1 //===- ARMInstrNEON.td - NEON support for ARM -----------------------------===//
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 ARM NEON instruction set.
12 //===----------------------------------------------------------------------===//
14 //===----------------------------------------------------------------------===//
15 // NEON-specific DAG Nodes.
16 //===----------------------------------------------------------------------===//
18 def SDTARMVCMP : SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisSameAs<1, 2>]>;
20 def NEONvceq : SDNode<"ARMISD::VCEQ", SDTARMVCMP>;
21 def NEONvcge : SDNode<"ARMISD::VCGE", SDTARMVCMP>;
22 def NEONvcgeu : SDNode<"ARMISD::VCGEU", SDTARMVCMP>;
23 def NEONvcgt : SDNode<"ARMISD::VCGT", SDTARMVCMP>;
24 def NEONvcgtu : SDNode<"ARMISD::VCGTU", SDTARMVCMP>;
25 def NEONvtst : SDNode<"ARMISD::VTST", SDTARMVCMP>;
27 // Types for vector shift by immediates. The "SHX" version is for long and
28 // narrow operations where the source and destination vectors have different
29 // types. The "SHINS" version is for shift and insert operations.
30 def SDTARMVSH : SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisSameAs<0, 1>,
32 def SDTARMVSHX : SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisInt<1>,
34 def SDTARMVSHINS : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<0, 1>,
35 SDTCisSameAs<0, 2>, SDTCisVT<3, i32>]>;
37 def NEONvshl : SDNode<"ARMISD::VSHL", SDTARMVSH>;
38 def NEONvshrs : SDNode<"ARMISD::VSHRs", SDTARMVSH>;
39 def NEONvshru : SDNode<"ARMISD::VSHRu", SDTARMVSH>;
40 def NEONvshlls : SDNode<"ARMISD::VSHLLs", SDTARMVSHX>;
41 def NEONvshllu : SDNode<"ARMISD::VSHLLu", SDTARMVSHX>;
42 def NEONvshlli : SDNode<"ARMISD::VSHLLi", SDTARMVSHX>;
43 def NEONvshrn : SDNode<"ARMISD::VSHRN", SDTARMVSHX>;
45 def NEONvrshrs : SDNode<"ARMISD::VRSHRs", SDTARMVSH>;
46 def NEONvrshru : SDNode<"ARMISD::VRSHRu", SDTARMVSH>;
47 def NEONvrshrn : SDNode<"ARMISD::VRSHRN", SDTARMVSHX>;
49 def NEONvqshls : SDNode<"ARMISD::VQSHLs", SDTARMVSH>;
50 def NEONvqshlu : SDNode<"ARMISD::VQSHLu", SDTARMVSH>;
51 def NEONvqshlsu : SDNode<"ARMISD::VQSHLsu", SDTARMVSH>;
52 def NEONvqshrns : SDNode<"ARMISD::VQSHRNs", SDTARMVSHX>;
53 def NEONvqshrnu : SDNode<"ARMISD::VQSHRNu", SDTARMVSHX>;
54 def NEONvqshrnsu : SDNode<"ARMISD::VQSHRNsu", SDTARMVSHX>;
56 def NEONvqrshrns : SDNode<"ARMISD::VQRSHRNs", SDTARMVSHX>;
57 def NEONvqrshrnu : SDNode<"ARMISD::VQRSHRNu", SDTARMVSHX>;
58 def NEONvqrshrnsu : SDNode<"ARMISD::VQRSHRNsu", SDTARMVSHX>;
60 def NEONvsli : SDNode<"ARMISD::VSLI", SDTARMVSHINS>;
61 def NEONvsri : SDNode<"ARMISD::VSRI", SDTARMVSHINS>;
63 def SDTARMVGETLN : SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisInt<1>,
65 def NEONvgetlaneu : SDNode<"ARMISD::VGETLANEu", SDTARMVGETLN>;
66 def NEONvgetlanes : SDNode<"ARMISD::VGETLANEs", SDTARMVGETLN>;
68 def NEONvdup : SDNode<"ARMISD::VDUP", SDTypeProfile<1, 1, [SDTCisVec<0>]>>;
70 // VDUPLANE can produce a quad-register result from a double-register source,
71 // so the result is not constrained to match the source.
72 def NEONvduplane : SDNode<"ARMISD::VDUPLANE",
73 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
76 def SDTARMVLD2 : SDTypeProfile<2, 1, [SDTCisSameAs<0, 1>, SDTCisPtrTy<2>]>;
77 def SDTARMVLD3 : SDTypeProfile<3, 1, [SDTCisSameAs<0, 1>,
78 SDTCisSameAs<0, 2>, SDTCisPtrTy<3>]>;
79 def SDTARMVLD4 : SDTypeProfile<4, 1, [SDTCisSameAs<0, 1>,
81 SDTCisSameAs<0, 3>, SDTCisPtrTy<4>]>;
82 def NEONvld2d : SDNode<"ARMISD::VLD2D", SDTARMVLD2,
83 [SDNPHasChain, SDNPMayLoad]>;
84 def NEONvld3d : SDNode<"ARMISD::VLD3D", SDTARMVLD3,
85 [SDNPHasChain, SDNPMayLoad]>;
86 def NEONvld4d : SDNode<"ARMISD::VLD4D", SDTARMVLD4,
87 [SDNPHasChain, SDNPMayLoad]>;
89 def SDTARMVST2 : SDTypeProfile<0, 3, [SDTCisPtrTy<0>, SDTCisSameAs<1, 2>]>;
90 def SDTARMVST3 : SDTypeProfile<0, 4, [SDTCisPtrTy<0>, SDTCisSameAs<1, 2>,
92 def SDTARMVST4 : SDTypeProfile<0, 5, [SDTCisPtrTy<0>, SDTCisSameAs<1, 2>,
96 def NEONvst2d : SDNode<"ARMISD::VST2D", SDTARMVST2,
97 [SDNPHasChain, SDNPMayStore]>;
98 def NEONvst3d : SDNode<"ARMISD::VST3D", SDTARMVST3,
99 [SDNPHasChain, SDNPMayStore]>;
100 def NEONvst4d : SDNode<"ARMISD::VST4D", SDTARMVST4,
101 [SDNPHasChain, SDNPMayStore]>;
103 def SDTARMVSHUF : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisSameAs<0, 1>]>;
104 def NEONvrev64 : SDNode<"ARMISD::VREV64", SDTARMVSHUF>;
105 def NEONvrev32 : SDNode<"ARMISD::VREV32", SDTARMVSHUF>;
106 def NEONvrev16 : SDNode<"ARMISD::VREV16", SDTARMVSHUF>;
108 //===----------------------------------------------------------------------===//
109 // NEON operand definitions
110 //===----------------------------------------------------------------------===//
112 // addrmode_neonldstm := reg
114 /* TODO: Take advantage of vldm.
115 def addrmode_neonldstm : Operand<i32>,
116 ComplexPattern<i32, 2, "SelectAddrModeNeonLdStM", []> {
117 let PrintMethod = "printAddrNeonLdStMOperand";
118 let MIOperandInfo = (ops GPR, i32imm);
122 //===----------------------------------------------------------------------===//
123 // NEON load / store instructions
124 //===----------------------------------------------------------------------===//
126 /* TODO: Take advantage of vldm.
128 def VLDMD : NI<(outs),
129 (ins addrmode_neonldstm:$addr, reglist:$dst1, variable_ops),
131 "vldm${addr:submode} ${addr:base}, $dst1",
133 let Inst{27-25} = 0b110;
135 let Inst{11-9} = 0b101;
138 def VLDMS : NI<(outs),
139 (ins addrmode_neonldstm:$addr, reglist:$dst1, variable_ops),
141 "vldm${addr:submode} ${addr:base}, $dst1",
143 let Inst{27-25} = 0b110;
145 let Inst{11-9} = 0b101;
150 // Use vldmia to load a Q register as a D register pair.
151 def VLDRQ : NI4<(outs QPR:$dst), (ins addrmode4:$addr),
153 "vldmia $addr, ${dst:dregpair}",
154 [(set QPR:$dst, (v2f64 (load addrmode4:$addr)))]> {
155 let Inst{27-25} = 0b110;
156 let Inst{24} = 0; // P bit
157 let Inst{23} = 1; // U bit
159 let Inst{11-9} = 0b101;
162 // Use vstmia to store a Q register as a D register pair.
163 def VSTRQ : NI4<(outs), (ins QPR:$src, addrmode4:$addr),
165 "vstmia $addr, ${src:dregpair}",
166 [(store (v2f64 QPR:$src), addrmode4:$addr)]> {
167 let Inst{27-25} = 0b110;
168 let Inst{24} = 0; // P bit
169 let Inst{23} = 1; // U bit
171 let Inst{11-9} = 0b101;
174 // VLD1 : Vector Load (multiple single elements)
175 class VLD1D<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
176 : NLdSt<(outs DPR:$dst), (ins addrmode6:$addr),
178 !strconcat(OpcodeStr, "\t\\{$dst\\}, $addr"),
179 [(set DPR:$dst, (Ty (IntOp addrmode6:$addr)))]>;
180 class VLD1Q<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
181 : NLdSt<(outs QPR:$dst), (ins addrmode6:$addr),
183 !strconcat(OpcodeStr, "\t${dst:dregpair}, $addr"),
184 [(set QPR:$dst, (Ty (IntOp addrmode6:$addr)))]>;
186 def VLD1d8 : VLD1D<"vld1.8", v8i8, int_arm_neon_vld1>;
187 def VLD1d16 : VLD1D<"vld1.16", v4i16, int_arm_neon_vld1>;
188 def VLD1d32 : VLD1D<"vld1.32", v2i32, int_arm_neon_vld1>;
189 def VLD1df : VLD1D<"vld1.32", v2f32, int_arm_neon_vld1>;
190 def VLD1d64 : VLD1D<"vld1.64", v1i64, int_arm_neon_vld1>;
192 def VLD1q8 : VLD1Q<"vld1.8", v16i8, int_arm_neon_vld1>;
193 def VLD1q16 : VLD1Q<"vld1.16", v8i16, int_arm_neon_vld1>;
194 def VLD1q32 : VLD1Q<"vld1.32", v4i32, int_arm_neon_vld1>;
195 def VLD1qf : VLD1Q<"vld1.32", v4f32, int_arm_neon_vld1>;
196 def VLD1q64 : VLD1Q<"vld1.64", v2i64, int_arm_neon_vld1>;
200 // VLD2 : Vector Load (multiple 2-element structures)
201 class VLD2D<string OpcodeStr>
202 : NLdSt<(outs DPR:$dst1, DPR:$dst2), (ins addrmode6:$addr),
204 !strconcat(OpcodeStr, "\t\\{$dst1,$dst2\\}, $addr"), []>;
206 def VLD2d8 : VLD2D<"vld2.8">;
207 def VLD2d16 : VLD2D<"vld2.16">;
208 def VLD2d32 : VLD2D<"vld2.32">;
210 // VLD3 : Vector Load (multiple 3-element structures)
211 class VLD3D<string OpcodeStr>
212 : NLdSt<(outs DPR:$dst1, DPR:$dst2, DPR:$dst3), (ins addrmode6:$addr),
214 !strconcat(OpcodeStr, "\t\\{$dst1,$dst2,$dst3\\}, $addr"), []>;
216 def VLD3d8 : VLD3D<"vld3.8">;
217 def VLD3d16 : VLD3D<"vld3.16">;
218 def VLD3d32 : VLD3D<"vld3.32">;
220 // VLD4 : Vector Load (multiple 4-element structures)
221 class VLD4D<string OpcodeStr>
222 : NLdSt<(outs DPR:$dst1, DPR:$dst2, DPR:$dst3, DPR:$dst4),
223 (ins addrmode6:$addr),
225 !strconcat(OpcodeStr, "\t\\{$dst1,$dst2,$dst3,$dst4\\}, $addr"), []>;
227 def VLD4d8 : VLD4D<"vld4.8">;
228 def VLD4d16 : VLD4D<"vld4.16">;
229 def VLD4d32 : VLD4D<"vld4.32">;
232 // VST1 : Vector Store (multiple single elements)
233 class VST1D<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
234 : NLdSt<(outs), (ins addrmode6:$addr, DPR:$src),
236 !strconcat(OpcodeStr, "\t\\{$src\\}, $addr"),
237 [(IntOp addrmode6:$addr, (Ty DPR:$src))]>;
238 class VST1Q<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
239 : NLdSt<(outs), (ins addrmode6:$addr, QPR:$src),
241 !strconcat(OpcodeStr, "\t${src:dregpair}, $addr"),
242 [(IntOp addrmode6:$addr, (Ty QPR:$src))]>;
244 def VST1d8 : VST1D<"vst1.8", v8i8, int_arm_neon_vst1>;
245 def VST1d16 : VST1D<"vst1.16", v4i16, int_arm_neon_vst1>;
246 def VST1d32 : VST1D<"vst1.32", v2i32, int_arm_neon_vst1>;
247 def VST1df : VST1D<"vst1.32", v2f32, int_arm_neon_vst1>;
248 def VST1d64 : VST1D<"vst1.64", v1i64, int_arm_neon_vst1>;
250 def VST1q8 : VST1Q<"vst1.8", v16i8, int_arm_neon_vst1>;
251 def VST1q16 : VST1Q<"vst1.16", v8i16, int_arm_neon_vst1>;
252 def VST1q32 : VST1Q<"vst1.32", v4i32, int_arm_neon_vst1>;
253 def VST1qf : VST1Q<"vst1.32", v4f32, int_arm_neon_vst1>;
254 def VST1q64 : VST1Q<"vst1.64", v2i64, int_arm_neon_vst1>;
256 let mayStore = 1 in {
258 // VST2 : Vector Store (multiple 2-element structures)
259 class VST2D<string OpcodeStr>
260 : NLdSt<(outs), (ins addrmode6:$addr, DPR:$src1, DPR:$src2), NoItinerary,
261 !strconcat(OpcodeStr, "\t\\{$src1,$src2\\}, $addr"), []>;
263 def VST2d8 : VST2D<"vst2.8">;
264 def VST2d16 : VST2D<"vst2.16">;
265 def VST2d32 : VST2D<"vst2.32">;
267 // VST3 : Vector Store (multiple 3-element structures)
268 class VST3D<string OpcodeStr>
269 : NLdSt<(outs), (ins addrmode6:$addr, DPR:$src1, DPR:$src2, DPR:$src3),
271 !strconcat(OpcodeStr, "\t\\{$src1,$src2,$src3\\}, $addr"), []>;
273 def VST3d8 : VST3D<"vst3.8">;
274 def VST3d16 : VST3D<"vst3.16">;
275 def VST3d32 : VST3D<"vst3.32">;
277 // VST4 : Vector Store (multiple 4-element structures)
278 class VST4D<string OpcodeStr>
279 : NLdSt<(outs), (ins addrmode6:$addr,
280 DPR:$src1, DPR:$src2, DPR:$src3, DPR:$src4), NoItinerary,
281 !strconcat(OpcodeStr, "\t\\{$src1,$src2,$src3,$src4\\}, $addr"), []>;
283 def VST4d8 : VST4D<"vst4.8">;
284 def VST4d16 : VST4D<"vst4.16">;
285 def VST4d32 : VST4D<"vst4.32">;
289 //===----------------------------------------------------------------------===//
290 // NEON pattern fragments
291 //===----------------------------------------------------------------------===//
293 // Extract D sub-registers of Q registers.
294 // (arm_dsubreg_0 is 5; arm_dsubreg_1 is 6)
295 def DSubReg_i8_reg : SDNodeXForm<imm, [{
296 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 8, MVT::i32);
298 def DSubReg_i16_reg : SDNodeXForm<imm, [{
299 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 4, MVT::i32);
301 def DSubReg_i32_reg : SDNodeXForm<imm, [{
302 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 2, MVT::i32);
304 def DSubReg_f64_reg : SDNodeXForm<imm, [{
305 return CurDAG->getTargetConstant(5 + N->getZExtValue(), MVT::i32);
308 // Extract S sub-registers of Q registers.
309 // (arm_ssubreg_0 is 1; arm_ssubreg_1 is 2; etc.)
310 def SSubReg_f32_reg : SDNodeXForm<imm, [{
311 return CurDAG->getTargetConstant(1 + N->getZExtValue(), MVT::i32);
314 // Translate lane numbers from Q registers to D subregs.
315 def SubReg_i8_lane : SDNodeXForm<imm, [{
316 return CurDAG->getTargetConstant(N->getZExtValue() & 7, MVT::i32);
318 def SubReg_i16_lane : SDNodeXForm<imm, [{
319 return CurDAG->getTargetConstant(N->getZExtValue() & 3, MVT::i32);
321 def SubReg_i32_lane : SDNodeXForm<imm, [{
322 return CurDAG->getTargetConstant(N->getZExtValue() & 1, MVT::i32);
325 //===----------------------------------------------------------------------===//
326 // Instruction Classes
327 //===----------------------------------------------------------------------===//
329 // Basic 2-register operations, both double- and quad-register.
