1 //===-- AArch64InstrNEON.td - NEON support for AArch64 -----*- tablegen -*-===//
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 AArch64 NEON instruction set.
12 //===----------------------------------------------------------------------===//
14 //===----------------------------------------------------------------------===//
15 // NEON-specific DAG Nodes.
16 //===----------------------------------------------------------------------===//
18 // (outs Result), (ins Imm, OpCmode)
19 def SDT_Neon_movi : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVT<1, i32>]>;
21 def Neon_movi : SDNode<"AArch64ISD::NEON_MOVIMM", SDT_Neon_movi>;
23 def Neon_mvni : SDNode<"AArch64ISD::NEON_MVNIMM", SDT_Neon_movi>;
25 // (outs Result), (ins Imm)
26 def Neon_fmovi : SDNode<"AArch64ISD::NEON_FMOVIMM", SDTypeProfile<1, 1,
27 [SDTCisVec<0>, SDTCisVT<1, i32>]>>;
29 // (outs Result), (ins LHS, RHS, CondCode)
30 def Neon_cmp : SDNode<"AArch64ISD::NEON_CMP", SDTypeProfile<1, 3,
31 [SDTCisVec<0>, SDTCisSameAs<1, 2>]>>;
33 // (outs Result), (ins LHS, 0/0.0 constant, CondCode)
34 def Neon_cmpz : SDNode<"AArch64ISD::NEON_CMPZ", SDTypeProfile<1, 3,
35 [SDTCisVec<0>, SDTCisVec<1>]>>;
37 // (outs Result), (ins LHS, RHS)
38 def Neon_tst : SDNode<"AArch64ISD::NEON_TST", SDTypeProfile<1, 2,
39 [SDTCisVec<0>, SDTCisSameAs<1, 2>]>>;
41 def SDTARMVSH : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0, 1>,
43 def Neon_sqrshlImm : SDNode<"AArch64ISD::NEON_QSHLs", SDTARMVSH>;
44 def Neon_uqrshlImm : SDNode<"AArch64ISD::NEON_QSHLu", SDTARMVSH>;
46 def SDTPERMUTE : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0, 1>,
48 def Neon_uzp1 : SDNode<"AArch64ISD::NEON_UZP1", SDTPERMUTE>;
49 def Neon_uzp2 : SDNode<"AArch64ISD::NEON_UZP2", SDTPERMUTE>;
50 def Neon_zip1 : SDNode<"AArch64ISD::NEON_ZIP1", SDTPERMUTE>;
51 def Neon_zip2 : SDNode<"AArch64ISD::NEON_ZIP2", SDTPERMUTE>;
52 def Neon_trn1 : SDNode<"AArch64ISD::NEON_TRN1", SDTPERMUTE>;
53 def Neon_trn2 : SDNode<"AArch64ISD::NEON_TRN2", SDTPERMUTE>;
55 def SDTVSHUF : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisSameAs<0, 1>]>;
56 def Neon_rev64 : SDNode<"AArch64ISD::NEON_REV64", SDTVSHUF>;
57 def Neon_rev32 : SDNode<"AArch64ISD::NEON_REV32", SDTVSHUF>;
58 def Neon_rev16 : SDNode<"AArch64ISD::NEON_REV16", SDTVSHUF>;
59 def Neon_vdup : SDNode<"AArch64ISD::NEON_VDUP", SDTypeProfile<1, 1,
61 def Neon_vduplane : SDNode<"AArch64ISD::NEON_VDUPLANE", SDTypeProfile<1, 2,
62 [SDTCisVec<0>, SDTCisVec<1>, SDTCisVT<2, i64>]>>;
63 def Neon_vextract : SDNode<"AArch64ISD::NEON_VEXTRACT", SDTypeProfile<1, 3,
64 [SDTCisVec<0>, SDTCisSameAs<0, 1>,
65 SDTCisSameAs<0, 2>, SDTCisVT<3, i64>]>>;
67 def SDT_assertext : SDTypeProfile<1, 1,
68 [SDTCisInt<0>, SDTCisInt<1>, SDTCisSameAs<1, 0>]>;
69 def assertsext : SDNode<"ISD::AssertSext", SDT_assertext>;
70 def assertzext : SDNode<"ISD::AssertZext", SDT_assertext>;
72 //===----------------------------------------------------------------------===//
74 //===----------------------------------------------------------------------===//
