 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
687 def : Pat<(v4i32 (X86Movsd VR128:$src1, VR128:$src2)),
688 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
691 def : Pat<(v4i64 (X86Movsd VR256:$src1, VR256:$src2)),
692 (SUBREG_TO_REG (i32 0),
693 (VMOVSDrr (EXTRACT_SUBREG (v4i64 VR256:$src1), sub_xmm),
694 (EXTRACT_SUBREG (v4i64 VR256:$src2), sub_xmm)),
696 def : Pat<(v4f64 (X86Movsd VR256:$src1, VR256:$src2)),
697 (SUBREG_TO_REG (i32 0),
698 (VMOVSDrr (EXTRACT_SUBREG (v4f64 VR256:$src1), sub_xmm),
699 (EXTRACT_SUBREG (v4f64 VR256:$src2), sub_xmm)),
702 // FIXME: Instead of a X86Movlps there should be a X86Movsd here, the problem
703 // is during lowering, where it's not possible to recognize the fold cause
704 // it has two uses through a bitcast. One use disappears at isel time and the
705 // fold opportunity reappears.
706 def : Pat<(v2f64 (X86Movlpd VR128:$src1, VR128:$src2)),
707 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
708 def : Pat<(v2i64 (X86Movlpd VR128:$src1, VR128:$src2)),
709 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
710 def : Pat<(v4f32 (X86Movlps VR128:$src1, VR128:$src2)),
711 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
712 def : Pat<(v4i32 (X86Movlps VR128:$src1, VR128:$src2)),
713 (VMOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
716 let Predicates = [UseSSE1] in {
717 let Predicates = [NoSSE41], AddedComplexity = 15 in {
718 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
719 // MOVSS to the lower bits.
720 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
721 (MOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
722 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
723 (MOVSSrr (v4f32 (V_SET0)), (COPY_TO_REGCLASS VR128:$src, FR32))>;
724 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
725 (MOVSSrr (v4i32 (V_SET0)), (COPY_TO_REGCLASS VR128:$src, FR32))>;
728 let AddedComplexity = 20 in {
729 // MOVSSrm already zeros the high parts of the register.
730 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
731 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
732 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
733 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
734 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
735 (COPY_TO_REGCLASS (MOVSSrm addr:$src), VR128)>;
738 // Extract and store.
739 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
741 (MOVSSmr addr:$dst, (COPY_TO_REGCLASS VR128:$src, FR32))>;
743 // Shuffle with MOVSS
744 def : Pat<(v4i32 (X86Movss VR128:$src1, VR128:$src2)),
745 (MOVSSrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR32))>;
746 def : Pat<(v4f32 (X86Movss VR128:$src1, VR128:$src2)),
747 (MOVSSrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR32))>;
750 let Predicates = [UseSSE2] in {
751 let Predicates = [NoSSE41], AddedComplexity = 15 in {
752 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
753 // MOVSD to the lower bits.
754 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
755 (MOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
758 let AddedComplexity = 20 in {
759 // MOVSDrm already zeros the high parts of the register.
760 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
761 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
762 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
763 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
764 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
765 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
766 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
767 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
768 def : Pat<(v2f64 (X86vzload addr:$src)),
769 (COPY_TO_REGCLASS (MOVSDrm addr:$src), VR128)>;
772 // Extract and store.
773 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
775 (MOVSDmr addr:$dst, (COPY_TO_REGCLASS VR128:$src, FR64))>;
777 // Shuffle with MOVSD
778 def : Pat<(v2i64 (X86Movsd VR128:$src1, VR128:$src2)),
779 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
780 def : Pat<(v2f64 (X86Movsd VR128:$src1, VR128:$src2)),
781 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
782 def : Pat<(v4f32 (X86Movsd VR128:$src1, VR128:$src2)),
783 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
784 def : Pat<(v4i32 (X86Movsd VR128:$src1, VR128:$src2)),
785 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
787 // FIXME: Instead of a X86Movlps there should be a X86Movsd here, the problem
788 // is during lowering, where it's not possible to recognize the fold because
789 // it has two uses through a bitcast. One use disappears at isel time and the
790 // fold opportunity reappears.
791 def : Pat<(v2f64 (X86Movlpd VR128:$src1, VR128:$src2)),
792 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
793 def : Pat<(v2i64 (X86Movlpd VR128:$src1, VR128:$src2)),
794 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
795 def : Pat<(v4f32 (X86Movlps VR128:$src1, VR128:$src2)),
796 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
797 def : Pat<(v4i32 (X86Movlps VR128:$src1, VR128:$src2)),
798 (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
801 //===----------------------------------------------------------------------===//
802 // SSE 1 & 2 - Move Aligned/Unaligned FP Instructions
803 //===----------------------------------------------------------------------===//
805 multiclass sse12_mov_packed<bits<8> opc, RegisterClass RC,
806 X86MemOperand x86memop, PatFrag ld_frag,
807 string asm, Domain d,
809 bit IsReMaterializable = 1> {
810 let hasSideEffects = 0 in
811 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
812 !strconcat(asm, "\t{$src, $dst|$dst, $src}"), [], itins.rr, d>,
813 Sched<[WriteFShuffle]>;
814 let canFoldAsLoad = 1, isReMaterializable = IsReMaterializable in
815 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
816 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
817 [(set RC:$dst, (ld_frag addr:$src))], itins.rm, d>,
821 let Predicates = [HasAVX, NoVLX] in {
822 defm VMOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
823 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
825 defm VMOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
826 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
828 defm VMOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
829 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
831 defm VMOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
832 "movupd", SSEPackedDouble, SSE_MOVU_ITINS, 0>,
835 defm VMOVAPSY : sse12_mov_packed<0x28, VR256, f256mem, alignedloadv8f32,
836 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
838 defm VMOVAPDY : sse12_mov_packed<0x28, VR256, f256mem, alignedloadv4f64,
839 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
841 defm VMOVUPSY : sse12_mov_packed<0x10, VR256, f256mem, loadv8f32,
842 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
844 defm VMOVUPDY : sse12_mov_packed<0x10, VR256, f256mem, loadv4f64,
845 "movupd", SSEPackedDouble, SSE_MOVU_ITINS, 0>,
849 let Predicates = [UseSSE1] in {
850 defm MOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
851 "movaps", SSEPackedSingle, SSE_MOVA_ITINS>,
853 defm MOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
854 "movups", SSEPackedSingle, SSE_MOVU_ITINS>,
857 let Predicates = [UseSSE2] in {
858 defm MOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
859 "movapd", SSEPackedDouble, SSE_MOVA_ITINS>,
861 defm MOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
862 "movupd", SSEPackedDouble, SSE_MOVU_ITINS, 0>,
866 let SchedRW = [WriteStore], Predicates = [HasAVX, NoVLX] in {
867 def VMOVAPSmr : VPSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
868 "movaps\t{$src, $dst|$dst, $src}",
869 [(alignedstore (v4f32 VR128:$src), addr:$dst)],
870 IIC_SSE_MOVA_P_MR>, VEX;
871 def VMOVAPDmr : VPDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
872 "movapd\t{$src, $dst|$dst, $src}",
873 [(alignedstore (v2f64 VR128:$src), addr:$dst)],
874 IIC_SSE_MOVA_P_MR>, VEX;
875 def VMOVUPSmr : VPSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
876 "movups\t{$src, $dst|$dst, $src}",
877 [(store (v4f32 VR128:$src), addr:$dst)],
878 IIC_SSE_MOVU_P_MR>, VEX;
879 def VMOVUPDmr : VPDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
880 "movupd\t{$src, $dst|$dst, $src}",
881 [(store (v2f64 VR128:$src), addr:$dst)],
882 IIC_SSE_MOVU_P_MR>, VEX;
883 def VMOVAPSYmr : VPSI<0x29, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
884 "movaps\t{$src, $dst|$dst, $src}",
885 [(alignedstore256 (v8f32 VR256:$src), addr:$dst)],
886 IIC_SSE_MOVA_P_MR>, VEX, VEX_L;
887 def VMOVAPDYmr : VPDI<0x29, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
888 "movapd\t{$src, $dst|$dst, $src}",
889 [(alignedstore256 (v4f64 VR256:$src), addr:$dst)],
890 IIC_SSE_MOVA_P_MR>, VEX, VEX_L;
891 def VMOVUPSYmr : VPSI<0x11, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
892 "movups\t{$src, $dst|$dst, $src}",
893 [(store (v8f32 VR256:$src), addr:$dst)],
894 IIC_SSE_MOVU_P_MR>, VEX, VEX_L;
895 def VMOVUPDYmr : VPDI<0x11, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
896 "movupd\t{$src, $dst|$dst, $src}",
897 [(store (v4f64 VR256:$src), addr:$dst)],
898 IIC_SSE_MOVU_P_MR>, VEX, VEX_L;
902 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
903 SchedRW = [WriteFShuffle] in {
904 def VMOVAPSrr_REV : VPSI<0x29, MRMDestReg, (outs VR128:$dst),
906 "movaps\t{$src, $dst|$dst, $src}", [],
907 IIC_SSE_MOVA_P_RR>, VEX;
908 def VMOVAPDrr_REV : VPDI<0x29, MRMDestReg, (outs VR128:$dst),
910 "movapd\t{$src, $dst|$dst, $src}", [],
911 IIC_SSE_MOVA_P_RR>, VEX;
912 def VMOVUPSrr_REV : VPSI<0x11, MRMDestReg, (outs VR128:$dst),
914 "movups\t{$src, $dst|$dst, $src}", [],
915 IIC_SSE_MOVU_P_RR>, VEX;
916 def VMOVUPDrr_REV : VPDI<0x11, MRMDestReg, (outs VR128:$dst),
918 "movupd\t{$src, $dst|$dst, $src}", [],
919 IIC_SSE_MOVU_P_RR>, VEX;
920 def VMOVAPSYrr_REV : VPSI<0x29, MRMDestReg, (outs VR256:$dst),
922 "movaps\t{$src, $dst|$dst, $src}", [],
923 IIC_SSE_MOVA_P_RR>, VEX, VEX_L;
924 def VMOVAPDYrr_REV : VPDI<0x29, MRMDestReg, (outs VR256:$dst),
926 "movapd\t{$src, $dst|$dst, $src}", [],
927 IIC_SSE_MOVA_P_RR>, VEX, VEX_L;
928 def VMOVUPSYrr_REV : VPSI<0x11, MRMDestReg, (outs VR256:$dst),
930 "movups\t{$src, $dst|$dst, $src}", [],
931 IIC_SSE_MOVU_P_RR>, VEX, VEX_L;
932 def VMOVUPDYrr_REV : VPDI<0x11, MRMDestReg, (outs VR256:$dst),
934 "movupd\t{$src, $dst|$dst, $src}", [],
935 IIC_SSE_MOVU_P_RR>, VEX, VEX_L;
938 let Predicates = [HasAVX] in {
939 def : Pat<(v8i32 (X86vzmovl
940 (insert_subvector undef, (v4i32 VR128:$src), (iPTR 0)))),
941 (SUBREG_TO_REG (i32 0), (VMOVAPSrr VR128:$src), sub_xmm)>;
942 def : Pat<(v4i64 (X86vzmovl
943 (insert_subvector undef, (v2i64 VR128:$src), (iPTR 0)))),
944 (SUBREG_TO_REG (i32 0), (VMOVAPSrr VR128:$src), sub_xmm)>;
945 def : Pat<(v8f32 (X86vzmovl
946 (insert_subvector undef, (v4f32 VR128:$src), (iPTR 0)))),
947 (SUBREG_TO_REG (i32 0), (VMOVAPSrr VR128:$src), sub_xmm)>;
948 def : Pat<(v4f64 (X86vzmovl
949 (insert_subvector undef, (v2f64 VR128:$src), (iPTR 0)))),
950 (SUBREG_TO_REG (i32 0), (VMOVAPSrr VR128:$src), sub_xmm)>;
954 def : Pat<(int_x86_avx_storeu_ps_256 addr:$dst, VR256:$src),
955 (VMOVUPSYmr addr:$dst, VR256:$src)>;
956 def : Pat<(int_x86_avx_storeu_pd_256 addr:$dst, VR256:$src),
957 (VMOVUPDYmr addr:$dst, VR256:$src)>;
959 let SchedRW = [WriteStore] in {
960 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
961 "movaps\t{$src, $dst|$dst, $src}",
962 [(alignedstore (v4f32 VR128:$src), addr:$dst)],
964 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
965 "movapd\t{$src, $dst|$dst, $src}",
966 [(alignedstore (v2f64 VR128:$src), addr:$dst)],
968 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
969 "movups\t{$src, $dst|$dst, $src}",
970 [(store (v4f32 VR128:$src), addr:$dst)],
972 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
973 "movupd\t{$src, $dst|$dst, $src}",
974 [(store (v2f64 VR128:$src), addr:$dst)],
979 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
980 SchedRW = [WriteFShuffle] in {
981 def MOVAPSrr_REV : PSI<0x29, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
982 "movaps\t{$src, $dst|$dst, $src}", [],
984 def MOVAPDrr_REV : PDI<0x29, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
985 "movapd\t{$src, $dst|$dst, $src}", [],
987 def MOVUPSrr_REV : PSI<0x11, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
988 "movups\t{$src, $dst|$dst, $src}", [],
990 def MOVUPDrr_REV : PDI<0x11, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
991 "movupd\t{$src, $dst|$dst, $src}", [],
995 let Predicates = [HasAVX] in {
996 def : Pat<(int_x86_sse_storeu_ps addr:$dst, VR128:$src),
997 (VMOVUPSmr addr:$dst, VR128:$src)>;
998 def : Pat<(int_x86_sse2_storeu_pd addr:$dst, VR128:$src),
999 (VMOVUPDmr addr:$dst, VR128:$src)>;
1002 let Predicates = [UseSSE1] in
1003 def : Pat<(int_x86_sse_storeu_ps addr:$dst, VR128:$src),
1004 (MOVUPSmr addr:$dst, VR128:$src)>;
1005 let Predicates = [UseSSE2] in
1006 def : Pat<(int_x86_sse2_storeu_pd addr:$dst, VR128:$src),
1007 (MOVUPDmr addr:$dst, VR128:$src)>;
1009 // Use vmovaps/vmovups for AVX integer load/store.
1010 let Predicates = [HasAVX, NoVLX] in {
1011 // 128-bit load/store
1012 def : Pat<(alignedloadv2i64 addr:$src),
1013 (VMOVAPSrm addr:$src)>;
1014 def : Pat<(loadv2i64 addr:$src),
1015 (VMOVUPSrm addr:$src)>;
1017 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
1018 (VMOVAPSmr addr:$dst, VR128:$src)>;
1019 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
1020 (VMOVAPSmr addr:$dst, VR128:$src)>;
1021 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
1022 (VMOVAPSmr addr:$dst, VR128:$src)>;
1023 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
1024 (VMOVAPSmr addr:$dst, VR128:$src)>;
1025 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
1026 (VMOVUPSmr addr:$dst, VR128:$src)>;
1027 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
1028 (VMOVUPSmr addr:$dst, VR128:$src)>;
1029 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
1030 (VMOVUPSmr addr:$dst, VR128:$src)>;
1031 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
1032 (VMOVUPSmr addr:$dst, VR128:$src)>;
1034 // 256-bit load/store
1035 def : Pat<(alignedloadv4i64 addr:$src),
1036 (VMOVAPSYrm addr:$src)>;
1037 def : Pat<(loadv4i64 addr:$src),
1038 (VMOVUPSYrm addr:$src)>;
1039 def : Pat<(alignedstore256 (v4i64 VR256:$src), addr:$dst),
1040 (VMOVAPSYmr addr:$dst, VR256:$src)>;
1041 def : Pat<(alignedstore256 (v8i32 VR256:$src), addr:$dst),
1042 (VMOVAPSYmr addr:$dst, VR256:$src)>;
1043 def : Pat<(alignedstore256 (v16i16 VR256:$src), addr:$dst),
1044 (VMOVAPSYmr addr:$dst, VR256:$src)>;
1045 def : Pat<(alignedstore256 (v32i8 VR256:$src), addr:$dst),
1046 (VMOVAPSYmr addr:$dst, VR256:$src)>;
1047 def : Pat<(store (v4i64 VR256:$src), addr:$dst),
1048 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1049 def : Pat<(store (v8i32 VR256:$src), addr:$dst),
1050 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1051 def : Pat<(store (v16i16 VR256:$src), addr:$dst),
1052 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1053 def : Pat<(store (v32i8 VR256:$src), addr:$dst),
1054 (VMOVUPSYmr addr:$dst, VR256:$src)>;
1056 // Special patterns for storing subvector extracts of lower 128-bits
1057 // Its cheaper to just use VMOVAPS/VMOVUPS instead of VEXTRACTF128mr
1058 def : Pat<(alignedstore (v2f64 (extract_subvector
1059 (v4f64 VR256:$src), (iPTR 0))), addr:$dst),
1060 (VMOVAPDmr addr:$dst, (v2f64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1061 def : Pat<(alignedstore (v4f32 (extract_subvector
1062 (v8f32 VR256:$src), (iPTR 0))), addr:$dst),
1063 (VMOVAPSmr addr:$dst, (v4f32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1064 def : Pat<(alignedstore (v2i64 (extract_subvector
1065 (v4i64 VR256:$src), (iPTR 0))), addr:$dst),
1066 (VMOVAPDmr addr:$dst, (v2i64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1067 def : Pat<(alignedstore (v4i32 (extract_subvector
1068 (v8i32 VR256:$src), (iPTR 0))), addr:$dst),
1069 (VMOVAPSmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1070 def : Pat<(alignedstore (v8i16 (extract_subvector
1071 (v16i16 VR256:$src), (iPTR 0))), addr:$dst),
1072 (VMOVAPSmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1073 def : Pat<(alignedstore (v16i8 (extract_subvector
1074 (v32i8 VR256:$src), (iPTR 0))), addr:$dst),
1075 (VMOVAPSmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1077 def : Pat<(store (v2f64 (extract_subvector
1078 (v4f64 VR256:$src), (iPTR 0))), addr:$dst),
1079 (VMOVUPDmr addr:$dst, (v2f64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1080 def : Pat<(store (v4f32 (extract_subvector
1081 (v8f32 VR256:$src), (iPTR 0))), addr:$dst),
1082 (VMOVUPSmr addr:$dst, (v4f32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1083 def : Pat<(store (v2i64 (extract_subvector
1084 (v4i64 VR256:$src), (iPTR 0))), addr:$dst),
1085 (VMOVUPDmr addr:$dst, (v2i64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1086 def : Pat<(store (v4i32 (extract_subvector
1087 (v8i32 VR256:$src), (iPTR 0))), addr:$dst),
1088 (VMOVUPSmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1089 def : Pat<(store (v8i16 (extract_subvector
1090 (v16i16 VR256:$src), (iPTR 0))), addr:$dst),
1091 (VMOVUPSmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1092 def : Pat<(store (v16i8 (extract_subvector
1093 (v32i8 VR256:$src), (iPTR 0))), addr:$dst),
1094 (VMOVUPSmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
1097 // Use movaps / movups for SSE integer load / store (one byte shorter).
1098 // The instructions selected below are then converted to MOVDQA/MOVDQU
1099 // during the SSE domain pass.
1100 let Predicates = [UseSSE1] in {
1101 def : Pat<(alignedloadv2i64 addr:$src),
1102 (MOVAPSrm addr:$src)>;
1103 def : Pat<(loadv2i64 addr:$src),
1104 (MOVUPSrm addr:$src)>;
1106 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
1107 (MOVAPSmr addr:$dst, VR128:$src)>;
1108 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
1109 (MOVAPSmr addr:$dst, VR128:$src)>;
1110 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
1111 (MOVAPSmr addr:$dst, VR128:$src)>;
1112 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
1113 (MOVAPSmr addr:$dst, VR128:$src)>;
1114 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
1115 (MOVUPSmr addr:$dst, VR128:$src)>;
1116 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
1117 (MOVUPSmr addr:$dst, VR128:$src)>;
1118 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
1119 (MOVUPSmr addr:$dst, VR128:$src)>;
1120 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
1121 (MOVUPSmr addr:$dst, VR128:$src)>;
1124 // Alias instruction to load FR32 or FR64 from f128mem using movaps. Upper
1125 // bits are disregarded. FIXME: Set encoding to pseudo!
1126 let canFoldAsLoad = 1, isReMaterializable = 1, SchedRW = [WriteLoad] in {
1127 let isCodeGenOnly = 1 in {
1128 def FsVMOVAPSrm : VPSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
1129 "movaps\t{$src, $dst|$dst, $src}",
1130 [(set FR32:$dst, (alignedloadfsf32 addr:$src))],
1131 IIC_SSE_MOVA_P_RM>, VEX;
1132 def FsVMOVAPDrm : VPDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1133 "movapd\t{$src, $dst|$dst, $src}",
1134 [(set FR64:$dst, (alignedloadfsf64 addr:$src))],
1135 IIC_SSE_MOVA_P_RM>, VEX;
1136 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
1137 "movaps\t{$src, $dst|$dst, $src}",
1138 [(set FR32:$dst, (alignedloadfsf32 addr:$src))],
1140 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1141 "movapd\t{$src, $dst|$dst, $src}",
1142 [(set FR64:$dst, (alignedloadfsf64 addr:$src))],
1147 //===----------------------------------------------------------------------===//
1148 // SSE 1 & 2 - Move Low packed FP Instructions
1149 //===----------------------------------------------------------------------===//
1151 multiclass sse12_mov_hilo_packed_base<bits<8>opc, SDNode psnode, SDNode pdnode,
1152 string base_opc, string asm_opr,
1153 InstrItinClass itin> {
1154 def PSrm : PI<opc, MRMSrcMem,
1155 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1156 !strconcat(base_opc, "s", asm_opr),
1158 (psnode VR128:$src1,
1159 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))],
1160 itin, SSEPackedSingle>, PS,
1161 Sched<[WriteFShuffleLd, ReadAfterLd]>;
1163 def PDrm : PI<opc, MRMSrcMem,
1164 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1165 !strconcat(base_opc, "d", asm_opr),
1166 [(set VR128:$dst, (v2f64 (pdnode VR128:$src1,
1167 (scalar_to_vector (loadf64 addr:$src2)))))],
1168 itin, SSEPackedDouble>, PD,
1169 Sched<[WriteFShuffleLd, ReadAfterLd]>;
1173 multiclass sse12_mov_hilo_packed<bits<8>opc, SDNode psnode, SDNode pdnode,
1174 string base_opc, InstrItinClass itin> {
1175 defm V#NAME : sse12_mov_hilo_packed_base<opc, psnode, pdnode, base_opc,
1176 "\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1179 let Constraints = "$src1 = $dst" in
1180 defm NAME : sse12_mov_hilo_packed_base<opc, psnode, pdnode, base_opc,
1181 "\t{$src2, $dst|$dst, $src2}",
1185 let AddedComplexity = 20 in {
1186 defm MOVL : sse12_mov_hilo_packed<0x12, X86Movlps, X86Movlpd, "movlp",
1190 let SchedRW = [WriteStore] in {
1191 def VMOVLPSmr : VPSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1192 "movlps\t{$src, $dst|$dst, $src}",
1193 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
1194 (iPTR 0))), addr:$dst)],
1195 IIC_SSE_MOV_LH>, VEX;
1196 def VMOVLPDmr : VPDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1197 "movlpd\t{$src, $dst|$dst, $src}",
1198 [(store (f64 (vector_extract (v2f64 VR128:$src),
1199 (iPTR 0))), addr:$dst)],
1200 IIC_SSE_MOV_LH>, VEX;
1201 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1202 "movlps\t{$src, $dst|$dst, $src}",
1203 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
1204 (iPTR 0))), addr:$dst)],
1206 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1207 "movlpd\t{$src, $dst|$dst, $src}",
1208 [(store (f64 (vector_extract (v2f64 VR128:$src),
1209 (iPTR 0))), addr:$dst)],
1213 let Predicates = [HasAVX] in {
1214 // Shuffle with VMOVLPS
1215 def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))),
1216 (VMOVLPSrm VR128:$src1, addr:$src2)>;
1217 def : Pat<(v4i32 (X86Movlps VR128:$src1, (load addr:$src2))),
1218 (VMOVLPSrm VR128:$src1, addr:$src2)>;
1220 // Shuffle with VMOVLPD
1221 def : Pat<(v2f64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1222 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1223 def : Pat<(v2i64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1224 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1225 def : Pat<(v2f64 (X86Movsd VR128:$src1,
1226 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))),
1227 (VMOVLPDrm VR128:$src1, addr:$src2)>;
1230 def : Pat<(store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)),
1232 (VMOVLPSmr addr:$src1, VR128:$src2)>;
1233 def : Pat<(store (v4i32 (X86Movlps
1234 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)), addr:$src1),
1235 (VMOVLPSmr addr:$src1, VR128:$src2)>;
1236 def : Pat<(store (v2f64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1238 (VMOVLPDmr addr:$src1, VR128:$src2)>;
1239 def : Pat<(store (v2i64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1241 (VMOVLPDmr addr:$src1, VR128:$src2)>;
1244 let Predicates = [UseSSE1] in {
1245 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
1246 def : Pat<(store (i64 (vector_extract (bc_v2i64 (v4f32 VR128:$src2)),
1247 (iPTR 0))), addr:$src1),
1248 (MOVLPSmr addr:$src1, VR128:$src2)>;
1250 // Shuffle with MOVLPS
1251 def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))),
1252 (MOVLPSrm VR128:$src1, addr:$src2)>;
1253 def : Pat<(v4i32 (X86Movlps VR128:$src1, (load addr:$src2))),
1254 (MOVLPSrm VR128:$src1, addr:$src2)>;
1255 def : Pat<(X86Movlps VR128:$src1,
1256 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1257 (MOVLPSrm VR128:$src1, addr:$src2)>;
1260 def : Pat<(store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)),
1262 (MOVLPSmr addr:$src1, VR128:$src2)>;
1263 def : Pat<(store (v4i32 (X86Movlps
1264 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
1266 (MOVLPSmr addr:$src1, VR128:$src2)>;
1269 let Predicates = [UseSSE2] in {
1270 // Shuffle with MOVLPD
1271 def : Pat<(v2f64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1272 (MOVLPDrm VR128:$src1, addr:$src2)>;
1273 def : Pat<(v2i64 (X86Movlpd VR128:$src1, (load addr:$src2))),
1274 (MOVLPDrm VR128:$src1, addr:$src2)>;
1275 def : Pat<(v2f64 (X86Movsd VR128:$src1,
1276 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))),
1277 (MOVLPDrm VR128:$src1, addr:$src2)>;
1280 def : Pat<(store (v2f64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1282 (MOVLPDmr addr:$src1, VR128:$src2)>;
1283 def : Pat<(store (v2i64 (X86Movlpd (load addr:$src1), VR128:$src2)),
1285 (MOVLPDmr addr:$src1, VR128:$src2)>;
1288 //===----------------------------------------------------------------------===//
1289 // SSE 1 & 2 - Move Hi packed FP Instructions
1290 //===----------------------------------------------------------------------===//
1292 let AddedComplexity = 20 in {
1293 defm MOVH : sse12_mov_hilo_packed<0x16, X86Movlhps, X86Movlhpd, "movhp",
1297 let SchedRW = [WriteStore] in {
1298 // v2f64 extract element 1 is always custom lowered to unpack high to low
1299 // and extract element 0 so the non-store version isn't too horrible.
1300 def VMOVHPSmr : VPSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1301 "movhps\t{$src, $dst|$dst, $src}",
1302 [(store (f64 (vector_extract
1303 (X86Unpckh (bc_v2f64 (v4f32 VR128:$src)),
1304 (bc_v2f64 (v4f32 VR128:$src))),
1305 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>, VEX;
1306 def VMOVHPDmr : VPDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1307 "movhpd\t{$src, $dst|$dst, $src}",
1308 [(store (f64 (vector_extract
1309 (v2f64 (X86Unpckh VR128:$src, VR128:$src)),
1310 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>, VEX;
1311 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1312 "movhps\t{$src, $dst|$dst, $src}",
1313 [(store (f64 (vector_extract
1314 (X86Unpckh (bc_v2f64 (v4f32 VR128:$src)),
1315 (bc_v2f64 (v4f32 VR128:$src))),
1316 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>;
1317 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1318 "movhpd\t{$src, $dst|$dst, $src}",
1319 [(store (f64 (vector_extract
1320 (v2f64 (X86Unpckh VR128:$src, VR128:$src)),
1321 (iPTR 0))), addr:$dst)], IIC_SSE_MOV_LH>;
1324 let Predicates = [HasAVX] in {
1326 def : Pat<(X86Movlhps VR128:$src1,
1327 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1328 (VMOVHPSrm VR128:$src1, addr:$src2)>;
1329 def : Pat<(X86Movlhps VR128:$src1,
1330 (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
1331 (VMOVHPSrm VR128:$src1, addr:$src2)>;
1335 // FIXME: Instead of X86Unpckl, there should be a X86Movlhpd here, the problem
1336 // is during lowering, where it's not possible to recognize the load fold
1337 // cause it has two uses through a bitcast. One use disappears at isel time
1338 // and the fold opportunity reappears.
1339 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1340 (scalar_to_vector (loadf64 addr:$src2)))),
1341 (VMOVHPDrm VR128:$src1, addr:$src2)>;
1342 // Also handle an i64 load because that may get selected as a faster way to
1344 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1345 (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src2)))))),
1346 (VMOVHPDrm VR128:$src1, addr:$src2)>;
1348 def : Pat<(store (f64 (vector_extract
1349 (v2f64 (X86VPermilpi VR128:$src, (i8 1))),
1350 (iPTR 0))), addr:$dst),
1351 (VMOVHPDmr addr:$dst, VR128:$src)>;
1354 let Predicates = [UseSSE1] in {
1356 def : Pat<(X86Movlhps VR128:$src1,
1357 (bc_v4f32 (v2i64 (scalar_to_vector (loadi64 addr:$src2))))),
1358 (MOVHPSrm VR128:$src1, addr:$src2)>;
1359 def : Pat<(X86Movlhps VR128:$src1,
1360 (bc_v4f32 (v2i64 (X86vzload addr:$src2)))),
1361 (MOVHPSrm VR128:$src1, addr:$src2)>;
1364 let Predicates = [UseSSE2] in {
1367 // FIXME: Instead of X86Unpckl, there should be a X86Movlhpd here, the problem
1368 // is during lowering, where it's not possible to recognize the load fold
1369 // cause it has two uses through a bitcast. One use disappears at isel time
1370 // and the fold opportunity reappears.
1371 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1372 (scalar_to_vector (loadf64 addr:$src2)))),
1373 (MOVHPDrm VR128:$src1, addr:$src2)>;
1374 // Also handle an i64 load because that may get selected as a faster way to
1376 def : Pat<(v2f64 (X86Unpckl VR128:$src1,
1377 (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src2)))))),
1378 (MOVHPDrm VR128:$src1, addr:$src2)>;
1380 def : Pat<(store (f64 (vector_extract
1381 (v2f64 (X86Shufp VR128:$src, VR128:$src, (i8 1))),
1382 (iPTR 0))), addr:$dst),
1383 (MOVHPDmr addr:$dst, VR128:$src)>;
1386 //===----------------------------------------------------------------------===//
1387 // SSE 1 & 2 - Move Low to High and High to Low packed FP Instructions
1388 //===----------------------------------------------------------------------===//
1390 let AddedComplexity = 20, Predicates = [UseAVX] in {
1391 def VMOVLHPSrr : VPSI<0x16, MRMSrcReg, (outs VR128:$dst),
1392 (ins VR128:$src1, VR128:$src2),
1393 "movlhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1395 (v4f32 (X86Movlhps VR128:$src1, VR128:$src2)))],
1397 VEX_4V, Sched<[WriteFShuffle]>;
1398 def VMOVHLPSrr : VPSI<0x12, MRMSrcReg, (outs VR128:$dst),
1399 (ins VR128:$src1, VR128:$src2),
1400 "movhlps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1402 (v4f32 (X86Movhlps VR128:$src1, VR128:$src2)))],
1404 VEX_4V, Sched<[WriteFShuffle]>;
1406 let Constraints = "$src1 = $dst", AddedComplexity = 20 in {
1407 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
1408 (ins VR128:$src1, VR128:$src2),
1409 "movlhps\t{$src2, $dst|$dst, $src2}",
1411 (v4f32 (X86Movlhps VR128:$src1, VR128:$src2)))],
1412 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
1413 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
1414 (ins VR128:$src1, VR128:$src2),
1415 "movhlps\t{$src2, $dst|$dst, $src2}",
1417 (v4f32 (X86Movhlps VR128:$src1, VR128:$src2)))],
1418 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
1421 let Predicates = [UseAVX] in {
1423 def : Pat<(v4i32 (X86Movlhps VR128:$src1, VR128:$src2)),
1424 (VMOVLHPSrr VR128:$src1, VR128:$src2)>;
1425 def : Pat<(v2i64 (X86Movlhps VR128:$src1, VR128:$src2)),
1426 (VMOVLHPSrr (v2i64 VR128:$src1), VR128:$src2)>;
1429 def : Pat<(v4i32 (X86Movhlps VR128:$src1, VR128:$src2)),
1430 (VMOVHLPSrr VR128:$src1, VR128:$src2)>;
1433 let Predicates = [UseSSE1] in {
1435 def : Pat<(v4i32 (X86Movlhps VR128:$src1, VR128:$src2)),
1436 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
1437 def : Pat<(v2i64 (X86Movlhps VR128:$src1, VR128:$src2)),
1438 (MOVLHPSrr (v2i64 VR128:$src1), VR128:$src2)>;
1441 def : Pat<(v4i32 (X86Movhlps VR128:$src1, VR128:$src2)),
1442 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
1445 //===----------------------------------------------------------------------===//
1446 // SSE 1 & 2 - Conversion Instructions
1447 //===----------------------------------------------------------------------===//
1449 def SSE_CVT_PD : OpndItins<
1450 IIC_SSE_CVT_PD_RR, IIC_SSE_CVT_PD_RM
1453 let Sched = WriteCvtI2F in
1454 def SSE_CVT_PS : OpndItins<
1455 IIC_SSE_CVT_PS_RR, IIC_SSE_CVT_PS_RM
1458 let Sched = WriteCvtI2F in
1459 def SSE_CVT_Scalar : OpndItins<
1460 IIC_SSE_CVT_Scalar_RR, IIC_SSE_CVT_Scalar_RM
1463 let Sched = WriteCvtF2I in
1464 def SSE_CVT_SS2SI_32 : OpndItins<
1465 IIC_SSE_CVT_SS2SI32_RR, IIC_SSE_CVT_SS2SI32_RM
1468 let Sched = WriteCvtF2I in
1469 def SSE_CVT_SS2SI_64 : OpndItins<
1470 IIC_SSE_CVT_SS2SI64_RR, IIC_SSE_CVT_SS2SI64_RM
1473 let Sched = WriteCvtF2I in
1474 def SSE_CVT_SD2SI : OpndItins<
1475 IIC_SSE_CVT_SD2SI_RR, IIC_SSE_CVT_SD2SI_RM
1478 multiclass sse12_cvt_s<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1479 SDNode OpNode, X86MemOperand x86memop, PatFrag ld_frag,
1480 string asm, OpndItins itins> {
1481 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
1482 [(set DstRC:$dst, (OpNode SrcRC:$src))],
1483 itins.rr>, Sched<[itins.Sched]>;
1484 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
1485 [(set DstRC:$dst, (OpNode (ld_frag addr:$src)))],
1486 itins.rm>, Sched<[itins.Sched.Folded]>;
1489 multiclass sse12_cvt_p<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1490 X86MemOperand x86memop, string asm, Domain d,
1492 let hasSideEffects = 0 in {
1493 def rr : I<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
1494 [], itins.rr, d>, Sched<[itins.Sched]>;
1496 def rm : I<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
1497 [], itins.rm, d>, Sched<[itins.Sched.Folded]>;
1501 multiclass sse12_vcvt_avx<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1502 X86MemOperand x86memop, string asm> {
1503 let hasSideEffects = 0, Predicates = [UseAVX] in {
1504 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src),
1505 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>,
1506 Sched<[WriteCvtI2F]>;
1508 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
1509 (ins DstRC:$src1, x86memop:$src),
1510 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>,
1511 Sched<[WriteCvtI2FLd, ReadAfterLd]>;
1512 } // hasSideEffects = 0
1515 let Predicates = [UseAVX] in {
1516 defm VCVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
1517 "cvttss2si\t{$src, $dst|$dst, $src}",
1520 defm VCVTTSS2SI64 : sse12_cvt_s<0x2C, FR32, GR64, fp_to_sint, f32mem, loadf32,
1521 "cvttss2si\t{$src, $dst|$dst, $src}",
1523 XS, VEX, VEX_W, VEX_LIG;
1524 defm VCVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
1525 "cvttsd2si\t{$src, $dst|$dst, $src}",
1528 defm VCVTTSD2SI64 : sse12_cvt_s<0x2C, FR64, GR64, fp_to_sint, f64mem, loadf64,
1529 "cvttsd2si\t{$src, $dst|$dst, $src}",
1531 XD, VEX, VEX_W, VEX_LIG;
1533 def : InstAlias<"vcvttss2si{l}\t{$src, $dst|$dst, $src}",
1534 (VCVTTSS2SIrr GR32:$dst, FR32:$src), 0>;
1535 def : InstAlias<"vcvttss2si{l}\t{$src, $dst|$dst, $src}",
1536 (VCVTTSS2SIrm GR32:$dst, f32mem:$src), 0>;
1537 def : InstAlias<"vcvttsd2si{l}\t{$src, $dst|$dst, $src}",
1538 (VCVTTSD2SIrr GR32:$dst, FR64:$src), 0>;
1539 def : InstAlias<"vcvttsd2si{l}\t{$src, $dst|$dst, $src}",
1540 (VCVTTSD2SIrm GR32:$dst, f64mem:$src), 0>;
1541 def : InstAlias<"vcvttss2si{q}\t{$src, $dst|$dst, $src}",
1542 (VCVTTSS2SI64rr GR64:$dst, FR32:$src), 0>;
1543 def : InstAlias<"vcvttss2si{q}\t{$src, $dst|$dst, $src}",
1544 (VCVTTSS2SI64rm GR64:$dst, f32mem:$src), 0>;
1545 def : InstAlias<"vcvttsd2si{q}\t{$src, $dst|$dst, $src}",
1546 (VCVTTSD2SI64rr GR64:$dst, FR64:$src), 0>;
1547 def : InstAlias<"vcvttsd2si{q}\t{$src, $dst|$dst, $src}",
1548 (VCVTTSD2SI64rm GR64:$dst, f64mem:$src), 0>;
1550 // The assembler can recognize rr 64-bit instructions by seeing a rxx
1551 // register, but the same isn't true when only using memory operands,
1552 // provide other assembly "l" and "q" forms to address this explicitly
1553 // where appropriate to do so.
1554 defm VCVTSI2SS : sse12_vcvt_avx<0x2A, GR32, FR32, i32mem, "cvtsi2ss{l}">,
1555 XS, VEX_4V, VEX_LIG;
1556 defm VCVTSI2SS64 : sse12_vcvt_avx<0x2A, GR64, FR32, i64mem, "cvtsi2ss{q}">,
1557 XS, VEX_4V, VEX_W, VEX_LIG;
1558 defm VCVTSI2SD : sse12_vcvt_avx<0x2A, GR32, FR64, i32mem, "cvtsi2sd{l}">,
1559 XD, VEX_4V, VEX_LIG;
1560 defm VCVTSI2SD64 : sse12_vcvt_avx<0x2A, GR64, FR64, i64mem, "cvtsi2sd{q}">,
1561 XD, VEX_4V, VEX_W, VEX_LIG;
1563 let Predicates = [UseAVX] in {
1564 def : InstAlias<"vcvtsi2ss\t{$src, $src1, $dst|$dst, $src1, $src}",
1565 (VCVTSI2SSrm FR64:$dst, FR64:$src1, i32mem:$src), 0>;
1566 def : InstAlias<"vcvtsi2sd\t{$src, $src1, $dst|$dst, $src1, $src}",
1567 (VCVTSI2SDrm FR64:$dst, FR64:$src1, i32mem:$src), 0>;
1569 def : Pat<(f32 (sint_to_fp (loadi32 addr:$src))),
1570 (VCVTSI2SSrm (f32 (IMPLICIT_DEF)), addr:$src)>;
1571 def : Pat<(f32 (sint_to_fp (loadi64 addr:$src))),
1572 (VCVTSI2SS64rm (f32 (IMPLICIT_DEF)), addr:$src)>;
1573 def : Pat<(f64 (sint_to_fp (loadi32 addr:$src))),
1574 (VCVTSI2SDrm (f64 (IMPLICIT_DEF)), addr:$src)>;
1575 def : Pat<(f64 (sint_to_fp (loadi64 addr:$src))),
1576 (VCVTSI2SD64rm (f64 (IMPLICIT_DEF)), addr:$src)>;
1578 def : Pat<(f32 (sint_to_fp GR32:$src)),
1579 (VCVTSI2SSrr (f32 (IMPLICIT_DEF)), GR32:$src)>;
1580 def : Pat<(f32 (sint_to_fp GR64:$src)),
1581 (VCVTSI2SS64rr (f32 (IMPLICIT_DEF)), GR64:$src)>;
1582 def : Pat<(f64 (sint_to_fp GR32:$src)),
1583 (VCVTSI2SDrr (f64 (IMPLICIT_DEF)), GR32:$src)>;
1584 def : Pat<(f64 (sint_to_fp GR64:$src)),
1585 (VCVTSI2SD64rr (f64 (IMPLICIT_DEF)), GR64:$src)>;
1588 defm CVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
1589 "cvttss2si\t{$src, $dst|$dst, $src}",
1590 SSE_CVT_SS2SI_32>, XS;
1591 defm CVTTSS2SI64 : sse12_cvt_s<0x2C, FR32, GR64, fp_to_sint, f32mem, loadf32,
1592 "cvttss2si\t{$src, $dst|$dst, $src}",
1593 SSE_CVT_SS2SI_64>, XS, REX_W;
1594 defm CVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
1595 "cvttsd2si\t{$src, $dst|$dst, $src}",
1597 defm CVTTSD2SI64 : sse12_cvt_s<0x2C, FR64, GR64, fp_to_sint, f64mem, loadf64,
1598 "cvttsd2si\t{$src, $dst|$dst, $src}",
1599 SSE_CVT_SD2SI>, XD, REX_W;
1600 defm CVTSI2SS : sse12_cvt_s<0x2A, GR32, FR32, sint_to_fp, i32mem, loadi32,
1601 "cvtsi2ss{l}\t{$src, $dst|$dst, $src}",
1602 SSE_CVT_Scalar>, XS;
1603 defm CVTSI2SS64 : sse12_cvt_s<0x2A, GR64, FR32, sint_to_fp, i64mem, loadi64,
1604 "cvtsi2ss{q}\t{$src, $dst|$dst, $src}",
1605 SSE_CVT_Scalar>, XS, REX_W;
1606 defm CVTSI2SD : sse12_cvt_s<0x2A, GR32, FR64, sint_to_fp, i32mem, loadi32,
1607 "cvtsi2sd{l}\t{$src, $dst|$dst, $src}",
1608 SSE_CVT_Scalar>, XD;
1609 defm CVTSI2SD64 : sse12_cvt_s<0x2A, GR64, FR64, sint_to_fp, i64mem, loadi64,
1610 "cvtsi2sd{q}\t{$src, $dst|$dst, $src}",
1611 SSE_CVT_Scalar>, XD, REX_W;
1613 def : InstAlias<"cvttss2si{l}\t{$src, $dst|$dst, $src}",
1614 (CVTTSS2SIrr GR32:$dst, FR32:$src), 0>;
1615 def : InstAlias<"cvttss2si{l}\t{$src, $dst|$dst, $src}",
1616 (CVTTSS2SIrm GR32:$dst, f32mem:$src), 0>;
1617 def : InstAlias<"cvttsd2si{l}\t{$src, $dst|$dst, $src}",
1618 (CVTTSD2SIrr GR32:$dst, FR64:$src), 0>;
1619 def : InstAlias<"cvttsd2si{l}\t{$src, $dst|$dst, $src}",
1620 (CVTTSD2SIrm GR32:$dst, f64mem:$src), 0>;
1621 def : InstAlias<"cvttss2si{q}\t{$src, $dst|$dst, $src}",
1622 (CVTTSS2SI64rr GR64:$dst, FR32:$src), 0>;
1623 def : InstAlias<"cvttss2si{q}\t{$src, $dst|$dst, $src}",
1624 (CVTTSS2SI64rm GR64:$dst, f32mem:$src), 0>;
1625 def : InstAlias<"cvttsd2si{q}\t{$src, $dst|$dst, $src}",
1626 (CVTTSD2SI64rr GR64:$dst, FR64:$src), 0>;
1627 def : InstAlias<"cvttsd2si{q}\t{$src, $dst|$dst, $src}",
1628 (CVTTSD2SI64rm GR64:$dst, f64mem:$src), 0>;
1630 def : InstAlias<"cvtsi2ss\t{$src, $dst|$dst, $src}",
1631 (CVTSI2SSrm FR64:$dst, i32mem:$src), 0>;
1632 def : InstAlias<"cvtsi2sd\t{$src, $dst|$dst, $src}",
1633 (CVTSI2SDrm FR64:$dst, i32mem:$src), 0>;
1635 // Conversion Instructions Intrinsics - Match intrinsics which expect MM
1636 // and/or XMM operand(s).
