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, $src}",
842 SSEPackedSingle>, TB; /* PD SSE3 form is avaiable */
847 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
848 "cvtsd2ss\t{$src, $dst|$dst, $src}",
849 [(set FR32:$dst, (fround FR64:$src))]>;
850 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
851 "cvtsd2ss\t{$src, $dst|$dst, $src}",
852 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
853 Requires<[HasSSE2, OptForSize]>;
855 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
856 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
857 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
858 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
860 // SSE2 instructions with XS prefix
861 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
862 "cvtss2sd\t{$src, $dst|$dst, $src}",
863 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
865 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
866 "cvtss2sd\t{$src, $dst|$dst, $src}",
867 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
868 Requires<[HasSSE2, OptForSize]>;
870 def : Pat<(extloadf32 addr:$src),
871 (CVTSS2SDrr (MOVSSrm addr:$src))>,
872 Requires<[HasSSE2, OptForSpeed]>;
874 // SSE2 instructions without OpSize prefix
875 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
876 "cvtdq2ps\t{$src, $dst|$dst, $src}",
877 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
878 TB, Requires<[HasSSE2]>;
879 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
880 "cvtdq2ps\t{$src, $dst|$dst, $src}",
881 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
882 (bitconvert (memopv2i64 addr:$src))))]>,
883 TB, Requires<[HasSSE2]>;
885 // SSE2 instructions with XS prefix
886 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
887 "cvtdq2pd\t{$src, $dst|$dst, $src}",
888 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
889 XS, Requires<[HasSSE2]>;
890 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
891 "cvtdq2pd\t{$src, $dst|$dst, $src}",
892 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
893 (bitconvert (memopv2i64 addr:$src))))]>,
894 XS, Requires<[HasSSE2]>;
896 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
897 "cvtps2dq\t{$src, $dst|$dst, $src}",
898 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
899 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
900 "cvtps2dq\t{$src, $dst|$dst, $src}",
901 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
902 (memop addr:$src)))]>;
903 // SSE2 packed instructions with XS prefix
904 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
905 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
906 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
907 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
909 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
910 "cvttps2dq\t{$src, $dst|$dst, $src}",
912 (int_x86_sse2_cvttps2dq VR128:$src))]>,
913 XS, Requires<[HasSSE2]>;
914 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
915 "cvttps2dq\t{$src, $dst|$dst, $src}",
916 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
917 (memop addr:$src)))]>,
918 XS, Requires<[HasSSE2]>;
920 // SSE2 packed instructions with XD prefix
921 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
922 "cvtpd2dq\t{$src, $dst|$dst, $src}",
923 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
924 XD, Requires<[HasSSE2]>;
925 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
926 "cvtpd2dq\t{$src, $dst|$dst, $src}",
927 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
928 (memop addr:$src)))]>,
929 XD, Requires<[HasSSE2]>;
931 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
932 "cvttpd2dq\t{$src, $dst|$dst, $src}",
933 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
934 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
935 "cvttpd2dq\t{$src, $dst|$dst, $src}",
936 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
937 (memop addr:$src)))]>;
939 // SSE2 instructions without OpSize prefix
940 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
941 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
942 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
943 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
945 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
946 "cvtps2pd\t{$src, $dst|$dst, $src}",
947 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
948 TB, Requires<[HasSSE2]>;
949 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
950 "cvtps2pd\t{$src, $dst|$dst, $src}",
951 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
952 (load addr:$src)))]>,
953 TB, Requires<[HasSSE2]>;
955 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
956 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
957 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
958 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
961 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
962 "cvtpd2ps\t{$src, $dst|$dst, $src}",
963 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
964 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
965 "cvtpd2ps\t{$src, $dst|$dst, $src}",
966 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
967 (memop addr:$src)))]>;
969 // Match intrinsics which expect XMM operand(s).
970 // Aliases for intrinsics
971 let Constraints = "$src1 = $dst" in {
972 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
973 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
974 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
975 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
977 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
978 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
979 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
980 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
981 (load addr:$src2)))]>;
982 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
983 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
984 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
985 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
988 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
989 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
990 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
991 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
992 (load addr:$src2)))]>, XS,
996 //===----------------------------------------------------------------------===//
997 // SSE 1 & 2 - Compare Instructions
998 //===----------------------------------------------------------------------===//
1000 // sse12_cmp_scalar - sse 1 & 2 compare scalar instructions
1001 multiclass sse12_cmp_scalar<RegisterClass RC, X86MemOperand x86memop,
1002 string asm, string asm_alt> {
1003 def rr : SIi8<0xC2, MRMSrcReg,
1004 (outs RC:$dst), (ins RC:$src1, RC:$src, SSECC:$cc),
1007 def rm : SIi8<0xC2, MRMSrcMem,
1008 (outs RC:$dst), (ins RC:$src1, x86memop:$src, SSECC:$cc),
1010 // Accept explicit immediate argument form instead of comparison code.
1011 let isAsmParserOnly = 1 in {
1012 def rr_alt : SIi8<0xC2, MRMSrcReg,
1013 (outs RC:$dst), (ins RC:$src1, RC:$src, i8imm:$src2),
1016 def rm_alt : SIi8<0xC2, MRMSrcMem,
1017 (outs RC:$dst), (ins RC:$src1, x86memop:$src, i8imm:$src2),
1022 let neverHasSideEffects = 1, isAsmParserOnly = 1 in {
1023 defm VCMPSS : sse12_cmp_scalar<FR32, f32mem,
1024 "cmp${cc}ss\t{$src, $src1, $dst|$dst, $src1, $src}",
1025 "cmpss\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}">,
1027 defm VCMPSD : sse12_cmp_scalar<FR64, f64mem,
1028 "cmp${cc}sd\t{$src, $src1, $dst|$dst, $src1, $src}",
1029 "cmpsd\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}">,
1033 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1034 defm CMPSS : sse12_cmp_scalar<FR32, f32mem,
1035 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
1036 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}">, XS;
1037 defm CMPSD : sse12_cmp_scalar<FR64, f64mem,
1038 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1039 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}">, XD;
1042 multiclass sse12_cmp_scalar_int<RegisterClass RC, X86MemOperand x86memop,
1043 Intrinsic Int, string asm> {
1044 def rr : SIi8<0xC2, MRMSrcReg, (outs VR128:$dst),
1045 (ins VR128:$src1, VR128:$src, SSECC:$cc), asm,
1046 [(set VR128:$dst, (Int VR128:$src1,
1047 VR128:$src, imm:$cc))]>;
1048 def rm : SIi8<0xC2, MRMSrcMem, (outs VR128:$dst),
1049 (ins VR128:$src1, f32mem:$src, SSECC:$cc), asm,
1050 [(set VR128:$dst, (Int VR128:$src1,
1051 (load addr:$src), imm:$cc))]>;
1054 // Aliases to match intrinsics which expect XMM operand(s).
1055 let isAsmParserOnly = 1 in {
1056 defm Int_VCMPSS : sse12_cmp_scalar_int<VR128, f32mem, int_x86_sse_cmp_ss,
1057 "cmp${cc}ss\t{$src, $src1, $dst|$dst, $src1, $src}">,
1059 defm Int_VCMPSD : sse12_cmp_scalar_int<VR128, f64mem, int_x86_sse2_cmp_sd,
1060 "cmp${cc}sd\t{$src, $src1, $dst|$dst, $src1, $src}">,
1063 let Constraints = "$src1 = $dst" in {
1064 defm Int_CMPSS : sse12_cmp_scalar_int<VR128, f32mem, int_x86_sse_cmp_ss,
1065 "cmp${cc}ss\t{$src, $dst|$dst, $src}">, XS;
1066 defm Int_CMPSD : sse12_cmp_scalar_int<VR128, f64mem, int_x86_sse2_cmp_sd,
1067 "cmp${cc}sd\t{$src, $dst|$dst, $src}">, XD;
1071 // sse12_ord_cmp - Unordered/Ordered scalar fp compare and set EFLAGS
1072 multiclass sse12_ord_cmp<bits<8> opc, RegisterClass RC, SDNode OpNode,
1073 ValueType vt, X86MemOperand x86memop,
1074 PatFrag ld_frag, string OpcodeStr, Domain d> {
1075 def rr: PI<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
1076 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
1077 [(set EFLAGS, (OpNode (vt RC:$src1), RC:$src2))], d>;
1078 def rm: PI<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
1079 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
1080 [(set EFLAGS, (OpNode (vt RC:$src1),
1081 (ld_frag addr:$src2)))], d>;
1084 let Defs = [EFLAGS] in {
1085 let isAsmParserOnly = 1 in {
1086 defm VUCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
1087 "ucomiss", SSEPackedSingle>, VEX;
1088 defm VUCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
1089 "ucomisd", SSEPackedDouble>, OpSize, VEX;
1090 let Pattern = []<dag> in {
1091 defm VCOMISS : sse12_ord_cmp<0x2F, VR128, undef, v4f32, f128mem, load,
1092 "comiss", SSEPackedSingle>, VEX;
1093 defm VCOMISD : sse12_ord_cmp<0x2F, VR128, undef, v2f64, f128mem, load,
1094 "comisd", SSEPackedDouble>, OpSize, VEX;
1097 defm Int_VUCOMISS : sse12_ord_cmp<0x2E, VR128, X86ucomi, v4f32, f128mem,
1098 load, "ucomiss", SSEPackedSingle>, VEX;
1099 defm Int_VUCOMISD : sse12_ord_cmp<0x2E, VR128, X86ucomi, v2f64, f128mem,
1100 load, "ucomisd", SSEPackedDouble>, OpSize, VEX;
1102 defm Int_VCOMISS : sse12_ord_cmp<0x2F, VR128, X86comi, v4f32, f128mem,
1103 load, "comiss", SSEPackedSingle>, VEX;
1104 defm Int_VCOMISD : sse12_ord_cmp<0x2F, VR128, X86comi, v2f64, f128mem,
1105 load, "comisd", SSEPackedDouble>, OpSize, VEX;
1107 defm UCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
1108 "ucomiss", SSEPackedSingle>, TB;
1109 defm UCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
1110 "ucomisd", SSEPackedDouble>, TB, OpSize;
1112 let Pattern = []<dag> in {
1113 defm COMISS : sse12_ord_cmp<0x2F, VR128, undef, v4f32, f128mem, load,
1114 "comiss", SSEPackedSingle>, TB;
1115 defm COMISD : sse12_ord_cmp<0x2F, VR128, undef, v2f64, f128mem, load,
1116 "comisd", SSEPackedDouble>, TB, OpSize;
1119 defm Int_UCOMISS : sse12_ord_cmp<0x2E, VR128, X86ucomi, v4f32, f128mem,
1120 load, "ucomiss", SSEPackedSingle>, TB;
1121 defm Int_UCOMISD : sse12_ord_cmp<0x2E, VR128, X86ucomi, v2f64, f128mem,
1122 load, "ucomisd", SSEPackedDouble>, TB, OpSize;
1124 defm Int_COMISS : sse12_ord_cmp<0x2F, VR128, X86comi, v4f32, f128mem, load,
1125 "comiss", SSEPackedSingle>, TB;
1126 defm Int_COMISD : sse12_ord_cmp<0x2F, VR128, X86comi, v2f64, f128mem, load,
1127 "comisd", SSEPackedDouble>, TB, OpSize;
1128 } // Defs = [EFLAGS]
1130 // sse12_cmp_packed - sse 1 & 2 compared packed instructions
1131 multiclass sse12_cmp_packed<RegisterClass RC, X86MemOperand x86memop,
1132 Intrinsic Int, string asm, string asm_alt,
1134 def rri : PIi8<0xC2, MRMSrcReg,
1135 (outs RC:$dst), (ins RC:$src1, RC:$src, SSECC:$cc), asm,
1136 [(set RC:$dst, (Int RC:$src1, RC:$src, imm:$cc))], d>;
1137 def rmi : PIi8<0xC2, MRMSrcMem,
1138 (outs RC:$dst), (ins RC:$src1, f128mem:$src, SSECC:$cc), asm,
1139 [(set RC:$dst, (Int RC:$src1, (memop addr:$src), imm:$cc))], d>;
1140 // Accept explicit immediate argument form instead of comparison code.
