1 //====- X86InstrSSE.td - Describe the X86 Instruction Set --*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the X86 SSE instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
17 //===----------------------------------------------------------------------===//
18 // SSE specific DAG Nodes.
19 //===----------------------------------------------------------------------===//
21 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
22 SDTCisFP<0>, SDTCisInt<2> ]>;
23 def SDTX86VFCMP : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>,
24 SDTCisFP<1>, SDTCisVT<3, i8>]>;
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 X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
35 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
36 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
37 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>;
38 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>;
39 def X86pshufb : SDNode<"X86ISD::PSHUFB",
40 SDTypeProfile<1, 2, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
42 def X86pextrb : SDNode<"X86ISD::PEXTRB",
43 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
44 def X86pextrw : SDNode<"X86ISD::PEXTRW",
45 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
46 def X86pinsrb : SDNode<"X86ISD::PINSRB",
47 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
48 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
49 def X86pinsrw : SDNode<"X86ISD::PINSRW",
50 SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
51 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
52 def X86insrtps : SDNode<"X86ISD::INSERTPS",
53 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
54 SDTCisVT<2, v4f32>, SDTCisPtrTy<3>]>>;
55 def X86vzmovl : SDNode<"X86ISD::VZEXT_MOVL",
56 SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>;
57 def X86vzload : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
58 [SDNPHasChain, SDNPMayLoad]>;
59 def X86vshl : SDNode<"X86ISD::VSHL", SDTIntShiftOp>;
60 def X86vshr : SDNode<"X86ISD::VSRL", SDTIntShiftOp>;
61 def X86cmpps : SDNode<"X86ISD::CMPPS", SDTX86VFCMP>;
62 def X86cmppd : SDNode<"X86ISD::CMPPD", SDTX86VFCMP>;
63 def X86pcmpeqb : SDNode<"X86ISD::PCMPEQB", SDTIntBinOp, [SDNPCommutative]>;
64 def X86pcmpeqw : SDNode<"X86ISD::PCMPEQW", SDTIntBinOp, [SDNPCommutative]>;
65 def X86pcmpeqd : SDNode<"X86ISD::PCMPEQD", SDTIntBinOp, [SDNPCommutative]>;
66 def X86pcmpeqq : SDNode<"X86ISD::PCMPEQQ", SDTIntBinOp, [SDNPCommutative]>;
67 def X86pcmpgtb : SDNode<"X86ISD::PCMPGTB", SDTIntBinOp>;
68 def X86pcmpgtw : SDNode<"X86ISD::PCMPGTW", SDTIntBinOp>;
69 def X86pcmpgtd : SDNode<"X86ISD::PCMPGTD", SDTIntBinOp>;
70 def X86pcmpgtq : SDNode<"X86ISD::PCMPGTQ", SDTIntBinOp>;
72 def SDTX86CmpPTest : SDTypeProfile<1, 2, [SDTCisVT<0, i32>,
75 def X86ptest : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
77 //===----------------------------------------------------------------------===//
78 // SSE Complex Patterns
79 //===----------------------------------------------------------------------===//
81 // These are 'extloads' from a scalar to the low element of a vector, zeroing
82 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
84 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
85 [SDNPHasChain, SDNPMayLoad]>;
86 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
87 [SDNPHasChain, SDNPMayLoad]>;
89 def ssmem : Operand<v4f32> {
90 let PrintMethod = "printf32mem";
91 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
92 let ParserMatchClass = X86MemAsmOperand;
94 def sdmem : Operand<v2f64> {
95 let PrintMethod = "printf64mem";
96 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
97 let ParserMatchClass = X86MemAsmOperand;
100 //===----------------------------------------------------------------------===//
101 // SSE pattern fragments
102 //===----------------------------------------------------------------------===//
104 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
105 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
106 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
107 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
109 // Like 'store', but always requires vector alignment.
110 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
111 (store node:$val, node:$ptr), [{
112 return cast<StoreSDNode>(N)->getAlignment() >= 16;
115 // Like 'load', but always requires vector alignment.
116 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
117 return cast<LoadSDNode>(N)->getAlignment() >= 16;
120 def alignedloadfsf32 : PatFrag<(ops node:$ptr),
121 (f32 (alignedload node:$ptr))>;
122 def alignedloadfsf64 : PatFrag<(ops node:$ptr),
123 (f64 (alignedload node:$ptr))>;
124 def alignedloadv4f32 : PatFrag<(ops node:$ptr),
125 (v4f32 (alignedload node:$ptr))>;
126 def alignedloadv2f64 : PatFrag<(ops node:$ptr),
127 (v2f64 (alignedload node:$ptr))>;
128 def alignedloadv4i32 : PatFrag<(ops node:$ptr),
129 (v4i32 (alignedload node:$ptr))>;
130 def alignedloadv2i64 : PatFrag<(ops node:$ptr),
131 (v2i64 (alignedload node:$ptr))>;
133 // Like 'load', but uses special alignment checks suitable for use in
134 // memory operands in most SSE instructions, which are required to
135 // be naturally aligned on some targets but not on others. If the subtarget
136 // allows unaligned accesses, match any load, though this may require
137 // setting a feature bit in the processor (on startup, for example).
138 // Opteron 10h and later implement such a feature.
139 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
140 return Subtarget->hasVectorUAMem()
141 || cast<LoadSDNode>(N)->getAlignment() >= 16;
144 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
145 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
146 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
147 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
148 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
149 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
150 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
152 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
154 // FIXME: 8 byte alignment for mmx reads is not required
155 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
156 return cast<LoadSDNode>(N)->getAlignment() >= 8;
159 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
160 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
161 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
162 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
165 // Like 'store', but requires the non-temporal bit to be set
166 def nontemporalstore : PatFrag<(ops node:$val, node:$ptr),
167 (st node:$val, node:$ptr), [{
168 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
169 return ST->isNonTemporal();
173 def alignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
174 (st node:$val, node:$ptr), [{
175 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
176 return ST->isNonTemporal() && !ST->isTruncatingStore() &&
177 ST->getAddressingMode() == ISD::UNINDEXED &&
178 ST->getAlignment() >= 16;
182 def unalignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
183 (st node:$val, node:$ptr), [{
184 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
185 return ST->isNonTemporal() &&
186 ST->getAlignment() < 16;
190 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
191 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
192 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
193 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
194 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
195 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
197 def vzmovl_v2i64 : PatFrag<(ops node:$src),
198 (bitconvert (v2i64 (X86vzmovl
199 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
200 def vzmovl_v4i32 : PatFrag<(ops node:$src),
201 (bitconvert (v4i32 (X86vzmovl
202 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
204 def vzload_v2i64 : PatFrag<(ops node:$src),
205 (bitconvert (v2i64 (X86vzload node:$src)))>;
208 def fp32imm0 : PatLeaf<(f32 fpimm), [{
209 return N->isExactlyValue(+0.0);
212 // BYTE_imm - Transform bit immediates into byte immediates.
213 def BYTE_imm : SDNodeXForm<imm, [{
214 // Transformation function: imm >> 3
215 return getI32Imm(N->getZExtValue() >> 3);
218 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
220 def SHUFFLE_get_shuf_imm : SDNodeXForm<vector_shuffle, [{
221 return getI8Imm(X86::getShuffleSHUFImmediate(N));
224 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
226 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<vector_shuffle, [{
227 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
230 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
232 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<vector_shuffle, [{
233 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
236 // SHUFFLE_get_palign_imm xform function: convert vector_shuffle mask to
238 def SHUFFLE_get_palign_imm : SDNodeXForm<vector_shuffle, [{
239 return getI8Imm(X86::getShufflePALIGNRImmediate(N));
242 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
243 (vector_shuffle node:$lhs, node:$rhs), [{
244 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
245 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
248 def movddup : PatFrag<(ops node:$lhs, node:$rhs),
249 (vector_shuffle node:$lhs, node:$rhs), [{
250 return X86::isMOVDDUPMask(cast<ShuffleVectorSDNode>(N));
253 def movhlps : PatFrag<(ops node:$lhs, node:$rhs),
254 (vector_shuffle node:$lhs, node:$rhs), [{
255 return X86::isMOVHLPSMask(cast<ShuffleVectorSDNode>(N));
258 def movhlps_undef : PatFrag<(ops node:$lhs, node:$rhs),
259 (vector_shuffle node:$lhs, node:$rhs), [{
260 return X86::isMOVHLPS_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
263 def movlhps : PatFrag<(ops node:$lhs, node:$rhs),
264 (vector_shuffle node:$lhs, node:$rhs), [{
265 return X86::isMOVLHPSMask(cast<ShuffleVectorSDNode>(N));
268 def movlp : PatFrag<(ops node:$lhs, node:$rhs),
269 (vector_shuffle node:$lhs, node:$rhs), [{
270 return X86::isMOVLPMask(cast<ShuffleVectorSDNode>(N));
273 def movl : PatFrag<(ops node:$lhs, node:$rhs),
274 (vector_shuffle node:$lhs, node:$rhs), [{
275 return X86::isMOVLMask(cast<ShuffleVectorSDNode>(N));
278 def movshdup : PatFrag<(ops node:$lhs, node:$rhs),
279 (vector_shuffle node:$lhs, node:$rhs), [{
280 return X86::isMOVSHDUPMask(cast<ShuffleVectorSDNode>(N));
283 def movsldup : PatFrag<(ops node:$lhs, node:$rhs),
284 (vector_shuffle node:$lhs, node:$rhs), [{
285 return X86::isMOVSLDUPMask(cast<ShuffleVectorSDNode>(N));
288 def unpckl : PatFrag<(ops node:$lhs, node:$rhs),
289 (vector_shuffle node:$lhs, node:$rhs), [{
290 return X86::isUNPCKLMask(cast<ShuffleVectorSDNode>(N));
293 def unpckh : PatFrag<(ops node:$lhs, node:$rhs),
294 (vector_shuffle node:$lhs, node:$rhs), [{
295 return X86::isUNPCKHMask(cast<ShuffleVectorSDNode>(N));
298 def unpckl_undef : PatFrag<(ops node:$lhs, node:$rhs),
299 (vector_shuffle node:$lhs, node:$rhs), [{
300 return X86::isUNPCKL_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
303 def unpckh_undef : PatFrag<(ops node:$lhs, node:$rhs),
304 (vector_shuffle node:$lhs, node:$rhs), [{
305 return X86::isUNPCKH_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
308 def pshufd : PatFrag<(ops node:$lhs, node:$rhs),
309 (vector_shuffle node:$lhs, node:$rhs), [{
310 return X86::isPSHUFDMask(cast<ShuffleVectorSDNode>(N));
311 }], SHUFFLE_get_shuf_imm>;
313 def shufp : PatFrag<(ops node:$lhs, node:$rhs),
314 (vector_shuffle node:$lhs, node:$rhs), [{
315 return X86::isSHUFPMask(cast<ShuffleVectorSDNode>(N));
316 }], SHUFFLE_get_shuf_imm>;
318 def pshufhw : PatFrag<(ops node:$lhs, node:$rhs),
319 (vector_shuffle node:$lhs, node:$rhs), [{
320 return X86::isPSHUFHWMask(cast<ShuffleVectorSDNode>(N));
321 }], SHUFFLE_get_pshufhw_imm>;
323 def pshuflw : PatFrag<(ops node:$lhs, node:$rhs),
324 (vector_shuffle node:$lhs, node:$rhs), [{
325 return X86::isPSHUFLWMask(cast<ShuffleVectorSDNode>(N));
326 }], SHUFFLE_get_pshuflw_imm>;
328 def palign : PatFrag<(ops node:$lhs, node:$rhs),
329 (vector_shuffle node:$lhs, node:$rhs), [{
330 return X86::isPALIGNRMask(cast<ShuffleVectorSDNode>(N));
331 }], SHUFFLE_get_palign_imm>;
333 //===----------------------------------------------------------------------===//
334 // SSE scalar FP Instructions
335 //===----------------------------------------------------------------------===//
337 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded after
338 // instruction selection into a branch sequence.
