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:$cond,
370 //===----------------------------------------------------------------------===//
371 // SSE 1 & 2 Instructions Classes
372 //===----------------------------------------------------------------------===//
374 /// sse12_fp_scalar - SSE 1 & 2 scalar instructions class
375 multiclass sse12_fp_scalar<bits<8> opc, string OpcodeStr, SDNode OpNode,
376 RegisterClass RC, X86MemOperand x86memop> {
377 let isCommutable = 1 in {
378 def rr : SI<opc, MRMSrcReg, (outs RC:$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:$src1, x86memop:$src2),
382 OpcodeStr, [(set RC:$dst, (OpNode RC:$src1, (load addr:$src2)))]>;
385 /// sse12_fp_scalar_int - SSE 1 & 2 scalar instructions intrinsics class
386 multiclass sse12_fp_scalar_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
387 string asm, string SSEVer, string FPSizeStr,
388 Operand memopr, ComplexPattern mem_cpat> {
389 def rr_Int : SI<opc, MRMSrcReg, (outs RC:$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:$src1, memopr:$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:$src2), OpcodeStr, pat_rm, d>;
428 /// sse12_fp_packed_int - SSE 1 & 2 packed instructions intrinsics class
429 multiclass sse12_fp_packed_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
430 string asm, string SSEVer, string FPSizeStr,
431 X86MemOperand x86memop, PatFrag mem_frag,
433 def rr_Int : PI<opc, MRMSrcReg, (outs RC:$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:$src1, x86memop:$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:$src2)))], d>;
447 //===----------------------------------------------------------------------===//
448 // SSE 1 & 2 - Move Instructions
449 //===----------------------------------------------------------------------===//
451 class sse12_move_rr<RegisterClass RC, ValueType vt, string asm> :
452 SI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, RC:$src2), asm,
453 [(set (vt VR128:$dst), (movl VR128:$src1, (scalar_to_vector RC:$src2)))]>;
455 // Loading from memory automatically zeroing upper bits.
456 class sse12_move_rm<RegisterClass RC, X86MemOperand x86memop,
457 PatFrag mem_pat, string OpcodeStr> :
458 SI<0x10, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
459 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
460 [(set RC:$dst, (mem_pat addr:$src))]>;
462 // Move Instructions. Register-to-register movss/movsd is not used for FR32/64
463 // register copies because it's a partial register update; FsMOVAPSrr/FsMOVAPDrr
464 // is used instead. Register-to-register movss/movsd is not modeled as an
465 // INSERT_SUBREG because INSERT_SUBREG requires that the insert be implementable
466 // in terms of a copy, and just mentioned, we don't use movss/movsd for copies.
467 let isAsmParserOnly = 1 in {
468 def VMOVSSrr : sse12_move_rr<FR32, v4f32,
469 "movss\t{$src2, $src1, $dst|$dst, $src1, $src2}">, XS, VEX_4V;
470 def VMOVSDrr : sse12_move_rr<FR64, v2f64,
471 "movsd\t{$src2, $src1, $dst|$dst, $src1, $src2}">, XD, VEX_4V;
473 let canFoldAsLoad = 1, isReMaterializable = 1 in {
474 def VMOVSSrm : sse12_move_rm<FR32, f32mem, loadf32, "movss">, XS, VEX;
476 let AddedComplexity = 20 in
477 def VMOVSDrm : sse12_move_rm<FR64, f64mem, loadf64, "movsd">, XD, VEX;
481 let Constraints = "$src1 = $dst" in {
482 def MOVSSrr : sse12_move_rr<FR32, v4f32,
483 "movss\t{$src2, $dst|$dst, $src2}">, XS;
484 def MOVSDrr : sse12_move_rr<FR64, v2f64,
485 "movsd\t{$src2, $dst|$dst, $src2}">, XD;
488 let canFoldAsLoad = 1, isReMaterializable = 1 in {
489 def MOVSSrm : sse12_move_rm<FR32, f32mem, loadf32, "movss">, XS;
491 let AddedComplexity = 20 in
492 def MOVSDrm : sse12_move_rm<FR64, f64mem, loadf64, "movsd">, XD;
495 let AddedComplexity = 15 in {
496 // Extract the low 32-bit value from one vector and insert it into another.
497 def : Pat<(v4f32 (movl VR128:$src1, VR128:$src2)),
498 (MOVSSrr (v4f32 VR128:$src1),
499 (EXTRACT_SUBREG (v4f32 VR128:$src2), sub_ss))>;
500 // Extract the low 64-bit value from one vector and insert it into another.
501 def : Pat<(v2f64 (movl VR128:$src1, VR128:$src2)),
502 (MOVSDrr (v2f64 VR128:$src1),
503 (EXTRACT_SUBREG (v2f64 VR128:$src2), sub_sd))>;
506 // Implicitly promote a 32-bit scalar to a vector.
507 def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
508 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), FR32:$src, sub_ss)>;
509 // Implicitly promote a 64-bit scalar to a vector.
510 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
511 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), FR64:$src, sub_sd)>;
513 let AddedComplexity = 20 in {
514 // MOVSSrm zeros the high parts of the register; represent this
515 // with SUBREG_TO_REG.
516 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
517 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
518 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
519 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
520 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
521 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
522 // MOVSDrm zeros the high parts of the register; represent this
523 // with SUBREG_TO_REG.
524 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
525 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
526 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
527 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
528 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
529 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
530 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
531 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
532 def : Pat<(v2f64 (X86vzload addr:$src)),
533 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
536 // Store scalar value to memory.
537 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
538 "movss\t{$src, $dst|$dst, $src}",
539 [(store FR32:$src, addr:$dst)]>;
540 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
541 "movsd\t{$src, $dst|$dst, $src}",
542 [(store FR64:$src, addr:$dst)]>;
544 let isAsmParserOnly = 1 in {
545 def VMOVSSmr : SI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
546 "movss\t{$src, $dst|$dst, $src}",
547 [(store FR32:$src, addr:$dst)]>, XS, VEX_4V;
548 def VMOVSDmr : SI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
549 "movsd\t{$src, $dst|$dst, $src}",
550 [(store FR64:$src, addr:$dst)]>, XD, VEX_4V;
553 // Extract and store.
554 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
557 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
558 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
561 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
563 // Move Aligned/Unaligned floating point values
564 multiclass sse12_mov_packed<bits<8> opc, RegisterClass RC,
565 X86MemOperand x86memop, PatFrag ld_frag,
566 string asm, Domain d,
567 bit IsReMaterializable = 1> {
568 let neverHasSideEffects = 1 in
569 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src), asm, [], d>;
570 let canFoldAsLoad = 1, isReMaterializable = IsReMaterializable in
571 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src), asm,
572 [(set RC:$dst, (ld_frag addr:$src))], d>;
575 defm MOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
576 "movaps\t{$src, $dst|$dst, $src}",
577 SSEPackedSingle>, TB;
578 defm MOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
579 "movapd\t{$src, $dst|$dst, $src}",
580 SSEPackedDouble>, TB, OpSize;
581 defm MOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
582 "movups\t{$src, $dst|$dst, $src}",
583 SSEPackedSingle>, TB;
584 defm MOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
585 "movupd\t{$src, $dst|$dst, $src}",
586 SSEPackedDouble, 0>, TB, OpSize;
588 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
589 "movaps\t{$src, $dst|$dst, $src}",
590 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
591 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
592 "movapd\t{$src, $dst|$dst, $src}",
593 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
594 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
595 "movups\t{$src, $dst|$dst, $src}",
596 [(store (v4f32 VR128:$src), addr:$dst)]>;
597 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
598 "movupd\t{$src, $dst|$dst, $src}",
599 [(store (v2f64 VR128:$src), addr:$dst)]>;
601 // Intrinsic forms of MOVUPS/D load and store
602 let canFoldAsLoad = 1, isReMaterializable = 1 in
603 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
604 "movups\t{$src, $dst|$dst, $src}",
605 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
606 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
607 "movupd\t{$src, $dst|$dst, $src}",
608 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
610 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
611 "movups\t{$src, $dst|$dst, $src}",
612 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
613 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
614 "movupd\t{$src, $dst|$dst, $src}",
615 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
617 // Move Low/High packed floating point values
618 multiclass sse12_mov_hilo_packed<bits<8>opc, RegisterClass RC,
619 PatFrag mov_frag, string base_opc,
621 def PSrm : PI<opc, MRMSrcMem,
622 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
623 !strconcat(!strconcat(base_opc,"s"), asm_opr),
626 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))],
627 SSEPackedSingle>, TB;
629 def PDrm : PI<opc, MRMSrcMem,
630 (outs RC:$dst), (ins RC:$src1, f64mem:$src2),
631 !strconcat(!strconcat(base_opc,"d"), asm_opr),
632 [(set RC:$dst, (v2f64 (mov_frag RC:$src1,
633 (scalar_to_vector (loadf64 addr:$src2)))))],
634 SSEPackedDouble>, TB, OpSize;
637 let Constraints = "$src1 = $dst", AddedComplexity = 20 in {
638 defm MOVL : sse12_mov_hilo_packed<0x12, VR128, movlp, "movlp",
639 "\t{$src2, $dst|$dst, $src2}">;
640 defm MOVH : sse12_mov_hilo_packed<0x16, VR128, movlhps, "movhp",
641 "\t{$src2, $dst|$dst, $src2}">;
644 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
645 "movlps\t{$src, $dst|$dst, $src}",
646 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
647 (iPTR 0))), addr:$dst)]>;
648 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
649 "movlpd\t{$src, $dst|$dst, $src}",
650 [(store (f64 (vector_extract (v2f64 VR128:$src),
651 (iPTR 0))), addr:$dst)]>;
653 // v2f64 extract element 1 is always custom lowered to unpack high to low
654 // and extract element 0 so the non-store version isn't too horrible.
655 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
656 "movhps\t{$src, $dst|$dst, $src}",
657 [(store (f64 (vector_extract
658 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
659 (undef)), (iPTR 0))), addr:$dst)]>;
660 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
661 "movhpd\t{$src, $dst|$dst, $src}",
662 [(store (f64 (vector_extract
663 (v2f64 (unpckh VR128:$src, (undef))),
664 (iPTR 0))), addr:$dst)]>;
666 let Constraints = "$src1 = $dst", AddedComplexity = 20 in {
667 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
668 (ins VR128:$src1, VR128:$src2),
669 "movlhps\t{$src2, $dst|$dst, $src2}",
671 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
672 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
673 (ins VR128:$src1, VR128:$src2),
674 "movhlps\t{$src2, $dst|$dst, $src2}",
676 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
679 def : Pat<(movlhps VR128:$src1, (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
680 (MOVHPSrm (v4i32 VR128:$src1), addr:$src2)>;
681 let AddedComplexity = 20 in {
682 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
683 (MOVLHPSrr (v4f32 VR128:$src), (v4f32 VR128:$src))>;
684 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
685 (MOVLHPSrr (v2i64 VR128:$src), (v2i64 VR128:$src))>;
688 //===----------------------------------------------------------------------===//
689 // SSE 1 & 2 - Conversion Instructions
690 //===----------------------------------------------------------------------===//
692 multiclass sse12_cvt_s<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
693 SDNode OpNode, X86MemOperand x86memop, PatFrag ld_frag,
695 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
696 [(set DstRC:$dst, (OpNode SrcRC:$src))]>;
697 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
698 [(set DstRC:$dst, (OpNode (ld_frag addr:$src)))]>;
701 multiclass sse12_cvt_p<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
702 SDNode OpNode, X86MemOperand x86memop, PatFrag ld_frag,
703 string asm, Domain d> {
704 def rr : PI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
705 [(set DstRC:$dst, (OpNode SrcRC:$src))], d>;
706 def rm : PI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
707 [(set DstRC:$dst, (OpNode (ld_frag addr:$src)))], d>;
710 multiclass sse12_vcvt_avx<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
711 SDNode OpNode, X86MemOperand x86memop, PatFrag ld_frag,
713 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src),
715 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
716 (ins DstRC:$src1, x86memop:$src), asm, []>;
719 let isAsmParserOnly = 1 in {
720 defm VCVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
721 "cvttss2si\t{$src, $dst|$dst, $src}">, XS, VEX;
722 defm VCVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
723 "cvttsd2si\t{$src, $dst|$dst, $src}">, XD, VEX;
724 defm VCVTSI2SS : sse12_vcvt_avx<0x2A, GR32, FR32, sint_to_fp, i32mem, loadi32,
725 "cvtsi2ss\t{$src, $src1, $dst|$dst, $src1, $src}">, XS,
727 defm VCVTSI2SD : sse12_vcvt_avx<0x2A, GR32, FR64, sint_to_fp, i32mem, loadi32,
728 "cvtsi2sd\t{$src, $src1, $dst|$dst, $src1, $src}">, XD,
732 defm CVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
733 "cvttss2si\t{$src, $dst|$dst, $src}">, XS;
734 defm CVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
735 "cvttsd2si\t{$src, $dst|$dst, $src}">, XD;
736 defm CVTSI2SS : sse12_cvt_s<0x2A, GR32, FR32, sint_to_fp, i32mem, loadi32,
737 "cvtsi2ss\t{$src, $dst|$dst, $src}">, XS;
738 defm CVTSI2SD : sse12_cvt_s<0x2A, GR32, FR64, sint_to_fp, i32mem, loadi32,
739 "cvtsi2sd\t{$src, $dst|$dst, $src}">, XD;
741 // Conversion Instructions Intrinsics - Match intrinsics which expect MM
742 // and/or XMM operand(s).
