1 //====- X86InstrSSE.td - Describe the X86 Instruction Set -------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the Evan Cheng and is distributed under
6 // the University of Illinois Open Source 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 //===----------------------------------------------------------------------===//
17 //===----------------------------------------------------------------------===//
18 // SSE specific DAG Nodes.
19 //===----------------------------------------------------------------------===//
21 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
22 SDTCisFP<0>, SDTCisInt<2> ]>;
24 def X86loadp : SDNode<"X86ISD::LOAD_PACK", SDTLoad, [SDNPHasChain]>;
25 def X86loadu : SDNode<"X86ISD::LOAD_UA", SDTLoad, [SDNPHasChain]>;
26 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
27 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
28 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
29 [SDNPCommutative, SDNPAssociative]>;
30 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
31 [SDNPCommutative, SDNPAssociative]>;
32 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
33 [SDNPCommutative, SDNPAssociative]>;
34 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
35 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest,
36 [SDNPHasChain, SDNPOutFlag]>;
37 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest,
38 [SDNPHasChain, SDNPOutFlag]>;
39 def X86s2vec : SDNode<"X86ISD::S2VEC", SDTypeProfile<1, 1, []>, []>;
40 def X86pextrw : SDNode<"X86ISD::PEXTRW", SDTypeProfile<1, 2, []>, []>;
41 def X86pinsrw : SDNode<"X86ISD::PINSRW", SDTypeProfile<1, 3, []>, []>;
43 //===----------------------------------------------------------------------===//
44 // SSE Complex Patterns
45 //===----------------------------------------------------------------------===//
47 // These are 'extloads' from a scalar to the low element of a vector, zeroing
48 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
50 def sse_load_f32 : ComplexPattern<v4f32, 4, "SelectScalarSSELoad", [],
52 def sse_load_f64 : ComplexPattern<v2f64, 4, "SelectScalarSSELoad", [],
55 def ssmem : Operand<v4f32> {
56 let PrintMethod = "printf32mem";
57 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
59 def sdmem : Operand<v2f64> {
60 let PrintMethod = "printf64mem";
61 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc, i32imm);
64 //===----------------------------------------------------------------------===//
65 // SSE pattern fragments
66 //===----------------------------------------------------------------------===//
68 def X86loadpf32 : PatFrag<(ops node:$ptr), (f32 (X86loadp node:$ptr))>;
69 def X86loadpf64 : PatFrag<(ops node:$ptr), (f64 (X86loadp node:$ptr))>;
71 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
72 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
73 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
75 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
76 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
77 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
78 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
79 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
80 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
82 def fp32imm0 : PatLeaf<(f32 fpimm), [{
83 return N->isExactlyValue(+0.0);
86 def PSxLDQ_imm : SDNodeXForm<imm, [{
87 // Transformation function: imm >> 3
88 return getI32Imm(N->getValue() >> 3);
91 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
93 def SHUFFLE_get_shuf_imm : SDNodeXForm<build_vector, [{
94 return getI8Imm(X86::getShuffleSHUFImmediate(N));
97 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
99 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<build_vector, [{
100 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
103 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
105 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<build_vector, [{
106 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
109 def SSE_splat_mask : PatLeaf<(build_vector), [{
110 return X86::isSplatMask(N);
111 }], SHUFFLE_get_shuf_imm>;
113 def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
114 return X86::isSplatLoMask(N);
117 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
118 return X86::isMOVHLPSMask(N);
121 def MOVHLPS_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
122 return X86::isMOVHLPS_v_undef_Mask(N);
125 def MOVHP_shuffle_mask : PatLeaf<(build_vector), [{
126 return X86::isMOVHPMask(N);
129 def MOVLP_shuffle_mask : PatLeaf<(build_vector), [{
130 return X86::isMOVLPMask(N);
133 def MOVL_shuffle_mask : PatLeaf<(build_vector), [{
134 return X86::isMOVLMask(N);
137 def MOVSHDUP_shuffle_mask : PatLeaf<(build_vector), [{
138 return X86::isMOVSHDUPMask(N);
141 def MOVSLDUP_shuffle_mask : PatLeaf<(build_vector), [{
142 return X86::isMOVSLDUPMask(N);
145 def UNPCKL_shuffle_mask : PatLeaf<(build_vector), [{
146 return X86::isUNPCKLMask(N);
149 def UNPCKH_shuffle_mask : PatLeaf<(build_vector), [{
150 return X86::isUNPCKHMask(N);
153 def UNPCKL_v_undef_shuffle_mask : PatLeaf<(build_vector), [{
154 return X86::isUNPCKL_v_undef_Mask(N);
157 def PSHUFD_shuffle_mask : PatLeaf<(build_vector), [{
158 return X86::isPSHUFDMask(N);
159 }], SHUFFLE_get_shuf_imm>;
161 def PSHUFHW_shuffle_mask : PatLeaf<(build_vector), [{
162 return X86::isPSHUFHWMask(N);
163 }], SHUFFLE_get_pshufhw_imm>;
165 def PSHUFLW_shuffle_mask : PatLeaf<(build_vector), [{
166 return X86::isPSHUFLWMask(N);
167 }], SHUFFLE_get_pshuflw_imm>;
169 def SHUFP_unary_shuffle_mask : PatLeaf<(build_vector), [{
170 return X86::isPSHUFDMask(N);
171 }], SHUFFLE_get_shuf_imm>;
173 def SHUFP_shuffle_mask : PatLeaf<(build_vector), [{
174 return X86::isSHUFPMask(N);
175 }], SHUFFLE_get_shuf_imm>;
177 def PSHUFD_binary_shuffle_mask : PatLeaf<(build_vector), [{
178 return X86::isSHUFPMask(N);
179 }], SHUFFLE_get_shuf_imm>;
181 //===----------------------------------------------------------------------===//
182 // SSE scalar FP Instructions
183 //===----------------------------------------------------------------------===//
185 // Instruction templates
186 // SSI - SSE1 instructions with XS prefix.
187 // SDI - SSE2 instructions with XD prefix.
188 // PSI - SSE1 instructions with TB prefix.
189 // PDI - SSE2 instructions with TB and OpSize prefixes.
190 // PSIi8 - SSE1 instructions with ImmT == Imm8 and TB prefix.
191 // PDIi8 - SSE2 instructions with ImmT == Imm8 and TB and OpSize prefixes.
192 // S3I - SSE3 instructions with TB and OpSize prefixes.
193 // S3SI - SSE3 instructions with XS prefix.
194 // S3DI - SSE3 instructions with XD prefix.
195 // SS38I - SSSE3 instructions with T8 and OpSize prefixes.
196 // SS3AI - SSSE3 instructions with TA and OpSize prefixes.
197 class SSI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
198 : I<o, F, ops, asm, pattern>, XS, Requires<[HasSSE1]>;
199 class SDI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
200 : I<o, F, ops, asm, pattern>, XD, Requires<[HasSSE2]>;
201 class PSI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
202 : I<o, F, ops, asm, pattern>, TB, Requires<[HasSSE1]>;
203 class PDI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
204 : I<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE2]>;
205 class PSIi8<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
206 : Ii8<o, F, ops, asm, pattern>, TB, Requires<[HasSSE1]>;
207 class PDIi8<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
208 : Ii8<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE2]>;
210 class S3SI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
211 : I<o, F, ops, asm, pattern>, XS, Requires<[HasSSE3]>;
212 class S3DI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
213 : I<o, F, ops, asm, pattern>, XD, Requires<[HasSSE3]>;
214 class S3I<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
215 : I<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE3]>;
217 class SS38I<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
218 : I<o, F, ops, asm, pattern>, T8, OpSize, Requires<[HasSSSE3]>;
219 class SS3AI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
220 : I<o, F, ops, asm, pattern>, TA, OpSize, Requires<[HasSSSE3]>;
222 //===----------------------------------------------------------------------===//
223 // Helpers for defining instructions that directly correspond to intrinsics.
225 multiclass SS_IntUnary<bits<8> o, string OpcodeStr, Intrinsic IntId> {
226 def r : SSI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
227 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
228 [(set VR128:$dst, (v4f32 (IntId VR128:$src)))]>;
229 def m : SSI<o, MRMSrcMem, (ops VR128:$dst, ssmem:$src),
230 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
231 [(set VR128:$dst, (v4f32 (IntId sse_load_f32:$src)))]>;
234 multiclass SD_IntUnary<bits<8> o, string OpcodeStr, Intrinsic IntId> {
235 def r : SDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
236 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
237 [(set VR128:$dst, (v2f64 (IntId VR128:$src)))]>;
238 def m : SDI<o, MRMSrcMem, (ops VR128:$dst, sdmem:$src),
239 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
240 [(set VR128:$dst, (v2f64 (IntId sse_load_f64:$src)))]>;
243 class PS_Intr<bits<8> o, string OpcodeStr, Intrinsic IntId>
244 : PSI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
245 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
246 [(set VR128:$dst, (IntId VR128:$src))]>;
247 class PS_Intm<bits<8> o, string OpcodeStr, Intrinsic IntId>
248 : PSI<o, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
249 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
250 [(set VR128:$dst, (IntId (load addr:$src)))]>;
251 class PD_Intr<bits<8> o, string OpcodeStr, Intrinsic IntId>
252 : PDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
253 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
254 [(set VR128:$dst, (IntId VR128:$src))]>;
255 class PD_Intm<bits<8> o, string OpcodeStr, Intrinsic IntId>
256 : PDI<o, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
257 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
258 [(set VR128:$dst, (IntId (load addr:$src)))]>;
260 class PS_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
261 : PSI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
262 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
263 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
264 class PS_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
265 : PSI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f32mem:$src2),
266 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
267 [(set VR128:$dst, (IntId VR128:$src1, (load addr:$src2)))]>;
268 class PD_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
269 : PDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
270 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
271 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
272 class PD_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
273 : PDI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
274 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
275 [(set VR128:$dst, (IntId VR128:$src1, (load addr:$src2)))]>;
277 // Some 'special' instructions
278 def IMPLICIT_DEF_FR32 : I<0, Pseudo, (ops FR32:$dst),
279 "#IMPLICIT_DEF $dst",
280 [(set FR32:$dst, (undef))]>, Requires<[HasSSE2]>;
281 def IMPLICIT_DEF_FR64 : I<0, Pseudo, (ops FR64:$dst),
282 "#IMPLICIT_DEF $dst",
283 [(set FR64:$dst, (undef))]>, Requires<[HasSSE2]>;
285 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
286 // scheduler into a branch sequence.