330 class N2VD<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
331 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
332 ValueType ResTy, ValueType OpTy, SDNode OpNode>
333 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
334 (ins DPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
335 [(set DPR:$dst, (ResTy (OpNode (OpTy DPR:$src))))]>;
336 class N2VQ<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
337 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
338 ValueType ResTy, ValueType OpTy, SDNode OpNode>
339 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
340 (ins QPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
341 [(set QPR:$dst, (ResTy (OpNode (OpTy QPR:$src))))]>;
343 // Basic 2-register operations, scalar single-precision.
344 class N2VDs<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
345 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
346 ValueType ResTy, ValueType OpTy, SDNode OpNode>
347 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4,
348 (outs DPR_VFP2:$dst), (ins DPR_VFP2:$src),
349 NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "", []>;
351 class N2VDsPat<SDNode OpNode, ValueType ResTy, ValueType OpTy, NeonI Inst>
352 : NEONFPPat<(ResTy (OpNode SPR:$a)),
354 (Inst (INSERT_SUBREG (OpTy (IMPLICIT_DEF)), SPR:$a, arm_ssubreg_0)),
357 // Basic 2-register intrinsics, both double- and quad-register.
358 class N2VDInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
359 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
360 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
361 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
362 (ins DPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
363 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src))))]>;
364 class N2VQInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
365 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
366 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
367 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
368 (ins QPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
369 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src))))]>;
371 // Basic 2-register intrinsics, scalar single-precision
372 class N2VDInts<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
373 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
374 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
375 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4,
376 (outs DPR_VFP2:$dst), (ins DPR_VFP2:$src), NoItinerary,
377 !strconcat(OpcodeStr, "\t$dst, $src"), "", []>;
379 class N2VDIntsPat<SDNode OpNode, NeonI Inst>
380 : NEONFPPat<(f32 (OpNode SPR:$a)),
382 (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$a, arm_ssubreg_0)),
385 // Narrow 2-register intrinsics.
386 class N2VNInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
387 bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
388 string OpcodeStr, ValueType TyD, ValueType TyQ, Intrinsic IntOp>
389 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, op6, op4, (outs DPR:$dst),
390 (ins QPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
391 [(set DPR:$dst, (TyD (IntOp (TyQ QPR:$src))))]>;
393 // Long 2-register intrinsics. (This is currently only used for VMOVL and is
394 // derived from N2VImm instead of N2V because of the way the size is encoded.)
395 class N2VLInt<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
396 bit op6, bit op4, string OpcodeStr, ValueType TyQ, ValueType TyD,
398 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4, (outs QPR:$dst),
399 (ins DPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
400 [(set QPR:$dst, (TyQ (IntOp (TyD DPR:$src))))]>;
402 // 2-register shuffles (VTRN/VZIP/VUZP), both double- and quad-register.
403 class N2VDShuffle<bits<2> op19_18, bits<5> op11_7, string OpcodeStr>
404 : N2V<0b11, 0b11, op19_18, 0b10, op11_7, 0, 0, (outs DPR:$dst1, DPR:$dst2),
405 (ins DPR:$src1, DPR:$src2), NoItinerary,
406 !strconcat(OpcodeStr, "\t$dst1, $dst2"),
407 "$src1 = $dst1, $src2 = $dst2", []>;
408 class N2VQShuffle<bits<2> op19_18, bits<5> op11_7, string OpcodeStr>
409 : N2V<0b11, 0b11, op19_18, 0b10, op11_7, 1, 0, (outs QPR:$dst1, QPR:$dst2),
410 (ins QPR:$src1, QPR:$src2), NoItinerary,
411 !strconcat(OpcodeStr, "\t$dst1, $dst2"),
412 "$src1 = $dst1, $src2 = $dst2", []>;
414 // Basic 3-register operations, both double- and quad-register.
415 class N3VD<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
416 string OpcodeStr, ValueType ResTy, ValueType OpTy,
417 SDNode OpNode, bit Commutable>
418 : N3V<op24, op23, op21_20, op11_8, 0, op4,
419 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2), NoItinerary,
420 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
421 [(set DPR:$dst, (ResTy (OpNode (OpTy DPR:$src1), (OpTy DPR:$src2))))]> {
422 let isCommutable = Commutable;
424 class N3VQ<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
425 string OpcodeStr, ValueType ResTy, ValueType OpTy,
426 SDNode OpNode, bit Commutable>
427 : N3V<op24, op23, op21_20, op11_8, 1, op4,
428 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2), NoItinerary,
429 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
430 [(set QPR:$dst, (ResTy (OpNode (OpTy QPR:$src1), (OpTy QPR:$src2))))]> {
431 let isCommutable = Commutable;
434 // Basic 3-register operations, scalar single-precision
435 class N3VDs<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
436 string OpcodeStr, ValueType ResTy, ValueType OpTy,
437 SDNode OpNode, bit Commutable>
438 : N3V<op24, op23, op21_20, op11_8, 0, op4,
439 (outs DPR_VFP2:$dst), (ins DPR_VFP2:$src1, DPR_VFP2:$src2), NoItinerary,
440 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "", []> {
441 let isCommutable = Commutable;
443 class N3VDsPat<SDNode OpNode, NeonI Inst>
444 : NEONFPPat<(f32 (OpNode SPR:$a, SPR:$b)),
446 (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$a, arm_ssubreg_0),
447 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$b, arm_ssubreg_0)),
450 // Basic 3-register intrinsics, both double- and quad-register.
451 class N3VDInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
452 string OpcodeStr, ValueType ResTy, ValueType OpTy,
453 Intrinsic IntOp, bit Commutable>
454 : N3V<op24, op23, op21_20, op11_8, 0, op4,
455 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2), NoItinerary,
456 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
457 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src1), (OpTy DPR:$src2))))]> {
458 let isCommutable = Commutable;
460 class N3VQInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
461 string OpcodeStr, ValueType ResTy, ValueType OpTy,
462 Intrinsic IntOp, bit Commutable>
463 : N3V<op24, op23, op21_20, op11_8, 1, op4,
464 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2), NoItinerary,
465 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
466 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src1), (OpTy QPR:$src2))))]> {
467 let isCommutable = Commutable;
470 // Multiply-Add/Sub operations, both double- and quad-register.
471 class N3VDMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
472 string OpcodeStr, ValueType Ty, SDNode MulOp, SDNode OpNode>
473 : N3V<op24, op23, op21_20, op11_8, 0, op4,
474 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, DPR:$src3), NoItinerary,
475 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
476 [(set DPR:$dst, (Ty (OpNode DPR:$src1,
477 (Ty (MulOp DPR:$src2, DPR:$src3)))))]>;
478 class N3VQMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
479 string OpcodeStr, ValueType Ty, SDNode MulOp, SDNode OpNode>
480 : N3V<op24, op23, op21_20, op11_8, 1, op4,
481 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, QPR:$src3), NoItinerary,
482 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
483 [(set QPR:$dst, (Ty (OpNode QPR:$src1,
484 (Ty (MulOp QPR:$src2, QPR:$src3)))))]>;
486 // Multiply-Add/Sub operations, scalar single-precision
487 class N3VDMulOps<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
488 string OpcodeStr, ValueType Ty, SDNode MulOp, SDNode OpNode>
489 : N3V<op24, op23, op21_20, op11_8, 0, op4,
490 (outs DPR_VFP2:$dst),
491 (ins DPR_VFP2:$src1, DPR_VFP2:$src2, DPR_VFP2:$src3), NoItinerary,
492 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst", []>;
494 class N3VDMulOpsPat<SDNode MulNode, SDNode OpNode, NeonI Inst>
495 : NEONFPPat<(f32 (OpNode SPR:$acc, (f32 (MulNode SPR:$a, SPR:$b)))),
497 (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$acc, arm_ssubreg_0),
498 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$a, arm_ssubreg_0),
499 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$b, arm_ssubreg_0)),
502 // Neon 3-argument intrinsics, both double- and quad-register.
503 // The destination register is also used as the first source operand register.
504 class N3VDInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
505 string OpcodeStr, ValueType ResTy, ValueType OpTy,
507 : N3V<op24, op23, op21_20, op11_8, 0, op4,
508 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, DPR:$src3), NoItinerary,
509 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
510 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src1),
511 (OpTy DPR:$src2), (OpTy DPR:$src3))))]>;
512 class N3VQInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
513 string OpcodeStr, ValueType ResTy, ValueType OpTy,
515 : N3V<op24, op23, op21_20, op11_8, 1, op4,
516 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, QPR:$src3), NoItinerary,
517 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
518 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src1),
519 (OpTy QPR:$src2), (OpTy QPR:$src3))))]>;
521 // Neon Long 3-argument intrinsic. The destination register is
522 // a quad-register and is also used as the first source operand register.
523 class N3VLInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
524 string OpcodeStr, ValueType TyQ, ValueType TyD, Intrinsic IntOp>
525 : N3V<op24, op23, op21_20, op11_8, 0, op4,
526 (outs QPR:$dst), (ins QPR:$src1, DPR:$src2, DPR:$src3), NoItinerary,
527 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
529 (TyQ (IntOp (TyQ QPR:$src1), (TyD DPR:$src2), (TyD DPR:$src3))))]>;
531 // Narrowing 3-register intrinsics.
532 class N3VNInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
533 string OpcodeStr, ValueType TyD, ValueType TyQ,
534 Intrinsic IntOp, bit Commutable>
535 : N3V<op24, op23, op21_20, op11_8, 0, op4,
536 (outs DPR:$dst), (ins QPR:$src1, QPR:$src2), NoItinerary,
537 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
538 [(set DPR:$dst, (TyD (IntOp (TyQ QPR:$src1), (TyQ QPR:$src2))))]> {
539 let isCommutable = Commutable;
542 // Long 3-register intrinsics.
543 class N3VLInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
544 string OpcodeStr, ValueType TyQ, ValueType TyD,
545 Intrinsic IntOp, bit Commutable>
546 : N3V<op24, op23, op21_20, op11_8, 0, op4,
547 (outs QPR:$dst), (ins DPR:$src1, DPR:$src2), NoItinerary,
548 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
549 [(set QPR:$dst, (TyQ (IntOp (TyD DPR:$src1), (TyD DPR:$src2))))]> {
550 let isCommutable = Commutable;
553 // Wide 3-register intrinsics.
554 class N3VWInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
555 string OpcodeStr, ValueType TyQ, ValueType TyD,
556 Intrinsic IntOp, bit Commutable>
557 : N3V<op24, op23, op21_20, op11_8, 0, op4,
558 (outs QPR:$dst), (ins QPR:$src1, DPR:$src2), NoItinerary,
559 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
560 [(set QPR:$dst, (TyQ (IntOp (TyQ QPR:$src1), (TyD DPR:$src2))))]> {
561 let isCommutable = Commutable;
564 // Pairwise long 2-register intrinsics, both double- and quad-register.
565 class N2VDPLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
566 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
567 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
568 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
569 (ins DPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
570 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src))))]>;
571 class N2VQPLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
572 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
573 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
574 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
575 (ins QPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
576 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src))))]>;
578 // Pairwise long 2-register accumulate intrinsics,
579 // both double- and quad-register.
580 // The destination register is also used as the first source operand register.
581 class N2VDPLInt2<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
582 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
583 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
584 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4,
585 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2), NoItinerary,
586 !strconcat(OpcodeStr, "\t$dst, $src2"), "$src1 = $dst",
587 [(set DPR:$dst, (ResTy (IntOp (ResTy DPR:$src1), (OpTy DPR:$src2))))]>;
588 class N2VQPLInt2<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
589 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
590 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
591 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4,
592 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2), NoItinerary,
593 !strconcat(OpcodeStr, "\t$dst, $src2"), "$src1 = $dst",
594 [(set QPR:$dst, (ResTy (IntOp (ResTy QPR:$src1), (OpTy QPR:$src2))))]>;
596 // Shift by immediate,
597 // both double- and quad-register.
598 class N2VDSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
599 bit op4, string OpcodeStr, ValueType Ty, SDNode OpNode>
600 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
601 (outs DPR:$dst), (ins DPR:$src, i32imm:$SIMM), NoItinerary,
602 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
603 [(set DPR:$dst, (Ty (OpNode (Ty DPR:$src), (i32 imm:$SIMM))))]>;
604 class N2VQSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
605 bit op4, string OpcodeStr, ValueType Ty, SDNode OpNode>
606 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
607 (outs QPR:$dst), (ins QPR:$src, i32imm:$SIMM), NoItinerary,
608 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
609 [(set QPR:$dst, (Ty (OpNode (Ty QPR:$src), (i32 imm:$SIMM))))]>;
611 // Long shift by immediate.
612 class N2VLSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
613 bit op6, bit op4, string OpcodeStr, ValueType ResTy,
614 ValueType OpTy, SDNode OpNode>
615 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4,
616 (outs QPR:$dst), (ins DPR:$src, i32imm:$SIMM), NoItinerary,
617 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
618 [(set QPR:$dst, (ResTy (OpNode (OpTy DPR:$src),
619 (i32 imm:$SIMM))))]>;
621 // Narrow shift by immediate.
622 class N2VNSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
623 bit op6, bit op4, string OpcodeStr, ValueType ResTy,
624 ValueType OpTy, SDNode OpNode>
625 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4,
626 (outs DPR:$dst), (ins QPR:$src, i32imm:$SIMM), NoItinerary,
627 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
628 [(set DPR:$dst, (ResTy (OpNode (OpTy QPR:$src),
629 (i32 imm:$SIMM))))]>;
631 // Shift right by immediate and accumulate,
632 // both double- and quad-register.
633 class N2VDShAdd<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
634 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
635 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
636 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, i32imm:$SIMM),
638 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
639 [(set DPR:$dst, (Ty (add DPR:$src1,
640 (Ty (ShOp DPR:$src2, (i32 imm:$SIMM))))))]>;
641 class N2VQShAdd<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
642 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
643 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
644 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, i32imm:$SIMM),
646 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
647 [(set QPR:$dst, (Ty (add QPR:$src1,
648 (Ty (ShOp QPR:$src2, (i32 imm:$SIMM))))))]>;
650 // Shift by immediate and insert,
651 // both double- and quad-register.
652 class N2VDShIns<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
653 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
654 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
655 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, i32imm:$SIMM),
657 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
658 [(set DPR:$dst, (Ty (ShOp DPR:$src1, DPR:$src2, (i32 imm:$SIMM))))]>;
659 class N2VQShIns<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
660 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
661 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
662 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, i32imm:$SIMM),
664 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
665 [(set QPR:$dst, (Ty (ShOp QPR:$src1, QPR:$src2, (i32 imm:$SIMM))))]>;
667 // Convert, with fractional bits immediate,
668 // both double- and quad-register.
669 class N2VCvtD<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
670 bit op4, string OpcodeStr, ValueType ResTy, ValueType OpTy,
672 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
673 (outs DPR:$dst), (ins DPR:$src, i32imm:$SIMM), NoItinerary,
674 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
675 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src), (i32 imm:$SIMM))))]>;
676 class N2VCvtQ<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
677 bit op4, string OpcodeStr, ValueType ResTy, ValueType OpTy,
679 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
680 (outs QPR:$dst), (ins QPR:$src, i32imm:$SIMM), NoItinerary,
681 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
682 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src), (i32 imm:$SIMM))))]>;
684 //===----------------------------------------------------------------------===//
686 //===----------------------------------------------------------------------===//
688 // Neon 3-register vector operations.
690 // First with only element sizes of 8, 16 and 32 bits:
691 multiclass N3V_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
692 string OpcodeStr, SDNode OpNode, bit Commutable = 0> {
693 // 64-bit vector types.
694 def v8i8 : N3VD<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
695 v8i8, v8i8, OpNode, Commutable>;
696 def v4i16 : N3VD<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr, "16"),
697 v4i16, v4i16, OpNode, Commutable>;
698 def v2i32 : N3VD<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr, "32"),
699 v2i32, v2i32, OpNode, Commutable>;
701 // 128-bit vector types.
702 def v16i8 : N3VQ<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
703 v16i8, v16i8, OpNode, Commutable>;
704 def v8i16 : N3VQ<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr, "16"),
705 v8i16, v8i16, OpNode, Commutable>;
706 def v4i32 : N3VQ<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr, "32"),
707 v4i32, v4i32, OpNode, Commutable>;
710 // ....then also with element size 64 bits:
711 multiclass N3V_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
712 string OpcodeStr, SDNode OpNode, bit Commutable = 0>
713 : N3V_QHS<op24, op23, op11_8, op4, OpcodeStr, OpNode, Commutable> {
714 def v1i64 : N3VD<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr, "64"),
715 v1i64, v1i64, OpNode, Commutable>;
716 def v2i64 : N3VQ<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr, "64"),
717 v2i64, v2i64, OpNode, Commutable>;
721 // Neon Narrowing 2-register vector intrinsics,
722 // source operand element sizes of 16, 32 and 64 bits:
723 multiclass N2VNInt_HSD<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
724 bits<5> op11_7, bit op6, bit op4, string OpcodeStr,
726 def v8i8 : N2VNInt<op24_23, op21_20, 0b00, op17_16, op11_7, op6, op4,
727 !strconcat(OpcodeStr, "16"), v8i8, v8i16, IntOp>;
728 def v4i16 : N2VNInt<op24_23, op21_20, 0b01, op17_16, op11_7, op6, op4,
729 !strconcat(OpcodeStr, "32"), v4i16, v4i32, IntOp>;
730 def v2i32 : N2VNInt<op24_23, op21_20, 0b10, op17_16, op11_7, op6, op4,
731 !strconcat(OpcodeStr, "64"), v2i32, v2i64, IntOp>;
735 // Neon Lengthening 2-register vector intrinsic (currently specific to VMOVL).
