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 // VDUPLANE can produce a quad-register result from a double-register source,
69 // so the result is not constrained to match the source.
70 def NEONvduplane : SDNode<"ARMISD::VDUPLANE",
71 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
74 def SDTARMVLD2 : SDTypeProfile<2, 1, [SDTCisSameAs<0, 1>, SDTCisPtrTy<2>]>;
75 def SDTARMVLD3 : SDTypeProfile<3, 1, [SDTCisSameAs<0, 1>,
76 SDTCisSameAs<0, 2>, SDTCisPtrTy<3>]>;
77 def SDTARMVLD4 : SDTypeProfile<4, 1, [SDTCisSameAs<0, 1>,
79 SDTCisSameAs<0, 3>, SDTCisPtrTy<4>]>;
80 def NEONvld2d : SDNode<"ARMISD::VLD2D", SDTARMVLD2,
81 [SDNPHasChain, SDNPMayLoad]>;
82 def NEONvld3d : SDNode<"ARMISD::VLD3D", SDTARMVLD3,
83 [SDNPHasChain, SDNPMayLoad]>;
84 def NEONvld4d : SDNode<"ARMISD::VLD4D", SDTARMVLD4,
85 [SDNPHasChain, SDNPMayLoad]>;
87 def SDTARMVST2 : SDTypeProfile<0, 3, [SDTCisPtrTy<0>, SDTCisSameAs<1, 2>]>;
88 def SDTARMVST3 : SDTypeProfile<0, 4, [SDTCisPtrTy<0>, SDTCisSameAs<1, 2>,
90 def SDTARMVST4 : SDTypeProfile<0, 5, [SDTCisPtrTy<0>, SDTCisSameAs<1, 2>,
94 def NEONvst2d : SDNode<"ARMISD::VST2D", SDTARMVST2,
95 [SDNPHasChain, SDNPMayStore]>;
96 def NEONvst3d : SDNode<"ARMISD::VST3D", SDTARMVST3,
97 [SDNPHasChain, SDNPMayStore]>;
98 def NEONvst4d : SDNode<"ARMISD::VST4D", SDTARMVST4,
99 [SDNPHasChain, SDNPMayStore]>;
101 def SDTARMVSHUF : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisSameAs<0, 1>]>;
102 def NEONvrev64 : SDNode<"ARMISD::VREV64", SDTARMVSHUF>;
103 def NEONvrev32 : SDNode<"ARMISD::VREV32", SDTARMVSHUF>;
104 def NEONvrev16 : SDNode<"ARMISD::VREV16", SDTARMVSHUF>;
106 //===----------------------------------------------------------------------===//
107 // NEON operand definitions
108 //===----------------------------------------------------------------------===//
110 // addrmode_neonldstm := reg
112 /* TODO: Take advantage of vldm.
113 def addrmode_neonldstm : Operand<i32>,
114 ComplexPattern<i32, 2, "SelectAddrModeNeonLdStM", []> {
115 let PrintMethod = "printAddrNeonLdStMOperand";
116 let MIOperandInfo = (ops GPR, i32imm);
120 //===----------------------------------------------------------------------===//
121 // NEON load / store instructions
122 //===----------------------------------------------------------------------===//
124 /* TODO: Take advantage of vldm.
126 def VLDMD : NI<(outs),
127 (ins addrmode_neonldstm:$addr, reglist:$dst1, variable_ops),
129 "vldm${addr:submode} ${addr:base}, $dst1",
131 let Inst{27-25} = 0b110;
133 let Inst{11-9} = 0b101;
136 def VLDMS : NI<(outs),
137 (ins addrmode_neonldstm:$addr, reglist:$dst1, variable_ops),
139 "vldm${addr:submode} ${addr:base}, $dst1",
141 let Inst{27-25} = 0b110;
143 let Inst{11-9} = 0b101;
148 // Use vldmia to load a Q register as a D register pair.
149 def VLDRQ : NI4<(outs QPR:$dst), (ins addrmode4:$addr),
151 "vldmia $addr, ${dst:dregpair}",
152 [(set QPR:$dst, (v2f64 (load addrmode4:$addr)))]> {
153 let Inst{27-25} = 0b110;
154 let Inst{24} = 0; // P bit
155 let Inst{23} = 1; // U bit
157 let Inst{11-9} = 0b101;
160 // Use vstmia to store a Q register as a D register pair.
161 def VSTRQ : NI4<(outs), (ins QPR:$src, addrmode4:$addr),
163 "vstmia $addr, ${src:dregpair}",
164 [(store (v2f64 QPR:$src), addrmode4:$addr)]> {
165 let Inst{27-25} = 0b110;
166 let Inst{24} = 0; // P bit
167 let Inst{23} = 1; // U bit
169 let Inst{11-9} = 0b101;
172 // VLD1 : Vector Load (multiple single elements)
173 class VLD1D<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
174 : NLdSt<(outs DPR:$dst), (ins addrmode6:$addr),
176 !strconcat(OpcodeStr, "\t\\{$dst\\}, $addr"),
177 [(set DPR:$dst, (Ty (IntOp addrmode6:$addr)))]>;
178 class VLD1Q<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
179 : NLdSt<(outs QPR:$dst), (ins addrmode6:$addr),
181 !strconcat(OpcodeStr, "\t${dst:dregpair}, $addr"),
182 [(set QPR:$dst, (Ty (IntOp addrmode6:$addr)))]>;
184 def VLD1d8 : VLD1D<"vld1.8", v8i8, int_arm_neon_vld1>;
185 def VLD1d16 : VLD1D<"vld1.16", v4i16, int_arm_neon_vld1>;
186 def VLD1d32 : VLD1D<"vld1.32", v2i32, int_arm_neon_vld1>;
187 def VLD1df : VLD1D<"vld1.32", v2f32, int_arm_neon_vld1>;
188 def VLD1d64 : VLD1D<"vld1.64", v1i64, int_arm_neon_vld1>;
190 def VLD1q8 : VLD1Q<"vld1.8", v16i8, int_arm_neon_vld1>;
191 def VLD1q16 : VLD1Q<"vld1.16", v8i16, int_arm_neon_vld1>;
192 def VLD1q32 : VLD1Q<"vld1.32", v4i32, int_arm_neon_vld1>;
193 def VLD1qf : VLD1Q<"vld1.32", v4f32, int_arm_neon_vld1>;
194 def VLD1q64 : VLD1Q<"vld1.64", v2i64, int_arm_neon_vld1>;
198 // VLD2 : Vector Load (multiple 2-element structures)
199 class VLD2D<string OpcodeStr>
200 : NLdSt<(outs DPR:$dst1, DPR:$dst2), (ins addrmode6:$addr),
202 !strconcat(OpcodeStr, "\t\\{$dst1,$dst2\\}, $addr"), []>;
204 def VLD2d8 : VLD2D<"vld2.8">;
205 def VLD2d16 : VLD2D<"vld2.16">;
206 def VLD2d32 : VLD2D<"vld2.32">;
208 // VLD3 : Vector Load (multiple 3-element structures)
209 class VLD3D<string OpcodeStr>
210 : NLdSt<(outs DPR:$dst1, DPR:$dst2, DPR:$dst3), (ins addrmode6:$addr),
212 !strconcat(OpcodeStr, "\t\\{$dst1,$dst2,$dst3\\}, $addr"), []>;
214 def VLD3d8 : VLD3D<"vld3.8">;
215 def VLD3d16 : VLD3D<"vld3.16">;
216 def VLD3d32 : VLD3D<"vld3.32">;
218 // VLD4 : Vector Load (multiple 4-element structures)
219 class VLD4D<string OpcodeStr>
220 : NLdSt<(outs DPR:$dst1, DPR:$dst2, DPR:$dst3, DPR:$dst4),
221 (ins addrmode6:$addr),
223 !strconcat(OpcodeStr, "\t\\{$dst1,$dst2,$dst3,$dst4\\}, $addr"), []>;
225 def VLD4d8 : VLD4D<"vld4.8">;
226 def VLD4d16 : VLD4D<"vld4.16">;
227 def VLD4d32 : VLD4D<"vld4.32">;
230 // VST1 : Vector Store (multiple single elements)
231 class VST1D<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
232 : NLdSt<(outs), (ins addrmode6:$addr, DPR:$src),
234 !strconcat(OpcodeStr, "\t\\{$src\\}, $addr"),
235 [(IntOp addrmode6:$addr, (Ty DPR:$src))]>;
236 class VST1Q<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
237 : NLdSt<(outs), (ins addrmode6:$addr, QPR:$src),
239 !strconcat(OpcodeStr, "\t${src:dregpair}, $addr"),
240 [(IntOp addrmode6:$addr, (Ty QPR:$src))]>;
242 def VST1d8 : VST1D<"vst1.8", v8i8, int_arm_neon_vst1>;
243 def VST1d16 : VST1D<"vst1.16", v4i16, int_arm_neon_vst1>;
244 def VST1d32 : VST1D<"vst1.32", v2i32, int_arm_neon_vst1>;
245 def VST1df : VST1D<"vst1.32", v2f32, int_arm_neon_vst1>;
246 def VST1d64 : VST1D<"vst1.64", v1i64, int_arm_neon_vst1>;
248 def VST1q8 : VST1Q<"vst1.8", v16i8, int_arm_neon_vst1>;
249 def VST1q16 : VST1Q<"vst1.16", v8i16, int_arm_neon_vst1>;
250 def VST1q32 : VST1Q<"vst1.32", v4i32, int_arm_neon_vst1>;
251 def VST1qf : VST1Q<"vst1.32", v4f32, int_arm_neon_vst1>;
252 def VST1q64 : VST1Q<"vst1.64", v2i64, int_arm_neon_vst1>;
254 let mayStore = 1 in {
256 // VST2 : Vector Store (multiple 2-element structures)
257 class VST2D<string OpcodeStr>
258 : NLdSt<(outs), (ins addrmode6:$addr, DPR:$src1, DPR:$src2), NoItinerary,
259 !strconcat(OpcodeStr, "\t\\{$src1,$src2\\}, $addr"), []>;
261 def VST2d8 : VST2D<"vst2.8">;
262 def VST2d16 : VST2D<"vst2.16">;
263 def VST2d32 : VST2D<"vst2.32">;
265 // VST3 : Vector Store (multiple 3-element structures)
266 class VST3D<string OpcodeStr>
267 : NLdSt<(outs), (ins addrmode6:$addr, DPR:$src1, DPR:$src2, DPR:$src3),
269 !strconcat(OpcodeStr, "\t\\{$src1,$src2,$src3\\}, $addr"), []>;
271 def VST3d8 : VST3D<"vst3.8">;
272 def VST3d16 : VST3D<"vst3.16">;
273 def VST3d32 : VST3D<"vst3.32">;
275 // VST4 : Vector Store (multiple 4-element structures)
276 class VST4D<string OpcodeStr>
277 : NLdSt<(outs), (ins addrmode6:$addr,
278 DPR:$src1, DPR:$src2, DPR:$src3, DPR:$src4), NoItinerary,
279 !strconcat(OpcodeStr, "\t\\{$src1,$src2,$src3,$src4\\}, $addr"), []>;
281 def VST4d8 : VST4D<"vst4.8">;
282 def VST4d16 : VST4D<"vst4.16">;
283 def VST4d32 : VST4D<"vst4.32">;
287 //===----------------------------------------------------------------------===//
288 // NEON pattern fragments
289 //===----------------------------------------------------------------------===//
291 // Extract D sub-registers of Q registers.
292 // (arm_dsubreg_0 is 5; arm_dsubreg_1 is 6)
293 def DSubReg_i8_reg : SDNodeXForm<imm, [{
294 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 8, MVT::i32);
296 def DSubReg_i16_reg : SDNodeXForm<imm, [{
297 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 4, MVT::i32);
299 def DSubReg_i32_reg : SDNodeXForm<imm, [{
300 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 2, MVT::i32);
302 def DSubReg_f64_reg : SDNodeXForm<imm, [{
303 return CurDAG->getTargetConstant(5 + N->getZExtValue(), MVT::i32);
306 // Extract S sub-registers of Q registers.
307 // (arm_ssubreg_0 is 1; arm_ssubreg_1 is 2; etc.)
308 def SSubReg_f32_reg : SDNodeXForm<imm, [{
309 return CurDAG->getTargetConstant(1 + N->getZExtValue(), MVT::i32);
312 // Translate lane numbers from Q registers to D subregs.
313 def SubReg_i8_lane : SDNodeXForm<imm, [{
314 return CurDAG->getTargetConstant(N->getZExtValue() & 7, MVT::i32);
316 def SubReg_i16_lane : SDNodeXForm<imm, [{
317 return CurDAG->getTargetConstant(N->getZExtValue() & 3, MVT::i32);
319 def SubReg_i32_lane : SDNodeXForm<imm, [{
320 return CurDAG->getTargetConstant(N->getZExtValue() & 1, MVT::i32);
323 //===----------------------------------------------------------------------===//
324 // Instruction Classes
325 //===----------------------------------------------------------------------===//
327 // Basic 2-register operations, both double- and quad-register.
328 class N2VD<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
329 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
330 ValueType ResTy, ValueType OpTy, SDNode OpNode>
331 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
332 (ins DPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
333 [(set DPR:$dst, (ResTy (OpNode (OpTy DPR:$src))))]>;
334 class N2VQ<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
335 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
336 ValueType ResTy, ValueType OpTy, SDNode OpNode>
337 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
338 (ins QPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
339 [(set QPR:$dst, (ResTy (OpNode (OpTy QPR:$src))))]>;
341 // Basic 2-register operations, scalar single-precision.
342 class N2VDs<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
343 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
344 ValueType ResTy, ValueType OpTy, SDNode OpNode>
345 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4,
346 (outs DPR_VFP2:$dst), (ins DPR_VFP2:$src),
347 NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "", []>;
349 class N2VDsPat<SDNode OpNode, ValueType ResTy, ValueType OpTy, NeonI Inst>
350 : NEONFPPat<(ResTy (OpNode SPR:$a)),
352 (Inst (INSERT_SUBREG (OpTy (IMPLICIT_DEF)), SPR:$a, arm_ssubreg_0)),
355 // Basic 2-register intrinsics, both double- and quad-register.
356 class N2VDInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
357 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
358 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
359 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
360 (ins DPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
361 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src))))]>;
362 class N2VQInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
363 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
364 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
365 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
366 (ins QPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
367 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src))))]>;
369 // Basic 2-register intrinsics, scalar single-precision
370 class N2VDInts<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
371 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
372 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
373 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4,
374 (outs DPR_VFP2:$dst), (ins DPR_VFP2:$src), NoItinerary,
375 !strconcat(OpcodeStr, "\t$dst, $src"), "", []>;
377 class N2VDIntsPat<SDNode OpNode, NeonI Inst>
378 : NEONFPPat<(f32 (OpNode SPR:$a)),
380 (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$a, arm_ssubreg_0)),
383 // Narrow 2-register intrinsics.
384 class N2VNInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
385 bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
386 string OpcodeStr, ValueType TyD, ValueType TyQ, Intrinsic IntOp>
387 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, op6, op4, (outs DPR:$dst),
388 (ins QPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
389 [(set DPR:$dst, (TyD (IntOp (TyQ QPR:$src))))]>;
391 // Long 2-register intrinsics. (This is currently only used for VMOVL and is
392 // derived from N2VImm instead of N2V because of the way the size is encoded.)
393 class N2VLInt<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
394 bit op6, bit op4, string OpcodeStr, ValueType TyQ, ValueType TyD,
396 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4, (outs QPR:$dst),
397 (ins DPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
398 [(set QPR:$dst, (TyQ (IntOp (TyD DPR:$src))))]>;
400 // 2-register shuffles (VTRN/VZIP/VUZP), both double- and quad-register.
401 class N2VDShuffle<bits<2> op19_18, bits<5> op11_7, string OpcodeStr>
402 : N2V<0b11, 0b11, op19_18, 0b10, op11_7, 0, 0, (outs DPR:$dst1, DPR:$dst2),
403 (ins DPR:$src1, DPR:$src2), NoItinerary,
404 !strconcat(OpcodeStr, "\t$dst1, $dst2"),
405 "$src1 = $dst1, $src2 = $dst2", []>;
406 class N2VQShuffle<bits<2> op19_18, bits<5> op11_7, string OpcodeStr>
407 : N2V<0b11, 0b11, op19_18, 0b10, op11_7, 1, 0, (outs QPR:$dst1, QPR:$dst2),
408 (ins QPR:$src1, QPR:$src2), NoItinerary,
409 !strconcat(OpcodeStr, "\t$dst1, $dst2"),
410 "$src1 = $dst1, $src2 = $dst2", []>;
412 // Basic 3-register operations, both double- and quad-register.
