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 //===----------------------------------------------------------------------===//
72 // NEON operand definitions
73 //===----------------------------------------------------------------------===//
75 // addrmode_neonldstm := reg
77 /* TODO: Take advantage of vldm.
78 def addrmode_neonldstm : Operand<i32>,
79 ComplexPattern<i32, 2, "SelectAddrModeNeonLdStM", []> {
80 let PrintMethod = "printAddrNeonLdStMOperand";
81 let MIOperandInfo = (ops GPR, i32imm);
85 //===----------------------------------------------------------------------===//
86 // NEON load / store instructions
87 //===----------------------------------------------------------------------===//
89 /* TODO: Take advantage of vldm.
91 def VLDMD : NI<(outs),
92 (ins addrmode_neonldstm:$addr, reglist:$dst1, variable_ops),
93 "vldm${addr:submode} ${addr:base}, $dst1",
96 def VLDMS : NI<(outs),
97 (ins addrmode_neonldstm:$addr, reglist:$dst1, variable_ops),
98 "vldm${addr:submode} ${addr:base}, $dst1",
103 // Use vldmia to load a Q register as a D register pair.
104 def VLDRQ : NI<(outs QPR:$dst), (ins GPR:$addr),
105 "vldmia $addr, ${dst:dregpair}",
106 [(set QPR:$dst, (v2f64 (load GPR:$addr)))]>;
108 // Use vstmia to store a Q register as a D register pair.
109 def VSTRQ : NI<(outs), (ins QPR:$src, GPR:$addr),
110 "vstmia $addr, ${src:dregpair}",
111 [(store (v2f64 QPR:$src), GPR:$addr)]>;
114 //===----------------------------------------------------------------------===//
115 // NEON pattern fragments
116 //===----------------------------------------------------------------------===//
118 // Extract D sub-registers of Q registers.
119 // (arm_dsubreg_0 is 5; arm_dsubreg_1 is 6)
120 def SubReg_i8_reg : SDNodeXForm<imm, [{
121 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 8, MVT::i32);
123 def SubReg_i16_reg : SDNodeXForm<imm, [{
124 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 4, MVT::i32);
126 def SubReg_i32_reg : SDNodeXForm<imm, [{
127 return CurDAG->getTargetConstant(5 + N->getZExtValue() / 2, MVT::i32);
129 def SubReg_f64_reg : SDNodeXForm<imm, [{
130 return CurDAG->getTargetConstant(5 + N->getZExtValue(), MVT::i32);
133 // Translate lane numbers from Q registers to D subregs.
134 def SubReg_i8_lane : SDNodeXForm<imm, [{
135 return CurDAG->getTargetConstant(N->getZExtValue() & 7, MVT::i32);
137 def SubReg_i16_lane : SDNodeXForm<imm, [{
138 return CurDAG->getTargetConstant(N->getZExtValue() & 3, MVT::i32);
140 def SubReg_i32_lane : SDNodeXForm<imm, [{
141 return CurDAG->getTargetConstant(N->getZExtValue() & 1, MVT::i32);
144 //===----------------------------------------------------------------------===//
145 // Instruction Classes
146 //===----------------------------------------------------------------------===//
148 // Basic 2-register operations, both double- and quad-register.
149 class N2VD<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
150 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
151 ValueType ResTy, ValueType OpTy, SDNode OpNode>
152 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
153 (ins DPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
154 [(set DPR:$dst, (ResTy (OpNode (OpTy DPR:$src))))]>;
155 class N2VQ<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
156 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
157 ValueType ResTy, ValueType OpTy, SDNode OpNode>
158 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
159 (ins QPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
160 [(set QPR:$dst, (ResTy (OpNode (OpTy QPR:$src))))]>;
162 // Basic 2-register intrinsics, both double- and quad-register.
163 class N2VDInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
164 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
165 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
166 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
167 (ins DPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
168 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src))))]>;
169 class N2VQInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
170 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
171 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
172 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
173 (ins QPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
174 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src))))]>;
176 // Narrow 2-register intrinsics.
177 class N2VNInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
178 bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
179 string OpcodeStr, ValueType TyD, ValueType TyQ, Intrinsic IntOp>
180 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, op6, op4, (outs DPR:$dst),
181 (ins QPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
182 [(set DPR:$dst, (TyD (IntOp (TyQ QPR:$src))))]>;
184 // Long 2-register intrinsics. (This is currently only used for VMOVL and is
185 // derived from N2VImm instead of N2V because of the way the size is encoded.)
186 class N2VLInt<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
187 bit op6, bit op4, string OpcodeStr, ValueType TyQ, ValueType TyD,
189 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4, (outs QPR:$dst),
190 (ins DPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
191 [(set QPR:$dst, (TyQ (IntOp (TyD DPR:$src))))]>;
193 // Basic 3-register operations, both double- and quad-register.
194 class N3VD<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
195 string OpcodeStr, ValueType ResTy, ValueType OpTy,
196 SDNode OpNode, bit Commutable>
197 : N3V<op24, op23, op21_20, op11_8, 0, op4,
198 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2),
199 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
200 [(set DPR:$dst, (ResTy (OpNode (OpTy DPR:$src1), (OpTy DPR:$src2))))]> {
201 let isCommutable = Commutable;
203 class N3VQ<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
204 string OpcodeStr, ValueType ResTy, ValueType OpTy,
205 SDNode OpNode, bit Commutable>
206 : N3V<op24, op23, op21_20, op11_8, 1, op4,
207 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2),
208 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
209 [(set QPR:$dst, (ResTy (OpNode (OpTy QPR:$src1), (OpTy QPR:$src2))))]> {
210 let isCommutable = Commutable;
213 // Basic 3-register intrinsics, both double- and quad-register.
214 class N3VDInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
215 string OpcodeStr, ValueType ResTy, ValueType OpTy,
216 Intrinsic IntOp, bit Commutable>
217 : N3V<op24, op23, op21_20, op11_8, 0, op4,
218 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2),
219 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
220 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src1), (OpTy DPR:$src2))))]> {
221 let isCommutable = Commutable;
223 class N3VQInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
224 string OpcodeStr, ValueType ResTy, ValueType OpTy,
225 Intrinsic IntOp, bit Commutable>
226 : N3V<op24, op23, op21_20, op11_8, 1, op4,
227 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2),
228 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
229 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src1), (OpTy QPR:$src2))))]> {
230 let isCommutable = Commutable;
233 // Multiply-Add/Sub operations, both double- and quad-register.
234 class N3VDMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
235 string OpcodeStr, ValueType Ty, SDNode MulOp, SDNode OpNode>
236 : N3V<op24, op23, op21_20, op11_8, 0, op4,
237 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, DPR:$src3),
238 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
239 [(set DPR:$dst, (Ty (OpNode DPR:$src1,
240 (Ty (MulOp DPR:$src2, DPR:$src3)))))]>;
241 class N3VQMulOp<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
242 string OpcodeStr, ValueType Ty, SDNode MulOp, SDNode OpNode>
243 : N3V<op24, op23, op21_20, op11_8, 1, op4,
244 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, QPR:$src3),
245 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
246 [(set QPR:$dst, (Ty (OpNode QPR:$src1,
247 (Ty (MulOp QPR:$src2, QPR:$src3)))))]>;
249 // Neon 3-argument intrinsics, both double- and quad-register.
250 // The destination register is also used as the first source operand register.
251 class N3VDInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
252 string OpcodeStr, ValueType ResTy, ValueType OpTy,
254 : N3V<op24, op23, op21_20, op11_8, 0, op4,
255 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, DPR:$src3),
256 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
257 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src1),
258 (OpTy DPR:$src2), (OpTy DPR:$src3))))]>;
259 class N3VQInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
260 string OpcodeStr, ValueType ResTy, ValueType OpTy,
262 : N3V<op24, op23, op21_20, op11_8, 1, op4,
263 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, QPR:$src3),
264 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
265 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src1),
266 (OpTy QPR:$src2), (OpTy QPR:$src3))))]>;
268 // Neon Long 3-argument intrinsic. The destination register is
269 // a quad-register and is also used as the first source operand register.
270 class N3VLInt3<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
271 string OpcodeStr, ValueType TyQ, ValueType TyD, Intrinsic IntOp>
272 : N3V<op24, op23, op21_20, op11_8, 0, op4,
273 (outs QPR:$dst), (ins QPR:$src1, DPR:$src2, DPR:$src3),
274 !strconcat(OpcodeStr, "\t$dst, $src2, $src3"), "$src1 = $dst",
276 (TyQ (IntOp (TyQ QPR:$src1), (TyD DPR:$src2), (TyD DPR:$src3))))]>;
278 // Narrowing 3-register intrinsics.
279 class N3VNInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
280 string OpcodeStr, ValueType TyD, ValueType TyQ,
281 Intrinsic IntOp, bit Commutable>
282 : N3V<op24, op23, op21_20, op11_8, 0, op4,
283 (outs DPR:$dst), (ins QPR:$src1, QPR:$src2),
284 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
285 [(set DPR:$dst, (TyD (IntOp (TyQ QPR:$src1), (TyQ QPR:$src2))))]> {
286 let isCommutable = Commutable;
289 // Long 3-register intrinsics.
290 class N3VLInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
291 string OpcodeStr, ValueType TyQ, ValueType TyD,
292 Intrinsic IntOp, bit Commutable>
293 : N3V<op24, op23, op21_20, op11_8, 0, op4,
294 (outs QPR:$dst), (ins DPR:$src1, DPR:$src2),
295 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
296 [(set QPR:$dst, (TyQ (IntOp (TyD DPR:$src1), (TyD DPR:$src2))))]> {
297 let isCommutable = Commutable;
300 // Wide 3-register intrinsics.
301 class N3VWInt<bit op24, bit op23, bits<2> op21_20, bits<4> op11_8, bit op4,
302 string OpcodeStr, ValueType TyQ, ValueType TyD,
303 Intrinsic IntOp, bit Commutable>
304 : N3V<op24, op23, op21_20, op11_8, 0, op4,
305 (outs QPR:$dst), (ins QPR:$src1, DPR:$src2),
306 !strconcat(OpcodeStr, "\t$dst, $src1, $src2"), "",
307 [(set QPR:$dst, (TyQ (IntOp (TyQ QPR:$src1), (TyD DPR:$src2))))]> {
308 let isCommutable = Commutable;
311 // Pairwise long 2-register intrinsics, both double- and quad-register.
312 class N2VDPLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
313 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
314 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
315 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4, (outs DPR:$dst),
316 (ins DPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
317 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src))))]>;
318 class N2VQPLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
319 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
320 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
321 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4, (outs QPR:$dst),
322 (ins QPR:$src), !strconcat(OpcodeStr, "\t$dst, $src"), "",
323 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src))))]>;
325 // Pairwise long 2-register accumulate intrinsics,
326 // both double- and quad-register.
327 // The destination register is also used as the first source operand register.
328 class N2VDPLInt2<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
329 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
330 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
331 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 0, op4,
332 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2),
333 !strconcat(OpcodeStr, "\t$dst, $src2"), "$src1 = $dst",
334 [(set DPR:$dst, (ResTy (IntOp (ResTy DPR:$src1), (OpTy DPR:$src2))))]>;
335 class N2VQPLInt2<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
336 bits<2> op17_16, bits<5> op11_7, bit op4, string OpcodeStr,
337 ValueType ResTy, ValueType OpTy, Intrinsic IntOp>
338 : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, 1, op4,
339 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2),
340 !strconcat(OpcodeStr, "\t$dst, $src2"), "$src1 = $dst",
341 [(set QPR:$dst, (ResTy (IntOp (ResTy QPR:$src1), (OpTy QPR:$src2))))]>;
343 // Shift by immediate,
344 // both double- and quad-register.
345 class N2VDSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
346 bit op4, string OpcodeStr, ValueType Ty, SDNode OpNode>
347 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
348 (outs DPR:$dst), (ins DPR:$src, i32imm:$SIMM),
349 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
350 [(set DPR:$dst, (Ty (OpNode (Ty DPR:$src), (i32 imm:$SIMM))))]>;
351 class N2VQSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
352 bit op4, string OpcodeStr, ValueType Ty, SDNode OpNode>
353 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
354 (outs QPR:$dst), (ins QPR:$src, i32imm:$SIMM),
355 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
356 [(set QPR:$dst, (Ty (OpNode (Ty QPR:$src), (i32 imm:$SIMM))))]>;
358 // Long shift by immediate.
359 class N2VLSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
360 bit op6, bit op4, string OpcodeStr, ValueType ResTy,
361 ValueType OpTy, SDNode OpNode>
362 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4,
363 (outs QPR:$dst), (ins DPR:$src, i32imm:$SIMM),
364 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
365 [(set QPR:$dst, (ResTy (OpNode (OpTy DPR:$src),
366 (i32 imm:$SIMM))))]>;
368 // Narrow shift by immediate.
369 class N2VNSh<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
370 bit op6, bit op4, string OpcodeStr, ValueType ResTy,
371 ValueType OpTy, SDNode OpNode>
372 : N2VImm<op24, op23, op21_16, op11_8, op7, op6, op4,
373 (outs DPR:$dst), (ins QPR:$src, i32imm:$SIMM),
374 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
375 [(set DPR:$dst, (ResTy (OpNode (OpTy QPR:$src),
376 (i32 imm:$SIMM))))]>;
378 // Shift right by immediate and accumulate,
379 // both double- and quad-register.
