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 class SSI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
196 : I<o, F, ops, asm, pattern>, XS, Requires<[HasSSE1]>;
197 class SDI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
198 : I<o, F, ops, asm, pattern>, XD, Requires<[HasSSE2]>;
199 class PSI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
200 : I<o, F, ops, asm, pattern>, TB, Requires<[HasSSE1]>;
201 class PDI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
202 : I<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE2]>;
203 class PSIi8<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
204 : Ii8<o, F, ops, asm, pattern>, TB, Requires<[HasSSE1]>;
205 class PDIi8<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
206 : Ii8<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE2]>;
208 class S3SI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
209 : I<o, F, ops, asm, pattern>, XS, Requires<[HasSSE3]>;
210 class S3DI<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
211 : I<o, F, ops, asm, pattern>, XD, Requires<[HasSSE3]>;
212 class S3I<bits<8> o, Format F, dag ops, string asm, list<dag> pattern>
213 : I<o, F, ops, asm, pattern>, TB, OpSize, Requires<[HasSSE3]>;
215 //===----------------------------------------------------------------------===//
216 // Helpers for defining instructions that directly correspond to intrinsics.
218 multiclass SS_IntUnary<bits<8> o, string OpcodeStr, Intrinsic IntId> {
219 def r : SSI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
220 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src"),
221 [(set VR128:$dst, (v4f32 (IntId VR128:$src)))]>;
222 def m : SSI<o, MRMSrcMem, (ops VR128:$dst, ssmem:$src),
223 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src"),
224 [(set VR128:$dst, (v4f32 (IntId sse_load_f32:$src)))]>;
227 multiclass SD_IntUnary<bits<8> o, string OpcodeStr, Intrinsic IntId> {
228 def r : SDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
229 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src"),
230 [(set VR128:$dst, (v2f64 (IntId VR128:$src)))]>;
231 def m : SDI<o, MRMSrcMem, (ops VR128:$dst, sdmem:$src),
232 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src"),
233 [(set VR128:$dst, (v2f64 (IntId sse_load_f64:$src)))]>;
236 class PS_Intr<bits<8> o, string OpcodeStr, Intrinsic IntId>
237 : PSI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
238 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
239 [(set VR128:$dst, (IntId VR128:$src))]>;
240 class PS_Intm<bits<8> o, string OpcodeStr, Intrinsic IntId>
241 : PSI<o, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
242 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
243 [(set VR128:$dst, (IntId (load addr:$src)))]>;
244 class PD_Intr<bits<8> o, string OpcodeStr, Intrinsic IntId>
245 : PDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src),
246 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
247 [(set VR128:$dst, (IntId VR128:$src))]>;
248 class PD_Intm<bits<8> o, string OpcodeStr, Intrinsic IntId>
249 : PDI<o, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
250 !strconcat(OpcodeStr, " {$src, $dst|$dst, $src}"),
251 [(set VR128:$dst, (IntId (load addr:$src)))]>;
253 class PS_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
254 : PSI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
255 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
256 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
257 class PS_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
258 : PSI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f32mem:$src2),
259 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
260 [(set VR128:$dst, (IntId VR128:$src1, (load addr:$src2)))]>;
261 class PD_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
262 : PDI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
263 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
264 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
265 class PD_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
266 : PDI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
267 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
268 [(set VR128:$dst, (IntId VR128:$src1, (load addr:$src2)))]>;
270 // Some 'special' instructions
271 def IMPLICIT_DEF_FR32 : I<0, Pseudo, (ops FR32:$dst),
272 "#IMPLICIT_DEF $dst",
273 [(set FR32:$dst, (undef))]>, Requires<[HasSSE2]>;
274 def IMPLICIT_DEF_FR64 : I<0, Pseudo, (ops FR64:$dst),
275 "#IMPLICIT_DEF $dst",
276 [(set FR64:$dst, (undef))]>, Requires<[HasSSE2]>;
278 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded by the
279 // scheduler into a branch sequence.
280 let usesCustomDAGSchedInserter = 1 in { // Expanded by the scheduler.
281 def CMOV_FR32 : I<0, Pseudo,
282 (ops FR32:$dst, FR32:$t, FR32:$f, i8imm:$cond),
283 "#CMOV_FR32 PSEUDO!",
284 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond))]>;
285 def CMOV_FR64 : I<0, Pseudo,
286 (ops FR64:$dst, FR64:$t, FR64:$f, i8imm:$cond),
287 "#CMOV_FR64 PSEUDO!",
288 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond))]>;
289 def CMOV_V4F32 : I<0, Pseudo,
290 (ops VR128:$dst, VR128:$t, VR128:$f, i8imm:$cond),
291 "#CMOV_V4F32 PSEUDO!",
293 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
294 def CMOV_V2F64 : I<0, Pseudo,
295 (ops VR128:$dst, VR128:$t, VR128:$f, i8imm:$cond),
296 "#CMOV_V2F64 PSEUDO!",
298 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
299 def CMOV_V2I64 : I<0, Pseudo,
300 (ops VR128:$dst, VR128:$t, VR128:$f, i8imm:$cond),
301 "#CMOV_V2I64 PSEUDO!",
303 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond)))]>;
307 def MOVSSrr : SSI<0x10, MRMSrcReg, (ops FR32:$dst, FR32:$src),
308 "movss {$src, $dst|$dst, $src}", []>;
309 def MOVSSrm : SSI<0x10, MRMSrcMem, (ops FR32:$dst, f32mem:$src),
310 "movss {$src, $dst|$dst, $src}",
311 [(set FR32:$dst, (loadf32 addr:$src))]>;
312 def MOVSDrr : SDI<0x10, MRMSrcReg, (ops FR64:$dst, FR64:$src),
313 "movsd {$src, $dst|$dst, $src}", []>;
314 def MOVSDrm : SDI<0x10, MRMSrcMem, (ops FR64:$dst, f64mem:$src),
315 "movsd {$src, $dst|$dst, $src}",
316 [(set FR64:$dst, (loadf64 addr:$src))]>;
318 def MOVSSmr : SSI<0x11, MRMDestMem, (ops f32mem:$dst, FR32:$src),
319 "movss {$src, $dst|$dst, $src}",
320 [(store FR32:$src, addr:$dst)]>;
321 def MOVSDmr : SDI<0x11, MRMDestMem, (ops f64mem:$dst, FR64:$src),
322 "movsd {$src, $dst|$dst, $src}",
323 [(store FR64:$src, addr:$dst)]>;
325 /// scalar_sse12_fp_binop_rm - Scalar SSE binops come in four basic forms:
326 /// 1. f32 vs f64 - These come in SSE1/SSE2 forms for float/doubles.
327 /// 2. rr vs rm - They include a reg+reg form and a ref+mem form.
329 /// In addition, scalar SSE ops have an intrinsic form. This form is unlike the
330 /// normal form, in that they take an entire vector (instead of a scalar) and
331 /// leave the top elements undefined. This adds another two variants of the
332 /// above permutations, giving us 8 forms for 'instruction'.
334 let isTwoAddress = 1 in {
335 multiclass scalar_sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
336 SDNode OpNode, Intrinsic F32Int,
337 Intrinsic F64Int, bit Commutable = 0> {
338 // Scalar operation, reg+reg.
339 def SSrr : SSI<opc, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
340 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2"),
341 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
342 let isCommutable = Commutable;
344 def SDrr : SDI<opc, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
345 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2"),
346 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
347 let isCommutable = Commutable;
349 // Scalar operation, reg+mem.
350 def SSrm : SSI<opc, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f32mem:$src2),
351 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2"),
352 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
353 def SDrm : SDI<opc, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f64mem:$src2),
354 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2"),
355 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
357 // Vector intrinsic operation, reg+reg.
358 def SSrr_Int : SSI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
359 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2"),
360 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2))]> {
361 let isCommutable = Commutable;
363 def SDrr_Int : SDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
364 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2"),
365 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2))]> {
366 let isCommutable = Commutable;
368 // Vector intrinsic operation, reg+mem.
369 def SSrm_Int : SSI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, ssmem:$src2),
370 !strconcat(OpcodeStr, "ss {$src2, $dst|$dst, $src2"),
371 [(set VR128:$dst, (F32Int VR128:$src1,
372 sse_load_f32:$src2))]>;
373 def SDrm_Int : SDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, sdmem:$src2),
374 !strconcat(OpcodeStr, "sd {$src2, $dst|$dst, $src2"),
375 [(set VR128:$dst, (F64Int VR128:$src1,
376 sse_load_f64:$src2))]>;
380 // Arithmetic instructions
382 defm ADD : scalar_sse12_fp_binop_rm<0x58, "add", fadd,
383 int_x86_sse_add_ss, int_x86_sse2_add_sd, 1>;
384 defm MUL : scalar_sse12_fp_binop_rm<0x59, "mul", fmul,
385 int_x86_sse_mul_ss, int_x86_sse2_mul_sd, 1>;
386 defm SUB : scalar_sse12_fp_binop_rm<0x5C, "sub", fsub,
387 int_x86_sse_sub_ss, int_x86_sse2_sub_sd>;
388 defm DIV : scalar_sse12_fp_binop_rm<0x5E, "div", fdiv,
389 int_x86_sse_div_ss, int_x86_sse2_div_sd>;
391 defm MAX : scalar_sse12_fp_binop_rm<0x5F, "max", X86fmax,
392 int_x86_sse_max_ss, int_x86_sse2_max_sd>;
393 defm MIN : scalar_sse12_fp_binop_rm<0x5D, "min", X86fmin,
394 int_x86_sse_min_ss, int_x86_sse2_min_sd>;
397 def SQRTSSr : SSI<0x51, MRMSrcReg, (ops FR32:$dst, FR32:$src),
398 "sqrtss {$src, $dst|$dst, $src}",
399 [(set FR32:$dst, (fsqrt FR32:$src))]>;
400 def SQRTSSm : SSI<0x51, MRMSrcMem, (ops FR32:$dst, f32mem:$src),
401 "sqrtss {$src, $dst|$dst, $src}",
402 [(set FR32:$dst, (fsqrt (loadf32 addr:$src)))]>;
403 def SQRTSDr : SDI<0x51, MRMSrcReg, (ops FR64:$dst, FR64:$src),
404 "sqrtsd {$src, $dst|$dst, $src}",
405 [(set FR64:$dst, (fsqrt FR64:$src))]>;
406 def SQRTSDm : SDI<0x51, MRMSrcMem, (ops FR64:$dst, f64mem:$src),
407 "sqrtsd {$src, $dst|$dst, $src}",
408 [(set FR64:$dst, (fsqrt (loadf64 addr:$src)))]>;
410 // Aliases to match intrinsics which expect XMM operand(s).
