1 //===- ARMInstrNEON.td - NEON support for ARM -----------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the ARM NEON instruction set.
12 //===----------------------------------------------------------------------===//
14 //===----------------------------------------------------------------------===//
15 // NEON-specific DAG Nodes.
16 //===----------------------------------------------------------------------===//
18 def SDTARMVCMP : SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisSameAs<1, 2>]>;
20 def NEONvceq : SDNode<"ARMISD::VCEQ", SDTARMVCMP>;
21 def NEONvcge : SDNode<"ARMISD::VCGE", SDTARMVCMP>;
22 def NEONvcgeu : SDNode<"ARMISD::VCGEU", SDTARMVCMP>;
23 def NEONvcgt : SDNode<"ARMISD::VCGT", SDTARMVCMP>;
24 def NEONvcgtu : SDNode<"ARMISD::VCGTU", SDTARMVCMP>;
25 def NEONvtst : SDNode<"ARMISD::VTST", SDTARMVCMP>;
27 // Types for vector shift by immediates. The "SHX" version is for long and
28 // narrow operations where the source and destination vectors have different
29 // types. The "SHINS" version is for shift and insert operations.
30 def SDTARMVSH : SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisSameAs<0, 1>,
32 def SDTARMVSHX : SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisInt<1>,
34 def SDTARMVSHINS : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<0, 1>,
35 SDTCisSameAs<0, 2>, SDTCisVT<3, i32>]>;
37 def NEONvshl : SDNode<"ARMISD::VSHL", SDTARMVSH>;
38 def NEONvshrs : SDNode<"ARMISD::VSHRs", SDTARMVSH>;
39 def NEONvshru : SDNode<"ARMISD::VSHRu", SDTARMVSH>;
40 def NEONvshlls : SDNode<"ARMISD::VSHLLs", SDTARMVSHX>;
41 def NEONvshllu : SDNode<"ARMISD::VSHLLu", SDTARMVSHX>;
42 def NEONvshlli : SDNode<"ARMISD::VSHLLi", SDTARMVSHX>;
43 def NEONvshrn : SDNode<"ARMISD::VSHRN", SDTARMVSHX>;
45 def NEONvrshrs : SDNode<"ARMISD::VRSHRs", SDTARMVSH>;
46 def NEONvrshru : SDNode<"ARMISD::VRSHRu", SDTARMVSH>;
47 def NEONvrshrn : SDNode<"ARMISD::VRSHRN", SDTARMVSHX>;
49 def NEONvqshls : SDNode<"ARMISD::VQSHLs", SDTARMVSH>;
50 def NEONvqshlu : SDNode<"ARMISD::VQSHLu", SDTARMVSH>;
51 def NEONvqshlsu : SDNode<"ARMISD::VQSHLsu", SDTARMVSH>;
52 def NEONvqshrns : SDNode<"ARMISD::VQSHRNs", SDTARMVSHX>;
53 def NEONvqshrnu : SDNode<"ARMISD::VQSHRNu", SDTARMVSHX>;
54 def NEONvqshrnsu : SDNode<"ARMISD::VQSHRNsu", SDTARMVSHX>;
56 def NEONvqrshrns : SDNode<"ARMISD::VQRSHRNs", SDTARMVSHX>;
57 def NEONvqrshrnu : SDNode<"ARMISD::VQRSHRNu", SDTARMVSHX>;
58 def NEONvqrshrnsu : SDNode<"ARMISD::VQRSHRNsu", SDTARMVSHX>;
60 def NEONvsli : SDNode<"ARMISD::VSLI", SDTARMVSHINS>;
61 def NEONvsri : SDNode<"ARMISD::VSRI", SDTARMVSHINS>;
63 def SDTARMVGETLN : SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisInt<1>,
65 def NEONvgetlaneu : SDNode<"ARMISD::VGETLANEu", SDTARMVGETLN>;
66 def NEONvgetlanes : SDNode<"ARMISD::VGETLANEs", SDTARMVGETLN>;
68 def NEONvduplaneq : SDNode<"ARMISD::VDUPLANEQ",
69 SDTypeProfile<1, 2, [SDTCisVT<2, i32>]>>;
71 def SDTARMVLD2 : SDTypeProfile<2, 1, [SDTCisSameAs<0, 1>, SDTCisPtrTy<2>]>;
72 def SDTARMVLD3 : SDTypeProfile<3, 1, [SDTCisSameAs<0, 1>,
73 SDTCisSameAs<0, 2>, SDTCisPtrTy<3>]>;
74 def SDTARMVLD4 : SDTypeProfile<4, 1, [SDTCisSameAs<0, 1>,
76 SDTCisSameAs<0, 3>, SDTCisPtrTy<4>]>;
77 def NEONvld2d : SDNode<"ARMISD::VLD2D", SDTARMVLD2,
78 [SDNPHasChain, SDNPMayLoad]>;
79 def NEONvld3d : SDNode<"ARMISD::VLD3D", SDTARMVLD3,
80 [SDNPHasChain, SDNPMayLoad]>;
81 def NEONvld4d : SDNode<"ARMISD::VLD4D", SDTARMVLD4,
82 [SDNPHasChain, SDNPMayLoad]>;
84 //===----------------------------------------------------------------------===//
85 // NEON operand definitions
86 //===----------------------------------------------------------------------===//
88 // addrmode_neonldstm := reg
90 /* TODO: Take advantage of vldm.
91 def addrmode_neonldstm : Operand<i32>,
92 ComplexPattern<i32, 2, "SelectAddrModeNeonLdStM", []> {
93 let PrintMethod = "printAddrNeonLdStMOperand";
94 let MIOperandInfo = (ops GPR, i32imm);
98 //===----------------------------------------------------------------------===//
99 // NEON load / store instructions
100 //===----------------------------------------------------------------------===//
102 /* TODO: Take advantage of vldm.
104 def VLDMD : NI<(outs),
105 (ins addrmode_neonldstm:$addr, reglist:$dst1, variable_ops),
107 "vldm${addr:submode} ${addr:base}, $dst1",
109 let Inst{27-25} = 0b110;
111 let Inst{11-9} = 0b101;
114 def VLDMS : NI<(outs),
115 (ins addrmode_neonldstm:$addr, reglist:$dst1, variable_ops),
117 "vldm${addr:submode} ${addr:base}, $dst1",
119 let Inst{27-25} = 0b110;
121 let Inst{11-9} = 0b101;
126 // Use vldmia to load a Q register as a D register pair.
127 def VLDRQ : NI<(outs QPR:$dst), (ins GPR:$addr),
129 "vldmia $addr, ${dst:dregpair}",
130 [(set QPR:$dst, (v2f64 (load GPR:$addr)))]> {
131 let Inst{27-25} = 0b110;
132 let Inst{24} = 0; // P bit
133 let Inst{23} = 1; // U bit
135 let Inst{11-9} = 0b101;
138 // Use vstmia to store a Q register as a D register pair.
139 def VSTRQ : NI<(outs), (ins QPR:$src, GPR:$addr),
141 "vstmia $addr, ${src:dregpair}",
142 [(store (v2f64 QPR:$src), GPR:$addr)]> {
143 let Inst{27-25} = 0b110;
144 let Inst{24} = 0; // P bit
145 let Inst{23} = 1; // U bit
147 let Inst{11-9} = 0b101;
151 // VLD1 : Vector Load (multiple single elements)
152 class VLD1D<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
153 : NLdSt<(outs DPR:$dst), (ins addrmode6:$addr),
155 !strconcat(OpcodeStr, "\t\\{$dst\\}, $addr"),
156 [(set DPR:$dst, (Ty (IntOp addrmode6:$addr)))]>;
157 class VLD1Q<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
158 : NLdSt<(outs QPR:$dst), (ins addrmode6:$addr),
160 !strconcat(OpcodeStr, "\t${dst:dregpair}, $addr"),
161 [(set QPR:$dst, (Ty (IntOp addrmode6:$addr)))]>;
163 def VLD1d8 : VLD1D<"vld1.8", v8i8, int_arm_neon_vld1i>;
164 def VLD1d16 : VLD1D<"vld1.16", v4i16, int_arm_neon_vld1i>;
165 def VLD1d32 : VLD1D<"vld1.32", v2i32, int_arm_neon_vld1i>;
166 def VLD1df : VLD1D<"vld1.32", v2f32, int_arm_neon_vld1f>;
167 def VLD1d64 : VLD1D<"vld1.64", v1i64, int_arm_neon_vld1i>;
169 def VLD1q8 : VLD1Q<"vld1.8", v16i8, int_arm_neon_vld1i>;
170 def VLD1q16 : VLD1Q<"vld1.16", v8i16, int_arm_neon_vld1i>;
171 def VLD1q32 : VLD1Q<"vld1.32", v4i32, int_arm_neon_vld1i>;
172 def VLD1qf : VLD1Q<"vld1.32", v4f32, int_arm_neon_vld1f>;
173 def VLD1q64 : VLD1Q<"vld1.64", v2i64, int_arm_neon_vld1i>;
175 // VST1 : Vector Store (multiple single elements)
176 class VST1D<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
177 : NLdSt<(outs), (ins addrmode6:$addr, DPR:$src),
179 !strconcat(OpcodeStr, "\t\\{$src\\}, $addr"),
180 [(IntOp addrmode6:$addr, (Ty DPR:$src))]>;
181 class VST1Q<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
182 : NLdSt<(outs), (ins addrmode6:$addr, QPR:$src),
184 !strconcat(OpcodeStr, "\t${src:dregpair}, $addr"),
185 [(IntOp addrmode6:$addr, (Ty QPR:$src))]>;
187 def VST1d8 : VST1D<"vst1.8", v8i8, int_arm_neon_vst1i>;
188 def VST1d16 : VST1D<"vst1.16", v4i16, int_arm_neon_vst1i>;
189 def VST1d32 : VST1D<"vst1.32", v2i32, int_arm_neon_vst1i>;
190 def VST1df : VST1D<"vst1.32", v2f32, int_arm_neon_vst1f>;
191 def VST1d64 : VST1D<"vst1.64", v1i64, int_arm_neon_vst1i>;
193 def VST1q8 : VST1Q<"vst1.8", v16i8, int_arm_neon_vst1i>;
194 def VST1q16 : VST1Q<"vst1.16", v8i16, int_arm_neon_vst1i>;
195 def VST1q32 : VST1Q<"vst1.32", v4i32, int_arm_neon_vst1i>;
196 def VST1qf : VST1Q<"vst1.32", v4f32, int_arm_neon_vst1f>;
197 def VST1q64 : VST1Q<"vst1.64", v2i64, int_arm_neon_vst1i>;
199 // VLD2 : Vector Load (multiple 2-element structures)
200 class VLD2D<string OpcodeStr>
201 : NLdSt<(outs DPR:$dst1, DPR:$dst2), (ins addrmode6:$addr),
203 !strconcat(OpcodeStr, "\t\\{$dst1,$dst2\\}, $addr"), []>;
205 def VLD2d8 : VLD2D<"vld2.8">;
206 def VLD2d16 : VLD2D<"vld2.16">;
207 def VLD2d32 : VLD2D<"vld2.32">;
209 // VLD3 : Vector Load (multiple 3-element structures)
210 class VLD3D<string OpcodeStr>
211 : NLdSt<(outs DPR:$dst1, DPR:$dst2, DPR:$dst3), (ins addrmode6:$addr),
213 !strconcat(OpcodeStr, "\t\\{$dst1,$dst2,$dst3\\}, $addr"), []>;
215 def VLD3d8 : VLD3D<"vld3.8">;
216 def VLD3d16 : VLD3D<"vld3.16">;
217 def VLD3d32 : VLD3D<"vld3.32">;
219 // VLD4 : Vector Load (multiple 4-element structures)
220 class VLD4D<string OpcodeStr>
221 : NLdSt<(outs DPR:$dst1, DPR:$dst2, DPR:$dst3, DPR:$dst4),
222 (ins addrmode6:$addr),
224 !strconcat(OpcodeStr, "\t\\{$dst1,$dst2,$dst3,$dst4\\}, $addr"), []>;
226 def VLD4d8 : VLD4D<"vld4.8">;
227 def VLD4d16 : VLD4D<"vld4.16">;
228 def VLD4d32 : VLD4D<"vld4.32">;
231 //===----------------------------------------------------------------------===//
232 // NEON pattern fragments
233 //===----------------------------------------------------------------------===//
235 // Extract D sub-registers of Q registers.
236 // (arm_dsubreg_0 is 5; arm_dsubreg_1 is 6)
237 def SubReg_i8_reg : SDNodeXForm<imm, [{
238 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 8, MVT::i32);
240 def SubReg_i16_reg : SDNodeXForm<imm, [{
241 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 4, MVT::i32);
243 def SubReg_i32_reg : SDNodeXForm<imm, [{
244 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 2, MVT::i32);
246 def SubReg_f64_reg : SDNodeXForm<imm, [{
247 return CurDAG->getTargetConstant(5 + N->getZExtValue(), MVT::i32);
250 // Translate lane numbers from Q registers to D subregs.
251 def SubReg_i8_lane : SDNodeXForm<imm, [{
252 return CurDAG->getTargetConstant(N->getZExtValue() & 7, MVT::i32);
254 def SubReg_i16_lane : SDNodeXForm<imm, [{
255 return CurDAG->getTargetConstant(N->getZExtValue() & 3, MVT::i32);
257 def SubReg_i32_lane : SDNodeXForm<imm, [{
258 return CurDAG->getTargetConstant(N->getZExtValue() & 1, MVT::i32);
261 //===----------------------------------------------------------------------===//
262 // Instruction Classes
263 //===----------------------------------------------------------------------===//
265 // Basic 2-register operations, both double- and quad-register.
266 class N2VD<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
267 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
268 ValueType ResTy, ValueType OpTy, SDNode OpNode>
269 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
270 (ins DPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
271 [(set DPR:$dst, (ResTy (OpNode (OpTy DPR:$src))))]>;
272 class N2VQ<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
273 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
274 ValueType ResTy, ValueType OpTy, SDNode OpNode>
275 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
276 (ins QPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
277 [(set QPR:$dst, (ResTy (OpNode (OpTy QPR:$src))))]>;
279 // Basic 2-register intrinsics, both double- and quad-register.
280 class N2VDInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
281 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
282 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
283 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
284 (ins DPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
285 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src))))]>;
286 class N2VQInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
287 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
288 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
289 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
290 (ins QPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
291 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src))))]>;
293 // Basic 2-register operations, scalar single-precision
294 class N2VDInts<SDNode OpNode, NeonI Inst>
295 : NEONFPPat<(f32 (OpNode SPR:$a)),
296 (EXTRACT_SUBREG (COPY_TO_REGCLASS
297 (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)),
298 SPR:$a, arm_ssubreg_0)),
302 // Narrow 2-register intrinsics.
303 class N2VNInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
304 bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
305 string OpcodeStr, ValueType TyD, ValueType TyQ, Intrinsic IntOp>
306 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, op6, op4, (outs DPR:$dst),
307 (ins QPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
308 [(set DPR:$dst, (TyD (IntOp (TyQ QPR:$src))))]>;
310 // Long 2-register intrinsics. (This is currently only used for VMOVL and is
311 // derived from N2VImm instead of N2V because of the way the size is encoded.)