380 class N2VDShAdd<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
381 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
382 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
383 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, i32imm:$SIMM),
384 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
385 [(set DPR:$dst, (Ty (add DPR:$src1,
386 (Ty (ShOp DPR:$src2, (i32 imm:$SIMM))))))]>;
387 class N2VQShAdd<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
388 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
389 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
390 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, i32imm:$SIMM),
391 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
392 [(set QPR:$dst, (Ty (add QPR:$src1,
393 (Ty (ShOp QPR:$src2, (i32 imm:$SIMM))))))]>;
395 // Shift by immediate and insert,
396 // both double- and quad-register.
397 class N2VDShIns<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
398 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
399 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
400 (outs DPR:$dst), (ins DPR:$src1, DPR:$src2, i32imm:$SIMM),
401 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
402 [(set DPR:$dst, (Ty (ShOp DPR:$src1, DPR:$src2, (i32 imm:$SIMM))))]>;
403 class N2VQShIns<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
404 bit op4, string OpcodeStr, ValueType Ty, SDNode ShOp>
405 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
406 (outs QPR:$dst), (ins QPR:$src1, QPR:$src2, i32imm:$SIMM),
407 !strconcat(OpcodeStr, "\t$dst, $src2, $SIMM"), "$src1 = $dst",
408 [(set QPR:$dst, (Ty (ShOp QPR:$src1, QPR:$src2, (i32 imm:$SIMM))))]>;
410 // Convert, with fractional bits immediate,
411 // both double- and quad-register.
412 class N2VCvtD<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
413 bit op4, string OpcodeStr, ValueType ResTy, ValueType OpTy,
415 : N2VImm<op24, op23, op21_16, op11_8, op7, 0, op4,
416 (outs DPR:$dst), (ins DPR:$src, i32imm:$SIMM),
417 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
418 [(set DPR:$dst, (ResTy (IntOp (OpTy DPR:$src), (i32 imm:$SIMM))))]>;
419 class N2VCvtQ<bit op24, bit op23, bits<6> op21_16, bits<4> op11_8, bit op7,
420 bit op4, string OpcodeStr, ValueType ResTy, ValueType OpTy,
422 : N2VImm<op24, op23, op21_16, op11_8, op7, 1, op4,
423 (outs QPR:$dst), (ins QPR:$src, i32imm:$SIMM),
424 !strconcat(OpcodeStr, "\t$dst, $src, $SIMM"), "",
425 [(set QPR:$dst, (ResTy (IntOp (OpTy QPR:$src), (i32 imm:$SIMM))))]>;
427 //===----------------------------------------------------------------------===//
429 //===----------------------------------------------------------------------===//
431 // Neon 3-register vector operations.
433 // First with only element sizes of 8, 16 and 32 bits:
434 multiclass N3V_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
435 string OpcodeStr, SDNode OpNode, bit Commutable = 0> {
436 // 64-bit vector types.
437 def v8i8 : N3VD<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
438 v8i8, v8i8, OpNode, Commutable>;
439 def v4i16 : N3VD<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr, "16"),
440 v4i16, v4i16, OpNode, Commutable>;
441 def v2i32 : N3VD<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr, "32"),
442 v2i32, v2i32, OpNode, Commutable>;
444 // 128-bit vector types.
445 def v16i8 : N3VQ<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
446 v16i8, v16i8, OpNode, Commutable>;
447 def v8i16 : N3VQ<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr, "16"),
448 v8i16, v8i16, OpNode, Commutable>;
449 def v4i32 : N3VQ<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr, "32"),
450 v4i32, v4i32, OpNode, Commutable>;
453 // ....then also with element size 64 bits:
454 multiclass N3V_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
455 string OpcodeStr, SDNode OpNode, bit Commutable = 0>
456 : N3V_QHS<op24, op23, op11_8, op4, OpcodeStr, OpNode, Commutable> {
457 def v1i64 : N3VD<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr, "64"),
458 v1i64, v1i64, OpNode, Commutable>;
459 def v2i64 : N3VQ<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr, "64"),
460 v2i64, v2i64, OpNode, Commutable>;
464 // Neon Narrowing 2-register vector intrinsics,
465 // source operand element sizes of 16, 32 and 64 bits:
466 multiclass N2VNInt_HSD<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
467 bits<5> op11_7, bit op6, bit op4, string OpcodeStr,
469 def v8i8 : N2VNInt<op24_23, op21_20, 0b00, op17_16, op11_7, op6, op4,
470 !strconcat(OpcodeStr, "16"), v8i8, v8i16, IntOp>;
471 def v4i16 : N2VNInt<op24_23, op21_20, 0b01, op17_16, op11_7, op6, op4,
472 !strconcat(OpcodeStr, "32"), v4i16, v4i32, IntOp>;
473 def v2i32 : N2VNInt<op24_23, op21_20, 0b10, op17_16, op11_7, op6, op4,
474 !strconcat(OpcodeStr, "64"), v2i32, v2i64, IntOp>;
478 // Neon Lengthening 2-register vector intrinsic (currently specific to VMOVL).
479 // source operand element sizes of 16, 32 and 64 bits:
480 multiclass N2VLInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op7, bit op6,
481 bit op4, string OpcodeStr, Intrinsic IntOp> {
482 def v8i16 : N2VLInt<op24, op23, 0b001000, op11_8, op7, op6, op4,
483 !strconcat(OpcodeStr, "8"), v8i16, v8i8, IntOp>;
484 def v4i32 : N2VLInt<op24, op23, 0b010000, op11_8, op7, op6, op4,
485 !strconcat(OpcodeStr, "16"), v4i32, v4i16, IntOp>;
486 def v2i64 : N2VLInt<op24, op23, 0b100000, op11_8, op7, op6, op4,
487 !strconcat(OpcodeStr, "32"), v2i64, v2i32, IntOp>;
491 // Neon 3-register vector intrinsics.
493 // First with only element sizes of 16 and 32 bits:
494 multiclass N3VInt_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
495 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
496 // 64-bit vector types.
497 def v4i16 : N3VDInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
498 v4i16, v4i16, IntOp, Commutable>;
499 def v2i32 : N3VDInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
500 v2i32, v2i32, IntOp, Commutable>;
502 // 128-bit vector types.
503 def v8i16 : N3VQInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
504 v8i16, v8i16, IntOp, Commutable>;
505 def v4i32 : N3VQInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
506 v4i32, v4i32, IntOp, Commutable>;
509 // ....then also with element size of 8 bits:
510 multiclass N3VInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
511 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
512 : N3VInt_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
513 def v8i8 : N3VDInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
514 v8i8, v8i8, IntOp, Commutable>;
515 def v16i8 : N3VQInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
516 v16i8, v16i8, IntOp, Commutable>;
519 // ....then also with element size of 64 bits:
520 multiclass N3VInt_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
521 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
522 : N3VInt_QHS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
523 def v1i64 : N3VDInt<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr,"64"),
524 v1i64, v1i64, IntOp, Commutable>;
525 def v2i64 : N3VQInt<op24, op23, 0b11, op11_8, op4, !strconcat(OpcodeStr,"64"),
526 v2i64, v2i64, IntOp, Commutable>;
530 // Neon Narrowing 3-register vector intrinsics,
531 // source operand element sizes of 16, 32 and 64 bits:
532 multiclass N3VNInt_HSD<bit op24, bit op23, bits<4> op11_8, bit op4,
533 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
534 def v8i8 : N3VNInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr,"16"),
535 v8i8, v8i16, IntOp, Commutable>;
536 def v4i16 : N3VNInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"32"),
537 v4i16, v4i32, IntOp, Commutable>;
538 def v2i32 : N3VNInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"64"),
539 v2i32, v2i64, IntOp, Commutable>;
543 // Neon Long 3-register vector intrinsics.
545 // First with only element sizes of 16 and 32 bits:
546 multiclass N3VLInt_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
547 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
548 def v4i32 : N3VLInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
549 v4i32, v4i16, IntOp, Commutable>;
550 def v2i64 : N3VLInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
551 v2i64, v2i32, IntOp, Commutable>;
554 // ....then also with element size of 8 bits:
555 multiclass N3VLInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
556 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0>
557 : N3VLInt_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp, Commutable> {
558 def v8i16 : N3VLInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
559 v8i16, v8i8, IntOp, Commutable>;
563 // Neon Wide 3-register vector intrinsics,
564 // source operand element sizes of 8, 16 and 32 bits:
565 multiclass N3VWInt_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
566 string OpcodeStr, Intrinsic IntOp, bit Commutable = 0> {
567 def v8i16 : N3VWInt<op24, op23, 0b00, op11_8, op4, !strconcat(OpcodeStr, "8"),
568 v8i16, v8i8, IntOp, Commutable>;
569 def v4i32 : N3VWInt<op24, op23, 0b01, op11_8, op4, !strconcat(OpcodeStr,"16"),
570 v4i32, v4i16, IntOp, Commutable>;
571 def v2i64 : N3VWInt<op24, op23, 0b10, op11_8, op4, !strconcat(OpcodeStr,"32"),
572 v2i64, v2i32, IntOp, Commutable>;
576 // Neon Multiply-Op vector operations,
577 // element sizes of 8, 16 and 32 bits:
578 multiclass N3VMulOp_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
579 string OpcodeStr, SDNode OpNode> {
580 // 64-bit vector types.
581 def v8i8 : N3VDMulOp<op24, op23, 0b00, op11_8, op4,
582 !strconcat(OpcodeStr, "8"), v8i8, mul, OpNode>;
583 def v4i16 : N3VDMulOp<op24, op23, 0b01, op11_8, op4,
584 !strconcat(OpcodeStr, "16"), v4i16, mul, OpNode>;
585 def v2i32 : N3VDMulOp<op24, op23, 0b10, op11_8, op4,
586 !strconcat(OpcodeStr, "32"), v2i32, mul, OpNode>;
588 // 128-bit vector types.
589 def v16i8 : N3VQMulOp<op24, op23, 0b00, op11_8, op4,
590 !strconcat(OpcodeStr, "8"), v16i8, mul, OpNode>;
591 def v8i16 : N3VQMulOp<op24, op23, 0b01, op11_8, op4,
592 !strconcat(OpcodeStr, "16"), v8i16, mul, OpNode>;
593 def v4i32 : N3VQMulOp<op24, op23, 0b10, op11_8, op4,
594 !strconcat(OpcodeStr, "32"), v4i32, mul, OpNode>;
598 // Neon 3-argument intrinsics,
599 // element sizes of 8, 16 and 32 bits:
600 multiclass N3VInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
601 string OpcodeStr, Intrinsic IntOp> {
602 // 64-bit vector types.
603 def v8i8 : N3VDInt3<op24, op23, 0b00, op11_8, op4,
604 !strconcat(OpcodeStr, "8"), v8i8, v8i8, IntOp>;
605 def v4i16 : N3VDInt3<op24, op23, 0b01, op11_8, op4,
606 !strconcat(OpcodeStr, "16"), v4i16, v4i16, IntOp>;
607 def v2i32 : N3VDInt3<op24, op23, 0b10, op11_8, op4,
608 !strconcat(OpcodeStr, "32"), v2i32, v2i32, IntOp>;
610 // 128-bit vector types.
611 def v16i8 : N3VQInt3<op24, op23, 0b00, op11_8, op4,
612 !strconcat(OpcodeStr, "8"), v16i8, v16i8, IntOp>;
613 def v8i16 : N3VQInt3<op24, op23, 0b01, op11_8, op4,
614 !strconcat(OpcodeStr, "16"), v8i16, v8i16, IntOp>;
615 def v4i32 : N3VQInt3<op24, op23, 0b10, op11_8, op4,
616 !strconcat(OpcodeStr, "32"), v4i32, v4i32, IntOp>;
620 // Neon Long 3-argument intrinsics.
622 // First with only element sizes of 16 and 32 bits:
623 multiclass N3VLInt3_HS<bit op24, bit op23, bits<4> op11_8, bit op4,
624 string OpcodeStr, Intrinsic IntOp> {
625 def v4i32 : N3VLInt3<op24, op23, 0b01, op11_8, op4,
626 !strconcat(OpcodeStr, "16"), v4i32, v4i16, IntOp>;
627 def v2i64 : N3VLInt3<op24, op23, 0b10, op11_8, op4,
628 !strconcat(OpcodeStr, "32"), v2i64, v2i32, IntOp>;
631 // ....then also with element size of 8 bits:
632 multiclass N3VLInt3_QHS<bit op24, bit op23, bits<4> op11_8, bit op4,
633 string OpcodeStr, Intrinsic IntOp>
634 : N3VLInt3_HS<op24, op23, op11_8, op4, OpcodeStr, IntOp> {
635 def v8i16 : N3VLInt3<op24, op23, 0b01, op11_8, op4,
636 !strconcat(OpcodeStr, "8"), v8i16, v8i8, IntOp>;
640 // Neon 2-register vector intrinsics,
641 // element sizes of 8, 16 and 32 bits:
642 multiclass N2VInt_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
643 bits<5> op11_7, bit op4, string OpcodeStr,
645 // 64-bit vector types.