339 let Uses = [EFLAGS], usesCustomInserter = 1 in {
340 def CMOV_FR32 : I<0, Pseudo,
341 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
342 "#CMOV_FR32 PSEUDO!",
343 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
345 def CMOV_FR64 : I<0, Pseudo,
346 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
347 "#CMOV_FR64 PSEUDO!",
348 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
350 def CMOV_V4F32 : I<0, Pseudo,
351 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
352 "#CMOV_V4F32 PSEUDO!",
354 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
356 def CMOV_V2F64 : I<0, Pseudo,
357 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
358 "#CMOV_V2F64 PSEUDO!",
360 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
362 def CMOV_V2I64 : I<0, Pseudo,
363 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
364 "#CMOV_V2I64 PSEUDO!",
366 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
370 //===----------------------------------------------------------------------===//
371 // SSE 1 & 2 Instructions Classes
372 //===----------------------------------------------------------------------===//
374 /// sse12_fp_scalar - SSE 1 & 2 scalar instructions class
375 multiclass sse12_fp_scalar<bits<8> opc, string OpcodeStr, SDNode OpNode,
376 RegisterClass RC, X86MemOperand memop> {
377 let isCommutable = 1 in {
378 def rr : SI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
379 OpcodeStr, [(set RC:$dst, (OpNode RC:$src1, RC:$src2))]>;
381 def rm : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memop:$src2),
382 OpcodeStr, [(set RC:$dst, (OpNode RC:$src1, (load addr:$src2)))]>;
385 /// sse12_fp_scalar_int - SSE 1 & 2 scalar instructions intrinsics class
386 multiclass sse12_fp_scalar_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
387 string asm, string SSEVer, string FPSizeStr,
388 Operand memop, ComplexPattern mem_cpat> {
389 def rr_Int : SI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
390 asm, [(set RC:$dst, (
391 !nameconcat<Intrinsic>("int_x86_sse",
392 !strconcat(SSEVer, !strconcat("_",
393 !strconcat(OpcodeStr, FPSizeStr))))
394 RC:$src1, RC:$src2))]>;
395 def rm_Int : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memop:$src2),
396 asm, [(set RC:$dst, (
397 !nameconcat<Intrinsic>("int_x86_sse",
398 !strconcat(SSEVer, !strconcat("_",
399 !strconcat(OpcodeStr, FPSizeStr))))
400 RC:$src1, mem_cpat:$src2))]>;
403 /// sse12_fp_packed - SSE 1 & 2 packed instructions class
404 multiclass sse12_fp_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
405 RegisterClass RC, ValueType vt,
406 X86MemOperand x86memop, PatFrag mem_frag,
407 Domain d, bit MayLoad = 0> {
408 let isCommutable = 1 in
409 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
410 OpcodeStr, [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))],d>;
411 let mayLoad = MayLoad in
412 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
413 OpcodeStr, [(set RC:$dst, (OpNode RC:$src1,
414 (mem_frag addr:$src2)))],d>;
417 /// sse12_fp_packed_logical_rm - SSE 1 & 2 packed instructions class
418 multiclass sse12_fp_packed_logical_rm<bits<8> opc, RegisterClass RC, Domain d,
419 string OpcodeStr, X86MemOperand x86memop,
420 list<dag> pat_rr, list<dag> pat_rm> {
421 let isCommutable = 1 in
422 def rr : PI<opc, MRMSrcReg, (outs RC:$dst),
423 (ins RC:$src1, RC:$src2), OpcodeStr, pat_rr, d>;
424 def rm : PI<opc, MRMSrcMem, (outs RC:$dst),
425 (ins RC:$src1, x86memop:$src2), OpcodeStr, pat_rm, d>;
428 /// sse12_fp_packed_int - SSE 1 & 2 packed instructions intrinsics class
429 multiclass sse12_fp_packed_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
430 string asm, string SSEVer, string FPSizeStr,
431 X86MemOperand memop, PatFrag mem_frag,
433 def rr_Int : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
434 asm, [(set RC:$dst, (
435 !nameconcat<Intrinsic>("int_x86_sse",
436 !strconcat(SSEVer, !strconcat("_",
437 !strconcat(OpcodeStr, FPSizeStr))))
438 RC:$src1, RC:$src2))], d>;
439 def rm_Int : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memop:$src2),
440 asm, [(set RC:$dst, (
441 !nameconcat<Intrinsic>("int_x86_sse",
442 !strconcat(SSEVer, !strconcat("_",
443 !strconcat(OpcodeStr, FPSizeStr))))
444 RC:$src1, (mem_frag addr:$src2)))], d>;
447 //===----------------------------------------------------------------------===//
449 //===----------------------------------------------------------------------===//
451 // Move Instructions. Register-to-register movss is not used for FR32
452 // register copies because it's a partial register update; FsMOVAPSrr is
453 // used instead. Register-to-register movss is not modeled as an INSERT_SUBREG
454 // because INSERT_SUBREG requires that the insert be implementable in terms of
455 // a copy, and just mentioned, we don't use movss for copies.
456 let Constraints = "$src1 = $dst" in
457 def MOVSSrr : SSI<0x10, MRMSrcReg,
458 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
459 "movss\t{$src2, $dst|$dst, $src2}",
460 [(set (v4f32 VR128:$dst),
461 (movl VR128:$src1, (scalar_to_vector FR32:$src2)))]>;
463 // Extract the low 32-bit value from one vector and insert it into another.
464 let AddedComplexity = 15 in
465 def : Pat<(v4f32 (movl VR128:$src1, VR128:$src2)),
466 (MOVSSrr (v4f32 VR128:$src1),
467 (EXTRACT_SUBREG (v4f32 VR128:$src2), sub_ss))>;
469 // Implicitly promote a 32-bit scalar to a vector.
470 def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
471 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), FR32:$src, sub_ss)>;
473 // Loading from memory automatically zeroing upper bits.
474 let canFoldAsLoad = 1, isReMaterializable = 1 in
475 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
476 "movss\t{$src, $dst|$dst, $src}",
477 [(set FR32:$dst, (loadf32 addr:$src))]>;
479 // MOVSSrm zeros the high parts of the register; represent this
480 // with SUBREG_TO_REG.
481 let AddedComplexity = 20 in {
482 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
483 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
484 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
485 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
486 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
487 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
490 // Store scalar value to memory.
491 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
492 "movss\t{$src, $dst|$dst, $src}",
493 [(store FR32:$src, addr:$dst)]>;
495 // Extract and store.
496 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
499 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
501 // Conversion instructions
502 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
503 "cvttss2si\t{$src, $dst|$dst, $src}",
504 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
505 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
506 "cvttss2si\t{$src, $dst|$dst, $src}",
507 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
508 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
509 "cvtsi2ss\t{$src, $dst|$dst, $src}",
510 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
511 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
512 "cvtsi2ss\t{$src, $dst|$dst, $src}",
513 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
515 // Match intrinsics which expect XMM operand(s).
516 def CVTSS2SIrr: SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
517 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
518 def CVTSS2SIrm: SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
519 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
521 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
522 "cvtss2si\t{$src, $dst|$dst, $src}",
523 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
524 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
525 "cvtss2si\t{$src, $dst|$dst, $src}",
526 [(set GR32:$dst, (int_x86_sse_cvtss2si
527 (load addr:$src)))]>;
529 // Match intrinsics which expect MM and XMM operand(s).
530 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
531 "cvtps2pi\t{$src, $dst|$dst, $src}",
532 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
533 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
534 "cvtps2pi\t{$src, $dst|$dst, $src}",
535 [(set VR64:$dst, (int_x86_sse_cvtps2pi
536 (load addr:$src)))]>;
537 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
538 "cvttps2pi\t{$src, $dst|$dst, $src}",
539 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
540 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
541 "cvttps2pi\t{$src, $dst|$dst, $src}",
542 [(set VR64:$dst, (int_x86_sse_cvttps2pi
543 (load addr:$src)))]>;
544 let Constraints = "$src1 = $dst" in {
545 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
546 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
547 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
548 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
550 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
551 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
552 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
553 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
554 (load addr:$src2)))]>;
557 // Aliases for intrinsics
558 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
559 "cvttss2si\t{$src, $dst|$dst, $src}",
561 (int_x86_sse_cvttss2si VR128:$src))]>;
562 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
563 "cvttss2si\t{$src, $dst|$dst, $src}",
565 (int_x86_sse_cvttss2si(load addr:$src)))]>;
567 let Constraints = "$src1 = $dst" in {
568 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
569 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
570 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
571 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
573 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
574 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
575 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
576 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
577 (loadi32 addr:$src2)))]>;
580 // Comparison instructions
581 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
582 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
583 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
584 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
586 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
587 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
588 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
590 // Accept explicit immediate argument form instead of comparison code.
591 let isAsmParserOnly = 1 in {
592 def CMPSSrr_alt : SSIi8<0xC2, MRMSrcReg,
593 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, i8imm:$src2),
594 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
596 def CMPSSrm_alt : SSIi8<0xC2, MRMSrcMem,
597 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, i8imm:$src2),
598 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
602 let Defs = [EFLAGS] in {
603 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
604 "ucomiss\t{$src2, $src1|$src1, $src2}",
605 [(set EFLAGS, (X86cmp FR32:$src1, FR32:$src2))]>;
606 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
607 "ucomiss\t{$src2, $src1|$src1, $src2}",
608 [(set EFLAGS, (X86cmp FR32:$src1, (loadf32 addr:$src2)))]>;
610 def COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
611 "comiss\t{$src2, $src1|$src1, $src2}", []>;
612 def COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
613 "comiss\t{$src2, $src1|$src1, $src2}", []>;
617 // Aliases to match intrinsics which expect XMM operand(s).
618 let Constraints = "$src1 = $dst" in {
619 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
621 (ins VR128:$src1, VR128:$src, SSECC:$cc),
622 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
623 [(set VR128:$dst, (int_x86_sse_cmp_ss
625 VR128:$src, imm:$cc))]>;
626 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
628 (ins VR128:$src1, f32mem:$src, SSECC:$cc),
629 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
630 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
631 (load addr:$src), imm:$cc))]>;
634 let Defs = [EFLAGS] in {
635 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
636 "ucomiss\t{$src2, $src1|$src1, $src2}",
637 [(set EFLAGS, (X86ucomi (v4f32 VR128:$src1),
639 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
640 "ucomiss\t{$src2, $src1|$src1, $src2}",
641 [(set EFLAGS, (X86ucomi (v4f32 VR128:$src1),
642 (load addr:$src2)))]>;
644 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
645 "comiss\t{$src2, $src1|$src1, $src2}",
646 [(set EFLAGS, (X86comi (v4f32 VR128:$src1),
648 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
649 "comiss\t{$src2, $src1|$src1, $src2}",
650 [(set EFLAGS, (X86comi (v4f32 VR128:$src1),
651 (load addr:$src2)))]>;
654 // Aliases of packed SSE1 instructions for scalar use. These all have names
655 // that start with 'Fs'.
657 // Alias instructions that map fld0 to pxor for sse.
658 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
660 // FIXME: Set encoding to pseudo!
661 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
662 [(set FR32:$dst, fp32imm0)]>,
663 Requires<[HasSSE1]>, TB, OpSize;
665 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
667 let neverHasSideEffects = 1 in
668 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
669 "movaps\t{$src, $dst|$dst, $src}", []>;
671 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
673 let canFoldAsLoad = 1, isReMaterializable = 1 in
674 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
675 "movaps\t{$src, $dst|$dst, $src}",
676 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
678 /// sse12_fp_alias_pack_logical - SSE 1 & 2 aliased packed FP logical ops
680 multiclass sse12_fp_alias_pack_logical<bits<8> opc, string OpcodeStr,
681 SDNode OpNode, bit MayLoad = 0> {
682 let isAsmParserOnly = 1 in {
683 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
684 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode, FR32,
685 f32, f128mem, memopfsf32, SSEPackedSingle, MayLoad>, VEX_4V;
687 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
688 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode, FR64,
689 f64, f128mem, memopfsf64, SSEPackedDouble, MayLoad>, OpSize,
693 let Constraints = "$src1 = $dst" in {
694 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
695 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, FR32, f32,
696 f128mem, memopfsf32, SSEPackedSingle, MayLoad>, TB;
698 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
699 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, FR64, f64,
700 f128mem, memopfsf64, SSEPackedDouble, MayLoad>, TB, OpSize;
704 // Alias bitwise logical operations using SSE logical ops on packed FP values.
705 defm FsAND : sse12_fp_alias_pack_logical<0x54, "and", X86fand>;
706 defm FsOR : sse12_fp_alias_pack_logical<0x56, "or", X86for>;
707 defm FsXOR : sse12_fp_alias_pack_logical<0x57, "xor", X86fxor>;
709 let neverHasSideEffects = 1, Pattern = []<dag>, isCommutable = 0 in
710 defm FsANDN : sse12_fp_alias_pack_logical<0x55, "andn", undef, 1>;
712 /// basic_sse12_fp_binop_rm - SSE 1 & 2 binops come in both scalar and
715 /// In addition, we also have a special variant of the scalar form here to
716 /// represent the associated intrinsic operation. This form is unlike the
717 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
718 /// and leaves the top elements unmodified (therefore these cannot be commuted).