413 class N3VD<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
414 string OpcodeStr, ValueType ResTy, ValueType OpTy,
415 SDNode OpNode, bit Commutable>
416 : N3V<op24, op23, op21_20, op11_8, 0, op4,
417 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2), NoItinerary,
418 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
419 [(set DPR:$dst, (ResTy (OpNode (OpTy DPR:$src1), (OpTy DPR:$src2))))]> {
420 let isCommutable = Commutable;
422 class N3VQ<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
423 string OpcodeStr, ValueType ResTy, ValueType OpTy,
424 SDNode OpNode, bit Commutable>
425 : N3V<op24, op23, op21_20, op11_8, 1, op4,
426 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2), NoItinerary,
427 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
428 [(set QPR:$dst, (ResTy (OpNode (OpTy QPR:$src1), (OpTy QPR:$src2))))]> {
429 let isCommutable = Commutable;
432 // Basic 3-register operations, scalar single-precision
433 class N3VDs<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
434 string OpcodeStr, ValueType ResTy, ValueType OpTy,
435 SDNode OpNode, bit Commutable>
436 : N3V<op24, op23, op21_20, op11_8, 0, op4,
437 (outs DPR_VFP2:$dst), (ins DPR_VFP2:$src1, DPR_VFP2:$src2), NoItinerary,
438 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "", []> {
439 let isCommutable = Commutable;
441 class N3VDsPat<SDNode OpNode, NeonI Inst>
442 : NEONFPPat<(f32 (OpNode SPR:$a, SPR:$b)),
444 (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$a, arm_ssubreg_0),
445 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$b, arm_ssubreg_0)),
448 // Basic 3-register intrinsics, both double- and quad-register.
449 class N3VDInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
450 string OpcodeStr, ValueType ResTy, ValueType OpTy,
451 Intrinsic IntOp, bit Commutable>
452 : N3V<op24, op23, op21_20, op11_8, 0, op4,
453 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2), NoItinerary,
454 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
455 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src1), (OpTy DPR:$src2))))]> {
456 let isCommutable = Commutable;
458 class N3VQInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
459 string OpcodeStr, ValueType ResTy, ValueType OpTy,
460 Intrinsic IntOp, bit Commutable>
461 : N3V<op24, op23, op21_20, op11_8, 1, op4,
462 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2), NoItinerary,
463 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
464 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src1), (OpTy QPR:$src2))))]> {
465 let isCommutable = Commutable;
468 // Multiply-Add/Sub operations, both double- and quad-register.
469 class N3VDMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
470 string OpcodeStr, ValueType Ty, SDNode MulOp, SDNode OpNode>
471 : N3V<op24, op23, op21_20, op11_8, 0, op4,
472 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, DPR:$src3), NoItinerary,
473 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
474 [(set DPR:$dst, (Ty (OpNode DPR:$src1,
475 (Ty (MulOp DPR:$src2, DPR:$src3)))))]>;
476 class N3VQMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
477 string OpcodeStr, ValueType Ty, SDNode MulOp, SDNode OpNode>
478 : N3V<op24, op23, op21_20, op11_8, 1, op4,
479 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, QPR:$src3), NoItinerary,
480 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
481 [(set QPR:$dst, (Ty (OpNode QPR:$src1,
482 (Ty (MulOp QPR:$src2, QPR:$src3)))))]>;
484 // Multiply-Add/Sub operations, scalar single-precision
485 class N3VDMulOps<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
486 string OpcodeStr, ValueType Ty, SDNode MulOp, SDNode OpNode>
487 : N3V<op24, op23, op21_20, op11_8, 0, op4,
488 (outs DPR_VFP2:$dst),
489 (ins DPR_VFP2:$src1, DPR_VFP2:$src2, DPR_VFP2:$src3), NoItinerary,
490 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst", []>;
492 class N3VDMulOpsPat<SDNode MulNode, SDNode OpNode, NeonI Inst>
493 : NEONFPPat<(f32 (OpNode SPR:$acc, (f32 (MulNode SPR:$a, SPR:$b)))),
495 (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$acc, arm_ssubreg_0),
496 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$a, arm_ssubreg_0),
497 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$b, arm_ssubreg_0)),
500 // Neon 3-argument intrinsics, both double- and quad-register.
501 // The destination register is also used as the first source operand register.
502 class N3VDInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
503 string OpcodeStr, ValueType ResTy, ValueType OpTy,
505 : N3V<op24, op23, op21_20, op11_8, 0, op4,
506 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, DPR:$src3), NoItinerary,
507 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
508 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src1),
509 (OpTy DPR:$src2), (OpTy DPR:$src3))))]>;
510 class N3VQInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
511 string OpcodeStr, ValueType ResTy, ValueType OpTy,
513 : N3V<op24, op23, op21_20, op11_8, 1, op4,
514 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, QPR:$src3), NoItinerary,
515 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
516 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src1),
517 (OpTy QPR:$src2), (OpTy QPR:$src3))))]>;
519 // Neon Long 3-argument intrinsic. The destination register is
520 // a quad-register and is also used as the first source operand register.
521 class N3VLInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
522 string OpcodeStr, ValueType TyQ, ValueType TyD, Intrinsic IntOp>
523 : N3V<op24, op23, op21_20, op11_8, 0, op4,
524 (outs QPR:$dst), (ins QPR:$src1, DPR:$src2, DPR:$src3), NoItinerary,
525 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
527 (TyQ (IntOp (TyQ QPR:$src1), (TyD DPR:$src2), (TyD DPR:$src3))))]>;
529 // Narrowing 3-register intrinsics.
530 class N3VNInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
531 string OpcodeStr, ValueType TyD, ValueType TyQ,
532 Intrinsic IntOp, bit Commutable>
533 : N3V<op24, op23, op21_20, op11_8, 0, op4,
534 (outs DPR:$dst), (ins QPR:$src1, QPR:$src2), NoItinerary,
535 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
536 [(set DPR:$dst, (TyD (IntOp (TyQ QPR:$src1), (TyQ QPR:$src2))))]> {
537 let isCommutable = Commutable;
540 // Long 3-register intrinsics.
541 class N3VLInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
542 string OpcodeStr, ValueType TyQ, ValueType TyD,
543 Intrinsic IntOp, bit Commutable>
544 : N3V<op24, op23, op21_20, op11_8, 0, op4,
545 (outs QPR:$dst), (ins DPR:$src1, DPR:$src2), NoItinerary,
546 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
547 [(set QPR:$dst, (TyQ (IntOp (TyD DPR:$src1), (TyD DPR:$src2))))]> {
548 let isCommutable = Commutable;
551 // Wide 3-register intrinsics.
552 class N3VWInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
553 string OpcodeStr, ValueType TyQ, ValueType TyD,
554 Intrinsic IntOp, bit Commutable>
555 : N3V<op24, op23, op21_20, op11_8, 0, op4,
556 (outs QPR:$dst), (ins QPR:$src1, DPR:$src2), NoItinerary,
557 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
558 [(set QPR:$dst, (TyQ (IntOp (TyQ QPR:$src1), (TyD DPR:$src2))))]> {
559 let isCommutable = Commutable;
562 // Pairwise long 2-register intrinsics, both double- and quad-register.
563 class N2VDPLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
564 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
565 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
566 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
567 (ins DPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
568 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src))))]>;
569 class N2VQPLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
570 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
571 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
572 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
573 (ins QPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
574 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src))))]>;
576 // Pairwise long 2-register accumulate intrinsics,
577 // both double- and quad-register.
578 // The destination register is also used as the first source operand register.
579 class N2VDPLInt2<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
580 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
581 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
582 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4,
583 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2), NoItinerary,
584 !strconcat(OpcodeStr, "\t$dst, $src2"), "$src1 = $dst",
585 [(set DPR:$dst, (ResTy (IntOp (ResTy DPR:$src1), (OpTy DPR:$src2))))]>;
586 class N2VQPLInt2<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
587 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
588 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
589 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4,
590 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2), NoItinerary,
591 !strconcat(OpcodeStr, "\t$dst, $src2"), "$src1 = $dst",
592 [(set QPR:$dst, (ResTy (IntOp (ResTy QPR:$src1), (OpTy QPR:$src2))))]>;
594 // Shift by immediate,
595 // both double- and quad-register.
596 class N2VDSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
597 bit op4, string OpcodeStr, ValueType Ty, SDNode OpNode>
598 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
599 (outs DPR:$dst), (ins DPR:$src, i32imm:$SIMM), NoItinerary,
600 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
601 [(set DPR:$dst, (Ty (OpNode (Ty DPR:$src), (i32 imm:$SIMM))))]>;
602 class N2VQSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
603 bit op4, string OpcodeStr, ValueType Ty, SDNode OpNode>
604 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
605 (outs QPR:$dst), (ins QPR:$src, i32imm:$SIMM), NoItinerary,
606 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
607 [(set QPR:$dst, (Ty (OpNode (Ty QPR:$src), (i32 imm:$SIMM))))]>;
609 // Long shift by immediate.
610 class N2VLSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
611 bit op6, bit op4, string OpcodeStr, ValueType ResTy,
612 ValueType OpTy, SDNode OpNode>
613 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4,
614 (outs QPR:$dst), (ins DPR:$src, i32imm:$SIMM), NoItinerary,
615 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
616 [(set QPR:$dst, (ResTy (OpNode (OpTy DPR:$src),
617 (i32 imm:$SIMM))))]>;
619 // Narrow shift by immediate.
620 class N2VNSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
621 bit op6, bit op4, string OpcodeStr, ValueType ResTy,
622 ValueType OpTy, SDNode OpNode>
623 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4,
624 (outs DPR:$dst), (ins QPR:$src, i32imm:$SIMM), NoItinerary,
625 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
626 [(set DPR:$dst, (ResTy (OpNode (OpTy QPR:$src),
627 (i32 imm:$SIMM))))]>;
629 // Shift right by immediate and accumulate,
630 // both double- and quad-register.
631 class N2VDShAdd<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
632 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
633 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
634 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, i32imm:$SIMM),
636 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
637 [(set DPR:$dst, (Ty (add DPR:$src1,
638 (Ty (ShOp DPR:$src2, (i32 imm:$SIMM))))))]>;
639 class N2VQShAdd<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
640 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
641 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
642 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, i32imm:$SIMM),
644 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
645 [(set QPR:$dst, (Ty (add QPR:$src1,
646 (Ty (ShOp QPR:$src2, (i32 imm:$SIMM))))))]>;
648 // Shift by immediate and insert,
649 // both double- and quad-register.
650 class N2VDShIns<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
651 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
652 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
653 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, i32imm:$SIMM),
655 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
656 [(set DPR:$dst, (Ty (ShOp DPR:$src1, DPR:$src2, (i32 imm:$SIMM))))]>;
657 class N2VQShIns<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
658 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
659 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
660 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, i32imm:$SIMM),
662 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
663 [(set QPR:$dst, (Ty (ShOp QPR:$src1, QPR:$src2, (i32 imm:$SIMM))))]>;
665 // Convert, with fractional bits immediate,
666 // both double- and quad-register.
667 class N2VCvtD<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
668 bit op4, string OpcodeStr, ValueType ResTy, ValueType OpTy,
670 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
671 (outs DPR:$dst), (ins DPR:$src, i32imm:$SIMM), NoItinerary,
672 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
673 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src), (i32 imm:$SIMM))))]>;
674 class N2VCvtQ<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
675 bit op4, string OpcodeStr, ValueType ResTy, ValueType OpTy,
677 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
678 (outs QPR:$dst), (ins QPR:$src, i32imm:$SIMM), NoItinerary,
679 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
680 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src), (i32 imm:$SIMM))))]>;
682 //===----------------------------------------------------------------------===//
684 //===----------------------------------------------------------------------===//
686 // Neon 3-register vector operations.
688 // First with only element sizes of 8, 16 and 32 bits:
689 multiclass N3V_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
690 string OpcodeStr, SDNode OpNode, bit Commutable = 0> {
691 // 64-bit vector types.
692 def v8i8 : N3VD<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
693 v8i8, v8i8, OpNode, Commutable>;
694 def v4i16 : N3VD<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr, "16"),
695 v4i16, v4i16, OpNode, Commutable>;
696 def v2i32 : N3VD<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr, "32"),
697 v2i32, v2i32, OpNode, Commutable>;
699 // 128-bit vector types.
700 def v16i8 : N3VQ<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
701 v16i8, v16i8, OpNode, Commutable>;
702 def v8i16 : N3VQ<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr, "16"),
703 v8i16, v8i16, OpNode, Commutable>;
704 def v4i32 : N3VQ<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr, "32"),
705 v4i32, v4i32, OpNode, Commutable>;
708 // ....then also with element size 64 bits:
709 multiclass N3V_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
710 string OpcodeStr, SDNode OpNode, bit Commutable = 0>
711 : N3V_QHS<op24, op23, op11_8, op4, OpcodeStr, OpNode, Commutable> {
712 def v1i64 : N3VD<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr, "64"),
713 v1i64, v1i64, OpNode, Commutable>;
714 def v2i64 : N3VQ<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr, "64"),
715 v2i64, v2i64, OpNode, Commutable>;
719 // Neon Narrowing 2-register vector intrinsics,
720 // source operand element sizes of 16, 32 and 64 bits:
721 multiclass N2VNInt_HSD<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
722 bits<5> op11_7, bit op6, bit op4, string OpcodeStr,
724 def v8i8 : N2VNInt<op24_23, op21_20, 0b00, op17_16, op11_7, op6, op4,
725 !strconcat(OpcodeStr, "16"), v8i8, v8i16, IntOp>;
726 def v4i16 : N2VNInt<op24_23, op21_20, 0b01, op17_16, op11_7, op6, op4,
727 !strconcat(OpcodeStr, "32"), v4i16, v4i32, IntOp>;
728 def v2i32 : N2VNInt<op24_23, op21_20, 0b10, op17_16, op11_7, op6, op4,
729 !strconcat(OpcodeStr, "64"), v2i32, v2i64, IntOp>;
733 // Neon Lengthening 2-register vector intrinsic (currently specific to VMOVL).
734 // source operand element sizes of 16, 32 and 64 bits:
735 multiclass N2VLInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op7, bit op6,
736 bit op4, string OpcodeStr, Intrinsic IntOp> {
737 def v8i16 : N2VLInt<op24, op23, 0b001000, op11_8, op7, op6, op4,
738 !strconcat(OpcodeStr, "8"), v8i16, v8i8, IntOp>;
739 def v4i32 : N2VLInt<op24, op23, 0b010000, op11_8, op7, op6, op4,
740 !strconcat(OpcodeStr, "16"), v4i32, v4i16, IntOp>;
741 def v2i64 : N2VLInt<op24, op23, 0b100000, op11_8, op7, op6, op4,
742 !strconcat(OpcodeStr, "32"), v2i64, v2i32, IntOp>;
746 // Neon 3-register vector intrinsics.
748 // First with only element sizes of 16 and 32 bits:
749 multiclass N3VInt_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
750 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
751 // 64-bit vector types.
752 def v4i16 : N3VDInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
753 v4i16, v4i16, IntOp, Commutable>;
754 def v2i32 : N3VDInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
755 v2i32, v2i32, IntOp, Commutable>;
757 // 128-bit vector types.
758 def v8i16 : N3VQInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
759 v8i16, v8i16, IntOp, Commutable>;
760 def v4i32 : N3VQInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
761 v4i32, v4i32, IntOp, Commutable>;
764 // ....then also with element size of 8 bits:
765 multiclass N3VInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
766 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
767 : N3VInt_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
768 def v8i8 : N3VDInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
769 v8i8, v8i8, IntOp, Commutable>;
770 def v16i8 : N3VQInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
771 v16i8, v16i8, IntOp, Commutable>;
774 // ....then also with element size of 64 bits:
775 multiclass N3VInt_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
776 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
777 : N3VInt_QHS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
778 def v1i64 : N3VDInt<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr,"64"),
779 v1i64, v1i64, IntOp, Commutable>;
780 def v2i64 : N3VQInt<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr,"64"),
781 v2i64, v2i64, IntOp, Commutable>;
785 // Neon Narrowing 3-register vector intrinsics,
786 // source operand element sizes of 16, 32 and 64 bits:
787 multiclass N3VNInt_HSD<bit op24, bit op23, bits<4> op11_8, bit op4,
788 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
789 def v8i8 : N3VNInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr,"16"),
790 v8i8, v8i16, IntOp, Commutable>;
791 def v4i16 : N3VNInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"32"),
792 v4i16, v4i32, IntOp, Commutable>;
793 def v2i32 : N3VNInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"64"),
794 v2i32, v2i64, IntOp, Commutable>;
798 // Neon Long 3-register vector intrinsics.
