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 SDTARMVLD : SDTypeProfile<0, 1, [SDTCisPtrTy<0>]>;
72 def NEONvld2d : SDNode<"ARMISD::VLD2D", SDTARMVLD,
73 [SDNPHasChain, SDNPOutFlag, SDNPMayLoad]>;
74 def NEONvld3d : SDNode<"ARMISD::VLD3D", SDTARMVLD,
75 [SDNPHasChain, SDNPOutFlag, SDNPMayLoad]>;
76 def NEONvld4d : SDNode<"ARMISD::VLD4D", SDTARMVLD,
77 [SDNPHasChain, SDNPOutFlag, SDNPMayLoad]>;
79 //===----------------------------------------------------------------------===//
80 // NEON operand definitions
81 //===----------------------------------------------------------------------===//
83 // addrmode_neonldstm := reg
85 /* TODO: Take advantage of vldm.
86 def addrmode_neonldstm : Operand<i32>,
87 ComplexPattern<i32, 2, "SelectAddrModeNeonLdStM", []> {
88 let PrintMethod = "printAddrNeonLdStMOperand";
89 let MIOperandInfo = (ops GPR, i32imm);
93 //===----------------------------------------------------------------------===//
94 // NEON load / store instructions
95 //===----------------------------------------------------------------------===//
97 /* TODO: Take advantage of vldm.
99 def VLDMD : NI<(outs),
100 (ins addrmode_neonldstm:$addr, reglist:$dst1, variable_ops),
101 "vldm${addr:submode} ${addr:base}, $dst1",
103 let Inst{27-25} = 0b110;
105 let Inst{11-9} = 0b101;
108 def VLDMS : NI<(outs),
109 (ins addrmode_neonldstm:$addr, reglist:$dst1, variable_ops),
110 "vldm${addr:submode} ${addr:base}, $dst1",
112 let Inst{27-25} = 0b110;
114 let Inst{11-9} = 0b101;
119 // Use vldmia to load a Q register as a D register pair.
120 def VLDRQ : NI<(outs QPR:$dst), (ins GPR:$addr),
121 "vldmia $addr, ${dst:dregpair}",
122 [(set QPR:$dst, (v2f64 (load GPR:$addr)))]> {
123 let Inst{27-25} = 0b110;
124 let Inst{24} = 0; // P bit
125 let Inst{23} = 1; // U bit
127 let Inst{11-9} = 0b101;
130 // Use vstmia to store a Q register as a D register pair.
131 def VSTRQ : NI<(outs), (ins QPR:$src, GPR:$addr),
132 "vstmia $addr, ${src:dregpair}",
133 [(store (v2f64 QPR:$src), GPR:$addr)]> {
134 let Inst{27-25} = 0b110;
135 let Inst{24} = 0; // P bit
136 let Inst{23} = 1; // U bit
138 let Inst{11-9} = 0b101;
142 // VLD1 : Vector Load (multiple single elements)
143 class VLD1D<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
144 : NLdSt<(outs DPR:$dst), (ins addrmode6:$addr),
145 !strconcat(OpcodeStr, "\t${dst:dregsingle}, $addr"),
146 [(set DPR:$dst, (Ty (IntOp addrmode6:$addr)))]>;
147 class VLD1Q<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
148 : NLdSt<(outs QPR:$dst), (ins addrmode6:$addr),
149 !strconcat(OpcodeStr, "\t${dst:dregpair}, $addr"),
150 [(set QPR:$dst, (Ty (IntOp addrmode6:$addr)))]>;
152 def VLD1d8 : VLD1D<"vld1.8", v8i8, int_arm_neon_vld1i>;
153 def VLD1d16 : VLD1D<"vld1.16", v4i16, int_arm_neon_vld1i>;
154 def VLD1d32 : VLD1D<"vld1.32", v2i32, int_arm_neon_vld1i>;
155 def VLD1df : VLD1D<"vld1.32", v2f32, int_arm_neon_vld1f>;
156 def VLD1d64 : VLD1D<"vld1.64", v1i64, int_arm_neon_vld1i>;
158 def VLD1q8 : VLD1Q<"vld1.8", v16i8, int_arm_neon_vld1i>;
159 def VLD1q16 : VLD1Q<"vld1.16", v8i16, int_arm_neon_vld1i>;
160 def VLD1q32 : VLD1Q<"vld1.32", v4i32, int_arm_neon_vld1i>;
161 def VLD1qf : VLD1Q<"vld1.32", v4f32, int_arm_neon_vld1f>;
162 def VLD1q64 : VLD1Q<"vld1.64", v2i64, int_arm_neon_vld1i>;
164 // VST1 : Vector Store (multiple single elements)
165 class VST1D<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
166 : NLdSt<(outs), (ins addrmode6:$addr, DPR:$src),
167 !strconcat(OpcodeStr, "\t${src:dregsingle}, $addr"),
168 [(IntOp addrmode6:$addr, (Ty DPR:$src))]>;
169 class VST1Q<string OpcodeStr, ValueType Ty, Intrinsic IntOp>
170 : NLdSt<(outs), (ins addrmode6:$addr, QPR:$src),
171 !strconcat(OpcodeStr, "\t${src:dregpair}, $addr"),
172 [(IntOp addrmode6:$addr, (Ty QPR:$src))]>;
174 def VST1d8 : VST1D<"vst1.8", v8i8, int_arm_neon_vst1i>;
175 def VST1d16 : VST1D<"vst1.16", v4i16, int_arm_neon_vst1i>;
176 def VST1d32 : VST1D<"vst1.32", v2i32, int_arm_neon_vst1i>;
177 def VST1df : VST1D<"vst1.32", v2f32, int_arm_neon_vst1f>;
178 def VST1d64 : VST1D<"vst1.64", v1i64, int_arm_neon_vst1i>;
180 def VST1q8 : VST1Q<"vst1.8", v16i8, int_arm_neon_vst1i>;
181 def VST1q16 : VST1Q<"vst1.16", v8i16, int_arm_neon_vst1i>;
182 def VST1q32 : VST1Q<"vst1.32", v4i32, int_arm_neon_vst1i>;
183 def VST1qf : VST1Q<"vst1.32", v4f32, int_arm_neon_vst1f>;
184 def VST1q64 : VST1Q<"vst1.64", v2i64, int_arm_neon_vst1i>;
187 //===----------------------------------------------------------------------===//
188 // NEON pattern fragments
189 //===----------------------------------------------------------------------===//
191 // Extract D sub-registers of Q registers.
192 // (arm_dsubreg_0 is 5; arm_dsubreg_1 is 6)
193 def SubReg_i8_reg : SDNodeXForm<imm, [{
194 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 8, MVT::i32);
196 def SubReg_i16_reg : SDNodeXForm<imm, [{
197 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 4, MVT::i32);
199 def SubReg_i32_reg : SDNodeXForm<imm, [{
200 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 2, MVT::i32);
202 def SubReg_f64_reg : SDNodeXForm<imm, [{
203 return CurDAG->getTargetConstant(5 + N->getZExtValue(), MVT::i32);
206 // Translate lane numbers from Q registers to D subregs.
207 def SubReg_i8_lane : SDNodeXForm<imm, [{
208 return CurDAG->getTargetConstant(N->getZExtValue() & 7, MVT::i32);
210 def SubReg_i16_lane : SDNodeXForm<imm, [{
211 return CurDAG->getTargetConstant(N->getZExtValue() & 3, MVT::i32);
213 def SubReg_i32_lane : SDNodeXForm<imm, [{
214 return CurDAG->getTargetConstant(N->getZExtValue() & 1, MVT::i32);
217 //===----------------------------------------------------------------------===//
218 // Instruction Classes
219 //===----------------------------------------------------------------------===//
221 // Basic 2-register operations, both double- and quad-register.
222 class N2VD<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
223 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
224 ValueType ResTy, ValueType OpTy, SDNode OpNode>
225 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
226 (ins DPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
227 [(set DPR:$dst, (ResTy (OpNode (OpTy DPR:$src))))]>;
228 class N2VQ<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
229 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
230 ValueType ResTy, ValueType OpTy, SDNode OpNode>
231 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
232 (ins QPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
233 [(set QPR:$dst, (ResTy (OpNode (OpTy QPR:$src))))]>;
235 // Basic 2-register intrinsics, both double- and quad-register.
236 class N2VDInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
237 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
238 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
239 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
240 (ins DPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
241 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src))))]>;
242 class N2VQInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
243 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
244 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
245 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
246 (ins QPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
247 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src))))]>;
249 // Narrow 2-register intrinsics.
250 class N2VNInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
251 bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
252 string OpcodeStr, ValueType TyD, ValueType TyQ, Intrinsic IntOp>
253 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, op6, op4, (outs DPR:$dst),
254 (ins QPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
255 [(set DPR:$dst, (TyD (IntOp (TyQ QPR:$src))))]>;
257 // Long 2-register intrinsics. (This is currently only used for VMOVL and is
258 // derived from N2VImm instead of N2V because of the way the size is encoded.)
259 class N2VLInt<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
260 bit op6, bit op4, string OpcodeStr, ValueType TyQ, ValueType TyD,
262 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4, (outs QPR:$dst),
263 (ins DPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
264 [(set QPR:$dst, (TyQ (IntOp (TyD DPR:$src))))]>;
266 // Basic 3-register operations, both double- and quad-register.
267 class N3VD<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
268 string OpcodeStr, ValueType ResTy, ValueType OpTy,
269 SDNode OpNode, bit Commutable>
270 : N3V<op24, op23, op21_20, op11_8, 0, op4,
271 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2),
272 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
273 [(set DPR:$dst, (ResTy (OpNode (OpTy DPR:$src1), (OpTy DPR:$src2))))]> {
274 let isCommutable = Commutable;
276 class N3VQ<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
277 string OpcodeStr, ValueType ResTy, ValueType OpTy,
278 SDNode OpNode, bit Commutable>
279 : N3V<op24, op23, op21_20, op11_8, 1, op4,
280 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2),
281 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
282 [(set QPR:$dst, (ResTy (OpNode (OpTy QPR:$src1), (OpTy QPR:$src2))))]> {
283 let isCommutable = Commutable;
286 // Basic 3-register operations, scalar single-precision
287 class N3VDs<SDNode OpNode, NeonI Inst>
288 : NEONFPPat<(f32 (OpNode SPR:$a, SPR:$b)),
289 (EXTRACT_SUBREG (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$a, arm_ssubreg_0),
290 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$b, arm_ssubreg_0)),
293 // Basic 3-register intrinsics, both double- and quad-register.
294 class N3VDInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
295 string OpcodeStr, ValueType ResTy, ValueType OpTy,
296 Intrinsic IntOp, bit Commutable>
297 : N3V<op24, op23, op21_20, op11_8, 0, op4,
298 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2),
299 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
300 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src1), (OpTy DPR:$src2))))]> {
301 let isCommutable = Commutable;
303 class N3VQInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
304 string OpcodeStr, ValueType ResTy, ValueType OpTy,
305 Intrinsic IntOp, bit Commutable>
306 : N3V<op24, op23, op21_20, op11_8, 1, op4,
307 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2),
308 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
309 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src1), (OpTy QPR:$src2))))]> {
310 let isCommutable = Commutable;
313 // Multiply-Add/Sub operations, both double- and quad-register.
314 class N3VDMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
315 string OpcodeStr, ValueType Ty, SDNode MulOp, SDNode OpNode>
316 : N3V<op24, op23, op21_20, op11_8, 0, op4,
317 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, DPR:$src3),
318 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
319 [(set DPR:$dst, (Ty (OpNode DPR:$src1,
320 (Ty (MulOp DPR:$src2, DPR:$src3)))))]>;
321 class N3VQMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
322 string OpcodeStr, ValueType Ty, SDNode MulOp, SDNode OpNode>
323 : N3V<op24, op23, op21_20, op11_8, 1, op4,
324 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, QPR:$src3),
325 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
326 [(set QPR:$dst, (Ty (OpNode QPR:$src1,
327 (Ty (MulOp QPR:$src2, QPR:$src3)))))]>;
329 // Multiply-Add/Sub operations, scalar single-precision
330 class N3VDMulOps<SDNode MulNode, SDNode OpNode, NeonI Inst>
331 : NEONFPPat<(f32 (OpNode SPR:$acc,
332 (f32 (MulNode SPR:$a, SPR:$b)))),
333 (EXTRACT_SUBREG (Inst (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$acc, arm_ssubreg_0),
334 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$a, arm_ssubreg_0),
335 (INSERT_SUBREG (v2f32 (IMPLICIT_DEF)), SPR:$b, arm_ssubreg_0)),
338 // Neon 3-argument intrinsics, both double- and quad-register.
339 // The destination register is also used as the first source operand register.
340 class N3VDInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
341 string OpcodeStr, ValueType ResTy, ValueType OpTy,
343 : N3V<op24, op23, op21_20, op11_8, 0, op4,
344 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, DPR:$src3),
345 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
346 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src1),
347 (OpTy DPR:$src2), (OpTy DPR:$src3))))]>;
348 class N3VQInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
349 string OpcodeStr, ValueType ResTy, ValueType OpTy,
351 : N3V<op24, op23, op21_20, op11_8, 1, op4,
352 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, QPR:$src3),
353 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
354 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src1),
355 (OpTy QPR:$src2), (OpTy QPR:$src3))))]>;
357 // Neon Long 3-argument intrinsic. The destination register is
358 // a quad-register and is also used as the first source operand register.
359 class N3VLInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
360 string OpcodeStr, ValueType TyQ, ValueType TyD, Intrinsic IntOp>
361 : N3V<op24, op23, op21_20, op11_8, 0, op4,
362 (outs QPR:$dst), (ins QPR:$src1, DPR:$src2, DPR:$src3),
363 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
365 (TyQ (IntOp (TyQ QPR:$src1), (TyD DPR:$src2), (TyD DPR:$src3))))]>;
367 // Narrowing 3-register intrinsics.
368 class N3VNInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
369 string OpcodeStr, ValueType TyD, ValueType TyQ,
370 Intrinsic IntOp, bit Commutable>
371 : N3V<op24, op23, op21_20, op11_8, 0, op4,
372 (outs DPR:$dst), (ins QPR:$src1, QPR:$src2),
373 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
374 [(set DPR:$dst, (TyD (IntOp (TyQ QPR:$src1), (TyQ QPR:$src2))))]> {
375 let isCommutable = Commutable;
378 // Long 3-register intrinsics.