412 defm SQRTSS_Int : SS_IntUnary<0x51, "sqrtss" , int_x86_sse_sqrt_ss>;
413 defm SQRTSD_Int : SD_IntUnary<0x51, "sqrtsd" , int_x86_sse2_sqrt_sd>;
414 defm RSQRTSS_Int : SS_IntUnary<0x52, "rsqrtss", int_x86_sse_rsqrt_ss>;
415 defm RCPSS_Int : SS_IntUnary<0x53, "rcpss" , int_x86_sse_rcp_ss>;
417 // Conversion instructions
418 def CVTTSS2SIrr: SSI<0x2C, MRMSrcReg, (ops GR32:$dst, FR32:$src),
419 "cvttss2si {$src, $dst|$dst, $src}",
420 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
421 def CVTTSS2SIrm: SSI<0x2C, MRMSrcMem, (ops GR32:$dst, f32mem:$src),
422 "cvttss2si {$src, $dst|$dst, $src}",
423 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
424 def CVTTSD2SIrr: SDI<0x2C, MRMSrcReg, (ops GR32:$dst, FR64:$src),
425 "cvttsd2si {$src, $dst|$dst, $src}",
426 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
427 def CVTTSD2SIrm: SDI<0x2C, MRMSrcMem, (ops GR32:$dst, f64mem:$src),
428 "cvttsd2si {$src, $dst|$dst, $src}",
429 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
430 def CVTSD2SSrr: SDI<0x5A, MRMSrcReg, (ops FR32:$dst, FR64:$src),
431 "cvtsd2ss {$src, $dst|$dst, $src}",
432 [(set FR32:$dst, (fround FR64:$src))]>;
433 def CVTSD2SSrm: SDI<0x5A, MRMSrcMem, (ops FR32:$dst, f64mem:$src),
434 "cvtsd2ss {$src, $dst|$dst, $src}",
435 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
436 def CVTSI2SSrr: SSI<0x2A, MRMSrcReg, (ops FR32:$dst, GR32:$src),
437 "cvtsi2ss {$src, $dst|$dst, $src}",
438 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
439 def CVTSI2SSrm: SSI<0x2A, MRMSrcMem, (ops FR32:$dst, i32mem:$src),
440 "cvtsi2ss {$src, $dst|$dst, $src}",
441 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
442 def CVTSI2SDrr: SDI<0x2A, MRMSrcReg, (ops FR64:$dst, GR32:$src),
443 "cvtsi2sd {$src, $dst|$dst, $src}",
444 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
445 def CVTSI2SDrm: SDI<0x2A, MRMSrcMem, (ops FR64:$dst, i32mem:$src),
446 "cvtsi2sd {$src, $dst|$dst, $src}",
447 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
449 // SSE2 instructions with XS prefix
450 def CVTSS2SDrr: I<0x5A, MRMSrcReg, (ops FR64:$dst, FR32:$src),
451 "cvtss2sd {$src, $dst|$dst, $src}",
452 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
454 def CVTSS2SDrm: I<0x5A, MRMSrcMem, (ops FR64:$dst, f32mem:$src),
455 "cvtss2sd {$src, $dst|$dst, $src}",
456 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
459 // Match intrinsics which expect XMM operand(s).
460 def Int_CVTSS2SIrr: SSI<0x2D, MRMSrcReg, (ops GR32:$dst, VR128:$src),
461 "cvtss2si {$src, $dst|$dst, $src}",
462 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
463 def Int_CVTSS2SIrm: SSI<0x2D, MRMSrcMem, (ops GR32:$dst, f32mem:$src),
464 "cvtss2si {$src, $dst|$dst, $src}",
465 [(set GR32:$dst, (int_x86_sse_cvtss2si
466 (load addr:$src)))]>;
467 def Int_CVTSD2SIrr: SDI<0x2D, MRMSrcReg, (ops GR32:$dst, VR128:$src),
468 "cvtsd2si {$src, $dst|$dst, $src}",
469 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
470 def Int_CVTSD2SIrm: SDI<0x2D, MRMSrcMem, (ops GR32:$dst, f128mem:$src),
471 "cvtsd2si {$src, $dst|$dst, $src}",
472 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
473 (load addr:$src)))]>;
475 // Aliases for intrinsics
476 def Int_CVTTSS2SIrr: SSI<0x2C, MRMSrcReg, (ops GR32:$dst, VR128:$src),
477 "cvttss2si {$src, $dst|$dst, $src}",
478 [(set GR32:$dst, (int_x86_sse_cvttss2si VR128:$src))]>;
479 def Int_CVTTSS2SIrm: SSI<0x2C, MRMSrcMem, (ops GR32:$dst, f32mem:$src),
480 "cvttss2si {$src, $dst|$dst, $src}",
481 [(set GR32:$dst, (int_x86_sse_cvttss2si(load addr:$src)))]>;
482 def Int_CVTTSD2SIrr: SDI<0x2C, MRMSrcReg, (ops GR32:$dst, VR128:$src),
483 "cvttsd2si {$src, $dst|$dst, $src}",
484 [(set GR32:$dst, (int_x86_sse2_cvttsd2si VR128:$src))]>;
485 def Int_CVTTSD2SIrm: SDI<0x2C, MRMSrcMem, (ops GR32:$dst, f128mem:$src),
486 "cvttsd2si {$src, $dst|$dst, $src}",
487 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
488 (load addr:$src)))]>;
490 let isTwoAddress = 1 in {
491 def Int_CVTSI2SSrr: SSI<0x2A, MRMSrcReg,
492 (ops VR128:$dst, VR128:$src1, GR32:$src2),
493 "cvtsi2ss {$src2, $dst|$dst, $src2}",
494 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
496 def Int_CVTSI2SSrm: SSI<0x2A, MRMSrcMem,
497 (ops VR128:$dst, VR128:$src1, i32mem:$src2),
498 "cvtsi2ss {$src2, $dst|$dst, $src2}",
499 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
500 (loadi32 addr:$src2)))]>;
503 // Comparison instructions
504 let isTwoAddress = 1 in {
505 def CMPSSrr : SSI<0xC2, MRMSrcReg,
506 (ops FR32:$dst, FR32:$src1, FR32:$src, SSECC:$cc),
507 "cmp${cc}ss {$src, $dst|$dst, $src}",
509 def CMPSSrm : SSI<0xC2, MRMSrcMem,
510 (ops FR32:$dst, FR32:$src1, f32mem:$src, SSECC:$cc),
511 "cmp${cc}ss {$src, $dst|$dst, $src}", []>;
512 def CMPSDrr : SDI<0xC2, MRMSrcReg,
513 (ops FR64:$dst, FR64:$src1, FR64:$src, SSECC:$cc),
514 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
515 def CMPSDrm : SDI<0xC2, MRMSrcMem,
516 (ops FR64:$dst, FR64:$src1, f64mem:$src, SSECC:$cc),
517 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
520 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (ops FR32:$src1, FR32:$src2),
521 "ucomiss {$src2, $src1|$src1, $src2}",
522 [(X86cmp FR32:$src1, FR32:$src2)]>;
523 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (ops FR32:$src1, f32mem:$src2),
524 "ucomiss {$src2, $src1|$src1, $src2}",
525 [(X86cmp FR32:$src1, (loadf32 addr:$src2))]>;
526 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (ops FR64:$src1, FR64:$src2),
527 "ucomisd {$src2, $src1|$src1, $src2}",
528 [(X86cmp FR64:$src1, FR64:$src2)]>;
529 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (ops FR64:$src1, f64mem:$src2),
530 "ucomisd {$src2, $src1|$src1, $src2}",
531 [(X86cmp FR64:$src1, (loadf64 addr:$src2))]>;
533 // Aliases to match intrinsics which expect XMM operand(s).
534 let isTwoAddress = 1 in {
535 def Int_CMPSSrr : SSI<0xC2, MRMSrcReg,
536 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
537 "cmp${cc}ss {$src, $dst|$dst, $src}",
538 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
539 VR128:$src, imm:$cc))]>;
540 def Int_CMPSSrm : SSI<0xC2, MRMSrcMem,
541 (ops VR128:$dst, VR128:$src1, f32mem:$src, SSECC:$cc),
542 "cmp${cc}ss {$src, $dst|$dst, $src}",
543 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
544 (load addr:$src), imm:$cc))]>;
545 def Int_CMPSDrr : SDI<0xC2, MRMSrcReg,
546 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
547 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
548 def Int_CMPSDrm : SDI<0xC2, MRMSrcMem,
549 (ops VR128:$dst, VR128:$src1, f64mem:$src, SSECC:$cc),
550 "cmp${cc}sd {$src, $dst|$dst, $src}", []>;
553 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
554 "ucomiss {$src2, $src1|$src1, $src2}",
555 [(X86ucomi (v4f32 VR128:$src1), VR128:$src2)]>;
556 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
557 "ucomiss {$src2, $src1|$src1, $src2}",
558 [(X86ucomi (v4f32 VR128:$src1), (load addr:$src2))]>;
559 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
560 "ucomisd {$src2, $src1|$src1, $src2}",
561 [(X86ucomi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
562 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
563 "ucomisd {$src2, $src1|$src1, $src2}",
564 [(X86ucomi (v2f64 VR128:$src1), (load addr:$src2))]>;
566 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
567 "comiss {$src2, $src1|$src1, $src2}",
568 [(X86comi (v4f32 VR128:$src1), VR128:$src2)]>;
569 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
570 "comiss {$src2, $src1|$src1, $src2}",
571 [(X86comi (v4f32 VR128:$src1), (load addr:$src2))]>;
572 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (ops VR128:$src1, VR128:$src2),
573 "comisd {$src2, $src1|$src1, $src2}",
574 [(X86comi (v2f64 VR128:$src1), (v2f64 VR128:$src2))]>;
575 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (ops VR128:$src1, f128mem:$src2),
576 "comisd {$src2, $src1|$src1, $src2}",
577 [(X86comi (v2f64 VR128:$src1), (load addr:$src2))]>;
579 // Aliases of packed instructions for scalar use. These all have names that
582 // Alias instructions that map fld0 to pxor for sse.
583 def FsFLD0SS : I<0xEF, MRMInitReg, (ops FR32:$dst),
584 "pxor $dst, $dst", [(set FR32:$dst, fp32imm0)]>,
585 Requires<[HasSSE1]>, TB, OpSize;
586 def FsFLD0SD : I<0xEF, MRMInitReg, (ops FR64:$dst),
587 "pxor $dst, $dst", [(set FR64:$dst, fp64imm0)]>,
588 Requires<[HasSSE2]>, TB, OpSize;
590 // Alias instructions to do FR32 / FR64 reg-to-reg copy using movaps / movapd.
591 // Upper bits are disregarded.
592 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (ops FR32:$dst, FR32:$src),
593 "movaps {$src, $dst|$dst, $src}", []>;
594 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (ops FR64:$dst, FR64:$src),
595 "movapd {$src, $dst|$dst, $src}", []>;
597 // Alias instructions to load FR32 / FR64 from f128mem using movaps / movapd.