800 // First with only element sizes of 16 and 32 bits:
801 multiclass N3VLInt_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
802 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
803 def v4i32 : N3VLInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
804 v4i32, v4i16, IntOp, Commutable>;
805 def v2i64 : N3VLInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
806 v2i64, v2i32, IntOp, Commutable>;
809 // ....then also with element size of 8 bits:
810 multiclass N3VLInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
811 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
812 : N3VLInt_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
813 def v8i16 : N3VLInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
814 v8i16, v8i8, IntOp, Commutable>;
818 // Neon Wide 3-register vector intrinsics,
819 // source operand element sizes of 8, 16 and 32 bits:
820 multiclass N3VWInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
821 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
822 def v8i16 : N3VWInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
823 v8i16, v8i8, IntOp, Commutable>;
824 def v4i32 : N3VWInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
825 v4i32, v4i16, IntOp, Commutable>;
826 def v2i64 : N3VWInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
827 v2i64, v2i32, IntOp, Commutable>;
831 // Neon Multiply-Op vector operations,
832 // element sizes of 8, 16 and 32 bits:
833 multiclass N3VMulOp_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
834 string OpcodeStr, SDNode OpNode> {
835 // 64-bit vector types.
836 def v8i8 : N3VDMulOp<op24, op23, 0b00, op11_8, op4,
837 !strconcat(OpcodeStr, "8"), v8i8, mul, OpNode>;
838 def v4i16 : N3VDMulOp<op24, op23, 0b01, op11_8, op4,
839 !strconcat(OpcodeStr, "16"), v4i16, mul, OpNode>;
840 def v2i32 : N3VDMulOp<op24, op23, 0b10, op11_8, op4,
841 !strconcat(OpcodeStr, "32"), v2i32, mul, OpNode>;
843 // 128-bit vector types.
844 def v16i8 : N3VQMulOp<op24, op23, 0b00, op11_8, op4,
845 !strconcat(OpcodeStr, "8"), v16i8, mul, OpNode>;
846 def v8i16 : N3VQMulOp<op24, op23, 0b01, op11_8, op4,
847 !strconcat(OpcodeStr, "16"), v8i16, mul, OpNode>;
848 def v4i32 : N3VQMulOp<op24, op23, 0b10, op11_8, op4,
849 !strconcat(OpcodeStr, "32"), v4i32, mul, OpNode>;
853 // Neon 3-argument intrinsics,
854 // element sizes of 8, 16 and 32 bits:
855 multiclass N3VInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
856 string OpcodeStr, Intrinsic IntOp> {
857 // 64-bit vector types.
858 def v8i8 : N3VDInt3<op24, op23, 0b00, op11_8, op4,
859 !strconcat(OpcodeStr, "8"), v8i8, v8i8, IntOp>;
860 def v4i16 : N3VDInt3<op24, op23, 0b01, op11_8, op4,
861 !strconcat(OpcodeStr, "16"), v4i16, v4i16, IntOp>;
862 def v2i32 : N3VDInt3<op24, op23, 0b10, op11_8, op4,
863 !strconcat(OpcodeStr, "32"), v2i32, v2i32, IntOp>;
865 // 128-bit vector types.
866 def v16i8 : N3VQInt3<op24, op23, 0b00, op11_8, op4,
867 !strconcat(OpcodeStr, "8"), v16i8, v16i8, IntOp>;
868 def v8i16 : N3VQInt3<op24, op23, 0b01, op11_8, op4,
869 !strconcat(OpcodeStr, "16"), v8i16, v8i16, IntOp>;
870 def v4i32 : N3VQInt3<op24, op23, 0b10, op11_8, op4,
871 !strconcat(OpcodeStr, "32"), v4i32, v4i32, IntOp>;
875 // Neon Long 3-argument intrinsics.
877 // First with only element sizes of 16 and 32 bits:
878 multiclass N3VLInt3_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
879 string OpcodeStr, Intrinsic IntOp> {
880 def v4i32 : N3VLInt3<op24, op23, 0b01, op11_8, op4,
881 !strconcat(OpcodeStr, "16"), v4i32, v4i16, IntOp>;
882 def v2i64 : N3VLInt3<op24, op23, 0b10, op11_8, op4,
883 !strconcat(OpcodeStr, "32"), v2i64, v2i32, IntOp>;
886 // ....then also with element size of 8 bits:
887 multiclass N3VLInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
888 string OpcodeStr, Intrinsic IntOp>
889 : N3VLInt3_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp> {
890 def v8i16 : N3VLInt3<op24, op23, 0b01, op11_8, op4,
891 !strconcat(OpcodeStr, "8"), v8i16, v8i8, IntOp>;
895 // Neon 2-register vector intrinsics,
896 // element sizes of 8, 16 and 32 bits:
897 multiclass N2VInt_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
898 bits<5> op11_7, bit op4, string OpcodeStr,
900 // 64-bit vector types.
901 def v8i8 : N2VDInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
902 !strconcat(OpcodeStr, "8"), v8i8, v8i8, IntOp>;
903 def v4i16 : N2VDInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
904 !strconcat(OpcodeStr, "16"), v4i16, v4i16, IntOp>;
905 def v2i32 : N2VDInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
906 !strconcat(OpcodeStr, "32"), v2i32, v2i32, IntOp>;
908 // 128-bit vector types.
909 def v16i8 : N2VQInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
910 !strconcat(OpcodeStr, "8"), v16i8, v16i8, IntOp>;
911 def v8i16 : N2VQInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
912 !strconcat(OpcodeStr, "16"), v8i16, v8i16, IntOp>;
913 def v4i32 : N2VQInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
914 !strconcat(OpcodeStr, "32"), v4i32, v4i32, IntOp>;
918 // Neon Pairwise long 2-register intrinsics,
919 // element sizes of 8, 16 and 32 bits:
920 multiclass N2VPLInt_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
921 bits<5> op11_7, bit op4,
922 string OpcodeStr, Intrinsic IntOp> {
923 // 64-bit vector types.
924 def v8i8 : N2VDPLInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
925 !strconcat(OpcodeStr, "8"), v4i16, v8i8, IntOp>;
926 def v4i16 : N2VDPLInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
927 !strconcat(OpcodeStr, "16"), v2i32, v4i16, IntOp>;
928 def v2i32 : N2VDPLInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
929 !strconcat(OpcodeStr, "32"), v1i64, v2i32, IntOp>;
931 // 128-bit vector types.
932 def v16i8 : N2VQPLInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
933 !strconcat(OpcodeStr, "8"), v8i16, v16i8, IntOp>;
934 def v8i16 : N2VQPLInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
935 !strconcat(OpcodeStr, "16"), v4i32, v8i16, IntOp>;
936 def v4i32 : N2VQPLInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
937 !strconcat(OpcodeStr, "32"), v2i64, v4i32, IntOp>;
941 // Neon Pairwise long 2-register accumulate intrinsics,
942 // element sizes of 8, 16 and 32 bits:
943 multiclass N2VPLInt2_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
944 bits<5> op11_7, bit op4,
945 string OpcodeStr, Intrinsic IntOp> {
946 // 64-bit vector types.
947 def v8i8 : N2VDPLInt2<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
948 !strconcat(OpcodeStr, "8"), v4i16, v8i8, IntOp>;
949 def v4i16 : N2VDPLInt2<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
950 !strconcat(OpcodeStr, "16"), v2i32, v4i16, IntOp>;
951 def v2i32 : N2VDPLInt2<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
952 !strconcat(OpcodeStr, "32"), v1i64, v2i32, IntOp>;
954 // 128-bit vector types.
955 def v16i8 : N2VQPLInt2<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
956 !strconcat(OpcodeStr, "8"), v8i16, v16i8, IntOp>;
957 def v8i16 : N2VQPLInt2<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
958 !strconcat(OpcodeStr, "16"), v4i32, v8i16, IntOp>;
959 def v4i32 : N2VQPLInt2<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
960 !strconcat(OpcodeStr, "32"), v2i64, v4i32, IntOp>;
964 // Neon 2-register vector shift by immediate,
965 // element sizes of 8, 16, 32 and 64 bits:
966 multiclass N2VSh_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
967 string OpcodeStr, SDNode OpNode> {
968 // 64-bit vector types.
969 def v8i8 : N2VDSh<op24, op23, 0b001000, op11_8, 0, op4,
970 !strconcat(OpcodeStr, "8"), v8i8, OpNode>;
971 def v4i16 : N2VDSh<op24, op23, 0b010000, op11_8, 0, op4,
972 !strconcat(OpcodeStr, "16"), v4i16, OpNode>;
973 def v2i32 : N2VDSh<op24, op23, 0b100000, op11_8, 0, op4,
974 !strconcat(OpcodeStr, "32"), v2i32, OpNode>;
975 def v1i64 : N2VDSh<op24, op23, 0b000000, op11_8, 1, op4,
976 !strconcat(OpcodeStr, "64"), v1i64, OpNode>;
978 // 128-bit vector types.
979 def v16i8 : N2VQSh<op24, op23, 0b001000, op11_8, 0, op4,
980 !strconcat(OpcodeStr, "8"), v16i8, OpNode>;
981 def v8i16 : N2VQSh<op24, op23, 0b010000, op11_8, 0, op4,
982 !strconcat(OpcodeStr, "16"), v8i16, OpNode>;
983 def v4i32 : N2VQSh<op24, op23, 0b100000, op11_8, 0, op4,
984 !strconcat(OpcodeStr, "32"), v4i32, OpNode>;
985 def v2i64 : N2VQSh<op24, op23, 0b000000, op11_8, 1, op4,
986 !strconcat(OpcodeStr, "64"), v2i64, OpNode>;
990 // Neon Shift-Accumulate vector operations,
991 // element sizes of 8, 16, 32 and 64 bits:
992 multiclass N2VShAdd_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
993 string OpcodeStr, SDNode ShOp> {
994 // 64-bit vector types.
995 def v8i8 : N2VDShAdd<op24, op23, 0b001000, op11_8, 0, op4,
996 !strconcat(OpcodeStr, "8"), v8i8, ShOp>;
997 def v4i16 : N2VDShAdd<op24, op23, 0b010000, op11_8, 0, op4,
998 !strconcat(OpcodeStr, "16"), v4i16, ShOp>;
999 def v2i32 : N2VDShAdd<op24, op23, 0b100000, op11_8, 0, op4,
1000 !strconcat(OpcodeStr, "32"), v2i32, ShOp>;
1001 def v1i64 : N2VDShAdd<op24, op23, 0b000000, op11_8, 1, op4,
1002 !strconcat(OpcodeStr, "64"), v1i64, ShOp>;
1004 // 128-bit vector types.
1005 def v16i8 : N2VQShAdd<op24, op23, 0b001000, op11_8, 0, op4,
1006 !strconcat(OpcodeStr, "8"), v16i8, ShOp>;
1007 def v8i16 : N2VQShAdd<op24, op23, 0b010000, op11_8, 0, op4,
1008 !strconcat(OpcodeStr, "16"), v8i16, ShOp>;
1009 def v4i32 : N2VQShAdd<op24, op23, 0b100000, op11_8, 0, op4,
1010 !strconcat(OpcodeStr, "32"), v4i32, ShOp>;
1011 def v2i64 : N2VQShAdd<op24, op23, 0b000000, op11_8, 1, op4,
1012 !strconcat(OpcodeStr, "64"), v2i64, ShOp>;
1016 // Neon Shift-Insert vector operations,
1017 // element sizes of 8, 16, 32 and 64 bits:
1018 multiclass N2VShIns_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
1019 string OpcodeStr, SDNode ShOp> {
1020 // 64-bit vector types.
1021 def v8i8 : N2VDShIns<op24, op23, 0b001000, op11_8, 0, op4,
1022 !strconcat(OpcodeStr, "8"), v8i8, ShOp>;
1023 def v4i16 : N2VDShIns<op24, op23, 0b010000, op11_8, 0, op4,
1024 !strconcat(OpcodeStr, "16"), v4i16, ShOp>;
1025 def v2i32 : N2VDShIns<op24, op23, 0b100000, op11_8, 0, op4,
1026 !strconcat(OpcodeStr, "32"), v2i32, ShOp>;
1027 def v1i64 : N2VDShIns<op24, op23, 0b000000, op11_8, 1, op4,
1028 !strconcat(OpcodeStr, "64"), v1i64, ShOp>;
1030 // 128-bit vector types.
1031 def v16i8 : N2VQShIns<op24, op23, 0b001000, op11_8, 0, op4,
1032 !strconcat(OpcodeStr, "8"), v16i8, ShOp>;
1033 def v8i16 : N2VQShIns<op24, op23, 0b010000, op11_8, 0, op4,
1034 !strconcat(OpcodeStr, "16"), v8i16, ShOp>;
1035 def v4i32 : N2VQShIns<op24, op23, 0b100000, op11_8, 0, op4,
1036 !strconcat(OpcodeStr, "32"), v4i32, ShOp>;
1037 def v2i64 : N2VQShIns<op24, op23, 0b000000, op11_8, 1, op4,
1038 !strconcat(OpcodeStr, "64"), v2i64, ShOp>;
1041 //===----------------------------------------------------------------------===//
1042 // Instruction Definitions.
1043 //===----------------------------------------------------------------------===//
1045 // Vector Add Operations.
1047 // VADD : Vector Add (integer and floating-point)
1048 defm VADD : N3V_QHSD<0, 0, 0b1000, 0, "vadd.i", add, 1>;
1049 def VADDfd : N3VD<0, 0, 0b00, 0b1101, 0, "vadd.f32", v2f32, v2f32, fadd, 1>;
1050 def VADDfq : N3VQ<0, 0, 0b00, 0b1101, 0, "vadd.f32", v4f32, v4f32, fadd, 1>;
1051 // VADDL : Vector Add Long (Q = D + D)
1052 defm VADDLs : N3VLInt_QHS<0,1,0b0000,0, "vaddl.s", int_arm_neon_vaddls, 1>;
1053 defm VADDLu : N3VLInt_QHS<1,1,0b0000,0, "vaddl.u", int_arm_neon_vaddlu, 1>;
1054 // VADDW : Vector Add Wide (Q = Q + D)
1055 defm VADDWs : N3VWInt_QHS<0,1,0b0001,0, "vaddw.s", int_arm_neon_vaddws, 0>;
1056 defm VADDWu : N3VWInt_QHS<1,1,0b0001,0, "vaddw.u", int_arm_neon_vaddwu, 0>;
1057 // VHADD : Vector Halving Add
1058 defm VHADDs : N3VInt_QHS<0,0,0b0000,0, "vhadd.s", int_arm_neon_vhadds, 1>;
1059 defm VHADDu : N3VInt_QHS<1,0,0b0000,0, "vhadd.u", int_arm_neon_vhaddu, 1>;
1060 // VRHADD : Vector Rounding Halving Add
1061 defm VRHADDs : N3VInt_QHS<0,0,0b0001,0, "vrhadd.s", int_arm_neon_vrhadds, 1>;
1062 defm VRHADDu : N3VInt_QHS<1,0,0b0001,0, "vrhadd.u", int_arm_neon_vrhaddu, 1>;
1063 // VQADD : Vector Saturating Add
1064 defm VQADDs : N3VInt_QHSD<0,0,0b0000,1, "vqadd.s", int_arm_neon_vqadds, 1>;
1065 defm VQADDu : N3VInt_QHSD<1,0,0b0000,1, "vqadd.u", int_arm_neon_vqaddu, 1>;
1066 // VADDHN : Vector Add and Narrow Returning High Half (D = Q + Q)
1067 defm VADDHN : N3VNInt_HSD<0,1,0b0100,0, "vaddhn.i", int_arm_neon_vaddhn, 1>;
1068 // VRADDHN : Vector Rounding Add and Narrow Returning High Half (D = Q + Q)
1069 defm VRADDHN : N3VNInt_HSD<1,1,0b0100,0, "vraddhn.i", int_arm_neon_vraddhn, 1>;
1071 // Vector Multiply Operations.