1638 multiclass sse12_cvt_sint<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
1639 Intrinsic Int, Operand memop, ComplexPattern mem_cpat,
1640 string asm, OpndItins itins> {
1641 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src),
1642 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1643 [(set DstRC:$dst, (Int SrcRC:$src))], itins.rr>,
1644 Sched<[itins.Sched]>;
1645 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins memop:$src),
1646 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1647 [(set DstRC:$dst, (Int mem_cpat:$src))], itins.rm>,
1648 Sched<[itins.Sched.Folded]>;
1651 multiclass sse12_cvt_sint_3addr<bits<8> opc, RegisterClass SrcRC,
1652 RegisterClass DstRC, Intrinsic Int, X86MemOperand x86memop,
1653 PatFrag ld_frag, string asm, OpndItins itins,
1655 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src2),
1657 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
1658 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
1659 [(set DstRC:$dst, (Int DstRC:$src1, SrcRC:$src2))],
1660 itins.rr>, Sched<[itins.Sched]>;
1661 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
1662 (ins DstRC:$src1, x86memop:$src2),
1664 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
1665 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
1666 [(set DstRC:$dst, (Int DstRC:$src1, (ld_frag addr:$src2)))],
1667 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
1670 let Predicates = [UseAVX] in {
1671 defm VCVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32,
1672 int_x86_sse2_cvtsd2si, sdmem, sse_load_f64, "cvtsd2si",
1673 SSE_CVT_SD2SI>, XD, VEX, VEX_LIG;
1674 defm VCVTSD2SI64 : sse12_cvt_sint<0x2D, VR128, GR64,
1675 int_x86_sse2_cvtsd2si64, sdmem, sse_load_f64, "cvtsd2si",
1676 SSE_CVT_SD2SI>, XD, VEX, VEX_W, VEX_LIG;
1678 defm CVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse2_cvtsd2si,
1679 sdmem, sse_load_f64, "cvtsd2si", SSE_CVT_SD2SI>, XD;
1680 defm CVTSD2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse2_cvtsd2si64,
1681 sdmem, sse_load_f64, "cvtsd2si", SSE_CVT_SD2SI>, XD, REX_W;
1684 let isCodeGenOnly = 1 in {
1685 let Predicates = [UseAVX] in {
1686 defm Int_VCVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1687 int_x86_sse_cvtsi2ss, i32mem, loadi32, "cvtsi2ss{l}",
1688 SSE_CVT_Scalar, 0>, XS, VEX_4V;
1689 defm Int_VCVTSI2SS64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1690 int_x86_sse_cvtsi642ss, i64mem, loadi64, "cvtsi2ss{q}",
1691 SSE_CVT_Scalar, 0>, XS, VEX_4V,
1693 defm Int_VCVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1694 int_x86_sse2_cvtsi2sd, i32mem, loadi32, "cvtsi2sd{l}",
1695 SSE_CVT_Scalar, 0>, XD, VEX_4V;
1696 defm Int_VCVTSI2SD64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1697 int_x86_sse2_cvtsi642sd, i64mem, loadi64, "cvtsi2sd{q}",
1698 SSE_CVT_Scalar, 0>, XD,
1701 let Constraints = "$src1 = $dst" in {
1702 defm Int_CVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1703 int_x86_sse_cvtsi2ss, i32mem, loadi32,
1704 "cvtsi2ss{l}", SSE_CVT_Scalar>, XS;
1705 defm Int_CVTSI2SS64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1706 int_x86_sse_cvtsi642ss, i64mem, loadi64,
1707 "cvtsi2ss{q}", SSE_CVT_Scalar>, XS, REX_W;
1708 defm Int_CVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
1709 int_x86_sse2_cvtsi2sd, i32mem, loadi32,
1710 "cvtsi2sd{l}", SSE_CVT_Scalar>, XD;
1711 defm Int_CVTSI2SD64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
1712 int_x86_sse2_cvtsi642sd, i64mem, loadi64,
1713 "cvtsi2sd{q}", SSE_CVT_Scalar>, XD, REX_W;
1715 } // isCodeGenOnly = 1
1719 // Aliases for intrinsics
1720 let isCodeGenOnly = 1 in {
1721 let Predicates = [UseAVX] in {
1722 defm Int_VCVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
1723 ssmem, sse_load_f32, "cvttss2si",
1724 SSE_CVT_SS2SI_32>, XS, VEX;
1725 defm Int_VCVTTSS2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1726 int_x86_sse_cvttss2si64, ssmem, sse_load_f32,
1727 "cvttss2si", SSE_CVT_SS2SI_64>,
1729 defm Int_VCVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
1730 sdmem, sse_load_f64, "cvttsd2si",
1731 SSE_CVT_SD2SI>, XD, VEX;
1732 defm Int_VCVTTSD2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1733 int_x86_sse2_cvttsd2si64, sdmem, sse_load_f64,
1734 "cvttsd2si", SSE_CVT_SD2SI>,
1737 defm Int_CVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
1738 ssmem, sse_load_f32, "cvttss2si",
1739 SSE_CVT_SS2SI_32>, XS;
1740 defm Int_CVTTSS2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1741 int_x86_sse_cvttss2si64, ssmem, sse_load_f32,
1742 "cvttss2si", SSE_CVT_SS2SI_64>, XS, REX_W;
1743 defm Int_CVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
1744 sdmem, sse_load_f64, "cvttsd2si",
1746 defm Int_CVTTSD2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
1747 int_x86_sse2_cvttsd2si64, sdmem, sse_load_f64,
1748 "cvttsd2si", SSE_CVT_SD2SI>, XD, REX_W;
1749 } // isCodeGenOnly = 1
1751 let Predicates = [UseAVX] in {
1752 defm VCVTSS2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse_cvtss2si,
1753 ssmem, sse_load_f32, "cvtss2si",
1754 SSE_CVT_SS2SI_32>, XS, VEX, VEX_LIG;
1755 defm VCVTSS2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse_cvtss2si64,
1756 ssmem, sse_load_f32, "cvtss2si",
1757 SSE_CVT_SS2SI_64>, XS, VEX, VEX_W, VEX_LIG;
1759 defm CVTSS2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse_cvtss2si,
1760 ssmem, sse_load_f32, "cvtss2si",
1761 SSE_CVT_SS2SI_32>, XS;
1762 defm CVTSS2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse_cvtss2si64,
1763 ssmem, sse_load_f32, "cvtss2si",
1764 SSE_CVT_SS2SI_64>, XS, REX_W;
1766 defm VCVTDQ2PS : sse12_cvt_p<0x5B, VR128, VR128, i128mem,
1767 "vcvtdq2ps\t{$src, $dst|$dst, $src}",
1768 SSEPackedSingle, SSE_CVT_PS>,
1769 PS, VEX, Requires<[HasAVX]>;
1770 defm VCVTDQ2PSY : sse12_cvt_p<0x5B, VR256, VR256, i256mem,
1771 "vcvtdq2ps\t{$src, $dst|$dst, $src}",
1772 SSEPackedSingle, SSE_CVT_PS>,
1773 PS, VEX, VEX_L, Requires<[HasAVX]>;
1775 defm CVTDQ2PS : sse12_cvt_p<0x5B, VR128, VR128, i128mem,
1776 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1777 SSEPackedSingle, SSE_CVT_PS>,
1778 PS, Requires<[UseSSE2]>;
1780 let Predicates = [UseAVX] in {
1781 def : InstAlias<"vcvtss2si{l}\t{$src, $dst|$dst, $src}",
1782 (VCVTSS2SIrr GR32:$dst, VR128:$src), 0>;
1783 def : InstAlias<"vcvtss2si{l}\t{$src, $dst|$dst, $src}",
1784 (VCVTSS2SIrm GR32:$dst, ssmem:$src), 0>;
1785 def : InstAlias<"vcvtsd2si{l}\t{$src, $dst|$dst, $src}",
1786 (VCVTSD2SIrr GR32:$dst, VR128:$src), 0>;
1787 def : InstAlias<"vcvtsd2si{l}\t{$src, $dst|$dst, $src}",
1788 (VCVTSD2SIrm GR32:$dst, sdmem:$src), 0>;
1789 def : InstAlias<"vcvtss2si{q}\t{$src, $dst|$dst, $src}",
1790 (VCVTSS2SI64rr GR64:$dst, VR128:$src), 0>;
1791 def : InstAlias<"vcvtss2si{q}\t{$src, $dst|$dst, $src}",
1792 (VCVTSS2SI64rm GR64:$dst, ssmem:$src), 0>;
1793 def : InstAlias<"vcvtsd2si{q}\t{$src, $dst|$dst, $src}",
1794 (VCVTSD2SI64rr GR64:$dst, VR128:$src), 0>;
1795 def : InstAlias<"vcvtsd2si{q}\t{$src, $dst|$dst, $src}",
1796 (VCVTSD2SI64rm GR64:$dst, sdmem:$src), 0>;
1799 def : InstAlias<"cvtss2si{l}\t{$src, $dst|$dst, $src}",
1800 (CVTSS2SIrr GR32:$dst, VR128:$src), 0>;
1801 def : InstAlias<"cvtss2si{l}\t{$src, $dst|$dst, $src}",
1802 (CVTSS2SIrm GR32:$dst, ssmem:$src), 0>;
1803 def : InstAlias<"cvtsd2si{l}\t{$src, $dst|$dst, $src}",
1804 (CVTSD2SIrr GR32:$dst, VR128:$src), 0>;
1805 def : InstAlias<"cvtsd2si{l}\t{$src, $dst|$dst, $src}",
1806 (CVTSD2SIrm GR32:$dst, sdmem:$src), 0>;
1807 def : InstAlias<"cvtss2si{q}\t{$src, $dst|$dst, $src}",
1808 (CVTSS2SI64rr GR64:$dst, VR128:$src), 0>;
1809 def : InstAlias<"cvtss2si{q}\t{$src, $dst|$dst, $src}",
1810 (CVTSS2SI64rm GR64:$dst, ssmem:$src), 0>;
1811 def : InstAlias<"cvtsd2si{q}\t{$src, $dst|$dst, $src}",
1812 (CVTSD2SI64rr GR64:$dst, VR128:$src), 0>;
1813 def : InstAlias<"cvtsd2si{q}\t{$src, $dst|$dst, $src}",
1814 (CVTSD2SI64rm GR64:$dst, sdmem:$src)>;
1818 // Convert scalar double to scalar single
1819 let hasSideEffects = 0, Predicates = [UseAVX] in {
1820 def VCVTSD2SSrr : VSDI<0x5A, MRMSrcReg, (outs FR32:$dst),
1821 (ins FR64:$src1, FR64:$src2),
1822 "cvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}", [],
1823 IIC_SSE_CVT_Scalar_RR>, VEX_4V, VEX_LIG,
1824 Sched<[WriteCvtF2F]>;
1826 def VCVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst),
1827 (ins FR64:$src1, f64mem:$src2),
1828 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1829 [], IIC_SSE_CVT_Scalar_RM>,
1830 XD, Requires<[HasAVX, OptForSize]>, VEX_4V, VEX_LIG,
1831 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1834 def : Pat<(f32 (fround FR64:$src)), (VCVTSD2SSrr FR64:$src, FR64:$src)>,
1837 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1838 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1839 [(set FR32:$dst, (fround FR64:$src))],
1840 IIC_SSE_CVT_Scalar_RR>, Sched<[WriteCvtF2F]>;
1841 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1842 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1843 [(set FR32:$dst, (fround (loadf64 addr:$src)))],
1844 IIC_SSE_CVT_Scalar_RM>,
1846 Requires<[UseSSE2, OptForSize]>, Sched<[WriteCvtF2FLd]>;
1848 let isCodeGenOnly = 1 in {
1849 def Int_VCVTSD2SSrr: I<0x5A, MRMSrcReg,
1850 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1851 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1853 (int_x86_sse2_cvtsd2ss VR128:$src1, VR128:$src2))],
1854 IIC_SSE_CVT_Scalar_RR>, XD, VEX_4V, Requires<[UseAVX]>,
1855 Sched<[WriteCvtF2F]>;
1856 def Int_VCVTSD2SSrm: I<0x5A, MRMSrcReg,
1857 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1858 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1859 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss
1860 VR128:$src1, sse_load_f64:$src2))],
1861 IIC_SSE_CVT_Scalar_RM>, XD, VEX_4V, Requires<[UseAVX]>,
1862 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1864 let Constraints = "$src1 = $dst" in {
1865 def Int_CVTSD2SSrr: I<0x5A, MRMSrcReg,
1866 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1867 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1869 (int_x86_sse2_cvtsd2ss VR128:$src1, VR128:$src2))],
1870 IIC_SSE_CVT_Scalar_RR>, XD, Requires<[UseSSE2]>,
1871 Sched<[WriteCvtF2F]>;
1872 def Int_CVTSD2SSrm: I<0x5A, MRMSrcReg,
1873 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2),
1874 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1875 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss
1876 VR128:$src1, sse_load_f64:$src2))],
1877 IIC_SSE_CVT_Scalar_RM>, XD, Requires<[UseSSE2]>,
1878 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1880 } // isCodeGenOnly = 1
1882 // Convert scalar single to scalar double
1883 // SSE2 instructions with XS prefix
1884 let hasSideEffects = 0, Predicates = [UseAVX] in {
1885 def VCVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst),
1886 (ins FR32:$src1, FR32:$src2),
1887 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1888 [], IIC_SSE_CVT_Scalar_RR>,
1889 XS, Requires<[HasAVX]>, VEX_4V, VEX_LIG,
1890 Sched<[WriteCvtF2F]>;
1892 def VCVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst),
1893 (ins FR32:$src1, f32mem:$src2),
1894 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1895 [], IIC_SSE_CVT_Scalar_RM>,
1896 XS, VEX_4V, VEX_LIG, Requires<[HasAVX, OptForSize]>,
1897 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1900 def : Pat<(f64 (fextend FR32:$src)),
1901 (VCVTSS2SDrr FR32:$src, FR32:$src)>, Requires<[UseAVX]>;
1902 def : Pat<(fextend (loadf32 addr:$src)),
1903 (VCVTSS2SDrm (f32 (IMPLICIT_DEF)), addr:$src)>, Requires<[UseAVX]>;
1905 def : Pat<(extloadf32 addr:$src),
1906 (VCVTSS2SDrm (f32 (IMPLICIT_DEF)), addr:$src)>,
1907 Requires<[UseAVX, OptForSize]>;
1908 def : Pat<(extloadf32 addr:$src),
1909 (VCVTSS2SDrr (f32 (IMPLICIT_DEF)), (VMOVSSrm addr:$src))>,
1910 Requires<[UseAVX, OptForSpeed]>;
1912 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1913 "cvtss2sd\t{$src, $dst|$dst, $src}",
1914 [(set FR64:$dst, (fextend FR32:$src))],
1915 IIC_SSE_CVT_Scalar_RR>, XS,
1916 Requires<[UseSSE2]>, Sched<[WriteCvtF2F]>;
1917 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1918 "cvtss2sd\t{$src, $dst|$dst, $src}",
1919 [(set FR64:$dst, (extloadf32 addr:$src))],
1920 IIC_SSE_CVT_Scalar_RM>, XS,
1921 Requires<[UseSSE2, OptForSize]>, Sched<[WriteCvtF2FLd]>;
1923 // extload f32 -> f64. This matches load+fextend because we have a hack in
1924 // the isel (PreprocessForFPConvert) that can introduce loads after dag
1926 // Since these loads aren't folded into the fextend, we have to match it
1928 def : Pat<(fextend (loadf32 addr:$src)),
1929 (CVTSS2SDrm addr:$src)>, Requires<[UseSSE2]>;
1930 def : Pat<(extloadf32 addr:$src),
1931 (CVTSS2SDrr (MOVSSrm addr:$src))>, Requires<[UseSSE2, OptForSpeed]>;
1933 let isCodeGenOnly = 1 in {
1934 def Int_VCVTSS2SDrr: I<0x5A, MRMSrcReg,
1935 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1936 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1938 (int_x86_sse2_cvtss2sd VR128:$src1, VR128:$src2))],
1939 IIC_SSE_CVT_Scalar_RR>, XS, VEX_4V, Requires<[UseAVX]>,
1940 Sched<[WriteCvtF2F]>;
1941 def Int_VCVTSS2SDrm: I<0x5A, MRMSrcMem,
1942 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
1943 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1945 (int_x86_sse2_cvtss2sd VR128:$src1, sse_load_f32:$src2))],
1946 IIC_SSE_CVT_Scalar_RM>, XS, VEX_4V, Requires<[UseAVX]>,
1947 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1948 let Constraints = "$src1 = $dst" in { // SSE2 instructions with XS prefix
1949 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1950 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1951 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1953 (int_x86_sse2_cvtss2sd VR128:$src1, VR128:$src2))],
1954 IIC_SSE_CVT_Scalar_RR>, XS, Requires<[UseSSE2]>,
1955 Sched<[WriteCvtF2F]>;
1956 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1957 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2),
1958 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1960 (int_x86_sse2_cvtss2sd VR128:$src1, sse_load_f32:$src2))],
1961 IIC_SSE_CVT_Scalar_RM>, XS, Requires<[UseSSE2]>,
1962 Sched<[WriteCvtF2FLd, ReadAfterLd]>;
1964 } // isCodeGenOnly = 1
1966 // Convert packed single/double fp to doubleword
1967 def VCVTPS2DQrr : VPDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1968 "cvtps2dq\t{$src, $dst|$dst, $src}",
1969 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))],
1970 IIC_SSE_CVT_PS_RR>, VEX, Sched<[WriteCvtF2I]>;
1971 def VCVTPS2DQrm : VPDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1972 "cvtps2dq\t{$src, $dst|$dst, $src}",
1974 (int_x86_sse2_cvtps2dq (loadv4f32 addr:$src)))],
1975 IIC_SSE_CVT_PS_RM>, VEX, Sched<[WriteCvtF2ILd]>;
1976 def VCVTPS2DQYrr : VPDI<0x5B, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
1977 "cvtps2dq\t{$src, $dst|$dst, $src}",
1979 (int_x86_avx_cvt_ps2dq_256 VR256:$src))],
1980 IIC_SSE_CVT_PS_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>;
1981 def VCVTPS2DQYrm : VPDI<0x5B, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
1982 "cvtps2dq\t{$src, $dst|$dst, $src}",
1984 (int_x86_avx_cvt_ps2dq_256 (loadv8f32 addr:$src)))],
1985 IIC_SSE_CVT_PS_RM>, VEX, VEX_L, Sched<[WriteCvtF2ILd]>;
1986 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1987 "cvtps2dq\t{$src, $dst|$dst, $src}",
1988 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))],
1989 IIC_SSE_CVT_PS_RR>, Sched<[WriteCvtF2I]>;
1990 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1991 "cvtps2dq\t{$src, $dst|$dst, $src}",
1993 (int_x86_sse2_cvtps2dq (memopv4f32 addr:$src)))],
1994 IIC_SSE_CVT_PS_RM>, Sched<[WriteCvtF2ILd]>;
1997 // Convert Packed Double FP to Packed DW Integers
1998 let Predicates = [HasAVX] in {
1999 // The assembler can recognize rr 256-bit instructions by seeing a ymm
2000 // register, but the same isn't true when using memory operands instead.
2001 // Provide other assembly rr and rm forms to address this explicitly.
2002 def VCVTPD2DQrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2003 "vcvtpd2dq\t{$src, $dst|$dst, $src}",
2004 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
2005 VEX, Sched<[WriteCvtF2I]>;
2008 def : InstAlias<"vcvtpd2dqx\t{$src, $dst|$dst, $src}",
2009 (VCVTPD2DQrr VR128:$dst, VR128:$src), 0>;
2010 def VCVTPD2DQXrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2011 "vcvtpd2dqx\t{$src, $dst|$dst, $src}",
2013 (int_x86_sse2_cvtpd2dq (loadv2f64 addr:$src)))]>, VEX,
2014 Sched<[WriteCvtF2ILd]>;
2017 def VCVTPD2DQYrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2018 "vcvtpd2dq{y}\t{$src, $dst|$dst, $src}",
2020 (int_x86_avx_cvt_pd2dq_256 VR256:$src))]>, VEX, VEX_L,
2021 Sched<[WriteCvtF2I]>;
2022 def VCVTPD2DQYrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2023 "vcvtpd2dq{y}\t{$src, $dst|$dst, $src}",
2025 (int_x86_avx_cvt_pd2dq_256 (loadv4f64 addr:$src)))]>,
2026 VEX, VEX_L, Sched<[WriteCvtF2ILd]>;
2027 def : InstAlias<"vcvtpd2dq\t{$src, $dst|$dst, $src}",
2028 (VCVTPD2DQYrr VR128:$dst, VR256:$src), 0>;
2031 def CVTPD2DQrm : SDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2032 "cvtpd2dq\t{$src, $dst|$dst, $src}",
2034 (int_x86_sse2_cvtpd2dq (memopv2f64 addr:$src)))],
2035 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtF2ILd]>;
2036 def CVTPD2DQrr : SDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2037 "cvtpd2dq\t{$src, $dst|$dst, $src}",
2038 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))],
2039 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2I]>;
2041 // Convert with truncation packed single/double fp to doubleword
2042 // SSE2 packed instructions with XS prefix
2043 def VCVTTPS2DQrr : VS2SI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2044 "cvttps2dq\t{$src, $dst|$dst, $src}",
2046 (int_x86_sse2_cvttps2dq VR128:$src))],
2047 IIC_SSE_CVT_PS_RR>, VEX, Sched<[WriteCvtF2I]>;
2048 def VCVTTPS2DQrm : VS2SI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2049 "cvttps2dq\t{$src, $dst|$dst, $src}",
2050 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
2051 (loadv4f32 addr:$src)))],
2052 IIC_SSE_CVT_PS_RM>, VEX, Sched<[WriteCvtF2ILd]>;
2053 def VCVTTPS2DQYrr : VS2SI<0x5B, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
2054 "cvttps2dq\t{$src, $dst|$dst, $src}",
2056 (int_x86_avx_cvtt_ps2dq_256 VR256:$src))],
2057 IIC_SSE_CVT_PS_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>;
2058 def VCVTTPS2DQYrm : VS2SI<0x5B, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
2059 "cvttps2dq\t{$src, $dst|$dst, $src}",
2060 [(set VR256:$dst, (int_x86_avx_cvtt_ps2dq_256
2061 (loadv8f32 addr:$src)))],
2062 IIC_SSE_CVT_PS_RM>, VEX, VEX_L,
2063 Sched<[WriteCvtF2ILd]>;
2065 def CVTTPS2DQrr : S2SI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2066 "cvttps2dq\t{$src, $dst|$dst, $src}",
2067 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))],
2068 IIC_SSE_CVT_PS_RR>, Sched<[WriteCvtF2I]>;
2069 def CVTTPS2DQrm : S2SI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2070 "cvttps2dq\t{$src, $dst|$dst, $src}",
2072 (int_x86_sse2_cvttps2dq (memopv4f32 addr:$src)))],
2073 IIC_SSE_CVT_PS_RM>, Sched<[WriteCvtF2ILd]>;
2075 let Predicates = [HasAVX] in {
2076 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
2077 (VCVTDQ2PSrr VR128:$src)>;
2078 def : Pat<(v4f32 (sint_to_fp (bc_v4i32 (loadv2i64 addr:$src)))),
2079 (VCVTDQ2PSrm addr:$src)>;
2081 def : Pat<(int_x86_sse2_cvtdq2ps VR128:$src),
2082 (VCVTDQ2PSrr VR128:$src)>;
2083 def : Pat<(int_x86_sse2_cvtdq2ps (bc_v4i32 (loadv2i64 addr:$src))),
2084 (VCVTDQ2PSrm addr:$src)>;
2086 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
2087 (VCVTTPS2DQrr VR128:$src)>;
2088 def : Pat<(v4i32 (fp_to_sint (loadv4f32 addr:$src))),
2089 (VCVTTPS2DQrm addr:$src)>;
2091 def : Pat<(v8f32 (sint_to_fp (v8i32 VR256:$src))),
2092 (VCVTDQ2PSYrr VR256:$src)>;
2093 def : Pat<(v8f32 (sint_to_fp (bc_v8i32 (loadv4i64 addr:$src)))),
2094 (VCVTDQ2PSYrm addr:$src)>;
2096 def : Pat<(v8i32 (fp_to_sint (v8f32 VR256:$src))),
2097 (VCVTTPS2DQYrr VR256:$src)>;
2098 def : Pat<(v8i32 (fp_to_sint (loadv8f32 addr:$src))),
2099 (VCVTTPS2DQYrm addr:$src)>;
2102 let Predicates = [UseSSE2] in {
2103 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
2104 (CVTDQ2PSrr VR128:$src)>;
2105 def : Pat<(v4f32 (sint_to_fp (bc_v4i32 (memopv2i64 addr:$src)))),
2106 (CVTDQ2PSrm addr:$src)>;
2108 def : Pat<(int_x86_sse2_cvtdq2ps VR128:$src),
2109 (CVTDQ2PSrr VR128:$src)>;
2110 def : Pat<(int_x86_sse2_cvtdq2ps (bc_v4i32 (memopv2i64 addr:$src))),
2111 (CVTDQ2PSrm addr:$src)>;
2113 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
2114 (CVTTPS2DQrr VR128:$src)>;
2115 def : Pat<(v4i32 (fp_to_sint (memopv4f32 addr:$src))),
2116 (CVTTPS2DQrm addr:$src)>;
2119 def VCVTTPD2DQrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2120 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2122 (int_x86_sse2_cvttpd2dq VR128:$src))],
2123 IIC_SSE_CVT_PD_RR>, VEX, Sched<[WriteCvtF2I]>;
2125 // The assembler can recognize rr 256-bit instructions by seeing a ymm
2126 // register, but the same isn't true when using memory operands instead.
2127 // Provide other assembly rr and rm forms to address this explicitly.
2130 def : InstAlias<"vcvttpd2dqx\t{$src, $dst|$dst, $src}",
2131 (VCVTTPD2DQrr VR128:$dst, VR128:$src), 0>;
2132 def VCVTTPD2DQXrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2133 "cvttpd2dqx\t{$src, $dst|$dst, $src}",
2134 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
2135 (loadv2f64 addr:$src)))],
2136 IIC_SSE_CVT_PD_RM>, VEX, Sched<[WriteCvtF2ILd]>;
2139 def VCVTTPD2DQYrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2140 "cvttpd2dq{y}\t{$src, $dst|$dst, $src}",
2142 (int_x86_avx_cvtt_pd2dq_256 VR256:$src))],
2143 IIC_SSE_CVT_PD_RR>, VEX, VEX_L, Sched<[WriteCvtF2I]>;
2144 def VCVTTPD2DQYrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2145 "cvttpd2dq{y}\t{$src, $dst|$dst, $src}",
2147 (int_x86_avx_cvtt_pd2dq_256 (loadv4f64 addr:$src)))],
2148 IIC_SSE_CVT_PD_RM>, VEX, VEX_L, Sched<[WriteCvtF2ILd]>;
2149 def : InstAlias<"vcvttpd2dq\t{$src, $dst|$dst, $src}",
2150 (VCVTTPD2DQYrr VR128:$dst, VR256:$src), 0>;
2152 let Predicates = [HasAVX] in {
2153 def : Pat<(v4i32 (fp_to_sint (v4f64 VR256:$src))),
2154 (VCVTTPD2DQYrr VR256:$src)>;
2155 def : Pat<(v4i32 (fp_to_sint (loadv4f64 addr:$src))),
2156 (VCVTTPD2DQYrm addr:$src)>;
2157 } // Predicates = [HasAVX]
2159 def CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2160 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2161 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))],
2162 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2I]>;
2163 def CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
2164 "cvttpd2dq\t{$src, $dst|$dst, $src}",
2165 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
2166 (memopv2f64 addr:$src)))],
2168 Sched<[WriteCvtF2ILd]>;
2170 // Convert packed single to packed double
2171 let Predicates = [HasAVX] in {
2172 // SSE2 instructions without OpSize prefix
2173 def VCVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2174 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2175 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))],
2176 IIC_SSE_CVT_PD_RR>, PS, VEX, Sched<[WriteCvtF2F]>;
2177 def VCVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2178 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2179 [(set VR128:$dst, (v2f64 (extloadv2f32 addr:$src)))],
2180 IIC_SSE_CVT_PD_RM>, PS, VEX, Sched<[WriteCvtF2FLd]>;
2181 def VCVTPS2PDYrr : I<0x5A, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
2182 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2184 (int_x86_avx_cvt_ps2_pd_256 VR128:$src))],
2185 IIC_SSE_CVT_PD_RR>, PS, VEX, VEX_L, Sched<[WriteCvtF2F]>;
2186 def VCVTPS2PDYrm : I<0x5A, MRMSrcMem, (outs VR256:$dst), (ins f128mem:$src),
2187 "vcvtps2pd\t{$src, $dst|$dst, $src}",
2189 (int_x86_avx_cvt_ps2_pd_256 (loadv4f32 addr:$src)))],
2190 IIC_SSE_CVT_PD_RM>, PS, VEX, VEX_L, Sched<[WriteCvtF2FLd]>;
2193 let Predicates = [UseSSE2] in {
2194 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2195 "cvtps2pd\t{$src, $dst|$dst, $src}",
2196 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))],
2197 IIC_SSE_CVT_PD_RR>, PS, Sched<[WriteCvtF2F]>;
2198 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2199 "cvtps2pd\t{$src, $dst|$dst, $src}",
2200 [(set VR128:$dst, (v2f64 (extloadv2f32 addr:$src)))],
2201 IIC_SSE_CVT_PD_RM>, PS, Sched<[WriteCvtF2FLd]>;
2204 // Convert Packed DW Integers to Packed Double FP
2205 let Predicates = [HasAVX] in {
2206 let hasSideEffects = 0, mayLoad = 1 in
2207 def VCVTDQ2PDrm : S2SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2208 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2209 []>, VEX, Sched<[WriteCvtI2FLd]>;
2210 def VCVTDQ2PDrr : S2SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2211 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2213 (int_x86_sse2_cvtdq2pd VR128:$src))]>, VEX,
2214 Sched<[WriteCvtI2F]>;
2215 def VCVTDQ2PDYrm : S2SI<0xE6, MRMSrcMem, (outs VR256:$dst), (ins i128mem:$src),
2216 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2218 (int_x86_avx_cvtdq2_pd_256
2219 (bitconvert (loadv2i64 addr:$src))))]>, VEX, VEX_L,
2220 Sched<[WriteCvtI2FLd]>;
2221 def VCVTDQ2PDYrr : S2SI<0xE6, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
2222 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
2224 (int_x86_avx_cvtdq2_pd_256 VR128:$src))]>, VEX, VEX_L,
2225 Sched<[WriteCvtI2F]>;
2228 let hasSideEffects = 0, mayLoad = 1 in
2229 def CVTDQ2PDrm : S2SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2230 "cvtdq2pd\t{$src, $dst|$dst, $src}", [],
2231 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtI2FLd]>;
2232 def CVTDQ2PDrr : S2SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2233 "cvtdq2pd\t{$src, $dst|$dst, $src}",
2234 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))],
2235 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtI2F]>;
2237 // AVX 256-bit register conversion intrinsics
2238 let Predicates = [HasAVX] in {
2239 def : Pat<(v4f64 (sint_to_fp (v4i32 VR128:$src))),
2240 (VCVTDQ2PDYrr VR128:$src)>;
2241 def : Pat<(v4f64 (sint_to_fp (bc_v4i32 (loadv2i64 addr:$src)))),
2242 (VCVTDQ2PDYrm addr:$src)>;
2243 } // Predicates = [HasAVX]
2245 // Convert packed double to packed single
2246 // The assembler can recognize rr 256-bit instructions by seeing a ymm
2247 // register, but the same isn't true when using memory operands instead.
2248 // Provide other assembly rr and rm forms to address this explicitly.
2249 def VCVTPD2PSrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2250 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2251 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))],
2252 IIC_SSE_CVT_PD_RR>, VEX, Sched<[WriteCvtF2F]>;
2255 def : InstAlias<"vcvtpd2psx\t{$src, $dst|$dst, $src}",
2256 (VCVTPD2PSrr VR128:$dst, VR128:$src), 0>;
2257 def VCVTPD2PSXrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2258 "cvtpd2psx\t{$src, $dst|$dst, $src}",
2260 (int_x86_sse2_cvtpd2ps (loadv2f64 addr:$src)))],
2261 IIC_SSE_CVT_PD_RM>, VEX, Sched<[WriteCvtF2FLd]>;
2264 def VCVTPD2PSYrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
2265 "cvtpd2ps{y}\t{$src, $dst|$dst, $src}",
2267 (int_x86_avx_cvt_pd2_ps_256 VR256:$src))],
2268 IIC_SSE_CVT_PD_RR>, VEX, VEX_L, Sched<[WriteCvtF2F]>;
2269 def VCVTPD2PSYrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
2270 "cvtpd2ps{y}\t{$src, $dst|$dst, $src}",
2272 (int_x86_avx_cvt_pd2_ps_256 (loadv4f64 addr:$src)))],
2273 IIC_SSE_CVT_PD_RM>, VEX, VEX_L, Sched<[WriteCvtF2FLd]>;
2274 def : InstAlias<"vcvtpd2ps\t{$src, $dst|$dst, $src}",
2275 (VCVTPD2PSYrr VR128:$dst, VR256:$src), 0>;
2277 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2278 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2279 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))],
2280 IIC_SSE_CVT_PD_RR>, Sched<[WriteCvtF2F]>;
2281 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2282 "cvtpd2ps\t{$src, $dst|$dst, $src}",
2284 (int_x86_sse2_cvtpd2ps (memopv2f64 addr:$src)))],
2285 IIC_SSE_CVT_PD_RM>, Sched<[WriteCvtF2FLd]>;
2288 // AVX 256-bit register conversion intrinsics
2289 // FIXME: Migrate SSE conversion intrinsics matching to use patterns as below
2290 // whenever possible to avoid declaring two versions of each one.
2291 let Predicates = [HasAVX] in {
2292 def : Pat<(int_x86_avx_cvtdq2_ps_256 VR256:$src),
2293 (VCVTDQ2PSYrr VR256:$src)>;
2294 def : Pat<(int_x86_avx_cvtdq2_ps_256 (bitconvert (loadv4i64 addr:$src))),
2295 (VCVTDQ2PSYrm addr:$src)>;
2297 // Match fround and fextend for 128/256-bit conversions
2298 def : Pat<(v4f32 (X86vfpround (v2f64 VR128:$src))),
2299 (VCVTPD2PSrr VR128:$src)>;
2300 def : Pat<(v4f32 (X86vfpround (loadv2f64 addr:$src))),
2301 (VCVTPD2PSXrm addr:$src)>;
2302 def : Pat<(v4f32 (fround (v4f64 VR256:$src))),
2303 (VCVTPD2PSYrr VR256:$src)>;
2304 def : Pat<(v4f32 (fround (loadv4f64 addr:$src))),
2305 (VCVTPD2PSYrm addr:$src)>;
2307 def : Pat<(v2f64 (X86vfpext (v4f32 VR128:$src))),
2308 (VCVTPS2PDrr VR128:$src)>;
2309 def : Pat<(v4f64 (fextend (v4f32 VR128:$src))),
2310 (VCVTPS2PDYrr VR128:$src)>;
2311 def : Pat<(v4f64 (extloadv4f32 addr:$src)),
2312 (VCVTPS2PDYrm addr:$src)>;
2315 let Predicates = [UseSSE2] in {
2316 // Match fround and fextend for 128 conversions
2317 def : Pat<(v4f32 (X86vfpround (v2f64 VR128:$src))),
2318 (CVTPD2PSrr VR128:$src)>;
2319 def : Pat<(v4f32 (X86vfpround (memopv2f64 addr:$src))),
2320 (CVTPD2PSrm addr:$src)>;
2322 def : Pat<(v2f64 (X86vfpext (v4f32 VR128:$src))),
2323 (CVTPS2PDrr VR128:$src)>;
2326 //===----------------------------------------------------------------------===//
2327 // SSE 1 & 2 - Compare Instructions
2328 //===----------------------------------------------------------------------===//
2330 // sse12_cmp_scalar - sse 1 & 2 compare scalar instructions
2331 multiclass sse12_cmp_scalar<RegisterClass RC, X86MemOperand x86memop,
2332 Operand CC, SDNode OpNode, ValueType VT,
2333 PatFrag ld_frag, string asm, string asm_alt,
2334 OpndItins itins, ImmLeaf immLeaf> {
2335 def rr : SIi8<0xC2, MRMSrcReg,
2336 (outs RC:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
2337 [(set RC:$dst, (OpNode (VT RC:$src1), RC:$src2, immLeaf:$cc))],
2338 itins.rr>, Sched<[itins.Sched]>;
2339 def rm : SIi8<0xC2, MRMSrcMem,
2340 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
2341 [(set RC:$dst, (OpNode (VT RC:$src1),
2342 (ld_frag addr:$src2), immLeaf:$cc))],
2344 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2346 // Accept explicit immediate argument form instead of comparison code.
2347 let isAsmParserOnly = 1, hasSideEffects = 0 in {
2348 def rr_alt : SIi8<0xC2, MRMSrcReg, (outs RC:$dst),
2349 (ins RC:$src1, RC:$src2, u8imm:$cc), asm_alt, [],
2350 IIC_SSE_ALU_F32S_RR>, Sched<[itins.Sched]>;
2352 def rm_alt : SIi8<0xC2, MRMSrcMem, (outs RC:$dst),
2353 (ins RC:$src1, x86memop:$src2, u8imm:$cc), asm_alt, [],
2354 IIC_SSE_ALU_F32S_RM>,
2355 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2359 defm VCMPSS : sse12_cmp_scalar<FR32, f32mem, AVXCC, X86cmps, f32, loadf32,
2360 "cmp${cc}ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2361 "cmpss\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2362 SSE_ALU_F32S, i8immZExt5>, XS, VEX_4V, VEX_LIG;
2363 defm VCMPSD : sse12_cmp_scalar<FR64, f64mem, AVXCC, X86cmps, f64, loadf64,
2364 "cmp${cc}sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2365 "cmpsd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2366 SSE_ALU_F32S, i8immZExt5>, // same latency as 32 bit compare
2367 XD, VEX_4V, VEX_LIG;
2369 let Constraints = "$src1 = $dst" in {
2370 defm CMPSS : sse12_cmp_scalar<FR32, f32mem, SSECC, X86cmps, f32, loadf32,
2371 "cmp${cc}ss\t{$src2, $dst|$dst, $src2}",
2372 "cmpss\t{$cc, $src2, $dst|$dst, $src2, $cc}", SSE_ALU_F32S,
2374 defm CMPSD : sse12_cmp_scalar<FR64, f64mem, SSECC, X86cmps, f64, loadf64,
2375 "cmp${cc}sd\t{$src2, $dst|$dst, $src2}",
2376 "cmpsd\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2377 SSE_ALU_F64S, i8immZExt3>, XD;
2380 multiclass sse12_cmp_scalar_int<X86MemOperand x86memop, Operand CC,
2381 Intrinsic Int, string asm, OpndItins itins,
2383 def rr : SIi8<0xC2, MRMSrcReg, (outs VR128:$dst),
2384 (ins VR128:$src1, VR128:$src, CC:$cc), asm,
2385 [(set VR128:$dst, (Int VR128:$src1,
2386 VR128:$src, immLeaf:$cc))],
2388 Sched<[itins.Sched]>;
2389 def rm : SIi8<0xC2, MRMSrcMem, (outs VR128:$dst),
2390 (ins VR128:$src1, x86memop:$src, CC:$cc), asm,
2391 [(set VR128:$dst, (Int VR128:$src1,
2392 (load addr:$src), immLeaf:$cc))],
2394 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2397 let isCodeGenOnly = 1 in {
2398 // Aliases to match intrinsics which expect XMM operand(s).
2399 defm Int_VCMPSS : sse12_cmp_scalar_int<f32mem, AVXCC, int_x86_sse_cmp_ss,
2400 "cmp${cc}ss\t{$src, $src1, $dst|$dst, $src1, $src}",
2401 SSE_ALU_F32S, i8immZExt5>,
2403 defm Int_VCMPSD : sse12_cmp_scalar_int<f64mem, AVXCC, int_x86_sse2_cmp_sd,
2404 "cmp${cc}sd\t{$src, $src1, $dst|$dst, $src1, $src}",
2405 SSE_ALU_F32S, i8immZExt5>, // same latency as f32
2407 let Constraints = "$src1 = $dst" in {
2408 defm Int_CMPSS : sse12_cmp_scalar_int<f32mem, SSECC, int_x86_sse_cmp_ss,
2409 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
2410 SSE_ALU_F32S, i8immZExt3>, XS;
2411 defm Int_CMPSD : sse12_cmp_scalar_int<f64mem, SSECC, int_x86_sse2_cmp_sd,
2412 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
2413 SSE_ALU_F64S, i8immZExt3>,
2419 // sse12_ord_cmp - Unordered/Ordered scalar fp compare and set EFLAGS
2420 multiclass sse12_ord_cmp<bits<8> opc, RegisterClass RC, SDNode OpNode,
2421 ValueType vt, X86MemOperand x86memop,
2422 PatFrag ld_frag, string OpcodeStr> {
2423 def rr: SI<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
2424 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2425 [(set EFLAGS, (OpNode (vt RC:$src1), RC:$src2))],
2428 def rm: SI<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
2429 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
2430 [(set EFLAGS, (OpNode (vt RC:$src1),
2431 (ld_frag addr:$src2)))],
2433 Sched<[WriteFAddLd, ReadAfterLd]>;
2436 let Defs = [EFLAGS] in {
2437 defm VUCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
2438 "ucomiss">, PS, VEX, VEX_LIG;
2439 defm VUCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
2440 "ucomisd">, PD, VEX, VEX_LIG;
2441 let Pattern = []<dag> in {
2442 defm VCOMISS : sse12_ord_cmp<0x2F, VR128, undef, v4f32, f128mem, load,
2443 "comiss">, PS, VEX, VEX_LIG;
2444 defm VCOMISD : sse12_ord_cmp<0x2F, VR128, undef, v2f64, f128mem, load,
2445 "comisd">, PD, VEX, VEX_LIG;
2448 let isCodeGenOnly = 1 in {
2449 defm Int_VUCOMISS : sse12_ord_cmp<0x2E, VR128, X86ucomi, v4f32, f128mem,
2450 load, "ucomiss">, PS, VEX;
2451 defm Int_VUCOMISD : sse12_ord_cmp<0x2E, VR128, X86ucomi, v2f64, f128mem,
2452 load, "ucomisd">, PD, VEX;
2454 defm Int_VCOMISS : sse12_ord_cmp<0x2F, VR128, X86comi, v4f32, f128mem,
2455 load, "comiss">, PS, VEX;
2456 defm Int_VCOMISD : sse12_ord_cmp<0x2F, VR128, X86comi, v2f64, f128mem,
2457 load, "comisd">, PD, VEX;
2459 defm UCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
2461 defm UCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
2464 let Pattern = []<dag> in {
2465 defm COMISS : sse12_ord_cmp<0x2F, VR128, undef, v4f32, f128mem, load,
2467 defm COMISD : sse12_ord_cmp<0x2F, VR128, undef, v2f64, f128mem, load,
2471 let isCodeGenOnly = 1 in {
2472 defm Int_UCOMISS : sse12_ord_cmp<0x2E, VR128, X86ucomi, v4f32, f128mem,
2473 load, "ucomiss">, PS;
2474 defm Int_UCOMISD : sse12_ord_cmp<0x2E, VR128, X86ucomi, v2f64, f128mem,
2475 load, "ucomisd">, PD;
2477 defm Int_COMISS : sse12_ord_cmp<0x2F, VR128, X86comi, v4f32, f128mem, load,
2479 defm Int_COMISD : sse12_ord_cmp<0x2F, VR128, X86comi, v2f64, f128mem, load,
2482 } // Defs = [EFLAGS]
2484 // sse12_cmp_packed - sse 1 & 2 compare packed instructions
2485 multiclass sse12_cmp_packed<RegisterClass RC, X86MemOperand x86memop,
2486 Operand CC, Intrinsic Int, string asm,
2487 string asm_alt, Domain d, ImmLeaf immLeaf,
2488 PatFrag ld_frag, OpndItins itins = SSE_ALU_F32P> {
2489 let isCommutable = 1 in
2490 def rri : PIi8<0xC2, MRMSrcReg,
2491 (outs RC:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
2492 [(set RC:$dst, (Int RC:$src1, RC:$src2, immLeaf:$cc))],
2495 def rmi : PIi8<0xC2, MRMSrcMem,
2496 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
2497 [(set RC:$dst, (Int RC:$src1, (ld_frag addr:$src2), immLeaf:$cc))],
2499 Sched<[WriteFAddLd, ReadAfterLd]>;
2501 // Accept explicit immediate argument form instead of comparison code.
2502 let isAsmParserOnly = 1, hasSideEffects = 0 in {
2503 def rri_alt : PIi8<0xC2, MRMSrcReg,
2504 (outs RC:$dst), (ins RC:$src1, RC:$src2, u8imm:$cc),
2505 asm_alt, [], itins.rr, d>, Sched<[WriteFAdd]>;
2507 def rmi_alt : PIi8<0xC2, MRMSrcMem,
2508 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, u8imm:$cc),
2509 asm_alt, [], itins.rm, d>,
2510 Sched<[WriteFAddLd, ReadAfterLd]>;
2514 defm VCMPPS : sse12_cmp_packed<VR128, f128mem, AVXCC, int_x86_sse_cmp_ps,
2515 "cmp${cc}ps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2516 "cmpps\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2517 SSEPackedSingle, i8immZExt5, loadv4f32>, PS, VEX_4V;
2518 defm VCMPPD : sse12_cmp_packed<VR128, f128mem, AVXCC, int_x86_sse2_cmp_pd,
2519 "cmp${cc}pd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2520 "cmppd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2521 SSEPackedDouble, i8immZExt5, loadv2f64>, PD, VEX_4V;
2522 defm VCMPPSY : sse12_cmp_packed<VR256, f256mem, AVXCC, int_x86_avx_cmp_ps_256,
2523 "cmp${cc}ps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2524 "cmpps\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2525 SSEPackedSingle, i8immZExt5, loadv8f32>, PS, VEX_4V, VEX_L;
2526 defm VCMPPDY : sse12_cmp_packed<VR256, f256mem, AVXCC, int_x86_avx_cmp_pd_256,
2527 "cmp${cc}pd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2528 "cmppd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}",
2529 SSEPackedDouble, i8immZExt5, loadv4f64>, PD, VEX_4V, VEX_L;
2530 let Constraints = "$src1 = $dst" in {
2531 defm CMPPS : sse12_cmp_packed<VR128, f128mem, SSECC, int_x86_sse_cmp_ps,
2532 "cmp${cc}ps\t{$src2, $dst|$dst, $src2}",
2533 "cmpps\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2534 SSEPackedSingle, i8immZExt5, memopv4f32, SSE_ALU_F32P>, PS;
2535 defm CMPPD : sse12_cmp_packed<VR128, f128mem, SSECC, int_x86_sse2_cmp_pd,
2536 "cmp${cc}pd\t{$src2, $dst|$dst, $src2}",
2537 "cmppd\t{$cc, $src2, $dst|$dst, $src2, $cc}",
2538 SSEPackedDouble, i8immZExt5, memopv2f64, SSE_ALU_F64P>, PD;
2541 let Predicates = [HasAVX] in {
2542 def : Pat<(v4i32 (X86cmpp (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
2543 (VCMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
2544 def : Pat<(v4i32 (X86cmpp (v4f32 VR128:$src1), (loadv4f32 addr:$src2), imm:$cc)),
2545 (VCMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
2546 def : Pat<(v2i64 (X86cmpp (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
2547 (VCMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
2548 def : Pat<(v2i64 (X86cmpp (v2f64 VR128:$src1), (loadv2f64 addr:$src2), imm:$cc)),
2549 (VCMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
2551 def : Pat<(v8i32 (X86cmpp (v8f32 VR256:$src1), VR256:$src2, imm:$cc)),
2552 (VCMPPSYrri (v8f32 VR256:$src1), (v8f32 VR256:$src2), imm:$cc)>;
2553 def : Pat<(v8i32 (X86cmpp (v8f32 VR256:$src1), (loadv8f32 addr:$src2), imm:$cc)),
2554 (VCMPPSYrmi (v8f32 VR256:$src1), addr:$src2, imm:$cc)>;
2555 def : Pat<(v4i64 (X86cmpp (v4f64 VR256:$src1), VR256:$src2, imm:$cc)),
2556 (VCMPPDYrri VR256:$src1, VR256:$src2, imm:$cc)>;
2557 def : Pat<(v4i64 (X86cmpp (v4f64 VR256:$src1), (loadv4f64 addr:$src2), imm:$cc)),
2558 (VCMPPDYrmi VR256:$src1, addr:$src2, imm:$cc)>;
2561 let Predicates = [UseSSE1] in {
2562 def : Pat<(v4i32 (X86cmpp (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
2563 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
2564 def : Pat<(v4i32 (X86cmpp (v4f32 VR128:$src1), (memopv4f32 addr:$src2), imm:$cc)),
2565 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
2568 let Predicates = [UseSSE2] in {
2569 def : Pat<(v2i64 (X86cmpp (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
2570 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
2571 def : Pat<(v2i64 (X86cmpp (v2f64 VR128:$src1), (memopv2f64 addr:$src2), imm:$cc)),
2572 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
2575 //===----------------------------------------------------------------------===//
2576 // SSE 1 & 2 - Shuffle Instructions
2577 //===----------------------------------------------------------------------===//
2579 /// sse12_shuffle - sse 1 & 2 fp shuffle instructions
2580 multiclass sse12_shuffle<RegisterClass RC, X86MemOperand x86memop,
2581 ValueType vt, string asm, PatFrag mem_frag,
2583 def rmi : PIi8<0xC6, MRMSrcMem, (outs RC:$dst),
2584 (ins RC:$src1, x86memop:$src2, u8imm:$src3), asm,
2585 [(set RC:$dst, (vt (X86Shufp RC:$src1, (mem_frag addr:$src2),
2586 (i8 imm:$src3))))], IIC_SSE_SHUFP, d>,
2587 Sched<[WriteFShuffleLd, ReadAfterLd]>;
2588 def rri : PIi8<0xC6, MRMSrcReg, (outs RC:$dst),
2589 (ins RC:$src1, RC:$src2, u8imm:$src3), asm,
2590 [(set RC:$dst, (vt (X86Shufp RC:$src1, RC:$src2,
2591 (i8 imm:$src3))))], IIC_SSE_SHUFP, d>,
2592 Sched<[WriteFShuffle]>;
2595 defm VSHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
2596 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2597 loadv4f32, SSEPackedSingle>, PS, VEX_4V;
2598 defm VSHUFPSY : sse12_shuffle<VR256, f256mem, v8f32,
2599 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2600 loadv8f32, SSEPackedSingle>, PS, VEX_4V, VEX_L;
2601 defm VSHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
2602 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2603 loadv2f64, SSEPackedDouble>, PD, VEX_4V;
2604 defm VSHUFPDY : sse12_shuffle<VR256, f256mem, v4f64,
2605 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2606 loadv4f64, SSEPackedDouble>, PD, VEX_4V, VEX_L;
2608 let Constraints = "$src1 = $dst" in {
2609 defm SHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
2610 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2611 memopv4f32, SSEPackedSingle>, PS;
2612 defm SHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
2613 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2614 memopv2f64, SSEPackedDouble>, PD;
2617 let Predicates = [HasAVX] in {
2618 def : Pat<(v4i32 (X86Shufp VR128:$src1,
2619 (bc_v4i32 (loadv2i64 addr:$src2)), (i8 imm:$imm))),
2620 (VSHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>;
2621 def : Pat<(v4i32 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2622 (VSHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>;
2624 def : Pat<(v2i64 (X86Shufp VR128:$src1,
2625 (loadv2i64 addr:$src2), (i8 imm:$imm))),
2626 (VSHUFPDrmi VR128:$src1, addr:$src2, imm:$imm)>;
2627 def : Pat<(v2i64 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2628 (VSHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>;
2631 def : Pat<(v8i32 (X86Shufp VR256:$src1, VR256:$src2, (i8 imm:$imm))),
2632 (VSHUFPSYrri VR256:$src1, VR256:$src2, imm:$imm)>;
2633 def : Pat<(v8i32 (X86Shufp VR256:$src1,
2634 (bc_v8i32 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
2635 (VSHUFPSYrmi VR256:$src1, addr:$src2, imm:$imm)>;
2637 def : Pat<(v4i64 (X86Shufp VR256:$src1, VR256:$src2, (i8 imm:$imm))),
2638 (VSHUFPDYrri VR256:$src1, VR256:$src2, imm:$imm)>;
2639 def : Pat<(v4i64 (X86Shufp VR256:$src1,
2640 (loadv4i64 addr:$src2), (i8 imm:$imm))),
2641 (VSHUFPDYrmi VR256:$src1, addr:$src2, imm:$imm)>;
2644 let Predicates = [UseSSE1] in {
2645 def : Pat<(v4i32 (X86Shufp VR128:$src1,
2646 (bc_v4i32 (memopv2i64 addr:$src2)), (i8 imm:$imm))),
2647 (SHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>;
2648 def : Pat<(v4i32 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2649 (SHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>;
2652 let Predicates = [UseSSE2] in {
2653 // Generic SHUFPD patterns
2654 def : Pat<(v2i64 (X86Shufp VR128:$src1,
2655 (memopv2i64 addr:$src2), (i8 imm:$imm))),
2656 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$imm)>;
2657 def : Pat<(v2i64 (X86Shufp VR128:$src1, VR128:$src2, (i8 imm:$imm))),
2658 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>;
2661 //===----------------------------------------------------------------------===//
2662 // SSE 1 & 2 - Unpack FP Instructions
2663 //===----------------------------------------------------------------------===//
2665 /// sse12_unpack_interleave - sse 1 & 2 fp unpack and interleave
2666 multiclass sse12_unpack_interleave<bits<8> opc, SDNode OpNode, ValueType vt,
2667 PatFrag mem_frag, RegisterClass RC,
2668 X86MemOperand x86memop, string asm,
2670 def rr : PI<opc, MRMSrcReg,
2671 (outs RC:$dst), (ins RC:$src1, RC:$src2),
2673 (vt (OpNode RC:$src1, RC:$src2)))],
2674 IIC_SSE_UNPCK, d>, Sched<[WriteFShuffle]>;
2675 def rm : PI<opc, MRMSrcMem,
2676 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
2678 (vt (OpNode RC:$src1,
2679 (mem_frag addr:$src2))))],
2681 Sched<[WriteFShuffleLd, ReadAfterLd]>;
2684 defm VUNPCKHPS: sse12_unpack_interleave<0x15, X86Unpckh, v4f32, loadv4f32,
2685 VR128, f128mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2686 SSEPackedSingle>, PS, VEX_4V;
2687 defm VUNPCKHPD: sse12_unpack_interleave<0x15, X86Unpckh, v2f64, loadv2f64,
2688 VR128, f128mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2689 SSEPackedDouble>, PD, VEX_4V;
2690 defm VUNPCKLPS: sse12_unpack_interleave<0x14, X86Unpckl, v4f32, loadv4f32,
2691 VR128, f128mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2692 SSEPackedSingle>, PS, VEX_4V;
2693 defm VUNPCKLPD: sse12_unpack_interleave<0x14, X86Unpckl, v2f64, loadv2f64,
2694 VR128, f128mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2695 SSEPackedDouble>, PD, VEX_4V;
2697 defm VUNPCKHPSY: sse12_unpack_interleave<0x15, X86Unpckh, v8f32, loadv8f32,
2698 VR256, f256mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2699 SSEPackedSingle>, PS, VEX_4V, VEX_L;
2700 defm VUNPCKHPDY: sse12_unpack_interleave<0x15, X86Unpckh, v4f64, loadv4f64,
2701 VR256, f256mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2702 SSEPackedDouble>, PD, VEX_4V, VEX_L;
2703 defm VUNPCKLPSY: sse12_unpack_interleave<0x14, X86Unpckl, v8f32, loadv8f32,
2704 VR256, f256mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2705 SSEPackedSingle>, PS, VEX_4V, VEX_L;
2706 defm VUNPCKLPDY: sse12_unpack_interleave<0x14, X86Unpckl, v4f64, loadv4f64,
2707 VR256, f256mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2708 SSEPackedDouble>, PD, VEX_4V, VEX_L;
2710 let Constraints = "$src1 = $dst" in {
2711 defm UNPCKHPS: sse12_unpack_interleave<0x15, X86Unpckh, v4f32, memopv4f32,
2712 VR128, f128mem, "unpckhps\t{$src2, $dst|$dst, $src2}",
2713 SSEPackedSingle>, PS;
2714 defm UNPCKHPD: sse12_unpack_interleave<0x15, X86Unpckh, v2f64, memopv2f64,
2715 VR128, f128mem, "unpckhpd\t{$src2, $dst|$dst, $src2}",
2716 SSEPackedDouble>, PD;
2717 defm UNPCKLPS: sse12_unpack_interleave<0x14, X86Unpckl, v4f32, memopv4f32,
2718 VR128, f128mem, "unpcklps\t{$src2, $dst|$dst, $src2}",
2719 SSEPackedSingle>, PS;
2720 defm UNPCKLPD: sse12_unpack_interleave<0x14, X86Unpckl, v2f64, memopv2f64,
2721 VR128, f128mem, "unpcklpd\t{$src2, $dst|$dst, $src2}",
2722 SSEPackedDouble>, PD;
2723 } // Constraints = "$src1 = $dst"
2725 let Predicates = [HasAVX1Only] in {
2726 def : Pat<(v8i32 (X86Unpckl VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
2727 (VUNPCKLPSYrm VR256:$src1, addr:$src2)>;
2728 def : Pat<(v8i32 (X86Unpckl VR256:$src1, VR256:$src2)),
2729 (VUNPCKLPSYrr VR256:$src1, VR256:$src2)>;
2730 def : Pat<(v8i32 (X86Unpckh VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
2731 (VUNPCKHPSYrm VR256:$src1, addr:$src2)>;
2732 def : Pat<(v8i32 (X86Unpckh VR256:$src1, VR256:$src2)),
2733 (VUNPCKHPSYrr VR256:$src1, VR256:$src2)>;
2735 def : Pat<(v4i64 (X86Unpckl VR256:$src1, (loadv4i64 addr:$src2))),
2736 (VUNPCKLPDYrm VR256:$src1, addr:$src2)>;
2737 def : Pat<(v4i64 (X86Unpckl VR256:$src1, VR256:$src2)),
2738 (VUNPCKLPDYrr VR256:$src1, VR256:$src2)>;
2739 def : Pat<(v4i64 (X86Unpckh VR256:$src1, (loadv4i64 addr:$src2))),
2740 (VUNPCKHPDYrm VR256:$src1, addr:$src2)>;
2741 def : Pat<(v4i64 (X86Unpckh VR256:$src1, VR256:$src2)),
2742 (VUNPCKHPDYrr VR256:$src1, VR256:$src2)>;
2745 //===----------------------------------------------------------------------===//
2746 // SSE 1 & 2 - Extract Floating-Point Sign mask
2747 //===----------------------------------------------------------------------===//
2749 /// sse12_extr_sign_mask - sse 1 & 2 unpack and interleave
2750 multiclass sse12_extr_sign_mask<RegisterClass RC, Intrinsic Int, string asm,
2752 def rr : PI<0x50, MRMSrcReg, (outs GR32orGR64:$dst), (ins RC:$src),
2753 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
2754 [(set GR32orGR64:$dst, (Int RC:$src))], IIC_SSE_MOVMSK, d>,
2755 Sched<[WriteVecLogic]>;
2758 let Predicates = [HasAVX] in {
2759 defm VMOVMSKPS : sse12_extr_sign_mask<VR128, int_x86_sse_movmsk_ps,
2760 "movmskps", SSEPackedSingle>, PS, VEX;
2761 defm VMOVMSKPD : sse12_extr_sign_mask<VR128, int_x86_sse2_movmsk_pd,
2762 "movmskpd", SSEPackedDouble>, PD, VEX;
2763 defm VMOVMSKPSY : sse12_extr_sign_mask<VR256, int_x86_avx_movmsk_ps_256,
2764 "movmskps", SSEPackedSingle>, PS,
2766 defm VMOVMSKPDY : sse12_extr_sign_mask<VR256, int_x86_avx_movmsk_pd_256,
2767 "movmskpd", SSEPackedDouble>, PD,
2770 def : Pat<(i32 (X86fgetsign FR32:$src)),
2771 (VMOVMSKPSrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
2772 def : Pat<(i64 (X86fgetsign FR32:$src)),
2773 (SUBREG_TO_REG (i64 0),
2774 (VMOVMSKPSrr (COPY_TO_REGCLASS FR32:$src, VR128)), sub_32bit)>;
2775 def : Pat<(i32 (X86fgetsign FR64:$src)),
2776 (VMOVMSKPDrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
2777 def : Pat<(i64 (X86fgetsign FR64:$src)),
2778 (SUBREG_TO_REG (i64 0),
2779 (VMOVMSKPDrr (COPY_TO_REGCLASS FR64:$src, VR128)), sub_32bit)>;
2782 defm MOVMSKPS : sse12_extr_sign_mask<VR128, int_x86_sse_movmsk_ps, "movmskps",
2783 SSEPackedSingle>, PS;
2784 defm MOVMSKPD : sse12_extr_sign_mask<VR128, int_x86_sse2_movmsk_pd, "movmskpd",
2785 SSEPackedDouble>, PD;
2787 def : Pat<(i32 (X86fgetsign FR32:$src)),
2788 (MOVMSKPSrr (COPY_TO_REGCLASS FR32:$src, VR128))>,
2789 Requires<[UseSSE1]>;
2790 def : Pat<(i64 (X86fgetsign FR32:$src)),
2791 (SUBREG_TO_REG (i64 0),
2792 (MOVMSKPSrr (COPY_TO_REGCLASS FR32:$src, VR128)), sub_32bit)>,
2793 Requires<[UseSSE1]>;
2794 def : Pat<(i32 (X86fgetsign FR64:$src)),
2795 (MOVMSKPDrr (COPY_TO_REGCLASS FR64:$src, VR128))>,
2796 Requires<[UseSSE2]>;
2797 def : Pat<(i64 (X86fgetsign FR64:$src)),
2798 (SUBREG_TO_REG (i64 0),
2799 (MOVMSKPDrr (COPY_TO_REGCLASS FR64:$src, VR128)), sub_32bit)>,
2800 Requires<[UseSSE2]>;
2802 //===---------------------------------------------------------------------===//
2803 // SSE2 - Packed Integer Logical Instructions
2804 //===---------------------------------------------------------------------===//
2806 let ExeDomain = SSEPackedInt in { // SSE integer instructions
2808 /// PDI_binop_rm - Simple SSE2 binary operator.
