1 //====- X86InstrSSE.td - Describe the X86 Instruction Set --*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the X86 SSE instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
17 //===----------------------------------------------------------------------===//
18 // SSE specific DAG Nodes.
19 //===----------------------------------------------------------------------===//
21 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
22 SDTCisFP<0>, SDTCisInt<2> ]>;
23 def SDTX86VFCMP : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>,
24 SDTCisFP<1>, SDTCisVT<3, i8>]>;
26 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
27 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
28 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
29 [SDNPCommutative, SDNPAssociative]>;
30 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
31 [SDNPCommutative, SDNPAssociative]>;
32 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
33 [SDNPCommutative, SDNPAssociative]>;
34 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
35 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
36 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
37 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>;
38 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>;
39 def X86pshufb : SDNode<"X86ISD::PSHUFB",
40 SDTypeProfile<1, 2, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
42 def X86pextrb : SDNode<"X86ISD::PEXTRB",
43 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
44 def X86pextrw : SDNode<"X86ISD::PEXTRW",
45 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
46 def X86pinsrb : SDNode<"X86ISD::PINSRB",
47 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
48 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
49 def X86pinsrw : SDNode<"X86ISD::PINSRW",
50 SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
51 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
52 def X86insrtps : SDNode<"X86ISD::INSERTPS",
53 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
54 SDTCisVT<2, v4f32>, SDTCisPtrTy<3>]>>;
55 def X86vzmovl : SDNode<"X86ISD::VZEXT_MOVL",
56 SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>;
57 def X86vzload : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
58 [SDNPHasChain, SDNPMayLoad]>;
59 def X86vshl : SDNode<"X86ISD::VSHL", SDTIntShiftOp>;
60 def X86vshr : SDNode<"X86ISD::VSRL", SDTIntShiftOp>;
61 def X86cmpps : SDNode<"X86ISD::CMPPS", SDTX86VFCMP>;
62 def X86cmppd : SDNode<"X86ISD::CMPPD", SDTX86VFCMP>;
63 def X86pcmpeqb : SDNode<"X86ISD::PCMPEQB", SDTIntBinOp, [SDNPCommutative]>;
64 def X86pcmpeqw : SDNode<"X86ISD::PCMPEQW", SDTIntBinOp, [SDNPCommutative]>;
65 def X86pcmpeqd : SDNode<"X86ISD::PCMPEQD", SDTIntBinOp, [SDNPCommutative]>;
66 def X86pcmpeqq : SDNode<"X86ISD::PCMPEQQ", SDTIntBinOp, [SDNPCommutative]>;
67 def X86pcmpgtb : SDNode<"X86ISD::PCMPGTB", SDTIntBinOp>;
68 def X86pcmpgtw : SDNode<"X86ISD::PCMPGTW", SDTIntBinOp>;
69 def X86pcmpgtd : SDNode<"X86ISD::PCMPGTD", SDTIntBinOp>;
70 def X86pcmpgtq : SDNode<"X86ISD::PCMPGTQ", SDTIntBinOp>;
72 def SDTX86CmpPTest : SDTypeProfile<1, 2, [SDTCisVT<0, i32>,
75 def X86ptest : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
77 //===----------------------------------------------------------------------===//
78 // SSE Complex Patterns
79 //===----------------------------------------------------------------------===//
81 // These are 'extloads' from a scalar to the low element of a vector, zeroing
82 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
84 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
85 [SDNPHasChain, SDNPMayLoad]>;
86 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
87 [SDNPHasChain, SDNPMayLoad]>;
89 def ssmem : Operand<v4f32> {
90 let PrintMethod = "printf32mem";
91 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
92 let ParserMatchClass = X86MemAsmOperand;
94 def sdmem : Operand<v2f64> {
95 let PrintMethod = "printf64mem";
96 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
97 let ParserMatchClass = X86MemAsmOperand;
100 //===----------------------------------------------------------------------===//
101 // SSE pattern fragments
102 //===----------------------------------------------------------------------===//
104 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
105 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
106 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
107 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
109 // Like 'store', but always requires vector alignment.
110 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
111 (store node:$val, node:$ptr), [{
112 return cast<StoreSDNode>(N)->getAlignment() >= 16;
115 // Like 'load', but always requires vector alignment.
116 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
117 return cast<LoadSDNode>(N)->getAlignment() >= 16;
120 def alignedloadfsf32 : PatFrag<(ops node:$ptr),
121 (f32 (alignedload node:$ptr))>;
122 def alignedloadfsf64 : PatFrag<(ops node:$ptr),
123 (f64 (alignedload node:$ptr))>;
124 def alignedloadv4f32 : PatFrag<(ops node:$ptr),
125 (v4f32 (alignedload node:$ptr))>;
126 def alignedloadv2f64 : PatFrag<(ops node:$ptr),
127 (v2f64 (alignedload node:$ptr))>;
128 def alignedloadv4i32 : PatFrag<(ops node:$ptr),
129 (v4i32 (alignedload node:$ptr))>;
130 def alignedloadv2i64 : PatFrag<(ops node:$ptr),
131 (v2i64 (alignedload node:$ptr))>;
133 // Like 'load', but uses special alignment checks suitable for use in
134 // memory operands in most SSE instructions, which are required to
135 // be naturally aligned on some targets but not on others. If the subtarget
136 // allows unaligned accesses, match any load, though this may require
137 // setting a feature bit in the processor (on startup, for example).
138 // Opteron 10h and later implement such a feature.
139 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
140 return Subtarget->hasVectorUAMem()
141 || cast<LoadSDNode>(N)->getAlignment() >= 16;
144 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
145 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
146 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
147 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
148 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
149 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
150 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
152 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
154 // FIXME: 8 byte alignment for mmx reads is not required
155 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
156 return cast<LoadSDNode>(N)->getAlignment() >= 8;
159 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
160 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
161 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
162 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
165 // Like 'store', but requires the non-temporal bit to be set
166 def nontemporalstore : PatFrag<(ops node:$val, node:$ptr),
167 (st node:$val, node:$ptr), [{
168 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
169 return ST->isNonTemporal();
173 def alignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
174 (st node:$val, node:$ptr), [{
175 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
176 return ST->isNonTemporal() && !ST->isTruncatingStore() &&
177 ST->getAddressingMode() == ISD::UNINDEXED &&
178 ST->getAlignment() >= 16;
182 def unalignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
183 (st node:$val, node:$ptr), [{
184 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
185 return ST->isNonTemporal() &&
186 ST->getAlignment() < 16;
190 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
191 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
192 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
193 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
194 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
195 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
197 def vzmovl_v2i64 : PatFrag<(ops node:$src),
198 (bitconvert (v2i64 (X86vzmovl
199 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
200 def vzmovl_v4i32 : PatFrag<(ops node:$src),
201 (bitconvert (v4i32 (X86vzmovl
202 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
204 def vzload_v2i64 : PatFrag<(ops node:$src),
205 (bitconvert (v2i64 (X86vzload node:$src)))>;
208 def fp32imm0 : PatLeaf<(f32 fpimm), [{
209 return N->isExactlyValue(+0.0);
212 // BYTE_imm - Transform bit immediates into byte immediates.
213 def BYTE_imm : SDNodeXForm<imm, [{
214 // Transformation function: imm >> 3
215 return getI32Imm(N->getZExtValue() >> 3);
218 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
220 def SHUFFLE_get_shuf_imm : SDNodeXForm<vector_shuffle, [{
221 return getI8Imm(X86::getShuffleSHUFImmediate(N));
224 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
226 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<vector_shuffle, [{
227 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
230 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
232 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<vector_shuffle, [{
233 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
236 // SHUFFLE_get_palign_imm xform function: convert vector_shuffle mask to
238 def SHUFFLE_get_palign_imm : SDNodeXForm<vector_shuffle, [{
239 return getI8Imm(X86::getShufflePALIGNRImmediate(N));
242 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
243 (vector_shuffle node:$lhs, node:$rhs), [{
244 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
245 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
248 def movddup : PatFrag<(ops node:$lhs, node:$rhs),
249 (vector_shuffle node:$lhs, node:$rhs), [{
250 return X86::isMOVDDUPMask(cast<ShuffleVectorSDNode>(N));
253 def movhlps : PatFrag<(ops node:$lhs, node:$rhs),
254 (vector_shuffle node:$lhs, node:$rhs), [{
255 return X86::isMOVHLPSMask(cast<ShuffleVectorSDNode>(N));
258 def movhlps_undef : PatFrag<(ops node:$lhs, node:$rhs),
259 (vector_shuffle node:$lhs, node:$rhs), [{
260 return X86::isMOVHLPS_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
263 def movlhps : PatFrag<(ops node:$lhs, node:$rhs),
264 (vector_shuffle node:$lhs, node:$rhs), [{
265 return X86::isMOVLHPSMask(cast<ShuffleVectorSDNode>(N));
268 def movlp : PatFrag<(ops node:$lhs, node:$rhs),
269 (vector_shuffle node:$lhs, node:$rhs), [{
270 return X86::isMOVLPMask(cast<ShuffleVectorSDNode>(N));
273 def movl : PatFrag<(ops node:$lhs, node:$rhs),
274 (vector_shuffle node:$lhs, node:$rhs), [{
275 return X86::isMOVLMask(cast<ShuffleVectorSDNode>(N));
278 def movshdup : PatFrag<(ops node:$lhs, node:$rhs),
279 (vector_shuffle node:$lhs, node:$rhs), [{
280 return X86::isMOVSHDUPMask(cast<ShuffleVectorSDNode>(N));
283 def movsldup : PatFrag<(ops node:$lhs, node:$rhs),
284 (vector_shuffle node:$lhs, node:$rhs), [{
285 return X86::isMOVSLDUPMask(cast<ShuffleVectorSDNode>(N));
288 def unpckl : PatFrag<(ops node:$lhs, node:$rhs),
289 (vector_shuffle node:$lhs, node:$rhs), [{
290 return X86::isUNPCKLMask(cast<ShuffleVectorSDNode>(N));
293 def unpckh : PatFrag<(ops node:$lhs, node:$rhs),
294 (vector_shuffle node:$lhs, node:$rhs), [{
295 return X86::isUNPCKHMask(cast<ShuffleVectorSDNode>(N));
298 def unpckl_undef : PatFrag<(ops node:$lhs, node:$rhs),
299 (vector_shuffle node:$lhs, node:$rhs), [{
300 return X86::isUNPCKL_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
303 def unpckh_undef : PatFrag<(ops node:$lhs, node:$rhs),
304 (vector_shuffle node:$lhs, node:$rhs), [{
305 return X86::isUNPCKH_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
308 def pshufd : PatFrag<(ops node:$lhs, node:$rhs),
309 (vector_shuffle node:$lhs, node:$rhs), [{
310 return X86::isPSHUFDMask(cast<ShuffleVectorSDNode>(N));
311 }], SHUFFLE_get_shuf_imm>;
313 def shufp : PatFrag<(ops node:$lhs, node:$rhs),
314 (vector_shuffle node:$lhs, node:$rhs), [{
315 return X86::isSHUFPMask(cast<ShuffleVectorSDNode>(N));
316 }], SHUFFLE_get_shuf_imm>;
318 def pshufhw : PatFrag<(ops node:$lhs, node:$rhs),
319 (vector_shuffle node:$lhs, node:$rhs), [{
320 return X86::isPSHUFHWMask(cast<ShuffleVectorSDNode>(N));
321 }], SHUFFLE_get_pshufhw_imm>;
323 def pshuflw : PatFrag<(ops node:$lhs, node:$rhs),
324 (vector_shuffle node:$lhs, node:$rhs), [{
325 return X86::isPSHUFLWMask(cast<ShuffleVectorSDNode>(N));
326 }], SHUFFLE_get_pshuflw_imm>;
328 def palign : PatFrag<(ops node:$lhs, node:$rhs),
329 (vector_shuffle node:$lhs, node:$rhs), [{
330 return X86::isPALIGNRMask(cast<ShuffleVectorSDNode>(N));
331 }], SHUFFLE_get_palign_imm>;
333 //===----------------------------------------------------------------------===//
334 // SSE scalar FP Instructions
335 //===----------------------------------------------------------------------===//
337 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded after
338 // instruction selection into a branch sequence.