1141 let isAsmParserOnly = 1 in {
1142 def rri_alt : PIi8<0xC2, MRMSrcReg,
1143 (outs RC:$dst), (ins RC:$src1, RC:$src, i8imm:$src2),
1145 def rmi_alt : PIi8<0xC2, MRMSrcMem,
1146 (outs RC:$dst), (ins RC:$src1, f128mem:$src, i8imm:$src2),
1151 let isAsmParserOnly = 1 in {
1152 defm VCMPPS : sse12_cmp_packed<VR128, f128mem, int_x86_sse_cmp_ps,
1153 "cmp${cc}ps\t{$src, $src1, $dst|$dst, $src1, $src}",
1154 "cmpps\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}",
1155 SSEPackedSingle>, VEX_4V;
1156 defm VCMPPD : sse12_cmp_packed<VR128, f128mem, int_x86_sse2_cmp_pd,
1157 "cmp${cc}pd\t{$src, $src1, $dst|$dst, $src1, $src}",
1158 "cmppd\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}",
1159 SSEPackedDouble>, OpSize, VEX_4V;
1161 let Constraints = "$src1 = $dst" in {
1162 defm CMPPS : sse12_cmp_packed<VR128, f128mem, int_x86_sse_cmp_ps,
1163 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1164 "cmpps\t{$src2, $src, $dst|$dst, $src, $src2}",
1165 SSEPackedSingle>, TB;
1166 defm CMPPD : sse12_cmp_packed<VR128, f128mem, int_x86_sse2_cmp_pd,
1167 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1168 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}",
1169 SSEPackedDouble>, TB, OpSize;
1172 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1173 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
1174 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1175 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
1176 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1177 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1178 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1179 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1181 //===----------------------------------------------------------------------===//
1182 // SSE 1 & 2 - Shuffle Instructions
1183 //===----------------------------------------------------------------------===//
1185 /// sse12_shuffle - sse 1 & 2 shuffle instructions
1186 multiclass sse12_shuffle<RegisterClass RC, X86MemOperand x86memop,
1187 ValueType vt, string asm, PatFrag mem_frag,
1188 Domain d, bit IsConvertibleToThreeAddress = 0> {
1189 def rmi : PIi8<0xC6, MRMSrcMem, (outs VR128:$dst),
1190 (ins VR128:$src1, f128mem:$src2, i8imm:$src3), asm,
1191 [(set VR128:$dst, (vt (shufp:$src3
1192 VR128:$src1, (mem_frag addr:$src2))))], d>;
1193 let isConvertibleToThreeAddress = IsConvertibleToThreeAddress in
1194 def rri : PIi8<0xC6, MRMSrcReg, (outs VR128:$dst),
1195 (ins VR128:$src1, VR128:$src2, i8imm:$src3), asm,
1197 (vt (shufp:$src3 VR128:$src1, VR128:$src2)))], d>;
1200 let isAsmParserOnly = 1 in {
1201 defm VSHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
1202 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
1203 memopv4f32, SSEPackedSingle>, VEX_4V;
1204 defm VSHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
1205 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src2, $src2, $src3}",
1206 memopv2f64, SSEPackedDouble>, OpSize, VEX_4V;
1209 let Constraints = "$src1 = $dst" in {
1210 defm SHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
1211 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1212 memopv4f32, SSEPackedSingle, 1 /* cvt to pshufd */>,
1214 defm SHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
1215 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1216 memopv2f64, SSEPackedDouble>, TB, OpSize;
1219 //===----------------------------------------------------------------------===//
1220 // SSE 1 & 2 - Unpack Instructions
1221 //===----------------------------------------------------------------------===//
1223 /// sse12_unpack_interleave - sse 1 & 2 unpack and interleave
1224 multiclass sse12_unpack_interleave<bits<8> opc, PatFrag OpNode, ValueType vt,
1225 PatFrag mem_frag, RegisterClass RC,
1226 X86MemOperand x86memop, string asm,
1228 def rr : PI<opc, MRMSrcReg,
1229 (outs RC:$dst), (ins RC:$src1, RC:$src2),
1231 (vt (OpNode RC:$src1, RC:$src2)))], d>;
1232 def rm : PI<opc, MRMSrcMem,
1233 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
1235 (vt (OpNode RC:$src1,
1236 (mem_frag addr:$src2))))], d>;
1239 let AddedComplexity = 10 in {
1240 let isAsmParserOnly = 1 in {
1241 defm VUNPCKHPS: sse12_unpack_interleave<0x15, unpckh, v4f32, memopv4f32,
1242 VR128, f128mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1243 SSEPackedSingle>, VEX_4V;
1244 defm VUNPCKHPD: sse12_unpack_interleave<0x15, unpckh, v2f64, memopv2f64,
1245 VR128, f128mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1246 SSEPackedDouble>, OpSize, VEX_4V;
1247 defm VUNPCKLPS: sse12_unpack_interleave<0x14, unpckl, v4f32, memopv4f32,
1248 VR128, f128mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1249 SSEPackedSingle>, VEX_4V;
1250 defm VUNPCKLPD: sse12_unpack_interleave<0x14, unpckl, v2f64, memopv2f64,
1251 VR128, f128mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1252 SSEPackedDouble>, OpSize, VEX_4V;
1255 let Constraints = "$src1 = $dst" in {
1256 defm UNPCKHPS: sse12_unpack_interleave<0x15, unpckh, v4f32, memopv4f32,
1257 VR128, f128mem, "unpckhps\t{$src2, $dst|$dst, $src2}",
1258 SSEPackedSingle>, TB;
1259 defm UNPCKHPD: sse12_unpack_interleave<0x15, unpckh, v2f64, memopv2f64,
1260 VR128, f128mem, "unpckhpd\t{$src2, $dst|$dst, $src2}",
1261 SSEPackedDouble>, TB, OpSize;
1262 defm UNPCKLPS: sse12_unpack_interleave<0x14, unpckl, v4f32, memopv4f32,
1263 VR128, f128mem, "unpcklps\t{$src2, $dst|$dst, $src2}",
1264 SSEPackedSingle>, TB;
1265 defm UNPCKLPD: sse12_unpack_interleave<0x14, unpckl, v2f64, memopv2f64,
1266 VR128, f128mem, "unpcklpd\t{$src2, $dst|$dst, $src2}",
1267 SSEPackedDouble>, TB, OpSize;
1268 } // Constraints = "$src1 = $dst"
1269 } // AddedComplexity
1271 //===----------------------------------------------------------------------===//
1272 // SSE 1 & 2 - Extract Floating-Point Sign mask
1273 //===----------------------------------------------------------------------===//
1275 /// sse12_extr_sign_mask - sse 1 & 2 unpack and interleave
1276 multiclass sse12_extr_sign_mask<RegisterClass RC, Intrinsic Int, string asm,
1278 def rr : PI<0x50, MRMSrcReg, (outs GR32:$dst), (ins RC:$src),
1279 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1280 [(set GR32:$dst, (Int RC:$src))], d>;
1284 defm MOVMSKPS : sse12_extr_sign_mask<VR128, int_x86_sse_movmsk_ps, "movmskps",
1285 SSEPackedSingle>, TB;
1286 defm MOVMSKPD : sse12_extr_sign_mask<VR128, int_x86_sse2_movmsk_pd, "movmskpd",
1287 SSEPackedDouble>, TB, OpSize;
1289 let isAsmParserOnly = 1 in {
1290 defm VMOVMSKPS : sse12_extr_sign_mask<VR128, int_x86_sse_movmsk_ps,
1291 "movmskps", SSEPackedSingle>, VEX;
1292 defm VMOVMSKPD : sse12_extr_sign_mask<VR128, int_x86_sse2_movmsk_pd,
1293 "movmskpd", SSEPackedDouble>, OpSize,
1297 //===----------------------------------------------------------------------===//
1298 // SSE 1 & 2 - Misc aliasing of packed SSE 1 & 2 instructions
1299 //===----------------------------------------------------------------------===//
1301 // Aliases of packed SSE1 & SSE2 instructions for scalar use. These all have
1302 // names that start with 'Fs'.
1304 // Alias instructions that map fld0 to pxor for sse.
1305 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1306 canFoldAsLoad = 1 in {
1307 // FIXME: Set encoding to pseudo!