312 class N2VLInt<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
313 bit op6, bit op4, string OpcodeStr, ValueType TyQ, ValueType TyD,
315 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4, (outs QPR:$dst),
316 (ins DPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
317 [(set QPR:$dst, (TyQ (IntOp (TyD DPR:$src))))]>;
319 // Basic 3-register operations, both double- and quad-register.
320 class N3VD<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
321 string OpcodeStr, ValueType ResTy, ValueType OpTy,
322 SDNode OpNode, bit Commutable>
323 : N3V<op24, op23, op21_20, op11_8, 0, op4,
324 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2), NoItinerary,
325 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
326 [(set DPR:$dst, (ResTy (OpNode (OpTy DPR:$src1), (OpTy DPR:$src2))))]> {
327 let isCommutable = Commutable;
329 class N3VQ<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
330 string OpcodeStr, ValueType ResTy, ValueType OpTy,
331 SDNode OpNode, bit Commutable>
332 : N3V<op24, op23, op21_20, op11_8, 1, op4,
333 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2), NoItinerary,
334 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
335 [(set QPR:$dst, (ResTy (OpNode (OpTy QPR:$src1), (OpTy QPR:$src2))))]> {
336 let isCommutable = Commutable;
339 // Basic 3-register operations, scalar single-precision
340 class N3VDs<SDNode OpNode, NeonI Inst>
341 : NEONFPPat<(f32 (OpNode SPR:$a, SPR:$b)),
342 (EXTRACT_SUBREG (COPY_TO_REGCLASS
343 (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)),
344 SPR:$a, arm_ssubreg_0),
345 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)),
346 SPR:$b, arm_ssubreg_0)),
350 // Basic 3-register intrinsics, both double- and quad-register.
351 class N3VDInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
352 string OpcodeStr, ValueType ResTy, ValueType OpTy,
353 Intrinsic IntOp, bit Commutable>
354 : N3V<op24, op23, op21_20, op11_8, 0, op4,
355 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2), NoItinerary,
356 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
357 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src1), (OpTy DPR:$src2))))]> {
358 let isCommutable = Commutable;
360 class N3VQInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
361 string OpcodeStr, ValueType ResTy, ValueType OpTy,
362 Intrinsic IntOp, bit Commutable>
363 : N3V<op24, op23, op21_20, op11_8, 1, op4,
364 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2), NoItinerary,
365 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
366 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src1), (OpTy QPR:$src2))))]> {
367 let isCommutable = Commutable;
370 // Multiply-Add/Sub operations, both double- and quad-register.
371 class N3VDMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
372 string OpcodeStr, ValueType Ty, SDNode MulOp, SDNode OpNode>
373 : N3V<op24, op23, op21_20, op11_8, 0, op4,
374 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, DPR:$src3), NoItinerary,
375 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
376 [(set DPR:$dst, (Ty (OpNode DPR:$src1,
377 (Ty (MulOp DPR:$src2, DPR:$src3)))))]>;
378 class N3VQMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
379 string OpcodeStr, ValueType Ty, SDNode MulOp, SDNode OpNode>
380 : N3V<op24, op23, op21_20, op11_8, 1, op4,
381 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, QPR:$src3), NoItinerary,
382 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
383 [(set QPR:$dst, (Ty (OpNode QPR:$src1,
384 (Ty (MulOp QPR:$src2, QPR:$src3)))))]>;
386 // Multiply-Add/Sub operations, scalar single-precision
387 class N3VDMulOps<SDNode MulNode, SDNode OpNode, NeonI Inst>
388 : NEONFPPat<(f32 (OpNode SPR:$acc,
389 (f32 (MulNode SPR:$a, SPR:$b)))),
390 (EXTRACT_SUBREG (COPY_TO_REGCLASS
391 (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)),
392 SPR:$acc, arm_ssubreg_0),
393 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)),
394 SPR:$a, arm_ssubreg_0),
395 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)),
396 SPR:$b, arm_ssubreg_0)),
400 // Neon 3-argument intrinsics, both double- and quad-register.
401 // The destination register is also used as the first source operand register.
402 class N3VDInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
403 string OpcodeStr, ValueType ResTy, ValueType OpTy,
405 : N3V<op24, op23, op21_20, op11_8, 0, op4,
406 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, DPR:$src3), NoItinerary,
407 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
408 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src1),
409 (OpTy DPR:$src2), (OpTy DPR:$src3))))]>;
410 class N3VQInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
411 string OpcodeStr, ValueType ResTy, ValueType OpTy,
413 : N3V<op24, op23, op21_20, op11_8, 1, op4,
414 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, QPR:$src3), NoItinerary,
415 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
416 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src1),
417 (OpTy QPR:$src2), (OpTy QPR:$src3))))]>;
419 // Neon Long 3-argument intrinsic. The destination register is
420 // a quad-register and is also used as the first source operand register.
421 class N3VLInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
422 string OpcodeStr, ValueType TyQ, ValueType TyD, Intrinsic IntOp>
423 : N3V<op24, op23, op21_20, op11_8, 0, op4,
424 (outs QPR:$dst), (ins QPR:$src1, DPR:$src2, DPR:$src3), NoItinerary,
425 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
427 (TyQ (IntOp (TyQ QPR:$src1), (TyD DPR:$src2), (TyD DPR:$src3))))]>;
429 // Narrowing 3-register intrinsics.
430 class N3VNInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
431 string OpcodeStr, ValueType TyD, ValueType TyQ,
432 Intrinsic IntOp, bit Commutable>
433 : N3V<op24, op23, op21_20, op11_8, 0, op4,
434 (outs DPR:$dst), (ins QPR:$src1, QPR:$src2), NoItinerary,
435 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
436 [(set DPR:$dst, (TyD (IntOp (TyQ QPR:$src1), (TyQ QPR:$src2))))]> {
437 let isCommutable = Commutable;
440 // Long 3-register intrinsics.
441 class N3VLInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
442 string OpcodeStr, ValueType TyQ, ValueType TyD,
443 Intrinsic IntOp, bit Commutable>
444 : N3V<op24, op23, op21_20, op11_8, 0, op4,
445 (outs QPR:$dst), (ins DPR:$src1, DPR:$src2), NoItinerary,
446 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
447 [(set QPR:$dst, (TyQ (IntOp (TyD DPR:$src1), (TyD DPR:$src2))))]> {
448 let isCommutable = Commutable;
451 // Wide 3-register intrinsics.
452 class N3VWInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
453 string OpcodeStr, ValueType TyQ, ValueType TyD,
454 Intrinsic IntOp, bit Commutable>
455 : N3V<op24, op23, op21_20, op11_8, 0, op4,
456 (outs QPR:$dst), (ins QPR:$src1, DPR:$src2), NoItinerary,
457 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
458 [(set QPR:$dst, (TyQ (IntOp (TyQ QPR:$src1), (TyD DPR:$src2))))]> {
459 let isCommutable = Commutable;
462 // Pairwise long 2-register intrinsics, both double- and quad-register.
463 class N2VDPLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
464 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
465 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
466 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
467 (ins DPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
468 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src))))]>;
469 class N2VQPLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
470 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
471 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
472 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
473 (ins QPR:$src), NoItinerary, !strconcat(OpcodeStr, "\t$dst, $src"), "",
474 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src))))]>;
476 // Pairwise long 2-register accumulate intrinsics,
477 // both double- and quad-register.
478 // The destination register is also used as the first source operand register.
479 class N2VDPLInt2<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
480 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
481 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
482 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4,
483 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2), NoItinerary,
484 !strconcat(OpcodeStr, "\t$dst, $src2"), "$src1 = $dst",
485 [(set DPR:$dst, (ResTy (IntOp (ResTy DPR:$src1), (OpTy DPR:$src2))))]>;
486 class N2VQPLInt2<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
487 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
488 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
489 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4,
490 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2), NoItinerary,
491 !strconcat(OpcodeStr, "\t$dst, $src2"), "$src1 = $dst",
492 [(set QPR:$dst, (ResTy (IntOp (ResTy QPR:$src1), (OpTy QPR:$src2))))]>;
494 // Shift by immediate,
495 // both double- and quad-register.
496 class N2VDSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
497 bit op4, string OpcodeStr, ValueType Ty, SDNode OpNode>
498 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
499 (outs DPR:$dst), (ins DPR:$src, i32imm:$SIMM), NoItinerary,
500 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
501 [(set DPR:$dst, (Ty (OpNode (Ty DPR:$src), (i32 imm:$SIMM))))]>;
502 class N2VQSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
503 bit op4, string OpcodeStr, ValueType Ty, SDNode OpNode>
504 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
505 (outs QPR:$dst), (ins QPR:$src, i32imm:$SIMM), NoItinerary,
506 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
507 [(set QPR:$dst, (Ty (OpNode (Ty QPR:$src), (i32 imm:$SIMM))))]>;
509 // Long shift by immediate.
510 class N2VLSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
511 bit op6, bit op4, string OpcodeStr, ValueType ResTy,
512 ValueType OpTy, SDNode OpNode>
513 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4,
514 (outs QPR:$dst), (ins DPR:$src, i32imm:$SIMM), NoItinerary,
515 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
516 [(set QPR:$dst, (ResTy (OpNode (OpTy DPR:$src),
517 (i32 imm:$SIMM))))]>;
519 // Narrow shift by immediate.
520 class N2VNSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
521 bit op6, bit op4, string OpcodeStr, ValueType ResTy,
522 ValueType OpTy, SDNode OpNode>
523 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4,
524 (outs DPR:$dst), (ins QPR:$src, i32imm:$SIMM), NoItinerary,
525 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
526 [(set DPR:$dst, (ResTy (OpNode (OpTy QPR:$src),
527 (i32 imm:$SIMM))))]>;
529 // Shift right by immediate and accumulate,
530 // both double- and quad-register.
531 class N2VDShAdd<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
532 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
533 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
534 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, i32imm:$SIMM),
536 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
537 [(set DPR:$dst, (Ty (add DPR:$src1,
538 (Ty (ShOp DPR:$src2, (i32 imm:$SIMM))))))]>;
539 class N2VQShAdd<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
540 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
541 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
542 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, i32imm:$SIMM),
544 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
545 [(set QPR:$dst, (Ty (add QPR:$src1,
546 (Ty (ShOp QPR:$src2, (i32 imm:$SIMM))))))]>;
548 // Shift by immediate and insert,
549 // both double- and quad-register.
550 class N2VDShIns<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
551 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
552 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
553 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, i32imm:$SIMM),
555 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
556 [(set DPR:$dst, (Ty (ShOp DPR:$src1, DPR:$src2, (i32 imm:$SIMM))))]>;
557 class N2VQShIns<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
558 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
559 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
560 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, i32imm:$SIMM),
562 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
563 [(set QPR:$dst, (Ty (ShOp QPR:$src1, QPR:$src2, (i32 imm:$SIMM))))]>;
565 // Convert, with fractional bits immediate,
566 // both double- and quad-register.
567 class N2VCvtD<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
568 bit op4, string OpcodeStr, ValueType ResTy, ValueType OpTy,
570 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
571 (outs DPR:$dst), (ins DPR:$src, i32imm:$SIMM), NoItinerary,
572 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
573 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src), (i32 imm:$SIMM))))]>;
574 class N2VCvtQ<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
575 bit op4, string OpcodeStr, ValueType ResTy, ValueType OpTy,
577 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
578 (outs QPR:$dst), (ins QPR:$src, i32imm:$SIMM), NoItinerary,
579 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
580 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src), (i32 imm:$SIMM))))]>;
582 //===----------------------------------------------------------------------===//
584 //===----------------------------------------------------------------------===//
586 // Neon 3-register vector operations.
588 // First with only element sizes of 8, 16 and 32 bits:
589 multiclass N3V_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
590 string OpcodeStr, SDNode OpNode, bit Commutable = 0> {
591 // 64-bit vector types.
592 def v8i8 : N3VD<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
593 v8i8, v8i8, OpNode, Commutable>;
594 def v4i16 : N3VD<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr, "16"),
595 v4i16, v4i16, OpNode, Commutable>;
596 def v2i32 : N3VD<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr, "32"),
597 v2i32, v2i32, OpNode, Commutable>;
599 // 128-bit vector types.
600 def v16i8 : N3VQ<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
601 v16i8, v16i8, OpNode, Commutable>;
602 def v8i16 : N3VQ<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr, "16"),
603 v8i16, v8i16, OpNode, Commutable>;
604 def v4i32 : N3VQ<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr, "32"),
605 v4i32, v4i32, OpNode, Commutable>;
608 // ....then also with element size 64 bits:
609 multiclass N3V_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
610 string OpcodeStr, SDNode OpNode, bit Commutable = 0>
611 : N3V_QHS<op24, op23, op11_8, op4, OpcodeStr, OpNode, Commutable> {
612 def v1i64 : N3VD<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr, "64"),
613 v1i64, v1i64, OpNode, Commutable>;
614 def v2i64 : N3VQ<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr, "64"),
615 v2i64, v2i64, OpNode, Commutable>;
619 // Neon Narrowing 2-register vector intrinsics,
620 // source operand element sizes of 16, 32 and 64 bits:
621 multiclass N2VNInt_HSD<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
622 bits<5> op11_7, bit op6, bit op4, string OpcodeStr,
624 def v8i8 : N2VNInt<op24_23, op21_20, 0b00, op17_16, op11_7, op6, op4,
625 !strconcat(OpcodeStr, "16"), v8i8, v8i16, IntOp>;
626 def v4i16 : N2VNInt<op24_23, op21_20, 0b01, op17_16, op11_7, op6, op4,
627 !strconcat(OpcodeStr, "32"), v4i16, v4i32, IntOp>;
628 def v2i32 : N2VNInt<op24_23, op21_20, 0b10, op17_16, op11_7, op6, op4,
629 !strconcat(OpcodeStr, "64"), v2i32, v2i64, IntOp>;
633 // Neon Lengthening 2-register vector intrinsic (currently specific to VMOVL).
634 // source operand element sizes of 16, 32 and 64 bits:
635 multiclass N2VLInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op7, bit op6,
636 bit op4, string OpcodeStr, Intrinsic IntOp> {
637 def v8i16 : N2VLInt<op24, op23, 0b001000, op11_8, op7, op6, op4,
638 !strconcat(OpcodeStr, "8"), v8i16, v8i8, IntOp>;
639 def v4i32 : N2VLInt<op24, op23, 0b010000, op11_8, op7, op6, op4,
640 !strconcat(OpcodeStr, "16"), v4i32, v4i16, IntOp>;
641 def v2i64 : N2VLInt<op24, op23, 0b100000, op11_8, op7, op6, op4,
642 !strconcat(OpcodeStr, "32"), v2i64, v2i32, IntOp>;
646 // Neon 3-register vector intrinsics.
648 // First with only element sizes of 16 and 32 bits:
649 multiclass N3VInt_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
650 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
651 // 64-bit vector types.
652 def v4i16 : N3VDInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
653 v4i16, v4i16, IntOp, Commutable>;
654 def v2i32 : N3VDInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
655 v2i32, v2i32, IntOp, Commutable>;
657 // 128-bit vector types.