379 class N3VLInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
380 string OpcodeStr, ValueType TyQ, ValueType TyD,
381 Intrinsic IntOp, bit Commutable>
382 : N3V<op24, op23, op21_20, op11_8, 0, op4,
383 (outs QPR:$dst), (ins DPR:$src1, DPR:$src2),
384 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
385 [(set QPR:$dst, (TyQ (IntOp (TyD DPR:$src1), (TyD DPR:$src2))))]> {
386 let isCommutable = Commutable;
389 // Wide 3-register intrinsics.
390 class N3VWInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
391 string OpcodeStr, ValueType TyQ, ValueType TyD,
392 Intrinsic IntOp, bit Commutable>
393 : N3V<op24, op23, op21_20, op11_8, 0, op4,
394 (outs QPR:$dst), (ins QPR:$src1, DPR:$src2),
395 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
396 [(set QPR:$dst, (TyQ (IntOp (TyQ QPR:$src1), (TyD DPR:$src2))))]> {
397 let isCommutable = Commutable;
400 // Pairwise long 2-register intrinsics, both double- and quad-register.
401 class N2VDPLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
402 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
403 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
404 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
405 (ins DPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
406 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src))))]>;
407 class N2VQPLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
408 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
409 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
410 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
411 (ins QPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
412 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src))))]>;
414 // Pairwise long 2-register accumulate intrinsics,
415 // both double- and quad-register.
416 // The destination register is also used as the first source operand register.
417 class N2VDPLInt2<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
418 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
419 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
420 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4,
421 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2),
422 !strconcat(OpcodeStr, "\t$dst, $src2"), "$src1 = $dst",
423 [(set DPR:$dst, (ResTy (IntOp (ResTy DPR:$src1), (OpTy DPR:$src2))))]>;
424 class N2VQPLInt2<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
425 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
426 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
427 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4,
428 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2),
429 !strconcat(OpcodeStr, "\t$dst, $src2"), "$src1 = $dst",
430 [(set QPR:$dst, (ResTy (IntOp (ResTy QPR:$src1), (OpTy QPR:$src2))))]>;
432 // Shift by immediate,
433 // both double- and quad-register.
434 class N2VDSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
435 bit op4, string OpcodeStr, ValueType Ty, SDNode OpNode>
436 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
437 (outs DPR:$dst), (ins DPR:$src, i32imm:$SIMM),
438 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
439 [(set DPR:$dst, (Ty (OpNode (Ty DPR:$src), (i32 imm:$SIMM))))]>;
440 class N2VQSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
441 bit op4, string OpcodeStr, ValueType Ty, SDNode OpNode>
442 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
443 (outs QPR:$dst), (ins QPR:$src, i32imm:$SIMM),
444 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
445 [(set QPR:$dst, (Ty (OpNode (Ty QPR:$src), (i32 imm:$SIMM))))]>;
447 // Long shift by immediate.
448 class N2VLSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
449 bit op6, bit op4, string OpcodeStr, ValueType ResTy,
450 ValueType OpTy, SDNode OpNode>
451 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4,
452 (outs QPR:$dst), (ins DPR:$src, i32imm:$SIMM),
453 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
454 [(set QPR:$dst, (ResTy (OpNode (OpTy DPR:$src),
455 (i32 imm:$SIMM))))]>;
457 // Narrow shift by immediate.
458 class N2VNSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
459 bit op6, bit op4, string OpcodeStr, ValueType ResTy,
460 ValueType OpTy, SDNode OpNode>
461 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4,
462 (outs DPR:$dst), (ins QPR:$src, i32imm:$SIMM),
463 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
464 [(set DPR:$dst, (ResTy (OpNode (OpTy QPR:$src),
465 (i32 imm:$SIMM))))]>;
467 // Shift right by immediate and accumulate,
468 // both double- and quad-register.
469 class N2VDShAdd<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
470 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
471 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
472 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, i32imm:$SIMM),
473 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
474 [(set DPR:$dst, (Ty (add DPR:$src1,
475 (Ty (ShOp DPR:$src2, (i32 imm:$SIMM))))))]>;
476 class N2VQShAdd<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
477 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
478 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
479 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, i32imm:$SIMM),
480 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
481 [(set QPR:$dst, (Ty (add QPR:$src1,
482 (Ty (ShOp QPR:$src2, (i32 imm:$SIMM))))))]>;
484 // Shift by immediate and insert,
485 // both double- and quad-register.
486 class N2VDShIns<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
487 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
488 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
489 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, i32imm:$SIMM),
490 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
491 [(set DPR:$dst, (Ty (ShOp DPR:$src1, DPR:$src2, (i32 imm:$SIMM))))]>;
492 class N2VQShIns<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
493 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
494 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
495 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, i32imm:$SIMM),
496 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
497 [(set QPR:$dst, (Ty (ShOp QPR:$src1, QPR:$src2, (i32 imm:$SIMM))))]>;
499 // Convert, with fractional bits immediate,
500 // both double- and quad-register.
501 class N2VCvtD<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
502 bit op4, string OpcodeStr, ValueType ResTy, ValueType OpTy,
504 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
505 (outs DPR:$dst), (ins DPR:$src, i32imm:$SIMM),
506 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
507 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src), (i32 imm:$SIMM))))]>;
508 class N2VCvtQ<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
509 bit op4, string OpcodeStr, ValueType ResTy, ValueType OpTy,
511 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
512 (outs QPR:$dst), (ins QPR:$src, i32imm:$SIMM),
513 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
514 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src), (i32 imm:$SIMM))))]>;
516 //===----------------------------------------------------------------------===//
518 //===----------------------------------------------------------------------===//
520 // Neon 3-register vector operations.
522 // First with only element sizes of 8, 16 and 32 bits:
523 multiclass N3V_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
524 string OpcodeStr, SDNode OpNode, bit Commutable = 0> {
525 // 64-bit vector types.
526 def v8i8 : N3VD<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
527 v8i8, v8i8, OpNode, Commutable>;
528 def v4i16 : N3VD<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr, "16"),
529 v4i16, v4i16, OpNode, Commutable>;
530 def v2i32 : N3VD<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr, "32"),
531 v2i32, v2i32, OpNode, Commutable>;
533 // 128-bit vector types.
534 def v16i8 : N3VQ<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
535 v16i8, v16i8, OpNode, Commutable>;
536 def v8i16 : N3VQ<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr, "16"),
537 v8i16, v8i16, OpNode, Commutable>;
538 def v4i32 : N3VQ<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr, "32"),
539 v4i32, v4i32, OpNode, Commutable>;
542 // ....then also with element size 64 bits:
543 multiclass N3V_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
544 string OpcodeStr, SDNode OpNode, bit Commutable = 0>
545 : N3V_QHS<op24, op23, op11_8, op4, OpcodeStr, OpNode, Commutable> {
546 def v1i64 : N3VD<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr, "64"),
547 v1i64, v1i64, OpNode, Commutable>;
548 def v2i64 : N3VQ<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr, "64"),
549 v2i64, v2i64, OpNode, Commutable>;
553 // Neon Narrowing 2-register vector intrinsics,
554 // source operand element sizes of 16, 32 and 64 bits:
555 multiclass N2VNInt_HSD<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
556 bits<5> op11_7, bit op6, bit op4, string OpcodeStr,
558 def v8i8 : N2VNInt<op24_23, op21_20, 0b00, op17_16, op11_7, op6, op4,
559 !strconcat(OpcodeStr, "16"), v8i8, v8i16, IntOp>;
560 def v4i16 : N2VNInt<op24_23, op21_20, 0b01, op17_16, op11_7, op6, op4,
561 !strconcat(OpcodeStr, "32"), v4i16, v4i32, IntOp>;
562 def v2i32 : N2VNInt<op24_23, op21_20, 0b10, op17_16, op11_7, op6, op4,
563 !strconcat(OpcodeStr, "64"), v2i32, v2i64, IntOp>;
567 // Neon Lengthening 2-register vector intrinsic (currently specific to VMOVL).
568 // source operand element sizes of 16, 32 and 64 bits:
569 multiclass N2VLInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op7, bit op6,
570 bit op4, string OpcodeStr, Intrinsic IntOp> {
571 def v8i16 : N2VLInt<op24, op23, 0b001000, op11_8, op7, op6, op4,
572 !strconcat(OpcodeStr, "8"), v8i16, v8i8, IntOp>;
573 def v4i32 : N2VLInt<op24, op23, 0b010000, op11_8, op7, op6, op4,
574 !strconcat(OpcodeStr, "16"), v4i32, v4i16, IntOp>;
575 def v2i64 : N2VLInt<op24, op23, 0b100000, op11_8, op7, op6, op4,
576 !strconcat(OpcodeStr, "32"), v2i64, v2i32, IntOp>;
580 // Neon 3-register vector intrinsics.
582 // First with only element sizes of 16 and 32 bits:
583 multiclass N3VInt_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
584 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
585 // 64-bit vector types.
586 def v4i16 : N3VDInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
587 v4i16, v4i16, IntOp, Commutable>;
588 def v2i32 : N3VDInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
589 v2i32, v2i32, IntOp, Commutable>;
591 // 128-bit vector types.
592 def v8i16 : N3VQInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
593 v8i16, v8i16, IntOp, Commutable>;
594 def v4i32 : N3VQInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
595 v4i32, v4i32, IntOp, Commutable>;
598 // ....then also with element size of 8 bits:
599 multiclass N3VInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
600 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
601 : N3VInt_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
602 def v8i8 : N3VDInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
603 v8i8, v8i8, IntOp, Commutable>;
604 def v16i8 : N3VQInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
605 v16i8, v16i8, IntOp, Commutable>;
608 // ....then also with element size of 64 bits:
609 multiclass N3VInt_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
610 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
611 : N3VInt_QHS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
612 def v1i64 : N3VDInt<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr,"64"),
613 v1i64, v1i64, IntOp, Commutable>;
614 def v2i64 : N3VQInt<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr,"64"),
615 v2i64, v2i64, IntOp, Commutable>;
619 // Neon Narrowing 3-register vector intrinsics,
620 // source operand element sizes of 16, 32 and 64 bits:
621 multiclass N3VNInt_HSD<bit op24, bit op23, bits<4> op11_8, bit op4,
622 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
623 def v8i8 : N3VNInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr,"16"),
624 v8i8, v8i16, IntOp, Commutable>;
625 def v4i16 : N3VNInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"32"),
626 v4i16, v4i32, IntOp, Commutable>;
627 def v2i32 : N3VNInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"64"),
628 v2i32, v2i64, IntOp, Commutable>;
632 // Neon Long 3-register vector intrinsics.
634 // First with only element sizes of 16 and 32 bits:
635 multiclass N3VLInt_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
636 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
637 def v4i32 : N3VLInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
638 v4i32, v4i16, IntOp, Commutable>;
639 def v2i64 : N3VLInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
640 v2i64, v2i32, IntOp, Commutable>;
643 // ....then also with element size of 8 bits:
644 multiclass N3VLInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
645 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
646 : N3VLInt_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
647 def v8i16 : N3VLInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
648 v8i16, v8i8, IntOp, Commutable>;
652 // Neon Wide 3-register vector intrinsics,
653 // source operand element sizes of 8, 16 and 32 bits:
654 multiclass N3VWInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
655 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
656 def v8i16 : N3VWInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
657 v8i16, v8i8, IntOp, Commutable>;
658 def v4i32 : N3VWInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
659 v4i32, v4i16, IntOp, Commutable>;
660 def v2i64 : N3VWInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
661 v2i64, v2i32, IntOp, Commutable>;
665 // Neon Multiply-Op vector operations,
666 // element sizes of 8, 16 and 32 bits:
667 multiclass N3VMulOp_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
668 string OpcodeStr, SDNode OpNode> {
669 // 64-bit vector types.
670 def v8i8 : N3VDMulOp<op24, op23, 0b00, op11_8, op4,
671 !strconcat(OpcodeStr, "8"), v8i8, mul, OpNode>;
672 def v4i16 : N3VDMulOp<op24, op23, 0b01, op11_8, op4,
673 !strconcat(OpcodeStr, "16"), v4i16, mul, OpNode>;
674 def v2i32 : N3VDMulOp<op24, op23, 0b10, op11_8, op4,
675 !strconcat(OpcodeStr, "32"), v2i32, mul, OpNode>;
677 // 128-bit vector types.
678 def v16i8 : N3VQMulOp<op24, op23, 0b00, op11_8, op4,
679 !strconcat(OpcodeStr, "8"), v16i8, mul, OpNode>;
680 def v8i16 : N3VQMulOp<op24, op23, 0b01, op11_8, op4,
681 !strconcat(OpcodeStr, "16"), v8i16, mul, OpNode>;
682 def v4i32 : N3VQMulOp<op24, op23, 0b10, op11_8, op4,
683 !strconcat(OpcodeStr, "32"), v4i32, mul, OpNode>;
687 // Neon 3-argument intrinsics,
688 // element sizes of 8, 16 and 32 bits:
689 multiclass N3VInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
690 string OpcodeStr, Intrinsic IntOp> {
691 // 64-bit vector types.
692 def v8i8 : N3VDInt3<op24, op23, 0b00, op11_8, op4,
693 !strconcat(OpcodeStr, "8"), v8i8, v8i8, IntOp>;
694 def v4i16 : N3VDInt3<op24, op23, 0b01, op11_8, op4,
695 !strconcat(OpcodeStr, "16"), v4i16, v4i16, IntOp>;
696 def v2i32 : N3VDInt3<op24, op23, 0b10, op11_8, op4,
697 !strconcat(OpcodeStr, "32"), v2i32, v2i32, IntOp>;
699 // 128-bit vector types.