646 def v8i8 : N2VDInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
647 !strconcat(OpcodeStr, "8"), v8i8, v8i8, IntOp>;
648 def v4i16 : N2VDInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
649 !strconcat(OpcodeStr, "16"), v4i16, v4i16, IntOp>;
650 def v2i32 : N2VDInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
651 !strconcat(OpcodeStr, "32"), v2i32, v2i32, IntOp>;
653 // 128-bit vector types.
654 def v16i8 : N2VQInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
655 !strconcat(OpcodeStr, "8"), v16i8, v16i8, IntOp>;
656 def v8i16 : N2VQInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
657 !strconcat(OpcodeStr, "16"), v8i16, v8i16, IntOp>;
658 def v4i32 : N2VQInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
659 !strconcat(OpcodeStr, "32"), v4i32, v4i32, IntOp>;
663 // Neon Pairwise long 2-register intrinsics,
664 // element sizes of 8, 16 and 32 bits:
665 multiclass N2VPLInt_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
666 bits<5> op11_7, bit op4,
667 string OpcodeStr, Intrinsic IntOp> {
668 // 64-bit vector types.
669 def v8i8 : N2VDPLInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
670 !strconcat(OpcodeStr, "8"), v4i16, v8i8, IntOp>;
671 def v4i16 : N2VDPLInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
672 !strconcat(OpcodeStr, "16"), v2i32, v4i16, IntOp>;
673 def v2i32 : N2VDPLInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
674 !strconcat(OpcodeStr, "32"), v1i64, v2i32, IntOp>;
676 // 128-bit vector types.
677 def v16i8 : N2VQPLInt<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
678 !strconcat(OpcodeStr, "8"), v8i16, v16i8, IntOp>;
679 def v8i16 : N2VQPLInt<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
680 !strconcat(OpcodeStr, "16"), v4i32, v8i16, IntOp>;
681 def v4i32 : N2VQPLInt<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
682 !strconcat(OpcodeStr, "32"), v2i64, v4i32, IntOp>;
686 // Neon Pairwise long 2-register accumulate intrinsics,
687 // element sizes of 8, 16 and 32 bits:
688 multiclass N2VPLInt2_QHS<bits<2> op24_23, bits<2> op21_20, bits<2> op17_16,
689 bits<5> op11_7, bit op4,
690 string OpcodeStr, Intrinsic IntOp> {
691 // 64-bit vector types.
692 def v8i8 : N2VDPLInt2<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
693 !strconcat(OpcodeStr, "8"), v4i16, v8i8, IntOp>;
694 def v4i16 : N2VDPLInt2<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
695 !strconcat(OpcodeStr, "16"), v2i32, v4i16, IntOp>;
696 def v2i32 : N2VDPLInt2<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
697 !strconcat(OpcodeStr, "32"), v1i64, v2i32, IntOp>;
699 // 128-bit vector types.
700 def v16i8 : N2VQPLInt2<op24_23, op21_20, 0b00, op17_16, op11_7, op4,
701 !strconcat(OpcodeStr, "8"), v8i16, v16i8, IntOp>;
702 def v8i16 : N2VQPLInt2<op24_23, op21_20, 0b01, op17_16, op11_7, op4,
703 !strconcat(OpcodeStr, "16"), v4i32, v8i16, IntOp>;
704 def v4i32 : N2VQPLInt2<op24_23, op21_20, 0b10, op17_16, op11_7, op4,
705 !strconcat(OpcodeStr, "32"), v2i64, v4i32, IntOp>;
709 // Neon 2-register vector shift by immediate,
710 // element sizes of 8, 16, 32 and 64 bits:
711 multiclass N2VSh_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
712 string OpcodeStr, SDNode OpNode> {
713 // 64-bit vector types.
714 def v8i8 : N2VDSh<op24, op23, 0b001000, op11_8, 0, op4,
715 !strconcat(OpcodeStr, "8"), v8i8, OpNode>;
716 def v4i16 : N2VDSh<op24, op23, 0b010000, op11_8, 0, op4,
717 !strconcat(OpcodeStr, "16"), v4i16, OpNode>;
718 def v2i32 : N2VDSh<op24, op23, 0b100000, op11_8, 0, op4,
719 !strconcat(OpcodeStr, "32"), v2i32, OpNode>;
720 def v1i64 : N2VDSh<op24, op23, 0b000000, op11_8, 1, op4,
721 !strconcat(OpcodeStr, "64"), v1i64, OpNode>;
723 // 128-bit vector types.
724 def v16i8 : N2VQSh<op24, op23, 0b001000, op11_8, 0, op4,
725 !strconcat(OpcodeStr, "8"), v16i8, OpNode>;
726 def v8i16 : N2VQSh<op24, op23, 0b010000, op11_8, 0, op4,
727 !strconcat(OpcodeStr, "16"), v8i16, OpNode>;
728 def v4i32 : N2VQSh<op24, op23, 0b100000, op11_8, 0, op4,
729 !strconcat(OpcodeStr, "32"), v4i32, OpNode>;
730 def v2i64 : N2VQSh<op24, op23, 0b000000, op11_8, 1, op4,
731 !strconcat(OpcodeStr, "64"), v2i64, OpNode>;
735 // Neon Shift-Accumulate vector operations,
736 // element sizes of 8, 16, 32 and 64 bits:
737 multiclass N2VShAdd_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
738 string OpcodeStr, SDNode ShOp> {
739 // 64-bit vector types.
740 def v8i8 : N2VDShAdd<op24, op23, 0b001000, op11_8, 0, op4,
741 !strconcat(OpcodeStr, "8"), v8i8, ShOp>;
742 def v4i16 : N2VDShAdd<op24, op23, 0b010000, op11_8, 0, op4,
743 !strconcat(OpcodeStr, "16"), v4i16, ShOp>;
744 def v2i32 : N2VDShAdd<op24, op23, 0b100000, op11_8, 0, op4,
745 !strconcat(OpcodeStr, "32"), v2i32, ShOp>;
746 def v1i64 : N2VDShAdd<op24, op23, 0b000000, op11_8, 1, op4,
747 !strconcat(OpcodeStr, "64"), v1i64, ShOp>;
749 // 128-bit vector types.
750 def v16i8 : N2VQShAdd<op24, op23, 0b001000, op11_8, 0, op4,
751 !strconcat(OpcodeStr, "8"), v16i8, ShOp>;
752 def v8i16 : N2VQShAdd<op24, op23, 0b010000, op11_8, 0, op4,
753 !strconcat(OpcodeStr, "16"), v8i16, ShOp>;
754 def v4i32 : N2VQShAdd<op24, op23, 0b100000, op11_8, 0, op4,
755 !strconcat(OpcodeStr, "32"), v4i32, ShOp>;
756 def v2i64 : N2VQShAdd<op24, op23, 0b000000, op11_8, 1, op4,
757 !strconcat(OpcodeStr, "64"), v2i64, ShOp>;
761 // Neon Shift-Insert vector operations,
762 // element sizes of 8, 16, 32 and 64 bits:
763 multiclass N2VShIns_QHSD<bit op24, bit op23, bits<4> op11_8, bit op4,
764 string OpcodeStr, SDNode ShOp> {
765 // 64-bit vector types.
766 def v8i8 : N2VDShIns<op24, op23, 0b001000, op11_8, 0, op4,
767 !strconcat(OpcodeStr, "8"), v8i8, ShOp>;
768 def v4i16 : N2VDShIns<op24, op23, 0b010000, op11_8, 0, op4,
769 !strconcat(OpcodeStr, "16"), v4i16, ShOp>;
770 def v2i32 : N2VDShIns<op24, op23, 0b100000, op11_8, 0, op4,
771 !strconcat(OpcodeStr, "32"), v2i32, ShOp>;
772 def v1i64 : N2VDShIns<op24, op23, 0b000000, op11_8, 1, op4,
773 !strconcat(OpcodeStr, "64"), v1i64, ShOp>;
775 // 128-bit vector types.
776 def v16i8 : N2VQShIns<op24, op23, 0b001000, op11_8, 0, op4,
777 !strconcat(OpcodeStr, "8"), v16i8, ShOp>;
778 def v8i16 : N2VQShIns<op24, op23, 0b010000, op11_8, 0, op4,
779 !strconcat(OpcodeStr, "16"), v8i16, ShOp>;
780 def v4i32 : N2VQShIns<op24, op23, 0b100000, op11_8, 0, op4,
781 !strconcat(OpcodeStr, "32"), v4i32, ShOp>;
782 def v2i64 : N2VQShIns<op24, op23, 0b000000, op11_8, 1, op4,
783 !strconcat(OpcodeStr, "64"), v2i64, ShOp>;
786 //===----------------------------------------------------------------------===//
787 // Instruction Definitions.
788 //===----------------------------------------------------------------------===//
790 // Vector Add Operations.
792 // VADD : Vector Add (integer and floating-point)
793 defm VADD : N3V_QHSD<0, 0, 0b1000, 0, "vadd.i", add, 1>;
794 def VADDfd : N3VD<0, 0, 0b00, 0b1101, 0, "vadd.f32", v2f32, v2f32, fadd, 1>;
795 def VADDfq : N3VQ<0, 0, 0b00, 0b1101, 0, "vadd.f32", v4f32, v4f32, fadd, 1>;
796 // VADDL : Vector Add Long (Q = D + D)
797 defm VADDLs : N3VLInt_QHS<0,1,0b0000,0, "vaddl.s", int_arm_neon_vaddls, 1>;
798 defm VADDLu : N3VLInt_QHS<1,1,0b0000,0, "vaddl.u", int_arm_neon_vaddlu, 1>;
799 // VADDW : Vector Add Wide (Q = Q + D)
800 defm VADDWs : N3VWInt_QHS<0,1,0b0001,0, "vaddw.s", int_arm_neon_vaddws, 0>;
801 defm VADDWu : N3VWInt_QHS<1,1,0b0001,0, "vaddw.u", int_arm_neon_vaddwu, 0>;
802 // VHADD : Vector Halving Add
803 defm VHADDs : N3VInt_QHS<0,0,0b0000,0, "vhadd.s", int_arm_neon_vhadds, 1>;
804 defm VHADDu : N3VInt_QHS<1,0,0b0000,0, "vhadd.u", int_arm_neon_vhaddu, 1>;
805 // VRHADD : Vector Rounding Halving Add
806 defm VRHADDs : N3VInt_QHS<0,0,0b0001,0, "vrhadd.s", int_arm_neon_vrhadds, 1>;
807 defm VRHADDu : N3VInt_QHS<1,0,0b0001,0, "vrhadd.u", int_arm_neon_vrhaddu, 1>;
808 // VQADD : Vector Saturating Add
809 defm VQADDs : N3VInt_QHSD<0,0,0b0000,1, "vqadd.s", int_arm_neon_vqadds, 1>;
810 defm VQADDu : N3VInt_QHSD<1,0,0b0000,1, "vqadd.u", int_arm_neon_vqaddu, 1>;
811 // VADDHN : Vector Add and Narrow Returning High Half (D = Q + Q)
812 defm VADDHN : N3VNInt_HSD<0,1,0b0100,0, "vaddhn.i", int_arm_neon_vaddhn, 1>;
813 // VRADDHN : Vector Rounding Add and Narrow Returning High Half (D = Q + Q)
814 defm VRADDHN : N3VNInt_HSD<1,1,0b0100,0, "vraddhn.i", int_arm_neon_vraddhn, 1>;
816 // Vector Multiply Operations.
818 // VMUL : Vector Multiply (integer, polynomial and floating-point)
819 defm VMUL : N3V_QHS<0, 0, 0b1001, 1, "vmul.i", mul, 1>;
820 def VMULpd : N3VDInt<1, 0, 0b00, 0b1001, 1, "vmul.p8", v8i8, v8i8,
821 int_arm_neon_vmulp, 1>;
822 def VMULpq : N3VQInt<1, 0, 0b00, 0b1001, 1, "vmul.p8", v16i8, v16i8,
823 int_arm_neon_vmulp, 1>;
824 def VMULfd : N3VD<1, 0, 0b00, 0b1101, 1, "vmul.f32", v2f32, v2f32, fmul, 1>;
825 def VMULfq : N3VQ<1, 0, 0b00, 0b1101, 1, "vmul.f32", v4f32, v4f32, fmul, 1>;
826 // VQDMULH : Vector Saturating Doubling Multiply Returning High Half
827 defm VQDMULH : N3VInt_HS<0,0,0b1011,0, "vqdmulh.s", int_arm_neon_vqdmulh, 1>;
828 // VQRDMULH : Vector Rounding Saturating Doubling Multiply Returning High Half
829 defm VQRDMULH : N3VInt_HS<1,0,0b1011,0, "vqrdmulh.s", int_arm_neon_vqrdmulh, 1>;
830 // VMULL : Vector Multiply Long (integer and polynomial) (Q = D * D)
831 defm VMULLs : N3VLInt_QHS<0,1,0b1100,0, "vmull.s", int_arm_neon_vmulls, 1>;
832 defm VMULLu : N3VLInt_QHS<1,1,0b1100,0, "vmull.u", int_arm_neon_vmullu, 1>;
833 def VMULLp : N3VLInt<0, 1, 0b00, 0b1110, 0, "vmull.p8", v8i16, v8i8,
834 int_arm_neon_vmullp, 1>;
835 // VQDMULL : Vector Saturating Doubling Multiply Long (Q = D * D)
836 defm VQDMULL : N3VLInt_HS<0,1,0b1101,0, "vqdmull.s", int_arm_neon_vqdmull, 1>;
838 // Vector Multiply-Accumulate and Multiply-Subtract Operations.