598 // Upper bits are disregarded.
599 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (ops FR32:$dst, f128mem:$src),
600 "movaps {$src, $dst|$dst, $src}",
601 [(set FR32:$dst, (X86loadpf32 addr:$src))]>;
602 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (ops FR64:$dst, f128mem:$src),
603 "movapd {$src, $dst|$dst, $src}",
604 [(set FR64:$dst, (X86loadpf64 addr:$src))]>;
606 // Alias bitwise logical operations using SSE logical ops on packed FP values.
607 let isTwoAddress = 1 in {
608 let isCommutable = 1 in {
609 def FsANDPSrr : PSI<0x54, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
610 "andps {$src2, $dst|$dst, $src2}",
611 [(set FR32:$dst, (X86fand FR32:$src1, FR32:$src2))]>;
612 def FsANDPDrr : PDI<0x54, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
613 "andpd {$src2, $dst|$dst, $src2}",
614 [(set FR64:$dst, (X86fand FR64:$src1, FR64:$src2))]>;
615 def FsORPSrr : PSI<0x56, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
616 "orps {$src2, $dst|$dst, $src2}",
617 [(set FR32:$dst, (X86for FR32:$src1, FR32:$src2))]>;
618 def FsORPDrr : PDI<0x56, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
619 "orpd {$src2, $dst|$dst, $src2}",
620 [(set FR64:$dst, (X86for FR64:$src1, FR64:$src2))]>;
621 def FsXORPSrr : PSI<0x57, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
622 "xorps {$src2, $dst|$dst, $src2}",
623 [(set FR32:$dst, (X86fxor FR32:$src1, FR32:$src2))]>;
624 def FsXORPDrr : PDI<0x57, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
625 "xorpd {$src2, $dst|$dst, $src2}",
626 [(set FR64:$dst, (X86fxor FR64:$src1, FR64:$src2))]>;
628 def FsANDPSrm : PSI<0x54, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
629 "andps {$src2, $dst|$dst, $src2}",
630 [(set FR32:$dst, (X86fand FR32:$src1,
631 (X86loadpf32 addr:$src2)))]>;
632 def FsANDPDrm : PDI<0x54, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
633 "andpd {$src2, $dst|$dst, $src2}",
634 [(set FR64:$dst, (X86fand FR64:$src1,
635 (X86loadpf64 addr:$src2)))]>;
636 def FsORPSrm : PSI<0x56, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
637 "orps {$src2, $dst|$dst, $src2}",
638 [(set FR32:$dst, (X86for FR32:$src1,
639 (X86loadpf32 addr:$src2)))]>;
640 def FsORPDrm : PDI<0x56, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
641 "orpd {$src2, $dst|$dst, $src2}",
642 [(set FR64:$dst, (X86for FR64:$src1,
643 (X86loadpf64 addr:$src2)))]>;
644 def FsXORPSrm : PSI<0x57, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
645 "xorps {$src2, $dst|$dst, $src2}",
646 [(set FR32:$dst, (X86fxor FR32:$src1,
647 (X86loadpf32 addr:$src2)))]>;
648 def FsXORPDrm : PDI<0x57, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
649 "xorpd {$src2, $dst|$dst, $src2}",
650 [(set FR64:$dst, (X86fxor FR64:$src1,
651 (X86loadpf64 addr:$src2)))]>;
653 def FsANDNPSrr : PSI<0x55, MRMSrcReg, (ops FR32:$dst, FR32:$src1, FR32:$src2),
654 "andnps {$src2, $dst|$dst, $src2}", []>;
655 def FsANDNPSrm : PSI<0x55, MRMSrcMem, (ops FR32:$dst, FR32:$src1, f128mem:$src2),
656 "andnps {$src2, $dst|$dst, $src2}", []>;
657 def FsANDNPDrr : PDI<0x55, MRMSrcReg, (ops FR64:$dst, FR64:$src1, FR64:$src2),
658 "andnpd {$src2, $dst|$dst, $src2}", []>;
659 def FsANDNPDrm : PDI<0x55, MRMSrcMem, (ops FR64:$dst, FR64:$src1, f128mem:$src2),
660 "andnpd {$src2, $dst|$dst, $src2}", []>;
663 //===----------------------------------------------------------------------===//
664 // SSE packed FP Instructions
665 //===----------------------------------------------------------------------===//
667 // Some 'special' instructions
668 def IMPLICIT_DEF_VR128 : I<0, Pseudo, (ops VR128:$dst),
669 "#IMPLICIT_DEF $dst",
670 [(set VR128:$dst, (v4f32 (undef)))]>,
674 def MOVAPSrr : PSI<0x28, MRMSrcReg, (ops VR128:$dst, VR128:$src),
675 "movaps {$src, $dst|$dst, $src}", []>;
676 def MOVAPSrm : PSI<0x28, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
677 "movaps {$src, $dst|$dst, $src}",
678 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
679 def MOVAPDrr : PDI<0x28, MRMSrcReg, (ops VR128:$dst, VR128:$src),
680 "movapd {$src, $dst|$dst, $src}", []>;
681 def MOVAPDrm : PDI<0x28, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
682 "movapd {$src, $dst|$dst, $src}",
683 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
685 def MOVAPSmr : PSI<0x29, MRMDestMem, (ops f128mem:$dst, VR128:$src),
686 "movaps {$src, $dst|$dst, $src}",
687 [(store (v4f32 VR128:$src), addr:$dst)]>;
688 def MOVAPDmr : PDI<0x29, MRMDestMem, (ops f128mem:$dst, VR128:$src),
689 "movapd {$src, $dst|$dst, $src}",
690 [(store (v2f64 VR128:$src), addr:$dst)]>;
692 def MOVUPSrr : PSI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src),
693 "movups {$src, $dst|$dst, $src}", []>;
694 def MOVUPSrm : PSI<0x10, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
695 "movups {$src, $dst|$dst, $src}",
696 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
697 def MOVUPSmr : PSI<0x11, MRMDestMem, (ops f128mem:$dst, VR128:$src),
698 "movups {$src, $dst|$dst, $src}",
699 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
700 def MOVUPDrr : PDI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src),
701 "movupd {$src, $dst|$dst, $src}", []>;
702 def MOVUPDrm : PDI<0x10, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
703 "movupd {$src, $dst|$dst, $src}",
704 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
705 def MOVUPDmr : PDI<0x11, MRMDestMem, (ops f128mem:$dst, VR128:$src),
706 "movupd {$src, $dst|$dst, $src}",
707 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
709 let isTwoAddress = 1 in {
710 let AddedComplexity = 20 in {
711 def MOVLPSrm : PSI<0x12, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
712 "movlps {$src2, $dst|$dst, $src2}",
714 (v4f32 (vector_shuffle VR128:$src1,
715 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
716 MOVLP_shuffle_mask)))]>;
717 def MOVLPDrm : PDI<0x12, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
718 "movlpd {$src2, $dst|$dst, $src2}",
720 (v2f64 (vector_shuffle VR128:$src1,
721 (scalar_to_vector (loadf64 addr:$src2)),
722 MOVLP_shuffle_mask)))]>;
723 def MOVHPSrm : PSI<0x16, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
724 "movhps {$src2, $dst|$dst, $src2}",
726 (v4f32 (vector_shuffle VR128:$src1,
727 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2)))),
728 MOVHP_shuffle_mask)))]>;
729 def MOVHPDrm : PDI<0x16, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f64mem:$src2),
730 "movhpd {$src2, $dst|$dst, $src2}",
732 (v2f64 (vector_shuffle VR128:$src1,
733 (scalar_to_vector (loadf64 addr:$src2)),
734 MOVHP_shuffle_mask)))]>;
738 def MOVLPSmr : PSI<0x13, MRMDestMem, (ops f64mem:$dst, VR128:$src),
739 "movlps {$src, $dst|$dst, $src}",
740 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
741 (iPTR 0))), addr:$dst)]>;
742 def MOVLPDmr : PDI<0x13, MRMDestMem, (ops f64mem:$dst, VR128:$src),
743 "movlpd {$src, $dst|$dst, $src}",
744 [(store (f64 (vector_extract (v2f64 VR128:$src),
745 (iPTR 0))), addr:$dst)]>;
747 // v2f64 extract element 1 is always custom lowered to unpack high to low
748 // and extract element 0 so the non-store version isn't too horrible.