700 def v16i8 : N3VQInt3<op24, op23, 0b00, op11_8, op4,
701 !strconcat(OpcodeStr, "8"), v16i8, v16i8, IntOp>;
702 def v8i16 : N3VQInt3<op24, op23, 0b01, op11_8, op4,
703 !strconcat(OpcodeStr, "16"), v8i16, v8i16, IntOp>;
704 def v4i32 : N3VQInt3<op24, op23, 0b10, op11_8, op4,
705 !strconcat(OpcodeStr, "32"), v4i32, v4i32, IntOp>;
709 // Neon Long 3-argument intrinsics.
711 // First with only element sizes of 16 and 32 bits:
712 multiclass N3VLInt3_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
713 string OpcodeStr, Intrinsic IntOp> {
714 def v4i32 : N3VLInt3<op24, op23, 0b01, op11_8, op4,
715 !strconcat(OpcodeStr, "16"), v4i32, v4i16, IntOp>;
716 def v2i64 : N3VLInt3<op24, op23, 0b10, op11_8, op4,
717 !strconcat(OpcodeStr, "32"), v2i64, v2i32, IntOp>;
720 // ....then also with element size of 8 bits:
721 multiclass N3VLInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
722 string OpcodeStr, Intrinsic IntOp>
723 : N3VLInt3_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp> {
724 def v8i16 : N3VLInt3<op24, op23, 0b01, op11_8, op4,
725 !strconcat(OpcodeStr, "8"), v8i16, v8i8, IntOp>;
729 // Neon 2-register vector intrinsics,
730 // element sizes of 8, 16 and 32 bits:
731 multiclass N2VInt_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
732 bits<5> op11_7, bit op4, string OpcodeStr,
734 // 64-bit vector types.
735 def v8i8 : N2VDInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
736 !strconcat(OpcodeStr, "8"), v8i8, v8i8, IntOp>;
737 def v4i16 : N2VDInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
738 !strconcat(OpcodeStr, "16"), v4i16, v4i16, IntOp>;
739 def v2i32 : N2VDInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
740 !strconcat(OpcodeStr, "32"), v2i32, v2i32, IntOp>;
742 // 128-bit vector types.
743 def v16i8 : N2VQInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
744 !strconcat(OpcodeStr, "8"), v16i8, v16i8, IntOp>;
745 def v8i16 : N2VQInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
746 !strconcat(OpcodeStr, "16"), v8i16, v8i16, IntOp>;
747 def v4i32 : N2VQInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
748 !strconcat(OpcodeStr, "32"), v4i32, v4i32, IntOp>;
752 // Neon Pairwise long 2-register intrinsics,
753 // element sizes of 8, 16 and 32 bits:
754 multiclass N2VPLInt_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
755 bits<5> op11_7, bit op4,
756 string OpcodeStr, Intrinsic IntOp> {
757 // 64-bit vector types.
758 def v8i8 : N2VDPLInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
759 !strconcat(OpcodeStr, "8"), v4i16, v8i8, IntOp>;
760 def v4i16 : N2VDPLInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
761 !strconcat(OpcodeStr, "16"), v2i32, v4i16, IntOp>;
762 def v2i32 : N2VDPLInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
763 !strconcat(OpcodeStr, "32"), v1i64, v2i32, IntOp>;
765 // 128-bit vector types.
766 def v16i8 : N2VQPLInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
767 !strconcat(OpcodeStr, "8"), v8i16, v16i8, IntOp>;
768 def v8i16 : N2VQPLInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
769 !strconcat(OpcodeStr, "16"), v4i32, v8i16, IntOp>;
770 def v4i32 : N2VQPLInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
771 !strconcat(OpcodeStr, "32"), v2i64, v4i32, IntOp>;
775 // Neon Pairwise long 2-register accumulate intrinsics,
776 // element sizes of 8, 16 and 32 bits:
777 multiclass N2VPLInt2_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
778 bits<5> op11_7, bit op4,
779 string OpcodeStr, Intrinsic IntOp> {
780 // 64-bit vector types.
781 def v8i8 : N2VDPLInt2<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
782 !strconcat(OpcodeStr, "8"), v4i16, v8i8, IntOp>;
783 def v4i16 : N2VDPLInt2<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
784 !strconcat(OpcodeStr, "16"), v2i32, v4i16, IntOp>;
785 def v2i32 : N2VDPLInt2<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
786 !strconcat(OpcodeStr, "32"), v1i64, v2i32, IntOp>;
788 // 128-bit vector types.
789 def v16i8 : N2VQPLInt2<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
790 !strconcat(OpcodeStr, "8"), v8i16, v16i8, IntOp>;
791 def v8i16 : N2VQPLInt2<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
792 !strconcat(OpcodeStr, "16"), v4i32, v8i16, IntOp>;
793 def v4i32 : N2VQPLInt2<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
794 !strconcat(OpcodeStr, "32"), v2i64, v4i32, IntOp>;
798 // Neon 2-register vector shift by immediate,
799 // element sizes of 8, 16, 32 and 64 bits:
800 multiclass N2VSh_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
801 string OpcodeStr, SDNode OpNode> {
802 // 64-bit vector types.
803 def v8i8 : N2VDSh<op24, op23, 0b001000, op11_8, 0, op4,
804 !strconcat(OpcodeStr, "8"), v8i8, OpNode>;
805 def v4i16 : N2VDSh<op24, op23, 0b010000, op11_8, 0, op4,
806 !strconcat(OpcodeStr, "16"), v4i16, OpNode>;
807 def v2i32 : N2VDSh<op24, op23, 0b100000, op11_8, 0, op4,
808 !strconcat(OpcodeStr, "32"), v2i32, OpNode>;
809 def v1i64 : N2VDSh<op24, op23, 0b000000, op11_8, 1, op4,
810 !strconcat(OpcodeStr, "64"), v1i64, OpNode>;
812 // 128-bit vector types.
813 def v16i8 : N2VQSh<op24, op23, 0b001000, op11_8, 0, op4,
814 !strconcat(OpcodeStr, "8"), v16i8, OpNode>;
815 def v8i16 : N2VQSh<op24, op23, 0b010000, op11_8, 0, op4,
816 !strconcat(OpcodeStr, "16"), v8i16, OpNode>;
817 def v4i32 : N2VQSh<op24, op23, 0b100000, op11_8, 0, op4,
818 !strconcat(OpcodeStr, "32"), v4i32, OpNode>;
819 def v2i64 : N2VQSh<op24, op23, 0b000000, op11_8, 1, op4,
820 !strconcat(OpcodeStr, "64"), v2i64, OpNode>;
824 // Neon Shift-Accumulate vector operations,
825 // element sizes of 8, 16, 32 and 64 bits:
826 multiclass N2VShAdd_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
827 string OpcodeStr, SDNode ShOp> {
828 // 64-bit vector types.
829 def v8i8 : N2VDShAdd<op24, op23, 0b001000, op11_8, 0, op4,
830 !strconcat(OpcodeStr, "8"), v8i8, ShOp>;
831 def v4i16 : N2VDShAdd<op24, op23, 0b010000, op11_8, 0, op4,
832 !strconcat(OpcodeStr, "16"), v4i16, ShOp>;
833 def v2i32 : N2VDShAdd<op24, op23, 0b100000, op11_8, 0, op4,
834 !strconcat(OpcodeStr, "32"), v2i32, ShOp>;
835 def v1i64 : N2VDShAdd<op24, op23, 0b000000, op11_8, 1, op4,
836 !strconcat(OpcodeStr, "64"), v1i64, ShOp>;
838 // 128-bit vector types.
839 def v16i8 : N2VQShAdd<op24, op23, 0b001000, op11_8, 0, op4,
840 !strconcat(OpcodeStr, "8"), v16i8, ShOp>;
841 def v8i16 : N2VQShAdd<op24, op23, 0b010000, op11_8, 0, op4,
842 !strconcat(OpcodeStr, "16"), v8i16, ShOp>;
843 def v4i32 : N2VQShAdd<op24, op23, 0b100000, op11_8, 0, op4,
844 !strconcat(OpcodeStr, "32"), v4i32, ShOp>;
845 def v2i64 : N2VQShAdd<op24, op23, 0b000000, op11_8, 1, op4,
846 !strconcat(OpcodeStr, "64"), v2i64, ShOp>;
850 // Neon Shift-Insert vector operations,
851 // element sizes of 8, 16, 32 and 64 bits:
852 multiclass N2VShIns_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
853 string OpcodeStr, SDNode ShOp> {
854 // 64-bit vector types.
855 def v8i8 : N2VDShIns<op24, op23, 0b001000, op11_8, 0, op4,
856 !strconcat(OpcodeStr, "8"), v8i8, ShOp>;
857 def v4i16 : N2VDShIns<op24, op23, 0b010000, op11_8, 0, op4,
858 !strconcat(OpcodeStr, "16"), v4i16, ShOp>;
859 def v2i32 : N2VDShIns<op24, op23, 0b100000, op11_8, 0, op4,
860 !strconcat(OpcodeStr, "32"), v2i32, ShOp>;
861 def v1i64 : N2VDShIns<op24, op23, 0b000000, op11_8, 1, op4,
862 !strconcat(OpcodeStr, "64"), v1i64, ShOp>;
864 // 128-bit vector types.
865 def v16i8 : N2VQShIns<op24, op23, 0b001000, op11_8, 0, op4,
866 !strconcat(OpcodeStr, "8"), v16i8, ShOp>;
867 def v8i16 : N2VQShIns<op24, op23, 0b010000, op11_8, 0, op4,
868 !strconcat(OpcodeStr, "16"), v8i16, ShOp>;
869 def v4i32 : N2VQShIns<op24, op23, 0b100000, op11_8, 0, op4,
870 !strconcat(OpcodeStr, "32"), v4i32, ShOp>;
871 def v2i64 : N2VQShIns<op24, op23, 0b000000, op11_8, 1, op4,
872 !strconcat(OpcodeStr, "64"), v2i64, ShOp>;
875 //===----------------------------------------------------------------------===//
876 // Instruction Definitions.
877 //===----------------------------------------------------------------------===//
879 // Vector Add Operations.
881 // VADD : Vector Add (integer and floating-point)
882 defm VADD : N3V_QHSD<0, 0, 0b1000, 0, "vadd.i", add, 1>;
883 def VADDfd : N3VD<0, 0, 0b00, 0b1101, 0, "vadd.f32", v2f32, v2f32, fadd, 1>;
884 def VADDfq : N3VQ<0, 0, 0b00, 0b1101, 0, "vadd.f32", v4f32, v4f32, fadd, 1>;
885 // VADDL : Vector Add Long (Q = D + D)
886 defm VADDLs : N3VLInt_QHS<0,1,0b0000,0, "vaddl.s", int_arm_neon_vaddls, 1>;
887 defm VADDLu : N3VLInt_QHS<1,1,0b0000,0, "vaddl.u", int_arm_neon_vaddlu, 1>;
888 // VADDW : Vector Add Wide (Q = Q + D)
889 defm VADDWs : N3VWInt_QHS<0,1,0b0001,0, "vaddw.s", int_arm_neon_vaddws, 0>;
890 defm VADDWu : N3VWInt_QHS<1,1,0b0001,0, "vaddw.u", int_arm_neon_vaddwu, 0>;
891 // VHADD : Vector Halving Add
892 defm VHADDs : N3VInt_QHS<0,0,0b0000,0, "vhadd.s", int_arm_neon_vhadds, 1>;
893 defm VHADDu : N3VInt_QHS<1,0,0b0000,0, "vhadd.u", int_arm_neon_vhaddu, 1>;
894 // VRHADD : Vector Rounding Halving Add
895 defm VRHADDs : N3VInt_QHS<0,0,0b0001,0, "vrhadd.s", int_arm_neon_vrhadds, 1>;
896 defm VRHADDu : N3VInt_QHS<1,0,0b0001,0, "vrhadd.u", int_arm_neon_vrhaddu, 1>;
897 // VQADD : Vector Saturating Add
898 defm VQADDs : N3VInt_QHSD<0,0,0b0000,1, "vqadd.s", int_arm_neon_vqadds, 1>;
899 defm VQADDu : N3VInt_QHSD<1,0,0b0000,1, "vqadd.u", int_arm_neon_vqaddu, 1>;
900 // VADDHN : Vector Add and Narrow Returning High Half (D = Q + Q)
901 defm VADDHN : N3VNInt_HSD<0,1,0b0100,0, "vaddhn.i", int_arm_neon_vaddhn, 1>;
902 // VRADDHN : Vector Rounding Add and Narrow Returning High Half (D = Q + Q)
903 defm VRADDHN : N3VNInt_HSD<1,1,0b0100,0, "vraddhn.i", int_arm_neon_vraddhn, 1>;
905 // Vector Add Operations used for single-precision FP
906 def : N3VDs<fadd, VADDfd>;
908 // Vector Multiply Operations.
910 // VMUL : Vector Multiply (integer, polynomial and floating-point)
911 defm VMUL : N3V_QHS<0, 0, 0b1001, 1, "vmul.i", mul, 1>;
912 def VMULpd : N3VDInt<1, 0, 0b00, 0b1001, 1, "vmul.p8", v8i8, v8i8,
913 int_arm_neon_vmulp, 1>;
914 def VMULpq : N3VQInt<1, 0, 0b00, 0b1001, 1, "vmul.p8", v16i8, v16i8,
915 int_arm_neon_vmulp, 1>;
916 def VMULfd : N3VD<1, 0, 0b00, 0b1101, 1, "vmul.f32", v2f32, v2f32, fmul, 1>;
917 def VMULfq : N3VQ<1, 0, 0b00, 0b1101, 1, "vmul.f32", v4f32, v4f32, fmul, 1>;
918 // VQDMULH : Vector Saturating Doubling Multiply Returning High Half
919 defm VQDMULH : N3VInt_HS<0,0,0b1011,0, "vqdmulh.s", int_arm_neon_vqdmulh, 1>;
920 // VQRDMULH : Vector Rounding Saturating Doubling Multiply Returning High Half
921 defm VQRDMULH : N3VInt_HS<1,0,0b1011,0, "vqrdmulh.s", int_arm_neon_vqrdmulh, 1>;
922 // VMULL : Vector Multiply Long (integer and polynomial) (Q = D * D)
923 defm VMULLs : N3VLInt_QHS<0,1,0b1100,0, "vmull.s", int_arm_neon_vmulls, 1>;
924 defm VMULLu : N3VLInt_QHS<1,1,0b1100,0, "vmull.u", int_arm_neon_vmullu, 1>;
925 def VMULLp : N3VLInt<0, 1, 0b00, 0b1110, 0, "vmull.p8", v8i16, v8i8,
926 int_arm_neon_vmullp, 1>;
927 // VQDMULL : Vector Saturating Doubling Multiply Long (Q = D * D)
928 defm VQDMULL : N3VLInt_HS<0,1,0b1101,0, "vqdmull.s", int_arm_neon_vqdmull, 1>;
930 // Vector Multiply Operations used for single-precision FP
931 def : N3VDs<fmul, VMULfd>;
933 // Vector Multiply-Accumulate and Multiply-Subtract Operations.