720 /// These three forms can each be reg+reg or reg+mem, so there are a total of
721 /// six "instructions".
723 multiclass basic_sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
726 let isAsmParserOnly = 1 in {
727 defm V#NAME#SS : sse12_fp_scalar<opc,
728 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
729 OpNode, FR32, f32mem>, XS, VEX_4V;
731 defm V#NAME#SD : sse12_fp_scalar<opc,
732 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
733 OpNode, FR64, f64mem>, XD, VEX_4V;
735 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
736 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
737 VR128, v4f32, f128mem, memopv4f32, SSEPackedSingle>,
740 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
741 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
742 VR128, v2f64, f128mem, memopv2f64, SSEPackedDouble>,
745 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
746 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
747 "", "_ss", ssmem, sse_load_f32>, XS, VEX_4V;
749 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
750 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
751 "2", "_sd", sdmem, sse_load_f64>, XD, VEX_4V;
754 let Constraints = "$src1 = $dst" in {
755 defm SS : sse12_fp_scalar<opc,
756 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
757 OpNode, FR32, f32mem>, XS;
759 defm SD : sse12_fp_scalar<opc,
760 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
761 OpNode, FR64, f64mem>, XD;
763 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
764 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v4f32,
765 f128mem, memopv4f32, SSEPackedSingle>, TB;
767 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
768 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v2f64,
769 f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
771 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
772 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
773 "", "_ss", ssmem, sse_load_f32>, XS;
775 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
776 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
777 "2", "_sd", sdmem, sse_load_f64>, XD;
781 // Arithmetic instructions
782 defm ADD : basic_sse12_fp_binop_rm<0x58, "add", fadd>;
783 defm MUL : basic_sse12_fp_binop_rm<0x59, "mul", fmul>;
785 let isCommutable = 0 in {
786 defm SUB : basic_sse12_fp_binop_rm<0x5C, "sub", fsub>;
787 defm DIV : basic_sse12_fp_binop_rm<0x5E, "div", fdiv>;
790 /// sse12_fp_binop_rm - Other SSE 1 & 2 binops
792 /// This multiclass is like basic_sse12_fp_binop_rm, with the addition of
793 /// instructions for a full-vector intrinsic form. Operations that map
794 /// onto C operators don't use this form since they just use the plain
795 /// vector form instead of having a separate vector intrinsic form.
797 /// This provides a total of eight "instructions".
799 multiclass sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
802 let isAsmParserOnly = 1 in {
803 // Scalar operation, reg+reg.
804 defm V#NAME#SS : sse12_fp_scalar<opc,
805 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
806 OpNode, FR32, f32mem>, XS, VEX_4V;
808 defm V#NAME#SD : sse12_fp_scalar<opc,
809 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
810 OpNode, FR64, f64mem>, XD, VEX_4V;
812 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
813 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
814 VR128, v4f32, f128mem, memopv4f32, SSEPackedSingle>,
817 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
818 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
819 VR128, v2f64, f128mem, memopv2f64, SSEPackedDouble>,
822 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
823 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
824 "", "_ss", ssmem, sse_load_f32>, XS, VEX_4V;
826 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
827 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
828 "2", "_sd", sdmem, sse_load_f64>, XD, VEX_4V;
830 defm V#NAME#PS : sse12_fp_packed_int<opc, OpcodeStr, VR128,
831 !strconcat(OpcodeStr, "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
832 "", "_ps", f128mem, memopv4f32, SSEPackedSingle>, VEX_4V;
834 defm V#NAME#PD : sse12_fp_packed_int<opc, OpcodeStr, VR128,
835 !strconcat(OpcodeStr, "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
836 "2", "_pd", f128mem, memopv2f64, SSEPackedDouble>, OpSize,
840 let Constraints = "$src1 = $dst" in {
841 // Scalar operation, reg+reg.
842 defm SS : sse12_fp_scalar<opc,
843 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
844 OpNode, FR32, f32mem>, XS;
845 defm SD : sse12_fp_scalar<opc,
846 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
847 OpNode, FR64, f64mem>, XD;
848 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
849 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v4f32,
850 f128mem, memopv4f32, SSEPackedSingle>, TB;
852 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
853 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v2f64,
854 f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
856 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
857 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
858 "", "_ss", ssmem, sse_load_f32>, XS;
860 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
861 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
862 "2", "_sd", sdmem, sse_load_f64>, XD;
864 defm PS : sse12_fp_packed_int<opc, OpcodeStr, VR128,
865 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
866 "", "_ps", f128mem, memopv4f32, SSEPackedSingle>, TB;
868 defm PD : sse12_fp_packed_int<opc, OpcodeStr, VR128,
869 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
870 "2", "_pd", f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
874 let isCommutable = 0 in {
875 defm MAX : sse12_fp_binop_rm<0x5F, "max", X86fmax>;
876 defm MIN : sse12_fp_binop_rm<0x5D, "min", X86fmin>;
879 //===----------------------------------------------------------------------===//
880 // SSE packed FP Instructions
883 let neverHasSideEffects = 1 in
884 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
885 "movaps\t{$src, $dst|$dst, $src}", []>;
886 let canFoldAsLoad = 1, isReMaterializable = 1 in
887 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
888 "movaps\t{$src, $dst|$dst, $src}",
889 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
891 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
892 "movaps\t{$src, $dst|$dst, $src}",
893 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
895 let neverHasSideEffects = 1 in
896 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
897 "movups\t{$src, $dst|$dst, $src}", []>;
898 let canFoldAsLoad = 1, isReMaterializable = 1 in
899 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
900 "movups\t{$src, $dst|$dst, $src}",
901 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
902 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
903 "movups\t{$src, $dst|$dst, $src}",
904 [(store (v4f32 VR128:$src), addr:$dst)]>;
906 // Intrinsic forms of MOVUPS load and store
907 let canFoldAsLoad = 1, isReMaterializable = 1 in
908 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
909 "movups\t{$src, $dst|$dst, $src}",
910 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
911 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
912 "movups\t{$src, $dst|$dst, $src}",
913 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
915 let Constraints = "$src1 = $dst" in {
916 let AddedComplexity = 20 in {
917 def MOVLPSrm : PSI<0x12, MRMSrcMem,
918 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
919 "movlps\t{$src2, $dst|$dst, $src2}",
922 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
923 def MOVHPSrm : PSI<0x16, MRMSrcMem,
924 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
925 "movhps\t{$src2, $dst|$dst, $src2}",
927 (movlhps VR128:$src1,
928 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
930 } // Constraints = "$src1 = $dst"
933 def : Pat<(movlhps VR128:$src1, (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
934 (MOVHPSrm (v4i32 VR128:$src1), addr:$src2)>;
936 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
937 "movlps\t{$src, $dst|$dst, $src}",
938 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
939 (iPTR 0))), addr:$dst)]>;
941 // v2f64 extract element 1 is always custom lowered to unpack high to low
942 // and extract element 0 so the non-store version isn't too horrible.
943 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
944 "movhps\t{$src, $dst|$dst, $src}",
945 [(store (f64 (vector_extract
946 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
947 (undef)), (iPTR 0))), addr:$dst)]>;
949 let Constraints = "$src1 = $dst" in {
950 let AddedComplexity = 20 in {
951 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
952 (ins VR128:$src1, VR128:$src2),
953 "movlhps\t{$src2, $dst|$dst, $src2}",
955 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
957 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
958 (ins VR128:$src1, VR128:$src2),
959 "movhlps\t{$src2, $dst|$dst, $src2}",
961 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
963 } // Constraints = "$src1 = $dst"
965 let AddedComplexity = 20 in {
966 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
967 (MOVLHPSrr (v4f32 VR128:$src), (v4f32 VR128:$src))>;
968 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
969 (MOVLHPSrr (v2i64 VR128:$src), (v2i64 VR128:$src))>;
976 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
978 /// In addition, we also have a special variant of the scalar form here to
979 /// represent the associated intrinsic operation. This form is unlike the
980 /// plain scalar form, in that it takes an entire vector (instead of a
981 /// scalar) and leaves the top elements undefined.
983 /// And, we have a special variant form for a full-vector intrinsic form.
985 /// These four forms can each have a reg or a mem operand, so there are a
986 /// total of eight "instructions".
988 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
992 bit Commutable = 0> {
993 // Scalar operation, reg.
994 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
995 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
996 [(set FR32:$dst, (OpNode FR32:$src))]> {
997 let isCommutable = Commutable;
1000 // Scalar operation, mem.
1001 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
1002 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1003 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
1004 Requires<[HasSSE1, OptForSize]>;
1006 // Vector operation, reg.
1007 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1008 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1009 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
1010 let isCommutable = Commutable;
1013 // Vector operation, mem.
1014 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1015 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1016 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
1018 // Intrinsic operation, reg.
1019 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1020 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1021 [(set VR128:$dst, (F32Int VR128:$src))]> {
1022 let isCommutable = Commutable;
1025 // Intrinsic operation, mem.
1026 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
1027 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1028 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
1030 // Vector intrinsic operation, reg
1031 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1032 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1033 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
1034 let isCommutable = Commutable;
1037 // Vector intrinsic operation, mem
1038 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1039 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1040 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
1044 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
1045 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
1047 // Reciprocal approximations. Note that these typically require refinement
1048 // in order to obtain suitable precision.
1049 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
1050 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
1051 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
1052 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
1054 /// sse12_fp_packed_logical - SSE 1 & 2 packed FP logical ops
1056 multiclass sse12_fp_packed_logical<bits<8> opc, string OpcodeStr,
1057 SDNode OpNode, int HasPat = 0,
1058 list<list<dag>> Pattern = []> {
1059 let Constraints = "$src1 = $dst" in {
1060 defm PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
1061 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"), f128mem,
1062 !if(HasPat, Pattern[0], // rr
1063 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1,
1065 !if(HasPat, Pattern[2], // rm
1066 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
1067 (memopv2i64 addr:$src2)))])>, TB;
1069 defm PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
1070 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"), f128mem,
1071 !if(HasPat, Pattern[1], // rr
1072 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1075 !if(HasPat, Pattern[3], // rm
1076 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1077 (memopv2i64 addr:$src2)))])>,
1083 defm AND : sse12_fp_packed_logical<0x54, "and", and>;
1084 defm OR : sse12_fp_packed_logical<0x56, "or", or>;
1085 defm XOR : sse12_fp_packed_logical<0x57, "xor", xor>;
1086 let isCommutable = 0 in
1087 defm ANDN : sse12_fp_packed_logical<0x55, "andn", undef /* dummy */, 1, [
1089 [(set VR128:$dst, (v2i64 (and (xor VR128:$src1,
1090 (bc_v2i64 (v4i32 immAllOnesV))),
1093 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1094 (bc_v2i64 (v2f64 VR128:$src2))))],
1096 [(set VR128:$dst, (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
1097 (bc_v2i64 (v4i32 immAllOnesV))),
1098 (memopv2i64 addr:$src2))))],
1100 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1101 (memopv2i64 addr:$src2)))]]>;
1103 let Constraints = "$src1 = $dst" in {
1104 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
1105 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1106 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1107 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1108 VR128:$src, imm:$cc))]>;
1109 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
1110 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1111 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1112 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1113 (memop addr:$src), imm:$cc))]>;
1115 // Accept explicit immediate argument form instead of comparison code.
1116 let isAsmParserOnly = 1 in {
1117 def CMPPSrri_alt : PSIi8<0xC2, MRMSrcReg,
1118 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1119 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1120 def CMPPSrmi_alt : PSIi8<0xC2, MRMSrcMem,
1121 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1122 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1125 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1126 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
1127 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1128 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
1130 // Shuffle and unpack instructions
1131 let Constraints = "$src1 = $dst" in {
1132 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
1133 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
1134 (outs VR128:$dst), (ins VR128:$src1,
1135 VR128:$src2, i8imm:$src3),
1136 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1138 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1139 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
1140 (outs VR128:$dst), (ins VR128:$src1,
1141 f128mem:$src2, i8imm:$src3),
1142 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1145 VR128:$src1, (memopv4f32 addr:$src2))))]>;
1147 let AddedComplexity = 10 in {
1148 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
1149 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1150 "unpckhps\t{$src2, $dst|$dst, $src2}",
1152 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
1153 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
1154 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1155 "unpckhps\t{$src2, $dst|$dst, $src2}",
1157 (v4f32 (unpckh VR128:$src1,
1158 (memopv4f32 addr:$src2))))]>;
1160 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
1161 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1162 "unpcklps\t{$src2, $dst|$dst, $src2}",
1164 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
1165 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
1166 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1167 "unpcklps\t{$src2, $dst|$dst, $src2}",
1169 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
1170 } // AddedComplexity
1171 } // Constraints = "$src1 = $dst"
1174 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1175 "movmskps\t{$src, $dst|$dst, $src}",
1176 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
1177 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1178 "movmskpd\t{$src, $dst|$dst, $src}",
1179 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1181 // Prefetch intrinsic.