2809 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
2810 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
2811 X86MemOperand x86memop, OpndItins itins,
2812 bit IsCommutable, bit Is2Addr> {
2813 let isCommutable = IsCommutable in
2814 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
2815 (ins RC:$src1, RC:$src2),
2817 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2818 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2819 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))], itins.rr>,
2820 Sched<[itins.Sched]>;
2821 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
2822 (ins RC:$src1, x86memop:$src2),
2824 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2825 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2826 [(set RC:$dst, (OpVT (OpNode RC:$src1,
2827 (bitconvert (memop_frag addr:$src2)))))],
2829 Sched<[itins.Sched.Folded, ReadAfterLd]>;
2831 } // ExeDomain = SSEPackedInt
2833 multiclass PDI_binop_all<bits<8> opc, string OpcodeStr, SDNode Opcode,
2834 ValueType OpVT128, ValueType OpVT256,
2835 OpndItins itins, bit IsCommutable = 0> {
2836 let Predicates = [HasAVX, NoVLX] in
2837 defm V#NAME : PDI_binop_rm<opc, !strconcat("v", OpcodeStr), Opcode, OpVT128,
2838 VR128, loadv2i64, i128mem, itins, IsCommutable, 0>, VEX_4V;
2840 let Constraints = "$src1 = $dst" in
2841 defm NAME : PDI_binop_rm<opc, OpcodeStr, Opcode, OpVT128, VR128,
2842 memopv2i64, i128mem, itins, IsCommutable, 1>;
2844 let Predicates = [HasAVX2, NoVLX] in
2845 defm V#NAME#Y : PDI_binop_rm<opc, !strconcat("v", OpcodeStr), Opcode,
2846 OpVT256, VR256, loadv4i64, i256mem, itins,
2847 IsCommutable, 0>, VEX_4V, VEX_L;
2850 // These are ordered here for pattern ordering requirements with the fp versions
2852 defm PAND : PDI_binop_all<0xDB, "pand", and, v2i64, v4i64,
2853 SSE_VEC_BIT_ITINS_P, 1>;
2854 defm POR : PDI_binop_all<0xEB, "por", or, v2i64, v4i64,
2855 SSE_VEC_BIT_ITINS_P, 1>;
2856 defm PXOR : PDI_binop_all<0xEF, "pxor", xor, v2i64, v4i64,
2857 SSE_VEC_BIT_ITINS_P, 1>;
2858 defm PANDN : PDI_binop_all<0xDF, "pandn", X86andnp, v2i64, v4i64,
2859 SSE_VEC_BIT_ITINS_P, 0>;
2861 //===----------------------------------------------------------------------===//
2862 // SSE 1 & 2 - Logical Instructions
2863 //===----------------------------------------------------------------------===//
2865 // Multiclass for scalars using the X86 logical operation aliases for FP.
2866 multiclass sse12_fp_packed_scalar_logical_alias<
2867 bits<8> opc, string OpcodeStr, SDNode OpNode, OpndItins itins> {
2868 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode,
2869 FR32, f32, f128mem, loadf32_128, SSEPackedSingle, itins, 0>,
2872 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode,
2873 FR64, f64, f128mem, loadf64_128, SSEPackedDouble, itins, 0>,
2876 let Constraints = "$src1 = $dst" in {
2877 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode, FR32,
2878 f32, f128mem, memopfsf32_128, SSEPackedSingle, itins>, PS;
2880 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode, FR64,
2881 f64, f128mem, memopfsf64_128, SSEPackedDouble, itins>, PD;
2885 let isCodeGenOnly = 1 in {
2886 defm FsAND : sse12_fp_packed_scalar_logical_alias<0x54, "and", X86fand,
2888 defm FsOR : sse12_fp_packed_scalar_logical_alias<0x56, "or", X86for,
2890 defm FsXOR : sse12_fp_packed_scalar_logical_alias<0x57, "xor", X86fxor,
2893 let isCommutable = 0 in
2894 defm FsANDN : sse12_fp_packed_scalar_logical_alias<0x55, "andn", X86fandn,
2898 // Multiclass for vectors using the X86 logical operation aliases for FP.
2899 multiclass sse12_fp_packed_vector_logical_alias<
2900 bits<8> opc, string OpcodeStr, SDNode OpNode, OpndItins itins> {
2901 let Predicates = [HasAVX, NoVLX] in {
2902 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode,
2903 VR128, v4f32, f128mem, loadv4f32, SSEPackedSingle, itins, 0>,
2906 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode,
2907 VR128, v2f64, f128mem, loadv2f64, SSEPackedDouble, itins, 0>,
2911 let Constraints = "$src1 = $dst" in {
2912 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode, VR128,
2913 v4f32, f128mem, memopv4f32, SSEPackedSingle, itins>,
2916 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode, VR128,
2917 v2f64, f128mem, memopv2f64, SSEPackedDouble, itins>,
2922 let isCodeGenOnly = 1 in {
2923 defm FvAND : sse12_fp_packed_vector_logical_alias<0x54, "and", X86fand,
2925 defm FvOR : sse12_fp_packed_vector_logical_alias<0x56, "or", X86for,
2927 defm FvXOR : sse12_fp_packed_vector_logical_alias<0x57, "xor", X86fxor,
2930 let isCommutable = 0 in
2931 defm FvANDN : sse12_fp_packed_vector_logical_alias<0x55, "andn", X86fandn,
2935 /// sse12_fp_packed_logical - SSE 1 & 2 packed FP logical ops
2937 multiclass sse12_fp_packed_logical<bits<8> opc, string OpcodeStr,
2939 let Predicates = [HasAVX, NoVLX] in {
2940 defm V#NAME#PSY : sse12_fp_packed_logical_rm<opc, VR256, SSEPackedSingle,
2941 !strconcat(OpcodeStr, "ps"), f256mem,
2942 [(set VR256:$dst, (v4i64 (OpNode VR256:$src1, VR256:$src2)))],
2943 [(set VR256:$dst, (OpNode (bc_v4i64 (v8f32 VR256:$src1)),
2944 (loadv4i64 addr:$src2)))], 0>, PS, VEX_4V, VEX_L;
2946 defm V#NAME#PDY : sse12_fp_packed_logical_rm<opc, VR256, SSEPackedDouble,
2947 !strconcat(OpcodeStr, "pd"), f256mem,
2948 [(set VR256:$dst, (OpNode (bc_v4i64 (v4f64 VR256:$src1)),
2949 (bc_v4i64 (v4f64 VR256:$src2))))],
2950 [(set VR256:$dst, (OpNode (bc_v4i64 (v4f64 VR256:$src1)),
2951 (loadv4i64 addr:$src2)))], 0>,
2954 // In AVX no need to add a pattern for 128-bit logical rr ps, because they
2955 // are all promoted to v2i64, and the patterns are covered by the int
2956 // version. This is needed in SSE only, because v2i64 isn't supported on
2957 // SSE1, but only on SSE2.
2958 defm V#NAME#PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
2959 !strconcat(OpcodeStr, "ps"), f128mem, [],
2960 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2961 (loadv2i64 addr:$src2)))], 0>, PS, VEX_4V;
2963 defm V#NAME#PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
2964 !strconcat(OpcodeStr, "pd"), f128mem,
2965 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2966 (bc_v2i64 (v2f64 VR128:$src2))))],
2967 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2968 (loadv2i64 addr:$src2)))], 0>,
2972 let Constraints = "$src1 = $dst" in {
2973 defm PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
2974 !strconcat(OpcodeStr, "ps"), f128mem,
2975 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))],
2976 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
2977 (memopv2i64 addr:$src2)))]>, PS;
2979 defm PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
2980 !strconcat(OpcodeStr, "pd"), f128mem,
2981 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2982 (bc_v2i64 (v2f64 VR128:$src2))))],
2983 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
2984 (memopv2i64 addr:$src2)))]>, PD;
2988 defm AND : sse12_fp_packed_logical<0x54, "and", and>;
2989 defm OR : sse12_fp_packed_logical<0x56, "or", or>;
2990 defm XOR : sse12_fp_packed_logical<0x57, "xor", xor>;
2991 let isCommutable = 0 in
2992 defm ANDN : sse12_fp_packed_logical<0x55, "andn", X86andnp>;
2994 // AVX1 requires type coercions in order to fold loads directly into logical
2996 let Predicates = [HasAVX1Only] in {
2997 def : Pat<(bc_v8f32 (and VR256:$src1, (loadv4i64 addr:$src2))),
2998 (VANDPSYrm VR256:$src1, addr:$src2)>;
2999 def : Pat<(bc_v8f32 (or VR256:$src1, (loadv4i64 addr:$src2))),
3000 (VORPSYrm VR256:$src1, addr:$src2)>;
3001 def : Pat<(bc_v8f32 (xor VR256:$src1, (loadv4i64 addr:$src2))),
3002 (VXORPSYrm VR256:$src1, addr:$src2)>;
3003 def : Pat<(bc_v8f32 (X86andnp VR256:$src1, (loadv4i64 addr:$src2))),
3004 (VANDNPSYrm VR256:$src1, addr:$src2)>;
3007 //===----------------------------------------------------------------------===//
3008 // SSE 1 & 2 - Arithmetic Instructions
3009 //===----------------------------------------------------------------------===//
3011 /// basic_sse12_fp_binop_xxx - SSE 1 & 2 binops come in both scalar and
3014 /// In addition, we also have a special variant of the scalar form here to
3015 /// represent the associated intrinsic operation. This form is unlike the
3016 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
3017 /// and leaves the top elements unmodified (therefore these cannot be commuted).
3019 /// These three forms can each be reg+reg or reg+mem.
3022 /// FIXME: once all 256-bit intrinsics are matched, cleanup and refactor those
3024 multiclass basic_sse12_fp_binop_p<bits<8> opc, string OpcodeStr,
3025 SDNode OpNode, SizeItins itins> {
3026 let Predicates = [HasAVX, NoVLX] in {
3027 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode,
3028 VR128, v4f32, f128mem, loadv4f32,
3029 SSEPackedSingle, itins.s, 0>, PS, VEX_4V;
3030 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode,
3031 VR128, v2f64, f128mem, loadv2f64,
3032 SSEPackedDouble, itins.d, 0>, PD, VEX_4V;
3034 defm V#NAME#PSY : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"),
3035 OpNode, VR256, v8f32, f256mem, loadv8f32,
3036 SSEPackedSingle, itins.s, 0>, PS, VEX_4V, VEX_L;
3037 defm V#NAME#PDY : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"),
3038 OpNode, VR256, v4f64, f256mem, loadv4f64,
3039 SSEPackedDouble, itins.d, 0>, PD, VEX_4V, VEX_L;
3042 let Constraints = "$src1 = $dst" in {
3043 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode, VR128,
3044 v4f32, f128mem, memopv4f32, SSEPackedSingle,
3046 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode, VR128,
3047 v2f64, f128mem, memopv2f64, SSEPackedDouble,
3052 multiclass basic_sse12_fp_binop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
3054 defm V#NAME#SS : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "ss"),
3055 OpNode, FR32, f32mem, SSEPackedSingle, itins.s, 0>,
3056 XS, VEX_4V, VEX_LIG;
3057 defm V#NAME#SD : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "sd"),
3058 OpNode, FR64, f64mem, SSEPackedDouble, itins.d, 0>,
3059 XD, VEX_4V, VEX_LIG;
3061 let Constraints = "$src1 = $dst" in {
3062 defm SS : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "ss"),
3063 OpNode, FR32, f32mem, SSEPackedSingle,
3065 defm SD : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "sd"),
3066 OpNode, FR64, f64mem, SSEPackedDouble,
3071 multiclass basic_sse12_fp_binop_s_int<bits<8> opc, string OpcodeStr,
3073 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
3074 !strconcat(OpcodeStr, "ss"), "", "_ss", ssmem, sse_load_f32,
3075 SSEPackedSingle, itins.s, 0>, XS, VEX_4V, VEX_LIG;
3076 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
3077 !strconcat(OpcodeStr, "sd"), "2", "_sd", sdmem, sse_load_f64,
3078 SSEPackedDouble, itins.d, 0>, XD, VEX_4V, VEX_LIG;
3080 let Constraints = "$src1 = $dst" in {
3081 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
3082 !strconcat(OpcodeStr, "ss"), "", "_ss", ssmem, sse_load_f32,
3083 SSEPackedSingle, itins.s>, XS;
3084 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
3085 !strconcat(OpcodeStr, "sd"), "2", "_sd", sdmem, sse_load_f64,
3086 SSEPackedDouble, itins.d>, XD;
3090 // Binary Arithmetic instructions
3091 defm ADD : basic_sse12_fp_binop_p<0x58, "add", fadd, SSE_ALU_ITINS_P>,
3092 basic_sse12_fp_binop_s<0x58, "add", fadd, SSE_ALU_ITINS_S>,
3093 basic_sse12_fp_binop_s_int<0x58, "add", SSE_ALU_ITINS_S>;
3094 defm MUL : basic_sse12_fp_binop_p<0x59, "mul", fmul, SSE_MUL_ITINS_P>,
3095 basic_sse12_fp_binop_s<0x59, "mul", fmul, SSE_MUL_ITINS_S>,
3096 basic_sse12_fp_binop_s_int<0x59, "mul", SSE_MUL_ITINS_S>;
3097 let isCommutable = 0 in {
3098 defm SUB : basic_sse12_fp_binop_p<0x5C, "sub", fsub, SSE_ALU_ITINS_P>,
3099 basic_sse12_fp_binop_s<0x5C, "sub", fsub, SSE_ALU_ITINS_S>,
3100 basic_sse12_fp_binop_s_int<0x5C, "sub", SSE_ALU_ITINS_S>;
3101 defm DIV : basic_sse12_fp_binop_p<0x5E, "div", fdiv, SSE_DIV_ITINS_P>,
3102 basic_sse12_fp_binop_s<0x5E, "div", fdiv, SSE_DIV_ITINS_S>,
3103 basic_sse12_fp_binop_s_int<0x5E, "div", SSE_DIV_ITINS_S>;
3104 defm MAX : basic_sse12_fp_binop_p<0x5F, "max", X86fmax, SSE_ALU_ITINS_P>,
3105 basic_sse12_fp_binop_s<0x5F, "max", X86fmax, SSE_ALU_ITINS_S>,
3106 basic_sse12_fp_binop_s_int<0x5F, "max", SSE_ALU_ITINS_S>;
3107 defm MIN : basic_sse12_fp_binop_p<0x5D, "min", X86fmin, SSE_ALU_ITINS_P>,
3108 basic_sse12_fp_binop_s<0x5D, "min", X86fmin, SSE_ALU_ITINS_S>,
3109 basic_sse12_fp_binop_s_int<0x5D, "min", SSE_ALU_ITINS_S>;
3112 let isCodeGenOnly = 1 in {
3113 defm MAXC: basic_sse12_fp_binop_p<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_P>,
3114 basic_sse12_fp_binop_s<0x5F, "max", X86fmaxc, SSE_ALU_ITINS_S>;
3115 defm MINC: basic_sse12_fp_binop_p<0x5D, "min", X86fminc, SSE_ALU_ITINS_P>,
3116 basic_sse12_fp_binop_s<0x5D, "min", X86fminc, SSE_ALU_ITINS_S>;
3119 // Patterns used to select SSE scalar fp arithmetic instructions from
3122 // (1) a scalar fp operation followed by a blend
3124 // The effect is that the backend no longer emits unnecessary vector
3125 // insert instructions immediately after SSE scalar fp instructions
3126 // like addss or mulss.
3128 // For example, given the following code:
3129 // __m128 foo(__m128 A, __m128 B) {
3134 // Previously we generated:
3135 // addss %xmm0, %xmm1
3136 // movss %xmm1, %xmm0
3139 // addss %xmm1, %xmm0
3141 // (2) a vector packed single/double fp operation followed by a vector insert
3143 // The effect is that the backend converts the packed fp instruction
3144 // followed by a vector insert into a single SSE scalar fp instruction.
3146 // For example, given the following code:
3147 // __m128 foo(__m128 A, __m128 B) {
3148 // __m128 C = A + B;
3149 // return (__m128) {c[0], a[1], a[2], a[3]};
3152 // Previously we generated:
3153 // addps %xmm0, %xmm1
3154 // movss %xmm1, %xmm0
3157 // addss %xmm1, %xmm0
3159 // TODO: Some canonicalization in lowering would simplify the number of
3160 // patterns we have to try to match.
3161 multiclass scalar_math_f32_patterns<SDNode Op, string OpcPrefix> {
3162 let Predicates = [UseSSE1] in {
3163 // extracted scalar math op with insert via movss
3164 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3165 (Op (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3167 (!cast<I>(OpcPrefix#SSrr_Int) v4f32:$dst,
3168 (COPY_TO_REGCLASS FR32:$src, VR128))>;
3170 // vector math op with insert via movss
3171 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3172 (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3173 (!cast<I>(OpcPrefix#SSrr_Int) v4f32:$dst, v4f32:$src)>;
3176 // With SSE 4.1, blendi is preferred to movsd, so match that too.
3177 let Predicates = [UseSSE41] in {
3178 // extracted scalar math op with insert via blend
3179 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3180 (Op (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3181 FR32:$src))), (i8 1))),
3182 (!cast<I>(OpcPrefix#SSrr_Int) v4f32:$dst,
3183 (COPY_TO_REGCLASS FR32:$src, VR128))>;
3185 // vector math op with insert via blend
3186 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$dst),
3187 (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)), (i8 1))),
3188 (!cast<I>(OpcPrefix#SSrr_Int)v4f32:$dst, v4f32:$src)>;
3192 // Repeat everything for AVX, except for the movss + scalar combo...
3193 // because that one shouldn't occur with AVX codegen?
3194 let Predicates = [HasAVX] in {
3195 // extracted scalar math op with insert via blend
3196 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
3197 (Op (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
3198 FR32:$src))), (i8 1))),
3199 (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst,
3200 (COPY_TO_REGCLASS FR32:$src, VR128))>;
3202 // vector math op with insert via movss
3203 def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
3204 (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
3205 (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst, v4f32:$src)>;
3207 // vector math op with insert via blend
3208 def : Pat<(v4f32 (X86Blendi (v4f32 VR128:$dst),
3209 (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)), (i8 1))),
3210 (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst, v4f32:$src)>;
3214 defm : scalar_math_f32_patterns<fadd, "ADD">;
3215 defm : scalar_math_f32_patterns<fsub, "SUB">;
3216 defm : scalar_math_f32_patterns<fmul, "MUL">;
3217 defm : scalar_math_f32_patterns<fdiv, "DIV">;
3219 multiclass scalar_math_f64_patterns<SDNode Op, string OpcPrefix> {
3220 let Predicates = [UseSSE2] in {
3221 // extracted scalar math op with insert via movsd
3222 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
3223 (Op (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3225 (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst,
3226 (COPY_TO_REGCLASS FR64:$src, VR128))>;
3228 // vector math op with insert via movsd
3229 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3230 (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3231 (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
3234 // With SSE 4.1, blendi is preferred to movsd, so match those too.
3235 let Predicates = [UseSSE41] in {
3236 // extracted scalar math op with insert via blend
3237 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
3238 (Op (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3239 FR64:$src))), (i8 1))),
3240 (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst,
3241 (COPY_TO_REGCLASS FR64:$src, VR128))>;
3243 // vector math op with insert via blend
3244 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst),
3245 (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)), (i8 1))),
3246 (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
3249 // Repeat everything for AVX.
3250 let Predicates = [HasAVX] in {
3251 // extracted scalar math op with insert via movsd
3252 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
3253 (Op (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3255 (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst,
3256 (COPY_TO_REGCLASS FR64:$src, VR128))>;
3258 // extracted scalar math op with insert via blend
3259 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
3260 (Op (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
3261 FR64:$src))), (i8 1))),
3262 (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst,
3263 (COPY_TO_REGCLASS FR64:$src, VR128))>;
3265 // vector math op with insert via movsd
3266 def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
3267 (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
3268 (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
3270 // vector math op with insert via blend
3271 def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst),
3272 (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)), (i8 1))),
3273 (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
3277 defm : scalar_math_f64_patterns<fadd, "ADD">;
3278 defm : scalar_math_f64_patterns<fsub, "SUB">;
3279 defm : scalar_math_f64_patterns<fmul, "MUL">;
3280 defm : scalar_math_f64_patterns<fdiv, "DIV">;
3284 /// In addition, we also have a special variant of the scalar form here to
3285 /// represent the associated intrinsic operation. This form is unlike the
3286 /// plain scalar form, in that it takes an entire vector (instead of a
3287 /// scalar) and leaves the top elements undefined.
3289 /// And, we have a special variant form for a full-vector intrinsic form.
3291 let Sched = WriteFSqrt in {
3292 def SSE_SQRTPS : OpndItins<
3293 IIC_SSE_SQRTPS_RR, IIC_SSE_SQRTPS_RM
3296 def SSE_SQRTSS : OpndItins<
3297 IIC_SSE_SQRTSS_RR, IIC_SSE_SQRTSS_RM
3300 def SSE_SQRTPD : OpndItins<
3301 IIC_SSE_SQRTPD_RR, IIC_SSE_SQRTPD_RM
3304 def SSE_SQRTSD : OpndItins<
3305 IIC_SSE_SQRTSD_RR, IIC_SSE_SQRTSD_RM
3309 let Sched = WriteFRsqrt in {
3310 def SSE_RSQRTPS : OpndItins<
3311 IIC_SSE_RSQRTPS_RR, IIC_SSE_RSQRTPS_RM
3314 def SSE_RSQRTSS : OpndItins<
3315 IIC_SSE_RSQRTSS_RR, IIC_SSE_RSQRTSS_RM
3319 let Sched = WriteFRcp in {
3320 def SSE_RCPP : OpndItins<
3321 IIC_SSE_RCPP_RR, IIC_SSE_RCPP_RM
3324 def SSE_RCPS : OpndItins<
3325 IIC_SSE_RCPS_RR, IIC_SSE_RCPS_RM
3329 /// sse_fp_unop_s - SSE1 unops in scalar form
3330 /// For the non-AVX defs, we need $src1 to be tied to $dst because
3331 /// the HW instructions are 2 operand / destructive.
3332 multiclass sse_fp_unop_s<bits<8> opc, string OpcodeStr, RegisterClass RC,
3333 ValueType vt, ValueType ScalarVT,
3334 X86MemOperand x86memop, Operand vec_memop,
3335 ComplexPattern mem_cpat, Intrinsic Intr,
3336 SDNode OpNode, Domain d, OpndItins itins,
3337 Predicate target, string Suffix> {
3338 let hasSideEffects = 0 in {
3339 def r : I<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1),
3340 !strconcat(OpcodeStr, "\t{$src1, $dst|$dst, $src1}"),
3341 [(set RC:$dst, (OpNode RC:$src1))], itins.rr, d>, Sched<[itins.Sched]>,
3344 def m : I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src1),
3345 !strconcat(OpcodeStr, "\t{$src1, $dst|$dst, $src1}"),
3346 [(set RC:$dst, (OpNode (load addr:$src1)))], itins.rm, d>,
3347 Sched<[itins.Sched.Folded, ReadAfterLd]>,
3348 Requires<[target, OptForSize]>;
3350 let isCodeGenOnly = 1, Constraints = "$src1 = $dst" in {
3351 def r_Int : I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
3352 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3353 []>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3355 def m_Int : I<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, vec_memop:$src2),
3356 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3357 []>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3361 let Predicates = [target] in {
3362 def : Pat<(vt (OpNode mem_cpat:$src)),
3363 (vt (COPY_TO_REGCLASS (vt (!cast<Instruction>(NAME#Suffix##m_Int)
3364 (vt (IMPLICIT_DEF)), mem_cpat:$src)), RC))>;
3365 // These are unary operations, but they are modeled as having 2 source operands
3366 // because the high elements of the destination are unchanged in SSE.
3367 def : Pat<(Intr VR128:$src),
3368 (!cast<Instruction>(NAME#Suffix##r_Int) VR128:$src, VR128:$src)>;
3369 def : Pat<(Intr (load addr:$src)),
3370 (vt (COPY_TO_REGCLASS(!cast<Instruction>(NAME#Suffix##m)
3371 addr:$src), VR128))>;
3372 def : Pat<(Intr mem_cpat:$src),
3373 (!cast<Instruction>(NAME#Suffix##m_Int)
3374 (vt (IMPLICIT_DEF)), mem_cpat:$src)>;
3378 multiclass avx_fp_unop_s<bits<8> opc, string OpcodeStr, RegisterClass RC,
3379 ValueType vt, ValueType ScalarVT,
3380 X86MemOperand x86memop, Operand vec_memop,
3381 ComplexPattern mem_cpat,
3382 Intrinsic Intr, SDNode OpNode, Domain d,
3383 OpndItins itins, Predicate target, string Suffix> {
3384 let hasSideEffects = 0 in {
3385 def r : I<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
3386 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3387 [], itins.rr, d>, Sched<[itins.Sched]>;
3389 def m : I<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
3390 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3391 [], itins.rm, d>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3392 let isCodeGenOnly = 1 in {
3393 def r_Int : I<opc, MRMSrcReg, (outs VR128:$dst),
3394 (ins VR128:$src1, VR128:$src2),
3395 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3396 []>, Sched<[itins.Sched.Folded]>;
3398 def m_Int : I<opc, MRMSrcMem, (outs VR128:$dst),
3399 (ins VR128:$src1, vec_memop:$src2),
3400 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3401 []>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3405 let Predicates = [target] in {
3406 def : Pat<(OpNode RC:$src), (!cast<Instruction>("V"#NAME#Suffix##r)
3407 (ScalarVT (IMPLICIT_DEF)), RC:$src)>;
3409 def : Pat<(vt (OpNode mem_cpat:$src)),
3410 (!cast<Instruction>("V"#NAME#Suffix##m_Int) (vt (IMPLICIT_DEF)),
3413 def : Pat<(Intr VR128:$src),
3414 (vt (COPY_TO_REGCLASS(
3415 !cast<Instruction>("V"#NAME#Suffix##r) (ScalarVT (IMPLICIT_DEF)),
3416 (ScalarVT (COPY_TO_REGCLASS VR128:$src, RC))), VR128))>;
3418 def : Pat<(Intr mem_cpat:$src),
3419 (!cast<Instruction>("V"#NAME#Suffix##m_Int)
3420 (vt (IMPLICIT_DEF)), mem_cpat:$src)>;
3422 let Predicates = [target, OptForSize] in
3423 def : Pat<(ScalarVT (OpNode (load addr:$src))),
3424 (!cast<Instruction>("V"#NAME#Suffix##m) (ScalarVT (IMPLICIT_DEF)),
3428 /// sse1_fp_unop_p - SSE1 unops in packed form.
3429 multiclass sse1_fp_unop_p<bits<8> opc, string OpcodeStr, SDNode OpNode,
3431 let Predicates = [HasAVX] in {
3432 def V#NAME#PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3433 !strconcat("v", OpcodeStr,
3434 "ps\t{$src, $dst|$dst, $src}"),
3435 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))],
3436 itins.rr>, VEX, Sched<[itins.Sched]>;
3437 def V#NAME#PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3438 !strconcat("v", OpcodeStr,
3439 "ps\t{$src, $dst|$dst, $src}"),
3440 [(set VR128:$dst, (OpNode (loadv4f32 addr:$src)))],
3441 itins.rm>, VEX, Sched<[itins.Sched.Folded]>;
3442 def V#NAME#PSYr : PSI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3443 !strconcat("v", OpcodeStr,
3444 "ps\t{$src, $dst|$dst, $src}"),
3445 [(set VR256:$dst, (v8f32 (OpNode VR256:$src)))],
3446 itins.rr>, VEX, VEX_L, Sched<[itins.Sched]>;
3447 def V#NAME#PSYm : PSI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
3448 !strconcat("v", OpcodeStr,
3449 "ps\t{$src, $dst|$dst, $src}"),
3450 [(set VR256:$dst, (OpNode (loadv8f32 addr:$src)))],
3451 itins.rm>, VEX, VEX_L, Sched<[itins.Sched.Folded]>;
3454 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3455 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3456 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))], itins.rr>,
3457 Sched<[itins.Sched]>;
3458 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3459 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3460 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))], itins.rm>,
3461 Sched<[itins.Sched.Folded]>;
3464 /// sse2_fp_unop_p - SSE2 unops in vector forms.
3465 multiclass sse2_fp_unop_p<bits<8> opc, string OpcodeStr,
3466 SDNode OpNode, OpndItins itins> {
3467 let Predicates = [HasAVX] in {
3468 def V#NAME#PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3469 !strconcat("v", OpcodeStr,
3470 "pd\t{$src, $dst|$dst, $src}"),
3471 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))],
3472 itins.rr>, VEX, Sched<[itins.Sched]>;
3473 def V#NAME#PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3474 !strconcat("v", OpcodeStr,
3475 "pd\t{$src, $dst|$dst, $src}"),
3476 [(set VR128:$dst, (OpNode (loadv2f64 addr:$src)))],
3477 itins.rm>, VEX, Sched<[itins.Sched.Folded]>;
3478 def V#NAME#PDYr : PDI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3479 !strconcat("v", OpcodeStr,
3480 "pd\t{$src, $dst|$dst, $src}"),
3481 [(set VR256:$dst, (v4f64 (OpNode VR256:$src)))],
3482 itins.rr>, VEX, VEX_L, Sched<[itins.Sched]>;
3483 def V#NAME#PDYm : PDI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
3484 !strconcat("v", OpcodeStr,
3485 "pd\t{$src, $dst|$dst, $src}"),
3486 [(set VR256:$dst, (OpNode (loadv4f64 addr:$src)))],
3487 itins.rm>, VEX, VEX_L, Sched<[itins.Sched.Folded]>;
3490 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3491 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
3492 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))], itins.rr>,
3493 Sched<[itins.Sched]>;
3494 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3495 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
3496 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))], itins.rm>,
3497 Sched<[itins.Sched.Folded]>;
3500 multiclass sse1_fp_unop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
3502 defm SS : sse_fp_unop_s<opc, OpcodeStr##ss, FR32, v4f32, f32, f32mem,
3503 ssmem, sse_load_f32,
3504 !cast<Intrinsic>("int_x86_sse_"##OpcodeStr##_ss), OpNode,
3505 SSEPackedSingle, itins, UseSSE1, "SS">, XS;
3506 defm V#NAME#SS : avx_fp_unop_s<opc, "v"#OpcodeStr##ss, FR32, v4f32, f32,
3507 f32mem, ssmem, sse_load_f32,
3508 !cast<Intrinsic>("int_x86_sse_"##OpcodeStr##_ss), OpNode,
3509 SSEPackedSingle, itins, UseAVX, "SS">, XS, VEX_4V, VEX_LIG;
3512 multiclass sse2_fp_unop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
3514 defm SD : sse_fp_unop_s<opc, OpcodeStr##sd, FR64, v2f64, f64, f64mem,
3515 sdmem, sse_load_f64,
3516 !cast<Intrinsic>("int_x86_sse2_"##OpcodeStr##_sd),
3517 OpNode, SSEPackedDouble, itins, UseSSE2, "SD">, XD;
3518 defm V#NAME#SD : avx_fp_unop_s<opc, "v"#OpcodeStr##sd, FR64, v2f64, f64,
3519 f64mem, sdmem, sse_load_f64,
3520 !cast<Intrinsic>("int_x86_sse2_"##OpcodeStr##_sd),
3521 OpNode, SSEPackedDouble, itins, UseAVX, "SD">,
3522 XD, VEX_4V, VEX_LIG;
3526 defm SQRT : sse1_fp_unop_s<0x51, "sqrt", fsqrt, SSE_SQRTSS>,
3527 sse1_fp_unop_p<0x51, "sqrt", fsqrt, SSE_SQRTPS>,
3528 sse2_fp_unop_s<0x51, "sqrt", fsqrt, SSE_SQRTSD>,
3529 sse2_fp_unop_p<0x51, "sqrt", fsqrt, SSE_SQRTPD>;
3531 // Reciprocal approximations. Note that these typically require refinement
3532 // in order to obtain suitable precision.
3533 defm RSQRT : sse1_fp_unop_s<0x52, "rsqrt", X86frsqrt, SSE_RSQRTSS>,
3534 sse1_fp_unop_p<0x52, "rsqrt", X86frsqrt, SSE_RSQRTPS>;
3535 defm RCP : sse1_fp_unop_s<0x53, "rcp", X86frcp, SSE_RCPS>,
3536 sse1_fp_unop_p<0x53, "rcp", X86frcp, SSE_RCPP>;
3538 // There is no f64 version of the reciprocal approximation instructions.
3540 // TODO: We should add *scalar* op patterns for these just like we have for
3541 // the binops above. If the binop and unop patterns could all be unified
3542 // that would be even better.
3544 multiclass scalar_unary_math_patterns<Intrinsic Intr, string OpcPrefix,
3545 SDNode Move, ValueType VT,
3546 Predicate BasePredicate> {
3547 let Predicates = [BasePredicate] in {
3548 def : Pat<(VT (Move VT:$dst, (Intr VT:$src))),
3549 (!cast<I>(OpcPrefix#r_Int) VT:$dst, VT:$src)>;
3552 // With SSE 4.1, blendi is preferred to movs*, so match that too.
3553 let Predicates = [UseSSE41] in {
3554 def : Pat<(VT (X86Blendi VT:$dst, (Intr VT:$src), (i8 1))),
3555 (!cast<I>(OpcPrefix#r_Int) VT:$dst, VT:$src)>;
3558 // Repeat for AVX versions of the instructions.
3559 let Predicates = [HasAVX] in {
3560 def : Pat<(VT (Move VT:$dst, (Intr VT:$src))),
3561 (!cast<I>("V"#OpcPrefix#r_Int) VT:$dst, VT:$src)>;
3563 def : Pat<(VT (X86Blendi VT:$dst, (Intr VT:$src), (i8 1))),
3564 (!cast<I>("V"#OpcPrefix#r_Int) VT:$dst, VT:$src)>;
3568 defm : scalar_unary_math_patterns<int_x86_sse_rcp_ss, "RCPSS", X86Movss,
3570 defm : scalar_unary_math_patterns<int_x86_sse_rsqrt_ss, "RSQRTSS", X86Movss,
3572 defm : scalar_unary_math_patterns<int_x86_sse_sqrt_ss, "SQRTSS", X86Movss,
3574 defm : scalar_unary_math_patterns<int_x86_sse2_sqrt_sd, "SQRTSD", X86Movsd,
3578 //===----------------------------------------------------------------------===//
3579 // SSE 1 & 2 - Non-temporal stores
3580 //===----------------------------------------------------------------------===//
3582 let AddedComplexity = 400 in { // Prefer non-temporal versions
3583 let SchedRW = [WriteStore] in {
3584 let Predicates = [HasAVX, NoVLX] in {
3585 def VMOVNTPSmr : VPSI<0x2B, MRMDestMem, (outs),
3586 (ins f128mem:$dst, VR128:$src),
3587 "movntps\t{$src, $dst|$dst, $src}",
3588 [(alignednontemporalstore (v4f32 VR128:$src),
3590 IIC_SSE_MOVNT>, VEX;
3591 def VMOVNTPDmr : VPDI<0x2B, MRMDestMem, (outs),
3592 (ins f128mem:$dst, VR128:$src),
3593 "movntpd\t{$src, $dst|$dst, $src}",
3594 [(alignednontemporalstore (v2f64 VR128:$src),
3596 IIC_SSE_MOVNT>, VEX;
3598 let ExeDomain = SSEPackedInt in
3599 def VMOVNTDQmr : VPDI<0xE7, MRMDestMem, (outs),
3600 (ins f128mem:$dst, VR128:$src),
3601 "movntdq\t{$src, $dst|$dst, $src}",
3602 [(alignednontemporalstore (v2i64 VR128:$src),
3604 IIC_SSE_MOVNT>, VEX;
3606 def VMOVNTPSYmr : VPSI<0x2B, MRMDestMem, (outs),
3607 (ins f256mem:$dst, VR256:$src),
3608 "movntps\t{$src, $dst|$dst, $src}",
3609 [(alignednontemporalstore (v8f32 VR256:$src),
3611 IIC_SSE_MOVNT>, VEX, VEX_L;
3612 def VMOVNTPDYmr : VPDI<0x2B, MRMDestMem, (outs),
3613 (ins f256mem:$dst, VR256:$src),
3614 "movntpd\t{$src, $dst|$dst, $src}",
3615 [(alignednontemporalstore (v4f64 VR256:$src),
3617 IIC_SSE_MOVNT>, VEX, VEX_L;
3618 let ExeDomain = SSEPackedInt in
3619 def VMOVNTDQYmr : VPDI<0xE7, MRMDestMem, (outs),
3620 (ins f256mem:$dst, VR256:$src),
3621 "movntdq\t{$src, $dst|$dst, $src}",
3622 [(alignednontemporalstore (v4i64 VR256:$src),
3624 IIC_SSE_MOVNT>, VEX, VEX_L;
3627 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3628 "movntps\t{$src, $dst|$dst, $src}",
3629 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)],
3631 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3632 "movntpd\t{$src, $dst|$dst, $src}",
3633 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)],
3636 let ExeDomain = SSEPackedInt in
3637 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
3638 "movntdq\t{$src, $dst|$dst, $src}",
3639 [(alignednontemporalstore (v2i64 VR128:$src), addr:$dst)],
3642 // There is no AVX form for instructions below this point
3643 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
3644 "movnti{l}\t{$src, $dst|$dst, $src}",
3645 [(nontemporalstore (i32 GR32:$src), addr:$dst)],
3647 PS, Requires<[HasSSE2]>;
3648 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
3649 "movnti{q}\t{$src, $dst|$dst, $src}",
3650 [(nontemporalstore (i64 GR64:$src), addr:$dst)],
3652 PS, Requires<[HasSSE2]>;
3653 } // SchedRW = [WriteStore]
3655 let Predicates = [HasAVX2, NoVLX] in {
3656 def : Pat<(alignednontemporalstore (v8i32 VR256:$src), addr:$dst),
3657 (VMOVNTDQYmr addr:$dst, VR256:$src)>;
3658 def : Pat<(alignednontemporalstore (v16i16 VR256:$src), addr:$dst),
3659 (VMOVNTDQYmr addr:$dst, VR256:$src)>;
3660 def : Pat<(alignednontemporalstore (v32i8 VR256:$src), addr:$dst),
3661 (VMOVNTDQYmr addr:$dst, VR256:$src)>;
3664 let Predicates = [HasAVX, NoVLX] in {
3665 def : Pat<(alignednontemporalstore (v4i32 VR128:$src), addr:$dst),
3666 (VMOVNTDQmr addr:$dst, VR128:$src)>;
3667 def : Pat<(alignednontemporalstore (v8i16 VR128:$src), addr:$dst),
3668 (VMOVNTDQmr addr:$dst, VR128:$src)>;
3669 def : Pat<(alignednontemporalstore (v16i8 VR128:$src), addr:$dst),
3670 (VMOVNTDQmr addr:$dst, VR128:$src)>;
3673 def : Pat<(alignednontemporalstore (v4i32 VR128:$src), addr:$dst),
3674 (MOVNTDQmr addr:$dst, VR128:$src)>;
3675 def : Pat<(alignednontemporalstore (v8i16 VR128:$src), addr:$dst),
3676 (MOVNTDQmr addr:$dst, VR128:$src)>;
3677 def : Pat<(alignednontemporalstore (v16i8 VR128:$src), addr:$dst),
3678 (MOVNTDQmr addr:$dst, VR128:$src)>;
3680 } // AddedComplexity
3682 //===----------------------------------------------------------------------===//
3683 // SSE 1 & 2 - Prefetch and memory fence
3684 //===----------------------------------------------------------------------===//
3686 // Prefetch intrinsic.
3687 let Predicates = [HasSSE1], SchedRW = [WriteLoad] in {
3688 def PREFETCHT0 : I<0x18, MRM1m, (outs), (ins i8mem:$src),
3689 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3), (i32 1))],
3690 IIC_SSE_PREFETCH>, TB;
3691 def PREFETCHT1 : I<0x18, MRM2m, (outs), (ins i8mem:$src),
3692 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2), (i32 1))],
3693 IIC_SSE_PREFETCH>, TB;
3694 def PREFETCHT2 : I<0x18, MRM3m, (outs), (ins i8mem:$src),
3695 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1), (i32 1))],
3696 IIC_SSE_PREFETCH>, TB;
3697 def PREFETCHNTA : I<0x18, MRM0m, (outs), (ins i8mem:$src),
3698 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0), (i32 1))],
3699 IIC_SSE_PREFETCH>, TB;
3702 // FIXME: How should flush instruction be modeled?
3703 let SchedRW = [WriteLoad] in {
3705 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
3706 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)],
3707 IIC_SSE_PREFETCH>, PS, Requires<[HasSSE2]>;
3710 let SchedRW = [WriteNop] in {
3711 // Pause. This "instruction" is encoded as "rep; nop", so even though it
3712 // was introduced with SSE2, it's backward compatible.