935 // VMLA : Vector Multiply Accumulate (integer and floating-point)
936 defm VMLA : N3VMulOp_QHS<0, 0, 0b1001, 0, "vmla.i", add>;
937 def VMLAfd : N3VDMulOp<0, 0, 0b00, 0b1101, 1, "vmla.f32", v2f32, fmul, fadd>;
938 def VMLAfq : N3VQMulOp<0, 0, 0b00, 0b1101, 1, "vmla.f32", v4f32, fmul, fadd>;
939 // VMLAL : Vector Multiply Accumulate Long (Q += D * D)
940 defm VMLALs : N3VLInt3_QHS<0,1,0b1000,0, "vmlal.s", int_arm_neon_vmlals>;
941 defm VMLALu : N3VLInt3_QHS<1,1,0b1000,0, "vmlal.u", int_arm_neon_vmlalu>;
942 // VQDMLAL : Vector Saturating Doubling Multiply Accumulate Long (Q += D * D)
943 defm VQDMLAL : N3VLInt3_HS<0, 1, 0b1001, 0, "vqdmlal.s", int_arm_neon_vqdmlal>;
944 // VMLS : Vector Multiply Subtract (integer and floating-point)
945 defm VMLS : N3VMulOp_QHS<0, 0, 0b1001, 0, "vmls.i", sub>;
946 def VMLSfd : N3VDMulOp<0, 0, 0b10, 0b1101, 1, "vmls.f32", v2f32, fmul, fsub>;
947 def VMLSfq : N3VQMulOp<0, 0, 0b10, 0b1101, 1, "vmls.f32", v4f32, fmul, fsub>;
948 // VMLSL : Vector Multiply Subtract Long (Q -= D * D)
949 defm VMLSLs : N3VLInt3_QHS<0,1,0b1010,0, "vmlsl.s", int_arm_neon_vmlsls>;
950 defm VMLSLu : N3VLInt3_QHS<1,1,0b1010,0, "vmlsl.u", int_arm_neon_vmlslu>;
951 // VQDMLSL : Vector Saturating Doubling Multiply Subtract Long (Q -= D * D)
952 defm VQDMLSL : N3VLInt3_HS<0, 1, 0b1011, 0, "vqdmlsl.s", int_arm_neon_vqdmlsl>;
954 // Vector Multiply-Accumulate/Subtract used for single-precision FP
955 def : N3VDMulOps<fmul, fadd, VMLAfd>;
956 def : N3VDMulOps<fmul, fsub, VMLSfd>;
958 // Vector Subtract Operations.
960 // VSUB : Vector Subtract (integer and floating-point)
961 defm VSUB : N3V_QHSD<1, 0, 0b1000, 0, "vsub.i", sub, 0>;
962 def VSUBfd : N3VD<0, 0, 0b10, 0b1101, 0, "vsub.f32", v2f32, v2f32, fsub, 0>;
963 def VSUBfq : N3VQ<0, 0, 0b10, 0b1101, 0, "vsub.f32", v4f32, v4f32, fsub, 0>;
964 // VSUBL : Vector Subtract Long (Q = D - D)
965 defm VSUBLs : N3VLInt_QHS<0,1,0b0010,0, "vsubl.s", int_arm_neon_vsubls, 1>;
966 defm VSUBLu : N3VLInt_QHS<1,1,0b0010,0, "vsubl.u", int_arm_neon_vsublu, 1>;
967 // VSUBW : Vector Subtract Wide (Q = Q - D)
968 defm VSUBWs : N3VWInt_QHS<0,1,0b0011,0, "vsubw.s", int_arm_neon_vsubws, 0>;
969 defm VSUBWu : N3VWInt_QHS<1,1,0b0011,0, "vsubw.u", int_arm_neon_vsubwu, 0>;
970 // VHSUB : Vector Halving Subtract
971 defm VHSUBs : N3VInt_QHS<0, 0, 0b0010, 0, "vhsub.s", int_arm_neon_vhsubs, 0>;
972 defm VHSUBu : N3VInt_QHS<1, 0, 0b0010, 0, "vhsub.u", int_arm_neon_vhsubu, 0>;
973 // VQSUB : Vector Saturing Subtract
974 defm VQSUBs : N3VInt_QHSD<0, 0, 0b0010, 1, "vqsub.s", int_arm_neon_vqsubs, 0>;
975 defm VQSUBu : N3VInt_QHSD<1, 0, 0b0010, 1, "vqsub.u", int_arm_neon_vqsubu, 0>;
976 // VSUBHN : Vector Subtract and Narrow Returning High Half (D = Q - Q)
977 defm VSUBHN : N3VNInt_HSD<0,1,0b0110,0, "vsubhn.i", int_arm_neon_vsubhn, 0>;
978 // VRSUBHN : Vector Rounding Subtract and Narrow Returning High Half (D=Q-Q)
979 defm VRSUBHN : N3VNInt_HSD<1,1,0b0110,0, "vrsubhn.i", int_arm_neon_vrsubhn, 0>;
981 // Vector Sub Operations used for single-precision FP
982 def : N3VDs<fsub, VSUBfd>;
984 // Vector Comparisons.
986 // VCEQ : Vector Compare Equal
987 defm VCEQ : N3V_QHS<1, 0, 0b1000, 1, "vceq.i", NEONvceq, 1>;
988 def VCEQfd : N3VD<0,0,0b00,0b1110,0, "vceq.f32", v2i32, v2f32, NEONvceq, 1>;
989 def VCEQfq : N3VQ<0,0,0b00,0b1110,0, "vceq.f32", v4i32, v4f32, NEONvceq, 1>;
990 // VCGE : Vector Compare Greater Than or Equal
991 defm VCGEs : N3V_QHS<0, 0, 0b0011, 1, "vcge.s", NEONvcge, 0>;
992 defm VCGEu : N3V_QHS<1, 0, 0b0011, 1, "vcge.u", NEONvcgeu, 0>;
993 def VCGEfd : N3VD<1,0,0b00,0b1110,0, "vcge.f32", v2i32, v2f32, NEONvcge, 0>;
994 def VCGEfq : N3VQ<1,0,0b00,0b1110,0, "vcge.f32", v4i32, v4f32, NEONvcge, 0>;
995 // VCGT : Vector Compare Greater Than
996 defm VCGTs : N3V_QHS<0, 0, 0b0011, 0, "vcgt.s", NEONvcgt, 0>;
997 defm VCGTu : N3V_QHS<1, 0, 0b0011, 0, "vcgt.u", NEONvcgtu, 0>;
998 def VCGTfd : N3VD<1,0,0b10,0b1110,0, "vcgt.f32", v2i32, v2f32, NEONvcgt, 0>;
999 def VCGTfq : N3VQ<1,0,0b10,0b1110,0, "vcgt.f32", v4i32, v4f32, NEONvcgt, 0>;
1000 // VACGE : Vector Absolute Compare Greater Than or Equal (aka VCAGE)
1001 def VACGEd : N3VDInt<1, 0, 0b00, 0b1110, 1, "vacge.f32", v2i32, v2f32,
1002 int_arm_neon_vacged, 0>;
1003 def VACGEq : N3VQInt<1, 0, 0b00, 0b1110, 1, "vacge.f32", v4i32, v4f32,
1004 int_arm_neon_vacgeq, 0>;
1005 // VACGT : Vector Absolute Compare Greater Than (aka VCAGT)
1006 def VACGTd : N3VDInt<1, 0, 0b10, 0b1110, 1, "vacgt.f32", v2i32, v2f32,
1007 int_arm_neon_vacgtd, 0>;
1008 def VACGTq : N3VQInt<1, 0, 0b10, 0b1110, 1, "vacgt.f32", v4i32, v4f32,
1009 int_arm_neon_vacgtq, 0>;
1010 // VTST : Vector Test Bits
1011 defm VTST : N3V_QHS<0, 0, 0b1000, 1, "vtst.i", NEONvtst, 1>;
1013 // Vector Bitwise Operations.
1015 // VAND : Vector Bitwise AND
1016 def VANDd : N3VD<0, 0, 0b00, 0b0001, 1, "vand", v2i32, v2i32, and, 1>;
1017 def VANDq : N3VQ<0, 0, 0b00, 0b0001, 1, "vand", v4i32, v4i32, and, 1>;
1019 // VEOR : Vector Bitwise Exclusive OR
1020 def VEORd : N3VD<1, 0, 0b00, 0b0001, 1, "veor", v2i32, v2i32, xor, 1>;
1021 def VEORq : N3VQ<1, 0, 0b00, 0b0001, 1, "veor", v4i32, v4i32, xor, 1>;
1023 // VORR : Vector Bitwise OR
1024 def VORRd : N3VD<0, 0, 0b10, 0b0001, 1, "vorr", v2i32, v2i32, or, 1>;
1025 def VORRq : N3VQ<0, 0, 0b10, 0b0001, 1, "vorr", v4i32, v4i32, or, 1>;
1027 // VBIC : Vector Bitwise Bit Clear (AND NOT)
1028 def VBICd : N3V<0, 0, 0b01, 0b0001, 0, 1, (outs DPR:$dst),
1029 (ins DPR:$src1, DPR:$src2), "vbic\t$dst, $src1, $src2", "",
1030 [(set DPR:$dst, (v2i32 (and DPR:$src1,(vnot DPR:$src2))))]>;
1031 def VBICq : N3V<0, 0, 0b01, 0b0001, 1, 1, (outs QPR:$dst),
1032 (ins QPR:$src1, QPR:$src2), "vbic\t$dst, $src1, $src2", "",
1033 [(set QPR:$dst, (v4i32 (and QPR:$src1,(vnot QPR:$src2))))]>;
1035 // VORN : Vector Bitwise OR NOT
1036 def VORNd : N3V<0, 0, 0b11, 0b0001, 0, 1, (outs DPR:$dst),
1037 (ins DPR:$src1, DPR:$src2), "vorn\t$dst, $src1, $src2", "",
1038 [(set DPR:$dst, (v2i32 (or DPR:$src1, (vnot DPR:$src2))))]>;
1039 def VORNq : N3V<0, 0, 0b11, 0b0001, 1, 1, (outs QPR:$dst),
1040 (ins QPR:$src1, QPR:$src2), "vorn\t$dst, $src1, $src2", "",
1041 [(set QPR:$dst, (v4i32 (or QPR:$src1, (vnot QPR:$src2))))]>;
1043 // VMVN : Vector Bitwise NOT
1044 def VMVNd : N2V<0b11, 0b11, 0b00, 0b00, 0b01011, 0, 0,
1045 (outs DPR:$dst), (ins DPR:$src), "vmvn\t$dst, $src", "",
1046 [(set DPR:$dst, (v2i32 (vnot DPR:$src)))]>;
1047 def VMVNq : N2V<0b11, 0b11, 0b00, 0b00, 0b01011, 1, 0,
1048 (outs QPR:$dst), (ins QPR:$src), "vmvn\t$dst, $src", "",
1049 [(set QPR:$dst, (v4i32 (vnot QPR:$src)))]>;
1050 def : Pat<(v2i32 (vnot_conv DPR:$src)), (VMVNd DPR:$src)>;
1051 def : Pat<(v4i32 (vnot_conv QPR:$src)), (VMVNq QPR:$src)>;
1053 // VBSL : Vector Bitwise Select
1054 def VBSLd : N3V<1, 0, 0b01, 0b0001, 0, 1, (outs DPR:$dst),
1055 (ins DPR:$src1, DPR:$src2, DPR:$src3),
1056 "vbsl\t$dst, $src2, $src3", "$src1 = $dst",
1058 (v2i32 (or (and DPR:$src2, DPR:$src1),
1059 (and DPR:$src3, (vnot DPR:$src1)))))]>;
1060 def VBSLq : N3V<1, 0, 0b01, 0b0001, 1, 1, (outs QPR:$dst),
1061 (ins QPR:$src1, QPR:$src2, QPR:$src3),
1062 "vbsl\t$dst, $src2, $src3", "$src1 = $dst",
1064 (v4i32 (or (and QPR:$src2, QPR:$src1),
1065 (and QPR:$src3, (vnot QPR:$src1)))))]>;
1067 // VBIF : Vector Bitwise Insert if False
1068 // like VBSL but with: "vbif\t$dst, $src3, $src1", "$src2 = $dst",
1069 // VBIT : Vector Bitwise Insert if True
1070 // like VBSL but with: "vbit\t$dst, $src2, $src1", "$src3 = $dst",
1071 // These are not yet implemented. The TwoAddress pass will not go looking
1072 // for equivalent operations with different register constraints; it just
1075 // Vector Absolute Differences.