287 let usesCustomDAGSchedInserter = 1 in { // Expanded by the scheduler.
288 def CMOV_FR32 : I<0, Pseudo,
289 (ops FR32:$dst, FR32:$t, FR32:$f, i8imm:$cond),
290 "#CMOV_FR32 PSEUDO!",
291 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond))]>;
292 def CMOV_FR64 : I<0, Pseudo,
293 (ops FR64:$dst, FR64:$t, FR64:$f, i8imm:$cond),
294 "#CMOV_FR64 PSEUDO!",
295 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond))]>;
296 def CMOV_V4F32 : I<0, Pseudo,
297 (ops VR128:$dst, VR128:$t, VR128:$f, i8imm:$cond),
298 "#CMOV_V4F32 PSEUDO!",
300 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
301 def CMOV_V2F64 : I<0, Pseudo,
302 (ops VR128:$dst, VR128:$t, VR128:$f, i8imm:$cond),
303 "#CMOV_V2F64 PSEUDO!",
305 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
306 def CMOV_V2I64 : I<0, Pseudo,
307 (ops VR128:$dst, VR128:$t, VR128:$f, i8imm:$cond),
308 "#CMOV_V2I64 PSEUDO!",
310 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
314 def MOVSSrr : SSI<0x10, MRMSrcReg, (ops FR32:$dst, FR32:$src),
315 "movss {$src, $dst|$dst, $src}", []>;
316 def MOVSSrm : SSI<0x10, MRMSrcMem, (ops FR32:$dst, f32mem:$src),
317 "movss {$src, $dst|$dst, $src}",
318 [(set FR32:$dst, (loadf32 addr:$src))]>;
319 def MOVSDrr : SDI<0x10, MRMSrcReg, (ops FR64:$dst, FR64:$src),
320 "movsd {$src, $dst|$dst, $src}", []>;
321 def MOVSDrm : SDI<0x10, MRMSrcMem, (ops FR64:$dst, f64mem:$src),
322 "movsd {$src, $dst|$dst, $src}",
323 [(set FR64:$dst, (loadf64 addr:$src))]>;
325 def MOVSSmr : SSI<0x11, MRMDestMem, (ops f32mem:$dst, FR32:$src),
326 "movss {$src, $dst|$dst, $src}",
327 [(store FR32:$src, addr:$dst)]>;
328 def MOVSDmr : SDI<0x11, MRMDestMem, (ops f64mem:$dst, FR64:$src),
329 "movsd {$src, $dst|$dst, $src}",
330 [(store FR64:$src, addr:$dst)]>;
332 /// scalar_sse12_fp_binop_rm - Scalar SSE binops come in four basic forms:
333 /// 1. f32 vs f64 - These come in SSE1/SSE2 forms for float/doubles.
334 /// 2. rr vs rm - They include a reg+reg form and a ref+mem form.
336 /// In addition, scalar SSE ops have an intrinsic form. This form is unlike the
337 /// normal form, in that they take an entire vector (instead of a scalar) and
338 /// leave the top elements undefined. This adds another two variants of the
339 /// above permutations, giving us 8 forms for 'instruction'.
341 let isTwoAddress = 1 in {
342 multiclass scalar_sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
343 SDNode OpNode, Intrinsic F32Int,
344 Intrinsic F64Int, bit Commutable = 0> {
345 // Scalar operation, reg+reg.
346 def SSrr : SSI<opc, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
347 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
348 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
349 let isCommutable = Commutable;
351 def SDrr : SDI<opc, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
352 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
353 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
354 let isCommutable = Commutable;
356 // Scalar operation, reg+mem.
357 def SSrm : SSI<opc, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f32mem:$src2),
358 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
359 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
360 def SDrm : SDI<opc, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f64mem:$src2),
361 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
362 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
364 // Vector intrinsic operation, reg+reg.
365 def SSrr_Int : SSI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
366 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
367 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
368 let isCommutable = Commutable;
370 def SDrr_Int : SDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
371 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
372 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
373 let isCommutable = Commutable;
375 // Vector intrinsic operation, reg+mem.
376 def SSrm_Int : SSI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, ssmem:$src2),
377 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2}"),
378 [(set VR128:$dst, (F32Int VR128:$src1,
379 sse_load_f32:$src2))]>;
380 def SDrm_Int : SDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, sdmem:$src2),
381 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2}"),
382 [(set VR128:$dst, (F64Int VR128:$src1,
383 sse_load_f64:$src2))]>;
387 // Arithmetic instructions
389 defm ADD : scalar_sse12_fp_binop_rm<0x58, "add", fadd,
390 int_x86_sse_add_ss, int_x86_sse2_add_sd, 1>;
391 defm MUL : scalar_sse12_fp_binop_rm<0x59, "mul", fmul,
392 int_x86_sse_mul_ss, int_x86_sse2_mul_sd, 1>;
393 defm SUB : scalar_sse12_fp_binop_rm<0x5C, "sub", fsub,
394 int_x86_sse_sub_ss, int_x86_sse2_sub_sd>;
395 defm DIV : scalar_sse12_fp_binop_rm<0x5E, "div", fdiv,
396 int_x86_sse_div_ss, int_x86_sse2_div_sd>;
398 defm MAX : scalar_sse12_fp_binop_rm<0x5F, "max", X86fmax,
399 int_x86_sse_max_ss, int_x86_sse2_max_sd>;
400 defm MIN : scalar_sse12_fp_binop_rm<0x5D, "min", X86fmin,
401 int_x86_sse_min_ss, int_x86_sse2_min_sd>;
404 def SQRTSSr : SSI<0x51, MRMSrcReg, (ops FR32:$dst, FR32:$src),
405 "sqrtss {$src, $dst|$dst, $src}",
406 [(set FR32:$dst, (fsqrt FR32:$src))]>;
407 def SQRTSSm : SSI<0x51, MRMSrcMem, (ops FR32:$dst, f32mem:$src),
408 "sqrtss {$src, $dst|$dst, $src}",
409 [(set FR32:$dst, (fsqrt (loadf32 addr:$src)))]>;
410 def SQRTSDr : SDI<0x51, MRMSrcReg, (ops FR64:$dst, FR64:$src),
411 "sqrtsd {$src, $dst|$dst, $src}",
412 [(set FR64:$dst, (fsqrt FR64:$src))]>;
413 def SQRTSDm : SDI<0x51, MRMSrcMem, (ops FR64:$dst, f64mem:$src),
414 "sqrtsd {$src, $dst|$dst, $src}",
415 [(set FR64:$dst, (fsqrt (loadf64 addr:$src)))]>;
417 // Aliases to match intrinsics which expect XMM operand(s).
419 defm SQRTSS_Int : SS_IntUnary<0x51, "sqrtss" , int_x86_sse_sqrt_ss>;
420 defm SQRTSD_Int : SD_IntUnary<0x51, "sqrtsd" , int_x86_sse2_sqrt_sd>;
421 defm RSQRTSS_Int : SS_IntUnary<0x52, "rsqrtss", int_x86_sse_rsqrt_ss>;
422 defm RCPSS_Int : SS_IntUnary<0x53, "rcpss" , int_x86_sse_rcp_ss>;
424 // Conversion instructions
425 def CVTTSS2SIrr: SSI<0x2C, MRMSrcReg, (ops GR32:$dst, FR32:$src),
426 "cvttss2si {$src, $dst|$dst, $src}",
427 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
428 def CVTTSS2SIrm: SSI<0x2C, MRMSrcMem, (ops GR32:$dst, f32mem:$src),
429 "cvttss2si {$src, $dst|$dst, $src}",
430 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
431 def CVTTSD2SIrr: SDI<0x2C, MRMSrcReg, (ops GR32:$dst, FR64:$src),
432 "cvttsd2si {$src, $dst|$dst, $src}",
433 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
434 def CVTTSD2SIrm: SDI<0x2C, MRMSrcMem, (ops GR32:$dst, f64mem:$src),
435 "cvttsd2si {$src, $dst|$dst, $src}",
436 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
437 def CVTSD2SSrr: SDI<0x5A, MRMSrcReg, (ops FR32:$dst, FR64:$src),
438 "cvtsd2ss {$src, $dst|$dst, $src}",
439 [(set FR32:$dst, (fround FR64:$src))]>;
440 def CVTSD2SSrm: SDI<0x5A, MRMSrcMem, (ops FR32:$dst, f64mem:$src),
441 "cvtsd2ss {$src, $dst|$dst, $src}",
442 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
443 def CVTSI2SSrr: SSI<0x2A, MRMSrcReg, (ops FR32:$dst, GR32:$src),
444 "cvtsi2ss {$src, $dst|$dst, $src}",
445 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
446 def CVTSI2SSrm: SSI<0x2A, MRMSrcMem, (ops FR32:$dst, i32mem:$src),
447 "cvtsi2ss {$src, $dst|$dst, $src}",
448 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
449 def CVTSI2SDrr: SDI<0x2A, MRMSrcReg, (ops FR64:$dst, GR32:$src),
450 "cvtsi2sd {$src, $dst|$dst, $src}",
451 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
452 def CVTSI2SDrm: SDI<0x2A, MRMSrcMem, (ops FR64:$dst, i32mem:$src),
453 "cvtsi2sd {$src, $dst|$dst, $src}",
454 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
456 // SSE2 instructions with XS prefix
457 def CVTSS2SDrr: I<0x5A, MRMSrcReg, (ops FR64:$dst, FR32:$src),
458 "cvtss2sd {$src, $dst|$dst, $src}",
459 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
461 def CVTSS2SDrm: I<0x5A, MRMSrcMem, (ops FR64:$dst, f32mem:$src),
462 "cvtss2sd {$src, $dst|$dst, $src}",
463 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
466 // Match intrinsics which expect XMM operand(s).