339 let Uses = [EFLAGS], usesCustomInserter = 1 in {
340 def CMOV_FR32 : I<0, Pseudo,
341 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
342 "#CMOV_FR32 PSEUDO!",
343 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
345 def CMOV_FR64 : I<0, Pseudo,
346 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
347 "#CMOV_FR64 PSEUDO!",
348 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
350 def CMOV_V4F32 : I<0, Pseudo,
351 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
352 "#CMOV_V4F32 PSEUDO!",
354 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
356 def CMOV_V2F64 : I<0, Pseudo,
357 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
358 "#CMOV_V2F64 PSEUDO!",
360 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
362 def CMOV_V2I64 : I<0, Pseudo,
363 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
364 "#CMOV_V2I64 PSEUDO!",
366 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
370 //===----------------------------------------------------------------------===//
371 // SSE 1 & 2 Instructions Classes
372 //===----------------------------------------------------------------------===//
374 /// sse12_fp_scalar - SSE 1 & 2 scalar instructions class
375 multiclass sse12_fp_scalar<bits<8> opc, string OpcodeStr, SDNode OpNode,
376 RegisterClass RC, X86MemOperand memop> {
377 let isCommutable = 1 in {
378 def rr : SI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
379 OpcodeStr, [(set RC:$dst, (OpNode RC:$src1, RC:$src2))]>;
381 def rm : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memop:$src2),
382 OpcodeStr, [(set RC:$dst, (OpNode RC:$src1, (load addr:$src2)))]>;
385 /// sse12_fp_scalar_int - SSE 1 & 2 scalar instructions intrinsics class
386 multiclass sse12_fp_scalar_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
387 string asm, string SSEVer, string FPSizeStr,
388 Operand memop, ComplexPattern mem_cpat> {
389 def rr_Int : SI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
390 asm, [(set RC:$dst, (
391 !nameconcat<Intrinsic>("int_x86_sse",
392 !strconcat(SSEVer, !strconcat("_",
393 !strconcat(OpcodeStr, FPSizeStr))))
394 RC:$src1, RC:$src2))]>;
395 def rm_Int : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memop:$src2),
396 asm, [(set RC:$dst, (
397 !nameconcat<Intrinsic>("int_x86_sse",
398 !strconcat(SSEVer, !strconcat("_",
399 !strconcat(OpcodeStr, FPSizeStr))))
400 RC:$src1, mem_cpat:$src2))]>;
403 /// sse12_fp_packed - SSE 1 & 2 packed instructions class
404 multiclass sse12_fp_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
405 RegisterClass RC, ValueType vt,
406 X86MemOperand x86memop, PatFrag mem_frag,
407 Domain d, bit MayLoad = 0> {
408 let isCommutable = 1 in
409 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
410 OpcodeStr, [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))],d>;
411 let mayLoad = MayLoad in
412 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
413 OpcodeStr, [(set RC:$dst, (OpNode RC:$src1,
414 (mem_frag addr:$src2)))],d>;
417 /// sse12_fp_packed_int - SSE 1 & 2 packed instructions intrinsics class
418 multiclass sse12_fp_packed_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
419 string asm, string SSEVer, string FPSizeStr,
420 X86MemOperand memop, PatFrag mem_frag,
422 def rr_Int : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
423 asm, [(set RC:$dst, (
424 !nameconcat<Intrinsic>("int_x86_sse",
425 !strconcat(SSEVer, !strconcat("_",
426 !strconcat(OpcodeStr, FPSizeStr))))
427 RC:$src1, RC:$src2))], d>;
428 def rm_Int : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memop:$src2),
429 asm, [(set RC:$dst, (
430 !nameconcat<Intrinsic>("int_x86_sse",
431 !strconcat(SSEVer, !strconcat("_",
432 !strconcat(OpcodeStr, FPSizeStr))))
433 RC:$src1, (mem_frag addr:$src2)))], d>;
436 //===----------------------------------------------------------------------===//
438 //===----------------------------------------------------------------------===//
440 // Move Instructions. Register-to-register movss is not used for FR32
441 // register copies because it's a partial register update; FsMOVAPSrr is
442 // used instead. Register-to-register movss is not modeled as an INSERT_SUBREG
443 // because INSERT_SUBREG requires that the insert be implementable in terms of
444 // a copy, and just mentioned, we don't use movss for copies.
445 let Constraints = "$src1 = $dst" in
446 def MOVSSrr : SSI<0x10, MRMSrcReg,
447 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
448 "movss\t{$src2, $dst|$dst, $src2}",
449 [(set (v4f32 VR128:$dst),
450 (movl VR128:$src1, (scalar_to_vector FR32:$src2)))]>;
452 // Extract the low 32-bit value from one vector and insert it into another.
453 let AddedComplexity = 15 in
454 def : Pat<(v4f32 (movl VR128:$src1, VR128:$src2)),
455 (MOVSSrr (v4f32 VR128:$src1),
456 (EXTRACT_SUBREG (v4f32 VR128:$src2), sub_ss))>;
458 // Implicitly promote a 32-bit scalar to a vector.
459 def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
460 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), FR32:$src, sub_ss)>;
462 // Loading from memory automatically zeroing upper bits.
463 let canFoldAsLoad = 1, isReMaterializable = 1 in
464 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
465 "movss\t{$src, $dst|$dst, $src}",
466 [(set FR32:$dst, (loadf32 addr:$src))]>;
468 // MOVSSrm zeros the high parts of the register; represent this
469 // with SUBREG_TO_REG.
470 let AddedComplexity = 20 in {
471 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
472 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
473 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
474 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
475 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
476 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
479 // Store scalar value to memory.
480 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
481 "movss\t{$src, $dst|$dst, $src}",
482 [(store FR32:$src, addr:$dst)]>;
484 // Extract and store.
485 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
488 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
490 // Conversion instructions
491 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
492 "cvttss2si\t{$src, $dst|$dst, $src}",
493 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
494 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
495 "cvttss2si\t{$src, $dst|$dst, $src}",
496 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
497 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
498 "cvtsi2ss\t{$src, $dst|$dst, $src}",
499 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
500 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
501 "cvtsi2ss\t{$src, $dst|$dst, $src}",
502 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
504 // Match intrinsics which expect XMM operand(s).
505 def CVTSS2SIrr: SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
506 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
507 def CVTSS2SIrm: SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
508 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
510 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
511 "cvtss2si\t{$src, $dst|$dst, $src}",
512 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
513 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
514 "cvtss2si\t{$src, $dst|$dst, $src}",
515 [(set GR32:$dst, (int_x86_sse_cvtss2si
516 (load addr:$src)))]>;
518 // Match intrinsics which expect MM and XMM operand(s).
519 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
520 "cvtps2pi\t{$src, $dst|$dst, $src}",
521 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
522 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
523 "cvtps2pi\t{$src, $dst|$dst, $src}",
524 [(set VR64:$dst, (int_x86_sse_cvtps2pi
525 (load addr:$src)))]>;
526 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
527 "cvttps2pi\t{$src, $dst|$dst, $src}",
528 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
529 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
530 "cvttps2pi\t{$src, $dst|$dst, $src}",
531 [(set VR64:$dst, (int_x86_sse_cvttps2pi
532 (load addr:$src)))]>;
533 let Constraints = "$src1 = $dst" in {
534 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
535 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
536 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
537 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
539 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
540 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
541 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
542 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
543 (load addr:$src2)))]>;
546 // Aliases for intrinsics
547 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
548 "cvttss2si\t{$src, $dst|$dst, $src}",
550 (int_x86_sse_cvttss2si VR128:$src))]>;
551 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
552 "cvttss2si\t{$src, $dst|$dst, $src}",
554 (int_x86_sse_cvttss2si(load addr:$src)))]>;
556 let Constraints = "$src1 = $dst" in {
557 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
558 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
559 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
560 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
562 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
563 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
564 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
565 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
566 (loadi32 addr:$src2)))]>;
569 // Comparison instructions
570 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
571 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
572 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
573 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
575 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
576 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
577 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
579 // Accept explicit immediate argument form instead of comparison code.
580 let isAsmParserOnly = 1 in {
581 def CMPSSrr_alt : SSIi8<0xC2, MRMSrcReg,
582 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, i8imm:$src2),
583 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
585 def CMPSSrm_alt : SSIi8<0xC2, MRMSrcMem,
586 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, i8imm:$src2),
587 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
591 let Defs = [EFLAGS] in {
592 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
593 "ucomiss\t{$src2, $src1|$src1, $src2}",
594 [(set EFLAGS, (X86cmp FR32:$src1, FR32:$src2))]>;
595 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
596 "ucomiss\t{$src2, $src1|$src1, $src2}",
597 [(set EFLAGS, (X86cmp FR32:$src1, (loadf32 addr:$src2)))]>;
599 def COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
600 "comiss\t{$src2, $src1|$src1, $src2}", []>;
601 def COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
602 "comiss\t{$src2, $src1|$src1, $src2}", []>;
606 // Aliases to match intrinsics which expect XMM operand(s).
607 let Constraints = "$src1 = $dst" in {
608 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
610 (ins VR128:$src1, VR128:$src, SSECC:$cc),
611 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
612 [(set VR128:$dst, (int_x86_sse_cmp_ss
614 VR128:$src, imm:$cc))]>;
615 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
617 (ins VR128:$src1, f32mem:$src, SSECC:$cc),
618 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
619 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
620 (load addr:$src), imm:$cc))]>;
623 let Defs = [EFLAGS] in {
624 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
625 "ucomiss\t{$src2, $src1|$src1, $src2}",
626 [(set EFLAGS, (X86ucomi (v4f32 VR128:$src1),
628 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
629 "ucomiss\t{$src2, $src1|$src1, $src2}",
630 [(set EFLAGS, (X86ucomi (v4f32 VR128:$src1),
631 (load addr:$src2)))]>;
633 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
634 "comiss\t{$src2, $src1|$src1, $src2}",
635 [(set EFLAGS, (X86comi (v4f32 VR128:$src1),
637 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
638 "comiss\t{$src2, $src1|$src1, $src2}",
639 [(set EFLAGS, (X86comi (v4f32 VR128:$src1),
640 (load addr:$src2)))]>;
643 // Aliases of packed SSE1 instructions for scalar use. These all have names
644 // that start with 'Fs'.
646 // Alias instructions that map fld0 to pxor for sse.
647 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
649 // FIXME: Set encoding to pseudo!
650 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
651 [(set FR32:$dst, fp32imm0)]>,
652 Requires<[HasSSE1]>, TB, OpSize;
654 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
656 let neverHasSideEffects = 1 in
657 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
658 "movaps\t{$src, $dst|$dst, $src}", []>;
660 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
662 let canFoldAsLoad = 1, isReMaterializable = 1 in
663 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
664 "movaps\t{$src, $dst|$dst, $src}",
665 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
667 /// sse12_fp_alias_pack_logical - SSE 1 & 2 aliased packed FP logical ops
669 multiclass sse12_fp_alias_pack_logical<bits<8> opc, string OpcodeStr,
670 SDNode OpNode, bit MayLoad = 0> {
671 let isAsmParserOnly = 1 in {
672 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
673 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode, FR32,
674 f32, f128mem, memopfsf32, SSEPackedSingle, MayLoad>, VEX_4V;
676 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
677 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode, FR64,
678 f64, f128mem, memopfsf64, SSEPackedDouble, MayLoad>, OpSize,
682 let Constraints = "$src1 = $dst" in {
683 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
684 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, FR32, f32,
685 f128mem, memopfsf32, SSEPackedSingle, MayLoad>, TB;
687 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
688 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, FR64, f64,
689 f128mem, memopfsf64, SSEPackedDouble, MayLoad>, TB, OpSize;
693 // Alias bitwise logical operations using SSE logical ops on packed FP values.
694 defm FsAND : sse12_fp_alias_pack_logical<0x54, "and", X86fand>;
695 defm FsOR : sse12_fp_alias_pack_logical<0x56, "or", X86for>;
696 defm FsXOR : sse12_fp_alias_pack_logical<0x57, "xor", X86fxor>;
698 let neverHasSideEffects = 1, Pattern = []<dag>, isCommutable = 0 in
699 defm FsANDN : sse12_fp_alias_pack_logical<0x55, "andn", undef, 1>;
701 /// basic_sse12_fp_binop_rm - SSE 1 & 2 binops come in both scalar and
704 /// In addition, we also have a special variant of the scalar form here to
705 /// represent the associated intrinsic operation. This form is unlike the
706 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
707 /// and leaves the top elements unmodified (therefore these cannot be commuted).
709 /// These three forms can each be reg+reg or reg+mem, so there are a total of
710 /// six "instructions".