3713 def PAUSE : I<0x90, RawFrm, (outs), (ins),
3714 "pause", [(int_x86_sse2_pause)], IIC_SSE_PAUSE>,
3715 OBXS, Requires<[HasSSE2]>;
3718 let SchedRW = [WriteFence] in {
3719 // Load, store, and memory fence
3720 def SFENCE : I<0xAE, MRM_F8, (outs), (ins),
3721 "sfence", [(int_x86_sse_sfence)], IIC_SSE_SFENCE>,
3722 PS, Requires<[HasSSE1]>;
3723 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
3724 "lfence", [(int_x86_sse2_lfence)], IIC_SSE_LFENCE>,
3725 TB, Requires<[HasSSE2]>;
3726 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
3727 "mfence", [(int_x86_sse2_mfence)], IIC_SSE_MFENCE>,
3728 TB, Requires<[HasSSE2]>;
3731 def : Pat<(X86SFence), (SFENCE)>;
3732 def : Pat<(X86LFence), (LFENCE)>;
3733 def : Pat<(X86MFence), (MFENCE)>;
3735 //===----------------------------------------------------------------------===//
3736 // SSE 1 & 2 - Load/Store XCSR register
3737 //===----------------------------------------------------------------------===//
3739 def VLDMXCSR : VPSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
3740 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)],
3741 IIC_SSE_LDMXCSR>, VEX, Sched<[WriteLoad]>;
3742 def VSTMXCSR : VPSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
3743 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)],
3744 IIC_SSE_STMXCSR>, VEX, Sched<[WriteStore]>;
3746 let Predicates = [UseSSE1] in {
3747 def LDMXCSR : I<0xAE, MRM2m, (outs), (ins i32mem:$src),
3748 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)],
3749 IIC_SSE_LDMXCSR>, TB, Sched<[WriteLoad]>;
3750 def STMXCSR : I<0xAE, MRM3m, (outs), (ins i32mem:$dst),
3751 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)],
3752 IIC_SSE_STMXCSR>, TB, Sched<[WriteStore]>;
3755 //===---------------------------------------------------------------------===//
3756 // SSE2 - Move Aligned/Unaligned Packed Integer Instructions
3757 //===---------------------------------------------------------------------===//
3759 let ExeDomain = SSEPackedInt in { // SSE integer instructions
3761 let hasSideEffects = 0, SchedRW = [WriteMove] in {
3762 def VMOVDQArr : VPDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3763 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>,
3765 def VMOVDQAYrr : VPDI<0x6F, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3766 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>,
3768 def VMOVDQUrr : VSSI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3769 "movdqu\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVU_P_RR>,
3771 def VMOVDQUYrr : VSSI<0x6F, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3772 "movdqu\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVU_P_RR>,
3777 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
3778 SchedRW = [WriteMove] in {
3779 def VMOVDQArr_REV : VPDI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3780 "movdqa\t{$src, $dst|$dst, $src}", [],
3783 def VMOVDQAYrr_REV : VPDI<0x7F, MRMDestReg, (outs VR256:$dst), (ins VR256:$src),
3784 "movdqa\t{$src, $dst|$dst, $src}", [],
3785 IIC_SSE_MOVA_P_RR>, VEX, VEX_L;
3786 def VMOVDQUrr_REV : VSSI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3787 "movdqu\t{$src, $dst|$dst, $src}", [],
3790 def VMOVDQUYrr_REV : VSSI<0x7F, MRMDestReg, (outs VR256:$dst), (ins VR256:$src),
3791 "movdqu\t{$src, $dst|$dst, $src}", [],
3792 IIC_SSE_MOVU_P_RR>, VEX, VEX_L;
3795 let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1,
3796 hasSideEffects = 0, SchedRW = [WriteLoad] in {
3797 def VMOVDQArm : VPDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3798 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RM>,
3800 def VMOVDQAYrm : VPDI<0x6F, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
3801 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RM>,
3803 let Predicates = [HasAVX] in {
3804 def VMOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3805 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_RM>,
3807 def VMOVDQUYrm : I<0x6F, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
3808 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_RM>,
3813 let mayStore = 1, hasSideEffects = 0, SchedRW = [WriteStore] in {
3814 def VMOVDQAmr : VPDI<0x7F, MRMDestMem, (outs),
3815 (ins i128mem:$dst, VR128:$src),
3816 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_MR>,
3818 def VMOVDQAYmr : VPDI<0x7F, MRMDestMem, (outs),
3819 (ins i256mem:$dst, VR256:$src),
3820 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_MR>,
3822 let Predicates = [HasAVX] in {
3823 def VMOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3824 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_MR>,
3826 def VMOVDQUYmr : I<0x7F, MRMDestMem, (outs), (ins i256mem:$dst, VR256:$src),
3827 "vmovdqu\t{$src, $dst|$dst, $src}",[], IIC_SSE_MOVU_P_MR>,
3832 let SchedRW = [WriteMove] in {
3833 let hasSideEffects = 0 in
3834 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3835 "movdqa\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVA_P_RR>;
3837 def MOVDQUrr : I<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3838 "movdqu\t{$src, $dst|$dst, $src}",
3839 [], IIC_SSE_MOVU_P_RR>, XS, Requires<[UseSSE2]>;
3842 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in {
3843 def MOVDQArr_REV : PDI<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3844 "movdqa\t{$src, $dst|$dst, $src}", [],
3847 def MOVDQUrr_REV : I<0x7F, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
3848 "movdqu\t{$src, $dst|$dst, $src}",
3849 [], IIC_SSE_MOVU_P_RR>, XS, Requires<[UseSSE2]>;
3853 let canFoldAsLoad = 1, mayLoad = 1, isReMaterializable = 1,
3854 hasSideEffects = 0, SchedRW = [WriteLoad] in {
3855 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3856 "movdqa\t{$src, $dst|$dst, $src}",
3857 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/],
3859 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3860 "movdqu\t{$src, $dst|$dst, $src}",
3861 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/],
3863 XS, Requires<[UseSSE2]>;
3866 let mayStore = 1, hasSideEffects = 0, SchedRW = [WriteStore] in {
3867 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3868 "movdqa\t{$src, $dst|$dst, $src}",
3869 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/],
3871 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
3872 "movdqu\t{$src, $dst|$dst, $src}",
3873 [/*(store (v2i64 VR128:$src), addr:$dst)*/],
3875 XS, Requires<[UseSSE2]>;
3878 } // ExeDomain = SSEPackedInt
3880 let Predicates = [HasAVX] in {
3881 def : Pat<(int_x86_sse2_storeu_dq addr:$dst, VR128:$src),
3882 (VMOVDQUmr addr:$dst, VR128:$src)>;
3883 def : Pat<(int_x86_avx_storeu_dq_256 addr:$dst, VR256:$src),
3884 (VMOVDQUYmr addr:$dst, VR256:$src)>;
3886 let Predicates = [UseSSE2] in
3887 def : Pat<(int_x86_sse2_storeu_dq addr:$dst, VR128:$src),
3888 (MOVDQUmr addr:$dst, VR128:$src)>;
3890 //===---------------------------------------------------------------------===//
3891 // SSE2 - Packed Integer Arithmetic Instructions
3892 //===---------------------------------------------------------------------===//
3894 let Sched = WriteVecIMul in
3895 def SSE_PMADD : OpndItins<
3896 IIC_SSE_PMADD, IIC_SSE_PMADD
3899 let ExeDomain = SSEPackedInt in { // SSE integer instructions
3901 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
3902 RegisterClass RC, PatFrag memop_frag,
3903 X86MemOperand x86memop,
3905 bit IsCommutable = 0,
3907 let isCommutable = IsCommutable in
3908 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
3909 (ins RC:$src1, RC:$src2),
3911 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3912 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3913 [(set RC:$dst, (IntId RC:$src1, RC:$src2))], itins.rr>,
3914 Sched<[itins.Sched]>;
3915 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
3916 (ins RC:$src1, x86memop:$src2),
3918 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3919 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3920 [(set RC:$dst, (IntId RC:$src1, (bitconvert (memop_frag addr:$src2))))],
3921 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
3924 multiclass PDI_binop_all_int<bits<8> opc, string OpcodeStr, Intrinsic IntId128,
3925 Intrinsic IntId256, OpndItins itins,
3926 bit IsCommutable = 0> {
3927 let Predicates = [HasAVX] in
3928 defm V#NAME : PDI_binop_rm_int<opc, !strconcat("v", OpcodeStr), IntId128,
3929 VR128, loadv2i64, i128mem, itins,
3930 IsCommutable, 0>, VEX_4V;
3932 let Constraints = "$src1 = $dst" in
3933 defm NAME : PDI_binop_rm_int<opc, OpcodeStr, IntId128, VR128, memopv2i64,
3934 i128mem, itins, IsCommutable, 1>;
3936 let Predicates = [HasAVX2] in
3937 defm V#NAME#Y : PDI_binop_rm_int<opc, !strconcat("v", OpcodeStr), IntId256,
3938 VR256, loadv4i64, i256mem, itins,
3939 IsCommutable, 0>, VEX_4V, VEX_L;
3942 multiclass PDI_binop_rmi<bits<8> opc, bits<8> opc2, Format ImmForm,
3943 string OpcodeStr, SDNode OpNode,
3944 SDNode OpNode2, RegisterClass RC,
3945 ValueType DstVT, ValueType SrcVT, PatFrag bc_frag,
3946 PatFrag ld_frag, ShiftOpndItins itins,
3948 // src2 is always 128-bit
3949 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
3950 (ins RC:$src1, VR128:$src2),
3952 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3953 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3954 [(set RC:$dst, (DstVT (OpNode RC:$src1, (SrcVT VR128:$src2))))],
3955 itins.rr>, Sched<[WriteVecShift]>;
3956 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
3957 (ins RC:$src1, i128mem:$src2),
3959 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3960 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3961 [(set RC:$dst, (DstVT (OpNode RC:$src1,
3962 (bc_frag (ld_frag addr:$src2)))))], itins.rm>,
3963 Sched<[WriteVecShiftLd, ReadAfterLd]>;
3964 def ri : PDIi8<opc2, ImmForm, (outs RC:$dst),
3965 (ins RC:$src1, u8imm:$src2),
3967 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3968 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3969 [(set RC:$dst, (DstVT (OpNode2 RC:$src1, (i8 imm:$src2))))], itins.ri>,
3970 Sched<[WriteVecShift]>;
3973 /// PDI_binop_rm2 - Simple SSE2 binary operator with different src and dst types
3974 multiclass PDI_binop_rm2<bits<8> opc, string OpcodeStr, SDNode OpNode,
3975 ValueType DstVT, ValueType SrcVT, RegisterClass RC,
3976 PatFrag memop_frag, X86MemOperand x86memop,
3978 bit IsCommutable = 0, bit Is2Addr = 1> {
3979 let isCommutable = IsCommutable in
3980 def rr : PDI<opc, MRMSrcReg, (outs RC:$dst),
3981 (ins RC:$src1, RC:$src2),
3983 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3984 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3985 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1), RC:$src2)))]>,
3986 Sched<[itins.Sched]>;
3987 def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
3988 (ins RC:$src1, x86memop:$src2),
3990 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3991 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3992 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1),
3993 (bitconvert (memop_frag addr:$src2)))))]>,
3994 Sched<[itins.Sched.Folded, ReadAfterLd]>;
3996 } // ExeDomain = SSEPackedInt
3998 defm PADDB : PDI_binop_all<0xFC, "paddb", add, v16i8, v32i8,
3999 SSE_INTALU_ITINS_P, 1>;
4000 defm PADDW : PDI_binop_all<0xFD, "paddw", add, v8i16, v16i16,
4001 SSE_INTALU_ITINS_P, 1>;
4002 defm PADDD : PDI_binop_all<0xFE, "paddd", add, v4i32, v8i32,
4003 SSE_INTALU_ITINS_P, 1>;
4004 defm PADDQ : PDI_binop_all<0xD4, "paddq", add, v2i64, v4i64,
4005 SSE_INTALUQ_ITINS_P, 1>;
4006 defm PMULLW : PDI_binop_all<0xD5, "pmullw", mul, v8i16, v16i16,
4007 SSE_INTMUL_ITINS_P, 1>;
4008 defm PMULHUW : PDI_binop_all<0xE4, "pmulhuw", mulhu, v8i16, v16i16,
4009 SSE_INTMUL_ITINS_P, 1>;
4010 defm PMULHW : PDI_binop_all<0xE5, "pmulhw", mulhs, v8i16, v16i16,
4011 SSE_INTMUL_ITINS_P, 1>;
4012 defm PSUBB : PDI_binop_all<0xF8, "psubb", sub, v16i8, v32i8,
4013 SSE_INTALU_ITINS_P, 0>;
4014 defm PSUBW : PDI_binop_all<0xF9, "psubw", sub, v8i16, v16i16,
4015 SSE_INTALU_ITINS_P, 0>;
4016 defm PSUBD : PDI_binop_all<0xFA, "psubd", sub, v4i32, v8i32,
4017 SSE_INTALU_ITINS_P, 0>;
4018 defm PSUBQ : PDI_binop_all<0xFB, "psubq", sub, v2i64, v4i64,
4019 SSE_INTALUQ_ITINS_P, 0>;
4020 defm PSUBUSB : PDI_binop_all<0xD8, "psubusb", X86subus, v16i8, v32i8,
4021 SSE_INTALU_ITINS_P, 0>;
4022 defm PSUBUSW : PDI_binop_all<0xD9, "psubusw", X86subus, v8i16, v16i16,
4023 SSE_INTALU_ITINS_P, 0>;
4024 defm PMINUB : PDI_binop_all<0xDA, "pminub", X86umin, v16i8, v32i8,
4025 SSE_INTALU_ITINS_P, 1>;
4026 defm PMINSW : PDI_binop_all<0xEA, "pminsw", X86smin, v8i16, v16i16,
4027 SSE_INTALU_ITINS_P, 1>;
4028 defm PMAXUB : PDI_binop_all<0xDE, "pmaxub", X86umax, v16i8, v32i8,
4029 SSE_INTALU_ITINS_P, 1>;
4030 defm PMAXSW : PDI_binop_all<0xEE, "pmaxsw", X86smax, v8i16, v16i16,
4031 SSE_INTALU_ITINS_P, 1>;
4034 defm PSUBSB : PDI_binop_all_int<0xE8, "psubsb", int_x86_sse2_psubs_b,
4035 int_x86_avx2_psubs_b, SSE_INTALU_ITINS_P, 0>;
4036 defm PSUBSW : PDI_binop_all_int<0xE9, "psubsw" , int_x86_sse2_psubs_w,
4037 int_x86_avx2_psubs_w, SSE_INTALU_ITINS_P, 0>;
4038 defm PADDSB : PDI_binop_all_int<0xEC, "paddsb" , int_x86_sse2_padds_b,
4039 int_x86_avx2_padds_b, SSE_INTALU_ITINS_P, 1>;
4040 defm PADDSW : PDI_binop_all_int<0xED, "paddsw" , int_x86_sse2_padds_w,
4041 int_x86_avx2_padds_w, SSE_INTALU_ITINS_P, 1>;
4042 defm PADDUSB : PDI_binop_all_int<0xDC, "paddusb", int_x86_sse2_paddus_b,
4043 int_x86_avx2_paddus_b, SSE_INTALU_ITINS_P, 1>;
4044 defm PADDUSW : PDI_binop_all_int<0xDD, "paddusw", int_x86_sse2_paddus_w,
4045 int_x86_avx2_paddus_w, SSE_INTALU_ITINS_P, 1>;
4046 defm PMADDWD : PDI_binop_all_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd,
4047 int_x86_avx2_pmadd_wd, SSE_PMADD, 1>;
4048 defm PAVGB : PDI_binop_all_int<0xE0, "pavgb", int_x86_sse2_pavg_b,
4049 int_x86_avx2_pavg_b, SSE_INTALU_ITINS_P, 1>;
4050 defm PAVGW : PDI_binop_all_int<0xE3, "pavgw", int_x86_sse2_pavg_w,
4051 int_x86_avx2_pavg_w, SSE_INTALU_ITINS_P, 1>;
4052 defm PSADBW : PDI_binop_all_int<0xF6, "psadbw", int_x86_sse2_psad_bw,
4053 int_x86_avx2_psad_bw, SSE_PMADD, 1>;
4055 let Predicates = [HasAVX] in
4056 defm VPMULUDQ : PDI_binop_rm2<0xF4, "vpmuludq", X86pmuludq, v2i64, v4i32, VR128,
4057 loadv2i64, i128mem, SSE_INTMUL_ITINS_P, 1, 0>,
4059 let Predicates = [HasAVX2] in
4060 defm VPMULUDQY : PDI_binop_rm2<0xF4, "vpmuludq", X86pmuludq, v4i64, v8i32,
4061 VR256, loadv4i64, i256mem,
4062 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V, VEX_L;
4063 let Constraints = "$src1 = $dst" in
4064 defm PMULUDQ : PDI_binop_rm2<0xF4, "pmuludq", X86pmuludq, v2i64, v4i32, VR128,
4065 memopv2i64, i128mem, SSE_INTMUL_ITINS_P, 1>;
4067 //===---------------------------------------------------------------------===//
4068 // SSE2 - Packed Integer Logical Instructions
4069 //===---------------------------------------------------------------------===//
4071 let Predicates = [HasAVX, NoVLX] in {
4072 defm VPSLLW : PDI_binop_rmi<0xF1, 0x71, MRM6r, "vpsllw", X86vshl, X86vshli,
4073 VR128, v8i16, v8i16, bc_v8i16, loadv2i64,
4074 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4075 defm VPSLLD : PDI_binop_rmi<0xF2, 0x72, MRM6r, "vpslld", X86vshl, X86vshli,
4076 VR128, v4i32, v4i32, bc_v4i32, loadv2i64,
4077 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4078 defm VPSLLQ : PDI_binop_rmi<0xF3, 0x73, MRM6r, "vpsllq", X86vshl, X86vshli,
4079 VR128, v2i64, v2i64, bc_v2i64, loadv2i64,
4080 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4082 defm VPSRLW : PDI_binop_rmi<0xD1, 0x71, MRM2r, "vpsrlw", X86vsrl, X86vsrli,
4083 VR128, v8i16, v8i16, bc_v8i16, loadv2i64,
4084 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4085 defm VPSRLD : PDI_binop_rmi<0xD2, 0x72, MRM2r, "vpsrld", X86vsrl, X86vsrli,
4086 VR128, v4i32, v4i32, bc_v4i32, loadv2i64,
4087 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4088 defm VPSRLQ : PDI_binop_rmi<0xD3, 0x73, MRM2r, "vpsrlq", X86vsrl, X86vsrli,
4089 VR128, v2i64, v2i64, bc_v2i64, loadv2i64,
4090 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4092 defm VPSRAW : PDI_binop_rmi<0xE1, 0x71, MRM4r, "vpsraw", X86vsra, X86vsrai,
4093 VR128, v8i16, v8i16, bc_v8i16, loadv2i64,
4094 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4095 defm VPSRAD : PDI_binop_rmi<0xE2, 0x72, MRM4r, "vpsrad", X86vsra, X86vsrai,
4096 VR128, v4i32, v4i32, bc_v4i32, loadv2i64,
4097 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V;
4099 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift] in {
4100 // 128-bit logical shifts.
4101 def VPSLLDQri : PDIi8<0x73, MRM7r,
4102 (outs VR128:$dst), (ins VR128:$src1, u8imm:$src2),
4103 "vpslldq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4105 (v2i64 (X86vshldq VR128:$src1, (i8 imm:$src2))))]>,
4107 def VPSRLDQri : PDIi8<0x73, MRM3r,
4108 (outs VR128:$dst), (ins VR128:$src1, u8imm:$src2),
4109 "vpsrldq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4111 (v2i64 (X86vshrdq VR128:$src1, (i8 imm:$src2))))]>,
4113 // PSRADQri doesn't exist in SSE[1-3].
4115 } // Predicates = [HasAVX]
4117 let Predicates = [HasAVX2, NoVLX] in {
4118 defm VPSLLWY : PDI_binop_rmi<0xF1, 0x71, MRM6r, "vpsllw", X86vshl, X86vshli,
4119 VR256, v16i16, v8i16, bc_v8i16, loadv2i64,
4120 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4121 defm VPSLLDY : PDI_binop_rmi<0xF2, 0x72, MRM6r, "vpslld", X86vshl, X86vshli,
4122 VR256, v8i32, v4i32, bc_v4i32, loadv2i64,
4123 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4124 defm VPSLLQY : PDI_binop_rmi<0xF3, 0x73, MRM6r, "vpsllq", X86vshl, X86vshli,
4125 VR256, v4i64, v2i64, bc_v2i64, loadv2i64,
4126 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4128 defm VPSRLWY : PDI_binop_rmi<0xD1, 0x71, MRM2r, "vpsrlw", X86vsrl, X86vsrli,
4129 VR256, v16i16, v8i16, bc_v8i16, loadv2i64,
4130 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4131 defm VPSRLDY : PDI_binop_rmi<0xD2, 0x72, MRM2r, "vpsrld", X86vsrl, X86vsrli,
4132 VR256, v8i32, v4i32, bc_v4i32, loadv2i64,
4133 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4134 defm VPSRLQY : PDI_binop_rmi<0xD3, 0x73, MRM2r, "vpsrlq", X86vsrl, X86vsrli,
4135 VR256, v4i64, v2i64, bc_v2i64, loadv2i64,
4136 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4138 defm VPSRAWY : PDI_binop_rmi<0xE1, 0x71, MRM4r, "vpsraw", X86vsra, X86vsrai,
4139 VR256, v16i16, v8i16, bc_v8i16, loadv2i64,
4140 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4141 defm VPSRADY : PDI_binop_rmi<0xE2, 0x72, MRM4r, "vpsrad", X86vsra, X86vsrai,
4142 VR256, v8i32, v4i32, bc_v4i32, loadv2i64,
4143 SSE_INTSHIFT_ITINS_P, 0>, VEX_4V, VEX_L;
4145 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift], hasSideEffects = 0 in {
4146 // 256-bit logical shifts.
4147 def VPSLLDQYri : PDIi8<0x73, MRM7r,
4148 (outs VR256:$dst), (ins VR256:$src1, u8imm:$src2),
4149 "vpslldq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4151 (v4i64 (X86vshldq VR256:$src1, (i8 imm:$src2))))]>,
4153 def VPSRLDQYri : PDIi8<0x73, MRM3r,
4154 (outs VR256:$dst), (ins VR256:$src1, u8imm:$src2),
4155 "vpsrldq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4157 (v4i64 (X86vshrdq VR256:$src1, (i8 imm:$src2))))]>,
4159 // PSRADQYri doesn't exist in SSE[1-3].
4161 } // Predicates = [HasAVX2]
4163 let Constraints = "$src1 = $dst" in {
4164 defm PSLLW : PDI_binop_rmi<0xF1, 0x71, MRM6r, "psllw", X86vshl, X86vshli,
4165 VR128, v8i16, v8i16, bc_v8i16, memopv2i64,
4166 SSE_INTSHIFT_ITINS_P>;
4167 defm PSLLD : PDI_binop_rmi<0xF2, 0x72, MRM6r, "pslld", X86vshl, X86vshli,
4168 VR128, v4i32, v4i32, bc_v4i32, memopv2i64,
4169 SSE_INTSHIFT_ITINS_P>;
4170 defm PSLLQ : PDI_binop_rmi<0xF3, 0x73, MRM6r, "psllq", X86vshl, X86vshli,
4171 VR128, v2i64, v2i64, bc_v2i64, memopv2i64,
4172 SSE_INTSHIFT_ITINS_P>;
4174 defm PSRLW : PDI_binop_rmi<0xD1, 0x71, MRM2r, "psrlw", X86vsrl, X86vsrli,
4175 VR128, v8i16, v8i16, bc_v8i16, memopv2i64,
4176 SSE_INTSHIFT_ITINS_P>;
4177 defm PSRLD : PDI_binop_rmi<0xD2, 0x72, MRM2r, "psrld", X86vsrl, X86vsrli,
4178 VR128, v4i32, v4i32, bc_v4i32, memopv2i64,
4179 SSE_INTSHIFT_ITINS_P>;
4180 defm PSRLQ : PDI_binop_rmi<0xD3, 0x73, MRM2r, "psrlq", X86vsrl, X86vsrli,
4181 VR128, v2i64, v2i64, bc_v2i64, memopv2i64,
4182 SSE_INTSHIFT_ITINS_P>;
4184 defm PSRAW : PDI_binop_rmi<0xE1, 0x71, MRM4r, "psraw", X86vsra, X86vsrai,
4185 VR128, v8i16, v8i16, bc_v8i16, memopv2i64,
4186 SSE_INTSHIFT_ITINS_P>;
4187 defm PSRAD : PDI_binop_rmi<0xE2, 0x72, MRM4r, "psrad", X86vsra, X86vsrai,
4188 VR128, v4i32, v4i32, bc_v4i32, memopv2i64,
4189 SSE_INTSHIFT_ITINS_P>;
4191 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecShift], hasSideEffects = 0 in {
4192 // 128-bit logical shifts.
4193 def PSLLDQri : PDIi8<0x73, MRM7r,
4194 (outs VR128:$dst), (ins VR128:$src1, u8imm:$src2),
4195 "pslldq\t{$src2, $dst|$dst, $src2}",
4197 (v2i64 (X86vshldq VR128:$src1, (i8 imm:$src2))))],
4198 IIC_SSE_INTSHDQ_P_RI>;
4199 def PSRLDQri : PDIi8<0x73, MRM3r,
4200 (outs VR128:$dst), (ins VR128:$src1, u8imm:$src2),
4201 "psrldq\t{$src2, $dst|$dst, $src2}",
4203 (v2i64 (X86vshrdq VR128:$src1, (i8 imm:$src2))))],
4204 IIC_SSE_INTSHDQ_P_RI>;
4205 // PSRADQri doesn't exist in SSE[1-3].
4207 } // Constraints = "$src1 = $dst"
4209 let Predicates = [HasAVX] in {
4210 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
4211 (VPSRLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4214 let Predicates = [UseSSE2] in {
4215 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
4216 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2))>;
4219 //===---------------------------------------------------------------------===//
4220 // SSE2 - Packed Integer Comparison Instructions
4221 //===---------------------------------------------------------------------===//
4223 defm PCMPEQB : PDI_binop_all<0x74, "pcmpeqb", X86pcmpeq, v16i8, v32i8,
4224 SSE_INTALU_ITINS_P, 1>;
4225 defm PCMPEQW : PDI_binop_all<0x75, "pcmpeqw", X86pcmpeq, v8i16, v16i16,
4226 SSE_INTALU_ITINS_P, 1>;
4227 defm PCMPEQD : PDI_binop_all<0x76, "pcmpeqd", X86pcmpeq, v4i32, v8i32,
4228 SSE_INTALU_ITINS_P, 1>;
4229 defm PCMPGTB : PDI_binop_all<0x64, "pcmpgtb", X86pcmpgt, v16i8, v32i8,
4230 SSE_INTALU_ITINS_P, 0>;
4231 defm PCMPGTW : PDI_binop_all<0x65, "pcmpgtw", X86pcmpgt, v8i16, v16i16,
4232 SSE_INTALU_ITINS_P, 0>;
4233 defm PCMPGTD : PDI_binop_all<0x66, "pcmpgtd", X86pcmpgt, v4i32, v8i32,
4234 SSE_INTALU_ITINS_P, 0>;
4236 //===---------------------------------------------------------------------===//
4237 // SSE2 - Packed Integer Shuffle Instructions
4238 //===---------------------------------------------------------------------===//
4240 let ExeDomain = SSEPackedInt in {
4241 multiclass sse2_pshuffle<string OpcodeStr, ValueType vt128, ValueType vt256,
4243 let Predicates = [HasAVX] in {
4244 def V#NAME#ri : Ii8<0x70, MRMSrcReg, (outs VR128:$dst),
4245 (ins VR128:$src1, u8imm:$src2),
4246 !strconcat("v", OpcodeStr,
4247 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4249 (vt128 (OpNode VR128:$src1, (i8 imm:$src2))))],
4250 IIC_SSE_PSHUF_RI>, VEX, Sched<[WriteShuffle]>;
4251 def V#NAME#mi : Ii8<0x70, MRMSrcMem, (outs VR128:$dst),
4252 (ins i128mem:$src1, u8imm:$src2),
4253 !strconcat("v", OpcodeStr,
4254 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4256 (vt128 (OpNode (bitconvert (loadv2i64 addr:$src1)),
4257 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>, VEX,
4258 Sched<[WriteShuffleLd]>;
4261 let Predicates = [HasAVX2] in {
4262 def V#NAME#Yri : Ii8<0x70, MRMSrcReg, (outs VR256:$dst),
4263 (ins VR256:$src1, u8imm:$src2),
4264 !strconcat("v", OpcodeStr,
4265 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4267 (vt256 (OpNode VR256:$src1, (i8 imm:$src2))))],
4268 IIC_SSE_PSHUF_RI>, VEX, VEX_L, Sched<[WriteShuffle]>;
4269 def V#NAME#Ymi : Ii8<0x70, MRMSrcMem, (outs VR256:$dst),
4270 (ins i256mem:$src1, u8imm:$src2),
4271 !strconcat("v", OpcodeStr,
4272 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4274 (vt256 (OpNode (bitconvert (loadv4i64 addr:$src1)),
4275 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>, VEX, VEX_L,
4276 Sched<[WriteShuffleLd]>;
4279 let Predicates = [UseSSE2] in {
4280 def ri : Ii8<0x70, MRMSrcReg,
4281 (outs VR128:$dst), (ins VR128:$src1, u8imm:$src2),
4282 !strconcat(OpcodeStr,
4283 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4285 (vt128 (OpNode VR128:$src1, (i8 imm:$src2))))],
4286 IIC_SSE_PSHUF_RI>, Sched<[WriteShuffle]>;
4287 def mi : Ii8<0x70, MRMSrcMem,
4288 (outs VR128:$dst), (ins i128mem:$src1, u8imm:$src2),
4289 !strconcat(OpcodeStr,
4290 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4292 (vt128 (OpNode (bitconvert (memopv2i64 addr:$src1)),
4293 (i8 imm:$src2))))], IIC_SSE_PSHUF_MI>,
4294 Sched<[WriteShuffleLd, ReadAfterLd]>;
4297 } // ExeDomain = SSEPackedInt
4299 defm PSHUFD : sse2_pshuffle<"pshufd", v4i32, v8i32, X86PShufd>, PD;
4300 defm PSHUFHW : sse2_pshuffle<"pshufhw", v8i16, v16i16, X86PShufhw>, XS;
4301 defm PSHUFLW : sse2_pshuffle<"pshuflw", v8i16, v16i16, X86PShuflw>, XD;
4303 let Predicates = [HasAVX] in {
4304 def : Pat<(v4f32 (X86PShufd (loadv4f32 addr:$src1), (i8 imm:$imm))),
4305 (VPSHUFDmi addr:$src1, imm:$imm)>;
4306 def : Pat<(v4f32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
4307 (VPSHUFDri VR128:$src1, imm:$imm)>;
4310 let Predicates = [UseSSE2] in {
4311 def : Pat<(v4f32 (X86PShufd (memopv4f32 addr:$src1), (i8 imm:$imm))),
4312 (PSHUFDmi addr:$src1, imm:$imm)>;
4313 def : Pat<(v4f32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
4314 (PSHUFDri VR128:$src1, imm:$imm)>;
4317 //===---------------------------------------------------------------------===//
4318 // Packed Integer Pack Instructions (SSE & AVX)
4319 //===---------------------------------------------------------------------===//
4321 let ExeDomain = SSEPackedInt in {
4322 multiclass sse2_pack<bits<8> opc, string OpcodeStr, ValueType OutVT,
4323 ValueType ArgVT, SDNode OpNode, PatFrag bc_frag,
4324 PatFrag ld_frag, bit Is2Addr = 1> {
4325 def rr : PDI<opc, MRMSrcReg,
4326 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4328 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4329 !strconcat(OpcodeStr,
4330 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4332 (OutVT (OpNode (ArgVT VR128:$src1), VR128:$src2)))]>,
4333 Sched<[WriteShuffle]>;
4334 def rm : PDI<opc, MRMSrcMem,
4335 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4337 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4338 !strconcat(OpcodeStr,
4339 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4341 (OutVT (OpNode VR128:$src1,
4342 (bc_frag (ld_frag addr:$src2)))))]>,
4343 Sched<[WriteShuffleLd, ReadAfterLd]>;
4346 multiclass sse2_pack_y<bits<8> opc, string OpcodeStr, ValueType OutVT,
4347 ValueType ArgVT, SDNode OpNode, PatFrag bc_frag> {
4348 def Yrr : PDI<opc, MRMSrcReg,
4349 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4350 !strconcat(OpcodeStr,
4351 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4353 (OutVT (OpNode (ArgVT VR256:$src1), VR256:$src2)))]>,
4354 Sched<[WriteShuffle]>;
4355 def Yrm : PDI<opc, MRMSrcMem,
4356 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4357 !strconcat(OpcodeStr,
4358 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4360 (OutVT (OpNode VR256:$src1,
4361 (bc_frag (loadv4i64 addr:$src2)))))]>,
4362 Sched<[WriteShuffleLd, ReadAfterLd]>;
4365 multiclass sse4_pack<bits<8> opc, string OpcodeStr, ValueType OutVT,
4366 ValueType ArgVT, SDNode OpNode, PatFrag bc_frag,
4367 PatFrag ld_frag, bit Is2Addr = 1> {
4368 def rr : SS48I<opc, MRMSrcReg,
4369 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4371 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4372 !strconcat(OpcodeStr,
4373 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4375 (OutVT (OpNode (ArgVT VR128:$src1), VR128:$src2)))]>,
4376 Sched<[WriteShuffle]>;
4377 def rm : SS48I<opc, MRMSrcMem,
4378 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4380 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4381 !strconcat(OpcodeStr,
4382 "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4384 (OutVT (OpNode VR128:$src1,
4385 (bc_frag (ld_frag addr:$src2)))))]>,
4386 Sched<[WriteShuffleLd, ReadAfterLd]>;
4389 multiclass sse4_pack_y<bits<8> opc, string OpcodeStr, ValueType OutVT,
4390 ValueType ArgVT, SDNode OpNode, PatFrag bc_frag> {
4391 def Yrr : SS48I<opc, MRMSrcReg,
4392 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4393 !strconcat(OpcodeStr,
4394 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4396 (OutVT (OpNode (ArgVT VR256:$src1), VR256:$src2)))]>,
4397 Sched<[WriteShuffle]>;
4398 def Yrm : SS48I<opc, MRMSrcMem,
4399 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4400 !strconcat(OpcodeStr,
4401 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4403 (OutVT (OpNode VR256:$src1,
4404 (bc_frag (loadv4i64 addr:$src2)))))]>,
4405 Sched<[WriteShuffleLd, ReadAfterLd]>;
4408 let Predicates = [HasAVX] in {
4409 defm VPACKSSWB : sse2_pack<0x63, "vpacksswb", v16i8, v8i16, X86Packss,
4410 bc_v8i16, loadv2i64, 0>, VEX_4V;
4411 defm VPACKSSDW : sse2_pack<0x6B, "vpackssdw", v8i16, v4i32, X86Packss,
4412 bc_v4i32, loadv2i64, 0>, VEX_4V;
4414 defm VPACKUSWB : sse2_pack<0x67, "vpackuswb", v16i8, v8i16, X86Packus,
4415 bc_v8i16, loadv2i64, 0>, VEX_4V;
4416 defm VPACKUSDW : sse4_pack<0x2B, "vpackusdw", v8i16, v4i32, X86Packus,
4417 bc_v4i32, loadv2i64, 0>, VEX_4V;
4420 let Predicates = [HasAVX2] in {
4421 defm VPACKSSWB : sse2_pack_y<0x63, "vpacksswb", v32i8, v16i16, X86Packss,
4422 bc_v16i16>, VEX_4V, VEX_L;
4423 defm VPACKSSDW : sse2_pack_y<0x6B, "vpackssdw", v16i16, v8i32, X86Packss,
4424 bc_v8i32>, VEX_4V, VEX_L;
4426 defm VPACKUSWB : sse2_pack_y<0x67, "vpackuswb", v32i8, v16i16, X86Packus,
4427 bc_v16i16>, VEX_4V, VEX_L;
4428 defm VPACKUSDW : sse4_pack_y<0x2B, "vpackusdw", v16i16, v8i32, X86Packus,
4429 bc_v8i32>, VEX_4V, VEX_L;
4432 let Constraints = "$src1 = $dst" in {
4433 defm PACKSSWB : sse2_pack<0x63, "packsswb", v16i8, v8i16, X86Packss,
4434 bc_v8i16, memopv2i64>;
4435 defm PACKSSDW : sse2_pack<0x6B, "packssdw", v8i16, v4i32, X86Packss,
4436 bc_v4i32, memopv2i64>;
4438 defm PACKUSWB : sse2_pack<0x67, "packuswb", v16i8, v8i16, X86Packus,
4439 bc_v8i16, memopv2i64>;
4441 let Predicates = [HasSSE41] in
4442 defm PACKUSDW : sse4_pack<0x2B, "packusdw", v8i16, v4i32, X86Packus,
4443 bc_v4i32, memopv2i64>;
4445 } // ExeDomain = SSEPackedInt
4447 //===---------------------------------------------------------------------===//
4448 // SSE2 - Packed Integer Unpack Instructions
4449 //===---------------------------------------------------------------------===//
4451 let ExeDomain = SSEPackedInt in {
4452 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
4453 SDNode OpNode, PatFrag bc_frag, PatFrag ld_frag,
4455 def rr : PDI<opc, MRMSrcReg,
4456 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
4458 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
4459 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4460 [(set VR128:$dst, (vt (OpNode VR128:$src1, VR128:$src2)))],
4461 IIC_SSE_UNPCK>, Sched<[WriteShuffle]>;
4462 def rm : PDI<opc, MRMSrcMem,
4463 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
4465 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
4466 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4467 [(set VR128:$dst, (OpNode VR128:$src1,
4468 (bc_frag (ld_frag addr:$src2))))],
4470 Sched<[WriteShuffleLd, ReadAfterLd]>;
4473 multiclass sse2_unpack_y<bits<8> opc, string OpcodeStr, ValueType vt,
4474 SDNode OpNode, PatFrag bc_frag> {
4475 def Yrr : PDI<opc, MRMSrcReg,
4476 (outs VR256:$dst), (ins VR256:$src1, VR256:$src2),
4477 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4478 [(set VR256:$dst, (vt (OpNode VR256:$src1, VR256:$src2)))]>,
4479 Sched<[WriteShuffle]>;
4480 def Yrm : PDI<opc, MRMSrcMem,
4481 (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
4482 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4483 [(set VR256:$dst, (OpNode VR256:$src1,
4484 (bc_frag (loadv4i64 addr:$src2))))]>,
4485 Sched<[WriteShuffleLd, ReadAfterLd]>;
4488 let Predicates = [HasAVX] in {
4489 defm VPUNPCKLBW : sse2_unpack<0x60, "vpunpcklbw", v16i8, X86Unpckl,
4490 bc_v16i8, loadv2i64, 0>, VEX_4V;
4491 defm VPUNPCKLWD : sse2_unpack<0x61, "vpunpcklwd", v8i16, X86Unpckl,
4492 bc_v8i16, loadv2i64, 0>, VEX_4V;
4493 defm VPUNPCKLDQ : sse2_unpack<0x62, "vpunpckldq", v4i32, X86Unpckl,
4494 bc_v4i32, loadv2i64, 0>, VEX_4V;
4495 defm VPUNPCKLQDQ : sse2_unpack<0x6C, "vpunpcklqdq", v2i64, X86Unpckl,
4496 bc_v2i64, loadv2i64, 0>, VEX_4V;
4498 defm VPUNPCKHBW : sse2_unpack<0x68, "vpunpckhbw", v16i8, X86Unpckh,
4499 bc_v16i8, loadv2i64, 0>, VEX_4V;
4500 defm VPUNPCKHWD : sse2_unpack<0x69, "vpunpckhwd", v8i16, X86Unpckh,
4501 bc_v8i16, loadv2i64, 0>, VEX_4V;
4502 defm VPUNPCKHDQ : sse2_unpack<0x6A, "vpunpckhdq", v4i32, X86Unpckh,
4503 bc_v4i32, loadv2i64, 0>, VEX_4V;
4504 defm VPUNPCKHQDQ : sse2_unpack<0x6D, "vpunpckhqdq", v2i64, X86Unpckh,
4505 bc_v2i64, loadv2i64, 0>, VEX_4V;
4508 let Predicates = [HasAVX2] in {
4509 defm VPUNPCKLBW : sse2_unpack_y<0x60, "vpunpcklbw", v32i8, X86Unpckl,
4510 bc_v32i8>, VEX_4V, VEX_L;
4511 defm VPUNPCKLWD : sse2_unpack_y<0x61, "vpunpcklwd", v16i16, X86Unpckl,
4512 bc_v16i16>, VEX_4V, VEX_L;
4513 defm VPUNPCKLDQ : sse2_unpack_y<0x62, "vpunpckldq", v8i32, X86Unpckl,
4514 bc_v8i32>, VEX_4V, VEX_L;
4515 defm VPUNPCKLQDQ : sse2_unpack_y<0x6C, "vpunpcklqdq", v4i64, X86Unpckl,
4516 bc_v4i64>, VEX_4V, VEX_L;
4518 defm VPUNPCKHBW : sse2_unpack_y<0x68, "vpunpckhbw", v32i8, X86Unpckh,
4519 bc_v32i8>, VEX_4V, VEX_L;
4520 defm VPUNPCKHWD : sse2_unpack_y<0x69, "vpunpckhwd", v16i16, X86Unpckh,
4521 bc_v16i16>, VEX_4V, VEX_L;
4522 defm VPUNPCKHDQ : sse2_unpack_y<0x6A, "vpunpckhdq", v8i32, X86Unpckh,
4523 bc_v8i32>, VEX_4V, VEX_L;
4524 defm VPUNPCKHQDQ : sse2_unpack_y<0x6D, "vpunpckhqdq", v4i64, X86Unpckh,
4525 bc_v4i64>, VEX_4V, VEX_L;
4528 let Constraints = "$src1 = $dst" in {
4529 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, X86Unpckl,
4530 bc_v16i8, memopv2i64>;
4531 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, X86Unpckl,
4532 bc_v8i16, memopv2i64>;
4533 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, X86Unpckl,
4534 bc_v4i32, memopv2i64>;
4535 defm PUNPCKLQDQ : sse2_unpack<0x6C, "punpcklqdq", v2i64, X86Unpckl,
4536 bc_v2i64, memopv2i64>;
4538 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, X86Unpckh,
4539 bc_v16i8, memopv2i64>;
4540 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, X86Unpckh,
4541 bc_v8i16, memopv2i64>;
4542 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, X86Unpckh,
4543 bc_v4i32, memopv2i64>;
4544 defm PUNPCKHQDQ : sse2_unpack<0x6D, "punpckhqdq", v2i64, X86Unpckh,
4545 bc_v2i64, memopv2i64>;
4547 } // ExeDomain = SSEPackedInt
4549 //===---------------------------------------------------------------------===//
4550 // SSE2 - Packed Integer Extract and Insert
4551 //===---------------------------------------------------------------------===//
4553 let ExeDomain = SSEPackedInt in {
4554 multiclass sse2_pinsrw<bit Is2Addr = 1> {
4555 def rri : Ii8<0xC4, MRMSrcReg,
4556 (outs VR128:$dst), (ins VR128:$src1,
4557 GR32orGR64:$src2, u8imm:$src3),
4559 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
4560 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4562 (X86pinsrw VR128:$src1, GR32orGR64:$src2, imm:$src3))],
4563 IIC_SSE_PINSRW>, Sched<[WriteShuffle]>;
4564 def rmi : Ii8<0xC4, MRMSrcMem,
4565 (outs VR128:$dst), (ins VR128:$src1,
4566 i16mem:$src2, u8imm:$src3),
4568 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
4569 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4571 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
4572 imm:$src3))], IIC_SSE_PINSRW>,
4573 Sched<[WriteShuffleLd, ReadAfterLd]>;
4577 let Predicates = [HasAVX] in
4578 def VPEXTRWri : Ii8<0xC5, MRMSrcReg,
4579 (outs GR32orGR64:$dst), (ins VR128:$src1, u8imm:$src2),
4580 "vpextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4581 [(set GR32orGR64:$dst, (X86pextrw (v8i16 VR128:$src1),
4582 imm:$src2))]>, PD, VEX,
4583 Sched<[WriteShuffle]>;
4584 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
4585 (outs GR32orGR64:$dst), (ins VR128:$src1, u8imm:$src2),
4586 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
4587 [(set GR32orGR64:$dst, (X86pextrw (v8i16 VR128:$src1),
4588 imm:$src2))], IIC_SSE_PEXTRW>,
4589 Sched<[WriteShuffleLd, ReadAfterLd]>;
4592 let Predicates = [HasAVX] in
4593 defm VPINSRW : sse2_pinsrw<0>, PD, VEX_4V;
4595 let Predicates = [UseSSE2], Constraints = "$src1 = $dst" in
4596 defm PINSRW : sse2_pinsrw, PD;
4598 } // ExeDomain = SSEPackedInt
4600 //===---------------------------------------------------------------------===//
4601 // SSE2 - Packed Mask Creation
4602 //===---------------------------------------------------------------------===//
4604 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecLogic] in {
4606 def VPMOVMSKBrr : VPDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst),
4608 "pmovmskb\t{$src, $dst|$dst, $src}",
4609 [(set GR32orGR64:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))],
4610 IIC_SSE_MOVMSK>, VEX;
4612 let Predicates = [HasAVX2] in {
4613 def VPMOVMSKBYrr : VPDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst),
4615 "pmovmskb\t{$src, $dst|$dst, $src}",
4616 [(set GR32orGR64:$dst, (int_x86_avx2_pmovmskb VR256:$src))]>,
4620 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst), (ins VR128:$src),
4621 "pmovmskb\t{$src, $dst|$dst, $src}",
4622 [(set GR32orGR64:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))],
4625 } // ExeDomain = SSEPackedInt
4627 //===---------------------------------------------------------------------===//
4628 // SSE2 - Conditional Store
4629 //===---------------------------------------------------------------------===//
4631 let ExeDomain = SSEPackedInt, SchedRW = [WriteStore] in {
4633 let Uses = [EDI], Predicates = [HasAVX,Not64BitMode] in
4634 def VMASKMOVDQU : VPDI<0xF7, MRMSrcReg, (outs),
4635 (ins VR128:$src, VR128:$mask),
4636 "maskmovdqu\t{$mask, $src|$src, $mask}",
4637 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)],
4638 IIC_SSE_MASKMOV>, VEX;
4639 let Uses = [RDI], Predicates = [HasAVX,In64BitMode] in
4640 def VMASKMOVDQU64 : VPDI<0xF7, MRMSrcReg, (outs),
4641 (ins VR128:$src, VR128:$mask),
4642 "maskmovdqu\t{$mask, $src|$src, $mask}",
4643 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)],
4644 IIC_SSE_MASKMOV>, VEX;
4646 let Uses = [EDI], Predicates = [UseSSE2,Not64BitMode] in
4647 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
4648 "maskmovdqu\t{$mask, $src|$src, $mask}",
4649 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)],
4651 let Uses = [RDI], Predicates = [UseSSE2,In64BitMode] in
4652 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
4653 "maskmovdqu\t{$mask, $src|$src, $mask}",
4654 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)],
4657 } // ExeDomain = SSEPackedInt
4659 //===---------------------------------------------------------------------===//
4660 // SSE2 - Move Doubleword
4661 //===---------------------------------------------------------------------===//
4663 //===---------------------------------------------------------------------===//
4664 // Move Int Doubleword to Packed Double Int
4666 def VMOVDI2PDIrr : VS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
4667 "movd\t{$src, $dst|$dst, $src}",
4669 (v4i32 (scalar_to_vector GR32:$src)))], IIC_SSE_MOVDQ>,
4670 VEX, Sched<[WriteMove]>;
4671 def VMOVDI2PDIrm : VS2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
4672 "movd\t{$src, $dst|$dst, $src}",
4674 (v4i32 (scalar_to_vector (loadi32 addr:$src))))],
4676 VEX, Sched<[WriteLoad]>;
4677 def VMOV64toPQIrr : VRS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4678 "movq\t{$src, $dst|$dst, $src}",
4680 (v2i64 (scalar_to_vector GR64:$src)))],
4681 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4682 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0, mayLoad = 1 in
4683 def VMOV64toPQIrm : VRS2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4684 "movq\t{$src, $dst|$dst, $src}",
4685 [], IIC_SSE_MOVDQ>, VEX, Sched<[WriteLoad]>;
4686 let isCodeGenOnly = 1 in
4687 def VMOV64toSDrr : VRS2I<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
4688 "movq\t{$src, $dst|$dst, $src}",
4689 [(set FR64:$dst, (bitconvert GR64:$src))],
4690 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4692 def MOVDI2PDIrr : S2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
4693 "movd\t{$src, $dst|$dst, $src}",
4695 (v4i32 (scalar_to_vector GR32:$src)))], IIC_SSE_MOVDQ>,
4697 def MOVDI2PDIrm : S2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
4698 "movd\t{$src, $dst|$dst, $src}",
4700 (v4i32 (scalar_to_vector (loadi32 addr:$src))))],
4701 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4702 def MOV64toPQIrr : RS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4703 "mov{d|q}\t{$src, $dst|$dst, $src}",
4705 (v2i64 (scalar_to_vector GR64:$src)))],
4706 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4707 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0, mayLoad = 1 in
4708 def MOV64toPQIrm : RS2I<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4709 "mov{d|q}\t{$src, $dst|$dst, $src}",
4710 [], IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4711 let isCodeGenOnly = 1 in
4712 def MOV64toSDrr : RS2I<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
4713 "mov{d|q}\t{$src, $dst|$dst, $src}",
4714 [(set FR64:$dst, (bitconvert GR64:$src))],
4715 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4717 //===---------------------------------------------------------------------===//
4718 // Move Int Doubleword to Single Scalar
4720 let isCodeGenOnly = 1 in {
4721 def VMOVDI2SSrr : VS2I<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
4722 "movd\t{$src, $dst|$dst, $src}",
4723 [(set FR32:$dst, (bitconvert GR32:$src))],
4724 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4726 def VMOVDI2SSrm : VS2I<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
4727 "movd\t{$src, $dst|$dst, $src}",
4728 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))],
4730 VEX, Sched<[WriteLoad]>;
4731 def MOVDI2SSrr : S2I<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
4732 "movd\t{$src, $dst|$dst, $src}",
4733 [(set FR32:$dst, (bitconvert GR32:$src))],
4734 IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
4736 def MOVDI2SSrm : S2I<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
4737 "movd\t{$src, $dst|$dst, $src}",
4738 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))],
4739 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4742 //===---------------------------------------------------------------------===//
4743 // Move Packed Doubleword Int to Packed Double Int
4745 def VMOVPDI2DIrr : VS2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
4746 "movd\t{$src, $dst|$dst, $src}",
4747 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
4748 (iPTR 0)))], IIC_SSE_MOVD_ToGP>, VEX,
4750 def VMOVPDI2DImr : VS2I<0x7E, MRMDestMem, (outs),
4751 (ins i32mem:$dst, VR128:$src),
4752 "movd\t{$src, $dst|$dst, $src}",
4753 [(store (i32 (vector_extract (v4i32 VR128:$src),
4754 (iPTR 0))), addr:$dst)], IIC_SSE_MOVDQ>,
4755 VEX, Sched<[WriteStore]>;
4756 def MOVPDI2DIrr : S2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
4757 "movd\t{$src, $dst|$dst, $src}",
4758 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
4759 (iPTR 0)))], IIC_SSE_MOVD_ToGP>,
4761 def MOVPDI2DImr : S2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
4762 "movd\t{$src, $dst|$dst, $src}",
4763 [(store (i32 (vector_extract (v4i32 VR128:$src),
4764 (iPTR 0))), addr:$dst)],
4765 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4767 def : Pat<(v8i32 (X86Vinsert (v8i32 immAllZerosV), GR32:$src2, (iPTR 0))),
4768 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrr GR32:$src2), sub_xmm)>;
4770 def : Pat<(v4i64 (X86Vinsert (bc_v4i64 (v8i32 immAllZerosV)), GR64:$src2, (iPTR 0))),
4771 (SUBREG_TO_REG (i32 0), (VMOV64toPQIrr GR64:$src2), sub_xmm)>;
4773 def : Pat<(v8i32 (X86Vinsert undef, GR32:$src2, (iPTR 0))),
4774 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrr GR32:$src2), sub_xmm)>;
4776 def : Pat<(v4i64 (X86Vinsert undef, GR64:$src2, (iPTR 0))),
4777 (SUBREG_TO_REG (i32 0), (VMOV64toPQIrr GR64:$src2), sub_xmm)>;
4779 //===---------------------------------------------------------------------===//
4780 // Move Packed Doubleword Int first element to Doubleword Int
4782 let SchedRW = [WriteMove] in {
4783 def VMOVPQIto64rr : VRS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
4784 "movq\t{$src, $dst|$dst, $src}",
4785 [(set GR64:$dst, (vector_extract (v2i64 VR128:$src),
4790 def MOVPQIto64rr : RS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
4791 "mov{d|q}\t{$src, $dst|$dst, $src}",
4792 [(set GR64:$dst, (vector_extract (v2i64 VR128:$src),
4797 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0, mayStore = 1 in
4798 def VMOVPQIto64rm : VRS2I<0x7E, MRMDestMem, (outs i64mem:$dst),
4799 (ins VR128:$src), "movq\t{$src, $dst|$dst, $src}",
4800 [], IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4801 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0, mayStore = 1 in
4802 def MOVPQIto64rm : RS2I<0x7E, MRMDestMem, (outs i64mem:$dst), (ins VR128:$src),
4803 "mov{d|q}\t{$src, $dst|$dst, $src}",
4804 [], IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4806 //===---------------------------------------------------------------------===//
4807 // Bitcast FR64 <-> GR64
4809 let isCodeGenOnly = 1 in {
4810 let Predicates = [UseAVX] in
4811 def VMOV64toSDrm : VS2SI<0x7E, MRMSrcMem, (outs FR64:$dst), (ins i64mem:$src),
4812 "movq\t{$src, $dst|$dst, $src}",
4813 [(set FR64:$dst, (bitconvert (loadi64 addr:$src)))]>,
4814 VEX, Sched<[WriteLoad]>;
4815 def VMOVSDto64rr : VRS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins FR64:$src),
4816 "movq\t{$src, $dst|$dst, $src}",
4817 [(set GR64:$dst, (bitconvert FR64:$src))],
4818 IIC_SSE_MOVDQ>, VEX, Sched<[WriteMove]>;
4819 def VMOVSDto64mr : VRS2I<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, FR64:$src),
4820 "movq\t{$src, $dst|$dst, $src}",
4821 [(store (i64 (bitconvert FR64:$src)), addr:$dst)],
4822 IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4824 def MOV64toSDrm : S2SI<0x7E, MRMSrcMem, (outs FR64:$dst), (ins i64mem:$src),
4825 "movq\t{$src, $dst|$dst, $src}",
4826 [(set FR64:$dst, (bitconvert (loadi64 addr:$src)))],
4827 IIC_SSE_MOVDQ>, Sched<[WriteLoad]>;
4828 def MOVSDto64rr : RS2I<0x7E, MRMDestReg, (outs GR64:$dst), (ins FR64:$src),
4829 "mov{d|q}\t{$src, $dst|$dst, $src}",
4830 [(set GR64:$dst, (bitconvert FR64:$src))],
4831 IIC_SSE_MOVD_ToGP>, Sched<[WriteMove]>;
4832 def MOVSDto64mr : RS2I<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, FR64:$src),
4833 "movq\t{$src, $dst|$dst, $src}",
4834 [(store (i64 (bitconvert FR64:$src)), addr:$dst)],
4835 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4838 //===---------------------------------------------------------------------===//
4839 // Move Scalar Single to Double Int
4841 let isCodeGenOnly = 1 in {
4842 def VMOVSS2DIrr : VS2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
4843 "movd\t{$src, $dst|$dst, $src}",
4844 [(set GR32:$dst, (bitconvert FR32:$src))],
4845 IIC_SSE_MOVD_ToGP>, VEX, Sched<[WriteMove]>;
4846 def VMOVSS2DImr : VS2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
4847 "movd\t{$src, $dst|$dst, $src}",
4848 [(store (i32 (bitconvert FR32:$src)), addr:$dst)],
4849 IIC_SSE_MOVDQ>, VEX, Sched<[WriteStore]>;
4850 def MOVSS2DIrr : S2I<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
4851 "movd\t{$src, $dst|$dst, $src}",
4852 [(set GR32:$dst, (bitconvert FR32:$src))],
4853 IIC_SSE_MOVD_ToGP>, Sched<[WriteMove]>;
4854 def MOVSS2DImr : S2I<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
4855 "movd\t{$src, $dst|$dst, $src}",
4856 [(store (i32 (bitconvert FR32:$src)), addr:$dst)],
4857 IIC_SSE_MOVDQ>, Sched<[WriteStore]>;
4860 //===---------------------------------------------------------------------===//
4861 // Patterns and instructions to describe movd/movq to XMM register zero-extends
4863 let isCodeGenOnly = 1, SchedRW = [WriteMove] in {
4864 let AddedComplexity = 15 in {
4865 def VMOVZQI2PQIrr : VS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4866 "movq\t{$src, $dst|$dst, $src}", // X86-64 only
4867 [(set VR128:$dst, (v2i64 (X86vzmovl
4868 (v2i64 (scalar_to_vector GR64:$src)))))],
4871 def MOVZQI2PQIrr : RS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
4872 "mov{d|q}\t{$src, $dst|$dst, $src}", // X86-64 only
4873 [(set VR128:$dst, (v2i64 (X86vzmovl
4874 (v2i64 (scalar_to_vector GR64:$src)))))],
4877 } // isCodeGenOnly, SchedRW
4879 let Predicates = [UseAVX] in {
4880 let AddedComplexity = 15 in
4881 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector GR32:$src)))),
4882 (VMOVDI2PDIrr GR32:$src)>;
4884 // AVX 128-bit movd/movq instructions write zeros in the high 128-bit part.