749 def MOVHPSmr : PSI<0x17, MRMDestMem, (ops f64mem:$dst, VR128:$src),
750 "movhps {$src, $dst|$dst, $src}",
751 [(store (f64 (vector_extract
752 (v2f64 (vector_shuffle
753 (bc_v2f64 (v4f32 VR128:$src)), (undef),
754 UNPCKH_shuffle_mask)), (iPTR 0))),
756 def MOVHPDmr : PDI<0x17, MRMDestMem, (ops f64mem:$dst, VR128:$src),
757 "movhpd {$src, $dst|$dst, $src}",
758 [(store (f64 (vector_extract
759 (v2f64 (vector_shuffle VR128:$src, (undef),
760 UNPCKH_shuffle_mask)), (iPTR 0))),
763 let isTwoAddress = 1 in {
764 let AddedComplexity = 15 in {
765 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
766 "movlhps {$src2, $dst|$dst, $src2}",
768 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
769 MOVHP_shuffle_mask)))]>;
771 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
772 "movhlps {$src2, $dst|$dst, $src2}",
774 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
775 MOVHLPS_shuffle_mask)))]>;
779 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (ops VR128:$dst, VR128:$src),
780 "movshdup {$src, $dst|$dst, $src}",
781 [(set VR128:$dst, (v4f32 (vector_shuffle
783 MOVSHDUP_shuffle_mask)))]>;
784 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
785 "movshdup {$src, $dst|$dst, $src}",
786 [(set VR128:$dst, (v4f32 (vector_shuffle
787 (loadv4f32 addr:$src), (undef),
788 MOVSHDUP_shuffle_mask)))]>;
790 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src),
791 "movsldup {$src, $dst|$dst, $src}",
792 [(set VR128:$dst, (v4f32 (vector_shuffle
794 MOVSLDUP_shuffle_mask)))]>;
795 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
796 "movsldup {$src, $dst|$dst, $src}",
797 [(set VR128:$dst, (v4f32 (vector_shuffle
798 (loadv4f32 addr:$src), (undef),
799 MOVSLDUP_shuffle_mask)))]>;
801 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (ops VR128:$dst, VR128:$src),
802 "movddup {$src, $dst|$dst, $src}",
803 [(set VR128:$dst, (v2f64 (vector_shuffle
805 SSE_splat_lo_mask)))]>;
806 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
807 "movddup {$src, $dst|$dst, $src}",
808 [(set VR128:$dst, (v2f64 (vector_shuffle
809 (scalar_to_vector (loadf64 addr:$src)),
811 SSE_splat_lo_mask)))]>;
813 // SSE2 instructions without OpSize prefix
814 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
815 "cvtdq2ps {$src, $dst|$dst, $src}",
816 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
817 TB, Requires<[HasSSE2]>;
818 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
819 "cvtdq2ps {$src, $dst|$dst, $src}",
820 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
821 (bitconvert (loadv2i64 addr:$src))))]>,
822 TB, Requires<[HasSSE2]>;
824 // SSE2 instructions with XS prefix
825 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
826 "cvtdq2pd {$src, $dst|$dst, $src}",
827 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
828 XS, Requires<[HasSSE2]>;
829 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
830 "cvtdq2pd {$src, $dst|$dst, $src}",
831 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
832 (bitconvert (loadv2i64 addr:$src))))]>,
833 XS, Requires<[HasSSE2]>;
835 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
836 "cvtps2dq {$src, $dst|$dst, $src}",
837 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
838 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
839 "cvtps2dq {$src, $dst|$dst, $src}",
840 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
841 (load addr:$src)))]>;
842 // SSE2 packed instructions with XS prefix
843 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
844 "cvttps2dq {$src, $dst|$dst, $src}",
845 [(set VR128:$dst, (int_x86_sse2_cvttps2dq VR128:$src))]>,
846 XS, Requires<[HasSSE2]>;
847 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
848 "cvttps2dq {$src, $dst|$dst, $src}",
849 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
850 (load addr:$src)))]>,
851 XS, Requires<[HasSSE2]>;
853 // SSE2 packed instructions with XD prefix
854 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
855 "cvtpd2dq {$src, $dst|$dst, $src}",
856 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
857 XD, Requires<[HasSSE2]>;
858 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
859 "cvtpd2dq {$src, $dst|$dst, $src}",
860 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
861 (load addr:$src)))]>,
862 XD, Requires<[HasSSE2]>;
863 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (ops VR128:$dst, VR128:$src),
864 "cvttpd2dq {$src, $dst|$dst, $src}",
865 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
866 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (ops VR128:$dst, f128mem:$src),
867 "cvttpd2dq {$src, $dst|$dst, $src}",
868 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
869 (load addr:$src)))]>;
871 // SSE2 instructions without OpSize prefix
872 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (ops VR128:$dst, VR128:$src),
873 "cvtps2pd {$src, $dst|$dst, $src}",
874 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
875 TB, Requires<[HasSSE2]>;
876 def Int_CVTPS2PDrm : I<0x5A, MRMSrcReg, (ops VR128:$dst, f64mem:$src),
877 "cvtps2pd {$src, $dst|$dst, $src}",
878 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
879 (load addr:$src)))]>,
880 TB, Requires<[HasSSE2]>;
882 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (ops VR128:$dst, VR128:$src),
883 "cvtpd2ps {$src, $dst|$dst, $src}",
884 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
885 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcReg, (ops VR128:$dst, f128mem:$src),
886 "cvtpd2ps {$src, $dst|$dst, $src}",
887 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
888 (load addr:$src)))]>;
890 // Match intrinsics which expect XMM operand(s).
891 // Aliases for intrinsics
892 let isTwoAddress = 1 in {
893 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
894 (ops VR128:$dst, VR128:$src1, GR32:$src2),
895 "cvtsi2sd {$src2, $dst|$dst, $src2}",
896 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
898 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
899 (ops VR128:$dst, VR128:$src1, i32mem:$src2),
900 "cvtsi2sd {$src2, $dst|$dst, $src2}",
901 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
902 (loadi32 addr:$src2)))]>;
903 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
904 (ops VR128:$dst, VR128:$src1, VR128:$src2),
905 "cvtsd2ss {$src2, $dst|$dst, $src2}",
906 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
908 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
909 (ops VR128:$dst, VR128:$src1, f64mem:$src2),
910 "cvtsd2ss {$src2, $dst|$dst, $src2}",
911 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
912 (load addr:$src2)))]>;
913 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
914 (ops VR128:$dst, VR128:$src1, VR128:$src2),
915 "cvtss2sd {$src2, $dst|$dst, $src2}",
916 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
919 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
920 (ops VR128:$dst, VR128:$src1, f32mem:$src2),
921 "cvtss2sd {$src2, $dst|$dst, $src2}",
922 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
923 (load addr:$src2)))]>, XS,
927 /// packed_sse12_fp_binop_rm - Packed SSE binops come in four basic forms:
928 /// 1. v4f32 vs v2f64 - These come in SSE1/SSE2 forms for float/doubles.
929 /// 2. rr vs rm - They include a reg+reg form and a ref+mem form.
931 let isTwoAddress = 1 in {
932 multiclass packed_sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
933 SDNode OpNode, bit Commutable = 0> {
934 // Packed operation, reg+reg.
935 def PSrr : PSI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
936 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2"),
937 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
938 let isCommutable = Commutable;
940 def PDrr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
941 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2"),
942 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
943 let isCommutable = Commutable;
945 // Packed operation, reg+mem.
946 def PSrm : PSI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
947 !strconcat(OpcodeStr, "ps {$src2, $dst|$dst, $src2"),
948 [(set VR128:$dst, (OpNode VR128:$src1, (loadv4f32 addr:$src2)))]>;
949 def PDrm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
950 !strconcat(OpcodeStr, "pd {$src2, $dst|$dst, $src2"),
951 [(set VR128:$dst, (OpNode VR128:$src1, (loadv2f64 addr:$src2)))]>;
955 defm ADD : packed_sse12_fp_binop_rm<0x58, "add", fadd, 1>;
956 defm MUL : packed_sse12_fp_binop_rm<0x59, "mul", fmul, 1>;
957 defm DIV : packed_sse12_fp_binop_rm<0x5E, "div", fdiv>;
958 defm SUB : packed_sse12_fp_binop_rm<0x5C, "sub", fsub>;
961 let isTwoAddress = 1 in {
962 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
963 (ops VR128:$dst, VR128:$src1, VR128:$src2),
964 "addsubps {$src2, $dst|$dst, $src2}",
965 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
967 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
968 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
969 "addsubps {$src2, $dst|$dst, $src2}",
970 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
971 (load addr:$src2)))]>;
972 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
973 (ops VR128:$dst, VR128:$src1, VR128:$src2),
974 "addsubpd {$src2, $dst|$dst, $src2}",
975 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
977 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
978 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
979 "addsubpd {$src2, $dst|$dst, $src2}",
980 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
981 (load addr:$src2)))]>;
984 def SQRTPSr : PS_Intr<0x51, "sqrtps", int_x86_sse_sqrt_ps>;
985 def SQRTPSm : PS_Intm<0x51, "sqrtps", int_x86_sse_sqrt_ps>;
986 def SQRTPDr : PD_Intr<0x51, "sqrtpd", int_x86_sse2_sqrt_pd>;
987 def SQRTPDm : PD_Intm<0x51, "sqrtpd", int_x86_sse2_sqrt_pd>;
989 def RSQRTPSr : PS_Intr<0x52, "rsqrtps", int_x86_sse_rsqrt_ps>;
990 def RSQRTPSm : PS_Intm<0x52, "rsqrtps", int_x86_sse_rsqrt_ps>;
991 def RCPPSr : PS_Intr<0x53, "rcpps", int_x86_sse_rcp_ps>;
992 def RCPPSm : PS_Intm<0x53, "rcpps", int_x86_sse_rcp_ps>;
994 let isTwoAddress = 1 in {
995 let isCommutable = 1 in {
996 def MAXPSrr : PS_Intrr<0x5F, "maxps", int_x86_sse_max_ps>;
997 def MAXPDrr : PD_Intrr<0x5F, "maxpd", int_x86_sse2_max_pd>;
998 def MINPSrr : PS_Intrr<0x5D, "minps", int_x86_sse_min_ps>;
999 def MINPDrr : PD_Intrr<0x5D, "minpd", int_x86_sse2_min_pd>;
1001 def MAXPSrm : PS_Intrm<0x5F, "maxps", int_x86_sse_max_ps>;
1002 def MAXPDrm : PD_Intrm<0x5F, "maxpd", int_x86_sse2_max_pd>;
1003 def MINPSrm : PS_Intrm<0x5D, "minps", int_x86_sse_min_ps>;
1004 def MINPDrm : PD_Intrm<0x5D, "minpd", int_x86_sse2_min_pd>;
1008 let isTwoAddress = 1 in {
1009 let isCommutable = 1 in {
1010 def ANDPSrr : PSI<0x54, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1011 "andps {$src2, $dst|$dst, $src2}",
1012 [(set VR128:$dst, (v2i64 (and VR128:$src1, VR128:$src2)))]>;
1013 def ANDPDrr : PDI<0x54, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1014 "andpd {$src2, $dst|$dst, $src2}",
1016 (and (bc_v2i64 (v2f64 VR128:$src1)),
1017 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1018 def ORPSrr : PSI<0x56, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1019 "orps {$src2, $dst|$dst, $src2}",
1020 [(set VR128:$dst, (v2i64 (or VR128:$src1, VR128:$src2)))]>;
1021 def ORPDrr : PDI<0x56, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1022 "orpd {$src2, $dst|$dst, $src2}",
1024 (or (bc_v2i64 (v2f64 VR128:$src1)),
1025 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1026 def XORPSrr : PSI<0x57, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1027 "xorps {$src2, $dst|$dst, $src2}",
1028 [(set VR128:$dst, (v2i64 (xor VR128:$src1, VR128:$src2)))]>;
1029 def XORPDrr : PDI<0x57, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1030 "xorpd {$src2, $dst|$dst, $src2}",
1032 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1033 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1035 def ANDPSrm : PSI<0x54, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1036 "andps {$src2, $dst|$dst, $src2}",
1037 [(set VR128:$dst, (and VR128:$src1,
1038 (bc_v2i64 (loadv4f32 addr:$src2))))]>;
1039 def ANDPDrm : PDI<0x54, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1040 "andpd {$src2, $dst|$dst, $src2}",
1042 (and (bc_v2i64 (v2f64 VR128:$src1)),
1043 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1044 def ORPSrm : PSI<0x56, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1045 "orps {$src2, $dst|$dst, $src2}",
1046 [(set VR128:$dst, (or VR128:$src1,
1047 (bc_v2i64 (loadv4f32 addr:$src2))))]>;
1048 def ORPDrm : PDI<0x56, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1049 "orpd {$src2, $dst|$dst, $src2}",
1051 (or (bc_v2i64 (v2f64 VR128:$src1)),
1052 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1053 def XORPSrm : PSI<0x57, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1054 "xorps {$src2, $dst|$dst, $src2}",
1055 [(set VR128:$dst, (xor VR128:$src1,
1056 (bc_v2i64 (loadv4f32 addr:$src2))))]>;
1057 def XORPDrm : PDI<0x57, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1058 "xorpd {$src2, $dst|$dst, $src2}",
1060 (xor (bc_v2i64 (v2f64 VR128:$src1)),
1061 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1062 def ANDNPSrr : PSI<0x55, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1063 "andnps {$src2, $dst|$dst, $src2}",
1064 [(set VR128:$dst, (v2i64 (and (xor VR128:$src1,
1065 (bc_v2i64 (v4i32 immAllOnesV))),
1067 def ANDNPSrm : PSI<0x55, MRMSrcMem, (ops VR128:$dst, VR128:$src1,f128mem:$src2),