467 def Int_CVTSS2SIrr: SSI<0x2D, MRMSrcReg, (ops GR32:$dst, VR128:$src),
468 "cvtss2si {$src, $dst|$dst, $src}",
469 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
470 def Int_CVTSS2SIrm: SSI<0x2D, MRMSrcMem, (ops GR32:$dst, f32mem:$src),
471 "cvtss2si {$src, $dst|$dst, $src}",
472 [(set GR32:$dst, (int_x86_sse_cvtss2si
473 (load addr:$src)))]>;
474 def Int_CVTSD2SIrr: SDI<0x2D, MRMSrcReg, (ops GR32:$dst, VR128:$src),
475 "cvtsd2si {$src, $dst|$dst, $src}",
476 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
477 def Int_CVTSD2SIrm: SDI<0x2D, MRMSrcMem, (ops GR32:$dst, f128mem:$src),
478 "cvtsd2si {$src, $dst|$dst, $src}",
479 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
480 (load addr:$src)))]>;
482 // Aliases for intrinsics
483 def Int_CVTTSS2SIrr: SSI<0x2C, MRMSrcReg, (ops GR32:$dst, VR128:$src),
484 "cvttss2si {$src, $dst|$dst, $src}",
485 [(set GR32:$dst, (int_x86_sse_cvttss2si VR128:$src))]>;
486 def Int_CVTTSS2SIrm: SSI<0x2C, MRMSrcMem, (ops GR32:$dst, f32mem:$src),
487 "cvttss2si {$src, $dst|$dst, $src}",
488 [(set GR32:$dst, (int_x86_sse_cvttss2si(load addr:$src)))]>;
489 def Int_CVTTSD2SIrr: SDI<0x2C, MRMSrcReg, (ops GR32:$dst, VR128:$src),
490 "cvttsd2si {$src, $dst|$dst, $src}",
491 [(set GR32:$dst, (int_x86_sse2_cvttsd2si VR128:$src))]>;
492 def Int_CVTTSD2SIrm: SDI<0x2C, MRMSrcMem, (ops GR32:$dst, f128mem:$src),
493 "cvttsd2si {$src, $dst|$dst, $src}",
494 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
495 (load addr:$src)))]>;
497 let isTwoAddress = 1 in {
498 def Int_CVTSI2SSrr: SSI<0x2A, MRMSrcReg,
499 (ops VR128:$dst, VR128:$src1, GR32:$src2),
500 "cvtsi2ss {$src2, $dst|$dst, $src2}",
501 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
503 def Int_CVTSI2SSrm: SSI<0x2A, MRMSrcMem,
504 (ops VR128:$dst, VR128:$src1, i32mem:$src2),
505 "cvtsi2ss {$src2, $dst|$dst, $src2}",
506 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
507 (loadi32 addr:$src2)))]>;
510 // Comparison instructions
511 let isTwoAddress = 1 in {
512 def CMPSSrr : SSI<0xC2, MRMSrcReg,
513 (ops FR32:$dst, FR32:$src1, FR32:$src, SSECC:$cc),
514 "cmp${cc}ss {$src, $dst|$dst, $src}",
516 def CMPSSrm : SSI<0xC2, MRMSrcMem,
517 (ops FR32:$dst, FR32:$src1, f32mem:$src, SSECC:$cc),
518 "cmp${cc}ss {$src, $dst|$dst, $src}", []>;
519 def CMPSDrr : SDI<0xC2, MRMSrcReg,
520 (ops FR64:$dst, FR64:$src1, FR64:$src, SSECC:$cc),
521 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
522 def CMPSDrm : SDI<0xC2, MRMSrcMem,
523 (ops FR64:$dst, FR64:$src1, f64mem:$src, SSECC:$cc),
524 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
527 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (ops FR32:$src1, FR32:$src2),
528 "ucomiss {$src2, $src1|$src1, $src2}",
529 [(X86cmp FR32:$src1, FR32:$src2)]>;
530 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (ops FR32:$src1, f32mem:$src2),
531 "ucomiss {$src2, $src1|$src1, $src2}",
532 [(X86cmp FR32:$src1, (loadf32 addr:$src2))]>;
533 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (ops FR64:$src1, FR64:$src2),
534 "ucomisd {$src2, $src1|$src1, $src2}",
535 [(X86cmp FR64:$src1, FR64:$src2)]>;
536 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (ops FR64:$src1, f64mem:$src2),
537 "ucomisd {$src2, $src1|$src1, $src2}",
538 [(X86cmp FR64:$src1, (loadf64 addr:$src2))]>;
540 // Aliases to match intrinsics which expect XMM operand(s).
541 let isTwoAddress = 1 in {
542 def Int_CMPSSrr : SSI<0xC2, MRMSrcReg,
543 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
544 "cmp${cc}ss {$src, $dst|$dst, $src}",
545 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
546 VR128:$src, imm:$cc))]>;
547 def Int_CMPSSrm : SSI<0xC2, MRMSrcMem,
548 (ops VR128:$dst, VR128:$src1, f32mem:$src, SSECC:$cc),
549 "cmp${cc}ss {$src, $dst|$dst, $src}",
550 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
551 (load addr:$src), imm:$cc))]>;
552 def Int_CMPSDrr : SDI<0xC2, MRMSrcReg,
553 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
554 "cmp${cc}sd {$src, $dst|$dst, $src}",
555 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
556 VR128:$src, imm:$cc))]>;
557 def Int_CMPSDrm : SDI<0xC2, MRMSrcMem,
558 (ops VR128:$dst, VR128:$src1, f64mem:$src, SSECC:$cc),
559 "cmp${cc}sd {$src, $dst|$dst, $src}",
560 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
561 (load addr:$src), imm:$cc))]>;
564 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
565 "ucomiss {$src2, $src1|$src1, $src2}",
566 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2)]>;
567 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
568 "ucomiss {$src2, $src1|$src1, $src2}",
569 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2))]>;
570 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
571 "ucomisd {$src2, $src1|$src1, $src2}",
572 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
573 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
574 "ucomisd {$src2, $src1|$src1, $src2}",
575 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2))]>;
577 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
578 "comiss {$src2, $src1|$src1, $src2}",
579 [(X86comi (v4f32 VR128:$src1), VR128:$src2)]>;
580 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
581 "comiss {$src2, $src1|$src1, $src2}",
582 [(X86comi (v4f32 VR128:$src1), (load addr:$src2))]>;
583 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
584 "comisd {$src2, $src1|$src1, $src2}",
585 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
586 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
587 "comisd {$src2, $src1|$src1, $src2}",
588 [(X86comi (v2f64 VR128:$src1), (load addr:$src2))]>;
590 // Aliases of packed instructions for scalar use. These all have names that
593 // Alias instructions that map fld0 to pxor for sse.
594 def FsFLD0SS : I<0xEF, MRMInitReg, (ops FR32:$dst),
595 "pxor $dst, $dst", [(set FR32:$dst, fp32imm0)]>,
596 Requires<[HasSSE1]>, TB, OpSize;
597 def FsFLD0SD : I<0xEF, MRMInitReg, (ops FR64:$dst),
598 "pxor $dst, $dst", [(set FR64:$dst, fp64imm0)]>,
599 Requires<[HasSSE2]>, TB, OpSize;
601 // Alias instructions to do FR32 / FR64 reg-to-reg copy using movaps / movapd.
602 // Upper bits are disregarded.
603 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (ops FR32:$dst, FR32:$src),
604 "movaps {$src, $dst|$dst, $src}", []>;
605 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (ops FR64:$dst, FR64:$src),
606 "movapd {$src, $dst|$dst, $src}", []>;
608 // Alias instructions to load FR32 / FR64 from f128mem using movaps / movapd.
609 // Upper bits are disregarded.
610 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (ops FR32:$dst, f128mem:$src),
611 "movaps {$src, $dst|$dst, $src}",
612 [(set FR32:$dst, (X86loadpf32 addr:$src))]>;
613 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (ops FR64:$dst, f128mem:$src),
614 "movapd {$src, $dst|$dst, $src}",
615 [(set FR64:$dst, (X86loadpf64 addr:$src))]>;
617 // Alias bitwise logical operations using SSE logical ops on packed FP values.