743 multiclass sse12_cvt_pint<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
744 Intrinsic Int, X86MemOperand x86memop, PatFrag ld_frag,
745 string asm, Domain d> {
746 def rr : PI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
747 [(set DstRC:$dst, (Int SrcRC:$src))], d>;
748 def rm : PI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
749 [(set DstRC:$dst, (Int (ld_frag addr:$src)))], d>;
752 multiclass sse12_cvt_sint<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
753 Intrinsic Int, X86MemOperand x86memop, PatFrag ld_frag,
755 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
756 [(set DstRC:$dst, (Int SrcRC:$src))]>;
757 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
758 [(set DstRC:$dst, (Int (ld_frag addr:$src)))]>;
761 multiclass sse12_cvt_pint_3addr<bits<8> opc, RegisterClass SrcRC,
762 RegisterClass DstRC, Intrinsic Int, X86MemOperand x86memop,
763 PatFrag ld_frag, string asm, Domain d> {
764 def rr : PI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src2),
765 asm, [(set DstRC:$dst, (Int DstRC:$src1, SrcRC:$src2))], d>;
766 def rm : PI<opc, MRMSrcMem, (outs DstRC:$dst),
767 (ins DstRC:$src1, x86memop:$src2), asm,
768 [(set DstRC:$dst, (Int DstRC:$src1, (ld_frag addr:$src2)))], d>;
771 multiclass sse12_cvt_sint_3addr<bits<8> opc, RegisterClass SrcRC,
772 RegisterClass DstRC, Intrinsic Int, X86MemOperand x86memop,
773 PatFrag ld_frag, string asm> {
774 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src2),
775 asm, [(set DstRC:$dst, (Int DstRC:$src1, SrcRC:$src2))]>;
776 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
777 (ins DstRC:$src1, x86memop:$src2), asm,
778 [(set DstRC:$dst, (Int DstRC:$src1, (ld_frag addr:$src2)))]>;
781 let isAsmParserOnly = 1 in {
782 defm Int_VCVTSS2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse_cvtss2si,
783 f32mem, load, "cvtss2si\t{$src, $dst|$dst, $src}">, XS,
785 defm Int_VCVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse2_cvtsd2si,
786 f128mem, load, "cvtsd2si\t{$src, $dst|$dst, $src}">, XD,
789 defm Int_CVTSS2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse_cvtss2si,
790 f32mem, load, "cvtss2si\t{$src, $dst|$dst, $src}">, XS;
791 defm Int_CVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse2_cvtsd2si,
792 f128mem, load, "cvtsd2si\t{$src, $dst|$dst, $src}">, XD;
795 let Constraints = "$src1 = $dst" in {
796 defm Int_CVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
797 int_x86_sse_cvtsi2ss, i32mem, loadi32,
798 "cvtsi2ss\t{$src2, $dst|$dst, $src2}">, XS;
799 defm Int_CVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
800 int_x86_sse2_cvtsi2sd, i32mem, loadi32,
801 "cvtsi2ss\t{$src2, $dst|$dst, $src2}">, XD;
804 // Instructions below don't have an AVX form.
805 defm Int_CVTPS2PI : sse12_cvt_pint<0x2D, VR128, VR64, int_x86_sse_cvtps2pi,
806 f64mem, load, "cvtps2pi\t{$src, $dst|$dst, $src}",
807 SSEPackedSingle>, TB;
808 defm Int_CVTPD2PI : sse12_cvt_pint<0x2D, VR128, VR64, int_x86_sse_cvtpd2pi,
809 f128mem, memop, "cvtpd2pi\t{$src, $dst|$dst, $src}",
810 SSEPackedDouble>, TB, OpSize;
811 defm Int_CVTTPS2PI : sse12_cvt_pint<0x2C, VR128, VR64, int_x86_sse_cvttps2pi,
812 f64mem, load, "cvttps2pi\t{$src, $dst|$dst, $src}",
813 SSEPackedSingle>, TB;
814 defm Int_CVTTPD2PI : sse12_cvt_pint<0x2C, VR128, VR64, int_x86_sse_cvttpd2pi,
815 f128mem, memop, "cvttpd2pi\t{$src, $dst|$dst, $src}",
816 SSEPackedDouble>, TB, OpSize;
817 defm Int_CVTPI2PD : sse12_cvt_pint<0x2A, VR64, VR128, int_x86_sse_cvtpi2pd,
818 i64mem, load, "cvtpi2pd\t{$src, $dst|$dst, $src}",
819 SSEPackedDouble>, TB, OpSize;
820 let Constraints = "$src1 = $dst" in {
821 defm Int_CVTPI2PS : sse12_cvt_pint_3addr<0x2A, VR64, VR128,
822 int_x86_sse_cvtpi2ps,
823 i64mem, load, "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
824 SSEPackedSingle>, TB;
829 // Aliases for intrinsics
830 defm Int_CVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
831 f32mem, load, "cvttss2si\t{$src, $dst|$dst, $src}">,
833 defm Int_CVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
834 f128mem, load, "cvttss2si\t{$src, $dst|$dst, $src}">,
837 let Pattern = []<dag> in {
838 defm CVTSS2SI : sse12_cvt_s<0x2D, FR32, GR32, undef, f32mem, load /*dummy*/,
839 "cvtss2si{l}\t{$src, $dst|$dst, $src}">, XS;
840 defm CVTDQ2PS : sse12_cvt_p<0x5B, VR128, VR128, undef, f128mem, load /*dummy*/,
841 "cvtdq2ps\t{$src, $dst|$dst, $src}",
842 SSEPackedSingle>, TB; /* PD SSE3 form is avaiable */
845 //===----------------------------------------------------------------------===//
846 // SSE 1 & 2 - Compare Instructions
847 //===----------------------------------------------------------------------===//
849 // sse12_cmp_scalar - sse 1 & 2 compare scalar instructions
850 multiclass sse12_cmp_scalar<RegisterClass RC, X86MemOperand x86memop,
851 string asm, string asm_alt> {
852 def rr : SIi8<0xC2, MRMSrcReg,
853 (outs RC:$dst), (ins RC:$src1, RC:$src, SSECC:$cc),
856 def rm : SIi8<0xC2, MRMSrcMem,
857 (outs RC:$dst), (ins RC:$src1, x86memop:$src, SSECC:$cc),
859 // Accept explicit immediate argument form instead of comparison code.
860 let isAsmParserOnly = 1 in {
861 def rr_alt : SIi8<0xC2, MRMSrcReg,
862 (outs RC:$dst), (ins RC:$src1, RC:$src, i8imm:$src2),
865 def rm_alt : SIi8<0xC2, MRMSrcMem,
866 (outs RC:$dst), (ins RC:$src1, x86memop:$src, i8imm:$src2),
871 let neverHasSideEffects = 1, isAsmParserOnly = 1 in {
872 defm VCMPSS : sse12_cmp_scalar<FR32, f32mem,
873 "cmp${cc}ss\t{$src, $src1, $dst|$dst, $src1, $src}",
874 "cmpss\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}">,
876 defm VCMPSD : sse12_cmp_scalar<FR64, f64mem,
877 "cmp${cc}sd\t{$src, $src1, $dst|$dst, $src1, $src}",
878 "cmpsd\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}">,
882 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
883 defm CMPSS : sse12_cmp_scalar<FR32, f32mem,
884 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
885 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}">, XS;
886 defm CMPSD : sse12_cmp_scalar<FR64, f64mem,
887 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
888 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}">, XD;
891 multiclass sse12_cmp_scalar_int<RegisterClass RC, X86MemOperand x86memop,
892 Intrinsic Int, string asm> {
893 def rr : SIi8<0xC2, MRMSrcReg, (outs VR128:$dst),
894 (ins VR128:$src1, VR128:$src, SSECC:$cc), asm,
895 [(set VR128:$dst, (Int VR128:$src1,
896 VR128:$src, imm:$cc))]>;
897 def rm : SIi8<0xC2, MRMSrcMem, (outs VR128:$dst),
898 (ins VR128:$src1, f32mem:$src, SSECC:$cc), asm,
899 [(set VR128:$dst, (Int VR128:$src1,
900 (load addr:$src), imm:$cc))]>;
903 // Aliases to match intrinsics which expect XMM operand(s).
904 let isAsmParserOnly = 1 in {
905 defm Int_VCMPSS : sse12_cmp_scalar_int<VR128, f32mem, int_x86_sse_cmp_ss,
906 "cmp${cc}ss\t{$src, $src1, $dst|$dst, $src1, $src}">,
908 defm Int_VCMPSD : sse12_cmp_scalar_int<VR128, f64mem, int_x86_sse2_cmp_sd,
909 "cmp${cc}sd\t{$src, $src1, $dst|$dst, $src1, $src}">,
912 let Constraints = "$src1 = $dst" in {
913 defm Int_CMPSS : sse12_cmp_scalar_int<VR128, f32mem, int_x86_sse_cmp_ss,
914 "cmp${cc}ss\t{$src, $dst|$dst, $src}">, XS;
915 defm Int_CMPSD : sse12_cmp_scalar_int<VR128, f64mem, int_x86_sse2_cmp_sd,
916 "cmp${cc}sd\t{$src, $dst|$dst, $src}">, XD;
920 // sse12_ord_cmp - Unordered/Ordered scalar fp compare and set EFLAGS
921 multiclass sse12_ord_cmp<bits<8> opc, RegisterClass RC, SDNode OpNode,
922 ValueType vt, X86MemOperand x86memop,
923 PatFrag ld_frag, string OpcodeStr, Domain d> {
924 def rr: PI<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
925 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
926 [(set EFLAGS, (OpNode (vt RC:$src1), RC:$src2))], d>;
927 def rm: PI<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
928 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
929 [(set EFLAGS, (OpNode (vt RC:$src1),
930 (ld_frag addr:$src2)))], d>;
933 let Defs = [EFLAGS] in {
934 let isAsmParserOnly = 1 in {
935 defm VUCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
936 "ucomiss", SSEPackedSingle>, VEX;
937 defm VUCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
938 "ucomisd", SSEPackedDouble>, OpSize, VEX;
939 let Pattern = []<dag> in {
940 defm VCOMISS : sse12_ord_cmp<0x2F, VR128, undef, v4f32, f128mem, load,
941 "comiss", SSEPackedSingle>, VEX;
942 defm VCOMISD : sse12_ord_cmp<0x2F, VR128, undef, v2f64, f128mem, load,
943 "comisd", SSEPackedDouble>, OpSize, VEX;
946 defm Int_VUCOMISS : sse12_ord_cmp<0x2E, VR128, X86ucomi, v4f32, f128mem,
947 load, "ucomiss", SSEPackedSingle>, VEX;
948 defm Int_VUCOMISD : sse12_ord_cmp<0x2E, VR128, X86ucomi, v2f64, f128mem,
949 load, "ucomisd", SSEPackedDouble>, OpSize, VEX;
951 defm Int_VCOMISS : sse12_ord_cmp<0x2F, VR128, X86comi, v4f32, f128mem,
952 load, "comiss", SSEPackedSingle>, VEX;
953 defm Int_VCOMISD : sse12_ord_cmp<0x2F, VR128, X86comi, v2f64, f128mem,
954 load, "comisd", SSEPackedDouble>, OpSize, VEX;
956 defm UCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
957 "ucomiss", SSEPackedSingle>, TB;
958 defm UCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
959 "ucomisd", SSEPackedDouble>, TB, OpSize;
961 let Pattern = []<dag> in {
962 defm COMISS : sse12_ord_cmp<0x2F, VR128, undef, v4f32, f128mem, load,
963 "comiss", SSEPackedSingle>, TB;
964 defm COMISD : sse12_ord_cmp<0x2F, VR128, undef, v2f64, f128mem, load,
965 "comisd", SSEPackedDouble>, TB, OpSize;
968 defm Int_UCOMISS : sse12_ord_cmp<0x2E, VR128, X86ucomi, v4f32, f128mem,
969 load, "ucomiss", SSEPackedSingle>, TB;
970 defm Int_UCOMISD : sse12_ord_cmp<0x2E, VR128, X86ucomi, v2f64, f128mem,
971 load, "ucomisd", SSEPackedDouble>, TB, OpSize;
973 defm Int_COMISS : sse12_ord_cmp<0x2F, VR128, X86comi, v4f32, f128mem, load,
974 "comiss", SSEPackedSingle>, TB;
975 defm Int_COMISD : sse12_ord_cmp<0x2F, VR128, X86comi, v2f64, f128mem, load,
976 "comisd", SSEPackedDouble>, TB, OpSize;
979 // sse12_cmp_packed - sse 1 & 2 compared packed instructions
980 multiclass sse12_cmp_packed<RegisterClass RC, X86MemOperand x86memop,
981 Intrinsic Int, string asm, string asm_alt,
983 def rri : PIi8<0xC2, MRMSrcReg,
984 (outs RC:$dst), (ins RC:$src1, RC:$src, SSECC:$cc), asm,
985 [(set RC:$dst, (Int RC:$src1, RC:$src, imm:$cc))], d>;
986 def rmi : PIi8<0xC2, MRMSrcMem,
987 (outs RC:$dst), (ins RC:$src1, f128mem:$src, SSECC:$cc), asm,
988 [(set RC:$dst, (Int RC:$src1, (memop addr:$src), imm:$cc))], d>;
989 // Accept explicit immediate argument form instead of comparison code.