1073 // VMUL : Vector Multiply (integer, polynomial and floating-point)
1074 defm VMUL : N3V_QHS<0, 0, 0b1001, 1, "vmul.i", mul, 1>;
1075 def VMULpd : N3VDInt<1, 0, 0b00, 0b1001, 1, "vmul.p8", v8i8, v8i8,
1076 int_arm_neon_vmulp, 1>;
1077 def VMULpq : N3VQInt<1, 0, 0b00, 0b1001, 1, "vmul.p8", v16i8, v16i8,
1078 int_arm_neon_vmulp, 1>;
1079 def VMULfd : N3VD<1, 0, 0b00, 0b1101, 1, "vmul.f32", v2f32, v2f32, fmul, 1>;
1080 def VMULfq : N3VQ<1, 0, 0b00, 0b1101, 1, "vmul.f32", v4f32, v4f32, fmul, 1>;
1081 // VQDMULH : Vector Saturating Doubling Multiply Returning High Half
1082 defm VQDMULH : N3VInt_HS<0,0,0b1011,0, "vqdmulh.s", int_arm_neon_vqdmulh, 1>;
1083 // VQRDMULH : Vector Rounding Saturating Doubling Multiply Returning High Half
1084 defm VQRDMULH : N3VInt_HS<1,0,0b1011,0, "vqrdmulh.s", int_arm_neon_vqrdmulh, 1>;
1085 // VMULL : Vector Multiply Long (integer and polynomial) (Q = D * D)
1086 defm VMULLs : N3VLInt_QHS<0,1,0b1100,0, "vmull.s", int_arm_neon_vmulls, 1>;
1087 defm VMULLu : N3VLInt_QHS<1,1,0b1100,0, "vmull.u", int_arm_neon_vmullu, 1>;
1088 def VMULLp : N3VLInt<0, 1, 0b00, 0b1110, 0, "vmull.p8", v8i16, v8i8,
1089 int_arm_neon_vmullp, 1>;
1090 // VQDMULL : Vector Saturating Doubling Multiply Long (Q = D * D)
1091 defm VQDMULL : N3VLInt_HS<0,1,0b1101,0, "vqdmull.s", int_arm_neon_vqdmull, 1>;
1093 // Vector Multiply-Accumulate and Multiply-Subtract Operations.
1095 // VMLA : Vector Multiply Accumulate (integer and floating-point)
1096 defm VMLA : N3VMulOp_QHS<0, 0, 0b1001, 0, "vmla.i", add>;
1097 def VMLAfd : N3VDMulOp<0, 0, 0b00, 0b1101, 1, "vmla.f32", v2f32, fmul, fadd>;
1098 def VMLAfq : N3VQMulOp<0, 0, 0b00, 0b1101, 1, "vmla.f32", v4f32, fmul, fadd>;
1099 // VMLAL : Vector Multiply Accumulate Long (Q += D * D)
1100 defm VMLALs : N3VLInt3_QHS<0,1,0b1000,0, "vmlal.s", int_arm_neon_vmlals>;
1101 defm VMLALu : N3VLInt3_QHS<1,1,0b1000,0, "vmlal.u", int_arm_neon_vmlalu>;
1102 // VQDMLAL : Vector Saturating Doubling Multiply Accumulate Long (Q += D * D)
1103 defm VQDMLAL : N3VLInt3_HS<0, 1, 0b1001, 0, "vqdmlal.s", int_arm_neon_vqdmlal>;
1104 // VMLS : Vector Multiply Subtract (integer and floating-point)
1105 defm VMLS : N3VMulOp_QHS<0, 0, 0b1001, 0, "vmls.i", sub>;
1106 def VMLSfd : N3VDMulOp<0, 0, 0b10, 0b1101, 1, "vmls.f32", v2f32, fmul, fsub>;
1107 def VMLSfq : N3VQMulOp<0, 0, 0b10, 0b1101, 1, "vmls.f32", v4f32, fmul, fsub>;
1108 // VMLSL : Vector Multiply Subtract Long (Q -= D * D)
1109 defm VMLSLs : N3VLInt3_QHS<0,1,0b1010,0, "vmlsl.s", int_arm_neon_vmlsls>;
1110 defm VMLSLu : N3VLInt3_QHS<1,1,0b1010,0, "vmlsl.u", int_arm_neon_vmlslu>;
1111 // VQDMLSL : Vector Saturating Doubling Multiply Subtract Long (Q -= D * D)
1112 defm VQDMLSL : N3VLInt3_HS<0, 1, 0b1011, 0, "vqdmlsl.s", int_arm_neon_vqdmlsl>;
1114 // Vector Subtract Operations.
1116 // VSUB : Vector Subtract (integer and floating-point)
1117 defm VSUB : N3V_QHSD<1, 0, 0b1000, 0, "vsub.i", sub, 0>;
1118 def VSUBfd : N3VD<0, 0, 0b10, 0b1101, 0, "vsub.f32", v2f32, v2f32, fsub, 0>;
1119 def VSUBfq : N3VQ<0, 0, 0b10, 0b1101, 0, "vsub.f32", v4f32, v4f32, fsub, 0>;
1120 // VSUBL : Vector Subtract Long (Q = D - D)
1121 defm VSUBLs : N3VLInt_QHS<0,1,0b0010,0, "vsubl.s", int_arm_neon_vsubls, 1>;
1122 defm VSUBLu : N3VLInt_QHS<1,1,0b0010,0, "vsubl.u", int_arm_neon_vsublu, 1>;
1123 // VSUBW : Vector Subtract Wide (Q = Q - D)
1124 defm VSUBWs : N3VWInt_QHS<0,1,0b0011,0, "vsubw.s", int_arm_neon_vsubws, 0>;
1125 defm VSUBWu : N3VWInt_QHS<1,1,0b0011,0, "vsubw.u", int_arm_neon_vsubwu, 0>;
1126 // VHSUB : Vector Halving Subtract
1127 defm VHSUBs : N3VInt_QHS<0, 0, 0b0010, 0, "vhsub.s", int_arm_neon_vhsubs, 0>;
1128 defm VHSUBu : N3VInt_QHS<1, 0, 0b0010, 0, "vhsub.u", int_arm_neon_vhsubu, 0>;
1129 // VQSUB : Vector Saturing Subtract
1130 defm VQSUBs : N3VInt_QHSD<0, 0, 0b0010, 1, "vqsub.s", int_arm_neon_vqsubs, 0>;
1131 defm VQSUBu : N3VInt_QHSD<1, 0, 0b0010, 1, "vqsub.u", int_arm_neon_vqsubu, 0>;
1132 // VSUBHN : Vector Subtract and Narrow Returning High Half (D = Q - Q)
1133 defm VSUBHN : N3VNInt_HSD<0,1,0b0110,0, "vsubhn.i", int_arm_neon_vsubhn, 0>;
1134 // VRSUBHN : Vector Rounding Subtract and Narrow Returning High Half (D=Q-Q)
1135 defm VRSUBHN : N3VNInt_HSD<1,1,0b0110,0, "vrsubhn.i", int_arm_neon_vrsubhn, 0>;
1137 // Vector Comparisons.
1139 // VCEQ : Vector Compare Equal
1140 defm VCEQ : N3V_QHS<1, 0, 0b1000, 1, "vceq.i", NEONvceq, 1>;
1141 def VCEQfd : N3VD<0,0,0b00,0b1110,0, "vceq.f32", v2i32, v2f32, NEONvceq, 1>;
1142 def VCEQfq : N3VQ<0,0,0b00,0b1110,0, "vceq.f32", v4i32, v4f32, NEONvceq, 1>;
1143 // VCGE : Vector Compare Greater Than or Equal
1144 defm VCGEs : N3V_QHS<0, 0, 0b0011, 1, "vcge.s", NEONvcge, 0>;
1145 defm VCGEu : N3V_QHS<1, 0, 0b0011, 1, "vcge.u", NEONvcgeu, 0>;
1146 def VCGEfd : N3VD<1,0,0b00,0b1110,0, "vcge.f32", v2i32, v2f32, NEONvcge, 0>;
1147 def VCGEfq : N3VQ<1,0,0b00,0b1110,0, "vcge.f32", v4i32, v4f32, NEONvcge, 0>;
1148 // VCGT : Vector Compare Greater Than
1149 defm VCGTs : N3V_QHS<0, 0, 0b0011, 0, "vcgt.s", NEONvcgt, 0>;
1150 defm VCGTu : N3V_QHS<1, 0, 0b0011, 0, "vcgt.u", NEONvcgtu, 0>;
1151 def VCGTfd : N3VD<1,0,0b10,0b1110,0, "vcgt.f32", v2i32, v2f32, NEONvcgt, 0>;
1152 def VCGTfq : N3VQ<1,0,0b10,0b1110,0, "vcgt.f32", v4i32, v4f32, NEONvcgt, 0>;
1153 // VACGE : Vector Absolute Compare Greater Than or Equal (aka VCAGE)
1154 def VACGEd : N3VDInt<1, 0, 0b00, 0b1110, 1, "vacge.f32", v2i32, v2f32,
1155 int_arm_neon_vacged, 0>;
1156 def VACGEq : N3VQInt<1, 0, 0b00, 0b1110, 1, "vacge.f32", v4i32, v4f32,
1157 int_arm_neon_vacgeq, 0>;
1158 // VACGT : Vector Absolute Compare Greater Than (aka VCAGT)
1159 def VACGTd : N3VDInt<1, 0, 0b10, 0b1110, 1, "vacgt.f32", v2i32, v2f32,
1160 int_arm_neon_vacgtd, 0>;
1161 def VACGTq : N3VQInt<1, 0, 0b10, 0b1110, 1, "vacgt.f32", v4i32, v4f32,
1162 int_arm_neon_vacgtq, 0>;
1163 // VTST : Vector Test Bits
1164 defm VTST : N3V_QHS<0, 0, 0b1000, 1, "vtst.i", NEONvtst, 1>;
1166 // Vector Bitwise Operations.
1168 // VAND : Vector Bitwise AND
1169 def VANDd : N3VD<0, 0, 0b00, 0b0001, 1, "vand", v2i32, v2i32, and, 1>;
1170 def VANDq : N3VQ<0, 0, 0b00, 0b0001, 1, "vand", v4i32, v4i32, and, 1>;
1172 // VEOR : Vector Bitwise Exclusive OR
1173 def VEORd : N3VD<1, 0, 0b00, 0b0001, 1, "veor", v2i32, v2i32, xor, 1>;
1174 def VEORq : N3VQ<1, 0, 0b00, 0b0001, 1, "veor", v4i32, v4i32, xor, 1>;
1176 // VORR : Vector Bitwise OR
1177 def VORRd : N3VD<0, 0, 0b10, 0b0001, 1, "vorr", v2i32, v2i32, or, 1>;
1178 def VORRq : N3VQ<0, 0, 0b10, 0b0001, 1, "vorr", v4i32, v4i32, or, 1>;
1180 // VBIC : Vector Bitwise Bit Clear (AND NOT)
1181 def VBICd : N3V<0, 0, 0b01, 0b0001, 0, 1, (outs DPR:$dst),
1182 (ins DPR:$src1, DPR:$src2), NoItinerary,
1183 "vbic\t$dst, $src1, $src2", "",
1184 [(set DPR:$dst, (v2i32 (and DPR:$src1,(vnot DPR:$src2))))]>;
1185 def VBICq : N3V<0, 0, 0b01, 0b0001, 1, 1, (outs QPR:$dst),
1186 (ins QPR:$src1, QPR:$src2), NoItinerary,
1187 "vbic\t$dst, $src1, $src2", "",
1188 [(set QPR:$dst, (v4i32 (and QPR:$src1,(vnot QPR:$src2))))]>;
1190 // VORN : Vector Bitwise OR NOT
1191 def VORNd : N3V<0, 0, 0b11, 0b0001, 0, 1, (outs DPR:$dst),
1192 (ins DPR:$src1, DPR:$src2), NoItinerary,
1193 "vorn\t$dst, $src1, $src2", "",
1194 [(set DPR:$dst, (v2i32 (or DPR:$src1, (vnot DPR:$src2))))]>;
1195 def VORNq : N3V<0, 0, 0b11, 0b0001, 1, 1, (outs QPR:$dst),
1196 (ins QPR:$src1, QPR:$src2), NoItinerary,
1197 "vorn\t$dst, $src1, $src2", "",
1198 [(set QPR:$dst, (v4i32 (or QPR:$src1, (vnot QPR:$src2))))]>;
1200 // VMVN : Vector Bitwise NOT
1201 def VMVNd : N2V<0b11, 0b11, 0b00, 0b00, 0b01011, 0, 0,
1202 (outs DPR:$dst), (ins DPR:$src), NoItinerary,
1203 "vmvn\t$dst, $src", "",
1204 [(set DPR:$dst, (v2i32 (vnot DPR:$src)))]>;
1205 def VMVNq : N2V<0b11, 0b11, 0b00, 0b00, 0b01011, 1, 0,
1206 (outs QPR:$dst), (ins QPR:$src), NoItinerary,
1207 "vmvn\t$dst, $src", "",
1208 [(set QPR:$dst, (v4i32 (vnot QPR:$src)))]>;
1209 def : Pat<(v2i32 (vnot_conv DPR:$src)), (VMVNd DPR:$src)>;
1210 def : Pat<(v4i32 (vnot_conv QPR:$src)), (VMVNq QPR:$src)>;
1212 // VBSL : Vector Bitwise Select
1213 def VBSLd : N3V<1, 0, 0b01, 0b0001, 0, 1, (outs DPR:$dst),
1214 (ins DPR:$src1, DPR:$src2, DPR:$src3), NoItinerary,
1215 "vbsl\t$dst, $src2, $src3", "$src1 = $dst",
1217 (v2i32 (or (and DPR:$src2, DPR:$src1),
1218 (and DPR:$src3, (vnot DPR:$src1)))))]>;
1219 def VBSLq : N3V<1, 0, 0b01, 0b0001, 1, 1, (outs QPR:$dst),
1220 (ins QPR:$src1, QPR:$src2, QPR:$src3), NoItinerary,
1221 "vbsl\t$dst, $src2, $src3", "$src1 = $dst",
1223 (v4i32 (or (and QPR:$src2, QPR:$src1),
1224 (and QPR:$src3, (vnot QPR:$src1)))))]>;
1226 // VBIF : Vector Bitwise Insert if False
1227 // like VBSL but with: "vbif\t$dst, $src3, $src1", "$src2 = $dst",
1228 // VBIT : Vector Bitwise Insert if True
1229 // like VBSL but with: "vbit\t$dst, $src2, $src1", "$src3 = $dst",
1230 // These are not yet implemented. The TwoAddress pass will not go looking
1231 // for equivalent operations with different register constraints; it just
1234 // Vector Absolute Differences.
1236 // VABD : Vector Absolute Difference
1237 defm VABDs : N3VInt_QHS<0, 0, 0b0111, 0, "vabd.s", int_arm_neon_vabds, 0>;
1238 defm VABDu : N3VInt_QHS<1, 0, 0b0111, 0, "vabd.u", int_arm_neon_vabdu, 0>;
1239 def VABDfd : N3VDInt<1, 0, 0b10, 0b1101, 0, "vabd.f32", v2f32, v2f32,
1240 int_arm_neon_vabds, 0>;
1241 def VABDfq : N3VQInt<1, 0, 0b10, 0b1101, 0, "vabd.f32", v4f32, v4f32,
1242 int_arm_neon_vabds, 0>;
1244 // VABDL : Vector Absolute Difference Long (Q = | D - D |)
1245 defm VABDLs : N3VLInt_QHS<0,1,0b0111,0, "vabdl.s", int_arm_neon_vabdls, 0>;
1246 defm VABDLu : N3VLInt_QHS<1,1,0b0111,0, "vabdl.u", int_arm_neon_vabdlu, 0>;
1248 // VABA : Vector Absolute Difference and Accumulate
1249 defm VABAs : N3VInt3_QHS<0,1,0b0101,0, "vaba.s", int_arm_neon_vabas>;
1250 defm VABAu : N3VInt3_QHS<1,1,0b0101,0, "vaba.u", int_arm_neon_vabau>;
1252 // VABAL : Vector Absolute Difference and Accumulate Long (Q += | D - D |)
1253 defm VABALs : N3VLInt3_QHS<0,1,0b0101,0, "vabal.s", int_arm_neon_vabals>;
1254 defm VABALu : N3VLInt3_QHS<1,1,0b0101,0, "vabal.u", int_arm_neon_vabalu>;
1256 // Vector Maximum and Minimum.
1258 // VMAX : Vector Maximum
1259 defm VMAXs : N3VInt_QHS<0, 0, 0b0110, 0, "vmax.s", int_arm_neon_vmaxs, 1>;
1260 defm VMAXu : N3VInt_QHS<1, 0, 0b0110, 0, "vmax.u", int_arm_neon_vmaxu, 1>;
1261 def VMAXfd : N3VDInt<0, 0, 0b00, 0b1111, 0, "vmax.f32", v2f32, v2f32,
1262 int_arm_neon_vmaxs, 1>;
1263 def VMAXfq : N3VQInt<0, 0, 0b00, 0b1111, 0, "vmax.f32", v4f32, v4f32,
1264 int_arm_neon_vmaxs, 1>;
1266 // VMIN : Vector Minimum
1267 defm VMINs : N3VInt_QHS<0, 0, 0b0110, 1, "vmin.s", int_arm_neon_vmins, 1>;
1268 defm VMINu : N3VInt_QHS<1, 0, 0b0110, 1, "vmin.u", int_arm_neon_vminu, 1>;
1269 def VMINfd : N3VDInt<0, 0, 0b10, 0b1111, 0, "vmin.f32", v2f32, v2f32,
1270 int_arm_neon_vmins, 1>;
1271 def VMINfq : N3VQInt<0, 0, 0b10, 0b1111, 0, "vmin.f32", v4f32, v4f32,
1272 int_arm_neon_vmins, 1>;
1274 // Vector Pairwise Operations.