76 multiclass NeonI_3VSame_B_sizes<bit u, bits<2> size, bits<5> opcode,
77 string asmop, SDPatternOperator opnode8B,
78 SDPatternOperator opnode16B,
80 let isCommutable = Commutable in {
81 def _8B : NeonI_3VSame<0b0, u, size, opcode,
82 (outs VPR64:$Rd), (ins VPR64:$Rn, VPR64:$Rm),
83 asmop # "\t$Rd.8b, $Rn.8b, $Rm.8b",
84 [(set (v8i8 VPR64:$Rd),
85 (v8i8 (opnode8B (v8i8 VPR64:$Rn), (v8i8 VPR64:$Rm))))],
88 def _16B : NeonI_3VSame<0b1, u, size, opcode,
89 (outs VPR128:$Rd), (ins VPR128:$Rn, VPR128:$Rm),
90 asmop # "\t$Rd.16b, $Rn.16b, $Rm.16b",
91 [(set (v16i8 VPR128:$Rd),
92 (v16i8 (opnode16B (v16i8 VPR128:$Rn), (v16i8 VPR128:$Rm))))],
98 multiclass NeonI_3VSame_HS_sizes<bit u, bits<5> opcode,
99 string asmop, SDPatternOperator opnode,
100 bit Commutable = 0> {
101 let isCommutable = Commutable in {
102 def _4H : NeonI_3VSame<0b0, u, 0b01, opcode,
103 (outs VPR64:$Rd), (ins VPR64:$Rn, VPR64:$Rm),
104 asmop # "\t$Rd.4h, $Rn.4h, $Rm.4h",
105 [(set (v4i16 VPR64:$Rd),
106 (v4i16 (opnode (v4i16 VPR64:$Rn), (v4i16 VPR64:$Rm))))],
109 def _8H : NeonI_3VSame<0b1, u, 0b01, opcode,
110 (outs VPR128:$Rd), (ins VPR128:$Rn, VPR128:$Rm),
111 asmop # "\t$Rd.8h, $Rn.8h, $Rm.8h",
112 [(set (v8i16 VPR128:$Rd),
113 (v8i16 (opnode (v8i16 VPR128:$Rn), (v8i16 VPR128:$Rm))))],
116 def _2S : NeonI_3VSame<0b0, u, 0b10, opcode,
117 (outs VPR64:$Rd), (ins VPR64:$Rn, VPR64:$Rm),
118 asmop # "\t$Rd.2s, $Rn.2s, $Rm.2s",
119 [(set (v2i32 VPR64:$Rd),
120 (v2i32 (opnode (v2i32 VPR64:$Rn), (v2i32 VPR64:$Rm))))],
123 def _4S : NeonI_3VSame<0b1, u, 0b10, opcode,
124 (outs VPR128:$Rd), (ins VPR128:$Rn, VPR128:$Rm),
125 asmop # "\t$Rd.4s, $Rn.4s, $Rm.4s",
126 [(set (v4i32 VPR128:$Rd),
127 (v4i32 (opnode (v4i32 VPR128:$Rn), (v4i32 VPR128:$Rm))))],
131 multiclass NeonI_3VSame_BHS_sizes<bit u, bits<5> opcode,
132 string asmop, SDPatternOperator opnode,
134 : NeonI_3VSame_HS_sizes<u, opcode, asmop, opnode, Commutable> {
135 let isCommutable = Commutable in {
136 def _8B : NeonI_3VSame<0b0, u, 0b00, opcode,
137 (outs VPR64:$Rd), (ins VPR64:$Rn, VPR64:$Rm),
138 asmop # "\t$Rd.8b, $Rn.8b, $Rm.8b",
139 [(set (v8i8 VPR64:$Rd),
140 (v8i8 (opnode (v8i8 VPR64:$Rn), (v8i8 VPR64:$Rm))))],
143 def _16B : NeonI_3VSame<0b1, u, 0b00, opcode,
144 (outs VPR128:$Rd), (ins VPR128:$Rn, VPR128:$Rm),
145 asmop # "\t$Rd.16b, $Rn.16b, $Rm.16b",
146 [(set (v16i8 VPR128:$Rd),
147 (v16i8 (opnode (v16i8 VPR128:$Rn), (v16i8 VPR128:$Rm))))],
152 multiclass NeonI_3VSame_BHSD_sizes<bit u, bits<5> opcode,
153 string asmop, SDPatternOperator opnode,
155 : NeonI_3VSame_BHS_sizes<u, opcode, asmop, opnode, Commutable> {
156 let isCommutable = Commutable in {
157 def _2D : NeonI_3VSame<0b1, u, 0b11, opcode,
158 (outs VPR128:$Rd), (ins VPR128:$Rn, VPR128:$Rm),
159 asmop # "\t$Rd.2d, $Rn.2d, $Rm.2d",
160 [(set (v2i64 VPR128:$Rd),
161 (v2i64 (opnode (v2i64 VPR128:$Rn), (v2i64 VPR128:$Rm))))],
166 // Multiclass NeonI_3VSame_SD_sizes: Operand types are floating point types,
167 // but Result types can be integer or floating point types.