990 let isAsmParserOnly = 1 in {
991 def rri_alt : PIi8<0xC2, MRMSrcReg,
992 (outs RC:$dst), (ins RC:$src1, RC:$src, i8imm:$src2),
994 def rmi_alt : PIi8<0xC2, MRMSrcMem,
995 (outs RC:$dst), (ins RC:$src1, f128mem:$src, i8imm:$src2),
1000 let isAsmParserOnly = 1 in {
1001 defm VCMPPS : sse12_cmp_packed<VR128, f128mem, int_x86_sse_cmp_ps,
1002 "cmp${cc}ps\t{$src, $src1, $dst|$dst, $src1, $src}",
1003 "cmpps\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}",
1004 SSEPackedSingle>, VEX_4V;
1005 defm VCMPPD : sse12_cmp_packed<VR128, f128mem, int_x86_sse2_cmp_pd,
1006 "cmp${cc}pd\t{$src, $src1, $dst|$dst, $src1, $src}",
1007 "cmppd\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}",
1008 SSEPackedDouble>, OpSize, VEX_4V;
1010 let Constraints = "$src1 = $dst" in {
1011 defm CMPPS : sse12_cmp_packed<VR128, f128mem, int_x86_sse_cmp_ps,
1012 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1013 "cmpps\t{$src2, $src, $dst|$dst, $src, $src2}",
1014 SSEPackedSingle>, TB;
1015 defm CMPPD : sse12_cmp_packed<VR128, f128mem, int_x86_sse2_cmp_pd,
1016 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1017 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}",
1018 SSEPackedDouble>, TB, OpSize;
1021 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1022 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
1023 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1024 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
1025 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1026 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1027 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1028 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1030 //===----------------------------------------------------------------------===//
1031 // SSE 1 & 2 - Shuffle Instructions
1032 //===----------------------------------------------------------------------===//
1034 /// sse12_shuffle - sse 1 & 2 shuffle instructions
1035 multiclass sse12_shuffle<RegisterClass RC, X86MemOperand x86memop,
1036 ValueType vt, string asm, PatFrag mem_frag,
1037 Domain d, bit IsConvertibleToThreeAddress = 0> {
1038 def rmi : PIi8<0xC6, MRMSrcMem, (outs VR128:$dst),
1039 (ins VR128:$src1, f128mem:$src2, i8imm:$src3), asm,
1040 [(set VR128:$dst, (vt (shufp:$src3
1041 VR128:$src1, (mem_frag addr:$src2))))], d>;
1042 let isConvertibleToThreeAddress = IsConvertibleToThreeAddress in
1043 def rri : PIi8<0xC6, MRMSrcReg, (outs VR128:$dst),
1044 (ins VR128:$src1, VR128:$src2, i8imm:$src3), asm,
1046 (vt (shufp:$src3 VR128:$src1, VR128:$src2)))], d>;
1049 let isAsmParserOnly = 1 in {
1050 defm VSHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
1051 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
1052 memopv4f32, SSEPackedSingle>, VEX_4V;
1053 defm VSHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
1054 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src2, $src2, $src3}",
1055 memopv2f64, SSEPackedDouble>, OpSize, VEX_4V;
1058 let Constraints = "$src1 = $dst" in {
1059 defm SHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
1060 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1061 memopv4f32, SSEPackedSingle, 1 /* cvt to pshufd */>,
1063 defm SHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
1064 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1065 memopv2f64, SSEPackedDouble>, TB, OpSize;
1068 //===----------------------------------------------------------------------===//
1069 // SSE 1 & 2 - Unpack Instructions
1070 //===----------------------------------------------------------------------===//
1072 /// sse12_unpack_interleave - sse 1 & 2 unpack and interleave
1073 multiclass sse12_unpack_interleave<bits<8> opc, PatFrag OpNode, ValueType vt,
1074 PatFrag mem_frag, RegisterClass RC,
1075 X86MemOperand x86memop, string asm,
1077 def rr : PI<opc, MRMSrcReg,
1078 (outs RC:$dst), (ins RC:$src1, RC:$src2),
1080 (vt (OpNode RC:$src1, RC:$src2)))], d>;
1081 def rm : PI<opc, MRMSrcMem,
1082 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
1084 (vt (OpNode RC:$src1,
1085 (mem_frag addr:$src2))))], d>;
1088 let AddedComplexity = 10 in {
1089 let isAsmParserOnly = 1 in {
1090 defm VUNPCKHPS: sse12_unpack_interleave<0x15, unpckh, v4f32, memopv4f32,
1091 VR128, f128mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1092 SSEPackedSingle>, VEX_4V;
1093 defm VUNPCKHPD: sse12_unpack_interleave<0x15, unpckh, v2f64, memopv2f64,
1094 VR128, f128mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1095 SSEPackedDouble>, OpSize, VEX_4V;
1096 defm VUNPCKLPS: sse12_unpack_interleave<0x14, unpckl, v4f32, memopv4f32,
1097 VR128, f128mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1098 SSEPackedSingle>, VEX_4V;
1099 defm VUNPCKLPD: sse12_unpack_interleave<0x14, unpckl, v2f64, memopv2f64,
1100 VR128, f128mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1101 SSEPackedDouble>, OpSize, VEX_4V;
1104 let Constraints = "$src1 = $dst" in {
1105 defm UNPCKHPS: sse12_unpack_interleave<0x15, unpckh, v4f32, memopv4f32,
1106 VR128, f128mem, "unpckhps\t{$src2, $dst|$dst, $src2}",
1107 SSEPackedSingle>, TB;
1108 defm UNPCKHPD: sse12_unpack_interleave<0x15, unpckh, v2f64, memopv2f64,
1109 VR128, f128mem, "unpckhpd\t{$src2, $dst|$dst, $src2}",
1110 SSEPackedDouble>, TB, OpSize;
1111 defm UNPCKLPS: sse12_unpack_interleave<0x14, unpckl, v4f32, memopv4f32,
1112 VR128, f128mem, "unpcklps\t{$src2, $dst|$dst, $src2}",
1113 SSEPackedSingle>, TB;
1114 defm UNPCKLPD: sse12_unpack_interleave<0x14, unpckl, v2f64, memopv2f64,
1115 VR128, f128mem, "unpcklpd\t{$src2, $dst|$dst, $src2}",
1116 SSEPackedDouble>, TB, OpSize;
1117 } // Constraints = "$src1 = $dst"
1118 } // AddedComplexity
1120 //===----------------------------------------------------------------------===//
1121 // SSE 1 & 2 - Extract Floating-Point Sign mask
1122 //===----------------------------------------------------------------------===//
1124 /// sse12_extr_sign_mask - sse 1 & 2 unpack and interleave
1125 multiclass sse12_extr_sign_mask<RegisterClass RC, Intrinsic Int, string asm,
1127 def rr : PI<0x50, MRMSrcReg, (outs GR32:$dst), (ins RC:$src),
1128 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1129 [(set GR32:$dst, (Int RC:$src))], d>;
1133 defm MOVMSKPS : sse12_extr_sign_mask<VR128, int_x86_sse_movmsk_ps, "movmskps",
1134 SSEPackedSingle>, TB;
1135 defm MOVMSKPD : sse12_extr_sign_mask<VR128, int_x86_sse2_movmsk_pd, "movmskpd",
1136 SSEPackedDouble>, TB, OpSize;
1138 let isAsmParserOnly = 1 in {
1139 defm VMOVMSKPS : sse12_extr_sign_mask<VR128, int_x86_sse_movmsk_ps,
1140 "movmskps", SSEPackedSingle>, VEX;
1141 defm VMOVMSKPD : sse12_extr_sign_mask<VR128, int_x86_sse2_movmsk_pd,
1142 "movmskpd", SSEPackedDouble>, OpSize,
1146 //===----------------------------------------------------------------------===//
1147 // SSE 1 & 2 - Misc aliasing of packed SSE 1 & 2 instructions
1148 //===----------------------------------------------------------------------===//
1150 // Aliases of packed SSE1 & SSE2 instructions for scalar use. These all have
1151 // names that start with 'Fs'.
1153 // Alias instructions that map fld0 to pxor for sse.
1154 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1155 canFoldAsLoad = 1 in {
1156 // FIXME: Set encoding to pseudo!
1157 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
1158 [(set FR32:$dst, fp32imm0)]>,
1159 Requires<[HasSSE1]>, TB, OpSize;
1160 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1161 [(set FR64:$dst, fpimm0)]>,
1162 Requires<[HasSSE2]>, TB, OpSize;
1165 // Alias instruction to do FR32 or FR64 reg-to-reg copy using movaps. Upper
1166 // bits are disregarded.
1167 let neverHasSideEffects = 1 in {
1168 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
1169 "movaps\t{$src, $dst|$dst, $src}", []>;
1170 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1171 "movapd\t{$src, $dst|$dst, $src}", []>;
1174 // Alias instruction to load FR32 or FR64 from f128mem using movaps. Upper
1175 // bits are disregarded.