4885 // These instructions also write zeros in the high part of a 256-bit register.
4886 let AddedComplexity = 20 in {
4887 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
4888 (VMOVDI2PDIrm addr:$src)>;
4889 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
4890 (VMOVDI2PDIrm addr:$src)>;
4891 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
4892 (VMOVDI2PDIrm addr:$src)>;
4893 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
4894 (v4i32 (scalar_to_vector (loadi32 addr:$src))), (iPTR 0)))),
4895 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrm addr:$src), sub_xmm)>;
4897 // Use regular 128-bit instructions to match 256-bit scalar_to_vec+zext.
4898 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
4899 (v4i32 (scalar_to_vector GR32:$src)),(iPTR 0)))),
4900 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrr GR32:$src), sub_xmm)>;
4901 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
4902 (v2i64 (scalar_to_vector GR64:$src)),(iPTR 0)))),
4903 (SUBREG_TO_REG (i64 0), (VMOVZQI2PQIrr GR64:$src), sub_xmm)>;
4906 let Predicates = [UseSSE2] in {
4907 let AddedComplexity = 15 in
4908 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector GR32:$src)))),
4909 (MOVDI2PDIrr GR32:$src)>;
4911 let AddedComplexity = 20 in {
4912 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
4913 (MOVDI2PDIrm addr:$src)>;
4914 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
4915 (MOVDI2PDIrm addr:$src)>;
4916 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
4917 (MOVDI2PDIrm addr:$src)>;
4921 // These are the correct encodings of the instructions so that we know how to
4922 // read correct assembly, even though we continue to emit the wrong ones for
4923 // compatibility with Darwin's buggy assembler.
4924 def : InstAlias<"movq\t{$src, $dst|$dst, $src}",
4925 (MOV64toPQIrr VR128:$dst, GR64:$src), 0>;
4926 def : InstAlias<"movq\t{$src, $dst|$dst, $src}",
4927 (MOVPQIto64rr GR64:$dst, VR128:$src), 0>;
4928 // Allow "vmovd" but print "vmovq" since we don't need compatibility for AVX.
4929 def : InstAlias<"vmovd\t{$src, $dst|$dst, $src}",
4930 (VMOV64toPQIrr VR128:$dst, GR64:$src), 0>;
4931 def : InstAlias<"vmovd\t{$src, $dst|$dst, $src}",
4932 (VMOVPQIto64rr GR64:$dst, VR128:$src), 0>;
4934 //===---------------------------------------------------------------------===//
4935 // SSE2 - Move Quadword
4936 //===---------------------------------------------------------------------===//
4938 //===---------------------------------------------------------------------===//
4939 // Move Quadword Int to Packed Quadword Int
4942 let ExeDomain = SSEPackedInt, SchedRW = [WriteLoad] in {
4943 def VMOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4944 "vmovq\t{$src, $dst|$dst, $src}",
4946 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
4947 VEX, Requires<[UseAVX]>;
4948 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4949 "movq\t{$src, $dst|$dst, $src}",
4951 (v2i64 (scalar_to_vector (loadi64 addr:$src))))],
4953 Requires<[UseSSE2]>; // SSE2 instruction with XS Prefix
4954 } // ExeDomain, SchedRW
4956 //===---------------------------------------------------------------------===//
4957 // Move Packed Quadword Int to Quadword Int
4959 let ExeDomain = SSEPackedInt, SchedRW = [WriteStore] in {
4960 def VMOVPQI2QImr : VS2I<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
4961 "movq\t{$src, $dst|$dst, $src}",
4962 [(store (i64 (vector_extract (v2i64 VR128:$src),
4963 (iPTR 0))), addr:$dst)],
4964 IIC_SSE_MOVDQ>, VEX;
4965 def MOVPQI2QImr : S2I<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
4966 "movq\t{$src, $dst|$dst, $src}",
4967 [(store (i64 (vector_extract (v2i64 VR128:$src),
4968 (iPTR 0))), addr:$dst)],
4970 } // ExeDomain, SchedRW
4972 // For disassembler only
4973 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0,
4974 SchedRW = [WriteVecLogic] in {
4975 def VMOVPQI2QIrr : VS2I<0xD6, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
4976 "movq\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVQ_RR>, VEX;
4977 def MOVPQI2QIrr : S2I<0xD6, MRMDestReg, (outs VR128:$dst), (ins VR128:$src),
4978 "movq\t{$src, $dst|$dst, $src}", [], IIC_SSE_MOVQ_RR>;
4981 //===---------------------------------------------------------------------===//
4982 // Store / copy lower 64-bits of a XMM register.
4984 let Predicates = [UseAVX] in
4985 def : Pat<(int_x86_sse2_storel_dq addr:$dst, VR128:$src),
4986 (VMOVPQI2QImr addr:$dst, VR128:$src)>;
4987 let Predicates = [UseSSE2] in
4988 def : Pat<(int_x86_sse2_storel_dq addr:$dst, VR128:$src),
4989 (MOVPQI2QImr addr:$dst, VR128:$src)>;
4991 let ExeDomain = SSEPackedInt, isCodeGenOnly = 1, AddedComplexity = 20 in {
4992 def VMOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4993 "vmovq\t{$src, $dst|$dst, $src}",
4995 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
4996 (loadi64 addr:$src))))))],
4998 XS, VEX, Requires<[UseAVX]>, Sched<[WriteLoad]>;
5000 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
5001 "movq\t{$src, $dst|$dst, $src}",
5003 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
5004 (loadi64 addr:$src))))))],
5006 XS, Requires<[UseSSE2]>, Sched<[WriteLoad]>;
5007 } // ExeDomain, isCodeGenOnly, AddedComplexity
5009 let Predicates = [UseAVX], AddedComplexity = 20 in {
5010 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
5011 (VMOVZQI2PQIrm addr:$src)>;
5012 def : Pat<(v2i64 (X86vzload addr:$src)),
5013 (VMOVZQI2PQIrm addr:$src)>;
5014 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
5015 (v2i64 (scalar_to_vector (loadi64 addr:$src))), (iPTR 0)))),
5016 (SUBREG_TO_REG (i64 0), (VMOVZQI2PQIrm addr:$src), sub_xmm)>;
5019 let Predicates = [UseSSE2], AddedComplexity = 20 in {
5020 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
5021 (MOVZQI2PQIrm addr:$src)>;
5022 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
5025 let Predicates = [HasAVX] in {
5026 def : Pat<(v4i64 (alignedX86vzload addr:$src)),
5027 (SUBREG_TO_REG (i32 0), (VMOVAPSrm addr:$src), sub_xmm)>;
5028 def : Pat<(v4i64 (X86vzload addr:$src)),
5029 (SUBREG_TO_REG (i32 0), (VMOVUPSrm addr:$src), sub_xmm)>;
5032 //===---------------------------------------------------------------------===//
5033 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
5034 // IA32 document. movq xmm1, xmm2 does clear the high bits.
5036 let ExeDomain = SSEPackedInt, SchedRW = [WriteVecLogic] in {
5037 let AddedComplexity = 15 in
5038 def VMOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5039 "vmovq\t{$src, $dst|$dst, $src}",
5040 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))],
5042 XS, VEX, Requires<[UseAVX]>;
5043 let AddedComplexity = 15 in
5044 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5045 "movq\t{$src, $dst|$dst, $src}",
5046 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))],
5048 XS, Requires<[UseSSE2]>;
5049 } // ExeDomain, SchedRW
5051 let ExeDomain = SSEPackedInt, isCodeGenOnly = 1, SchedRW = [WriteVecLogicLd] in {
5052 let AddedComplexity = 20 in
5053 def VMOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5054 "vmovq\t{$src, $dst|$dst, $src}",
5055 [(set VR128:$dst, (v2i64 (X86vzmovl
5056 (loadv2i64 addr:$src))))],
5058 XS, VEX, Requires<[UseAVX]>;
5059 let AddedComplexity = 20 in {
5060 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5061 "movq\t{$src, $dst|$dst, $src}",
5062 [(set VR128:$dst, (v2i64 (X86vzmovl
5063 (loadv2i64 addr:$src))))],
5065 XS, Requires<[UseSSE2]>;
5067 } // ExeDomain, isCodeGenOnly, SchedRW
5069 let AddedComplexity = 20 in {
5070 let Predicates = [UseAVX] in {
5071 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
5072 (VMOVZPQILo2PQIrr VR128:$src)>;
5074 let Predicates = [UseSSE2] in {
5075 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
5076 (MOVZPQILo2PQIrr VR128:$src)>;
5080 //===---------------------------------------------------------------------===//
5081 // SSE3 - Replicate Single FP - MOVSHDUP and MOVSLDUP
5082 //===---------------------------------------------------------------------===//
5083 multiclass sse3_replicate_sfp<bits<8> op, SDNode OpNode, string OpcodeStr,
5084 ValueType vt, RegisterClass RC, PatFrag mem_frag,
5085 X86MemOperand x86memop> {
5086 def rr : S3SI<op, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
5087 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5088 [(set RC:$dst, (vt (OpNode RC:$src)))],
5089 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
5090 def rm : S3SI<op, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
5091 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5092 [(set RC:$dst, (OpNode (mem_frag addr:$src)))],
5093 IIC_SSE_MOV_LH>, Sched<[WriteLoad]>;
5096 let Predicates = [HasAVX] in {
5097 defm VMOVSHDUP : sse3_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
5098 v4f32, VR128, loadv4f32, f128mem>, VEX;
5099 defm VMOVSLDUP : sse3_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
5100 v4f32, VR128, loadv4f32, f128mem>, VEX;
5101 defm VMOVSHDUPY : sse3_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
5102 v8f32, VR256, loadv8f32, f256mem>, VEX, VEX_L;
5103 defm VMOVSLDUPY : sse3_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
5104 v8f32, VR256, loadv8f32, f256mem>, VEX, VEX_L;
5106 defm MOVSHDUP : sse3_replicate_sfp<0x16, X86Movshdup, "movshdup", v4f32, VR128,
5107 memopv4f32, f128mem>;
5108 defm MOVSLDUP : sse3_replicate_sfp<0x12, X86Movsldup, "movsldup", v4f32, VR128,
5109 memopv4f32, f128mem>;
5111 let Predicates = [HasAVX] in {
5112 def : Pat<(v4i32 (X86Movshdup VR128:$src)),
5113 (VMOVSHDUPrr VR128:$src)>;
5114 def : Pat<(v4i32 (X86Movshdup (bc_v4i32 (loadv2i64 addr:$src)))),
5115 (VMOVSHDUPrm addr:$src)>;
5116 def : Pat<(v4i32 (X86Movsldup VR128:$src)),
5117 (VMOVSLDUPrr VR128:$src)>;
5118 def : Pat<(v4i32 (X86Movsldup (bc_v4i32 (loadv2i64 addr:$src)))),
5119 (VMOVSLDUPrm addr:$src)>;
5120 def : Pat<(v8i32 (X86Movshdup VR256:$src)),
5121 (VMOVSHDUPYrr VR256:$src)>;
5122 def : Pat<(v8i32 (X86Movshdup (bc_v8i32 (loadv4i64 addr:$src)))),
5123 (VMOVSHDUPYrm addr:$src)>;
5124 def : Pat<(v8i32 (X86Movsldup VR256:$src)),
5125 (VMOVSLDUPYrr VR256:$src)>;
5126 def : Pat<(v8i32 (X86Movsldup (bc_v8i32 (loadv4i64 addr:$src)))),
5127 (VMOVSLDUPYrm addr:$src)>;
5130 let Predicates = [UseSSE3] in {
5131 def : Pat<(v4i32 (X86Movshdup VR128:$src)),
5132 (MOVSHDUPrr VR128:$src)>;
5133 def : Pat<(v4i32 (X86Movshdup (bc_v4i32 (memopv2i64 addr:$src)))),
5134 (MOVSHDUPrm addr:$src)>;
5135 def : Pat<(v4i32 (X86Movsldup VR128:$src)),
5136 (MOVSLDUPrr VR128:$src)>;
5137 def : Pat<(v4i32 (X86Movsldup (bc_v4i32 (memopv2i64 addr:$src)))),
5138 (MOVSLDUPrm addr:$src)>;
5141 //===---------------------------------------------------------------------===//
5142 // SSE3 - Replicate Double FP - MOVDDUP
5143 //===---------------------------------------------------------------------===//
5145 multiclass sse3_replicate_dfp<string OpcodeStr> {
5146 def rr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
5147 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5148 [(set VR128:$dst, (v2f64 (X86Movddup VR128:$src)))],
5149 IIC_SSE_MOV_LH>, Sched<[WriteFShuffle]>;
5150 def rm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
5151 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5154 (scalar_to_vector (loadf64 addr:$src)))))],
5155 IIC_SSE_MOV_LH>, Sched<[WriteLoad]>;
5158 // FIXME: Merge with above classe when there're patterns for the ymm version
5159 multiclass sse3_replicate_dfp_y<string OpcodeStr> {
5160 def rr : S3DI<0x12, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
5161 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5162 [(set VR256:$dst, (v4f64 (X86Movddup VR256:$src)))]>,
5163 Sched<[WriteFShuffle]>;
5164 def rm : S3DI<0x12, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
5165 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5168 (scalar_to_vector (loadf64 addr:$src)))))]>,
5172 let Predicates = [HasAVX] in {
5173 defm VMOVDDUP : sse3_replicate_dfp<"vmovddup">, VEX;
5174 defm VMOVDDUPY : sse3_replicate_dfp_y<"vmovddup">, VEX, VEX_L;
5177 defm MOVDDUP : sse3_replicate_dfp<"movddup">;
5179 let Predicates = [HasAVX] in {
5180 def : Pat<(X86Movddup (loadv2f64 addr:$src)),
5181 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5182 def : Pat<(X86Movddup (bc_v2f64 (loadv4f32 addr:$src))),
5183 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5184 def : Pat<(X86Movddup (bc_v2f64 (loadv2i64 addr:$src))),
5185 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5186 def : Pat<(X86Movddup (bc_v2f64
5187 (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
5188 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5191 def : Pat<(X86Movddup (loadv4f64 addr:$src)),
5192 (VMOVDDUPYrm addr:$src)>;
5193 def : Pat<(X86Movddup (loadv4i64 addr:$src)),
5194 (VMOVDDUPYrm addr:$src)>;
5195 def : Pat<(X86Movddup (v4i64 (scalar_to_vector (loadi64 addr:$src)))),
5196 (VMOVDDUPYrm addr:$src)>;
5197 def : Pat<(X86Movddup (v4i64 VR256:$src)),
5198 (VMOVDDUPYrr VR256:$src)>;
5201 let Predicates = [UseAVX, OptForSize] in {
5202 def : Pat<(v2f64 (X86VBroadcast (loadf64 addr:$src))),
5203 (VMOVDDUPrm addr:$src)>;
5204 def : Pat<(v2i64 (X86VBroadcast (loadi64 addr:$src))),
5205 (VMOVDDUPrm addr:$src)>;
5208 let Predicates = [UseSSE3] in {
5209 def : Pat<(X86Movddup (memopv2f64 addr:$src)),
5210 (MOVDDUPrm addr:$src)>;
5211 def : Pat<(X86Movddup (bc_v2f64 (memopv4f32 addr:$src))),
5212 (MOVDDUPrm addr:$src)>;
5213 def : Pat<(X86Movddup (bc_v2f64 (memopv2i64 addr:$src))),
5214 (MOVDDUPrm addr:$src)>;
5215 def : Pat<(X86Movddup (bc_v2f64
5216 (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
5217 (MOVDDUPrm addr:$src)>;
5220 //===---------------------------------------------------------------------===//
5221 // SSE3 - Move Unaligned Integer
5222 //===---------------------------------------------------------------------===//
5224 let SchedRW = [WriteLoad] in {
5225 let Predicates = [HasAVX] in {
5226 def VLDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5227 "vlddqu\t{$src, $dst|$dst, $src}",
5228 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>, VEX;
5229 def VLDDQUYrm : S3DI<0xF0, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
5230 "vlddqu\t{$src, $dst|$dst, $src}",
5231 [(set VR256:$dst, (int_x86_avx_ldu_dq_256 addr:$src))]>,
5234 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5235 "lddqu\t{$src, $dst|$dst, $src}",
5236 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))],
5240 //===---------------------------------------------------------------------===//
5241 // SSE3 - Arithmetic
5242 //===---------------------------------------------------------------------===//
5244 multiclass sse3_addsub<Intrinsic Int, string OpcodeStr, RegisterClass RC,
5245 X86MemOperand x86memop, OpndItins itins,
5246 PatFrag ld_frag, bit Is2Addr = 1> {
5247 def rr : I<0xD0, MRMSrcReg,
5248 (outs RC:$dst), (ins RC:$src1, RC:$src2),
5250 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5251 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5252 [(set RC:$dst, (Int RC:$src1, RC:$src2))], itins.rr>,
5253 Sched<[itins.Sched]>;
5254 def rm : I<0xD0, MRMSrcMem,
5255 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5257 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5258 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5259 [(set RC:$dst, (Int RC:$src1, (ld_frag addr:$src2)))], itins.rr>,
5260 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5263 let Predicates = [HasAVX] in {
5264 let ExeDomain = SSEPackedSingle in {
5265 defm VADDSUBPS : sse3_addsub<int_x86_sse3_addsub_ps, "vaddsubps", VR128,
5266 f128mem, SSE_ALU_F32P, loadv4f32, 0>, XD, VEX_4V;
5267 defm VADDSUBPSY : sse3_addsub<int_x86_avx_addsub_ps_256, "vaddsubps", VR256,
5268 f256mem, SSE_ALU_F32P, loadv8f32, 0>, XD, VEX_4V, VEX_L;
5270 let ExeDomain = SSEPackedDouble in {
5271 defm VADDSUBPD : sse3_addsub<int_x86_sse3_addsub_pd, "vaddsubpd", VR128,
5272 f128mem, SSE_ALU_F64P, loadv2f64, 0>, PD, VEX_4V;
5273 defm VADDSUBPDY : sse3_addsub<int_x86_avx_addsub_pd_256, "vaddsubpd", VR256,
5274 f256mem, SSE_ALU_F64P, loadv4f64, 0>, PD, VEX_4V, VEX_L;
5277 let Constraints = "$src1 = $dst", Predicates = [UseSSE3] in {
5278 let ExeDomain = SSEPackedSingle in
5279 defm ADDSUBPS : sse3_addsub<int_x86_sse3_addsub_ps, "addsubps", VR128,
5280 f128mem, SSE_ALU_F32P, memopv4f32>, XD;
5281 let ExeDomain = SSEPackedDouble in
5282 defm ADDSUBPD : sse3_addsub<int_x86_sse3_addsub_pd, "addsubpd", VR128,
5283 f128mem, SSE_ALU_F64P, memopv2f64>, PD;
5286 // Patterns used to select 'addsub' instructions.
5287 let Predicates = [HasAVX] in {
5288 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 VR128:$rhs))),
5289 (VADDSUBPSrr VR128:$lhs, VR128:$rhs)>;
5290 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (loadv4f32 addr:$rhs))),
5291 (VADDSUBPSrm VR128:$lhs, f128mem:$rhs)>;
5292 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 VR128:$rhs))),
5293 (VADDSUBPDrr VR128:$lhs, VR128:$rhs)>;
5294 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (loadv2f64 addr:$rhs))),
5295 (VADDSUBPDrm VR128:$lhs, f128mem:$rhs)>;
5297 def : Pat<(v8f32 (X86Addsub (v8f32 VR256:$lhs), (v8f32 VR256:$rhs))),
5298 (VADDSUBPSYrr VR256:$lhs, VR256:$rhs)>;
5299 def : Pat<(v8f32 (X86Addsub (v8f32 VR256:$lhs), (loadv8f32 addr:$rhs))),
5300 (VADDSUBPSYrm VR256:$lhs, f256mem:$rhs)>;
5301 def : Pat<(v4f64 (X86Addsub (v4f64 VR256:$lhs), (v4f64 VR256:$rhs))),
5302 (VADDSUBPDYrr VR256:$lhs, VR256:$rhs)>;
5303 def : Pat<(v4f64 (X86Addsub (v4f64 VR256:$lhs), (loadv4f64 addr:$rhs))),
5304 (VADDSUBPDYrm VR256:$lhs, f256mem:$rhs)>;
5307 let Predicates = [UseSSE3] in {
5308 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (v4f32 VR128:$rhs))),
5309 (ADDSUBPSrr VR128:$lhs, VR128:$rhs)>;
5310 def : Pat<(v4f32 (X86Addsub (v4f32 VR128:$lhs), (memopv4f32 addr:$rhs))),
5311 (ADDSUBPSrm VR128:$lhs, f128mem:$rhs)>;
5312 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (v2f64 VR128:$rhs))),
5313 (ADDSUBPDrr VR128:$lhs, VR128:$rhs)>;
5314 def : Pat<(v2f64 (X86Addsub (v2f64 VR128:$lhs), (memopv2f64 addr:$rhs))),
5315 (ADDSUBPDrm VR128:$lhs, f128mem:$rhs)>;
5318 //===---------------------------------------------------------------------===//
5319 // SSE3 Instructions
5320 //===---------------------------------------------------------------------===//
5323 multiclass S3D_Int<bits<8> o, string OpcodeStr, ValueType vt, RegisterClass RC,
5324 X86MemOperand x86memop, SDNode OpNode, PatFrag ld_frag,
5326 def rr : S3DI<o, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
5328 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5329 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5330 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], IIC_SSE_HADDSUB_RR>,
5333 def rm : S3DI<o, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5335 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5336 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5337 [(set RC:$dst, (vt (OpNode RC:$src1, (ld_frag addr:$src2))))],
5338 IIC_SSE_HADDSUB_RM>, Sched<[WriteFAddLd, ReadAfterLd]>;
5340 multiclass S3_Int<bits<8> o, string OpcodeStr, ValueType vt, RegisterClass RC,
5341 X86MemOperand x86memop, SDNode OpNode, PatFrag ld_frag,
5343 def rr : S3I<o, MRMSrcReg, (outs RC:$dst), (ins RC:$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, (vt (OpNode RC:$src1, RC:$src2)))], IIC_SSE_HADDSUB_RR>,
5350 def rm : S3I<o, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
5352 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5353 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5354 [(set RC:$dst, (vt (OpNode RC:$src1, (ld_frag addr:$src2))))],
5355 IIC_SSE_HADDSUB_RM>, Sched<[WriteFAddLd, ReadAfterLd]>;
5358 let Predicates = [HasAVX] in {
5359 let ExeDomain = SSEPackedSingle in {
5360 defm VHADDPS : S3D_Int<0x7C, "vhaddps", v4f32, VR128, f128mem,
5361 X86fhadd, loadv4f32, 0>, VEX_4V;
5362 defm VHSUBPS : S3D_Int<0x7D, "vhsubps", v4f32, VR128, f128mem,
5363 X86fhsub, loadv4f32, 0>, VEX_4V;
5364 defm VHADDPSY : S3D_Int<0x7C, "vhaddps", v8f32, VR256, f256mem,
5365 X86fhadd, loadv8f32, 0>, VEX_4V, VEX_L;
5366 defm VHSUBPSY : S3D_Int<0x7D, "vhsubps", v8f32, VR256, f256mem,
5367 X86fhsub, loadv8f32, 0>, VEX_4V, VEX_L;
5369 let ExeDomain = SSEPackedDouble in {
5370 defm VHADDPD : S3_Int <0x7C, "vhaddpd", v2f64, VR128, f128mem,
5371 X86fhadd, loadv2f64, 0>, VEX_4V;
5372 defm VHSUBPD : S3_Int <0x7D, "vhsubpd", v2f64, VR128, f128mem,
5373 X86fhsub, loadv2f64, 0>, VEX_4V;
5374 defm VHADDPDY : S3_Int <0x7C, "vhaddpd", v4f64, VR256, f256mem,
5375 X86fhadd, loadv4f64, 0>, VEX_4V, VEX_L;
5376 defm VHSUBPDY : S3_Int <0x7D, "vhsubpd", v4f64, VR256, f256mem,
5377 X86fhsub, loadv4f64, 0>, VEX_4V, VEX_L;
5381 let Constraints = "$src1 = $dst" in {
5382 let ExeDomain = SSEPackedSingle in {
5383 defm HADDPS : S3D_Int<0x7C, "haddps", v4f32, VR128, f128mem, X86fhadd,
5385 defm HSUBPS : S3D_Int<0x7D, "hsubps", v4f32, VR128, f128mem, X86fhsub,
5388 let ExeDomain = SSEPackedDouble in {
5389 defm HADDPD : S3_Int<0x7C, "haddpd", v2f64, VR128, f128mem, X86fhadd,
5391 defm HSUBPD : S3_Int<0x7D, "hsubpd", v2f64, VR128, f128mem, X86fhsub,
5396 //===---------------------------------------------------------------------===//
5397 // SSSE3 - Packed Absolute Instructions
5398 //===---------------------------------------------------------------------===//
5401 /// SS3I_unop_rm_int - Simple SSSE3 unary op whose type can be v*{i8,i16,i32}.
5402 multiclass SS3I_unop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId128,
5404 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
5406 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5407 [(set VR128:$dst, (IntId128 VR128:$src))], IIC_SSE_PABS_RR>,
5408 Sched<[WriteVecALU]>;
5410 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
5412 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5415 (bitconvert (ld_frag addr:$src))))], IIC_SSE_PABS_RM>,
5416 Sched<[WriteVecALULd]>;
5419 /// SS3I_unop_rm_int_y - Simple SSSE3 unary op whose type can be v*{i8,i16,i32}.
5420 multiclass SS3I_unop_rm_int_y<bits<8> opc, string OpcodeStr,
5421 Intrinsic IntId256> {
5422 def rr256 : SS38I<opc, MRMSrcReg, (outs VR256:$dst),
5424 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5425 [(set VR256:$dst, (IntId256 VR256:$src))]>,
5426 Sched<[WriteVecALU]>;
5428 def rm256 : SS38I<opc, MRMSrcMem, (outs VR256:$dst),
5430 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5433 (bitconvert (loadv4i64 addr:$src))))]>,
5434 Sched<[WriteVecALULd]>;
5437 // Helper fragments to match sext vXi1 to vXiY.
5438 def v16i1sextv16i8 : PatLeaf<(v16i8 (X86pcmpgt (bc_v16i8 (v4i32 immAllZerosV)),
5440 def v8i1sextv8i16 : PatLeaf<(v8i16 (X86vsrai VR128:$src, (i8 15)))>;
5441 def v4i1sextv4i32 : PatLeaf<(v4i32 (X86vsrai VR128:$src, (i8 31)))>;
5442 def v32i1sextv32i8 : PatLeaf<(v32i8 (X86pcmpgt (bc_v32i8 (v8i32 immAllZerosV)),
5444 def v16i1sextv16i16: PatLeaf<(v16i16 (X86vsrai VR256:$src, (i8 15)))>;
5445 def v8i1sextv8i32 : PatLeaf<(v8i32 (X86vsrai VR256:$src, (i8 31)))>;
5447 let Predicates = [HasAVX] in {
5448 defm VPABSB : SS3I_unop_rm_int<0x1C, "vpabsb", int_x86_ssse3_pabs_b_128,
5450 defm VPABSW : SS3I_unop_rm_int<0x1D, "vpabsw", int_x86_ssse3_pabs_w_128,
5452 defm VPABSD : SS3I_unop_rm_int<0x1E, "vpabsd", int_x86_ssse3_pabs_d_128,
5456 (bc_v2i64 (v16i1sextv16i8)),
5457 (bc_v2i64 (add (v16i8 VR128:$src), (v16i1sextv16i8)))),
5458 (VPABSBrr128 VR128:$src)>;
5460 (bc_v2i64 (v8i1sextv8i16)),
5461 (bc_v2i64 (add (v8i16 VR128:$src), (v8i1sextv8i16)))),
5462 (VPABSWrr128 VR128:$src)>;
5464 (bc_v2i64 (v4i1sextv4i32)),
5465 (bc_v2i64 (add (v4i32 VR128:$src), (v4i1sextv4i32)))),
5466 (VPABSDrr128 VR128:$src)>;
5469 let Predicates = [HasAVX2] in {
5470 defm VPABSB : SS3I_unop_rm_int_y<0x1C, "vpabsb",
5471 int_x86_avx2_pabs_b>, VEX, VEX_L;
5472 defm VPABSW : SS3I_unop_rm_int_y<0x1D, "vpabsw",
5473 int_x86_avx2_pabs_w>, VEX, VEX_L;
5474 defm VPABSD : SS3I_unop_rm_int_y<0x1E, "vpabsd",
5475 int_x86_avx2_pabs_d>, VEX, VEX_L;
5478 (bc_v4i64 (v32i1sextv32i8)),
5479 (bc_v4i64 (add (v32i8 VR256:$src), (v32i1sextv32i8)))),
5480 (VPABSBrr256 VR256:$src)>;
5482 (bc_v4i64 (v16i1sextv16i16)),
5483 (bc_v4i64 (add (v16i16 VR256:$src), (v16i1sextv16i16)))),
5484 (VPABSWrr256 VR256:$src)>;
5486 (bc_v4i64 (v8i1sextv8i32)),
5487 (bc_v4i64 (add (v8i32 VR256:$src), (v8i1sextv8i32)))),
5488 (VPABSDrr256 VR256:$src)>;
5491 defm PABSB : SS3I_unop_rm_int<0x1C, "pabsb", int_x86_ssse3_pabs_b_128,
5493 defm PABSW : SS3I_unop_rm_int<0x1D, "pabsw", int_x86_ssse3_pabs_w_128,
5495 defm PABSD : SS3I_unop_rm_int<0x1E, "pabsd", int_x86_ssse3_pabs_d_128,
5498 let Predicates = [HasSSSE3] in {
5500 (bc_v2i64 (v16i1sextv16i8)),
5501 (bc_v2i64 (add (v16i8 VR128:$src), (v16i1sextv16i8)))),
5502 (PABSBrr128 VR128:$src)>;
5504 (bc_v2i64 (v8i1sextv8i16)),
5505 (bc_v2i64 (add (v8i16 VR128:$src), (v8i1sextv8i16)))),
5506 (PABSWrr128 VR128:$src)>;
5508 (bc_v2i64 (v4i1sextv4i32)),
5509 (bc_v2i64 (add (v4i32 VR128:$src), (v4i1sextv4i32)))),
5510 (PABSDrr128 VR128:$src)>;
5513 //===---------------------------------------------------------------------===//
5514 // SSSE3 - Packed Binary Operator Instructions
5515 //===---------------------------------------------------------------------===//
5517 let Sched = WriteVecALU in {
5518 def SSE_PHADDSUBD : OpndItins<
5519 IIC_SSE_PHADDSUBD_RR, IIC_SSE_PHADDSUBD_RM
5521 def SSE_PHADDSUBSW : OpndItins<
5522 IIC_SSE_PHADDSUBSW_RR, IIC_SSE_PHADDSUBSW_RM
5524 def SSE_PHADDSUBW : OpndItins<
5525 IIC_SSE_PHADDSUBW_RR, IIC_SSE_PHADDSUBW_RM
5528 let Sched = WriteShuffle in
5529 def SSE_PSHUFB : OpndItins<
5530 IIC_SSE_PSHUFB_RR, IIC_SSE_PSHUFB_RM
5532 let Sched = WriteVecALU in
5533 def SSE_PSIGN : OpndItins<
5534 IIC_SSE_PSIGN_RR, IIC_SSE_PSIGN_RM
5536 let Sched = WriteVecIMul in
5537 def SSE_PMULHRSW : OpndItins<
5538 IIC_SSE_PMULHRSW, IIC_SSE_PMULHRSW
5541 /// SS3I_binop_rm - Simple SSSE3 bin op
5542 multiclass SS3I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
5543 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
5544 X86MemOperand x86memop, OpndItins itins,
5546 let isCommutable = 1 in
5547 def rr : SS38I<opc, MRMSrcReg, (outs RC:$dst),
5548 (ins RC:$src1, RC:$src2),
5550 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5551 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5552 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))], itins.rr>,
5553 Sched<[itins.Sched]>;
5554 def rm : SS38I<opc, MRMSrcMem, (outs RC:$dst),
5555 (ins RC:$src1, x86memop:$src2),
5557 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5558 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5560 (OpVT (OpNode RC:$src1,
5561 (bitconvert (memop_frag addr:$src2)))))], itins.rm>,
5562 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5565 /// SS3I_binop_rm_int - Simple SSSE3 bin op whose type can be v*{i8,i16,i32}.
5566 multiclass SS3I_binop_rm_int<bits<8> opc, string OpcodeStr,
5567 Intrinsic IntId128, OpndItins itins,
5568 PatFrag ld_frag, bit Is2Addr = 1> {
5569 let isCommutable = 1 in
5570 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
5571 (ins VR128:$src1, VR128:$src2),
5573 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5574 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5575 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
5576 Sched<[itins.Sched]>;
5577 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
5578 (ins VR128:$src1, i128mem:$src2),
5580 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5581 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5583 (IntId128 VR128:$src1,
5584 (bitconvert (ld_frag addr:$src2))))]>,
5585 Sched<[itins.Sched.Folded, ReadAfterLd]>;
5588 multiclass SS3I_binop_rm_int_y<bits<8> opc, string OpcodeStr,
5590 X86FoldableSchedWrite Sched> {
5591 let isCommutable = 1 in
5592 def rr256 : SS38I<opc, MRMSrcReg, (outs VR256:$dst),
5593 (ins VR256:$src1, VR256:$src2),
5594 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5595 [(set VR256:$dst, (IntId256 VR256:$src1, VR256:$src2))]>,
5597 def rm256 : SS38I<opc, MRMSrcMem, (outs VR256:$dst),
5598 (ins VR256:$src1, i256mem:$src2),
5599 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5601 (IntId256 VR256:$src1, (bitconvert (loadv4i64 addr:$src2))))]>,
5602 Sched<[Sched.Folded, ReadAfterLd]>;
5605 let ImmT = NoImm, Predicates = [HasAVX] in {
5606 let isCommutable = 0 in {
5607 defm VPHADDW : SS3I_binop_rm<0x01, "vphaddw", X86hadd, v8i16, VR128,
5609 SSE_PHADDSUBW, 0>, VEX_4V;
5610 defm VPHADDD : SS3I_binop_rm<0x02, "vphaddd", X86hadd, v4i32, VR128,
5612 SSE_PHADDSUBD, 0>, VEX_4V;
5613 defm VPHSUBW : SS3I_binop_rm<0x05, "vphsubw", X86hsub, v8i16, VR128,
5615 SSE_PHADDSUBW, 0>, VEX_4V;
5616 defm VPHSUBD : SS3I_binop_rm<0x06, "vphsubd", X86hsub, v4i32, VR128,
5618 SSE_PHADDSUBD, 0>, VEX_4V;
5619 defm VPSIGNB : SS3I_binop_rm<0x08, "vpsignb", X86psign, v16i8, VR128,
5621 SSE_PSIGN, 0>, VEX_4V;
5622 defm VPSIGNW : SS3I_binop_rm<0x09, "vpsignw", X86psign, v8i16, VR128,
5624 SSE_PSIGN, 0>, VEX_4V;
5625 defm VPSIGND : SS3I_binop_rm<0x0A, "vpsignd", X86psign, v4i32, VR128,
5627 SSE_PSIGN, 0>, VEX_4V;
5628 defm VPSHUFB : SS3I_binop_rm<0x00, "vpshufb", X86pshufb, v16i8, VR128,
5630 SSE_PSHUFB, 0>, VEX_4V;
5631 defm VPHADDSW : SS3I_binop_rm_int<0x03, "vphaddsw",
5632 int_x86_ssse3_phadd_sw_128,
5633 SSE_PHADDSUBSW, loadv2i64, 0>, VEX_4V;
5634 defm VPHSUBSW : SS3I_binop_rm_int<0x07, "vphsubsw",
5635 int_x86_ssse3_phsub_sw_128,
5636 SSE_PHADDSUBSW, loadv2i64, 0>, VEX_4V;
5637 defm VPMADDUBSW : SS3I_binop_rm_int<0x04, "vpmaddubsw",
5638 int_x86_ssse3_pmadd_ub_sw_128,
5639 SSE_PMADD, loadv2i64, 0>, VEX_4V;
5641 defm VPMULHRSW : SS3I_binop_rm_int<0x0B, "vpmulhrsw",
5642 int_x86_ssse3_pmul_hr_sw_128,
5643 SSE_PMULHRSW, loadv2i64, 0>, VEX_4V;
5646 let ImmT = NoImm, Predicates = [HasAVX2] in {
5647 let isCommutable = 0 in {
5648 defm VPHADDWY : SS3I_binop_rm<0x01, "vphaddw", X86hadd, v16i16, VR256,
5650 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5651 defm VPHADDDY : SS3I_binop_rm<0x02, "vphaddd", X86hadd, v8i32, VR256,
5653 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5654 defm VPHSUBWY : SS3I_binop_rm<0x05, "vphsubw", X86hsub, v16i16, VR256,
5656 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5657 defm VPHSUBDY : SS3I_binop_rm<0x06, "vphsubd", X86hsub, v8i32, VR256,
5659 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5660 defm VPSIGNBY : SS3I_binop_rm<0x08, "vpsignb", X86psign, v32i8, VR256,
5662 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5663 defm VPSIGNWY : SS3I_binop_rm<0x09, "vpsignw", X86psign, v16i16, VR256,
5665 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5666 defm VPSIGNDY : SS3I_binop_rm<0x0A, "vpsignd", X86psign, v8i32, VR256,
5668 SSE_PHADDSUBW, 0>, VEX_4V, VEX_L;
5669 defm VPSHUFBY : SS3I_binop_rm<0x00, "vpshufb", X86pshufb, v32i8, VR256,
5671 SSE_PSHUFB, 0>, VEX_4V, VEX_L;
5672 defm VPHADDSW : SS3I_binop_rm_int_y<0x03, "vphaddsw",
5673 int_x86_avx2_phadd_sw,
5674 WriteVecALU>, VEX_4V, VEX_L;
5675 defm VPHSUBSW : SS3I_binop_rm_int_y<0x07, "vphsubsw",
5676 int_x86_avx2_phsub_sw,
5677 WriteVecALU>, VEX_4V, VEX_L;
5678 defm VPMADDUBSW : SS3I_binop_rm_int_y<0x04, "vpmaddubsw",
5679 int_x86_avx2_pmadd_ub_sw,
5680 WriteVecIMul>, VEX_4V, VEX_L;
5682 defm VPMULHRSW : SS3I_binop_rm_int_y<0x0B, "vpmulhrsw",
5683 int_x86_avx2_pmul_hr_sw,
5684 WriteVecIMul>, VEX_4V, VEX_L;
5687 // None of these have i8 immediate fields.
5688 let ImmT = NoImm, Constraints = "$src1 = $dst" in {
5689 let isCommutable = 0 in {
5690 defm PHADDW : SS3I_binop_rm<0x01, "phaddw", X86hadd, v8i16, VR128,
5691 memopv2i64, i128mem, SSE_PHADDSUBW>;
5692 defm PHADDD : SS3I_binop_rm<0x02, "phaddd", X86hadd, v4i32, VR128,
5693 memopv2i64, i128mem, SSE_PHADDSUBD>;
5694 defm PHSUBW : SS3I_binop_rm<0x05, "phsubw", X86hsub, v8i16, VR128,
5695 memopv2i64, i128mem, SSE_PHADDSUBW>;
5696 defm PHSUBD : SS3I_binop_rm<0x06, "phsubd", X86hsub, v4i32, VR128,
5697 memopv2i64, i128mem, SSE_PHADDSUBD>;
5698 defm PSIGNB : SS3I_binop_rm<0x08, "psignb", X86psign, v16i8, VR128,
5699 memopv2i64, i128mem, SSE_PSIGN>;
5700 defm PSIGNW : SS3I_binop_rm<0x09, "psignw", X86psign, v8i16, VR128,
5701 memopv2i64, i128mem, SSE_PSIGN>;
5702 defm PSIGND : SS3I_binop_rm<0x0A, "psignd", X86psign, v4i32, VR128,
5703 memopv2i64, i128mem, SSE_PSIGN>;
5704 defm PSHUFB : SS3I_binop_rm<0x00, "pshufb", X86pshufb, v16i8, VR128,
5705 memopv2i64, i128mem, SSE_PSHUFB>;
5706 defm PHADDSW : SS3I_binop_rm_int<0x03, "phaddsw",
5707 int_x86_ssse3_phadd_sw_128,
5708 SSE_PHADDSUBSW, memopv2i64>;
5709 defm PHSUBSW : SS3I_binop_rm_int<0x07, "phsubsw",
5710 int_x86_ssse3_phsub_sw_128,
5711 SSE_PHADDSUBSW, memopv2i64>;
5712 defm PMADDUBSW : SS3I_binop_rm_int<0x04, "pmaddubsw",
5713 int_x86_ssse3_pmadd_ub_sw_128,
5714 SSE_PMADD, memopv2i64>;
5716 defm PMULHRSW : SS3I_binop_rm_int<0x0B, "pmulhrsw",
5717 int_x86_ssse3_pmul_hr_sw_128,
5718 SSE_PMULHRSW, memopv2i64>;
5721 //===---------------------------------------------------------------------===//
5722 // SSSE3 - Packed Align Instruction Patterns
5723 //===---------------------------------------------------------------------===//
5725 multiclass ssse3_palignr<string asm, bit Is2Addr = 1> {
5726 let hasSideEffects = 0 in {
5727 def R128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
5728 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
5730 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
5732 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
5733 [], IIC_SSE_PALIGNRR>, Sched<[WriteShuffle]>;
5735 def R128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
5736 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
5738 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
5740 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
5741 [], IIC_SSE_PALIGNRM>, Sched<[WriteShuffleLd, ReadAfterLd]>;
5745 multiclass ssse3_palignr_y<string asm, bit Is2Addr = 1> {
5746 let hasSideEffects = 0 in {
5747 def R256rr : SS3AI<0x0F, MRMSrcReg, (outs VR256:$dst),
5748 (ins VR256:$src1, VR256:$src2, u8imm:$src3),
5750 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
5751 []>, Sched<[WriteShuffle]>;
5753 def R256rm : SS3AI<0x0F, MRMSrcMem, (outs VR256:$dst),
5754 (ins VR256:$src1, i256mem:$src2, u8imm:$src3),
5756 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
5757 []>, Sched<[WriteShuffleLd, ReadAfterLd]>;
5761 let Predicates = [HasAVX] in
5762 defm VPALIGN : ssse3_palignr<"vpalignr", 0>, VEX_4V;
5763 let Predicates = [HasAVX2] in
5764 defm VPALIGN : ssse3_palignr_y<"vpalignr", 0>, VEX_4V, VEX_L;
5765 let Constraints = "$src1 = $dst", Predicates = [UseSSSE3] in
5766 defm PALIGN : ssse3_palignr<"palignr">;
5768 let Predicates = [HasAVX2] in {
5769 def : Pat<(v8i32 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5770 (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
5771 def : Pat<(v8f32 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5772 (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
5773 def : Pat<(v16i16 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5774 (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
5775 def : Pat<(v32i8 (X86PAlignr VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5776 (VPALIGNR256rr VR256:$src2, VR256:$src1, imm:$imm)>;
5779 let Predicates = [HasAVX] in {
5780 def : Pat<(v4i32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5781 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5782 def : Pat<(v4f32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5783 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5784 def : Pat<(v8i16 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5785 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5786 def : Pat<(v16i8 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5787 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5790 let Predicates = [UseSSSE3] in {
5791 def : Pat<(v4i32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5792 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5793 def : Pat<(v4f32 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5794 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5795 def : Pat<(v8i16 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5796 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5797 def : Pat<(v16i8 (X86PAlignr VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5798 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
5801 //===---------------------------------------------------------------------===//
5802 // SSSE3 - Thread synchronization
5803 //===---------------------------------------------------------------------===//
5805 let SchedRW = [WriteSystem] in {
5806 let usesCustomInserter = 1 in {
5807 def MONITOR : PseudoI<(outs), (ins i32mem:$src1, GR32:$src2, GR32:$src3),
5808 [(int_x86_sse3_monitor addr:$src1, GR32:$src2, GR32:$src3)]>,
5809 Requires<[HasSSE3]>;
5812 let Uses = [EAX, ECX, EDX] in
5813 def MONITORrrr : I<0x01, MRM_C8, (outs), (ins), "monitor", [], IIC_SSE_MONITOR>,
5814 TB, Requires<[HasSSE3]>;
5815 let Uses = [ECX, EAX] in
5816 def MWAITrr : I<0x01, MRM_C9, (outs), (ins), "mwait",
5817 [(int_x86_sse3_mwait ECX, EAX)], IIC_SSE_MWAIT>,
5818 TB, Requires<[HasSSE3]>;
5821 def : InstAlias<"mwait\t{%eax, %ecx|ecx, eax}", (MWAITrr)>, Requires<[Not64BitMode]>;
5822 def : InstAlias<"mwait\t{%rax, %rcx|rcx, rax}", (MWAITrr)>, Requires<[In64BitMode]>;
5824 def : InstAlias<"monitor\t{%eax, %ecx, %edx|edx, ecx, eax}", (MONITORrrr)>,
5825 Requires<[Not64BitMode]>;
5826 def : InstAlias<"monitor\t{%rax, %rcx, %rdx|rdx, rcx, rax}", (MONITORrrr)>,
5827 Requires<[In64BitMode]>;
5829 //===----------------------------------------------------------------------===//
5830 // SSE4.1 - Packed Move with Sign/Zero Extend
5831 //===----------------------------------------------------------------------===//
5833 multiclass SS41I_pmovx_rrrm<bits<8> opc, string OpcodeStr, X86MemOperand MemOp,
5834 RegisterClass OutRC, RegisterClass InRC,
5836 def rr : SS48I<opc, MRMSrcReg, (outs OutRC:$dst), (ins InRC:$src),
5837 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5839 Sched<[itins.Sched]>;
5841 def rm : SS48I<opc, MRMSrcMem, (outs OutRC:$dst), (ins MemOp:$src),
5842 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5844 itins.rm>, Sched<[itins.Sched.Folded]>;
5847 multiclass SS41I_pmovx_rm_all<bits<8> opc, string OpcodeStr,
5848 X86MemOperand MemOp, X86MemOperand MemYOp,
5849 OpndItins SSEItins, OpndItins AVXItins,
5850 OpndItins AVX2Itins> {
5851 defm NAME : SS41I_pmovx_rrrm<opc, OpcodeStr, MemOp, VR128, VR128, SSEItins>;
5852 let Predicates = [HasAVX] in
5853 defm V#NAME : SS41I_pmovx_rrrm<opc, !strconcat("v", OpcodeStr), MemOp,
5854 VR128, VR128, AVXItins>, VEX;
5855 let Predicates = [HasAVX2] in
5856 defm V#NAME#Y : SS41I_pmovx_rrrm<opc, !strconcat("v", OpcodeStr), MemYOp,
5857 VR256, VR128, AVX2Itins>, VEX, VEX_L;
5860 multiclass SS41I_pmovx_rm<bits<8> opc, string OpcodeStr,
5861 X86MemOperand MemOp, X86MemOperand MemYOp> {
5862 defm PMOVSX#NAME : SS41I_pmovx_rm_all<opc, !strconcat("pmovsx", OpcodeStr),
5864 SSE_INTALU_ITINS_SHUFF_P,
5865 DEFAULT_ITINS_SHUFFLESCHED,
5866 DEFAULT_ITINS_SHUFFLESCHED>;
5867 defm PMOVZX#NAME : SS41I_pmovx_rm_all<!add(opc, 0x10),
5868 !strconcat("pmovzx", OpcodeStr),
5870 SSE_INTALU_ITINS_SHUFF_P,
5871 DEFAULT_ITINS_SHUFFLESCHED,
5872 DEFAULT_ITINS_SHUFFLESCHED>;
5875 defm BW : SS41I_pmovx_rm<0x20, "bw", i64mem, i128mem>;
5876 defm WD : SS41I_pmovx_rm<0x23, "wd", i64mem, i128mem>;
5877 defm DQ : SS41I_pmovx_rm<0x25, "dq", i64mem, i128mem>;
5879 defm BD : SS41I_pmovx_rm<0x21, "bd", i32mem, i64mem>;
5880 defm WQ : SS41I_pmovx_rm<0x24, "wq", i32mem, i64mem>;
5882 defm BQ : SS41I_pmovx_rm<0x22, "bq", i16mem, i32mem>;
5885 multiclass SS41I_pmovx_avx2_patterns<string OpcPrefix, string ExtTy, SDNode ExtOp> {
5886 // Register-Register patterns
5887 def : Pat<(v16i16 (ExtOp (v16i8 VR128:$src))),
5888 (!cast<I>(OpcPrefix#BWYrr) VR128:$src)>;
5889 def : Pat<(v8i32 (ExtOp (v16i8 VR128:$src))),
5890 (!cast<I>(OpcPrefix#BDYrr) VR128:$src)>;
5891 def : Pat<(v4i64 (ExtOp (v16i8 VR128:$src))),
5892 (!cast<I>(OpcPrefix#BQYrr) VR128:$src)>;
5894 def : Pat<(v8i32 (ExtOp (v8i16 VR128:$src))),
5895 (!cast<I>(OpcPrefix#WDYrr) VR128:$src)>;
5896 def : Pat<(v4i64 (ExtOp (v8i16 VR128:$src))),
5897 (!cast<I>(OpcPrefix#WQYrr) VR128:$src)>;
5899 def : Pat<(v4i64 (ExtOp (v4i32 VR128:$src))),
5900 (!cast<I>(OpcPrefix#DQYrr) VR128:$src)>;
5902 // On AVX2, we also support 256bit inputs.