1182 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1183 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1184 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1185 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1186 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1187 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1188 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1189 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1191 // Non-temporal stores
1192 def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1193 "movntps\t{$src, $dst|$dst, $src}",
1194 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1196 let AddedComplexity = 400 in { // Prefer non-temporal versions
1197 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1198 "movntps\t{$src, $dst|$dst, $src}",
1199 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
1201 def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1202 "movntdq\t{$src, $dst|$dst, $src}",
1203 [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
1205 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
1206 "movnti\t{$src, $dst|$dst, $src}",
1207 [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
1208 TB, Requires<[HasSSE2]>;
1210 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
1211 "movnti\t{$src, $dst|$dst, $src}",
1212 [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
1213 TB, Requires<[HasSSE2]>;
1216 // Load, store, and memory fence
1217 def SFENCE : I<0xAE, MRM_F8, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>,
1218 TB, Requires<[HasSSE1]>;
1221 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1222 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1223 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1224 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1226 // Alias instructions that map zero vector to pxor / xorp* for sse.
1227 // We set canFoldAsLoad because this can be converted to a constant-pool
1228 // load of an all-zeros value if folding it would be beneficial.
1229 // FIXME: Change encoding to pseudo!
1230 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1231 isCodeGenOnly = 1 in {
1232 def V_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1233 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
1234 def V_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1235 [(set VR128:$dst, (v2f64 immAllZerosV))]>;
1236 let ExeDomain = SSEPackedInt in
1237 def V_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
1238 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1241 def : Pat<(v2i64 immAllZerosV), (V_SET0PI)>;
1242 def : Pat<(v8i16 immAllZerosV), (V_SET0PI)>;
1243 def : Pat<(v16i8 immAllZerosV), (V_SET0PI)>;
1245 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1246 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
1248 //===---------------------------------------------------------------------===//
1249 // SSE2 Instructions
1250 //===---------------------------------------------------------------------===//
1252 // Move Instructions. Register-to-register movsd is not used for FR64
1253 // register copies because it's a partial register update; FsMOVAPDrr is
1254 // used instead. Register-to-register movsd is not modeled as an INSERT_SUBREG
1255 // because INSERT_SUBREG requires that the insert be implementable in terms of
1256 // a copy, and just mentioned, we don't use movsd for copies.
1257 let Constraints = "$src1 = $dst" in
1258 def MOVSDrr : SDI<0x10, MRMSrcReg,
1259 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
1260 "movsd\t{$src2, $dst|$dst, $src2}",
1261 [(set (v2f64 VR128:$dst),
1262 (movl VR128:$src1, (scalar_to_vector FR64:$src2)))]>;
1264 // Extract the low 64-bit value from one vector and insert it into another.
1265 let AddedComplexity = 15 in
1266 def : Pat<(v2f64 (movl VR128:$src1, VR128:$src2)),
1267 (MOVSDrr (v2f64 VR128:$src1),
1268 (EXTRACT_SUBREG (v2f64 VR128:$src2), sub_sd))>;
1270 // Implicitly promote a 64-bit scalar to a vector.
1271 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
1272 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), FR64:$src, sub_sd)>;
1274 // Loading from memory automatically zeroing upper bits.
1275 let canFoldAsLoad = 1, isReMaterializable = 1, AddedComplexity = 20 in
1276 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1277 "movsd\t{$src, $dst|$dst, $src}",
1278 [(set FR64:$dst, (loadf64 addr:$src))]>;
1280 // MOVSDrm zeros the high parts of the register; represent this
1281 // with SUBREG_TO_REG.
1282 let AddedComplexity = 20 in {
1283 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
1284 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1285 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
1286 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1287 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
1288 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1289 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
1290 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1291 def : Pat<(v2f64 (X86vzload addr:$src)),
1292 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1295 // Store scalar value to memory.
1296 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1297 "movsd\t{$src, $dst|$dst, $src}",
1298 [(store FR64:$src, addr:$dst)]>;
1300 // Extract and store.
1301 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
1304 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
1306 // Conversion instructions
1307 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1308 "cvttsd2si\t{$src, $dst|$dst, $src}",
1309 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1310 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1311 "cvttsd2si\t{$src, $dst|$dst, $src}",
1312 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1313 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1314 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1315 [(set FR32:$dst, (fround FR64:$src))]>;
1316 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1317 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1318 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
1319 Requires<[HasSSE2, OptForSize]>;
1320 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1321 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1322 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1323 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1324 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1325 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1327 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1328 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1329 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1330 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1331 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1332 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1333 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1334 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1335 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1336 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1337 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1338 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1339 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1340 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1341 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1342 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1343 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1344 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1345 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1346 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1348 // SSE2 instructions with XS prefix
1349 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1350 "cvtss2sd\t{$src, $dst|$dst, $src}",
1351 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1352 Requires<[HasSSE2]>;
1353 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1354 "cvtss2sd\t{$src, $dst|$dst, $src}",
1355 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1356 Requires<[HasSSE2, OptForSize]>;
1358 def : Pat<(extloadf32 addr:$src),
1359 (CVTSS2SDrr (MOVSSrm addr:$src))>,
1360 Requires<[HasSSE2, OptForSpeed]>;
1362 // Match intrinsics which expect XMM operand(s).
1363 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1364 "cvtsd2si\t{$src, $dst|$dst, $src}",
1365 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1366 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1367 "cvtsd2si\t{$src, $dst|$dst, $src}",
1368 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1369 (load addr:$src)))]>;
1371 // Match intrinsics which expect MM and XMM operand(s).
1372 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1373 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1374 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1375 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1376 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1377 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1378 (memop addr:$src)))]>;
1379 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1380 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1381 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1382 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1383 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1384 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1385 (memop addr:$src)))]>;
1386 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1387 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1388 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1389 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1390 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1391 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1392 (load addr:$src)))]>;
1394 // Aliases for intrinsics
1395 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1396 "cvttsd2si\t{$src, $dst|$dst, $src}",
1398 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1399 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1400 "cvttsd2si\t{$src, $dst|$dst, $src}",
1401 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1402 (load addr:$src)))]>;
1404 // Comparison instructions
1405 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1406 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1407 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1408 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1410 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1411 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1412 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1414 // Accept explicit immediate argument form instead of comparison code.
1415 let isAsmParserOnly = 1 in {
1416 def CMPSDrr_alt : SDIi8<0xC2, MRMSrcReg,
1417 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, i8imm:$src2),
1418 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1420 def CMPSDrm_alt : SDIi8<0xC2, MRMSrcMem,
1421 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, i8imm:$src2),
1422 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1426 let Defs = [EFLAGS] in {
1427 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1428 "ucomisd\t{$src2, $src1|$src1, $src2}",
1429 [(set EFLAGS, (X86cmp FR64:$src1, FR64:$src2))]>;
1430 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1431 "ucomisd\t{$src2, $src1|$src1, $src2}",
1432 [(set EFLAGS, (X86cmp FR64:$src1, (loadf64 addr:$src2)))]>;
1433 } // Defs = [EFLAGS]
1435 // Aliases to match intrinsics which expect XMM operand(s).
1436 let Constraints = "$src1 = $dst" in {
1437 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1439 (ins VR128:$src1, VR128:$src, SSECC:$cc),
1440 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1441 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1442 VR128:$src, imm:$cc))]>;
1443 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1445 (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1446 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1447 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1448 (load addr:$src), imm:$cc))]>;
1451 let Defs = [EFLAGS] in {
1452 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1453 "ucomisd\t{$src2, $src1|$src1, $src2}",
1454 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1456 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1457 "ucomisd\t{$src2, $src1|$src1, $src2}",
1458 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1459 (load addr:$src2)))]>;
1461 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1462 "comisd\t{$src2, $src1|$src1, $src2}",
1463 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1465 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1466 "comisd\t{$src2, $src1|$src1, $src2}",
1467 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1468 (load addr:$src2)))]>;
1469 } // Defs = [EFLAGS]
1471 // Aliases of packed SSE2 instructions for scalar use. These all have names
1472 // that start with 'Fs'.
1474 // Alias instructions that map fld0 to pxor for sse.
1475 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1476 canFoldAsLoad = 1 in
1477 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1478 [(set FR64:$dst, fpimm0)]>,
1479 Requires<[HasSSE2]>, TB, OpSize;
1481 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1483 let neverHasSideEffects = 1 in
1484 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1485 "movapd\t{$src, $dst|$dst, $src}", []>;
1487 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1489 let canFoldAsLoad = 1, isReMaterializable = 1 in
1490 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1491 "movapd\t{$src, $dst|$dst, $src}",
1492 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1494 //===---------------------------------------------------------------------===//
1495 // SSE packed FP Instructions
1497 // Move Instructions
1498 let neverHasSideEffects = 1 in
1499 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1500 "movapd\t{$src, $dst|$dst, $src}", []>;
1501 let canFoldAsLoad = 1, isReMaterializable = 1 in
1502 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1503 "movapd\t{$src, $dst|$dst, $src}",
1504 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1506 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1507 "movapd\t{$src, $dst|$dst, $src}",
1508 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1510 let neverHasSideEffects = 1 in
1511 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1512 "movupd\t{$src, $dst|$dst, $src}", []>;
1513 let canFoldAsLoad = 1 in
1514 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1515 "movupd\t{$src, $dst|$dst, $src}",
1516 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1517 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1518 "movupd\t{$src, $dst|$dst, $src}",
1519 [(store (v2f64 VR128:$src), addr:$dst)]>;
1521 // Intrinsic forms of MOVUPD load and store
1522 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1523 "movupd\t{$src, $dst|$dst, $src}",
1524 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1525 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1526 "movupd\t{$src, $dst|$dst, $src}",
1527 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1529 let Constraints = "$src1 = $dst" in {
1530 let AddedComplexity = 20 in {
1531 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1532 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1533 "movlpd\t{$src2, $dst|$dst, $src2}",
1535 (v2f64 (movlp VR128:$src1,
1536 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1537 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1538 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1539 "movhpd\t{$src2, $dst|$dst, $src2}",
1541 (v2f64 (movlhps VR128:$src1,
1542 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1543 } // AddedComplexity
1544 } // Constraints = "$src1 = $dst"
1546 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1547 "movlpd\t{$src, $dst|$dst, $src}",
1548 [(store (f64 (vector_extract (v2f64 VR128:$src),
1549 (iPTR 0))), addr:$dst)]>;
1551 // v2f64 extract element 1 is always custom lowered to unpack high to low
1552 // and extract element 0 so the non-store version isn't too horrible.
1553 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1554 "movhpd\t{$src, $dst|$dst, $src}",
1555 [(store (f64 (vector_extract
1556 (v2f64 (unpckh VR128:$src, (undef))),
1557 (iPTR 0))), addr:$dst)]>;
1559 // SSE2 instructions without OpSize prefix
1560 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1561 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1562 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1563 TB, Requires<[HasSSE2]>;
1564 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1565 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1566 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1567 (bitconvert (memopv2i64 addr:$src))))]>,
1568 TB, Requires<[HasSSE2]>;
1570 // SSE2 instructions with XS prefix
1571 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1572 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1573 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1574 XS, Requires<[HasSSE2]>;
1575 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1576 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1577 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1578 (bitconvert (memopv2i64 addr:$src))))]>,
1579 XS, Requires<[HasSSE2]>;
1581 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1582 "cvtps2dq\t{$src, $dst|$dst, $src}",
1583 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1584 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1585 "cvtps2dq\t{$src, $dst|$dst, $src}",
1586 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1587 (memop addr:$src)))]>;
1588 // SSE2 packed instructions with XS prefix
1589 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1590 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1591 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1592 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1594 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1595 "cvttps2dq\t{$src, $dst|$dst, $src}",
1597 (int_x86_sse2_cvttps2dq VR128:$src))]>,
1598 XS, Requires<[HasSSE2]>;
1599 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1600 "cvttps2dq\t{$src, $dst|$dst, $src}",
1601 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1602 (memop addr:$src)))]>,
1603 XS, Requires<[HasSSE2]>;
1605 // SSE2 packed instructions with XD prefix
1606 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1607 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1608 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1609 XD, Requires<[HasSSE2]>;
1610 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1611 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1612 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1613 (memop addr:$src)))]>,
1614 XD, Requires<[HasSSE2]>;
1616 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1617 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1618 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1619 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1620 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1621 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1622 (memop addr:$src)))]>;
1624 // SSE2 instructions without OpSize prefix
1625 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1626 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1627 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1628 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1630 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1631 "cvtps2pd\t{$src, $dst|$dst, $src}",
1632 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1633 TB, Requires<[HasSSE2]>;
1634 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1635 "cvtps2pd\t{$src, $dst|$dst, $src}",
1636 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1637 (load addr:$src)))]>,
1638 TB, Requires<[HasSSE2]>;
1640 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1641 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1642 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1643 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1646 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1647 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1648 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1649 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1650 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1651 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1652 (memop addr:$src)))]>;
1654 // Match intrinsics which expect XMM operand(s).