1068 "andnps {$src2, $dst|$dst, $src2}",
1069 [(set VR128:$dst, (v2i64 (and (xor VR128:$src1,
1070 (bc_v2i64 (v4i32 immAllOnesV))),
1071 (bc_v2i64 (loadv4f32 addr:$src2)))))]>;
1072 def ANDNPDrr : PDI<0x55, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1073 "andnpd {$src2, $dst|$dst, $src2}",
1075 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1076 (bc_v2i64 (v2f64 VR128:$src2))))]>;
1077 def ANDNPDrm : PDI<0x55, MRMSrcMem, (ops VR128:$dst, VR128:$src1,f128mem:$src2),
1078 "andnpd {$src2, $dst|$dst, $src2}",
1080 (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1081 (bc_v2i64 (loadv2f64 addr:$src2))))]>;
1084 let isTwoAddress = 1 in {
1085 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
1086 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
1087 "cmp${cc}ps {$src, $dst|$dst, $src}",
1088 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1089 VR128:$src, imm:$cc))]>;
1090 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
1091 (ops VR128:$dst, VR128:$src1, f128mem:$src, SSECC:$cc),
1092 "cmp${cc}ps {$src, $dst|$dst, $src}",
1093 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1094 (load addr:$src), imm:$cc))]>;
1095 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1096 (ops VR128:$dst, VR128:$src1, VR128:$src, SSECC:$cc),
1097 "cmp${cc}pd {$src, $dst|$dst, $src}",
1098 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1099 VR128:$src, imm:$cc))]>;
1100 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1101 (ops VR128:$dst, VR128:$src1, f128mem:$src, SSECC:$cc),
1102 "cmp${cc}pd {$src, $dst|$dst, $src}",
1103 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1104 (load addr:$src), imm:$cc))]>;
1107 // Shuffle and unpack instructions
1108 let isTwoAddress = 1 in {
1109 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
1110 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
1111 (ops VR128:$dst, VR128:$src1, VR128:$src2, i32i8imm:$src3),
1112 "shufps {$src3, $src2, $dst|$dst, $src2, $src3}",
1113 [(set VR128:$dst, (v4f32 (vector_shuffle
1114 VR128:$src1, VR128:$src2,
1115 SHUFP_shuffle_mask:$src3)))]>;
1116 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
1117 (ops VR128:$dst, VR128:$src1, f128mem:$src2, i32i8imm:$src3),
1118 "shufps {$src3, $src2, $dst|$dst, $src2, $src3}",
1119 [(set VR128:$dst, (v4f32 (vector_shuffle
1120 VR128:$src1, (load addr:$src2),
1121 SHUFP_shuffle_mask:$src3)))]>;
1122 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1123 (ops VR128:$dst, VR128:$src1, VR128:$src2, i8imm:$src3),
1124 "shufpd {$src3, $src2, $dst|$dst, $src2, $src3}",
1125 [(set VR128:$dst, (v2f64 (vector_shuffle
1126 VR128:$src1, VR128:$src2,
1127 SHUFP_shuffle_mask:$src3)))]>;
1128 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1129 (ops VR128:$dst, VR128:$src1, f128mem:$src2, i8imm:$src3),
1130 "shufpd {$src3, $src2, $dst|$dst, $src2, $src3}",
1131 [(set VR128:$dst, (v2f64 (vector_shuffle
1132 VR128:$src1, (load addr:$src2),
1133 SHUFP_shuffle_mask:$src3)))]>;
1135 let AddedComplexity = 10 in {
1136 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
1137 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1138 "unpckhps {$src2, $dst|$dst, $src2}",
1139 [(set VR128:$dst, (v4f32 (vector_shuffle
1140 VR128:$src1, VR128:$src2,
1141 UNPCKH_shuffle_mask)))]>;
1142 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
1143 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1144 "unpckhps {$src2, $dst|$dst, $src2}",
1145 [(set VR128:$dst, (v4f32 (vector_shuffle
1146 VR128:$src1, (load addr:$src2),
1147 UNPCKH_shuffle_mask)))]>;
1148 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1149 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1150 "unpckhpd {$src2, $dst|$dst, $src2}",
1151 [(set VR128:$dst, (v2f64 (vector_shuffle
1152 VR128:$src1, VR128:$src2,
1153 UNPCKH_shuffle_mask)))]>;
1154 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1155 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1156 "unpckhpd {$src2, $dst|$dst, $src2}",
1157 [(set VR128:$dst, (v2f64 (vector_shuffle
1158 VR128:$src1, (load addr:$src2),
1159 UNPCKH_shuffle_mask)))]>;
1161 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
1162 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1163 "unpcklps {$src2, $dst|$dst, $src2}",
1164 [(set VR128:$dst, (v4f32 (vector_shuffle
1165 VR128:$src1, VR128:$src2,
1166 UNPCKL_shuffle_mask)))]>;
1167 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
1168 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1169 "unpcklps {$src2, $dst|$dst, $src2}",
1170 [(set VR128:$dst, (v4f32 (vector_shuffle
1171 VR128:$src1, (load addr:$src2),
1172 UNPCKL_shuffle_mask)))]>;
1173 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1174 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1175 "unpcklpd {$src2, $dst|$dst, $src2}",
1176 [(set VR128:$dst, (v2f64 (vector_shuffle
1177 VR128:$src1, VR128:$src2,
1178 UNPCKL_shuffle_mask)))]>;
1179 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1180 (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1181 "unpcklpd {$src2, $dst|$dst, $src2}",
1182 [(set VR128:$dst, (v2f64 (vector_shuffle
1183 VR128:$src1, (load addr:$src2),
1184 UNPCKL_shuffle_mask)))]>;
1185 } // AddedComplexity
1190 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
1191 : S3DI<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1192 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1193 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
1194 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
1195 : S3DI<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1196 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1197 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (load addr:$src2))))]>;
1198 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
1199 : S3I<o, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1200 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1201 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
1202 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
1203 : S3I<o, MRMSrcMem, (ops VR128:$dst, VR128:$src1, f128mem:$src2),
1204 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2}"),
1205 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (load addr:$src2))))]>;
1207 let isTwoAddress = 1 in {
1208 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
1209 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
1210 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
1211 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
1212 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
1213 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
1214 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
1215 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
1218 //===----------------------------------------------------------------------===//
1219 // SSE integer instructions
1220 //===----------------------------------------------------------------------===//
1222 // Move Instructions
1223 def MOVDQArr : PDI<0x6F, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1224 "movdqa {$src, $dst|$dst, $src}", []>;
1225 def MOVDQArm : PDI<0x6F, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1226 "movdqa {$src, $dst|$dst, $src}",
1227 [(set VR128:$dst, (loadv2i64 addr:$src))]>;
1228 def MOVDQAmr : PDI<0x7F, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1229 "movdqa {$src, $dst|$dst, $src}",
1230 [(store (v2i64 VR128:$src), addr:$dst)]>;
1231 def MOVDQUrm : I<0x6F, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1232 "movdqu {$src, $dst|$dst, $src}",
1233 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1234 XS, Requires<[HasSSE2]>;
1235 def MOVDQUmr : I<0x7F, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1236 "movdqu {$src, $dst|$dst, $src}",
1237 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1238 XS, Requires<[HasSSE2]>;
1239 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (ops VR128:$dst, i128mem:$src),
1240 "lddqu {$src, $dst|$dst, $src}",
1241 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
1244 let isTwoAddress = 1 in {
1245 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1246 bit Commutable = 0> {
1247 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1248 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2"),
1249 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1250 let isCommutable = Commutable;
1252 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1253 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2"),
1254 [(set VR128:$dst, (IntId VR128:$src1,
1255 (bitconvert (loadv2i64 addr:$src2))))]>;
1259 let isTwoAddress = 1 in {
1260 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1261 string OpcodeStr, Intrinsic IntId> {
1262 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1263 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2"),
1264 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1265 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1266 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2"),
1267 [(set VR128:$dst, (IntId VR128:$src1,
1268 (bitconvert (loadv2i64 addr:$src2))))]>;
1269 def ri : PDIi8<opc2, ImmForm, (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1270 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2"),
1271 [(set VR128:$dst, (IntId VR128:$src1,
1272 (scalar_to_vector (i32 imm:$src2))))]>;
1277 let isTwoAddress = 1 in {
1278 /// PDI_binop_rm - Simple SSE2 binary operator.
1279 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1280 ValueType OpVT, bit Commutable = 0> {
1281 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1282 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2"),
1283 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1284 let isCommutable = Commutable;
1286 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1287 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2"),
1288 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1289 (bitconvert (loadv2i64 addr:$src2)))))]>;
1292 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1294 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1295 /// to collapse (bitconvert VT to VT) into its operand.
1297 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1298 bit Commutable = 0> {
1299 def rr : PDI<opc, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1300 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2"),
1301 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1302 let isCommutable = Commutable;
1304 def rm : PDI<opc, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1305 !strconcat(OpcodeStr, " {$src2, $dst|$dst, $src2"),
1306 [(set VR128:$dst, (OpNode VR128:$src1,(loadv2i64 addr:$src2)))]>;
1311 // 128-bit Integer Arithmetic
1313 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1314 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1315 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1316 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1318 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1319 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1320 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1321 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1323 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1324 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1325 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1326 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1328 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1329 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1330 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1331 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1333 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1335 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1336 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1337 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1339 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1341 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1342 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1345 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1346 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1347 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1348 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1349 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
1352 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw", int_x86_sse2_psll_w>;
1353 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld", int_x86_sse2_psll_d>;
1354 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq", int_x86_sse2_psll_q>;
1356 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw", int_x86_sse2_psrl_w>;
1357 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld", int_x86_sse2_psrl_d>;
1358 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq", int_x86_sse2_psrl_q>;
1360 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw", int_x86_sse2_psra_w>;
1361 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad", int_x86_sse2_psra_d>;
1362 // PSRAQ doesn't exist in SSE[1-3].
1365 // 128-bit logical shifts.
1366 let isTwoAddress = 1 in {
1367 def PSLLDQri : PDIi8<0x73, MRM7r, (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1368 "pslldq {$src2, $dst|$dst, $src2}", []>;
1369 def PSRLDQri : PDIi8<0x73, MRM3r, (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1370 "psrldq {$src2, $dst|$dst, $src2}", []>;
1371 // PSRADQri doesn't exist in SSE[1-3].