5903 def : Pat<(v16i16 (ExtOp (v32i8 VR256:$src))),
5904 (!cast<I>(OpcPrefix#BWYrr) (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
5905 def : Pat<(v8i32 (ExtOp (v32i8 VR256:$src))),
5906 (!cast<I>(OpcPrefix#BDYrr) (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
5907 def : Pat<(v4i64 (ExtOp (v32i8 VR256:$src))),
5908 (!cast<I>(OpcPrefix#BQYrr) (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
5910 def : Pat<(v8i32 (ExtOp (v16i16 VR256:$src))),
5911 (!cast<I>(OpcPrefix#WDYrr) (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
5912 def : Pat<(v4i64 (ExtOp (v16i16 VR256:$src))),
5913 (!cast<I>(OpcPrefix#WQYrr) (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
5915 def : Pat<(v4i64 (ExtOp (v8i32 VR256:$src))),
5916 (!cast<I>(OpcPrefix#DQYrr) (EXTRACT_SUBREG VR256:$src, sub_xmm))>;
5918 // Simple Register-Memory patterns
5919 def : Pat<(v16i16 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5920 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5921 def : Pat<(v8i32 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5922 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5923 def : Pat<(v4i64 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
5924 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5926 def : Pat<(v8i32 (!cast<PatFrag>(ExtTy#"extloadvi16") addr:$src)),
5927 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5928 def : Pat<(v4i64 (!cast<PatFrag>(ExtTy#"extloadvi16") addr:$src)),
5929 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5931 def : Pat<(v4i64 (!cast<PatFrag>(ExtTy#"extloadvi32") addr:$src)),
5932 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5934 // AVX2 Register-Memory patterns
5935 def : Pat<(v16i16 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5936 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5937 def : Pat<(v16i16 (ExtOp (v16i8 (vzmovl_v2i64 addr:$src)))),
5938 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5939 def : Pat<(v16i16 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5940 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5941 def : Pat<(v16i16 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5942 (!cast<I>(OpcPrefix#BWYrm) addr:$src)>;
5944 def : Pat<(v8i32 (ExtOp (bc_v16i8 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
5945 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5946 def : Pat<(v8i32 (ExtOp (v16i8 (vzmovl_v2i64 addr:$src)))),
5947 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5948 def : Pat<(v8i32 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5949 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5950 def : Pat<(v8i32 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5951 (!cast<I>(OpcPrefix#BDYrm) addr:$src)>;
5953 def : Pat<(v4i64 (ExtOp (bc_v16i8 (v4i32 (scalar_to_vector (loadi32 addr:$src)))))),
5954 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5955 def : Pat<(v4i64 (ExtOp (v16i8 (vzmovl_v4i32 addr:$src)))),
5956 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5957 def : Pat<(v4i64 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
5958 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5959 def : Pat<(v4i64 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
5960 (!cast<I>(OpcPrefix#BQYrm) addr:$src)>;
5962 def : Pat<(v8i32 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
5963 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5964 def : Pat<(v8i32 (ExtOp (v8i16 (vzmovl_v2i64 addr:$src)))),
5965 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5966 def : Pat<(v8i32 (ExtOp (v8i16 (vzload_v2i64 addr:$src)))),
5967 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5968 def : Pat<(v8i32 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
5969 (!cast<I>(OpcPrefix#WDYrm) addr:$src)>;
5971 def : Pat<(v4i64 (ExtOp (bc_v8i16 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
5972 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5973 def : Pat<(v4i64 (ExtOp (v8i16 (vzmovl_v2i64 addr:$src)))),
5974 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5975 def : Pat<(v4i64 (ExtOp (v8i16 (vzload_v2i64 addr:$src)))),
5976 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5977 def : Pat<(v4i64 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
5978 (!cast<I>(OpcPrefix#WQYrm) addr:$src)>;
5980 def : Pat<(v4i64 (ExtOp (bc_v4i32 (loadv2i64 addr:$src)))),
5981 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5982 def : Pat<(v4i64 (ExtOp (v4i32 (vzmovl_v2i64 addr:$src)))),
5983 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5984 def : Pat<(v4i64 (ExtOp (v4i32 (vzload_v2i64 addr:$src)))),
5985 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5986 def : Pat<(v4i64 (ExtOp (bc_v4i32 (loadv2i64 addr:$src)))),
5987 (!cast<I>(OpcPrefix#DQYrm) addr:$src)>;
5990 let Predicates = [HasAVX2] in {
5991 defm : SS41I_pmovx_avx2_patterns<"VPMOVSX", "s", X86vsext>;
5992 defm : SS41I_pmovx_avx2_patterns<"VPMOVZX", "z", X86vzext>;
5995 // SSE4.1/AVX patterns.
5996 multiclass SS41I_pmovx_patterns<string OpcPrefix, string ExtTy,
5997 SDNode ExtOp, PatFrag ExtLoad16> {
5998 def : Pat<(v8i16 (ExtOp (v16i8 VR128:$src))),
5999 (!cast<I>(OpcPrefix#BWrr) VR128:$src)>;
6000 def : Pat<(v4i32 (ExtOp (v16i8 VR128:$src))),
6001 (!cast<I>(OpcPrefix#BDrr) VR128:$src)>;
6002 def : Pat<(v2i64 (ExtOp (v16i8 VR128:$src))),
6003 (!cast<I>(OpcPrefix#BQrr) VR128:$src)>;
6005 def : Pat<(v4i32 (ExtOp (v8i16 VR128:$src))),
6006 (!cast<I>(OpcPrefix#WDrr) VR128:$src)>;
6007 def : Pat<(v2i64 (ExtOp (v8i16 VR128:$src))),
6008 (!cast<I>(OpcPrefix#WQrr) VR128:$src)>;
6010 def : Pat<(v2i64 (ExtOp (v4i32 VR128:$src))),
6011 (!cast<I>(OpcPrefix#DQrr) VR128:$src)>;
6013 def : Pat<(v8i16 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
6014 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
6015 def : Pat<(v4i32 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
6016 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
6017 def : Pat<(v2i64 (!cast<PatFrag>(ExtTy#"extloadvi8") addr:$src)),
6018 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
6020 def : Pat<(v4i32 (!cast<PatFrag>(ExtTy#"extloadvi16") addr:$src)),
6021 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
6022 def : Pat<(v2i64 (!cast<PatFrag>(ExtTy#"extloadvi16") addr:$src)),
6023 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
6025 def : Pat<(v2i64 (!cast<PatFrag>(ExtTy#"extloadvi32") addr:$src)),
6026 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
6028 def : Pat<(v8i16 (ExtOp (bc_v16i8 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
6029 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
6030 def : Pat<(v8i16 (ExtOp (bc_v16i8 (v2f64 (scalar_to_vector (loadf64 addr:$src)))))),
6031 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
6032 def : Pat<(v8i16 (ExtOp (v16i8 (vzmovl_v2i64 addr:$src)))),
6033 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
6034 def : Pat<(v8i16 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
6035 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
6036 def : Pat<(v8i16 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
6037 (!cast<I>(OpcPrefix#BWrm) addr:$src)>;
6039 def : Pat<(v4i32 (ExtOp (bc_v16i8 (v4i32 (scalar_to_vector (loadi32 addr:$src)))))),
6040 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
6041 def : Pat<(v4i32 (ExtOp (v16i8 (vzmovl_v4i32 addr:$src)))),
6042 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
6043 def : Pat<(v4i32 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
6044 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
6045 def : Pat<(v4i32 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
6046 (!cast<I>(OpcPrefix#BDrm) addr:$src)>;
6048 def : Pat<(v2i64 (ExtOp (bc_v16i8 (v4i32 (scalar_to_vector (ExtLoad16 addr:$src)))))),
6049 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
6050 def : Pat<(v2i64 (ExtOp (v16i8 (vzmovl_v4i32 addr:$src)))),
6051 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
6052 def : Pat<(v2i64 (ExtOp (v16i8 (vzload_v2i64 addr:$src)))),
6053 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
6054 def : Pat<(v2i64 (ExtOp (bc_v16i8 (loadv2i64 addr:$src)))),
6055 (!cast<I>(OpcPrefix#BQrm) addr:$src)>;
6057 def : Pat<(v4i32 (ExtOp (bc_v8i16 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
6058 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
6059 def : Pat<(v4i32 (ExtOp (bc_v8i16 (v2f64 (scalar_to_vector (loadf64 addr:$src)))))),
6060 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
6061 def : Pat<(v4i32 (ExtOp (v8i16 (vzmovl_v2i64 addr:$src)))),
6062 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
6063 def : Pat<(v4i32 (ExtOp (v8i16 (vzload_v2i64 addr:$src)))),
6064 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
6065 def : Pat<(v4i32 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
6066 (!cast<I>(OpcPrefix#WDrm) addr:$src)>;
6068 def : Pat<(v2i64 (ExtOp (bc_v8i16 (v4i32 (scalar_to_vector (loadi32 addr:$src)))))),
6069 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
6070 def : Pat<(v2i64 (ExtOp (v8i16 (vzmovl_v4i32 addr:$src)))),
6071 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
6072 def : Pat<(v2i64 (ExtOp (v8i16 (vzload_v2i64 addr:$src)))),
6073 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
6074 def : Pat<(v2i64 (ExtOp (bc_v8i16 (loadv2i64 addr:$src)))),
6075 (!cast<I>(OpcPrefix#WQrm) addr:$src)>;
6077 def : Pat<(v2i64 (ExtOp (bc_v4i32 (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
6078 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
6079 def : Pat<(v2i64 (ExtOp (bc_v4i32 (v2f64 (scalar_to_vector (loadf64 addr:$src)))))),
6080 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
6081 def : Pat<(v2i64 (ExtOp (v4i32 (vzmovl_v2i64 addr:$src)))),
6082 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
6083 def : Pat<(v2i64 (ExtOp (v4i32 (vzload_v2i64 addr:$src)))),
6084 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
6085 def : Pat<(v2i64 (ExtOp (bc_v4i32 (loadv2i64 addr:$src)))),
6086 (!cast<I>(OpcPrefix#DQrm) addr:$src)>;
6089 let Predicates = [HasAVX] in {
6090 defm : SS41I_pmovx_patterns<"VPMOVSX", "s", X86vsext, extloadi32i16>;
6091 defm : SS41I_pmovx_patterns<"VPMOVZX", "z", X86vzext, loadi16_anyext>;
6094 let Predicates = [UseSSE41] in {
6095 defm : SS41I_pmovx_patterns<"PMOVSX", "s", X86vsext, extloadi32i16>;
6096 defm : SS41I_pmovx_patterns<"PMOVZX", "z", X86vzext, loadi16_anyext>;
6099 //===----------------------------------------------------------------------===//
6100 // SSE4.1 - Extract Instructions
6101 //===----------------------------------------------------------------------===//
6103 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
6104 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
6105 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
6106 (ins VR128:$src1, u8imm:$src2),
6107 !strconcat(OpcodeStr,
6108 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6109 [(set GR32orGR64:$dst, (X86pextrb (v16i8 VR128:$src1),
6111 Sched<[WriteShuffle]>;
6112 let hasSideEffects = 0, mayStore = 1,
6113 SchedRW = [WriteShuffleLd, WriteRMW] in
6114 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6115 (ins i8mem:$dst, VR128:$src1, u8imm:$src2),
6116 !strconcat(OpcodeStr,
6117 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6118 [(store (i8 (trunc (assertzext (X86pextrb (v16i8 VR128:$src1),
6119 imm:$src2)))), addr:$dst)]>;
6122 let Predicates = [HasAVX] in
6123 defm VPEXTRB : SS41I_extract8<0x14, "vpextrb">, VEX;
6125 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
6128 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
6129 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
6130 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in
6131 def rr_REV : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
6132 (ins VR128:$src1, u8imm:$src2),
6133 !strconcat(OpcodeStr,
6134 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6135 []>, Sched<[WriteShuffle]>;
6137 let hasSideEffects = 0, mayStore = 1,
6138 SchedRW = [WriteShuffleLd, WriteRMW] in
6139 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6140 (ins i16mem:$dst, VR128:$src1, u8imm:$src2),
6141 !strconcat(OpcodeStr,
6142 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6143 [(store (i16 (trunc (assertzext (X86pextrw (v8i16 VR128:$src1),
6144 imm:$src2)))), addr:$dst)]>;
6147 let Predicates = [HasAVX] in
6148 defm VPEXTRW : SS41I_extract16<0x15, "vpextrw">, VEX;
6150 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
6153 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
6154 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
6155 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
6156 (ins VR128:$src1, u8imm:$src2),
6157 !strconcat(OpcodeStr,
6158 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6160 (extractelt (v4i32 VR128:$src1), imm:$src2))]>,
6161 Sched<[WriteShuffle]>;
6162 let SchedRW = [WriteShuffleLd, WriteRMW] in
6163 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6164 (ins i32mem:$dst, VR128:$src1, u8imm:$src2),
6165 !strconcat(OpcodeStr,
6166 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6167 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
6171 let Predicates = [HasAVX] in
6172 defm VPEXTRD : SS41I_extract32<0x16, "vpextrd">, VEX;
6174 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
6176 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
6177 multiclass SS41I_extract64<bits<8> opc, string OpcodeStr> {
6178 def rr : SS4AIi8<opc, MRMDestReg, (outs GR64:$dst),
6179 (ins VR128:$src1, u8imm:$src2),
6180 !strconcat(OpcodeStr,
6181 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6183 (extractelt (v2i64 VR128:$src1), imm:$src2))]>,
6184 Sched<[WriteShuffle]>, REX_W;
6185 let SchedRW = [WriteShuffleLd, WriteRMW] in
6186 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6187 (ins i64mem:$dst, VR128:$src1, u8imm:$src2),
6188 !strconcat(OpcodeStr,
6189 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6190 [(store (extractelt (v2i64 VR128:$src1), imm:$src2),
6191 addr:$dst)]>, REX_W;
6194 let Predicates = [HasAVX] in
6195 defm VPEXTRQ : SS41I_extract64<0x16, "vpextrq">, VEX, VEX_W;
6197 defm PEXTRQ : SS41I_extract64<0x16, "pextrq">;
6199 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
6201 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr,
6202 OpndItins itins = DEFAULT_ITINS> {
6203 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32orGR64:$dst),
6204 (ins VR128:$src1, u8imm:$src2),
6205 !strconcat(OpcodeStr,
6206 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6207 [(set GR32orGR64:$dst,
6208 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))],
6209 itins.rr>, Sched<[WriteFBlend]>;
6210 let SchedRW = [WriteFBlendLd, WriteRMW] in
6211 def mr : SS4AIi8<opc, MRMDestMem, (outs),
6212 (ins f32mem:$dst, VR128:$src1, u8imm:$src2),
6213 !strconcat(OpcodeStr,
6214 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6215 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
6216 addr:$dst)], itins.rm>;
6219 let ExeDomain = SSEPackedSingle in {
6220 let Predicates = [UseAVX] in
6221 defm VEXTRACTPS : SS41I_extractf32<0x17, "vextractps">, VEX;
6222 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps", SSE_EXTRACT_ITINS>;
6225 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
6226 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
6229 (VEXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
6231 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
6234 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
6235 Requires<[UseSSE41]>;
6237 //===----------------------------------------------------------------------===//
6238 // SSE4.1 - Insert Instructions
6239 //===----------------------------------------------------------------------===//
6241 multiclass SS41I_insert8<bits<8> opc, string asm, bit Is2Addr = 1> {
6242 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6243 (ins VR128:$src1, GR32orGR64:$src2, u8imm:$src3),
6245 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6247 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6249 (X86pinsrb VR128:$src1, GR32orGR64:$src2, imm:$src3))]>,
6250 Sched<[WriteShuffle]>;
6251 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6252 (ins VR128:$src1, i8mem:$src2, u8imm:$src3),
6254 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6256 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6258 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
6259 imm:$src3))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6262 let Predicates = [HasAVX] in
6263 defm VPINSRB : SS41I_insert8<0x20, "vpinsrb", 0>, VEX_4V;
6264 let Constraints = "$src1 = $dst" in
6265 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
6267 multiclass SS41I_insert32<bits<8> opc, string asm, bit Is2Addr = 1> {
6268 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6269 (ins VR128:$src1, GR32:$src2, u8imm:$src3),
6271 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6273 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6275 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
6276 Sched<[WriteShuffle]>;
6277 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6278 (ins VR128:$src1, i32mem:$src2, u8imm:$src3),
6280 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6282 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6284 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
6285 imm:$src3)))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6288 let Predicates = [HasAVX] in
6289 defm VPINSRD : SS41I_insert32<0x22, "vpinsrd", 0>, VEX_4V;
6290 let Constraints = "$src1 = $dst" in
6291 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
6293 multiclass SS41I_insert64<bits<8> opc, string asm, bit Is2Addr = 1> {
6294 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6295 (ins VR128:$src1, GR64:$src2, u8imm:$src3),
6297 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6299 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6301 (v2i64 (insertelt VR128:$src1, GR64:$src2, imm:$src3)))]>,
6302 Sched<[WriteShuffle]>;
6303 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6304 (ins VR128:$src1, i64mem:$src2, u8imm:$src3),
6306 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6308 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6310 (v2i64 (insertelt VR128:$src1, (loadi64 addr:$src2),
6311 imm:$src3)))]>, Sched<[WriteShuffleLd, ReadAfterLd]>;
6314 let Predicates = [HasAVX] in
6315 defm VPINSRQ : SS41I_insert64<0x22, "vpinsrq", 0>, VEX_4V, VEX_W;
6316 let Constraints = "$src1 = $dst" in
6317 defm PINSRQ : SS41I_insert64<0x22, "pinsrq">, REX_W;
6319 // insertps has a few different modes, there's the first two here below which
6320 // are optimized inserts that won't zero arbitrary elements in the destination
6321 // vector. The next one matches the intrinsic and could zero arbitrary elements
6322 // in the target vector.
6323 multiclass SS41I_insertf32<bits<8> opc, string asm, bit Is2Addr = 1,
6324 OpndItins itins = DEFAULT_ITINS> {
6325 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
6326 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
6328 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6330 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6332 (X86insertps VR128:$src1, VR128:$src2, imm:$src3))], itins.rr>,
6333 Sched<[WriteFShuffle]>;
6334 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
6335 (ins VR128:$src1, f32mem:$src2, u8imm:$src3),
6337 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6339 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6341 (X86insertps VR128:$src1,
6342 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
6343 imm:$src3))], itins.rm>,
6344 Sched<[WriteFShuffleLd, ReadAfterLd]>;
6347 let ExeDomain = SSEPackedSingle in {
6348 let Predicates = [UseAVX] in
6349 defm VINSERTPS : SS41I_insertf32<0x21, "vinsertps", 0>, VEX_4V;
6350 let Constraints = "$src1 = $dst" in
6351 defm INSERTPS : SS41I_insertf32<0x21, "insertps", 1, SSE_INSERT_ITINS>;
6354 let Predicates = [UseSSE41] in {
6355 // If we're inserting an element from a load or a null pshuf of a load,
6356 // fold the load into the insertps instruction.
6357 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd (v4f32
6358 (scalar_to_vector (loadf32 addr:$src2))), (i8 0)),
6360 (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6361 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd
6362 (loadv4f32 addr:$src2), (i8 0)), imm:$src3)),
6363 (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6366 let Predicates = [UseAVX] in {
6367 // If we're inserting an element from a vbroadcast of a load, fold the
6368 // load into the X86insertps instruction.
6369 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
6370 (X86VBroadcast (loadf32 addr:$src2)), imm:$src3)),
6371 (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6372 def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
6373 (X86VBroadcast (loadv4f32 addr:$src2)), imm:$src3)),
6374 (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
6377 //===----------------------------------------------------------------------===//
6378 // SSE4.1 - Round Instructions
6379 //===----------------------------------------------------------------------===//
6381 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd, string OpcodeStr,
6382 X86MemOperand x86memop, RegisterClass RC,
6383 PatFrag mem_frag32, PatFrag mem_frag64,
6384 Intrinsic V4F32Int, Intrinsic V2F64Int> {
6385 let ExeDomain = SSEPackedSingle in {
6386 // Intrinsic operation, reg.
6387 // Vector intrinsic operation, reg
6388 def PSr : SS4AIi8<opcps, MRMSrcReg,
6389 (outs RC:$dst), (ins RC:$src1, i32u8imm:$src2),
6390 !strconcat(OpcodeStr,
6391 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6392 [(set RC:$dst, (V4F32Int RC:$src1, imm:$src2))],
6393 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAdd]>;
6395 // Vector intrinsic operation, mem
6396 def PSm : SS4AIi8<opcps, MRMSrcMem,
6397 (outs RC:$dst), (ins x86memop:$src1, i32u8imm:$src2),
6398 !strconcat(OpcodeStr,
6399 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6401 (V4F32Int (mem_frag32 addr:$src1),imm:$src2))],
6402 IIC_SSE_ROUNDPS_MEM>, Sched<[WriteFAddLd]>;
6403 } // ExeDomain = SSEPackedSingle
6405 let ExeDomain = SSEPackedDouble in {
6406 // Vector intrinsic operation, reg
6407 def PDr : SS4AIi8<opcpd, MRMSrcReg,
6408 (outs RC:$dst), (ins RC:$src1, i32u8imm:$src2),
6409 !strconcat(OpcodeStr,
6410 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6411 [(set RC:$dst, (V2F64Int RC:$src1, imm:$src2))],
6412 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAdd]>;
6414 // Vector intrinsic operation, mem
6415 def PDm : SS4AIi8<opcpd, MRMSrcMem,
6416 (outs RC:$dst), (ins x86memop:$src1, i32u8imm:$src2),
6417 !strconcat(OpcodeStr,
6418 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
6420 (V2F64Int (mem_frag64 addr:$src1),imm:$src2))],
6421 IIC_SSE_ROUNDPS_REG>, Sched<[WriteFAddLd]>;
6422 } // ExeDomain = SSEPackedDouble
6425 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
6428 Intrinsic F64Int, bit Is2Addr = 1> {
6429 let ExeDomain = GenericDomain in {
6431 let hasSideEffects = 0 in
6432 def SSr : SS4AIi8<opcss, MRMSrcReg,
6433 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2, i32u8imm:$src3),
6435 !strconcat(OpcodeStr,
6436 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6437 !strconcat(OpcodeStr,
6438 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6439 []>, Sched<[WriteFAdd]>;
6441 // Intrinsic operation, reg.
6442 let isCodeGenOnly = 1 in
6443 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
6444 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32u8imm:$src3),
6446 !strconcat(OpcodeStr,
6447 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6448 !strconcat(OpcodeStr,
6449 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6450 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
6453 // Intrinsic operation, mem.
6454 def SSm : SS4AIi8<opcss, MRMSrcMem,
6455 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2, i32u8imm:$src3),
6457 !strconcat(OpcodeStr,
6458 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6459 !strconcat(OpcodeStr,
6460 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6462 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
6463 Sched<[WriteFAddLd, ReadAfterLd]>;
6466 let hasSideEffects = 0 in
6467 def SDr : SS4AIi8<opcsd, MRMSrcReg,
6468 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2, i32u8imm:$src3),
6470 !strconcat(OpcodeStr,
6471 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6472 !strconcat(OpcodeStr,
6473 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6474 []>, Sched<[WriteFAdd]>;
6476 // Intrinsic operation, reg.
6477 let isCodeGenOnly = 1 in
6478 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
6479 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32u8imm:$src3),
6481 !strconcat(OpcodeStr,
6482 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6483 !strconcat(OpcodeStr,
6484 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6485 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
6488 // Intrinsic operation, mem.
6489 def SDm : SS4AIi8<opcsd, MRMSrcMem,
6490 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2, i32u8imm:$src3),
6492 !strconcat(OpcodeStr,
6493 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6494 !strconcat(OpcodeStr,
6495 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6497 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
6498 Sched<[WriteFAddLd, ReadAfterLd]>;
6499 } // ExeDomain = GenericDomain
6502 // FP round - roundss, roundps, roundsd, roundpd
6503 let Predicates = [HasAVX] in {
6505 defm VROUND : sse41_fp_unop_rm<0x08, 0x09, "vround", f128mem, VR128,
6506 loadv4f32, loadv2f64,
6507 int_x86_sse41_round_ps,
6508 int_x86_sse41_round_pd>, VEX;
6509 defm VROUNDY : sse41_fp_unop_rm<0x08, 0x09, "vround", f256mem, VR256,
6510 loadv8f32, loadv4f64,
6511 int_x86_avx_round_ps_256,
6512 int_x86_avx_round_pd_256>, VEX, VEX_L;
6513 defm VROUND : sse41_fp_binop_rm<0x0A, 0x0B, "vround",
6514 int_x86_sse41_round_ss,
6515 int_x86_sse41_round_sd, 0>, VEX_4V, VEX_LIG;
6518 let Predicates = [UseAVX] in {
6519 def : Pat<(ffloor FR32:$src),
6520 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x1))>;
6521 def : Pat<(f64 (ffloor FR64:$src)),
6522 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x1))>;
6523 def : Pat<(f32 (fnearbyint FR32:$src)),
6524 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0xC))>;
6525 def : Pat<(f64 (fnearbyint FR64:$src)),
6526 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0xC))>;
6527 def : Pat<(f32 (fceil FR32:$src)),
6528 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x2))>;
6529 def : Pat<(f64 (fceil FR64:$src)),
6530 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x2))>;
6531 def : Pat<(f32 (frint FR32:$src)),
6532 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x4))>;
6533 def : Pat<(f64 (frint FR64:$src)),
6534 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x4))>;
6535 def : Pat<(f32 (ftrunc FR32:$src)),
6536 (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x3))>;
6537 def : Pat<(f64 (ftrunc FR64:$src)),
6538 (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x3))>;
6541 let Predicates = [HasAVX] in {
6542 def : Pat<(v4f32 (ffloor VR128:$src)),
6543 (VROUNDPSr VR128:$src, (i32 0x1))>;
6544 def : Pat<(v4f32 (fnearbyint VR128:$src)),
6545 (VROUNDPSr VR128:$src, (i32 0xC))>;
6546 def : Pat<(v4f32 (fceil VR128:$src)),
6547 (VROUNDPSr VR128:$src, (i32 0x2))>;
6548 def : Pat<(v4f32 (frint VR128:$src)),
6549 (VROUNDPSr VR128:$src, (i32 0x4))>;
6550 def : Pat<(v4f32 (ftrunc VR128:$src)),
6551 (VROUNDPSr VR128:$src, (i32 0x3))>;
6553 def : Pat<(v2f64 (ffloor VR128:$src)),
6554 (VROUNDPDr VR128:$src, (i32 0x1))>;
6555 def : Pat<(v2f64 (fnearbyint VR128:$src)),
6556 (VROUNDPDr VR128:$src, (i32 0xC))>;
6557 def : Pat<(v2f64 (fceil VR128:$src)),
6558 (VROUNDPDr VR128:$src, (i32 0x2))>;
6559 def : Pat<(v2f64 (frint VR128:$src)),
6560 (VROUNDPDr VR128:$src, (i32 0x4))>;
6561 def : Pat<(v2f64 (ftrunc VR128:$src)),
6562 (VROUNDPDr VR128:$src, (i32 0x3))>;
6564 def : Pat<(v8f32 (ffloor VR256:$src)),
6565 (VROUNDYPSr VR256:$src, (i32 0x1))>;
6566 def : Pat<(v8f32 (fnearbyint VR256:$src)),
6567 (VROUNDYPSr VR256:$src, (i32 0xC))>;
6568 def : Pat<(v8f32 (fceil VR256:$src)),
6569 (VROUNDYPSr VR256:$src, (i32 0x2))>;
6570 def : Pat<(v8f32 (frint VR256:$src)),
6571 (VROUNDYPSr VR256:$src, (i32 0x4))>;
6572 def : Pat<(v8f32 (ftrunc VR256:$src)),
6573 (VROUNDYPSr VR256:$src, (i32 0x3))>;
6575 def : Pat<(v4f64 (ffloor VR256:$src)),
6576 (VROUNDYPDr VR256:$src, (i32 0x1))>;
6577 def : Pat<(v4f64 (fnearbyint VR256:$src)),
6578 (VROUNDYPDr VR256:$src, (i32 0xC))>;
6579 def : Pat<(v4f64 (fceil VR256:$src)),
6580 (VROUNDYPDr VR256:$src, (i32 0x2))>;
6581 def : Pat<(v4f64 (frint VR256:$src)),
6582 (VROUNDYPDr VR256:$src, (i32 0x4))>;
6583 def : Pat<(v4f64 (ftrunc VR256:$src)),
6584 (VROUNDYPDr VR256:$src, (i32 0x3))>;
6587 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round", f128mem, VR128,
6588 memopv4f32, memopv2f64,
6589 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
6590 let Constraints = "$src1 = $dst" in
6591 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
6592 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
6594 let Predicates = [UseSSE41] in {
6595 def : Pat<(ffloor FR32:$src),
6596 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x1))>;
6597 def : Pat<(f64 (ffloor FR64:$src)),
6598 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x1))>;
6599 def : Pat<(f32 (fnearbyint FR32:$src)),
6600 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0xC))>;
6601 def : Pat<(f64 (fnearbyint FR64:$src)),
6602 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0xC))>;
6603 def : Pat<(f32 (fceil FR32:$src)),
6604 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x2))>;
6605 def : Pat<(f64 (fceil FR64:$src)),
6606 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x2))>;
6607 def : Pat<(f32 (frint FR32:$src)),
6608 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x4))>;
6609 def : Pat<(f64 (frint FR64:$src)),
6610 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x4))>;
6611 def : Pat<(f32 (ftrunc FR32:$src)),
6612 (ROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src, (i32 0x3))>;
6613 def : Pat<(f64 (ftrunc FR64:$src)),
6614 (ROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src, (i32 0x3))>;
6616 def : Pat<(v4f32 (ffloor VR128:$src)),
6617 (ROUNDPSr VR128:$src, (i32 0x1))>;
6618 def : Pat<(v4f32 (fnearbyint VR128:$src)),
6619 (ROUNDPSr VR128:$src, (i32 0xC))>;
6620 def : Pat<(v4f32 (fceil VR128:$src)),
6621 (ROUNDPSr VR128:$src, (i32 0x2))>;
6622 def : Pat<(v4f32 (frint VR128:$src)),
6623 (ROUNDPSr VR128:$src, (i32 0x4))>;
6624 def : Pat<(v4f32 (ftrunc VR128:$src)),
6625 (ROUNDPSr VR128:$src, (i32 0x3))>;
6627 def : Pat<(v2f64 (ffloor VR128:$src)),
6628 (ROUNDPDr VR128:$src, (i32 0x1))>;
6629 def : Pat<(v2f64 (fnearbyint VR128:$src)),
6630 (ROUNDPDr VR128:$src, (i32 0xC))>;
6631 def : Pat<(v2f64 (fceil VR128:$src)),
6632 (ROUNDPDr VR128:$src, (i32 0x2))>;
6633 def : Pat<(v2f64 (frint VR128:$src)),
6634 (ROUNDPDr VR128:$src, (i32 0x4))>;
6635 def : Pat<(v2f64 (ftrunc VR128:$src)),
6636 (ROUNDPDr VR128:$src, (i32 0x3))>;
6639 //===----------------------------------------------------------------------===//
6640 // SSE4.1 - Packed Bit Test
6641 //===----------------------------------------------------------------------===//
6643 // ptest instruction we'll lower to this in X86ISelLowering primarily from
6644 // the intel intrinsic that corresponds to this.
6645 let Defs = [EFLAGS], Predicates = [HasAVX] in {
6646 def VPTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
6647 "vptest\t{$src2, $src1|$src1, $src2}",
6648 [(set EFLAGS, (X86ptest VR128:$src1, (v2i64 VR128:$src2)))]>,
6649 Sched<[WriteVecLogic]>, VEX;
6650 def VPTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
6651 "vptest\t{$src2, $src1|$src1, $src2}",
6652 [(set EFLAGS,(X86ptest VR128:$src1, (loadv2i64 addr:$src2)))]>,
6653 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX;
6655 def VPTESTYrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR256:$src1, VR256:$src2),
6656 "vptest\t{$src2, $src1|$src1, $src2}",
6657 [(set EFLAGS, (X86ptest VR256:$src1, (v4i64 VR256:$src2)))]>,
6658 Sched<[WriteVecLogic]>, VEX, VEX_L;
6659 def VPTESTYrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR256:$src1, i256mem:$src2),
6660 "vptest\t{$src2, $src1|$src1, $src2}",
6661 [(set EFLAGS,(X86ptest VR256:$src1, (loadv4i64 addr:$src2)))]>,
6662 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX, VEX_L;
6665 let Defs = [EFLAGS] in {
6666 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
6667 "ptest\t{$src2, $src1|$src1, $src2}",
6668 [(set EFLAGS, (X86ptest VR128:$src1, (v2i64 VR128:$src2)))]>,
6669 Sched<[WriteVecLogic]>;
6670 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
6671 "ptest\t{$src2, $src1|$src1, $src2}",
6672 [(set EFLAGS, (X86ptest VR128:$src1, (memopv2i64 addr:$src2)))]>,
6673 Sched<[WriteVecLogicLd, ReadAfterLd]>;
6676 // The bit test instructions below are AVX only
6677 multiclass avx_bittest<bits<8> opc, string OpcodeStr, RegisterClass RC,
6678 X86MemOperand x86memop, PatFrag mem_frag, ValueType vt> {
6679 def rr : SS48I<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
6680 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
6681 [(set EFLAGS, (X86testp RC:$src1, (vt RC:$src2)))]>,
6682 Sched<[WriteVecLogic]>, VEX;
6683 def rm : SS48I<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
6684 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
6685 [(set EFLAGS, (X86testp RC:$src1, (mem_frag addr:$src2)))]>,
6686 Sched<[WriteVecLogicLd, ReadAfterLd]>, VEX;
6689 let Defs = [EFLAGS], Predicates = [HasAVX] in {
6690 let ExeDomain = SSEPackedSingle in {
6691 defm VTESTPS : avx_bittest<0x0E, "vtestps", VR128, f128mem, loadv4f32, v4f32>;
6692 defm VTESTPSY : avx_bittest<0x0E, "vtestps", VR256, f256mem, loadv8f32, v8f32>,
6695 let ExeDomain = SSEPackedDouble in {
6696 defm VTESTPD : avx_bittest<0x0F, "vtestpd", VR128, f128mem, loadv2f64, v2f64>;
6697 defm VTESTPDY : avx_bittest<0x0F, "vtestpd", VR256, f256mem, loadv4f64, v4f64>,
6702 //===----------------------------------------------------------------------===//
6703 // SSE4.1 - Misc Instructions
6704 //===----------------------------------------------------------------------===//
6706 let Defs = [EFLAGS], Predicates = [HasPOPCNT] in {
6707 def POPCNT16rr : I<0xB8, MRMSrcReg, (outs GR16:$dst), (ins GR16:$src),
6708 "popcnt{w}\t{$src, $dst|$dst, $src}",
6709 [(set GR16:$dst, (ctpop GR16:$src)), (implicit EFLAGS)],
6710 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>,
6712 def POPCNT16rm : I<0xB8, MRMSrcMem, (outs GR16:$dst), (ins i16mem:$src),
6713 "popcnt{w}\t{$src, $dst|$dst, $src}",
6714 [(set GR16:$dst, (ctpop (loadi16 addr:$src))),
6715 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
6716 Sched<[WriteFAddLd]>, OpSize16, XS;
6718 def POPCNT32rr : I<0xB8, MRMSrcReg, (outs GR32:$dst), (ins GR32:$src),
6719 "popcnt{l}\t{$src, $dst|$dst, $src}",
6720 [(set GR32:$dst, (ctpop GR32:$src)), (implicit EFLAGS)],
6721 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>,
6724 def POPCNT32rm : I<0xB8, MRMSrcMem, (outs GR32:$dst), (ins i32mem:$src),
6725 "popcnt{l}\t{$src, $dst|$dst, $src}",
6726 [(set GR32:$dst, (ctpop (loadi32 addr:$src))),
6727 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
6728 Sched<[WriteFAddLd]>, OpSize32, XS;
6730 def POPCNT64rr : RI<0xB8, MRMSrcReg, (outs GR64:$dst), (ins GR64:$src),
6731 "popcnt{q}\t{$src, $dst|$dst, $src}",
6732 [(set GR64:$dst, (ctpop GR64:$src)), (implicit EFLAGS)],
6733 IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>, XS;
6734 def POPCNT64rm : RI<0xB8, MRMSrcMem, (outs GR64:$dst), (ins i64mem:$src),
6735 "popcnt{q}\t{$src, $dst|$dst, $src}",
6736 [(set GR64:$dst, (ctpop (loadi64 addr:$src))),
6737 (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
6738 Sched<[WriteFAddLd]>, XS;
6743 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
6744 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
6745 Intrinsic IntId128, PatFrag ld_frag,
6746 X86FoldableSchedWrite Sched> {
6747 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
6749 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6750 [(set VR128:$dst, (IntId128 VR128:$src))]>,
6752 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
6754 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
6756 (IntId128 (bitconvert (ld_frag addr:$src))))]>,
6757 Sched<[Sched.Folded]>;
6760 // PHMIN has the same profile as PSAD, thus we use the same scheduling
6761 // model, although the naming is misleading.
6762 let Predicates = [HasAVX] in
6763 defm VPHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "vphminposuw",
6764 int_x86_sse41_phminposuw, loadv2i64,
6766 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
6767 int_x86_sse41_phminposuw, memopv2i64,
6770 /// SS48I_binop_rm - Simple SSE41 binary operator.
6771 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
6772 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
6773 X86MemOperand x86memop, bit Is2Addr = 1,
6774 OpndItins itins = SSE_INTALU_ITINS_P> {
6775 let isCommutable = 1 in
6776 def rr : SS48I<opc, MRMSrcReg, (outs RC:$dst),
6777 (ins RC:$src1, RC:$src2),
6779 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
6780 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
6781 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))]>,
6782 Sched<[itins.Sched]>;
6783 def rm : SS48I<opc, MRMSrcMem, (outs RC:$dst),
6784 (ins RC:$src1, x86memop:$src2),
6786 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
6787 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
6789 (OpVT (OpNode RC:$src1, (bitconvert (memop_frag addr:$src2)))))]>,
6790 Sched<[itins.Sched.Folded, ReadAfterLd]>;
6793 /// SS48I_binop_rm2 - Simple SSE41 binary operator with different src and dst
6795 multiclass SS48I_binop_rm2<bits<8> opc, string OpcodeStr, SDNode OpNode,
6796 ValueType DstVT, ValueType SrcVT, RegisterClass RC,
6797 PatFrag memop_frag, X86MemOperand x86memop,
6799 bit IsCommutable = 0, bit Is2Addr = 1> {
6800 let isCommutable = IsCommutable in
6801 def rr : SS48I<opc, MRMSrcReg, (outs RC:$dst),
6802 (ins RC:$src1, RC:$src2),
6804 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
6805 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
6806 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1), RC:$src2)))]>,
6807 Sched<[itins.Sched]>;
6808 def rm : SS48I<opc, MRMSrcMem, (outs RC:$dst),
6809 (ins RC:$src1, x86memop:$src2),
6811 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
6812 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
6813 [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1),
6814 (bitconvert (memop_frag addr:$src2)))))]>,
6815 Sched<[itins.Sched.Folded, ReadAfterLd]>;
6818 let Predicates = [HasAVX, NoVLX] in {
6819 let isCommutable = 0 in
6820 defm VPMINSB : SS48I_binop_rm<0x38, "vpminsb", X86smin, v16i8, VR128,
6821 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6823 defm VPMINSD : SS48I_binop_rm<0x39, "vpminsd", X86smin, v4i32, VR128,
6824 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6826 defm VPMINUD : SS48I_binop_rm<0x3B, "vpminud", X86umin, v4i32, VR128,
6827 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6829 defm VPMINUW : SS48I_binop_rm<0x3A, "vpminuw", X86umin, v8i16, VR128,
6830 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6832 defm VPMAXSB : SS48I_binop_rm<0x3C, "vpmaxsb", X86smax, v16i8, VR128,
6833 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6835 defm VPMAXSD : SS48I_binop_rm<0x3D, "vpmaxsd", X86smax, v4i32, VR128,
6836 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6838 defm VPMAXUD : SS48I_binop_rm<0x3F, "vpmaxud", X86umax, v4i32, VR128,
6839 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6841 defm VPMAXUW : SS48I_binop_rm<0x3E, "vpmaxuw", X86umax, v8i16, VR128,
6842 loadv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6844 defm VPMULDQ : SS48I_binop_rm2<0x28, "vpmuldq", X86pmuldq, v2i64, v4i32,
6845 VR128, loadv2i64, i128mem,
6846 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V;
6849 let Predicates = [HasAVX2, NoVLX] in {
6850 let isCommutable = 0 in
6851 defm VPMINSBY : SS48I_binop_rm<0x38, "vpminsb", X86smin, v32i8, VR256,
6852 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6854 defm VPMINSDY : SS48I_binop_rm<0x39, "vpminsd", X86smin, v8i32, VR256,
6855 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6857 defm VPMINUDY : SS48I_binop_rm<0x3B, "vpminud", X86umin, v8i32, VR256,
6858 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6860 defm VPMINUWY : SS48I_binop_rm<0x3A, "vpminuw", X86umin, v16i16, VR256,
6861 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6863 defm VPMAXSBY : SS48I_binop_rm<0x3C, "vpmaxsb", X86smax, v32i8, VR256,
6864 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6866 defm VPMAXSDY : SS48I_binop_rm<0x3D, "vpmaxsd", X86smax, v8i32, VR256,
6867 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6869 defm VPMAXUDY : SS48I_binop_rm<0x3F, "vpmaxud", X86umax, v8i32, VR256,
6870 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6872 defm VPMAXUWY : SS48I_binop_rm<0x3E, "vpmaxuw", X86umax, v16i16, VR256,
6873 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6875 defm VPMULDQY : SS48I_binop_rm2<0x28, "vpmuldq", X86pmuldq, v4i64, v8i32,
6876 VR256, loadv4i64, i256mem,
6877 SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V, VEX_L;
6880 let Constraints = "$src1 = $dst" in {
6881 let isCommutable = 0 in
6882 defm PMINSB : SS48I_binop_rm<0x38, "pminsb", X86smin, v16i8, VR128,
6883 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6884 defm PMINSD : SS48I_binop_rm<0x39, "pminsd", X86smin, v4i32, VR128,
6885 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6886 defm PMINUD : SS48I_binop_rm<0x3B, "pminud", X86umin, v4i32, VR128,
6887 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6888 defm PMINUW : SS48I_binop_rm<0x3A, "pminuw", X86umin, v8i16, VR128,
6889 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6890 defm PMAXSB : SS48I_binop_rm<0x3C, "pmaxsb", X86smax, v16i8, VR128,
6891 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6892 defm PMAXSD : SS48I_binop_rm<0x3D, "pmaxsd", X86smax, v4i32, VR128,
6893 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6894 defm PMAXUD : SS48I_binop_rm<0x3F, "pmaxud", X86umax, v4i32, VR128,
6895 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6896 defm PMAXUW : SS48I_binop_rm<0x3E, "pmaxuw", X86umax, v8i16, VR128,
6897 memopv2i64, i128mem, 1, SSE_INTALU_ITINS_P>;
6898 defm PMULDQ : SS48I_binop_rm2<0x28, "pmuldq", X86pmuldq, v2i64, v4i32,
6899 VR128, memopv2i64, i128mem,
6900 SSE_INTMUL_ITINS_P, 1>;
6903 let Predicates = [HasAVX, NoVLX] in {
6904 defm VPMULLD : SS48I_binop_rm<0x40, "vpmulld", mul, v4i32, VR128,
6905 memopv2i64, i128mem, 0, SSE_PMULLD_ITINS>,
6907 defm VPCMPEQQ : SS48I_binop_rm<0x29, "vpcmpeqq", X86pcmpeq, v2i64, VR128,
6908 memopv2i64, i128mem, 0, SSE_INTALU_ITINS_P>,
6911 let Predicates = [HasAVX2] in {
6912 defm VPMULLDY : SS48I_binop_rm<0x40, "vpmulld", mul, v8i32, VR256,
6913 loadv4i64, i256mem, 0, SSE_PMULLD_ITINS>,
6915 defm VPCMPEQQY : SS48I_binop_rm<0x29, "vpcmpeqq", X86pcmpeq, v4i64, VR256,
6916 loadv4i64, i256mem, 0, SSE_INTALU_ITINS_P>,
6920 let Constraints = "$src1 = $dst" in {
6921 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, VR128,
6922 memopv2i64, i128mem, 1, SSE_PMULLD_ITINS>;
6923 defm PCMPEQQ : SS48I_binop_rm<0x29, "pcmpeqq", X86pcmpeq, v2i64, VR128,
6924 memopv2i64, i128mem, 1, SSE_INTALUQ_ITINS_P>;
6927 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
6928 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
6929 Intrinsic IntId, RegisterClass RC, PatFrag memop_frag,
6930 X86MemOperand x86memop, bit Is2Addr = 1,
6931 OpndItins itins = DEFAULT_ITINS> {
6932 let isCommutable = 1 in
6933 def rri : SS4AIi8<opc, MRMSrcReg, (outs RC:$dst),
6934 (ins RC:$src1, RC:$src2, u8imm:$src3),
6936 !strconcat(OpcodeStr,
6937 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6938 !strconcat(OpcodeStr,
6939 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6940 [(set RC:$dst, (IntId RC:$src1, RC:$src2, imm:$src3))], itins.rr>,
6941 Sched<[itins.Sched]>;
6942 def rmi : SS4AIi8<opc, MRMSrcMem, (outs RC:$dst),
6943 (ins RC:$src1, x86memop:$src2, u8imm:$src3),
6945 !strconcat(OpcodeStr,
6946 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6947 !strconcat(OpcodeStr,
6948 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6951 (bitconvert (memop_frag addr:$src2)), imm:$src3))], itins.rm>,
6952 Sched<[itins.Sched.Folded, ReadAfterLd]>;
6955 /// SS41I_binop_rmi - SSE 4.1 binary operator with 8-bit immediate
6956 multiclass SS41I_binop_rmi<bits<8> opc, string OpcodeStr, SDNode OpNode,
6957 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
6958 X86MemOperand x86memop, bit Is2Addr = 1,
6959 OpndItins itins = DEFAULT_ITINS> {
6960 let isCommutable = 1 in
6961 def rri : SS4AIi8<opc, MRMSrcReg, (outs RC:$dst),
6962 (ins RC:$src1, RC:$src2, u8imm:$src3),
6964 !strconcat(OpcodeStr,
6965 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6966 !strconcat(OpcodeStr,
6967 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6968 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2, imm:$src3)))],
6969 itins.rr>, Sched<[itins.Sched]>;
6970 def rmi : SS4AIi8<opc, MRMSrcMem, (outs RC:$dst),
6971 (ins RC:$src1, x86memop:$src2, u8imm:$src3),
6973 !strconcat(OpcodeStr,
6974 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
6975 !strconcat(OpcodeStr,
6976 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
6978 (OpVT (OpNode RC:$src1,
6979 (bitconvert (memop_frag addr:$src2)), imm:$src3)))], itins.rm>,
6980 Sched<[itins.Sched.Folded, ReadAfterLd]>;
6983 let Predicates = [HasAVX] in {
6984 let isCommutable = 0 in {
6985 defm VMPSADBW : SS41I_binop_rmi_int<0x42, "vmpsadbw", int_x86_sse41_mpsadbw,
6986 VR128, loadv2i64, i128mem, 0,
6987 DEFAULT_ITINS_MPSADSCHED>, VEX_4V;
6990 let ExeDomain = SSEPackedSingle in {
6991 defm VBLENDPS : SS41I_binop_rmi<0x0C, "vblendps", X86Blendi, v4f32,
6992 VR128, loadv4f32, f128mem, 0,
6993 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V;
6994 defm VBLENDPSY : SS41I_binop_rmi<0x0C, "vblendps", X86Blendi, v8f32,
6995 VR256, loadv8f32, f256mem, 0,
6996 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V, VEX_L;
6998 let ExeDomain = SSEPackedDouble in {
6999 defm VBLENDPD : SS41I_binop_rmi<0x0D, "vblendpd", X86Blendi, v2f64,
7000 VR128, loadv2f64, f128mem, 0,
7001 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V;
7002 defm VBLENDPDY : SS41I_binop_rmi<0x0D, "vblendpd", X86Blendi, v4f64,
7003 VR256, loadv4f64, f256mem, 0,
7004 DEFAULT_ITINS_FBLENDSCHED>, VEX_4V, VEX_L;
7006 defm VPBLENDW : SS41I_binop_rmi<0x0E, "vpblendw", X86Blendi, v8i16,
7007 VR128, loadv2i64, i128mem, 0,
7008 DEFAULT_ITINS_BLENDSCHED>, VEX_4V;
7010 let ExeDomain = SSEPackedSingle in
7011 defm VDPPS : SS41I_binop_rmi_int<0x40, "vdpps", int_x86_sse41_dpps,
7012 VR128, loadv4f32, f128mem, 0,
7013 SSE_DPPS_ITINS>, VEX_4V;
7014 let ExeDomain = SSEPackedDouble in
7015 defm VDPPD : SS41I_binop_rmi_int<0x41, "vdppd", int_x86_sse41_dppd,
7016 VR128, loadv2f64, f128mem, 0,
7017 SSE_DPPS_ITINS>, VEX_4V;
7018 let ExeDomain = SSEPackedSingle in
7019 defm VDPPSY : SS41I_binop_rmi_int<0x40, "vdpps", int_x86_avx_dp_ps_256,
7020 VR256, loadv8f32, i256mem, 0,
7021 SSE_DPPS_ITINS>, VEX_4V, VEX_L;
7024 let Predicates = [HasAVX2] in {
7025 let isCommutable = 0 in {
7026 defm VMPSADBWY : SS41I_binop_rmi_int<0x42, "vmpsadbw", int_x86_avx2_mpsadbw,
7027 VR256, loadv4i64, i256mem, 0,
7028 DEFAULT_ITINS_MPSADSCHED>, VEX_4V, VEX_L;
7030 defm VPBLENDWY : SS41I_binop_rmi<0x0E, "vpblendw", X86Blendi, v16i16,
7031 VR256, loadv4i64, i256mem, 0,
7032 DEFAULT_ITINS_BLENDSCHED>, VEX_4V, VEX_L;
7035 let Constraints = "$src1 = $dst" in {
7036 let isCommutable = 0 in {
7037 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw", int_x86_sse41_mpsadbw,
7038 VR128, memopv2i64, i128mem,
7039 1, SSE_MPSADBW_ITINS>;
7041 let ExeDomain = SSEPackedSingle in
7042 defm BLENDPS : SS41I_binop_rmi<0x0C, "blendps", X86Blendi, v4f32,
7043 VR128, memopv4f32, f128mem,
7044 1, SSE_INTALU_ITINS_FBLEND_P>;
7045 let ExeDomain = SSEPackedDouble in
7046 defm BLENDPD : SS41I_binop_rmi<0x0D, "blendpd", X86Blendi, v2f64,
7047 VR128, memopv2f64, f128mem,
7048 1, SSE_INTALU_ITINS_FBLEND_P>;
7049 defm PBLENDW : SS41I_binop_rmi<0x0E, "pblendw", X86Blendi, v8i16,
7050 VR128, memopv2i64, i128mem,
7051 1, SSE_INTALU_ITINS_BLEND_P>;
7052 let ExeDomain = SSEPackedSingle in
7053 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps", int_x86_sse41_dpps,
7054 VR128, memopv4f32, f128mem, 1,
7056 let ExeDomain = SSEPackedDouble in
7057 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd", int_x86_sse41_dppd,
7058 VR128, memopv2f64, f128mem, 1,
7062 /// SS41I_quaternary_int_avx - AVX SSE 4.1 with 4 operators
7063 multiclass SS41I_quaternary_int_avx<bits<8> opc, string OpcodeStr,
7064 RegisterClass RC, X86MemOperand x86memop,
7065 PatFrag mem_frag, Intrinsic IntId,
7066 X86FoldableSchedWrite Sched> {
7067 def rr : Ii8<opc, MRMSrcReg, (outs RC:$dst),
7068 (ins RC:$src1, RC:$src2, RC:$src3),
7069 !strconcat(OpcodeStr,
7070 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
7071 [(set RC:$dst, (IntId RC:$src1, RC:$src2, RC:$src3))],
7072 NoItinerary, SSEPackedInt>, TAPD, VEX_4V, VEX_I8IMM,
7075 def rm : Ii8<opc, MRMSrcMem, (outs RC:$dst),
7076 (ins RC:$src1, x86memop:$src2, RC:$src3),
7077 !strconcat(OpcodeStr,
7078 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
7080 (IntId RC:$src1, (bitconvert (mem_frag addr:$src2)),
7082 NoItinerary, SSEPackedInt>, TAPD, VEX_4V, VEX_I8IMM,
7083 Sched<[Sched.Folded, ReadAfterLd]>;
7086 let Predicates = [HasAVX] in {
7087 let ExeDomain = SSEPackedDouble in {
7088 defm VBLENDVPD : SS41I_quaternary_int_avx<0x4B, "vblendvpd", VR128, f128mem,
7089 loadv2f64, int_x86_sse41_blendvpd,
7091 defm VBLENDVPDY : SS41I_quaternary_int_avx<0x4B, "vblendvpd", VR256, f256mem,
7092 loadv4f64, int_x86_avx_blendv_pd_256,
7093 WriteFVarBlend>, VEX_L;
7094 } // ExeDomain = SSEPackedDouble
7095 let ExeDomain = SSEPackedSingle in {
7096 defm VBLENDVPS : SS41I_quaternary_int_avx<0x4A, "vblendvps", VR128, f128mem,
7097 loadv4f32, int_x86_sse41_blendvps,
7099 defm VBLENDVPSY : SS41I_quaternary_int_avx<0x4A, "vblendvps", VR256, f256mem,
7100 loadv8f32, int_x86_avx_blendv_ps_256,
7101 WriteFVarBlend>, VEX_L;
7102 } // ExeDomain = SSEPackedSingle
7103 defm VPBLENDVB : SS41I_quaternary_int_avx<0x4C, "vpblendvb", VR128, i128mem,
7104 loadv2i64, int_x86_sse41_pblendvb,
7108 let Predicates = [HasAVX2] in {
7109 defm VPBLENDVBY : SS41I_quaternary_int_avx<0x4C, "vpblendvb", VR256, i256mem,
7110 loadv4i64, int_x86_avx2_pblendvb,
7111 WriteVarBlend>, VEX_L;
7114 let Predicates = [HasAVX] in {
7115 def : Pat<(v16i8 (vselect (v16i8 VR128:$mask), (v16i8 VR128:$src1),
7116 (v16i8 VR128:$src2))),
7117 (VPBLENDVBrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7118 def : Pat<(v4i32 (vselect (v4i32 VR128:$mask), (v4i32 VR128:$src1),
7119 (v4i32 VR128:$src2))),
7120 (VBLENDVPSrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7121 def : Pat<(v4f32 (vselect (v4i32 VR128:$mask), (v4f32 VR128:$src1),
7122 (v4f32 VR128:$src2))),
7123 (VBLENDVPSrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7124 def : Pat<(v2i64 (vselect (v2i64 VR128:$mask), (v2i64 VR128:$src1),
7125 (v2i64 VR128:$src2))),
7126 (VBLENDVPDrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7127 def : Pat<(v2f64 (vselect (v2i64 VR128:$mask), (v2f64 VR128:$src1),
7128 (v2f64 VR128:$src2))),
7129 (VBLENDVPDrr VR128:$src2, VR128:$src1, VR128:$mask)>;
7130 def : Pat<(v8i32 (vselect (v8i32 VR256:$mask), (v8i32 VR256:$src1),
7131 (v8i32 VR256:$src2))),
7132 (VBLENDVPSYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7133 def : Pat<(v8f32 (vselect (v8i32 VR256:$mask), (v8f32 VR256:$src1),
7134 (v8f32 VR256:$src2))),
7135 (VBLENDVPSYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7136 def : Pat<(v4i64 (vselect (v4i64 VR256:$mask), (v4i64 VR256:$src1),
7137 (v4i64 VR256:$src2))),
7138 (VBLENDVPDYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7139 def : Pat<(v4f64 (vselect (v4i64 VR256:$mask), (v4f64 VR256:$src1),
7140 (v4f64 VR256:$src2))),
7141 (VBLENDVPDYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7144 let Predicates = [HasAVX2] in {
7145 def : Pat<(v32i8 (vselect (v32i8 VR256:$mask), (v32i8 VR256:$src1),
7146 (v32i8 VR256:$src2))),
7147 (VPBLENDVBYrr VR256:$src2, VR256:$src1, VR256:$mask)>;
7151 // FIXME: Prefer a movss or movsd over a blendps when optimizing for size or
7152 // on targets where they have equal performance. These were changed to use
7153 // blends because blends have better throughput on SandyBridge and Haswell, but
7154 // movs[s/d] are 1-2 byte shorter instructions.