1655 // Aliases for intrinsics
1656 let Constraints = "$src1 = $dst" in {
1657 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1658 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1659 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1660 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1662 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1663 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1664 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1665 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1666 (loadi32 addr:$src2)))]>;
1667 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1668 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1669 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1670 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1672 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1673 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1674 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1675 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1676 (load addr:$src2)))]>;
1677 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1678 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1679 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1680 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1681 VR128:$src2))]>, XS,
1682 Requires<[HasSSE2]>;
1683 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1684 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1685 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1686 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1687 (load addr:$src2)))]>, XS,
1688 Requires<[HasSSE2]>;
1693 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1695 /// In addition, we also have a special variant of the scalar form here to
1696 /// represent the associated intrinsic operation. This form is unlike the
1697 /// plain scalar form, in that it takes an entire vector (instead of a
1698 /// scalar) and leaves the top elements undefined.
1700 /// And, we have a special variant form for a full-vector intrinsic form.
1702 /// These four forms can each have a reg or a mem operand, so there are a
1703 /// total of eight "instructions".
1705 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1709 bit Commutable = 0> {
1710 // Scalar operation, reg.
1711 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1712 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1713 [(set FR64:$dst, (OpNode FR64:$src))]> {
1714 let isCommutable = Commutable;
1717 // Scalar operation, mem.
1718 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1719 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1720 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1722 // Vector operation, reg.
1723 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1724 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1725 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1726 let isCommutable = Commutable;
1729 // Vector operation, mem.
1730 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1731 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1732 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1734 // Intrinsic operation, reg.
1735 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1736 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1737 [(set VR128:$dst, (F64Int VR128:$src))]> {
1738 let isCommutable = Commutable;
1741 // Intrinsic operation, mem.
1742 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1743 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1744 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1746 // Vector intrinsic operation, reg
1747 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1748 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1749 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1750 let isCommutable = Commutable;
1753 // Vector intrinsic operation, mem
1754 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1755 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1756 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1760 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1761 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1763 // There is no f64 version of the reciprocal approximation instructions.
1765 let Constraints = "$src1 = $dst" in {
1766 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1767 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1768 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1769 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1770 VR128:$src, imm:$cc))]>;
1771 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1772 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1773 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1774 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1775 (memop addr:$src), imm:$cc))]>;
1777 // Accept explicit immediate argument form instead of comparison code.
1778 let isAsmParserOnly = 1 in {
1779 def CMPPDrri_alt : PDIi8<0xC2, MRMSrcReg,
1780 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1781 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1782 def CMPPDrmi_alt : PDIi8<0xC2, MRMSrcMem,
1783 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1784 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1787 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1788 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1789 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1790 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1792 // Shuffle and unpack instructions
1793 let Constraints = "$src1 = $dst" in {
1794 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1795 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1796 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1798 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1799 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1800 (outs VR128:$dst), (ins VR128:$src1,
1801 f128mem:$src2, i8imm:$src3),
1802 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1805 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1807 let AddedComplexity = 10 in {
1808 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1809 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1810 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1812 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1813 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1814 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1815 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1817 (v2f64 (unpckh VR128:$src1,
1818 (memopv2f64 addr:$src2))))]>;
1820 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1821 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1822 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1824 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1825 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1826 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1827 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1829 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1830 } // AddedComplexity
1831 } // Constraints = "$src1 = $dst"
1834 //===---------------------------------------------------------------------===//
1835 // SSE integer instructions
1836 let ExeDomain = SSEPackedInt in {
1838 // Move Instructions
1839 let neverHasSideEffects = 1 in
1840 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1841 "movdqa\t{$src, $dst|$dst, $src}", []>;
1842 let canFoldAsLoad = 1, mayLoad = 1 in
1843 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1844 "movdqa\t{$src, $dst|$dst, $src}",
1845 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1847 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1848 "movdqa\t{$src, $dst|$dst, $src}",
1849 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1850 let canFoldAsLoad = 1, mayLoad = 1 in
1851 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1852 "movdqu\t{$src, $dst|$dst, $src}",
1853 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1854 XS, Requires<[HasSSE2]>;
1856 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1857 "movdqu\t{$src, $dst|$dst, $src}",
1858 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1859 XS, Requires<[HasSSE2]>;
1861 // Intrinsic forms of MOVDQU load and store
1862 let canFoldAsLoad = 1 in
1863 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1864 "movdqu\t{$src, $dst|$dst, $src}",
1865 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1866 XS, Requires<[HasSSE2]>;
1867 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1868 "movdqu\t{$src, $dst|$dst, $src}",
1869 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1870 XS, Requires<[HasSSE2]>;
1872 let Constraints = "$src1 = $dst" in {
1874 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1875 bit Commutable = 0> {
1876 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1877 (ins VR128:$src1, VR128:$src2),
1878 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1879 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1880 let isCommutable = Commutable;
1882 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1883 (ins VR128:$src1, i128mem:$src2),
1884 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1885 [(set VR128:$dst, (IntId VR128:$src1,
1886 (bitconvert (memopv2i64
1890 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1892 Intrinsic IntId, Intrinsic IntId2> {
1893 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1894 (ins VR128:$src1, VR128:$src2),
1895 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1896 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1897 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1898 (ins VR128:$src1, i128mem:$src2),
1899 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1900 [(set VR128:$dst, (IntId VR128:$src1,
1901 (bitconvert (memopv2i64 addr:$src2))))]>;
1902 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
1903 (ins VR128:$src1, i32i8imm:$src2),
1904 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1905 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1908 /// PDI_binop_rm - Simple SSE2 binary operator.
1909 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1910 ValueType OpVT, bit Commutable = 0> {
1911 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1912 (ins VR128:$src1, VR128:$src2),
1913 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1914 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1915 let isCommutable = Commutable;
1917 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1918 (ins VR128:$src1, i128mem:$src2),
1919 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1920 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1921 (bitconvert (memopv2i64 addr:$src2)))))]>;
1924 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1926 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1927 /// to collapse (bitconvert VT to VT) into its operand.
1929 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1930 bit Commutable = 0> {
1931 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1932 (ins VR128:$src1, VR128:$src2),
1933 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1934 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1935 let isCommutable = Commutable;
1937 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1938 (ins VR128:$src1, i128mem:$src2),
1939 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1940 [(set VR128:$dst, (OpNode VR128:$src1,
1941 (memopv2i64 addr:$src2)))]>;
1944 } // Constraints = "$src1 = $dst"
1945 } // ExeDomain = SSEPackedInt
1947 // 128-bit Integer Arithmetic
1949 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1950 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1951 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1952 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1954 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1955 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1956 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1957 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1959 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1960 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1961 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1962 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1964 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1965 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1966 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1967 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1969 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1971 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1972 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1973 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1975 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1977 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1978 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1981 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1982 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1983 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1984 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1985 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
1988 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
1989 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
1990 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
1991 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
1992 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
1993 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
1995 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
1996 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
1997 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
1998 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
1999 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2000 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2002 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2003 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2004 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2005 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2007 // 128-bit logical shifts.
2008 let Constraints = "$src1 = $dst", neverHasSideEffects = 1,
2009 ExeDomain = SSEPackedInt in {
2010 def PSLLDQri : PDIi8<0x73, MRM7r,
2011 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2012 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2013 def PSRLDQri : PDIi8<0x73, MRM3r,
2014 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2015 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2016 // PSRADQri doesn't exist in SSE[1-3].
2019 let Predicates = [HasSSE2] in {
2020 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2021 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2022 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2023 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2024 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2025 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2026 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2027 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2028 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2029 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2031 // Shift up / down and insert zero's.
2032 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2033 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2034 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2035 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2039 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2040 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2041 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2043 let Constraints = "$src1 = $dst", ExeDomain = SSEPackedInt in {
2044 def PANDNrr : PDI<0xDF, MRMSrcReg,
2045 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2046 "pandn\t{$src2, $dst|$dst, $src2}",
2047 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2050 def PANDNrm : PDI<0xDF, MRMSrcMem,
2051 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2052 "pandn\t{$src2, $dst|$dst, $src2}",
2053 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2054 (memopv2i64 addr:$src2))))]>;
2057 // SSE2 Integer comparison
2058 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2059 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2060 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2061 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2062 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2063 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2065 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2066 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2067 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2068 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2069 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2070 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2071 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2072 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2073 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2074 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2075 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2076 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2078 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2079 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2080 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2081 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2082 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2083 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2084 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2085 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2086 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2087 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2088 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2089 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2092 // Pack instructions
2093 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2094 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2095 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2097 let ExeDomain = SSEPackedInt in {
2099 // Shuffle and unpack instructions
2100 let AddedComplexity = 5 in {
2101 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2102 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2103 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2104 [(set VR128:$dst, (v4i32 (pshufd:$src2
2105 VR128:$src1, (undef))))]>;
2106 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2107 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2108 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2109 [(set VR128:$dst, (v4i32 (pshufd:$src2
2110 (bc_v4i32 (memopv2i64 addr:$src1)),
2114 // SSE2 with ImmT == Imm8 and XS prefix.
2115 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2116 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2117 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2118 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2120 XS, Requires<[HasSSE2]>;
2121 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2122 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2123 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2124 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2125 (bc_v8i16 (memopv2i64 addr:$src1)),
2127 XS, Requires<[HasSSE2]>;
2129 // SSE2 with ImmT == Imm8 and XD prefix.
2130 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2131 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2132 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2133 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2135 XD, Requires<[HasSSE2]>;
2136 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2137 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2138 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2139 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2140 (bc_v8i16 (memopv2i64 addr:$src1)),
2142 XD, Requires<[HasSSE2]>;
2144 // Unpack instructions
2145 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
2146 PatFrag unp_frag, PatFrag bc_frag> {
2147 def rr : PDI<opc, MRMSrcReg,
2148 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2149 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2150 [(set VR128:$dst, (vt (unp_frag VR128:$src1, VR128:$src2)))]>;
2151 def rm : PDI<opc, MRMSrcMem,
2152 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2153 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2154 [(set VR128:$dst, (unp_frag VR128:$src1,
2155 (bc_frag (memopv2i64
2159 let Constraints = "$src1 = $dst" in {
2160 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, unpckl, bc_v16i8>;
2161 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, unpckl, bc_v8i16>;
2162 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, unpckl, bc_v4i32>;
2164 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2165 /// knew to collapse (bitconvert VT to VT) into its operand.
2166 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2167 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2168 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2170 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2171 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2172 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2173 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2175 (v2i64 (unpckl VR128:$src1,
2176 (memopv2i64 addr:$src2))))]>;
2178 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, unpckh, bc_v16i8>;
2179 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, unpckh, bc_v8i16>;
2180 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, unpckh, bc_v4i32>;
2182 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2183 /// knew to collapse (bitconvert VT to VT) into its operand.
2184 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2185 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2186 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2188 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2189 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2190 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2191 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2193 (v2i64 (unpckh VR128:$src1,
2194 (memopv2i64 addr:$src2))))]>;
2198 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2199 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2200 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2201 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2203 let Constraints = "$src1 = $dst" in {
2204 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2205 (outs VR128:$dst), (ins VR128:$src1,
2206 GR32:$src2, i32i8imm:$src3),
2207 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2209 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2210 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2211 (outs VR128:$dst), (ins VR128:$src1,
2212 i16mem:$src2, i32i8imm:$src3),
2213 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2215 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2220 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2221 "pmovmskb\t{$src, $dst|$dst, $src}",
2222 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2224 // Conditional store
2226 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2227 "maskmovdqu\t{$mask, $src|$src, $mask}",
2228 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2231 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2232 "maskmovdqu\t{$mask, $src|$src, $mask}",
2233 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2235 } // ExeDomain = SSEPackedInt
2237 // Non-temporal stores
2238 def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2239 "movntpd\t{$src, $dst|$dst, $src}",
2240 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2241 let ExeDomain = SSEPackedInt in
2242 def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2243 "movntdq\t{$src, $dst|$dst, $src}",
2244 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2245 def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2246 "movnti\t{$src, $dst|$dst, $src}",
2247 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2248 TB, Requires<[HasSSE2]>;
2250 let AddedComplexity = 400 in { // Prefer non-temporal versions
2251 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2252 "movntpd\t{$src, $dst|$dst, $src}",
2253 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
2255 let ExeDomain = SSEPackedInt in
2256 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2257 "movntdq\t{$src, $dst|$dst, $src}",
2258 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2262 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2263 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2264 TB, Requires<[HasSSE2]>;
2266 // Load, store, and memory fence
2267 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
2268 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2269 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
2270 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2272 // Pause. This "instruction" is encoded as "rep; nop", so even though it
2273 // was introduced with SSE2, it's backward compatible.