840 // VMLA : Vector Multiply Accumulate (integer and floating-point)
841 defm VMLA : N3VMulOp_QHS<0, 0, 0b1001, 0, "vmla.i", add>;
842 def VMLAfd : N3VDMulOp<0, 0, 0b00, 0b1101, 1, "vmla.f32", v2f32, fmul, fadd>;
843 def VMLAfq : N3VQMulOp<0, 0, 0b00, 0b1101, 1, "vmla.f32", v4f32, fmul, fadd>;
844 // VMLAL : Vector Multiply Accumulate Long (Q += D * D)
845 defm VMLALs : N3VLInt3_QHS<0,1,0b1000,0, "vmlal.s", int_arm_neon_vmlals>;
846 defm VMLALu : N3VLInt3_QHS<1,1,0b1000,0, "vmlal.u", int_arm_neon_vmlalu>;
847 // VQDMLAL : Vector Saturating Doubling Multiply Accumulate Long (Q += D * D)
848 defm VQDMLAL : N3VLInt3_HS<0, 1, 0b1001, 0, "vqdmlal.s", int_arm_neon_vqdmlal>;
849 // VMLS : Vector Multiply Subtract (integer and floating-point)
850 defm VMLS : N3VMulOp_QHS<0, 0, 0b1001, 0, "vmls.i", sub>;
851 def VMLSfd : N3VDMulOp<0, 0, 0b10, 0b1101, 1, "vmls.f32", v2f32, fmul, fsub>;
852 def VMLSfq : N3VQMulOp<0, 0, 0b10, 0b1101, 1, "vmls.f32", v4f32, fmul, fsub>;
853 // VMLSL : Vector Multiply Subtract Long (Q -= D * D)
854 defm VMLSLs : N3VLInt3_QHS<0,1,0b1010,0, "vmlsl.s", int_arm_neon_vmlsls>;
855 defm VMLSLu : N3VLInt3_QHS<1,1,0b1010,0, "vmlsl.u", int_arm_neon_vmlslu>;
856 // VQDMLSL : Vector Saturating Doubling Multiply Subtract Long (Q -= D * D)
857 defm VQDMLSL : N3VLInt3_HS<0, 1, 0b1011, 0, "vqdmlsl.s", int_arm_neon_vqdmlsl>;
859 // Vector Subtract Operations.
861 // VSUB : Vector Subtract (integer and floating-point)
862 defm VSUB : N3V_QHSD<1, 0, 0b1000, 0, "vsub.i", sub, 0>;
863 def VSUBfd : N3VD<0, 0, 0b10, 0b1101, 0, "vsub.f32", v2f32, v2f32, fsub, 0>;
864 def VSUBfq : N3VQ<0, 0, 0b10, 0b1101, 0, "vsub.f32", v4f32, v4f32, fsub, 0>;
865 // VSUBL : Vector Subtract Long (Q = D - D)
866 defm VSUBLs : N3VLInt_QHS<0,1,0b0010,0, "vsubl.s", int_arm_neon_vsubls, 1>;
867 defm VSUBLu : N3VLInt_QHS<1,1,0b0010,0, "vsubl.u", int_arm_neon_vsublu, 1>;
868 // VSUBW : Vector Subtract Wide (Q = Q - D)
869 defm VSUBWs : N3VWInt_QHS<0,1,0b0011,0, "vsubw.s", int_arm_neon_vsubws, 0>;
870 defm VSUBWu : N3VWInt_QHS<1,1,0b0011,0, "vsubw.u", int_arm_neon_vsubwu, 0>;
871 // VHSUB : Vector Halving Subtract
872 defm VHSUBs : N3VInt_QHS<0, 0, 0b0010, 0, "vhsub.s", int_arm_neon_vhsubs, 0>;
873 defm VHSUBu : N3VInt_QHS<1, 0, 0b0010, 0, "vhsub.u", int_arm_neon_vhsubu, 0>;
874 // VQSUB : Vector Saturing Subtract
875 defm VQSUBs : N3VInt_QHSD<0, 0, 0b0010, 1, "vqsub.s", int_arm_neon_vqsubs, 0>;
876 defm VQSUBu : N3VInt_QHSD<1, 0, 0b0010, 1, "vqsub.u", int_arm_neon_vqsubu, 0>;
877 // VSUBHN : Vector Subtract and Narrow Returning High Half (D = Q - Q)
878 defm VSUBHN : N3VNInt_HSD<0,1,0b0110,0, "vsubhn.i", int_arm_neon_vsubhn, 0>;
879 // VRSUBHN : Vector Rounding Subtract and Narrow Returning High Half (D=Q-Q)
880 defm VRSUBHN : N3VNInt_HSD<1,1,0b0110,0, "vrsubhn.i", int_arm_neon_vrsubhn, 0>;
882 // Vector Comparisons.
884 // VCEQ : Vector Compare Equal
885 defm VCEQ : N3V_QHS<1, 0, 0b1000, 1, "vceq.i", NEONvceq, 1>;
886 def VCEQfd : N3VD<0,0,0b00,0b1110,0, "vceq.f32", v2i32, v2f32, NEONvceq, 1>;
887 def VCEQfq : N3VQ<0,0,0b00,0b1110,0, "vceq.f32", v4i32, v4f32, NEONvceq, 1>;
888 // VCGE : Vector Compare Greater Than or Equal
889 defm VCGEs : N3V_QHS<0, 0, 0b0011, 1, "vcge.s", NEONvcge, 0>;
890 defm VCGEu : N3V_QHS<1, 0, 0b0011, 1, "vcge.u", NEONvcgeu, 0>;
891 def VCGEfd : N3VD<1,0,0b00,0b1110,0, "vcge.f32", v2i32, v2f32, NEONvcge, 0>;
892 def VCGEfq : N3VQ<1,0,0b00,0b1110,0, "vcge.f32", v4i32, v4f32, NEONvcge, 0>;
893 // VCGT : Vector Compare Greater Than
894 defm VCGTs : N3V_QHS<0, 0, 0b0011, 0, "vcgt.s", NEONvcgt, 0>;
895 defm VCGTu : N3V_QHS<1, 0, 0b0011, 0, "vcgt.u", NEONvcgtu, 0>;
896 def VCGTfd : N3VD<1,0,0b10,0b1110,0, "vcgt.f32", v2i32, v2f32, NEONvcgt, 0>;
897 def VCGTfq : N3VQ<1,0,0b10,0b1110,0, "vcgt.f32", v4i32, v4f32, NEONvcgt, 0>;
898 // VACGE : Vector Absolute Compare Greater Than or Equal (aka VCAGE)
899 def VACGEd : N3VDInt<1, 0, 0b00, 0b1110, 1, "vacge.f32", v2i32, v2f32,
900 int_arm_neon_vacged, 0>;
901 def VACGEq : N3VQInt<1, 0, 0b00, 0b1110, 1, "vacge.f32", v4i32, v4f32,
902 int_arm_neon_vacgeq, 0>;
903 // VACGT : Vector Absolute Compare Greater Than (aka VCAGT)
904 def VACGTd : N3VDInt<1, 0, 0b10, 0b1110, 1, "vacgt.f32", v2i32, v2f32,
905 int_arm_neon_vacgtd, 0>;
906 def VACGTq : N3VQInt<1, 0, 0b10, 0b1110, 1, "vacgt.f32", v4i32, v4f32,
907 int_arm_neon_vacgtq, 0>;
908 // VTST : Vector Test Bits
909 defm VTST : N3V_QHS<0, 0, 0b1000, 1, "vtst.i", NEONvtst, 1>;
911 // Vector Bitwise Operations.
913 // VAND : Vector Bitwise AND
914 def VANDd : N3VD<0, 0, 0b00, 0b0001, 1, "vand", v2i32, v2i32, and, 1>;
915 def VANDq : N3VQ<0, 0, 0b00, 0b0001, 1, "vand", v4i32, v4i32, and, 1>;
917 // VEOR : Vector Bitwise Exclusive OR
918 def VEORd : N3VD<1, 0, 0b00, 0b0001, 1, "veor", v2i32, v2i32, xor, 1>;
919 def VEORq : N3VQ<1, 0, 0b00, 0b0001, 1, "veor", v4i32, v4i32, xor, 1>;
921 // VORR : Vector Bitwise OR
922 def VORRd : N3VD<0, 0, 0b10, 0b0001, 1, "vorr", v2i32, v2i32, or, 1>;
923 def VORRq : N3VQ<0, 0, 0b10, 0b0001, 1, "vorr", v4i32, v4i32, or, 1>;
925 // VBIC : Vector Bitwise Bit Clear (AND NOT)
926 def VBICd : N3V<0, 0, 0b01, 0b0001, 0, 1, (outs DPR:$dst),
927 (ins DPR:$src1, DPR:$src2), "vbic\t$dst, $src1, $src2", "",
928 [(set DPR:$dst, (v2i32 (and DPR:$src1,(vnot DPR:$src2))))]>;
929 def VBICq : N3V<0, 0, 0b01, 0b0001, 1, 1, (outs QPR:$dst),
930 (ins QPR:$src1, QPR:$src2), "vbic\t$dst, $src1, $src2", "",
931 [(set QPR:$dst, (v4i32 (and QPR:$src1,(vnot QPR:$src2))))]>;
933 // VORN : Vector Bitwise OR NOT
934 def VORNd : N3V<0, 0, 0b11, 0b0001, 0, 1, (outs DPR:$dst),
935 (ins DPR:$src1, DPR:$src2), "vorn\t$dst, $src1, $src2", "",
936 [(set DPR:$dst, (v2i32 (or DPR:$src1, (vnot DPR:$src2))))]>;
937 def VORNq : N3V<0, 0, 0b11, 0b0001, 1, 1, (outs QPR:$dst),
938 (ins QPR:$src1, QPR:$src2), "vorn\t$dst, $src1, $src2", "",
939 [(set QPR:$dst, (v4i32 (or QPR:$src1, (vnot QPR:$src2))))]>;
941 // VMVN : Vector Bitwise NOT
942 def VMVNd : N2V<0b11, 0b11, 0b00, 0b00, 0b01011, 0, 0,
943 (outs DPR:$dst), (ins DPR:$src), "vmvn\t$dst, $src", "",
944 [(set DPR:$dst, (v2i32 (vnot DPR:$src)))]>;
945 def VMVNq : N2V<0b11, 0b11, 0b00, 0b00, 0b01011, 1, 0,
946 (outs QPR:$dst), (ins QPR:$src), "vmvn\t$dst, $src", "",
947 [(set QPR:$dst, (v4i32 (vnot QPR:$src)))]>;
948 def : Pat<(v2i32 (vnot_conv DPR:$src)), (VMVNd DPR:$src)>;
949 def : Pat<(v4i32 (vnot_conv QPR:$src)), (VMVNq QPR:$src)>;
951 // VBSL : Vector Bitwise Select
952 def VBSLd : N3V<1, 0, 0b01, 0b0001, 0, 1, (outs DPR:$dst),
953 (ins DPR:$src1, DPR:$src2, DPR:$src3),
954 "vbsl\t$dst, $src2, $src3", "$src1 = $dst",
956 (v2i32 (or (and DPR:$src2, DPR:$src1),
957 (and DPR:$src3, (vnot DPR:$src1)))))]>;
958 def VBSLq : N3V<1, 0, 0b01, 0b0001, 1, 1, (outs QPR:$dst),
959 (ins QPR:$src1, QPR:$src2, QPR:$src3),
960 "vbsl\t$dst, $src2, $src3", "$src1 = $dst",
962 (v4i32 (or (and QPR:$src2, QPR:$src1),
963 (and QPR:$src3, (vnot QPR:$src1)))))]>;
965 // VBIF : Vector Bitwise Insert if False
966 // like VBSL but with: "vbif\t$dst, $src3, $src1", "$src2 = $dst",
967 // VBIT : Vector Bitwise Insert if True
968 // like VBSL but with: "vbit\t$dst, $src2, $src1", "$src3 = $dst",
969 // These are not yet implemented. The TwoAddress pass will not go looking
970 // for equivalent operations with different register constraints; it just
973 // Vector Absolute Differences.
975 // VABD : Vector Absolute Difference
976 defm VABDs : N3VInt_QHS<0, 0, 0b0111, 0, "vabd.s", int_arm_neon_vabds, 0>;
977 defm VABDu : N3VInt_QHS<1, 0, 0b0111, 0, "vabd.u", int_arm_neon_vabdu, 0>;
978 def VABDfd : N3VDInt<1, 0, 0b10, 0b1101, 0, "vabd.f32", v2f32, v2f32,
979 int_arm_neon_vabdf, 0>;
980 def VABDfq : N3VQInt<1, 0, 0b10, 0b1101, 0, "vabd.f32", v4f32, v4f32,
981 int_arm_neon_vabdf, 0>;
983 // VABDL : Vector Absolute Difference Long (Q = | D - D |)
984 defm VABDLs : N3VLInt_QHS<0,1,0b0111,0, "vabdl.s", int_arm_neon_vabdls, 0>;
985 defm VABDLu : N3VLInt_QHS<1,1,0b0111,0, "vabdl.u", int_arm_neon_vabdlu, 0>;
987 // VABA : Vector Absolute Difference and Accumulate
988 defm VABAs : N3VInt3_QHS<0,1,0b0101,0, "vaba.s", int_arm_neon_vabas>;
989 defm VABAu : N3VInt3_QHS<1,1,0b0101,0, "vaba.u", int_arm_neon_vabau>;
991 // VABAL : Vector Absolute Difference and Accumulate Long (Q += | D - D |)
992 defm VABALs : N3VLInt3_QHS<0,1,0b0101,0, "vabal.s", int_arm_neon_vabals>;
993 defm VABALu : N3VLInt3_QHS<1,1,0b0101,0, "vabal.u", int_arm_neon_vabalu>;
995 // Vector Maximum and Minimum.