168 multiclass NeonI_3VSame_SD_sizes<bit u, bit size, bits<5> opcode,
169 string asmop, SDPatternOperator opnode2S,
170 SDPatternOperator opnode4S,
171 SDPatternOperator opnode2D,
172 ValueType ResTy2S, ValueType ResTy4S,
173 ValueType ResTy2D, bit Commutable = 0> {
174 let isCommutable = Commutable in {
175 def _2S : NeonI_3VSame<0b0, u, {size, 0b0}, opcode,
176 (outs VPR64:$Rd), (ins VPR64:$Rn, VPR64:$Rm),
177 asmop # "\t$Rd.2s, $Rn.2s, $Rm.2s",
178 [(set (ResTy2S VPR64:$Rd),
179 (ResTy2S (opnode2S (v2f32 VPR64:$Rn), (v2f32 VPR64:$Rm))))],
182 def _4S : NeonI_3VSame<0b1, u, {size, 0b0}, opcode,
183 (outs VPR128:$Rd), (ins VPR128:$Rn, VPR128:$Rm),
184 asmop # "\t$Rd.4s, $Rn.4s, $Rm.4s",
185 [(set (ResTy4S VPR128:$Rd),
186 (ResTy4S (opnode4S (v4f32 VPR128:$Rn), (v4f32 VPR128:$Rm))))],
189 def _2D : NeonI_3VSame<0b1, u, {size, 0b1}, opcode,
190 (outs VPR128:$Rd), (ins VPR128:$Rn, VPR128:$Rm),
191 asmop # "\t$Rd.2d, $Rn.2d, $Rm.2d",
192 [(set (ResTy2D VPR128:$Rd),
193 (ResTy2D (opnode2D (v2f64 VPR128:$Rn), (v2f64 VPR128:$Rm))))],
198 //===----------------------------------------------------------------------===//
199 // Instruction Definitions
200 //===----------------------------------------------------------------------===//
202 // Vector Arithmetic Instructions
204 // Vector Add (Integer and Floating-Point)
206 defm ADDvvv : NeonI_3VSame_BHSD_sizes<0b0, 0b10000, "add", add, 1>;
207 defm FADDvvv : NeonI_3VSame_SD_sizes<0b0, 0b0, 0b11010, "fadd", fadd, fadd, fadd,
208 v2f32, v4f32, v2f64, 1>;
210 // Vector Sub (Integer and Floating-Point)
212 defm SUBvvv : NeonI_3VSame_BHSD_sizes<0b1, 0b10000, "sub", sub, 0>;
213 defm FSUBvvv : NeonI_3VSame_SD_sizes<0b0, 0b1, 0b11010, "fsub", fsub, fsub, fsub,
214 v2f32, v4f32, v2f64, 0>;
216 // Vector Multiply (Integer and Floating-Point)
218 defm MULvvv : NeonI_3VSame_BHS_sizes<0b0, 0b10011, "mul", mul, 1>;
219 defm FMULvvv : NeonI_3VSame_SD_sizes<0b1, 0b0, 0b11011, "fmul", fmul, fmul, fmul,
220 v2f32, v4f32, v2f64, 1>;
222 // Vector Multiply (Polynomial)
224 defm PMULvvv : NeonI_3VSame_B_sizes<0b1, 0b00, 0b10011, "pmul",
225 int_arm_neon_vmulp, int_arm_neon_vmulp, 1>;
227 // Vector Multiply-accumulate and Multiply-subtract (Integer)
229 // class NeonI_3VSame_Constraint_impl: NeonI_3VSame with no data type and
230 // two operands constraints.