1077 // VABD : Vector Absolute Difference
1078 defm VABDs : N3VInt_QHS<0, 0, 0b0111, 0, "vabd.s", int_arm_neon_vabds, 0>;
1079 defm VABDu : N3VInt_QHS<1, 0, 0b0111, 0, "vabd.u", int_arm_neon_vabdu, 0>;
1080 def VABDfd : N3VDInt<1, 0, 0b10, 0b1101, 0, "vabd.f32", v2f32, v2f32,
1081 int_arm_neon_vabdf, 0>;
1082 def VABDfq : N3VQInt<1, 0, 0b10, 0b1101, 0, "vabd.f32", v4f32, v4f32,
1083 int_arm_neon_vabdf, 0>;
1085 // VABDL : Vector Absolute Difference Long (Q = | D - D |)
1086 defm VABDLs : N3VLInt_QHS<0,1,0b0111,0, "vabdl.s", int_arm_neon_vabdls, 0>;
1087 defm VABDLu : N3VLInt_QHS<1,1,0b0111,0, "vabdl.u", int_arm_neon_vabdlu, 0>;
1089 // VABA : Vector Absolute Difference and Accumulate
1090 defm VABAs : N3VInt3_QHS<0,1,0b0101,0, "vaba.s", int_arm_neon_vabas>;
1091 defm VABAu : N3VInt3_QHS<1,1,0b0101,0, "vaba.u", int_arm_neon_vabau>;
1093 // VABAL : Vector Absolute Difference and Accumulate Long (Q += | D - D |)
1094 defm VABALs : N3VLInt3_QHS<0,1,0b0101,0, "vabal.s", int_arm_neon_vabals>;
1095 defm VABALu : N3VLInt3_QHS<1,1,0b0101,0, "vabal.u", int_arm_neon_vabalu>;
1097 // Vector Maximum and Minimum.
1099 // VMAX : Vector Maximum
1100 defm VMAXs : N3VInt_QHS<0, 0, 0b0110, 0, "vmax.s", int_arm_neon_vmaxs, 1>;
1101 defm VMAXu : N3VInt_QHS<1, 0, 0b0110, 0, "vmax.u", int_arm_neon_vmaxu, 1>;
1102 def VMAXfd : N3VDInt<0, 0, 0b00, 0b1111, 0, "vmax.f32", v2f32, v2f32,
1103 int_arm_neon_vmaxf, 1>;
1104 def VMAXfq : N3VQInt<0, 0, 0b00, 0b1111, 0, "vmax.f32", v4f32, v4f32,
1105 int_arm_neon_vmaxf, 1>;
1107 // VMIN : Vector Minimum
1108 defm VMINs : N3VInt_QHS<0, 0, 0b0110, 1, "vmin.s", int_arm_neon_vmins, 1>;
1109 defm VMINu : N3VInt_QHS<1, 0, 0b0110, 1, "vmin.u", int_arm_neon_vminu, 1>;
1110 def VMINfd : N3VDInt<0, 0, 0b10, 0b1111, 0, "vmin.f32", v2f32, v2f32,
1111 int_arm_neon_vminf, 1>;
1112 def VMINfq : N3VQInt<0, 0, 0b10, 0b1111, 0, "vmin.f32", v4f32, v4f32,
1113 int_arm_neon_vminf, 1>;
1115 // Vector Pairwise Operations.
1117 // VPADD : Vector Pairwise Add
1118 def VPADDi8 : N3VDInt<0, 0, 0b00, 0b1011, 1, "vpadd.i8", v8i8, v8i8,
1119 int_arm_neon_vpaddi, 0>;
1120 def VPADDi16 : N3VDInt<0, 0, 0b01, 0b1011, 1, "vpadd.i16", v4i16, v4i16,
1121 int_arm_neon_vpaddi, 0>;
1122 def VPADDi32 : N3VDInt<0, 0, 0b10, 0b1011, 1, "vpadd.i32", v2i32, v2i32,
1123 int_arm_neon_vpaddi, 0>;
1124 def VPADDf : N3VDInt<1, 0, 0b00, 0b1101, 0, "vpadd.f32", v2f32, v2f32,
1125 int_arm_neon_vpaddf, 0>;
1127 // VPADDL : Vector Pairwise Add Long
1128 defm VPADDLs : N2VPLInt_QHS<0b11, 0b11, 0b00, 0b00100, 0, "vpaddl.s",
1129 int_arm_neon_vpaddls>;
1130 defm VPADDLu : N2VPLInt_QHS<0b11, 0b11, 0b00, 0b00101, 0, "vpaddl.u",
1131 int_arm_neon_vpaddlu>;
1133 // VPADAL : Vector Pairwise Add and Accumulate Long
1134 defm VPADALs : N2VPLInt2_QHS<0b11, 0b11, 0b00, 0b00100, 0, "vpadal.s",
1135 int_arm_neon_vpadals>;
1136 defm VPADALu : N2VPLInt2_QHS<0b11, 0b11, 0b00, 0b00101, 0, "vpadal.u",
1137 int_arm_neon_vpadalu>;
1139 // VPMAX : Vector Pairwise Maximum
1140 def VPMAXs8 : N3VDInt<0, 0, 0b00, 0b1010, 0, "vpmax.s8", v8i8, v8i8,
1141 int_arm_neon_vpmaxs, 0>;
1142 def VPMAXs16 : N3VDInt<0, 0, 0b01, 0b1010, 0, "vpmax.s16", v4i16, v4i16,
1143 int_arm_neon_vpmaxs, 0>;
1144 def VPMAXs32 : N3VDInt<0, 0, 0b10, 0b1010, 0, "vpmax.s32", v2i32, v2i32,
1145 int_arm_neon_vpmaxs, 0>;
1146 def VPMAXu8 : N3VDInt<1, 0, 0b00, 0b1010, 0, "vpmax.u8", v8i8, v8i8,
1147 int_arm_neon_vpmaxu, 0>;
1148 def VPMAXu16 : N3VDInt<1, 0, 0b01, 0b1010, 0, "vpmax.u16", v4i16, v4i16,
1149 int_arm_neon_vpmaxu, 0>;
1150 def VPMAXu32 : N3VDInt<1, 0, 0b10, 0b1010, 0, "vpmax.u32", v2i32, v2i32,
1151 int_arm_neon_vpmaxu, 0>;
1152 def VPMAXf : N3VDInt<1, 0, 0b00, 0b1111, 0, "vpmax.f32", v2f32, v2f32,
1153 int_arm_neon_vpmaxf, 0>;
1155 // VPMIN : Vector Pairwise Minimum
1156 def VPMINs8 : N3VDInt<0, 0, 0b00, 0b1010, 1, "vpmin.s8", v8i8, v8i8,
1157 int_arm_neon_vpmins, 0>;
1158 def VPMINs16 : N3VDInt<0, 0, 0b01, 0b1010, 1, "vpmin.s16", v4i16, v4i16,
1159 int_arm_neon_vpmins, 0>;
1160 def VPMINs32 : N3VDInt<0, 0, 0b10, 0b1010, 1, "vpmin.s32", v2i32, v2i32,
1161 int_arm_neon_vpmins, 0>;
1162 def VPMINu8 : N3VDInt<1, 0, 0b00, 0b1010, 1, "vpmin.u8", v8i8, v8i8,
1163 int_arm_neon_vpminu, 0>;
1164 def VPMINu16 : N3VDInt<1, 0, 0b01, 0b1010, 1, "vpmin.u16", v4i16, v4i16,
1165 int_arm_neon_vpminu, 0>;
1166 def VPMINu32 : N3VDInt<1, 0, 0b10, 0b1010, 1, "vpmin.u32", v2i32, v2i32,
1167 int_arm_neon_vpminu, 0>;
1168 def VPMINf : N3VDInt<1, 0, 0b10, 0b1111, 0, "vpmin.f32", v2f32, v2f32,
1169 int_arm_neon_vpminf, 0>;
1171 // Vector Reciprocal and Reciprocal Square Root Estimate and Step.
1173 // VRECPE : Vector Reciprocal Estimate
1174 def VRECPEd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01000, 0, "vrecpe.u32",
1175 v2i32, v2i32, int_arm_neon_vrecpe>;
1176 def VRECPEq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01000, 0, "vrecpe.u32",
1177 v4i32, v4i32, int_arm_neon_vrecpe>;
1178 def VRECPEfd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01010, 0, "vrecpe.f32",
1179 v2f32, v2f32, int_arm_neon_vrecpef>;
1180 def VRECPEfq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01010, 0, "vrecpe.f32",
1181 v4f32, v4f32, int_arm_neon_vrecpef>;
1183 // VRECPS : Vector Reciprocal Step
1184 def VRECPSfd : N3VDInt<0, 0, 0b00, 0b1111, 1, "vrecps.f32", v2f32, v2f32,
1185 int_arm_neon_vrecps, 1>;
1186 def VRECPSfq : N3VQInt<0, 0, 0b00, 0b1111, 1, "vrecps.f32", v4f32, v4f32,
1187 int_arm_neon_vrecps, 1>;
1189 // VRSQRTE : Vector Reciprocal Square Root Estimate
1190 def VRSQRTEd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01001, 0, "vrsqrte.u32",
1191 v2i32, v2i32, int_arm_neon_vrsqrte>;
1192 def VRSQRTEq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01001, 0, "vrsqrte.u32",
1193 v4i32, v4i32, int_arm_neon_vrsqrte>;
1194 def VRSQRTEfd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01011, 0, "vrsqrte.f32",
1195 v2f32, v2f32, int_arm_neon_vrsqrtef>;
1196 def VRSQRTEfq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01011, 0, "vrsqrte.f32",
1197 v4f32, v4f32, int_arm_neon_vrsqrtef>;
1199 // VRSQRTS : Vector Reciprocal Square Root Step
1200 def VRSQRTSfd : N3VDInt<0, 0, 0b10, 0b1111, 1, "vrsqrts.f32", v2f32, v2f32,
1201 int_arm_neon_vrsqrts, 1>;
1202 def VRSQRTSfq : N3VQInt<0, 0, 0b10, 0b1111, 1, "vrsqrts.f32", v4f32, v4f32,
1203 int_arm_neon_vrsqrts, 1>;
1207 // VSHL : Vector Shift
1208 defm VSHLs : N3VInt_QHSD<0, 0, 0b0100, 0, "vshl.s", int_arm_neon_vshifts, 0>;
1209 defm VSHLu : N3VInt_QHSD<1, 0, 0b0100, 0, "vshl.u", int_arm_neon_vshiftu, 0>;
1210 // VSHL : Vector Shift Left (Immediate)
1211 defm VSHLi : N2VSh_QHSD<0, 1, 0b0111, 1, "vshl.i", NEONvshl>;
1212 // VSHR : Vector Shift Right (Immediate)
1213 defm VSHRs : N2VSh_QHSD<0, 1, 0b0000, 1, "vshr.s", NEONvshrs>;
1214 defm VSHRu : N2VSh_QHSD<1, 1, 0b0000, 1, "vshr.u", NEONvshru>;
1216 // VSHLL : Vector Shift Left Long
1217 def VSHLLs8 : N2VLSh<0, 1, 0b001000, 0b1010, 0, 0, 1, "vshll.s8",
1218 v8i16, v8i8, NEONvshlls>;
1219 def VSHLLs16 : N2VLSh<0, 1, 0b010000, 0b1010, 0, 0, 1, "vshll.s16",
1220 v4i32, v4i16, NEONvshlls>;
1221 def VSHLLs32 : N2VLSh<0, 1, 0b100000, 0b1010, 0, 0, 1, "vshll.s32",
1222 v2i64, v2i32, NEONvshlls>;
1223 def VSHLLu8 : N2VLSh<1, 1, 0b001000, 0b1010, 0, 0, 1, "vshll.u8",
1224 v8i16, v8i8, NEONvshllu>;
1225 def VSHLLu16 : N2VLSh<1, 1, 0b010000, 0b1010, 0, 0, 1, "vshll.u16",
1226 v4i32, v4i16, NEONvshllu>;
1227 def VSHLLu32 : N2VLSh<1, 1, 0b100000, 0b1010, 0, 0, 1, "vshll.u32",
1228 v2i64, v2i32, NEONvshllu>;
1230 // VSHLL : Vector Shift Left Long (with maximum shift count)
1231 def VSHLLi8 : N2VLSh<1, 1, 0b110010, 0b0011, 0, 0, 0, "vshll.i8",
1232 v8i16, v8i8, NEONvshlli>;
1233 def VSHLLi16 : N2VLSh<1, 1, 0b110110, 0b0011, 0, 0, 0, "vshll.i16",
1234 v4i32, v4i16, NEONvshlli>;
1235 def VSHLLi32 : N2VLSh<1, 1, 0b111010, 0b0011, 0, 0, 0, "vshll.i32",
1236 v2i64, v2i32, NEONvshlli>;
1238 // VSHRN : Vector Shift Right and Narrow
1239 def VSHRN16 : N2VNSh<0, 1, 0b001000, 0b1000, 0, 0, 1, "vshrn.i16",
1240 v8i8, v8i16, NEONvshrn>;
1241 def VSHRN32 : N2VNSh<0, 1, 0b010000, 0b1000, 0, 0, 1, "vshrn.i32",
1242 v4i16, v4i32, NEONvshrn>;
1243 def VSHRN64 : N2VNSh<0, 1, 0b100000, 0b1000, 0, 0, 1, "vshrn.i64",
1244 v2i32, v2i64, NEONvshrn>;
1246 // VRSHL : Vector Rounding Shift
1247 defm VRSHLs : N3VInt_QHSD<0,0,0b0101,0, "vrshl.s", int_arm_neon_vrshifts, 0>;
1248 defm VRSHLu : N3VInt_QHSD<1,0,0b0101,0, "vrshl.u", int_arm_neon_vrshiftu, 0>;
1249 // VRSHR : Vector Rounding Shift Right
1250 defm VRSHRs : N2VSh_QHSD<0, 1, 0b0010, 1, "vrshr.s", NEONvrshrs>;
1251 defm VRSHRu : N2VSh_QHSD<1, 1, 0b0010, 1, "vrshr.u", NEONvrshru>;
1253 // VRSHRN : Vector Rounding Shift Right and Narrow