997 // VMAX : Vector Maximum
998 defm VMAXs : N3VInt_QHS<0, 0, 0b0110, 0, "vmax.s", int_arm_neon_vmaxs, 1>;
999 defm VMAXu : N3VInt_QHS<1, 0, 0b0110, 0, "vmax.u", int_arm_neon_vmaxu, 1>;
1000 def VMAXfd : N3VDInt<0, 0, 0b00, 0b1111, 0, "vmax.f32", v2f32, v2f32,
1001 int_arm_neon_vmaxf, 1>;
1002 def VMAXfq : N3VQInt<0, 0, 0b00, 0b1111, 0, "vmax.f32", v4f32, v4f32,
1003 int_arm_neon_vmaxf, 1>;
1005 // VMIN : Vector Minimum
1006 defm VMINs : N3VInt_QHS<0, 0, 0b0110, 1, "vmin.s", int_arm_neon_vmins, 1>;
1007 defm VMINu : N3VInt_QHS<1, 0, 0b0110, 1, "vmin.u", int_arm_neon_vminu, 1>;
1008 def VMINfd : N3VDInt<0, 0, 0b10, 0b1111, 0, "vmin.f32", v2f32, v2f32,
1009 int_arm_neon_vminf, 1>;
1010 def VMINfq : N3VQInt<0, 0, 0b10, 0b1111, 0, "vmin.f32", v4f32, v4f32,
1011 int_arm_neon_vminf, 1>;
1013 // Vector Pairwise Operations.
1015 // VPADD : Vector Pairwise Add
1016 def VPADDi8 : N3VDInt<0, 0, 0b00, 0b1011, 1, "vpadd.i8", v8i8, v8i8,
1017 int_arm_neon_vpaddi, 0>;
1018 def VPADDi16 : N3VDInt<0, 0, 0b01, 0b1011, 1, "vpadd.i16", v4i16, v4i16,
1019 int_arm_neon_vpaddi, 0>;
1020 def VPADDi32 : N3VDInt<0, 0, 0b10, 0b1011, 1, "vpadd.i32", v2i32, v2i32,
1021 int_arm_neon_vpaddi, 0>;
1022 def VPADDf : N3VDInt<1, 0, 0b00, 0b1101, 0, "vpadd.f32", v2f32, v2f32,
1023 int_arm_neon_vpaddf, 0>;
1025 // VPADDL : Vector Pairwise Add Long
1026 defm VPADDLs : N2VPLInt_QHS<0b11, 0b11, 0b00, 0b00100, 0, "vpaddl.s",
1027 int_arm_neon_vpaddls>;
1028 defm VPADDLu : N2VPLInt_QHS<0b11, 0b11, 0b00, 0b00101, 0, "vpaddl.u",
1029 int_arm_neon_vpaddlu>;
1031 // VPADAL : Vector Pairwise Add and Accumulate Long
1032 defm VPADALs : N2VPLInt2_QHS<0b11, 0b11, 0b00, 0b00100, 0, "vpadal.s",
1033 int_arm_neon_vpadals>;
1034 defm VPADALu : N2VPLInt2_QHS<0b11, 0b11, 0b00, 0b00101, 0, "vpadal.u",
1035 int_arm_neon_vpadalu>;
1037 // VPMAX : Vector Pairwise Maximum
1038 def VPMAXs8 : N3VDInt<0, 0, 0b00, 0b1010, 0, "vpmax.s8", v8i8, v8i8,
1039 int_arm_neon_vpmaxs, 0>;
1040 def VPMAXs16 : N3VDInt<0, 0, 0b01, 0b1010, 0, "vpmax.s16", v4i16, v4i16,
1041 int_arm_neon_vpmaxs, 0>;
1042 def VPMAXs32 : N3VDInt<0, 0, 0b10, 0b1010, 0, "vpmax.s32", v2i32, v2i32,
1043 int_arm_neon_vpmaxs, 0>;
1044 def VPMAXu8 : N3VDInt<1, 0, 0b00, 0b1010, 0, "vpmax.u8", v8i8, v8i8,
1045 int_arm_neon_vpmaxu, 0>;
1046 def VPMAXu16 : N3VDInt<1, 0, 0b01, 0b1010, 0, "vpmax.u16", v4i16, v4i16,
1047 int_arm_neon_vpmaxu, 0>;
1048 def VPMAXu32 : N3VDInt<1, 0, 0b10, 0b1010, 0, "vpmax.u32", v2i32, v2i32,
1049 int_arm_neon_vpmaxu, 0>;
1050 def VPMAXf : N3VDInt<1, 0, 0b00, 0b1111, 0, "vpmax.f32", v2f32, v2f32,
1051 int_arm_neon_vpmaxf, 0>;
1053 // VPMIN : Vector Pairwise Minimum
1054 def VPMINs8 : N3VDInt<0, 0, 0b00, 0b1010, 1, "vpmin.s8", v8i8, v8i8,
1055 int_arm_neon_vpmins, 0>;
1056 def VPMINs16 : N3VDInt<0, 0, 0b01, 0b1010, 1, "vpmin.s16", v4i16, v4i16,
1057 int_arm_neon_vpmins, 0>;
1058 def VPMINs32 : N3VDInt<0, 0, 0b10, 0b1010, 1, "vpmin.s32", v2i32, v2i32,
1059 int_arm_neon_vpmins, 0>;
1060 def VPMINu8 : N3VDInt<1, 0, 0b00, 0b1010, 1, "vpmin.u8", v8i8, v8i8,
1061 int_arm_neon_vpminu, 0>;
1062 def VPMINu16 : N3VDInt<1, 0, 0b01, 0b1010, 1, "vpmin.u16", v4i16, v4i16,
1063 int_arm_neon_vpminu, 0>;
1064 def VPMINu32 : N3VDInt<1, 0, 0b10, 0b1010, 1, "vpmin.u32", v2i32, v2i32,
1065 int_arm_neon_vpminu, 0>;
1066 def VPMINf : N3VDInt<1, 0, 0b10, 0b1111, 0, "vpmin.f32", v2f32, v2f32,
1067 int_arm_neon_vpminf, 0>;
1069 // Vector Reciprocal and Reciprocal Square Root Estimate and Step.
1071 // VRECPE : Vector Reciprocal Estimate
1072 def VRECPEd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01000, 0, "vrecpe.u32",
1073 v2i32, v2i32, int_arm_neon_vrecpe>;
1074 def VRECPEq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01000, 0, "vrecpe.u32",
1075 v4i32, v4i32, int_arm_neon_vrecpe>;
1076 def VRECPEfd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01010, 0, "vrecpe.f32",
1077 v2f32, v2f32, int_arm_neon_vrecpef>;
1078 def VRECPEfq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01010, 0, "vrecpe.f32",
1079 v4f32, v4f32, int_arm_neon_vrecpef>;
1081 // VRECPS : Vector Reciprocal Step
1082 def VRECPSfd : N3VDInt<0, 0, 0b00, 0b1111, 1, "vrecps.f32", v2f32, v2f32,
1083 int_arm_neon_vrecps, 1>;
1084 def VRECPSfq : N3VQInt<0, 0, 0b00, 0b1111, 1, "vrecps.f32", v4f32, v4f32,
1085 int_arm_neon_vrecps, 1>;
1087 // VRSQRTE : Vector Reciprocal Square Root Estimate
1088 def VRSQRTEd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01001, 0, "vrsqrte.u32",
1089 v2i32, v2i32, int_arm_neon_vrsqrte>;
1090 def VRSQRTEq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01001, 0, "vrsqrte.u32",
1091 v4i32, v4i32, int_arm_neon_vrsqrte>;
1092 def VRSQRTEfd : N2VDInt<0b11, 0b11, 0b10, 0b11, 0b01011, 0, "vrsqrte.f32",
1093 v2f32, v2f32, int_arm_neon_vrsqrtef>;
1094 def VRSQRTEfq : N2VQInt<0b11, 0b11, 0b10, 0b11, 0b01011, 0, "vrsqrte.f32",
1095 v4f32, v4f32, int_arm_neon_vrsqrtef>;
1097 // VRSQRTS : Vector Reciprocal Square Root Step
1098 def VRSQRTSfd : N3VDInt<0, 0, 0b10, 0b1111, 1, "vrsqrts.f32", v2f32, v2f32,
1099 int_arm_neon_vrsqrts, 1>;
1100 def VRSQRTSfq : N3VQInt<0, 0, 0b10, 0b1111, 1, "vrsqrts.f32", v4f32, v4f32,
1101 int_arm_neon_vrsqrts, 1>;
1105 // VSHL : Vector Shift
1106 defm VSHLs : N3VInt_QHSD<0, 0, 0b0100, 0, "vshl.s", int_arm_neon_vshifts, 0>;
1107 defm VSHLu : N3VInt_QHSD<1, 0, 0b0100, 0, "vshl.u", int_arm_neon_vshiftu, 0>;
1108 // VSHL : Vector Shift Left (Immediate)
1109 defm VSHLi : N2VSh_QHSD<0, 1, 0b0111, 1, "vshl.i", NEONvshl>;
1110 // VSHR : Vector Shift Right (Immediate)
1111 defm VSHRs : N2VSh_QHSD<0, 1, 0b0000, 1, "vshr.s", NEONvshrs>;
1112 defm VSHRu : N2VSh_QHSD<1, 1, 0b0000, 1, "vshr.u", NEONvshru>;
1114 // VSHLL : Vector Shift Left Long
1115 def VSHLLs8 : N2VLSh<0, 1, 0b001000, 0b1010, 0, 0, 1, "vshll.s8",
1116 v8i16, v8i8, NEONvshlls>;
1117 def VSHLLs16 : N2VLSh<0, 1, 0b010000, 0b1010, 0, 0, 1, "vshll.s16",
1118 v4i32, v4i16, NEONvshlls>;
1119 def VSHLLs32 : N2VLSh<0, 1, 0b100000, 0b1010, 0, 0, 1, "vshll.s32",
1120 v2i64, v2i32, NEONvshlls>;
1121 def VSHLLu8 : N2VLSh<1, 1, 0b001000, 0b1010, 0, 0, 1, "vshll.u8",
1122 v8i16, v8i8, NEONvshllu>;
1123 def VSHLLu16 : N2VLSh<1, 1, 0b010000, 0b1010, 0, 0, 1, "vshll.u16",
1124 v4i32, v4i16, NEONvshllu>;
1125 def VSHLLu32 : N2VLSh<1, 1, 0b100000, 0b1010, 0, 0, 1, "vshll.u32",
1126 v2i64, v2i32, NEONvshllu>;
1128 // VSHLL : Vector Shift Left Long (with maximum shift count)
1129 def VSHLLi8 : N2VLSh<1, 1, 0b110010, 0b0011, 0, 0, 0, "vshll.i8",
1130 v8i16, v8i8, NEONvshlli>;
1131 def VSHLLi16 : N2VLSh<1, 1, 0b110110, 0b0011, 0, 0, 0, "vshll.i16",
1132 v4i32, v4i16, NEONvshlli>;
1133 def VSHLLi32 : N2VLSh<1, 1, 0b111010, 0b0011, 0, 0, 0, "vshll.i32",
1134 v2i64, v2i32, NEONvshlli>;
1136 // VSHRN : Vector Shift Right and Narrow
1137 def VSHRN16 : N2VNSh<0, 1, 0b001000, 0b1000, 0, 0, 1, "vshrn.i16",
1138 v8i8, v8i16, NEONvshrn>;
1139 def VSHRN32 : N2VNSh<0, 1, 0b010000, 0b1000, 0, 0, 1, "vshrn.i32",
1140 v4i16, v4i32, NEONvshrn>;
1141 def VSHRN64 : N2VNSh<0, 1, 0b100000, 0b1000, 0, 0, 1, "vshrn.i64",
1142 v2i32, v2i64, NEONvshrn>;
1144 // VRSHL : Vector Rounding Shift
1145 defm VRSHLs : N3VInt_QHSD<0,0,0b0101,0, "vrshl.s", int_arm_neon_vrshifts, 0>;
1146 defm VRSHLu : N3VInt_QHSD<1,0,0b0101,0, "vrshl.u", int_arm_neon_vrshiftu, 0>;
1147 // VRSHR : Vector Rounding Shift Right
1148 defm VRSHRs : N2VSh_QHSD<0, 1, 0b0010, 1, "vrshr.s", NEONvrshrs>;
1149 defm VRSHRu : N2VSh_QHSD<1, 1, 0b0010, 1, "vrshr.u", NEONvrshru>;
1151 // VRSHRN : Vector Rounding Shift Right and Narrow
1152 def VRSHRN16 : N2VNSh<0, 1, 0b001000, 0b1000, 0, 1, 1, "vrshrn.i16",
1153 v8i8, v8i16, NEONvrshrn>;
1154 def VRSHRN32 : N2VNSh<0, 1, 0b010000, 0b1000, 0, 1, 1, "vrshrn.i32",
1155 v4i16, v4i32, NEONvrshrn>;