1176 let canFoldAsLoad = 1, isReMaterializable = 1 in {
1177 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
1178 "movaps\t{$src, $dst|$dst, $src}",
1179 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
1180 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1181 "movapd\t{$src, $dst|$dst, $src}",
1182 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1185 //===----------------------------------------------------------------------===//
1186 // SSE 1 & 2 - Logical Instructions
1187 //===----------------------------------------------------------------------===//
1189 /// sse12_fp_alias_pack_logical - SSE 1 & 2 aliased packed FP logical ops
1191 multiclass sse12_fp_alias_pack_logical<bits<8> opc, string OpcodeStr,
1192 SDNode OpNode, bit MayLoad = 0> {
1193 let isAsmParserOnly = 1 in {
1194 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1195 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode, FR32,
1196 f32, f128mem, memopfsf32, SSEPackedSingle, MayLoad>, VEX_4V;
1198 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1199 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode, FR64,
1200 f64, f128mem, memopfsf64, SSEPackedDouble, MayLoad>, OpSize,
1204 let Constraints = "$src1 = $dst" in {
1205 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1206 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, FR32, f32,
1207 f128mem, memopfsf32, SSEPackedSingle, MayLoad>, TB;
1209 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1210 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, FR64, f64,
1211 f128mem, memopfsf64, SSEPackedDouble, MayLoad>, TB, OpSize;
1215 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1216 defm FsAND : sse12_fp_alias_pack_logical<0x54, "and", X86fand>;
1217 defm FsOR : sse12_fp_alias_pack_logical<0x56, "or", X86for>;
1218 defm FsXOR : sse12_fp_alias_pack_logical<0x57, "xor", X86fxor>;
1220 let neverHasSideEffects = 1, Pattern = []<dag>, isCommutable = 0 in
1221 defm FsANDN : sse12_fp_alias_pack_logical<0x55, "andn", undef, 1>;
1223 /// sse12_fp_packed_logical - SSE 1 & 2 packed FP logical ops
1225 multiclass sse12_fp_packed_logical<bits<8> opc, string OpcodeStr,
1226 SDNode OpNode, int HasPat = 0,
1227 list<list<dag>> Pattern = []> {
1228 let isAsmParserOnly = 1 in {
1229 defm V#NAME#PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
1230 !strconcat(OpcodeStr, "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1232 !if(HasPat, Pattern[0], // rr
1233 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1,
1235 !if(HasPat, Pattern[2], // rm
1236 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
1237 (memopv2i64 addr:$src2)))])>,
1240 defm V#NAME#PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
1241 !strconcat(OpcodeStr, "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1243 !if(HasPat, Pattern[1], // rr
1244 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1247 !if(HasPat, Pattern[3], // rm
1248 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1249 (memopv2i64 addr:$src2)))])>,
1252 let Constraints = "$src1 = $dst" in {
1253 defm PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
1254 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"), f128mem,
1255 !if(HasPat, Pattern[0], // rr
1256 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1,
1258 !if(HasPat, Pattern[2], // rm
1259 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
1260 (memopv2i64 addr:$src2)))])>, TB;
1262 defm PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
1263 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"), f128mem,
1264 !if(HasPat, Pattern[1], // rr
1265 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1268 !if(HasPat, Pattern[3], // rm
1269 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1270 (memopv2i64 addr:$src2)))])>,
1275 defm AND : sse12_fp_packed_logical<0x54, "and", and>;
1276 defm OR : sse12_fp_packed_logical<0x56, "or", or>;
1277 defm XOR : sse12_fp_packed_logical<0x57, "xor", xor>;
1278 let isCommutable = 0 in
1279 defm ANDN : sse12_fp_packed_logical<0x55, "andn", undef /* dummy */, 1, [
1281 [(set VR128:$dst, (v2i64 (and (xor VR128:$src1,
1282 (bc_v2i64 (v4i32 immAllOnesV))),
1285 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1286 (bc_v2i64 (v2f64 VR128:$src2))))],
1288 [(set VR128:$dst, (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
1289 (bc_v2i64 (v4i32 immAllOnesV))),
1290 (memopv2i64 addr:$src2))))],
1292 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1293 (memopv2i64 addr:$src2)))]]>;
1295 //===----------------------------------------------------------------------===//
1296 // SSE 1 & 2 - Arithmetic Instructions
1297 //===----------------------------------------------------------------------===//
1299 /// basic_sse12_fp_binop_rm - SSE 1 & 2 binops come in both scalar and
1302 /// In addition, we also have a special variant of the scalar form here to
1303 /// represent the associated intrinsic operation. This form is unlike the
1304 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1305 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1307 /// These three forms can each be reg+reg or reg+mem.
1309 multiclass basic_sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
1312 let isAsmParserOnly = 1 in {
1313 defm V#NAME#SS : sse12_fp_scalar<opc,
1314 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1315 OpNode, FR32, f32mem>, XS, VEX_4V;
1317 defm V#NAME#SD : sse12_fp_scalar<opc,
1318 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1319 OpNode, FR64, f64mem>, XD, VEX_4V;
1321 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1322 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
1323 VR128, v4f32, f128mem, memopv4f32, SSEPackedSingle>,
1326 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1327 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
1328 VR128, v2f64, f128mem, memopv2f64, SSEPackedDouble>,
1331 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1332 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1333 "", "_ss", ssmem, sse_load_f32>, XS, VEX_4V;
1335 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1336 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1337 "2", "_sd", sdmem, sse_load_f64>, XD, VEX_4V;
1340 let Constraints = "$src1 = $dst" in {
1341 defm SS : sse12_fp_scalar<opc,
1342 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
1343 OpNode, FR32, f32mem>, XS;
1345 defm SD : sse12_fp_scalar<opc,
1346 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1347 OpNode, FR64, f64mem>, XD;
1349 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1350 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v4f32,
1351 f128mem, memopv4f32, SSEPackedSingle>, TB;
1353 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1354 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v2f64,
1355 f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
1357 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1358 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
1359 "", "_ss", ssmem, sse_load_f32>, XS;
1361 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1362 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1363 "2", "_sd", sdmem, sse_load_f64>, XD;
1367 // Arithmetic instructions
1368 defm ADD : basic_sse12_fp_binop_rm<0x58, "add", fadd>;
1369 defm MUL : basic_sse12_fp_binop_rm<0x59, "mul", fmul>;
1371 let isCommutable = 0 in {
1372 defm SUB : basic_sse12_fp_binop_rm<0x5C, "sub", fsub>;
1373 defm DIV : basic_sse12_fp_binop_rm<0x5E, "div", fdiv>;
1376 /// sse12_fp_binop_rm - Other SSE 1 & 2 binops
1378 /// This multiclass is like basic_sse12_fp_binop_rm, with the addition of
1379 /// instructions for a full-vector intrinsic form. Operations that map
1380 /// onto C operators don't use this form since they just use the plain
1381 /// vector form instead of having a separate vector intrinsic form.
1383 multiclass sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
1386 let isAsmParserOnly = 1 in {
1387 // Scalar operation, reg+reg.
1388 defm V#NAME#SS : sse12_fp_scalar<opc,
1389 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1390 OpNode, FR32, f32mem>, XS, VEX_4V;
1392 defm V#NAME#SD : sse12_fp_scalar<opc,
1393 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1394 OpNode, FR64, f64mem>, XD, VEX_4V;
1396 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1397 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
1398 VR128, v4f32, f128mem, memopv4f32, SSEPackedSingle>,
1401 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1402 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
1403 VR128, v2f64, f128mem, memopv2f64, SSEPackedDouble>,
1406 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1407 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1408 "", "_ss", ssmem, sse_load_f32>, XS, VEX_4V;
1410 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1411 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1412 "2", "_sd", sdmem, sse_load_f64>, XD, VEX_4V;
1414 defm V#NAME#PS : sse12_fp_packed_int<opc, OpcodeStr, VR128,
1415 !strconcat(OpcodeStr, "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1416 "", "_ps", f128mem, memopv4f32, SSEPackedSingle>, VEX_4V;
1418 defm V#NAME#PD : sse12_fp_packed_int<opc, OpcodeStr, VR128,
1419 !strconcat(OpcodeStr, "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1420 "2", "_pd", f128mem, memopv2f64, SSEPackedDouble>, OpSize,
1424 let Constraints = "$src1 = $dst" in {
1425 // Scalar operation, reg+reg.
1426 defm SS : sse12_fp_scalar<opc,
1427 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
1428 OpNode, FR32, f32mem>, XS;
1429 defm SD : sse12_fp_scalar<opc,
1430 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1431 OpNode, FR64, f64mem>, XD;
1432 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1433 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v4f32,
1434 f128mem, memopv4f32, SSEPackedSingle>, TB;
1436 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1437 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v2f64,
1438 f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
1440 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1441 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
1442 "", "_ss", ssmem, sse_load_f32>, XS;
1444 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1445 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1446 "2", "_sd", sdmem, sse_load_f64>, XD;
1448 defm PS : sse12_fp_packed_int<opc, OpcodeStr, VR128,
1449 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
1450 "", "_ps", f128mem, memopv4f32, SSEPackedSingle>, TB;
1452 defm PD : sse12_fp_packed_int<opc, OpcodeStr, VR128,
1453 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1454 "2", "_pd", f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
1458 let isCommutable = 0 in {
1459 defm MAX : sse12_fp_binop_rm<0x5F, "max", X86fmax>;
1460 defm MIN : sse12_fp_binop_rm<0x5D, "min", X86fmin>;
1465 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
1467 /// In addition, we also have a special variant of the scalar form here to
1468 /// represent the associated intrinsic operation. This form is unlike the
1469 /// plain scalar form, in that it takes an entire vector (instead of a
1470 /// scalar) and leaves the top elements undefined.
1472 /// And, we have a special variant form for a full-vector intrinsic form.
1474 /// These four forms can each have a reg or a mem operand, so there are a
1475 /// total of eight "instructions".
1477 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
1481 bit Commutable = 0> {
1482 // Scalar operation, reg.
1483 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
1484 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1485 [(set FR32:$dst, (OpNode FR32:$src))]> {
1486 let isCommutable = Commutable;
1489 // Scalar operation, mem.
1490 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
1491 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1492 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
1493 Requires<[HasSSE1, OptForSize]>;
1495 // Vector operation, reg.
1496 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1497 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1498 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
1499 let isCommutable = Commutable;
1502 // Vector operation, mem.
1503 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1504 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1505 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
1507 // Intrinsic operation, reg.
1508 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1509 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1510 [(set VR128:$dst, (F32Int VR128:$src))]> {
1511 let isCommutable = Commutable;
1514 // Intrinsic operation, mem.
1515 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
1516 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1517 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
1519 // Vector intrinsic operation, reg
1520 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1521 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1522 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
1523 let isCommutable = Commutable;
1526 // Vector intrinsic operation, mem
1527 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1528 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1529 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
1533 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
1534 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
1536 // Reciprocal approximations. Note that these typically require refinement
1537 // in order to obtain suitable precision.
1538 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
1539 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
1540 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
1541 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
1543 // Prefetch intrinsic.
1544 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1545 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1546 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1547 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1548 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1549 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1550 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1551 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1553 // Non-temporal stores
1554 def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1555 "movntps\t{$src, $dst|$dst, $src}",
1556 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1558 let AddedComplexity = 400 in { // Prefer non-temporal versions
1559 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1560 "movntps\t{$src, $dst|$dst, $src}",
1561 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
1563 def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1564 "movntdq\t{$src, $dst|$dst, $src}",
1565 [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
1567 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
1568 "movnti\t{$src, $dst|$dst, $src}",
1569 [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
1570 TB, Requires<[HasSSE2]>;
1572 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
1573 "movnti\t{$src, $dst|$dst, $src}",
1574 [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
1575 TB, Requires<[HasSSE2]>;
1578 // Load, store, and memory fence
1579 def SFENCE : I<0xAE, MRM_F8, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>,
1580 TB, Requires<[HasSSE1]>;
1583 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1584 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1585 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1586 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1588 // Alias instructions that map zero vector to pxor / xorp* for sse.
1589 // We set canFoldAsLoad because this can be converted to a constant-pool
1590 // load of an all-zeros value if folding it would be beneficial.
1591 // FIXME: Change encoding to pseudo!
1592 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1593 isCodeGenOnly = 1 in {
1594 def V_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1595 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
1596 def V_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1597 [(set VR128:$dst, (v2f64 immAllZerosV))]>;
1598 let ExeDomain = SSEPackedInt in
1599 def V_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
1600 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1603 def : Pat<(v2i64 immAllZerosV), (V_SET0PI)>;
1604 def : Pat<(v8i16 immAllZerosV), (V_SET0PI)>;
1605 def : Pat<(v16i8 immAllZerosV), (V_SET0PI)>;
1607 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1608 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
1610 //===---------------------------------------------------------------------===//
1611 // SSE2 Instructions
1612 //===---------------------------------------------------------------------===//
1614 // Conversion instructions
1615 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1616 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1617 [(set FR32:$dst, (fround FR64:$src))]>;
1618 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1619 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1620 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
1621 Requires<[HasSSE2, OptForSize]>;
1623 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1624 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1625 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1626 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1628 // SSE2 instructions with XS prefix
1629 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1630 "cvtss2sd\t{$src, $dst|$dst, $src}",
1631 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1632 Requires<[HasSSE2]>;
1633 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1634 "cvtss2sd\t{$src, $dst|$dst, $src}",
1635 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1636 Requires<[HasSSE2, OptForSize]>;
1638 def : Pat<(extloadf32 addr:$src),
1639 (CVTSS2SDrr (MOVSSrm addr:$src))>,
1640 Requires<[HasSSE2, OptForSpeed]>;
1642 // SSE2 instructions without OpSize prefix
1643 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1644 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1645 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1646 TB, Requires<[HasSSE2]>;
1647 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1648 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1649 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1650 (bitconvert (memopv2i64 addr:$src))))]>,
1651 TB, Requires<[HasSSE2]>;
1653 // SSE2 instructions with XS prefix
1654 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1655 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1656 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1657 XS, Requires<[HasSSE2]>;
1658 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1659 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1660 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1661 (bitconvert (memopv2i64 addr:$src))))]>,
1662 XS, Requires<[HasSSE2]>;
1664 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1665 "cvtps2dq\t{$src, $dst|$dst, $src}",
1666 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1667 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1668 "cvtps2dq\t{$src, $dst|$dst, $src}",
1669 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1670 (memop addr:$src)))]>;
1671 // SSE2 packed instructions with XS prefix
1672 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1673 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1674 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1675 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1677 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1678 "cvttps2dq\t{$src, $dst|$dst, $src}",
1680 (int_x86_sse2_cvttps2dq VR128:$src))]>,
1681 XS, Requires<[HasSSE2]>;
1682 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1683 "cvttps2dq\t{$src, $dst|$dst, $src}",
1684 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1685 (memop addr:$src)))]>,
1686 XS, Requires<[HasSSE2]>;
1688 // SSE2 packed instructions with XD prefix
1689 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1690 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1691 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1692 XD, Requires<[HasSSE2]>;
1693 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1694 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1695 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1696 (memop addr:$src)))]>,
1697 XD, Requires<[HasSSE2]>;
1699 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1700 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1701 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1702 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1703 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1704 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1705 (memop addr:$src)))]>;
1707 // SSE2 instructions without OpSize prefix
1708 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1709 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1710 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1711 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1713 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1714 "cvtps2pd\t{$src, $dst|$dst, $src}",
1715 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1716 TB, Requires<[HasSSE2]>;
1717 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1718 "cvtps2pd\t{$src, $dst|$dst, $src}",
1719 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1720 (load addr:$src)))]>,
1721 TB, Requires<[HasSSE2]>;
1723 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1724 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1725 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1726 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1729 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1730 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1731 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1732 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1733 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1734 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1735 (memop addr:$src)))]>;
1737 // Match intrinsics which expect XMM operand(s).