1254 def VRSHRN16 : N2VNSh<0, 1, 0b001000, 0b1000, 0, 1, 1, "vrshrn.i16",
1255 v8i8, v8i16, NEONvrshrn>;
1256 def VRSHRN32 : N2VNSh<0, 1, 0b010000, 0b1000, 0, 1, 1, "vrshrn.i32",
1257 v4i16, v4i32, NEONvrshrn>;
1258 def VRSHRN64 : N2VNSh<0, 1, 0b100000, 0b1000, 0, 1, 1, "vrshrn.i64",
1259 v2i32, v2i64, NEONvrshrn>;
1261 // VQSHL : Vector Saturating Shift
1262 defm VQSHLs : N3VInt_QHSD<0,0,0b0100,1, "vqshl.s", int_arm_neon_vqshifts, 0>;
1263 defm VQSHLu : N3VInt_QHSD<1,0,0b0100,1, "vqshl.u", int_arm_neon_vqshiftu, 0>;
1264 // VQSHL : Vector Saturating Shift Left (Immediate)
1265 defm VQSHLsi : N2VSh_QHSD<0, 1, 0b0111, 1, "vqshl.s", NEONvqshls>;
1266 defm VQSHLui : N2VSh_QHSD<1, 1, 0b0111, 1, "vqshl.u", NEONvqshlu>;
1267 // VQSHLU : Vector Saturating Shift Left (Immediate, Unsigned)
1268 defm VQSHLsu : N2VSh_QHSD<1, 1, 0b0110, 1, "vqshlu.s", NEONvqshlsu>;
1270 // VQSHRN : Vector Saturating Shift Right and Narrow
1271 def VQSHRNs16 : N2VNSh<0, 1, 0b001000, 0b1001, 0, 0, 1, "vqshrn.s16",
1272 v8i8, v8i16, NEONvqshrns>;
1273 def VQSHRNs32 : N2VNSh<0, 1, 0b010000, 0b1001, 0, 0, 1, "vqshrn.s32",
1274 v4i16, v4i32, NEONvqshrns>;
1275 def VQSHRNs64 : N2VNSh<0, 1, 0b100000, 0b1001, 0, 0, 1, "vqshrn.s64",
1276 v2i32, v2i64, NEONvqshrns>;
1277 def VQSHRNu16 : N2VNSh<1, 1, 0b001000, 0b1001, 0, 0, 1, "vqshrn.u16",
1278 v8i8, v8i16, NEONvqshrnu>;
1279 def VQSHRNu32 : N2VNSh<1, 1, 0b010000, 0b1001, 0, 0, 1, "vqshrn.u32",
1280 v4i16, v4i32, NEONvqshrnu>;
1281 def VQSHRNu64 : N2VNSh<1, 1, 0b100000, 0b1001, 0, 0, 1, "vqshrn.u64",
1282 v2i32, v2i64, NEONvqshrnu>;
1284 // VQSHRUN : Vector Saturating Shift Right and Narrow (Unsigned)
1285 def VQSHRUN16 : N2VNSh<1, 1, 0b001000, 0b1000, 0, 0, 1, "vqshrun.s16",
1286 v8i8, v8i16, NEONvqshrnsu>;
1287 def VQSHRUN32 : N2VNSh<1, 1, 0b010000, 0b1000, 0, 0, 1, "vqshrun.s32",
1288 v4i16, v4i32, NEONvqshrnsu>;
1289 def VQSHRUN64 : N2VNSh<1, 1, 0b100000, 0b1000, 0, 0, 1, "vqshrun.s64",
1290 v2i32, v2i64, NEONvqshrnsu>;
1292 // VQRSHL : Vector Saturating Rounding Shift
1293 defm VQRSHLs : N3VInt_QHSD<0, 0, 0b0101, 1, "vqrshl.s",
1294 int_arm_neon_vqrshifts, 0>;
1295 defm VQRSHLu : N3VInt_QHSD<1, 0, 0b0101, 1, "vqrshl.u",
1296 int_arm_neon_vqrshiftu, 0>;
1298 // VQRSHRN : Vector Saturating Rounding Shift Right and Narrow
1299 def VQRSHRNs16: N2VNSh<0, 1, 0b001000, 0b1001, 0, 1, 1, "vqrshrn.s16",
1300 v8i8, v8i16, NEONvqrshrns>;
1301 def VQRSHRNs32: N2VNSh<0, 1, 0b010000, 0b1001, 0, 1, 1, "vqrshrn.s32",
1302 v4i16, v4i32, NEONvqrshrns>;
1303 def VQRSHRNs64: N2VNSh<0, 1, 0b100000, 0b1001, 0, 1, 1, "vqrshrn.s64",
1304 v2i32, v2i64, NEONvqrshrns>;
1305 def VQRSHRNu16: N2VNSh<1, 1, 0b001000, 0b1001, 0, 1, 1, "vqrshrn.u16",
1306 v8i8, v8i16, NEONvqrshrnu>;
1307 def VQRSHRNu32: N2VNSh<1, 1, 0b010000, 0b1001, 0, 1, 1, "vqrshrn.u32",
1308 v4i16, v4i32, NEONvqrshrnu>;
1309 def VQRSHRNu64: N2VNSh<1, 1, 0b100000, 0b1001, 0, 1, 1, "vqrshrn.u64",
1310 v2i32, v2i64, NEONvqrshrnu>;
1312 // VQRSHRUN : Vector Saturating Rounding Shift Right and Narrow (Unsigned)
1313 def VQRSHRUN16: N2VNSh<1, 1, 0b001000, 0b1000, 0, 1, 1, "vqrshrun.s16",
1314 v8i8, v8i16, NEONvqrshrnsu>;
1315 def VQRSHRUN32: N2VNSh<1, 1, 0b010000, 0b1000, 0, 1, 1, "vqrshrun.s32",
1316 v4i16, v4i32, NEONvqrshrnsu>;
1317 def VQRSHRUN64: N2VNSh<1, 1, 0b100000, 0b1000, 0, 1, 1, "vqrshrun.s64",
1318 v2i32, v2i64, NEONvqrshrnsu>;
1320 // VSRA : Vector Shift Right and Accumulate
1321 defm VSRAs : N2VShAdd_QHSD<0, 1, 0b0001, 1, "vsra.s", NEONvshrs>;
1322 defm VSRAu : N2VShAdd_QHSD<1, 1, 0b0001, 1, "vsra.u", NEONvshru>;
1323 // VRSRA : Vector Rounding Shift Right and Accumulate
1324 defm VRSRAs : N2VShAdd_QHSD<0, 1, 0b0011, 1, "vrsra.s", NEONvrshrs>;
1325 defm VRSRAu : N2VShAdd_QHSD<1, 1, 0b0011, 1, "vrsra.u", NEONvrshru>;
1327 // VSLI : Vector Shift Left and Insert
1328 defm VSLI : N2VShIns_QHSD<1, 1, 0b0101, 1, "vsli.", NEONvsli>;
1329 // VSRI : Vector Shift Right and Insert
1330 defm VSRI : N2VShIns_QHSD<1, 1, 0b0100, 1, "vsri.", NEONvsri>;
1332 // Vector Absolute and Saturating Absolute.
1334 // VABS : Vector Absolute Value
1335 defm VABS : N2VInt_QHS<0b11, 0b11, 0b01, 0b00110, 0, "vabs.s",
1337 def VABSfd : N2VDInt<0b11, 0b11, 0b10, 0b01, 0b01110, 0, "vabs.f32",
1338 v2f32, v2f32, int_arm_neon_vabsf>;
1339 def VABSfq : N2VQInt<0b11, 0b11, 0b10, 0b01, 0b01110, 0, "vabs.f32",
1340 v4f32, v4f32, int_arm_neon_vabsf>;
1342 // VQABS : Vector Saturating Absolute Value
1343 defm VQABS : N2VInt_QHS<0b11, 0b11, 0b00, 0b01110, 0, "vqabs.s",
1344 int_arm_neon_vqabs>;
1348 def vneg : PatFrag<(ops node:$in), (sub immAllZerosV, node:$in)>;
1349 def vneg_conv : PatFrag<(ops node:$in), (sub immAllZerosV_bc, node:$in)>;
1351 class VNEGD<bits<2> size, string OpcodeStr, ValueType Ty>
1352 : N2V<0b11, 0b11, size, 0b01, 0b00111, 0, 0, (outs DPR:$dst), (ins DPR:$src),
1353 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1354 [(set DPR:$dst, (Ty (vneg DPR:$src)))]>;
1355 class VNEGQ<bits<2> size, string OpcodeStr, ValueType Ty>
1356 : N2V<0b11, 0b11, size, 0b01, 0b00111, 1, 0, (outs QPR:$dst), (ins QPR:$src),
1357 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1358 [(set QPR:$dst, (Ty (vneg QPR:$src)))]>;
1360 // VNEG : Vector Negate
1361 def VNEGs8d : VNEGD<0b00, "vneg.s8", v8i8>;
1362 def VNEGs16d : VNEGD<0b01, "vneg.s16", v4i16>;
1363 def VNEGs32d : VNEGD<0b10, "vneg.s32", v2i32>;
1364 def VNEGs8q : VNEGQ<0b00, "vneg.s8", v16i8>;
1365 def VNEGs16q : VNEGQ<0b01, "vneg.s16", v8i16>;
1366 def VNEGs32q : VNEGQ<0b10, "vneg.s32", v4i32>;
1368 // VNEG : Vector Negate (floating-point)
1369 def VNEGf32d : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 0, 0,
1370 (outs DPR:$dst), (ins DPR:$src), "vneg.f32\t$dst, $src", "",
1371 [(set DPR:$dst, (v2f32 (fneg DPR:$src)))]>;
1372 def VNEGf32q : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 1, 0,
1373 (outs QPR:$dst), (ins QPR:$src), "vneg.f32\t$dst, $src", "",
1374 [(set QPR:$dst, (v4f32 (fneg QPR:$src)))]>;
1376 def : Pat<(v8i8 (vneg_conv DPR:$src)), (VNEGs8d DPR:$src)>;
1377 def : Pat<(v4i16 (vneg_conv DPR:$src)), (VNEGs16d DPR:$src)>;
1378 def : Pat<(v2i32 (vneg_conv DPR:$src)), (VNEGs32d DPR:$src)>;
1379 def : Pat<(v16i8 (vneg_conv QPR:$src)), (VNEGs8q QPR:$src)>;
1380 def : Pat<(v8i16 (vneg_conv QPR:$src)), (VNEGs16q QPR:$src)>;
1381 def : Pat<(v4i32 (vneg_conv QPR:$src)), (VNEGs32q QPR:$src)>;
1383 // VQNEG : Vector Saturating Negate
1384 defm VQNEG : N2VInt_QHS<0b11, 0b11, 0b00, 0b01111, 0, "vqneg.s",
1385 int_arm_neon_vqneg>;
1387 // Vector Bit Counting Operations.
1389 // VCLS : Vector Count Leading Sign Bits
1390 defm VCLS : N2VInt_QHS<0b11, 0b11, 0b00, 0b01000, 0, "vcls.s",
1392 // VCLZ : Vector Count Leading Zeros
1393 defm VCLZ : N2VInt_QHS<0b11, 0b11, 0b00, 0b01001, 0, "vclz.i",
1395 // VCNT : Vector Count One Bits
1396 def VCNTd : N2VDInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0, "vcnt.8",
1397 v8i8, v8i8, int_arm_neon_vcnt>;
1398 def VCNTq : N2VQInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0, "vcnt.8",
1399 v16i8, v16i8, int_arm_neon_vcnt>;
1401 // Vector Move Operations.
1403 // VMOV : Vector Move (Register)
1405 def VMOVD : N3V<0, 0, 0b10, 0b0001, 0, 1, (outs DPR:$dst), (ins DPR:$src),
1406 "vmov\t$dst, $src", "", []>;
1407 def VMOVQ : N3V<0, 0, 0b10, 0b0001, 1, 1, (outs QPR:$dst), (ins QPR:$src),
1408 "vmov\t$dst, $src", "", []>;
1410 // VMOV : Vector Move (Immediate)
1412 // VMOV_get_imm8 xform function: convert build_vector to VMOV.i8 imm.
1413 def VMOV_get_imm8 : SDNodeXForm<build_vector, [{
1414 return ARM::getVMOVImm(N, 1, *CurDAG);
1416 def vmovImm8 : PatLeaf<(build_vector), [{
1417 return ARM::getVMOVImm(N, 1, *CurDAG).getNode() != 0;
1420 // VMOV_get_imm16 xform function: convert build_vector to VMOV.i16 imm.
1421 def VMOV_get_imm16 : SDNodeXForm<build_vector, [{
1422 return ARM::getVMOVImm(N, 2, *CurDAG);
1424 def vmovImm16 : PatLeaf<(build_vector), [{
1425 return ARM::getVMOVImm(N, 2, *CurDAG).getNode() != 0;
1426 }], VMOV_get_imm16>;
1428 // VMOV_get_imm32 xform function: convert build_vector to VMOV.i32 imm.
1429 def VMOV_get_imm32 : SDNodeXForm<build_vector, [{
1430 return ARM::getVMOVImm(N, 4, *CurDAG);
1432 def vmovImm32 : PatLeaf<(build_vector), [{
1433 return ARM::getVMOVImm(N, 4, *CurDAG).getNode() != 0;
1434 }], VMOV_get_imm32>;
1436 // VMOV_get_imm64 xform function: convert build_vector to VMOV.i64 imm.
1437 def VMOV_get_imm64 : SDNodeXForm<build_vector, [{
1438 return ARM::getVMOVImm(N, 8, *CurDAG);
1440 def vmovImm64 : PatLeaf<(build_vector), [{
1441 return ARM::getVMOVImm(N, 8, *CurDAG).getNode() != 0;
1442 }], VMOV_get_imm64>;
1444 // Note: Some of the cmode bits in the following VMOV instructions need to
1445 // be encoded based on the immed values.