1276 // VPADD : Vector Pairwise Add
1277 def VPADDi8 : N3VDInt<0, 0, 0b00, 0b1011, 1, "vpadd.i8", v8i8, v8i8,
1278 int_arm_neon_vpadd, 0>;
1279 def VPADDi16 : N3VDInt<0, 0, 0b01, 0b1011, 1, "vpadd.i16", v4i16, v4i16,
1280 int_arm_neon_vpadd, 0>;
1281 def VPADDi32 : N3VDInt<0, 0, 0b10, 0b1011, 1, "vpadd.i32", v2i32, v2i32,
1282 int_arm_neon_vpadd, 0>;
1283 def VPADDf : N3VDInt<1, 0, 0b00, 0b1101, 0, "vpadd.f32", v2f32, v2f32,
1284 int_arm_neon_vpadd, 0>;
1286 // VPADDL : Vector Pairwise Add Long
1287 defm VPADDLs : N2VPLInt_QHS<0b11, 0b11, 0b00, 0b00100, 0, "vpaddl.s",
1288 int_arm_neon_vpaddls>;
1289 defm VPADDLu : N2VPLInt_QHS<0b11, 0b11, 0b00, 0b00101, 0, "vpaddl.u",
1290 int_arm_neon_vpaddlu>;
1292 // VPADAL : Vector Pairwise Add and Accumulate Long
1293 defm VPADALs : N2VPLInt2_QHS<0b11, 0b11, 0b00, 0b00100, 0, "vpadal.s",
1294 int_arm_neon_vpadals>;
1295 defm VPADALu : N2VPLInt2_QHS<0b11, 0b11, 0b00, 0b00101, 0, "vpadal.u",
1296 int_arm_neon_vpadalu>;
1298 // VPMAX : Vector Pairwise Maximum
1299 def VPMAXs8 : N3VDInt<0, 0, 0b00, 0b1010, 0, "vpmax.s8", v8i8, v8i8,
1300 int_arm_neon_vpmaxs, 0>;
1301 def VPMAXs16 : N3VDInt<0, 0, 0b01, 0b1010, 0, "vpmax.s16", v4i16, v4i16,
1302 int_arm_neon_vpmaxs, 0>;
1303 def VPMAXs32 : N3VDInt<0, 0, 0b10, 0b1010, 0, "vpmax.s32", v2i32, v2i32,
1304 int_arm_neon_vpmaxs, 0>;
1305 def VPMAXu8 : N3VDInt<1, 0, 0b00, 0b1010, 0, "vpmax.u8", v8i8, v8i8,
1306 int_arm_neon_vpmaxu, 0>;
1307 def VPMAXu16 : N3VDInt<1, 0, 0b01, 0b1010, 0, "vpmax.u16", v4i16, v4i16,
1308 int_arm_neon_vpmaxu, 0>;
1309 def VPMAXu32 : N3VDInt<1, 0, 0b10, 0b1010, 0, "vpmax.u32", v2i32, v2i32,
1310 int_arm_neon_vpmaxu, 0>;
1311 def VPMAXf : N3VDInt<1, 0, 0b00, 0b1111, 0, "vpmax.f32", v2f32, v2f32,
1312 int_arm_neon_vpmaxs, 0>;
1314 // VPMIN : Vector Pairwise Minimum
1315 def VPMINs8 : N3VDInt<0, 0, 0b00, 0b1010, 1, "vpmin.s8", v8i8, v8i8,
1316 int_arm_neon_vpmins, 0>;
1317 def VPMINs16 : N3VDInt<0, 0, 0b01, 0b1010, 1, "vpmin.s16", v4i16, v4i16,
1318 int_arm_neon_vpmins, 0>;
1319 def VPMINs32 : N3VDInt<0, 0, 0b10, 0b1010, 1, "vpmin.s32", v2i32, v2i32,
1320 int_arm_neon_vpmins, 0>;
1321 def VPMINu8 : N3VDInt<1, 0, 0b00, 0b1010, 1, "vpmin.u8", v8i8, v8i8,
1322 int_arm_neon_vpminu, 0>;
1323 def VPMINu16 : N3VDInt<1, 0, 0b01, 0b1010, 1, "vpmin.u16", v4i16, v4i16,
1324 int_arm_neon_vpminu, 0>;
1325 def VPMINu32 : N3VDInt<1, 0, 0b10, 0b1010, 1, "vpmin.u32", v2i32, v2i32,
1326 int_arm_neon_vpminu, 0>;
1327 def VPMINf : N3VDInt<1, 0, 0b10, 0b1111, 0, "vpmin.f32", v2f32, v2f32,
1328 int_arm_neon_vpmins, 0>;
1330 // Vector Reciprocal and Reciprocal Square Root Estimate and Step.
1332 // VRECPE : Vector Reciprocal Estimate
1333 def VRECPEd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01000, 0, "vrecpe.u32",
1334 v2i32, v2i32, int_arm_neon_vrecpe>;
1335 def VRECPEq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01000, 0, "vrecpe.u32",
1336 v4i32, v4i32, int_arm_neon_vrecpe>;
1337 def VRECPEfd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01010, 0, "vrecpe.f32",
1338 v2f32, v2f32, int_arm_neon_vrecpe>;
1339 def VRECPEfq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01010, 0, "vrecpe.f32",
1340 v4f32, v4f32, int_arm_neon_vrecpe>;
1342 // VRECPS : Vector Reciprocal Step
1343 def VRECPSfd : N3VDInt<0, 0, 0b00, 0b1111, 1, "vrecps.f32", v2f32, v2f32,
1344 int_arm_neon_vrecps, 1>;
1345 def VRECPSfq : N3VQInt<0, 0, 0b00, 0b1111, 1, "vrecps.f32", v4f32, v4f32,
1346 int_arm_neon_vrecps, 1>;
1348 // VRSQRTE : Vector Reciprocal Square Root Estimate
1349 def VRSQRTEd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01001, 0, "vrsqrte.u32",
1350 v2i32, v2i32, int_arm_neon_vrsqrte>;
1351 def VRSQRTEq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01001, 0, "vrsqrte.u32",
1352 v4i32, v4i32, int_arm_neon_vrsqrte>;
1353 def VRSQRTEfd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01011, 0, "vrsqrte.f32",
1354 v2f32, v2f32, int_arm_neon_vrsqrte>;
1355 def VRSQRTEfq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01011, 0, "vrsqrte.f32",
1356 v4f32, v4f32, int_arm_neon_vrsqrte>;
1358 // VRSQRTS : Vector Reciprocal Square Root Step
1359 def VRSQRTSfd : N3VDInt<0, 0, 0b10, 0b1111, 1, "vrsqrts.f32", v2f32, v2f32,
1360 int_arm_neon_vrsqrts, 1>;
1361 def VRSQRTSfq : N3VQInt<0, 0, 0b10, 0b1111, 1, "vrsqrts.f32", v4f32, v4f32,
1362 int_arm_neon_vrsqrts, 1>;
1366 // VSHL : Vector Shift
1367 defm VSHLs : N3VInt_QHSD<0, 0, 0b0100, 0, "vshl.s", int_arm_neon_vshifts, 0>;
1368 defm VSHLu : N3VInt_QHSD<1, 0, 0b0100, 0, "vshl.u", int_arm_neon_vshiftu, 0>;
1369 // VSHL : Vector Shift Left (Immediate)
1370 defm VSHLi : N2VSh_QHSD<0, 1, 0b0111, 1, "vshl.i", NEONvshl>;
1371 // VSHR : Vector Shift Right (Immediate)
1372 defm VSHRs : N2VSh_QHSD<0, 1, 0b0000, 1, "vshr.s", NEONvshrs>;
1373 defm VSHRu : N2VSh_QHSD<1, 1, 0b0000, 1, "vshr.u", NEONvshru>;
1375 // VSHLL : Vector Shift Left Long
1376 def VSHLLs8 : N2VLSh<0, 1, 0b001000, 0b1010, 0, 0, 1, "vshll.s8",
1377 v8i16, v8i8, NEONvshlls>;
1378 def VSHLLs16 : N2VLSh<0, 1, 0b010000, 0b1010, 0, 0, 1, "vshll.s16",
1379 v4i32, v4i16, NEONvshlls>;
1380 def VSHLLs32 : N2VLSh<0, 1, 0b100000, 0b1010, 0, 0, 1, "vshll.s32",
1381 v2i64, v2i32, NEONvshlls>;
1382 def VSHLLu8 : N2VLSh<1, 1, 0b001000, 0b1010, 0, 0, 1, "vshll.u8",
1383 v8i16, v8i8, NEONvshllu>;
1384 def VSHLLu16 : N2VLSh<1, 1, 0b010000, 0b1010, 0, 0, 1, "vshll.u16",
1385 v4i32, v4i16, NEONvshllu>;
1386 def VSHLLu32 : N2VLSh<1, 1, 0b100000, 0b1010, 0, 0, 1, "vshll.u32",
1387 v2i64, v2i32, NEONvshllu>;
1389 // VSHLL : Vector Shift Left Long (with maximum shift count)
1390 def VSHLLi8 : N2VLSh<1, 1, 0b110010, 0b0011, 0, 0, 0, "vshll.i8",
1391 v8i16, v8i8, NEONvshlli>;
1392 def VSHLLi16 : N2VLSh<1, 1, 0b110110, 0b0011, 0, 0, 0, "vshll.i16",
1393 v4i32, v4i16, NEONvshlli>;
1394 def VSHLLi32 : N2VLSh<1, 1, 0b111010, 0b0011, 0, 0, 0, "vshll.i32",
1395 v2i64, v2i32, NEONvshlli>;
1397 // VSHRN : Vector Shift Right and Narrow
1398 def VSHRN16 : N2VNSh<0, 1, 0b001000, 0b1000, 0, 0, 1, "vshrn.i16",
1399 v8i8, v8i16, NEONvshrn>;
1400 def VSHRN32 : N2VNSh<0, 1, 0b010000, 0b1000, 0, 0, 1, "vshrn.i32",
1401 v4i16, v4i32, NEONvshrn>;
1402 def VSHRN64 : N2VNSh<0, 1, 0b100000, 0b1000, 0, 0, 1, "vshrn.i64",
1403 v2i32, v2i64, NEONvshrn>;
1405 // VRSHL : Vector Rounding Shift
1406 defm VRSHLs : N3VInt_QHSD<0,0,0b0101,0, "vrshl.s", int_arm_neon_vrshifts, 0>;
1407 defm VRSHLu : N3VInt_QHSD<1,0,0b0101,0, "vrshl.u", int_arm_neon_vrshiftu, 0>;
1408 // VRSHR : Vector Rounding Shift Right
1409 defm VRSHRs : N2VSh_QHSD<0, 1, 0b0010, 1, "vrshr.s", NEONvrshrs>;
1410 defm VRSHRu : N2VSh_QHSD<1, 1, 0b0010, 1, "vrshr.u", NEONvrshru>;
1412 // VRSHRN : Vector Rounding Shift Right and Narrow
1413 def VRSHRN16 : N2VNSh<0, 1, 0b001000, 0b1000, 0, 1, 1, "vrshrn.i16",
1414 v8i8, v8i16, NEONvrshrn>;
1415 def VRSHRN32 : N2VNSh<0, 1, 0b010000, 0b1000, 0, 1, 1, "vrshrn.i32",
1416 v4i16, v4i32, NEONvrshrn>;
1417 def VRSHRN64 : N2VNSh<0, 1, 0b100000, 0b1000, 0, 1, 1, "vrshrn.i64",
1418 v2i32, v2i64, NEONvrshrn>;
1420 // VQSHL : Vector Saturating Shift
1421 defm VQSHLs : N3VInt_QHSD<0,0,0b0100,1, "vqshl.s", int_arm_neon_vqshifts, 0>;
1422 defm VQSHLu : N3VInt_QHSD<1,0,0b0100,1, "vqshl.u", int_arm_neon_vqshiftu, 0>;
1423 // VQSHL : Vector Saturating Shift Left (Immediate)
1424 defm VQSHLsi : N2VSh_QHSD<0, 1, 0b0111, 1, "vqshl.s", NEONvqshls>;
1425 defm VQSHLui : N2VSh_QHSD<1, 1, 0b0111, 1, "vqshl.u", NEONvqshlu>;
1426 // VQSHLU : Vector Saturating Shift Left (Immediate, Unsigned)
1427 defm VQSHLsu : N2VSh_QHSD<1, 1, 0b0110, 1, "vqshlu.s", NEONvqshlsu>;
1429 // VQSHRN : Vector Saturating Shift Right and Narrow
1430 def VQSHRNs16 : N2VNSh<0, 1, 0b001000, 0b1001, 0, 0, 1, "vqshrn.s16",
1431 v8i8, v8i16, NEONvqshrns>;
1432 def VQSHRNs32 : N2VNSh<0, 1, 0b010000, 0b1001, 0, 0, 1, "vqshrn.s32",
1433 v4i16, v4i32, NEONvqshrns>;
1434 def VQSHRNs64 : N2VNSh<0, 1, 0b100000, 0b1001, 0, 0, 1, "vqshrn.s64",
1435 v2i32, v2i64, NEONvqshrns>;
1436 def VQSHRNu16 : N2VNSh<1, 1, 0b001000, 0b1001, 0, 0, 1, "vqshrn.u16",
1437 v8i8, v8i16, NEONvqshrnu>;
1438 def VQSHRNu32 : N2VNSh<1, 1, 0b010000, 0b1001, 0, 0, 1, "vqshrn.u32",
1439 v4i16, v4i32, NEONvqshrnu>;
1440 def VQSHRNu64 : N2VNSh<1, 1, 0b100000, 0b1001, 0, 0, 1, "vqshrn.u64",
1441 v2i32, v2i64, NEONvqshrnu>;
1443 // VQSHRUN : Vector Saturating Shift Right and Narrow (Unsigned)
1444 def VQSHRUN16 : N2VNSh<1, 1, 0b001000, 0b1000, 0, 0, 1, "vqshrun.s16",
1445 v8i8, v8i16, NEONvqshrnsu>;
1446 def VQSHRUN32 : N2VNSh<1, 1, 0b010000, 0b1000, 0, 0, 1, "vqshrun.s32",
1447 v4i16, v4i32, NEONvqshrnsu>;
1448 def VQSHRUN64 : N2VNSh<1, 1, 0b100000, 0b1000, 0, 0, 1, "vqshrun.s64",
1449 v2i32, v2i64, NEONvqshrnsu>;
1451 // VQRSHL : Vector Saturating Rounding Shift
1452 defm VQRSHLs : N3VInt_QHSD<0, 0, 0b0101, 1, "vqrshl.s",
1453 int_arm_neon_vqrshifts, 0>;
1454 defm VQRSHLu : N3VInt_QHSD<1, 0, 0b0101, 1, "vqrshl.u",
1455 int_arm_neon_vqrshiftu, 0>;
1457 // VQRSHRN : Vector Saturating Rounding Shift Right and Narrow
1458 def VQRSHRNs16: N2VNSh<0, 1, 0b001000, 0b1001, 0, 1, 1, "vqrshrn.s16",
1459 v8i8, v8i16, NEONvqrshrns>;
1460 def VQRSHRNs32: N2VNSh<0, 1, 0b010000, 0b1001, 0, 1, 1, "vqrshrn.s32",
1461 v4i16, v4i32, NEONvqrshrns>;
1462 def VQRSHRNs64: N2VNSh<0, 1, 0b100000, 0b1001, 0, 1, 1, "vqrshrn.s64",
1463 v2i32, v2i64, NEONvqrshrns>;
1464 def VQRSHRNu16: N2VNSh<1, 1, 0b001000, 0b1001, 0, 1, 1, "vqrshrn.u16",
1465 v8i8, v8i16, NEONvqrshrnu>;
1466 def VQRSHRNu32: N2VNSh<1, 1, 0b010000, 0b1001, 0, 1, 1, "vqrshrn.u32",
1467 v4i16, v4i32, NEONvqrshrnu>;
1468 def VQRSHRNu64: N2VNSh<1, 1, 0b100000, 0b1001, 0, 1, 1, "vqrshrn.u64",
1469 v2i32, v2i64, NEONvqrshrnu>;
1471 // VQRSHRUN : Vector Saturating Rounding Shift Right and Narrow (Unsigned)
1472 def VQRSHRUN16: N2VNSh<1, 1, 0b001000, 0b1000, 0, 1, 1, "vqrshrun.s16",
1473 v8i8, v8i16, NEONvqrshrnsu>;
1474 def VQRSHRUN32: N2VNSh<1, 1, 0b010000, 0b1000, 0, 1, 1, "vqrshrun.s32",
1475 v4i16, v4i32, NEONvqrshrnsu>;
1476 def VQRSHRUN64: N2VNSh<1, 1, 0b100000, 0b1000, 0, 1, 1, "vqrshrun.s64",
1477 v2i32, v2i64, NEONvqrshrnsu>;
1479 // VSRA : Vector Shift Right and Accumulate
1480 defm VSRAs : N2VShAdd_QHSD<0, 1, 0b0001, 1, "vsra.s", NEONvshrs>;
1481 defm VSRAu : N2VShAdd_QHSD<1, 1, 0b0001, 1, "vsra.u", NEONvshru>;
1482 // VRSRA : Vector Rounding Shift Right and Accumulate
1483 defm VRSRAs : N2VShAdd_QHSD<0, 1, 0b0011, 1, "vrsra.s", NEONvrshrs>;
1484 defm VRSRAu : N2VShAdd_QHSD<1, 1, 0b0011, 1, "vrsra.u", NEONvrshru>;
1486 // VSLI : Vector Shift Left and Insert
1487 defm VSLI : N2VShIns_QHSD<1, 1, 0b0101, 1, "vsli.", NEONvsli>;
1488 // VSRI : Vector Shift Right and Insert
1489 defm VSRI : N2VShIns_QHSD<1, 1, 0b0100, 1, "vsri.", NEONvsri>;
1491 // Vector Absolute and Saturating Absolute.