658 def v8i16 : N3VQInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
659 v8i16, v8i16, IntOp, Commutable>;
660 def v4i32 : N3VQInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
661 v4i32, v4i32, IntOp, Commutable>;
664 // ....then also with element size of 8 bits:
665 multiclass N3VInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
666 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
667 : N3VInt_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
668 def v8i8 : N3VDInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
669 v8i8, v8i8, IntOp, Commutable>;
670 def v16i8 : N3VQInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
671 v16i8, v16i8, IntOp, Commutable>;
674 // ....then also with element size of 64 bits:
675 multiclass N3VInt_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
676 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
677 : N3VInt_QHS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
678 def v1i64 : N3VDInt<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr,"64"),
679 v1i64, v1i64, IntOp, Commutable>;
680 def v2i64 : N3VQInt<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr,"64"),
681 v2i64, v2i64, IntOp, Commutable>;
685 // Neon Narrowing 3-register vector intrinsics,
686 // source operand element sizes of 16, 32 and 64 bits:
687 multiclass N3VNInt_HSD<bit op24, bit op23, bits<4> op11_8, bit op4,
688 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
689 def v8i8 : N3VNInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr,"16"),
690 v8i8, v8i16, IntOp, Commutable>;
691 def v4i16 : N3VNInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"32"),
692 v4i16, v4i32, IntOp, Commutable>;
693 def v2i32 : N3VNInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"64"),
694 v2i32, v2i64, IntOp, Commutable>;
698 // Neon Long 3-register vector intrinsics.
700 // First with only element sizes of 16 and 32 bits:
701 multiclass N3VLInt_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
702 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
703 def v4i32 : N3VLInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
704 v4i32, v4i16, IntOp, Commutable>;
705 def v2i64 : N3VLInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
706 v2i64, v2i32, IntOp, Commutable>;
709 // ....then also with element size of 8 bits:
710 multiclass N3VLInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
711 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
712 : N3VLInt_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
713 def v8i16 : N3VLInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
714 v8i16, v8i8, IntOp, Commutable>;
718 // Neon Wide 3-register vector intrinsics,
719 // source operand element sizes of 8, 16 and 32 bits:
720 multiclass N3VWInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
721 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
722 def v8i16 : N3VWInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
723 v8i16, v8i8, IntOp, Commutable>;
724 def v4i32 : N3VWInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
725 v4i32, v4i16, IntOp, Commutable>;
726 def v2i64 : N3VWInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
727 v2i64, v2i32, IntOp, Commutable>;
731 // Neon Multiply-Op vector operations,
732 // element sizes of 8, 16 and 32 bits:
733 multiclass N3VMulOp_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
734 string OpcodeStr, SDNode OpNode> {
735 // 64-bit vector types.
736 def v8i8 : N3VDMulOp<op24, op23, 0b00, op11_8, op4,
737 !strconcat(OpcodeStr, "8"), v8i8, mul, OpNode>;
738 def v4i16 : N3VDMulOp<op24, op23, 0b01, op11_8, op4,
739 !strconcat(OpcodeStr, "16"), v4i16, mul, OpNode>;
740 def v2i32 : N3VDMulOp<op24, op23, 0b10, op11_8, op4,
741 !strconcat(OpcodeStr, "32"), v2i32, mul, OpNode>;
743 // 128-bit vector types.
744 def v16i8 : N3VQMulOp<op24, op23, 0b00, op11_8, op4,
745 !strconcat(OpcodeStr, "8"), v16i8, mul, OpNode>;
746 def v8i16 : N3VQMulOp<op24, op23, 0b01, op11_8, op4,
747 !strconcat(OpcodeStr, "16"), v8i16, mul, OpNode>;
748 def v4i32 : N3VQMulOp<op24, op23, 0b10, op11_8, op4,
749 !strconcat(OpcodeStr, "32"), v4i32, mul, OpNode>;
753 // Neon 3-argument intrinsics,
754 // element sizes of 8, 16 and 32 bits:
755 multiclass N3VInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
756 string OpcodeStr, Intrinsic IntOp> {
757 // 64-bit vector types.
758 def v8i8 : N3VDInt3<op24, op23, 0b00, op11_8, op4,
759 !strconcat(OpcodeStr, "8"), v8i8, v8i8, IntOp>;
760 def v4i16 : N3VDInt3<op24, op23, 0b01, op11_8, op4,
761 !strconcat(OpcodeStr, "16"), v4i16, v4i16, IntOp>;
762 def v2i32 : N3VDInt3<op24, op23, 0b10, op11_8, op4,
763 !strconcat(OpcodeStr, "32"), v2i32, v2i32, IntOp>;
765 // 128-bit vector types.
766 def v16i8 : N3VQInt3<op24, op23, 0b00, op11_8, op4,
767 !strconcat(OpcodeStr, "8"), v16i8, v16i8, IntOp>;
768 def v8i16 : N3VQInt3<op24, op23, 0b01, op11_8, op4,
769 !strconcat(OpcodeStr, "16"), v8i16, v8i16, IntOp>;
770 def v4i32 : N3VQInt3<op24, op23, 0b10, op11_8, op4,
771 !strconcat(OpcodeStr, "32"), v4i32, v4i32, IntOp>;
775 // Neon Long 3-argument intrinsics.
777 // First with only element sizes of 16 and 32 bits:
778 multiclass N3VLInt3_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
779 string OpcodeStr, Intrinsic IntOp> {
780 def v4i32 : N3VLInt3<op24, op23, 0b01, op11_8, op4,
781 !strconcat(OpcodeStr, "16"), v4i32, v4i16, IntOp>;
782 def v2i64 : N3VLInt3<op24, op23, 0b10, op11_8, op4,
783 !strconcat(OpcodeStr, "32"), v2i64, v2i32, IntOp>;
786 // ....then also with element size of 8 bits:
787 multiclass N3VLInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
788 string OpcodeStr, Intrinsic IntOp>
789 : N3VLInt3_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp> {
790 def v8i16 : N3VLInt3<op24, op23, 0b01, op11_8, op4,
791 !strconcat(OpcodeStr, "8"), v8i16, v8i8, IntOp>;
795 // Neon 2-register vector intrinsics,
796 // element sizes of 8, 16 and 32 bits:
797 multiclass N2VInt_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
798 bits<5> op11_7, bit op4, string OpcodeStr,
800 // 64-bit vector types.
801 def v8i8 : N2VDInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
802 !strconcat(OpcodeStr, "8"), v8i8, v8i8, IntOp>;
803 def v4i16 : N2VDInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
804 !strconcat(OpcodeStr, "16"), v4i16, v4i16, IntOp>;
805 def v2i32 : N2VDInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
806 !strconcat(OpcodeStr, "32"), v2i32, v2i32, IntOp>;
808 // 128-bit vector types.
809 def v16i8 : N2VQInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
810 !strconcat(OpcodeStr, "8"), v16i8, v16i8, IntOp>;
811 def v8i16 : N2VQInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
812 !strconcat(OpcodeStr, "16"), v8i16, v8i16, IntOp>;
813 def v4i32 : N2VQInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
814 !strconcat(OpcodeStr, "32"), v4i32, v4i32, IntOp>;
818 // Neon Pairwise long 2-register intrinsics,
819 // element sizes of 8, 16 and 32 bits:
820 multiclass N2VPLInt_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
821 bits<5> op11_7, bit op4,
822 string OpcodeStr, Intrinsic IntOp> {
823 // 64-bit vector types.
824 def v8i8 : N2VDPLInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
825 !strconcat(OpcodeStr, "8"), v4i16, v8i8, IntOp>;
826 def v4i16 : N2VDPLInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
827 !strconcat(OpcodeStr, "16"), v2i32, v4i16, IntOp>;
828 def v2i32 : N2VDPLInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
829 !strconcat(OpcodeStr, "32"), v1i64, v2i32, IntOp>;
831 // 128-bit vector types.
832 def v16i8 : N2VQPLInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
833 !strconcat(OpcodeStr, "8"), v8i16, v16i8, IntOp>;
834 def v8i16 : N2VQPLInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
835 !strconcat(OpcodeStr, "16"), v4i32, v8i16, IntOp>;
836 def v4i32 : N2VQPLInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
837 !strconcat(OpcodeStr, "32"), v2i64, v4i32, IntOp>;
841 // Neon Pairwise long 2-register accumulate intrinsics,
842 // element sizes of 8, 16 and 32 bits:
843 multiclass N2VPLInt2_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
844 bits<5> op11_7, bit op4,
845 string OpcodeStr, Intrinsic IntOp> {
846 // 64-bit vector types.
847 def v8i8 : N2VDPLInt2<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
848 !strconcat(OpcodeStr, "8"), v4i16, v8i8, IntOp>;
849 def v4i16 : N2VDPLInt2<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
850 !strconcat(OpcodeStr, "16"), v2i32, v4i16, IntOp>;
851 def v2i32 : N2VDPLInt2<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
852 !strconcat(OpcodeStr, "32"), v1i64, v2i32, IntOp>;
854 // 128-bit vector types.
855 def v16i8 : N2VQPLInt2<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
856 !strconcat(OpcodeStr, "8"), v8i16, v16i8, IntOp>;
857 def v8i16 : N2VQPLInt2<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
858 !strconcat(OpcodeStr, "16"), v4i32, v8i16, IntOp>;
859 def v4i32 : N2VQPLInt2<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
860 !strconcat(OpcodeStr, "32"), v2i64, v4i32, IntOp>;
864 // Neon 2-register vector shift by immediate,
865 // element sizes of 8, 16, 32 and 64 bits:
866 multiclass N2VSh_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
867 string OpcodeStr, SDNode OpNode> {
868 // 64-bit vector types.
869 def v8i8 : N2VDSh<op24, op23, 0b001000, op11_8, 0, op4,
870 !strconcat(OpcodeStr, "8"), v8i8, OpNode>;
871 def v4i16 : N2VDSh<op24, op23, 0b010000, op11_8, 0, op4,
872 !strconcat(OpcodeStr, "16"), v4i16, OpNode>;
873 def v2i32 : N2VDSh<op24, op23, 0b100000, op11_8, 0, op4,
874 !strconcat(OpcodeStr, "32"), v2i32, OpNode>;
875 def v1i64 : N2VDSh<op24, op23, 0b000000, op11_8, 1, op4,
876 !strconcat(OpcodeStr, "64"), v1i64, OpNode>;
878 // 128-bit vector types.
879 def v16i8 : N2VQSh<op24, op23, 0b001000, op11_8, 0, op4,
880 !strconcat(OpcodeStr, "8"), v16i8, OpNode>;
881 def v8i16 : N2VQSh<op24, op23, 0b010000, op11_8, 0, op4,
882 !strconcat(OpcodeStr, "16"), v8i16, OpNode>;
883 def v4i32 : N2VQSh<op24, op23, 0b100000, op11_8, 0, op4,
884 !strconcat(OpcodeStr, "32"), v4i32, OpNode>;
885 def v2i64 : N2VQSh<op24, op23, 0b000000, op11_8, 1, op4,
886 !strconcat(OpcodeStr, "64"), v2i64, OpNode>;
890 // Neon Shift-Accumulate vector operations,
891 // element sizes of 8, 16, 32 and 64 bits:
892 multiclass N2VShAdd_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
893 string OpcodeStr, SDNode ShOp> {
894 // 64-bit vector types.
895 def v8i8 : N2VDShAdd<op24, op23, 0b001000, op11_8, 0, op4,
896 !strconcat(OpcodeStr, "8"), v8i8, ShOp>;
897 def v4i16 : N2VDShAdd<op24, op23, 0b010000, op11_8, 0, op4,
898 !strconcat(OpcodeStr, "16"), v4i16, ShOp>;
899 def v2i32 : N2VDShAdd<op24, op23, 0b100000, op11_8, 0, op4,
900 !strconcat(OpcodeStr, "32"), v2i32, ShOp>;
901 def v1i64 : N2VDShAdd<op24, op23, 0b000000, op11_8, 1, op4,
902 !strconcat(OpcodeStr, "64"), v1i64, ShOp>;
904 // 128-bit vector types.
905 def v16i8 : N2VQShAdd<op24, op23, 0b001000, op11_8, 0, op4,
906 !strconcat(OpcodeStr, "8"), v16i8, ShOp>;
907 def v8i16 : N2VQShAdd<op24, op23, 0b010000, op11_8, 0, op4,
908 !strconcat(OpcodeStr, "16"), v8i16, ShOp>;
909 def v4i32 : N2VQShAdd<op24, op23, 0b100000, op11_8, 0, op4,
910 !strconcat(OpcodeStr, "32"), v4i32, ShOp>;
911 def v2i64 : N2VQShAdd<op24, op23, 0b000000, op11_8, 1, op4,
912 !strconcat(OpcodeStr, "64"), v2i64, ShOp>;
916 // Neon Shift-Insert vector operations,
917 // element sizes of 8, 16, 32 and 64 bits:
918 multiclass N2VShIns_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
919 string OpcodeStr, SDNode ShOp> {
920 // 64-bit vector types.
921 def v8i8 : N2VDShIns<op24, op23, 0b001000, op11_8, 0, op4,
922 !strconcat(OpcodeStr, "8"), v8i8, ShOp>;
923 def v4i16 : N2VDShIns<op24, op23, 0b010000, op11_8, 0, op4,
924 !strconcat(OpcodeStr, "16"), v4i16, ShOp>;
925 def v2i32 : N2VDShIns<op24, op23, 0b100000, op11_8, 0, op4,
926 !strconcat(OpcodeStr, "32"), v2i32, ShOp>;
927 def v1i64 : N2VDShIns<op24, op23, 0b000000, op11_8, 1, op4,
928 !strconcat(OpcodeStr, "64"), v1i64, ShOp>;
930 // 128-bit vector types.
931 def v16i8 : N2VQShIns<op24, op23, 0b001000, op11_8, 0, op4,
932 !strconcat(OpcodeStr, "8"), v16i8, ShOp>;
933 def v8i16 : N2VQShIns<op24, op23, 0b010000, op11_8, 0, op4,
934 !strconcat(OpcodeStr, "16"), v8i16, ShOp>;
935 def v4i32 : N2VQShIns<op24, op23, 0b100000, op11_8, 0, op4,
936 !strconcat(OpcodeStr, "32"), v4i32, ShOp>;
937 def v2i64 : N2VQShIns<op24, op23, 0b000000, op11_8, 1, op4,
938 !strconcat(OpcodeStr, "64"), v2i64, ShOp>;
941 //===----------------------------------------------------------------------===//
942 // Instruction Definitions.
943 //===----------------------------------------------------------------------===//
945 // Vector Add Operations.