2274 def PAUSE : I<0x90, RawFrm, (outs), (ins), "pause", []>, REP;
2276 //TODO: custom lower this so as to never even generate the noop
2277 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2279 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2280 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2281 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2284 // Alias instructions that map zero vector to pxor / xorp* for sse.
2285 // We set canFoldAsLoad because this can be converted to a constant-pool
2286 // load of an all-ones value if folding it would be beneficial.
2287 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2288 isCodeGenOnly = 1, ExeDomain = SSEPackedInt in
2289 // FIXME: Change encoding to pseudo.
2290 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
2291 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2293 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2294 "movd\t{$src, $dst|$dst, $src}",
2296 (v4i32 (scalar_to_vector GR32:$src)))]>;
2297 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2298 "movd\t{$src, $dst|$dst, $src}",
2300 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2302 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2303 "movd\t{$src, $dst|$dst, $src}",
2304 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2306 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2307 "movd\t{$src, $dst|$dst, $src}",
2308 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2310 // SSE2 instructions with XS prefix
2311 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2312 "movq\t{$src, $dst|$dst, $src}",
2314 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2315 Requires<[HasSSE2]>;
2316 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2317 "movq\t{$src, $dst|$dst, $src}",
2318 [(store (i64 (vector_extract (v2i64 VR128:$src),
2319 (iPTR 0))), addr:$dst)]>;
2321 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
2322 (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
2324 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2325 "movd\t{$src, $dst|$dst, $src}",
2326 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2328 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2329 "movd\t{$src, $dst|$dst, $src}",
2330 [(store (i32 (vector_extract (v4i32 VR128:$src),
2331 (iPTR 0))), addr:$dst)]>;
2333 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2334 "movd\t{$src, $dst|$dst, $src}",
2335 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2336 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2337 "movd\t{$src, $dst|$dst, $src}",
2338 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2340 // Store / copy lower 64-bits of a XMM register.
2341 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2342 "movq\t{$src, $dst|$dst, $src}",
2343 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2345 // movd / movq to XMM register zero-extends
2346 let AddedComplexity = 15 in {
2347 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2348 "movd\t{$src, $dst|$dst, $src}",
2349 [(set VR128:$dst, (v4i32 (X86vzmovl
2350 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2351 // This is X86-64 only.
2352 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2353 "mov{d|q}\t{$src, $dst|$dst, $src}",
2354 [(set VR128:$dst, (v2i64 (X86vzmovl
2355 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2358 let AddedComplexity = 20 in {
2359 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2360 "movd\t{$src, $dst|$dst, $src}",
2362 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2363 (loadi32 addr:$src))))))]>;
2365 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2366 (MOVZDI2PDIrm addr:$src)>;
2367 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2368 (MOVZDI2PDIrm addr:$src)>;
2369 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2370 (MOVZDI2PDIrm addr:$src)>;
2372 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2373 "movq\t{$src, $dst|$dst, $src}",
2375 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2376 (loadi64 addr:$src))))))]>, XS,
2377 Requires<[HasSSE2]>;
2379 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2380 (MOVZQI2PQIrm addr:$src)>;
2381 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2382 (MOVZQI2PQIrm addr:$src)>;
2383 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2386 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2387 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2388 let AddedComplexity = 15 in
2389 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2390 "movq\t{$src, $dst|$dst, $src}",
2391 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2392 XS, Requires<[HasSSE2]>;
2394 let AddedComplexity = 20 in {
2395 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2396 "movq\t{$src, $dst|$dst, $src}",
2397 [(set VR128:$dst, (v2i64 (X86vzmovl
2398 (loadv2i64 addr:$src))))]>,
2399 XS, Requires<[HasSSE2]>;
2401 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2402 (MOVZPQILo2PQIrm addr:$src)>;
2405 // Instructions for the disassembler
2406 // xr = XMM register
2409 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2410 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2412 //===---------------------------------------------------------------------===//
2413 // SSE3 Instructions
2414 //===---------------------------------------------------------------------===//
2416 // Move Instructions
2417 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2418 "movshdup\t{$src, $dst|$dst, $src}",
2419 [(set VR128:$dst, (v4f32 (movshdup
2420 VR128:$src, (undef))))]>;
2421 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2422 "movshdup\t{$src, $dst|$dst, $src}",
2423 [(set VR128:$dst, (movshdup
2424 (memopv4f32 addr:$src), (undef)))]>;
2426 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2427 "movsldup\t{$src, $dst|$dst, $src}",
2428 [(set VR128:$dst, (v4f32 (movsldup
2429 VR128:$src, (undef))))]>;
2430 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2431 "movsldup\t{$src, $dst|$dst, $src}",
2432 [(set VR128:$dst, (movsldup
2433 (memopv4f32 addr:$src), (undef)))]>;
2435 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2436 "movddup\t{$src, $dst|$dst, $src}",
2437 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2438 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2439 "movddup\t{$src, $dst|$dst, $src}",
2441 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2444 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2446 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2448 let AddedComplexity = 5 in {
2449 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2450 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2451 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2452 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2453 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2454 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2455 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2456 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2460 let Constraints = "$src1 = $dst" in {
2461 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2462 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2463 "addsubps\t{$src2, $dst|$dst, $src2}",
2464 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2466 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2467 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2468 "addsubps\t{$src2, $dst|$dst, $src2}",
2469 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2470 (memop addr:$src2)))]>;
2471 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2472 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2473 "addsubpd\t{$src2, $dst|$dst, $src2}",
2474 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2476 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2477 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2478 "addsubpd\t{$src2, $dst|$dst, $src2}",
2479 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2480 (memop addr:$src2)))]>;
2483 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2484 "lddqu\t{$src, $dst|$dst, $src}",
2485 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2488 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2489 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2490 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2491 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2492 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2493 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2494 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2495 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2496 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2497 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2498 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2499 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2500 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2501 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2502 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2503 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2505 let Constraints = "$src1 = $dst" in {
2506 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2507 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2508 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2509 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2510 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2511 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2512 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2513 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2516 // Thread synchronization
2517 def MONITOR : I<0x01, MRM_C8, (outs), (ins), "monitor",
2518 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2519 def MWAIT : I<0x01, MRM_C9, (outs), (ins), "mwait",
2520 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2522 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2523 let AddedComplexity = 15 in
2524 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2525 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2526 let AddedComplexity = 20 in
2527 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2528 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2530 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2531 let AddedComplexity = 15 in
2532 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2533 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2534 let AddedComplexity = 20 in
2535 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2536 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2538 //===---------------------------------------------------------------------===//
2539 // SSSE3 Instructions
2540 //===---------------------------------------------------------------------===//
2542 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2543 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2544 Intrinsic IntId64, Intrinsic IntId128> {
2545 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2546 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2547 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2549 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2550 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2552 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2554 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2556 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2557 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2560 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2562 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2565 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2568 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2569 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2570 Intrinsic IntId64, Intrinsic IntId128> {
2571 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2573 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2574 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2576 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2578 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2581 (bitconvert (memopv4i16 addr:$src))))]>;
2583 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2585 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2586 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2589 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2591 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2594 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2597 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2598 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2599 Intrinsic IntId64, Intrinsic IntId128> {
2600 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2602 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2603 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2605 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2607 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2610 (bitconvert (memopv2i32 addr:$src))))]>;
2612 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2614 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2615 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2618 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2620 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2623 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2626 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2627 int_x86_ssse3_pabs_b,
2628 int_x86_ssse3_pabs_b_128>;
2629 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2630 int_x86_ssse3_pabs_w,
2631 int_x86_ssse3_pabs_w_128>;
2632 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2633 int_x86_ssse3_pabs_d,
2634 int_x86_ssse3_pabs_d_128>;
2636 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2637 let Constraints = "$src1 = $dst" in {
2638 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2639 Intrinsic IntId64, Intrinsic IntId128,
2640 bit Commutable = 0> {
2641 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2642 (ins VR64:$src1, VR64:$src2),
2643 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2644 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2645 let isCommutable = Commutable;
2647 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2648 (ins VR64:$src1, i64mem:$src2),
2649 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2651 (IntId64 VR64:$src1,
2652 (bitconvert (memopv8i8 addr:$src2))))]>;
2654 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2655 (ins VR128:$src1, VR128:$src2),
2656 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2657 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2659 let isCommutable = Commutable;
2661 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2662 (ins VR128:$src1, i128mem:$src2),
2663 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2665 (IntId128 VR128:$src1,
2666 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2670 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2671 let Constraints = "$src1 = $dst" in {
2672 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2673 Intrinsic IntId64, Intrinsic IntId128,
2674 bit Commutable = 0> {
2675 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2676 (ins VR64:$src1, VR64:$src2),
2677 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2678 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2679 let isCommutable = Commutable;
2681 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2682 (ins VR64:$src1, i64mem:$src2),
2683 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2685 (IntId64 VR64:$src1,
2686 (bitconvert (memopv4i16 addr:$src2))))]>;
2688 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2689 (ins VR128:$src1, VR128:$src2),
2690 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2691 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2693 let isCommutable = Commutable;
2695 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2696 (ins VR128:$src1, i128mem:$src2),
2697 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2699 (IntId128 VR128:$src1,
2700 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2704 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2705 let Constraints = "$src1 = $dst" in {
2706 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2707 Intrinsic IntId64, Intrinsic IntId128,
2708 bit Commutable = 0> {
2709 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2710 (ins VR64:$src1, VR64:$src2),
2711 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2712 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2713 let isCommutable = Commutable;
2715 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2716 (ins VR64:$src1, i64mem:$src2),
2717 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2719 (IntId64 VR64:$src1,
2720 (bitconvert (memopv2i32 addr:$src2))))]>;
2722 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2723 (ins VR128:$src1, VR128:$src2),
2724 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2725 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2727 let isCommutable = Commutable;
2729 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2730 (ins VR128:$src1, i128mem:$src2),
2731 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2733 (IntId128 VR128:$src1,
2734 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2738 let ImmT = NoImm in { // None of these have i8 immediate fields.
2739 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2740 int_x86_ssse3_phadd_w,
2741 int_x86_ssse3_phadd_w_128>;
2742 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2743 int_x86_ssse3_phadd_d,
2744 int_x86_ssse3_phadd_d_128>;
2745 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2746 int_x86_ssse3_phadd_sw,
2747 int_x86_ssse3_phadd_sw_128>;
2748 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2749 int_x86_ssse3_phsub_w,
2750 int_x86_ssse3_phsub_w_128>;
2751 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2752 int_x86_ssse3_phsub_d,
2753 int_x86_ssse3_phsub_d_128>;
2754 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2755 int_x86_ssse3_phsub_sw,
2756 int_x86_ssse3_phsub_sw_128>;
2757 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2758 int_x86_ssse3_pmadd_ub_sw,
2759 int_x86_ssse3_pmadd_ub_sw_128>;
2760 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2761 int_x86_ssse3_pmul_hr_sw,
2762 int_x86_ssse3_pmul_hr_sw_128, 1>;
2764 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2765 int_x86_ssse3_pshuf_b,
2766 int_x86_ssse3_pshuf_b_128>;
2767 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2768 int_x86_ssse3_psign_b,
2769 int_x86_ssse3_psign_b_128>;
2770 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2771 int_x86_ssse3_psign_w,
2772 int_x86_ssse3_psign_w_128>;
2773 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2774 int_x86_ssse3_psign_d,
2775 int_x86_ssse3_psign_d_128>;
2778 // palignr patterns.