1374 let Predicates = [HasSSE2] in {
1375 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
1376 (v2i64 (PSLLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1377 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
1378 (v2i64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1379 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
1380 (v2f64 (PSRLDQri VR128:$src1, (PSxLDQ_imm imm:$src2)))>;
1384 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
1385 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
1386 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
1388 let isTwoAddress = 1 in {
1389 def PANDNrr : PDI<0xDF, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1390 "pandn {$src2, $dst|$dst, $src2}",
1391 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1394 def PANDNrm : PDI<0xDF, MRMSrcMem, (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1395 "pandn {$src2, $dst|$dst, $src2}",
1396 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
1397 (load addr:$src2))))]>;
1400 // SSE2 Integer comparison
1401 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
1402 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
1403 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
1404 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
1405 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
1406 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
1408 // Pack instructions
1409 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
1410 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
1411 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
1413 // Shuffle and unpack instructions
1414 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
1415 (ops VR128:$dst, VR128:$src1, i8imm:$src2),
1416 "pshufd {$src2, $src1, $dst|$dst, $src1, $src2}",
1417 [(set VR128:$dst, (v4i32 (vector_shuffle
1418 VR128:$src1, (undef),
1419 PSHUFD_shuffle_mask:$src2)))]>;
1420 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
1421 (ops VR128:$dst, i128mem:$src1, i8imm:$src2),
1422 "pshufd {$src2, $src1, $dst|$dst, $src1, $src2}",
1423 [(set VR128:$dst, (v4i32 (vector_shuffle
1424 (bc_v4i32(loadv2i64 addr:$src1)),
1426 PSHUFD_shuffle_mask:$src2)))]>;
1428 // SSE2 with ImmT == Imm8 and XS prefix.
1429 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
1430 (ops VR128:$dst, VR128:$src1, i8imm:$src2),
1431 "pshufhw {$src2, $src1, $dst|$dst, $src1, $src2}",
1432 [(set VR128:$dst, (v8i16 (vector_shuffle
1433 VR128:$src1, (undef),
1434 PSHUFHW_shuffle_mask:$src2)))]>,
1435 XS, Requires<[HasSSE2]>;
1436 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
1437 (ops VR128:$dst, i128mem:$src1, i8imm:$src2),
1438 "pshufhw {$src2, $src1, $dst|$dst, $src1, $src2}",
1439 [(set VR128:$dst, (v8i16 (vector_shuffle
1440 (bc_v8i16 (loadv2i64 addr:$src1)),
1442 PSHUFHW_shuffle_mask:$src2)))]>,
1443 XS, Requires<[HasSSE2]>;
1445 // SSE2 with ImmT == Imm8 and XD prefix.
1446 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
1447 (ops VR128:$dst, VR128:$src1, i32i8imm:$src2),
1448 "pshuflw {$src2, $src1, $dst|$dst, $src1, $src2}",
1449 [(set VR128:$dst, (v8i16 (vector_shuffle
1450 VR128:$src1, (undef),
1451 PSHUFLW_shuffle_mask:$src2)))]>,
1452 XD, Requires<[HasSSE2]>;
1453 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
1454 (ops VR128:$dst, i128mem:$src1, i32i8imm:$src2),
1455 "pshuflw {$src2, $src1, $dst|$dst, $src1, $src2}",
1456 [(set VR128:$dst, (v8i16 (vector_shuffle
1457 (bc_v8i16 (loadv2i64 addr:$src1)),
1459 PSHUFLW_shuffle_mask:$src2)))]>,
1460 XD, Requires<[HasSSE2]>;
1462 let isTwoAddress = 1 in {
1463 def PUNPCKLBWrr : PDI<0x60, MRMSrcReg,
1464 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1465 "punpcklbw {$src2, $dst|$dst, $src2}",
1467 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1468 UNPCKL_shuffle_mask)))]>;
1469 def PUNPCKLBWrm : PDI<0x60, MRMSrcMem,
1470 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1471 "punpcklbw {$src2, $dst|$dst, $src2}",
1473 (v16i8 (vector_shuffle VR128:$src1,
1474 (bc_v16i8 (loadv2i64 addr:$src2)),
1475 UNPCKL_shuffle_mask)))]>;
1476 def PUNPCKLWDrr : PDI<0x61, MRMSrcReg,
1477 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1478 "punpcklwd {$src2, $dst|$dst, $src2}",
1480 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1481 UNPCKL_shuffle_mask)))]>;
1482 def PUNPCKLWDrm : PDI<0x61, MRMSrcMem,
1483 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1484 "punpcklwd {$src2, $dst|$dst, $src2}",
1486 (v8i16 (vector_shuffle VR128:$src1,
1487 (bc_v8i16 (loadv2i64 addr:$src2)),
1488 UNPCKL_shuffle_mask)))]>;
1489 def PUNPCKLDQrr : PDI<0x62, MRMSrcReg,
1490 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1491 "punpckldq {$src2, $dst|$dst, $src2}",
1493 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1494 UNPCKL_shuffle_mask)))]>;
1495 def PUNPCKLDQrm : PDI<0x62, MRMSrcMem,
1496 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1497 "punpckldq {$src2, $dst|$dst, $src2}",
1499 (v4i32 (vector_shuffle VR128:$src1,
1500 (bc_v4i32 (loadv2i64 addr:$src2)),
1501 UNPCKL_shuffle_mask)))]>;
1502 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
1503 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1504 "punpcklqdq {$src2, $dst|$dst, $src2}",
1506 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1507 UNPCKL_shuffle_mask)))]>;
1508 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
1509 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1510 "punpcklqdq {$src2, $dst|$dst, $src2}",
1512 (v2i64 (vector_shuffle VR128:$src1,
1513 (loadv2i64 addr:$src2),
1514 UNPCKL_shuffle_mask)))]>;
1516 def PUNPCKHBWrr : PDI<0x68, MRMSrcReg,
1517 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1518 "punpckhbw {$src2, $dst|$dst, $src2}",
1520 (v16i8 (vector_shuffle VR128:$src1, VR128:$src2,
1521 UNPCKH_shuffle_mask)))]>;
1522 def PUNPCKHBWrm : PDI<0x68, MRMSrcMem,
1523 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1524 "punpckhbw {$src2, $dst|$dst, $src2}",
1526 (v16i8 (vector_shuffle VR128:$src1,
1527 (bc_v16i8 (loadv2i64 addr:$src2)),
1528 UNPCKH_shuffle_mask)))]>;
1529 def PUNPCKHWDrr : PDI<0x69, MRMSrcReg,
1530 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1531 "punpckhwd {$src2, $dst|$dst, $src2}",
1533 (v8i16 (vector_shuffle VR128:$src1, VR128:$src2,
1534 UNPCKH_shuffle_mask)))]>;
1535 def PUNPCKHWDrm : PDI<0x69, MRMSrcMem,
1536 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1537 "punpckhwd {$src2, $dst|$dst, $src2}",
1539 (v8i16 (vector_shuffle VR128:$src1,
1540 (bc_v8i16 (loadv2i64 addr:$src2)),
1541 UNPCKH_shuffle_mask)))]>;
1542 def PUNPCKHDQrr : PDI<0x6A, MRMSrcReg,
1543 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1544 "punpckhdq {$src2, $dst|$dst, $src2}",
1546 (v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1547 UNPCKH_shuffle_mask)))]>;
1548 def PUNPCKHDQrm : PDI<0x6A, MRMSrcMem,
1549 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1550 "punpckhdq {$src2, $dst|$dst, $src2}",
1552 (v4i32 (vector_shuffle VR128:$src1,
1553 (bc_v4i32 (loadv2i64 addr:$src2)),
1554 UNPCKH_shuffle_mask)))]>;
1555 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
1556 (ops VR128:$dst, VR128:$src1, VR128:$src2),
1557 "punpckhqdq {$src2, $dst|$dst, $src2}",
1559 (v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
1560 UNPCKH_shuffle_mask)))]>;
1561 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
1562 (ops VR128:$dst, VR128:$src1, i128mem:$src2),
1563 "punpckhqdq {$src2, $dst|$dst, $src2}",
1565 (v2i64 (vector_shuffle VR128:$src1,
1566 (loadv2i64 addr:$src2),
1567 UNPCKH_shuffle_mask)))]>;
1571 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
1572 (ops GR32:$dst, VR128:$src1, i32i8imm:$src2),
1573 "pextrw {$src2, $src1, $dst|$dst, $src1, $src2}",
1574 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
1575 (iPTR imm:$src2)))]>;
1576 let isTwoAddress = 1 in {
1577 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
1578 (ops VR128:$dst, VR128:$src1, GR32:$src2, i32i8imm:$src3),
1579 "pinsrw {$src3, $src2, $dst|$dst, $src2, $src3}",
1580 [(set VR128:$dst, (v8i16 (X86pinsrw (v8i16 VR128:$src1),
1581 GR32:$src2, (iPTR imm:$src3))))]>;
1582 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
1583 (ops VR128:$dst, VR128:$src1, i16mem:$src2, i32i8imm:$src3),
1584 "pinsrw {$src3, $src2, $dst|$dst, $src2, $src3}",
1586 (v8i16 (X86pinsrw (v8i16 VR128:$src1),
1587 (i32 (anyext (loadi16 addr:$src2))),
1588 (iPTR imm:$src3))))]>;
1591 //===----------------------------------------------------------------------===//
1592 // Miscellaneous Instructions
1593 //===----------------------------------------------------------------------===//
1596 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (ops GR32:$dst, VR128:$src),
1597 "movmskps {$src, $dst|$dst, $src}",
1598 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
1599 def MOVMSKPDrr : PSI<0x50, MRMSrcReg, (ops GR32:$dst, VR128:$src),
1600 "movmskpd {$src, $dst|$dst, $src}",
1601 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1603 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (ops GR32:$dst, VR128:$src),
1604 "pmovmskb {$src, $dst|$dst, $src}",
1605 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
1607 // Conditional store
1608 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (ops VR128:$src, VR128:$mask),
1609 "maskmovdqu {$mask, $src|$src, $mask}",
1610 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>,
1613 // Prefetching loads.
1614 // TODO: no intrinsics for these?