947 // VADD : Vector Add (integer and floating-point)
948 defm VADD : N3V_QHSD<0, 0, 0b1000, 0, "vadd.i", add, 1>;
949 def VADDfd : N3VD<0, 0, 0b00, 0b1101, 0, "vadd.f32", v2f32, v2f32, fadd, 1>;
950 def VADDfq : N3VQ<0, 0, 0b00, 0b1101, 0, "vadd.f32", v4f32, v4f32, fadd, 1>;
951 // VADDL : Vector Add Long (Q = D + D)
952 defm VADDLs : N3VLInt_QHS<0,1,0b0000,0, "vaddl.s", int_arm_neon_vaddls, 1>;
953 defm VADDLu : N3VLInt_QHS<1,1,0b0000,0, "vaddl.u", int_arm_neon_vaddlu, 1>;
954 // VADDW : Vector Add Wide (Q = Q + D)
955 defm VADDWs : N3VWInt_QHS<0,1,0b0001,0, "vaddw.s", int_arm_neon_vaddws, 0>;
956 defm VADDWu : N3VWInt_QHS<1,1,0b0001,0, "vaddw.u", int_arm_neon_vaddwu, 0>;
957 // VHADD : Vector Halving Add
958 defm VHADDs : N3VInt_QHS<0,0,0b0000,0, "vhadd.s", int_arm_neon_vhadds, 1>;
959 defm VHADDu : N3VInt_QHS<1,0,0b0000,0, "vhadd.u", int_arm_neon_vhaddu, 1>;
960 // VRHADD : Vector Rounding Halving Add
961 defm VRHADDs : N3VInt_QHS<0,0,0b0001,0, "vrhadd.s", int_arm_neon_vrhadds, 1>;
962 defm VRHADDu : N3VInt_QHS<1,0,0b0001,0, "vrhadd.u", int_arm_neon_vrhaddu, 1>;
963 // VQADD : Vector Saturating Add
964 defm VQADDs : N3VInt_QHSD<0,0,0b0000,1, "vqadd.s", int_arm_neon_vqadds, 1>;
965 defm VQADDu : N3VInt_QHSD<1,0,0b0000,1, "vqadd.u", int_arm_neon_vqaddu, 1>;
966 // VADDHN : Vector Add and Narrow Returning High Half (D = Q + Q)
967 defm VADDHN : N3VNInt_HSD<0,1,0b0100,0, "vaddhn.i", int_arm_neon_vaddhn, 1>;
968 // VRADDHN : Vector Rounding Add and Narrow Returning High Half (D = Q + Q)
969 defm VRADDHN : N3VNInt_HSD<1,1,0b0100,0, "vraddhn.i", int_arm_neon_vraddhn, 1>;
971 // Vector Add Operations used for single-precision FP
972 def : N3VDs<fadd, VADDfd>;
974 // Vector Multiply Operations.
976 // VMUL : Vector Multiply (integer, polynomial and floating-point)
977 defm VMUL : N3V_QHS<0, 0, 0b1001, 1, "vmul.i", mul, 1>;
978 def VMULpd : N3VDInt<1, 0, 0b00, 0b1001, 1, "vmul.p8", v8i8, v8i8,
979 int_arm_neon_vmulp, 1>;
980 def VMULpq : N3VQInt<1, 0, 0b00, 0b1001, 1, "vmul.p8", v16i8, v16i8,
981 int_arm_neon_vmulp, 1>;
982 def VMULfd : N3VD<1, 0, 0b00, 0b1101, 1, "vmul.f32", v2f32, v2f32, fmul, 1>;
983 def VMULfq : N3VQ<1, 0, 0b00, 0b1101, 1, "vmul.f32", v4f32, v4f32, fmul, 1>;
984 // VQDMULH : Vector Saturating Doubling Multiply Returning High Half
985 defm VQDMULH : N3VInt_HS<0,0,0b1011,0, "vqdmulh.s", int_arm_neon_vqdmulh, 1>;
986 // VQRDMULH : Vector Rounding Saturating Doubling Multiply Returning High Half
987 defm VQRDMULH : N3VInt_HS<1,0,0b1011,0, "vqrdmulh.s", int_arm_neon_vqrdmulh, 1>;
988 // VMULL : Vector Multiply Long (integer and polynomial) (Q = D * D)
989 defm VMULLs : N3VLInt_QHS<0,1,0b1100,0, "vmull.s", int_arm_neon_vmulls, 1>;
990 defm VMULLu : N3VLInt_QHS<1,1,0b1100,0, "vmull.u", int_arm_neon_vmullu, 1>;
991 def VMULLp : N3VLInt<0, 1, 0b00, 0b1110, 0, "vmull.p8", v8i16, v8i8,
992 int_arm_neon_vmullp, 1>;
993 // VQDMULL : Vector Saturating Doubling Multiply Long (Q = D * D)
994 defm VQDMULL : N3VLInt_HS<0,1,0b1101,0, "vqdmull.s", int_arm_neon_vqdmull, 1>;
996 // Vector Multiply Operations used for single-precision FP
997 def : N3VDs<fmul, VMULfd>;
999 // Vector Multiply-Accumulate and Multiply-Subtract Operations.
1001 // VMLA : Vector Multiply Accumulate (integer and floating-point)
1002 defm VMLA : N3VMulOp_QHS<0, 0, 0b1001, 0, "vmla.i", add>;
1003 def VMLAfd : N3VDMulOp<0, 0, 0b00, 0b1101, 1, "vmla.f32", v2f32, fmul, fadd>;
1004 def VMLAfq : N3VQMulOp<0, 0, 0b00, 0b1101, 1, "vmla.f32", v4f32, fmul, fadd>;
1005 // VMLAL : Vector Multiply Accumulate Long (Q += D * D)
1006 defm VMLALs : N3VLInt3_QHS<0,1,0b1000,0, "vmlal.s", int_arm_neon_vmlals>;
1007 defm VMLALu : N3VLInt3_QHS<1,1,0b1000,0, "vmlal.u", int_arm_neon_vmlalu>;
1008 // VQDMLAL : Vector Saturating Doubling Multiply Accumulate Long (Q += D * D)
1009 defm VQDMLAL : N3VLInt3_HS<0, 1, 0b1001, 0, "vqdmlal.s", int_arm_neon_vqdmlal>;
1010 // VMLS : Vector Multiply Subtract (integer and floating-point)
1011 defm VMLS : N3VMulOp_QHS<0, 0, 0b1001, 0, "vmls.i", sub>;
1012 def VMLSfd : N3VDMulOp<0, 0, 0b10, 0b1101, 1, "vmls.f32", v2f32, fmul, fsub>;
1013 def VMLSfq : N3VQMulOp<0, 0, 0b10, 0b1101, 1, "vmls.f32", v4f32, fmul, fsub>;
1014 // VMLSL : Vector Multiply Subtract Long (Q -= D * D)
1015 defm VMLSLs : N3VLInt3_QHS<0,1,0b1010,0, "vmlsl.s", int_arm_neon_vmlsls>;
1016 defm VMLSLu : N3VLInt3_QHS<1,1,0b1010,0, "vmlsl.u", int_arm_neon_vmlslu>;
1017 // VQDMLSL : Vector Saturating Doubling Multiply Subtract Long (Q -= D * D)
1018 defm VQDMLSL : N3VLInt3_HS<0, 1, 0b1011, 0, "vqdmlsl.s", int_arm_neon_vqdmlsl>;
1020 // Vector Multiply-Accumulate/Subtract used for single-precision FP
1021 def : N3VDMulOps<fmul, fadd, VMLAfd>;
1022 def : N3VDMulOps<fmul, fsub, VMLSfd>;
1024 // Vector Subtract Operations.
1026 // VSUB : Vector Subtract (integer and floating-point)
1027 defm VSUB : N3V_QHSD<1, 0, 0b1000, 0, "vsub.i", sub, 0>;
1028 def VSUBfd : N3VD<0, 0, 0b10, 0b1101, 0, "vsub.f32", v2f32, v2f32, fsub, 0>;
1029 def VSUBfq : N3VQ<0, 0, 0b10, 0b1101, 0, "vsub.f32", v4f32, v4f32, fsub, 0>;
1030 // VSUBL : Vector Subtract Long (Q = D - D)
1031 defm VSUBLs : N3VLInt_QHS<0,1,0b0010,0, "vsubl.s", int_arm_neon_vsubls, 1>;
1032 defm VSUBLu : N3VLInt_QHS<1,1,0b0010,0, "vsubl.u", int_arm_neon_vsublu, 1>;
1033 // VSUBW : Vector Subtract Wide (Q = Q - D)
1034 defm VSUBWs : N3VWInt_QHS<0,1,0b0011,0, "vsubw.s", int_arm_neon_vsubws, 0>;
1035 defm VSUBWu : N3VWInt_QHS<1,1,0b0011,0, "vsubw.u", int_arm_neon_vsubwu, 0>;
1036 // VHSUB : Vector Halving Subtract
1037 defm VHSUBs : N3VInt_QHS<0, 0, 0b0010, 0, "vhsub.s", int_arm_neon_vhsubs, 0>;
1038 defm VHSUBu : N3VInt_QHS<1, 0, 0b0010, 0, "vhsub.u", int_arm_neon_vhsubu, 0>;
1039 // VQSUB : Vector Saturing Subtract
1040 defm VQSUBs : N3VInt_QHSD<0, 0, 0b0010, 1, "vqsub.s", int_arm_neon_vqsubs, 0>;
1041 defm VQSUBu : N3VInt_QHSD<1, 0, 0b0010, 1, "vqsub.u", int_arm_neon_vqsubu, 0>;
1042 // VSUBHN : Vector Subtract and Narrow Returning High Half (D = Q - Q)
1043 defm VSUBHN : N3VNInt_HSD<0,1,0b0110,0, "vsubhn.i", int_arm_neon_vsubhn, 0>;
1044 // VRSUBHN : Vector Rounding Subtract and Narrow Returning High Half (D=Q-Q)
1045 defm VRSUBHN : N3VNInt_HSD<1,1,0b0110,0, "vrsubhn.i", int_arm_neon_vrsubhn, 0>;
1047 // Vector Sub Operations used for single-precision FP
1048 def : N3VDs<fsub, VSUBfd>;
1050 // Vector Comparisons.
1052 // VCEQ : Vector Compare Equal
1053 defm VCEQ : N3V_QHS<1, 0, 0b1000, 1, "vceq.i", NEONvceq, 1>;
1054 def VCEQfd : N3VD<0,0,0b00,0b1110,0, "vceq.f32", v2i32, v2f32, NEONvceq, 1>;
1055 def VCEQfq : N3VQ<0,0,0b00,0b1110,0, "vceq.f32", v4i32, v4f32, NEONvceq, 1>;
1056 // VCGE : Vector Compare Greater Than or Equal
1057 defm VCGEs : N3V_QHS<0, 0, 0b0011, 1, "vcge.s", NEONvcge, 0>;
1058 defm VCGEu : N3V_QHS<1, 0, 0b0011, 1, "vcge.u", NEONvcgeu, 0>;
1059 def VCGEfd : N3VD<1,0,0b00,0b1110,0, "vcge.f32", v2i32, v2f32, NEONvcge, 0>;
1060 def VCGEfq : N3VQ<1,0,0b00,0b1110,0, "vcge.f32", v4i32, v4f32, NEONvcge, 0>;
1061 // VCGT : Vector Compare Greater Than
1062 defm VCGTs : N3V_QHS<0, 0, 0b0011, 0, "vcgt.s", NEONvcgt, 0>;
1063 defm VCGTu : N3V_QHS<1, 0, 0b0011, 0, "vcgt.u", NEONvcgtu, 0>;
1064 def VCGTfd : N3VD<1,0,0b10,0b1110,0, "vcgt.f32", v2i32, v2f32, NEONvcgt, 0>;
1065 def VCGTfq : N3VQ<1,0,0b10,0b1110,0, "vcgt.f32", v4i32, v4f32, NEONvcgt, 0>;
1066 // VACGE : Vector Absolute Compare Greater Than or Equal (aka VCAGE)
1067 def VACGEd : N3VDInt<1, 0, 0b00, 0b1110, 1, "vacge.f32", v2i32, v2f32,
1068 int_arm_neon_vacged, 0>;
1069 def VACGEq : N3VQInt<1, 0, 0b00, 0b1110, 1, "vacge.f32", v4i32, v4f32,
1070 int_arm_neon_vacgeq, 0>;
1071 // VACGT : Vector Absolute Compare Greater Than (aka VCAGT)
1072 def VACGTd : N3VDInt<1, 0, 0b10, 0b1110, 1, "vacgt.f32", v2i32, v2f32,
1073 int_arm_neon_vacgtd, 0>;
1074 def VACGTq : N3VQInt<1, 0, 0b10, 0b1110, 1, "vacgt.f32", v4i32, v4f32,
1075 int_arm_neon_vacgtq, 0>;
1076 // VTST : Vector Test Bits
1077 defm VTST : N3V_QHS<0, 0, 0b1000, 1, "vtst.i", NEONvtst, 1>;
1079 // Vector Bitwise Operations.
1081 // VAND : Vector Bitwise AND
1082 def VANDd : N3VD<0, 0, 0b00, 0b0001, 1, "vand", v2i32, v2i32, and, 1>;
1083 def VANDq : N3VQ<0, 0, 0b00, 0b0001, 1, "vand", v4i32, v4i32, and, 1>;
1085 // VEOR : Vector Bitwise Exclusive OR
1086 def VEORd : N3VD<1, 0, 0b00, 0b0001, 1, "veor", v2i32, v2i32, xor, 1>;
1087 def VEORq : N3VQ<1, 0, 0b00, 0b0001, 1, "veor", v4i32, v4i32, xor, 1>;
1089 // VORR : Vector Bitwise OR
1090 def VORRd : N3VD<0, 0, 0b10, 0b0001, 1, "vorr", v2i32, v2i32, or, 1>;
1091 def VORRq : N3VQ<0, 0, 0b10, 0b0001, 1, "vorr", v4i32, v4i32, or, 1>;
1093 // VBIC : Vector Bitwise Bit Clear (AND NOT)
1094 def VBICd : N3V<0, 0, 0b01, 0b0001, 0, 1, (outs DPR:$dst),
1095 (ins DPR:$src1, DPR:$src2), NoItinerary,
1096 "vbic\t$dst, $src1, $src2", "",
1097 [(set DPR:$dst, (v2i32 (and DPR:$src1,(vnot DPR:$src2))))]>;
1098 def VBICq : N3V<0, 0, 0b01, 0b0001, 1, 1, (outs QPR:$dst),
1099 (ins QPR:$src1, QPR:$src2), NoItinerary,
1100 "vbic\t$dst, $src1, $src2", "",
1101 [(set QPR:$dst, (v4i32 (and QPR:$src1,(vnot QPR:$src2))))]>;
1103 // VORN : Vector Bitwise OR NOT
1104 def VORNd : N3V<0, 0, 0b11, 0b0001, 0, 1, (outs DPR:$dst),
1105 (ins DPR:$src1, DPR:$src2), NoItinerary,
1106 "vorn\t$dst, $src1, $src2", "",
1107 [(set DPR:$dst, (v2i32 (or DPR:$src1, (vnot DPR:$src2))))]>;
1108 def VORNq : N3V<0, 0, 0b11, 0b0001, 1, 1, (outs QPR:$dst),
1109 (ins QPR:$src1, QPR:$src2), NoItinerary,
1110 "vorn\t$dst, $src1, $src2", "",
1111 [(set QPR:$dst, (v4i32 (or QPR:$src1, (vnot QPR:$src2))))]>;
1113 // VMVN : Vector Bitwise NOT
1114 def VMVNd : N2V<0b11, 0b11, 0b00, 0b00, 0b01011, 0, 0,
1115 (outs DPR:$dst), (ins DPR:$src), NoItinerary,
1116 "vmvn\t$dst, $src", "",
1117 [(set DPR:$dst, (v2i32 (vnot DPR:$src)))]>;
1118 def VMVNq : N2V<0b11, 0b11, 0b00, 0b00, 0b01011, 1, 0,
1119 (outs QPR:$dst), (ins QPR:$src), NoItinerary,
1120 "vmvn\t$dst, $src", "",
1121 [(set QPR:$dst, (v4i32 (vnot QPR:$src)))]>;
1122 def : Pat<(v2i32 (vnot_conv DPR:$src)), (VMVNd DPR:$src)>;
1123 def : Pat<(v4i32 (vnot_conv QPR:$src)), (VMVNq QPR:$src)>;
1125 // VBSL : Vector Bitwise Select
1126 def VBSLd : N3V<1, 0, 0b01, 0b0001, 0, 1, (outs DPR:$dst),
1127 (ins DPR:$src1, DPR:$src2, DPR:$src3), NoItinerary,
1128 "vbsl\t$dst, $src2, $src3", "$src1 = $dst",
1130 (v2i32 (or (and DPR:$src2, DPR:$src1),
1131 (and DPR:$src3, (vnot DPR:$src1)))))]>;
1132 def VBSLq : N3V<1, 0, 0b01, 0b0001, 1, 1, (outs QPR:$dst),
1133 (ins QPR:$src1, QPR:$src2, QPR:$src3), NoItinerary,
1134 "vbsl\t$dst, $src2, $src3", "$src1 = $dst",
1136 (v4i32 (or (and QPR:$src2, QPR:$src1),
1137 (and QPR:$src3, (vnot QPR:$src1)))))]>;
1139 // VBIF : Vector Bitwise Insert if False
1140 // like VBSL but with: "vbif\t$dst, $src3, $src1", "$src2 = $dst",
1141 // VBIT : Vector Bitwise Insert if True
1142 // like VBSL but with: "vbit\t$dst, $src2, $src1", "$src3 = $dst",
1143 // These are not yet implemented. The TwoAddress pass will not go looking
1144 // for equivalent operations with different register constraints; it just
1147 // Vector Absolute Differences.