1308 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
1309 [(set FR32:$dst, fp32imm0)]>,
1310 Requires<[HasSSE1]>, TB, OpSize;
1311 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1312 [(set FR64:$dst, fpimm0)]>,
1313 Requires<[HasSSE2]>, TB, OpSize;
1316 // Alias instruction to do FR32 or FR64 reg-to-reg copy using movaps. Upper
1317 // bits are disregarded.
1318 let neverHasSideEffects = 1 in {
1319 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
1320 "movaps\t{$src, $dst|$dst, $src}", []>;
1321 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1322 "movapd\t{$src, $dst|$dst, $src}", []>;
1325 // Alias instruction to load FR32 or FR64 from f128mem using movaps. Upper
1326 // bits are disregarded.
1327 let canFoldAsLoad = 1, isReMaterializable = 1 in {
1328 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
1329 "movaps\t{$src, $dst|$dst, $src}",
1330 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
1331 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1332 "movapd\t{$src, $dst|$dst, $src}",
1333 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1336 //===----------------------------------------------------------------------===//
1337 // SSE 1 & 2 - Logical Instructions
1338 //===----------------------------------------------------------------------===//
1340 /// sse12_fp_alias_pack_logical - SSE 1 & 2 aliased packed FP logical ops
1342 multiclass sse12_fp_alias_pack_logical<bits<8> opc, string OpcodeStr,
1343 SDNode OpNode, bit MayLoad = 0> {
1344 let isAsmParserOnly = 1 in {
1345 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1346 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode, FR32,
1347 f32, f128mem, memopfsf32, SSEPackedSingle, MayLoad>, VEX_4V;
1349 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1350 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode, FR64,
1351 f64, f128mem, memopfsf64, SSEPackedDouble, MayLoad>, OpSize,
1355 let Constraints = "$src1 = $dst" in {
1356 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1357 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, FR32, f32,
1358 f128mem, memopfsf32, SSEPackedSingle, MayLoad>, TB;
1360 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1361 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, FR64, f64,
1362 f128mem, memopfsf64, SSEPackedDouble, MayLoad>, TB, OpSize;
1366 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1367 defm FsAND : sse12_fp_alias_pack_logical<0x54, "and", X86fand>;
1368 defm FsOR : sse12_fp_alias_pack_logical<0x56, "or", X86for>;
1369 defm FsXOR : sse12_fp_alias_pack_logical<0x57, "xor", X86fxor>;
1371 let neverHasSideEffects = 1, Pattern = []<dag>, isCommutable = 0 in
1372 defm FsANDN : sse12_fp_alias_pack_logical<0x55, "andn", undef, 1>;
1374 /// sse12_fp_packed_logical - SSE 1 & 2 packed FP logical ops
1376 multiclass sse12_fp_packed_logical<bits<8> opc, string OpcodeStr,
1377 SDNode OpNode, int HasPat = 0,
1378 list<list<dag>> Pattern = []> {
1379 let isAsmParserOnly = 1 in {
1380 defm V#NAME#PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
1381 !strconcat(OpcodeStr, "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1383 !if(HasPat, Pattern[0], // rr
1384 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1,
1386 !if(HasPat, Pattern[2], // rm
1387 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
1388 (memopv2i64 addr:$src2)))])>,
1391 defm V#NAME#PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
1392 !strconcat(OpcodeStr, "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1394 !if(HasPat, Pattern[1], // rr
1395 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1398 !if(HasPat, Pattern[3], // rm
1399 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1400 (memopv2i64 addr:$src2)))])>,
1403 let Constraints = "$src1 = $dst" in {
1404 defm PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
1405 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"), f128mem,
1406 !if(HasPat, Pattern[0], // rr
1407 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1,
1409 !if(HasPat, Pattern[2], // rm
1410 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
1411 (memopv2i64 addr:$src2)))])>, TB;
1413 defm PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
1414 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"), f128mem,
1415 !if(HasPat, Pattern[1], // rr
1416 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1419 !if(HasPat, Pattern[3], // rm
1420 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1421 (memopv2i64 addr:$src2)))])>,
1426 defm AND : sse12_fp_packed_logical<0x54, "and", and>;
1427 defm OR : sse12_fp_packed_logical<0x56, "or", or>;
1428 defm XOR : sse12_fp_packed_logical<0x57, "xor", xor>;
1429 let isCommutable = 0 in
1430 defm ANDN : sse12_fp_packed_logical<0x55, "andn", undef /* dummy */, 1, [
1432 [(set VR128:$dst, (v2i64 (and (xor VR128:$src1,
1433 (bc_v2i64 (v4i32 immAllOnesV))),
1436 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1437 (bc_v2i64 (v2f64 VR128:$src2))))],
1439 [(set VR128:$dst, (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
1440 (bc_v2i64 (v4i32 immAllOnesV))),
1441 (memopv2i64 addr:$src2))))],
1443 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1444 (memopv2i64 addr:$src2)))]]>;
1446 //===----------------------------------------------------------------------===//
1447 // SSE 1 & 2 - Arithmetic Instructions
1448 //===----------------------------------------------------------------------===//
1450 /// basic_sse12_fp_binop_rm - SSE 1 & 2 binops come in both scalar and
1453 /// In addition, we also have a special variant of the scalar form here to
1454 /// represent the associated intrinsic operation. This form is unlike the
1455 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1456 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1458 /// These three forms can each be reg+reg or reg+mem.
1460 multiclass basic_sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
1463 let isAsmParserOnly = 1 in {
1464 defm V#NAME#SS : sse12_fp_scalar<opc,
1465 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1466 OpNode, FR32, f32mem>, XS, VEX_4V;
1468 defm V#NAME#SD : sse12_fp_scalar<opc,
1469 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1470 OpNode, FR64, f64mem>, XD, VEX_4V;
1472 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1473 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
1474 VR128, v4f32, f128mem, memopv4f32, SSEPackedSingle>,
1477 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1478 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
1479 VR128, v2f64, f128mem, memopv2f64, SSEPackedDouble>,
1482 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1483 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1484 "", "_ss", ssmem, sse_load_f32>, XS, VEX_4V;
1486 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1487 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1488 "2", "_sd", sdmem, sse_load_f64>, XD, VEX_4V;
1491 let Constraints = "$src1 = $dst" in {
1492 defm SS : sse12_fp_scalar<opc,
1493 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
1494 OpNode, FR32, f32mem>, XS;
1496 defm SD : sse12_fp_scalar<opc,
1497 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1498 OpNode, FR64, f64mem>, XD;
1500 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1501 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v4f32,
1502 f128mem, memopv4f32, SSEPackedSingle>, TB;
1504 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1505 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v2f64,
1506 f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
1508 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1509 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
1510 "", "_ss", ssmem, sse_load_f32>, XS;
1512 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1513 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1514 "2", "_sd", sdmem, sse_load_f64>, XD;
1518 // Arithmetic instructions
1519 defm ADD : basic_sse12_fp_binop_rm<0x58, "add", fadd>;
1520 defm MUL : basic_sse12_fp_binop_rm<0x59, "mul", fmul>;
1522 let isCommutable = 0 in {
1523 defm SUB : basic_sse12_fp_binop_rm<0x5C, "sub", fsub>;
1524 defm DIV : basic_sse12_fp_binop_rm<0x5E, "div", fdiv>;
1527 /// sse12_fp_binop_rm - Other SSE 1 & 2 binops
1529 /// This multiclass is like basic_sse12_fp_binop_rm, with the addition of
1530 /// instructions for a full-vector intrinsic form. Operations that map
1531 /// onto C operators don't use this form since they just use the plain
1532 /// vector form instead of having a separate vector intrinsic form.
1534 multiclass sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
1537 let isAsmParserOnly = 1 in {
1538 // Scalar operation, reg+reg.
1539 defm V#NAME#SS : sse12_fp_scalar<opc,
1540 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1541 OpNode, FR32, f32mem>, XS, VEX_4V;
1543 defm V#NAME#SD : sse12_fp_scalar<opc,
1544 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1545 OpNode, FR64, f64mem>, XD, VEX_4V;
1547 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1548 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
1549 VR128, v4f32, f128mem, memopv4f32, SSEPackedSingle>,
1552 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1553 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
1554 VR128, v2f64, f128mem, memopv2f64, SSEPackedDouble>,
1557 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1558 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1559 "", "_ss", ssmem, sse_load_f32>, XS, VEX_4V;
1561 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1562 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1563 "2", "_sd", sdmem, sse_load_f64>, XD, VEX_4V;
1565 defm V#NAME#PS : sse12_fp_packed_int<opc, OpcodeStr, VR128,
1566 !strconcat(OpcodeStr, "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1567 "", "_ps", f128mem, memopv4f32, SSEPackedSingle>, VEX_4V;
1569 defm V#NAME#PD : sse12_fp_packed_int<opc, OpcodeStr, VR128,
1570 !strconcat(OpcodeStr, "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1571 "2", "_pd", f128mem, memopv2f64, SSEPackedDouble>, OpSize,
1575 let Constraints = "$src1 = $dst" in {
1576 // Scalar operation, reg+reg.
1577 defm SS : sse12_fp_scalar<opc,
1578 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
1579 OpNode, FR32, f32mem>, XS;
1580 defm SD : sse12_fp_scalar<opc,
1581 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1582 OpNode, FR64, f64mem>, XD;
1583 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1584 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v4f32,
1585 f128mem, memopv4f32, SSEPackedSingle>, TB;
1587 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
1588 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v2f64,
1589 f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
1591 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1592 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
1593 "", "_ss", ssmem, sse_load_f32>, XS;
1595 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1596 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
1597 "2", "_sd", sdmem, sse_load_f64>, XD;
1599 defm PS : sse12_fp_packed_int<opc, OpcodeStr, VR128,
1600 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
1601 "", "_ps", f128mem, memopv4f32, SSEPackedSingle>, TB;
1603 defm PD : sse12_fp_packed_int<opc, OpcodeStr, VR128,
1604 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1605 "2", "_pd", f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
1609 let isCommutable = 0 in {
1610 defm MAX : sse12_fp_binop_rm<0x5F, "max", X86fmax>;
1611 defm MIN : sse12_fp_binop_rm<0x5D, "min", X86fmin>;
1616 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
1618 /// In addition, we also have a special variant of the scalar form here to
1619 /// represent the associated intrinsic operation. This form is unlike the
1620 /// plain scalar form, in that it takes an entire vector (instead of a
1621 /// scalar) and leaves the top elements undefined.