2779 let Constraints = "$src1 = $dst" in {
2780 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2781 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2782 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2784 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2785 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2786 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2789 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2790 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2791 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2793 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2794 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2795 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2799 let AddedComplexity = 5 in {
2801 def : Pat<(v1i64 (palign:$src3 VR64:$src1, VR64:$src2)),
2802 (PALIGNR64rr VR64:$src2, VR64:$src1,
2803 (SHUFFLE_get_palign_imm VR64:$src3))>,
2804 Requires<[HasSSSE3]>;
2805 def : Pat<(v2i32 (palign:$src3 VR64:$src1, VR64:$src2)),
2806 (PALIGNR64rr VR64:$src2, VR64:$src1,
2807 (SHUFFLE_get_palign_imm VR64:$src3))>,
2808 Requires<[HasSSSE3]>;
2809 def : Pat<(v2f32 (palign:$src3 VR64:$src1, VR64:$src2)),
2810 (PALIGNR64rr VR64:$src2, VR64:$src1,
2811 (SHUFFLE_get_palign_imm VR64:$src3))>,
2812 Requires<[HasSSSE3]>;
2813 def : Pat<(v4i16 (palign:$src3 VR64:$src1, VR64:$src2)),
2814 (PALIGNR64rr VR64:$src2, VR64:$src1,
2815 (SHUFFLE_get_palign_imm VR64:$src3))>,
2816 Requires<[HasSSSE3]>;
2817 def : Pat<(v8i8 (palign:$src3 VR64:$src1, VR64:$src2)),
2818 (PALIGNR64rr VR64:$src2, VR64:$src1,
2819 (SHUFFLE_get_palign_imm VR64:$src3))>,
2820 Requires<[HasSSSE3]>;
2822 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2823 (PALIGNR128rr VR128:$src2, VR128:$src1,
2824 (SHUFFLE_get_palign_imm VR128:$src3))>,
2825 Requires<[HasSSSE3]>;
2826 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2827 (PALIGNR128rr VR128:$src2, VR128:$src1,
2828 (SHUFFLE_get_palign_imm VR128:$src3))>,
2829 Requires<[HasSSSE3]>;
2830 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2831 (PALIGNR128rr VR128:$src2, VR128:$src1,
2832 (SHUFFLE_get_palign_imm VR128:$src3))>,
2833 Requires<[HasSSSE3]>;
2834 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2835 (PALIGNR128rr VR128:$src2, VR128:$src1,
2836 (SHUFFLE_get_palign_imm VR128:$src3))>,
2837 Requires<[HasSSSE3]>;
2840 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2841 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2842 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2843 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2845 //===---------------------------------------------------------------------===//
2846 // Non-Instruction Patterns
2847 //===---------------------------------------------------------------------===//
2849 // extload f32 -> f64. This matches load+fextend because we have a hack in
2850 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2852 // Since these loads aren't folded into the fextend, we have to match it
2854 let Predicates = [HasSSE2] in
2855 def : Pat<(fextend (loadf32 addr:$src)),
2856 (CVTSS2SDrm addr:$src)>;
2859 let Predicates = [HasSSE2] in {
2860 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2861 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2862 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2863 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2864 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2865 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2866 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2867 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2868 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2869 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2870 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2871 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2872 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2873 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2874 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2875 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2876 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2877 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2878 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2879 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2880 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2881 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2882 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2883 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2884 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2885 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2886 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2887 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2888 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2889 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2892 // Move scalar to XMM zero-extended
2893 // movd to XMM register zero-extends
2894 let AddedComplexity = 15 in {
2895 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2896 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2897 (MOVSDrr (v2f64 (V_SET0PS)), FR64:$src)>;
2898 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2899 (MOVSSrr (v4f32 (V_SET0PS)), FR32:$src)>;
2900 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2901 (MOVSSrr (v4f32 (V_SET0PS)),
2902 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss)))>;
2903 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2904 (MOVSSrr (v4i32 (V_SET0PI)),
2905 (EXTRACT_SUBREG (v4i32 VR128:$src), sub_ss))>;
2908 // Splat v2f64 / v2i64
2909 let AddedComplexity = 10 in {
2910 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
2911 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2912 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
2913 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2914 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
2915 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2916 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
2917 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2920 // Special unary SHUFPSrri case.
2921 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
2922 (SHUFPSrri VR128:$src1, VR128:$src1,
2923 (SHUFFLE_get_shuf_imm VR128:$src3))>;
2924 let AddedComplexity = 5 in
2925 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
2926 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2927 Requires<[HasSSE2]>;
2928 // Special unary SHUFPDrri case.
2929 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
2930 (SHUFPDrri VR128:$src1, VR128:$src1,
2931 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2932 Requires<[HasSSE2]>;
2933 // Special unary SHUFPDrri case.
2934 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
2935 (SHUFPDrri VR128:$src1, VR128:$src1,
2936 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2937 Requires<[HasSSE2]>;
2938 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2939 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
2940 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2941 Requires<[HasSSE2]>;
2943 // Special binary v4i32 shuffle cases with SHUFPS.
2944 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
2945 (SHUFPSrri VR128:$src1, VR128:$src2,
2946 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2947 Requires<[HasSSE2]>;
2948 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
2949 (SHUFPSrmi VR128:$src1, addr:$src2,
2950 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2951 Requires<[HasSSE2]>;
2952 // Special binary v2i64 shuffle cases using SHUFPDrri.
2953 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
2954 (SHUFPDrri VR128:$src1, VR128:$src2,
2955 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2956 Requires<[HasSSE2]>;
2958 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2959 let AddedComplexity = 15 in {
2960 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
2961 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2962 Requires<[OptForSpeed, HasSSE2]>;
2963 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
2964 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2965 Requires<[OptForSpeed, HasSSE2]>;
2967 let AddedComplexity = 10 in {
2968 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
2969 (UNPCKLPSrr VR128:$src, VR128:$src)>;
2970 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
2971 (PUNPCKLBWrr VR128:$src, VR128:$src)>;
2972 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
2973 (PUNPCKLWDrr VR128:$src, VR128:$src)>;
2974 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
2975 (PUNPCKLDQrr VR128:$src, VR128:$src)>;
2978 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2979 let AddedComplexity = 15 in {
2980 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
2981 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2982 Requires<[OptForSpeed, HasSSE2]>;
2983 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
2984 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2985 Requires<[OptForSpeed, HasSSE2]>;
2987 let AddedComplexity = 10 in {
2988 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
2989 (UNPCKHPSrr VR128:$src, VR128:$src)>;
2990 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
2991 (PUNPCKHBWrr VR128:$src, VR128:$src)>;
2992 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
2993 (PUNPCKHWDrr VR128:$src, VR128:$src)>;
2994 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
2995 (PUNPCKHDQrr VR128:$src, VR128:$src)>;
2998 let AddedComplexity = 20 in {
2999 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3000 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3001 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3003 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3004 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3005 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3007 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3008 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3009 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3010 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3011 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3014 let AddedComplexity = 20 in {
3015 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3016 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3017 (MOVLPSrm VR128:$src1, addr:$src2)>;
3018 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3019 (MOVLPDrm VR128:$src1, addr:$src2)>;
3020 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3021 (MOVLPSrm VR128:$src1, addr:$src2)>;
3022 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3023 (MOVLPDrm VR128:$src1, addr:$src2)>;
3026 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3027 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3028 (MOVLPSmr addr:$src1, VR128:$src2)>;
3029 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3030 (MOVLPDmr addr:$src1, VR128:$src2)>;
3031 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3033 (MOVLPSmr addr:$src1, VR128:$src2)>;
3034 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3035 (MOVLPDmr addr:$src1, VR128:$src2)>;
3037 let AddedComplexity = 15 in {
3038 // Setting the lowest element in the vector.
3039 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3040 (MOVSSrr (v4i32 VR128:$src1),
3041 (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_ss))>;
3042 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3043 (MOVSDrr (v2i64 VR128:$src1),
3044 (EXTRACT_SUBREG (v2i64 VR128:$src2), sub_sd))>;
3046 // vector_shuffle v1, v2 <4, 5, 2, 3> using movsd
3047 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3048 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3049 Requires<[HasSSE2]>;
3050 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3051 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3052 Requires<[HasSSE2]>;
3055 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3056 // fall back to this for SSE1)
3057 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3058 (SHUFPSrri VR128:$src2, VR128:$src1,
3059 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3061 // Set lowest element and zero upper elements.
3062 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3063 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3065 // Some special case pandn patterns.
3066 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3068 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3069 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3071 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3072 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3074 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3076 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3077 (memop addr:$src2))),
3078 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3079 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3080 (memop addr:$src2))),
3081 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3082 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3083 (memop addr:$src2))),
3084 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3086 // vector -> vector casts
3087 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3088 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3089 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3090 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3091 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3092 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3093 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3094 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3096 // Use movaps / movups for SSE integer load / store (one byte shorter).
3097 def : Pat<(alignedloadv4i32 addr:$src),
3098 (MOVAPSrm addr:$src)>;
3099 def : Pat<(loadv4i32 addr:$src),
3100 (MOVUPSrm addr:$src)>;
3101 def : Pat<(alignedloadv2i64 addr:$src),
3102 (MOVAPSrm addr:$src)>;
3103 def : Pat<(loadv2i64 addr:$src),
3104 (MOVUPSrm addr:$src)>;
3106 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3107 (MOVAPSmr addr:$dst, VR128:$src)>;
3108 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3109 (MOVAPSmr addr:$dst, VR128:$src)>;
3110 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3111 (MOVAPSmr addr:$dst, VR128:$src)>;
3112 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3113 (MOVAPSmr addr:$dst, VR128:$src)>;
3114 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3115 (MOVUPSmr addr:$dst, VR128:$src)>;
3116 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3117 (MOVUPSmr addr:$dst, VR128:$src)>;
3118 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3119 (MOVUPSmr addr:$dst, VR128:$src)>;
3120 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3121 (MOVUPSmr addr:$dst, VR128:$src)>;
3123 //===----------------------------------------------------------------------===//
3124 // SSE4.1 Instructions
3125 //===----------------------------------------------------------------------===//
3127 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3130 Intrinsic V2F64Int> {
3131 // Intrinsic operation, reg.
3132 // Vector intrinsic operation, reg
3133 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3134 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3135 !strconcat(OpcodeStr,
3136 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3137 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3140 // Vector intrinsic operation, mem
3141 def PSm_Int : Ii8<opcps, MRMSrcMem,
3142 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3143 !strconcat(OpcodeStr,
3144 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3146 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3148 Requires<[HasSSE41]>;
3150 // Vector intrinsic operation, reg
3151 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3152 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3153 !strconcat(OpcodeStr,
3154 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3155 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3158 // Vector intrinsic operation, mem
3159 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3160 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3161 !strconcat(OpcodeStr,
3162 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3164 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3168 let Constraints = "$src1 = $dst" in {
3169 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3173 // Intrinsic operation, reg.
3174 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3176 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3177 !strconcat(OpcodeStr,
3178 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3180 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3183 // Intrinsic operation, mem.
3184 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3186 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3187 !strconcat(OpcodeStr,
3188 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3190 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3193 // Intrinsic operation, reg.
3194 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3196 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3197 !strconcat(OpcodeStr,
3198 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3200 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3203 // Intrinsic operation, mem.
3204 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3206 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3207 !strconcat(OpcodeStr,
3208 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3210 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3215 // FP round - roundss, roundps, roundsd, roundpd
3216 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3217 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3218 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3219 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3221 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3222 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3223 Intrinsic IntId128> {
3224 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3226 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3227 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3228 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3230 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3233 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3236 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3237 int_x86_sse41_phminposuw>;
3239 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3240 let Constraints = "$src1 = $dst" in {
3241 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3242 Intrinsic IntId128, bit Commutable = 0> {
3243 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3244 (ins VR128:$src1, VR128:$src2),
3245 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3246 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3248 let isCommutable = Commutable;
3250 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3251 (ins VR128:$src1, i128mem:$src2),
3252 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3254 (IntId128 VR128:$src1,
3255 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3259 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3260 int_x86_sse41_pcmpeqq, 1>;
3261 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3262 int_x86_sse41_packusdw, 0>;
3263 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3264 int_x86_sse41_pminsb, 1>;
3265 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3266 int_x86_sse41_pminsd, 1>;
3267 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3268 int_x86_sse41_pminud, 1>;
3269 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3270 int_x86_sse41_pminuw, 1>;
3271 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3272 int_x86_sse41_pmaxsb, 1>;
3273 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3274 int_x86_sse41_pmaxsd, 1>;
3275 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3276 int_x86_sse41_pmaxud, 1>;
3277 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3278 int_x86_sse41_pmaxuw, 1>;
3280 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3282 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3283 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3284 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3285 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3287 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3288 let Constraints = "$src1 = $dst" in {
3289 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3290 SDNode OpNode, Intrinsic IntId128,
3291 bit Commutable = 0> {
3292 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3293 (ins VR128:$src1, VR128:$src2),
3294 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3295 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3296 VR128:$src2))]>, OpSize {
3297 let isCommutable = Commutable;
3299 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3300 (ins VR128:$src1, VR128:$src2),
3301 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3302 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3304 let isCommutable = Commutable;
3306 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3307 (ins VR128:$src1, i128mem:$src2),
3308 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3310 (OpVT (OpNode VR128:$src1, (memop addr:$src2))))]>, OpSize;
3311 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3312 (ins VR128:$src1, i128mem:$src2),
3313 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3315 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3320 /// SS48I_binop_rm - Simple SSE41 binary operator.