736 // source operand element sizes of 16, 32 and 64 bits:
737 multiclass N2VLInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op7, bit op6,
738 bit op4, string OpcodeStr, Intrinsic IntOp> {
739 def v8i16 : N2VLInt<op24, op23, 0b001000, op11_8, op7, op6, op4,
740 !strconcat(OpcodeStr, "8"), v8i16, v8i8, IntOp>;
741 def v4i32 : N2VLInt<op24, op23, 0b010000, op11_8, op7, op6, op4,
742 !strconcat(OpcodeStr, "16"), v4i32, v4i16, IntOp>;
743 def v2i64 : N2VLInt<op24, op23, 0b100000, op11_8, op7, op6, op4,
744 !strconcat(OpcodeStr, "32"), v2i64, v2i32, IntOp>;
748 // Neon 3-register vector intrinsics.
750 // First with only element sizes of 16 and 32 bits:
751 multiclass N3VInt_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
752 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
753 // 64-bit vector types.
754 def v4i16 : N3VDInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
755 v4i16, v4i16, IntOp, Commutable>;
756 def v2i32 : N3VDInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
757 v2i32, v2i32, IntOp, Commutable>;
759 // 128-bit vector types.
760 def v8i16 : N3VQInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
761 v8i16, v8i16, IntOp, Commutable>;
762 def v4i32 : N3VQInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
763 v4i32, v4i32, IntOp, Commutable>;
766 // ....then also with element size of 8 bits:
767 multiclass N3VInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
768 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
769 : N3VInt_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
770 def v8i8 : N3VDInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
771 v8i8, v8i8, IntOp, Commutable>;
772 def v16i8 : N3VQInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
773 v16i8, v16i8, IntOp, Commutable>;
776 // ....then also with element size of 64 bits:
777 multiclass N3VInt_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
778 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
779 : N3VInt_QHS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
780 def v1i64 : N3VDInt<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr,"64"),
781 v1i64, v1i64, IntOp, Commutable>;
782 def v2i64 : N3VQInt<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr,"64"),
783 v2i64, v2i64, IntOp, Commutable>;
787 // Neon Narrowing 3-register vector intrinsics,
788 // source operand element sizes of 16, 32 and 64 bits:
789 multiclass N3VNInt_HSD<bit op24, bit op23, bits<4> op11_8, bit op4,
790 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
791 def v8i8 : N3VNInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr,"16"),
792 v8i8, v8i16, IntOp, Commutable>;
793 def v4i16 : N3VNInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"32"),
794 v4i16, v4i32, IntOp, Commutable>;
795 def v2i32 : N3VNInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"64"),
796 v2i32, v2i64, IntOp, Commutable>;
800 // Neon Long 3-register vector intrinsics.
802 // First with only element sizes of 16 and 32 bits:
803 multiclass N3VLInt_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
804 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
805 def v4i32 : N3VLInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
806 v4i32, v4i16, IntOp, Commutable>;
807 def v2i64 : N3VLInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
808 v2i64, v2i32, IntOp, Commutable>;
811 // ....then also with element size of 8 bits:
812 multiclass N3VLInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
813 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
814 : N3VLInt_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
815 def v8i16 : N3VLInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
816 v8i16, v8i8, IntOp, Commutable>;
820 // Neon Wide 3-register vector intrinsics,
821 // source operand element sizes of 8, 16 and 32 bits:
822 multiclass N3VWInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
823 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
824 def v8i16 : N3VWInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
825 v8i16, v8i8, IntOp, Commutable>;
826 def v4i32 : N3VWInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
827 v4i32, v4i16, IntOp, Commutable>;
828 def v2i64 : N3VWInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
829 v2i64, v2i32, IntOp, Commutable>;
833 // Neon Multiply-Op vector operations,
834 // element sizes of 8, 16 and 32 bits:
835 multiclass N3VMulOp_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
836 string OpcodeStr, SDNode OpNode> {
837 // 64-bit vector types.
838 def v8i8 : N3VDMulOp<op24, op23, 0b00, op11_8, op4,
839 !strconcat(OpcodeStr, "8"), v8i8, mul, OpNode>;
840 def v4i16 : N3VDMulOp<op24, op23, 0b01, op11_8, op4,
841 !strconcat(OpcodeStr, "16"), v4i16, mul, OpNode>;
842 def v2i32 : N3VDMulOp<op24, op23, 0b10, op11_8, op4,
843 !strconcat(OpcodeStr, "32"), v2i32, mul, OpNode>;
845 // 128-bit vector types.
846 def v16i8 : N3VQMulOp<op24, op23, 0b00, op11_8, op4,
847 !strconcat(OpcodeStr, "8"), v16i8, mul, OpNode>;
848 def v8i16 : N3VQMulOp<op24, op23, 0b01, op11_8, op4,
849 !strconcat(OpcodeStr, "16"), v8i16, mul, OpNode>;
850 def v4i32 : N3VQMulOp<op24, op23, 0b10, op11_8, op4,
851 !strconcat(OpcodeStr, "32"), v4i32, mul, OpNode>;
855 // Neon 3-argument intrinsics,
856 // element sizes of 8, 16 and 32 bits:
857 multiclass N3VInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
858 string OpcodeStr, Intrinsic IntOp> {
859 // 64-bit vector types.
860 def v8i8 : N3VDInt3<op24, op23, 0b00, op11_8, op4,
861 !strconcat(OpcodeStr, "8"), v8i8, v8i8, IntOp>;
862 def v4i16 : N3VDInt3<op24, op23, 0b01, op11_8, op4,
863 !strconcat(OpcodeStr, "16"), v4i16, v4i16, IntOp>;
864 def v2i32 : N3VDInt3<op24, op23, 0b10, op11_8, op4,
865 !strconcat(OpcodeStr, "32"), v2i32, v2i32, IntOp>;
867 // 128-bit vector types.
868 def v16i8 : N3VQInt3<op24, op23, 0b00, op11_8, op4,
869 !strconcat(OpcodeStr, "8"), v16i8, v16i8, IntOp>;
870 def v8i16 : N3VQInt3<op24, op23, 0b01, op11_8, op4,
871 !strconcat(OpcodeStr, "16"), v8i16, v8i16, IntOp>;
872 def v4i32 : N3VQInt3<op24, op23, 0b10, op11_8, op4,
873 !strconcat(OpcodeStr, "32"), v4i32, v4i32, IntOp>;
877 // Neon Long 3-argument intrinsics.
879 // First with only element sizes of 16 and 32 bits:
880 multiclass N3VLInt3_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
881 string OpcodeStr, Intrinsic IntOp> {
882 def v4i32 : N3VLInt3<op24, op23, 0b01, op11_8, op4,
883 !strconcat(OpcodeStr, "16"), v4i32, v4i16, IntOp>;
884 def v2i64 : N3VLInt3<op24, op23, 0b10, op11_8, op4,
885 !strconcat(OpcodeStr, "32"), v2i64, v2i32, IntOp>;
888 // ....then also with element size of 8 bits:
889 multiclass N3VLInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
890 string OpcodeStr, Intrinsic IntOp>
891 : N3VLInt3_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp> {
892 def v8i16 : N3VLInt3<op24, op23, 0b01, op11_8, op4,
893 !strconcat(OpcodeStr, "8"), v8i16, v8i8, IntOp>;
897 // Neon 2-register vector intrinsics,
898 // element sizes of 8, 16 and 32 bits:
899 multiclass N2VInt_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
900 bits<5> op11_7, bit op4, string OpcodeStr,
902 // 64-bit vector types.
903 def v8i8 : N2VDInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
904 !strconcat(OpcodeStr, "8"), v8i8, v8i8, IntOp>;
905 def v4i16 : N2VDInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
906 !strconcat(OpcodeStr, "16"), v4i16, v4i16, IntOp>;
907 def v2i32 : N2VDInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
908 !strconcat(OpcodeStr, "32"), v2i32, v2i32, IntOp>;
910 // 128-bit vector types.
911 def v16i8 : N2VQInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
912 !strconcat(OpcodeStr, "8"), v16i8, v16i8, IntOp>;
913 def v8i16 : N2VQInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
914 !strconcat(OpcodeStr, "16"), v8i16, v8i16, IntOp>;
915 def v4i32 : N2VQInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
916 !strconcat(OpcodeStr, "32"), v4i32, v4i32, IntOp>;
920 // Neon Pairwise long 2-register intrinsics,
921 // element sizes of 8, 16 and 32 bits:
922 multiclass N2VPLInt_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
923 bits<5> op11_7, bit op4,
924 string OpcodeStr, Intrinsic IntOp> {
925 // 64-bit vector types.
926 def v8i8 : N2VDPLInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
927 !strconcat(OpcodeStr, "8"), v4i16, v8i8, IntOp>;
928 def v4i16 : N2VDPLInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
929 !strconcat(OpcodeStr, "16"), v2i32, v4i16, IntOp>;
930 def v2i32 : N2VDPLInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
931 !strconcat(OpcodeStr, "32"), v1i64, v2i32, IntOp>;
933 // 128-bit vector types.
934 def v16i8 : N2VQPLInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
935 !strconcat(OpcodeStr, "8"), v8i16, v16i8, IntOp>;
936 def v8i16 : N2VQPLInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
937 !strconcat(OpcodeStr, "16"), v4i32, v8i16, IntOp>;
938 def v4i32 : N2VQPLInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
939 !strconcat(OpcodeStr, "32"), v2i64, v4i32, IntOp>;
943 // Neon Pairwise long 2-register accumulate intrinsics,
944 // element sizes of 8, 16 and 32 bits:
945 multiclass N2VPLInt2_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
946 bits<5> op11_7, bit op4,
947 string OpcodeStr, Intrinsic IntOp> {
948 // 64-bit vector types.
949 def v8i8 : N2VDPLInt2<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
950 !strconcat(OpcodeStr, "8"), v4i16, v8i8, IntOp>;
951 def v4i16 : N2VDPLInt2<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
952 !strconcat(OpcodeStr, "16"), v2i32, v4i16, IntOp>;
953 def v2i32 : N2VDPLInt2<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
954 !strconcat(OpcodeStr, "32"), v1i64, v2i32, IntOp>;
956 // 128-bit vector types.
957 def v16i8 : N2VQPLInt2<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
958 !strconcat(OpcodeStr, "8"), v8i16, v16i8, IntOp>;
959 def v8i16 : N2VQPLInt2<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
960 !strconcat(OpcodeStr, "16"), v4i32, v8i16, IntOp>;
961 def v4i32 : N2VQPLInt2<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
962 !strconcat(OpcodeStr, "32"), v2i64, v4i32, IntOp>;
966 // Neon 2-register vector shift by immediate,
967 // element sizes of 8, 16, 32 and 64 bits:
968 multiclass N2VSh_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
969 string OpcodeStr, SDNode OpNode> {
970 // 64-bit vector types.
971 def v8i8 : N2VDSh<op24, op23, 0b001000, op11_8, 0, op4,
972 !strconcat(OpcodeStr, "8"), v8i8, OpNode>;
973 def v4i16 : N2VDSh<op24, op23, 0b010000, op11_8, 0, op4,
974 !strconcat(OpcodeStr, "16"), v4i16, OpNode>;
975 def v2i32 : N2VDSh<op24, op23, 0b100000, op11_8, 0, op4,
976 !strconcat(OpcodeStr, "32"), v2i32, OpNode>;
977 def v1i64 : N2VDSh<op24, op23, 0b000000, op11_8, 1, op4,
978 !strconcat(OpcodeStr, "64"), v1i64, OpNode>;
980 // 128-bit vector types.
981 def v16i8 : N2VQSh<op24, op23, 0b001000, op11_8, 0, op4,
982 !strconcat(OpcodeStr, "8"), v16i8, OpNode>;
983 def v8i16 : N2VQSh<op24, op23, 0b010000, op11_8, 0, op4,
984 !strconcat(OpcodeStr, "16"), v8i16, OpNode>;
985 def v4i32 : N2VQSh<op24, op23, 0b100000, op11_8, 0, op4,
986 !strconcat(OpcodeStr, "32"), v4i32, OpNode>;
987 def v2i64 : N2VQSh<op24, op23, 0b000000, op11_8, 1, op4,
988 !strconcat(OpcodeStr, "64"), v2i64, OpNode>;
992 // Neon Shift-Accumulate vector operations,
993 // element sizes of 8, 16, 32 and 64 bits:
994 multiclass N2VShAdd_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
995 string OpcodeStr, SDNode ShOp> {
996 // 64-bit vector types.
997 def v8i8 : N2VDShAdd<op24, op23, 0b001000, op11_8, 0, op4,
998 !strconcat(OpcodeStr, "8"), v8i8, ShOp>;
999 def v4i16 : N2VDShAdd<op24, op23, 0b010000, op11_8, 0, op4,
1000 !strconcat(OpcodeStr, "16"), v4i16, ShOp>;
1001 def v2i32 : N2VDShAdd<op24, op23, 0b100000, op11_8, 0, op4,
1002 !strconcat(OpcodeStr, "32"), v2i32, ShOp>;
1003 def v1i64 : N2VDShAdd<op24, op23, 0b000000, op11_8, 1, op4,
1004 !strconcat(OpcodeStr, "64"), v1i64, ShOp>;
1006 // 128-bit vector types.
1007 def v16i8 : N2VQShAdd<op24, op23, 0b001000, op11_8, 0, op4,
1008 !strconcat(OpcodeStr, "8"), v16i8, ShOp>;
1009 def v8i16 : N2VQShAdd<op24, op23, 0b010000, op11_8, 0, op4,
1010 !strconcat(OpcodeStr, "16"), v8i16, ShOp>;
1011 def v4i32 : N2VQShAdd<op24, op23, 0b100000, op11_8, 0, op4,
1012 !strconcat(OpcodeStr, "32"), v4i32, ShOp>;
1013 def v2i64 : N2VQShAdd<op24, op23, 0b000000, op11_8, 1, op4,
1014 !strconcat(OpcodeStr, "64"), v2i64, ShOp>;
1018 // Neon Shift-Insert vector operations,
1019 // element sizes of 8, 16, 32 and 64 bits:
1020 multiclass N2VShIns_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
1021 string OpcodeStr, SDNode ShOp> {
1022 // 64-bit vector types.
1023 def v8i8 : N2VDShIns<op24, op23, 0b001000, op11_8, 0, op4,
1024 !strconcat(OpcodeStr, "8"), v8i8, ShOp>;
1025 def v4i16 : N2VDShIns<op24, op23, 0b010000, op11_8, 0, op4,
1026 !strconcat(OpcodeStr, "16"), v4i16, ShOp>;
1027 def v2i32 : N2VDShIns<op24, op23, 0b100000, op11_8, 0, op4,
1028 !strconcat(OpcodeStr, "32"), v2i32, ShOp>;
1029 def v1i64 : N2VDShIns<op24, op23, 0b000000, op11_8, 1, op4,
1030 !strconcat(OpcodeStr, "64"), v1i64, ShOp>;
1032 // 128-bit vector types.
1033 def v16i8 : N2VQShIns<op24, op23, 0b001000, op11_8, 0, op4,
1034 !strconcat(OpcodeStr, "8"), v16i8, ShOp>;
1035 def v8i16 : N2VQShIns<op24, op23, 0b010000, op11_8, 0, op4,
1036 !strconcat(OpcodeStr, "16"), v8i16, ShOp>;
1037 def v4i32 : N2VQShIns<op24, op23, 0b100000, op11_8, 0, op4,
1038 !strconcat(OpcodeStr, "32"), v4i32, ShOp>;
1039 def v2i64 : N2VQShIns<op24, op23, 0b000000, op11_8, 1, op4,
1040 !strconcat(OpcodeStr, "64"), v2i64, ShOp>;
1043 //===----------------------------------------------------------------------===//
1044 // Instruction Definitions.
1045 //===----------------------------------------------------------------------===//
1047 // Vector Add Operations.