231 class NeonI_3VSame_Constraint_impl<string asmop, string asmlane,
232 RegisterOperand VPRC, ValueType OpTy, bit q, bit u, bits<2> size,
233 bits<5> opcode, SDPatternOperator opnode>
234 : NeonI_3VSame<q, u, size, opcode,
235 (outs VPRC:$Rd), (ins VPRC:$src, VPRC:$Rn, VPRC:$Rm),
236 asmop # "\t$Rd" # asmlane # ", $Rn" # asmlane # ", $Rm" # asmlane,
237 [(set (OpTy VPRC:$Rd),
238 (OpTy (opnode (OpTy VPRC:$src), (OpTy VPRC:$Rn), (OpTy VPRC:$Rm))))],
240 let Constraints = "$src = $Rd";
243 def Neon_mla : PatFrag<(ops node:$Ra, node:$Rn, node:$Rm),
244 (add node:$Ra, (mul node:$Rn, node:$Rm))>;
246 def Neon_mls : PatFrag<(ops node:$Ra, node:$Rn, node:$Rm),
247 (sub node:$Ra, (mul node:$Rn, node:$Rm))>;
250 def MLAvvv_8B: NeonI_3VSame_Constraint_impl<"mla", ".8b", VPR64, v8i8,
251 0b0, 0b0, 0b00, 0b10010, Neon_mla>;
252 def MLAvvv_16B: NeonI_3VSame_Constraint_impl<"mla", ".16b", VPR128, v16i8,
253 0b1, 0b0, 0b00, 0b10010, Neon_mla>;
254 def MLAvvv_4H: NeonI_3VSame_Constraint_impl<"mla", ".4h", VPR64, v4i16,
255 0b0, 0b0, 0b01, 0b10010, Neon_mla>;
256 def MLAvvv_8H: NeonI_3VSame_Constraint_impl<"mla", ".8h", VPR128, v8i16,
257 0b1, 0b0, 0b01, 0b10010, Neon_mla>;
258 def MLAvvv_2S: NeonI_3VSame_Constraint_impl<"mla", ".2s", VPR64, v2i32,
259 0b0, 0b0, 0b10, 0b10010, Neon_mla>;
260 def MLAvvv_4S: NeonI_3VSame_Constraint_impl<"mla", ".4s", VPR128, v4i32,
261 0b1, 0b0, 0b10, 0b10010, Neon_mla>;
263 def MLSvvv_8B: NeonI_3VSame_Constraint_impl<"mls", ".8b", VPR64, v8i8,
264 0b0, 0b1, 0b00, 0b10010, Neon_mls>;
265 def MLSvvv_16B: NeonI_3VSame_Constraint_impl<"mls", ".16b", VPR128, v16i8,
266 0b1, 0b1, 0b00, 0b10010, Neon_mls>;
267 def MLSvvv_4H: NeonI_3VSame_Constraint_impl<"mls", ".4h", VPR64, v4i16,
268 0b0, 0b1, 0b01, 0b10010, Neon_mls>;
269 def MLSvvv_8H: NeonI_3VSame_Constraint_impl<"mls", ".8h", VPR128, v8i16,
270 0b1, 0b1, 0b01, 0b10010, Neon_mls>;
271 def MLSvvv_2S: NeonI_3VSame_Constraint_impl<"mls", ".2s", VPR64, v2i32,
272 0b0, 0b1, 0b10, 0b10010, Neon_mls>;
273 def MLSvvv_4S: NeonI_3VSame_Constraint_impl<"mls", ".4s", VPR128, v4i32,
274 0b1, 0b1, 0b10, 0b10010, Neon_mls>;
276 // Vector Multiply-accumulate and Multiply-subtract (Floating Point)
278 def Neon_fmla : PatFrag<(ops node:$Ra, node:$Rn, node:$Rm),
279 (fadd node:$Ra, (fmul node:$Rn, node:$Rm))>;
281 def Neon_fmls : PatFrag<(ops node:$Ra, node:$Rn, node:$Rm),
282 (fsub node:$Ra, (fmul node:$Rn, node:$Rm))>;
284 let Predicates = [HasNEON, UseFusedMAC] in {
285 def FMLAvvv_2S: NeonI_3VSame_Constraint_impl<"fmla", ".2s", VPR64, v2f32,