1738 // Aliases for intrinsics
1739 let Constraints = "$src1 = $dst" in {
1740 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1741 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1742 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1743 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1745 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1746 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1747 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1748 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1749 (load addr:$src2)))]>;
1750 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1751 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1752 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1753 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1754 VR128:$src2))]>, XS,
1755 Requires<[HasSSE2]>;
1756 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1757 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1758 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1759 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1760 (load addr:$src2)))]>, XS,
1761 Requires<[HasSSE2]>;
1766 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1768 /// In addition, we also have a special variant of the scalar form here to
1769 /// represent the associated intrinsic operation. This form is unlike the
1770 /// plain scalar form, in that it takes an entire vector (instead of a
1771 /// scalar) and leaves the top elements undefined.
1773 /// And, we have a special variant form for a full-vector intrinsic form.
1775 /// These four forms can each have a reg or a mem operand, so there are a
1776 /// total of eight "instructions".
1778 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1782 bit Commutable = 0> {
1783 // Scalar operation, reg.
1784 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1785 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1786 [(set FR64:$dst, (OpNode FR64:$src))]> {
1787 let isCommutable = Commutable;
1790 // Scalar operation, mem.
1791 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1792 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1793 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1795 // Vector operation, reg.
1796 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1797 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1798 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1799 let isCommutable = Commutable;
1802 // Vector operation, mem.
1803 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1804 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1805 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1807 // Intrinsic operation, reg.
1808 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1809 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1810 [(set VR128:$dst, (F64Int VR128:$src))]> {
1811 let isCommutable = Commutable;
1814 // Intrinsic operation, mem.
1815 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1816 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1817 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1819 // Vector intrinsic operation, reg
1820 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1821 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1822 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1823 let isCommutable = Commutable;
1826 // Vector intrinsic operation, mem
1827 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1828 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1829 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1833 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1834 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1836 // There is no f64 version of the reciprocal approximation instructions.
1838 //===---------------------------------------------------------------------===//
1839 // SSE integer instructions
1840 let ExeDomain = SSEPackedInt in {
1842 // Move Instructions
1843 let neverHasSideEffects = 1 in
1844 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1845 "movdqa\t{$src, $dst|$dst, $src}", []>;
1846 let canFoldAsLoad = 1, mayLoad = 1 in
1847 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1848 "movdqa\t{$src, $dst|$dst, $src}",
1849 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1851 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1852 "movdqa\t{$src, $dst|$dst, $src}",
1853 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1854 let canFoldAsLoad = 1, mayLoad = 1 in
1855 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1856 "movdqu\t{$src, $dst|$dst, $src}",
1857 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1858 XS, Requires<[HasSSE2]>;
1860 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1861 "movdqu\t{$src, $dst|$dst, $src}",
1862 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1863 XS, Requires<[HasSSE2]>;
1865 // Intrinsic forms of MOVDQU load and store
1866 let canFoldAsLoad = 1 in
1867 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1868 "movdqu\t{$src, $dst|$dst, $src}",
1869 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1870 XS, Requires<[HasSSE2]>;
1871 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1872 "movdqu\t{$src, $dst|$dst, $src}",
1873 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1874 XS, Requires<[HasSSE2]>;
1876 let Constraints = "$src1 = $dst" in {
1878 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1879 bit Commutable = 0> {
1880 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1881 (ins VR128:$src1, VR128:$src2),
1882 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1883 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1884 let isCommutable = Commutable;
1886 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1887 (ins VR128:$src1, i128mem:$src2),
1888 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1889 [(set VR128:$dst, (IntId VR128:$src1,
1890 (bitconvert (memopv2i64
1894 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1896 Intrinsic IntId, Intrinsic IntId2> {
1897 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1898 (ins VR128:$src1, VR128:$src2),
1899 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1900 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1901 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1902 (ins VR128:$src1, i128mem:$src2),
1903 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1904 [(set VR128:$dst, (IntId VR128:$src1,
1905 (bitconvert (memopv2i64 addr:$src2))))]>;
1906 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
1907 (ins VR128:$src1, i32i8imm:$src2),
1908 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1909 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1912 /// PDI_binop_rm - Simple SSE2 binary operator.
1913 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1914 ValueType OpVT, bit Commutable = 0> {
1915 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1916 (ins VR128:$src1, VR128:$src2),
1917 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1918 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1919 let isCommutable = Commutable;
1921 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1922 (ins VR128:$src1, i128mem:$src2),
1923 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1924 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1925 (bitconvert (memopv2i64 addr:$src2)))))]>;
1928 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1930 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1931 /// to collapse (bitconvert VT to VT) into its operand.
1933 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1934 bit Commutable = 0> {
1935 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1936 (ins VR128:$src1, VR128:$src2),
1937 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1938 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1939 let isCommutable = Commutable;
1941 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1942 (ins VR128:$src1, i128mem:$src2),
1943 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1944 [(set VR128:$dst, (OpNode VR128:$src1,
1945 (memopv2i64 addr:$src2)))]>;
1948 } // Constraints = "$src1 = $dst"
1949 } // ExeDomain = SSEPackedInt
1951 // 128-bit Integer Arithmetic
1953 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1954 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1955 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1956 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1958 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1959 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1960 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1961 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1963 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1964 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1965 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1966 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1968 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1969 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1970 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1971 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1973 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1975 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1976 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1977 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1979 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1981 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1982 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1985 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1986 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1987 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1988 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1989 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
1992 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
1993 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
1994 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
1995 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
1996 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
1997 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
1999 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2000 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2001 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2002 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2003 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2004 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2006 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2007 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2008 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2009 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2011 // 128-bit logical shifts.
2012 let Constraints = "$src1 = $dst", neverHasSideEffects = 1,
2013 ExeDomain = SSEPackedInt in {
2014 def PSLLDQri : PDIi8<0x73, MRM7r,
2015 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2016 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2017 def PSRLDQri : PDIi8<0x73, MRM3r,
2018 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2019 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2020 // PSRADQri doesn't exist in SSE[1-3].
2023 let Predicates = [HasSSE2] in {
2024 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2025 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2026 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2027 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2028 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2029 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2030 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2031 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2032 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2033 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2035 // Shift up / down and insert zero's.
2036 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2037 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2038 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2039 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2043 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2044 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2045 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2047 let Constraints = "$src1 = $dst", ExeDomain = SSEPackedInt in {
2048 def PANDNrr : PDI<0xDF, MRMSrcReg,
2049 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2050 "pandn\t{$src2, $dst|$dst, $src2}",
2051 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2054 def PANDNrm : PDI<0xDF, MRMSrcMem,
2055 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2056 "pandn\t{$src2, $dst|$dst, $src2}",
2057 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2058 (memopv2i64 addr:$src2))))]>;
2061 // SSE2 Integer comparison
2062 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2063 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2064 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2065 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2066 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2067 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2069 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2070 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2071 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2072 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2073 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2074 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2075 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2076 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2077 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2078 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2079 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2080 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2082 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2083 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2084 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2085 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2086 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2087 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2088 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2089 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2090 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2091 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2092 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2093 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2096 // Pack instructions
2097 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2098 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2099 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2101 let ExeDomain = SSEPackedInt in {
2103 // Shuffle and unpack instructions
2104 let AddedComplexity = 5 in {
2105 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2106 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2107 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2108 [(set VR128:$dst, (v4i32 (pshufd:$src2
2109 VR128:$src1, (undef))))]>;
2110 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2111 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2112 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2113 [(set VR128:$dst, (v4i32 (pshufd:$src2
2114 (bc_v4i32 (memopv2i64 addr:$src1)),
2118 // SSE2 with ImmT == Imm8 and XS prefix.
2119 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2120 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2121 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2122 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2124 XS, Requires<[HasSSE2]>;
2125 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2126 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2127 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2128 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2129 (bc_v8i16 (memopv2i64 addr:$src1)),
2131 XS, Requires<[HasSSE2]>;
2133 // SSE2 with ImmT == Imm8 and XD prefix.
2134 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2135 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2136 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2137 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2139 XD, Requires<[HasSSE2]>;
2140 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2141 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2142 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2143 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2144 (bc_v8i16 (memopv2i64 addr:$src1)),
2146 XD, Requires<[HasSSE2]>;
2148 // Unpack instructions
2149 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
2150 PatFrag unp_frag, PatFrag bc_frag> {
2151 def rr : PDI<opc, MRMSrcReg,
2152 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2153 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2154 [(set VR128:$dst, (vt (unp_frag VR128:$src1, VR128:$src2)))]>;
2155 def rm : PDI<opc, MRMSrcMem,
2156 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2157 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2158 [(set VR128:$dst, (unp_frag VR128:$src1,
2159 (bc_frag (memopv2i64
2163 let Constraints = "$src1 = $dst" in {
2164 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, unpckl, bc_v16i8>;
2165 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, unpckl, bc_v8i16>;
2166 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, unpckl, bc_v4i32>;
2168 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2169 /// knew to collapse (bitconvert VT to VT) into its operand.
2170 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2171 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2172 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2174 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2175 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2176 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2177 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2179 (v2i64 (unpckl VR128:$src1,
2180 (memopv2i64 addr:$src2))))]>;
2182 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, unpckh, bc_v16i8>;
2183 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, unpckh, bc_v8i16>;
2184 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, unpckh, bc_v4i32>;
2186 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2187 /// knew to collapse (bitconvert VT to VT) into its operand.
2188 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2189 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2190 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2192 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2193 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2194 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2195 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2197 (v2i64 (unpckh VR128:$src1,
2198 (memopv2i64 addr:$src2))))]>;
2202 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2203 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2204 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2205 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2207 let Constraints = "$src1 = $dst" in {
2208 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2209 (outs VR128:$dst), (ins VR128:$src1,
2210 GR32:$src2, i32i8imm:$src3),
2211 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2213 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2214 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2215 (outs VR128:$dst), (ins VR128:$src1,
2216 i16mem:$src2, i32i8imm:$src3),
2217 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2219 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2224 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2225 "pmovmskb\t{$src, $dst|$dst, $src}",
2226 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2228 // Conditional store
2230 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2231 "maskmovdqu\t{$mask, $src|$src, $mask}",
2232 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2235 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2236 "maskmovdqu\t{$mask, $src|$src, $mask}",
2237 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2239 } // ExeDomain = SSEPackedInt
2241 // Non-temporal stores
2242 def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2243 "movntpd\t{$src, $dst|$dst, $src}",
2244 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2245 let ExeDomain = SSEPackedInt in
2246 def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2247 "movntdq\t{$src, $dst|$dst, $src}",
2248 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2249 def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2250 "movnti\t{$src, $dst|$dst, $src}",
2251 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2252 TB, Requires<[HasSSE2]>;
2254 let AddedComplexity = 400 in { // Prefer non-temporal versions
2255 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2256 "movntpd\t{$src, $dst|$dst, $src}",
2257 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
2259 let ExeDomain = SSEPackedInt in
2260 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2261 "movntdq\t{$src, $dst|$dst, $src}",
2262 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2266 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2267 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2268 TB, Requires<[HasSSE2]>;
2270 // Load, store, and memory fence
2271 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
2272 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2273 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
2274 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2276 // Pause. This "instruction" is encoded as "rep; nop", so even though it
2277 // was introduced with SSE2, it's backward compatible.