1049 // VADD : Vector Add (integer and floating-point)
1050 defm VADD : N3V_QHSD<0, 0, 0b1000, 0, "vadd.i", add, 1>;
1051 def VADDfd : N3VD<0, 0, 0b00, 0b1101, 0, "vadd.f32", v2f32, v2f32, fadd, 1>;
1052 def VADDfq : N3VQ<0, 0, 0b00, 0b1101, 0, "vadd.f32", v4f32, v4f32, fadd, 1>;
1053 // VADDL : Vector Add Long (Q = D + D)
1054 defm VADDLs : N3VLInt_QHS<0,1,0b0000,0, "vaddl.s", int_arm_neon_vaddls, 1>;
1055 defm VADDLu : N3VLInt_QHS<1,1,0b0000,0, "vaddl.u", int_arm_neon_vaddlu, 1>;
1056 // VADDW : Vector Add Wide (Q = Q + D)
1057 defm VADDWs : N3VWInt_QHS<0,1,0b0001,0, "vaddw.s", int_arm_neon_vaddws, 0>;
1058 defm VADDWu : N3VWInt_QHS<1,1,0b0001,0, "vaddw.u", int_arm_neon_vaddwu, 0>;
1059 // VHADD : Vector Halving Add
1060 defm VHADDs : N3VInt_QHS<0,0,0b0000,0, "vhadd.s", int_arm_neon_vhadds, 1>;
1061 defm VHADDu : N3VInt_QHS<1,0,0b0000,0, "vhadd.u", int_arm_neon_vhaddu, 1>;
1062 // VRHADD : Vector Rounding Halving Add
1063 defm VRHADDs : N3VInt_QHS<0,0,0b0001,0, "vrhadd.s", int_arm_neon_vrhadds, 1>;
1064 defm VRHADDu : N3VInt_QHS<1,0,0b0001,0, "vrhadd.u", int_arm_neon_vrhaddu, 1>;
1065 // VQADD : Vector Saturating Add
1066 defm VQADDs : N3VInt_QHSD<0,0,0b0000,1, "vqadd.s", int_arm_neon_vqadds, 1>;
1067 defm VQADDu : N3VInt_QHSD<1,0,0b0000,1, "vqadd.u", int_arm_neon_vqaddu, 1>;
1068 // VADDHN : Vector Add and Narrow Returning High Half (D = Q + Q)
1069 defm VADDHN : N3VNInt_HSD<0,1,0b0100,0, "vaddhn.i", int_arm_neon_vaddhn, 1>;
1070 // VRADDHN : Vector Rounding Add and Narrow Returning High Half (D = Q + Q)
1071 defm VRADDHN : N3VNInt_HSD<1,1,0b0100,0, "vraddhn.i", int_arm_neon_vraddhn, 1>;
1073 // Vector Multiply Operations.
1075 // VMUL : Vector Multiply (integer, polynomial and floating-point)
1076 defm VMUL : N3V_QHS<0, 0, 0b1001, 1, "vmul.i", mul, 1>;
1077 def VMULpd : N3VDInt<1, 0, 0b00, 0b1001, 1, "vmul.p8", v8i8, v8i8,
1078 int_arm_neon_vmulp, 1>;
1079 def VMULpq : N3VQInt<1, 0, 0b00, 0b1001, 1, "vmul.p8", v16i8, v16i8,
1080 int_arm_neon_vmulp, 1>;
1081 def VMULfd : N3VD<1, 0, 0b00, 0b1101, 1, "vmul.f32", v2f32, v2f32, fmul, 1>;
1082 def VMULfq : N3VQ<1, 0, 0b00, 0b1101, 1, "vmul.f32", v4f32, v4f32, fmul, 1>;
1083 // VQDMULH : Vector Saturating Doubling Multiply Returning High Half
1084 defm VQDMULH : N3VInt_HS<0,0,0b1011,0, "vqdmulh.s", int_arm_neon_vqdmulh, 1>;
1085 // VQRDMULH : Vector Rounding Saturating Doubling Multiply Returning High Half
1086 defm VQRDMULH : N3VInt_HS<1,0,0b1011,0, "vqrdmulh.s", int_arm_neon_vqrdmulh, 1>;
1087 // VMULL : Vector Multiply Long (integer and polynomial) (Q = D * D)
1088 defm VMULLs : N3VLInt_QHS<0,1,0b1100,0, "vmull.s", int_arm_neon_vmulls, 1>;
1089 defm VMULLu : N3VLInt_QHS<1,1,0b1100,0, "vmull.u", int_arm_neon_vmullu, 1>;
1090 def VMULLp : N3VLInt<0, 1, 0b00, 0b1110, 0, "vmull.p8", v8i16, v8i8,
1091 int_arm_neon_vmullp, 1>;
1092 // VQDMULL : Vector Saturating Doubling Multiply Long (Q = D * D)
1093 defm VQDMULL : N3VLInt_HS<0,1,0b1101,0, "vqdmull.s", int_arm_neon_vqdmull, 1>;
1095 // Vector Multiply-Accumulate and Multiply-Subtract Operations.
1097 // VMLA : Vector Multiply Accumulate (integer and floating-point)
1098 defm VMLA : N3VMulOp_QHS<0, 0, 0b1001, 0, "vmla.i", add>;
1099 def VMLAfd : N3VDMulOp<0, 0, 0b00, 0b1101, 1, "vmla.f32", v2f32, fmul, fadd>;
1100 def VMLAfq : N3VQMulOp<0, 0, 0b00, 0b1101, 1, "vmla.f32", v4f32, fmul, fadd>;
1101 // VMLAL : Vector Multiply Accumulate Long (Q += D * D)
1102 defm VMLALs : N3VLInt3_QHS<0,1,0b1000,0, "vmlal.s", int_arm_neon_vmlals>;
1103 defm VMLALu : N3VLInt3_QHS<1,1,0b1000,0, "vmlal.u", int_arm_neon_vmlalu>;
1104 // VQDMLAL : Vector Saturating Doubling Multiply Accumulate Long (Q += D * D)
1105 defm VQDMLAL : N3VLInt3_HS<0, 1, 0b1001, 0, "vqdmlal.s", int_arm_neon_vqdmlal>;
1106 // VMLS : Vector Multiply Subtract (integer and floating-point)
1107 defm VMLS : N3VMulOp_QHS<0, 0, 0b1001, 0, "vmls.i", sub>;
1108 def VMLSfd : N3VDMulOp<0, 0, 0b10, 0b1101, 1, "vmls.f32", v2f32, fmul, fsub>;
1109 def VMLSfq : N3VQMulOp<0, 0, 0b10, 0b1101, 1, "vmls.f32", v4f32, fmul, fsub>;
1110 // VMLSL : Vector Multiply Subtract Long (Q -= D * D)
1111 defm VMLSLs : N3VLInt3_QHS<0,1,0b1010,0, "vmlsl.s", int_arm_neon_vmlsls>;
1112 defm VMLSLu : N3VLInt3_QHS<1,1,0b1010,0, "vmlsl.u", int_arm_neon_vmlslu>;
1113 // VQDMLSL : Vector Saturating Doubling Multiply Subtract Long (Q -= D * D)
1114 defm VQDMLSL : N3VLInt3_HS<0, 1, 0b1011, 0, "vqdmlsl.s", int_arm_neon_vqdmlsl>;
1116 // Vector Subtract Operations.
1118 // VSUB : Vector Subtract (integer and floating-point)
1119 defm VSUB : N3V_QHSD<1, 0, 0b1000, 0, "vsub.i", sub, 0>;
1120 def VSUBfd : N3VD<0, 0, 0b10, 0b1101, 0, "vsub.f32", v2f32, v2f32, fsub, 0>;
1121 def VSUBfq : N3VQ<0, 0, 0b10, 0b1101, 0, "vsub.f32", v4f32, v4f32, fsub, 0>;
1122 // VSUBL : Vector Subtract Long (Q = D - D)
1123 defm VSUBLs : N3VLInt_QHS<0,1,0b0010,0, "vsubl.s", int_arm_neon_vsubls, 1>;
1124 defm VSUBLu : N3VLInt_QHS<1,1,0b0010,0, "vsubl.u", int_arm_neon_vsublu, 1>;
1125 // VSUBW : Vector Subtract Wide (Q = Q - D)
1126 defm VSUBWs : N3VWInt_QHS<0,1,0b0011,0, "vsubw.s", int_arm_neon_vsubws, 0>;
1127 defm VSUBWu : N3VWInt_QHS<1,1,0b0011,0, "vsubw.u", int_arm_neon_vsubwu, 0>;
1128 // VHSUB : Vector Halving Subtract
1129 defm VHSUBs : N3VInt_QHS<0, 0, 0b0010, 0, "vhsub.s", int_arm_neon_vhsubs, 0>;
1130 defm VHSUBu : N3VInt_QHS<1, 0, 0b0010, 0, "vhsub.u", int_arm_neon_vhsubu, 0>;
1131 // VQSUB : Vector Saturing Subtract
1132 defm VQSUBs : N3VInt_QHSD<0, 0, 0b0010, 1, "vqsub.s", int_arm_neon_vqsubs, 0>;
1133 defm VQSUBu : N3VInt_QHSD<1, 0, 0b0010, 1, "vqsub.u", int_arm_neon_vqsubu, 0>;
1134 // VSUBHN : Vector Subtract and Narrow Returning High Half (D = Q - Q)
1135 defm VSUBHN : N3VNInt_HSD<0,1,0b0110,0, "vsubhn.i", int_arm_neon_vsubhn, 0>;
1136 // VRSUBHN : Vector Rounding Subtract and Narrow Returning High Half (D=Q-Q)
1137 defm VRSUBHN : N3VNInt_HSD<1,1,0b0110,0, "vrsubhn.i", int_arm_neon_vrsubhn, 0>;
1139 // Vector Comparisons.
1141 // VCEQ : Vector Compare Equal
1142 defm VCEQ : N3V_QHS<1, 0, 0b1000, 1, "vceq.i", NEONvceq, 1>;
1143 def VCEQfd : N3VD<0,0,0b00,0b1110,0, "vceq.f32", v2i32, v2f32, NEONvceq, 1>;
1144 def VCEQfq : N3VQ<0,0,0b00,0b1110,0, "vceq.f32", v4i32, v4f32, NEONvceq, 1>;
1145 // VCGE : Vector Compare Greater Than or Equal
1146 defm VCGEs : N3V_QHS<0, 0, 0b0011, 1, "vcge.s", NEONvcge, 0>;
1147 defm VCGEu : N3V_QHS<1, 0, 0b0011, 1, "vcge.u", NEONvcgeu, 0>;
1148 def VCGEfd : N3VD<1,0,0b00,0b1110,0, "vcge.f32", v2i32, v2f32, NEONvcge, 0>;
1149 def VCGEfq : N3VQ<1,0,0b00,0b1110,0, "vcge.f32", v4i32, v4f32, NEONvcge, 0>;
1150 // VCGT : Vector Compare Greater Than
1151 defm VCGTs : N3V_QHS<0, 0, 0b0011, 0, "vcgt.s", NEONvcgt, 0>;
1152 defm VCGTu : N3V_QHS<1, 0, 0b0011, 0, "vcgt.u", NEONvcgtu, 0>;
1153 def VCGTfd : N3VD<1,0,0b10,0b1110,0, "vcgt.f32", v2i32, v2f32, NEONvcgt, 0>;
1154 def VCGTfq : N3VQ<1,0,0b10,0b1110,0, "vcgt.f32", v4i32, v4f32, NEONvcgt, 0>;
1155 // VACGE : Vector Absolute Compare Greater Than or Equal (aka VCAGE)
1156 def VACGEd : N3VDInt<1, 0, 0b00, 0b1110, 1, "vacge.f32", v2i32, v2f32,
1157 int_arm_neon_vacged, 0>;
1158 def VACGEq : N3VQInt<1, 0, 0b00, 0b1110, 1, "vacge.f32", v4i32, v4f32,
1159 int_arm_neon_vacgeq, 0>;
1160 // VACGT : Vector Absolute Compare Greater Than (aka VCAGT)
1161 def VACGTd : N3VDInt<1, 0, 0b10, 0b1110, 1, "vacgt.f32", v2i32, v2f32,
1162 int_arm_neon_vacgtd, 0>;
1163 def VACGTq : N3VQInt<1, 0, 0b10, 0b1110, 1, "vacgt.f32", v4i32, v4f32,
1164 int_arm_neon_vacgtq, 0>;
1165 // VTST : Vector Test Bits
1166 defm VTST : N3V_QHS<0, 0, 0b1000, 1, "vtst.i", NEONvtst, 1>;
1168 // Vector Bitwise Operations.
1170 // VAND : Vector Bitwise AND
1171 def VANDd : N3VD<0, 0, 0b00, 0b0001, 1, "vand", v2i32, v2i32, and, 1>;
1172 def VANDq : N3VQ<0, 0, 0b00, 0b0001, 1, "vand", v4i32, v4i32, and, 1>;
1174 // VEOR : Vector Bitwise Exclusive OR
1175 def VEORd : N3VD<1, 0, 0b00, 0b0001, 1, "veor", v2i32, v2i32, xor, 1>;
1176 def VEORq : N3VQ<1, 0, 0b00, 0b0001, 1, "veor", v4i32, v4i32, xor, 1>;
1178 // VORR : Vector Bitwise OR
1179 def VORRd : N3VD<0, 0, 0b10, 0b0001, 1, "vorr", v2i32, v2i32, or, 1>;
1180 def VORRq : N3VQ<0, 0, 0b10, 0b0001, 1, "vorr", v4i32, v4i32, or, 1>;
1182 // VBIC : Vector Bitwise Bit Clear (AND NOT)
1183 def VBICd : N3V<0, 0, 0b01, 0b0001, 0, 1, (outs DPR:$dst),
1184 (ins DPR:$src1, DPR:$src2), NoItinerary,
1185 "vbic\t$dst, $src1, $src2", "",
1186 [(set DPR:$dst, (v2i32 (and DPR:$src1,(vnot DPR:$src2))))]>;
1187 def VBICq : N3V<0, 0, 0b01, 0b0001, 1, 1, (outs QPR:$dst),
1188 (ins QPR:$src1, QPR:$src2), NoItinerary,
1189 "vbic\t$dst, $src1, $src2", "",
1190 [(set QPR:$dst, (v4i32 (and QPR:$src1,(vnot QPR:$src2))))]>;
1192 // VORN : Vector Bitwise OR NOT
1193 def VORNd : N3V<0, 0, 0b11, 0b0001, 0, 1, (outs DPR:$dst),
1194 (ins DPR:$src1, DPR:$src2), NoItinerary,
1195 "vorn\t$dst, $src1, $src2", "",
1196 [(set DPR:$dst, (v2i32 (or DPR:$src1, (vnot DPR:$src2))))]>;
1197 def VORNq : N3V<0, 0, 0b11, 0b0001, 1, 1, (outs QPR:$dst),
1198 (ins QPR:$src1, QPR:$src2), NoItinerary,
1199 "vorn\t$dst, $src1, $src2", "",
1200 [(set QPR:$dst, (v4i32 (or QPR:$src1, (vnot QPR:$src2))))]>;
1202 // VMVN : Vector Bitwise NOT
1203 def VMVNd : N2V<0b11, 0b11, 0b00, 0b00, 0b01011, 0, 0,
1204 (outs DPR:$dst), (ins DPR:$src), NoItinerary,
1205 "vmvn\t$dst, $src", "",
1206 [(set DPR:$dst, (v2i32 (vnot DPR:$src)))]>;
1207 def VMVNq : N2V<0b11, 0b11, 0b00, 0b00, 0b01011, 1, 0,
1208 (outs QPR:$dst), (ins QPR:$src), NoItinerary,
1209 "vmvn\t$dst, $src", "",
1210 [(set QPR:$dst, (v4i32 (vnot QPR:$src)))]>;
1211 def : Pat<(v2i32 (vnot_conv DPR:$src)), (VMVNd DPR:$src)>;
1212 def : Pat<(v4i32 (vnot_conv QPR:$src)), (VMVNq QPR:$src)>;
1214 // VBSL : Vector Bitwise Select
1215 def VBSLd : N3V<1, 0, 0b01, 0b0001, 0, 1, (outs DPR:$dst),
1216 (ins DPR:$src1, DPR:$src2, DPR:$src3), NoItinerary,
1217 "vbsl\t$dst, $src2, $src3", "$src1 = $dst",
1219 (v2i32 (or (and DPR:$src2, DPR:$src1),
1220 (and DPR:$src3, (vnot DPR:$src1)))))]>;
1221 def VBSLq : N3V<1, 0, 0b01, 0b0001, 1, 1, (outs QPR:$dst),
1222 (ins QPR:$src1, QPR:$src2, QPR:$src3), NoItinerary,
1223 "vbsl\t$dst, $src2, $src3", "$src1 = $dst",
1225 (v4i32 (or (and QPR:$src2, QPR:$src1),
1226 (and QPR:$src3, (vnot QPR:$src1)))))]>;
1228 // VBIF : Vector Bitwise Insert if False
1229 // like VBSL but with: "vbif\t$dst, $src3, $src1", "$src2 = $dst",
1230 // VBIT : Vector Bitwise Insert if True
1231 // like VBSL but with: "vbit\t$dst, $src2, $src1", "$src3 = $dst",
1232 // These are not yet implemented. The TwoAddress pass will not go looking
1233 // for equivalent operations with different register constraints; it just
1236 // Vector Absolute Differences.