1493 // VABS : Vector Absolute Value
1494 defm VABS : N2VInt_QHS<0b11, 0b11, 0b01, 0b00110, 0, "vabs.s",
1496 def VABSfd : N2VDInt<0b11, 0b11, 0b10, 0b01, 0b01110, 0, "vabs.f32",
1497 v2f32, v2f32, int_arm_neon_vabs>;
1498 def VABSfq : N2VQInt<0b11, 0b11, 0b10, 0b01, 0b01110, 0, "vabs.f32",
1499 v4f32, v4f32, int_arm_neon_vabs>;
1501 // VQABS : Vector Saturating Absolute Value
1502 defm VQABS : N2VInt_QHS<0b11, 0b11, 0b00, 0b01110, 0, "vqabs.s",
1503 int_arm_neon_vqabs>;
1507 def vneg : PatFrag<(ops node:$in), (sub immAllZerosV, node:$in)>;
1508 def vneg_conv : PatFrag<(ops node:$in), (sub immAllZerosV_bc, node:$in)>;
1510 class VNEGD<bits<2> size, string OpcodeStr, ValueType Ty>
1511 : N2V<0b11, 0b11, size, 0b01, 0b00111, 0, 0, (outs DPR:$dst), (ins DPR:$src),
1513 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1514 [(set DPR:$dst, (Ty (vneg DPR:$src)))]>;
1515 class VNEGQ<bits<2> size, string OpcodeStr, ValueType Ty>
1516 : N2V<0b11, 0b11, size, 0b01, 0b00111, 1, 0, (outs QPR:$dst), (ins QPR:$src),
1518 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1519 [(set QPR:$dst, (Ty (vneg QPR:$src)))]>;
1521 // VNEG : Vector Negate
1522 def VNEGs8d : VNEGD<0b00, "vneg.s8", v8i8>;
1523 def VNEGs16d : VNEGD<0b01, "vneg.s16", v4i16>;
1524 def VNEGs32d : VNEGD<0b10, "vneg.s32", v2i32>;
1525 def VNEGs8q : VNEGQ<0b00, "vneg.s8", v16i8>;
1526 def VNEGs16q : VNEGQ<0b01, "vneg.s16", v8i16>;
1527 def VNEGs32q : VNEGQ<0b10, "vneg.s32", v4i32>;
1529 // VNEG : Vector Negate (floating-point)
1530 def VNEGf32d : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 0, 0,
1531 (outs DPR:$dst), (ins DPR:$src), NoItinerary,
1532 "vneg.f32\t$dst, $src", "",
1533 [(set DPR:$dst, (v2f32 (fneg DPR:$src)))]>;
1534 def VNEGf32q : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 1, 0,
1535 (outs QPR:$dst), (ins QPR:$src), NoItinerary,
1536 "vneg.f32\t$dst, $src", "",
1537 [(set QPR:$dst, (v4f32 (fneg QPR:$src)))]>;
1539 def : Pat<(v8i8 (vneg_conv DPR:$src)), (VNEGs8d DPR:$src)>;
1540 def : Pat<(v4i16 (vneg_conv DPR:$src)), (VNEGs16d DPR:$src)>;
1541 def : Pat<(v2i32 (vneg_conv DPR:$src)), (VNEGs32d DPR:$src)>;
1542 def : Pat<(v16i8 (vneg_conv QPR:$src)), (VNEGs8q QPR:$src)>;
1543 def : Pat<(v8i16 (vneg_conv QPR:$src)), (VNEGs16q QPR:$src)>;
1544 def : Pat<(v4i32 (vneg_conv QPR:$src)), (VNEGs32q QPR:$src)>;
1546 // VQNEG : Vector Saturating Negate
1547 defm VQNEG : N2VInt_QHS<0b11, 0b11, 0b00, 0b01111, 0, "vqneg.s",
1548 int_arm_neon_vqneg>;
1550 // Vector Bit Counting Operations.
1552 // VCLS : Vector Count Leading Sign Bits
1553 defm VCLS : N2VInt_QHS<0b11, 0b11, 0b00, 0b01000, 0, "vcls.s",
1555 // VCLZ : Vector Count Leading Zeros
1556 defm VCLZ : N2VInt_QHS<0b11, 0b11, 0b00, 0b01001, 0, "vclz.i",
1558 // VCNT : Vector Count One Bits
1559 def VCNTd : N2VDInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0, "vcnt.8",
1560 v8i8, v8i8, int_arm_neon_vcnt>;
1561 def VCNTq : N2VQInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0, "vcnt.8",
1562 v16i8, v16i8, int_arm_neon_vcnt>;
1564 // Vector Move Operations.
1566 // VMOV : Vector Move (Register)
1568 def VMOVD : N3V<0, 0, 0b10, 0b0001, 0, 1, (outs DPR:$dst), (ins DPR:$src),
1569 NoItinerary, "vmov\t$dst, $src", "", []>;
1570 def VMOVQ : N3V<0, 0, 0b10, 0b0001, 1, 1, (outs QPR:$dst), (ins QPR:$src),
1571 NoItinerary, "vmov\t$dst, $src", "", []>;
1573 // VMOV : Vector Move (Immediate)
1575 // VMOV_get_imm8 xform function: convert build_vector to VMOV.i8 imm.
1576 def VMOV_get_imm8 : SDNodeXForm<build_vector, [{
1577 return ARM::getVMOVImm(N, 1, *CurDAG);
1579 def vmovImm8 : PatLeaf<(build_vector), [{
1580 return ARM::getVMOVImm(N, 1, *CurDAG).getNode() != 0;
1583 // VMOV_get_imm16 xform function: convert build_vector to VMOV.i16 imm.
1584 def VMOV_get_imm16 : SDNodeXForm<build_vector, [{
1585 return ARM::getVMOVImm(N, 2, *CurDAG);
1587 def vmovImm16 : PatLeaf<(build_vector), [{
1588 return ARM::getVMOVImm(N, 2, *CurDAG).getNode() != 0;
1589 }], VMOV_get_imm16>;
1591 // VMOV_get_imm32 xform function: convert build_vector to VMOV.i32 imm.
1592 def VMOV_get_imm32 : SDNodeXForm<build_vector, [{
1593 return ARM::getVMOVImm(N, 4, *CurDAG);
1595 def vmovImm32 : PatLeaf<(build_vector), [{
1596 return ARM::getVMOVImm(N, 4, *CurDAG).getNode() != 0;
1597 }], VMOV_get_imm32>;
1599 // VMOV_get_imm64 xform function: convert build_vector to VMOV.i64 imm.
1600 def VMOV_get_imm64 : SDNodeXForm<build_vector, [{
1601 return ARM::getVMOVImm(N, 8, *CurDAG);
1603 def vmovImm64 : PatLeaf<(build_vector), [{
1604 return ARM::getVMOVImm(N, 8, *CurDAG).getNode() != 0;
1605 }], VMOV_get_imm64>;
1607 // Note: Some of the cmode bits in the following VMOV instructions need to
1608 // be encoded based on the immed values.
1610 def VMOVv8i8 : N1ModImm<1, 0b000, 0b1110, 0, 0, 0, 1, (outs DPR:$dst),
1611 (ins i8imm:$SIMM), NoItinerary,
1612 "vmov.i8\t$dst, $SIMM", "",
1613 [(set DPR:$dst, (v8i8 vmovImm8:$SIMM))]>;
1614 def VMOVv16i8 : N1ModImm<1, 0b000, 0b1110, 0, 1, 0, 1, (outs QPR:$dst),
1615 (ins i8imm:$SIMM), NoItinerary,
1616 "vmov.i8\t$dst, $SIMM", "",
1617 [(set QPR:$dst, (v16i8 vmovImm8:$SIMM))]>;
1619 def VMOVv4i16 : N1ModImm<1, 0b000, 0b1000, 0, 0, 0, 1, (outs DPR:$dst),
1620 (ins i16imm:$SIMM), NoItinerary,
1621 "vmov.i16\t$dst, $SIMM", "",
1622 [(set DPR:$dst, (v4i16 vmovImm16:$SIMM))]>;
1623 def VMOVv8i16 : N1ModImm<1, 0b000, 0b1000, 0, 1, 0, 1, (outs QPR:$dst),
1624 (ins i16imm:$SIMM), NoItinerary,
1625 "vmov.i16\t$dst, $SIMM", "",
1626 [(set QPR:$dst, (v8i16 vmovImm16:$SIMM))]>;
1628 def VMOVv2i32 : N1ModImm<1, 0b000, 0b0000, 0, 0, 0, 1, (outs DPR:$dst),
1629 (ins i32imm:$SIMM), NoItinerary,
1630 "vmov.i32\t$dst, $SIMM", "",
1631 [(set DPR:$dst, (v2i32 vmovImm32:$SIMM))]>;
1632 def VMOVv4i32 : N1ModImm<1, 0b000, 0b0000, 0, 1, 0, 1, (outs QPR:$dst),
1633 (ins i32imm:$SIMM), NoItinerary,
1634 "vmov.i32\t$dst, $SIMM", "",
1635 [(set QPR:$dst, (v4i32 vmovImm32:$SIMM))]>;
1637 def VMOVv1i64 : N1ModImm<1, 0b000, 0b1110, 0, 0, 1, 1, (outs DPR:$dst),
1638 (ins i64imm:$SIMM), NoItinerary,
1639 "vmov.i64\t$dst, $SIMM", "",
1640 [(set DPR:$dst, (v1i64 vmovImm64:$SIMM))]>;
1641 def VMOVv2i64 : N1ModImm<1, 0b000, 0b1110, 0, 1, 1, 1, (outs QPR:$dst),
1642 (ins i64imm:$SIMM), NoItinerary,
1643 "vmov.i64\t$dst, $SIMM", "",
1644 [(set QPR:$dst, (v2i64 vmovImm64:$SIMM))]>;
1646 // VMOV : Vector Get Lane (move scalar to ARM core register)
1648 def VGETLNs8 : NVGetLane<0b11100101, 0b1011, 0b00,
1649 (outs GPR:$dst), (ins DPR:$src, lane_cst:$lane),
1650 NoItinerary, "vmov", ".s8\t$dst, $src[$lane]",
1651 [(set GPR:$dst, (NEONvgetlanes (v8i8 DPR:$src),
1653 def VGETLNs16 : NVGetLane<0b11100001, 0b1011, 0b01,
1654 (outs GPR:$dst), (ins DPR:$src, lane_cst:$lane),
1655 NoItinerary, "vmov", ".s16\t$dst, $src[$lane]",
1656 [(set GPR:$dst, (NEONvgetlanes (v4i16 DPR:$src),
1658 def VGETLNu8 : NVGetLane<0b11101101, 0b1011, 0b00,
1659 (outs GPR:$dst), (ins DPR:$src, lane_cst:$lane),
1660 NoItinerary, "vmov", ".u8\t$dst, $src[$lane]",
1661 [(set GPR:$dst, (NEONvgetlaneu (v8i8 DPR:$src),
1663 def VGETLNu16 : NVGetLane<0b11101001, 0b1011, 0b01,
1664 (outs GPR:$dst), (ins DPR:$src, lane_cst:$lane),
1665 NoItinerary, "vmov", ".u16\t$dst, $src[$lane]",
1666 [(set GPR:$dst, (NEONvgetlaneu (v4i16 DPR:$src),
1668 def VGETLNi32 : NVGetLane<0b11100001, 0b1011, 0b00,
1669 (outs GPR:$dst), (ins DPR:$src, lane_cst:$lane),
1670 NoItinerary, "vmov", ".32\t$dst, $src[$lane]",
1671 [(set GPR:$dst, (extractelt (v2i32 DPR:$src),
1673 // def VGETLNf32: see FMRDH and FMRDL in ARMInstrVFP.td
1674 def : Pat<(NEONvgetlanes (v16i8 QPR:$src), imm:$lane),
1675 (VGETLNs8 (v8i8 (EXTRACT_SUBREG QPR:$src,
1676 (DSubReg_i8_reg imm:$lane))),
1677 (SubReg_i8_lane imm:$lane))>;
1678 def : Pat<(NEONvgetlanes (v8i16 QPR:$src), imm:$lane),
1679 (VGETLNs16 (v4i16 (EXTRACT_SUBREG QPR:$src,
1680 (DSubReg_i16_reg imm:$lane))),
1681 (SubReg_i16_lane imm:$lane))>;
1682 def : Pat<(NEONvgetlaneu (v16i8 QPR:$src), imm:$lane),
1683 (VGETLNu8 (v8i8 (EXTRACT_SUBREG QPR:$src,
1684 (DSubReg_i8_reg imm:$lane))),
1685 (SubReg_i8_lane imm:$lane))>;
1686 def : Pat<(NEONvgetlaneu (v8i16 QPR:$src), imm:$lane),
1687 (VGETLNu16 (v4i16 (EXTRACT_SUBREG QPR:$src,
1688 (DSubReg_i16_reg imm:$lane))),
1689 (SubReg_i16_lane imm:$lane))>;
1690 def : Pat<(extractelt (v4i32 QPR:$src), imm:$lane),
1691 (VGETLNi32 (v2i32 (EXTRACT_SUBREG QPR:$src,
1692 (DSubReg_i32_reg imm:$lane))),
1693 (SubReg_i32_lane imm:$lane))>;
1694 def : Pat<(extractelt (v4f32 QPR:$src1), imm:$src2),
1695 (EXTRACT_SUBREG QPR:$src1, (SSubReg_f32_reg imm:$src2))>;
1696 //def : Pat<(extractelt (v2i64 QPR:$src1), imm:$src2),
1697 // (EXTRACT_SUBREG QPR:$src1, (DSubReg_f64_reg imm:$src2))>;
1698 def : Pat<(extractelt (v2f64 QPR:$src1), imm:$src2),
1699 (EXTRACT_SUBREG QPR:$src1, (DSubReg_f64_reg imm:$src2))>;
1702 // VMOV : Vector Set Lane (move ARM core register to scalar)
1704 let Constraints = "$src1 = $dst" in {
1705 def VSETLNi8 : NVSetLane<0b11100100, 0b1011, 0b00, (outs DPR:$dst),
1706 (ins DPR:$src1, GPR:$src2, lane_cst:$lane),
1707 NoItinerary, "vmov", ".8\t$dst[$lane], $src2",
1708 [(set DPR:$dst, (vector_insert (v8i8 DPR:$src1),
1709 GPR:$src2, imm:$lane))]>;
1710 def VSETLNi16 : NVSetLane<0b11100000, 0b1011, 0b01, (outs DPR:$dst),
1711 (ins DPR:$src1, GPR:$src2, lane_cst:$lane),
1712 NoItinerary, "vmov", ".16\t$dst[$lane], $src2",
1713 [(set DPR:$dst, (vector_insert (v4i16 DPR:$src1),
1714 GPR:$src2, imm:$lane))]>;
1715 def VSETLNi32 : NVSetLane<0b11100000, 0b1011, 0b00, (outs DPR:$dst),
1716 (ins DPR:$src1, GPR:$src2, lane_cst:$lane),
1717 NoItinerary, "vmov", ".32\t$dst[$lane], $src2",
1718 [(set DPR:$dst, (insertelt (v2i32 DPR:$src1),
1719 GPR:$src2, imm:$lane))]>;
1721 def : Pat<(vector_insert (v16i8 QPR:$src1), GPR:$src2, imm:$lane),
1722 (v16i8 (INSERT_SUBREG QPR:$src1,
1723 (VSETLNi8 (v8i8 (EXTRACT_SUBREG QPR:$src1,
1724 (DSubReg_i8_reg imm:$lane))),
1725 GPR:$src2, (SubReg_i8_lane imm:$lane)),
1726 (DSubReg_i8_reg imm:$lane)))>;
1727 def : Pat<(vector_insert (v8i16 QPR:$src1), GPR:$src2, imm:$lane),
1728 (v8i16 (INSERT_SUBREG QPR:$src1,
1729 (VSETLNi16 (v4i16 (EXTRACT_SUBREG QPR:$src1,
1730 (DSubReg_i16_reg imm:$lane))),
1731 GPR:$src2, (SubReg_i16_lane imm:$lane)),
1732 (DSubReg_i16_reg imm:$lane)))>;
1733 def : Pat<(insertelt (v4i32 QPR:$src1), GPR:$src2, imm:$lane),
1734 (v4i32 (INSERT_SUBREG QPR:$src1,
1735 (VSETLNi32 (v2i32 (EXTRACT_SUBREG QPR:$src1,
1736 (DSubReg_i32_reg imm:$lane))),
1737 GPR:$src2, (SubReg_i32_lane imm:$lane)),
1738 (DSubReg_i32_reg imm:$lane)))>;
1740 def : Pat<(v4f32 (insertelt QPR:$src1, SPR:$src2, imm:$src3)),
1741 (INSERT_SUBREG QPR:$src1, SPR:$src2, (SSubReg_f32_reg imm:$src3))>;
1743 //def : Pat<(v2i64 (insertelt QPR:$src1, DPR:$src2, imm:$src3)),
1744 // (INSERT_SUBREG QPR:$src1, DPR:$src2, (DSubReg_f64_reg imm:$src3))>;
1745 def : Pat<(v2f64 (insertelt QPR:$src1, DPR:$src2, imm:$src3)),
1746 (INSERT_SUBREG QPR:$src1, DPR:$src2, (DSubReg_f64_reg imm:$src3))>;