1149 // VABD : Vector Absolute Difference
1150 defm VABDs : N3VInt_QHS<0, 0, 0b0111, 0, "vabd.s", int_arm_neon_vabds, 0>;
1151 defm VABDu : N3VInt_QHS<1, 0, 0b0111, 0, "vabd.u", int_arm_neon_vabdu, 0>;
1152 def VABDfd : N3VDInt<1, 0, 0b10, 0b1101, 0, "vabd.f32", v2f32, v2f32,
1153 int_arm_neon_vabdf, 0>;
1154 def VABDfq : N3VQInt<1, 0, 0b10, 0b1101, 0, "vabd.f32", v4f32, v4f32,
1155 int_arm_neon_vabdf, 0>;
1157 // VABDL : Vector Absolute Difference Long (Q = | D - D |)
1158 defm VABDLs : N3VLInt_QHS<0,1,0b0111,0, "vabdl.s", int_arm_neon_vabdls, 0>;
1159 defm VABDLu : N3VLInt_QHS<1,1,0b0111,0, "vabdl.u", int_arm_neon_vabdlu, 0>;
1161 // VABA : Vector Absolute Difference and Accumulate
1162 defm VABAs : N3VInt3_QHS<0,1,0b0101,0, "vaba.s", int_arm_neon_vabas>;
1163 defm VABAu : N3VInt3_QHS<1,1,0b0101,0, "vaba.u", int_arm_neon_vabau>;
1165 // VABAL : Vector Absolute Difference and Accumulate Long (Q += | D - D |)
1166 defm VABALs : N3VLInt3_QHS<0,1,0b0101,0, "vabal.s", int_arm_neon_vabals>;
1167 defm VABALu : N3VLInt3_QHS<1,1,0b0101,0, "vabal.u", int_arm_neon_vabalu>;
1169 // Vector Maximum and Minimum.
1171 // VMAX : Vector Maximum
1172 defm VMAXs : N3VInt_QHS<0, 0, 0b0110, 0, "vmax.s", int_arm_neon_vmaxs, 1>;
1173 defm VMAXu : N3VInt_QHS<1, 0, 0b0110, 0, "vmax.u", int_arm_neon_vmaxu, 1>;
1174 def VMAXfd : N3VDInt<0, 0, 0b00, 0b1111, 0, "vmax.f32", v2f32, v2f32,
1175 int_arm_neon_vmaxf, 1>;
1176 def VMAXfq : N3VQInt<0, 0, 0b00, 0b1111, 0, "vmax.f32", v4f32, v4f32,
1177 int_arm_neon_vmaxf, 1>;
1179 // VMIN : Vector Minimum
1180 defm VMINs : N3VInt_QHS<0, 0, 0b0110, 1, "vmin.s", int_arm_neon_vmins, 1>;
1181 defm VMINu : N3VInt_QHS<1, 0, 0b0110, 1, "vmin.u", int_arm_neon_vminu, 1>;
1182 def VMINfd : N3VDInt<0, 0, 0b10, 0b1111, 0, "vmin.f32", v2f32, v2f32,
1183 int_arm_neon_vminf, 1>;
1184 def VMINfq : N3VQInt<0, 0, 0b10, 0b1111, 0, "vmin.f32", v4f32, v4f32,
1185 int_arm_neon_vminf, 1>;
1187 // Vector Pairwise Operations.
1189 // VPADD : Vector Pairwise Add
1190 def VPADDi8 : N3VDInt<0, 0, 0b00, 0b1011, 1, "vpadd.i8", v8i8, v8i8,
1191 int_arm_neon_vpaddi, 0>;
1192 def VPADDi16 : N3VDInt<0, 0, 0b01, 0b1011, 1, "vpadd.i16", v4i16, v4i16,
1193 int_arm_neon_vpaddi, 0>;
1194 def VPADDi32 : N3VDInt<0, 0, 0b10, 0b1011, 1, "vpadd.i32", v2i32, v2i32,
1195 int_arm_neon_vpaddi, 0>;
1196 def VPADDf : N3VDInt<1, 0, 0b00, 0b1101, 0, "vpadd.f32", v2f32, v2f32,
1197 int_arm_neon_vpaddf, 0>;
1199 // VPADDL : Vector Pairwise Add Long
1200 defm VPADDLs : N2VPLInt_QHS<0b11, 0b11, 0b00, 0b00100, 0, "vpaddl.s",
1201 int_arm_neon_vpaddls>;
1202 defm VPADDLu : N2VPLInt_QHS<0b11, 0b11, 0b00, 0b00101, 0, "vpaddl.u",
1203 int_arm_neon_vpaddlu>;
1205 // VPADAL : Vector Pairwise Add and Accumulate Long
1206 defm VPADALs : N2VPLInt2_QHS<0b11, 0b11, 0b00, 0b00100, 0, "vpadal.s",
1207 int_arm_neon_vpadals>;
1208 defm VPADALu : N2VPLInt2_QHS<0b11, 0b11, 0b00, 0b00101, 0, "vpadal.u",
1209 int_arm_neon_vpadalu>;
1211 // VPMAX : Vector Pairwise Maximum
1212 def VPMAXs8 : N3VDInt<0, 0, 0b00, 0b1010, 0, "vpmax.s8", v8i8, v8i8,
1213 int_arm_neon_vpmaxs, 0>;
1214 def VPMAXs16 : N3VDInt<0, 0, 0b01, 0b1010, 0, "vpmax.s16", v4i16, v4i16,
1215 int_arm_neon_vpmaxs, 0>;
1216 def VPMAXs32 : N3VDInt<0, 0, 0b10, 0b1010, 0, "vpmax.s32", v2i32, v2i32,
1217 int_arm_neon_vpmaxs, 0>;
1218 def VPMAXu8 : N3VDInt<1, 0, 0b00, 0b1010, 0, "vpmax.u8", v8i8, v8i8,
1219 int_arm_neon_vpmaxu, 0>;
1220 def VPMAXu16 : N3VDInt<1, 0, 0b01, 0b1010, 0, "vpmax.u16", v4i16, v4i16,
1221 int_arm_neon_vpmaxu, 0>;
1222 def VPMAXu32 : N3VDInt<1, 0, 0b10, 0b1010, 0, "vpmax.u32", v2i32, v2i32,
1223 int_arm_neon_vpmaxu, 0>;
1224 def VPMAXf : N3VDInt<1, 0, 0b00, 0b1111, 0, "vpmax.f32", v2f32, v2f32,
1225 int_arm_neon_vpmaxf, 0>;
1227 // VPMIN : Vector Pairwise Minimum
1228 def VPMINs8 : N3VDInt<0, 0, 0b00, 0b1010, 1, "vpmin.s8", v8i8, v8i8,
1229 int_arm_neon_vpmins, 0>;
1230 def VPMINs16 : N3VDInt<0, 0, 0b01, 0b1010, 1, "vpmin.s16", v4i16, v4i16,
1231 int_arm_neon_vpmins, 0>;
1232 def VPMINs32 : N3VDInt<0, 0, 0b10, 0b1010, 1, "vpmin.s32", v2i32, v2i32,
1233 int_arm_neon_vpmins, 0>;
1234 def VPMINu8 : N3VDInt<1, 0, 0b00, 0b1010, 1, "vpmin.u8", v8i8, v8i8,
1235 int_arm_neon_vpminu, 0>;
1236 def VPMINu16 : N3VDInt<1, 0, 0b01, 0b1010, 1, "vpmin.u16", v4i16, v4i16,
1237 int_arm_neon_vpminu, 0>;
1238 def VPMINu32 : N3VDInt<1, 0, 0b10, 0b1010, 1, "vpmin.u32", v2i32, v2i32,
1239 int_arm_neon_vpminu, 0>;
1240 def VPMINf : N3VDInt<1, 0, 0b10, 0b1111, 0, "vpmin.f32", v2f32, v2f32,
1241 int_arm_neon_vpminf, 0>;
1243 // Vector Reciprocal and Reciprocal Square Root Estimate and Step.
1245 // VRECPE : Vector Reciprocal Estimate
1246 def VRECPEd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01000, 0, "vrecpe.u32",
1247 v2i32, v2i32, int_arm_neon_vrecpe>;
1248 def VRECPEq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01000, 0, "vrecpe.u32",
1249 v4i32, v4i32, int_arm_neon_vrecpe>;
1250 def VRECPEfd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01010, 0, "vrecpe.f32",
1251 v2f32, v2f32, int_arm_neon_vrecpef>;
1252 def VRECPEfq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01010, 0, "vrecpe.f32",
1253 v4f32, v4f32, int_arm_neon_vrecpef>;
1255 // VRECPS : Vector Reciprocal Step
1256 def VRECPSfd : N3VDInt<0, 0, 0b00, 0b1111, 1, "vrecps.f32", v2f32, v2f32,
1257 int_arm_neon_vrecps, 1>;
1258 def VRECPSfq : N3VQInt<0, 0, 0b00, 0b1111, 1, "vrecps.f32", v4f32, v4f32,
1259 int_arm_neon_vrecps, 1>;
1261 // VRSQRTE : Vector Reciprocal Square Root Estimate
1262 def VRSQRTEd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01001, 0, "vrsqrte.u32",
1263 v2i32, v2i32, int_arm_neon_vrsqrte>;
1264 def VRSQRTEq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01001, 0, "vrsqrte.u32",
1265 v4i32, v4i32, int_arm_neon_vrsqrte>;
1266 def VRSQRTEfd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01011, 0, "vrsqrte.f32",
1267 v2f32, v2f32, int_arm_neon_vrsqrtef>;
1268 def VRSQRTEfq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01011, 0, "vrsqrte.f32",
1269 v4f32, v4f32, int_arm_neon_vrsqrtef>;
1271 // VRSQRTS : Vector Reciprocal Square Root Step
1272 def VRSQRTSfd : N3VDInt<0, 0, 0b10, 0b1111, 1, "vrsqrts.f32", v2f32, v2f32,
1273 int_arm_neon_vrsqrts, 1>;
1274 def VRSQRTSfq : N3VQInt<0, 0, 0b10, 0b1111, 1, "vrsqrts.f32", v4f32, v4f32,
1275 int_arm_neon_vrsqrts, 1>;
1279 // VSHL : Vector Shift
1280 defm VSHLs : N3VInt_QHSD<0, 0, 0b0100, 0, "vshl.s", int_arm_neon_vshifts, 0>;
1281 defm VSHLu : N3VInt_QHSD<1, 0, 0b0100, 0, "vshl.u", int_arm_neon_vshiftu, 0>;
1282 // VSHL : Vector Shift Left (Immediate)
1283 defm VSHLi : N2VSh_QHSD<0, 1, 0b0111, 1, "vshl.i", NEONvshl>;
1284 // VSHR : Vector Shift Right (Immediate)
1285 defm VSHRs : N2VSh_QHSD<0, 1, 0b0000, 1, "vshr.s", NEONvshrs>;
1286 defm VSHRu : N2VSh_QHSD<1, 1, 0b0000, 1, "vshr.u", NEONvshru>;
1288 // VSHLL : Vector Shift Left Long
1289 def VSHLLs8 : N2VLSh<0, 1, 0b001000, 0b1010, 0, 0, 1, "vshll.s8",
1290 v8i16, v8i8, NEONvshlls>;
1291 def VSHLLs16 : N2VLSh<0, 1, 0b010000, 0b1010, 0, 0, 1, "vshll.s16",
1292 v4i32, v4i16, NEONvshlls>;
1293 def VSHLLs32 : N2VLSh<0, 1, 0b100000, 0b1010, 0, 0, 1, "vshll.s32",
1294 v2i64, v2i32, NEONvshlls>;
1295 def VSHLLu8 : N2VLSh<1, 1, 0b001000, 0b1010, 0, 0, 1, "vshll.u8",
1296 v8i16, v8i8, NEONvshllu>;
1297 def VSHLLu16 : N2VLSh<1, 1, 0b010000, 0b1010, 0, 0, 1, "vshll.u16",
1298 v4i32, v4i16, NEONvshllu>;
1299 def VSHLLu32 : N2VLSh<1, 1, 0b100000, 0b1010, 0, 0, 1, "vshll.u32",
1300 v2i64, v2i32, NEONvshllu>;
1302 // VSHLL : Vector Shift Left Long (with maximum shift count)
1303 def VSHLLi8 : N2VLSh<1, 1, 0b110010, 0b0011, 0, 0, 0, "vshll.i8",
1304 v8i16, v8i8, NEONvshlli>;
1305 def VSHLLi16 : N2VLSh<1, 1, 0b110110, 0b0011, 0, 0, 0, "vshll.i16",
1306 v4i32, v4i16, NEONvshlli>;
1307 def VSHLLi32 : N2VLSh<1, 1, 0b111010, 0b0011, 0, 0, 0, "vshll.i32",
1308 v2i64, v2i32, NEONvshlli>;
1310 // VSHRN : Vector Shift Right and Narrow
1311 def VSHRN16 : N2VNSh<0, 1, 0b001000, 0b1000, 0, 0, 1, "vshrn.i16",
1312 v8i8, v8i16, NEONvshrn>;
1313 def VSHRN32 : N2VNSh<0, 1, 0b010000, 0b1000, 0, 0, 1, "vshrn.i32",
1314 v4i16, v4i32, NEONvshrn>;
1315 def VSHRN64 : N2VNSh<0, 1, 0b100000, 0b1000, 0, 0, 1, "vshrn.i64",
1316 v2i32, v2i64, NEONvshrn>;
1318 // VRSHL : Vector Rounding Shift
1319 defm VRSHLs : N3VInt_QHSD<0,0,0b0101,0, "vrshl.s", int_arm_neon_vrshifts, 0>;
1320 defm VRSHLu : N3VInt_QHSD<1,0,0b0101,0, "vrshl.u", int_arm_neon_vrshiftu, 0>;
1321 // VRSHR : Vector Rounding Shift Right
1322 defm VRSHRs : N2VSh_QHSD<0, 1, 0b0010, 1, "vrshr.s", NEONvrshrs>;
1323 defm VRSHRu : N2VSh_QHSD<1, 1, 0b0010, 1, "vrshr.u", NEONvrshru>;