2278 def PAUSE : I<0x90, RawFrm, (outs), (ins), "pause", []>, REP;
2280 //TODO: custom lower this so as to never even generate the noop
2281 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2283 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2284 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2285 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2288 // Alias instructions that map zero vector to pxor / xorp* for sse.
2289 // We set canFoldAsLoad because this can be converted to a constant-pool
2290 // load of an all-ones value if folding it would be beneficial.
2291 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2292 isCodeGenOnly = 1, ExeDomain = SSEPackedInt in
2293 // FIXME: Change encoding to pseudo.
2294 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
2295 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2297 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2298 "movd\t{$src, $dst|$dst, $src}",
2300 (v4i32 (scalar_to_vector GR32:$src)))]>;
2301 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2302 "movd\t{$src, $dst|$dst, $src}",
2304 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2306 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2307 "movd\t{$src, $dst|$dst, $src}",
2308 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2310 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2311 "movd\t{$src, $dst|$dst, $src}",
2312 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2314 // SSE2 instructions with XS prefix
2315 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2316 "movq\t{$src, $dst|$dst, $src}",
2318 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2319 Requires<[HasSSE2]>;
2320 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2321 "movq\t{$src, $dst|$dst, $src}",
2322 [(store (i64 (vector_extract (v2i64 VR128:$src),
2323 (iPTR 0))), addr:$dst)]>;
2325 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
2326 (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
2328 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2329 "movd\t{$src, $dst|$dst, $src}",
2330 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2332 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2333 "movd\t{$src, $dst|$dst, $src}",
2334 [(store (i32 (vector_extract (v4i32 VR128:$src),
2335 (iPTR 0))), addr:$dst)]>;
2337 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2338 "movd\t{$src, $dst|$dst, $src}",
2339 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2340 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2341 "movd\t{$src, $dst|$dst, $src}",
2342 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2344 // Store / copy lower 64-bits of a XMM register.
2345 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2346 "movq\t{$src, $dst|$dst, $src}",
2347 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2349 // movd / movq to XMM register zero-extends
2350 let AddedComplexity = 15 in {
2351 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2352 "movd\t{$src, $dst|$dst, $src}",
2353 [(set VR128:$dst, (v4i32 (X86vzmovl
2354 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2355 // This is X86-64 only.
2356 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2357 "mov{d|q}\t{$src, $dst|$dst, $src}",
2358 [(set VR128:$dst, (v2i64 (X86vzmovl
2359 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2362 let AddedComplexity = 20 in {
2363 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2364 "movd\t{$src, $dst|$dst, $src}",
2366 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2367 (loadi32 addr:$src))))))]>;
2369 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2370 (MOVZDI2PDIrm addr:$src)>;
2371 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2372 (MOVZDI2PDIrm addr:$src)>;
2373 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2374 (MOVZDI2PDIrm addr:$src)>;
2376 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2377 "movq\t{$src, $dst|$dst, $src}",
2379 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2380 (loadi64 addr:$src))))))]>, XS,
2381 Requires<[HasSSE2]>;
2383 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2384 (MOVZQI2PQIrm addr:$src)>;
2385 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2386 (MOVZQI2PQIrm addr:$src)>;
2387 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2390 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2391 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2392 let AddedComplexity = 15 in
2393 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2394 "movq\t{$src, $dst|$dst, $src}",
2395 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2396 XS, Requires<[HasSSE2]>;
2398 let AddedComplexity = 20 in {
2399 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2400 "movq\t{$src, $dst|$dst, $src}",
2401 [(set VR128:$dst, (v2i64 (X86vzmovl
2402 (loadv2i64 addr:$src))))]>,
2403 XS, Requires<[HasSSE2]>;
2405 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2406 (MOVZPQILo2PQIrm addr:$src)>;
2409 // Instructions for the disassembler
2410 // xr = XMM register
2413 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2414 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2416 //===---------------------------------------------------------------------===//
2417 // SSE3 Instructions
2418 //===---------------------------------------------------------------------===//
2420 // Conversion Instructions
2421 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2422 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
2423 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2424 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
2425 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2426 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
2427 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2428 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
2430 // Move Instructions
2431 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2432 "movshdup\t{$src, $dst|$dst, $src}",
2433 [(set VR128:$dst, (v4f32 (movshdup
2434 VR128:$src, (undef))))]>;
2435 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2436 "movshdup\t{$src, $dst|$dst, $src}",
2437 [(set VR128:$dst, (movshdup
2438 (memopv4f32 addr:$src), (undef)))]>;
2440 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2441 "movsldup\t{$src, $dst|$dst, $src}",
2442 [(set VR128:$dst, (v4f32 (movsldup
2443 VR128:$src, (undef))))]>;
2444 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2445 "movsldup\t{$src, $dst|$dst, $src}",
2446 [(set VR128:$dst, (movsldup
2447 (memopv4f32 addr:$src), (undef)))]>;
2449 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2450 "movddup\t{$src, $dst|$dst, $src}",
2451 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2452 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2453 "movddup\t{$src, $dst|$dst, $src}",
2455 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2458 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2460 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2462 let AddedComplexity = 5 in {
2463 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2464 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2465 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2466 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2467 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2468 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2469 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2470 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2474 let Constraints = "$src1 = $dst" in {
2475 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2476 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2477 "addsubps\t{$src2, $dst|$dst, $src2}",
2478 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2480 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2481 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2482 "addsubps\t{$src2, $dst|$dst, $src2}",
2483 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2484 (memop addr:$src2)))]>;
2485 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2486 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2487 "addsubpd\t{$src2, $dst|$dst, $src2}",
2488 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2490 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2491 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2492 "addsubpd\t{$src2, $dst|$dst, $src2}",
2493 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2494 (memop addr:$src2)))]>;
2497 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2498 "lddqu\t{$src, $dst|$dst, $src}",
2499 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2502 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2503 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2504 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2505 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2506 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2507 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2508 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2509 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2510 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2511 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2512 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2513 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2514 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2515 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2516 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2517 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2519 let Constraints = "$src1 = $dst" in {
2520 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2521 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2522 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2523 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2524 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2525 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2526 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2527 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2530 // Thread synchronization
2531 def MONITOR : I<0x01, MRM_C8, (outs), (ins), "monitor",
2532 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2533 def MWAIT : I<0x01, MRM_C9, (outs), (ins), "mwait",
2534 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2536 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2537 let AddedComplexity = 15 in
2538 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2539 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2540 let AddedComplexity = 20 in
2541 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2542 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2544 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2545 let AddedComplexity = 15 in
2546 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2547 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2548 let AddedComplexity = 20 in
2549 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2550 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2552 //===---------------------------------------------------------------------===//
2553 // SSSE3 Instructions
2554 //===---------------------------------------------------------------------===//
2556 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2557 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2558 Intrinsic IntId64, Intrinsic IntId128> {
2559 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2560 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2561 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2563 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2564 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2566 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2568 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2570 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2571 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2574 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2576 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2579 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2582 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2583 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2584 Intrinsic IntId64, Intrinsic IntId128> {
2585 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2587 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2588 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2590 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2592 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2595 (bitconvert (memopv4i16 addr:$src))))]>;
2597 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2599 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2600 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2603 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2605 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2608 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2611 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2612 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2613 Intrinsic IntId64, Intrinsic IntId128> {
2614 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2616 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2617 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2619 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2621 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2624 (bitconvert (memopv2i32 addr:$src))))]>;
2626 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2628 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2629 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2632 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2634 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2637 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2640 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2641 int_x86_ssse3_pabs_b,
2642 int_x86_ssse3_pabs_b_128>;
2643 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2644 int_x86_ssse3_pabs_w,
2645 int_x86_ssse3_pabs_w_128>;
2646 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2647 int_x86_ssse3_pabs_d,
2648 int_x86_ssse3_pabs_d_128>;
2650 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2651 let Constraints = "$src1 = $dst" in {
2652 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2653 Intrinsic IntId64, Intrinsic IntId128,
2654 bit Commutable = 0> {
2655 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2656 (ins VR64:$src1, VR64:$src2),
2657 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2658 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2659 let isCommutable = Commutable;
2661 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2662 (ins VR64:$src1, i64mem:$src2),
2663 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2665 (IntId64 VR64:$src1,
2666 (bitconvert (memopv8i8 addr:$src2))))]>;
2668 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2669 (ins VR128:$src1, VR128:$src2),
2670 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2671 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2673 let isCommutable = Commutable;
2675 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2676 (ins VR128:$src1, i128mem:$src2),
2677 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2679 (IntId128 VR128:$src1,
2680 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2684 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2685 let Constraints = "$src1 = $dst" in {
2686 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2687 Intrinsic IntId64, Intrinsic IntId128,
2688 bit Commutable = 0> {
2689 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2690 (ins VR64:$src1, VR64:$src2),
2691 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2692 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2693 let isCommutable = Commutable;
2695 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2696 (ins VR64:$src1, i64mem:$src2),
2697 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2699 (IntId64 VR64:$src1,
2700 (bitconvert (memopv4i16 addr:$src2))))]>;
2702 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2703 (ins VR128:$src1, VR128:$src2),
2704 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2705 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2707 let isCommutable = Commutable;
2709 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2710 (ins VR128:$src1, i128mem:$src2),
2711 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2713 (IntId128 VR128:$src1,
2714 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2718 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2719 let Constraints = "$src1 = $dst" in {
2720 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2721 Intrinsic IntId64, Intrinsic IntId128,
2722 bit Commutable = 0> {
2723 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2724 (ins VR64:$src1, VR64:$src2),
2725 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2726 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2727 let isCommutable = Commutable;
2729 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2730 (ins VR64:$src1, i64mem:$src2),
2731 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2733 (IntId64 VR64:$src1,
2734 (bitconvert (memopv2i32 addr:$src2))))]>;
2736 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2737 (ins VR128:$src1, VR128:$src2),
2738 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2739 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2741 let isCommutable = Commutable;
2743 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2744 (ins VR128:$src1, i128mem:$src2),
2745 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2747 (IntId128 VR128:$src1,
2748 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2752 let ImmT = NoImm in { // None of these have i8 immediate fields.
2753 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2754 int_x86_ssse3_phadd_w,
2755 int_x86_ssse3_phadd_w_128>;
2756 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2757 int_x86_ssse3_phadd_d,
2758 int_x86_ssse3_phadd_d_128>;
2759 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2760 int_x86_ssse3_phadd_sw,
2761 int_x86_ssse3_phadd_sw_128>;
2762 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2763 int_x86_ssse3_phsub_w,
2764 int_x86_ssse3_phsub_w_128>;
2765 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2766 int_x86_ssse3_phsub_d,
2767 int_x86_ssse3_phsub_d_128>;
2768 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2769 int_x86_ssse3_phsub_sw,
2770 int_x86_ssse3_phsub_sw_128>;
2771 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2772 int_x86_ssse3_pmadd_ub_sw,
2773 int_x86_ssse3_pmadd_ub_sw_128>;
2774 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2775 int_x86_ssse3_pmul_hr_sw,
2776 int_x86_ssse3_pmul_hr_sw_128, 1>;
2778 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2779 int_x86_ssse3_pshuf_b,
2780 int_x86_ssse3_pshuf_b_128>;
2781 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2782 int_x86_ssse3_psign_b,
2783 int_x86_ssse3_psign_b_128>;
2784 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2785 int_x86_ssse3_psign_w,
2786 int_x86_ssse3_psign_w_128>;
2787 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2788 int_x86_ssse3_psign_d,
2789 int_x86_ssse3_psign_d_128>;
2792 // palignr patterns.