286 0b0, 0b0, 0b00, 0b11001, Neon_fmla>;
287 def FMLAvvv_4S: NeonI_3VSame_Constraint_impl<"fmla", ".4s", VPR128, v4f32,
288 0b1, 0b0, 0b00, 0b11001, Neon_fmla>;
289 def FMLAvvv_2D: NeonI_3VSame_Constraint_impl<"fmla", ".2d", VPR128, v2f64,
290 0b1, 0b0, 0b01, 0b11001, Neon_fmla>;
292 def FMLSvvv_2S: NeonI_3VSame_Constraint_impl<"fmls", ".2s", VPR64, v2f32,
293 0b0, 0b0, 0b10, 0b11001, Neon_fmls>;
294 def FMLSvvv_4S: NeonI_3VSame_Constraint_impl<"fmls", ".4s", VPR128, v4f32,
295 0b1, 0b0, 0b10, 0b11001, Neon_fmls>;
296 def FMLSvvv_2D: NeonI_3VSame_Constraint_impl<"fmls", ".2d", VPR128, v2f64,
297 0b1, 0b0, 0b11, 0b11001, Neon_fmls>;
300 // We're also allowed to match the fma instruction regardless of compile
302 def : Pat<(v2f32 (fma VPR64:$Rn, VPR64:$Rm, VPR64:$Ra)),
303 (FMLAvvv_2S VPR64:$Ra, VPR64:$Rn, VPR64:$Rm)>;
304 def : Pat<(v4f32 (fma VPR128:$Rn, VPR128:$Rm, VPR128:$Ra)),
305 (FMLAvvv_4S VPR128:$Ra, VPR128:$Rn, VPR128:$Rm)>;
306 def : Pat<(v2f64 (fma VPR128:$Rn, VPR128:$Rm, VPR128:$Ra)),
307 (FMLAvvv_2D VPR128:$Ra, VPR128:$Rn, VPR128:$Rm)>;
309 def : Pat<(v2f32 (fma (fneg VPR64:$Rn), VPR64:$Rm, VPR64:$Ra)),
310 (FMLSvvv_2S VPR64:$Ra, VPR64:$Rn, VPR64:$Rm)>;
311 def : Pat<(v4f32 (fma (fneg VPR128:$Rn), VPR128:$Rm, VPR128:$Ra)),
312 (FMLSvvv_4S VPR128:$Ra, VPR128:$Rn, VPR128:$Rm)>;
313 def : Pat<(v2f64 (fma (fneg VPR128:$Rn), VPR128:$Rm, VPR128:$Ra)),
314 (FMLSvvv_2D VPR128:$Ra, VPR128:$Rn, VPR128:$Rm)>;
316 // Vector Divide (Floating-Point)
318 defm FDIVvvv : NeonI_3VSame_SD_sizes<0b1, 0b0, 0b11111, "fdiv", fdiv, fdiv, fdiv,
319 v2f32, v4f32, v2f64, 0>;
321 // Vector Bitwise Operations
323 // Vector Bitwise AND
325 defm ANDvvv : NeonI_3VSame_B_sizes<0b0, 0b00, 0b00011, "and", and, and, 1>;
327 // Vector Bitwise Exclusive OR
329 defm EORvvv : NeonI_3VSame_B_sizes<0b1, 0b00, 0b00011, "eor", xor, xor, 1>;
333 defm ORRvvv : NeonI_3VSame_B_sizes<0b0, 0b10, 0b00011, "orr", or, or, 1>;
335 // ORR disassembled as MOV if Vn==Vm
337 // Vector Move - register
338 // Alias for ORR if Vn=Vm.
339 // FIXME: This is actually the preferred syntax but TableGen can't deal with
340 // custom printing of aliases.
341 def : NeonInstAlias<"mov $Rd.8b, $Rn.8b",
342 (ORRvvv_8B VPR64:$Rd, VPR64:$Rn, VPR64:$Rn), 0>;
343 def : NeonInstAlias<"mov $Rd.16b, $Rn.16b",
344 (ORRvvv_16B VPR128:$Rd, VPR128:$Rn, VPR128:$Rn), 0>;
346 // The MOVI instruction takes two immediate operands. The first is the
347 // immediate encoding, while the second is the cmode. A cmode of 14, or
348 // 0b1110, produces a MOVI operation, rather than a MVNI, ORR, or BIC.