1447 def VMOVv8i8 : N1ModImm<1, 0b000, 0b1110, 0, 0, 0, 1, (outs DPR:$dst),
1448 (ins i8imm:$SIMM), "vmov.i8\t$dst, $SIMM", "",
1449 [(set DPR:$dst, (v8i8 vmovImm8:$SIMM))]>;
1450 def VMOVv16i8 : N1ModImm<1, 0b000, 0b1110, 0, 1, 0, 1, (outs QPR:$dst),
1451 (ins i8imm:$SIMM), "vmov.i8\t$dst, $SIMM", "",
1452 [(set QPR:$dst, (v16i8 vmovImm8:$SIMM))]>;
1454 def VMOVv4i16 : N1ModImm<1, 0b000, 0b1000, 0, 0, 0, 1, (outs DPR:$dst),
1455 (ins i16imm:$SIMM), "vmov.i16\t$dst, $SIMM", "",
1456 [(set DPR:$dst, (v4i16 vmovImm16:$SIMM))]>;
1457 def VMOVv8i16 : N1ModImm<1, 0b000, 0b1000, 0, 1, 0, 1, (outs QPR:$dst),
1458 (ins i16imm:$SIMM), "vmov.i16\t$dst, $SIMM", "",
1459 [(set QPR:$dst, (v8i16 vmovImm16:$SIMM))]>;
1461 def VMOVv2i32 : N1ModImm<1, 0b000, 0b0000, 0, 0, 0, 1, (outs DPR:$dst),
1462 (ins i32imm:$SIMM), "vmov.i32\t$dst, $SIMM", "",
1463 [(set DPR:$dst, (v2i32 vmovImm32:$SIMM))]>;
1464 def VMOVv4i32 : N1ModImm<1, 0b000, 0b0000, 0, 1, 0, 1, (outs QPR:$dst),
1465 (ins i32imm:$SIMM), "vmov.i32\t$dst, $SIMM", "",
1466 [(set QPR:$dst, (v4i32 vmovImm32:$SIMM))]>;
1468 def VMOVv1i64 : N1ModImm<1, 0b000, 0b1110, 0, 0, 1, 1, (outs DPR:$dst),
1469 (ins i64imm:$SIMM), "vmov.i64\t$dst, $SIMM", "",
1470 [(set DPR:$dst, (v1i64 vmovImm64:$SIMM))]>;
1471 def VMOVv2i64 : N1ModImm<1, 0b000, 0b1110, 0, 1, 1, 1, (outs QPR:$dst),
1472 (ins i64imm:$SIMM), "vmov.i64\t$dst, $SIMM", "",
1473 [(set QPR:$dst, (v2i64 vmovImm64:$SIMM))]>;
1475 // VMOV : Vector Get Lane (move scalar to ARM core register)
1477 def VGETLNs8 : NVGetLane<0b11100101, 0b1011, 0b00,
1478 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1479 "vmov", ".s8\t$dst, $src[$lane]",
1480 [(set GPR:$dst, (NEONvgetlanes (v8i8 DPR:$src),
1482 def VGETLNs16 : NVGetLane<0b11100001, 0b1011, 0b01,
1483 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1484 "vmov", ".s16\t$dst, $src[$lane]",
1485 [(set GPR:$dst, (NEONvgetlanes (v4i16 DPR:$src),
1487 def VGETLNu8 : NVGetLane<0b11101101, 0b1011, 0b00,
1488 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1489 "vmov", ".u8\t$dst, $src[$lane]",
1490 [(set GPR:$dst, (NEONvgetlaneu (v8i8 DPR:$src),
1492 def VGETLNu16 : NVGetLane<0b11101001, 0b1011, 0b01,
1493 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1494 "vmov", ".u16\t$dst, $src[$lane]",
1495 [(set GPR:$dst, (NEONvgetlaneu (v4i16 DPR:$src),
1497 def VGETLNi32 : NVGetLane<0b11100001, 0b1011, 0b00,
1498 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1499 "vmov", ".32\t$dst, $src[$lane]",
1500 [(set GPR:$dst, (extractelt (v2i32 DPR:$src),
1502 // def VGETLNf32: see FMRDH and FMRDL in ARMInstrVFP.td
1503 def : Pat<(NEONvgetlanes (v16i8 QPR:$src), imm:$lane),
1504 (VGETLNs8 (v8i8 (EXTRACT_SUBREG QPR:$src,
1505 (SubReg_i8_reg imm:$lane))),
1506 (SubReg_i8_lane imm:$lane))>;
1507 def : Pat<(NEONvgetlanes (v8i16 QPR:$src), imm:$lane),
1508 (VGETLNs16 (v4i16 (EXTRACT_SUBREG QPR:$src,
1509 (SubReg_i16_reg imm:$lane))),
1510 (SubReg_i16_lane imm:$lane))>;
1511 def : Pat<(NEONvgetlaneu (v16i8 QPR:$src), imm:$lane),
1512 (VGETLNu8 (v8i8 (EXTRACT_SUBREG QPR:$src,
1513 (SubReg_i8_reg imm:$lane))),
1514 (SubReg_i8_lane imm:$lane))>;
1515 def : Pat<(NEONvgetlaneu (v8i16 QPR:$src), imm:$lane),
1516 (VGETLNu16 (v4i16 (EXTRACT_SUBREG QPR:$src,
1517 (SubReg_i16_reg imm:$lane))),
1518 (SubReg_i16_lane imm:$lane))>;
1519 def : Pat<(extractelt (v4i32 QPR:$src), imm:$lane),
1520 (VGETLNi32 (v2i32 (EXTRACT_SUBREG QPR:$src,
1521 (SubReg_i32_reg imm:$lane))),
1522 (SubReg_i32_lane imm:$lane))>;
1523 //def : Pat<(extractelt (v2i64 QPR:$src1), imm:$src2),
1524 // (EXTRACT_SUBREG QPR:$src1, (SubReg_f64_reg imm:$src2))>;
1525 def : Pat<(extractelt (v2f64 QPR:$src1), imm:$src2),
1526 (EXTRACT_SUBREG QPR:$src1, (SubReg_f64_reg imm:$src2))>;
1529 // VMOV : Vector Set Lane (move ARM core register to scalar)
1531 let Constraints = "$src1 = $dst" in {
1532 def VSETLNi8 : NVSetLane<0b11100100, 0b1011, 0b00, (outs DPR:$dst),
1533 (ins DPR:$src1, GPR:$src2, i32imm:$lane),
1534 "vmov", ".8\t$dst[$lane], $src2",
1535 [(set DPR:$dst, (vector_insert (v8i8 DPR:$src1),
1536 GPR:$src2, imm:$lane))]>;
1537 def VSETLNi16 : NVSetLane<0b11100000, 0b1011, 0b01, (outs DPR:$dst),
1538 (ins DPR:$src1, GPR:$src2, i32imm:$lane),
1539 "vmov", ".16\t$dst[$lane], $src2",
1540 [(set DPR:$dst, (vector_insert (v4i16 DPR:$src1),
1541 GPR:$src2, imm:$lane))]>;
1542 def VSETLNi32 : NVSetLane<0b11100000, 0b1011, 0b00, (outs DPR:$dst),
1543 (ins DPR:$src1, GPR:$src2, i32imm:$lane),
1544 "vmov", ".32\t$dst[$lane], $src2",
1545 [(set DPR:$dst, (insertelt (v2i32 DPR:$src1),
1546 GPR:$src2, imm:$lane))]>;
1548 def : Pat<(vector_insert (v16i8 QPR:$src1), GPR:$src2, imm:$lane),
1549 (v16i8 (INSERT_SUBREG QPR:$src1,
1550 (VSETLNi8 (v8i8 (EXTRACT_SUBREG QPR:$src1,
1551 (SubReg_i8_reg imm:$lane))),
1552 GPR:$src2, (SubReg_i8_lane imm:$lane)),
1553 (SubReg_i8_reg imm:$lane)))>;
1554 def : Pat<(vector_insert (v8i16 QPR:$src1), GPR:$src2, imm:$lane),
1555 (v8i16 (INSERT_SUBREG QPR:$src1,
1556 (VSETLNi16 (v4i16 (EXTRACT_SUBREG QPR:$src1,
1557 (SubReg_i16_reg imm:$lane))),
1558 GPR:$src2, (SubReg_i16_lane imm:$lane)),
1559 (SubReg_i16_reg imm:$lane)))>;
1560 def : Pat<(insertelt (v4i32 QPR:$src1), GPR:$src2, imm:$lane),
1561 (v4i32 (INSERT_SUBREG QPR:$src1,
1562 (VSETLNi32 (v2i32 (EXTRACT_SUBREG QPR:$src1,
1563 (SubReg_i32_reg imm:$lane))),
1564 GPR:$src2, (SubReg_i32_lane imm:$lane)),
1565 (SubReg_i32_reg imm:$lane)))>;
1567 //def : Pat<(v2i64 (insertelt QPR:$src1, DPR:$src2, imm:$src3)),
1568 // (INSERT_SUBREG QPR:$src1, DPR:$src2, (SubReg_f64_reg imm:$src3))>;
1569 def : Pat<(v2f64 (insertelt QPR:$src1, DPR:$src2, imm:$src3)),
1570 (INSERT_SUBREG QPR:$src1, DPR:$src2, (SubReg_f64_reg imm:$src3))>;
1572 // VDUP : Vector Duplicate (from ARM core register to all elements)
1574 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
1575 (vector_shuffle node:$lhs, node:$rhs), [{
1576 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1577 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
1580 class VDUPD<bits<8> opcod1, bits<2> opcod3, string asmSize, ValueType Ty>
1581 : NVDup<opcod1, 0b1011, opcod3, (outs DPR:$dst), (ins GPR:$src),
1582 "vdup", !strconcat(asmSize, "\t$dst, $src"),
1583 [(set DPR:$dst, (Ty (splat_lo (scalar_to_vector GPR:$src), undef)))]>;
1584 class VDUPQ<bits<8> opcod1, bits<2> opcod3, string asmSize, ValueType Ty>
1585 : NVDup<opcod1, 0b1011, opcod3, (outs QPR:$dst), (ins GPR:$src),
1586 "vdup", !strconcat(asmSize, "\t$dst, $src"),
1587 [(set QPR:$dst, (Ty (splat_lo (scalar_to_vector GPR:$src), undef)))]>;
1589 def VDUP8d : VDUPD<0b11101100, 0b00, ".8", v8i8>;
1590 def VDUP16d : VDUPD<0b11101000, 0b01, ".16", v4i16>;
1591 def VDUP32d : VDUPD<0b11101000, 0b00, ".32", v2i32>;
1592 def VDUP8q : VDUPQ<0b11101110, 0b00, ".8", v16i8>;
1593 def VDUP16q : VDUPQ<0b11101010, 0b01, ".16", v8i16>;
1594 def VDUP32q : VDUPQ<0b11101010, 0b00, ".32", v4i32>;
1596 def VDUPfd : NVDup<0b11101000, 0b1011, 0b00, (outs DPR:$dst), (ins GPR:$src),
1597 "vdup", ".32\t$dst, $src",
1598 [(set DPR:$dst, (v2f32 (splat_lo
1600 (f32 (bitconvert GPR:$src))),
1602 def VDUPfq : NVDup<0b11101010, 0b1011, 0b00, (outs QPR:$dst), (ins GPR:$src),
1603 "vdup", ".32\t$dst, $src",
1604 [(set QPR:$dst, (v4f32 (splat_lo
1606 (f32 (bitconvert GPR:$src))),
1609 // VDUP : Vector Duplicate Lane (from scalar to all elements)
1611 def SHUFFLE_get_splat_lane : SDNodeXForm<vector_shuffle, [{
1612 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1613 return CurDAG->getTargetConstant(SVOp->getSplatIndex(), MVT::i32);
1616 def splat_lane : PatFrag<(ops node:$lhs, node:$rhs),
1617 (vector_shuffle node:$lhs, node:$rhs), [{
1618 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1619 return SVOp->isSplat();
1620 }], SHUFFLE_get_splat_lane>;
1622 class VDUPLND<bits<2> op19_18, bits<2> op17_16, string OpcodeStr, ValueType Ty>
1623 : N2V<0b11, 0b11, op19_18, op17_16, 0b11000, 0, 0,
1624 (outs DPR:$dst), (ins DPR:$src, i32imm:$lane),
1625 !strconcat(OpcodeStr, "\t$dst, $src[$lane]"), "",
1626 [(set DPR:$dst, (Ty (splat_lane:$lane DPR:$src, undef)))]>;
1628 // vector_shuffle requires that the source and destination types match, so
1629 // VDUP to a 128-bit result uses a target-specific VDUPLANEQ node.
1630 class VDUPLNQ<bits<2> op19_18, bits<2> op17_16, string OpcodeStr,
1631 ValueType ResTy, ValueType OpTy>
1632 : N2V<0b11, 0b11, op19_18, op17_16, 0b11000, 1, 0,
1633 (outs QPR:$dst), (ins DPR:$src, i32imm:$lane),
1634 !strconcat(OpcodeStr, "\t$dst, $src[$lane]"), "",
1635 [(set QPR:$dst, (ResTy (NEONvduplaneq (OpTy DPR:$src), imm:$lane)))]>;
1637 def VDUPLN8d : VDUPLND<0b00, 0b01, "vdup.8", v8i8>;
1638 def VDUPLN16d : VDUPLND<0b00, 0b10, "vdup.16", v4i16>;
1639 def VDUPLN32d : VDUPLND<0b01, 0b00, "vdup.32", v2i32>;
1640 def VDUPLNfd : VDUPLND<0b01, 0b00, "vdup.32", v2f32>;
1641 def VDUPLN8q : VDUPLNQ<0b00, 0b01, "vdup.8", v16i8, v8i8>;
1642 def VDUPLN16q : VDUPLNQ<0b00, 0b10, "vdup.16", v8i16, v4i16>;
1643 def VDUPLN32q : VDUPLNQ<0b01, 0b00, "vdup.32", v4i32, v2i32>;
1644 def VDUPLNfq : VDUPLNQ<0b01, 0b00, "vdup.32", v4f32, v2f32>;
1646 // VMOVN : Vector Narrowing Move
1647 defm VMOVN : N2VNInt_HSD<0b11,0b11,0b10,0b00100,0,0, "vmovn.i",
1648 int_arm_neon_vmovn>;
1649 // VQMOVN : Vector Saturating Narrowing Move
1650 defm VQMOVNs : N2VNInt_HSD<0b11,0b11,0b10,0b00101,0,0, "vqmovn.s",
1651 int_arm_neon_vqmovns>;
1652 defm VQMOVNu : N2VNInt_HSD<0b11,0b11,0b10,0b00101,1,0, "vqmovn.u",
1653 int_arm_neon_vqmovnu>;
1654 defm VQMOVNsu : N2VNInt_HSD<0b11,0b11,0b10,0b00100,1,0, "vqmovun.s",
1655 int_arm_neon_vqmovnsu>;
1656 // VMOVL : Vector Lengthening Move
1657 defm VMOVLs : N2VLInt_QHS<0,1,0b1010,0,0,1, "vmovl.s", int_arm_neon_vmovls>;
1658 defm VMOVLu : N2VLInt_QHS<1,1,0b1010,0,0,1, "vmovl.u", int_arm_neon_vmovlu>;
1660 // Vector Conversions.