712 multiclass basic_sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
715 let isAsmParserOnly = 1 in {
716 defm V#NAME#SS : sse12_fp_scalar<opc,
717 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
718 OpNode, FR32, f32mem>, XS, VEX_4V;
720 defm V#NAME#SD : sse12_fp_scalar<opc,
721 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
722 OpNode, FR64, f64mem>, XD, VEX_4V;
724 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
725 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
726 VR128, v4f32, f128mem, memopv4f32, SSEPackedSingle>,
729 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
730 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
731 VR128, v2f64, f128mem, memopv2f64, SSEPackedDouble>,
734 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
735 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
736 "", "_ss", ssmem, sse_load_f32>, XS, VEX_4V;
738 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
739 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
740 "2", "_sd", sdmem, sse_load_f64>, XD, VEX_4V;
743 let Constraints = "$src1 = $dst" in {
744 defm SS : sse12_fp_scalar<opc,
745 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
746 OpNode, FR32, f32mem>, XS;
748 defm SD : sse12_fp_scalar<opc,
749 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
750 OpNode, FR64, f64mem>, XD;
752 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
753 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v4f32,
754 f128mem, memopv4f32, SSEPackedSingle>, TB;
756 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
757 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v2f64,
758 f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
760 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
761 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
762 "", "_ss", ssmem, sse_load_f32>, XS;
764 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
765 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
766 "2", "_sd", sdmem, sse_load_f64>, XD;
770 // Arithmetic instructions
771 defm ADD : basic_sse12_fp_binop_rm<0x58, "add", fadd>;
772 defm MUL : basic_sse12_fp_binop_rm<0x59, "mul", fmul>;
774 let isCommutable = 0 in {
775 defm SUB : basic_sse12_fp_binop_rm<0x5C, "sub", fsub>;
776 defm DIV : basic_sse12_fp_binop_rm<0x5E, "div", fdiv>;
779 /// sse12_fp_binop_rm - Other SSE 1 & 2 binops
781 /// This multiclass is like basic_sse12_fp_binop_rm, with the addition of
782 /// instructions for a full-vector intrinsic form. Operations that map
783 /// onto C operators don't use this form since they just use the plain
784 /// vector form instead of having a separate vector intrinsic form.
786 /// This provides a total of eight "instructions".
788 multiclass sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
791 let isAsmParserOnly = 1 in {
792 // Scalar operation, reg+reg.
793 defm V#NAME#SS : sse12_fp_scalar<opc,
794 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
795 OpNode, FR32, f32mem>, XS, VEX_4V;
797 defm V#NAME#SD : sse12_fp_scalar<opc,
798 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
799 OpNode, FR64, f64mem>, XD, VEX_4V;
801 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
802 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
803 VR128, v4f32, f128mem, memopv4f32, SSEPackedSingle>,
806 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
807 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), OpNode,
808 VR128, v2f64, f128mem, memopv2f64, SSEPackedDouble>,
811 defm V#NAME#SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
812 !strconcat(OpcodeStr, "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
813 "", "_ss", ssmem, sse_load_f32>, XS, VEX_4V;
815 defm V#NAME#SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
816 !strconcat(OpcodeStr, "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
817 "2", "_sd", sdmem, sse_load_f64>, XD, VEX_4V;
819 defm V#NAME#PS : sse12_fp_packed_int<opc, OpcodeStr, VR128,
820 !strconcat(OpcodeStr, "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
821 "", "_ps", f128mem, memopv4f32, SSEPackedSingle>, VEX_4V;
823 defm V#NAME#PD : sse12_fp_packed_int<opc, OpcodeStr, VR128,
824 !strconcat(OpcodeStr, "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
825 "2", "_pd", f128mem, memopv2f64, SSEPackedDouble>, OpSize,
829 let Constraints = "$src1 = $dst" in {
830 // Scalar operation, reg+reg.
831 defm SS : sse12_fp_scalar<opc,
832 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
833 OpNode, FR32, f32mem>, XS;
834 defm SD : sse12_fp_scalar<opc,
835 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
836 OpNode, FR64, f64mem>, XD;
837 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr,
838 "ps\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v4f32,
839 f128mem, memopv4f32, SSEPackedSingle>, TB;
841 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr,
842 "pd\t{$src2, $dst|$dst, $src2}"), OpNode, VR128, v2f64,
843 f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
845 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
846 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
847 "", "_ss", ssmem, sse_load_f32>, XS;
849 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
850 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
851 "2", "_sd", sdmem, sse_load_f64>, XD;
853 defm PS : sse12_fp_packed_int<opc, OpcodeStr, VR128,
854 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
855 "", "_ps", f128mem, memopv4f32, SSEPackedSingle>, TB;
857 defm PD : sse12_fp_packed_int<opc, OpcodeStr, VR128,
858 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
859 "2", "_pd", f128mem, memopv2f64, SSEPackedDouble>, TB, OpSize;
863 let isCommutable = 0 in {
864 defm MAX : sse12_fp_binop_rm<0x5F, "max", X86fmax>;
865 defm MIN : sse12_fp_binop_rm<0x5D, "min", X86fmin>;
868 //===----------------------------------------------------------------------===//
869 // SSE packed FP Instructions
872 let neverHasSideEffects = 1 in
873 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
874 "movaps\t{$src, $dst|$dst, $src}", []>;
875 let canFoldAsLoad = 1, isReMaterializable = 1 in
876 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
877 "movaps\t{$src, $dst|$dst, $src}",
878 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
880 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
881 "movaps\t{$src, $dst|$dst, $src}",
882 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
884 let neverHasSideEffects = 1 in
885 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
886 "movups\t{$src, $dst|$dst, $src}", []>;
887 let canFoldAsLoad = 1, isReMaterializable = 1 in
888 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
889 "movups\t{$src, $dst|$dst, $src}",
890 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
891 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
892 "movups\t{$src, $dst|$dst, $src}",
893 [(store (v4f32 VR128:$src), addr:$dst)]>;
895 // Intrinsic forms of MOVUPS load and store
896 let canFoldAsLoad = 1, isReMaterializable = 1 in
897 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
898 "movups\t{$src, $dst|$dst, $src}",
899 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
900 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
901 "movups\t{$src, $dst|$dst, $src}",
902 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
904 let Constraints = "$src1 = $dst" in {
905 let AddedComplexity = 20 in {
906 def MOVLPSrm : PSI<0x12, MRMSrcMem,
907 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
908 "movlps\t{$src2, $dst|$dst, $src2}",
911 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
912 def MOVHPSrm : PSI<0x16, MRMSrcMem,
913 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
914 "movhps\t{$src2, $dst|$dst, $src2}",
916 (movlhps VR128:$src1,
917 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
919 } // Constraints = "$src1 = $dst"
922 def : Pat<(movlhps VR128:$src1, (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
923 (MOVHPSrm (v4i32 VR128:$src1), addr:$src2)>;
925 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
926 "movlps\t{$src, $dst|$dst, $src}",
927 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
928 (iPTR 0))), addr:$dst)]>;
930 // v2f64 extract element 1 is always custom lowered to unpack high to low
931 // and extract element 0 so the non-store version isn't too horrible.
932 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
933 "movhps\t{$src, $dst|$dst, $src}",
934 [(store (f64 (vector_extract
935 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
936 (undef)), (iPTR 0))), addr:$dst)]>;
938 let Constraints = "$src1 = $dst" in {
939 let AddedComplexity = 20 in {
940 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
941 (ins VR128:$src1, VR128:$src2),
942 "movlhps\t{$src2, $dst|$dst, $src2}",
944 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
946 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
947 (ins VR128:$src1, VR128:$src2),
948 "movhlps\t{$src2, $dst|$dst, $src2}",
950 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
952 } // Constraints = "$src1 = $dst"
954 let AddedComplexity = 20 in {
955 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
956 (MOVLHPSrr (v4f32 VR128:$src), (v4f32 VR128:$src))>;
957 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
958 (MOVLHPSrr (v2i64 VR128:$src), (v2i64 VR128:$src))>;
965 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
967 /// In addition, we also have a special variant of the scalar form here to
968 /// represent the associated intrinsic operation. This form is unlike the
969 /// plain scalar form, in that it takes an entire vector (instead of a
970 /// scalar) and leaves the top elements undefined.
972 /// And, we have a special variant form for a full-vector intrinsic form.
974 /// These four forms can each have a reg or a mem operand, so there are a
975 /// total of eight "instructions".
977 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
981 bit Commutable = 0> {
982 // Scalar operation, reg.
983 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
984 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
985 [(set FR32:$dst, (OpNode FR32:$src))]> {
986 let isCommutable = Commutable;
989 // Scalar operation, mem.
990 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
991 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
992 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
993 Requires<[HasSSE1, OptForSize]>;
995 // Vector operation, reg.
996 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
997 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
998 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
999 let isCommutable = Commutable;
1002 // Vector operation, mem.
1003 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1004 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1005 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
1007 // Intrinsic operation, reg.
1008 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1009 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1010 [(set VR128:$dst, (F32Int VR128:$src))]> {
1011 let isCommutable = Commutable;
1014 // Intrinsic operation, mem.
1015 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
1016 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1017 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
1019 // Vector intrinsic operation, reg
1020 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1021 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1022 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
1023 let isCommutable = Commutable;
1026 // Vector intrinsic operation, mem
1027 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1028 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1029 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
1033 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
1034 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
1036 // Reciprocal approximations. Note that these typically require refinement
1037 // in order to obtain suitable precision.
1038 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
1039 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
1040 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
1041 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
1043 /// sse12_fp_pack_logical - SSE 1 & 2 packed FP logical ops
1045 multiclass sse12_fp_pack_logical<bits<8> opc, string OpcodeStr,
1046 SDNode OpNode, int HasPat = 0,
1048 list<list<dag>> Pattern = []> {
1049 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
1050 (ins VR128:$src1, VR128:$src2),
1051 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
1052 !if(HasPat, Pattern[0],
1053 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1,
1055 { let isCommutable = Commutable; }
1057 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1058 (ins VR128:$src1, VR128:$src2),
1059 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1060 !if(HasPat, Pattern[1],
1061 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1062 (bc_v2i64 (v2f64 VR128:$src2))))])>
1063 { let isCommutable = Commutable; }
1065 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
1066 (ins VR128:$src1, f128mem:$src2),
1067 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
1068 !if(HasPat, Pattern[2],
1069 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
1070 (memopv2i64 addr:$src2)))])>;
1072 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1073 (ins VR128:$src1, f128mem:$src2),
1074 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1075 !if(HasPat, Pattern[3],
1076 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1077 (memopv2i64 addr:$src2)))])>;
1081 let Constraints = "$src1 = $dst" in {
1082 defm AND : sse12_fp_pack_logical<0x54, "and", and>;
1083 defm OR : sse12_fp_pack_logical<0x56, "or", or>;
1084 defm XOR : sse12_fp_pack_logical<0x57, "xor", xor>;
1085 defm ANDN : sse12_fp_pack_logical<0x55, "andn", undef /* dummy */, 1, 0, [
1087 [(set VR128:$dst, (v2i64 (and (xor VR128:$src1,
1088 (bc_v2i64 (v4i32 immAllOnesV))),
1091 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1092 (bc_v2i64 (v2f64 VR128:$src2))))],
1094 [(set VR128:$dst, (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
1095 (bc_v2i64 (v4i32 immAllOnesV))),
1096 (memopv2i64 addr:$src2))))],
1098 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1099 (memopv2i64 addr:$src2)))]]>;
1102 let Constraints = "$src1 = $dst" in {
1103 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
1104 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1105 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1106 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1107 VR128:$src, imm:$cc))]>;
1108 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
1109 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1110 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1111 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1112 (memop addr:$src), imm:$cc))]>;
1114 // Accept explicit immediate argument form instead of comparison code.
1115 let isAsmParserOnly = 1 in {
1116 def CMPPSrri_alt : PSIi8<0xC2, MRMSrcReg,
1117 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1118 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1119 def CMPPSrmi_alt : PSIi8<0xC2, MRMSrcMem,
1120 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1121 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1124 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1125 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
1126 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1127 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
1129 // Shuffle and unpack instructions
1130 let Constraints = "$src1 = $dst" in {
1131 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
1132 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
1133 (outs VR128:$dst), (ins VR128:$src1,
1134 VR128:$src2, i8imm:$src3),
1135 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1137 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1138 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
1139 (outs VR128:$dst), (ins VR128:$src1,
1140 f128mem:$src2, i8imm:$src3),
1141 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1144 VR128:$src1, (memopv4f32 addr:$src2))))]>;
1146 let AddedComplexity = 10 in {
1147 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
1148 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1149 "unpckhps\t{$src2, $dst|$dst, $src2}",
1151 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
1152 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
1153 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1154 "unpckhps\t{$src2, $dst|$dst, $src2}",
1156 (v4f32 (unpckh VR128:$src1,
1157 (memopv4f32 addr:$src2))))]>;
1159 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
1160 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1161 "unpcklps\t{$src2, $dst|$dst, $src2}",
1163 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
1164 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
1165 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1166 "unpcklps\t{$src2, $dst|$dst, $src2}",
1168 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
1169 } // AddedComplexity
1170 } // Constraints = "$src1 = $dst"
1173 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1174 "movmskps\t{$src, $dst|$dst, $src}",
1175 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
1176 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1177 "movmskpd\t{$src, $dst|$dst, $src}",
1178 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1180 // Prefetch intrinsic.