618 let isTwoAddress = 1 in {
619 let isCommutable = 1 in {
620 def FsANDPSrr : PSI<0x54, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
621 "andps {$src2, $dst|$dst, $src2}",
622 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
623 def FsANDPDrr : PDI<0x54, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
624 "andpd {$src2, $dst|$dst, $src2}",
625 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
626 def FsORPSrr : PSI<0x56, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
627 "orps {$src2, $dst|$dst, $src2}",
628 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
629 def FsORPDrr : PDI<0x56, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
630 "orpd {$src2, $dst|$dst, $src2}",
631 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
632 def FsXORPSrr : PSI<0x57, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
633 "xorps {$src2, $dst|$dst, $src2}",
634 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
635 def FsXORPDrr : PDI<0x57, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
636 "xorpd {$src2, $dst|$dst, $src2}",
637 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
639 def FsANDPSrm : PSI<0x54, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
640 "andps {$src2, $dst|$dst, $src2}",
641 [(set FR32:$dst, (X86fand FR32:$src1,
642 (X86loadpf32 addr:$src2)))]>;
643 def FsANDPDrm : PDI<0x54, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
644 "andpd {$src2, $dst|$dst, $src2}",
645 [(set FR64:$dst, (X86fand FR64:$src1,
646 (X86loadpf64 addr:$src2)))]>;
647 def FsORPSrm : PSI<0x56, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
648 "orps {$src2, $dst|$dst, $src2}",
649 [(set FR32:$dst, (X86for FR32:$src1,
650 (X86loadpf32 addr:$src2)))]>;
651 def FsORPDrm : PDI<0x56, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
652 "orpd {$src2, $dst|$dst, $src2}",
653 [(set FR64:$dst, (X86for FR64:$src1,
654 (X86loadpf64 addr:$src2)))]>;
655 def FsXORPSrm : PSI<0x57, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
656 "xorps {$src2, $dst|$dst, $src2}",
657 [(set FR32:$dst, (X86fxor FR32:$src1,
658 (X86loadpf32 addr:$src2)))]>;
659 def FsXORPDrm : PDI<0x57, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
660 "xorpd {$src2, $dst|$dst, $src2}",
661 [(set FR64:$dst, (X86fxor FR64:$src1,
662 (X86loadpf64 addr:$src2)))]>;
664 def FsANDNPSrr : PSI<0x55, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
665 "andnps {$src2, $dst|$dst, $src2}", []>;
666 def FsANDNPSrm : PSI<0x55, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
667 "andnps {$src2, $dst|$dst, $src2}", []>;
668 def FsANDNPDrr : PDI<0x55, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
669 "andnpd {$src2, $dst|$dst, $src2}", []>;
670 def FsANDNPDrm : PDI<0x55, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
671 "andnpd {$src2, $dst|$dst, $src2}", []>;
674 //===----------------------------------------------------------------------===//
675 // SSE packed FP Instructions
676 //===----------------------------------------------------------------------===//
678 // Some 'special' instructions
679 def IMPLICIT_DEF_VR128 : I<0, Pseudo, (ops VR128:$dst),
680 "#IMPLICIT_DEF $dst",
681 [(set VR128:$dst, (v4f32 (undef)))]>,
685 def MOVAPSrr : PSI<0x28, MRMSrcReg, (ops VR128:$dst, VR128:$src),
686 "movaps {$src, $dst|$dst, $src}", []>;
687 def MOVAPSrm : PSI<0x28, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
688 "movaps {$src, $dst|$dst, $src}",
689 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
690 def MOVAPDrr : PDI<0x28, MRMSrcReg, (ops VR128:$dst, VR128:$src),
691 "movapd {$src, $dst|$dst, $src}", []>;
692 def MOVAPDrm : PDI<0x28, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
693 "movapd {$src, $dst|$dst, $src}",
694 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
696 def MOVAPSmr : PSI<0x29, MRMDestMem, (ops f128mem:$dst, VR128:$src),
697 "movaps {$src, $dst|$dst, $src}",
698 [(store (v4f32 VR128:$src), addr:$dst)]>;
699 def MOVAPDmr : PDI<0x29, MRMDestMem, (ops f128mem:$dst, VR128:$src),
700 "movapd {$src, $dst|$dst, $src}",
701 [(store (v2f64 VR128:$src), addr:$dst)]>;
703 def MOVUPSrr : PSI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src),
704 "movups {$src, $dst|$dst, $src}", []>;
705 def MOVUPSrm : PSI<0x10, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
706 "movups {$src, $dst|$dst, $src}",
707 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
708 def MOVUPSmr : PSI<0x11, MRMDestMem, (ops f128mem:$dst, VR128:$src),
709 "movups {$src, $dst|$dst, $src}",
710 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
711 def MOVUPDrr : PDI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src),
712 "movupd {$src, $dst|$dst, $src}", []>;
713 def MOVUPDrm : PDI<0x10, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
714 "movupd {$src, $dst|$dst, $src}",
715 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
716 def MOVUPDmr : PDI<0x11, MRMDestMem, (ops f128mem:$dst, VR128:$src),
717 "movupd {$src, $dst|$dst, $src}",
718 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
720 let isTwoAddress = 1 in {
721 let AddedComplexity = 20 in {
722 def MOVLPSrm : PSI<0x12, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
723 "movlps {$src2, $dst|$dst, $src2}",
725 (v4f32 (vector_shuffle VR128:$src1,
726 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
727 MOVLP_shuffle_mask)))]>;
728 def MOVLPDrm : PDI<0x12, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
729 "movlpd {$src2, $dst|$dst, $src2}",
731 (v2f64 (vector_shuffle VR128:$src1,
732 (scalar_to_vector (loadf64 addr:$src2)),
733 MOVLP_shuffle_mask)))]>;
734 def MOVHPSrm : PSI<0x16, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
735 "movhps {$src2, $dst|$dst, $src2}",
737 (v4f32 (vector_shuffle VR128:$src1,
738 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
739 MOVHP_shuffle_mask)))]>;
740 def MOVHPDrm : PDI<0x16, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
741 "movhpd {$src2, $dst|$dst, $src2}",
743 (v2f64 (vector_shuffle VR128:$src1,
744 (scalar_to_vector (loadf64 addr:$src2)),
745 MOVHP_shuffle_mask)))]>;
749 def MOVLPSmr : PSI<0x13, MRMDestMem, (ops f64mem:$dst, VR128:$src),
750 "movlps {$src, $dst|$dst, $src}",
751 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
752 (iPTR 0))), addr:$dst)]>;
753 def MOVLPDmr : PDI<0x13, MRMDestMem, (ops f64mem:$dst, VR128:$src),
754 "movlpd {$src, $dst|$dst, $src}",
755 [(store (f64 (vector_extract (v2f64 VR128:$src),
756 (iPTR 0))), addr:$dst)]>;
758 // v2f64 extract element 1 is always custom lowered to unpack high to low
759 // and extract element 0 so the non-store version isn't too horrible.
760 def MOVHPSmr : PSI<0x17, MRMDestMem, (ops f64mem:$dst, VR128:$src),
761 "movhps {$src, $dst|$dst, $src}",
762 [(store (f64 (vector_extract
763 (v2f64 (vector_shuffle
764 (bc_v2f64 (v4f32 VR128:$src)), (undef),
765 UNPCKH_shuffle_mask)), (iPTR 0))),
767 def MOVHPDmr : PDI<0x17, MRMDestMem, (ops f64mem:$dst, VR128:$src),
768 "movhpd {$src, $dst|$dst, $src}",
769 [(store (f64 (vector_extract
770 (v2f64 (vector_shuffle VR128:$src, (undef),
771 UNPCKH_shuffle_mask)), (iPTR 0))),
774 let isTwoAddress = 1 in {
775 let AddedComplexity = 15 in {
776 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
777 "movlhps {$src2, $dst|$dst, $src2}",
779 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
780 MOVHP_shuffle_mask)))]>;
782 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
783 "movhlps {$src2, $dst|$dst, $src2}",
785 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
786 MOVHLPS_shuffle_mask)))]>;
790 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (ops VR128:$dst, VR128:$src),
791 "movshdup {$src, $dst|$dst, $src}",
792 [(set VR128:$dst, (v4f32 (vector_shuffle
794 MOVSHDUP_shuffle_mask)))]>;
795 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
796 "movshdup {$src, $dst|$dst, $src}",
797 [(set VR128:$dst, (v4f32 (vector_shuffle
798 (loadv4f32 addr:$src), (undef),
799 MOVSHDUP_shuffle_mask)))]>;
801 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src),
802 "movsldup {$src, $dst|$dst, $src}",
803 [(set VR128:$dst, (v4f32 (vector_shuffle
805 MOVSLDUP_shuffle_mask)))]>;
806 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
807 "movsldup {$src, $dst|$dst, $src}",
808 [(set VR128:$dst, (v4f32 (vector_shuffle
809 (loadv4f32 addr:$src), (undef),
810 MOVSLDUP_shuffle_mask)))]>;
812 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src),
813 "movddup {$src, $dst|$dst, $src}",
814 [(set VR128:$dst, (v2f64 (vector_shuffle
816 SSE_splat_lo_mask)))]>;
817 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
818 "movddup {$src, $dst|$dst, $src}",
819 [(set VR128:$dst, (v2f64 (vector_shuffle
820 (scalar_to_vector (loadf64 addr:$src)),
822 SSE_splat_lo_mask)))]>;
824 // SSE2 instructions without OpSize prefix
825 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
826 "cvtdq2ps {$src, $dst|$dst, $src}",
827 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
828 TB, Requires<[HasSSE2]>;
829 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
830 "cvtdq2ps {$src, $dst|$dst, $src}",
831 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
832 (bitconvert (loadv2i64 addr:$src))))]>,
833 TB, Requires<[HasSSE2]>;
835 // SSE2 instructions with XS prefix
836 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
837 "cvtdq2pd {$src, $dst|$dst, $src}",
838 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
839 XS, Requires<[HasSSE2]>;
840 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
841 "cvtdq2pd {$src, $dst|$dst, $src}",
842 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
843 (bitconvert (loadv2i64 addr:$src))))]>,
844 XS, Requires<[HasSSE2]>;
846 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
847 "cvtps2dq {$src, $dst|$dst, $src}",
848 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
849 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
850 "cvtps2dq {$src, $dst|$dst, $src}",
851 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
852 (load addr:$src)))]>;
853 // SSE2 packed instructions with XS prefix
854 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
855 "cvttps2dq {$src, $dst|$dst, $src}",
856 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
857 XS, Requires<[HasSSE2]>;
858 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
859 "cvttps2dq {$src, $dst|$dst, $src}",
860 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
861 (load addr:$src)))]>,
862 XS, Requires<[HasSSE2]>;
864 // SSE2 packed instructions with XD prefix
865 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
866 "cvtpd2dq {$src, $dst|$dst, $src}",
867 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
868 XD, Requires<[HasSSE2]>;
869 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
870 "cvtpd2dq {$src, $dst|$dst, $src}",
871 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
872 (load addr:$src)))]>,
873 XD, Requires<[HasSSE2]>;
874 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
875 "cvttpd2dq {$src, $dst|$dst, $src}",
876 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
877 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
878 "cvttpd2dq {$src, $dst|$dst, $src}",
879 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
880 (load addr:$src)))]>;
882 // SSE2 instructions without OpSize prefix
883 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (ops VR128:$dst, VR128:$src),
884 "cvtps2pd {$src, $dst|$dst, $src}",
885 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
886 TB, Requires<[HasSSE2]>;
887 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (ops VR128:$dst, f64mem:$src),
888 "cvtps2pd {$src, $dst|$dst, $src}",
889 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
890 (load addr:$src)))]>,
891 TB, Requires<[HasSSE2]>;
893 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (ops VR128:$dst, VR128:$src),
894 "cvtpd2ps {$src, $dst|$dst, $src}",
895 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
896 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (ops VR128:$dst, f128mem:$src),
897 "cvtpd2ps {$src, $dst|$dst, $src}",
898 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
899 (load addr:$src)))]>;
901 // Match intrinsics which expect XMM operand(s).
902 // Aliases for intrinsics
903 let isTwoAddress = 1 in {
904 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
905 (ops VR128:$dst, VR128:$src1, GR32:$src2),
906 "cvtsi2sd {$src2, $dst|$dst, $src2}",
907 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
909 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
910 (ops VR128:$dst, VR128:$src1, i32mem:$src2),
911 "cvtsi2sd {$src2, $dst|$dst, $src2}",
912 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
913 (loadi32 addr:$src2)))]>;
914 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
915 (ops VR128:$dst, VR128:$src1, VR128:$src2),
916 "cvtsd2ss {$src2, $dst|$dst, $src2}",
917 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
919 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
920 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
921 "cvtsd2ss {$src2, $dst|$dst, $src2}",
922 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
923 (load addr:$src2)))]>;
924 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
925 (ops VR128:$dst, VR128:$src1, VR128:$src2),
926 "cvtss2sd {$src2, $dst|$dst, $src2}",
927 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
930 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
931 (ops VR128:$dst, VR128:$src1, f32mem:$src2),
932 "cvtss2sd {$src2, $dst|$dst, $src2}",
933 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
934 (load addr:$src2)))]>, XS,
938 /// packed_sse12_fp_binop_rm - Packed SSE binops come in four basic forms:
939 /// 1. v4f32 vs v2f64 - These come in SSE1/SSE2 forms for float/doubles.
940 /// 2. rr vs rm - They include a reg+reg form and a ref+mem form.