1325 // VRSHRN : Vector Rounding Shift Right and Narrow
1326 def VRSHRN16 : N2VNSh<0, 1, 0b001000, 0b1000, 0, 1, 1, "vrshrn.i16",
1327 v8i8, v8i16, NEONvrshrn>;
1328 def VRSHRN32 : N2VNSh<0, 1, 0b010000, 0b1000, 0, 1, 1, "vrshrn.i32",
1329 v4i16, v4i32, NEONvrshrn>;
1330 def VRSHRN64 : N2VNSh<0, 1, 0b100000, 0b1000, 0, 1, 1, "vrshrn.i64",
1331 v2i32, v2i64, NEONvrshrn>;
1333 // VQSHL : Vector Saturating Shift
1334 defm VQSHLs : N3VInt_QHSD<0,0,0b0100,1, "vqshl.s", int_arm_neon_vqshifts, 0>;
1335 defm VQSHLu : N3VInt_QHSD<1,0,0b0100,1, "vqshl.u", int_arm_neon_vqshiftu, 0>;
1336 // VQSHL : Vector Saturating Shift Left (Immediate)
1337 defm VQSHLsi : N2VSh_QHSD<0, 1, 0b0111, 1, "vqshl.s", NEONvqshls>;
1338 defm VQSHLui : N2VSh_QHSD<1, 1, 0b0111, 1, "vqshl.u", NEONvqshlu>;
1339 // VQSHLU : Vector Saturating Shift Left (Immediate, Unsigned)
1340 defm VQSHLsu : N2VSh_QHSD<1, 1, 0b0110, 1, "vqshlu.s", NEONvqshlsu>;
1342 // VQSHRN : Vector Saturating Shift Right and Narrow
1343 def VQSHRNs16 : N2VNSh<0, 1, 0b001000, 0b1001, 0, 0, 1, "vqshrn.s16",
1344 v8i8, v8i16, NEONvqshrns>;
1345 def VQSHRNs32 : N2VNSh<0, 1, 0b010000, 0b1001, 0, 0, 1, "vqshrn.s32",
1346 v4i16, v4i32, NEONvqshrns>;
1347 def VQSHRNs64 : N2VNSh<0, 1, 0b100000, 0b1001, 0, 0, 1, "vqshrn.s64",
1348 v2i32, v2i64, NEONvqshrns>;
1349 def VQSHRNu16 : N2VNSh<1, 1, 0b001000, 0b1001, 0, 0, 1, "vqshrn.u16",
1350 v8i8, v8i16, NEONvqshrnu>;
1351 def VQSHRNu32 : N2VNSh<1, 1, 0b010000, 0b1001, 0, 0, 1, "vqshrn.u32",
1352 v4i16, v4i32, NEONvqshrnu>;
1353 def VQSHRNu64 : N2VNSh<1, 1, 0b100000, 0b1001, 0, 0, 1, "vqshrn.u64",
1354 v2i32, v2i64, NEONvqshrnu>;
1356 // VQSHRUN : Vector Saturating Shift Right and Narrow (Unsigned)
1357 def VQSHRUN16 : N2VNSh<1, 1, 0b001000, 0b1000, 0, 0, 1, "vqshrun.s16",
1358 v8i8, v8i16, NEONvqshrnsu>;
1359 def VQSHRUN32 : N2VNSh<1, 1, 0b010000, 0b1000, 0, 0, 1, "vqshrun.s32",
1360 v4i16, v4i32, NEONvqshrnsu>;
1361 def VQSHRUN64 : N2VNSh<1, 1, 0b100000, 0b1000, 0, 0, 1, "vqshrun.s64",
1362 v2i32, v2i64, NEONvqshrnsu>;
1364 // VQRSHL : Vector Saturating Rounding Shift
1365 defm VQRSHLs : N3VInt_QHSD<0, 0, 0b0101, 1, "vqrshl.s",
1366 int_arm_neon_vqrshifts, 0>;
1367 defm VQRSHLu : N3VInt_QHSD<1, 0, 0b0101, 1, "vqrshl.u",
1368 int_arm_neon_vqrshiftu, 0>;
1370 // VQRSHRN : Vector Saturating Rounding Shift Right and Narrow
1371 def VQRSHRNs16: N2VNSh<0, 1, 0b001000, 0b1001, 0, 1, 1, "vqrshrn.s16",
1372 v8i8, v8i16, NEONvqrshrns>;
1373 def VQRSHRNs32: N2VNSh<0, 1, 0b010000, 0b1001, 0, 1, 1, "vqrshrn.s32",
1374 v4i16, v4i32, NEONvqrshrns>;
1375 def VQRSHRNs64: N2VNSh<0, 1, 0b100000, 0b1001, 0, 1, 1, "vqrshrn.s64",
1376 v2i32, v2i64, NEONvqrshrns>;
1377 def VQRSHRNu16: N2VNSh<1, 1, 0b001000, 0b1001, 0, 1, 1, "vqrshrn.u16",
1378 v8i8, v8i16, NEONvqrshrnu>;
1379 def VQRSHRNu32: N2VNSh<1, 1, 0b010000, 0b1001, 0, 1, 1, "vqrshrn.u32",
1380 v4i16, v4i32, NEONvqrshrnu>;
1381 def VQRSHRNu64: N2VNSh<1, 1, 0b100000, 0b1001, 0, 1, 1, "vqrshrn.u64",
1382 v2i32, v2i64, NEONvqrshrnu>;
1384 // VQRSHRUN : Vector Saturating Rounding Shift Right and Narrow (Unsigned)
1385 def VQRSHRUN16: N2VNSh<1, 1, 0b001000, 0b1000, 0, 1, 1, "vqrshrun.s16",
1386 v8i8, v8i16, NEONvqrshrnsu>;
1387 def VQRSHRUN32: N2VNSh<1, 1, 0b010000, 0b1000, 0, 1, 1, "vqrshrun.s32",
1388 v4i16, v4i32, NEONvqrshrnsu>;
1389 def VQRSHRUN64: N2VNSh<1, 1, 0b100000, 0b1000, 0, 1, 1, "vqrshrun.s64",
1390 v2i32, v2i64, NEONvqrshrnsu>;
1392 // VSRA : Vector Shift Right and Accumulate
1393 defm VSRAs : N2VShAdd_QHSD<0, 1, 0b0001, 1, "vsra.s", NEONvshrs>;
1394 defm VSRAu : N2VShAdd_QHSD<1, 1, 0b0001, 1, "vsra.u", NEONvshru>;
1395 // VRSRA : Vector Rounding Shift Right and Accumulate
1396 defm VRSRAs : N2VShAdd_QHSD<0, 1, 0b0011, 1, "vrsra.s", NEONvrshrs>;
1397 defm VRSRAu : N2VShAdd_QHSD<1, 1, 0b0011, 1, "vrsra.u", NEONvrshru>;
1399 // VSLI : Vector Shift Left and Insert
1400 defm VSLI : N2VShIns_QHSD<1, 1, 0b0101, 1, "vsli.", NEONvsli>;
1401 // VSRI : Vector Shift Right and Insert
1402 defm VSRI : N2VShIns_QHSD<1, 1, 0b0100, 1, "vsri.", NEONvsri>;
1404 // Vector Absolute and Saturating Absolute.
1406 // VABS : Vector Absolute Value
1407 defm VABS : N2VInt_QHS<0b11, 0b11, 0b01, 0b00110, 0, "vabs.s",
1409 def VABSfd : N2VDInt<0b11, 0b11, 0b10, 0b01, 0b01110, 0, "vabs.f32",
1410 v2f32, v2f32, int_arm_neon_vabsf>;
1411 def VABSfq : N2VQInt<0b11, 0b11, 0b10, 0b01, 0b01110, 0, "vabs.f32",
1412 v4f32, v4f32, int_arm_neon_vabsf>;
1413 def : N2VDInts<fabs, VABSfd>;
1415 // VQABS : Vector Saturating Absolute Value
1416 defm VQABS : N2VInt_QHS<0b11, 0b11, 0b00, 0b01110, 0, "vqabs.s",
1417 int_arm_neon_vqabs>;
1421 def vneg : PatFrag<(ops node:$in), (sub immAllZerosV, node:$in)>;
1422 def vneg_conv : PatFrag<(ops node:$in), (sub immAllZerosV_bc, node:$in)>;
1424 class VNEGD<bits<2> size, string OpcodeStr, ValueType Ty>
1425 : N2V<0b11, 0b11, size, 0b01, 0b00111, 0, 0, (outs DPR:$dst), (ins DPR:$src),
1427 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1428 [(set DPR:$dst, (Ty (vneg DPR:$src)))]>;
1429 class VNEGQ<bits<2> size, string OpcodeStr, ValueType Ty>
1430 : N2V<0b11, 0b11, size, 0b01, 0b00111, 1, 0, (outs QPR:$dst), (ins QPR:$src),
1432 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1433 [(set QPR:$dst, (Ty (vneg QPR:$src)))]>;
1435 // VNEG : Vector Negate
1436 def VNEGs8d : VNEGD<0b00, "vneg.s8", v8i8>;
1437 def VNEGs16d : VNEGD<0b01, "vneg.s16", v4i16>;
1438 def VNEGs32d : VNEGD<0b10, "vneg.s32", v2i32>;
1439 def VNEGs8q : VNEGQ<0b00, "vneg.s8", v16i8>;
1440 def VNEGs16q : VNEGQ<0b01, "vneg.s16", v8i16>;
1441 def VNEGs32q : VNEGQ<0b10, "vneg.s32", v4i32>;
1443 // VNEG : Vector Negate (floating-point)
1444 def VNEGf32d : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 0, 0,
1445 (outs DPR:$dst), (ins DPR:$src), NoItinerary,
1446 "vneg.f32\t$dst, $src", "",
1447 [(set DPR:$dst, (v2f32 (fneg DPR:$src)))]>;
1448 def VNEGf32q : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 1, 0,
1449 (outs QPR:$dst), (ins QPR:$src), NoItinerary,
1450 "vneg.f32\t$dst, $src", "",
1451 [(set QPR:$dst, (v4f32 (fneg QPR:$src)))]>;
1452 def : N2VDInts<fneg, VNEGf32d>;
1454 def : Pat<(v8i8 (vneg_conv DPR:$src)), (VNEGs8d DPR:$src)>;
1455 def : Pat<(v4i16 (vneg_conv DPR:$src)), (VNEGs16d DPR:$src)>;
1456 def : Pat<(v2i32 (vneg_conv DPR:$src)), (VNEGs32d DPR:$src)>;
1457 def : Pat<(v16i8 (vneg_conv QPR:$src)), (VNEGs8q QPR:$src)>;
1458 def : Pat<(v8i16 (vneg_conv QPR:$src)), (VNEGs16q QPR:$src)>;
1459 def : Pat<(v4i32 (vneg_conv QPR:$src)), (VNEGs32q QPR:$src)>;
1461 // VQNEG : Vector Saturating Negate
1462 defm VQNEG : N2VInt_QHS<0b11, 0b11, 0b00, 0b01111, 0, "vqneg.s",
1463 int_arm_neon_vqneg>;
1465 // Vector Bit Counting Operations.
1467 // VCLS : Vector Count Leading Sign Bits
1468 defm VCLS : N2VInt_QHS<0b11, 0b11, 0b00, 0b01000, 0, "vcls.s",
1470 // VCLZ : Vector Count Leading Zeros
1471 defm VCLZ : N2VInt_QHS<0b11, 0b11, 0b00, 0b01001, 0, "vclz.i",
1473 // VCNT : Vector Count One Bits
1474 def VCNTd : N2VDInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0, "vcnt.8",
1475 v8i8, v8i8, int_arm_neon_vcnt>;
1476 def VCNTq : N2VQInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0, "vcnt.8",
1477 v16i8, v16i8, int_arm_neon_vcnt>;
1479 // Vector Move Operations.
1481 // VMOV : Vector Move (Register)
1483 def VMOVD : N3V<0, 0, 0b10, 0b0001, 0, 1, (outs DPR:$dst), (ins DPR:$src),
1484 NoItinerary, "vmov\t$dst, $src", "", []>;
1485 def VMOVQ : N3V<0, 0, 0b10, 0b0001, 1, 1, (outs QPR:$dst), (ins QPR:$src),
1486 NoItinerary, "vmov\t$dst, $src", "", []>;
1488 // VMOV : Vector Move (Immediate)
1490 // VMOV_get_imm8 xform function: convert build_vector to VMOV.i8 imm.
1491 def VMOV_get_imm8 : SDNodeXForm<build_vector, [{
1492 return ARM::getVMOVImm(N, 1, *CurDAG);
1494 def vmovImm8 : PatLeaf<(build_vector), [{
1495 return ARM::getVMOVImm(N, 1, *CurDAG).getNode() != 0;
1498 // VMOV_get_imm16 xform function: convert build_vector to VMOV.i16 imm.
1499 def VMOV_get_imm16 : SDNodeXForm<build_vector, [{
1500 return ARM::getVMOVImm(N, 2, *CurDAG);
1502 def vmovImm16 : PatLeaf<(build_vector), [{
1503 return ARM::getVMOVImm(N, 2, *CurDAG).getNode() != 0;
1504 }], VMOV_get_imm16>;
1506 // VMOV_get_imm32 xform function: convert build_vector to VMOV.i32 imm.
1507 def VMOV_get_imm32 : SDNodeXForm<build_vector, [{
1508 return ARM::getVMOVImm(N, 4, *CurDAG);
1510 def vmovImm32 : PatLeaf<(build_vector), [{
1511 return ARM::getVMOVImm(N, 4, *CurDAG).getNode() != 0;
1512 }], VMOV_get_imm32>;
1514 // VMOV_get_imm64 xform function: convert build_vector to VMOV.i64 imm.