1181 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1182 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1183 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1184 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1185 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1186 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1187 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1188 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1190 // Non-temporal stores
1191 def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1192 "movntps\t{$src, $dst|$dst, $src}",
1193 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1195 let AddedComplexity = 400 in { // Prefer non-temporal versions
1196 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1197 "movntps\t{$src, $dst|$dst, $src}",
1198 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
1200 def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1201 "movntdq\t{$src, $dst|$dst, $src}",
1202 [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
1204 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
1205 "movnti\t{$src, $dst|$dst, $src}",
1206 [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
1207 TB, Requires<[HasSSE2]>;
1209 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
1210 "movnti\t{$src, $dst|$dst, $src}",
1211 [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
1212 TB, Requires<[HasSSE2]>;
1215 // Load, store, and memory fence
1216 def SFENCE : I<0xAE, MRM_F8, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>,
1217 TB, Requires<[HasSSE1]>;
1220 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1221 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1222 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1223 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1225 // Alias instructions that map zero vector to pxor / xorp* for sse.
1226 // We set canFoldAsLoad because this can be converted to a constant-pool
1227 // load of an all-zeros value if folding it would be beneficial.
1228 // FIXME: Change encoding to pseudo!
1229 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1230 isCodeGenOnly = 1 in {
1231 def V_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1232 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
1233 def V_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1234 [(set VR128:$dst, (v2f64 immAllZerosV))]>;
1235 let ExeDomain = SSEPackedInt in
1236 def V_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
1237 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1240 def : Pat<(v2i64 immAllZerosV), (V_SET0PI)>;
1241 def : Pat<(v8i16 immAllZerosV), (V_SET0PI)>;
1242 def : Pat<(v16i8 immAllZerosV), (V_SET0PI)>;
1244 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1245 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
1247 //===---------------------------------------------------------------------===//
1248 // SSE2 Instructions
1249 //===---------------------------------------------------------------------===//
1251 // Move Instructions. Register-to-register movsd is not used for FR64
1252 // register copies because it's a partial register update; FsMOVAPDrr is
1253 // used instead. Register-to-register movsd is not modeled as an INSERT_SUBREG
1254 // because INSERT_SUBREG requires that the insert be implementable in terms of
1255 // a copy, and just mentioned, we don't use movsd for copies.
1256 let Constraints = "$src1 = $dst" in
1257 def MOVSDrr : SDI<0x10, MRMSrcReg,
1258 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
1259 "movsd\t{$src2, $dst|$dst, $src2}",
1260 [(set (v2f64 VR128:$dst),
1261 (movl VR128:$src1, (scalar_to_vector FR64:$src2)))]>;
1263 // Extract the low 64-bit value from one vector and insert it into another.
1264 let AddedComplexity = 15 in
1265 def : Pat<(v2f64 (movl VR128:$src1, VR128:$src2)),
1266 (MOVSDrr (v2f64 VR128:$src1),
1267 (EXTRACT_SUBREG (v2f64 VR128:$src2), sub_sd))>;
1269 // Implicitly promote a 64-bit scalar to a vector.
1270 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
1271 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), FR64:$src, sub_sd)>;
1273 // Loading from memory automatically zeroing upper bits.
1274 let canFoldAsLoad = 1, isReMaterializable = 1, AddedComplexity = 20 in
1275 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1276 "movsd\t{$src, $dst|$dst, $src}",
1277 [(set FR64:$dst, (loadf64 addr:$src))]>;
1279 // MOVSDrm zeros the high parts of the register; represent this
1280 // with SUBREG_TO_REG.
1281 let AddedComplexity = 20 in {
1282 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
1283 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1284 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
1285 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1286 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
1287 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1288 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
1289 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1290 def : Pat<(v2f64 (X86vzload addr:$src)),
1291 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1294 // Store scalar value to memory.
1295 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1296 "movsd\t{$src, $dst|$dst, $src}",
1297 [(store FR64:$src, addr:$dst)]>;
1299 // Extract and store.
1300 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
1303 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
1305 // Conversion instructions
1306 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1307 "cvttsd2si\t{$src, $dst|$dst, $src}",
1308 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1309 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1310 "cvttsd2si\t{$src, $dst|$dst, $src}",
1311 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1312 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1313 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1314 [(set FR32:$dst, (fround FR64:$src))]>;
1315 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1316 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1317 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
1318 Requires<[HasSSE2, OptForSize]>;
1319 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1320 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1321 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1322 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1323 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1324 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1326 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1327 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1328 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1329 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1330 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1331 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1332 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1333 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1334 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1335 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1336 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1337 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1338 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1339 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1340 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1341 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1342 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1343 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1344 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1345 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1347 // SSE2 instructions with XS prefix
1348 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1349 "cvtss2sd\t{$src, $dst|$dst, $src}",
1350 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1351 Requires<[HasSSE2]>;
1352 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1353 "cvtss2sd\t{$src, $dst|$dst, $src}",
1354 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1355 Requires<[HasSSE2, OptForSize]>;
1357 def : Pat<(extloadf32 addr:$src),
1358 (CVTSS2SDrr (MOVSSrm addr:$src))>,
1359 Requires<[HasSSE2, OptForSpeed]>;
1361 // Match intrinsics which expect XMM operand(s).
1362 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1363 "cvtsd2si\t{$src, $dst|$dst, $src}",
1364 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1365 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1366 "cvtsd2si\t{$src, $dst|$dst, $src}",
1367 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1368 (load addr:$src)))]>;
1370 // Match intrinsics which expect MM and XMM operand(s).
1371 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1372 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1373 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1374 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1375 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1376 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1377 (memop addr:$src)))]>;
1378 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1379 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1380 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1381 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1382 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1383 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1384 (memop addr:$src)))]>;
1385 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1386 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1387 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1388 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1389 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1390 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1391 (load addr:$src)))]>;
1393 // Aliases for intrinsics
1394 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1395 "cvttsd2si\t{$src, $dst|$dst, $src}",
1397 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1398 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1399 "cvttsd2si\t{$src, $dst|$dst, $src}",
1400 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1401 (load addr:$src)))]>;
1403 // Comparison instructions
1404 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1405 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1406 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1407 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1409 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1410 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1411 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1413 // Accept explicit immediate argument form instead of comparison code.
1414 let isAsmParserOnly = 1 in {
1415 def CMPSDrr_alt : SDIi8<0xC2, MRMSrcReg,
1416 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, i8imm:$src2),
1417 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1419 def CMPSDrm_alt : SDIi8<0xC2, MRMSrcMem,
1420 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, i8imm:$src2),
1421 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1425 let Defs = [EFLAGS] in {
1426 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1427 "ucomisd\t{$src2, $src1|$src1, $src2}",
1428 [(set EFLAGS, (X86cmp FR64:$src1, FR64:$src2))]>;
1429 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1430 "ucomisd\t{$src2, $src1|$src1, $src2}",
1431 [(set EFLAGS, (X86cmp FR64:$src1, (loadf64 addr:$src2)))]>;
1432 } // Defs = [EFLAGS]
1434 // Aliases to match intrinsics which expect XMM operand(s).
1435 let Constraints = "$src1 = $dst" in {
1436 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1438 (ins VR128:$src1, VR128:$src, SSECC:$cc),
1439 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1440 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1441 VR128:$src, imm:$cc))]>;
1442 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1444 (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1445 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1446 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1447 (load addr:$src), imm:$cc))]>;
1450 let Defs = [EFLAGS] in {
1451 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1452 "ucomisd\t{$src2, $src1|$src1, $src2}",
1453 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1455 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1456 "ucomisd\t{$src2, $src1|$src1, $src2}",
1457 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1458 (load addr:$src2)))]>;
1460 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1461 "comisd\t{$src2, $src1|$src1, $src2}",
1462 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1464 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1465 "comisd\t{$src2, $src1|$src1, $src2}",
1466 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1467 (load addr:$src2)))]>;
1468 } // Defs = [EFLAGS]
1470 // Aliases of packed SSE2 instructions for scalar use. These all have names
1471 // that start with 'Fs'.
1473 // Alias instructions that map fld0 to pxor for sse.
1474 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1475 canFoldAsLoad = 1 in
1476 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1477 [(set FR64:$dst, fpimm0)]>,
1478 Requires<[HasSSE2]>, TB, OpSize;
1480 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1482 let neverHasSideEffects = 1 in
1483 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1484 "movapd\t{$src, $dst|$dst, $src}", []>;
1486 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1488 let canFoldAsLoad = 1, isReMaterializable = 1 in
1489 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1490 "movapd\t{$src, $dst|$dst, $src}",
1491 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1493 //===---------------------------------------------------------------------===//
1494 // SSE packed FP Instructions
1496 // Move Instructions
1497 let neverHasSideEffects = 1 in
1498 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1499 "movapd\t{$src, $dst|$dst, $src}", []>;
1500 let canFoldAsLoad = 1, isReMaterializable = 1 in
1501 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1502 "movapd\t{$src, $dst|$dst, $src}",
1503 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1505 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1506 "movapd\t{$src, $dst|$dst, $src}",
1507 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1509 let neverHasSideEffects = 1 in
1510 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1511 "movupd\t{$src, $dst|$dst, $src}", []>;
1512 let canFoldAsLoad = 1 in
1513 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1514 "movupd\t{$src, $dst|$dst, $src}",
1515 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1516 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1517 "movupd\t{$src, $dst|$dst, $src}",
1518 [(store (v2f64 VR128:$src), addr:$dst)]>;
1520 // Intrinsic forms of MOVUPD load and store
1521 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1522 "movupd\t{$src, $dst|$dst, $src}",
1523 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1524 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1525 "movupd\t{$src, $dst|$dst, $src}",
1526 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1528 let Constraints = "$src1 = $dst" in {
1529 let AddedComplexity = 20 in {
1530 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1531 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1532 "movlpd\t{$src2, $dst|$dst, $src2}",
1534 (v2f64 (movlp VR128:$src1,
1535 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1536 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1537 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1538 "movhpd\t{$src2, $dst|$dst, $src2}",
1540 (v2f64 (movlhps VR128:$src1,
1541 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1542 } // AddedComplexity
1543 } // Constraints = "$src1 = $dst"
1545 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1546 "movlpd\t{$src, $dst|$dst, $src}",
1547 [(store (f64 (vector_extract (v2f64 VR128:$src),
1548 (iPTR 0))), addr:$dst)]>;
1550 // v2f64 extract element 1 is always custom lowered to unpack high to low
1551 // and extract element 0 so the non-store version isn't too horrible.
1552 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1553 "movhpd\t{$src, $dst|$dst, $src}",
1554 [(store (f64 (vector_extract
1555 (v2f64 (unpckh VR128:$src, (undef))),
1556 (iPTR 0))), addr:$dst)]>;
1558 // SSE2 instructions without OpSize prefix
1559 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1560 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1561 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1562 TB, Requires<[HasSSE2]>;
1563 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1564 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1565 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1566 (bitconvert (memopv2i64 addr:$src))))]>,
1567 TB, Requires<[HasSSE2]>;
1569 // SSE2 instructions with XS prefix
1570 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1571 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1572 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1573 XS, Requires<[HasSSE2]>;
1574 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1575 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1576 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1577 (bitconvert (memopv2i64 addr:$src))))]>,
1578 XS, Requires<[HasSSE2]>;
1580 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1581 "cvtps2dq\t{$src, $dst|$dst, $src}",
1582 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1583 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1584 "cvtps2dq\t{$src, $dst|$dst, $src}",
1585 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1586 (memop addr:$src)))]>;
1587 // SSE2 packed instructions with XS prefix
1588 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1589 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1590 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1591 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1593 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1594 "cvttps2dq\t{$src, $dst|$dst, $src}",
1596 (int_x86_sse2_cvttps2dq VR128:$src))]>,
1597 XS, Requires<[HasSSE2]>;
1598 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1599 "cvttps2dq\t{$src, $dst|$dst, $src}",
1600 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1601 (memop addr:$src)))]>,
1602 XS, Requires<[HasSSE2]>;
1604 // SSE2 packed instructions with XD prefix
1605 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1606 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1607 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1608 XD, Requires<[HasSSE2]>;
1609 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1610 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1611 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1612 (memop addr:$src)))]>,
1613 XD, Requires<[HasSSE2]>;
1615 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1616 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1617 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1618 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1619 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1620 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1621 (memop addr:$src)))]>;
1623 // SSE2 instructions without OpSize prefix
1624 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1625 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1626 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1627 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1629 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1630 "cvtps2pd\t{$src, $dst|$dst, $src}",
1631 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1632 TB, Requires<[HasSSE2]>;
1633 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1634 "cvtps2pd\t{$src, $dst|$dst, $src}",
1635 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1636 (load addr:$src)))]>,
1637 TB, Requires<[HasSSE2]>;
1639 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1640 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1641 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1642 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1645 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1646 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1647 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1648 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1649 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1650 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1651 (memop addr:$src)))]>;
1653 // Match intrinsics which expect XMM operand(s).