942 let isTwoAddress = 1 in {
943 multiclass packed_sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
944 SDNode OpNode, bit Commutable = 0> {
945 // Packed operation, reg+reg.
946 def PSrr : PSI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
947 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2}"),
948 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
949 let isCommutable = Commutable;
951 def PDrr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
952 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
953 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
954 let isCommutable = Commutable;
956 // Packed operation, reg+mem.
957 def PSrm : PSI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
958 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2}"),
959 [(set VR128:$dst, (OpNode VR128:$src1, (loadv4f32 addr:$src2)))]>;
960 def PDrm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
961 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2}"),
962 [(set VR128:$dst, (OpNode VR128:$src1, (loadv2f64 addr:$src2)))]>;
966 defm ADD : packed_sse12_fp_binop_rm<0x58, "add", fadd, 1>;
967 defm MUL : packed_sse12_fp_binop_rm<0x59, "mul", fmul, 1>;
968 defm DIV : packed_sse12_fp_binop_rm<0x5E, "div", fdiv>;
969 defm SUB : packed_sse12_fp_binop_rm<0x5C, "sub", fsub>;
972 let isTwoAddress = 1 in {
973 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
974 (ops VR128:$dst, VR128:$src1, VR128:$src2),
975 "addsubps {$src2, $dst|$dst, $src2}",
976 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
978 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
979 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
980 "addsubps {$src2, $dst|$dst, $src2}",
981 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
982 (load addr:$src2)))]>;
983 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
984 (ops VR128:$dst, VR128:$src1, VR128:$src2),
985 "addsubpd {$src2, $dst|$dst, $src2}",
986 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
988 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
989 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
990 "addsubpd {$src2, $dst|$dst, $src2}",
991 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
992 (load addr:$src2)))]>;
995 def SQRTPSr : PS_Intr<0x51, "sqrtps", int_x86_sse_sqrt_ps>;
996 def SQRTPSm : PS_Intm<0x51, "sqrtps", int_x86_sse_sqrt_ps>;
997 def SQRTPDr : PD_Intr<0x51, "sqrtpd", int_x86_sse2_sqrt_pd>;
998 def SQRTPDm : PD_Intm<0x51, "sqrtpd", int_x86_sse2_sqrt_pd>;
1000 def RSQRTPSr : PS_Intr<0x52, "rsqrtps", int_x86_sse_rsqrt_ps>;
1001 def RSQRTPSm : PS_Intm<0x52, "rsqrtps", int_x86_sse_rsqrt_ps>;
1002 def RCPPSr : PS_Intr<0x53, "rcpps", int_x86_sse_rcp_ps>;
1003 def RCPPSm : PS_Intm<0x53, "rcpps", int_x86_sse_rcp_ps>;
1005 let isTwoAddress = 1 in {
1006 let isCommutable = 1 in {
1007 def MAXPSrr : PS_Intrr<0x5F, "maxps", int_x86_sse_max_ps>;
1008 def MAXPDrr : PD_Intrr<0x5F, "maxpd", int_x86_sse2_max_pd>;
1009 def MINPSrr : PS_Intrr<0x5D, "minps", int_x86_sse_min_ps>;
1010 def MINPDrr : PD_Intrr<0x5D, "minpd", int_x86_sse2_min_pd>;
1012 def MAXPSrm : PS_Intrm<0x5F, "maxps", int_x86_sse_max_ps>;
1013 def MAXPDrm : PD_Intrm<0x5F, "maxpd", int_x86_sse2_max_pd>;
1014 def MINPSrm : PS_Intrm<0x5D, "minps", int_x86_sse_min_ps>;
1015 def MINPDrm : PD_Intrm<0x5D, "minpd", int_x86_sse2_min_pd>;
1019 let isTwoAddress = 1 in {
1020 let isCommutable = 1 in {
1021 def ANDPSrr : PSI<0x54, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1022 "andps {$src2, $dst|$dst, $src2}",
1023 [(set VR128:$dst, (v2i64 (and VR128:$src1, VR128:$src2)))]>;
1024 def ANDPDrr : PDI<0x54, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1025 "andpd {$src2, $dst|$dst, $src2}",
1027 (and (bc_v2i64 (v2f64 VR128:$src1)),
1028 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1029 def ORPSrr : PSI<0x56, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1030 "orps {$src2, $dst|$dst, $src2}",
1031 [(set VR128:$dst, (v2i64 (or VR128:$src1, VR128:$src2)))]>;
1032 def ORPDrr : PDI<0x56, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1033 "orpd {$src2, $dst|$dst, $src2}",
1035 (or (bc_v2i64 (v2f64 VR128:$src1)),
1036 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1037 def XORPSrr : PSI<0x57, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1038 "xorps {$src2, $dst|$dst, $src2}",
1039 [(set VR128:$dst, (v2i64 (xor VR128:$src1, VR128:$src2)))]>;
1040 def XORPDrr : PDI<0x57, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1041 "xorpd {$src2, $dst|$dst, $src2}",
1043 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1044 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1046 def ANDPSrm : PSI<0x54, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1047 "andps {$src2, $dst|$dst, $src2}",
1048 [(set VR128:$dst, (and VR128:$src1,
1049 (bc_v2i64 (loadv4f32 addr:$src2))))]>;
1050 def ANDPDrm : PDI<0x54, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1051 "andpd {$src2, $dst|$dst, $src2}",
1053 (and (bc_v2i64 (v2f64 VR128:$src1)),
1054 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1055 def ORPSrm : PSI<0x56, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1056 "orps {$src2, $dst|$dst, $src2}",
1057 [(set VR128:$dst, (or VR128:$src1,
1058 (bc_v2i64 (loadv4f32 addr:$src2))))]>;
1059 def ORPDrm : PDI<0x56, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1060 "orpd {$src2, $dst|$dst, $src2}",
1062 (or (bc_v2i64 (v2f64 VR128:$src1)),
1063 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1064 def XORPSrm : PSI<0x57, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1065 "xorps {$src2, $dst|$dst, $src2}",
1066 [(set VR128:$dst, (xor VR128:$src1,
1067 (bc_v2i64 (loadv4f32 addr:$src2))))]>;
1068 def XORPDrm : PDI<0x57, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1069 "xorpd {$src2, $dst|$dst, $src2}",
1071 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1072 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1073 def ANDNPSrr : PSI<0x55, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1074 "andnps {$src2, $dst|$dst, $src2}",
1075 [(set VR128:$dst, (v2i64 (and (xor VR128:$src1,
1076 (bc_v2i64 (v4i32 immAllOnesV))),
1078 def ANDNPSrm : PSI<0x55, MRMSrcMem, (ops VR128:$dst, VR128:$src1,f128mem:$src2),
1079 "andnps {$src2, $dst|$dst, $src2}",
1080 [(set VR128:$dst, (v2i64 (and (xor VR128:$src1,
1081 (bc_v2i64 (v4i32 immAllOnesV))),
1082 (bc_v2i64 (loadv4f32 addr:$src2)))))]>;
1083 def ANDNPDrr : PDI<0x55, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1084 "andnpd {$src2, $dst|$dst, $src2}",
1086 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1087 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1088 def ANDNPDrm : PDI<0x55, MRMSrcMem, (ops VR128:$dst, VR128:$src1,f128mem:$src2),
1089 "andnpd {$src2, $dst|$dst, $src2}",
1091 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1092 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1095 let isTwoAddress = 1 in {
1096 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
1097 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
1098 "cmp${cc}ps {$src, $dst|$dst, $src}",
1099 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1100 VR128:$src, imm:$cc))]>;
1101 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
1102 (ops VR128:$dst, VR128:$src1, f128mem:$src, SSECC:$cc),
1103 "cmp${cc}ps {$src, $dst|$dst, $src}",
1104 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1105 (load addr:$src), imm:$cc))]>;
1106 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1107 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
1108 "cmp${cc}pd {$src, $dst|$dst, $src}",
1109 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1110 VR128:$src, imm:$cc))]>;
1111 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1112 (ops VR128:$dst, VR128:$src1, f128mem:$src, SSECC:$cc),
1113 "cmp${cc}pd {$src, $dst|$dst, $src}",
1114 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1115 (load addr:$src), imm:$cc))]>;
1118 // Shuffle and unpack instructions
1119 let isTwoAddress = 1 in {
1120 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
1121 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
1122 (ops VR128:$dst, VR128:$src1, VR128:$src2, i32i8imm:$src3),
1123 "shufps {$src3, $src2, $dst|$dst, $src2, $src3}",
1124 [(set VR128:$dst, (v4f32 (vector_shuffle
1125 VR128:$src1, VR128:$src2,
1126 SHUFP_shuffle_mask:$src3)))]>;
1127 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
1128 (ops VR128:$dst, VR128:$src1, f128mem:$src2, i32i8imm:$src3),
1129 "shufps {$src3, $src2, $dst|$dst, $src2, $src3}",
1130 [(set VR128:$dst, (v4f32 (vector_shuffle
1131 VR128:$src1, (load addr:$src2),
1132 SHUFP_shuffle_mask:$src3)))]>;
1133 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1134 (ops VR128:$dst, VR128:$src1, VR128:$src2, i8imm:$src3),
1135 "shufpd {$src3, $src2, $dst|$dst, $src2, $src3}",
1136 [(set VR128:$dst, (v2f64 (vector_shuffle
1137 VR128:$src1, VR128:$src2,
1138 SHUFP_shuffle_mask:$src3)))]>;
1139 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1140 (ops VR128:$dst, VR128:$src1, f128mem:$src2, i8imm:$src3),
1141 "shufpd {$src3, $src2, $dst|$dst, $src2, $src3}",
1142 [(set VR128:$dst, (v2f64 (vector_shuffle
1143 VR128:$src1, (load addr:$src2),
1144 SHUFP_shuffle_mask:$src3)))]>;
1146 let AddedComplexity = 10 in {
1147 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
1148 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1149 "unpckhps {$src2, $dst|$dst, $src2}",
1150 [(set VR128:$dst, (v4f32 (vector_shuffle
1151 VR128:$src1, VR128:$src2,
1152 UNPCKH_shuffle_mask)))]>;
1153 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
1154 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1155 "unpckhps {$src2, $dst|$dst, $src2}",
1156 [(set VR128:$dst, (v4f32 (vector_shuffle
1157 VR128:$src1, (load addr:$src2),
1158 UNPCKH_shuffle_mask)))]>;
1159 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1160 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1161 "unpckhpd {$src2, $dst|$dst, $src2}",
1162 [(set VR128:$dst, (v2f64 (vector_shuffle
1163 VR128:$src1, VR128:$src2,
1164 UNPCKH_shuffle_mask)))]>;
1165 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1166 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1167 "unpckhpd {$src2, $dst|$dst, $src2}",
1168 [(set VR128:$dst, (v2f64 (vector_shuffle
1169 VR128:$src1, (load addr:$src2),
1170 UNPCKH_shuffle_mask)))]>;
1172 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
1173 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1174 "unpcklps {$src2, $dst|$dst, $src2}",
1175 [(set VR128:$dst, (v4f32 (vector_shuffle
1176 VR128:$src1, VR128:$src2,
1177 UNPCKL_shuffle_mask)))]>;
1178 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
1179 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1180 "unpcklps {$src2, $dst|$dst, $src2}",
1181 [(set VR128:$dst, (v4f32 (vector_shuffle
1182 VR128:$src1, (load addr:$src2),
1183 UNPCKL_shuffle_mask)))]>;
1184 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1185 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1186 "unpcklpd {$src2, $dst|$dst, $src2}",
1187 [(set VR128:$dst, (v2f64 (vector_shuffle
1188 VR128:$src1, VR128:$src2,
1189 UNPCKL_shuffle_mask)))]>;
1190 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1191 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1192 "unpcklpd {$src2, $dst|$dst, $src2}",
1193 [(set VR128:$dst, (v2f64 (vector_shuffle
1194 VR128:$src1, (load addr:$src2),
1195 UNPCKL_shuffle_mask)))]>;
1196 } // AddedComplexity
1201 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