2793 let Constraints = "$src1 = $dst" in {
2794 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2795 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2796 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2798 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2799 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2800 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2803 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2804 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2805 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2807 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2808 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2809 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2813 let AddedComplexity = 5 in {
2815 def : Pat<(v1i64 (palign:$src3 VR64:$src1, VR64:$src2)),
2816 (PALIGNR64rr VR64:$src2, VR64:$src1,
2817 (SHUFFLE_get_palign_imm VR64:$src3))>,
2818 Requires<[HasSSSE3]>;
2819 def : Pat<(v2i32 (palign:$src3 VR64:$src1, VR64:$src2)),
2820 (PALIGNR64rr VR64:$src2, VR64:$src1,
2821 (SHUFFLE_get_palign_imm VR64:$src3))>,
2822 Requires<[HasSSSE3]>;
2823 def : Pat<(v2f32 (palign:$src3 VR64:$src1, VR64:$src2)),
2824 (PALIGNR64rr VR64:$src2, VR64:$src1,
2825 (SHUFFLE_get_palign_imm VR64:$src3))>,
2826 Requires<[HasSSSE3]>;
2827 def : Pat<(v4i16 (palign:$src3 VR64:$src1, VR64:$src2)),
2828 (PALIGNR64rr VR64:$src2, VR64:$src1,
2829 (SHUFFLE_get_palign_imm VR64:$src3))>,
2830 Requires<[HasSSSE3]>;
2831 def : Pat<(v8i8 (palign:$src3 VR64:$src1, VR64:$src2)),
2832 (PALIGNR64rr VR64:$src2, VR64:$src1,
2833 (SHUFFLE_get_palign_imm VR64:$src3))>,
2834 Requires<[HasSSSE3]>;
2836 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2837 (PALIGNR128rr VR128:$src2, VR128:$src1,
2838 (SHUFFLE_get_palign_imm VR128:$src3))>,
2839 Requires<[HasSSSE3]>;
2840 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2841 (PALIGNR128rr VR128:$src2, VR128:$src1,
2842 (SHUFFLE_get_palign_imm VR128:$src3))>,
2843 Requires<[HasSSSE3]>;
2844 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2845 (PALIGNR128rr VR128:$src2, VR128:$src1,
2846 (SHUFFLE_get_palign_imm VR128:$src3))>,
2847 Requires<[HasSSSE3]>;
2848 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2849 (PALIGNR128rr VR128:$src2, VR128:$src1,
2850 (SHUFFLE_get_palign_imm VR128:$src3))>,
2851 Requires<[HasSSSE3]>;
2854 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2855 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2856 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2857 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2859 //===---------------------------------------------------------------------===//
2860 // Non-Instruction Patterns
2861 //===---------------------------------------------------------------------===//
2863 // extload f32 -> f64. This matches load+fextend because we have a hack in
2864 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2866 // Since these loads aren't folded into the fextend, we have to match it
2868 let Predicates = [HasSSE2] in
2869 def : Pat<(fextend (loadf32 addr:$src)),
2870 (CVTSS2SDrm addr:$src)>;
2873 let Predicates = [HasSSE2] in {
2874 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2875 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2876 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2877 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2878 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2879 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2880 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2881 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2882 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2883 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2884 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2885 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2886 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2887 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2888 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2889 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2890 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2891 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2892 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2893 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2894 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2895 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2896 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2897 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2898 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2899 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2900 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2901 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2902 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2903 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2906 // Move scalar to XMM zero-extended
2907 // movd to XMM register zero-extends
2908 let AddedComplexity = 15 in {
2909 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2910 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2911 (MOVSDrr (v2f64 (V_SET0PS)), FR64:$src)>;
2912 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2913 (MOVSSrr (v4f32 (V_SET0PS)), FR32:$src)>;
2914 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2915 (MOVSSrr (v4f32 (V_SET0PS)),
2916 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss)))>;
2917 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2918 (MOVSSrr (v4i32 (V_SET0PI)),
2919 (EXTRACT_SUBREG (v4i32 VR128:$src), sub_ss))>;
2922 // Splat v2f64 / v2i64
2923 let AddedComplexity = 10 in {
2924 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
2925 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2926 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
2927 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2928 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
2929 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2930 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
2931 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2934 // Special unary SHUFPSrri case.
2935 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
2936 (SHUFPSrri VR128:$src1, VR128:$src1,
2937 (SHUFFLE_get_shuf_imm VR128:$src3))>;
2938 let AddedComplexity = 5 in
2939 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
2940 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2941 Requires<[HasSSE2]>;
2942 // Special unary SHUFPDrri case.
2943 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
2944 (SHUFPDrri VR128:$src1, VR128:$src1,
2945 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2946 Requires<[HasSSE2]>;
2947 // Special unary SHUFPDrri case.
2948 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
2949 (SHUFPDrri VR128:$src1, VR128:$src1,
2950 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2951 Requires<[HasSSE2]>;
2952 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2953 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
2954 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2955 Requires<[HasSSE2]>;
2957 // Special binary v4i32 shuffle cases with SHUFPS.
2958 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
2959 (SHUFPSrri VR128:$src1, VR128:$src2,
2960 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2961 Requires<[HasSSE2]>;
2962 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
2963 (SHUFPSrmi VR128:$src1, addr:$src2,
2964 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2965 Requires<[HasSSE2]>;
2966 // Special binary v2i64 shuffle cases using SHUFPDrri.
2967 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
2968 (SHUFPDrri VR128:$src1, VR128:$src2,
2969 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2970 Requires<[HasSSE2]>;
2972 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2973 let AddedComplexity = 15 in {
2974 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
2975 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2976 Requires<[OptForSpeed, HasSSE2]>;
2977 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
2978 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2979 Requires<[OptForSpeed, HasSSE2]>;
2981 let AddedComplexity = 10 in {
2982 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
2983 (UNPCKLPSrr VR128:$src, VR128:$src)>;
2984 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
2985 (PUNPCKLBWrr VR128:$src, VR128:$src)>;
2986 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
2987 (PUNPCKLWDrr VR128:$src, VR128:$src)>;
2988 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
2989 (PUNPCKLDQrr VR128:$src, VR128:$src)>;
2992 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2993 let AddedComplexity = 15 in {
2994 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
2995 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2996 Requires<[OptForSpeed, HasSSE2]>;
2997 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
2998 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2999 Requires<[OptForSpeed, HasSSE2]>;
3001 let AddedComplexity = 10 in {
3002 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
3003 (UNPCKHPSrr VR128:$src, VR128:$src)>;
3004 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
3005 (PUNPCKHBWrr VR128:$src, VR128:$src)>;
3006 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
3007 (PUNPCKHWDrr VR128:$src, VR128:$src)>;
3008 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
3009 (PUNPCKHDQrr VR128:$src, VR128:$src)>;
3012 let AddedComplexity = 20 in {
3013 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3014 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3015 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3017 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3018 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3019 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3021 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3022 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3023 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3024 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3025 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3028 let AddedComplexity = 20 in {
3029 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3030 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3031 (MOVLPSrm VR128:$src1, addr:$src2)>;
3032 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3033 (MOVLPDrm VR128:$src1, addr:$src2)>;
3034 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3035 (MOVLPSrm VR128:$src1, addr:$src2)>;
3036 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3037 (MOVLPDrm VR128:$src1, addr:$src2)>;
3040 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3041 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3042 (MOVLPSmr addr:$src1, VR128:$src2)>;
3043 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3044 (MOVLPDmr addr:$src1, VR128:$src2)>;
3045 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3047 (MOVLPSmr addr:$src1, VR128:$src2)>;
3048 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3049 (MOVLPDmr addr:$src1, VR128:$src2)>;
3051 let AddedComplexity = 15 in {
3052 // Setting the lowest element in the vector.
3053 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3054 (MOVSSrr (v4i32 VR128:$src1),
3055 (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_ss))>;
3056 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3057 (MOVSDrr (v2i64 VR128:$src1),
3058 (EXTRACT_SUBREG (v2i64 VR128:$src2), sub_sd))>;
3060 // vector_shuffle v1, v2 <4, 5, 2, 3> using movsd
3061 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3062 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3063 Requires<[HasSSE2]>;
3064 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3065 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3066 Requires<[HasSSE2]>;
3069 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3070 // fall back to this for SSE1)
3071 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3072 (SHUFPSrri VR128:$src2, VR128:$src1,
3073 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3075 // Set lowest element and zero upper elements.
3076 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3077 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3079 // Some special case pandn patterns.
3080 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3082 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3083 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3085 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3086 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3088 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3090 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3091 (memop addr:$src2))),
3092 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3093 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3094 (memop addr:$src2))),
3095 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3096 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3097 (memop addr:$src2))),
3098 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3100 // vector -> vector casts
3101 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3102 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3103 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3104 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3105 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3106 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3107 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3108 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3110 // Use movaps / movups for SSE integer load / store (one byte shorter).
3111 def : Pat<(alignedloadv4i32 addr:$src),
3112 (MOVAPSrm addr:$src)>;
3113 def : Pat<(loadv4i32 addr:$src),
3114 (MOVUPSrm addr:$src)>;
3115 def : Pat<(alignedloadv2i64 addr:$src),
3116 (MOVAPSrm addr:$src)>;
3117 def : Pat<(loadv2i64 addr:$src),
3118 (MOVUPSrm addr:$src)>;
3120 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3121 (MOVAPSmr addr:$dst, VR128:$src)>;
3122 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3123 (MOVAPSmr addr:$dst, VR128:$src)>;
3124 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3125 (MOVAPSmr addr:$dst, VR128:$src)>;
3126 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3127 (MOVAPSmr addr:$dst, VR128:$src)>;
3128 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3129 (MOVUPSmr addr:$dst, VR128:$src)>;
3130 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3131 (MOVUPSmr addr:$dst, VR128:$src)>;
3132 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3133 (MOVUPSmr addr:$dst, VR128:$src)>;
3134 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3135 (MOVUPSmr addr:$dst, VR128:$src)>;
3137 //===----------------------------------------------------------------------===//
3138 // SSE4.1 Instructions
3139 //===----------------------------------------------------------------------===//
3141 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3144 Intrinsic V2F64Int> {
3145 // Intrinsic operation, reg.
3146 // Vector intrinsic operation, reg
3147 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3148 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3149 !strconcat(OpcodeStr,
3150 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3151 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3154 // Vector intrinsic operation, mem
3155 def PSm_Int : Ii8<opcps, MRMSrcMem,
3156 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3157 !strconcat(OpcodeStr,
3158 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3160 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3162 Requires<[HasSSE41]>;
3164 // Vector intrinsic operation, reg
3165 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3166 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3167 !strconcat(OpcodeStr,
3168 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3169 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3172 // Vector intrinsic operation, mem
3173 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3174 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3175 !strconcat(OpcodeStr,
3176 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3178 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3182 let Constraints = "$src1 = $dst" in {
3183 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3187 // Intrinsic operation, reg.
3188 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3190 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3191 !strconcat(OpcodeStr,
3192 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3194 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3197 // Intrinsic operation, mem.
3198 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3200 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3201 !strconcat(OpcodeStr,
3202 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3204 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3207 // Intrinsic operation, reg.
3208 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3210 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3211 !strconcat(OpcodeStr,
3212 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3214 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3217 // Intrinsic operation, mem.
3218 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3220 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3221 !strconcat(OpcodeStr,
3222 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3224 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3229 // FP round - roundss, roundps, roundsd, roundpd
3230 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3231 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3232 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3233 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3235 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3236 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3237 Intrinsic IntId128> {
3238 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3240 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3241 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3242 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3244 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3247 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3250 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3251 int_x86_sse41_phminposuw>;
3253 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3254 let Constraints = "$src1 = $dst" in {
3255 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3256 Intrinsic IntId128, bit Commutable = 0> {
3257 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3258 (ins VR128:$src1, VR128:$src2),
3259 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3260 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3262 let isCommutable = Commutable;
3264 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3265 (ins VR128:$src1, i128mem:$src2),
3266 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3268 (IntId128 VR128:$src1,
3269 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3273 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3274 int_x86_sse41_pcmpeqq, 1>;
3275 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3276 int_x86_sse41_packusdw, 0>;
3277 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3278 int_x86_sse41_pminsb, 1>;
3279 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3280 int_x86_sse41_pminsd, 1>;
3281 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3282 int_x86_sse41_pminud, 1>;
3283 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3284 int_x86_sse41_pminuw, 1>;
3285 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3286 int_x86_sse41_pmaxsb, 1>;
3287 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3288 int_x86_sse41_pmaxsd, 1>;
3289 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3290 int_x86_sse41_pmaxud, 1>;
3291 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3292 int_x86_sse41_pmaxuw, 1>;
3294 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3296 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3297 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3298 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3299 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3301 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3302 let Constraints = "$src1 = $dst" in {
3303 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3304 SDNode OpNode, Intrinsic IntId128,
3305 bit Commutable = 0> {
3306 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3307 (ins VR128:$src1, VR128:$src2),
3308 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3309 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3310 VR128:$src2))]>, OpSize {
3311 let isCommutable = Commutable;
3313 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3314 (ins VR128:$src1, VR128:$src2),
3315 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3316 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3318 let isCommutable = Commutable;
3320 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3321 (ins VR128:$src1, i128mem:$src2),
3322 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3324 (OpVT (OpNode VR128:$src1, (memop addr:$src2))))]>, OpSize;
3325 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3326 (ins VR128:$src1, i128mem:$src2),
3327 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3329 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3334 /// SS48I_binop_rm - Simple SSE41 binary operator.