349 def Neon_AllZero : PatFrag<(ops), (Neon_movi (i32 0), (i32 14))>;
350 def Neon_AllOne : PatFrag<(ops), (Neon_movi (i32 255), (i32 14))>;
352 def Neon_not8B : PatFrag<(ops node:$in),
353 (xor node:$in, (bitconvert (v8i8 Neon_AllOne)))>;
354 def Neon_not16B : PatFrag<(ops node:$in),
355 (xor node:$in, (bitconvert (v16i8 Neon_AllOne)))>;
357 def Neon_orn8B : PatFrag<(ops node:$Rn, node:$Rm),
358 (or node:$Rn, (Neon_not8B node:$Rm))>;
360 def Neon_orn16B : PatFrag<(ops node:$Rn, node:$Rm),
361 (or node:$Rn, (Neon_not16B node:$Rm))>;
363 def Neon_bic8B : PatFrag<(ops node:$Rn, node:$Rm),
364 (and node:$Rn, (Neon_not8B node:$Rm))>;
366 def Neon_bic16B : PatFrag<(ops node:$Rn, node:$Rm),
367 (and node:$Rn, (Neon_not16B node:$Rm))>;
370 // Vector Bitwise OR NOT - register
372 defm ORNvvv : NeonI_3VSame_B_sizes<0b0, 0b11, 0b00011, "orn",
373 Neon_orn8B, Neon_orn16B, 0>;
375 // Vector Bitwise Bit Clear (AND NOT) - register
377 defm BICvvv : NeonI_3VSame_B_sizes<0b0, 0b01, 0b00011, "bic",
378 Neon_bic8B, Neon_bic16B, 0>;
380 multiclass Neon_bitwise2V_patterns<SDPatternOperator opnode8B,
381 SDPatternOperator opnode16B,
383 Instruction INST16B> {
384 def : Pat<(v2i32 (opnode8B VPR64:$Rn, VPR64:$Rm)),
385 (INST8B VPR64:$Rn, VPR64:$Rm)>;
386 def : Pat<(v4i16 (opnode8B VPR64:$Rn, VPR64:$Rm)),
387 (INST8B VPR64:$Rn, VPR64:$Rm)>;
388 def : Pat<(v1i64 (opnode8B VPR64:$Rn, VPR64:$Rm)),
389 (INST8B VPR64:$Rn, VPR64:$Rm)>;
390 def : Pat<(v4i32 (opnode16B VPR128:$Rn, VPR128:$Rm)),
391 (INST16B VPR128:$Rn, VPR128:$Rm)>;
392 def : Pat<(v8i16 (opnode16B VPR128:$Rn, VPR128:$Rm)),
393 (INST16B VPR128:$Rn, VPR128:$Rm)>;
394 def : Pat<(v2i64 (opnode16B VPR128:$Rn, VPR128:$Rm)),
395 (INST16B VPR128:$Rn, VPR128:$Rm)>;
398 // Additional patterns for bitwise instructions AND, EOR, ORR, BIC, ORN
399 defm : Neon_bitwise2V_patterns<and, and, ANDvvv_8B, ANDvvv_16B>;
400 defm : Neon_bitwise2V_patterns<or, or, ORRvvv_8B, ORRvvv_16B>;
401 defm : Neon_bitwise2V_patterns<xor, xor, EORvvv_8B, EORvvv_16B>;
402 defm : Neon_bitwise2V_patterns<Neon_bic8B, Neon_bic16B, BICvvv_8B, BICvvv_16B>;
403 defm : Neon_bitwise2V_patterns<Neon_orn8B, Neon_orn16B, ORNvvv_8B, ORNvvv_16B>;
405 // Vector Bitwise Select
406 def BSLvvv_8B : NeonI_3VSame_Constraint_impl<"bsl", ".8b", VPR64, v8i8,
407 0b0, 0b1, 0b01, 0b00011, vselect>;
409 def BSLvvv_16B : NeonI_3VSame_Constraint_impl<"bsl", ".16b", VPR128, v16i8,