1202 : S3DI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1203 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1204 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
1205 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
1206 : S3DI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1207 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1208 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
1209 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
1210 : S3I<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1211 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1212 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
1213 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
1214 : S3I<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1215 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1216 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
1218 let isTwoAddress = 1 in {
1219 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
1220 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
1221 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
1222 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
1223 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
1224 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
1225 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
1226 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
1229 //===----------------------------------------------------------------------===//
1230 // SSE integer instructions
1231 //===----------------------------------------------------------------------===//
1233 // Move Instructions
1234 def MOVDQArr : PDI<0x6F, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1235 "movdqa {$src, $dst|$dst, $src}", []>;
1236 def MOVDQArm : PDI<0x6F, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1237 "movdqa {$src, $dst|$dst, $src}",
1238 [(set VR128:$dst, (loadv2i64 addr:$src))]>;
1239 def MOVDQAmr : PDI<0x7F, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1240 "movdqa {$src, $dst|$dst, $src}",
1241 [(store (v2i64 VR128:$src), addr:$dst)]>;
1242 def MOVDQUrm : I<0x6F, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1243 "movdqu {$src, $dst|$dst, $src}",
1244 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1245 XS, Requires<[HasSSE2]>;
1246 def MOVDQUmr : I<0x7F, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1247 "movdqu {$src, $dst|$dst, $src}",
1248 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1249 XS, Requires<[HasSSE2]>;
1250 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1251 "lddqu {$src, $dst|$dst, $src}",
1252 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
1255 let isTwoAddress = 1 in {
1256 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1257 bit Commutable = 0> {
1258 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1259 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1260 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1261 let isCommutable = Commutable;
1263 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1264 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1265 [(set VR128:$dst, (IntId VR128:$src1,
1266 (bitconvert (loadv2i64 addr:$src2))))]>;
1270 let isTwoAddress = 1 in {
1271 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1272 string OpcodeStr, Intrinsic IntId> {
1273 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1274 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1275 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1276 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1277 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1278 [(set VR128:$dst, (IntId VR128:$src1,
1279 (bitconvert (loadv2i64 addr:$src2))))]>;
1280 def ri : PDIi8<opc2, ImmForm, (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1281 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1282 [(set VR128:$dst, (IntId VR128:$src1,
1283 (scalar_to_vector (i32 imm:$src2))))]>;
1288 let isTwoAddress = 1 in {
1289 /// PDI_binop_rm - Simple SSE2 binary operator.
1290 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1291 ValueType OpVT, bit Commutable = 0> {
1292 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1293 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1294 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1295 let isCommutable = Commutable;
1297 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1298 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1299 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1300 (bitconvert (loadv2i64 addr:$src2)))))]>;
1303 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1305 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1306 /// to collapse (bitconvert VT to VT) into its operand.
1308 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1309 bit Commutable = 0> {
1310 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1311 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1312 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1313 let isCommutable = Commutable;
1315 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1316 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1317 [(set VR128:$dst, (OpNode VR128:$src1,(loadv2i64 addr:$src2)))]>;
1321 /// SS3I_binop_rm_int - Simple SSSE3 binary operatr whose type is v2i64.
1322 let isTwoAddress = 1 in {
1323 multiclass SS3I_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1324 bit Commutable = 0> {
1325 def rr : SS38I<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1326 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1327 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1328 let isCommutable = Commutable;
1330 def rm : SS38I<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1331 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1334 (bitconvert (loadv2i64 addr:$src2))))]>;
1338 // 128-bit Integer Arithmetic
1340 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1341 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1342 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1343 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1345 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1346 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1347 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1348 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1350 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1351 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1352 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1353 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1355 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1356 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1357 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1358 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1360 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1362 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1363 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1364 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1366 defm PMULHRSW128 : SS3I_binop_rm_int<0x0B, "pmulhrsw",
1367 int_x86_ssse3_pmulhrsw_128, 1>;
1369 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1371 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1372 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1375 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1376 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1377 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1378 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1379 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1382 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw", int_x86_sse2_psll_w>;
1383 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld", int_x86_sse2_psll_d>;
1384 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq", int_x86_sse2_psll_q>;
1386 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw", int_x86_sse2_psrl_w>;
1387 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld", int_x86_sse2_psrl_d>;
1388 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq", int_x86_sse2_psrl_q>;
1390 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw", int_x86_sse2_psra_w>;
1391 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad", int_x86_sse2_psra_d>;
1392 // PSRAQ doesn't exist in SSE[1-3].
1395 // 128-bit logical shifts.
1396 let isTwoAddress = 1 in {
1397 def PSLLDQri : PDIi8<0x73, MRM7r, (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1398 "pslldq {$src2, $dst|$dst, $src2}", []>;
1399 def PSRLDQri : PDIi8<0x73, MRM3r, (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1400 "psrldq {$src2, $dst|$dst, $src2}", []>;
1401 // PSRADQri doesn't exist in SSE[1-3].
1404 let Predicates = [HasSSE2] in {
1405 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1406 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1407 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1408 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1409 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1410 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1414 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1415 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1416 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1418 let isTwoAddress = 1 in {
1419 def PANDNrr : PDI<0xDF, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1420 "pandn {$src2, $dst|$dst, $src2}",
1421 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1424 def PANDNrm : PDI<0xDF, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1425 "pandn {$src2, $dst|$dst, $src2}",
1426 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1427 (load addr:$src2))))]>;
1430 // SSE2 Integer comparison
1431 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1432 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1433 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1434 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1435 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1436 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1438 // Pack instructions
1439 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1440 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1441 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1443 // Shuffle and unpack instructions
1444 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1445 (ops VR128:$dst, VR128:$src1, i8imm:$src2),
1446 "pshufd {$src2, $src1, $dst|$dst, $src1, $src2}",
1447 [(set VR128:$dst, (v4i32 (vector_shuffle
1448 VR128:$src1, (undef),
1449 PSHUFD_shuffle_mask:$src2)))]>;
1450 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1451 (ops VR128:$dst, i128mem:$src1, i8imm:$src2),
1452 "pshufd {$src2, $src1, $dst|$dst, $src1, $src2}",
1453 [(set VR128:$dst, (v4i32 (vector_shuffle
1454 (bc_v4i32(loadv2i64 addr:$src1)),
1456 PSHUFD_shuffle_mask:$src2)))]>;
1458 // SSE2 with ImmT == Imm8 and XS prefix.
1459 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1460 (ops VR128:$dst, VR128:$src1, i8imm:$src2),
1461 "pshufhw {$src2, $src1, $dst|$dst, $src1, $src2}",
1462 [(set VR128:$dst, (v8i16 (vector_shuffle
1463 VR128:$src1, (undef),
1464 PSHUFHW_shuffle_mask:$src2)))]>,
1465 XS, Requires<[HasSSE2]>;
1466 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
1467 (ops VR128:$dst, i128mem:$src1, i8imm:$src2),
1468 "pshufhw {$src2, $src1, $dst|$dst, $src1, $src2}",
1469 [(set VR128:$dst, (v8i16 (vector_shuffle
1470 (bc_v8i16 (loadv2i64 addr:$src1)),
1472 PSHUFHW_shuffle_mask:$src2)))]>,
1473 XS, Requires<[HasSSE2]>;
1475 // SSE2 with ImmT == Imm8 and XD prefix.