1654 // Aliases for intrinsics
1655 let Constraints = "$src1 = $dst" in {
1656 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1657 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1658 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1659 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1661 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1662 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1663 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1664 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1665 (loadi32 addr:$src2)))]>;
1666 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1667 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1668 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1669 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1671 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1672 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1673 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1674 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1675 (load addr:$src2)))]>;
1676 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1677 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1678 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1679 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1680 VR128:$src2))]>, XS,
1681 Requires<[HasSSE2]>;
1682 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1683 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1684 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1685 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1686 (load addr:$src2)))]>, XS,
1687 Requires<[HasSSE2]>;
1692 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1694 /// In addition, we also have a special variant of the scalar form here to
1695 /// represent the associated intrinsic operation. This form is unlike the
1696 /// plain scalar form, in that it takes an entire vector (instead of a
1697 /// scalar) and leaves the top elements undefined.
1699 /// And, we have a special variant form for a full-vector intrinsic form.
1701 /// These four forms can each have a reg or a mem operand, so there are a
1702 /// total of eight "instructions".
1704 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1708 bit Commutable = 0> {
1709 // Scalar operation, reg.
1710 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1711 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1712 [(set FR64:$dst, (OpNode FR64:$src))]> {
1713 let isCommutable = Commutable;
1716 // Scalar operation, mem.
1717 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1718 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1719 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1721 // Vector operation, reg.
1722 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1723 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1724 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1725 let isCommutable = Commutable;
1728 // Vector operation, mem.
1729 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1730 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1731 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1733 // Intrinsic operation, reg.
1734 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1735 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1736 [(set VR128:$dst, (F64Int VR128:$src))]> {
1737 let isCommutable = Commutable;
1740 // Intrinsic operation, mem.
1741 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1742 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1743 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1745 // Vector intrinsic operation, reg
1746 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1747 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1748 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1749 let isCommutable = Commutable;
1752 // Vector intrinsic operation, mem
1753 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1754 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1755 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1759 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1760 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1762 // There is no f64 version of the reciprocal approximation instructions.
1764 let Constraints = "$src1 = $dst" in {
1765 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1766 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1767 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1768 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1769 VR128:$src, imm:$cc))]>;
1770 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1771 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1772 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1773 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1774 (memop addr:$src), imm:$cc))]>;
1776 // Accept explicit immediate argument form instead of comparison code.
1777 let isAsmParserOnly = 1 in {
1778 def CMPPDrri_alt : PDIi8<0xC2, MRMSrcReg,
1779 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1780 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1781 def CMPPDrmi_alt : PDIi8<0xC2, MRMSrcMem,
1782 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1783 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1786 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1787 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1788 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1789 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1791 // Shuffle and unpack instructions
1792 let Constraints = "$src1 = $dst" in {
1793 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1794 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1795 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1797 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1798 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1799 (outs VR128:$dst), (ins VR128:$src1,
1800 f128mem:$src2, i8imm:$src3),
1801 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1804 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1806 let AddedComplexity = 10 in {
1807 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1808 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1809 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1811 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1812 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1813 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1814 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1816 (v2f64 (unpckh VR128:$src1,
1817 (memopv2f64 addr:$src2))))]>;
1819 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1820 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1821 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1823 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1824 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1825 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1826 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1828 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1829 } // AddedComplexity
1830 } // Constraints = "$src1 = $dst"
1833 //===---------------------------------------------------------------------===//
1834 // SSE integer instructions
1835 let ExeDomain = SSEPackedInt in {
1837 // Move Instructions
1838 let neverHasSideEffects = 1 in
1839 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1840 "movdqa\t{$src, $dst|$dst, $src}", []>;
1841 let canFoldAsLoad = 1, mayLoad = 1 in
1842 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1843 "movdqa\t{$src, $dst|$dst, $src}",
1844 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1846 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1847 "movdqa\t{$src, $dst|$dst, $src}",
1848 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1849 let canFoldAsLoad = 1, mayLoad = 1 in
1850 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1851 "movdqu\t{$src, $dst|$dst, $src}",
1852 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1853 XS, Requires<[HasSSE2]>;
1855 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1856 "movdqu\t{$src, $dst|$dst, $src}",
1857 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1858 XS, Requires<[HasSSE2]>;
1860 // Intrinsic forms of MOVDQU load and store
1861 let canFoldAsLoad = 1 in
1862 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1863 "movdqu\t{$src, $dst|$dst, $src}",
1864 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1865 XS, Requires<[HasSSE2]>;
1866 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1867 "movdqu\t{$src, $dst|$dst, $src}",
1868 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1869 XS, Requires<[HasSSE2]>;
1871 let Constraints = "$src1 = $dst" in {
1873 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1874 bit Commutable = 0> {
1875 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1876 (ins VR128:$src1, VR128:$src2),
1877 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1878 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1879 let isCommutable = Commutable;
1881 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1882 (ins VR128:$src1, i128mem:$src2),
1883 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1884 [(set VR128:$dst, (IntId VR128:$src1,
1885 (bitconvert (memopv2i64
1889 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1891 Intrinsic IntId, Intrinsic IntId2> {
1892 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1893 (ins VR128:$src1, VR128:$src2),
1894 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1895 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1896 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1897 (ins VR128:$src1, i128mem:$src2),
1898 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1899 [(set VR128:$dst, (IntId VR128:$src1,
1900 (bitconvert (memopv2i64 addr:$src2))))]>;
1901 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
1902 (ins VR128:$src1, i32i8imm:$src2),
1903 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1904 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1907 /// PDI_binop_rm - Simple SSE2 binary operator.
1908 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1909 ValueType OpVT, bit Commutable = 0> {
1910 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1911 (ins VR128:$src1, VR128:$src2),
1912 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1913 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1914 let isCommutable = Commutable;
1916 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1917 (ins VR128:$src1, i128mem:$src2),
1918 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1919 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1920 (bitconvert (memopv2i64 addr:$src2)))))]>;
1923 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1925 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1926 /// to collapse (bitconvert VT to VT) into its operand.
1928 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1929 bit Commutable = 0> {
1930 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1931 (ins VR128:$src1, VR128:$src2),
1932 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1933 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1934 let isCommutable = Commutable;
1936 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1937 (ins VR128:$src1, i128mem:$src2),
1938 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1939 [(set VR128:$dst, (OpNode VR128:$src1,
1940 (memopv2i64 addr:$src2)))]>;
1943 } // Constraints = "$src1 = $dst"
1944 } // ExeDomain = SSEPackedInt
1946 // 128-bit Integer Arithmetic
1948 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1949 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1950 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1951 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1953 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1954 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1955 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1956 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1958 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1959 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1960 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1961 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1963 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1964 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1965 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1966 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1968 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
1970 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
1971 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
1972 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
1974 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
1976 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
1977 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
1980 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
1981 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
1982 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
1983 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
1984 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
1987 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
1988 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
1989 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
1990 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
1991 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
1992 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
1994 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
1995 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
1996 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
1997 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
1998 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
1999 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2001 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2002 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2003 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2004 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2006 // 128-bit logical shifts.
2007 let Constraints = "$src1 = $dst", neverHasSideEffects = 1,
2008 ExeDomain = SSEPackedInt in {
2009 def PSLLDQri : PDIi8<0x73, MRM7r,
2010 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2011 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2012 def PSRLDQri : PDIi8<0x73, MRM3r,
2013 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2014 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2015 // PSRADQri doesn't exist in SSE[1-3].
2018 let Predicates = [HasSSE2] in {
2019 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2020 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2021 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2022 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2023 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2024 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2025 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2026 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2027 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2028 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2030 // Shift up / down and insert zero's.
2031 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2032 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2033 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2034 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2038 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2039 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2040 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2042 let Constraints = "$src1 = $dst", ExeDomain = SSEPackedInt in {
2043 def PANDNrr : PDI<0xDF, MRMSrcReg,
2044 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2045 "pandn\t{$src2, $dst|$dst, $src2}",
2046 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2049 def PANDNrm : PDI<0xDF, MRMSrcMem,
2050 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2051 "pandn\t{$src2, $dst|$dst, $src2}",
2052 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2053 (memopv2i64 addr:$src2))))]>;
2056 // SSE2 Integer comparison
2057 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2058 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2059 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2060 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2061 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2062 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2064 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2065 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2066 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2067 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2068 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2069 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2070 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2071 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2072 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2073 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2074 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2075 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2077 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2078 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2079 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2080 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2081 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2082 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2083 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2084 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2085 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2086 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2087 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2088 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2091 // Pack instructions
2092 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2093 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2094 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2096 let ExeDomain = SSEPackedInt in {
2098 // Shuffle and unpack instructions
2099 let AddedComplexity = 5 in {
2100 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2101 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2102 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2103 [(set VR128:$dst, (v4i32 (pshufd:$src2
2104 VR128:$src1, (undef))))]>;
2105 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2106 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2107 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2108 [(set VR128:$dst, (v4i32 (pshufd:$src2
2109 (bc_v4i32 (memopv2i64 addr:$src1)),
2113 // SSE2 with ImmT == Imm8 and XS prefix.
2114 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2115 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2116 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2117 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2119 XS, Requires<[HasSSE2]>;
2120 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2121 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2122 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2123 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2124 (bc_v8i16 (memopv2i64 addr:$src1)),
2126 XS, Requires<[HasSSE2]>;
2128 // SSE2 with ImmT == Imm8 and XD prefix.
2129 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2130 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2131 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2132 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2134 XD, Requires<[HasSSE2]>;
2135 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2136 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2137 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2138 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2139 (bc_v8i16 (memopv2i64 addr:$src1)),
2141 XD, Requires<[HasSSE2]>;
2143 // Unpack instructions
2144 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
2145 PatFrag unp_frag, PatFrag bc_frag> {
2146 def rr : PDI<opc, MRMSrcReg,
2147 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2148 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2149 [(set VR128:$dst, (vt (unp_frag VR128:$src1, VR128:$src2)))]>;
2150 def rm : PDI<opc, MRMSrcMem,
2151 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2152 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2153 [(set VR128:$dst, (unp_frag VR128:$src1,
2154 (bc_frag (memopv2i64
2158 let Constraints = "$src1 = $dst" in {
2159 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, unpckl, bc_v16i8>;
2160 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, unpckl, bc_v8i16>;
2161 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, unpckl, bc_v4i32>;
2163 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2164 /// knew to collapse (bitconvert VT to VT) into its operand.
2165 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2166 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2167 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2169 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2170 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2171 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2172 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2174 (v2i64 (unpckl VR128:$src1,
2175 (memopv2i64 addr:$src2))))]>;
2177 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, unpckh, bc_v16i8>;
2178 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, unpckh, bc_v8i16>;
2179 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, unpckh, bc_v4i32>;
2181 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2182 /// knew to collapse (bitconvert VT to VT) into its operand.
2183 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2184 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2185 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2187 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2188 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2189 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2190 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2192 (v2i64 (unpckh VR128:$src1,
2193 (memopv2i64 addr:$src2))))]>;
2197 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2198 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2199 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2200 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2202 let Constraints = "$src1 = $dst" in {
2203 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2204 (outs VR128:$dst), (ins VR128:$src1,
2205 GR32:$src2, i32i8imm:$src3),
2206 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2208 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2209 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2210 (outs VR128:$dst), (ins VR128:$src1,
2211 i16mem:$src2, i32i8imm:$src3),
2212 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2214 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2219 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2220 "pmovmskb\t{$src, $dst|$dst, $src}",
2221 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2223 // Conditional store
2225 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2226 "maskmovdqu\t{$mask, $src|$src, $mask}",
2227 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2230 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2231 "maskmovdqu\t{$mask, $src|$src, $mask}",
2232 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2234 } // ExeDomain = SSEPackedInt
2236 // Non-temporal stores
2237 def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2238 "movntpd\t{$src, $dst|$dst, $src}",
2239 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2240 let ExeDomain = SSEPackedInt in
2241 def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2242 "movntdq\t{$src, $dst|$dst, $src}",
2243 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2244 def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2245 "movnti\t{$src, $dst|$dst, $src}",
2246 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2247 TB, Requires<[HasSSE2]>;
2249 let AddedComplexity = 400 in { // Prefer non-temporal versions
2250 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2251 "movntpd\t{$src, $dst|$dst, $src}",
2252 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
2254 let ExeDomain = SSEPackedInt in
2255 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2256 "movntdq\t{$src, $dst|$dst, $src}",
2257 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2261 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2262 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2263 TB, Requires<[HasSSE2]>;
2265 // Load, store, and memory fence
2266 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
2267 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2268 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
2269 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2271 // Pause. This "instruction" is encoded as "rep; nop", so even though it
2272 // was introduced with SSE2, it's backward compatible.