1238 // VABD : Vector Absolute Difference
1239 defm VABDs : N3VInt_QHS<0, 0, 0b0111, 0, "vabd.s", int_arm_neon_vabds, 0>;
1240 defm VABDu : N3VInt_QHS<1, 0, 0b0111, 0, "vabd.u", int_arm_neon_vabdu, 0>;
1241 def VABDfd : N3VDInt<1, 0, 0b10, 0b1101, 0, "vabd.f32", v2f32, v2f32,
1242 int_arm_neon_vabds, 0>;
1243 def VABDfq : N3VQInt<1, 0, 0b10, 0b1101, 0, "vabd.f32", v4f32, v4f32,
1244 int_arm_neon_vabds, 0>;
1246 // VABDL : Vector Absolute Difference Long (Q = | D - D |)
1247 defm VABDLs : N3VLInt_QHS<0,1,0b0111,0, "vabdl.s", int_arm_neon_vabdls, 0>;
1248 defm VABDLu : N3VLInt_QHS<1,1,0b0111,0, "vabdl.u", int_arm_neon_vabdlu, 0>;
1250 // VABA : Vector Absolute Difference and Accumulate
1251 defm VABAs : N3VInt3_QHS<0,1,0b0101,0, "vaba.s", int_arm_neon_vabas>;
1252 defm VABAu : N3VInt3_QHS<1,1,0b0101,0, "vaba.u", int_arm_neon_vabau>;
1254 // VABAL : Vector Absolute Difference and Accumulate Long (Q += | D - D |)
1255 defm VABALs : N3VLInt3_QHS<0,1,0b0101,0, "vabal.s", int_arm_neon_vabals>;
1256 defm VABALu : N3VLInt3_QHS<1,1,0b0101,0, "vabal.u", int_arm_neon_vabalu>;
1258 // Vector Maximum and Minimum.
1260 // VMAX : Vector Maximum
1261 defm VMAXs : N3VInt_QHS<0, 0, 0b0110, 0, "vmax.s", int_arm_neon_vmaxs, 1>;
1262 defm VMAXu : N3VInt_QHS<1, 0, 0b0110, 0, "vmax.u", int_arm_neon_vmaxu, 1>;
1263 def VMAXfd : N3VDInt<0, 0, 0b00, 0b1111, 0, "vmax.f32", v2f32, v2f32,
1264 int_arm_neon_vmaxs, 1>;
1265 def VMAXfq : N3VQInt<0, 0, 0b00, 0b1111, 0, "vmax.f32", v4f32, v4f32,
1266 int_arm_neon_vmaxs, 1>;
1268 // VMIN : Vector Minimum
1269 defm VMINs : N3VInt_QHS<0, 0, 0b0110, 1, "vmin.s", int_arm_neon_vmins, 1>;
1270 defm VMINu : N3VInt_QHS<1, 0, 0b0110, 1, "vmin.u", int_arm_neon_vminu, 1>;
1271 def VMINfd : N3VDInt<0, 0, 0b10, 0b1111, 0, "vmin.f32", v2f32, v2f32,
1272 int_arm_neon_vmins, 1>;
1273 def VMINfq : N3VQInt<0, 0, 0b10, 0b1111, 0, "vmin.f32", v4f32, v4f32,
1274 int_arm_neon_vmins, 1>;
1276 // Vector Pairwise Operations.
1278 // VPADD : Vector Pairwise Add
1279 def VPADDi8 : N3VDInt<0, 0, 0b00, 0b1011, 1, "vpadd.i8", v8i8, v8i8,
1280 int_arm_neon_vpadd, 0>;
1281 def VPADDi16 : N3VDInt<0, 0, 0b01, 0b1011, 1, "vpadd.i16", v4i16, v4i16,
1282 int_arm_neon_vpadd, 0>;
1283 def VPADDi32 : N3VDInt<0, 0, 0b10, 0b1011, 1, "vpadd.i32", v2i32, v2i32,
1284 int_arm_neon_vpadd, 0>;
1285 def VPADDf : N3VDInt<1, 0, 0b00, 0b1101, 0, "vpadd.f32", v2f32, v2f32,
1286 int_arm_neon_vpadd, 0>;
1288 // VPADDL : Vector Pairwise Add Long
1289 defm VPADDLs : N2VPLInt_QHS<0b11, 0b11, 0b00, 0b00100, 0, "vpaddl.s",
1290 int_arm_neon_vpaddls>;
1291 defm VPADDLu : N2VPLInt_QHS<0b11, 0b11, 0b00, 0b00101, 0, "vpaddl.u",
1292 int_arm_neon_vpaddlu>;
1294 // VPADAL : Vector Pairwise Add and Accumulate Long
1295 defm VPADALs : N2VPLInt2_QHS<0b11, 0b11, 0b00, 0b00100, 0, "vpadal.s",
1296 int_arm_neon_vpadals>;
1297 defm VPADALu : N2VPLInt2_QHS<0b11, 0b11, 0b00, 0b00101, 0, "vpadal.u",
1298 int_arm_neon_vpadalu>;
1300 // VPMAX : Vector Pairwise Maximum
1301 def VPMAXs8 : N3VDInt<0, 0, 0b00, 0b1010, 0, "vpmax.s8", v8i8, v8i8,
1302 int_arm_neon_vpmaxs, 0>;
1303 def VPMAXs16 : N3VDInt<0, 0, 0b01, 0b1010, 0, "vpmax.s16", v4i16, v4i16,
1304 int_arm_neon_vpmaxs, 0>;
1305 def VPMAXs32 : N3VDInt<0, 0, 0b10, 0b1010, 0, "vpmax.s32", v2i32, v2i32,
1306 int_arm_neon_vpmaxs, 0>;
1307 def VPMAXu8 : N3VDInt<1, 0, 0b00, 0b1010, 0, "vpmax.u8", v8i8, v8i8,
1308 int_arm_neon_vpmaxu, 0>;
1309 def VPMAXu16 : N3VDInt<1, 0, 0b01, 0b1010, 0, "vpmax.u16", v4i16, v4i16,
1310 int_arm_neon_vpmaxu, 0>;
1311 def VPMAXu32 : N3VDInt<1, 0, 0b10, 0b1010, 0, "vpmax.u32", v2i32, v2i32,
1312 int_arm_neon_vpmaxu, 0>;
1313 def VPMAXf : N3VDInt<1, 0, 0b00, 0b1111, 0, "vpmax.f32", v2f32, v2f32,
1314 int_arm_neon_vpmaxs, 0>;
1316 // VPMIN : Vector Pairwise Minimum
1317 def VPMINs8 : N3VDInt<0, 0, 0b00, 0b1010, 1, "vpmin.s8", v8i8, v8i8,
1318 int_arm_neon_vpmins, 0>;
1319 def VPMINs16 : N3VDInt<0, 0, 0b01, 0b1010, 1, "vpmin.s16", v4i16, v4i16,
1320 int_arm_neon_vpmins, 0>;
1321 def VPMINs32 : N3VDInt<0, 0, 0b10, 0b1010, 1, "vpmin.s32", v2i32, v2i32,
1322 int_arm_neon_vpmins, 0>;
1323 def VPMINu8 : N3VDInt<1, 0, 0b00, 0b1010, 1, "vpmin.u8", v8i8, v8i8,
1324 int_arm_neon_vpminu, 0>;
1325 def VPMINu16 : N3VDInt<1, 0, 0b01, 0b1010, 1, "vpmin.u16", v4i16, v4i16,
1326 int_arm_neon_vpminu, 0>;
1327 def VPMINu32 : N3VDInt<1, 0, 0b10, 0b1010, 1, "vpmin.u32", v2i32, v2i32,
1328 int_arm_neon_vpminu, 0>;
1329 def VPMINf : N3VDInt<1, 0, 0b10, 0b1111, 0, "vpmin.f32", v2f32, v2f32,
1330 int_arm_neon_vpmins, 0>;
1332 // Vector Reciprocal and Reciprocal Square Root Estimate and Step.
1334 // VRECPE : Vector Reciprocal Estimate
1335 def VRECPEd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01000, 0, "vrecpe.u32",
1336 v2i32, v2i32, int_arm_neon_vrecpe>;
1337 def VRECPEq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01000, 0, "vrecpe.u32",
1338 v4i32, v4i32, int_arm_neon_vrecpe>;
1339 def VRECPEfd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01010, 0, "vrecpe.f32",
1340 v2f32, v2f32, int_arm_neon_vrecpe>;
1341 def VRECPEfq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01010, 0, "vrecpe.f32",
1342 v4f32, v4f32, int_arm_neon_vrecpe>;
1344 // VRECPS : Vector Reciprocal Step
1345 def VRECPSfd : N3VDInt<0, 0, 0b00, 0b1111, 1, "vrecps.f32", v2f32, v2f32,
1346 int_arm_neon_vrecps, 1>;
1347 def VRECPSfq : N3VQInt<0, 0, 0b00, 0b1111, 1, "vrecps.f32", v4f32, v4f32,
1348 int_arm_neon_vrecps, 1>;
1350 // VRSQRTE : Vector Reciprocal Square Root Estimate
1351 def VRSQRTEd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01001, 0, "vrsqrte.u32",
1352 v2i32, v2i32, int_arm_neon_vrsqrte>;
1353 def VRSQRTEq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01001, 0, "vrsqrte.u32",
1354 v4i32, v4i32, int_arm_neon_vrsqrte>;
1355 def VRSQRTEfd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01011, 0, "vrsqrte.f32",
1356 v2f32, v2f32, int_arm_neon_vrsqrte>;
1357 def VRSQRTEfq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01011, 0, "vrsqrte.f32",
1358 v4f32, v4f32, int_arm_neon_vrsqrte>;
1360 // VRSQRTS : Vector Reciprocal Square Root Step
1361 def VRSQRTSfd : N3VDInt<0, 0, 0b10, 0b1111, 1, "vrsqrts.f32", v2f32, v2f32,
1362 int_arm_neon_vrsqrts, 1>;
1363 def VRSQRTSfq : N3VQInt<0, 0, 0b10, 0b1111, 1, "vrsqrts.f32", v4f32, v4f32,
1364 int_arm_neon_vrsqrts, 1>;
1368 // VSHL : Vector Shift
1369 defm VSHLs : N3VInt_QHSD<0, 0, 0b0100, 0, "vshl.s", int_arm_neon_vshifts, 0>;
1370 defm VSHLu : N3VInt_QHSD<1, 0, 0b0100, 0, "vshl.u", int_arm_neon_vshiftu, 0>;
1371 // VSHL : Vector Shift Left (Immediate)
1372 defm VSHLi : N2VSh_QHSD<0, 1, 0b0111, 1, "vshl.i", NEONvshl>;
1373 // VSHR : Vector Shift Right (Immediate)
1374 defm VSHRs : N2VSh_QHSD<0, 1, 0b0000, 1, "vshr.s", NEONvshrs>;
1375 defm VSHRu : N2VSh_QHSD<1, 1, 0b0000, 1, "vshr.u", NEONvshru>;
1377 // VSHLL : Vector Shift Left Long
1378 def VSHLLs8 : N2VLSh<0, 1, 0b001000, 0b1010, 0, 0, 1, "vshll.s8",
1379 v8i16, v8i8, NEONvshlls>;
1380 def VSHLLs16 : N2VLSh<0, 1, 0b010000, 0b1010, 0, 0, 1, "vshll.s16",
1381 v4i32, v4i16, NEONvshlls>;
1382 def VSHLLs32 : N2VLSh<0, 1, 0b100000, 0b1010, 0, 0, 1, "vshll.s32",
1383 v2i64, v2i32, NEONvshlls>;
1384 def VSHLLu8 : N2VLSh<1, 1, 0b001000, 0b1010, 0, 0, 1, "vshll.u8",
1385 v8i16, v8i8, NEONvshllu>;
1386 def VSHLLu16 : N2VLSh<1, 1, 0b010000, 0b1010, 0, 0, 1, "vshll.u16",
1387 v4i32, v4i16, NEONvshllu>;
1388 def VSHLLu32 : N2VLSh<1, 1, 0b100000, 0b1010, 0, 0, 1, "vshll.u32",
1389 v2i64, v2i32, NEONvshllu>;
1391 // VSHLL : Vector Shift Left Long (with maximum shift count)
1392 def VSHLLi8 : N2VLSh<1, 1, 0b110010, 0b0011, 0, 0, 0, "vshll.i8",
1393 v8i16, v8i8, NEONvshlli>;
1394 def VSHLLi16 : N2VLSh<1, 1, 0b110110, 0b0011, 0, 0, 0, "vshll.i16",
1395 v4i32, v4i16, NEONvshlli>;
1396 def VSHLLi32 : N2VLSh<1, 1, 0b111010, 0b0011, 0, 0, 0, "vshll.i32",
1397 v2i64, v2i32, NEONvshlli>;
1399 // VSHRN : Vector Shift Right and Narrow
1400 def VSHRN16 : N2VNSh<0, 1, 0b001000, 0b1000, 0, 0, 1, "vshrn.i16",
1401 v8i8, v8i16, NEONvshrn>;
1402 def VSHRN32 : N2VNSh<0, 1, 0b010000, 0b1000, 0, 0, 1, "vshrn.i32",
1403 v4i16, v4i32, NEONvshrn>;
1404 def VSHRN64 : N2VNSh<0, 1, 0b100000, 0b1000, 0, 0, 1, "vshrn.i64",
1405 v2i32, v2i64, NEONvshrn>;
1407 // VRSHL : Vector Rounding Shift
1408 defm VRSHLs : N3VInt_QHSD<0,0,0b0101,0, "vrshl.s", int_arm_neon_vrshifts, 0>;
1409 defm VRSHLu : N3VInt_QHSD<1,0,0b0101,0, "vrshl.u", int_arm_neon_vrshiftu, 0>;
1410 // VRSHR : Vector Rounding Shift Right
1411 defm VRSHRs : N2VSh_QHSD<0, 1, 0b0010, 1, "vrshr.s", NEONvrshrs>;
1412 defm VRSHRu : N2VSh_QHSD<1, 1, 0b0010, 1, "vrshr.u", NEONvrshru>;
1414 // VRSHRN : Vector Rounding Shift Right and Narrow
1415 def VRSHRN16 : N2VNSh<0, 1, 0b001000, 0b1000, 0, 1, 1, "vrshrn.i16",
1416 v8i8, v8i16, NEONvrshrn>;
1417 def VRSHRN32 : N2VNSh<0, 1, 0b010000, 0b1000, 0, 1, 1, "vrshrn.i32",
1418 v4i16, v4i32, NEONvrshrn>;
1419 def VRSHRN64 : N2VNSh<0, 1, 0b100000, 0b1000, 0, 1, 1, "vrshrn.i64",
1420 v2i32, v2i64, NEONvrshrn>;
1422 // VQSHL : Vector Saturating Shift
1423 defm VQSHLs : N3VInt_QHSD<0,0,0b0100,1, "vqshl.s", int_arm_neon_vqshifts, 0>;
1424 defm VQSHLu : N3VInt_QHSD<1,0,0b0100,1, "vqshl.u", int_arm_neon_vqshiftu, 0>;
1425 // VQSHL : Vector Saturating Shift Left (Immediate)
1426 defm VQSHLsi : N2VSh_QHSD<0, 1, 0b0111, 1, "vqshl.s", NEONvqshls>;
1427 defm VQSHLui : N2VSh_QHSD<1, 1, 0b0111, 1, "vqshl.u", NEONvqshlu>;
1428 // VQSHLU : Vector Saturating Shift Left (Immediate, Unsigned)
1429 defm VQSHLsu : N2VSh_QHSD<1, 1, 0b0110, 1, "vqshlu.s", NEONvqshlsu>;
1431 // VQSHRN : Vector Saturating Shift Right and Narrow
1432 def VQSHRNs16 : N2VNSh<0, 1, 0b001000, 0b1001, 0, 0, 1, "vqshrn.s16",
1433 v8i8, v8i16, NEONvqshrns>;
1434 def VQSHRNs32 : N2VNSh<0, 1, 0b010000, 0b1001, 0, 0, 1, "vqshrn.s32",
1435 v4i16, v4i32, NEONvqshrns>;
1436 def VQSHRNs64 : N2VNSh<0, 1, 0b100000, 0b1001, 0, 0, 1, "vqshrn.s64",
1437 v2i32, v2i64, NEONvqshrns>;
1438 def VQSHRNu16 : N2VNSh<1, 1, 0b001000, 0b1001, 0, 0, 1, "vqshrn.u16",
1439 v8i8, v8i16, NEONvqshrnu>;
1440 def VQSHRNu32 : N2VNSh<1, 1, 0b010000, 0b1001, 0, 0, 1, "vqshrn.u32",
1441 v4i16, v4i32, NEONvqshrnu>;
1442 def VQSHRNu64 : N2VNSh<1, 1, 0b100000, 0b1001, 0, 0, 1, "vqshrn.u64",
1443 v2i32, v2i64, NEONvqshrnu>;
1445 // VQSHRUN : Vector Saturating Shift Right and Narrow (Unsigned)
1446 def VQSHRUN16 : N2VNSh<1, 1, 0b001000, 0b1000, 0, 0, 1, "vqshrun.s16",
1447 v8i8, v8i16, NEONvqshrnsu>;
1448 def VQSHRUN32 : N2VNSh<1, 1, 0b010000, 0b1000, 0, 0, 1, "vqshrun.s32",
1449 v4i16, v4i32, NEONvqshrnsu>;
1450 def VQSHRUN64 : N2VNSh<1, 1, 0b100000, 0b1000, 0, 0, 1, "vqshrun.s64",
1451 v2i32, v2i64, NEONvqshrnsu>;
1453 // VQRSHL : Vector Saturating Rounding Shift
1454 defm VQRSHLs : N3VInt_QHSD<0, 0, 0b0101, 1, "vqrshl.s",
1455 int_arm_neon_vqrshifts, 0>;
1456 defm VQRSHLu : N3VInt_QHSD<1, 0, 0b0101, 1, "vqrshl.u",
1457 int_arm_neon_vqrshiftu, 0>;
1459 // VQRSHRN : Vector Saturating Rounding Shift Right and Narrow
1460 def VQRSHRNs16: N2VNSh<0, 1, 0b001000, 0b1001, 0, 1, 1, "vqrshrn.s16",
1461 v8i8, v8i16, NEONvqrshrns>;
1462 def VQRSHRNs32: N2VNSh<0, 1, 0b010000, 0b1001, 0, 1, 1, "vqrshrn.s32",
1463 v4i16, v4i32, NEONvqrshrns>;
1464 def VQRSHRNs64: N2VNSh<0, 1, 0b100000, 0b1001, 0, 1, 1, "vqrshrn.s64",
1465 v2i32, v2i64, NEONvqrshrns>;
1466 def VQRSHRNu16: N2VNSh<1, 1, 0b001000, 0b1001, 0, 1, 1, "vqrshrn.u16",
1467 v8i8, v8i16, NEONvqrshrnu>;
1468 def VQRSHRNu32: N2VNSh<1, 1, 0b010000, 0b1001, 0, 1, 1, "vqrshrn.u32",
1469 v4i16, v4i32, NEONvqrshrnu>;
1470 def VQRSHRNu64: N2VNSh<1, 1, 0b100000, 0b1001, 0, 1, 1, "vqrshrn.u64",
1471 v2i32, v2i64, NEONvqrshrnu>;
1473 // VQRSHRUN : Vector Saturating Rounding Shift Right and Narrow (Unsigned)
1474 def VQRSHRUN16: N2VNSh<1, 1, 0b001000, 0b1000, 0, 1, 1, "vqrshrun.s16",
1475 v8i8, v8i16, NEONvqrshrnsu>;
1476 def VQRSHRUN32: N2VNSh<1, 1, 0b010000, 0b1000, 0, 1, 1, "vqrshrun.s32",
1477 v4i16, v4i32, NEONvqrshrnsu>;
1478 def VQRSHRUN64: N2VNSh<1, 1, 0b100000, 0b1000, 0, 1, 1, "vqrshrun.s64",
1479 v2i32, v2i64, NEONvqrshrnsu>;
1481 // VSRA : Vector Shift Right and Accumulate
1482 defm VSRAs : N2VShAdd_QHSD<0, 1, 0b0001, 1, "vsra.s", NEONvshrs>;
1483 defm VSRAu : N2VShAdd_QHSD<1, 1, 0b0001, 1, "vsra.u", NEONvshru>;
1484 // VRSRA : Vector Rounding Shift Right and Accumulate
1485 defm VRSRAs : N2VShAdd_QHSD<0, 1, 0b0011, 1, "vrsra.s", NEONvrshrs>;
1486 defm VRSRAu : N2VShAdd_QHSD<1, 1, 0b0011, 1, "vrsra.u", NEONvrshru>;
1488 // VSLI : Vector Shift Left and Insert
1489 defm VSLI : N2VShIns_QHSD<1, 1, 0b0101, 1, "vsli.", NEONvsli>;
1490 // VSRI : Vector Shift Right and Insert
1491 defm VSRI : N2VShIns_QHSD<1, 1, 0b0100, 1, "vsri.", NEONvsri>;
1493 // Vector Absolute and Saturating Absolute.