3335 let Constraints = "$src1 = $dst" in {
3336 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
3337 ValueType OpVT, bit Commutable = 0> {
3338 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3339 (ins VR128:$src1, VR128:$src2),
3340 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3341 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]>,
3343 let isCommutable = Commutable;
3345 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3346 (ins VR128:$src1, i128mem:$src2),
3347 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3348 [(set VR128:$dst, (OpNode VR128:$src1,
3349 (bc_v4i32 (memopv2i64 addr:$src2))))]>,
3354 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, 1>;
3356 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3357 let Constraints = "$src1 = $dst" in {
3358 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3359 Intrinsic IntId128, bit Commutable = 0> {
3360 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3361 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3362 !strconcat(OpcodeStr,
3363 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3365 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3367 let isCommutable = Commutable;
3369 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3370 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3371 !strconcat(OpcodeStr,
3372 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3374 (IntId128 VR128:$src1,
3375 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3380 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3381 int_x86_sse41_blendps, 0>;
3382 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3383 int_x86_sse41_blendpd, 0>;
3384 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3385 int_x86_sse41_pblendw, 0>;
3386 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3387 int_x86_sse41_dpps, 1>;
3388 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3389 int_x86_sse41_dppd, 1>;
3390 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3391 int_x86_sse41_mpsadbw, 0>;
3394 /// SS41I_ternary_int - SSE 4.1 ternary operator
3395 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3396 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3397 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3398 (ins VR128:$src1, VR128:$src2),
3399 !strconcat(OpcodeStr,
3400 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3401 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3404 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3405 (ins VR128:$src1, i128mem:$src2),
3406 !strconcat(OpcodeStr,
3407 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3410 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3414 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3415 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3416 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3419 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3420 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3421 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3422 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3424 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3425 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3427 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3431 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3432 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3433 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3434 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3435 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3436 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3438 // Common patterns involving scalar load.
3439 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3440 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3441 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3442 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3444 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3445 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3446 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3447 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3449 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3450 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3451 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3452 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3454 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3455 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3456 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3457 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3459 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3460 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3461 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3462 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3464 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3465 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3466 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3467 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3470 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3471 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3472 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3473 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3475 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3476 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3478 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3482 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3483 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3484 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3485 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3487 // Common patterns involving scalar load
3488 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3489 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3490 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3491 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3493 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3494 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3495 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3496 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3499 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3500 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3501 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3502 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3504 // Expecting a i16 load any extended to i32 value.
3505 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3506 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3507 [(set VR128:$dst, (IntId (bitconvert
3508 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3512 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3513 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3515 // Common patterns involving scalar load
3516 def : Pat<(int_x86_sse41_pmovsxbq
3517 (bitconvert (v4i32 (X86vzmovl
3518 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3519 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3521 def : Pat<(int_x86_sse41_pmovzxbq
3522 (bitconvert (v4i32 (X86vzmovl
3523 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3524 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3527 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3528 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3529 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3530 (ins VR128:$src1, i32i8imm:$src2),
3531 !strconcat(OpcodeStr,
3532 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3533 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3535 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3536 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3537 !strconcat(OpcodeStr,
3538 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3541 // There's an AssertZext in the way of writing the store pattern
3542 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3545 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3548 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3549 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3550 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3551 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3552 !strconcat(OpcodeStr,
3553 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3556 // There's an AssertZext in the way of writing the store pattern
3557 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3560 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3563 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3564 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3565 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3566 (ins VR128:$src1, i32i8imm:$src2),
3567 !strconcat(OpcodeStr,
3568 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3570 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3571 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3572 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3573 !strconcat(OpcodeStr,
3574 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3575 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3576 addr:$dst)]>, OpSize;
3579 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3582 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3584 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3585 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3586 (ins VR128:$src1, i32i8imm:$src2),
3587 !strconcat(OpcodeStr,
3588 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3590 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3592 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3593 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3594 !strconcat(OpcodeStr,
3595 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3596 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3597 addr:$dst)]>, OpSize;
3600 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3602 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3603 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3606 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3607 Requires<[HasSSE41]>;
3609 let Constraints = "$src1 = $dst" in {
3610 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3611 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3612 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3613 !strconcat(OpcodeStr,
3614 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3616 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3617 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3618 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3619 !strconcat(OpcodeStr,
3620 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3622 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3623 imm:$src3))]>, OpSize;
3627 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3629 let Constraints = "$src1 = $dst" in {
3630 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3631 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3632 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3633 !strconcat(OpcodeStr,
3634 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3636 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3638 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3639 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3640 !strconcat(OpcodeStr,
3641 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3643 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3644 imm:$src3)))]>, OpSize;
3648 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3650 // insertps has a few different modes, there's the first two here below which
3651 // are optimized inserts that won't zero arbitrary elements in the destination
3652 // vector. The next one matches the intrinsic and could zero arbitrary elements
3653 // in the target vector.
3654 let Constraints = "$src1 = $dst" in {
3655 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3656 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3657 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3658 !strconcat(OpcodeStr,
3659 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3661 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3663 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3664 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3665 !strconcat(OpcodeStr,
3666 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3668 (X86insrtps VR128:$src1,
3669 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3670 imm:$src3))]>, OpSize;
3674 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3676 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3677 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3679 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3680 // the intel intrinsic that corresponds to this.
3681 let Defs = [EFLAGS] in {
3682 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3683 "ptest \t{$src2, $src1|$src1, $src2}",
3684 [(set EFLAGS, (X86ptest VR128:$src1, VR128:$src2))]>,
3686 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3687 "ptest \t{$src2, $src1|$src1, $src2}",
3688 [(set EFLAGS, (X86ptest VR128:$src1, (load addr:$src2)))]>,
3692 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3693 "movntdqa\t{$src, $dst|$dst, $src}",
3694 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
3698 //===----------------------------------------------------------------------===//
3699 // SSE4.2 Instructions
3700 //===----------------------------------------------------------------------===//
3702 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3703 let Constraints = "$src1 = $dst" in {
3704 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3705 Intrinsic IntId128, bit Commutable = 0> {
3706 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3707 (ins VR128:$src1, VR128:$src2),
3708 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3709 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3711 let isCommutable = Commutable;
3713 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3714 (ins VR128:$src1, i128mem:$src2),
3715 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3717 (IntId128 VR128:$src1,
3718 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3722 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3724 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3725 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3726 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3727 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3729 // crc intrinsic instruction
3730 // This set of instructions are only rm, the only difference is the size
3732 let Constraints = "$src1 = $dst" in {
3733 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3734 (ins GR32:$src1, i8mem:$src2),
3735 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3737 (int_x86_sse42_crc32_8 GR32:$src1,
3738 (load addr:$src2)))]>;
3739 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3740 (ins GR32:$src1, GR8:$src2),
3741 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3743 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>;
3744 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3745 (ins GR32:$src1, i16mem:$src2),
3746 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3748 (int_x86_sse42_crc32_16 GR32:$src1,
3749 (load addr:$src2)))]>,
3751 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3752 (ins GR32:$src1, GR16:$src2),
3753 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3755 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3757 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3758 (ins GR32:$src1, i32mem:$src2),
3759 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3761 (int_x86_sse42_crc32_32 GR32:$src1,
3762 (load addr:$src2)))]>;
3763 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3764 (ins GR32:$src1, GR32:$src2),
3765 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3767 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>;
3768 def CRC64m8 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3769 (ins GR64:$src1, i8mem:$src2),
3770 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3772 (int_x86_sse42_crc64_8 GR64:$src1,
3773 (load addr:$src2)))]>,
3775 def CRC64r8 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3776 (ins GR64:$src1, GR8:$src2),
3777 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3779 (int_x86_sse42_crc64_8 GR64:$src1, GR8:$src2))]>,
3781 def CRC64m64 : SS42FI<0xF1, MRMSrcMem, (outs GR64:$dst),
3782 (ins GR64:$src1, i64mem:$src2),
3783 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3785 (int_x86_sse42_crc64_64 GR64:$src1,
3786 (load addr:$src2)))]>,
3788 def CRC64r64 : SS42FI<0xF1, MRMSrcReg, (outs GR64:$dst),
3789 (ins GR64:$src1, GR64:$src2),
3790 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3792 (int_x86_sse42_crc64_64 GR64:$src1, GR64:$src2))]>,
3796 // String/text processing instructions.
3797 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3798 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3799 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3800 "#PCMPISTRM128rr PSEUDO!",
3801 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3802 imm:$src3))]>, OpSize;
3803 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3804 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3805 "#PCMPISTRM128rm PSEUDO!",
3806 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3807 imm:$src3))]>, OpSize;
3810 let Defs = [XMM0, EFLAGS] in {
3811 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3812 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3813 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3814 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3815 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3816 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3819 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3820 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3821 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3822 "#PCMPESTRM128rr PSEUDO!",
3824 (int_x86_sse42_pcmpestrm128
3825 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3827 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3828 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3829 "#PCMPESTRM128rm PSEUDO!",
3830 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3831 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3835 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3836 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3837 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3838 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3839 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3840 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3841 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3844 let Defs = [ECX, EFLAGS] in {
3845 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3846 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3847 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3848 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3849 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3850 (implicit EFLAGS)]>, OpSize;
3851 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3852 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3853 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3854 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3855 (implicit EFLAGS)]>, OpSize;
3859 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3860 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3861 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3862 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3863 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3864 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3866 let Defs = [ECX, EFLAGS] in {
3867 let Uses = [EAX, EDX] in {
3868 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3869 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3870 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3871 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3872 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3873 (implicit EFLAGS)]>, OpSize;
3874 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3875 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3876 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3878 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
3879 (implicit EFLAGS)]>, OpSize;
3884 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3885 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3886 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3887 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3888 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3889 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;
3891 //===----------------------------------------------------------------------===//
3892 // AES-NI Instructions
3893 //===----------------------------------------------------------------------===//
3895 let Constraints = "$src1 = $dst" in {
3896 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr,
3897 Intrinsic IntId128, bit Commutable = 0> {
3898 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
3899 (ins VR128:$src1, VR128:$src2),
3900 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3901 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3903 let isCommutable = Commutable;
3905 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
3906 (ins VR128:$src1, i128mem:$src2),
3907 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3909 (IntId128 VR128:$src1,
3910 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3914 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
3915 int_x86_aesni_aesenc>;
3916 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
3917 int_x86_aesni_aesenclast>;
3918 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
3919 int_x86_aesni_aesdec>;
3920 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
3921 int_x86_aesni_aesdeclast>;
3923 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, VR128:$src2)),
3924 (AESENCrr VR128:$src1, VR128:$src2)>;
3925 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, (memop addr:$src2))),
3926 (AESENCrm VR128:$src1, addr:$src2)>;
3927 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, VR128:$src2)),
3928 (AESENCLASTrr VR128:$src1, VR128:$src2)>;
3929 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, (memop addr:$src2))),
3930 (AESENCLASTrm VR128:$src1, addr:$src2)>;
3931 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, VR128:$src2)),
3932 (AESDECrr VR128:$src1, VR128:$src2)>;
3933 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, (memop addr:$src2))),
3934 (AESDECrm VR128:$src1, addr:$src2)>;
3935 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, VR128:$src2)),
3936 (AESDECLASTrr VR128:$src1, VR128:$src2)>;
3937 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, (memop addr:$src2))),
3938 (AESDECLASTrm VR128:$src1, addr:$src2)>;
3940 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
3942 "aesimc\t{$src1, $dst|$dst, $src1}",
3944 (int_x86_aesni_aesimc VR128:$src1))]>,
3947 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
3948 (ins i128mem:$src1),
3949 "aesimc\t{$src1, $dst|$dst, $src1}",
3951 (int_x86_aesni_aesimc (bitconvert (memopv2i64 addr:$src1))))]>,
3954 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
3955 (ins VR128:$src1, i8imm:$src2),
3956 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3958 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
3960 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
3961 (ins i128mem:$src1, i8imm:$src2),
3962 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3964 (int_x86_aesni_aeskeygenassist (bitconvert (memopv2i64 addr:$src1)),