1476 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
1477 (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1478 "pshuflw {$src2, $src1, $dst|$dst, $src1, $src2}",
1479 [(set VR128:$dst, (v8i16 (vector_shuffle
1480 VR128:$src1, (undef),
1481 PSHUFLW_shuffle_mask:$src2)))]>,
1482 XD, Requires<[HasSSE2]>;
1483 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
1484 (ops VR128:$dst, i128mem:$src1, i32i8imm:$src2),
1485 "pshuflw {$src2, $src1, $dst|$dst, $src1, $src2}",
1486 [(set VR128:$dst, (v8i16 (vector_shuffle
1487 (bc_v8i16 (loadv2i64 addr:$src1)),
1489 PSHUFLW_shuffle_mask:$src2)))]>,
1490 XD, Requires<[HasSSE2]>;
1492 let isTwoAddress = 1 in {
1493 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
1494 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1495 "punpcklbw {$src2, $dst|$dst, $src2}",
1497 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1498 UNPCKL_shuffle_mask)))]>;
1499 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
1500 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1501 "punpcklbw {$src2, $dst|$dst, $src2}",
1503 (v16i8 (vector_shuffle VR128:$src1,
1504 (bc_v16i8 (loadv2i64 addr:$src2)),
1505 UNPCKL_shuffle_mask)))]>;
1506 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
1507 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1508 "punpcklwd {$src2, $dst|$dst, $src2}",
1510 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1511 UNPCKL_shuffle_mask)))]>;
1512 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
1513 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1514 "punpcklwd {$src2, $dst|$dst, $src2}",
1516 (v8i16 (vector_shuffle VR128:$src1,
1517 (bc_v8i16 (loadv2i64 addr:$src2)),
1518 UNPCKL_shuffle_mask)))]>;
1519 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
1520 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1521 "punpckldq {$src2, $dst|$dst, $src2}",
1523 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1524 UNPCKL_shuffle_mask)))]>;
1525 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
1526 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1527 "punpckldq {$src2, $dst|$dst, $src2}",
1529 (v4i32 (vector_shuffle VR128:$src1,
1530 (bc_v4i32 (loadv2i64 addr:$src2)),
1531 UNPCKL_shuffle_mask)))]>;
1532 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
1533 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1534 "punpcklqdq {$src2, $dst|$dst, $src2}",
1536 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1537 UNPCKL_shuffle_mask)))]>;
1538 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
1539 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1540 "punpcklqdq {$src2, $dst|$dst, $src2}",
1542 (v2i64 (vector_shuffle VR128:$src1,
1543 (loadv2i64 addr:$src2),
1544 UNPCKL_shuffle_mask)))]>;
1546 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
1547 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1548 "punpckhbw {$src2, $dst|$dst, $src2}",
1550 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1551 UNPCKH_shuffle_mask)))]>;
1552 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
1553 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1554 "punpckhbw {$src2, $dst|$dst, $src2}",
1556 (v16i8 (vector_shuffle VR128:$src1,
1557 (bc_v16i8 (loadv2i64 addr:$src2)),
1558 UNPCKH_shuffle_mask)))]>;
1559 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
1560 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1561 "punpckhwd {$src2, $dst|$dst, $src2}",
1563 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1564 UNPCKH_shuffle_mask)))]>;
1565 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
1566 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1567 "punpckhwd {$src2, $dst|$dst, $src2}",
1569 (v8i16 (vector_shuffle VR128:$src1,
1570 (bc_v8i16 (loadv2i64 addr:$src2)),
1571 UNPCKH_shuffle_mask)))]>;
1572 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
1573 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1574 "punpckhdq {$src2, $dst|$dst, $src2}",
1576 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1577 UNPCKH_shuffle_mask)))]>;
1578 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
1579 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1580 "punpckhdq {$src2, $dst|$dst, $src2}",
1582 (v4i32 (vector_shuffle VR128:$src1,
1583 (bc_v4i32 (loadv2i64 addr:$src2)),
1584 UNPCKH_shuffle_mask)))]>;
1585 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
1586 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1587 "punpckhqdq {$src2, $dst|$dst, $src2}",
1589 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1590 UNPCKH_shuffle_mask)))]>;
1591 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
1592 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1593 "punpckhqdq {$src2, $dst|$dst, $src2}",
1595 (v2i64 (vector_shuffle VR128:$src1,
1596 (loadv2i64 addr:$src2),
1597 UNPCKH_shuffle_mask)))]>;
1601 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
1602 (ops GR32:$dst, VR128:$src1, i32i8imm:$src2),
1603 "pextrw {$src2, $src1, $dst|$dst, $src1, $src2}",
1604 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
1605 (iPTR imm:$src2)))]>;
1606 let isTwoAddress = 1 in {
1607 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
1608 (ops VR128:$dst, VR128:$src1, GR32:$src2, i32i8imm:$src3),
1609 "pinsrw {$src3, $src2, $dst|$dst, $src2, $src3}",
1610 [(set VR128:$dst, (v8i16 (X86pinsrw (v8i16 VR128:$src1),
1611 GR32:$src2, (iPTR imm:$src3))))]>;
1612 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
1613 (ops VR128:$dst, VR128:$src1, i16mem:$src2, i32i8imm:$src3),
1614 "pinsrw {$src3, $src2, $dst|$dst, $src2, $src3}",
1616 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
1617 (i32 (anyext (loadi16 addr:$src2))),
1618 (iPTR imm:$src3))))]>;
1621 //===----------------------------------------------------------------------===//
1622 // Miscellaneous Instructions
1623 //===----------------------------------------------------------------------===//
1626 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (ops GR32:$dst, VR128:$src),
1627 "movmskps {$src, $dst|$dst, $src}",
1628 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
1629 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (ops GR32:$dst, VR128:$src),
1630 "movmskpd {$src, $dst|$dst, $src}",
1631 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1633 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (ops GR32:$dst, VR128:$src),
1634 "pmovmskb {$src, $dst|$dst, $src}",
1635 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
1637 // Conditional store
1638 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (ops VR128:$src, VR128:$mask),
1639 "maskmovdqu {$mask, $src|$src, $mask}",
1640 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>,
1643 // Prefetching loads.
1644 // TODO: no intrinsics for these?
1645 def PREFETCHT0 : PSI<0x18, MRM1m, (ops i8mem:$src), "prefetcht0 $src", []>;
1646 def PREFETCHT1 : PSI<0x18, MRM2m, (ops i8mem:$src), "prefetcht1 $src", []>;
1647 def PREFETCHT2 : PSI<0x18, MRM3m, (ops i8mem:$src), "prefetcht2 $src", []>;
1648 def PREFETCHTNTA : PSI<0x18, MRM0m, (ops i8mem:$src), "prefetchtnta $src", []>;
1650 // Non-temporal stores
1651 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1652 "movntps {$src, $dst|$dst, $src}",
1653 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1654 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1655 "movntpd {$src, $dst|$dst, $src}",
1656 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
1657 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (ops f128mem:$dst, VR128:$src),
1658 "movntdq {$src, $dst|$dst, $src}",
1659 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
1660 def MOVNTImr : I<0xC3, MRMDestMem, (ops i32mem:$dst, GR32:$src),
1661 "movnti {$src, $dst|$dst, $src}",
1662 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
1663 TB, Requires<[HasSSE2]>;
1666 def CLFLUSH : I<0xAE, MRM7m, (ops i8mem:$src),
1667 "clflush $src", [(int_x86_sse2_clflush addr:$src)]>,
1668 TB, Requires<[HasSSE2]>;
1670 // Load, store, and memory fence
1671 def SFENCE : PSI<0xAE, MRM7m, (ops), "sfence", [(int_x86_sse_sfence)]>;
1672 def LFENCE : I<0xAE, MRM5m, (ops),
1673 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
1674 def MFENCE : I<0xAE, MRM6m, (ops),
1675 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
1678 def LDMXCSR : PSI<0xAE, MRM2m, (ops i32mem:$src),
1679 "ldmxcsr $src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1680 def STMXCSR : PSI<0xAE, MRM3m, (ops i32mem:$dst),
1681 "stmxcsr $dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1683 // Thread synchronization
1684 def MONITOR : I<0xC8, RawFrm, (ops), "monitor",
1685 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
1686 def MWAIT : I<0xC9, RawFrm, (ops), "mwait",
1687 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
1689 //===----------------------------------------------------------------------===//
1690 // Alias Instructions
1691 //===----------------------------------------------------------------------===//
1693 // Alias instructions that map zero vector to pxor / xorp* for sse.
1694 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
1695 let isReMaterializable = 1 in {
1696 def V_SET0 : PSI<0x57, MRMInitReg, (ops VR128:$dst),
1698 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
1700 def V_SETALLONES : PDI<0x76, MRMInitReg, (ops VR128:$dst),
1701 "pcmpeqd $dst, $dst",
1702 [(set VR128:$dst, (v2f64 immAllOnesV))]>;
1705 // FR32 / FR64 to 128-bit vector conversion.
1706 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (ops VR128:$dst, FR32:$src),
1707 "movss {$src, $dst|$dst, $src}",
1709 (v4f32 (scalar_to_vector FR32:$src)))]>;
1710 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
1711 "movss {$src, $dst|$dst, $src}",
1713 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
1714 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (ops VR128:$dst, FR64:$src),
1715 "movsd {$src, $dst|$dst, $src}",
1717 (v2f64 (scalar_to_vector FR64:$src)))]>;
1718 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1719 "movsd {$src, $dst|$dst, $src}",
1721 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
1723 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (ops VR128:$dst, GR32:$src),
1724 "movd {$src, $dst|$dst, $src}",
1726 (v4i32 (scalar_to_vector GR32:$src)))]>;
1727 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (ops VR128:$dst, i32mem:$src),
1728 "movd {$src, $dst|$dst, $src}",
1730 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
1732 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (ops FR32:$dst, GR32:$src),
1733 "movd {$src, $dst|$dst, $src}",
1734 [(set FR32:$dst, (bitconvert GR32:$src))]>;
1736 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (ops FR32:$dst, i32mem:$src),
1737 "movd {$src, $dst|$dst, $src}",
1738 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
1740 // SSE2 instructions with XS prefix
1741 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1742 "movq {$src, $dst|$dst, $src}",
1744 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
1745 Requires<[HasSSE2]>;
1746 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (ops i64mem:$dst, VR128:$src),
1747 "movq {$src, $dst|$dst, $src}",
1748 [(store (i64 (vector_extract (v2i64 VR128:$src),
1749 (iPTR 0))), addr:$dst)]>;
1751 // FIXME: may not be able to eliminate this movss with coalescing the src and
1752 // dest register classes are different. We really want to write this pattern
1754 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1755 // (f32 FR32:$src)>;
1756 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (ops FR32:$dst, VR128:$src),
1757 "movss {$src, $dst|$dst, $src}",
1758 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
1760 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (ops f32mem:$dst, VR128:$src),
1761 "movss {$src, $dst|$dst, $src}",
1762 [(store (f32 (vector_extract (v4f32 VR128:$src),
1763 (iPTR 0))), addr:$dst)]>;
1764 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (ops FR64:$dst, VR128:$src),
1765 "movsd {$src, $dst|$dst, $src}",
1766 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
1768 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (ops f64mem:$dst, VR128:$src),
1769 "movsd {$src, $dst|$dst, $src}",
1770 [(store (f64 (vector_extract (v2f64 VR128:$src),
1771 (iPTR 0))), addr:$dst)]>;
1772 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (ops GR32:$dst, VR128:$src),
1773 "movd {$src, $dst|$dst, $src}",
1774 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
1776 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (ops i32mem:$dst, VR128:$src),
1777 "movd {$src, $dst|$dst, $src}",
1778 [(store (i32 (vector_extract (v4i32 VR128:$src),
1779 (iPTR 0))), addr:$dst)]>;
1781 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (ops GR32:$dst, FR32:$src),
1782 "movd {$src, $dst|$dst, $src}",
1783 [(set GR32:$dst, (bitconvert FR32:$src))]>;
1784 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (ops i32mem:$dst, FR32:$src),
1785 "movd {$src, $dst|$dst, $src}",
1786 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
1789 // Move to lower bits of a VR128, leaving upper bits alone.