1615 def PREFETCHT0 : PSI<0x18, MRM1m, (ops i8mem:$src), "prefetcht0 $src", []>;
1616 def PREFETCHT1 : PSI<0x18, MRM2m, (ops i8mem:$src), "prefetcht1 $src", []>;
1617 def PREFETCHT2 : PSI<0x18, MRM3m, (ops i8mem:$src), "prefetcht2 $src", []>;
1618 def PREFETCHTNTA : PSI<0x18, MRM0m, (ops i8mem:$src), "prefetchtnta $src", []>;
1620 // Non-temporal stores
1621 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1622 "movntps {$src, $dst|$dst, $src}",
1623 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1624 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (ops i128mem:$dst, VR128:$src),
1625 "movntpd {$src, $dst|$dst, $src}",
1626 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
1627 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (ops f128mem:$dst, VR128:$src),
1628 "movntdq {$src, $dst|$dst, $src}",
1629 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
1630 def MOVNTImr : I<0xC3, MRMDestMem, (ops i32mem:$dst, GR32:$src),
1631 "movnti {$src, $dst|$dst, $src}",
1632 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
1633 TB, Requires<[HasSSE2]>;
1636 def CLFLUSH : I<0xAE, MRM7m, (ops i8mem:$src),
1637 "clflush $src", [(int_x86_sse2_clflush addr:$src)]>,
1638 TB, Requires<[HasSSE2]>;
1640 // Load, store, and memory fence
1641 def SFENCE : PSI<0xAE, MRM7m, (ops), "sfence", [(int_x86_sse_sfence)]>;
1642 def LFENCE : I<0xAE, MRM5m, (ops),
1643 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
1644 def MFENCE : I<0xAE, MRM6m, (ops),
1645 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
1648 def LDMXCSR : PSI<0xAE, MRM2m, (ops i32mem:$src),
1649 "ldmxcsr $src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1650 def STMXCSR : PSI<0xAE, MRM3m, (ops i32mem:$dst),
1651 "stmxcsr $dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1653 // Thread synchronization
1654 def MONITOR : I<0xC8, RawFrm, (ops), "monitor",
1655 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
1656 def MWAIT : I<0xC9, RawFrm, (ops), "mwait",
1657 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
1659 //===----------------------------------------------------------------------===//
1660 // Alias Instructions
1661 //===----------------------------------------------------------------------===//
1663 // Alias instructions that map zero vector to pxor / xorp* for sse.
1664 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
1665 def V_SET0 : PSI<0x57, MRMInitReg, (ops VR128:$dst),
1667 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
1669 def V_SETALLONES : PDI<0x76, MRMInitReg, (ops VR128:$dst),
1670 "pcmpeqd $dst, $dst",
1671 [(set VR128:$dst, (v2f64 immAllOnesV))]>;
1673 // FR32 / FR64 to 128-bit vector conversion.
1674 def MOVSS2PSrr : SSI<0x10, MRMSrcReg, (ops VR128:$dst, FR32:$src),
1675 "movss {$src, $dst|$dst, $src}",
1677 (v4f32 (scalar_to_vector FR32:$src)))]>;
1678 def MOVSS2PSrm : SSI<0x10, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
1679 "movss {$src, $dst|$dst, $src}",
1681 (v4f32 (scalar_to_vector (loadf32 addr:$src))))]>;
1682 def MOVSD2PDrr : SDI<0x10, MRMSrcReg, (ops VR128:$dst, FR64:$src),
1683 "movsd {$src, $dst|$dst, $src}",
1685 (v2f64 (scalar_to_vector FR64:$src)))]>;
1686 def MOVSD2PDrm : SDI<0x10, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1687 "movsd {$src, $dst|$dst, $src}",
1689 (v2f64 (scalar_to_vector (loadf64 addr:$src))))]>;
1691 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (ops VR128:$dst, GR32:$src),
1692 "movd {$src, $dst|$dst, $src}",
1694 (v4i32 (scalar_to_vector GR32:$src)))]>;
1695 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (ops VR128:$dst, i32mem:$src),
1696 "movd {$src, $dst|$dst, $src}",
1698 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
1700 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (ops FR32:$dst, GR32:$src),
1701 "movd {$src, $dst|$dst, $src}",
1702 [(set FR32:$dst, (bitconvert GR32:$src))]>;
1704 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (ops FR32:$dst, i32mem:$src),
1705 "movd {$src, $dst|$dst, $src}",
1706 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
1708 // SSE2 instructions with XS prefix
1709 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1710 "movq {$src, $dst|$dst, $src}",
1712 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
1713 Requires<[HasSSE2]>;
1714 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (ops i64mem:$dst, VR128:$src),
1715 "movq {$src, $dst|$dst, $src}",
1716 [(store (i64 (vector_extract (v2i64 VR128:$src),
1717 (iPTR 0))), addr:$dst)]>;
1719 // FIXME: may not be able to eliminate this movss with coalescing the src and
1720 // dest register classes are different. We really want to write this pattern
1722 // def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1723 // (f32 FR32:$src)>;
1724 def MOVPS2SSrr : SSI<0x10, MRMSrcReg, (ops FR32:$dst, VR128:$src),
1725 "movss {$src, $dst|$dst, $src}",
1726 [(set FR32:$dst, (vector_extract (v4f32 VR128:$src),
1728 def MOVPS2SSmr : SSI<0x11, MRMDestMem, (ops f32mem:$dst, VR128:$src),
1729 "movss {$src, $dst|$dst, $src}",
1730 [(store (f32 (vector_extract (v4f32 VR128:$src),
1731 (iPTR 0))), addr:$dst)]>;
1732 def MOVPD2SDrr : SDI<0x10, MRMSrcReg, (ops FR64:$dst, VR128:$src),
1733 "movsd {$src, $dst|$dst, $src}",
1734 [(set FR64:$dst, (vector_extract (v2f64 VR128:$src),
1736 def MOVPD2SDmr : SDI<0x11, MRMDestMem, (ops f64mem:$dst, VR128:$src),
1737 "movsd {$src, $dst|$dst, $src}",
1738 [(store (f64 (vector_extract (v2f64 VR128:$src),
1739 (iPTR 0))), addr:$dst)]>;
1740 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (ops GR32:$dst, VR128:$src),
1741 "movd {$src, $dst|$dst, $src}",
1742 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
1744 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (ops i32mem:$dst, VR128:$src),
1745 "movd {$src, $dst|$dst, $src}",
1746 [(store (i32 (vector_extract (v4i32 VR128:$src),
1747 (iPTR 0))), addr:$dst)]>;
1749 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (ops GR32:$dst, FR32:$src),
1750 "movd {$src, $dst|$dst, $src}",
1751 [(set GR32:$dst, (bitconvert FR32:$src))]>;
1752 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (ops i32mem:$dst, FR32:$src),
1753 "movd {$src, $dst|$dst, $src}",
1754 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
1757 // Move to lower bits of a VR128, leaving upper bits alone.
1758 // Three operand (but two address) aliases.
1759 let isTwoAddress = 1 in {
1760 def MOVLSS2PSrr : SSI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src1, FR32:$src2),
1761 "movss {$src2, $dst|$dst, $src2}", []>;
1762 def MOVLSD2PDrr : SDI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src1, FR64:$src2),
1763 "movsd {$src2, $dst|$dst, $src2}", []>;
1765 let AddedComplexity = 15 in {
1766 def MOVLPSrr : SSI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1767 "movss {$src2, $dst|$dst, $src2}",
1769 (v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
1770 MOVL_shuffle_mask)))]>;
1771 def MOVLPDrr : SDI<0x10, MRMSrcReg, (ops VR128:$dst, VR128:$src1, VR128:$src2),
1772 "movsd {$src2, $dst|$dst, $src2}",
1774 (v2f64 (vector_shuffle VR128:$src1, VR128:$src2,
1775 MOVL_shuffle_mask)))]>;
1779 // Store / copy lower 64-bits of a XMM register.
1780 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (ops i64mem:$dst, VR128:$src),
1781 "movq {$src, $dst|$dst, $src}",
1782 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
1784 // Move to lower bits of a VR128 and zeroing upper bits.
1785 // Loading from memory automatically zeroing upper bits.
1786 let AddedComplexity = 20 in {
1787 def MOVZSS2PSrm : SSI<0x10, MRMSrcMem, (ops VR128:$dst, f32mem:$src),
1788 "movss {$src, $dst|$dst, $src}",
1789 [(set VR128:$dst, (v4f32 (vector_shuffle immAllZerosV,
1790 (v4f32 (scalar_to_vector (loadf32 addr:$src))),
1791 MOVL_shuffle_mask)))]>;
1792 def MOVZSD2PDrm : SDI<0x10, MRMSrcMem, (ops VR128:$dst, f64mem:$src),
1793 "movsd {$src, $dst|$dst, $src}",
1794 [(set VR128:$dst, (v2f64 (vector_shuffle immAllZerosV,
1795 (v2f64 (scalar_to_vector (loadf64 addr:$src))),
1796 MOVL_shuffle_mask)))]>;
1798 let AddedComplexity = 15 in
1799 // movd / movq to XMM register zero-extends
1800 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (ops VR128:$dst, GR32:$src),
1801 "movd {$src, $dst|$dst, $src}",
1802 [(set VR128:$dst, (v4i32 (vector_shuffle immAllZerosV,
1803 (v4i32 (scalar_to_vector GR32:$src)),
1804 MOVL_shuffle_mask)))]>;
1805 let AddedComplexity = 20 in
1806 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (ops VR128:$dst, i32mem:$src),
1807 "movd {$src, $dst|$dst, $src}",
1808 [(set VR128:$dst, (v4i32 (vector_shuffle immAllZerosV,
1809 (v4i32 (scalar_to_vector (loadi32 addr:$src))),
1810 MOVL_shuffle_mask)))]>;
1811 // Moving from XMM to XMM but still clear upper 64 bits.
1812 let AddedComplexity = 15 in
1813 def MOVZQI2PQIrr : I<0x7E, MRMSrcReg, (ops VR128:$dst, VR128:$src),
1814 "movq {$src, $dst|$dst, $src}",
1815 [(set VR128:$dst, (int_x86_sse2_movl_dq VR128:$src))]>,
1816 XS, Requires<[HasSSE2]>;
1817 let AddedComplexity = 20 in
1818 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
1819 "movq {$src, $dst|$dst, $src}",
1820 [(set VR128:$dst, (int_x86_sse2_movl_dq
1821 (bitconvert (loadv2i64 addr:$src))))]>,
1822 XS, Requires<[HasSSE2]>;
1824 //===----------------------------------------------------------------------===//
1825 // Non-Instruction Patterns
1826 //===----------------------------------------------------------------------===//
1828 // 128-bit vector undef's.
1829 def : Pat<(v2f64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
1830 def : Pat<(v16i8 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
1831 def : Pat<(v8i16 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
1832 def : Pat<(v4i32 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
1833 def : Pat<(v2i64 (undef)), (IMPLICIT_DEF_VR128)>, Requires<[HasSSE2]>;
1835 // 128-bit vector all zero's.
1836 def : Pat<(v16i8 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
1837 def : Pat<(v8i16 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
1838 def : Pat<(v4i32 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
1839 def : Pat<(v2i64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
1840 def : Pat<(v2f64 immAllZerosV), (V_SET0)>, Requires<[HasSSE2]>;
1842 // 128-bit vector all one's.
1843 def : Pat<(v16i8 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
1844 def : Pat<(v8i16 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
1845 def : Pat<(v4i32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
1846 def : Pat<(v2i64 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE2]>;
1847 def : Pat<(v4f32 immAllOnesV), (V_SETALLONES)>, Requires<[HasSSE1]>;
1849 // Store 128-bit integer vector values.