7155 let Predicates = [UseAVX] in {
7156 let AddedComplexity = 15 in {
7157 // Move scalar to XMM zero-extended, zeroing a VR128 then do a
7158 // MOVS{S,D} to the lower bits.
7159 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
7160 (VMOVSSrr (v4f32 (V_SET0)), FR32:$src)>;
7161 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
7162 (VBLENDPSrri (v4f32 (V_SET0)), VR128:$src, (i8 1))>;
7163 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
7164 (VPBLENDWrri (v4i32 (V_SET0)), VR128:$src, (i8 3))>;
7165 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
7166 (VMOVSDrr (v2f64 (V_SET0)), FR64:$src)>;
7168 // Move low f32 and clear high bits.
7169 def : Pat<(v8f32 (X86vzmovl (v8f32 VR256:$src))),
7170 (VBLENDPSYrri (v8f32 (AVX_SET0)), VR256:$src, (i8 1))>;
7172 // Move low f64 and clear high bits.
7173 def : Pat<(v4f64 (X86vzmovl (v4f64 VR256:$src))),
7174 (VBLENDPDYrri (v4f64 (AVX_SET0)), VR256:$src, (i8 1))>;
7177 def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
7178 (v4f32 (scalar_to_vector FR32:$src)), (iPTR 0)))),
7179 (SUBREG_TO_REG (i32 0),
7180 (v4f32 (VMOVSSrr (v4f32 (V_SET0)), FR32:$src)),
7182 def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
7183 (v2f64 (scalar_to_vector FR64:$src)), (iPTR 0)))),
7184 (SUBREG_TO_REG (i64 0),
7185 (v2f64 (VMOVSDrr (v2f64 (V_SET0)), FR64:$src)),
7188 // These will incur an FP/int domain crossing penalty, but it may be the only
7189 // way without AVX2. Do not add any complexity because we may be able to match
7190 // more optimal patterns defined earlier in this file.
7191 def : Pat<(v8i32 (X86vzmovl (v8i32 VR256:$src))),
7192 (VBLENDPSYrri (v8i32 (AVX_SET0)), VR256:$src, (i8 1))>;
7193 def : Pat<(v4i64 (X86vzmovl (v4i64 VR256:$src))),
7194 (VBLENDPDYrri (v4i64 (AVX_SET0)), VR256:$src, (i8 1))>;
7197 // FIXME: Prefer a movss or movsd over a blendps when optimizing for size or
7198 // on targets where they have equal performance. These were changed to use
7199 // blends because blends have better throughput on SandyBridge and Haswell, but
7200 // movs[s/d] are 1-2 byte shorter instructions.
7201 let Predicates = [UseSSE41] in {
7202 // With SSE41 we can use blends for these patterns.
7203 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
7204 (BLENDPSrri (v4f32 (V_SET0)), VR128:$src, (i8 1))>;
7205 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
7206 (PBLENDWrri (v4i32 (V_SET0)), VR128:$src, (i8 3))>;
7207 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
7208 (BLENDPDrri (v2f64 (V_SET0)), VR128:$src, (i8 1))>;
7212 /// SS41I_ternary_int - SSE 4.1 ternary operator
7213 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
7214 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
7215 X86MemOperand x86memop, Intrinsic IntId,
7216 OpndItins itins = DEFAULT_ITINS> {
7217 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
7218 (ins VR128:$src1, VR128:$src2),
7219 !strconcat(OpcodeStr,
7220 "\t{$src2, $dst|$dst, $src2}"),
7221 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))],
7222 itins.rr>, Sched<[itins.Sched]>;
7224 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
7225 (ins VR128:$src1, x86memop:$src2),
7226 !strconcat(OpcodeStr,
7227 "\t{$src2, $dst|$dst, $src2}"),
7230 (bitconvert (mem_frag addr:$src2)), XMM0))],
7231 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>;
7235 let ExeDomain = SSEPackedDouble in
7236 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", memopv2f64, f128mem,
7237 int_x86_sse41_blendvpd,
7238 DEFAULT_ITINS_FBLENDSCHED>;
7239 let ExeDomain = SSEPackedSingle in
7240 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", memopv4f32, f128mem,
7241 int_x86_sse41_blendvps,
7242 DEFAULT_ITINS_FBLENDSCHED>;
7243 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", memopv2i64, i128mem,
7244 int_x86_sse41_pblendvb,
7245 DEFAULT_ITINS_VARBLENDSCHED>;
7247 // Aliases with the implicit xmm0 argument
7248 def : InstAlias<"blendvpd\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7249 (BLENDVPDrr0 VR128:$dst, VR128:$src2)>;
7250 def : InstAlias<"blendvpd\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7251 (BLENDVPDrm0 VR128:$dst, f128mem:$src2)>;
7252 def : InstAlias<"blendvps\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7253 (BLENDVPSrr0 VR128:$dst, VR128:$src2)>;
7254 def : InstAlias<"blendvps\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7255 (BLENDVPSrm0 VR128:$dst, f128mem:$src2)>;
7256 def : InstAlias<"pblendvb\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7257 (PBLENDVBrr0 VR128:$dst, VR128:$src2)>;
7258 def : InstAlias<"pblendvb\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7259 (PBLENDVBrm0 VR128:$dst, i128mem:$src2)>;
7261 let Predicates = [UseSSE41] in {
7262 def : Pat<(v16i8 (vselect (v16i8 XMM0), (v16i8 VR128:$src1),
7263 (v16i8 VR128:$src2))),
7264 (PBLENDVBrr0 VR128:$src2, VR128:$src1)>;
7265 def : Pat<(v4i32 (vselect (v4i32 XMM0), (v4i32 VR128:$src1),
7266 (v4i32 VR128:$src2))),
7267 (BLENDVPSrr0 VR128:$src2, VR128:$src1)>;
7268 def : Pat<(v4f32 (vselect (v4i32 XMM0), (v4f32 VR128:$src1),
7269 (v4f32 VR128:$src2))),
7270 (BLENDVPSrr0 VR128:$src2, VR128:$src1)>;
7271 def : Pat<(v2i64 (vselect (v2i64 XMM0), (v2i64 VR128:$src1),
7272 (v2i64 VR128:$src2))),
7273 (BLENDVPDrr0 VR128:$src2, VR128:$src1)>;
7274 def : Pat<(v2f64 (vselect (v2i64 XMM0), (v2f64 VR128:$src1),
7275 (v2f64 VR128:$src2))),
7276 (BLENDVPDrr0 VR128:$src2, VR128:$src1)>;
7279 let SchedRW = [WriteLoad] in {
7280 let Predicates = [HasAVX] in
7281 def VMOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
7282 "vmovntdqa\t{$src, $dst|$dst, $src}",
7283 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
7285 let Predicates = [HasAVX2] in
7286 def VMOVNTDQAYrm : SS48I<0x2A, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
7287 "vmovntdqa\t{$src, $dst|$dst, $src}",
7288 [(set VR256:$dst, (int_x86_avx2_movntdqa addr:$src))]>,
7290 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
7291 "movntdqa\t{$src, $dst|$dst, $src}",
7292 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>;
7295 //===----------------------------------------------------------------------===//
7296 // SSE4.2 - Compare Instructions
7297 //===----------------------------------------------------------------------===//
7299 /// SS42I_binop_rm - Simple SSE 4.2 binary operator
7300 multiclass SS42I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
7301 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
7302 X86MemOperand x86memop, bit Is2Addr = 1> {
7303 def rr : SS428I<opc, MRMSrcReg, (outs RC:$dst),
7304 (ins RC:$src1, RC:$src2),
7306 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7307 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7308 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2)))]>;
7309 def rm : SS428I<opc, MRMSrcMem, (outs RC:$dst),
7310 (ins RC:$src1, x86memop:$src2),
7312 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7313 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7315 (OpVT (OpNode RC:$src1, (memop_frag addr:$src2))))]>;
7318 let Predicates = [HasAVX] in
7319 defm VPCMPGTQ : SS42I_binop_rm<0x37, "vpcmpgtq", X86pcmpgt, v2i64, VR128,
7320 loadv2i64, i128mem, 0>, VEX_4V;
7322 let Predicates = [HasAVX2] in
7323 defm VPCMPGTQY : SS42I_binop_rm<0x37, "vpcmpgtq", X86pcmpgt, v4i64, VR256,
7324 loadv4i64, i256mem, 0>, VEX_4V, VEX_L;
7326 let Constraints = "$src1 = $dst" in
7327 defm PCMPGTQ : SS42I_binop_rm<0x37, "pcmpgtq", X86pcmpgt, v2i64, VR128,
7328 memopv2i64, i128mem>;
7330 //===----------------------------------------------------------------------===//
7331 // SSE4.2 - String/text Processing Instructions
7332 //===----------------------------------------------------------------------===//
7334 // Packed Compare Implicit Length Strings, Return Mask
7335 multiclass pseudo_pcmpistrm<string asm, PatFrag ld_frag> {
7336 def REG : PseudoI<(outs VR128:$dst),
7337 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7338 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
7340 def MEM : PseudoI<(outs VR128:$dst),
7341 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7342 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1,
7343 (bc_v16i8 (ld_frag addr:$src2)), imm:$src3))]>;
7346 let Defs = [EFLAGS], usesCustomInserter = 1 in {
7347 defm VPCMPISTRM128 : pseudo_pcmpistrm<"#VPCMPISTRM128", loadv2i64>,
7349 defm PCMPISTRM128 : pseudo_pcmpistrm<"#PCMPISTRM128", memopv2i64>,
7350 Requires<[UseSSE42]>;
7353 multiclass pcmpistrm_SS42AI<string asm> {
7354 def rr : SS42AI<0x62, MRMSrcReg, (outs),
7355 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7356 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7357 []>, Sched<[WritePCmpIStrM]>;
7359 def rm :SS42AI<0x62, MRMSrcMem, (outs),
7360 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7361 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7362 []>, Sched<[WritePCmpIStrMLd, ReadAfterLd]>;
7365 let Defs = [XMM0, EFLAGS], hasSideEffects = 0 in {
7366 let Predicates = [HasAVX] in
7367 defm VPCMPISTRM128 : pcmpistrm_SS42AI<"vpcmpistrm">, VEX;
7368 defm PCMPISTRM128 : pcmpistrm_SS42AI<"pcmpistrm"> ;
7371 // Packed Compare Explicit Length Strings, Return Mask
7372 multiclass pseudo_pcmpestrm<string asm, PatFrag ld_frag> {
7373 def REG : PseudoI<(outs VR128:$dst),
7374 (ins VR128:$src1, VR128:$src3, u8imm:$src5),
7375 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
7376 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>;
7377 def MEM : PseudoI<(outs VR128:$dst),
7378 (ins VR128:$src1, i128mem:$src3, u8imm:$src5),
7379 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128 VR128:$src1, EAX,
7380 (bc_v16i8 (ld_frag addr:$src3)), EDX, imm:$src5))]>;
7383 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
7384 defm VPCMPESTRM128 : pseudo_pcmpestrm<"#VPCMPESTRM128", loadv2i64>,
7386 defm PCMPESTRM128 : pseudo_pcmpestrm<"#PCMPESTRM128", memopv2i64>,
7387 Requires<[UseSSE42]>;
7390 multiclass SS42AI_pcmpestrm<string asm> {
7391 def rr : SS42AI<0x60, MRMSrcReg, (outs),
7392 (ins VR128:$src1, VR128:$src3, u8imm:$src5),
7393 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7394 []>, Sched<[WritePCmpEStrM]>;
7396 def rm : SS42AI<0x60, MRMSrcMem, (outs),
7397 (ins VR128:$src1, i128mem:$src3, u8imm:$src5),
7398 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7399 []>, Sched<[WritePCmpEStrMLd, ReadAfterLd]>;
7402 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX], hasSideEffects = 0 in {
7403 let Predicates = [HasAVX] in
7404 defm VPCMPESTRM128 : SS42AI_pcmpestrm<"vpcmpestrm">, VEX;
7405 defm PCMPESTRM128 : SS42AI_pcmpestrm<"pcmpestrm">;
7408 // Packed Compare Implicit Length Strings, Return Index
7409 multiclass pseudo_pcmpistri<string asm, PatFrag ld_frag> {
7410 def REG : PseudoI<(outs GR32:$dst),
7411 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7412 [(set GR32:$dst, EFLAGS,
7413 (X86pcmpistri VR128:$src1, VR128:$src2, imm:$src3))]>;
7414 def MEM : PseudoI<(outs GR32:$dst),
7415 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7416 [(set GR32:$dst, EFLAGS, (X86pcmpistri VR128:$src1,
7417 (bc_v16i8 (ld_frag addr:$src2)), imm:$src3))]>;
7420 let Defs = [EFLAGS], usesCustomInserter = 1 in {
7421 defm VPCMPISTRI : pseudo_pcmpistri<"#VPCMPISTRI", loadv2i64>,
7423 defm PCMPISTRI : pseudo_pcmpistri<"#PCMPISTRI", memopv2i64>,
7424 Requires<[UseSSE42]>;
7427 multiclass SS42AI_pcmpistri<string asm> {
7428 def rr : SS42AI<0x63, MRMSrcReg, (outs),
7429 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7430 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7431 []>, Sched<[WritePCmpIStrI]>;
7433 def rm : SS42AI<0x63, MRMSrcMem, (outs),
7434 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7435 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
7436 []>, Sched<[WritePCmpIStrILd, ReadAfterLd]>;
7439 let Defs = [ECX, EFLAGS], hasSideEffects = 0 in {
7440 let Predicates = [HasAVX] in
7441 defm VPCMPISTRI : SS42AI_pcmpistri<"vpcmpistri">, VEX;
7442 defm PCMPISTRI : SS42AI_pcmpistri<"pcmpistri">;
7445 // Packed Compare Explicit Length Strings, Return Index
7446 multiclass pseudo_pcmpestri<string asm, PatFrag ld_frag> {
7447 def REG : PseudoI<(outs GR32:$dst),
7448 (ins VR128:$src1, VR128:$src3, u8imm:$src5),
7449 [(set GR32:$dst, EFLAGS,
7450 (X86pcmpestri VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>;
7451 def MEM : PseudoI<(outs GR32:$dst),
7452 (ins VR128:$src1, i128mem:$src3, u8imm:$src5),
7453 [(set GR32:$dst, EFLAGS,
7454 (X86pcmpestri VR128:$src1, EAX, (bc_v16i8 (ld_frag addr:$src3)), EDX,
7458 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
7459 defm VPCMPESTRI : pseudo_pcmpestri<"#VPCMPESTRI", loadv2i64>,
7461 defm PCMPESTRI : pseudo_pcmpestri<"#PCMPESTRI", memopv2i64>,
7462 Requires<[UseSSE42]>;
7465 multiclass SS42AI_pcmpestri<string asm> {
7466 def rr : SS42AI<0x61, MRMSrcReg, (outs),
7467 (ins VR128:$src1, VR128:$src3, u8imm:$src5),
7468 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7469 []>, Sched<[WritePCmpEStrI]>;
7471 def rm : SS42AI<0x61, MRMSrcMem, (outs),
7472 (ins VR128:$src1, i128mem:$src3, u8imm:$src5),
7473 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
7474 []>, Sched<[WritePCmpEStrILd, ReadAfterLd]>;
7477 let Defs = [ECX, EFLAGS], Uses = [EAX, EDX], hasSideEffects = 0 in {
7478 let Predicates = [HasAVX] in
7479 defm VPCMPESTRI : SS42AI_pcmpestri<"vpcmpestri">, VEX;
7480 defm PCMPESTRI : SS42AI_pcmpestri<"pcmpestri">;
7483 //===----------------------------------------------------------------------===//
7484 // SSE4.2 - CRC Instructions
7485 //===----------------------------------------------------------------------===//
7487 // No CRC instructions have AVX equivalents
7489 // crc intrinsic instruction
7490 // This set of instructions are only rm, the only difference is the size
7492 class SS42I_crc32r<bits<8> opc, string asm, RegisterClass RCOut,
7493 RegisterClass RCIn, SDPatternOperator Int> :
7494 SS42FI<opc, MRMSrcReg, (outs RCOut:$dst), (ins RCOut:$src1, RCIn:$src2),
7495 !strconcat(asm, "\t{$src2, $src1|$src1, $src2}"),
7496 [(set RCOut:$dst, (Int RCOut:$src1, RCIn:$src2))], IIC_CRC32_REG>,
7499 class SS42I_crc32m<bits<8> opc, string asm, RegisterClass RCOut,
7500 X86MemOperand x86memop, SDPatternOperator Int> :
7501 SS42FI<opc, MRMSrcMem, (outs RCOut:$dst), (ins RCOut:$src1, x86memop:$src2),
7502 !strconcat(asm, "\t{$src2, $src1|$src1, $src2}"),
7503 [(set RCOut:$dst, (Int RCOut:$src1, (load addr:$src2)))],
7504 IIC_CRC32_MEM>, Sched<[WriteFAddLd, ReadAfterLd]>;
7506 let Constraints = "$src1 = $dst" in {
7507 def CRC32r32m8 : SS42I_crc32m<0xF0, "crc32{b}", GR32, i8mem,
7508 int_x86_sse42_crc32_32_8>;
7509 def CRC32r32r8 : SS42I_crc32r<0xF0, "crc32{b}", GR32, GR8,
7510 int_x86_sse42_crc32_32_8>;
7511 def CRC32r32m16 : SS42I_crc32m<0xF1, "crc32{w}", GR32, i16mem,
7512 int_x86_sse42_crc32_32_16>, OpSize16;
7513 def CRC32r32r16 : SS42I_crc32r<0xF1, "crc32{w}", GR32, GR16,
7514 int_x86_sse42_crc32_32_16>, OpSize16;
7515 def CRC32r32m32 : SS42I_crc32m<0xF1, "crc32{l}", GR32, i32mem,
7516 int_x86_sse42_crc32_32_32>, OpSize32;
7517 def CRC32r32r32 : SS42I_crc32r<0xF1, "crc32{l}", GR32, GR32,
7518 int_x86_sse42_crc32_32_32>, OpSize32;
7519 def CRC32r64m64 : SS42I_crc32m<0xF1, "crc32{q}", GR64, i64mem,
7520 int_x86_sse42_crc32_64_64>, REX_W;
7521 def CRC32r64r64 : SS42I_crc32r<0xF1, "crc32{q}", GR64, GR64,
7522 int_x86_sse42_crc32_64_64>, REX_W;
7523 let hasSideEffects = 0 in {
7525 def CRC32r64m8 : SS42I_crc32m<0xF0, "crc32{b}", GR64, i8mem,
7527 def CRC32r64r8 : SS42I_crc32r<0xF0, "crc32{b}", GR64, GR8,
7532 //===----------------------------------------------------------------------===//
7533 // SHA-NI Instructions
7534 //===----------------------------------------------------------------------===//
7536 multiclass SHAI_binop<bits<8> Opc, string OpcodeStr, Intrinsic IntId,
7538 def rr : I<Opc, MRMSrcReg, (outs VR128:$dst),
7539 (ins VR128:$src1, VR128:$src2),
7540 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7542 (set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0)),
7543 (set VR128:$dst, (IntId VR128:$src1, VR128:$src2)))]>, T8;
7545 def rm : I<Opc, MRMSrcMem, (outs VR128:$dst),
7546 (ins VR128:$src1, i128mem:$src2),
7547 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7549 (set VR128:$dst, (IntId VR128:$src1,
7550 (bc_v4i32 (memopv2i64 addr:$src2)), XMM0)),
7551 (set VR128:$dst, (IntId VR128:$src1,
7552 (bc_v4i32 (memopv2i64 addr:$src2)))))]>, T8;
7555 let Constraints = "$src1 = $dst", Predicates = [HasSHA] in {
7556 def SHA1RNDS4rri : Ii8<0xCC, MRMSrcReg, (outs VR128:$dst),
7557 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7558 "sha1rnds4\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7560 (int_x86_sha1rnds4 VR128:$src1, VR128:$src2,
7561 (i8 imm:$src3)))]>, TA;
7562 def SHA1RNDS4rmi : Ii8<0xCC, MRMSrcMem, (outs VR128:$dst),
7563 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7564 "sha1rnds4\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7566 (int_x86_sha1rnds4 VR128:$src1,
7567 (bc_v4i32 (memopv2i64 addr:$src2)),
7568 (i8 imm:$src3)))]>, TA;
7570 defm SHA1NEXTE : SHAI_binop<0xC8, "sha1nexte", int_x86_sha1nexte>;
7571 defm SHA1MSG1 : SHAI_binop<0xC9, "sha1msg1", int_x86_sha1msg1>;
7572 defm SHA1MSG2 : SHAI_binop<0xCA, "sha1msg2", int_x86_sha1msg2>;
7575 defm SHA256RNDS2 : SHAI_binop<0xCB, "sha256rnds2", int_x86_sha256rnds2, 1>;
7577 defm SHA256MSG1 : SHAI_binop<0xCC, "sha256msg1", int_x86_sha256msg1>;
7578 defm SHA256MSG2 : SHAI_binop<0xCD, "sha256msg2", int_x86_sha256msg2>;
7581 // Aliases with explicit %xmm0
7582 def : InstAlias<"sha256rnds2\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7583 (SHA256RNDS2rr VR128:$dst, VR128:$src2)>;
7584 def : InstAlias<"sha256rnds2\t{%xmm0, $src2, $dst|$dst, $src2, xmm0}",
7585 (SHA256RNDS2rm VR128:$dst, i128mem:$src2)>;
7587 //===----------------------------------------------------------------------===//
7588 // AES-NI Instructions
7589 //===----------------------------------------------------------------------===//
7591 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId128,
7592 PatFrag ld_frag, bit Is2Addr = 1> {
7593 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
7594 (ins VR128:$src1, VR128:$src2),
7596 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7597 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7598 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
7599 Sched<[WriteAESDecEnc]>;
7600 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
7601 (ins VR128:$src1, i128mem:$src2),
7603 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
7604 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
7606 (IntId128 VR128:$src1, (ld_frag addr:$src2)))]>,
7607 Sched<[WriteAESDecEncLd, ReadAfterLd]>;
7610 // Perform One Round of an AES Encryption/Decryption Flow
7611 let Predicates = [HasAVX, HasAES] in {
7612 defm VAESENC : AESI_binop_rm_int<0xDC, "vaesenc",
7613 int_x86_aesni_aesenc, loadv2i64, 0>, VEX_4V;
7614 defm VAESENCLAST : AESI_binop_rm_int<0xDD, "vaesenclast",
7615 int_x86_aesni_aesenclast, loadv2i64, 0>, VEX_4V;
7616 defm VAESDEC : AESI_binop_rm_int<0xDE, "vaesdec",
7617 int_x86_aesni_aesdec, loadv2i64, 0>, VEX_4V;
7618 defm VAESDECLAST : AESI_binop_rm_int<0xDF, "vaesdeclast",
7619 int_x86_aesni_aesdeclast, loadv2i64, 0>, VEX_4V;
7622 let Constraints = "$src1 = $dst" in {
7623 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
7624 int_x86_aesni_aesenc, memopv2i64>;
7625 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
7626 int_x86_aesni_aesenclast, memopv2i64>;
7627 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
7628 int_x86_aesni_aesdec, memopv2i64>;
7629 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
7630 int_x86_aesni_aesdeclast, memopv2i64>;
7633 // Perform the AES InvMixColumn Transformation
7634 let Predicates = [HasAVX, HasAES] in {
7635 def VAESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
7637 "vaesimc\t{$src1, $dst|$dst, $src1}",
7639 (int_x86_aesni_aesimc VR128:$src1))]>, Sched<[WriteAESIMC]>,
7641 def VAESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
7642 (ins i128mem:$src1),
7643 "vaesimc\t{$src1, $dst|$dst, $src1}",
7644 [(set VR128:$dst, (int_x86_aesni_aesimc (loadv2i64 addr:$src1)))]>,
7645 Sched<[WriteAESIMCLd]>, VEX;
7647 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
7649 "aesimc\t{$src1, $dst|$dst, $src1}",
7651 (int_x86_aesni_aesimc VR128:$src1))]>, Sched<[WriteAESIMC]>;
7652 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
7653 (ins i128mem:$src1),
7654 "aesimc\t{$src1, $dst|$dst, $src1}",
7655 [(set VR128:$dst, (int_x86_aesni_aesimc (memopv2i64 addr:$src1)))]>,
7656 Sched<[WriteAESIMCLd]>;
7658 // AES Round Key Generation Assist
7659 let Predicates = [HasAVX, HasAES] in {
7660 def VAESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
7661 (ins VR128:$src1, u8imm:$src2),
7662 "vaeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7664 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
7665 Sched<[WriteAESKeyGen]>, VEX;
7666 def VAESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
7667 (ins i128mem:$src1, u8imm:$src2),
7668 "vaeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7670 (int_x86_aesni_aeskeygenassist (loadv2i64 addr:$src1), imm:$src2))]>,
7671 Sched<[WriteAESKeyGenLd]>, VEX;
7673 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
7674 (ins VR128:$src1, u8imm:$src2),
7675 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7677 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
7678 Sched<[WriteAESKeyGen]>;
7679 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
7680 (ins i128mem:$src1, u8imm:$src2),
7681 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7683 (int_x86_aesni_aeskeygenassist (memopv2i64 addr:$src1), imm:$src2))]>,
7684 Sched<[WriteAESKeyGenLd]>;
7686 //===----------------------------------------------------------------------===//
7687 // PCLMUL Instructions
7688 //===----------------------------------------------------------------------===//
7690 // AVX carry-less Multiplication instructions
7691 let isCommutable = 1 in
7692 def VPCLMULQDQrr : AVXPCLMULIi8<0x44, MRMSrcReg, (outs VR128:$dst),
7693 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7694 "vpclmulqdq\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7696 (int_x86_pclmulqdq VR128:$src1, VR128:$src2, imm:$src3))]>,
7697 Sched<[WriteCLMul]>;
7699 def VPCLMULQDQrm : AVXPCLMULIi8<0x44, MRMSrcMem, (outs VR128:$dst),
7700 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7701 "vpclmulqdq\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7702 [(set VR128:$dst, (int_x86_pclmulqdq VR128:$src1,
7703 (loadv2i64 addr:$src2), imm:$src3))]>,
7704 Sched<[WriteCLMulLd, ReadAfterLd]>;
7706 // Carry-less Multiplication instructions
7707 let Constraints = "$src1 = $dst" in {
7708 let isCommutable = 1 in
7709 def PCLMULQDQrr : PCLMULIi8<0x44, MRMSrcReg, (outs VR128:$dst),
7710 (ins VR128:$src1, VR128:$src2, u8imm:$src3),
7711 "pclmulqdq\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7713 (int_x86_pclmulqdq VR128:$src1, VR128:$src2, imm:$src3))],
7714 IIC_SSE_PCLMULQDQ_RR>, Sched<[WriteCLMul]>;
7716 def PCLMULQDQrm : PCLMULIi8<0x44, MRMSrcMem, (outs VR128:$dst),
7717 (ins VR128:$src1, i128mem:$src2, u8imm:$src3),
7718 "pclmulqdq\t{$src3, $src2, $dst|$dst, $src2, $src3}",
7719 [(set VR128:$dst, (int_x86_pclmulqdq VR128:$src1,
7720 (memopv2i64 addr:$src2), imm:$src3))],
7721 IIC_SSE_PCLMULQDQ_RM>,
7722 Sched<[WriteCLMulLd, ReadAfterLd]>;
7723 } // Constraints = "$src1 = $dst"
7726 multiclass pclmul_alias<string asm, int immop> {
7727 def : InstAlias<!strconcat("pclmul", asm, "dq {$src, $dst|$dst, $src}"),
7728 (PCLMULQDQrr VR128:$dst, VR128:$src, immop), 0>;
7730 def : InstAlias<!strconcat("pclmul", asm, "dq {$src, $dst|$dst, $src}"),
7731 (PCLMULQDQrm VR128:$dst, i128mem:$src, immop), 0>;
7733 def : InstAlias<!strconcat("vpclmul", asm,
7734 "dq {$src2, $src1, $dst|$dst, $src1, $src2}"),
7735 (VPCLMULQDQrr VR128:$dst, VR128:$src1, VR128:$src2, immop),
7738 def : InstAlias<!strconcat("vpclmul", asm,
7739 "dq {$src2, $src1, $dst|$dst, $src1, $src2}"),
7740 (VPCLMULQDQrm VR128:$dst, VR128:$src1, i128mem:$src2, immop),
7743 defm : pclmul_alias<"hqhq", 0x11>;
7744 defm : pclmul_alias<"hqlq", 0x01>;
7745 defm : pclmul_alias<"lqhq", 0x10>;
7746 defm : pclmul_alias<"lqlq", 0x00>;
7748 //===----------------------------------------------------------------------===//
7749 // SSE4A Instructions
7750 //===----------------------------------------------------------------------===//
7752 let Predicates = [HasSSE4A] in {
7754 let Constraints = "$src = $dst" in {
7755 def EXTRQI : Ii8<0x78, MRMXr, (outs VR128:$dst),
7756 (ins VR128:$src, u8imm:$len, u8imm:$idx),
7757 "extrq\t{$idx, $len, $src|$src, $len, $idx}",
7758 [(set VR128:$dst, (int_x86_sse4a_extrqi VR128:$src, imm:$len,
7760 def EXTRQ : I<0x79, MRMSrcReg, (outs VR128:$dst),
7761 (ins VR128:$src, VR128:$mask),
7762 "extrq\t{$mask, $src|$src, $mask}",
7763 [(set VR128:$dst, (int_x86_sse4a_extrq VR128:$src,
7764 VR128:$mask))]>, PD;
7766 def INSERTQI : Ii8<0x78, MRMSrcReg, (outs VR128:$dst),
7767 (ins VR128:$src, VR128:$src2, u8imm:$len, u8imm:$idx),
7768 "insertq\t{$idx, $len, $src2, $src|$src, $src2, $len, $idx}",
7769 [(set VR128:$dst, (int_x86_sse4a_insertqi VR128:$src,
7770 VR128:$src2, imm:$len, imm:$idx))]>, XD;
7771 def INSERTQ : I<0x79, MRMSrcReg, (outs VR128:$dst),
7772 (ins VR128:$src, VR128:$mask),
7773 "insertq\t{$mask, $src|$src, $mask}",
7774 [(set VR128:$dst, (int_x86_sse4a_insertq VR128:$src,
7775 VR128:$mask))]>, XD;
7778 def MOVNTSS : I<0x2B, MRMDestMem, (outs), (ins f32mem:$dst, VR128:$src),
7779 "movntss\t{$src, $dst|$dst, $src}",
7780 [(int_x86_sse4a_movnt_ss addr:$dst, VR128:$src)]>, XS;
7782 def MOVNTSD : I<0x2B, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
7783 "movntsd\t{$src, $dst|$dst, $src}",
7784 [(int_x86_sse4a_movnt_sd addr:$dst, VR128:$src)]>, XD;
7787 //===----------------------------------------------------------------------===//
7789 //===----------------------------------------------------------------------===//
7791 //===----------------------------------------------------------------------===//
7792 // VBROADCAST - Load from memory and broadcast to all elements of the
7793 // destination operand
7795 class avx_broadcast<bits<8> opc, string OpcodeStr, RegisterClass RC,
7796 X86MemOperand x86memop, Intrinsic Int, SchedWrite Sched> :
7797 AVX8I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
7798 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
7799 [(set RC:$dst, (Int addr:$src))]>, Sched<[Sched]>, VEX;
7801 class avx_broadcast_no_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
7802 X86MemOperand x86memop, ValueType VT,
7803 PatFrag ld_frag, SchedWrite Sched> :
7804 AVX8I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
7805 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
7806 [(set RC:$dst, (VT (X86VBroadcast (ld_frag addr:$src))))]>,
7807 Sched<[Sched]>, VEX {
7811 // AVX2 adds register forms
7812 class avx2_broadcast_reg<bits<8> opc, string OpcodeStr, RegisterClass RC,
7813 Intrinsic Int, SchedWrite Sched> :
7814 AVX28I<opc, MRMSrcReg, (outs RC:$dst), (ins VR128:$src),
7815 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
7816 [(set RC:$dst, (Int VR128:$src))]>, Sched<[Sched]>, VEX;
7818 let ExeDomain = SSEPackedSingle in {
7819 def VBROADCASTSSrm : avx_broadcast_no_int<0x18, "vbroadcastss", VR128,
7820 f32mem, v4f32, loadf32, WriteLoad>;
7821 def VBROADCASTSSYrm : avx_broadcast_no_int<0x18, "vbroadcastss", VR256,
7822 f32mem, v8f32, loadf32,
7823 WriteFShuffleLd>, VEX_L;
7825 let ExeDomain = SSEPackedDouble in
7826 def VBROADCASTSDYrm : avx_broadcast_no_int<0x19, "vbroadcastsd", VR256, f64mem,
7827 v4f64, loadf64, WriteFShuffleLd>, VEX_L;
7828 def VBROADCASTF128 : avx_broadcast<0x1A, "vbroadcastf128", VR256, f128mem,
7829 int_x86_avx_vbroadcastf128_pd_256,
7830 WriteFShuffleLd>, VEX_L;
7832 let ExeDomain = SSEPackedSingle in {
7833 def VBROADCASTSSrr : avx2_broadcast_reg<0x18, "vbroadcastss", VR128,
7834 int_x86_avx2_vbroadcast_ss_ps,
7836 def VBROADCASTSSYrr : avx2_broadcast_reg<0x18, "vbroadcastss", VR256,
7837 int_x86_avx2_vbroadcast_ss_ps_256,
7838 WriteFShuffle256>, VEX_L;
7840 let ExeDomain = SSEPackedDouble in
7841 def VBROADCASTSDYrr : avx2_broadcast_reg<0x19, "vbroadcastsd", VR256,
7842 int_x86_avx2_vbroadcast_sd_pd_256,
7843 WriteFShuffle256>, VEX_L;
7845 let Predicates = [HasAVX2] in
7846 def VBROADCASTI128 : avx_broadcast_no_int<0x5A, "vbroadcasti128", VR256,
7847 i128mem, v4i64, loadv2i64,
7850 let Predicates = [HasAVX] in
7851 def : Pat<(int_x86_avx_vbroadcastf128_ps_256 addr:$src),
7852 (VBROADCASTF128 addr:$src)>;
7855 //===----------------------------------------------------------------------===//
7856 // VINSERTF128 - Insert packed floating-point values
7858 let hasSideEffects = 0, ExeDomain = SSEPackedSingle in {
7859 def VINSERTF128rr : AVXAIi8<0x18, MRMSrcReg, (outs VR256:$dst),
7860 (ins VR256:$src1, VR128:$src2, u8imm:$src3),
7861 "vinsertf128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7862 []>, Sched<[WriteFShuffle]>, VEX_4V, VEX_L;
7864 def VINSERTF128rm : AVXAIi8<0x18, MRMSrcMem, (outs VR256:$dst),
7865 (ins VR256:$src1, f128mem:$src2, u8imm:$src3),
7866 "vinsertf128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
7867 []>, Sched<[WriteFShuffleLd, ReadAfterLd]>, VEX_4V, VEX_L;
7870 let Predicates = [HasAVX] in {
7871 def : Pat<(vinsert128_insert:$ins (v8f32 VR256:$src1), (v4f32 VR128:$src2),
7873 (VINSERTF128rr VR256:$src1, VR128:$src2,
7874 (INSERT_get_vinsert128_imm VR256:$ins))>;
7875 def : Pat<(vinsert128_insert:$ins (v4f64 VR256:$src1), (v2f64 VR128:$src2),
7877 (VINSERTF128rr VR256:$src1, VR128:$src2,
7878 (INSERT_get_vinsert128_imm VR256:$ins))>;
7880 def : Pat<(vinsert128_insert:$ins (v8f32 VR256:$src1), (loadv4f32 addr:$src2),
7882 (VINSERTF128rm VR256:$src1, addr:$src2,
7883 (INSERT_get_vinsert128_imm VR256:$ins))>;
7884 def : Pat<(vinsert128_insert:$ins (v4f64 VR256:$src1), (loadv2f64 addr:$src2),
7886 (VINSERTF128rm VR256:$src1, addr:$src2,
7887 (INSERT_get_vinsert128_imm VR256:$ins))>;
7890 let Predicates = [HasAVX1Only] in {
7891 def : Pat<(vinsert128_insert:$ins (v4i64 VR256:$src1), (v2i64 VR128:$src2),
7893 (VINSERTF128rr VR256:$src1, VR128:$src2,
7894 (INSERT_get_vinsert128_imm VR256:$ins))>;
7895 def : Pat<(vinsert128_insert:$ins (v8i32 VR256:$src1), (v4i32 VR128:$src2),
7897 (VINSERTF128rr VR256:$src1, VR128:$src2,
7898 (INSERT_get_vinsert128_imm VR256:$ins))>;
7899 def : Pat<(vinsert128_insert:$ins (v32i8 VR256:$src1), (v16i8 VR128:$src2),
7901 (VINSERTF128rr VR256:$src1, VR128:$src2,
7902 (INSERT_get_vinsert128_imm VR256:$ins))>;
7903 def : Pat<(vinsert128_insert:$ins (v16i16 VR256:$src1), (v8i16 VR128:$src2),
7905 (VINSERTF128rr VR256:$src1, VR128:$src2,
7906 (INSERT_get_vinsert128_imm VR256:$ins))>;
7908 def : Pat<(vinsert128_insert:$ins (v4i64 VR256:$src1), (loadv2i64 addr:$src2),
7910 (VINSERTF128rm VR256:$src1, addr:$src2,
7911 (INSERT_get_vinsert128_imm VR256:$ins))>;
7912 def : Pat<(vinsert128_insert:$ins (v8i32 VR256:$src1),
7913 (bc_v4i32 (loadv2i64 addr:$src2)),
7915 (VINSERTF128rm VR256:$src1, addr:$src2,
7916 (INSERT_get_vinsert128_imm VR256:$ins))>;
7917 def : Pat<(vinsert128_insert:$ins (v32i8 VR256:$src1),
7918 (bc_v16i8 (loadv2i64 addr:$src2)),
7920 (VINSERTF128rm VR256:$src1, addr:$src2,
7921 (INSERT_get_vinsert128_imm VR256:$ins))>;
7922 def : Pat<(vinsert128_insert:$ins (v16i16 VR256:$src1),
7923 (bc_v8i16 (loadv2i64 addr:$src2)),
7925 (VINSERTF128rm VR256:$src1, addr:$src2,
7926 (INSERT_get_vinsert128_imm VR256:$ins))>;
7929 //===----------------------------------------------------------------------===//
7930 // VEXTRACTF128 - Extract packed floating-point values
7932 let hasSideEffects = 0, ExeDomain = SSEPackedSingle in {
7933 def VEXTRACTF128rr : AVXAIi8<0x19, MRMDestReg, (outs VR128:$dst),
7934 (ins VR256:$src1, u8imm:$src2),
7935 "vextractf128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7936 []>, Sched<[WriteFShuffle]>, VEX, VEX_L;
7938 def VEXTRACTF128mr : AVXAIi8<0x19, MRMDestMem, (outs),
7939 (ins f128mem:$dst, VR256:$src1, u8imm:$src2),
7940 "vextractf128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
7941 []>, Sched<[WriteStore]>, VEX, VEX_L;
7945 let Predicates = [HasAVX] in {
7946 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
7947 (v4f32 (VEXTRACTF128rr
7948 (v8f32 VR256:$src1),
7949 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
7950 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
7951 (v2f64 (VEXTRACTF128rr
7952 (v4f64 VR256:$src1),
7953 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
7955 def : Pat<(store (v4f32 (vextract128_extract:$ext (v8f32 VR256:$src1),
7956 (iPTR imm))), addr:$dst),
7957 (VEXTRACTF128mr addr:$dst, VR256:$src1,
7958 (EXTRACT_get_vextract128_imm VR128:$ext))>;
7959 def : Pat<(store (v2f64 (vextract128_extract:$ext (v4f64 VR256:$src1),
7960 (iPTR imm))), addr:$dst),
7961 (VEXTRACTF128mr addr:$dst, VR256:$src1,
7962 (EXTRACT_get_vextract128_imm VR128:$ext))>;
7965 let Predicates = [HasAVX1Only] in {
7966 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
7967 (v2i64 (VEXTRACTF128rr
7968 (v4i64 VR256:$src1),
7969 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
7970 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
7971 (v4i32 (VEXTRACTF128rr
7972 (v8i32 VR256:$src1),
7973 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
7974 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
7975 (v8i16 (VEXTRACTF128rr
7976 (v16i16 VR256:$src1),
7977 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
7978 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
7979 (v16i8 (VEXTRACTF128rr
7980 (v32i8 VR256:$src1),
7981 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
7983 def : Pat<(alignedstore (v2i64 (vextract128_extract:$ext (v4i64 VR256:$src1),
7984 (iPTR imm))), addr:$dst),
7985 (VEXTRACTF128mr addr:$dst, VR256:$src1,
7986 (EXTRACT_get_vextract128_imm VR128:$ext))>;
7987 def : Pat<(alignedstore (v4i32 (vextract128_extract:$ext (v8i32 VR256:$src1),
7988 (iPTR imm))), addr:$dst),
7989 (VEXTRACTF128mr addr:$dst, VR256:$src1,
7990 (EXTRACT_get_vextract128_imm VR128:$ext))>;
7991 def : Pat<(alignedstore (v8i16 (vextract128_extract:$ext (v16i16 VR256:$src1),
7992 (iPTR imm))), addr:$dst),
7993 (VEXTRACTF128mr addr:$dst, VR256:$src1,
7994 (EXTRACT_get_vextract128_imm VR128:$ext))>;
7995 def : Pat<(alignedstore (v16i8 (vextract128_extract:$ext (v32i8 VR256:$src1),
7996 (iPTR imm))), addr:$dst),
7997 (VEXTRACTF128mr addr:$dst, VR256:$src1,
7998 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8001 //===----------------------------------------------------------------------===//
8002 // VMASKMOV - Conditional SIMD Packed Loads and Stores
8004 multiclass avx_movmask_rm<bits<8> opc_rm, bits<8> opc_mr, string OpcodeStr,
8005 Intrinsic IntLd, Intrinsic IntLd256,
8006 Intrinsic IntSt, Intrinsic IntSt256> {
8007 def rm : AVX8I<opc_rm, MRMSrcMem, (outs VR128:$dst),
8008 (ins VR128:$src1, f128mem:$src2),
8009 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8010 [(set VR128:$dst, (IntLd addr:$src2, VR128:$src1))]>,
8012 def Yrm : AVX8I<opc_rm, MRMSrcMem, (outs VR256:$dst),
8013 (ins VR256:$src1, f256mem:$src2),
8014 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8015 [(set VR256:$dst, (IntLd256 addr:$src2, VR256:$src1))]>,
8017 def mr : AVX8I<opc_mr, MRMDestMem, (outs),
8018 (ins f128mem:$dst, VR128:$src1, VR128:$src2),
8019 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8020 [(IntSt addr:$dst, VR128:$src1, VR128:$src2)]>, VEX_4V;
8021 def Ymr : AVX8I<opc_mr, MRMDestMem, (outs),
8022 (ins f256mem:$dst, VR256:$src1, VR256:$src2),
8023 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8024 [(IntSt256 addr:$dst, VR256:$src1, VR256:$src2)]>, VEX_4V, VEX_L;
8027 let ExeDomain = SSEPackedSingle in
8028 defm VMASKMOVPS : avx_movmask_rm<0x2C, 0x2E, "vmaskmovps",
8029 int_x86_avx_maskload_ps,
8030 int_x86_avx_maskload_ps_256,
8031 int_x86_avx_maskstore_ps,
8032 int_x86_avx_maskstore_ps_256>;
8033 let ExeDomain = SSEPackedDouble in
8034 defm VMASKMOVPD : avx_movmask_rm<0x2D, 0x2F, "vmaskmovpd",
8035 int_x86_avx_maskload_pd,
8036 int_x86_avx_maskload_pd_256,
8037 int_x86_avx_maskstore_pd,
8038 int_x86_avx_maskstore_pd_256>;
8040 //===----------------------------------------------------------------------===//
8041 // VPERMIL - Permute Single and Double Floating-Point Values
8043 multiclass avx_permil<bits<8> opc_rm, bits<8> opc_rmi, string OpcodeStr,
8044 RegisterClass RC, X86MemOperand x86memop_f,
8045 X86MemOperand x86memop_i, PatFrag i_frag,
8046 Intrinsic IntVar, ValueType vt> {
8047 def rr : AVX8I<opc_rm, MRMSrcReg, (outs RC:$dst),
8048 (ins RC:$src1, RC:$src2),
8049 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8050 [(set RC:$dst, (IntVar RC:$src1, RC:$src2))]>, VEX_4V,
8051 Sched<[WriteFShuffle]>;
8052 def rm : AVX8I<opc_rm, MRMSrcMem, (outs RC:$dst),
8053 (ins RC:$src1, x86memop_i:$src2),
8054 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8055 [(set RC:$dst, (IntVar RC:$src1,
8056 (bitconvert (i_frag addr:$src2))))]>, VEX_4V,
8057 Sched<[WriteFShuffleLd, ReadAfterLd]>;
8059 def ri : AVXAIi8<opc_rmi, MRMSrcReg, (outs RC:$dst),
8060 (ins RC:$src1, u8imm:$src2),
8061 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8062 [(set RC:$dst, (vt (X86VPermilpi RC:$src1, (i8 imm:$src2))))]>, VEX,
8063 Sched<[WriteFShuffle]>;
8064 def mi : AVXAIi8<opc_rmi, MRMSrcMem, (outs RC:$dst),
8065 (ins x86memop_f:$src1, u8imm:$src2),
8066 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8068 (vt (X86VPermilpi (load addr:$src1), (i8 imm:$src2))))]>, VEX,
8069 Sched<[WriteFShuffleLd]>;
8072 let ExeDomain = SSEPackedSingle in {
8073 defm VPERMILPS : avx_permil<0x0C, 0x04, "vpermilps", VR128, f128mem, i128mem,
8074 loadv2i64, int_x86_avx_vpermilvar_ps, v4f32>;
8075 defm VPERMILPSY : avx_permil<0x0C, 0x04, "vpermilps", VR256, f256mem, i256mem,
8076 loadv4i64, int_x86_avx_vpermilvar_ps_256, v8f32>, VEX_L;
8078 let ExeDomain = SSEPackedDouble in {
8079 defm VPERMILPD : avx_permil<0x0D, 0x05, "vpermilpd", VR128, f128mem, i128mem,
8080 loadv2i64, int_x86_avx_vpermilvar_pd, v2f64>;
8081 defm VPERMILPDY : avx_permil<0x0D, 0x05, "vpermilpd", VR256, f256mem, i256mem,
8082 loadv4i64, int_x86_avx_vpermilvar_pd_256, v4f64>, VEX_L;
8085 let Predicates = [HasAVX] in {
8086 def : Pat<(v8f32 (X86VPermilpv VR256:$src1, (v8i32 VR256:$src2))),
8087 (VPERMILPSYrr VR256:$src1, VR256:$src2)>;
8088 def : Pat<(v8f32 (X86VPermilpv VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)))),
8089 (VPERMILPSYrm VR256:$src1, addr:$src2)>;
8090 def : Pat<(v4f64 (X86VPermilpv VR256:$src1, (v4i64 VR256:$src2))),
8091 (VPERMILPDYrr VR256:$src1, VR256:$src2)>;
8092 def : Pat<(v4f64 (X86VPermilpv VR256:$src1, (loadv4i64 addr:$src2))),
8093 (VPERMILPDYrm VR256:$src1, addr:$src2)>;
8095 def : Pat<(v8i32 (X86VPermilpi VR256:$src1, (i8 imm:$imm))),
8096 (VPERMILPSYri VR256:$src1, imm:$imm)>;
8097 def : Pat<(v4i64 (X86VPermilpi VR256:$src1, (i8 imm:$imm))),
8098 (VPERMILPDYri VR256:$src1, imm:$imm)>;
8099 def : Pat<(v8i32 (X86VPermilpi (bc_v8i32 (loadv4i64 addr:$src1)),
8101 (VPERMILPSYmi addr:$src1, imm:$imm)>;
8102 def : Pat<(v4i64 (X86VPermilpi (loadv4i64 addr:$src1), (i8 imm:$imm))),
8103 (VPERMILPDYmi addr:$src1, imm:$imm)>;
8105 def : Pat<(v4f32 (X86VPermilpv VR128:$src1, (v4i32 VR128:$src2))),
8106 (VPERMILPSrr VR128:$src1, VR128:$src2)>;
8107 def : Pat<(v4f32 (X86VPermilpv VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)))),
8108 (VPERMILPSrm VR128:$src1, addr:$src2)>;
8109 def : Pat<(v2f64 (X86VPermilpv VR128:$src1, (v2i64 VR128:$src2))),
8110 (VPERMILPDrr VR128:$src1, VR128:$src2)>;
8111 def : Pat<(v2f64 (X86VPermilpv VR128:$src1, (loadv2i64 addr:$src2))),
8112 (VPERMILPDrm VR128:$src1, addr:$src2)>;
8114 def : Pat<(v2i64 (X86VPermilpi VR128:$src1, (i8 imm:$imm))),
8115 (VPERMILPDri VR128:$src1, imm:$imm)>;
8116 def : Pat<(v2i64 (X86VPermilpi (loadv2i64 addr:$src1), (i8 imm:$imm))),
8117 (VPERMILPDmi addr:$src1, imm:$imm)>;
8120 //===----------------------------------------------------------------------===//
8121 // VPERM2F128 - Permute Floating-Point Values in 128-bit chunks
8123 let ExeDomain = SSEPackedSingle in {
8124 def VPERM2F128rr : AVXAIi8<0x06, MRMSrcReg, (outs VR256:$dst),
8125 (ins VR256:$src1, VR256:$src2, u8imm:$src3),
8126 "vperm2f128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8127 [(set VR256:$dst, (v8f32 (X86VPerm2x128 VR256:$src1, VR256:$src2,
8128 (i8 imm:$src3))))]>, VEX_4V, VEX_L,
8129 Sched<[WriteFShuffle]>;
8130 def VPERM2F128rm : AVXAIi8<0x06, MRMSrcMem, (outs VR256:$dst),
8131 (ins VR256:$src1, f256mem:$src2, u8imm:$src3),
8132 "vperm2f128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8133 [(set VR256:$dst, (X86VPerm2x128 VR256:$src1, (loadv8f32 addr:$src2),
8134 (i8 imm:$src3)))]>, VEX_4V, VEX_L,
8135 Sched<[WriteFShuffleLd, ReadAfterLd]>;
8138 let Predicates = [HasAVX] in {
8139 def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8140 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8141 def : Pat<(v4f64 (X86VPerm2x128 VR256:$src1,
8142 (loadv4f64 addr:$src2), (i8 imm:$imm))),
8143 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8146 let Predicates = [HasAVX1Only] in {
8147 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8148 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8149 def : Pat<(v4i64 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8150 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8151 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8152 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8153 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8154 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8156 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1,
8157 (bc_v8i32 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8158 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8159 def : Pat<(v4i64 (X86VPerm2x128 VR256:$src1,
8160 (loadv4i64 addr:$src2), (i8 imm:$imm))),
8161 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8162 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1,
8163 (bc_v32i8 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8164 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8165 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1,
8166 (bc_v16i16 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8167 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$imm)>;
8170 //===----------------------------------------------------------------------===//
8171 // VZERO - Zero YMM registers
8173 let Defs = [YMM0, YMM1, YMM2, YMM3, YMM4, YMM5, YMM6, YMM7,
8174 YMM8, YMM9, YMM10, YMM11, YMM12, YMM13, YMM14, YMM15] in {
8175 // Zero All YMM registers
8176 def VZEROALL : I<0x77, RawFrm, (outs), (ins), "vzeroall",
8177 [(int_x86_avx_vzeroall)]>, PS, VEX, VEX_L, Requires<[HasAVX]>;
8179 // Zero Upper bits of YMM registers
8180 def VZEROUPPER : I<0x77, RawFrm, (outs), (ins), "vzeroupper",
8181 [(int_x86_avx_vzeroupper)]>, PS, VEX, Requires<[HasAVX]>;
8184 //===----------------------------------------------------------------------===//
8185 // Half precision conversion instructions
8186 //===----------------------------------------------------------------------===//
8187 multiclass f16c_ph2ps<RegisterClass RC, X86MemOperand x86memop, Intrinsic Int> {
8188 def rr : I<0x13, MRMSrcReg, (outs RC:$dst), (ins VR128:$src),
8189 "vcvtph2ps\t{$src, $dst|$dst, $src}",
8190 [(set RC:$dst, (Int VR128:$src))]>,
8191 T8PD, VEX, Sched<[WriteCvtF2F]>;
8192 let hasSideEffects = 0, mayLoad = 1 in
8193 def rm : I<0x13, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
8194 "vcvtph2ps\t{$src, $dst|$dst, $src}", []>, T8PD, VEX,
8195 Sched<[WriteCvtF2FLd]>;
8198 multiclass f16c_ps2ph<RegisterClass RC, X86MemOperand x86memop, Intrinsic Int> {
8199 def rr : Ii8<0x1D, MRMDestReg, (outs VR128:$dst),
8200 (ins RC:$src1, i32u8imm:$src2),
8201 "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}",
8202 [(set VR128:$dst, (Int RC:$src1, imm:$src2))]>,
8203 TAPD, VEX, Sched<[WriteCvtF2F]>;
8204 let hasSideEffects = 0, mayStore = 1,
8205 SchedRW = [WriteCvtF2FLd, WriteRMW] in
8206 def mr : Ii8<0x1D, MRMDestMem, (outs),
8207 (ins x86memop:$dst, RC:$src1, i32u8imm:$src2),
8208 "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8212 let Predicates = [HasF16C] in {
8213 defm VCVTPH2PS : f16c_ph2ps<VR128, f64mem, int_x86_vcvtph2ps_128>;
8214 defm VCVTPH2PSY : f16c_ph2ps<VR256, f128mem, int_x86_vcvtph2ps_256>, VEX_L;
8215 defm VCVTPS2PH : f16c_ps2ph<VR128, f64mem, int_x86_vcvtps2ph_128>;
8216 defm VCVTPS2PHY : f16c_ps2ph<VR256, f128mem, int_x86_vcvtps2ph_256>, VEX_L;
8218 // Pattern match vcvtph2ps of a scalar i64 load.