1515 def VMOV_get_imm64 : SDNodeXForm<build_vector, [{
1516 return ARM::getVMOVImm(N, 8, *CurDAG);
1518 def vmovImm64 : PatLeaf<(build_vector), [{
1519 return ARM::getVMOVImm(N, 8, *CurDAG).getNode() != 0;
1520 }], VMOV_get_imm64>;
1522 // Note: Some of the cmode bits in the following VMOV instructions need to
1523 // be encoded based on the immed values.
1525 def VMOVv8i8 : N1ModImm<1, 0b000, 0b1110, 0, 0, 0, 1, (outs DPR:$dst),
1526 (ins i8imm:$SIMM), NoItinerary,
1527 "vmov.i8\t$dst, $SIMM", "",
1528 [(set DPR:$dst, (v8i8 vmovImm8:$SIMM))]>;
1529 def VMOVv16i8 : N1ModImm<1, 0b000, 0b1110, 0, 1, 0, 1, (outs QPR:$dst),
1530 (ins i8imm:$SIMM), NoItinerary,
1531 "vmov.i8\t$dst, $SIMM", "",
1532 [(set QPR:$dst, (v16i8 vmovImm8:$SIMM))]>;
1534 def VMOVv4i16 : N1ModImm<1, 0b000, 0b1000, 0, 0, 0, 1, (outs DPR:$dst),
1535 (ins i16imm:$SIMM), NoItinerary,
1536 "vmov.i16\t$dst, $SIMM", "",
1537 [(set DPR:$dst, (v4i16 vmovImm16:$SIMM))]>;
1538 def VMOVv8i16 : N1ModImm<1, 0b000, 0b1000, 0, 1, 0, 1, (outs QPR:$dst),
1539 (ins i16imm:$SIMM), NoItinerary,
1540 "vmov.i16\t$dst, $SIMM", "",
1541 [(set QPR:$dst, (v8i16 vmovImm16:$SIMM))]>;
1543 def VMOVv2i32 : N1ModImm<1, 0b000, 0b0000, 0, 0, 0, 1, (outs DPR:$dst),
1544 (ins i32imm:$SIMM), NoItinerary,
1545 "vmov.i32\t$dst, $SIMM", "",
1546 [(set DPR:$dst, (v2i32 vmovImm32:$SIMM))]>;
1547 def VMOVv4i32 : N1ModImm<1, 0b000, 0b0000, 0, 1, 0, 1, (outs QPR:$dst),
1548 (ins i32imm:$SIMM), NoItinerary,
1549 "vmov.i32\t$dst, $SIMM", "",
1550 [(set QPR:$dst, (v4i32 vmovImm32:$SIMM))]>;
1552 def VMOVv1i64 : N1ModImm<1, 0b000, 0b1110, 0, 0, 1, 1, (outs DPR:$dst),
1553 (ins i64imm:$SIMM), NoItinerary,
1554 "vmov.i64\t$dst, $SIMM", "",
1555 [(set DPR:$dst, (v1i64 vmovImm64:$SIMM))]>;
1556 def VMOVv2i64 : N1ModImm<1, 0b000, 0b1110, 0, 1, 1, 1, (outs QPR:$dst),
1557 (ins i64imm:$SIMM), NoItinerary,
1558 "vmov.i64\t$dst, $SIMM", "",
1559 [(set QPR:$dst, (v2i64 vmovImm64:$SIMM))]>;
1561 // VMOV : Vector Get Lane (move scalar to ARM core register)
1563 def VGETLNs8 : NVGetLane<0b11100101, 0b1011, 0b00,
1564 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1565 NoItinerary, "vmov", ".s8\t$dst, $src[$lane]",
1566 [(set GPR:$dst, (NEONvgetlanes (v8i8 DPR:$src),
1568 def VGETLNs16 : NVGetLane<0b11100001, 0b1011, 0b01,
1569 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1570 NoItinerary, "vmov", ".s16\t$dst, $src[$lane]",
1571 [(set GPR:$dst, (NEONvgetlanes (v4i16 DPR:$src),
1573 def VGETLNu8 : NVGetLane<0b11101101, 0b1011, 0b00,
1574 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1575 NoItinerary, "vmov", ".u8\t$dst, $src[$lane]",
1576 [(set GPR:$dst, (NEONvgetlaneu (v8i8 DPR:$src),
1578 def VGETLNu16 : NVGetLane<0b11101001, 0b1011, 0b01,
1579 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1580 NoItinerary, "vmov", ".u16\t$dst, $src[$lane]",
1581 [(set GPR:$dst, (NEONvgetlaneu (v4i16 DPR:$src),
1583 def VGETLNi32 : NVGetLane<0b11100001, 0b1011, 0b00,
1584 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1585 NoItinerary, "vmov", ".32\t$dst, $src[$lane]",
1586 [(set GPR:$dst, (extractelt (v2i32 DPR:$src),
1588 // def VGETLNf32: see FMRDH and FMRDL in ARMInstrVFP.td
1589 def : Pat<(NEONvgetlanes (v16i8 QPR:$src), imm:$lane),
1590 (VGETLNs8 (v8i8 (EXTRACT_SUBREG QPR:$src,
1591 (SubReg_i8_reg imm:$lane))),
1592 (SubReg_i8_lane imm:$lane))>;
1593 def : Pat<(NEONvgetlanes (v8i16 QPR:$src), imm:$lane),
1594 (VGETLNs16 (v4i16 (EXTRACT_SUBREG QPR:$src,
1595 (SubReg_i16_reg imm:$lane))),
1596 (SubReg_i16_lane imm:$lane))>;
1597 def : Pat<(NEONvgetlaneu (v16i8 QPR:$src), imm:$lane),
1598 (VGETLNu8 (v8i8 (EXTRACT_SUBREG QPR:$src,
1599 (SubReg_i8_reg imm:$lane))),
1600 (SubReg_i8_lane imm:$lane))>;
1601 def : Pat<(NEONvgetlaneu (v8i16 QPR:$src), imm:$lane),
1602 (VGETLNu16 (v4i16 (EXTRACT_SUBREG QPR:$src,
1603 (SubReg_i16_reg imm:$lane))),
1604 (SubReg_i16_lane imm:$lane))>;
1605 def : Pat<(extractelt (v4i32 QPR:$src), imm:$lane),
1606 (VGETLNi32 (v2i32 (EXTRACT_SUBREG QPR:$src,
1607 (SubReg_i32_reg imm:$lane))),
1608 (SubReg_i32_lane imm:$lane))>;
1609 //def : Pat<(extractelt (v2i64 QPR:$src1), imm:$src2),
1610 // (EXTRACT_SUBREG QPR:$src1, (SubReg_f64_reg imm:$src2))>;
1611 def : Pat<(extractelt (v2f64 QPR:$src1), imm:$src2),
1612 (EXTRACT_SUBREG QPR:$src1, (SubReg_f64_reg imm:$src2))>;
1615 // VMOV : Vector Set Lane (move ARM core register to scalar)
1617 let Constraints = "$src1 = $dst" in {
1618 def VSETLNi8 : NVSetLane<0b11100100, 0b1011, 0b00, (outs DPR:$dst),
1619 (ins DPR:$src1, GPR:$src2, i32imm:$lane),
1620 NoItinerary, "vmov", ".8\t$dst[$lane], $src2",
1621 [(set DPR:$dst, (vector_insert (v8i8 DPR:$src1),
1622 GPR:$src2, imm:$lane))]>;
1623 def VSETLNi16 : NVSetLane<0b11100000, 0b1011, 0b01, (outs DPR:$dst),
1624 (ins DPR:$src1, GPR:$src2, i32imm:$lane),
1625 NoItinerary, "vmov", ".16\t$dst[$lane], $src2",
1626 [(set DPR:$dst, (vector_insert (v4i16 DPR:$src1),
1627 GPR:$src2, imm:$lane))]>;
1628 def VSETLNi32 : NVSetLane<0b11100000, 0b1011, 0b00, (outs DPR:$dst),
1629 (ins DPR:$src1, GPR:$src2, i32imm:$lane),
1630 NoItinerary, "vmov", ".32\t$dst[$lane], $src2",
1631 [(set DPR:$dst, (insertelt (v2i32 DPR:$src1),
1632 GPR:$src2, imm:$lane))]>;
1634 def : Pat<(vector_insert (v16i8 QPR:$src1), GPR:$src2, imm:$lane),
1635 (v16i8 (INSERT_SUBREG QPR:$src1,
1636 (VSETLNi8 (v8i8 (EXTRACT_SUBREG QPR:$src1,
1637 (SubReg_i8_reg imm:$lane))),
1638 GPR:$src2, (SubReg_i8_lane imm:$lane)),
1639 (SubReg_i8_reg imm:$lane)))>;
1640 def : Pat<(vector_insert (v8i16 QPR:$src1), GPR:$src2, imm:$lane),
1641 (v8i16 (INSERT_SUBREG QPR:$src1,
1642 (VSETLNi16 (v4i16 (EXTRACT_SUBREG QPR:$src1,
1643 (SubReg_i16_reg imm:$lane))),
1644 GPR:$src2, (SubReg_i16_lane imm:$lane)),
1645 (SubReg_i16_reg imm:$lane)))>;
1646 def : Pat<(insertelt (v4i32 QPR:$src1), GPR:$src2, imm:$lane),
1647 (v4i32 (INSERT_SUBREG QPR:$src1,
1648 (VSETLNi32 (v2i32 (EXTRACT_SUBREG QPR:$src1,
1649 (SubReg_i32_reg imm:$lane))),
1650 GPR:$src2, (SubReg_i32_lane imm:$lane)),
1651 (SubReg_i32_reg imm:$lane)))>;
1653 //def : Pat<(v2i64 (insertelt QPR:$src1, DPR:$src2, imm:$src3)),
1654 // (INSERT_SUBREG QPR:$src1, DPR:$src2, (SubReg_f64_reg imm:$src3))>;
1655 def : Pat<(v2f64 (insertelt QPR:$src1, DPR:$src2, imm:$src3)),
1656 (INSERT_SUBREG QPR:$src1, DPR:$src2, (SubReg_f64_reg imm:$src3))>;
1658 // VDUP : Vector Duplicate (from ARM core register to all elements)
1660 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
1661 (vector_shuffle node:$lhs, node:$rhs), [{
1662 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1663 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
1666 class VDUPD<bits<8> opcod1, bits<2> opcod3, string asmSize, ValueType Ty>
1667 : NVDup<opcod1, 0b1011, opcod3, (outs DPR:$dst), (ins GPR:$src),
1668 NoItinerary, "vdup", !strconcat(asmSize, "\t$dst, $src"),
1669 [(set DPR:$dst, (Ty (splat_lo (scalar_to_vector GPR:$src), undef)))]>;
1670 class VDUPQ<bits<8> opcod1, bits<2> opcod3, string asmSize, ValueType Ty>
1671 : NVDup<opcod1, 0b1011, opcod3, (outs QPR:$dst), (ins GPR:$src),
1672 NoItinerary, "vdup", !strconcat(asmSize, "\t$dst, $src"),
1673 [(set QPR:$dst, (Ty (splat_lo (scalar_to_vector GPR:$src), undef)))]>;
1675 def VDUP8d : VDUPD<0b11101100, 0b00, ".8", v8i8>;
1676 def VDUP16d : VDUPD<0b11101000, 0b01, ".16", v4i16>;
1677 def VDUP32d : VDUPD<0b11101000, 0b00, ".32", v2i32>;
1678 def VDUP8q : VDUPQ<0b11101110, 0b00, ".8", v16i8>;
1679 def VDUP16q : VDUPQ<0b11101010, 0b01, ".16", v8i16>;
1680 def VDUP32q : VDUPQ<0b11101010, 0b00, ".32", v4i32>;
1682 def VDUPfd : NVDup<0b11101000, 0b1011, 0b00, (outs DPR:$dst), (ins GPR:$src),
1683 NoItinerary, "vdup", ".32\t$dst, $src",
1684 [(set DPR:$dst, (v2f32 (splat_lo
1686 (f32 (bitconvert GPR:$src))),
1688 def VDUPfq : NVDup<0b11101010, 0b1011, 0b00, (outs QPR:$dst), (ins GPR:$src),
1689 NoItinerary, "vdup", ".32\t$dst, $src",
1690 [(set QPR:$dst, (v4f32 (splat_lo
1692 (f32 (bitconvert GPR:$src))),
1695 // VDUP : Vector Duplicate Lane (from scalar to all elements)
1697 def SHUFFLE_get_splat_lane : SDNodeXForm<vector_shuffle, [{
1698 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1699 return CurDAG->getTargetConstant(SVOp->getSplatIndex(), MVT::i32);
1702 def splat_lane : PatFrag<(ops node:$lhs, node:$rhs),
1703 (vector_shuffle node:$lhs, node:$rhs), [{
1704 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1705 return SVOp->isSplat();
1706 }], SHUFFLE_get_splat_lane>;
1708 class VDUPLND<bits<2> op19_18, bits<2> op17_16, string OpcodeStr, ValueType Ty>
1709 : N2V<0b11, 0b11, op19_18, op17_16, 0b11000, 0, 0,
1710 (outs DPR:$dst), (ins DPR:$src, i32imm:$lane), NoItinerary,
1711 !strconcat(OpcodeStr, "\t$dst, $src[$lane]"), "",
1712 [(set DPR:$dst, (Ty (splat_lane:$lane DPR:$src, undef)))]>;
1714 // vector_shuffle requires that the source and destination types match, so
1715 // VDUP to a 128-bit result uses a target-specific VDUPLANEQ node.
1716 class VDUPLNQ<bits<2> op19_18, bits<2> op17_16, string OpcodeStr,
1717 ValueType ResTy, ValueType OpTy>
1718 : N2V<0b11, 0b11, op19_18, op17_16, 0b11000, 1, 0,
1719 (outs QPR:$dst), (ins DPR:$src, i32imm:$lane), NoItinerary,
1720 !strconcat(OpcodeStr, "\t$dst, $src[$lane]"), "",
1721 [(set QPR:$dst, (ResTy (NEONvduplaneq (OpTy DPR:$src), imm:$lane)))]>;
1723 def VDUPLN8d : VDUPLND<0b00, 0b01, "vdup.8", v8i8>;
1724 def VDUPLN16d : VDUPLND<0b00, 0b10, "vdup.16", v4i16>;
1725 def VDUPLN32d : VDUPLND<0b01, 0b00, "vdup.32", v2i32>;
1726 def VDUPLNfd : VDUPLND<0b01, 0b00, "vdup.32", v2f32>;
1727 def VDUPLN8q : VDUPLNQ<0b00, 0b01, "vdup.8", v16i8, v8i8>;
1728 def VDUPLN16q : VDUPLNQ<0b00, 0b10, "vdup.16", v8i16, v4i16>;
1729 def VDUPLN32q : VDUPLNQ<0b01, 0b00, "vdup.32", v4i32, v2i32>;
1730 def VDUPLNfq : VDUPLNQ<0b01, 0b00, "vdup.32", v4f32, v2f32>;
1732 // VMOVN : Vector Narrowing Move
1733 defm VMOVN : N2VNInt_HSD<0b11,0b11,0b10,0b00100,0,0, "vmovn.i",
1734 int_arm_neon_vmovn>;
1735 // VQMOVN : Vector Saturating Narrowing Move
1736 defm VQMOVNs : N2VNInt_HSD<0b11,0b11,0b10,0b00101,0,0, "vqmovn.s",
1737 int_arm_neon_vqmovns>;
1738 defm VQMOVNu : N2VNInt_HSD<0b11,0b11,0b10,0b00101,1,0, "vqmovn.u",
1739 int_arm_neon_vqmovnu>;
1740 defm VQMOVNsu : N2VNInt_HSD<0b11,0b11,0b10,0b00100,1,0, "vqmovun.s",
1741 int_arm_neon_vqmovnsu>;
1742 // VMOVL : Vector Lengthening Move
1743 defm VMOVLs : N2VLInt_QHS<0,1,0b1010,0,0,1, "vmovl.s", int_arm_neon_vmovls>;
1744 defm VMOVLu : N2VLInt_QHS<1,1,0b1010,0,0,1, "vmovl.u", int_arm_neon_vmovlu>;
1746 // Vector Conversions.