1623 /// And, we have a special variant form for a full-vector intrinsic form.
1625 /// These four forms can each have a reg or a mem operand, so there are a
1626 /// total of eight "instructions".
1628 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
1632 bit Commutable = 0> {
1633 // Scalar operation, reg.
1634 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
1635 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1636 [(set FR32:$dst, (OpNode FR32:$src))]> {
1637 let isCommutable = Commutable;
1640 // Scalar operation, mem.
1641 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
1642 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1643 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
1644 Requires<[HasSSE1, OptForSize]>;
1646 // Vector operation, reg.
1647 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1648 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1649 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
1650 let isCommutable = Commutable;
1653 // Vector operation, mem.
1654 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1655 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1656 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
1658 // Intrinsic operation, reg.
1659 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1660 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1661 [(set VR128:$dst, (F32Int VR128:$src))]> {
1662 let isCommutable = Commutable;
1665 // Intrinsic operation, mem.
1666 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
1667 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1668 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
1670 // Vector intrinsic operation, reg
1671 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1672 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1673 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
1674 let isCommutable = Commutable;
1677 // Vector intrinsic operation, mem
1678 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1679 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1680 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
1684 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
1685 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
1687 // Reciprocal approximations. Note that these typically require refinement
1688 // in order to obtain suitable precision.
1689 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
1690 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
1691 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
1692 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
1694 // Prefetch intrinsic.
1695 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1696 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1697 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1698 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1699 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1700 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1701 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1702 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1704 // Non-temporal stores
1705 def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1706 "movntps\t{$src, $dst|$dst, $src}",
1707 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1709 let AddedComplexity = 400 in { // Prefer non-temporal versions
1710 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1711 "movntps\t{$src, $dst|$dst, $src}",
1712 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
1714 def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1715 "movntdq\t{$src, $dst|$dst, $src}",
1716 [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
1718 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
1719 "movnti\t{$src, $dst|$dst, $src}",
1720 [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
1721 TB, Requires<[HasSSE2]>;
1723 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
1724 "movnti\t{$src, $dst|$dst, $src}",
1725 [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
1726 TB, Requires<[HasSSE2]>;
1729 // Load, store, and memory fence
1730 def SFENCE : I<0xAE, MRM_F8, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>,
1731 TB, Requires<[HasSSE1]>;
1734 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1735 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1736 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1737 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1739 // Alias instructions that map zero vector to pxor / xorp* for sse.
1740 // We set canFoldAsLoad because this can be converted to a constant-pool
1741 // load of an all-zeros value if folding it would be beneficial.
1742 // FIXME: Change encoding to pseudo!
1743 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1744 isCodeGenOnly = 1 in {
1745 def V_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1746 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
1747 def V_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1748 [(set VR128:$dst, (v2f64 immAllZerosV))]>;
1749 let ExeDomain = SSEPackedInt in
1750 def V_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
1751 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1754 def : Pat<(v2i64 immAllZerosV), (V_SET0PI)>;
1755 def : Pat<(v8i16 immAllZerosV), (V_SET0PI)>;
1756 def : Pat<(v16i8 immAllZerosV), (V_SET0PI)>;
1758 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1759 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
1761 //===---------------------------------------------------------------------===//
1762 // SSE2 Instructions
1763 //===---------------------------------------------------------------------===//
1767 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1769 /// In addition, we also have a special variant of the scalar form here to
1770 /// represent the associated intrinsic operation. This form is unlike the
1771 /// plain scalar form, in that it takes an entire vector (instead of a
1772 /// scalar) and leaves the top elements undefined.
1774 /// And, we have a special variant form for a full-vector intrinsic form.
1776 /// These four forms can each have a reg or a mem operand, so there are a
1777 /// total of eight "instructions".
1779 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1783 bit Commutable = 0> {
1784 // Scalar operation, reg.
1785 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1786 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1787 [(set FR64:$dst, (OpNode FR64:$src))]> {
1788 let isCommutable = Commutable;
1791 // Scalar operation, mem.
1792 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1793 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1794 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1796 // Vector operation, reg.
1797 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1798 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1799 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1800 let isCommutable = Commutable;
1803 // Vector operation, mem.
1804 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1805 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1806 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1808 // Intrinsic operation, reg.
1809 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1810 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1811 [(set VR128:$dst, (F64Int VR128:$src))]> {
1812 let isCommutable = Commutable;
1815 // Intrinsic operation, mem.
1816 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1817 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1818 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1820 // Vector intrinsic operation, reg
1821 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1822 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1823 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1824 let isCommutable = Commutable;
1827 // Vector intrinsic operation, mem
1828 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1829 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1830 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1834 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1835 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1837 // There is no f64 version of the reciprocal approximation instructions.
1839 //===---------------------------------------------------------------------===//
1840 // SSE integer instructions
1841 let ExeDomain = SSEPackedInt in {
1843 // Move Instructions
1844 let neverHasSideEffects = 1 in
1845 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1846 "movdqa\t{$src, $dst|$dst, $src}", []>;
1847 let canFoldAsLoad = 1, mayLoad = 1 in
1848 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1849 "movdqa\t{$src, $dst|$dst, $src}",
1850 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1852 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1853 "movdqa\t{$src, $dst|$dst, $src}",
1854 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1855 let canFoldAsLoad = 1, mayLoad = 1 in
1856 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1857 "movdqu\t{$src, $dst|$dst, $src}",
1858 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1859 XS, Requires<[HasSSE2]>;
1861 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1862 "movdqu\t{$src, $dst|$dst, $src}",
1863 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1864 XS, Requires<[HasSSE2]>;
1866 // Intrinsic forms of MOVDQU load and store
1867 let canFoldAsLoad = 1 in
1868 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1869 "movdqu\t{$src, $dst|$dst, $src}",
1870 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1871 XS, Requires<[HasSSE2]>;
1872 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1873 "movdqu\t{$src, $dst|$dst, $src}",
1874 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1875 XS, Requires<[HasSSE2]>;
1877 let Constraints = "$src1 = $dst" in {
1879 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1880 bit Commutable = 0> {
1881 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1882 (ins VR128:$src1, VR128:$src2),
1883 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1884 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1885 let isCommutable = Commutable;
1887 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1888 (ins VR128:$src1, i128mem:$src2),
1889 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1890 [(set VR128:$dst, (IntId VR128:$src1,
1891 (bitconvert (memopv2i64
1895 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1897 Intrinsic IntId, Intrinsic IntId2> {
1898 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1899 (ins VR128:$src1, VR128:$src2),
1900 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1901 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1902 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1903 (ins VR128:$src1, i128mem:$src2),
1904 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1905 [(set VR128:$dst, (IntId VR128:$src1,
1906 (bitconvert (memopv2i64 addr:$src2))))]>;
1907 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
1908 (ins VR128:$src1, i32i8imm:$src2),
1909 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1910 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1913 /// PDI_binop_rm - Simple SSE2 binary operator.
1914 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1915 ValueType OpVT, bit Commutable = 0> {
1916 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1917 (ins VR128:$src1, VR128:$src2),
1918 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1919 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1920 let isCommutable = Commutable;
1922 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1923 (ins VR128:$src1, i128mem:$src2),
1924 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1925 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1926 (bitconvert (memopv2i64 addr:$src2)))))]>;
1929 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1931 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1932 /// to collapse (bitconvert VT to VT) into its operand.
1934 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1935 bit Commutable = 0> {
1936 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1937 (ins VR128:$src1, VR128:$src2),
1938 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1939 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1940 let isCommutable = Commutable;
1942 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1943 (ins VR128:$src1, i128mem:$src2),
1944 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1945 [(set VR128:$dst, (OpNode VR128:$src1,
1946 (memopv2i64 addr:$src2)))]>;
1949 } // Constraints = "$src1 = $dst"
1950 } // ExeDomain = SSEPackedInt
1952 // 128-bit Integer Arithmetic
1954 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1955 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1956 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1957 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1959 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1960 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1961 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1962 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1964 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1965 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1966 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1967 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1969 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1970 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1971 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1972 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1974 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1976 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1977 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1978 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1980 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1982 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1983 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1986 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1987 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1988 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1989 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1990 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
1993 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
1994 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
1995 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
1996 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
1997 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
1998 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
2000 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2001 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2002 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2003 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2004 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2005 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2007 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2008 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2009 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2010 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2012 // 128-bit logical shifts.
2013 let Constraints = "$src1 = $dst", neverHasSideEffects = 1,
2014 ExeDomain = SSEPackedInt in {
2015 def PSLLDQri : PDIi8<0x73, MRM7r,
2016 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2017 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2018 def PSRLDQri : PDIi8<0x73, MRM3r,
2019 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2020 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2021 // PSRADQri doesn't exist in SSE[1-3].
2024 let Predicates = [HasSSE2] in {
2025 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2026 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2027 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2028 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2029 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2030 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2031 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2032 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2033 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2034 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2036 // Shift up / down and insert zero's.
2037 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2038 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2039 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2040 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2044 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2045 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2046 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2048 let Constraints = "$src1 = $dst", ExeDomain = SSEPackedInt in {
2049 def PANDNrr : PDI<0xDF, MRMSrcReg,
2050 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2051 "pandn\t{$src2, $dst|$dst, $src2}",
2052 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2055 def PANDNrm : PDI<0xDF, MRMSrcMem,
2056 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2057 "pandn\t{$src2, $dst|$dst, $src2}",
2058 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2059 (memopv2i64 addr:$src2))))]>;
2062 // SSE2 Integer comparison
2063 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b, 1>;
2064 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w, 1>;
2065 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d, 1>;
2066 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2067 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2068 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2070 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2071 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2072 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2073 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2074 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2075 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2076 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2077 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2078 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2079 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2080 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2081 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2083 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2084 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2085 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2086 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2087 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2088 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2089 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2090 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2091 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2092 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2093 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2094 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2097 // Pack instructions
2098 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2099 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2100 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2102 let ExeDomain = SSEPackedInt in {
2104 // Shuffle and unpack instructions
2105 let AddedComplexity = 5 in {
2106 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2107 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2108 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2109 [(set VR128:$dst, (v4i32 (pshufd:$src2
2110 VR128:$src1, (undef))))]>;
2111 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2112 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2113 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2114 [(set VR128:$dst, (v4i32 (pshufd:$src2
2115 (bc_v4i32 (memopv2i64 addr:$src1)),
2119 // SSE2 with ImmT == Imm8 and XS prefix.