1156 def VRSHRN64 : N2VNSh<0, 1, 0b100000, 0b1000, 0, 1, 1, "vrshrn.i64",
1157 v2i32, v2i64, NEONvrshrn>;
1159 // VQSHL : Vector Saturating Shift
1160 defm VQSHLs : N3VInt_QHSD<0,0,0b0100,1, "vqshl.s", int_arm_neon_vqshifts, 0>;
1161 defm VQSHLu : N3VInt_QHSD<1,0,0b0100,1, "vqshl.u", int_arm_neon_vqshiftu, 0>;
1162 // VQSHL : Vector Saturating Shift Left (Immediate)
1163 defm VQSHLsi : N2VSh_QHSD<0, 1, 0b0111, 1, "vqshl.s", NEONvqshls>;
1164 defm VQSHLui : N2VSh_QHSD<1, 1, 0b0111, 1, "vqshl.u", NEONvqshlu>;
1165 // VQSHLU : Vector Saturating Shift Left (Immediate, Unsigned)
1166 defm VQSHLsu : N2VSh_QHSD<1, 1, 0b0110, 1, "vqshlu.s", NEONvqshlsu>;
1168 // VQSHRN : Vector Saturating Shift Right and Narrow
1169 def VQSHRNs16 : N2VNSh<0, 1, 0b001000, 0b1001, 0, 0, 1, "vqshrn.s16",
1170 v8i8, v8i16, NEONvqshrns>;
1171 def VQSHRNs32 : N2VNSh<0, 1, 0b010000, 0b1001, 0, 0, 1, "vqshrn.s32",
1172 v4i16, v4i32, NEONvqshrns>;
1173 def VQSHRNs64 : N2VNSh<0, 1, 0b100000, 0b1001, 0, 0, 1, "vqshrn.s64",
1174 v2i32, v2i64, NEONvqshrns>;
1175 def VQSHRNu16 : N2VNSh<1, 1, 0b001000, 0b1001, 0, 0, 1, "vqshrn.u16",
1176 v8i8, v8i16, NEONvqshrnu>;
1177 def VQSHRNu32 : N2VNSh<1, 1, 0b010000, 0b1001, 0, 0, 1, "vqshrn.u32",
1178 v4i16, v4i32, NEONvqshrnu>;
1179 def VQSHRNu64 : N2VNSh<1, 1, 0b100000, 0b1001, 0, 0, 1, "vqshrn.u64",
1180 v2i32, v2i64, NEONvqshrnu>;
1182 // VQSHRUN : Vector Saturating Shift Right and Narrow (Unsigned)
1183 def VQSHRUN16 : N2VNSh<1, 1, 0b001000, 0b1000, 0, 0, 1, "vqshrun.s16",
1184 v8i8, v8i16, NEONvqshrnsu>;
1185 def VQSHRUN32 : N2VNSh<1, 1, 0b010000, 0b1000, 0, 0, 1, "vqshrun.s32",
1186 v4i16, v4i32, NEONvqshrnsu>;
1187 def VQSHRUN64 : N2VNSh<1, 1, 0b100000, 0b1000, 0, 0, 1, "vqshrun.s64",
1188 v2i32, v2i64, NEONvqshrnsu>;
1190 // VQRSHL : Vector Saturating Rounding Shift
1191 defm VQRSHLs : N3VInt_QHSD<0, 0, 0b0101, 1, "vqrshl.s",
1192 int_arm_neon_vqrshifts, 0>;
1193 defm VQRSHLu : N3VInt_QHSD<1, 0, 0b0101, 1, "vqrshl.u",
1194 int_arm_neon_vqrshiftu, 0>;
1196 // VQRSHRN : Vector Saturating Rounding Shift Right and Narrow
1197 def VQRSHRNs16: N2VNSh<0, 1, 0b001000, 0b1001, 0, 1, 1, "vqrshrn.s16",
1198 v8i8, v8i16, NEONvqrshrns>;
1199 def VQRSHRNs32: N2VNSh<0, 1, 0b010000, 0b1001, 0, 1, 1, "vqrshrn.s32",
1200 v4i16, v4i32, NEONvqrshrns>;
1201 def VQRSHRNs64: N2VNSh<0, 1, 0b100000, 0b1001, 0, 1, 1, "vqrshrn.s64",
1202 v2i32, v2i64, NEONvqrshrns>;
1203 def VQRSHRNu16: N2VNSh<1, 1, 0b001000, 0b1001, 0, 1, 1, "vqrshrn.u16",
1204 v8i8, v8i16, NEONvqrshrnu>;
1205 def VQRSHRNu32: N2VNSh<1, 1, 0b010000, 0b1001, 0, 1, 1, "vqrshrn.u32",
1206 v4i16, v4i32, NEONvqrshrnu>;
1207 def VQRSHRNu64: N2VNSh<1, 1, 0b100000, 0b1001, 0, 1, 1, "vqrshrn.u64",
1208 v2i32, v2i64, NEONvqrshrnu>;
1210 // VQRSHRUN : Vector Saturating Rounding Shift Right and Narrow (Unsigned)
1211 def VQRSHRUN16: N2VNSh<1, 1, 0b001000, 0b1000, 0, 1, 1, "vqrshrun.s16",
1212 v8i8, v8i16, NEONvqrshrnsu>;
1213 def VQRSHRUN32: N2VNSh<1, 1, 0b010000, 0b1000, 0, 1, 1, "vqrshrun.s32",
1214 v4i16, v4i32, NEONvqrshrnsu>;
1215 def VQRSHRUN64: N2VNSh<1, 1, 0b100000, 0b1000, 0, 1, 1, "vqrshrun.s64",
1216 v2i32, v2i64, NEONvqrshrnsu>;
1218 // VSRA : Vector Shift Right and Accumulate
1219 defm VSRAs : N2VShAdd_QHSD<0, 1, 0b0001, 1, "vsra.s", NEONvshrs>;
1220 defm VSRAu : N2VShAdd_QHSD<1, 1, 0b0001, 1, "vsra.u", NEONvshru>;
1221 // VRSRA : Vector Rounding Shift Right and Accumulate
1222 defm VRSRAs : N2VShAdd_QHSD<0, 1, 0b0011, 1, "vrsra.s", NEONvrshrs>;
1223 defm VRSRAu : N2VShAdd_QHSD<1, 1, 0b0011, 1, "vrsra.u", NEONvrshru>;
1225 // VSLI : Vector Shift Left and Insert
1226 defm VSLI : N2VShIns_QHSD<1, 1, 0b0101, 1, "vsli.", NEONvsli>;
1227 // VSRI : Vector Shift Right and Insert
1228 defm VSRI : N2VShIns_QHSD<1, 1, 0b0100, 1, "vsri.", NEONvsri>;
1230 // Vector Absolute and Saturating Absolute.
1232 // VABS : Vector Absolute Value
1233 defm VABS : N2VInt_QHS<0b11, 0b11, 0b01, 0b00110, 0, "vabs.s",
1235 def VABSfd : N2VDInt<0b11, 0b11, 0b10, 0b01, 0b01110, 0, "vabs.f32",
1236 v2f32, v2f32, int_arm_neon_vabsf>;
1237 def VABSfq : N2VQInt<0b11, 0b11, 0b10, 0b01, 0b01110, 0, "vabs.f32",
1238 v4f32, v4f32, int_arm_neon_vabsf>;
1240 // VQABS : Vector Saturating Absolute Value
1241 defm VQABS : N2VInt_QHS<0b11, 0b11, 0b00, 0b01110, 0, "vqabs.s",
1242 int_arm_neon_vqabs>;
1246 def vneg : PatFrag<(ops node:$in), (sub immAllZerosV, node:$in)>;
1247 def vneg_conv : PatFrag<(ops node:$in), (sub immAllZerosV_bc, node:$in)>;
1249 class VNEGD<bits<2> size, string OpcodeStr, ValueType Ty>
1250 : N2V<0b11, 0b11, size, 0b01, 0b00111, 0, 0, (outs DPR:$dst), (ins DPR:$src),
1251 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1252 [(set DPR:$dst, (Ty (vneg DPR:$src)))]>;
1253 class VNEGQ<bits<2> size, string OpcodeStr, ValueType Ty>
1254 : N2V<0b11, 0b11, size, 0b01, 0b00111, 1, 0, (outs QPR:$dst), (ins QPR:$src),
1255 !strconcat(OpcodeStr, "\t$dst, $src"), "",
1256 [(set QPR:$dst, (Ty (vneg QPR:$src)))]>;
1258 // VNEG : Vector Negate
1259 def VNEGs8d : VNEGD<0b00, "vneg.s8", v8i8>;
1260 def VNEGs16d : VNEGD<0b01, "vneg.s16", v4i16>;
1261 def VNEGs32d : VNEGD<0b10, "vneg.s32", v2i32>;
1262 def VNEGs8q : VNEGQ<0b00, "vneg.s8", v16i8>;
1263 def VNEGs16q : VNEGQ<0b01, "vneg.s16", v8i16>;
1264 def VNEGs32q : VNEGQ<0b10, "vneg.s32", v4i32>;
1266 // VNEG : Vector Negate (floating-point)
1267 def VNEGf32d : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 0, 0,
1268 (outs DPR:$dst), (ins DPR:$src), "vneg.f32\t$dst, $src", "",
1269 [(set DPR:$dst, (v2f32 (fneg DPR:$src)))]>;
1270 def VNEGf32q : N2V<0b11, 0b11, 0b10, 0b01, 0b01111, 1, 0,
1271 (outs QPR:$dst), (ins QPR:$src), "vneg.f32\t$dst, $src", "",
1272 [(set QPR:$dst, (v4f32 (fneg QPR:$src)))]>;
1274 def : Pat<(v8i8 (vneg_conv DPR:$src)), (VNEGs8d DPR:$src)>;
1275 def : Pat<(v4i16 (vneg_conv DPR:$src)), (VNEGs16d DPR:$src)>;
1276 def : Pat<(v2i32 (vneg_conv DPR:$src)), (VNEGs32d DPR:$src)>;
1277 def : Pat<(v16i8 (vneg_conv QPR:$src)), (VNEGs8q QPR:$src)>;
1278 def : Pat<(v8i16 (vneg_conv QPR:$src)), (VNEGs16q QPR:$src)>;
1279 def : Pat<(v4i32 (vneg_conv QPR:$src)), (VNEGs32q QPR:$src)>;
1281 // VQNEG : Vector Saturating Negate
1282 defm VQNEG : N2VInt_QHS<0b11, 0b11, 0b00, 0b01111, 0, "vqneg.s",
1283 int_arm_neon_vqneg>;
1285 // Vector Bit Counting Operations.
1287 // VCLS : Vector Count Leading Sign Bits
1288 defm VCLS : N2VInt_QHS<0b11, 0b11, 0b00, 0b01000, 0, "vcls.s",
1290 // VCLZ : Vector Count Leading Zeros
1291 defm VCLZ : N2VInt_QHS<0b11, 0b11, 0b00, 0b01001, 0, "vclz.i",
1293 // VCNT : Vector Count One Bits
1294 def VCNTd : N2VDInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0, "vcnt.8",
1295 v8i8, v8i8, int_arm_neon_vcnt>;
1296 def VCNTq : N2VQInt<0b11, 0b11, 0b00, 0b00, 0b01010, 0, "vcnt.8",
1297 v16i8, v16i8, int_arm_neon_vcnt>;
1299 // Vector Move Operations.
1301 // VMOV : Vector Move (Register)
1303 def VMOVD : N3V<0, 0, 0b10, 0b0001, 0, 1, (outs DPR:$dst), (ins DPR:$src),
1304 "vmov\t$dst, $src", "", []>;
1305 def VMOVQ : N3V<0, 0, 0b10, 0b0001, 1, 1, (outs QPR:$dst), (ins QPR:$src),
1306 "vmov\t$dst, $src", "", []>;
1308 // VMOV : Vector Move (Immediate)
1310 // VMOV_get_imm8 xform function: convert build_vector to VMOV.i8 imm.
1311 def VMOV_get_imm8 : SDNodeXForm<build_vector, [{
1312 return ARM::getVMOVImm(N, 1, *CurDAG);
1314 def vmovImm8 : PatLeaf<(build_vector), [{
1315 return ARM::getVMOVImm(N, 1, *CurDAG).getNode() != 0;
1318 // VMOV_get_imm16 xform function: convert build_vector to VMOV.i16 imm.
1319 def VMOV_get_imm16 : SDNodeXForm<build_vector, [{
1320 return ARM::getVMOVImm(N, 2, *CurDAG);
1322 def vmovImm16 : PatLeaf<(build_vector), [{
1323 return ARM::getVMOVImm(N, 2, *CurDAG).getNode() != 0;
1324 }], VMOV_get_imm16>;
1326 // VMOV_get_imm32 xform function: convert build_vector to VMOV.i32 imm.