2273 def PAUSE : I<0x90, RawFrm, (outs), (ins), "pause", []>, REP;
2275 //TODO: custom lower this so as to never even generate the noop
2276 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2278 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2279 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2280 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2283 // Alias instructions that map zero vector to pxor / xorp* for sse.
2284 // We set canFoldAsLoad because this can be converted to a constant-pool
2285 // load of an all-ones value if folding it would be beneficial.
2286 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2287 isCodeGenOnly = 1, ExeDomain = SSEPackedInt in
2288 // FIXME: Change encoding to pseudo.
2289 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
2290 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2292 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2293 "movd\t{$src, $dst|$dst, $src}",
2295 (v4i32 (scalar_to_vector GR32:$src)))]>;
2296 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2297 "movd\t{$src, $dst|$dst, $src}",
2299 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2301 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2302 "movd\t{$src, $dst|$dst, $src}",
2303 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2305 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2306 "movd\t{$src, $dst|$dst, $src}",
2307 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2309 // SSE2 instructions with XS prefix
2310 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2311 "movq\t{$src, $dst|$dst, $src}",
2313 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2314 Requires<[HasSSE2]>;
2315 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2316 "movq\t{$src, $dst|$dst, $src}",
2317 [(store (i64 (vector_extract (v2i64 VR128:$src),
2318 (iPTR 0))), addr:$dst)]>;
2320 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
2321 (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
2323 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2324 "movd\t{$src, $dst|$dst, $src}",
2325 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2327 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2328 "movd\t{$src, $dst|$dst, $src}",
2329 [(store (i32 (vector_extract (v4i32 VR128:$src),
2330 (iPTR 0))), addr:$dst)]>;
2332 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2333 "movd\t{$src, $dst|$dst, $src}",
2334 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2335 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2336 "movd\t{$src, $dst|$dst, $src}",
2337 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2339 // Store / copy lower 64-bits of a XMM register.
2340 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2341 "movq\t{$src, $dst|$dst, $src}",
2342 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2344 // movd / movq to XMM register zero-extends
2345 let AddedComplexity = 15 in {
2346 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2347 "movd\t{$src, $dst|$dst, $src}",
2348 [(set VR128:$dst, (v4i32 (X86vzmovl
2349 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2350 // This is X86-64 only.
2351 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2352 "mov{d|q}\t{$src, $dst|$dst, $src}",
2353 [(set VR128:$dst, (v2i64 (X86vzmovl
2354 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2357 let AddedComplexity = 20 in {
2358 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2359 "movd\t{$src, $dst|$dst, $src}",
2361 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2362 (loadi32 addr:$src))))))]>;
2364 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2365 (MOVZDI2PDIrm addr:$src)>;
2366 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2367 (MOVZDI2PDIrm addr:$src)>;
2368 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2369 (MOVZDI2PDIrm addr:$src)>;
2371 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2372 "movq\t{$src, $dst|$dst, $src}",
2374 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2375 (loadi64 addr:$src))))))]>, XS,
2376 Requires<[HasSSE2]>;
2378 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2379 (MOVZQI2PQIrm addr:$src)>;
2380 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2381 (MOVZQI2PQIrm addr:$src)>;
2382 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2385 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2386 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2387 let AddedComplexity = 15 in
2388 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2389 "movq\t{$src, $dst|$dst, $src}",
2390 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2391 XS, Requires<[HasSSE2]>;
2393 let AddedComplexity = 20 in {
2394 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2395 "movq\t{$src, $dst|$dst, $src}",
2396 [(set VR128:$dst, (v2i64 (X86vzmovl
2397 (loadv2i64 addr:$src))))]>,
2398 XS, Requires<[HasSSE2]>;
2400 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2401 (MOVZPQILo2PQIrm addr:$src)>;
2404 // Instructions for the disassembler
2405 // xr = XMM register
2408 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2409 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2411 //===---------------------------------------------------------------------===//
2412 // SSE3 Instructions
2413 //===---------------------------------------------------------------------===//
2415 // Move Instructions
2416 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2417 "movshdup\t{$src, $dst|$dst, $src}",
2418 [(set VR128:$dst, (v4f32 (movshdup
2419 VR128:$src, (undef))))]>;
2420 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2421 "movshdup\t{$src, $dst|$dst, $src}",
2422 [(set VR128:$dst, (movshdup
2423 (memopv4f32 addr:$src), (undef)))]>;
2425 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2426 "movsldup\t{$src, $dst|$dst, $src}",
2427 [(set VR128:$dst, (v4f32 (movsldup
2428 VR128:$src, (undef))))]>;
2429 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2430 "movsldup\t{$src, $dst|$dst, $src}",
2431 [(set VR128:$dst, (movsldup
2432 (memopv4f32 addr:$src), (undef)))]>;
2434 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2435 "movddup\t{$src, $dst|$dst, $src}",
2436 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2437 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2438 "movddup\t{$src, $dst|$dst, $src}",
2440 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2443 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2445 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2447 let AddedComplexity = 5 in {
2448 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2449 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2450 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2451 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2452 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2453 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2454 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2455 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2459 let Constraints = "$src1 = $dst" in {
2460 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2461 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2462 "addsubps\t{$src2, $dst|$dst, $src2}",
2463 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2465 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2466 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2467 "addsubps\t{$src2, $dst|$dst, $src2}",
2468 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2469 (memop addr:$src2)))]>;
2470 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2471 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2472 "addsubpd\t{$src2, $dst|$dst, $src2}",
2473 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2475 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2476 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2477 "addsubpd\t{$src2, $dst|$dst, $src2}",
2478 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2479 (memop addr:$src2)))]>;
2482 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2483 "lddqu\t{$src, $dst|$dst, $src}",
2484 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2487 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2488 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2489 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2490 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2491 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2492 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2493 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2494 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2495 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2496 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2497 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2498 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2499 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2500 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2501 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2502 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2504 let Constraints = "$src1 = $dst" in {
2505 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2506 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2507 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2508 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2509 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2510 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2511 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2512 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2515 // Thread synchronization
2516 def MONITOR : I<0x01, MRM_C8, (outs), (ins), "monitor",
2517 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2518 def MWAIT : I<0x01, MRM_C9, (outs), (ins), "mwait",
2519 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2521 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2522 let AddedComplexity = 15 in
2523 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2524 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2525 let AddedComplexity = 20 in
2526 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2527 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2529 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2530 let AddedComplexity = 15 in
2531 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2532 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2533 let AddedComplexity = 20 in
2534 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2535 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2537 //===---------------------------------------------------------------------===//
2538 // SSSE3 Instructions
2539 //===---------------------------------------------------------------------===//
2541 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2542 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2543 Intrinsic IntId64, Intrinsic IntId128> {
2544 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2545 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2546 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2548 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2549 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2551 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2553 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2555 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2556 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2559 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2561 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2564 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2567 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2568 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2569 Intrinsic IntId64, Intrinsic IntId128> {
2570 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2572 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2573 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2575 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2577 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2580 (bitconvert (memopv4i16 addr:$src))))]>;
2582 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2584 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2585 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2588 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2590 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2593 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2596 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2597 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2598 Intrinsic IntId64, Intrinsic IntId128> {
2599 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2601 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2602 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2604 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2606 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2609 (bitconvert (memopv2i32 addr:$src))))]>;
2611 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2613 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2614 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2617 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2619 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2622 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2625 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2626 int_x86_ssse3_pabs_b,
2627 int_x86_ssse3_pabs_b_128>;
2628 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2629 int_x86_ssse3_pabs_w,
2630 int_x86_ssse3_pabs_w_128>;
2631 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2632 int_x86_ssse3_pabs_d,
2633 int_x86_ssse3_pabs_d_128>;
2635 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2636 let Constraints = "$src1 = $dst" in {
2637 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2638 Intrinsic IntId64, Intrinsic IntId128,
2639 bit Commutable = 0> {
2640 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2641 (ins VR64:$src1, VR64:$src2),
2642 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2643 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2644 let isCommutable = Commutable;
2646 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2647 (ins VR64:$src1, i64mem:$src2),
2648 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2650 (IntId64 VR64:$src1,
2651 (bitconvert (memopv8i8 addr:$src2))))]>;
2653 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2654 (ins VR128:$src1, VR128:$src2),
2655 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2656 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2658 let isCommutable = Commutable;
2660 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2661 (ins VR128:$src1, i128mem:$src2),
2662 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2664 (IntId128 VR128:$src1,
2665 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2669 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2670 let Constraints = "$src1 = $dst" in {
2671 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2672 Intrinsic IntId64, Intrinsic IntId128,
2673 bit Commutable = 0> {
2674 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2675 (ins VR64:$src1, VR64:$src2),
2676 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2677 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2678 let isCommutable = Commutable;
2680 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2681 (ins VR64:$src1, i64mem:$src2),
2682 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2684 (IntId64 VR64:$src1,
2685 (bitconvert (memopv4i16 addr:$src2))))]>;
2687 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2688 (ins VR128:$src1, VR128:$src2),
2689 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2690 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2692 let isCommutable = Commutable;
2694 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2695 (ins VR128:$src1, i128mem:$src2),
2696 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2698 (IntId128 VR128:$src1,
2699 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2703 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2704 let Constraints = "$src1 = $dst" in {
2705 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2706 Intrinsic IntId64, Intrinsic IntId128,
2707 bit Commutable = 0> {
2708 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2709 (ins VR64:$src1, VR64:$src2),
2710 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2711 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2712 let isCommutable = Commutable;
2714 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2715 (ins VR64:$src1, i64mem:$src2),
2716 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2718 (IntId64 VR64:$src1,
2719 (bitconvert (memopv2i32 addr:$src2))))]>;
2721 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2722 (ins VR128:$src1, VR128:$src2),
2723 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2724 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2726 let isCommutable = Commutable;
2728 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2729 (ins VR128:$src1, i128mem:$src2),
2730 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2732 (IntId128 VR128:$src1,
2733 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2737 let ImmT = NoImm in { // None of these have i8 immediate fields.
2738 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2739 int_x86_ssse3_phadd_w,
2740 int_x86_ssse3_phadd_w_128>;
2741 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2742 int_x86_ssse3_phadd_d,
2743 int_x86_ssse3_phadd_d_128>;
2744 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2745 int_x86_ssse3_phadd_sw,
2746 int_x86_ssse3_phadd_sw_128>;
2747 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2748 int_x86_ssse3_phsub_w,
2749 int_x86_ssse3_phsub_w_128>;
2750 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2751 int_x86_ssse3_phsub_d,
2752 int_x86_ssse3_phsub_d_128>;
2753 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2754 int_x86_ssse3_phsub_sw,
2755 int_x86_ssse3_phsub_sw_128>;
2756 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2757 int_x86_ssse3_pmadd_ub_sw,
2758 int_x86_ssse3_pmadd_ub_sw_128>;
2759 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2760 int_x86_ssse3_pmul_hr_sw,
2761 int_x86_ssse3_pmul_hr_sw_128, 1>;
2763 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2764 int_x86_ssse3_pshuf_b,
2765 int_x86_ssse3_pshuf_b_128>;
2766 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2767 int_x86_ssse3_psign_b,
2768 int_x86_ssse3_psign_b_128>;
2769 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2770 int_x86_ssse3_psign_w,
2771 int_x86_ssse3_psign_w_128>;
2772 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2773 int_x86_ssse3_psign_d,
2774 int_x86_ssse3_psign_d_128>;
2777 // palignr patterns.