2120 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2121 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2122 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2123 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2125 XS, Requires<[HasSSE2]>;
2126 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2127 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2128 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2129 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2130 (bc_v8i16 (memopv2i64 addr:$src1)),
2132 XS, Requires<[HasSSE2]>;
2134 // SSE2 with ImmT == Imm8 and XD prefix.
2135 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2136 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2137 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2138 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2140 XD, Requires<[HasSSE2]>;
2141 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2142 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2143 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2144 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2145 (bc_v8i16 (memopv2i64 addr:$src1)),
2147 XD, Requires<[HasSSE2]>;
2149 // Unpack instructions
2150 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
2151 PatFrag unp_frag, PatFrag bc_frag> {
2152 def rr : PDI<opc, MRMSrcReg,
2153 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2154 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2155 [(set VR128:$dst, (vt (unp_frag VR128:$src1, VR128:$src2)))]>;
2156 def rm : PDI<opc, MRMSrcMem,
2157 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2158 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2159 [(set VR128:$dst, (unp_frag VR128:$src1,
2160 (bc_frag (memopv2i64
2164 let Constraints = "$src1 = $dst" in {
2165 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, unpckl, bc_v16i8>;
2166 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, unpckl, bc_v8i16>;
2167 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, unpckl, bc_v4i32>;
2169 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2170 /// knew to collapse (bitconvert VT to VT) into its operand.
2171 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2172 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2173 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2175 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2176 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2177 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2178 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2180 (v2i64 (unpckl VR128:$src1,
2181 (memopv2i64 addr:$src2))))]>;
2183 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, unpckh, bc_v16i8>;
2184 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, unpckh, bc_v8i16>;
2185 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, unpckh, bc_v4i32>;
2187 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2188 /// knew to collapse (bitconvert VT to VT) into its operand.
2189 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2190 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2191 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2193 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2194 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2195 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2196 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2198 (v2i64 (unpckh VR128:$src1,
2199 (memopv2i64 addr:$src2))))]>;
2203 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2204 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2205 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2206 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2208 let Constraints = "$src1 = $dst" in {
2209 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2210 (outs VR128:$dst), (ins VR128:$src1,
2211 GR32:$src2, i32i8imm:$src3),
2212 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2214 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2215 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2216 (outs VR128:$dst), (ins VR128:$src1,
2217 i16mem:$src2, i32i8imm:$src3),
2218 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2220 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2225 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2226 "pmovmskb\t{$src, $dst|$dst, $src}",
2227 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2229 // Conditional store
2231 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2232 "maskmovdqu\t{$mask, $src|$src, $mask}",
2233 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2236 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2237 "maskmovdqu\t{$mask, $src|$src, $mask}",
2238 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2240 } // ExeDomain = SSEPackedInt
2242 // Non-temporal stores
2243 def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2244 "movntpd\t{$src, $dst|$dst, $src}",
2245 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2246 let ExeDomain = SSEPackedInt in
2247 def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2248 "movntdq\t{$src, $dst|$dst, $src}",
2249 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2250 def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2251 "movnti\t{$src, $dst|$dst, $src}",
2252 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2253 TB, Requires<[HasSSE2]>;
2255 let AddedComplexity = 400 in { // Prefer non-temporal versions
2256 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2257 "movntpd\t{$src, $dst|$dst, $src}",
2258 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
2260 let ExeDomain = SSEPackedInt in
2261 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2262 "movntdq\t{$src, $dst|$dst, $src}",
2263 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2267 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2268 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2269 TB, Requires<[HasSSE2]>;
2271 // Load, store, and memory fence
2272 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
2273 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2274 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
2275 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2277 // Pause. This "instruction" is encoded as "rep; nop", so even though it
2278 // was introduced with SSE2, it's backward compatible.
2279 def PAUSE : I<0x90, RawFrm, (outs), (ins), "pause", []>, REP;
2281 //TODO: custom lower this so as to never even generate the noop
2282 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2284 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2285 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2286 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2289 // Alias instructions that map zero vector to pxor / xorp* for sse.
2290 // We set canFoldAsLoad because this can be converted to a constant-pool
2291 // load of an all-ones value if folding it would be beneficial.
2292 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2293 isCodeGenOnly = 1, ExeDomain = SSEPackedInt in
2294 // FIXME: Change encoding to pseudo.
2295 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
2296 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2298 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2299 "movd\t{$src, $dst|$dst, $src}",
2301 (v4i32 (scalar_to_vector GR32:$src)))]>;
2302 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2303 "movd\t{$src, $dst|$dst, $src}",
2305 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2307 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2308 "movd\t{$src, $dst|$dst, $src}",
2309 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2311 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2312 "movd\t{$src, $dst|$dst, $src}",
2313 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2315 // SSE2 instructions with XS prefix
2316 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2317 "movq\t{$src, $dst|$dst, $src}",
2319 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2320 Requires<[HasSSE2]>;
2321 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2322 "movq\t{$src, $dst|$dst, $src}",
2323 [(store (i64 (vector_extract (v2i64 VR128:$src),
2324 (iPTR 0))), addr:$dst)]>;
2326 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
2327 (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
2329 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2330 "movd\t{$src, $dst|$dst, $src}",
2331 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2333 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2334 "movd\t{$src, $dst|$dst, $src}",
2335 [(store (i32 (vector_extract (v4i32 VR128:$src),
2336 (iPTR 0))), addr:$dst)]>;
2338 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2339 "movd\t{$src, $dst|$dst, $src}",
2340 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2341 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2342 "movd\t{$src, $dst|$dst, $src}",
2343 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2345 // Store / copy lower 64-bits of a XMM register.
2346 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2347 "movq\t{$src, $dst|$dst, $src}",
2348 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2350 // movd / movq to XMM register zero-extends
2351 let AddedComplexity = 15 in {
2352 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2353 "movd\t{$src, $dst|$dst, $src}",
2354 [(set VR128:$dst, (v4i32 (X86vzmovl
2355 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2356 // This is X86-64 only.
2357 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2358 "mov{d|q}\t{$src, $dst|$dst, $src}",
2359 [(set VR128:$dst, (v2i64 (X86vzmovl
2360 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2363 let AddedComplexity = 20 in {
2364 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2365 "movd\t{$src, $dst|$dst, $src}",
2367 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2368 (loadi32 addr:$src))))))]>;
2370 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2371 (MOVZDI2PDIrm addr:$src)>;
2372 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2373 (MOVZDI2PDIrm addr:$src)>;
2374 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2375 (MOVZDI2PDIrm addr:$src)>;
2377 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2378 "movq\t{$src, $dst|$dst, $src}",
2380 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2381 (loadi64 addr:$src))))))]>, XS,
2382 Requires<[HasSSE2]>;
2384 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2385 (MOVZQI2PQIrm addr:$src)>;
2386 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2387 (MOVZQI2PQIrm addr:$src)>;
2388 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2391 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2392 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2393 let AddedComplexity = 15 in
2394 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2395 "movq\t{$src, $dst|$dst, $src}",
2396 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2397 XS, Requires<[HasSSE2]>;
2399 let AddedComplexity = 20 in {
2400 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2401 "movq\t{$src, $dst|$dst, $src}",
2402 [(set VR128:$dst, (v2i64 (X86vzmovl
2403 (loadv2i64 addr:$src))))]>,
2404 XS, Requires<[HasSSE2]>;
2406 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2407 (MOVZPQILo2PQIrm addr:$src)>;
2410 // Instructions for the disassembler
2411 // xr = XMM register
2414 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2415 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2417 //===---------------------------------------------------------------------===//
2418 // SSE3 Instructions
2419 //===---------------------------------------------------------------------===//
2421 // Conversion Instructions
2422 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2423 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
2424 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2425 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
2426 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2427 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
2428 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2429 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
2431 // Move Instructions
2432 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2433 "movshdup\t{$src, $dst|$dst, $src}",
2434 [(set VR128:$dst, (v4f32 (movshdup
2435 VR128:$src, (undef))))]>;
2436 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2437 "movshdup\t{$src, $dst|$dst, $src}",
2438 [(set VR128:$dst, (movshdup
2439 (memopv4f32 addr:$src), (undef)))]>;
2441 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2442 "movsldup\t{$src, $dst|$dst, $src}",
2443 [(set VR128:$dst, (v4f32 (movsldup
2444 VR128:$src, (undef))))]>;
2445 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2446 "movsldup\t{$src, $dst|$dst, $src}",
2447 [(set VR128:$dst, (movsldup
2448 (memopv4f32 addr:$src), (undef)))]>;
2450 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2451 "movddup\t{$src, $dst|$dst, $src}",
2452 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2453 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2454 "movddup\t{$src, $dst|$dst, $src}",
2456 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2459 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2461 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2463 let AddedComplexity = 5 in {
2464 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2465 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2466 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2467 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2468 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2469 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2470 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2471 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2475 let Constraints = "$src1 = $dst" in {
2476 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2477 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2478 "addsubps\t{$src2, $dst|$dst, $src2}",
2479 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2481 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2482 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2483 "addsubps\t{$src2, $dst|$dst, $src2}",
2484 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2485 (memop addr:$src2)))]>;
2486 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2487 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2488 "addsubpd\t{$src2, $dst|$dst, $src2}",
2489 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2491 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2492 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2493 "addsubpd\t{$src2, $dst|$dst, $src2}",
2494 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2495 (memop addr:$src2)))]>;
2498 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2499 "lddqu\t{$src, $dst|$dst, $src}",
2500 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2503 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2504 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2505 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2506 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2507 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2508 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2509 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2510 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2511 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2512 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2513 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2514 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2515 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2516 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2517 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2518 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2520 let Constraints = "$src1 = $dst" in {
2521 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2522 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2523 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2524 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2525 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2526 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2527 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2528 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2531 // Thread synchronization