3321 let Constraints = "$src1 = $dst" in {
3322 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
3323 ValueType OpVT, bit Commutable = 0> {
3324 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3325 (ins VR128:$src1, VR128:$src2),
3326 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3327 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]>,
3329 let isCommutable = Commutable;
3331 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3332 (ins VR128:$src1, i128mem:$src2),
3333 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3334 [(set VR128:$dst, (OpNode VR128:$src1,
3335 (bc_v4i32 (memopv2i64 addr:$src2))))]>,
3340 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, 1>;
3342 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3343 let Constraints = "$src1 = $dst" in {
3344 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3345 Intrinsic IntId128, bit Commutable = 0> {
3346 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3347 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3348 !strconcat(OpcodeStr,
3349 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3351 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3353 let isCommutable = Commutable;
3355 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3356 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3357 !strconcat(OpcodeStr,
3358 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3360 (IntId128 VR128:$src1,
3361 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3366 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3367 int_x86_sse41_blendps, 0>;
3368 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3369 int_x86_sse41_blendpd, 0>;
3370 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3371 int_x86_sse41_pblendw, 0>;
3372 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3373 int_x86_sse41_dpps, 1>;
3374 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3375 int_x86_sse41_dppd, 1>;
3376 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3377 int_x86_sse41_mpsadbw, 0>;
3380 /// SS41I_ternary_int - SSE 4.1 ternary operator
3381 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3382 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3383 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3384 (ins VR128:$src1, VR128:$src2),
3385 !strconcat(OpcodeStr,
3386 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3387 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3390 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3391 (ins VR128:$src1, i128mem:$src2),
3392 !strconcat(OpcodeStr,
3393 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3396 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3400 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3401 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3402 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3405 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3406 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3407 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3408 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3410 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3411 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3413 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3417 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3418 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3419 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3420 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3421 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3422 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3424 // Common patterns involving scalar load.
3425 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3426 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3427 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3428 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3430 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3431 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3432 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3433 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3435 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3436 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3437 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3438 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3440 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3441 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3442 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3443 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3445 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3446 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3447 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3448 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3450 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3451 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3452 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3453 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3456 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3457 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3458 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3459 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3461 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3462 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3464 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3468 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3469 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3470 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3471 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3473 // Common patterns involving scalar load
3474 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3475 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3476 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3477 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3479 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3480 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3481 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3482 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3485 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3486 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3487 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3488 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3490 // Expecting a i16 load any extended to i32 value.
3491 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3492 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3493 [(set VR128:$dst, (IntId (bitconvert
3494 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3498 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3499 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3501 // Common patterns involving scalar load
3502 def : Pat<(int_x86_sse41_pmovsxbq
3503 (bitconvert (v4i32 (X86vzmovl
3504 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3505 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3507 def : Pat<(int_x86_sse41_pmovzxbq
3508 (bitconvert (v4i32 (X86vzmovl
3509 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3510 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3513 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3514 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3515 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3516 (ins VR128:$src1, i32i8imm:$src2),
3517 !strconcat(OpcodeStr,
3518 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3519 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3521 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3522 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3523 !strconcat(OpcodeStr,
3524 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3527 // There's an AssertZext in the way of writing the store pattern
3528 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3531 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3534 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3535 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3536 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3537 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3538 !strconcat(OpcodeStr,
3539 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3542 // There's an AssertZext in the way of writing the store pattern
3543 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3546 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3549 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3550 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3551 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3552 (ins VR128:$src1, i32i8imm:$src2),
3553 !strconcat(OpcodeStr,
3554 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3556 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3557 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3558 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3559 !strconcat(OpcodeStr,
3560 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3561 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3562 addr:$dst)]>, OpSize;
3565 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3568 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3570 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3571 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3572 (ins VR128:$src1, i32i8imm:$src2),
3573 !strconcat(OpcodeStr,
3574 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3576 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3578 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3579 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3580 !strconcat(OpcodeStr,
3581 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3582 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3583 addr:$dst)]>, OpSize;
3586 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3588 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3589 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3592 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3593 Requires<[HasSSE41]>;
3595 let Constraints = "$src1 = $dst" in {
3596 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3597 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3598 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3599 !strconcat(OpcodeStr,
3600 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3602 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3603 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3604 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3605 !strconcat(OpcodeStr,
3606 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3608 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3609 imm:$src3))]>, OpSize;
3613 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3615 let Constraints = "$src1 = $dst" in {
3616 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3617 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3618 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3619 !strconcat(OpcodeStr,
3620 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3622 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3624 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3625 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3626 !strconcat(OpcodeStr,
3627 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3629 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3630 imm:$src3)))]>, OpSize;
3634 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3636 // insertps has a few different modes, there's the first two here below which
3637 // are optimized inserts that won't zero arbitrary elements in the destination
3638 // vector. The next one matches the intrinsic and could zero arbitrary elements
3639 // in the target vector.
3640 let Constraints = "$src1 = $dst" in {
3641 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3642 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3643 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3644 !strconcat(OpcodeStr,
3645 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3647 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3649 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3650 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3651 !strconcat(OpcodeStr,
3652 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3654 (X86insrtps VR128:$src1,
3655 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3656 imm:$src3))]>, OpSize;
3660 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3662 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3663 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3665 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3666 // the intel intrinsic that corresponds to this.
3667 let Defs = [EFLAGS] in {
3668 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3669 "ptest \t{$src2, $src1|$src1, $src2}",
3670 [(set EFLAGS, (X86ptest VR128:$src1, VR128:$src2))]>,
3672 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3673 "ptest \t{$src2, $src1|$src1, $src2}",
3674 [(set EFLAGS, (X86ptest VR128:$src1, (load addr:$src2)))]>,
3678 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3679 "movntdqa\t{$src, $dst|$dst, $src}",
3680 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
3684 //===----------------------------------------------------------------------===//
3685 // SSE4.2 Instructions
3686 //===----------------------------------------------------------------------===//
3688 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3689 let Constraints = "$src1 = $dst" in {
3690 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3691 Intrinsic IntId128, bit Commutable = 0> {
3692 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3693 (ins VR128:$src1, VR128:$src2),
3694 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3695 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3697 let isCommutable = Commutable;
3699 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3700 (ins VR128:$src1, i128mem:$src2),
3701 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3703 (IntId128 VR128:$src1,
3704 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3708 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3710 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3711 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3712 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3713 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3715 // crc intrinsic instruction
3716 // This set of instructions are only rm, the only difference is the size
3718 let Constraints = "$src1 = $dst" in {
3719 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3720 (ins GR32:$src1, i8mem:$src2),
3721 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3723 (int_x86_sse42_crc32_8 GR32:$src1,
3724 (load addr:$src2)))]>;
3725 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3726 (ins GR32:$src1, GR8:$src2),
3727 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3729 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>;
3730 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3731 (ins GR32:$src1, i16mem:$src2),
3732 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3734 (int_x86_sse42_crc32_16 GR32:$src1,
3735 (load addr:$src2)))]>,
3737 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3738 (ins GR32:$src1, GR16:$src2),
3739 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3741 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3743 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3744 (ins GR32:$src1, i32mem:$src2),
3745 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3747 (int_x86_sse42_crc32_32 GR32:$src1,
3748 (load addr:$src2)))]>;
3749 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3750 (ins GR32:$src1, GR32:$src2),
3751 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3753 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>;
3754 def CRC64m8 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3755 (ins GR64:$src1, i8mem:$src2),
3756 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3758 (int_x86_sse42_crc64_8 GR64:$src1,
3759 (load addr:$src2)))]>,
3761 def CRC64r8 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3762 (ins GR64:$src1, GR8:$src2),
3763 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3765 (int_x86_sse42_crc64_8 GR64:$src1, GR8:$src2))]>,
3767 def CRC64m64 : SS42FI<0xF1, MRMSrcMem, (outs GR64:$dst),
3768 (ins GR64:$src1, i64mem:$src2),
3769 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3771 (int_x86_sse42_crc64_64 GR64:$src1,
3772 (load addr:$src2)))]>,
3774 def CRC64r64 : SS42FI<0xF1, MRMSrcReg, (outs GR64:$dst),
3775 (ins GR64:$src1, GR64:$src2),
3776 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3778 (int_x86_sse42_crc64_64 GR64:$src1, GR64:$src2))]>,
3782 // String/text processing instructions.
3783 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3784 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3785 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3786 "#PCMPISTRM128rr PSEUDO!",
3787 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3788 imm:$src3))]>, OpSize;
3789 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3790 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3791 "#PCMPISTRM128rm PSEUDO!",
3792 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3793 imm:$src3))]>, OpSize;
3796 let Defs = [XMM0, EFLAGS] in {
3797 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3798 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3799 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3800 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3801 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3802 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3805 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3806 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3807 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3808 "#PCMPESTRM128rr PSEUDO!",
3810 (int_x86_sse42_pcmpestrm128
3811 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3813 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3814 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3815 "#PCMPESTRM128rm PSEUDO!",
3816 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3817 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3821 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3822 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3823 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3824 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3825 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3826 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3827 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3830 let Defs = [ECX, EFLAGS] in {
3831 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3832 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3833 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3834 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3835 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3836 (implicit EFLAGS)]>, OpSize;
3837 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3838 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3839 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3840 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3841 (implicit EFLAGS)]>, OpSize;
3845 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3846 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3847 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3848 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3849 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3850 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3852 let Defs = [ECX, EFLAGS] in {
3853 let Uses = [EAX, EDX] in {
3854 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3855 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3856 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3857 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3858 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3859 (implicit EFLAGS)]>, OpSize;
3860 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3861 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3862 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3864 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
3865 (implicit EFLAGS)]>, OpSize;
3870 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3871 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3872 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3873 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3874 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3875 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;
3877 //===----------------------------------------------------------------------===//
3878 // AES-NI Instructions
3879 //===----------------------------------------------------------------------===//
3881 let Constraints = "$src1 = $dst" in {
3882 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr,
3883 Intrinsic IntId128, bit Commutable = 0> {
3884 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
3885 (ins VR128:$src1, VR128:$src2),
3886 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3887 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3889 let isCommutable = Commutable;
3891 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
3892 (ins VR128:$src1, i128mem:$src2),
3893 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3895 (IntId128 VR128:$src1,
3896 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3900 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
3901 int_x86_aesni_aesenc>;
3902 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
3903 int_x86_aesni_aesenclast>;
3904 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
3905 int_x86_aesni_aesdec>;
3906 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
3907 int_x86_aesni_aesdeclast>;
3909 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, VR128:$src2)),
3910 (AESENCrr VR128:$src1, VR128:$src2)>;
3911 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, (memop addr:$src2))),
3912 (AESENCrm VR128:$src1, addr:$src2)>;
3913 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, VR128:$src2)),
3914 (AESENCLASTrr VR128:$src1, VR128:$src2)>;
3915 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, (memop addr:$src2))),
3916 (AESENCLASTrm VR128:$src1, addr:$src2)>;
3917 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, VR128:$src2)),
3918 (AESDECrr VR128:$src1, VR128:$src2)>;
3919 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, (memop addr:$src2))),
3920 (AESDECrm VR128:$src1, addr:$src2)>;
3921 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, VR128:$src2)),
3922 (AESDECLASTrr VR128:$src1, VR128:$src2)>;
3923 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, (memop addr:$src2))),
3924 (AESDECLASTrm VR128:$src1, addr:$src2)>;
3926 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
3928 "aesimc\t{$src1, $dst|$dst, $src1}",
3930 (int_x86_aesni_aesimc VR128:$src1))]>,
3933 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
3934 (ins i128mem:$src1),
3935 "aesimc\t{$src1, $dst|$dst, $src1}",
3937 (int_x86_aesni_aesimc (bitconvert (memopv2i64 addr:$src1))))]>,
3940 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
3941 (ins VR128:$src1, i8imm:$src2),
3942 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3944 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
3946 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
3947 (ins i128mem:$src1, i8imm:$src2),
3948 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3950 (int_x86_aesni_aeskeygenassist (bitconvert (memopv2i64 addr:$src1)),