1850 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
1851 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
1852 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
1853 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
1854 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
1855 (MOVDQAmr addr:$dst, VR128:$src)>, Requires<[HasSSE2]>;
1857 // Scalar to v8i16 / v16i8. The source may be a GR32, but only the lower 8 or
1859 def : Pat<(v8i16 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
1860 Requires<[HasSSE2]>;
1861 def : Pat<(v16i8 (X86s2vec GR32:$src)), (MOVDI2PDIrr GR32:$src)>,
1862 Requires<[HasSSE2]>;
1865 let Predicates = [HasSSE2] in {
1866 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
1867 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
1868 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
1869 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
1870 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
1871 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
1872 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
1873 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
1874 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
1875 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
1876 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
1877 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
1878 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
1879 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
1880 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
1881 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
1882 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
1883 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
1884 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
1885 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
1886 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
1887 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
1888 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
1889 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
1890 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
1891 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
1892 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
1893 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
1894 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
1895 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
1898 // Move scalar to XMM zero-extended
1899 // movd to XMM register zero-extends
1900 let AddedComplexity = 15 in {
1901 def : Pat<(v8i16 (vector_shuffle immAllZerosV,
1902 (v8i16 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
1903 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
1904 def : Pat<(v16i8 (vector_shuffle immAllZerosV,
1905 (v16i8 (X86s2vec GR32:$src)), MOVL_shuffle_mask)),
1906 (MOVZDI2PDIrr GR32:$src)>, Requires<[HasSSE2]>;
1907 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
1908 def : Pat<(v2f64 (vector_shuffle immAllZerosV,
1909 (v2f64 (scalar_to_vector FR64:$src)), MOVL_shuffle_mask)),
1910 (MOVLSD2PDrr (V_SET0), FR64:$src)>, Requires<[HasSSE2]>;
1911 def : Pat<(v4f32 (vector_shuffle immAllZerosV,
1912 (v4f32 (scalar_to_vector FR32:$src)), MOVL_shuffle_mask)),
1913 (MOVLSS2PSrr (V_SET0), FR32:$src)>, Requires<[HasSSE2]>;
1916 // Splat v2f64 / v2i64
1917 let AddedComplexity = 10 in {
1918 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
1919 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1920 def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
1921 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1922 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), SSE_splat_lo_mask:$sm),
1923 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1924 def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), UNPCKH_shuffle_mask:$sm),
1925 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1929 def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SSE_splat_mask:$sm),
1930 (SHUFPSrri VR128:$src, VR128:$src, SSE_splat_mask:$sm)>,
1931 Requires<[HasSSE1]>;
1933 // Special unary SHUFPSrri case.
1934 // FIXME: when we want non two-address code, then we should use PSHUFD?
1935 def : Pat<(vector_shuffle (v4f32 VR128:$src1), (undef),
1936 SHUFP_unary_shuffle_mask:$sm),
1937 (SHUFPSrri VR128:$src1, VR128:$src1, SHUFP_unary_shuffle_mask:$sm)>,
1938 Requires<[HasSSE1]>;
1939 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
1940 def : Pat<(vector_shuffle (loadv4f32 addr:$src1), (undef),
1941 SHUFP_unary_shuffle_mask:$sm),
1942 (PSHUFDmi addr:$src1, SHUFP_unary_shuffle_mask:$sm)>,
1943 Requires<[HasSSE2]>;
1944 // Special binary v4i32 shuffle cases with SHUFPS.
1945 def : Pat<(vector_shuffle (v4i32 VR128:$src1), (v4i32 VR128:$src2),
1946 PSHUFD_binary_shuffle_mask:$sm),
1947 (SHUFPSrri VR128:$src1, VR128:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
1948 Requires<[HasSSE2]>;
1949 def : Pat<(vector_shuffle (v4i32 VR128:$src1),
1950 (bc_v4i32 (loadv2i64 addr:$src2)), PSHUFD_binary_shuffle_mask:$sm),
1951 (SHUFPSrmi VR128:$src1, addr:$src2, PSHUFD_binary_shuffle_mask:$sm)>,
1952 Requires<[HasSSE2]>;
1954 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
1955 let AddedComplexity = 10 in {
1956 def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef),
1957 UNPCKL_v_undef_shuffle_mask)),
1958 (UNPCKLPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1959 def : Pat<(v16i8 (vector_shuffle VR128:$src, (undef),
1960 UNPCKL_v_undef_shuffle_mask)),
1961 (PUNPCKLBWrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1962 def : Pat<(v8i16 (vector_shuffle VR128:$src, (undef),
1963 UNPCKL_v_undef_shuffle_mask)),
1964 (PUNPCKLWDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
1965 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
1966 UNPCKL_v_undef_shuffle_mask)),
1967 (PUNPCKLDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
1970 let AddedComplexity = 15 in
1971 // vector_shuffle v1, <undef> <1, 1, 3, 3>
1972 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
1973 MOVSHDUP_shuffle_mask)),
1974 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
1975 let AddedComplexity = 20 in
1976 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (loadv2i64 addr:$src)), (undef),
1977 MOVSHDUP_shuffle_mask)),
1978 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
1980 // vector_shuffle v1, <undef> <0, 0, 2, 2>
1981 let AddedComplexity = 15 in
1982 def : Pat<(v4i32 (vector_shuffle VR128:$src, (undef),
1983 MOVSLDUP_shuffle_mask)),
1984 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
1985 let AddedComplexity = 20 in
1986 def : Pat<(v4i32 (vector_shuffle (bc_v4i32 (loadv2i64 addr:$src)), (undef),
1987 MOVSLDUP_shuffle_mask)),
1988 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
1990 let AddedComplexity = 15 in {
1991 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
1992 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1993 MOVHP_shuffle_mask)),
1994 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
1996 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
1997 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
1998 MOVHLPS_shuffle_mask)),
1999 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
2001 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
2002 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (undef),
2003 MOVHLPS_v_undef_shuffle_mask)),
2004 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2005 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (undef),
2006 MOVHLPS_v_undef_shuffle_mask)),
2007 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
2010 let AddedComplexity = 20 in {
2011 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
2012 // vector_shuffle v1, (load v2) <0, 1, 4, 5> using MOVHPS
2013 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (loadv4f32 addr:$src2),
2014 MOVLP_shuffle_mask)),
2015 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2016 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (loadv2f64 addr:$src2),
2017 MOVLP_shuffle_mask)),
2018 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2019 def : Pat<(v4f32 (vector_shuffle VR128:$src1, (loadv4f32 addr:$src2),
2020 MOVHP_shuffle_mask)),
2021 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2022 def : Pat<(v2f64 (vector_shuffle VR128:$src1, (loadv2f64 addr:$src2),
2023 MOVHP_shuffle_mask)),
2024 (MOVHPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2026 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)),
2027 MOVLP_shuffle_mask)),
2028 (MOVLPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2029 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (loadv2i64 addr:$src2),
2030 MOVLP_shuffle_mask)),
2031 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2032 def : Pat<(v4i32 (vector_shuffle VR128:$src1, (bc_v4i32 (loadv2i64 addr:$src2)),
2033 MOVHP_shuffle_mask)),
2034 (MOVHPSrm VR128:$src1, addr:$src2)>, Requires<[HasSSE1]>;
2035 def : Pat<(v2i64 (vector_shuffle VR128:$src1, (loadv2i64 addr:$src2),
2036 MOVLP_shuffle_mask)),
2037 (MOVLPDrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2040 let AddedComplexity = 15 in {
2041 // Setting the lowest element in the vector.
2042 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2043 MOVL_shuffle_mask)),
2044 (MOVLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2045 def : Pat<(v2i64 (vector_shuffle VR128:$src1, VR128:$src2,
2046 MOVL_shuffle_mask)),
2047 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2049 // vector_shuffle v1, v2 <4, 5, 2, 3> using MOVLPDrr (movsd)
2050 def : Pat<(v4f32 (vector_shuffle VR128:$src1, VR128:$src2,
2051 MOVLP_shuffle_mask)),
2052 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2053 def : Pat<(v4i32 (vector_shuffle VR128:$src1, VR128:$src2,
2054 MOVLP_shuffle_mask)),
2055 (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2058 // Set lowest element and zero upper elements.
2059 let AddedComplexity = 20 in
2060 def : Pat<(bc_v2i64 (vector_shuffle immAllZerosV,
2061 (v2f64 (scalar_to_vector (loadf64 addr:$src))),
2062 MOVL_shuffle_mask)),
2063 (MOVZQI2PQIrm addr:$src)>, Requires<[HasSSE2]>;
2065 // FIXME: Temporary workaround since 2-wide shuffle is broken.
2066 def : Pat<(int_x86_sse2_movs_d VR128:$src1, VR128:$src2),
2067 (v2f64 (MOVLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2068 def : Pat<(int_x86_sse2_loadh_pd VR128:$src1, addr:$src2),
2069 (v2f64 (MOVHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2070 def : Pat<(int_x86_sse2_loadl_pd VR128:$src1, addr:$src2),
2071 (v2f64 (MOVLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2072 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, VR128:$src2, imm:$src3),
2073 (v2f64 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$src3))>,
2074 Requires<[HasSSE2]>;
2075 def : Pat<(int_x86_sse2_shuf_pd VR128:$src1, (load addr:$src2), imm:$src3),
2076 (v2f64 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$src3))>,
2077 Requires<[HasSSE2]>;
2078 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, VR128:$src2),
2079 (v2f64 (UNPCKHPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2080 def : Pat<(int_x86_sse2_unpckh_pd VR128:$src1, (load addr:$src2)),
2081 (v2f64 (UNPCKHPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2082 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, VR128:$src2),
2083 (v2f64 (UNPCKLPDrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2084 def : Pat<(int_x86_sse2_unpckl_pd VR128:$src1, (load addr:$src2)),
2085 (v2f64 (UNPCKLPDrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2086 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, VR128:$src2),
2087 (v2i64 (PUNPCKHQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2088 def : Pat<(int_x86_sse2_punpckh_qdq VR128:$src1, (load addr:$src2)),
2089 (v2i64 (PUNPCKHQDQrm VR128:$src1, addr:$src2))>, Requires<[HasSSE2]>;
2090 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, VR128:$src2),
2091 (v2i64 (PUNPCKLQDQrr VR128:$src1, VR128:$src2))>, Requires<[HasSSE2]>;
2092 def : Pat<(int_x86_sse2_punpckl_qdq VR128:$src1, (load addr:$src2)),
2093 (PUNPCKLQDQrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2095 // Some special case pandn patterns.
2096 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2098 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2099 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2101 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2102 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2104 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
2106 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
2107 (load addr:$src2))),
2108 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2109 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
2110 (load addr:$src2))),
2111 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2112 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
2113 (load addr:$src2))),
2114 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
2117 def : Pat<(v4f32 (X86loadu addr:$src)), (MOVUPSrm addr:$src)>,
2118 Requires<[HasSSE1]>;