1327 def VMOV_get_imm32 : SDNodeXForm<build_vector, [{
1328 return ARM::getVMOVImm(N, 4, *CurDAG);
1330 def vmovImm32 : PatLeaf<(build_vector), [{
1331 return ARM::getVMOVImm(N, 4, *CurDAG).getNode() != 0;
1332 }], VMOV_get_imm32>;
1334 // VMOV_get_imm64 xform function: convert build_vector to VMOV.i64 imm.
1335 def VMOV_get_imm64 : SDNodeXForm<build_vector, [{
1336 return ARM::getVMOVImm(N, 8, *CurDAG);
1338 def vmovImm64 : PatLeaf<(build_vector), [{
1339 return ARM::getVMOVImm(N, 8, *CurDAG).getNode() != 0;
1340 }], VMOV_get_imm64>;
1342 // Note: Some of the cmode bits in the following VMOV instructions need to
1343 // be encoded based on the immed values.
1345 def VMOVv8i8 : N1ModImm<1, 0b000, 0b1110, 0, 0, 0, 1, (outs DPR:$dst),
1346 (ins i8imm:$SIMM), "vmov.i8\t$dst, $SIMM", "",
1347 [(set DPR:$dst, (v8i8 vmovImm8:$SIMM))]>;
1348 def VMOVv16i8 : N1ModImm<1, 0b000, 0b1110, 0, 1, 0, 1, (outs QPR:$dst),
1349 (ins i8imm:$SIMM), "vmov.i8\t$dst, $SIMM", "",
1350 [(set QPR:$dst, (v16i8 vmovImm8:$SIMM))]>;
1352 def VMOVv4i16 : N1ModImm<1, 0b000, 0b1000, 0, 0, 0, 1, (outs DPR:$dst),
1353 (ins i16imm:$SIMM), "vmov.i16\t$dst, $SIMM", "",
1354 [(set DPR:$dst, (v4i16 vmovImm16:$SIMM))]>;
1355 def VMOVv8i16 : N1ModImm<1, 0b000, 0b1000, 0, 1, 0, 1, (outs QPR:$dst),
1356 (ins i16imm:$SIMM), "vmov.i16\t$dst, $SIMM", "",
1357 [(set QPR:$dst, (v8i16 vmovImm16:$SIMM))]>;
1359 def VMOVv2i32 : N1ModImm<1, 0b000, 0b0000, 0, 0, 0, 1, (outs DPR:$dst),
1360 (ins i32imm:$SIMM), "vmov.i32\t$dst, $SIMM", "",
1361 [(set DPR:$dst, (v2i32 vmovImm32:$SIMM))]>;
1362 def VMOVv4i32 : N1ModImm<1, 0b000, 0b0000, 0, 1, 0, 1, (outs QPR:$dst),
1363 (ins i32imm:$SIMM), "vmov.i32\t$dst, $SIMM", "",
1364 [(set QPR:$dst, (v4i32 vmovImm32:$SIMM))]>;
1366 def VMOVv1i64 : N1ModImm<1, 0b000, 0b1110, 0, 0, 1, 1, (outs DPR:$dst),
1367 (ins i64imm:$SIMM), "vmov.i64\t$dst, $SIMM", "",
1368 [(set DPR:$dst, (v1i64 vmovImm64:$SIMM))]>;
1369 def VMOVv2i64 : N1ModImm<1, 0b000, 0b1110, 0, 1, 1, 1, (outs QPR:$dst),
1370 (ins i64imm:$SIMM), "vmov.i64\t$dst, $SIMM", "",
1371 [(set QPR:$dst, (v2i64 vmovImm64:$SIMM))]>;
1373 // VMOV : Vector Get Lane (move scalar to ARM core register)
1375 def VGETLNs8 : NVGetLane<0b11100101, 0b1011, 0b00,
1376 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1377 "vmov", ".s8\t$dst, $src[$lane]",
1378 [(set GPR:$dst, (NEONvgetlanes (v8i8 DPR:$src),
1380 def VGETLNs16 : NVGetLane<0b11100001, 0b1011, 0b01,
1381 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1382 "vmov", ".s16\t$dst, $src[$lane]",
1383 [(set GPR:$dst, (NEONvgetlanes (v4i16 DPR:$src),
1385 def VGETLNu8 : NVGetLane<0b11101101, 0b1011, 0b00,
1386 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1387 "vmov", ".u8\t$dst, $src[$lane]",
1388 [(set GPR:$dst, (NEONvgetlaneu (v8i8 DPR:$src),
1390 def VGETLNu16 : NVGetLane<0b11101001, 0b1011, 0b01,
1391 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1392 "vmov", ".u16\t$dst, $src[$lane]",
1393 [(set GPR:$dst, (NEONvgetlaneu (v4i16 DPR:$src),
1395 def VGETLNi32 : NVGetLane<0b11100001, 0b1011, 0b00,
1396 (outs GPR:$dst), (ins DPR:$src, i32imm:$lane),
1397 "vmov", ".32\t$dst, $src[$lane]",
1398 [(set GPR:$dst, (extractelt (v2i32 DPR:$src),
1400 // def VGETLNf32: see FMRDH and FMRDL in ARMInstrVFP.td
1401 def : Pat<(NEONvgetlanes (v16i8 QPR:$src), imm:$lane),
1402 (VGETLNs8 (v8i8 (EXTRACT_SUBREG QPR:$src,
1403 (SubReg_i8_reg imm:$lane))),
1404 (SubReg_i8_lane imm:$lane))>;
1405 def : Pat<(NEONvgetlanes (v8i16 QPR:$src), imm:$lane),
1406 (VGETLNs16 (v4i16 (EXTRACT_SUBREG QPR:$src,
1407 (SubReg_i16_reg imm:$lane))),
1408 (SubReg_i16_lane imm:$lane))>;
1409 def : Pat<(NEONvgetlaneu (v16i8 QPR:$src), imm:$lane),
1410 (VGETLNu8 (v8i8 (EXTRACT_SUBREG QPR:$src,
1411 (SubReg_i8_reg imm:$lane))),
1412 (SubReg_i8_lane imm:$lane))>;
1413 def : Pat<(NEONvgetlaneu (v8i16 QPR:$src), imm:$lane),
1414 (VGETLNu16 (v4i16 (EXTRACT_SUBREG QPR:$src,
1415 (SubReg_i16_reg imm:$lane))),
1416 (SubReg_i16_lane imm:$lane))>;
1417 def : Pat<(extractelt (v4i32 QPR:$src), imm:$lane),
1418 (VGETLNi32 (v2i32 (EXTRACT_SUBREG QPR:$src,
1419 (SubReg_i32_reg imm:$lane))),
1420 (SubReg_i32_lane imm:$lane))>;
1421 //def : Pat<(extractelt (v2i64 QPR:$src1), imm:$src2),
1422 // (EXTRACT_SUBREG QPR:$src1, (SubReg_f64_reg imm:$src2))>;
1423 def : Pat<(extractelt (v2f64 QPR:$src1), imm:$src2),
1424 (EXTRACT_SUBREG QPR:$src1, (SubReg_f64_reg imm:$src2))>;
1427 // VMOV : Vector Set Lane (move ARM core register to scalar)
1429 let Constraints = "$src1 = $dst" in {
1430 def VSETLNi8 : NVSetLane<0b11100100, 0b1011, 0b00, (outs DPR:$dst),
1431 (ins DPR:$src1, GPR:$src2, i32imm:$lane),
1432 "vmov", ".8\t$dst[$lane], $src2",
1433 [(set DPR:$dst, (vector_insert (v8i8 DPR:$src1),
1434 GPR:$src2, imm:$lane))]>;
1435 def VSETLNi16 : NVSetLane<0b11100000, 0b1011, 0b01, (outs DPR:$dst),
1436 (ins DPR:$src1, GPR:$src2, i32imm:$lane),
1437 "vmov", ".16\t$dst[$lane], $src2",
1438 [(set DPR:$dst, (vector_insert (v4i16 DPR:$src1),
1439 GPR:$src2, imm:$lane))]>;
1440 def VSETLNi32 : NVSetLane<0b11100000, 0b1011, 0b00, (outs DPR:$dst),
1441 (ins DPR:$src1, GPR:$src2, i32imm:$lane),
1442 "vmov", ".32\t$dst[$lane], $src2",
1443 [(set DPR:$dst, (insertelt (v2i32 DPR:$src1),
1444 GPR:$src2, imm:$lane))]>;
1446 def : Pat<(vector_insert (v16i8 QPR:$src1), GPR:$src2, imm:$lane),
1447 (v16i8 (INSERT_SUBREG QPR:$src1,
1448 (VSETLNi8 (v8i8 (EXTRACT_SUBREG QPR:$src1,
1449 (SubReg_i8_reg imm:$lane))),
1450 GPR:$src2, (SubReg_i8_lane imm:$lane)),
1451 (SubReg_i8_reg imm:$lane)))>;
1452 def : Pat<(vector_insert (v8i16 QPR:$src1), GPR:$src2, imm:$lane),
1453 (v8i16 (INSERT_SUBREG QPR:$src1,
1454 (VSETLNi16 (v4i16 (EXTRACT_SUBREG QPR:$src1,
1455 (SubReg_i16_reg imm:$lane))),
1456 GPR:$src2, (SubReg_i16_lane imm:$lane)),
1457 (SubReg_i16_reg imm:$lane)))>;
1458 def : Pat<(insertelt (v4i32 QPR:$src1), GPR:$src2, imm:$lane),
1459 (v4i32 (INSERT_SUBREG QPR:$src1,
1460 (VSETLNi32 (v2i32 (EXTRACT_SUBREG QPR:$src1,
1461 (SubReg_i32_reg imm:$lane))),
1462 GPR:$src2, (SubReg_i32_lane imm:$lane)),
1463 (SubReg_i32_reg imm:$lane)))>;
1465 //def : Pat<(v2i64 (insertelt QPR:$src1, DPR:$src2, imm:$src3)),
1466 // (INSERT_SUBREG QPR:$src1, DPR:$src2, (SubReg_f64_reg imm:$src3))>;
1467 def : Pat<(v2f64 (insertelt QPR:$src1, DPR:$src2, imm:$src3)),
1468 (INSERT_SUBREG QPR:$src1, DPR:$src2, (SubReg_f64_reg imm:$src3))>;
1470 // VDUP : Vector Duplicate (from ARM core register to all elements)
1472 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
1473 (vector_shuffle node:$lhs, node:$rhs), [{
1474 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1475 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
1478 class VDUPD<bits<8> opcod1, bits<2> opcod3, string asmSize, ValueType Ty>
1479 : NVDup<opcod1, 0b1011, opcod3, (outs DPR:$dst), (ins GPR:$src),
1480 "vdup", !strconcat(asmSize, "\t$dst, $src"),
1481 [(set DPR:$dst, (Ty (splat_lo (scalar_to_vector GPR:$src), undef)))]>;
1482 class VDUPQ<bits<8> opcod1, bits<2> opcod3, string asmSize, ValueType Ty>
1483 : NVDup<opcod1, 0b1011, opcod3, (outs QPR:$dst), (ins GPR:$src),
1484 "vdup", !strconcat(asmSize, "\t$dst, $src"),
1485 [(set QPR:$dst, (Ty (splat_lo (scalar_to_vector GPR:$src), undef)))]>;
1487 def VDUP8d : VDUPD<0b11101100, 0b00, ".8", v8i8>;
1488 def VDUP16d : VDUPD<0b11101000, 0b01, ".16", v4i16>;
1489 def VDUP32d : VDUPD<0b11101000, 0b00, ".32", v2i32>;
1490 def VDUP8q : VDUPQ<0b11101110, 0b00, ".8", v16i8>;
1491 def VDUP16q : VDUPQ<0b11101010, 0b01, ".16", v8i16>;
1492 def VDUP32q : VDUPQ<0b11101010, 0b00, ".32", v4i32>;
1494 def VDUPfd : NVDup<0b11101000, 0b1011, 0b00, (outs DPR:$dst), (ins GPR:$src),
1495 "vdup", ".32\t$dst, $src",
1496 [(set DPR:$dst, (v2f32 (splat_lo
1498 (f32 (bitconvert GPR:$src))),
1500 def VDUPfq : NVDup<0b11101010, 0b1011, 0b00, (outs QPR:$dst), (ins GPR:$src),
1501 "vdup", ".32\t$dst, $src",
1502 [(set QPR:$dst, (v4f32 (splat_lo
1504 (f32 (bitconvert GPR:$src))),
1507 // VDUP : Vector Duplicate Lane (from scalar to all elements)
1509 def SHUFFLE_get_splat_lane : SDNodeXForm<vector_shuffle, [{
1510 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1511 return CurDAG->getTargetConstant(SVOp->getSplatIndex(), MVT::i32);
1514 def splat_lane : PatFrag<(ops node:$lhs, node:$rhs),
1515 (vector_shuffle node:$lhs, node:$rhs), [{
1516 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1517 return SVOp->isSplat();
1518 }], SHUFFLE_get_splat_lane>;
1520 class VDUPLND<bits<2> op19_18, bits<2> op17_16, string OpcodeStr, ValueType Ty>
1521 : N2V<0b11, 0b11, op19_18, op17_16, 0b11000, 0, 0,
1522 (outs DPR:$dst), (ins DPR:$src, i32imm:$lane),
1523 !strconcat(OpcodeStr, "\t$dst, $src[$lane]"), "",
1524 [(set DPR:$dst, (Ty (splat_lane:$lane DPR:$src, undef)))]>;
1526 // vector_shuffle requires that the source and destination types match, so
1527 // VDUP to a 128-bit result uses a target-specific VDUPLANEQ node.