1495 // VABS : Vector Absolute Value
1496 defm VABS : N2VInt_QHS<0b11, 0b11, 0b01, 0b00110, 0, "vabs.s",
1498 def VABSfd : N2VDInt<0b11, 0b11, 0b10, 0b01, 0b01110, 0, "vabs.f32",
1499 v2f32, v2f32, int_arm_neon_vabs>;
1500 def VABSfq : N2VQInt<0b11, 0b11, 0b10, 0b01, 0b01110, 0, "vabs.f32",
1501 v4f32, v4f32, int_arm_neon_vabs>;
1503 // VQABS : Vector Saturating Absolute Value
1504 defm VQABS : N2VInt_QHS<0b11, 0b11, 0b00, 0b01110, 0, "vqabs.s",
1505 int_arm_neon_vqabs>;
1509 def vneg : PatFrag<(ops node:$in), (sub immAllZerosV, node:$in)>;
1510 def vneg_conv : PatFrag<(ops node:$in), (sub immAllZerosV_bc, node:$in)>;
1512 class VNEGD<bits<2> size, string OpcodeStr, ValueType Ty>
1513 : N2V<0b11, 0b11, size, 0b01, 0b00111, 0, 0, (outs DPR:$dst), (ins DPR:$src),
1515 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1516 [(set DPR:$dst, (Ty (vneg DPR:$src)))]>;
1517 class VNEGQ<bits<2> size, string OpcodeStr, ValueType Ty>
1518 : N2V<0b11, 0b11, size, 0b01, 0b00111, 1, 0, (outs QPR:$dst), (ins QPR:$src),
1520 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1521 [(set QPR:$dst, (Ty (vneg QPR:$src)))]>;
1523 // VNEG : Vector Negate
1524 def VNEGs8d : VNEGD<0b00, "vneg.s8", v8i8>;
1525 def VNEGs16d : VNEGD<0b01, "vneg.s16", v4i16>;
1526 def VNEGs32d : VNEGD<0b10, "vneg.s32", v2i32>;
1527 def VNEGs8q : VNEGQ<0b00, "vneg.s8", v16i8>;
1528 def VNEGs16q : VNEGQ<0b01, "vneg.s16", v8i16>;
1529 def VNEGs32q : VNEGQ<0b10, "vneg.s32", v4i32>;
1531 // VNEG : Vector Negate (floating-point)
1532 def VNEGf32d : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 0, 0,
1533 (outs DPR:$dst), (ins DPR:$src), NoItinerary,
1534 "vneg.f32\t$dst, $src", "",
1535 [(set DPR:$dst, (v2f32 (fneg DPR:$src)))]>;
1536 def VNEGf32q : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 1, 0,
1537 (outs QPR:$dst), (ins QPR:$src), NoItinerary,
1538 "vneg.f32\t$dst, $src", "",
1539 [(set QPR:$dst, (v4f32 (fneg QPR:$src)))]>;
1541 def : Pat<(v8i8 (vneg_conv DPR:$src)), (VNEGs8d DPR:$src)>;
1542 def : Pat<(v4i16 (vneg_conv DPR:$src)), (VNEGs16d DPR:$src)>;
1543 def : Pat<(v2i32 (vneg_conv DPR:$src)), (VNEGs32d DPR:$src)>;
1544 def : Pat<(v16i8 (vneg_conv QPR:$src)), (VNEGs8q QPR:$src)>;
1545 def : Pat<(v8i16 (vneg_conv QPR:$src)), (VNEGs16q QPR:$src)>;
1546 def : Pat<(v4i32 (vneg_conv QPR:$src)), (VNEGs32q QPR:$src)>;
1548 // VQNEG : Vector Saturating Negate
1549 defm VQNEG : N2VInt_QHS<0b11, 0b11, 0b00, 0b01111, 0, "vqneg.s",
1550 int_arm_neon_vqneg>;
1552 // Vector Bit Counting Operations.
1554 // VCLS : Vector Count Leading Sign Bits
1555 defm VCLS : N2VInt_QHS<0b11, 0b11, 0b00, 0b01000, 0, "vcls.s",
1557 // VCLZ : Vector Count Leading Zeros
1558 defm VCLZ : N2VInt_QHS<0b11, 0b11, 0b00, 0b01001, 0, "vclz.i",
1560 // VCNT : Vector Count One Bits
1561 def VCNTd : N2VDInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0, "vcnt.8",
1562 v8i8, v8i8, int_arm_neon_vcnt>;
1563 def VCNTq : N2VQInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0, "vcnt.8",
1564 v16i8, v16i8, int_arm_neon_vcnt>;
1566 // Vector Move Operations.
1568 // VMOV : Vector Move (Register)
1570 def VMOVD : N3V<0, 0, 0b10, 0b0001, 0, 1, (outs DPR:$dst), (ins DPR:$src),
1571 NoItinerary, "vmov\t$dst, $src", "", []>;
1572 def VMOVQ : N3V<0, 0, 0b10, 0b0001, 1, 1, (outs QPR:$dst), (ins QPR:$src),
1573 NoItinerary, "vmov\t$dst, $src", "", []>;
1575 // VMOV : Vector Move (Immediate)
1577 // VMOV_get_imm8 xform function: convert build_vector to VMOV.i8 imm.
1578 def VMOV_get_imm8 : SDNodeXForm<build_vector, [{
1579 return ARM::getVMOVImm(N, 1, *CurDAG);
1581 def vmovImm8 : PatLeaf<(build_vector), [{
1582 return ARM::getVMOVImm(N, 1, *CurDAG).getNode() != 0;
1585 // VMOV_get_imm16 xform function: convert build_vector to VMOV.i16 imm.
1586 def VMOV_get_imm16 : SDNodeXForm<build_vector, [{
1587 return ARM::getVMOVImm(N, 2, *CurDAG);
1589 def vmovImm16 : PatLeaf<(build_vector), [{
1590 return ARM::getVMOVImm(N, 2, *CurDAG).getNode() != 0;
1591 }], VMOV_get_imm16>;
1593 // VMOV_get_imm32 xform function: convert build_vector to VMOV.i32 imm.
1594 def VMOV_get_imm32 : SDNodeXForm<build_vector, [{
1595 return ARM::getVMOVImm(N, 4, *CurDAG);
1597 def vmovImm32 : PatLeaf<(build_vector), [{
1598 return ARM::getVMOVImm(N, 4, *CurDAG).getNode() != 0;
1599 }], VMOV_get_imm32>;
1601 // VMOV_get_imm64 xform function: convert build_vector to VMOV.i64 imm.
1602 def VMOV_get_imm64 : SDNodeXForm<build_vector, [{
1603 return ARM::getVMOVImm(N, 8, *CurDAG);
1605 def vmovImm64 : PatLeaf<(build_vector), [{
1606 return ARM::getVMOVImm(N, 8, *CurDAG).getNode() != 0;
1607 }], VMOV_get_imm64>;
1609 // Note: Some of the cmode bits in the following VMOV instructions need to
1610 // be encoded based on the immed values.
1612 def VMOVv8i8 : N1ModImm<1, 0b000, 0b1110, 0, 0, 0, 1, (outs DPR:$dst),
1613 (ins i8imm:$SIMM), NoItinerary,
1614 "vmov.i8\t$dst, $SIMM", "",
1615 [(set DPR:$dst, (v8i8 vmovImm8:$SIMM))]>;
1616 def VMOVv16i8 : N1ModImm<1, 0b000, 0b1110, 0, 1, 0, 1, (outs QPR:$dst),
1617 (ins i8imm:$SIMM), NoItinerary,
1618 "vmov.i8\t$dst, $SIMM", "",
1619 [(set QPR:$dst, (v16i8 vmovImm8:$SIMM))]>;
1621 def VMOVv4i16 : N1ModImm<1, 0b000, 0b1000, 0, 0, 0, 1, (outs DPR:$dst),
1622 (ins i16imm:$SIMM), NoItinerary,
1623 "vmov.i16\t$dst, $SIMM", "",
1624 [(set DPR:$dst, (v4i16 vmovImm16:$SIMM))]>;
1625 def VMOVv8i16 : N1ModImm<1, 0b000, 0b1000, 0, 1, 0, 1, (outs QPR:$dst),
1626 (ins i16imm:$SIMM), NoItinerary,
1627 "vmov.i16\t$dst, $SIMM", "",
1628 [(set QPR:$dst, (v8i16 vmovImm16:$SIMM))]>;
1630 def VMOVv2i32 : N1ModImm<1, 0b000, 0b0000, 0, 0, 0, 1, (outs DPR:$dst),
1631 (ins i32imm:$SIMM), NoItinerary,
1632 "vmov.i32\t$dst, $SIMM", "",
1633 [(set DPR:$dst, (v2i32 vmovImm32:$SIMM))]>;
1634 def VMOVv4i32 : N1ModImm<1, 0b000, 0b0000, 0, 1, 0, 1, (outs QPR:$dst),
1635 (ins i32imm:$SIMM), NoItinerary,
1636 "vmov.i32\t$dst, $SIMM", "",
1637 [(set QPR:$dst, (v4i32 vmovImm32:$SIMM))]>;
1639 def VMOVv1i64 : N1ModImm<1, 0b000, 0b1110, 0, 0, 1, 1, (outs DPR:$dst),
1640 (ins i64imm:$SIMM), NoItinerary,
1641 "vmov.i64\t$dst, $SIMM", "",
1642 [(set DPR:$dst, (v1i64 vmovImm64:$SIMM))]>;
1643 def VMOVv2i64 : N1ModImm<1, 0b000, 0b1110, 0, 1, 1, 1, (outs QPR:$dst),
1644 (ins i64imm:$SIMM), NoItinerary,
1645 "vmov.i64\t$dst, $SIMM", "",
1646 [(set QPR:$dst, (v2i64 vmovImm64:$SIMM))]>;
1648 // VMOV : Vector Get Lane (move scalar to ARM core register)
1650 def VGETLNs8 : NVGetLane<0b11100101, 0b1011, 0b00,
1651 (outs GPR:$dst), (ins DPR:$src, lane_cst:$lane),
1652 NoItinerary, "vmov", ".s8\t$dst, $src[$lane]",
1653 [(set GPR:$dst, (NEONvgetlanes (v8i8 DPR:$src),
1655 def VGETLNs16 : NVGetLane<0b11100001, 0b1011, 0b01,
1656 (outs GPR:$dst), (ins DPR:$src, lane_cst:$lane),
1657 NoItinerary, "vmov", ".s16\t$dst, $src[$lane]",
1658 [(set GPR:$dst, (NEONvgetlanes (v4i16 DPR:$src),
1660 def VGETLNu8 : NVGetLane<0b11101101, 0b1011, 0b00,
1661 (outs GPR:$dst), (ins DPR:$src, lane_cst:$lane),
1662 NoItinerary, "vmov", ".u8\t$dst, $src[$lane]",
1663 [(set GPR:$dst, (NEONvgetlaneu (v8i8 DPR:$src),
1665 def VGETLNu16 : NVGetLane<0b11101001, 0b1011, 0b01,
1666 (outs GPR:$dst), (ins DPR:$src, lane_cst:$lane),
1667 NoItinerary, "vmov", ".u16\t$dst, $src[$lane]",
1668 [(set GPR:$dst, (NEONvgetlaneu (v4i16 DPR:$src),
1670 def VGETLNi32 : NVGetLane<0b11100001, 0b1011, 0b00,
1671 (outs GPR:$dst), (ins DPR:$src, lane_cst:$lane),
1672 NoItinerary, "vmov", ".32\t$dst, $src[$lane]",
1673 [(set GPR:$dst, (extractelt (v2i32 DPR:$src),
1675 // def VGETLNf32: see FMRDH and FMRDL in ARMInstrVFP.td
1676 def : Pat<(NEONvgetlanes (v16i8 QPR:$src), imm:$lane),
1677 (VGETLNs8 (v8i8 (EXTRACT_SUBREG QPR:$src,
1678 (DSubReg_i8_reg imm:$lane))),
1679 (SubReg_i8_lane imm:$lane))>;
1680 def : Pat<(NEONvgetlanes (v8i16 QPR:$src), imm:$lane),
1681 (VGETLNs16 (v4i16 (EXTRACT_SUBREG QPR:$src,
1682 (DSubReg_i16_reg imm:$lane))),
1683 (SubReg_i16_lane imm:$lane))>;
1684 def : Pat<(NEONvgetlaneu (v16i8 QPR:$src), imm:$lane),
1685 (VGETLNu8 (v8i8 (EXTRACT_SUBREG QPR:$src,
1686 (DSubReg_i8_reg imm:$lane))),
1687 (SubReg_i8_lane imm:$lane))>;
1688 def : Pat<(NEONvgetlaneu (v8i16 QPR:$src), imm:$lane),
1689 (VGETLNu16 (v4i16 (EXTRACT_SUBREG QPR:$src,
1690 (DSubReg_i16_reg imm:$lane))),
1691 (SubReg_i16_lane imm:$lane))>;
1692 def : Pat<(extractelt (v4i32 QPR:$src), imm:$lane),
1693 (VGETLNi32 (v2i32 (EXTRACT_SUBREG QPR:$src,
1694 (DSubReg_i32_reg imm:$lane))),
1695 (SubReg_i32_lane imm:$lane))>;
1696 def : Pat<(extractelt (v4f32 QPR:$src1), imm:$src2),
1697 (EXTRACT_SUBREG QPR:$src1, (SSubReg_f32_reg imm:$src2))>;
1698 //def : Pat<(extractelt (v2i64 QPR:$src1), imm:$src2),
1699 // (EXTRACT_SUBREG QPR:$src1, (DSubReg_f64_reg imm:$src2))>;
1700 def : Pat<(extractelt (v2f64 QPR:$src1), imm:$src2),
1701 (EXTRACT_SUBREG QPR:$src1, (DSubReg_f64_reg imm:$src2))>;
1704 // VMOV : Vector Set Lane (move ARM core register to scalar)
1706 let Constraints = "$src1 = $dst" in {
1707 def VSETLNi8 : NVSetLane<0b11100100, 0b1011, 0b00, (outs DPR:$dst),
1708 (ins DPR:$src1, GPR:$src2, lane_cst:$lane),
1709 NoItinerary, "vmov", ".8\t$dst[$lane], $src2",
1710 [(set DPR:$dst, (vector_insert (v8i8 DPR:$src1),
1711 GPR:$src2, imm:$lane))]>;
1712 def VSETLNi16 : NVSetLane<0b11100000, 0b1011, 0b01, (outs DPR:$dst),
1713 (ins DPR:$src1, GPR:$src2, lane_cst:$lane),
1714 NoItinerary, "vmov", ".16\t$dst[$lane], $src2",
1715 [(set DPR:$dst, (vector_insert (v4i16 DPR:$src1),
1716 GPR:$src2, imm:$lane))]>;
1717 def VSETLNi32 : NVSetLane<0b11100000, 0b1011, 0b00, (outs DPR:$dst),
1718 (ins DPR:$src1, GPR:$src2, lane_cst:$lane),
1719 NoItinerary, "vmov", ".32\t$dst[$lane], $src2",
1720 [(set DPR:$dst, (insertelt (v2i32 DPR:$src1),
1721 GPR:$src2, imm:$lane))]>;
1723 def : Pat<(vector_insert (v16i8 QPR:$src1), GPR:$src2, imm:$lane),
1724 (v16i8 (INSERT_SUBREG QPR:$src1,
1725 (VSETLNi8 (v8i8 (EXTRACT_SUBREG QPR:$src1,
1726 (DSubReg_i8_reg imm:$lane))),
1727 GPR:$src2, (SubReg_i8_lane imm:$lane)),
1728 (DSubReg_i8_reg imm:$lane)))>;
1729 def : Pat<(vector_insert (v8i16 QPR:$src1), GPR:$src2, imm:$lane),
1730 (v8i16 (INSERT_SUBREG QPR:$src1,
1731 (VSETLNi16 (v4i16 (EXTRACT_SUBREG QPR:$src1,
1732 (DSubReg_i16_reg imm:$lane))),
1733 GPR:$src2, (SubReg_i16_lane imm:$lane)),
1734 (DSubReg_i16_reg imm:$lane)))>;
1735 def : Pat<(insertelt (v4i32 QPR:$src1), GPR:$src2, imm:$lane),
1736 (v4i32 (INSERT_SUBREG QPR:$src1,
1737 (VSETLNi32 (v2i32 (EXTRACT_SUBREG QPR:$src1,
1738 (DSubReg_i32_reg imm:$lane))),
1739 GPR:$src2, (SubReg_i32_lane imm:$lane)),
1740 (DSubReg_i32_reg imm:$lane)))>;
1742 def : Pat<(v4f32 (insertelt QPR:$src1, SPR:$src2, imm:$src3)),
1743 (INSERT_SUBREG QPR:$src1, SPR:$src2, (SSubReg_f32_reg imm:$src3))>;
1745 //def : Pat<(v2i64 (insertelt QPR:$src1, DPR:$src2, imm:$src3)),
1746 // (INSERT_SUBREG QPR:$src1, DPR:$src2, (DSubReg_f64_reg imm:$src3))>;