1662 // VCVT : Vector Convert Between Floating-Point and Integers
1663 def VCVTf2sd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt.s32.f32",
1664 v2i32, v2f32, fp_to_sint>;
1665 def VCVTf2ud : N2VD<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt.u32.f32",
1666 v2i32, v2f32, fp_to_uint>;
1667 def VCVTs2fd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt.f32.s32",
1668 v2f32, v2i32, sint_to_fp>;
1669 def VCVTu2fd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt.f32.u32",
1670 v2f32, v2i32, uint_to_fp>;
1672 def VCVTf2sq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt.s32.f32",
1673 v4i32, v4f32, fp_to_sint>;
1674 def VCVTf2uq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt.u32.f32",
1675 v4i32, v4f32, fp_to_uint>;
1676 def VCVTs2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt.f32.s32",
1677 v4f32, v4i32, sint_to_fp>;
1678 def VCVTu2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt.f32.u32",
1679 v4f32, v4i32, uint_to_fp>;
1681 // VCVT : Vector Convert Between Floating-Point and Fixed-Point.
1682 // Note: Some of the opcode bits in the following VCVT instructions need to
1683 // be encoded based on the immed values.
1684 def VCVTf2xsd : N2VCvtD<0, 1, 0b000000, 0b1111, 0, 1, "vcvt.s32.f32",
1685 v2i32, v2f32, int_arm_neon_vcvtfp2fxs>;
1686 def VCVTf2xud : N2VCvtD<1, 1, 0b000000, 0b1111, 0, 1, "vcvt.u32.f32",
1687 v2i32, v2f32, int_arm_neon_vcvtfp2fxu>;
1688 def VCVTxs2fd : N2VCvtD<0, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.s32",
1689 v2f32, v2i32, int_arm_neon_vcvtfxs2fp>;
1690 def VCVTxu2fd : N2VCvtD<1, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.u32",
1691 v2f32, v2i32, int_arm_neon_vcvtfxu2fp>;
1693 def VCVTf2xsq : N2VCvtQ<0, 1, 0b000000, 0b1111, 0, 1, "vcvt.s32.f32",
1694 v4i32, v4f32, int_arm_neon_vcvtfp2fxs>;
1695 def VCVTf2xuq : N2VCvtQ<1, 1, 0b000000, 0b1111, 0, 1, "vcvt.u32.f32",
1696 v4i32, v4f32, int_arm_neon_vcvtfp2fxu>;
1697 def VCVTxs2fq : N2VCvtQ<0, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.s32",
1698 v4f32, v4i32, int_arm_neon_vcvtfxs2fp>;
1699 def VCVTxu2fq : N2VCvtQ<1, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.u32",
1700 v4f32, v4i32, int_arm_neon_vcvtfxu2fp>;
1702 // VREV : Vector Reverse
1704 def vrev64_shuffle : PatFrag<(ops node:$in),
1705 (vector_shuffle node:$in, undef), [{
1706 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1707 return ARM::isVREVMask(SVOp, 64);
1710 def vrev32_shuffle : PatFrag<(ops node:$in),
1711 (vector_shuffle node:$in, undef), [{
1712 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1713 return ARM::isVREVMask(SVOp, 32);
1716 def vrev16_shuffle : PatFrag<(ops node:$in),
1717 (vector_shuffle node:$in, undef), [{
1718 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1719 return ARM::isVREVMask(SVOp, 16);
1722 // VREV64 : Vector Reverse elements within 64-bit doublewords
1724 class VREV64D<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1725 : N2V<0b11, 0b11, op19_18, 0b00, 0b00000, 0, 0, (outs DPR:$dst),
1726 (ins DPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
1727 [(set DPR:$dst, (Ty (vrev64_shuffle (Ty DPR:$src))))]>;
1728 class VREV64Q<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1729 : N2V<0b11, 0b11, op19_18, 0b00, 0b00000, 1, 0, (outs QPR:$dst),
1730 (ins QPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
1731 [(set QPR:$dst, (Ty (vrev64_shuffle (Ty QPR:$src))))]>;
1733 def VREV64d8 : VREV64D<0b00, "vrev64.8", v8i8>;
1734 def VREV64d16 : VREV64D<0b01, "vrev64.16", v4i16>;
1735 def VREV64d32 : VREV64D<0b10, "vrev64.32", v2i32>;
1736 def VREV64df : VREV64D<0b10, "vrev64.32", v2f32>;
1738 def VREV64q8 : VREV64Q<0b00, "vrev64.8", v16i8>;
1739 def VREV64q16 : VREV64Q<0b01, "vrev64.16", v8i16>;
1740 def VREV64q32 : VREV64Q<0b10, "vrev64.32", v4i32>;
1741 def VREV64qf : VREV64Q<0b10, "vrev64.32", v4f32>;
1743 // VREV32 : Vector Reverse elements within 32-bit words
1745 class VREV32D<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1746 : N2V<0b11, 0b11, op19_18, 0b00, 0b00001, 0, 0, (outs DPR:$dst),
1747 (ins DPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
1748 [(set DPR:$dst, (Ty (vrev32_shuffle (Ty DPR:$src))))]>;
1749 class VREV32Q<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1750 : N2V<0b11, 0b11, op19_18, 0b00, 0b00001, 1, 0, (outs QPR:$dst),
1751 (ins QPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
1752 [(set QPR:$dst, (Ty (vrev32_shuffle (Ty QPR:$src))))]>;
1754 def VREV32d8 : VREV32D<0b00, "vrev32.8", v8i8>;
1755 def VREV32d16 : VREV32D<0b01, "vrev32.16", v4i16>;
1757 def VREV32q8 : VREV32Q<0b00, "vrev32.8", v16i8>;
1758 def VREV32q16 : VREV32Q<0b01, "vrev32.16", v8i16>;
1760 // VREV16 : Vector Reverse elements within 16-bit halfwords
1762 class VREV16D<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1763 : N2V<0b11, 0b11, op19_18, 0b00, 0b00010, 0, 0, (outs DPR:$dst),
1764 (ins DPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
1765 [(set DPR:$dst, (Ty (vrev16_shuffle (Ty DPR:$src))))]>;
1766 class VREV16Q<bits<2> op19_18, string OpcodeStr, ValueType Ty>
1767 : N2V<0b11, 0b11, op19_18, 0b00, 0b00010, 1, 0, (outs QPR:$dst),
1768 (ins QPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
1769 [(set QPR:$dst, (Ty (vrev16_shuffle (Ty QPR:$src))))]>;
1771 def VREV16d8 : VREV16D<0b00, "vrev16.8", v8i8>;
1772 def VREV16q8 : VREV16Q<0b00, "vrev16.8", v16i8>;
1774 //===----------------------------------------------------------------------===//
1775 // Non-Instruction Patterns
1776 //===----------------------------------------------------------------------===//
1779 def : Pat<(v1i64 (bitconvert (v2i32 DPR:$src))), (v1i64 DPR:$src)>;
1780 def : Pat<(v1i64 (bitconvert (v4i16 DPR:$src))), (v1i64 DPR:$src)>;
1781 def : Pat<(v1i64 (bitconvert (v8i8 DPR:$src))), (v1i64 DPR:$src)>;
1782 def : Pat<(v1i64 (bitconvert (f64 DPR:$src))), (v1i64 DPR:$src)>;
1783 def : Pat<(v1i64 (bitconvert (v2f32 DPR:$src))), (v1i64 DPR:$src)>;
1784 def : Pat<(v2i32 (bitconvert (v1i64 DPR:$src))), (v2i32 DPR:$src)>;
1785 def : Pat<(v2i32 (bitconvert (v4i16 DPR:$src))), (v2i32 DPR:$src)>;
1786 def : Pat<(v2i32 (bitconvert (v8i8 DPR:$src))), (v2i32 DPR:$src)>;
1787 def : Pat<(v2i32 (bitconvert (f64 DPR:$src))), (v2i32 DPR:$src)>;
1788 def : Pat<(v2i32 (bitconvert (v2f32 DPR:$src))), (v2i32 DPR:$src)>;
1789 def : Pat<(v4i16 (bitconvert (v1i64 DPR:$src))), (v4i16 DPR:$src)>;
1790 def : Pat<(v4i16 (bitconvert (v2i32 DPR:$src))), (v4i16 DPR:$src)>;
1791 def : Pat<(v4i16 (bitconvert (v8i8 DPR:$src))), (v4i16 DPR:$src)>;
1792 def : Pat<(v4i16 (bitconvert (f64 DPR:$src))), (v4i16 DPR:$src)>;
1793 def : Pat<(v4i16 (bitconvert (v2f32 DPR:$src))), (v4i16 DPR:$src)>;
1794 def : Pat<(v8i8 (bitconvert (v1i64 DPR:$src))), (v8i8 DPR:$src)>;
1795 def : Pat<(v8i8 (bitconvert (v2i32 DPR:$src))), (v8i8 DPR:$src)>;
1796 def : Pat<(v8i8 (bitconvert (v4i16 DPR:$src))), (v8i8 DPR:$src)>;
1797 def : Pat<(v8i8 (bitconvert (f64 DPR:$src))), (v8i8 DPR:$src)>;
1798 def : Pat<(v8i8 (bitconvert (v2f32 DPR:$src))), (v8i8 DPR:$src)>;
1799 def : Pat<(f64 (bitconvert (v1i64 DPR:$src))), (f64 DPR:$src)>;
1800 def : Pat<(f64 (bitconvert (v2i32 DPR:$src))), (f64 DPR:$src)>;
1801 def : Pat<(f64 (bitconvert (v4i16 DPR:$src))), (f64 DPR:$src)>;
1802 def : Pat<(f64 (bitconvert (v8i8 DPR:$src))), (f64 DPR:$src)>;
1803 def : Pat<(f64 (bitconvert (v2f32 DPR:$src))), (f64 DPR:$src)>;
1804 def : Pat<(v2f32 (bitconvert (f64 DPR:$src))), (v2f32 DPR:$src)>;
1805 def : Pat<(v2f32 (bitconvert (v1i64 DPR:$src))), (v2f32 DPR:$src)>;
1806 def : Pat<(v2f32 (bitconvert (v2i32 DPR:$src))), (v2f32 DPR:$src)>;
1807 def : Pat<(v2f32 (bitconvert (v4i16 DPR:$src))), (v2f32 DPR:$src)>;
1808 def : Pat<(v2f32 (bitconvert (v8i8 DPR:$src))), (v2f32 DPR:$src)>;
1810 def : Pat<(v2i64 (bitconvert (v4i32 QPR:$src))), (v2i64 QPR:$src)>;
1811 def : Pat<(v2i64 (bitconvert (v8i16 QPR:$src))), (v2i64 QPR:$src)>;
1812 def : Pat<(v2i64 (bitconvert (v16i8 QPR:$src))), (v2i64 QPR:$src)>;
1813 def : Pat<(v2i64 (bitconvert (v2f64 QPR:$src))), (v2i64 QPR:$src)>;
1814 def : Pat<(v2i64 (bitconvert (v4f32 QPR:$src))), (v2i64 QPR:$src)>;
1815 def : Pat<(v4i32 (bitconvert (v2i64 QPR:$src))), (v4i32 QPR:$src)>;
1816 def : Pat<(v4i32 (bitconvert (v8i16 QPR:$src))), (v4i32 QPR:$src)>;
1817 def : Pat<(v4i32 (bitconvert (v16i8 QPR:$src))), (v4i32 QPR:$src)>;
1818 def : Pat<(v4i32 (bitconvert (v2f64 QPR:$src))), (v4i32 QPR:$src)>;
1819 def : Pat<(v4i32 (bitconvert (v4f32 QPR:$src))), (v4i32 QPR:$src)>;
1820 def : Pat<(v8i16 (bitconvert (v2i64 QPR:$src))), (v8i16 QPR:$src)>;
1821 def : Pat<(v8i16 (bitconvert (v4i32 QPR:$src))), (v8i16 QPR:$src)>;
1822 def : Pat<(v8i16 (bitconvert (v16i8 QPR:$src))), (v8i16 QPR:$src)>;
1823 def : Pat<(v8i16 (bitconvert (v2f64 QPR:$src))), (v8i16 QPR:$src)>;
1824 def : Pat<(v8i16 (bitconvert (v4f32 QPR:$src))), (v8i16 QPR:$src)>;
1825 def : Pat<(v16i8 (bitconvert (v2i64 QPR:$src))), (v16i8 QPR:$src)>;
1826 def : Pat<(v16i8 (bitconvert (v4i32 QPR:$src))), (v16i8 QPR:$src)>;
1827 def : Pat<(v16i8 (bitconvert (v8i16 QPR:$src))), (v16i8 QPR:$src)>;
1828 def : Pat<(v16i8 (bitconvert (v2f64 QPR:$src))), (v16i8 QPR:$src)>;
1829 def : Pat<(v16i8 (bitconvert (v4f32 QPR:$src))), (v16i8 QPR:$src)>;
1830 def : Pat<(v4f32 (bitconvert (v2i64 QPR:$src))), (v4f32 QPR:$src)>;
1831 def : Pat<(v4f32 (bitconvert (v4i32 QPR:$src))), (v4f32 QPR:$src)>;
1832 def : Pat<(v4f32 (bitconvert (v8i16 QPR:$src))), (v4f32 QPR:$src)>;
1833 def : Pat<(v4f32 (bitconvert (v16i8 QPR:$src))), (v4f32 QPR:$src)>;
1834 def : Pat<(v4f32 (bitconvert (v2f64 QPR:$src))), (v4f32 QPR:$src)>;
1835 def : Pat<(v2f64 (bitconvert (v2i64 QPR:$src))), (v2f64 QPR:$src)>;
1836 def : Pat<(v2f64 (bitconvert (v4i32 QPR:$src))), (v2f64 QPR:$src)>;
1837 def : Pat<(v2f64 (bitconvert (v8i16 QPR:$src))), (v2f64 QPR:$src)>;
1838 def : Pat<(v2f64 (bitconvert (v16i8 QPR:$src))), (v2f64 QPR:$src)>;
1839 def : Pat<(v2f64 (bitconvert (v4f32 QPR:$src))), (v2f64 QPR:$src)>;