8219 def : Pat<(int_x86_vcvtph2ps_128 (vzmovl_v2i64 addr:$src)),
8220 (VCVTPH2PSrm addr:$src)>;
8221 def : Pat<(int_x86_vcvtph2ps_128 (vzload_v2i64 addr:$src)),
8222 (VCVTPH2PSrm addr:$src)>;
8224 def : Pat<(store (f64 (vector_extract (bc_v2f64 (v8i16
8225 (int_x86_vcvtps2ph_128 VR128:$src1, i32:$src2))), (iPTR 0))),
8227 (VCVTPS2PHmr addr:$dst, VR128:$src1, imm:$src2)>;
8228 def : Pat<(store (i64 (vector_extract (bc_v2i64 (v8i16
8229 (int_x86_vcvtps2ph_128 VR128:$src1, i32:$src2))), (iPTR 0))),
8231 (VCVTPS2PHmr addr:$dst, VR128:$src1, imm:$src2)>;
8232 def : Pat<(store (v8i16 (int_x86_vcvtps2ph_256 VR256:$src1, i32:$src2)),
8234 (VCVTPS2PHYmr addr:$dst, VR256:$src1, imm:$src2)>;
8237 // Patterns for matching conversions from float to half-float and vice versa.
8238 let Predicates = [HasF16C] in {
8239 def : Pat<(fp_to_f16 FR32:$src),
8240 (i16 (EXTRACT_SUBREG (VMOVPDI2DIrr (VCVTPS2PHrr
8241 (COPY_TO_REGCLASS FR32:$src, VR128), 0)), sub_16bit))>;
8243 def : Pat<(f16_to_fp GR16:$src),
8244 (f32 (COPY_TO_REGCLASS (VCVTPH2PSrr
8245 (COPY_TO_REGCLASS (MOVSX32rr16 GR16:$src), VR128)), FR32)) >;
8247 def : Pat<(f16_to_fp (i16 (fp_to_f16 FR32:$src))),
8248 (f32 (COPY_TO_REGCLASS (VCVTPH2PSrr
8249 (VCVTPS2PHrr (COPY_TO_REGCLASS FR32:$src, VR128), 0)), FR32)) >;
8252 //===----------------------------------------------------------------------===//
8253 // AVX2 Instructions
8254 //===----------------------------------------------------------------------===//
8256 /// AVX2_binop_rmi - AVX2 binary operator with 8-bit immediate
8257 multiclass AVX2_binop_rmi<bits<8> opc, string OpcodeStr, SDNode OpNode,
8258 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
8259 X86MemOperand x86memop> {
8260 let isCommutable = 1 in
8261 def rri : AVX2AIi8<opc, MRMSrcReg, (outs RC:$dst),
8262 (ins RC:$src1, RC:$src2, u8imm:$src3),
8263 !strconcat(OpcodeStr,
8264 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
8265 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2, imm:$src3)))]>,
8266 Sched<[WriteBlend]>, VEX_4V;
8267 def rmi : AVX2AIi8<opc, MRMSrcMem, (outs RC:$dst),
8268 (ins RC:$src1, x86memop:$src2, u8imm:$src3),
8269 !strconcat(OpcodeStr,
8270 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
8272 (OpVT (OpNode RC:$src1,
8273 (bitconvert (memop_frag addr:$src2)), imm:$src3)))]>,
8274 Sched<[WriteBlendLd, ReadAfterLd]>, VEX_4V;
8277 defm VPBLENDD : AVX2_binop_rmi<0x02, "vpblendd", X86Blendi, v4i32,
8278 VR128, loadv2i64, i128mem>;
8279 defm VPBLENDDY : AVX2_binop_rmi<0x02, "vpblendd", X86Blendi, v8i32,
8280 VR256, loadv4i64, i256mem>, VEX_L;
8282 //===----------------------------------------------------------------------===//
8283 // VPBROADCAST - Load from memory and broadcast to all elements of the
8284 // destination operand
8286 multiclass avx2_broadcast<bits<8> opc, string OpcodeStr,
8287 X86MemOperand x86memop, PatFrag ld_frag,
8288 Intrinsic Int128, Intrinsic Int256> {
8289 def rr : AVX28I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
8290 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8291 [(set VR128:$dst, (Int128 VR128:$src))]>,
8292 Sched<[WriteShuffle]>, VEX;
8293 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst), (ins x86memop:$src),
8294 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8296 (Int128 (scalar_to_vector (ld_frag addr:$src))))]>,
8297 Sched<[WriteLoad]>, VEX;
8298 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
8299 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8300 [(set VR256:$dst, (Int256 VR128:$src))]>,
8301 Sched<[WriteShuffle256]>, VEX, VEX_L;
8302 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst), (ins x86memop:$src),
8303 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
8305 (Int256 (scalar_to_vector (ld_frag addr:$src))))]>,
8306 Sched<[WriteLoad]>, VEX, VEX_L;
8309 defm VPBROADCASTB : avx2_broadcast<0x78, "vpbroadcastb", i8mem, loadi8,
8310 int_x86_avx2_pbroadcastb_128,
8311 int_x86_avx2_pbroadcastb_256>;
8312 defm VPBROADCASTW : avx2_broadcast<0x79, "vpbroadcastw", i16mem, loadi16,
8313 int_x86_avx2_pbroadcastw_128,
8314 int_x86_avx2_pbroadcastw_256>;
8315 defm VPBROADCASTD : avx2_broadcast<0x58, "vpbroadcastd", i32mem, loadi32,
8316 int_x86_avx2_pbroadcastd_128,
8317 int_x86_avx2_pbroadcastd_256>;
8318 defm VPBROADCASTQ : avx2_broadcast<0x59, "vpbroadcastq", i64mem, loadi64,
8319 int_x86_avx2_pbroadcastq_128,
8320 int_x86_avx2_pbroadcastq_256>;
8322 let Predicates = [HasAVX2] in {
8323 def : Pat<(v16i8 (X86VBroadcast (loadi8 addr:$src))),
8324 (VPBROADCASTBrm addr:$src)>;
8325 def : Pat<(v32i8 (X86VBroadcast (loadi8 addr:$src))),
8326 (VPBROADCASTBYrm addr:$src)>;
8327 def : Pat<(v8i16 (X86VBroadcast (loadi16 addr:$src))),
8328 (VPBROADCASTWrm addr:$src)>;
8329 def : Pat<(v16i16 (X86VBroadcast (loadi16 addr:$src))),
8330 (VPBROADCASTWYrm addr:$src)>;
8331 def : Pat<(v4i32 (X86VBroadcast (loadi32 addr:$src))),
8332 (VPBROADCASTDrm addr:$src)>;
8333 def : Pat<(v8i32 (X86VBroadcast (loadi32 addr:$src))),
8334 (VPBROADCASTDYrm addr:$src)>;
8335 def : Pat<(v2i64 (X86VBroadcast (loadi64 addr:$src))),
8336 (VPBROADCASTQrm addr:$src)>;
8337 def : Pat<(v4i64 (X86VBroadcast (loadi64 addr:$src))),
8338 (VPBROADCASTQYrm addr:$src)>;
8340 def : Pat<(v16i8 (X86VBroadcast (v16i8 VR128:$src))),
8341 (VPBROADCASTBrr VR128:$src)>;
8342 def : Pat<(v32i8 (X86VBroadcast (v16i8 VR128:$src))),
8343 (VPBROADCASTBYrr VR128:$src)>;
8344 def : Pat<(v8i16 (X86VBroadcast (v8i16 VR128:$src))),
8345 (VPBROADCASTWrr VR128:$src)>;
8346 def : Pat<(v16i16 (X86VBroadcast (v8i16 VR128:$src))),
8347 (VPBROADCASTWYrr VR128:$src)>;
8348 def : Pat<(v4i32 (X86VBroadcast (v4i32 VR128:$src))),
8349 (VPBROADCASTDrr VR128:$src)>;
8350 def : Pat<(v8i32 (X86VBroadcast (v4i32 VR128:$src))),
8351 (VPBROADCASTDYrr VR128:$src)>;
8352 def : Pat<(v2i64 (X86VBroadcast (v2i64 VR128:$src))),
8353 (VPBROADCASTQrr VR128:$src)>;
8354 def : Pat<(v4i64 (X86VBroadcast (v2i64 VR128:$src))),
8355 (VPBROADCASTQYrr VR128:$src)>;
8356 def : Pat<(v4f32 (X86VBroadcast (v4f32 VR128:$src))),
8357 (VBROADCASTSSrr VR128:$src)>;
8358 def : Pat<(v8f32 (X86VBroadcast (v4f32 VR128:$src))),
8359 (VBROADCASTSSYrr VR128:$src)>;
8360 def : Pat<(v2f64 (X86VBroadcast (v2f64 VR128:$src))),
8361 (VPBROADCASTQrr VR128:$src)>;
8362 def : Pat<(v4f64 (X86VBroadcast (v2f64 VR128:$src))),
8363 (VBROADCASTSDYrr VR128:$src)>;
8365 // Provide aliases for broadcast from the same register class that
8366 // automatically does the extract.
8367 def : Pat<(v32i8 (X86VBroadcast (v32i8 VR256:$src))),
8368 (VPBROADCASTBYrr (v16i8 (EXTRACT_SUBREG (v32i8 VR256:$src),
8370 def : Pat<(v16i16 (X86VBroadcast (v16i16 VR256:$src))),
8371 (VPBROADCASTWYrr (v8i16 (EXTRACT_SUBREG (v16i16 VR256:$src),
8373 def : Pat<(v8i32 (X86VBroadcast (v8i32 VR256:$src))),
8374 (VPBROADCASTDYrr (v4i32 (EXTRACT_SUBREG (v8i32 VR256:$src),
8376 def : Pat<(v4i64 (X86VBroadcast (v4i64 VR256:$src))),
8377 (VPBROADCASTQYrr (v2i64 (EXTRACT_SUBREG (v4i64 VR256:$src),
8379 def : Pat<(v8f32 (X86VBroadcast (v8f32 VR256:$src))),
8380 (VBROADCASTSSYrr (v4f32 (EXTRACT_SUBREG (v8f32 VR256:$src),
8382 def : Pat<(v4f64 (X86VBroadcast (v4f64 VR256:$src))),
8383 (VBROADCASTSDYrr (v2f64 (EXTRACT_SUBREG (v4f64 VR256:$src),
8386 // Provide fallback in case the load node that is used in the patterns above
8387 // is used by additional users, which prevents the pattern selection.
8388 let AddedComplexity = 20 in {
8389 def : Pat<(v4f32 (X86VBroadcast FR32:$src)),
8390 (VBROADCASTSSrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
8391 def : Pat<(v8f32 (X86VBroadcast FR32:$src)),
8392 (VBROADCASTSSYrr (COPY_TO_REGCLASS FR32:$src, VR128))>;
8393 def : Pat<(v4f64 (X86VBroadcast FR64:$src)),
8394 (VBROADCASTSDYrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
8396 def : Pat<(v4i32 (X86VBroadcast GR32:$src)),
8397 (VBROADCASTSSrr (COPY_TO_REGCLASS GR32:$src, VR128))>;
8398 def : Pat<(v8i32 (X86VBroadcast GR32:$src)),
8399 (VBROADCASTSSYrr (COPY_TO_REGCLASS GR32:$src, VR128))>;
8400 def : Pat<(v4i64 (X86VBroadcast GR64:$src)),
8401 (VBROADCASTSDYrr (COPY_TO_REGCLASS GR64:$src, VR128))>;
8403 def : Pat<(v16i8 (X86VBroadcast GR8:$src)),
8404 (VPBROADCASTBrr (COPY_TO_REGCLASS
8405 (i32 (SUBREG_TO_REG (i32 0), GR8:$src, sub_8bit)),
8407 def : Pat<(v32i8 (X86VBroadcast GR8:$src)),
8408 (VPBROADCASTBYrr (COPY_TO_REGCLASS
8409 (i32 (SUBREG_TO_REG (i32 0), GR8:$src, sub_8bit)),
8412 def : Pat<(v8i16 (X86VBroadcast GR16:$src)),
8413 (VPBROADCASTWrr (COPY_TO_REGCLASS
8414 (i32 (SUBREG_TO_REG (i32 0), GR16:$src, sub_16bit)),
8416 def : Pat<(v16i16 (X86VBroadcast GR16:$src)),
8417 (VPBROADCASTWYrr (COPY_TO_REGCLASS
8418 (i32 (SUBREG_TO_REG (i32 0), GR16:$src, sub_16bit)),
8421 // The patterns for VPBROADCASTD are not needed because they would match
8422 // the exact same thing as VBROADCASTSS patterns.
8424 def : Pat<(v2i64 (X86VBroadcast GR64:$src)),
8425 (VPBROADCASTQrr (COPY_TO_REGCLASS GR64:$src, VR128))>;
8426 // The v4i64 pattern is not needed because VBROADCASTSDYrr already match.
8430 // AVX1 broadcast patterns
8431 let Predicates = [HasAVX1Only] in {
8432 def : Pat<(v8i32 (X86VBroadcast (loadi32 addr:$src))),
8433 (VBROADCASTSSYrm addr:$src)>;
8434 def : Pat<(v4i64 (X86VBroadcast (loadi64 addr:$src))),
8435 (VBROADCASTSDYrm addr:$src)>;
8436 def : Pat<(v4i32 (X86VBroadcast (loadi32 addr:$src))),
8437 (VBROADCASTSSrm addr:$src)>;
8440 let Predicates = [HasAVX] in {
8441 // Provide fallback in case the load node that is used in the patterns above
8442 // is used by additional users, which prevents the pattern selection.
8443 let AddedComplexity = 20 in {
8444 // 128bit broadcasts:
8445 def : Pat<(v4f32 (X86VBroadcast FR32:$src)),
8446 (VPSHUFDri (COPY_TO_REGCLASS FR32:$src, VR128), 0)>;
8447 def : Pat<(v8f32 (X86VBroadcast FR32:$src)),
8448 (VINSERTF128rr (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)),
8449 (VPSHUFDri (COPY_TO_REGCLASS FR32:$src, VR128), 0), sub_xmm),
8450 (VPSHUFDri (COPY_TO_REGCLASS FR32:$src, VR128), 0), 1)>;
8451 def : Pat<(v4f64 (X86VBroadcast FR64:$src)),
8452 (VINSERTF128rr (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)),
8453 (VPSHUFDri (COPY_TO_REGCLASS FR64:$src, VR128), 0x44), sub_xmm),
8454 (VPSHUFDri (COPY_TO_REGCLASS FR64:$src, VR128), 0x44), 1)>;
8456 def : Pat<(v4i32 (X86VBroadcast GR32:$src)),
8457 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0)>;
8458 def : Pat<(v8i32 (X86VBroadcast GR32:$src)),
8459 (VINSERTF128rr (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)),
8460 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0), sub_xmm),
8461 (VPSHUFDri (COPY_TO_REGCLASS GR32:$src, VR128), 0), 1)>;
8462 def : Pat<(v4i64 (X86VBroadcast GR64:$src)),
8463 (VINSERTF128rr (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)),
8464 (VPSHUFDri (COPY_TO_REGCLASS GR64:$src, VR128), 0x44), sub_xmm),
8465 (VPSHUFDri (COPY_TO_REGCLASS GR64:$src, VR128), 0x44), 1)>;
8468 def : Pat<(v2f64 (X86VBroadcast f64:$src)),
8469 (VMOVDDUPrr (COPY_TO_REGCLASS FR64:$src, VR128))>;
8470 def : Pat<(v2i64 (X86VBroadcast i64:$src)),
8471 (VMOVDDUPrr (COPY_TO_REGCLASS GR64:$src, VR128))>;
8474 //===----------------------------------------------------------------------===//
8475 // VPERM - Permute instructions
8478 multiclass avx2_perm<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
8479 ValueType OpVT, X86FoldableSchedWrite Sched> {
8480 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst),
8481 (ins VR256:$src1, VR256:$src2),
8482 !strconcat(OpcodeStr,
8483 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8485 (OpVT (X86VPermv VR256:$src1, VR256:$src2)))]>,
8486 Sched<[Sched]>, VEX_4V, VEX_L;
8487 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst),
8488 (ins VR256:$src1, i256mem:$src2),
8489 !strconcat(OpcodeStr,
8490 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8492 (OpVT (X86VPermv VR256:$src1,
8493 (bitconvert (mem_frag addr:$src2)))))]>,
8494 Sched<[Sched.Folded, ReadAfterLd]>, VEX_4V, VEX_L;
8497 defm VPERMD : avx2_perm<0x36, "vpermd", loadv4i64, v8i32, WriteShuffle256>;
8498 let ExeDomain = SSEPackedSingle in
8499 defm VPERMPS : avx2_perm<0x16, "vpermps", loadv8f32, v8f32, WriteFShuffle256>;
8501 multiclass avx2_perm_imm<bits<8> opc, string OpcodeStr, PatFrag mem_frag,
8502 ValueType OpVT, X86FoldableSchedWrite Sched> {
8503 def Yri : AVX2AIi8<opc, MRMSrcReg, (outs VR256:$dst),
8504 (ins VR256:$src1, u8imm:$src2),
8505 !strconcat(OpcodeStr,
8506 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8508 (OpVT (X86VPermi VR256:$src1, (i8 imm:$src2))))]>,
8509 Sched<[Sched]>, VEX, VEX_L;
8510 def Ymi : AVX2AIi8<opc, MRMSrcMem, (outs VR256:$dst),
8511 (ins i256mem:$src1, u8imm:$src2),
8512 !strconcat(OpcodeStr,
8513 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8515 (OpVT (X86VPermi (mem_frag addr:$src1),
8516 (i8 imm:$src2))))]>,
8517 Sched<[Sched.Folded, ReadAfterLd]>, VEX, VEX_L;
8520 defm VPERMQ : avx2_perm_imm<0x00, "vpermq", loadv4i64, v4i64,
8521 WriteShuffle256>, VEX_W;
8522 let ExeDomain = SSEPackedDouble in
8523 defm VPERMPD : avx2_perm_imm<0x01, "vpermpd", loadv4f64, v4f64,
8524 WriteFShuffle256>, VEX_W;
8526 //===----------------------------------------------------------------------===//
8527 // VPERM2I128 - Permute Floating-Point Values in 128-bit chunks
8529 def VPERM2I128rr : AVX2AIi8<0x46, MRMSrcReg, (outs VR256:$dst),
8530 (ins VR256:$src1, VR256:$src2, u8imm:$src3),
8531 "vperm2i128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8532 [(set VR256:$dst, (v4i64 (X86VPerm2x128 VR256:$src1, VR256:$src2,
8533 (i8 imm:$src3))))]>, Sched<[WriteShuffle256]>,
8535 def VPERM2I128rm : AVX2AIi8<0x46, MRMSrcMem, (outs VR256:$dst),
8536 (ins VR256:$src1, f256mem:$src2, u8imm:$src3),
8537 "vperm2i128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8538 [(set VR256:$dst, (X86VPerm2x128 VR256:$src1, (loadv4i64 addr:$src2),
8540 Sched<[WriteShuffle256Ld, ReadAfterLd]>, VEX_4V, VEX_L;
8542 let Predicates = [HasAVX2] in {
8543 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8544 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8545 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8546 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8547 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
8548 (VPERM2I128rr VR256:$src1, VR256:$src2, imm:$imm)>;
8550 def : Pat<(v32i8 (X86VPerm2x128 VR256:$src1, (bc_v32i8 (loadv4i64 addr:$src2)),
8552 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8553 def : Pat<(v16i16 (X86VPerm2x128 VR256:$src1,
8554 (bc_v16i16 (loadv4i64 addr:$src2)), (i8 imm:$imm))),
8555 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8556 def : Pat<(v8i32 (X86VPerm2x128 VR256:$src1, (bc_v8i32 (loadv4i64 addr:$src2)),
8558 (VPERM2I128rm VR256:$src1, addr:$src2, imm:$imm)>;
8562 //===----------------------------------------------------------------------===//
8563 // VINSERTI128 - Insert packed integer values
8565 let hasSideEffects = 0 in {
8566 def VINSERTI128rr : AVX2AIi8<0x38, MRMSrcReg, (outs VR256:$dst),
8567 (ins VR256:$src1, VR128:$src2, u8imm:$src3),
8568 "vinserti128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8569 []>, Sched<[WriteShuffle256]>, VEX_4V, VEX_L;
8571 def VINSERTI128rm : AVX2AIi8<0x38, MRMSrcMem, (outs VR256:$dst),
8572 (ins VR256:$src1, i128mem:$src2, u8imm:$src3),
8573 "vinserti128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
8574 []>, Sched<[WriteShuffle256Ld, ReadAfterLd]>, VEX_4V, VEX_L;
8577 let Predicates = [HasAVX2] in {
8578 def : Pat<(vinsert128_insert:$ins (v4i64 VR256:$src1), (v2i64 VR128:$src2),
8580 (VINSERTI128rr VR256:$src1, VR128:$src2,
8581 (INSERT_get_vinsert128_imm VR256:$ins))>;
8582 def : Pat<(vinsert128_insert:$ins (v8i32 VR256:$src1), (v4i32 VR128:$src2),
8584 (VINSERTI128rr VR256:$src1, VR128:$src2,
8585 (INSERT_get_vinsert128_imm VR256:$ins))>;
8586 def : Pat<(vinsert128_insert:$ins (v32i8 VR256:$src1), (v16i8 VR128:$src2),
8588 (VINSERTI128rr VR256:$src1, VR128:$src2,
8589 (INSERT_get_vinsert128_imm VR256:$ins))>;
8590 def : Pat<(vinsert128_insert:$ins (v16i16 VR256:$src1), (v8i16 VR128:$src2),
8592 (VINSERTI128rr VR256:$src1, VR128:$src2,
8593 (INSERT_get_vinsert128_imm VR256:$ins))>;
8595 def : Pat<(vinsert128_insert:$ins (v4i64 VR256:$src1), (loadv2i64 addr:$src2),
8597 (VINSERTI128rm VR256:$src1, addr:$src2,
8598 (INSERT_get_vinsert128_imm VR256:$ins))>;
8599 def : Pat<(vinsert128_insert:$ins (v8i32 VR256:$src1),
8600 (bc_v4i32 (loadv2i64 addr:$src2)),
8602 (VINSERTI128rm VR256:$src1, addr:$src2,
8603 (INSERT_get_vinsert128_imm VR256:$ins))>;
8604 def : Pat<(vinsert128_insert:$ins (v32i8 VR256:$src1),
8605 (bc_v16i8 (loadv2i64 addr:$src2)),
8607 (VINSERTI128rm VR256:$src1, addr:$src2,
8608 (INSERT_get_vinsert128_imm VR256:$ins))>;
8609 def : Pat<(vinsert128_insert:$ins (v16i16 VR256:$src1),
8610 (bc_v8i16 (loadv2i64 addr:$src2)),
8612 (VINSERTI128rm VR256:$src1, addr:$src2,
8613 (INSERT_get_vinsert128_imm VR256:$ins))>;
8616 //===----------------------------------------------------------------------===//
8617 // VEXTRACTI128 - Extract packed integer values
8619 def VEXTRACTI128rr : AVX2AIi8<0x39, MRMDestReg, (outs VR128:$dst),
8620 (ins VR256:$src1, u8imm:$src2),
8621 "vextracti128\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8622 Sched<[WriteShuffle256]>, VEX, VEX_L;
8623 let hasSideEffects = 0, mayStore = 1 in
8624 def VEXTRACTI128mr : AVX2AIi8<0x39, MRMDestMem, (outs),
8625 (ins i128mem:$dst, VR256:$src1, u8imm:$src2),
8626 "vextracti128\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
8627 Sched<[WriteStore]>, VEX, VEX_L;
8629 let Predicates = [HasAVX2] in {
8630 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8631 (v2i64 (VEXTRACTI128rr
8632 (v4i64 VR256:$src1),
8633 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8634 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8635 (v4i32 (VEXTRACTI128rr
8636 (v8i32 VR256:$src1),
8637 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8638 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8639 (v8i16 (VEXTRACTI128rr
8640 (v16i16 VR256:$src1),
8641 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8642 def : Pat<(vextract128_extract:$ext VR256:$src1, (iPTR imm)),
8643 (v16i8 (VEXTRACTI128rr
8644 (v32i8 VR256:$src1),
8645 (EXTRACT_get_vextract128_imm VR128:$ext)))>;
8647 def : Pat<(store (v2i64 (vextract128_extract:$ext (v4i64 VR256:$src1),
8648 (iPTR imm))), addr:$dst),
8649 (VEXTRACTI128mr addr:$dst, VR256:$src1,
8650 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8651 def : Pat<(store (v4i32 (vextract128_extract:$ext (v8i32 VR256:$src1),
8652 (iPTR imm))), addr:$dst),
8653 (VEXTRACTI128mr addr:$dst, VR256:$src1,
8654 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8655 def : Pat<(store (v8i16 (vextract128_extract:$ext (v16i16 VR256:$src1),
8656 (iPTR imm))), addr:$dst),
8657 (VEXTRACTI128mr addr:$dst, VR256:$src1,
8658 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8659 def : Pat<(store (v16i8 (vextract128_extract:$ext (v32i8 VR256:$src1),
8660 (iPTR imm))), addr:$dst),
8661 (VEXTRACTI128mr addr:$dst, VR256:$src1,
8662 (EXTRACT_get_vextract128_imm VR128:$ext))>;
8665 //===----------------------------------------------------------------------===//
8666 // VPMASKMOV - Conditional SIMD Integer Packed Loads and Stores
8668 multiclass avx2_pmovmask<string OpcodeStr,
8669 Intrinsic IntLd128, Intrinsic IntLd256,
8670 Intrinsic IntSt128, Intrinsic IntSt256> {
8671 def rm : AVX28I<0x8c, MRMSrcMem, (outs VR128:$dst),
8672 (ins VR128:$src1, i128mem:$src2),
8673 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8674 [(set VR128:$dst, (IntLd128 addr:$src2, VR128:$src1))]>, VEX_4V;
8675 def Yrm : AVX28I<0x8c, MRMSrcMem, (outs VR256:$dst),
8676 (ins VR256:$src1, i256mem:$src2),
8677 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8678 [(set VR256:$dst, (IntLd256 addr:$src2, VR256:$src1))]>,
8680 def mr : AVX28I<0x8e, MRMDestMem, (outs),
8681 (ins i128mem:$dst, VR128:$src1, VR128:$src2),
8682 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8683 [(IntSt128 addr:$dst, VR128:$src1, VR128:$src2)]>, VEX_4V;
8684 def Ymr : AVX28I<0x8e, MRMDestMem, (outs),
8685 (ins i256mem:$dst, VR256:$src1, VR256:$src2),
8686 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8687 [(IntSt256 addr:$dst, VR256:$src1, VR256:$src2)]>, VEX_4V, VEX_L;
8690 defm VPMASKMOVD : avx2_pmovmask<"vpmaskmovd",
8691 int_x86_avx2_maskload_d,
8692 int_x86_avx2_maskload_d_256,
8693 int_x86_avx2_maskstore_d,
8694 int_x86_avx2_maskstore_d_256>;
8695 defm VPMASKMOVQ : avx2_pmovmask<"vpmaskmovq",
8696 int_x86_avx2_maskload_q,
8697 int_x86_avx2_maskload_q_256,
8698 int_x86_avx2_maskstore_q,
8699 int_x86_avx2_maskstore_q_256>, VEX_W;
8701 def: Pat<(masked_store addr:$ptr, (v8i32 VR256:$mask), (v8f32 VR256:$src)),
8702 (VMASKMOVPSYmr addr:$ptr, VR256:$mask, VR256:$src)>;
8704 def: Pat<(masked_store addr:$ptr, (v8i32 VR256:$mask), (v8i32 VR256:$src)),
8705 (VPMASKMOVDYmr addr:$ptr, VR256:$mask, VR256:$src)>;
8707 def: Pat<(masked_store addr:$ptr, (v4i32 VR128:$mask), (v4f32 VR128:$src)),
8708 (VMASKMOVPSmr addr:$ptr, VR128:$mask, VR128:$src)>;
8710 def: Pat<(masked_store addr:$ptr, (v4i32 VR128:$mask), (v4i32 VR128:$src)),
8711 (VPMASKMOVDmr addr:$ptr, VR128:$mask, VR128:$src)>;
8713 def: Pat<(v8f32 (masked_load addr:$ptr, (v8i32 VR256:$mask), undef)),
8714 (VMASKMOVPSYrm VR256:$mask, addr:$ptr)>;
8716 def: Pat<(v8f32 (masked_load addr:$ptr, (v8i32 VR256:$mask),
8717 (bc_v8f32 (v8i32 immAllZerosV)))),
8718 (VMASKMOVPSYrm VR256:$mask, addr:$ptr)>;
8720 def: Pat<(v8f32 (masked_load addr:$ptr, (v8i32 VR256:$mask), (v8f32 VR256:$src0))),
8721 (VBLENDVPSYrr VR256:$src0, (VMASKMOVPSYrm VR256:$mask, addr:$ptr),
8724 def: Pat<(v8i32 (masked_load addr:$ptr, (v8i32 VR256:$mask), undef)),
8725 (VPMASKMOVDYrm VR256:$mask, addr:$ptr)>;
8727 def: Pat<(v8i32 (masked_load addr:$ptr, (v8i32 VR256:$mask), (v8i32 immAllZerosV))),
8728 (VPMASKMOVDYrm VR256:$mask, addr:$ptr)>;
8730 def: Pat<(v8i32 (masked_load addr:$ptr, (v8i32 VR256:$mask), (v8i32 VR256:$src0))),
8731 (VBLENDVPSYrr VR256:$src0, (VPMASKMOVDYrm VR256:$mask, addr:$ptr),
8734 def: Pat<(v4f32 (masked_load addr:$ptr, (v4i32 VR128:$mask), undef)),
8735 (VMASKMOVPSrm VR128:$mask, addr:$ptr)>;
8737 def: Pat<(v4f32 (masked_load addr:$ptr, (v4i32 VR128:$mask),
8738 (bc_v4f32 (v4i32 immAllZerosV)))),
8739 (VMASKMOVPSrm VR128:$mask, addr:$ptr)>;
8741 def: Pat<(v4f32 (masked_load addr:$ptr, (v4i32 VR128:$mask), (v4f32 VR128:$src0))),
8742 (VBLENDVPSrr VR128:$src0, (VMASKMOVPSrm VR128:$mask, addr:$ptr),
8745 def: Pat<(v4i32 (masked_load addr:$ptr, (v4i32 VR128:$mask), undef)),
8746 (VPMASKMOVDrm VR128:$mask, addr:$ptr)>;
8748 def: Pat<(v4i32 (masked_load addr:$ptr, (v4i32 VR128:$mask), (v4i32 immAllZerosV))),
8749 (VPMASKMOVDrm VR128:$mask, addr:$ptr)>;
8751 def: Pat<(v4i32 (masked_load addr:$ptr, (v4i32 VR128:$mask), (v4i32 VR128:$src0))),
8752 (VBLENDVPSrr VR128:$src0, (VPMASKMOVDrm VR128:$mask, addr:$ptr),
8755 def: Pat<(masked_store addr:$ptr, (v4i64 VR256:$mask), (v4f64 VR256:$src)),
8756 (VMASKMOVPDYmr addr:$ptr, VR256:$mask, VR256:$src)>;
8758 def: Pat<(masked_store addr:$ptr, (v4i64 VR256:$mask), (v4i64 VR256:$src)),
8759 (VPMASKMOVQYmr addr:$ptr, VR256:$mask, VR256:$src)>;
8761 def: Pat<(v4f64 (masked_load addr:$ptr, (v4i64 VR256:$mask), undef)),
8762 (VMASKMOVPDYrm VR256:$mask, addr:$ptr)>;
8764 def: Pat<(v4f64 (masked_load addr:$ptr, (v4i64 VR256:$mask),
8765 (v4f64 immAllZerosV))),
8766 (VMASKMOVPDYrm VR256:$mask, addr:$ptr)>;
8768 def: Pat<(v4f64 (masked_load addr:$ptr, (v4i64 VR256:$mask), (v4f64 VR256:$src0))),
8769 (VBLENDVPDYrr VR256:$src0, (VMASKMOVPDYrm VR256:$mask, addr:$ptr),
8772 def: Pat<(v4i64 (masked_load addr:$ptr, (v4i64 VR256:$mask), undef)),
8773 (VPMASKMOVQYrm VR256:$mask, addr:$ptr)>;
8775 def: Pat<(v4i64 (masked_load addr:$ptr, (v4i64 VR256:$mask),
8776 (bc_v4i64 (v8i32 immAllZerosV)))),
8777 (VPMASKMOVQYrm VR256:$mask, addr:$ptr)>;
8779 def: Pat<(v4i64 (masked_load addr:$ptr, (v4i64 VR256:$mask), (v4i64 VR256:$src0))),
8780 (VBLENDVPDYrr VR256:$src0, (VPMASKMOVQYrm VR256:$mask, addr:$ptr),
8783 def: Pat<(masked_store addr:$ptr, (v2i64 VR128:$mask), (v2f64 VR128:$src)),
8784 (VMASKMOVPDmr addr:$ptr, VR128:$mask, VR128:$src)>;
8786 def: Pat<(masked_store addr:$ptr, (v2i64 VR128:$mask), (v2i64 VR128:$src)),
8787 (VPMASKMOVQmr addr:$ptr, VR128:$mask, VR128:$src)>;
8789 def: Pat<(v2f64 (masked_load addr:$ptr, (v2i64 VR128:$mask), undef)),
8790 (VMASKMOVPDrm VR128:$mask, addr:$ptr)>;
8792 def: Pat<(v2f64 (masked_load addr:$ptr, (v2i64 VR128:$mask),
8793 (v2f64 immAllZerosV))),
8794 (VMASKMOVPDrm VR128:$mask, addr:$ptr)>;
8796 def: Pat<(v2f64 (masked_load addr:$ptr, (v2i64 VR128:$mask), (v2f64 VR128:$src0))),
8797 (VBLENDVPDrr VR128:$src0, (VMASKMOVPDrm VR128:$mask, addr:$ptr),
8800 def: Pat<(v2i64 (masked_load addr:$ptr, (v2i64 VR128:$mask), undef)),
8801 (VPMASKMOVQrm VR128:$mask, addr:$ptr)>;
8803 def: Pat<(v2i64 (masked_load addr:$ptr, (v2i64 VR128:$mask),
8804 (bc_v2i64 (v4i32 immAllZerosV)))),
8805 (VPMASKMOVQrm VR128:$mask, addr:$ptr)>;
8807 def: Pat<(v2i64 (masked_load addr:$ptr, (v2i64 VR128:$mask), (v2i64 VR128:$src0))),
8808 (VBLENDVPDrr VR128:$src0, (VPMASKMOVQrm VR128:$mask, addr:$ptr),
8811 //===----------------------------------------------------------------------===//
8812 // Variable Bit Shifts
8814 multiclass avx2_var_shift<bits<8> opc, string OpcodeStr, SDNode OpNode,
8815 ValueType vt128, ValueType vt256> {
8816 def rr : AVX28I<opc, MRMSrcReg, (outs VR128:$dst),
8817 (ins VR128:$src1, VR128:$src2),
8818 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8820 (vt128 (OpNode VR128:$src1, (vt128 VR128:$src2))))]>,
8821 VEX_4V, Sched<[WriteVarVecShift]>;
8822 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst),
8823 (ins VR128:$src1, i128mem:$src2),
8824 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8826 (vt128 (OpNode VR128:$src1,
8827 (vt128 (bitconvert (loadv2i64 addr:$src2))))))]>,
8828 VEX_4V, Sched<[WriteVarVecShiftLd, ReadAfterLd]>;
8829 def Yrr : AVX28I<opc, MRMSrcReg, (outs VR256:$dst),
8830 (ins VR256:$src1, VR256:$src2),
8831 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8833 (vt256 (OpNode VR256:$src1, (vt256 VR256:$src2))))]>,
8834 VEX_4V, VEX_L, Sched<[WriteVarVecShift]>;
8835 def Yrm : AVX28I<opc, MRMSrcMem, (outs VR256:$dst),
8836 (ins VR256:$src1, i256mem:$src2),
8837 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
8839 (vt256 (OpNode VR256:$src1,
8840 (vt256 (bitconvert (loadv4i64 addr:$src2))))))]>,
8841 VEX_4V, VEX_L, Sched<[WriteVarVecShiftLd, ReadAfterLd]>;
8844 defm VPSLLVD : avx2_var_shift<0x47, "vpsllvd", shl, v4i32, v8i32>;
8845 defm VPSLLVQ : avx2_var_shift<0x47, "vpsllvq", shl, v2i64, v4i64>, VEX_W;
8846 defm VPSRLVD : avx2_var_shift<0x45, "vpsrlvd", srl, v4i32, v8i32>;
8847 defm VPSRLVQ : avx2_var_shift<0x45, "vpsrlvq", srl, v2i64, v4i64>, VEX_W;
8848 defm VPSRAVD : avx2_var_shift<0x46, "vpsravd", sra, v4i32, v8i32>;
8850 //===----------------------------------------------------------------------===//
8851 // VGATHER - GATHER Operations
8852 multiclass avx2_gather<bits<8> opc, string OpcodeStr, RegisterClass RC256,
8853 X86MemOperand memop128, X86MemOperand memop256> {
8854 def rm : AVX28I<opc, MRMSrcMem, (outs VR128:$dst, VR128:$mask_wb),
8855 (ins VR128:$src1, memop128:$src2, VR128:$mask),
8856 !strconcat(OpcodeStr,
8857 "\t{$mask, $src2, $dst|$dst, $src2, $mask}"),
8859 def Yrm : AVX28I<opc, MRMSrcMem, (outs RC256:$dst, RC256:$mask_wb),
8860 (ins RC256:$src1, memop256:$src2, RC256:$mask),
8861 !strconcat(OpcodeStr,
8862 "\t{$mask, $src2, $dst|$dst, $src2, $mask}"),
8863 []>, VEX_4VOp3, VEX_L;
8866 let mayLoad = 1, Constraints
8867 = "@earlyclobber $dst,@earlyclobber $mask_wb, $src1 = $dst, $mask = $mask_wb"
8869 defm VPGATHERDQ : avx2_gather<0x90, "vpgatherdq", VR256, vx64mem, vx64mem>, VEX_W;
8870 defm VPGATHERQQ : avx2_gather<0x91, "vpgatherqq", VR256, vx64mem, vy64mem>, VEX_W;
8871 defm VPGATHERDD : avx2_gather<0x90, "vpgatherdd", VR256, vx32mem, vy32mem>;
8872 defm VPGATHERQD : avx2_gather<0x91, "vpgatherqd", VR128, vx32mem, vy32mem>;
8874 let ExeDomain = SSEPackedDouble in {
8875 defm VGATHERDPD : avx2_gather<0x92, "vgatherdpd", VR256, vx64mem, vx64mem>, VEX_W;
8876 defm VGATHERQPD : avx2_gather<0x93, "vgatherqpd", VR256, vx64mem, vy64mem>, VEX_W;
8879 let ExeDomain = SSEPackedSingle in {
8880 defm VGATHERDPS : avx2_gather<0x92, "vgatherdps", VR256, vx32mem, vy32mem>;
8881 defm VGATHERQPS : avx2_gather<0x93, "vgatherqps", VR128, vx32mem, vy32mem>;