2532 def MONITOR : I<0x01, MRM_C8, (outs), (ins), "monitor",
2533 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2534 def MWAIT : I<0x01, MRM_C9, (outs), (ins), "mwait",
2535 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2537 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2538 let AddedComplexity = 15 in
2539 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2540 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2541 let AddedComplexity = 20 in
2542 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2543 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2545 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2546 let AddedComplexity = 15 in
2547 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2548 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2549 let AddedComplexity = 20 in
2550 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2551 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2553 //===---------------------------------------------------------------------===//
2554 // SSSE3 Instructions
2555 //===---------------------------------------------------------------------===//
2557 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2558 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2559 Intrinsic IntId64, Intrinsic IntId128> {
2560 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2561 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2562 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2564 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2565 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2567 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2569 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2571 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2572 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2575 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2577 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2580 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2583 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2584 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2585 Intrinsic IntId64, Intrinsic IntId128> {
2586 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2588 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2589 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2591 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2593 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2596 (bitconvert (memopv4i16 addr:$src))))]>;
2598 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2600 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2601 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2604 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2606 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2609 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2612 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2613 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2614 Intrinsic IntId64, Intrinsic IntId128> {
2615 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2617 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2618 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2620 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2622 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2625 (bitconvert (memopv2i32 addr:$src))))]>;
2627 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2629 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2630 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2633 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2635 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2638 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2641 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2642 int_x86_ssse3_pabs_b,
2643 int_x86_ssse3_pabs_b_128>;
2644 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2645 int_x86_ssse3_pabs_w,
2646 int_x86_ssse3_pabs_w_128>;
2647 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2648 int_x86_ssse3_pabs_d,
2649 int_x86_ssse3_pabs_d_128>;
2651 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2652 let Constraints = "$src1 = $dst" in {
2653 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2654 Intrinsic IntId64, Intrinsic IntId128,
2655 bit Commutable = 0> {
2656 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2657 (ins VR64:$src1, VR64:$src2),
2658 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2659 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2660 let isCommutable = Commutable;
2662 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2663 (ins VR64:$src1, i64mem:$src2),
2664 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2666 (IntId64 VR64:$src1,
2667 (bitconvert (memopv8i8 addr:$src2))))]>;
2669 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2670 (ins VR128:$src1, VR128:$src2),
2671 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2672 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2674 let isCommutable = Commutable;
2676 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2677 (ins VR128:$src1, i128mem:$src2),
2678 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2680 (IntId128 VR128:$src1,
2681 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2685 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2686 let Constraints = "$src1 = $dst" in {
2687 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2688 Intrinsic IntId64, Intrinsic IntId128,
2689 bit Commutable = 0> {
2690 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2691 (ins VR64:$src1, VR64:$src2),
2692 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2693 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2694 let isCommutable = Commutable;
2696 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2697 (ins VR64:$src1, i64mem:$src2),
2698 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2700 (IntId64 VR64:$src1,
2701 (bitconvert (memopv4i16 addr:$src2))))]>;
2703 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2704 (ins VR128:$src1, VR128:$src2),
2705 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2706 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2708 let isCommutable = Commutable;
2710 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2711 (ins VR128:$src1, i128mem:$src2),
2712 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2714 (IntId128 VR128:$src1,
2715 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2719 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2720 let Constraints = "$src1 = $dst" in {
2721 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2722 Intrinsic IntId64, Intrinsic IntId128,
2723 bit Commutable = 0> {
2724 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2725 (ins VR64:$src1, VR64:$src2),
2726 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2727 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2728 let isCommutable = Commutable;
2730 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2731 (ins VR64:$src1, i64mem:$src2),
2732 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2734 (IntId64 VR64:$src1,
2735 (bitconvert (memopv2i32 addr:$src2))))]>;
2737 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2738 (ins VR128:$src1, VR128:$src2),
2739 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2740 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2742 let isCommutable = Commutable;
2744 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2745 (ins VR128:$src1, i128mem:$src2),
2746 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2748 (IntId128 VR128:$src1,
2749 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2753 let ImmT = NoImm in { // None of these have i8 immediate fields.
2754 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2755 int_x86_ssse3_phadd_w,
2756 int_x86_ssse3_phadd_w_128>;
2757 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2758 int_x86_ssse3_phadd_d,
2759 int_x86_ssse3_phadd_d_128>;
2760 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2761 int_x86_ssse3_phadd_sw,
2762 int_x86_ssse3_phadd_sw_128>;
2763 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2764 int_x86_ssse3_phsub_w,
2765 int_x86_ssse3_phsub_w_128>;
2766 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2767 int_x86_ssse3_phsub_d,
2768 int_x86_ssse3_phsub_d_128>;
2769 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2770 int_x86_ssse3_phsub_sw,
2771 int_x86_ssse3_phsub_sw_128>;
2772 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2773 int_x86_ssse3_pmadd_ub_sw,
2774 int_x86_ssse3_pmadd_ub_sw_128>;
2775 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2776 int_x86_ssse3_pmul_hr_sw,
2777 int_x86_ssse3_pmul_hr_sw_128, 1>;
2779 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2780 int_x86_ssse3_pshuf_b,
2781 int_x86_ssse3_pshuf_b_128>;
2782 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2783 int_x86_ssse3_psign_b,
2784 int_x86_ssse3_psign_b_128>;
2785 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2786 int_x86_ssse3_psign_w,
2787 int_x86_ssse3_psign_w_128>;
2788 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2789 int_x86_ssse3_psign_d,
2790 int_x86_ssse3_psign_d_128>;
2793 // palignr patterns.
2794 let Constraints = "$src1 = $dst" in {
2795 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2796 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2797 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2799 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2800 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2801 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2804 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2805 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2806 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2808 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2809 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2810 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2814 let AddedComplexity = 5 in {
2816 def : Pat<(v1i64 (palign:$src3 VR64:$src1, VR64:$src2)),
2817 (PALIGNR64rr VR64:$src2, VR64:$src1,
2818 (SHUFFLE_get_palign_imm VR64:$src3))>,
2819 Requires<[HasSSSE3]>;
2820 def : Pat<(v2i32 (palign:$src3 VR64:$src1, VR64:$src2)),
2821 (PALIGNR64rr VR64:$src2, VR64:$src1,
2822 (SHUFFLE_get_palign_imm VR64:$src3))>,
2823 Requires<[HasSSSE3]>;
2824 def : Pat<(v2f32 (palign:$src3 VR64:$src1, VR64:$src2)),
2825 (PALIGNR64rr VR64:$src2, VR64:$src1,
2826 (SHUFFLE_get_palign_imm VR64:$src3))>,
2827 Requires<[HasSSSE3]>;
2828 def : Pat<(v4i16 (palign:$src3 VR64:$src1, VR64:$src2)),
2829 (PALIGNR64rr VR64:$src2, VR64:$src1,
2830 (SHUFFLE_get_palign_imm VR64:$src3))>,
2831 Requires<[HasSSSE3]>;
2832 def : Pat<(v8i8 (palign:$src3 VR64:$src1, VR64:$src2)),
2833 (PALIGNR64rr VR64:$src2, VR64:$src1,
2834 (SHUFFLE_get_palign_imm VR64:$src3))>,
2835 Requires<[HasSSSE3]>;
2837 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2838 (PALIGNR128rr VR128:$src2, VR128:$src1,
2839 (SHUFFLE_get_palign_imm VR128:$src3))>,
2840 Requires<[HasSSSE3]>;
2841 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2842 (PALIGNR128rr VR128:$src2, VR128:$src1,
2843 (SHUFFLE_get_palign_imm VR128:$src3))>,
2844 Requires<[HasSSSE3]>;
2845 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2846 (PALIGNR128rr VR128:$src2, VR128:$src1,
2847 (SHUFFLE_get_palign_imm VR128:$src3))>,
2848 Requires<[HasSSSE3]>;
2849 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2850 (PALIGNR128rr VR128:$src2, VR128:$src1,
2851 (SHUFFLE_get_palign_imm VR128:$src3))>,
2852 Requires<[HasSSSE3]>;
2855 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2856 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2857 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2858 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2860 //===---------------------------------------------------------------------===//
2861 // Non-Instruction Patterns
2862 //===---------------------------------------------------------------------===//
2864 // extload f32 -> f64. This matches load+fextend because we have a hack in
2865 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2867 // Since these loads aren't folded into the fextend, we have to match it
2869 let Predicates = [HasSSE2] in
2870 def : Pat<(fextend (loadf32 addr:$src)),
2871 (CVTSS2SDrm addr:$src)>;
2874 let Predicates = [HasSSE2] in {
2875 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2876 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2877 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2878 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2879 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2880 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2881 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2882 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2883 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2884 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2885 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2886 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2887 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2888 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2889 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2890 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2891 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2892 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2893 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2894 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2895 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2896 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2897 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2898 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2899 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2900 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2901 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2902 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2903 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2904 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2907 // Move scalar to XMM zero-extended
2908 // movd to XMM register zero-extends
2909 let AddedComplexity = 15 in {
2910 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2911 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2912 (MOVSDrr (v2f64 (V_SET0PS)), FR64:$src)>;
2913 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2914 (MOVSSrr (v4f32 (V_SET0PS)), FR32:$src)>;
2915 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2916 (MOVSSrr (v4f32 (V_SET0PS)),
2917 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss)))>;
2918 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2919 (MOVSSrr (v4i32 (V_SET0PI)),
2920 (EXTRACT_SUBREG (v4i32 VR128:$src), sub_ss))>;
2923 // Splat v2f64 / v2i64
2924 let AddedComplexity = 10 in {
2925 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
2926 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2927 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
2928 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2929 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
2930 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2931 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
2932 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2935 // Special unary SHUFPSrri case.
2936 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
2937 (SHUFPSrri VR128:$src1, VR128:$src1,
2938 (SHUFFLE_get_shuf_imm VR128:$src3))>;
2939 let AddedComplexity = 5 in
2940 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
2941 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2942 Requires<[HasSSE2]>;
2943 // Special unary SHUFPDrri case.
2944 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
2945 (SHUFPDrri VR128:$src1, VR128:$src1,
2946 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2947 Requires<[HasSSE2]>;
2948 // Special unary SHUFPDrri case.
2949 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
2950 (SHUFPDrri VR128:$src1, VR128:$src1,
2951 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2952 Requires<[HasSSE2]>;
2953 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2954 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
2955 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2956 Requires<[HasSSE2]>;
2958 // Special binary v4i32 shuffle cases with SHUFPS.