1748 // VDUP : Vector Duplicate (from ARM core register to all elements)
1750 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
1751 (vector_shuffle node:$lhs, node:$rhs), [{
1752 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1753 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
1756 class VDUPD<bits<8> opcod1, bits<2> opcod3, string asmSize, ValueType Ty>
1757 : NVDup<opcod1, 0b1011, opcod3, (outs DPR:$dst), (ins GPR:$src),
1758 NoItinerary, "vdup", !strconcat(asmSize, "\t$dst, $src"),
1759 [(set DPR:$dst, (Ty (splat_lo (scalar_to_vector GPR:$src), undef)))]>;
1760 class VDUPQ<bits<8> opcod1, bits<2> opcod3, string asmSize, ValueType Ty>
1761 : NVDup<opcod1, 0b1011, opcod3, (outs QPR:$dst), (ins GPR:$src),
1762 NoItinerary, "vdup", !strconcat(asmSize, "\t$dst, $src"),
1763 [(set QPR:$dst, (Ty (splat_lo (scalar_to_vector GPR:$src), undef)))]>;
1765 def VDUP8d : VDUPD<0b11101100, 0b00, ".8", v8i8>;
1766 def VDUP16d : VDUPD<0b11101000, 0b01, ".16", v4i16>;
1767 def VDUP32d : VDUPD<0b11101000, 0b00, ".32", v2i32>;
1768 def VDUP8q : VDUPQ<0b11101110, 0b00, ".8", v16i8>;
1769 def VDUP16q : VDUPQ<0b11101010, 0b01, ".16", v8i16>;
1770 def VDUP32q : VDUPQ<0b11101010, 0b00, ".32", v4i32>;
1772 def VDUPfd : NVDup<0b11101000, 0b1011, 0b00, (outs DPR:$dst), (ins GPR:$src),
1773 NoItinerary, "vdup", ".32\t$dst, $src",
1774 [(set DPR:$dst, (v2f32 (splat_lo
1776 (f32 (bitconvert GPR:$src))),
1778 def VDUPfq : NVDup<0b11101010, 0b1011, 0b00, (outs QPR:$dst), (ins GPR:$src),
1779 NoItinerary, "vdup", ".32\t$dst, $src",
1780 [(set QPR:$dst, (v4f32 (splat_lo
1782 (f32 (bitconvert GPR:$src))),
1785 // VDUP : Vector Duplicate Lane (from scalar to all elements)
1787 class VDUPLND<bits<2> op19_18, bits<2> op17_16, string OpcodeStr, ValueType Ty>
1788 : N2V<0b11, 0b11, op19_18, op17_16, 0b11000, 0, 0,
1789 (outs DPR:$dst), (ins DPR:$src, lane_cst:$lane), NoItinerary,
1790 !strconcat(OpcodeStr, "\t$dst, $src[$lane]"), "",
1791 [(set DPR:$dst, (Ty (NEONvduplane (Ty DPR:$src), imm:$lane)))]>;
1793 class VDUPLNQ<bits<2> op19_18, bits<2> op17_16, string OpcodeStr,
1794 ValueType ResTy, ValueType OpTy>
1795 : N2V<0b11, 0b11, op19_18, op17_16, 0b11000, 1, 0,
1796 (outs QPR:$dst), (ins DPR:$src, lane_cst:$lane), NoItinerary,
1797 !strconcat(OpcodeStr, "\t$dst, $src[$lane]"), "",
1798 [(set QPR:$dst, (ResTy (NEONvduplane (OpTy DPR:$src), imm:$lane)))]>;
1800 def VDUPLN8d : VDUPLND<0b00, 0b01, "vdup.8", v8i8>;
1801 def VDUPLN16d : VDUPLND<0b00, 0b10, "vdup.16", v4i16>;
1802 def VDUPLN32d : VDUPLND<0b01, 0b00, "vdup.32", v2i32>;
1803 def VDUPLNfd : VDUPLND<0b01, 0b00, "vdup.32", v2f32>;
1804 def VDUPLN8q : VDUPLNQ<0b00, 0b01, "vdup.8", v16i8, v8i8>;
1805 def VDUPLN16q : VDUPLNQ<0b00, 0b10, "vdup.16", v8i16, v4i16>;
1806 def VDUPLN32q : VDUPLNQ<0b01, 0b00, "vdup.32", v4i32, v2i32>;
1807 def VDUPLNfq : VDUPLNQ<0b01, 0b00, "vdup.32", v4f32, v2f32>;
1809 def : Pat<(v16i8 (NEONvduplane (v16i8 QPR:$src), imm:$lane)),
1810 (v16i8 (VDUPLN8q (v8i8 (EXTRACT_SUBREG QPR:$src,
1811 (DSubReg_i8_reg imm:$lane))),
1812 (SubReg_i8_lane imm:$lane)))>;
1813 def : Pat<(v8i16 (NEONvduplane (v8i16 QPR:$src), imm:$lane)),
1814 (v8i16 (VDUPLN16q (v4i16 (EXTRACT_SUBREG QPR:$src,
1815 (DSubReg_i16_reg imm:$lane))),
1816 (SubReg_i16_lane imm:$lane)))>;
1817 def : Pat<(v4i32 (NEONvduplane (v4i32 QPR:$src), imm:$lane)),
1818 (v4i32 (VDUPLN32q (v2i32 (EXTRACT_SUBREG QPR:$src,
1819 (DSubReg_i32_reg imm:$lane))),
1820 (SubReg_i32_lane imm:$lane)))>;
1821 def : Pat<(v4f32 (NEONvduplane (v4f32 QPR:$src), imm:$lane)),
1822 (v4f32 (VDUPLNfq (v2f32 (EXTRACT_SUBREG QPR:$src,
1823 (DSubReg_i32_reg imm:$lane))),
1824 (SubReg_i32_lane imm:$lane)))>;
1826 def VDUPfdf : N2V<0b11, 0b11, 0b01, 0b00, 0b11000, 0, 0,
1827 (outs DPR:$dst), (ins SPR:$src),
1828 NoItinerary, "vdup.32\t$dst, ${src:lane}", "",
1829 [(set DPR:$dst, (v2f32 (splat_lo
1830 (scalar_to_vector SPR:$src),
1833 def VDUPfqf : N2V<0b11, 0b11, 0b01, 0b00, 0b11000, 1, 0,
1834 (outs QPR:$dst), (ins SPR:$src),
1835 NoItinerary, "vdup.32\t$dst, ${src:lane}", "",
1836 [(set QPR:$dst, (v4f32 (splat_lo
1837 (scalar_to_vector SPR:$src),
1840 // VMOVN : Vector Narrowing Move
1841 defm VMOVN : N2VNInt_HSD<0b11,0b11,0b10,0b00100,0,0, "vmovn.i",
1842 int_arm_neon_vmovn>;
1843 // VQMOVN : Vector Saturating Narrowing Move
1844 defm VQMOVNs : N2VNInt_HSD<0b11,0b11,0b10,0b00101,0,0, "vqmovn.s",
1845 int_arm_neon_vqmovns>;
1846 defm VQMOVNu : N2VNInt_HSD<0b11,0b11,0b10,0b00101,1,0, "vqmovn.u",
1847 int_arm_neon_vqmovnu>;
1848 defm VQMOVNsu : N2VNInt_HSD<0b11,0b11,0b10,0b00100,1,0, "vqmovun.s",
1849 int_arm_neon_vqmovnsu>;
1850 // VMOVL : Vector Lengthening Move
1851 defm VMOVLs : N2VLInt_QHS<0,1,0b1010,0,0,1, "vmovl.s", int_arm_neon_vmovls>;
1852 defm VMOVLu : N2VLInt_QHS<1,1,0b1010,0,0,1, "vmovl.u", int_arm_neon_vmovlu>;
1854 // Vector Conversions.
1856 // VCVT : Vector Convert Between Floating-Point and Integers
1857 def VCVTf2sd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt.s32.f32",
1858 v2i32, v2f32, fp_to_sint>;
1859 def VCVTf2ud : N2VD<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt.u32.f32",
1860 v2i32, v2f32, fp_to_uint>;
1861 def VCVTs2fd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt.f32.s32",
1862 v2f32, v2i32, sint_to_fp>;
1863 def VCVTu2fd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt.f32.u32",
1864 v2f32, v2i32, uint_to_fp>;
1866 def VCVTf2sq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt.s32.f32",
1867 v4i32, v4f32, fp_to_sint>;
1868 def VCVTf2uq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt.u32.f32",
1869 v4i32, v4f32, fp_to_uint>;
1870 def VCVTs2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt.f32.s32",
1871 v4f32, v4i32, sint_to_fp>;
1872 def VCVTu2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt.f32.u32",
1873 v4f32, v4i32, uint_to_fp>;
1875 // VCVT : Vector Convert Between Floating-Point and Fixed-Point.
1876 // Note: Some of the opcode bits in the following VCVT instructions need to
1877 // be encoded based on the immed values.
1878 def VCVTf2xsd : N2VCvtD<0, 1, 0b000000, 0b1111, 0, 1, "vcvt.s32.f32",
1879 v2i32, v2f32, int_arm_neon_vcvtfp2fxs>;
1880 def VCVTf2xud : N2VCvtD<1, 1, 0b000000, 0b1111, 0, 1, "vcvt.u32.f32",
1881 v2i32, v2f32, int_arm_neon_vcvtfp2fxu>;
1882 def VCVTxs2fd : N2VCvtD<0, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.s32",
1883 v2f32, v2i32, int_arm_neon_vcvtfxs2fp>;
1884 def VCVTxu2fd : N2VCvtD<1, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.u32",
1885 v2f32, v2i32, int_arm_neon_vcvtfxu2fp>;
1887 def VCVTf2xsq : N2VCvtQ<0, 1, 0b000000, 0b1111, 0, 1, "vcvt.s32.f32",
1888 v4i32, v4f32, int_arm_neon_vcvtfp2fxs>;
1889 def VCVTf2xuq : N2VCvtQ<1, 1, 0b000000, 0b1111, 0, 1, "vcvt.u32.f32",
1890 v4i32, v4f32, int_arm_neon_vcvtfp2fxu>;
1891 def VCVTxs2fq : N2VCvtQ<0, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.s32",
1892 v4f32, v4i32, int_arm_neon_vcvtfxs2fp>;
1893 def VCVTxu2fq : N2VCvtQ<1, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.u32",
1894 v4f32, v4i32, int_arm_neon_vcvtfxu2fp>;
1898 // VREV64 : Vector Reverse elements within 64-bit doublewords
1900 class VREV64D<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1901 : N2V<0b11, 0b11, op19_18, 0b00, 0b00000, 0, 0, (outs DPR:$dst),
1902 (ins DPR:$src), NoItinerary,
1903 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1904 [(set DPR:$dst, (Ty (NEONvrev64 (Ty DPR:$src))))]>;
1905 class VREV64Q<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1906 : N2V<0b11, 0b11, op19_18, 0b00, 0b00000, 1, 0, (outs QPR:$dst),
1907 (ins QPR:$src), NoItinerary,
1908 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1909 [(set QPR:$dst, (Ty (NEONvrev64 (Ty QPR:$src))))]>;
1911 def VREV64d8 : VREV64D<0b00, "vrev64.8", v8i8>;
1912 def VREV64d16 : VREV64D<0b01, "vrev64.16", v4i16>;
1913 def VREV64d32 : VREV64D<0b10, "vrev64.32", v2i32>;
1914 def VREV64df : VREV64D<0b10, "vrev64.32", v2f32>;
1916 def VREV64q8 : VREV64Q<0b00, "vrev64.8", v16i8>;
1917 def VREV64q16 : VREV64Q<0b01, "vrev64.16", v8i16>;
1918 def VREV64q32 : VREV64Q<0b10, "vrev64.32", v4i32>;
1919 def VREV64qf : VREV64Q<0b10, "vrev64.32", v4f32>;
1921 // VREV32 : Vector Reverse elements within 32-bit words
1923 class VREV32D<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1924 : N2V<0b11, 0b11, op19_18, 0b00, 0b00001, 0, 0, (outs DPR:$dst),
1925 (ins DPR:$src), NoItinerary,
1926 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1927 [(set DPR:$dst, (Ty (NEONvrev32 (Ty DPR:$src))))]>;
1928 class VREV32Q<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1929 : N2V<0b11, 0b11, op19_18, 0b00, 0b00001, 1, 0, (outs QPR:$dst),
1930 (ins QPR:$src), NoItinerary,
1931 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1932 [(set QPR:$dst, (Ty (NEONvrev32 (Ty QPR:$src))))]>;
1934 def VREV32d8 : VREV32D<0b00, "vrev32.8", v8i8>;
1935 def VREV32d16 : VREV32D<0b01, "vrev32.16", v4i16>;
1937 def VREV32q8 : VREV32Q<0b00, "vrev32.8", v16i8>;
1938 def VREV32q16 : VREV32Q<0b01, "vrev32.16", v8i16>;
1940 // VREV16 : Vector Reverse elements within 16-bit halfwords
1942 class VREV16D<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1943 : N2V<0b11, 0b11, op19_18, 0b00, 0b00010, 0, 0, (outs DPR:$dst),
1944 (ins DPR:$src), NoItinerary,
1945 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1946 [(set DPR:$dst, (Ty (NEONvrev16 (Ty DPR:$src))))]>;
1947 class VREV16Q<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1948 : N2V<0b11, 0b11, op19_18, 0b00, 0b00010, 1, 0, (outs QPR:$dst),
1949 (ins QPR:$src), NoItinerary,
1950 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1951 [(set QPR:$dst, (Ty (NEONvrev16 (Ty QPR:$src))))]>;
1953 def VREV16d8 : VREV16D<0b00, "vrev16.8", v8i8>;
1954 def VREV16q8 : VREV16Q<0b00, "vrev16.8", v16i8>;
1956 // VTRN : Vector Transpose
1958 def VTRNd8 : N2VDShuffle<0b00, 0b00001, "vtrn.8">;
1959 def VTRNd16 : N2VDShuffle<0b01, 0b00001, "vtrn.16">;
1960 def VTRNd32 : N2VDShuffle<0b10, 0b00001, "vtrn.32">;
1962 def VTRNq8 : N2VQShuffle<0b00, 0b00001, "vtrn.8">;
1963 def VTRNq16 : N2VQShuffle<0b01, 0b00001, "vtrn.16">;
1964 def VTRNq32 : N2VQShuffle<0b10, 0b00001, "vtrn.32">;
1966 // VUZP : Vector Unzip (Deinterleave)
1968 def VUZPd8 : N2VDShuffle<0b00, 0b00010, "vuzp.8">;
1969 def VUZPd16 : N2VDShuffle<0b01, 0b00010, "vuzp.16">;
1970 def VUZPd32 : N2VDShuffle<0b10, 0b00010, "vuzp.32">;
1972 def VUZPq8 : N2VQShuffle<0b00, 0b00010, "vuzp.8">;
1973 def VUZPq16 : N2VQShuffle<0b01, 0b00010, "vuzp.16">;
1974 def VUZPq32 : N2VQShuffle<0b10, 0b00010, "vuzp.32">;
1976 // VZIP : Vector Zip (Interleave)
1978 def VZIPd8 : N2VDShuffle<0b00, 0b00011, "vzip.8">;
1979 def VZIPd16 : N2VDShuffle<0b01, 0b00011, "vzip.16">;
1980 def VZIPd32 : N2VDShuffle<0b10, 0b00011, "vzip.32">;
1982 def VZIPq8 : N2VQShuffle<0b00, 0b00011, "vzip.8">;
1983 def VZIPq16 : N2VQShuffle<0b01, 0b00011, "vzip.16">;
1984 def VZIPq32 : N2VQShuffle<0b10, 0b00011, "vzip.32">;
1986 // Vector Table Lookup and Table Extension.