2778 let Constraints = "$src1 = $dst" in {
2779 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2780 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2781 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2783 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2784 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2785 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2788 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2789 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2790 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2792 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2793 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2794 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2798 let AddedComplexity = 5 in {
2800 def : Pat<(v1i64 (palign:$src3 VR64:$src1, VR64:$src2)),
2801 (PALIGNR64rr VR64:$src2, VR64:$src1,
2802 (SHUFFLE_get_palign_imm VR64:$src3))>,
2803 Requires<[HasSSSE3]>;
2804 def : Pat<(v2i32 (palign:$src3 VR64:$src1, VR64:$src2)),
2805 (PALIGNR64rr VR64:$src2, VR64:$src1,
2806 (SHUFFLE_get_palign_imm VR64:$src3))>,
2807 Requires<[HasSSSE3]>;
2808 def : Pat<(v2f32 (palign:$src3 VR64:$src1, VR64:$src2)),
2809 (PALIGNR64rr VR64:$src2, VR64:$src1,
2810 (SHUFFLE_get_palign_imm VR64:$src3))>,
2811 Requires<[HasSSSE3]>;
2812 def : Pat<(v4i16 (palign:$src3 VR64:$src1, VR64:$src2)),
2813 (PALIGNR64rr VR64:$src2, VR64:$src1,
2814 (SHUFFLE_get_palign_imm VR64:$src3))>,
2815 Requires<[HasSSSE3]>;
2816 def : Pat<(v8i8 (palign:$src3 VR64:$src1, VR64:$src2)),
2817 (PALIGNR64rr VR64:$src2, VR64:$src1,
2818 (SHUFFLE_get_palign_imm VR64:$src3))>,
2819 Requires<[HasSSSE3]>;
2821 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2822 (PALIGNR128rr VR128:$src2, VR128:$src1,
2823 (SHUFFLE_get_palign_imm VR128:$src3))>,
2824 Requires<[HasSSSE3]>;
2825 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2826 (PALIGNR128rr VR128:$src2, VR128:$src1,
2827 (SHUFFLE_get_palign_imm VR128:$src3))>,
2828 Requires<[HasSSSE3]>;
2829 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2830 (PALIGNR128rr VR128:$src2, VR128:$src1,
2831 (SHUFFLE_get_palign_imm VR128:$src3))>,
2832 Requires<[HasSSSE3]>;
2833 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2834 (PALIGNR128rr VR128:$src2, VR128:$src1,
2835 (SHUFFLE_get_palign_imm VR128:$src3))>,
2836 Requires<[HasSSSE3]>;
2839 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2840 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2841 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2842 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2844 //===---------------------------------------------------------------------===//
2845 // Non-Instruction Patterns
2846 //===---------------------------------------------------------------------===//
2848 // extload f32 -> f64. This matches load+fextend because we have a hack in
2849 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2851 // Since these loads aren't folded into the fextend, we have to match it
2853 let Predicates = [HasSSE2] in
2854 def : Pat<(fextend (loadf32 addr:$src)),
2855 (CVTSS2SDrm addr:$src)>;
2858 let Predicates = [HasSSE2] in {
2859 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2860 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2861 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2862 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2863 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2864 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2865 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2866 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2867 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2868 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2869 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2870 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2871 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2872 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2873 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2874 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2875 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2876 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2877 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2878 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2879 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2880 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2881 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2882 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2883 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2884 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2885 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2886 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2887 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2888 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2891 // Move scalar to XMM zero-extended
2892 // movd to XMM register zero-extends
2893 let AddedComplexity = 15 in {
2894 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2895 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2896 (MOVSDrr (v2f64 (V_SET0PS)), FR64:$src)>;
2897 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2898 (MOVSSrr (v4f32 (V_SET0PS)), FR32:$src)>;
2899 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2900 (MOVSSrr (v4f32 (V_SET0PS)),
2901 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss)))>;
2902 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2903 (MOVSSrr (v4i32 (V_SET0PI)),
2904 (EXTRACT_SUBREG (v4i32 VR128:$src), sub_ss))>;
2907 // Splat v2f64 / v2i64
2908 let AddedComplexity = 10 in {
2909 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
2910 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2911 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
2912 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2913 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
2914 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2915 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
2916 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2919 // Special unary SHUFPSrri case.
2920 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
2921 (SHUFPSrri VR128:$src1, VR128:$src1,
2922 (SHUFFLE_get_shuf_imm VR128:$src3))>;
2923 let AddedComplexity = 5 in
2924 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
2925 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2926 Requires<[HasSSE2]>;
2927 // Special unary SHUFPDrri case.
2928 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
2929 (SHUFPDrri VR128:$src1, VR128:$src1,
2930 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2931 Requires<[HasSSE2]>;
2932 // Special unary SHUFPDrri case.
2933 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
2934 (SHUFPDrri VR128:$src1, VR128:$src1,
2935 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2936 Requires<[HasSSE2]>;
2937 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2938 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
2939 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2940 Requires<[HasSSE2]>;
2942 // Special binary v4i32 shuffle cases with SHUFPS.
2943 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
2944 (SHUFPSrri VR128:$src1, VR128:$src2,
2945 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2946 Requires<[HasSSE2]>;
2947 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
2948 (SHUFPSrmi VR128:$src1, addr:$src2,
2949 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2950 Requires<[HasSSE2]>;
2951 // Special binary v2i64 shuffle cases using SHUFPDrri.
2952 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
2953 (SHUFPDrri VR128:$src1, VR128:$src2,
2954 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2955 Requires<[HasSSE2]>;
2957 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2958 let AddedComplexity = 15 in {
2959 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
2960 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2961 Requires<[OptForSpeed, HasSSE2]>;
2962 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
2963 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2964 Requires<[OptForSpeed, HasSSE2]>;
2966 let AddedComplexity = 10 in {
2967 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
2968 (UNPCKLPSrr VR128:$src, VR128:$src)>;
2969 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
2970 (PUNPCKLBWrr VR128:$src, VR128:$src)>;
2971 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
2972 (PUNPCKLWDrr VR128:$src, VR128:$src)>;
2973 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
2974 (PUNPCKLDQrr VR128:$src, VR128:$src)>;
2977 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
2978 let AddedComplexity = 15 in {
2979 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
2980 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2981 Requires<[OptForSpeed, HasSSE2]>;
2982 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
2983 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2984 Requires<[OptForSpeed, HasSSE2]>;
2986 let AddedComplexity = 10 in {
2987 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
2988 (UNPCKHPSrr VR128:$src, VR128:$src)>;
2989 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
2990 (PUNPCKHBWrr VR128:$src, VR128:$src)>;
2991 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
2992 (PUNPCKHWDrr VR128:$src, VR128:$src)>;
2993 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
2994 (PUNPCKHDQrr VR128:$src, VR128:$src)>;
2997 let AddedComplexity = 20 in {
2998 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
2999 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3000 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3002 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3003 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3004 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3006 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3007 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3008 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3009 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3010 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3013 let AddedComplexity = 20 in {
3014 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3015 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3016 (MOVLPSrm VR128:$src1, addr:$src2)>;
3017 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3018 (MOVLPDrm VR128:$src1, addr:$src2)>;
3019 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3020 (MOVLPSrm VR128:$src1, addr:$src2)>;
3021 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3022 (MOVLPDrm VR128:$src1, addr:$src2)>;
3025 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3026 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3027 (MOVLPSmr addr:$src1, VR128:$src2)>;
3028 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3029 (MOVLPDmr addr:$src1, VR128:$src2)>;
3030 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3032 (MOVLPSmr addr:$src1, VR128:$src2)>;
3033 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3034 (MOVLPDmr addr:$src1, VR128:$src2)>;
3036 let AddedComplexity = 15 in {
3037 // Setting the lowest element in the vector.
3038 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3039 (MOVSSrr (v4i32 VR128:$src1),
3040 (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_ss))>;
3041 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3042 (MOVSDrr (v2i64 VR128:$src1),
3043 (EXTRACT_SUBREG (v2i64 VR128:$src2), sub_sd))>;
3045 // vector_shuffle v1, v2 <4, 5, 2, 3> using movsd
3046 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3047 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3048 Requires<[HasSSE2]>;
3049 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3050 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3051 Requires<[HasSSE2]>;
3054 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3055 // fall back to this for SSE1)
3056 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3057 (SHUFPSrri VR128:$src2, VR128:$src1,
3058 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3060 // Set lowest element and zero upper elements.
3061 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3062 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3064 // Some special case pandn patterns.
3065 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3067 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3068 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3070 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3071 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3073 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3075 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3076 (memop addr:$src2))),
3077 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3078 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3079 (memop addr:$src2))),
3080 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3081 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3082 (memop addr:$src2))),
3083 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3085 // vector -> vector casts
3086 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3087 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3088 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3089 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3090 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3091 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3092 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3093 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3095 // Use movaps / movups for SSE integer load / store (one byte shorter).
3096 def : Pat<(alignedloadv4i32 addr:$src),
3097 (MOVAPSrm addr:$src)>;
3098 def : Pat<(loadv4i32 addr:$src),
3099 (MOVUPSrm addr:$src)>;
3100 def : Pat<(alignedloadv2i64 addr:$src),
3101 (MOVAPSrm addr:$src)>;
3102 def : Pat<(loadv2i64 addr:$src),
3103 (MOVUPSrm addr:$src)>;
3105 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3106 (MOVAPSmr addr:$dst, VR128:$src)>;
3107 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3108 (MOVAPSmr addr:$dst, VR128:$src)>;
3109 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3110 (MOVAPSmr addr:$dst, VR128:$src)>;
3111 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3112 (MOVAPSmr addr:$dst, VR128:$src)>;
3113 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3114 (MOVUPSmr addr:$dst, VR128:$src)>;
3115 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3116 (MOVUPSmr addr:$dst, VR128:$src)>;
3117 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3118 (MOVUPSmr addr:$dst, VR128:$src)>;
3119 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3120 (MOVUPSmr addr:$dst, VR128:$src)>;
3122 //===----------------------------------------------------------------------===//
3123 // SSE4.1 Instructions
3124 //===----------------------------------------------------------------------===//
3126 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3129 Intrinsic V2F64Int> {
3130 // Intrinsic operation, reg.
3131 // Vector intrinsic operation, reg
3132 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3133 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3134 !strconcat(OpcodeStr,
3135 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3136 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3139 // Vector intrinsic operation, mem
3140 def PSm_Int : Ii8<opcps, MRMSrcMem,
3141 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3142 !strconcat(OpcodeStr,
3143 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3145 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3147 Requires<[HasSSE41]>;
3149 // Vector intrinsic operation, reg
3150 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3151 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3152 !strconcat(OpcodeStr,
3153 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3154 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3157 // Vector intrinsic operation, mem
3158 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3159 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3160 !strconcat(OpcodeStr,
3161 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3163 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3167 let Constraints = "$src1 = $dst" in {
3168 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3172 // Intrinsic operation, reg.
3173 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3175 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3176 !strconcat(OpcodeStr,
3177 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3179 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3182 // Intrinsic operation, mem.
3183 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3185 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3186 !strconcat(OpcodeStr,
3187 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3189 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3192 // Intrinsic operation, reg.
3193 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3195 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3196 !strconcat(OpcodeStr,
3197 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3199 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3202 // Intrinsic operation, mem.
3203 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3205 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3206 !strconcat(OpcodeStr,
3207 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3209 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3214 // FP round - roundss, roundps, roundsd, roundpd
3215 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3216 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3217 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3218 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3220 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3221 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3222 Intrinsic IntId128> {
3223 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3225 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3226 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3227 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3229 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3232 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3235 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3236 int_x86_sse41_phminposuw>;
3238 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3239 let Constraints = "$src1 = $dst" in {
3240 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3241 Intrinsic IntId128, bit Commutable = 0> {
3242 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3243 (ins VR128:$src1, VR128:$src2),
3244 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3245 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3247 let isCommutable = Commutable;
3249 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3250 (ins VR128:$src1, i128mem:$src2),
3251 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3253 (IntId128 VR128:$src1,
3254 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3258 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3259 int_x86_sse41_pcmpeqq, 1>;
3260 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3261 int_x86_sse41_packusdw, 0>;
3262 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3263 int_x86_sse41_pminsb, 1>;
3264 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3265 int_x86_sse41_pminsd, 1>;
3266 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3267 int_x86_sse41_pminud, 1>;
3268 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3269 int_x86_sse41_pminuw, 1>;
3270 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3271 int_x86_sse41_pmaxsb, 1>;
3272 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3273 int_x86_sse41_pmaxsd, 1>;
3274 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3275 int_x86_sse41_pmaxud, 1>;
3276 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3277 int_x86_sse41_pmaxuw, 1>;
3279 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3281 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3282 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3283 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3284 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3286 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3287 let Constraints = "$src1 = $dst" in {
3288 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3289 SDNode OpNode, Intrinsic IntId128,
3290 bit Commutable = 0> {
3291 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3292 (ins VR128:$src1, VR128:$src2),
3293 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3294 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3295 VR128:$src2))]>, OpSize {
3296 let isCommutable = Commutable;
3298 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3299 (ins VR128:$src1, VR128:$src2),
3300 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3301 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3303 let isCommutable = Commutable;
3305 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3306 (ins VR128:$src1, i128mem:$src2),
3307 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3309 (OpVT (OpNode VR128:$src1, (memop addr:$src2))))]>, OpSize;
3310 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3311 (ins VR128:$src1, i128mem:$src2),
3312 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3314 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3319 /// SS48I_binop_rm - Simple SSE41 binary operator.