2959 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
2960 (SHUFPSrri VR128:$src1, VR128:$src2,
2961 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2962 Requires<[HasSSE2]>;
2963 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
2964 (SHUFPSrmi VR128:$src1, addr:$src2,
2965 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2966 Requires<[HasSSE2]>;
2967 // Special binary v2i64 shuffle cases using SHUFPDrri.
2968 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
2969 (SHUFPDrri VR128:$src1, VR128:$src2,
2970 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2971 Requires<[HasSSE2]>;
2973 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2974 let AddedComplexity = 15 in {
2975 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
2976 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2977 Requires<[OptForSpeed, HasSSE2]>;
2978 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
2979 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2980 Requires<[OptForSpeed, HasSSE2]>;
2982 let AddedComplexity = 10 in {
2983 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
2984 (UNPCKLPSrr VR128:$src, VR128:$src)>;
2985 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
2986 (PUNPCKLBWrr VR128:$src, VR128:$src)>;
2987 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
2988 (PUNPCKLWDrr VR128:$src, VR128:$src)>;
2989 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
2990 (PUNPCKLDQrr VR128:$src, VR128:$src)>;
2993 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2994 let AddedComplexity = 15 in {
2995 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
2996 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2997 Requires<[OptForSpeed, HasSSE2]>;
2998 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
2999 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3000 Requires<[OptForSpeed, HasSSE2]>;
3002 let AddedComplexity = 10 in {
3003 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
3004 (UNPCKHPSrr VR128:$src, VR128:$src)>;
3005 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
3006 (PUNPCKHBWrr VR128:$src, VR128:$src)>;
3007 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
3008 (PUNPCKHWDrr VR128:$src, VR128:$src)>;
3009 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
3010 (PUNPCKHDQrr VR128:$src, VR128:$src)>;
3013 let AddedComplexity = 20 in {
3014 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3015 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3016 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3018 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3019 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3020 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3022 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3023 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3024 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3025 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3026 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3029 let AddedComplexity = 20 in {
3030 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3031 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3032 (MOVLPSrm VR128:$src1, addr:$src2)>;
3033 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3034 (MOVLPDrm VR128:$src1, addr:$src2)>;
3035 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3036 (MOVLPSrm VR128:$src1, addr:$src2)>;
3037 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3038 (MOVLPDrm VR128:$src1, addr:$src2)>;
3041 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3042 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3043 (MOVLPSmr addr:$src1, VR128:$src2)>;
3044 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3045 (MOVLPDmr addr:$src1, VR128:$src2)>;
3046 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3048 (MOVLPSmr addr:$src1, VR128:$src2)>;
3049 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3050 (MOVLPDmr addr:$src1, VR128:$src2)>;
3052 let AddedComplexity = 15 in {
3053 // Setting the lowest element in the vector.
3054 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3055 (MOVSSrr (v4i32 VR128:$src1),
3056 (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_ss))>;
3057 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3058 (MOVSDrr (v2i64 VR128:$src1),
3059 (EXTRACT_SUBREG (v2i64 VR128:$src2), sub_sd))>;
3061 // vector_shuffle v1, v2 <4, 5, 2, 3> using movsd
3062 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3063 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3064 Requires<[HasSSE2]>;
3065 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3066 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3067 Requires<[HasSSE2]>;
3070 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3071 // fall back to this for SSE1)
3072 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3073 (SHUFPSrri VR128:$src2, VR128:$src1,
3074 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3076 // Set lowest element and zero upper elements.
3077 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3078 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3080 // Some special case pandn patterns.
3081 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3083 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3084 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3086 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3087 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3089 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3091 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3092 (memop addr:$src2))),
3093 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3094 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3095 (memop addr:$src2))),
3096 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3097 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3098 (memop addr:$src2))),
3099 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3101 // vector -> vector casts
3102 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3103 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3104 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3105 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3106 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3107 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3108 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3109 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3111 // Use movaps / movups for SSE integer load / store (one byte shorter).
3112 def : Pat<(alignedloadv4i32 addr:$src),
3113 (MOVAPSrm addr:$src)>;
3114 def : Pat<(loadv4i32 addr:$src),
3115 (MOVUPSrm addr:$src)>;
3116 def : Pat<(alignedloadv2i64 addr:$src),
3117 (MOVAPSrm addr:$src)>;
3118 def : Pat<(loadv2i64 addr:$src),
3119 (MOVUPSrm addr:$src)>;
3121 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3122 (MOVAPSmr addr:$dst, VR128:$src)>;
3123 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3124 (MOVAPSmr addr:$dst, VR128:$src)>;
3125 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3126 (MOVAPSmr addr:$dst, VR128:$src)>;
3127 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3128 (MOVAPSmr addr:$dst, VR128:$src)>;
3129 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3130 (MOVUPSmr addr:$dst, VR128:$src)>;
3131 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3132 (MOVUPSmr addr:$dst, VR128:$src)>;
3133 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3134 (MOVUPSmr addr:$dst, VR128:$src)>;
3135 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3136 (MOVUPSmr addr:$dst, VR128:$src)>;
3138 //===----------------------------------------------------------------------===//
3139 // SSE4.1 Instructions
3140 //===----------------------------------------------------------------------===//
3142 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3145 Intrinsic V2F64Int> {
3146 // Intrinsic operation, reg.
3147 // Vector intrinsic operation, reg
3148 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3149 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3150 !strconcat(OpcodeStr,
3151 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3152 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3155 // Vector intrinsic operation, mem
3156 def PSm_Int : Ii8<opcps, MRMSrcMem,
3157 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3158 !strconcat(OpcodeStr,
3159 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3161 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3163 Requires<[HasSSE41]>;
3165 // Vector intrinsic operation, reg
3166 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3167 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3168 !strconcat(OpcodeStr,
3169 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3170 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3173 // Vector intrinsic operation, mem
3174 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3175 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3176 !strconcat(OpcodeStr,
3177 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3179 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3183 let Constraints = "$src1 = $dst" in {
3184 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3188 // Intrinsic operation, reg.
3189 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3191 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3192 !strconcat(OpcodeStr,
3193 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3195 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3198 // Intrinsic operation, mem.
3199 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3201 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3202 !strconcat(OpcodeStr,
3203 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3205 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3208 // Intrinsic operation, reg.
3209 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3211 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3212 !strconcat(OpcodeStr,
3213 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3215 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3218 // Intrinsic operation, mem.
3219 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3221 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3222 !strconcat(OpcodeStr,
3223 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3225 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3230 // FP round - roundss, roundps, roundsd, roundpd
3231 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3232 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3233 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3234 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3236 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3237 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3238 Intrinsic IntId128> {
3239 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3241 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3242 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3243 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3245 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3248 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3251 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3252 int_x86_sse41_phminposuw>;
3254 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3255 let Constraints = "$src1 = $dst" in {
3256 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3257 Intrinsic IntId128, bit Commutable = 0> {
3258 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3259 (ins VR128:$src1, VR128:$src2),
3260 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3261 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3263 let isCommutable = Commutable;
3265 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3266 (ins VR128:$src1, i128mem:$src2),
3267 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3269 (IntId128 VR128:$src1,
3270 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3274 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3275 int_x86_sse41_pcmpeqq, 1>;
3276 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3277 int_x86_sse41_packusdw, 0>;
3278 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3279 int_x86_sse41_pminsb, 1>;
3280 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3281 int_x86_sse41_pminsd, 1>;
3282 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3283 int_x86_sse41_pminud, 1>;
3284 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3285 int_x86_sse41_pminuw, 1>;
3286 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3287 int_x86_sse41_pmaxsb, 1>;
3288 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3289 int_x86_sse41_pmaxsd, 1>;
3290 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3291 int_x86_sse41_pmaxud, 1>;
3292 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3293 int_x86_sse41_pmaxuw, 1>;
3295 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3297 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3298 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3299 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3300 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3302 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3303 let Constraints = "$src1 = $dst" in {
3304 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3305 SDNode OpNode, Intrinsic IntId128,
3306 bit Commutable = 0> {
3307 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3308 (ins VR128:$src1, VR128:$src2),
3309 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3310 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3311 VR128:$src2))]>, OpSize {
3312 let isCommutable = Commutable;
3314 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3315 (ins VR128:$src1, VR128:$src2),
3316 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3317 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3319 let isCommutable = Commutable;
3321 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3322 (ins VR128:$src1, i128mem:$src2),
3323 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3325 (OpVT (OpNode VR128:$src1, (memop addr:$src2))))]>, OpSize;
3326 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3327 (ins VR128:$src1, i128mem:$src2),
3328 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3330 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3335 /// SS48I_binop_rm - Simple SSE41 binary operator.
3336 let Constraints = "$src1 = $dst" in {
3337 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
3338 ValueType OpVT, bit Commutable = 0> {
3339 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3340 (ins VR128:$src1, VR128:$src2),
3341 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3342 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]>,
3344 let isCommutable = Commutable;
3346 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3347 (ins VR128:$src1, i128mem:$src2),
3348 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3349 [(set VR128:$dst, (OpNode VR128:$src1,
3350 (bc_v4i32 (memopv2i64 addr:$src2))))]>,
3355 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, 1>;
3357 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3358 let Constraints = "$src1 = $dst" in {
3359 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3360 Intrinsic IntId128, bit Commutable = 0> {
3361 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3362 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3363 !strconcat(OpcodeStr,
3364 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3366 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3368 let isCommutable = Commutable;
3370 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3371 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3372 !strconcat(OpcodeStr,
3373 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3375 (IntId128 VR128:$src1,
3376 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3381 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3382 int_x86_sse41_blendps, 0>;
3383 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3384 int_x86_sse41_blendpd, 0>;
3385 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3386 int_x86_sse41_pblendw, 0>;
3387 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3388 int_x86_sse41_dpps, 1>;
3389 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3390 int_x86_sse41_dppd, 1>;
3391 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3392 int_x86_sse41_mpsadbw, 0>;
3395 /// SS41I_ternary_int - SSE 4.1 ternary operator
3396 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3397 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3398 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3399 (ins VR128:$src1, VR128:$src2),
3400 !strconcat(OpcodeStr,
3401 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3402 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3405 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3406 (ins VR128:$src1, i128mem:$src2),
3407 !strconcat(OpcodeStr,
3408 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3411 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3415 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3416 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3417 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3420 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3421 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3422 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3423 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3425 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3426 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3428 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3432 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3433 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3434 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3435 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3436 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3437 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3439 // Common patterns involving scalar load.