1988 // VTBL : Vector Table Lookup
1990 : N3V<1,1,0b11,0b1000,0,0, (outs DPR:$dst),
1991 (ins DPR:$tbl1, DPR:$src), NoItinerary,
1992 "vtbl.8\t$dst, \\{$tbl1\\}, $src", "",
1993 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbl1 DPR:$tbl1, DPR:$src)))]>;
1995 : N3V<1,1,0b11,0b1001,0,0, (outs DPR:$dst),
1996 (ins DPR:$tbl1, DPR:$tbl2, DPR:$src), NoItinerary,
1997 "vtbl.8\t$dst, \\{$tbl1,$tbl2\\}, $src", "",
1998 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbl2
1999 DPR:$tbl1, DPR:$tbl2, DPR:$src)))]>;
2001 : N3V<1,1,0b11,0b1010,0,0, (outs DPR:$dst),
2002 (ins DPR:$tbl1, DPR:$tbl2, DPR:$tbl3, DPR:$src), NoItinerary,
2003 "vtbl.8\t$dst, \\{$tbl1,$tbl2,$tbl3\\}, $src", "",
2004 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbl3
2005 DPR:$tbl1, DPR:$tbl2, DPR:$tbl3, DPR:$src)))]>;
2007 : N3V<1,1,0b11,0b1011,0,0, (outs DPR:$dst),
2008 (ins DPR:$tbl1, DPR:$tbl2, DPR:$tbl3, DPR:$tbl4, DPR:$src), NoItinerary,
2009 "vtbl.8\t$dst, \\{$tbl1,$tbl2,$tbl3,$tbl4\\}, $src", "",
2010 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbl4 DPR:$tbl1, DPR:$tbl2,
2011 DPR:$tbl3, DPR:$tbl4, DPR:$src)))]>;
2013 // VTBX : Vector Table Extension
2015 : N3V<1,1,0b11,0b1000,1,0, (outs DPR:$dst),
2016 (ins DPR:$orig, DPR:$tbl1, DPR:$src), NoItinerary,
2017 "vtbx.8\t$dst, \\{$tbl1\\}, $src", "$orig = $dst",
2018 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbx1
2019 DPR:$orig, DPR:$tbl1, DPR:$src)))]>;
2021 : N3V<1,1,0b11,0b1001,1,0, (outs DPR:$dst),
2022 (ins DPR:$orig, DPR:$tbl1, DPR:$tbl2, DPR:$src), NoItinerary,
2023 "vtbx.8\t$dst, \\{$tbl1,$tbl2\\}, $src", "$orig = $dst",
2024 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbx2
2025 DPR:$orig, DPR:$tbl1, DPR:$tbl2, DPR:$src)))]>;
2027 : N3V<1,1,0b11,0b1010,1,0, (outs DPR:$dst),
2028 (ins DPR:$orig, DPR:$tbl1, DPR:$tbl2, DPR:$tbl3, DPR:$src), NoItinerary,
2029 "vtbx.8\t$dst, \\{$tbl1,$tbl2,$tbl3\\}, $src", "$orig = $dst",
2030 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbx3 DPR:$orig, DPR:$tbl1,
2031 DPR:$tbl2, DPR:$tbl3, DPR:$src)))]>;
2033 : N3V<1,1,0b11,0b1011,1,0, (outs DPR:$dst), (ins DPR:$orig, DPR:$tbl1,
2034 DPR:$tbl2, DPR:$tbl3, DPR:$tbl4, DPR:$src), NoItinerary,
2035 "vtbx.8\t$dst, \\{$tbl1,$tbl2,$tbl3,$tbl4\\}, $src", "$orig = $dst",
2036 [(set DPR:$dst, (v8i8 (int_arm_neon_vtbx4 DPR:$orig, DPR:$tbl1,
2037 DPR:$tbl2, DPR:$tbl3, DPR:$tbl4, DPR:$src)))]>;
2039 //===----------------------------------------------------------------------===//
2040 // NEON instructions for single-precision FP math
2041 //===----------------------------------------------------------------------===//
2043 // These need separate instructions because they must use DPR_VFP2 register
2044 // class which have SPR sub-registers.
2046 // Vector Add Operations used for single-precision FP
2047 let neverHasSideEffects = 1 in
2048 def VADDfd_sfp : N3VDs<0, 0, 0b00, 0b1101, 0, "vadd.f32", v2f32, v2f32, fadd,1>;
2049 def : N3VDsPat<fadd, VADDfd_sfp>;
2051 // Vector Sub Operations used for single-precision FP
2052 let neverHasSideEffects = 1 in
2053 def VSUBfd_sfp : N3VDs<0, 0, 0b10, 0b1101, 0, "vsub.f32", v2f32, v2f32, fsub,0>;
2054 def : N3VDsPat<fsub, VSUBfd_sfp>;
2056 // Vector Multiply Operations used for single-precision FP
2057 let neverHasSideEffects = 1 in
2058 def VMULfd_sfp : N3VDs<1, 0, 0b00, 0b1101, 1, "vmul.f32", v2f32, v2f32, fmul,1>;
2059 def : N3VDsPat<fmul, VMULfd_sfp>;
2061 // Vector Multiply-Accumulate/Subtract used for single-precision FP
2062 let neverHasSideEffects = 1 in
2063 def VMLAfd_sfp : N3VDMulOps<0, 0, 0b00, 0b1101, 1, "vmla.f32", v2f32,fmul,fadd>;
2064 def : N3VDMulOpsPat<fmul, fadd, VMLAfd_sfp>;
2066 let neverHasSideEffects = 1 in
2067 def VMLSfd_sfp : N3VDMulOps<0, 0, 0b10, 0b1101, 1, "vmls.f32", v2f32,fmul,fsub>;
2068 def : N3VDMulOpsPat<fmul, fsub, VMLSfd_sfp>;
2070 // Vector Absolute used for single-precision FP
2071 let neverHasSideEffects = 1 in
2072 def VABSfd_sfp : N2VDInts<0b11, 0b11, 0b10, 0b01, 0b01110, 0, "vabs.f32",
2073 v2f32, v2f32, int_arm_neon_vabs>;
2074 def : N2VDIntsPat<fabs, VABSfd_sfp>;
2076 // Vector Negate used for single-precision FP
2077 let neverHasSideEffects = 1 in
2078 def VNEGf32d_sfp : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 0, 0,
2079 (outs DPR_VFP2:$dst), (ins DPR_VFP2:$src), NoItinerary,
2080 "vneg.f32\t$dst, $src", "", []>;
2081 def : N2VDIntsPat<fneg, VNEGf32d_sfp>;
2083 // Vector Convert between single-precision FP and integer
2084 let neverHasSideEffects = 1 in
2085 def VCVTf2sd_sfp : N2VDs<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt.s32.f32",
2086 v2i32, v2f32, fp_to_sint>;
2087 def : N2VDsPat<arm_ftosi, f32, v2f32, VCVTf2sd_sfp>;
2089 let neverHasSideEffects = 1 in
2090 def VCVTf2ud_sfp : N2VDs<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt.u32.f32",
2091 v2i32, v2f32, fp_to_uint>;
2092 def : N2VDsPat<arm_ftoui, f32, v2f32, VCVTf2ud_sfp>;
2094 let neverHasSideEffects = 1 in
2095 def VCVTs2fd_sfp : N2VDs<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt.f32.s32",
2096 v2f32, v2i32, sint_to_fp>;
2097 def : N2VDsPat<arm_sitof, f32, v2i32, VCVTs2fd_sfp>;
2099 let neverHasSideEffects = 1 in
2100 def VCVTu2fd_sfp : N2VDs<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt.f32.u32",
2101 v2f32, v2i32, uint_to_fp>;
2102 def : N2VDsPat<arm_uitof, f32, v2i32, VCVTu2fd_sfp>;
2104 //===----------------------------------------------------------------------===//
2105 // Non-Instruction Patterns
2106 //===----------------------------------------------------------------------===//
2109 def : Pat<(v1i64 (bitconvert (v2i32 DPR:$src))), (v1i64 DPR:$src)>;
2110 def : Pat<(v1i64 (bitconvert (v4i16 DPR:$src))), (v1i64 DPR:$src)>;
2111 def : Pat<(v1i64 (bitconvert (v8i8 DPR:$src))), (v1i64 DPR:$src)>;
2112 def : Pat<(v1i64 (bitconvert (f64 DPR:$src))), (v1i64 DPR:$src)>;
2113 def : Pat<(v1i64 (bitconvert (v2f32 DPR:$src))), (v1i64 DPR:$src)>;
2114 def : Pat<(v2i32 (bitconvert (v1i64 DPR:$src))), (v2i32 DPR:$src)>;
2115 def : Pat<(v2i32 (bitconvert (v4i16 DPR:$src))), (v2i32 DPR:$src)>;
2116 def : Pat<(v2i32 (bitconvert (v8i8 DPR:$src))), (v2i32 DPR:$src)>;
2117 def : Pat<(v2i32 (bitconvert (f64 DPR:$src))), (v2i32 DPR:$src)>;
2118 def : Pat<(v2i32 (bitconvert (v2f32 DPR:$src))), (v2i32 DPR:$src)>;
2119 def : Pat<(v4i16 (bitconvert (v1i64 DPR:$src))), (v4i16 DPR:$src)>;
2120 def : Pat<(v4i16 (bitconvert (v2i32 DPR:$src))), (v4i16 DPR:$src)>;
2121 def : Pat<(v4i16 (bitconvert (v8i8 DPR:$src))), (v4i16 DPR:$src)>;
2122 def : Pat<(v4i16 (bitconvert (f64 DPR:$src))), (v4i16 DPR:$src)>;
2123 def : Pat<(v4i16 (bitconvert (v2f32 DPR:$src))), (v4i16 DPR:$src)>;
2124 def : Pat<(v8i8 (bitconvert (v1i64 DPR:$src))), (v8i8 DPR:$src)>;
2125 def : Pat<(v8i8 (bitconvert (v2i32 DPR:$src))), (v8i8 DPR:$src)>;
2126 def : Pat<(v8i8 (bitconvert (v4i16 DPR:$src))), (v8i8 DPR:$src)>;
2127 def : Pat<(v8i8 (bitconvert (f64 DPR:$src))), (v8i8 DPR:$src)>;
2128 def : Pat<(v8i8 (bitconvert (v2f32 DPR:$src))), (v8i8 DPR:$src)>;
2129 def : Pat<(f64 (bitconvert (v1i64 DPR:$src))), (f64 DPR:$src)>;
2130 def : Pat<(f64 (bitconvert (v2i32 DPR:$src))), (f64 DPR:$src)>;
2131 def : Pat<(f64 (bitconvert (v4i16 DPR:$src))), (f64 DPR:$src)>;
2132 def : Pat<(f64 (bitconvert (v8i8 DPR:$src))), (f64 DPR:$src)>;
2133 def : Pat<(f64 (bitconvert (v2f32 DPR:$src))), (f64 DPR:$src)>;
2134 def : Pat<(v2f32 (bitconvert (f64 DPR:$src))), (v2f32 DPR:$src)>;
2135 def : Pat<(v2f32 (bitconvert (v1i64 DPR:$src))), (v2f32 DPR:$src)>;
2136 def : Pat<(v2f32 (bitconvert (v2i32 DPR:$src))), (v2f32 DPR:$src)>;
2137 def : Pat<(v2f32 (bitconvert (v4i16 DPR:$src))), (v2f32 DPR:$src)>;
2138 def : Pat<(v2f32 (bitconvert (v8i8 DPR:$src))), (v2f32 DPR:$src)>;
2140 def : Pat<(v2i64 (bitconvert (v4i32 QPR:$src))), (v2i64 QPR:$src)>;
2141 def : Pat<(v2i64 (bitconvert (v8i16 QPR:$src))), (v2i64 QPR:$src)>;
2142 def : Pat<(v2i64 (bitconvert (v16i8 QPR:$src))), (v2i64 QPR:$src)>;
2143 def : Pat<(v2i64 (bitconvert (v2f64 QPR:$src))), (v2i64 QPR:$src)>;
2144 def : Pat<(v2i64 (bitconvert (v4f32 QPR:$src))), (v2i64 QPR:$src)>;
2145 def : Pat<(v4i32 (bitconvert (v2i64 QPR:$src))), (v4i32 QPR:$src)>;
2146 def : Pat<(v4i32 (bitconvert (v8i16 QPR:$src))), (v4i32 QPR:$src)>;
2147 def : Pat<(v4i32 (bitconvert (v16i8 QPR:$src))), (v4i32 QPR:$src)>;
2148 def : Pat<(v4i32 (bitconvert (v2f64 QPR:$src))), (v4i32 QPR:$src)>;
2149 def : Pat<(v4i32 (bitconvert (v4f32 QPR:$src))), (v4i32 QPR:$src)>;
2150 def : Pat<(v8i16 (bitconvert (v2i64 QPR:$src))), (v8i16 QPR:$src)>;
2151 def : Pat<(v8i16 (bitconvert (v4i32 QPR:$src))), (v8i16 QPR:$src)>;
2152 def : Pat<(v8i16 (bitconvert (v16i8 QPR:$src))), (v8i16 QPR:$src)>;
2153 def : Pat<(v8i16 (bitconvert (v2f64 QPR:$src))), (v8i16 QPR:$src)>;
2154 def : Pat<(v8i16 (bitconvert (v4f32 QPR:$src))), (v8i16 QPR:$src)>;
2155 def : Pat<(v16i8 (bitconvert (v2i64 QPR:$src))), (v16i8 QPR:$src)>;
2156 def : Pat<(v16i8 (bitconvert (v4i32 QPR:$src))), (v16i8 QPR:$src)>;
2157 def : Pat<(v16i8 (bitconvert (v8i16 QPR:$src))), (v16i8 QPR:$src)>;
2158 def : Pat<(v16i8 (bitconvert (v2f64 QPR:$src))), (v16i8 QPR:$src)>;
2159 def : Pat<(v16i8 (bitconvert (v4f32 QPR:$src))), (v16i8 QPR:$src)>;
2160 def : Pat<(v4f32 (bitconvert (v2i64 QPR:$src))), (v4f32 QPR:$src)>;
2161 def : Pat<(v4f32 (bitconvert (v4i32 QPR:$src))), (v4f32 QPR:$src)>;
2162 def : Pat<(v4f32 (bitconvert (v8i16 QPR:$src))), (v4f32 QPR:$src)>;
2163 def : Pat<(v4f32 (bitconvert (v16i8 QPR:$src))), (v4f32 QPR:$src)>;
2164 def : Pat<(v4f32 (bitconvert (v2f64 QPR:$src))), (v4f32 QPR:$src)>;
2165 def : Pat<(v2f64 (bitconvert (v2i64 QPR:$src))), (v2f64 QPR:$src)>;
2166 def : Pat<(v2f64 (bitconvert (v4i32 QPR:$src))), (v2f64 QPR:$src)>;
2167 def : Pat<(v2f64 (bitconvert (v8i16 QPR:$src))), (v2f64 QPR:$src)>;
2168 def : Pat<(v2f64 (bitconvert (v16i8 QPR:$src))), (v2f64 QPR:$src)>;
2169 def : Pat<(v2f64 (bitconvert (v4f32 QPR:$src))), (v2f64 QPR:$src)>;