3320 let Constraints = "$src1 = $dst" in {
3321 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
3322 ValueType OpVT, bit Commutable = 0> {
3323 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3324 (ins VR128:$src1, VR128:$src2),
3325 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3326 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]>,
3328 let isCommutable = Commutable;
3330 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3331 (ins VR128:$src1, i128mem:$src2),
3332 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3333 [(set VR128:$dst, (OpNode VR128:$src1,
3334 (bc_v4i32 (memopv2i64 addr:$src2))))]>,
3339 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, 1>;
3341 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3342 let Constraints = "$src1 = $dst" in {
3343 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3344 Intrinsic IntId128, bit Commutable = 0> {
3345 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3346 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3347 !strconcat(OpcodeStr,
3348 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3350 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3352 let isCommutable = Commutable;
3354 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3355 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3356 !strconcat(OpcodeStr,
3357 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3359 (IntId128 VR128:$src1,
3360 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3365 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3366 int_x86_sse41_blendps, 0>;
3367 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3368 int_x86_sse41_blendpd, 0>;
3369 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3370 int_x86_sse41_pblendw, 0>;
3371 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3372 int_x86_sse41_dpps, 1>;
3373 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3374 int_x86_sse41_dppd, 1>;
3375 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3376 int_x86_sse41_mpsadbw, 0>;
3379 /// SS41I_ternary_int - SSE 4.1 ternary operator
3380 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3381 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3382 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3383 (ins VR128:$src1, VR128:$src2),
3384 !strconcat(OpcodeStr,
3385 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3386 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3389 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3390 (ins VR128:$src1, i128mem:$src2),
3391 !strconcat(OpcodeStr,
3392 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3395 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3399 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3400 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3401 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3404 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3405 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3406 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3407 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3409 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3410 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3412 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3416 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3417 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3418 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3419 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3420 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3421 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3423 // Common patterns involving scalar load.
3424 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3425 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3426 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3427 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3429 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3430 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3431 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3432 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3434 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3435 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3436 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3437 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3439 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3440 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3441 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3442 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3444 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3445 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3446 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3447 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3449 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3450 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3451 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3452 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3455 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3456 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3457 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3458 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3460 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3461 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3463 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3467 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3468 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3469 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3470 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3472 // Common patterns involving scalar load
3473 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3474 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3475 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3476 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3478 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3479 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3480 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3481 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3484 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3485 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3486 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3487 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3489 // Expecting a i16 load any extended to i32 value.
3490 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3491 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3492 [(set VR128:$dst, (IntId (bitconvert
3493 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3497 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3498 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3500 // Common patterns involving scalar load
3501 def : Pat<(int_x86_sse41_pmovsxbq
3502 (bitconvert (v4i32 (X86vzmovl
3503 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3504 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3506 def : Pat<(int_x86_sse41_pmovzxbq
3507 (bitconvert (v4i32 (X86vzmovl
3508 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3509 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3512 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3513 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3514 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3515 (ins VR128:$src1, i32i8imm:$src2),
3516 !strconcat(OpcodeStr,
3517 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3518 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3520 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3521 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3522 !strconcat(OpcodeStr,
3523 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3526 // There's an AssertZext in the way of writing the store pattern
3527 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3530 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3533 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3534 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3535 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3536 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3537 !strconcat(OpcodeStr,
3538 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3541 // There's an AssertZext in the way of writing the store pattern
3542 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3545 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3548 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3549 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3550 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3551 (ins VR128:$src1, i32i8imm:$src2),
3552 !strconcat(OpcodeStr,
3553 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3555 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3556 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3557 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3558 !strconcat(OpcodeStr,
3559 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3560 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3561 addr:$dst)]>, OpSize;
3564 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3567 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3569 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3570 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3571 (ins VR128:$src1, i32i8imm:$src2),
3572 !strconcat(OpcodeStr,
3573 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3575 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3577 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3578 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3579 !strconcat(OpcodeStr,
3580 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3581 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3582 addr:$dst)]>, OpSize;
3585 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3587 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3588 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3591 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3592 Requires<[HasSSE41]>;
3594 let Constraints = "$src1 = $dst" in {
3595 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3596 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3597 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3598 !strconcat(OpcodeStr,
3599 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3601 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3602 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3603 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3604 !strconcat(OpcodeStr,
3605 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3607 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3608 imm:$src3))]>, OpSize;
3612 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3614 let Constraints = "$src1 = $dst" in {
3615 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3616 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3617 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3618 !strconcat(OpcodeStr,
3619 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3621 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3623 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3624 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3625 !strconcat(OpcodeStr,
3626 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3628 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3629 imm:$src3)))]>, OpSize;
3633 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3635 // insertps has a few different modes, there's the first two here below which
3636 // are optimized inserts that won't zero arbitrary elements in the destination
3637 // vector. The next one matches the intrinsic and could zero arbitrary elements
3638 // in the target vector.
3639 let Constraints = "$src1 = $dst" in {
3640 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3641 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3642 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3643 !strconcat(OpcodeStr,
3644 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3646 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3648 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3649 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3650 !strconcat(OpcodeStr,
3651 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3653 (X86insrtps VR128:$src1,
3654 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3655 imm:$src3))]>, OpSize;
3659 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3661 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3662 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3664 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3665 // the intel intrinsic that corresponds to this.
3666 let Defs = [EFLAGS] in {
3667 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3668 "ptest \t{$src2, $src1|$src1, $src2}",
3669 [(set EFLAGS, (X86ptest VR128:$src1, VR128:$src2))]>,
3671 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3672 "ptest \t{$src2, $src1|$src1, $src2}",
3673 [(set EFLAGS, (X86ptest VR128:$src1, (load addr:$src2)))]>,
3677 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3678 "movntdqa\t{$src, $dst|$dst, $src}",
3679 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
3683 //===----------------------------------------------------------------------===//
3684 // SSE4.2 Instructions
3685 //===----------------------------------------------------------------------===//
3687 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3688 let Constraints = "$src1 = $dst" in {
3689 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3690 Intrinsic IntId128, bit Commutable = 0> {
3691 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3692 (ins VR128:$src1, VR128:$src2),
3693 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3694 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3696 let isCommutable = Commutable;
3698 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3699 (ins VR128:$src1, i128mem:$src2),
3700 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3702 (IntId128 VR128:$src1,
3703 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3707 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3709 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3710 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3711 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3712 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3714 // crc intrinsic instruction
3715 // This set of instructions are only rm, the only difference is the size
3717 let Constraints = "$src1 = $dst" in {
3718 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3719 (ins GR32:$src1, i8mem:$src2),
3720 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3722 (int_x86_sse42_crc32_8 GR32:$src1,
3723 (load addr:$src2)))]>;
3724 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3725 (ins GR32:$src1, GR8:$src2),
3726 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3728 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>;
3729 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3730 (ins GR32:$src1, i16mem:$src2),
3731 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3733 (int_x86_sse42_crc32_16 GR32:$src1,
3734 (load addr:$src2)))]>,
3736 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3737 (ins GR32:$src1, GR16:$src2),
3738 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3740 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3742 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3743 (ins GR32:$src1, i32mem:$src2),
3744 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3746 (int_x86_sse42_crc32_32 GR32:$src1,
3747 (load addr:$src2)))]>;
3748 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3749 (ins GR32:$src1, GR32:$src2),
3750 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3752 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>;
3753 def CRC64m8 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3754 (ins GR64:$src1, i8mem:$src2),
3755 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3757 (int_x86_sse42_crc64_8 GR64:$src1,
3758 (load addr:$src2)))]>,
3760 def CRC64r8 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3761 (ins GR64:$src1, GR8:$src2),
3762 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3764 (int_x86_sse42_crc64_8 GR64:$src1, GR8:$src2))]>,
3766 def CRC64m64 : SS42FI<0xF1, MRMSrcMem, (outs GR64:$dst),
3767 (ins GR64:$src1, i64mem:$src2),
3768 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3770 (int_x86_sse42_crc64_64 GR64:$src1,
3771 (load addr:$src2)))]>,
3773 def CRC64r64 : SS42FI<0xF1, MRMSrcReg, (outs GR64:$dst),
3774 (ins GR64:$src1, GR64:$src2),
3775 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3777 (int_x86_sse42_crc64_64 GR64:$src1, GR64:$src2))]>,
3781 // String/text processing instructions.
3782 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3783 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3784 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3785 "#PCMPISTRM128rr PSEUDO!",
3786 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3787 imm:$src3))]>, OpSize;
3788 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3789 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3790 "#PCMPISTRM128rm PSEUDO!",
3791 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3792 imm:$src3))]>, OpSize;
3795 let Defs = [XMM0, EFLAGS] in {
3796 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3797 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3798 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3799 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3800 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3801 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3804 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3805 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3806 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3807 "#PCMPESTRM128rr PSEUDO!",
3809 (int_x86_sse42_pcmpestrm128
3810 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3812 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3813 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3814 "#PCMPESTRM128rm PSEUDO!",
3815 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3816 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3820 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3821 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3822 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3823 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3824 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3825 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3826 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3829 let Defs = [ECX, EFLAGS] in {
3830 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3831 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3832 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3833 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3834 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3835 (implicit EFLAGS)]>, OpSize;
3836 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3837 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3838 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3839 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3840 (implicit EFLAGS)]>, OpSize;
3844 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3845 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3846 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3847 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3848 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3849 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3851 let Defs = [ECX, EFLAGS] in {
3852 let Uses = [EAX, EDX] in {
3853 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3854 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3855 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3856 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3857 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3858 (implicit EFLAGS)]>, OpSize;
3859 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3860 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3861 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3863 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
3864 (implicit EFLAGS)]>, OpSize;
3869 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3870 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3871 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3872 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3873 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3874 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;
3876 //===----------------------------------------------------------------------===//
3877 // AES-NI Instructions
3878 //===----------------------------------------------------------------------===//
3880 let Constraints = "$src1 = $dst" in {
3881 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr,
3882 Intrinsic IntId128, bit Commutable = 0> {
3883 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
3884 (ins VR128:$src1, VR128:$src2),
3885 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3886 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3888 let isCommutable = Commutable;
3890 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
3891 (ins VR128:$src1, i128mem:$src2),
3892 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3894 (IntId128 VR128:$src1,
3895 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3899 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
3900 int_x86_aesni_aesenc>;
3901 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
3902 int_x86_aesni_aesenclast>;
3903 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
3904 int_x86_aesni_aesdec>;
3905 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
3906 int_x86_aesni_aesdeclast>;
3908 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, VR128:$src2)),
3909 (AESENCrr VR128:$src1, VR128:$src2)>;
3910 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, (memop addr:$src2))),
3911 (AESENCrm VR128:$src1, addr:$src2)>;
3912 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, VR128:$src2)),
3913 (AESENCLASTrr VR128:$src1, VR128:$src2)>;
3914 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, (memop addr:$src2))),
3915 (AESENCLASTrm VR128:$src1, addr:$src2)>;
3916 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, VR128:$src2)),
3917 (AESDECrr VR128:$src1, VR128:$src2)>;
3918 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, (memop addr:$src2))),
3919 (AESDECrm VR128:$src1, addr:$src2)>;
3920 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, VR128:$src2)),
3921 (AESDECLASTrr VR128:$src1, VR128:$src2)>;
3922 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, (memop addr:$src2))),
3923 (AESDECLASTrm VR128:$src1, addr:$src2)>;
3925 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
3927 "aesimc\t{$src1, $dst|$dst, $src1}",
3929 (int_x86_aesni_aesimc VR128:$src1))]>,
3932 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
3933 (ins i128mem:$src1),
3934 "aesimc\t{$src1, $dst|$dst, $src1}",
3936 (int_x86_aesni_aesimc (bitconvert (memopv2i64 addr:$src1))))]>,
3939 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
3940 (ins VR128:$src1, i8imm:$src2),
3941 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3943 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
3945 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
3946 (ins i128mem:$src1, i8imm:$src2),
3947 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3949 (int_x86_aesni_aeskeygenassist (bitconvert (memopv2i64 addr:$src1)),