3440 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3441 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3442 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3443 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3445 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3446 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3447 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3448 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3450 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3451 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3452 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3453 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3455 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3456 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3457 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3458 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3460 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3461 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3462 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3463 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3465 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3466 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3467 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3468 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3471 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3472 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3473 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3474 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3476 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3477 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3479 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3483 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3484 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3485 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3486 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3488 // Common patterns involving scalar load
3489 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3490 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3491 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3492 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3494 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3495 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3496 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3497 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3500 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3501 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3502 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3503 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3505 // Expecting a i16 load any extended to i32 value.
3506 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3507 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3508 [(set VR128:$dst, (IntId (bitconvert
3509 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3513 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3514 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3516 // Common patterns involving scalar load
3517 def : Pat<(int_x86_sse41_pmovsxbq
3518 (bitconvert (v4i32 (X86vzmovl
3519 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3520 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3522 def : Pat<(int_x86_sse41_pmovzxbq
3523 (bitconvert (v4i32 (X86vzmovl
3524 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3525 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3528 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3529 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3530 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3531 (ins VR128:$src1, i32i8imm:$src2),
3532 !strconcat(OpcodeStr,
3533 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3534 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3536 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3537 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3538 !strconcat(OpcodeStr,
3539 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3542 // There's an AssertZext in the way of writing the store pattern
3543 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3546 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3549 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3550 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3551 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3552 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3553 !strconcat(OpcodeStr,
3554 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3557 // There's an AssertZext in the way of writing the store pattern
3558 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3561 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3564 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3565 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3566 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3567 (ins VR128:$src1, i32i8imm:$src2),
3568 !strconcat(OpcodeStr,
3569 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3571 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3572 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3573 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3574 !strconcat(OpcodeStr,
3575 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3576 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3577 addr:$dst)]>, OpSize;
3580 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3583 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3585 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3586 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3587 (ins VR128:$src1, i32i8imm:$src2),
3588 !strconcat(OpcodeStr,
3589 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3591 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3593 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3594 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3595 !strconcat(OpcodeStr,
3596 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3597 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3598 addr:$dst)]>, OpSize;
3601 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3603 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3604 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3607 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3608 Requires<[HasSSE41]>;
3610 let Constraints = "$src1 = $dst" in {
3611 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3612 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3613 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3614 !strconcat(OpcodeStr,
3615 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3617 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3618 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3619 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3620 !strconcat(OpcodeStr,
3621 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3623 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3624 imm:$src3))]>, OpSize;
3628 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3630 let Constraints = "$src1 = $dst" in {
3631 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3632 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3633 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3634 !strconcat(OpcodeStr,
3635 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3637 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3639 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3640 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3641 !strconcat(OpcodeStr,
3642 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3644 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3645 imm:$src3)))]>, OpSize;
3649 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3651 // insertps has a few different modes, there's the first two here below which
3652 // are optimized inserts that won't zero arbitrary elements in the destination
3653 // vector. The next one matches the intrinsic and could zero arbitrary elements
3654 // in the target vector.
3655 let Constraints = "$src1 = $dst" in {
3656 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3657 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3658 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3659 !strconcat(OpcodeStr,
3660 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3662 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3664 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3665 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3666 !strconcat(OpcodeStr,
3667 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3669 (X86insrtps VR128:$src1,
3670 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3671 imm:$src3))]>, OpSize;
3675 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3677 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3678 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3680 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3681 // the intel intrinsic that corresponds to this.
3682 let Defs = [EFLAGS] in {
3683 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3684 "ptest \t{$src2, $src1|$src1, $src2}",
3685 [(set EFLAGS, (X86ptest VR128:$src1, VR128:$src2))]>,
3687 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3688 "ptest \t{$src2, $src1|$src1, $src2}",
3689 [(set EFLAGS, (X86ptest VR128:$src1, (load addr:$src2)))]>,
3693 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3694 "movntdqa\t{$src, $dst|$dst, $src}",
3695 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
3699 //===----------------------------------------------------------------------===//
3700 // SSE4.2 Instructions
3701 //===----------------------------------------------------------------------===//
3703 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3704 let Constraints = "$src1 = $dst" in {
3705 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3706 Intrinsic IntId128, bit Commutable = 0> {
3707 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3708 (ins VR128:$src1, VR128:$src2),
3709 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3710 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3712 let isCommutable = Commutable;
3714 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3715 (ins VR128:$src1, i128mem:$src2),
3716 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3718 (IntId128 VR128:$src1,
3719 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3723 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3725 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3726 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3727 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3728 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3730 // crc intrinsic instruction
3731 // This set of instructions are only rm, the only difference is the size
3733 let Constraints = "$src1 = $dst" in {
3734 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3735 (ins GR32:$src1, i8mem:$src2),
3736 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3738 (int_x86_sse42_crc32_8 GR32:$src1,
3739 (load addr:$src2)))]>;
3740 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3741 (ins GR32:$src1, GR8:$src2),
3742 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3744 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>;
3745 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3746 (ins GR32:$src1, i16mem:$src2),
3747 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3749 (int_x86_sse42_crc32_16 GR32:$src1,
3750 (load addr:$src2)))]>,
3752 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3753 (ins GR32:$src1, GR16:$src2),
3754 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3756 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3758 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3759 (ins GR32:$src1, i32mem:$src2),
3760 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3762 (int_x86_sse42_crc32_32 GR32:$src1,
3763 (load addr:$src2)))]>;
3764 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3765 (ins GR32:$src1, GR32:$src2),
3766 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3768 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>;
3769 def CRC64m8 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3770 (ins GR64:$src1, i8mem:$src2),
3771 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3773 (int_x86_sse42_crc64_8 GR64:$src1,
3774 (load addr:$src2)))]>,
3776 def CRC64r8 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3777 (ins GR64:$src1, GR8:$src2),
3778 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3780 (int_x86_sse42_crc64_8 GR64:$src1, GR8:$src2))]>,
3782 def CRC64m64 : SS42FI<0xF1, MRMSrcMem, (outs GR64:$dst),
3783 (ins GR64:$src1, i64mem:$src2),
3784 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3786 (int_x86_sse42_crc64_64 GR64:$src1,
3787 (load addr:$src2)))]>,
3789 def CRC64r64 : SS42FI<0xF1, MRMSrcReg, (outs GR64:$dst),
3790 (ins GR64:$src1, GR64:$src2),
3791 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3793 (int_x86_sse42_crc64_64 GR64:$src1, GR64:$src2))]>,
3797 // String/text processing instructions.
3798 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3799 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3800 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3801 "#PCMPISTRM128rr PSEUDO!",
3802 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3803 imm:$src3))]>, OpSize;
3804 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3805 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3806 "#PCMPISTRM128rm PSEUDO!",
3807 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3808 imm:$src3))]>, OpSize;
3811 let Defs = [XMM0, EFLAGS] in {
3812 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3813 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3814 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3815 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3816 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3817 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3820 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3821 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3822 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3823 "#PCMPESTRM128rr PSEUDO!",
3825 (int_x86_sse42_pcmpestrm128
3826 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3828 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3829 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3830 "#PCMPESTRM128rm PSEUDO!",
3831 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3832 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3836 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3837 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3838 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3839 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3840 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3841 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3842 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3845 let Defs = [ECX, EFLAGS] in {
3846 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3847 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3848 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3849 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3850 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3851 (implicit EFLAGS)]>, OpSize;
3852 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3853 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3854 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3855 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3856 (implicit EFLAGS)]>, OpSize;
3860 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3861 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3862 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3863 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3864 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3865 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3867 let Defs = [ECX, EFLAGS] in {
3868 let Uses = [EAX, EDX] in {
3869 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3870 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3871 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3872 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3873 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3874 (implicit EFLAGS)]>, OpSize;
3875 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3876 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3877 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3879 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
3880 (implicit EFLAGS)]>, OpSize;
3885 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3886 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3887 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3888 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3889 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3890 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;
3892 //===----------------------------------------------------------------------===//
3893 // AES-NI Instructions
3894 //===----------------------------------------------------------------------===//
3896 let Constraints = "$src1 = $dst" in {
3897 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr,
3898 Intrinsic IntId128, bit Commutable = 0> {
3899 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
3900 (ins VR128:$src1, VR128:$src2),
3901 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3902 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3904 let isCommutable = Commutable;
3906 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
3907 (ins VR128:$src1, i128mem:$src2),
3908 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3910 (IntId128 VR128:$src1,
3911 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3915 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
3916 int_x86_aesni_aesenc>;
3917 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
3918 int_x86_aesni_aesenclast>;
3919 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
3920 int_x86_aesni_aesdec>;
3921 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
3922 int_x86_aesni_aesdeclast>;
3924 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, VR128:$src2)),
3925 (AESENCrr VR128:$src1, VR128:$src2)>;
3926 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, (memop addr:$src2))),
3927 (AESENCrm VR128:$src1, addr:$src2)>;
3928 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, VR128:$src2)),
3929 (AESENCLASTrr VR128:$src1, VR128:$src2)>;
3930 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, (memop addr:$src2))),
3931 (AESENCLASTrm VR128:$src1, addr:$src2)>;
3932 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, VR128:$src2)),
3933 (AESDECrr VR128:$src1, VR128:$src2)>;
3934 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, (memop addr:$src2))),
3935 (AESDECrm VR128:$src1, addr:$src2)>;
3936 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, VR128:$src2)),
3937 (AESDECLASTrr VR128:$src1, VR128:$src2)>;
3938 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, (memop addr:$src2))),
3939 (AESDECLASTrm VR128:$src1, addr:$src2)>;
3941 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
3943 "aesimc\t{$src1, $dst|$dst, $src1}",
3945 (int_x86_aesni_aesimc VR128:$src1))]>,
3948 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
3949 (ins i128mem:$src1),
3950 "aesimc\t{$src1, $dst|$dst, $src1}",
3952 (int_x86_aesni_aesimc (bitconvert (memopv2i64 addr:$src1))))]>,
3955 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
3956 (ins VR128:$src1, i8imm:$src2),
3957 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3959 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
3961 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
3962 (ins i128mem:$src1, i8imm:$src2),
3963 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3965 (int_x86_aesni_aeskeygenassist (bitconvert (memopv2i64 addr:$src1)),