1747 def : Pat<(v2f64 (insertelt QPR:$src1, DPR:$src2, imm:$src3)),
1748 (INSERT_SUBREG QPR:$src1, DPR:$src2, (DSubReg_f64_reg imm:$src3))>;
1750 // VDUP : Vector Duplicate (from ARM core register to all elements)
1752 class VDUPD<bits<8> opcod1, bits<2> opcod3, string asmSize, ValueType Ty>
1753 : NVDup<opcod1, 0b1011, opcod3, (outs DPR:$dst), (ins GPR:$src),
1754 NoItinerary, "vdup", !strconcat(asmSize, "\t$dst, $src"),
1755 [(set DPR:$dst, (Ty (NEONvdup (i32 GPR:$src))))]>;
1756 class VDUPQ<bits<8> opcod1, bits<2> opcod3, string asmSize, ValueType Ty>
1757 : NVDup<opcod1, 0b1011, opcod3, (outs QPR:$dst), (ins GPR:$src),
1758 NoItinerary, "vdup", !strconcat(asmSize, "\t$dst, $src"),
1759 [(set QPR:$dst, (Ty (NEONvdup (i32 GPR:$src))))]>;
1761 def VDUP8d : VDUPD<0b11101100, 0b00, ".8", v8i8>;
1762 def VDUP16d : VDUPD<0b11101000, 0b01, ".16", v4i16>;
1763 def VDUP32d : VDUPD<0b11101000, 0b00, ".32", v2i32>;
1764 def VDUP8q : VDUPQ<0b11101110, 0b00, ".8", v16i8>;
1765 def VDUP16q : VDUPQ<0b11101010, 0b01, ".16", v8i16>;
1766 def VDUP32q : VDUPQ<0b11101010, 0b00, ".32", v4i32>;
1768 def VDUPfd : NVDup<0b11101000, 0b1011, 0b00, (outs DPR:$dst), (ins GPR:$src),
1769 NoItinerary, "vdup", ".32\t$dst, $src",
1770 [(set DPR:$dst, (v2f32 (NEONvdup
1771 (f32 (bitconvert GPR:$src)))))]>;
1772 def VDUPfq : NVDup<0b11101010, 0b1011, 0b00, (outs QPR:$dst), (ins GPR:$src),
1773 NoItinerary, "vdup", ".32\t$dst, $src",
1774 [(set QPR:$dst, (v4f32 (NEONvdup
1775 (f32 (bitconvert GPR:$src)))))]>;
1777 // VDUP : Vector Duplicate Lane (from scalar to all elements)
1779 class VDUPLND<bits<2> op19_18, bits<2> op17_16, string OpcodeStr, ValueType Ty>
1780 : N2V<0b11, 0b11, op19_18, op17_16, 0b11000, 0, 0,
1781 (outs DPR:$dst), (ins DPR:$src, lane_cst:$lane), NoItinerary,
1782 !strconcat(OpcodeStr, "\t$dst, $src[$lane]"), "",
1783 [(set DPR:$dst, (Ty (NEONvduplane (Ty DPR:$src), imm:$lane)))]>;
1785 class VDUPLNQ<bits<2> op19_18, bits<2> op17_16, string OpcodeStr,
1786 ValueType ResTy, ValueType OpTy>
1787 : N2V<0b11, 0b11, op19_18, op17_16, 0b11000, 1, 0,
1788 (outs QPR:$dst), (ins DPR:$src, lane_cst:$lane), NoItinerary,
1789 !strconcat(OpcodeStr, "\t$dst, $src[$lane]"), "",
1790 [(set QPR:$dst, (ResTy (NEONvduplane (OpTy DPR:$src), imm:$lane)))]>;
1792 def VDUPLN8d : VDUPLND<0b00, 0b01, "vdup.8", v8i8>;
1793 def VDUPLN16d : VDUPLND<0b00, 0b10, "vdup.16", v4i16>;
1794 def VDUPLN32d : VDUPLND<0b01, 0b00, "vdup.32", v2i32>;
1795 def VDUPLNfd : VDUPLND<0b01, 0b00, "vdup.32", v2f32>;
1796 def VDUPLN8q : VDUPLNQ<0b00, 0b01, "vdup.8", v16i8, v8i8>;
1797 def VDUPLN16q : VDUPLNQ<0b00, 0b10, "vdup.16", v8i16, v4i16>;
1798 def VDUPLN32q : VDUPLNQ<0b01, 0b00, "vdup.32", v4i32, v2i32>;
1799 def VDUPLNfq : VDUPLNQ<0b01, 0b00, "vdup.32", v4f32, v2f32>;
1801 def : Pat<(v16i8 (NEONvduplane (v16i8 QPR:$src), imm:$lane)),
1802 (v16i8 (VDUPLN8q (v8i8 (EXTRACT_SUBREG QPR:$src,
1803 (DSubReg_i8_reg imm:$lane))),
1804 (SubReg_i8_lane imm:$lane)))>;
1805 def : Pat<(v8i16 (NEONvduplane (v8i16 QPR:$src), imm:$lane)),
1806 (v8i16 (VDUPLN16q (v4i16 (EXTRACT_SUBREG QPR:$src,
1807 (DSubReg_i16_reg imm:$lane))),
1808 (SubReg_i16_lane imm:$lane)))>;
1809 def : Pat<(v4i32 (NEONvduplane (v4i32 QPR:$src), imm:$lane)),
1810 (v4i32 (VDUPLN32q (v2i32 (EXTRACT_SUBREG QPR:$src,
1811 (DSubReg_i32_reg imm:$lane))),
1812 (SubReg_i32_lane imm:$lane)))>;
1813 def : Pat<(v4f32 (NEONvduplane (v4f32 QPR:$src), imm:$lane)),
1814 (v4f32 (VDUPLNfq (v2f32 (EXTRACT_SUBREG QPR:$src,
1815 (DSubReg_i32_reg imm:$lane))),
1816 (SubReg_i32_lane imm:$lane)))>;
1818 def VDUPfdf : N2V<0b11, 0b11, 0b01, 0b00, 0b11000, 0, 0,
1819 (outs DPR:$dst), (ins SPR:$src),
1820 NoItinerary, "vdup.32\t$dst, ${src:lane}", "",
1821 [(set DPR:$dst, (v2f32 (NEONvdup (f32 SPR:$src))))]>;
1823 def VDUPfqf : N2V<0b11, 0b11, 0b01, 0b00, 0b11000, 1, 0,
1824 (outs QPR:$dst), (ins SPR:$src),
1825 NoItinerary, "vdup.32\t$dst, ${src:lane}", "",
1826 [(set QPR:$dst, (v4f32 (NEONvdup (f32 SPR:$src))))]>;
1828 // VMOVN : Vector Narrowing Move
1829 defm VMOVN : N2VNInt_HSD<0b11,0b11,0b10,0b00100,0,0, "vmovn.i",
1830 int_arm_neon_vmovn>;
1831 // VQMOVN : Vector Saturating Narrowing Move
1832 defm VQMOVNs : N2VNInt_HSD<0b11,0b11,0b10,0b00101,0,0, "vqmovn.s",
1833 int_arm_neon_vqmovns>;
1834 defm VQMOVNu : N2VNInt_HSD<0b11,0b11,0b10,0b00101,1,0, "vqmovn.u",
1835 int_arm_neon_vqmovnu>;
1836 defm VQMOVNsu : N2VNInt_HSD<0b11,0b11,0b10,0b00100,1,0, "vqmovun.s",
1837 int_arm_neon_vqmovnsu>;
1838 // VMOVL : Vector Lengthening Move
1839 defm VMOVLs : N2VLInt_QHS<0,1,0b1010,0,0,1, "vmovl.s", int_arm_neon_vmovls>;
1840 defm VMOVLu : N2VLInt_QHS<1,1,0b1010,0,0,1, "vmovl.u", int_arm_neon_vmovlu>;
1842 // Vector Conversions.
1844 // VCVT : Vector Convert Between Floating-Point and Integers
1845 def VCVTf2sd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt.s32.f32",
1846 v2i32, v2f32, fp_to_sint>;
1847 def VCVTf2ud : N2VD<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt.u32.f32",
1848 v2i32, v2f32, fp_to_uint>;
1849 def VCVTs2fd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt.f32.s32",
1850 v2f32, v2i32, sint_to_fp>;
1851 def VCVTu2fd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt.f32.u32",
1852 v2f32, v2i32, uint_to_fp>;
1854 def VCVTf2sq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt.s32.f32",
1855 v4i32, v4f32, fp_to_sint>;
1856 def VCVTf2uq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt.u32.f32",
1857 v4i32, v4f32, fp_to_uint>;
1858 def VCVTs2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt.f32.s32",
1859 v4f32, v4i32, sint_to_fp>;
1860 def VCVTu2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt.f32.u32",
1861 v4f32, v4i32, uint_to_fp>;
1863 // VCVT : Vector Convert Between Floating-Point and Fixed-Point.
1864 // Note: Some of the opcode bits in the following VCVT instructions need to
1865 // be encoded based on the immed values.
1866 def VCVTf2xsd : N2VCvtD<0, 1, 0b000000, 0b1111, 0, 1, "vcvt.s32.f32",
1867 v2i32, v2f32, int_arm_neon_vcvtfp2fxs>;
1868 def VCVTf2xud : N2VCvtD<1, 1, 0b000000, 0b1111, 0, 1, "vcvt.u32.f32",
1869 v2i32, v2f32, int_arm_neon_vcvtfp2fxu>;
1870 def VCVTxs2fd : N2VCvtD<0, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.s32",
1871 v2f32, v2i32, int_arm_neon_vcvtfxs2fp>;
1872 def VCVTxu2fd : N2VCvtD<1, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.u32",
1873 v2f32, v2i32, int_arm_neon_vcvtfxu2fp>;
1875 def VCVTf2xsq : N2VCvtQ<0, 1, 0b000000, 0b1111, 0, 1, "vcvt.s32.f32",
1876 v4i32, v4f32, int_arm_neon_vcvtfp2fxs>;
1877 def VCVTf2xuq : N2VCvtQ<1, 1, 0b000000, 0b1111, 0, 1, "vcvt.u32.f32",
1878 v4i32, v4f32, int_arm_neon_vcvtfp2fxu>;
1879 def VCVTxs2fq : N2VCvtQ<0, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.s32",
1880 v4f32, v4i32, int_arm_neon_vcvtfxs2fp>;
1881 def VCVTxu2fq : N2VCvtQ<1, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.u32",
1882 v4f32, v4i32, int_arm_neon_vcvtfxu2fp>;
1886 // VREV64 : Vector Reverse elements within 64-bit doublewords
1888 class VREV64D<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1889 : N2V<0b11, 0b11, op19_18, 0b00, 0b00000, 0, 0, (outs DPR:$dst),
1890 (ins DPR:$src), NoItinerary,
1891 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1892 [(set DPR:$dst, (Ty (NEONvrev64 (Ty DPR:$src))))]>;
1893 class VREV64Q<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1894 : N2V<0b11, 0b11, op19_18, 0b00, 0b00000, 1, 0, (outs QPR:$dst),
1895 (ins QPR:$src), NoItinerary,
1896 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1897 [(set QPR:$dst, (Ty (NEONvrev64 (Ty QPR:$src))))]>;
1899 def VREV64d8 : VREV64D<0b00, "vrev64.8", v8i8>;
1900 def VREV64d16 : VREV64D<0b01, "vrev64.16", v4i16>;
1901 def VREV64d32 : VREV64D<0b10, "vrev64.32", v2i32>;
1902 def VREV64df : VREV64D<0b10, "vrev64.32", v2f32>;
1904 def VREV64q8 : VREV64Q<0b00, "vrev64.8", v16i8>;
1905 def VREV64q16 : VREV64Q<0b01, "vrev64.16", v8i16>;
1906 def VREV64q32 : VREV64Q<0b10, "vrev64.32", v4i32>;
1907 def VREV64qf : VREV64Q<0b10, "vrev64.32", v4f32>;
1909 // VREV32 : Vector Reverse elements within 32-bit words
1911 class VREV32D<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1912 : N2V<0b11, 0b11, op19_18, 0b00, 0b00001, 0, 0, (outs DPR:$dst),
1913 (ins DPR:$src), NoItinerary,
1914 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1915 [(set DPR:$dst, (Ty (NEONvrev32 (Ty DPR:$src))))]>;
1916 class VREV32Q<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1917 : N2V<0b11, 0b11, op19_18, 0b00, 0b00001, 1, 0, (outs QPR:$dst),
1918 (ins QPR:$src), NoItinerary,
1919 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1920 [(set QPR:$dst, (Ty (NEONvrev32 (Ty QPR:$src))))]>;
1922 def VREV32d8 : VREV32D<0b00, "vrev32.8", v8i8>;
1923 def VREV32d16 : VREV32D<0b01, "vrev32.16", v4i16>;
1925 def VREV32q8 : VREV32Q<0b00, "vrev32.8", v16i8>;
1926 def VREV32q16 : VREV32Q<0b01, "vrev32.16", v8i16>;
1928 // VREV16 : Vector Reverse elements within 16-bit halfwords
1930 class VREV16D<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1931 : N2V<0b11, 0b11, op19_18, 0b00, 0b00010, 0, 0, (outs DPR:$dst),
1932 (ins DPR:$src), NoItinerary,
1933 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1934 [(set DPR:$dst, (Ty (NEONvrev16 (Ty DPR:$src))))]>;
1935 class VREV16Q<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1936 : N2V<0b11, 0b11, op19_18, 0b00, 0b00010, 1, 0, (outs QPR:$dst),
1937 (ins QPR:$src), NoItinerary,
1938 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1939 [(set QPR:$dst, (Ty (NEONvrev16 (Ty QPR:$src))))]>;
1941 def VREV16d8 : VREV16D<0b00, "vrev16.8", v8i8>;
1942 def VREV16q8 : VREV16Q<0b00, "vrev16.8", v16i8>;
1944 // VTRN : Vector Transpose
1946 def VTRNd8 : N2VDShuffle<0b00, 0b00001, "vtrn.8">;
1947 def VTRNd16 : N2VDShuffle<0b01, 0b00001, "vtrn.16">;
1948 def VTRNd32 : N2VDShuffle<0b10, 0b00001, "vtrn.32">;
1950 def VTRNq8 : N2VQShuffle<0b00, 0b00001, "vtrn.8">;
1951 def VTRNq16 : N2VQShuffle<0b01, 0b00001, "vtrn.16">;
1952 def VTRNq32 : N2VQShuffle<0b10, 0b00001, "vtrn.32">;
1954 // VUZP : Vector Unzip (Deinterleave)
1956 def VUZPd8 : N2VDShuffle<0b00, 0b00010, "vuzp.8">;
1957 def VUZPd16 : N2VDShuffle<0b01, 0b00010, "vuzp.16">;
1958 def VUZPd32 : N2VDShuffle<0b10, 0b00010, "vuzp.32">;
1960 def VUZPq8 : N2VQShuffle<0b00, 0b00010, "vuzp.8">;
1961 def VUZPq16 : N2VQShuffle<0b01, 0b00010, "vuzp.16">;
1962 def VUZPq32 : N2VQShuffle<0b10, 0b00010, "vuzp.32">;
1964 // VZIP : Vector Zip (Interleave)
1966 def VZIPd8 : N2VDShuffle<0b00, 0b00011, "vzip.8">;
1967 def VZIPd16 : N2VDShuffle<0b01, 0b00011, "vzip.16">;
1968 def VZIPd32 : N2VDShuffle<0b10, 0b00011, "vzip.32">;
1970 def VZIPq8 : N2VQShuffle<0b00, 0b00011, "vzip.8">;
1971 def VZIPq16 : N2VQShuffle<0b01, 0b00011, "vzip.16">;
1972 def VZIPq32 : N2VQShuffle<0b10, 0b00011, "vzip.32">;
1974 // Vector Table Lookup and Table Extension.