1748 // VCVT : Vector Convert Between Floating-Point and Integers
1749 def VCVTf2sd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt.s32.f32",
1750 v2i32, v2f32, fp_to_sint>;
1751 def VCVTf2ud : N2VD<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt.u32.f32",
1752 v2i32, v2f32, fp_to_uint>;
1753 def VCVTs2fd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt.f32.s32",
1754 v2f32, v2i32, sint_to_fp>;
1755 def VCVTu2fd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt.f32.u32",
1756 v2f32, v2i32, uint_to_fp>;
1758 def VCVTf2sq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt.s32.f32",
1759 v4i32, v4f32, fp_to_sint>;
1760 def VCVTf2uq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt.u32.f32",
1761 v4i32, v4f32, fp_to_uint>;
1762 def VCVTs2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt.f32.s32",
1763 v4f32, v4i32, sint_to_fp>;
1764 def VCVTu2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt.f32.u32",
1765 v4f32, v4i32, uint_to_fp>;
1767 // VCVT : Vector Convert Between Floating-Point and Fixed-Point.
1768 // Note: Some of the opcode bits in the following VCVT instructions need to
1769 // be encoded based on the immed values.
1770 def VCVTf2xsd : N2VCvtD<0, 1, 0b000000, 0b1111, 0, 1, "vcvt.s32.f32",
1771 v2i32, v2f32, int_arm_neon_vcvtfp2fxs>;
1772 def VCVTf2xud : N2VCvtD<1, 1, 0b000000, 0b1111, 0, 1, "vcvt.u32.f32",
1773 v2i32, v2f32, int_arm_neon_vcvtfp2fxu>;
1774 def VCVTxs2fd : N2VCvtD<0, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.s32",
1775 v2f32, v2i32, int_arm_neon_vcvtfxs2fp>;
1776 def VCVTxu2fd : N2VCvtD<1, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.u32",
1777 v2f32, v2i32, int_arm_neon_vcvtfxu2fp>;
1779 def VCVTf2xsq : N2VCvtQ<0, 1, 0b000000, 0b1111, 0, 1, "vcvt.s32.f32",
1780 v4i32, v4f32, int_arm_neon_vcvtfp2fxs>;
1781 def VCVTf2xuq : N2VCvtQ<1, 1, 0b000000, 0b1111, 0, 1, "vcvt.u32.f32",
1782 v4i32, v4f32, int_arm_neon_vcvtfp2fxu>;
1783 def VCVTxs2fq : N2VCvtQ<0, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.s32",
1784 v4f32, v4i32, int_arm_neon_vcvtfxs2fp>;
1785 def VCVTxu2fq : N2VCvtQ<1, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.u32",
1786 v4f32, v4i32, int_arm_neon_vcvtfxu2fp>;
1788 // VREV : Vector Reverse
1790 def vrev64_shuffle : PatFrag<(ops node:$in),
1791 (vector_shuffle node:$in, undef), [{
1792 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1793 return ARM::isVREVMask(SVOp, 64);
1796 def vrev32_shuffle : PatFrag<(ops node:$in),
1797 (vector_shuffle node:$in, undef), [{
1798 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1799 return ARM::isVREVMask(SVOp, 32);
1802 def vrev16_shuffle : PatFrag<(ops node:$in),
1803 (vector_shuffle node:$in, undef), [{
1804 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1805 return ARM::isVREVMask(SVOp, 16);
1808 // VREV64 : Vector Reverse elements within 64-bit doublewords
1810 class VREV64D<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1811 : N2V<0b11, 0b11, op19_18, 0b00, 0b00000, 0, 0, (outs DPR:$dst),
1812 (ins DPR:$src), NoItinerary,
1813 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1814 [(set DPR:$dst, (Ty (vrev64_shuffle (Ty DPR:$src))))]>;
1815 class VREV64Q<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1816 : N2V<0b11, 0b11, op19_18, 0b00, 0b00000, 1, 0, (outs QPR:$dst),
1817 (ins QPR:$src), NoItinerary,
1818 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1819 [(set QPR:$dst, (Ty (vrev64_shuffle (Ty QPR:$src))))]>;
1821 def VREV64d8 : VREV64D<0b00, "vrev64.8", v8i8>;
1822 def VREV64d16 : VREV64D<0b01, "vrev64.16", v4i16>;
1823 def VREV64d32 : VREV64D<0b10, "vrev64.32", v2i32>;
1824 def VREV64df : VREV64D<0b10, "vrev64.32", v2f32>;
1826 def VREV64q8 : VREV64Q<0b00, "vrev64.8", v16i8>;
1827 def VREV64q16 : VREV64Q<0b01, "vrev64.16", v8i16>;
1828 def VREV64q32 : VREV64Q<0b10, "vrev64.32", v4i32>;
1829 def VREV64qf : VREV64Q<0b10, "vrev64.32", v4f32>;
1831 // VREV32 : Vector Reverse elements within 32-bit words
1833 class VREV32D<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1834 : N2V<0b11, 0b11, op19_18, 0b00, 0b00001, 0, 0, (outs DPR:$dst),
1835 (ins DPR:$src), NoItinerary,
1836 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1837 [(set DPR:$dst, (Ty (vrev32_shuffle (Ty DPR:$src))))]>;
1838 class VREV32Q<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1839 : N2V<0b11, 0b11, op19_18, 0b00, 0b00001, 1, 0, (outs QPR:$dst),
1840 (ins QPR:$src), NoItinerary,
1841 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1842 [(set QPR:$dst, (Ty (vrev32_shuffle (Ty QPR:$src))))]>;
1844 def VREV32d8 : VREV32D<0b00, "vrev32.8", v8i8>;
1845 def VREV32d16 : VREV32D<0b01, "vrev32.16", v4i16>;
1847 def VREV32q8 : VREV32Q<0b00, "vrev32.8", v16i8>;
1848 def VREV32q16 : VREV32Q<0b01, "vrev32.16", v8i16>;
1850 // VREV16 : Vector Reverse elements within 16-bit halfwords
1852 class VREV16D<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1853 : N2V<0b11, 0b11, op19_18, 0b00, 0b00010, 0, 0, (outs DPR:$dst),
1854 (ins DPR:$src), NoItinerary,
1855 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1856 [(set DPR:$dst, (Ty (vrev16_shuffle (Ty DPR:$src))))]>;
1857 class VREV16Q<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1858 : N2V<0b11, 0b11, op19_18, 0b00, 0b00010, 1, 0, (outs QPR:$dst),
1859 (ins QPR:$src), NoItinerary,
1860 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1861 [(set QPR:$dst, (Ty (vrev16_shuffle (Ty QPR:$src))))]>;
1863 def VREV16d8 : VREV16D<0b00, "vrev16.8", v8i8>;
1864 def VREV16q8 : VREV16Q<0b00, "vrev16.8", v16i8>;
1866 //===----------------------------------------------------------------------===//
1867 // Non-Instruction Patterns
1868 //===----------------------------------------------------------------------===//
1871 def : Pat<(v1i64 (bitconvert (v2i32 DPR:$src))), (v1i64 DPR:$src)>;
1872 def : Pat<(v1i64 (bitconvert (v4i16 DPR:$src))), (v1i64 DPR:$src)>;
1873 def : Pat<(v1i64 (bitconvert (v8i8 DPR:$src))), (v1i64 DPR:$src)>;
1874 def : Pat<(v1i64 (bitconvert (f64 DPR:$src))), (v1i64 DPR:$src)>;
1875 def : Pat<(v1i64 (bitconvert (v2f32 DPR:$src))), (v1i64 DPR:$src)>;
1876 def : Pat<(v2i32 (bitconvert (v1i64 DPR:$src))), (v2i32 DPR:$src)>;
1877 def : Pat<(v2i32 (bitconvert (v4i16 DPR:$src))), (v2i32 DPR:$src)>;
1878 def : Pat<(v2i32 (bitconvert (v8i8 DPR:$src))), (v2i32 DPR:$src)>;
1879 def : Pat<(v2i32 (bitconvert (f64 DPR:$src))), (v2i32 DPR:$src)>;
1880 def : Pat<(v2i32 (bitconvert (v2f32 DPR:$src))), (v2i32 DPR:$src)>;
1881 def : Pat<(v4i16 (bitconvert (v1i64 DPR:$src))), (v4i16 DPR:$src)>;
1882 def : Pat<(v4i16 (bitconvert (v2i32 DPR:$src))), (v4i16 DPR:$src)>;
1883 def : Pat<(v4i16 (bitconvert (v8i8 DPR:$src))), (v4i16 DPR:$src)>;
1884 def : Pat<(v4i16 (bitconvert (f64 DPR:$src))), (v4i16 DPR:$src)>;
1885 def : Pat<(v4i16 (bitconvert (v2f32 DPR:$src))), (v4i16 DPR:$src)>;
1886 def : Pat<(v8i8 (bitconvert (v1i64 DPR:$src))), (v8i8 DPR:$src)>;
1887 def : Pat<(v8i8 (bitconvert (v2i32 DPR:$src))), (v8i8 DPR:$src)>;
1888 def : Pat<(v8i8 (bitconvert (v4i16 DPR:$src))), (v8i8 DPR:$src)>;
1889 def : Pat<(v8i8 (bitconvert (f64 DPR:$src))), (v8i8 DPR:$src)>;
1890 def : Pat<(v8i8 (bitconvert (v2f32 DPR:$src))), (v8i8 DPR:$src)>;
1891 def : Pat<(f64 (bitconvert (v1i64 DPR:$src))), (f64 DPR:$src)>;
1892 def : Pat<(f64 (bitconvert (v2i32 DPR:$src))), (f64 DPR:$src)>;
1893 def : Pat<(f64 (bitconvert (v4i16 DPR:$src))), (f64 DPR:$src)>;
1894 def : Pat<(f64 (bitconvert (v8i8 DPR:$src))), (f64 DPR:$src)>;
1895 def : Pat<(f64 (bitconvert (v2f32 DPR:$src))), (f64 DPR:$src)>;
1896 def : Pat<(v2f32 (bitconvert (f64 DPR:$src))), (v2f32 DPR:$src)>;
1897 def : Pat<(v2f32 (bitconvert (v1i64 DPR:$src))), (v2f32 DPR:$src)>;
1898 def : Pat<(v2f32 (bitconvert (v2i32 DPR:$src))), (v2f32 DPR:$src)>;
1899 def : Pat<(v2f32 (bitconvert (v4i16 DPR:$src))), (v2f32 DPR:$src)>;
1900 def : Pat<(v2f32 (bitconvert (v8i8 DPR:$src))), (v2f32 DPR:$src)>;
1902 def : Pat<(v2i64 (bitconvert (v4i32 QPR:$src))), (v2i64 QPR:$src)>;
1903 def : Pat<(v2i64 (bitconvert (v8i16 QPR:$src))), (v2i64 QPR:$src)>;
1904 def : Pat<(v2i64 (bitconvert (v16i8 QPR:$src))), (v2i64 QPR:$src)>;
1905 def : Pat<(v2i64 (bitconvert (v2f64 QPR:$src))), (v2i64 QPR:$src)>;
1906 def : Pat<(v2i64 (bitconvert (v4f32 QPR:$src))), (v2i64 QPR:$src)>;
1907 def : Pat<(v4i32 (bitconvert (v2i64 QPR:$src))), (v4i32 QPR:$src)>;
1908 def : Pat<(v4i32 (bitconvert (v8i16 QPR:$src))), (v4i32 QPR:$src)>;
1909 def : Pat<(v4i32 (bitconvert (v16i8 QPR:$src))), (v4i32 QPR:$src)>;
1910 def : Pat<(v4i32 (bitconvert (v2f64 QPR:$src))), (v4i32 QPR:$src)>;
1911 def : Pat<(v4i32 (bitconvert (v4f32 QPR:$src))), (v4i32 QPR:$src)>;
1912 def : Pat<(v8i16 (bitconvert (v2i64 QPR:$src))), (v8i16 QPR:$src)>;
1913 def : Pat<(v8i16 (bitconvert (v4i32 QPR:$src))), (v8i16 QPR:$src)>;
1914 def : Pat<(v8i16 (bitconvert (v16i8 QPR:$src))), (v8i16 QPR:$src)>;
1915 def : Pat<(v8i16 (bitconvert (v2f64 QPR:$src))), (v8i16 QPR:$src)>;
1916 def : Pat<(v8i16 (bitconvert (v4f32 QPR:$src))), (v8i16 QPR:$src)>;
1917 def : Pat<(v16i8 (bitconvert (v2i64 QPR:$src))), (v16i8 QPR:$src)>;
1918 def : Pat<(v16i8 (bitconvert (v4i32 QPR:$src))), (v16i8 QPR:$src)>;
1919 def : Pat<(v16i8 (bitconvert (v8i16 QPR:$src))), (v16i8 QPR:$src)>;
1920 def : Pat<(v16i8 (bitconvert (v2f64 QPR:$src))), (v16i8 QPR:$src)>;
1921 def : Pat<(v16i8 (bitconvert (v4f32 QPR:$src))), (v16i8 QPR:$src)>;
1922 def : Pat<(v4f32 (bitconvert (v2i64 QPR:$src))), (v4f32 QPR:$src)>;
1923 def : Pat<(v4f32 (bitconvert (v4i32 QPR:$src))), (v4f32 QPR:$src)>;
1924 def : Pat<(v4f32 (bitconvert (v8i16 QPR:$src))), (v4f32 QPR:$src)>;
1925 def : Pat<(v4f32 (bitconvert (v16i8 QPR:$src))), (v4f32 QPR:$src)>;
1926 def : Pat<(v4f32 (bitconvert (v2f64 QPR:$src))), (v4f32 QPR:$src)>;
1927 def : Pat<(v2f64 (bitconvert (v2i64 QPR:$src))), (v2f64 QPR:$src)>;
1928 def : Pat<(v2f64 (bitconvert (v4i32 QPR:$src))), (v2f64 QPR:$src)>;
1929 def : Pat<(v2f64 (bitconvert (v8i16 QPR:$src))), (v2f64 QPR:$src)>;
1930 def : Pat<(v2f64 (bitconvert (v16i8 QPR:$src))), (v2f64 QPR:$src)>;
1931 def : Pat<(v2f64 (bitconvert (v4f32 QPR:$src))), (v2f64 QPR:$src)>;