1790 // Three operand (but two address) aliases.
1791 let isTwoAddress = 1 in {
1792 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src1, FR32:$src2),
1793 "movss {$src2, $dst|$dst, $src2}", []>;
1794 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src1, FR64:$src2),
1795 "movsd {$src2, $dst|$dst, $src2}", []>;
1797 let AddedComplexity = 15 in {
1798 def MOVLPSrr : SSI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1799 "movss {$src2, $dst|$dst, $src2}",
1801 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
1802 MOVL_shuffle_mask)))]>;
1803 def MOVLPDrr : SDI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1804 "movsd {$src2, $dst|$dst, $src2}",
1806 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
1807 MOVL_shuffle_mask)))]>;
1811 // Store / copy lower 64-bits of a XMM register.
1812 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (ops i64mem:$dst, VR128:$src),
1813 "movq {$src, $dst|$dst, $src}",
1814 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
1816 // Move to lower bits of a VR128 and zeroing upper bits.
1817 // Loading from memory automatically zeroing upper bits.
1818 let AddedComplexity = 20 in {
1819 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
1820 "movss {$src, $dst|$dst, $src}",
1821 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV,
1822 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
1823 MOVL_shuffle_mask)))]>;
1824 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1825 "movsd {$src, $dst|$dst, $src}",
1826 [(set VR128:$dst, (v2f64 (vector_shuffle immAllZerosV,
1827 (v2f64 (scalar_to_vector (loadf64 addr:$src))),
1828 MOVL_shuffle_mask)))]>;
1830 let AddedComplexity = 15 in
1831 // movd / movq to XMM register zero-extends
1832 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (ops VR128:$dst, GR32:$src),
1833 "movd {$src, $dst|$dst, $src}",
1834 [(set VR128:$dst, (v4i32 (vector_shuffle immAllZerosV,
1835 (v4i32 (scalar_to_vector GR32:$src)),
1836 MOVL_shuffle_mask)))]>;
1837 let AddedComplexity = 20 in
1838 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (ops VR128:$dst, i32mem:$src),
1839 "movd {$src, $dst|$dst, $src}",
1840 [(set VR128:$dst, (v4i32 (vector_shuffle immAllZerosV,
1841 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
1842 MOVL_shuffle_mask)))]>;
1843 // Moving from XMM to XMM but still clear upper 64 bits.
1844 let AddedComplexity = 15 in
1845 def MOVZQI2PQIrr : I<0x7E, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1846 "movq {$src, $dst|$dst, $src}",
1847 [(set VR128:$dst, (int_x86_sse2_movl_dq VR128:$src))]>,
1848 XS, Requires<[HasSSE2]>;
1849 let AddedComplexity = 20 in
1850 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1851 "movq {$src, $dst|$dst, $src}",
1852 [(set VR128:$dst, (int_x86_sse2_movl_dq
1853 (bitconvert (loadv2i64 addr:$src))))]>,
1854 XS, Requires<[HasSSE2]>;
1856 //===----------------------------------------------------------------------===//
1857 // Non-Instruction Patterns
1858 //===----------------------------------------------------------------------===//
1860 // 128-bit vector undef's.
1861 def : Pat<(v2f64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
1862 def : Pat<(v16i8 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
1863 def : Pat<(v8i16 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
1864 def : Pat<(v4i32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
1865 def : Pat<(v2i64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
1867 // 128-bit vector all zero's.
1868 def : Pat<(v16i8 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
1869 def : Pat<(v8i16 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
1870 def : Pat<(v4i32 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
1871 def : Pat<(v2i64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
1872 def : Pat<(v2f64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
1874 // 128-bit vector all one's.
1875 def : Pat<(v16i8 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
1876 def : Pat<(v8i16 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
1877 def : Pat<(v4i32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
1878 def : Pat<(v2i64 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
1879 def : Pat<(v4f32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE1]>;
1881 // Store 128-bit integer vector values.
1882 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
1883 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
1884 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
1885 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
1886 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
1887 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
1889 // Scalar to v8i16 / v16i8. The source may be a GR32, but only the lower 8 or
1891 def : Pat<(v8i16 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
1892 Requires<[HasSSE2]>;
1893 def : Pat<(v16i8 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
1894 Requires<[HasSSE2]>;
1897 let Predicates = [HasSSE2] in {
1898 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
1899 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
1900 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
1901 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
1902 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
1903 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
1904 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
1905 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
1906 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
1907 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
1908 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
1909 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
1910 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
1911 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
1912 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
1913 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
1914 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
1915 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
1916 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
1917 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
1918 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
1919 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
1920 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
1921 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
1922 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
1923 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
1924 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
1925 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
1926 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
1927 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
1930 // Move scalar to XMM zero-extended
1931 // movd to XMM register zero-extends
1932 let AddedComplexity = 15 in {
1933 def : Pat<(v8i16 (vector_shuffle immAllZerosV,
1934 (v8i16 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
1935 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
1936 def : Pat<(v16i8 (vector_shuffle immAllZerosV,
1937 (v16i8 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
1938 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
1939 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
1940 def : Pat<(v2f64 (vector_shuffle immAllZerosV,
1941 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
1942 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
1943 def : Pat<(v4f32 (vector_shuffle immAllZerosV,
1944 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
1945 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
1948 // Splat v2f64 / v2i64
1949 let AddedComplexity = 10 in {
1950 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
1951 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1952 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
1953 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1954 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
1955 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1956 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
1957 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1961 def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SSE_splat_mask:$sm),
1962 (SHUFPSrri VR128:$src, VR128:$src, SSE_splat_mask:$sm)>,
1963 Requires<[HasSSE1]>;
1965 // Special unary SHUFPSrri case.
1966 // FIXME: when we want non two-address code, then we should use PSHUFD?
1967 def : Pat<(vector_shuffle (v4f32 VR128:$src1), (undef),
1968 SHUFP_unary_shuffle_mask:$sm),
1969 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
1970 Requires<[HasSSE1]>;
1971 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
1972 def : Pat<(vector_shuffle (loadv4f32 addr:$src1), (undef),
1973 SHUFP_unary_shuffle_mask:$sm),
1974 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
1975 Requires<[HasSSE2]>;
1976 // Special binary v4i32 shuffle cases with SHUFPS.
1977 def : Pat<(vector_shuffle (v4i32 VR128:$src1), (v4i32 VR128:$src2),
1978 PSHUFD_binary_shuffle_mask:$sm),
1979 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
1980 Requires<[HasSSE2]>;
1981 def : Pat<(vector_shuffle (v4i32 VR128:$src1),
1982 (bc_v4i32 (loadv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm),
1983 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
1984 Requires<[HasSSE2]>;
1986 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
1987 let AddedComplexity = 10 in {
1988 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
1989 UNPCKL_v_undef_shuffle_mask)),
1990 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1991 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
1992 UNPCKL_v_undef_shuffle_mask)),
1993 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1994 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
1995 UNPCKL_v_undef_shuffle_mask)),
1996 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1997 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
1998 UNPCKL_v_undef_shuffle_mask)),
1999 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
2002 let AddedComplexity = 15 in
2003 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2004 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2005 MOVSHDUP_shuffle_mask)),
2006 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2007 let AddedComplexity = 20 in
2008 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (loadv2i64 addr:$src)), (undef),
2009 MOVSHDUP_shuffle_mask)),
2010 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2012 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2013 let AddedComplexity = 15 in
2014 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
2015 MOVSLDUP_shuffle_mask)),
2016 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2017 let AddedComplexity = 20 in
2018 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (loadv2i64 addr:$src)), (undef),
2019 MOVSLDUP_shuffle_mask)),
2020 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2022 let AddedComplexity = 15 in {
2023 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2024 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2025 MOVHP_shuffle_mask)),
2026 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
2028 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
2029 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2030 MOVHLPS_shuffle_mask)),
2031 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2033 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2034 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2035 MOVHLPS_v_undef_shuffle_mask)),
2036 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2037 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2038 MOVHLPS_v_undef_shuffle_mask)),
2039 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2042 let AddedComplexity = 20 in {
2043 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2044 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2045 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (loadv4f32 addr:$src2),
2046 MOVLP_shuffle_mask)),
2047 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2048 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (loadv2f64 addr:$src2),
2049 MOVLP_shuffle_mask)),
2050 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2051 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (loadv4f32 addr:$src2),
2052 MOVHP_shuffle_mask)),
2053 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2054 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (loadv2f64 addr:$src2),
2055 MOVHP_shuffle_mask)),
2056 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2058 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)),
2059 MOVLP_shuffle_mask)),
2060 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2061 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (loadv2i64 addr:$src2),
2062 MOVLP_shuffle_mask)),
2063 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2064 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)),
2065 MOVHP_shuffle_mask)),
2066 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2067 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (loadv2i64 addr:$src2),
2068 MOVLP_shuffle_mask)),
2069 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2072 let AddedComplexity = 15 in {
2073 // Setting the lowest element in the vector.
2074 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2075 MOVL_shuffle_mask)),
2076 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2077 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2078 MOVL_shuffle_mask)),
2079 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2081 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2082 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2083 MOVLP_shuffle_mask)),
2084 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2085 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2086 MOVLP_shuffle_mask)),
2087 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2090 // Set lowest element and zero upper elements.
2091 let AddedComplexity = 20 in
2092 def : Pat<(bc_v2i64 (vector_shuffle immAllZerosV,
2093 (v2f64 (scalar_to_vector (loadf64 addr:$src))),
2094 MOVL_shuffle_mask)),
2095 (MOVZQI2PQIrm addr:$src)>, Requires<[HasSSE2]>;
2097 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2098 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2099 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2100 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2101 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2102 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2103 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2104 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2105 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2106 Requires<[HasSSE2]>;
2107 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2108 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2109 Requires<[HasSSE2]>;
2110 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2111 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2112 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2113 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2114 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2115 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2116 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2117 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2118 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
2119 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2120 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
2121 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2122 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
2123 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2124 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
2125 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2127 // Some special case pandn patterns.
2128 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2130 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2131 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2133 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2134 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2136 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2138 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2139 (load addr:$src2))),
2140 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2141 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2142 (load addr:$src2))),
2143 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2144 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2145 (load addr:$src2))),
2146 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2149 def : Pat<(v4f32 (X86loadu addr:$src)), (MOVUPSrm addr:$src)>,
2150 Requires<[HasSSE1]>;