1976 // VTBL : Vector Table Lookup
1978 : N3V<1,1,0b11,0b1000,0,0, (outs DPR:$dst),
1979 (ins DPR:$tbl1, DPR:$src), NoItinerary,
1980 "vtbl.8\t$dst, \\{$tbl1\\}, $src", "",
1981 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbl1 DPR:$tbl1, DPR:$src)))]>;
1983 : N3V<1,1,0b11,0b1001,0,0, (outs DPR:$dst),
1984 (ins DPR:$tbl1, DPR:$tbl2, DPR:$src), NoItinerary,
1985 "vtbl.8\t$dst, \\{$tbl1,$tbl2\\}, $src", "",
1986 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbl2
1987 DPR:$tbl1, DPR:$tbl2, DPR:$src)))]>;
1989 : N3V<1,1,0b11,0b1010,0,0, (outs DPR:$dst),
1990 (ins DPR:$tbl1, DPR:$tbl2, DPR:$tbl3, DPR:$src), NoItinerary,
1991 "vtbl.8\t$dst, \\{$tbl1,$tbl2,$tbl3\\}, $src", "",
1992 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbl3
1993 DPR:$tbl1, DPR:$tbl2, DPR:$tbl3, DPR:$src)))]>;
1995 : N3V<1,1,0b11,0b1011,0,0, (outs DPR:$dst),
1996 (ins DPR:$tbl1, DPR:$tbl2, DPR:$tbl3, DPR:$tbl4, DPR:$src), NoItinerary,
1997 "vtbl.8\t$dst, \\{$tbl1,$tbl2,$tbl3,$tbl4\\}, $src", "",
1998 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbl4 DPR:$tbl1, DPR:$tbl2,
1999 DPR:$tbl3, DPR:$tbl4, DPR:$src)))]>;
2001 // VTBX : Vector Table Extension
2003 : N3V<1,1,0b11,0b1000,1,0, (outs DPR:$dst),
2004 (ins DPR:$orig, DPR:$tbl1, DPR:$src), NoItinerary,
2005 "vtbx.8\t$dst, \\{$tbl1\\}, $src", "$orig = $dst",
2006 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbx1
2007 DPR:$orig, DPR:$tbl1, DPR:$src)))]>;
2009 : N3V<1,1,0b11,0b1001,1,0, (outs DPR:$dst),
2010 (ins DPR:$orig, DPR:$tbl1, DPR:$tbl2, DPR:$src), NoItinerary,
2011 "vtbx.8\t$dst, \\{$tbl1,$tbl2\\}, $src", "$orig = $dst",
2012 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbx2
2013 DPR:$orig, DPR:$tbl1, DPR:$tbl2, DPR:$src)))]>;
2015 : N3V<1,1,0b11,0b1010,1,0, (outs DPR:$dst),
2016 (ins DPR:$orig, DPR:$tbl1, DPR:$tbl2, DPR:$tbl3, DPR:$src), NoItinerary,
2017 "vtbx.8\t$dst, \\{$tbl1,$tbl2,$tbl3\\}, $src", "$orig = $dst",
2018 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbx3 DPR:$orig, DPR:$tbl1,
2019 DPR:$tbl2, DPR:$tbl3, DPR:$src)))]>;
2021 : N3V<1,1,0b11,0b1011,1,0, (outs DPR:$dst), (ins DPR:$orig, DPR:$tbl1,
2022 DPR:$tbl2, DPR:$tbl3, DPR:$tbl4, DPR:$src), NoItinerary,
2023 "vtbx.8\t$dst, \\{$tbl1,$tbl2,$tbl3,$tbl4\\}, $src", "$orig = $dst",
2024 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbx4 DPR:$orig, DPR:$tbl1,
2025 DPR:$tbl2, DPR:$tbl3, DPR:$tbl4, DPR:$src)))]>;
2027 //===----------------------------------------------------------------------===//
2028 // NEON instructions for single-precision FP math
2029 //===----------------------------------------------------------------------===//
2031 // These need separate instructions because they must use DPR_VFP2 register
2032 // class which have SPR sub-registers.
2034 // Vector Add Operations used for single-precision FP
2035 let neverHasSideEffects = 1 in
2036 def VADDfd_sfp : N3VDs<0, 0, 0b00, 0b1101, 0, "vadd.f32", v2f32, v2f32, fadd,1>;
2037 def : N3VDsPat<fadd, VADDfd_sfp>;
2039 // Vector Sub Operations used for single-precision FP
2040 let neverHasSideEffects = 1 in
2041 def VSUBfd_sfp : N3VDs<0, 0, 0b10, 0b1101, 0, "vsub.f32", v2f32, v2f32, fsub,0>;
2042 def : N3VDsPat<fsub, VSUBfd_sfp>;
2044 // Vector Multiply Operations used for single-precision FP
2045 let neverHasSideEffects = 1 in
2046 def VMULfd_sfp : N3VDs<1, 0, 0b00, 0b1101, 1, "vmul.f32", v2f32, v2f32, fmul,1>;
2047 def : N3VDsPat<fmul, VMULfd_sfp>;
2049 // Vector Multiply-Accumulate/Subtract used for single-precision FP
2050 let neverHasSideEffects = 1 in
2051 def VMLAfd_sfp : N3VDMulOps<0, 0, 0b00, 0b1101, 1, "vmla.f32", v2f32,fmul,fadd>;
2052 def : N3VDMulOpsPat<fmul, fadd, VMLAfd_sfp>;
2054 let neverHasSideEffects = 1 in
2055 def VMLSfd_sfp : N3VDMulOps<0, 0, 0b10, 0b1101, 1, "vmls.f32", v2f32,fmul,fsub>;
2056 def : N3VDMulOpsPat<fmul, fsub, VMLSfd_sfp>;
2058 // Vector Absolute used for single-precision FP
2059 let neverHasSideEffects = 1 in
2060 def VABSfd_sfp : N2VDInts<0b11, 0b11, 0b10, 0b01, 0b01110, 0, "vabs.f32",
2061 v2f32, v2f32, int_arm_neon_vabs>;
2062 def : N2VDIntsPat<fabs, VABSfd_sfp>;
2064 // Vector Negate used for single-precision FP
2065 let neverHasSideEffects = 1 in
2066 def VNEGf32d_sfp : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 0, 0,
2067 (outs DPR_VFP2:$dst), (ins DPR_VFP2:$src), NoItinerary,
2068 "vneg.f32\t$dst, $src", "", []>;
2069 def : N2VDIntsPat<fneg, VNEGf32d_sfp>;
2071 // Vector Convert between single-precision FP and integer
2072 let neverHasSideEffects = 1 in
2073 def VCVTf2sd_sfp : N2VDs<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt.s32.f32",
2074 v2i32, v2f32, fp_to_sint>;
2075 def : N2VDsPat<arm_ftosi, f32, v2f32, VCVTf2sd_sfp>;
2077 let neverHasSideEffects = 1 in
2078 def VCVTf2ud_sfp : N2VDs<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt.u32.f32",
2079 v2i32, v2f32, fp_to_uint>;
2080 def : N2VDsPat<arm_ftoui, f32, v2f32, VCVTf2ud_sfp>;
2082 let neverHasSideEffects = 1 in
2083 def VCVTs2fd_sfp : N2VDs<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt.f32.s32",
2084 v2f32, v2i32, sint_to_fp>;
2085 def : N2VDsPat<arm_sitof, f32, v2i32, VCVTs2fd_sfp>;
2087 let neverHasSideEffects = 1 in
2088 def VCVTu2fd_sfp : N2VDs<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt.f32.u32",
2089 v2f32, v2i32, uint_to_fp>;
2090 def : N2VDsPat<arm_uitof, f32, v2i32, VCVTu2fd_sfp>;
2092 //===----------------------------------------------------------------------===//
2093 // Non-Instruction Patterns
2094 //===----------------------------------------------------------------------===//
2097 def : Pat<(v1i64 (bitconvert (v2i32 DPR:$src))), (v1i64 DPR:$src)>;
2098 def : Pat<(v1i64 (bitconvert (v4i16 DPR:$src))), (v1i64 DPR:$src)>;
2099 def : Pat<(v1i64 (bitconvert (v8i8 DPR:$src))), (v1i64 DPR:$src)>;
2100 def : Pat<(v1i64 (bitconvert (f64 DPR:$src))), (v1i64 DPR:$src)>;
2101 def : Pat<(v1i64 (bitconvert (v2f32 DPR:$src))), (v1i64 DPR:$src)>;
2102 def : Pat<(v2i32 (bitconvert (v1i64 DPR:$src))), (v2i32 DPR:$src)>;
2103 def : Pat<(v2i32 (bitconvert (v4i16 DPR:$src))), (v2i32 DPR:$src)>;
2104 def : Pat<(v2i32 (bitconvert (v8i8 DPR:$src))), (v2i32 DPR:$src)>;
2105 def : Pat<(v2i32 (bitconvert (f64 DPR:$src))), (v2i32 DPR:$src)>;
2106 def : Pat<(v2i32 (bitconvert (v2f32 DPR:$src))), (v2i32 DPR:$src)>;
2107 def : Pat<(v4i16 (bitconvert (v1i64 DPR:$src))), (v4i16 DPR:$src)>;
2108 def : Pat<(v4i16 (bitconvert (v2i32 DPR:$src))), (v4i16 DPR:$src)>;
2109 def : Pat<(v4i16 (bitconvert (v8i8 DPR:$src))), (v4i16 DPR:$src)>;
2110 def : Pat<(v4i16 (bitconvert (f64 DPR:$src))), (v4i16 DPR:$src)>;
2111 def : Pat<(v4i16 (bitconvert (v2f32 DPR:$src))), (v4i16 DPR:$src)>;
2112 def : Pat<(v8i8 (bitconvert (v1i64 DPR:$src))), (v8i8 DPR:$src)>;
2113 def : Pat<(v8i8 (bitconvert (v2i32 DPR:$src))), (v8i8 DPR:$src)>;
2114 def : Pat<(v8i8 (bitconvert (v4i16 DPR:$src))), (v8i8 DPR:$src)>;
2115 def : Pat<(v8i8 (bitconvert (f64 DPR:$src))), (v8i8 DPR:$src)>;
2116 def : Pat<(v8i8 (bitconvert (v2f32 DPR:$src))), (v8i8 DPR:$src)>;
2117 def : Pat<(f64 (bitconvert (v1i64 DPR:$src))), (f64 DPR:$src)>;
2118 def : Pat<(f64 (bitconvert (v2i32 DPR:$src))), (f64 DPR:$src)>;
2119 def : Pat<(f64 (bitconvert (v4i16 DPR:$src))), (f64 DPR:$src)>;
2120 def : Pat<(f64 (bitconvert (v8i8 DPR:$src))), (f64 DPR:$src)>;
2121 def : Pat<(f64 (bitconvert (v2f32 DPR:$src))), (f64 DPR:$src)>;
2122 def : Pat<(v2f32 (bitconvert (f64 DPR:$src))), (v2f32 DPR:$src)>;
2123 def : Pat<(v2f32 (bitconvert (v1i64 DPR:$src))), (v2f32 DPR:$src)>;
2124 def : Pat<(v2f32 (bitconvert (v2i32 DPR:$src))), (v2f32 DPR:$src)>;
2125 def : Pat<(v2f32 (bitconvert (v4i16 DPR:$src))), (v2f32 DPR:$src)>;
2126 def : Pat<(v2f32 (bitconvert (v8i8 DPR:$src))), (v2f32 DPR:$src)>;
2128 def : Pat<(v2i64 (bitconvert (v4i32 QPR:$src))), (v2i64 QPR:$src)>;
2129 def : Pat<(v2i64 (bitconvert (v8i16 QPR:$src))), (v2i64 QPR:$src)>;
2130 def : Pat<(v2i64 (bitconvert (v16i8 QPR:$src))), (v2i64 QPR:$src)>;
2131 def : Pat<(v2i64 (bitconvert (v2f64 QPR:$src))), (v2i64 QPR:$src)>;
2132 def : Pat<(v2i64 (bitconvert (v4f32 QPR:$src))), (v2i64 QPR:$src)>;
2133 def : Pat<(v4i32 (bitconvert (v2i64 QPR:$src))), (v4i32 QPR:$src)>;
2134 def : Pat<(v4i32 (bitconvert (v8i16 QPR:$src))), (v4i32 QPR:$src)>;
2135 def : Pat<(v4i32 (bitconvert (v16i8 QPR:$src))), (v4i32 QPR:$src)>;
2136 def : Pat<(v4i32 (bitconvert (v2f64 QPR:$src))), (v4i32 QPR:$src)>;
2137 def : Pat<(v4i32 (bitconvert (v4f32 QPR:$src))), (v4i32 QPR:$src)>;
2138 def : Pat<(v8i16 (bitconvert (v2i64 QPR:$src))), (v8i16 QPR:$src)>;
2139 def : Pat<(v8i16 (bitconvert (v4i32 QPR:$src))), (v8i16 QPR:$src)>;
2140 def : Pat<(v8i16 (bitconvert (v16i8 QPR:$src))), (v8i16 QPR:$src)>;
2141 def : Pat<(v8i16 (bitconvert (v2f64 QPR:$src))), (v8i16 QPR:$src)>;
2142 def : Pat<(v8i16 (bitconvert (v4f32 QPR:$src))), (v8i16 QPR:$src)>;
2143 def : Pat<(v16i8 (bitconvert (v2i64 QPR:$src))), (v16i8 QPR:$src)>;
2144 def : Pat<(v16i8 (bitconvert (v4i32 QPR:$src))), (v16i8 QPR:$src)>;
2145 def : Pat<(v16i8 (bitconvert (v8i16 QPR:$src))), (v16i8 QPR:$src)>;
2146 def : Pat<(v16i8 (bitconvert (v2f64 QPR:$src))), (v16i8 QPR:$src)>;
2147 def : Pat<(v16i8 (bitconvert (v4f32 QPR:$src))), (v16i8 QPR:$src)>;
2148 def : Pat<(v4f32 (bitconvert (v2i64 QPR:$src))), (v4f32 QPR:$src)>;
2149 def : Pat<(v4f32 (bitconvert (v4i32 QPR:$src))), (v4f32 QPR:$src)>;
2150 def : Pat<(v4f32 (bitconvert (v8i16 QPR:$src))), (v4f32 QPR:$src)>;
2151 def : Pat<(v4f32 (bitconvert (v16i8 QPR:$src))), (v4f32 QPR:$src)>;
2152 def : Pat<(v4f32 (bitconvert (v2f64 QPR:$src))), (v4f32 QPR:$src)>;
2153 def : Pat<(v2f64 (bitconvert (v2i64 QPR:$src))), (v2f64 QPR:$src)>;
2154 def : Pat<(v2f64 (bitconvert (v4i32 QPR:$src))), (v2f64 QPR:$src)>;
2155 def : Pat<(v2f64 (bitconvert (v8i16 QPR:$src))), (v2f64 QPR:$src)>;
2156 def : Pat<(v2f64 (bitconvert (v16i8 QPR:$src))), (v2f64 QPR:$src)>;
2157 def : Pat<(v2f64 (bitconvert (v4f32 QPR:$src))), (v2f64 QPR:$src)>;