1528 class VDUPLNQ<bits<2> op19_18, bits<2> op17_16, string OpcodeStr,
1529 ValueType ResTy, ValueType OpTy>
1530 : N2V<0b11, 0b11, op19_18, op17_16, 0b11000, 1, 0,
1531 (outs QPR:$dst), (ins DPR:$src, i32imm:$lane),
1532 !strconcat(OpcodeStr, "\t$dst, $src[$lane]"), "",
1533 [(set QPR:$dst, (ResTy (NEONvduplaneq (OpTy DPR:$src), imm:$lane)))]>;
1535 def VDUPLN8d : VDUPLND<0b00, 0b01, "vdup.8", v8i8>;
1536 def VDUPLN16d : VDUPLND<0b00, 0b10, "vdup.16", v4i16>;
1537 def VDUPLN32d : VDUPLND<0b01, 0b00, "vdup.32", v2i32>;
1538 def VDUPLNfd : VDUPLND<0b01, 0b00, "vdup.32", v2f32>;
1539 def VDUPLN8q : VDUPLNQ<0b00, 0b01, "vdup.8", v16i8, v8i8>;
1540 def VDUPLN16q : VDUPLNQ<0b00, 0b10, "vdup.16", v8i16, v4i16>;
1541 def VDUPLN32q : VDUPLNQ<0b01, 0b00, "vdup.32", v4i32, v2i32>;
1542 def VDUPLNfq : VDUPLNQ<0b01, 0b00, "vdup.32", v4f32, v2f32>;
1544 // VMOVN : Vector Narrowing Move
1545 defm VMOVN : N2VNInt_HSD<0b11,0b11,0b10,0b00100,0,0, "vmovn.i",
1546 int_arm_neon_vmovn>;
1547 // VQMOVN : Vector Saturating Narrowing Move
1548 defm VQMOVNs : N2VNInt_HSD<0b11,0b11,0b10,0b00101,0,0, "vqmovn.s",
1549 int_arm_neon_vqmovns>;
1550 defm VQMOVNu : N2VNInt_HSD<0b11,0b11,0b10,0b00101,1,0, "vqmovn.u",
1551 int_arm_neon_vqmovnu>;
1552 defm VQMOVNsu : N2VNInt_HSD<0b11,0b11,0b10,0b00100,1,0, "vqmovun.s",
1553 int_arm_neon_vqmovnsu>;
1554 // VMOVL : Vector Lengthening Move
1555 defm VMOVLs : N2VLInt_QHS<0,1,0b1010,0,0,1, "vmovl.s", int_arm_neon_vmovls>;
1556 defm VMOVLu : N2VLInt_QHS<1,1,0b1010,0,0,1, "vmovl.u", int_arm_neon_vmovlu>;
1558 // Vector Conversions.
1560 // VCVT : Vector Convert Between Floating-Point and Integers
1561 def VCVTf2sd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt.s32.f32",
1562 v2i32, v2f32, fp_to_sint>;
1563 def VCVTf2ud : N2VD<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt.u32.f32",
1564 v2i32, v2f32, fp_to_uint>;
1565 def VCVTs2fd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt.f32.s32",
1566 v2f32, v2i32, sint_to_fp>;
1567 def VCVTu2fd : N2VD<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt.f32.u32",
1568 v2f32, v2i32, uint_to_fp>;
1570 def VCVTf2sq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01110, 0, "vcvt.s32.f32",
1571 v4i32, v4f32, fp_to_sint>;
1572 def VCVTf2uq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01111, 0, "vcvt.u32.f32",
1573 v4i32, v4f32, fp_to_uint>;
1574 def VCVTs2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01100, 0, "vcvt.f32.s32",
1575 v4f32, v4i32, sint_to_fp>;
1576 def VCVTu2fq : N2VQ<0b11, 0b11, 0b10, 0b11, 0b01101, 0, "vcvt.f32.u32",
1577 v4f32, v4i32, uint_to_fp>;
1579 // VCVT : Vector Convert Between Floating-Point and Fixed-Point.
1580 // Note: Some of the opcode bits in the following VCVT instructions need to
1581 // be encoded based on the immed values.
1582 def VCVTf2xsd : N2VCvtD<0, 1, 0b000000, 0b1111, 0, 1, "vcvt.s32.f32",
1583 v2i32, v2f32, int_arm_neon_vcvtfp2fxs>;
1584 def VCVTf2xud : N2VCvtD<1, 1, 0b000000, 0b1111, 0, 1, "vcvt.u32.f32",
1585 v2i32, v2f32, int_arm_neon_vcvtfp2fxu>;
1586 def VCVTxs2fd : N2VCvtD<0, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.s32",
1587 v2f32, v2i32, int_arm_neon_vcvtfxs2fp>;
1588 def VCVTxu2fd : N2VCvtD<1, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.u32",
1589 v2f32, v2i32, int_arm_neon_vcvtfxu2fp>;
1591 def VCVTf2xsq : N2VCvtQ<0, 1, 0b000000, 0b1111, 0, 1, "vcvt.s32.f32",
1592 v4i32, v4f32, int_arm_neon_vcvtfp2fxs>;
1593 def VCVTf2xuq : N2VCvtQ<1, 1, 0b000000, 0b1111, 0, 1, "vcvt.u32.f32",
1594 v4i32, v4f32, int_arm_neon_vcvtfp2fxu>;
1595 def VCVTxs2fq : N2VCvtQ<0, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.s32",
1596 v4f32, v4i32, int_arm_neon_vcvtfxs2fp>;
1597 def VCVTxu2fq : N2VCvtQ<1, 1, 0b000000, 0b1110, 0, 1, "vcvt.f32.u32",
1598 v4f32, v4i32, int_arm_neon_vcvtfxu2fp>;
1600 //===----------------------------------------------------------------------===//
1601 // Non-Instruction Patterns
1602 //===----------------------------------------------------------------------===//
1605 def : Pat<(v1i64 (bitconvert (v2i32 DPR:$src))), (v1i64 DPR:$src)>;
1606 def : Pat<(v1i64 (bitconvert (v4i16 DPR:$src))), (v1i64 DPR:$src)>;
1607 def : Pat<(v1i64 (bitconvert (v8i8 DPR:$src))), (v1i64 DPR:$src)>;
1608 def : Pat<(v1i64 (bitconvert (f64 DPR:$src))), (v1i64 DPR:$src)>;
1609 def : Pat<(v1i64 (bitconvert (v2f32 DPR:$src))), (v1i64 DPR:$src)>;
1610 def : Pat<(v2i32 (bitconvert (v1i64 DPR:$src))), (v2i32 DPR:$src)>;
1611 def : Pat<(v2i32 (bitconvert (v4i16 DPR:$src))), (v2i32 DPR:$src)>;
1612 def : Pat<(v2i32 (bitconvert (v8i8 DPR:$src))), (v2i32 DPR:$src)>;
1613 def : Pat<(v2i32 (bitconvert (f64 DPR:$src))), (v2i32 DPR:$src)>;
1614 def : Pat<(v2i32 (bitconvert (v2f32 DPR:$src))), (v2i32 DPR:$src)>;
1615 def : Pat<(v4i16 (bitconvert (v1i64 DPR:$src))), (v4i16 DPR:$src)>;
1616 def : Pat<(v4i16 (bitconvert (v2i32 DPR:$src))), (v4i16 DPR:$src)>;
1617 def : Pat<(v4i16 (bitconvert (v8i8 DPR:$src))), (v4i16 DPR:$src)>;
1618 def : Pat<(v4i16 (bitconvert (f64 DPR:$src))), (v4i16 DPR:$src)>;
1619 def : Pat<(v4i16 (bitconvert (v2f32 DPR:$src))), (v4i16 DPR:$src)>;
1620 def : Pat<(v8i8 (bitconvert (v1i64 DPR:$src))), (v8i8 DPR:$src)>;
1621 def : Pat<(v8i8 (bitconvert (v2i32 DPR:$src))), (v8i8 DPR:$src)>;
1622 def : Pat<(v8i8 (bitconvert (v4i16 DPR:$src))), (v8i8 DPR:$src)>;
1623 def : Pat<(v8i8 (bitconvert (f64 DPR:$src))), (v8i8 DPR:$src)>;
1624 def : Pat<(v8i8 (bitconvert (v2f32 DPR:$src))), (v8i8 DPR:$src)>;
1625 def : Pat<(f64 (bitconvert (v1i64 DPR:$src))), (f64 DPR:$src)>;
1626 def : Pat<(f64 (bitconvert (v2i32 DPR:$src))), (f64 DPR:$src)>;
1627 def : Pat<(f64 (bitconvert (v4i16 DPR:$src))), (f64 DPR:$src)>;
1628 def : Pat<(f64 (bitconvert (v8i8 DPR:$src))), (f64 DPR:$src)>;
1629 def : Pat<(f64 (bitconvert (v2f32 DPR:$src))), (f64 DPR:$src)>;
1630 def : Pat<(v2f32 (bitconvert (f64 DPR:$src))), (v2f32 DPR:$src)>;
1631 def : Pat<(v2f32 (bitconvert (v1i64 DPR:$src))), (v2f32 DPR:$src)>;
1632 def : Pat<(v2f32 (bitconvert (v2i32 DPR:$src))), (v2f32 DPR:$src)>;
1633 def : Pat<(v2f32 (bitconvert (v4i16 DPR:$src))), (v2f32 DPR:$src)>;
1634 def : Pat<(v2f32 (bitconvert (v8i8 DPR:$src))), (v2f32 DPR:$src)>;
1636 def : Pat<(v2i64 (bitconvert (v4i32 QPR:$src))), (v2i64 QPR:$src)>;
1637 def : Pat<(v2i64 (bitconvert (v8i16 QPR:$src))), (v2i64 QPR:$src)>;
1638 def : Pat<(v2i64 (bitconvert (v16i8 QPR:$src))), (v2i64 QPR:$src)>;
1639 def : Pat<(v2i64 (bitconvert (v2f64 QPR:$src))), (v2i64 QPR:$src)>;
1640 def : Pat<(v2i64 (bitconvert (v4f32 QPR:$src))), (v2i64 QPR:$src)>;
1641 def : Pat<(v4i32 (bitconvert (v2i64 QPR:$src))), (v4i32 QPR:$src)>;
1642 def : Pat<(v4i32 (bitconvert (v8i16 QPR:$src))), (v4i32 QPR:$src)>;
1643 def : Pat<(v4i32 (bitconvert (v16i8 QPR:$src))), (v4i32 QPR:$src)>;
1644 def : Pat<(v4i32 (bitconvert (v2f64 QPR:$src))), (v4i32 QPR:$src)>;
1645 def : Pat<(v4i32 (bitconvert (v4f32 QPR:$src))), (v4i32 QPR:$src)>;
1646 def : Pat<(v8i16 (bitconvert (v2i64 QPR:$src))), (v8i16 QPR:$src)>;
1647 def : Pat<(v8i16 (bitconvert (v4i32 QPR:$src))), (v8i16 QPR:$src)>;
1648 def : Pat<(v8i16 (bitconvert (v16i8 QPR:$src))), (v8i16 QPR:$src)>;
1649 def : Pat<(v8i16 (bitconvert (v2f64 QPR:$src))), (v8i16 QPR:$src)>;
1650 def : Pat<(v8i16 (bitconvert (v4f32 QPR:$src))), (v8i16 QPR:$src)>;
1651 def : Pat<(v16i8 (bitconvert (v2i64 QPR:$src))), (v16i8 QPR:$src)>;
1652 def : Pat<(v16i8 (bitconvert (v4i32 QPR:$src))), (v16i8 QPR:$src)>;
1653 def : Pat<(v16i8 (bitconvert (v8i16 QPR:$src))), (v16i8 QPR:$src)>;
1654 def : Pat<(v16i8 (bitconvert (v2f64 QPR:$src))), (v16i8 QPR:$src)>;
1655 def : Pat<(v16i8 (bitconvert (v4f32 QPR:$src))), (v16i8 QPR:$src)>;
1656 def : Pat<(v4f32 (bitconvert (v2i64 QPR:$src))), (v4f32 QPR:$src)>;
1657 def : Pat<(v4f32 (bitconvert (v4i32 QPR:$src))), (v4f32 QPR:$src)>;
1658 def : Pat<(v4f32 (bitconvert (v8i16 QPR:$src))), (v4f32 QPR:$src)>;
1659 def : Pat<(v4f32 (bitconvert (v16i8 QPR:$src))), (v4f32 QPR:$src)>;
1660 def : Pat<(v4f32 (bitconvert (v2f64 QPR:$src))), (v4f32 QPR:$src)>;
1661 def : Pat<(v2f64 (bitconvert (v2i64 QPR:$src))), (v2f64 QPR:$src)>;
1662 def : Pat<(v2f64 (bitconvert (v4i32 QPR:$src))), (v2f64 QPR:$src)>;
1663 def : Pat<(v2f64 (bitconvert (v8i16 QPR:$src))), (v2f64 QPR:$src)>;
1664 def : Pat<(v2f64 (bitconvert (v16i8 QPR:$src))), (v2f64 QPR:$src)>;
1665 def : Pat<(v2f64 (bitconvert (v4f32 QPR:$src))), (v2f64 QPR:$src)>;