410 0b1, 0b1, 0b01, 0b00011, vselect>;
412 multiclass Neon_bitwise3V_patterns<SDPatternOperator opnode,
414 Instruction INST16B> {
415 // Disassociate type from instruction definition
416 def : Pat<(v8i8 (opnode (v8i8 VPR64:$src), VPR64:$Rn, VPR64:$Rm)),
417 (INST8B VPR64:$src, VPR64:$Rn, VPR64:$Rm)>;
418 def : Pat<(v2i32 (opnode (v2i32 VPR64:$src), VPR64:$Rn, VPR64:$Rm)),
419 (INST8B VPR64:$src, VPR64:$Rn, VPR64:$Rm)>;
420 def : Pat<(v4i16 (opnode (v4i16 VPR64:$src), VPR64:$Rn, VPR64:$Rm)),
421 (INST8B VPR64:$src, VPR64:$Rn, VPR64:$Rm)>;
422 def : Pat<(v1i64 (opnode (v1i64 VPR64:$src), VPR64:$Rn, VPR64:$Rm)),
423 (INST8B VPR64:$src, VPR64:$Rn, VPR64:$Rm)>;
424 def : Pat<(v16i8 (opnode (v16i8 VPR128:$src), VPR128:$Rn, VPR128:$Rm)),
425 (INST16B VPR128:$src, VPR128:$Rn, VPR128:$Rm)>;
426 def : Pat<(v4i32 (opnode (v4i32 VPR128:$src), VPR128:$Rn, VPR128:$Rm)),
427 (INST16B VPR128:$src, VPR128:$Rn, VPR128:$Rm)>;
428 def : Pat<(v8i16 (opnode (v8i16 VPR128:$src), VPR128:$Rn, VPR128:$Rm)),
429 (INST16B VPR128:$src, VPR128:$Rn, VPR128:$Rm)>;
430 def : Pat<(v2i64 (opnode (v2i64 VPR128:$src), VPR128:$Rn, VPR128:$Rm)),
431 (INST16B VPR128:$src, VPR128:$Rn, VPR128:$Rm)>;
433 // Allow to match BSL instruction pattern with non-constant operand
434 def : Pat<(v8i8 (or (and VPR64:$Rn, VPR64:$Rd),
435 (and VPR64:$Rm, (Neon_not8B VPR64:$Rd)))),
436 (INST8B VPR64:$Rd, VPR64:$Rn, VPR64:$Rm)>;
437 def : Pat<(v4i16 (or (and VPR64:$Rn, VPR64:$Rd),
438 (and VPR64:$Rm, (Neon_not8B VPR64:$Rd)))),
439 (INST8B VPR64:$Rd, VPR64:$Rn, VPR64:$Rm)>;
440 def : Pat<(v2i32 (or (and VPR64:$Rn, VPR64:$Rd),
441 (and VPR64:$Rm, (Neon_not8B VPR64:$Rd)))),
442 (INST8B VPR64:$Rd, VPR64:$Rn, VPR64:$Rm)>;
443 def : Pat<(v1i64 (or (and VPR64:$Rn, VPR64:$Rd),
444 (and VPR64:$Rm, (Neon_not8B VPR64:$Rd)))),
445 (INST8B VPR64:$Rd, VPR64:$Rn, VPR64:$Rm)>;
446 def : Pat<(v16i8 (or (and VPR128:$Rn, VPR128:$Rd),
447 (and VPR128:$Rm, (Neon_not16B VPR128:$Rd)))),
448 (INST16B VPR128:$Rd, VPR128:$Rn, VPR128:$Rm)>;
449 def : Pat<(v8i16 (or (and VPR128:$Rn, VPR128:$Rd),
450 (and VPR128:$Rm, (Neon_not16B VPR128:$Rd)))),
451 (INST16B VPR128:$Rd, VPR128:$Rn, VPR128:$Rm)>;
452 def : Pat<(v4i32 (or (and VPR128:$Rn, VPR128:$Rd),
453 (and VPR128:$Rm, (Neon_not16B VPR128:$Rd)))),
454 (INST16B VPR128:$Rd, VPR128:$Rn, VPR128:$Rm)>;
455 def : Pat<(v2i64 (or (and VPR128:$Rn, VPR128:$Rd),
456 (and VPR128:$Rm, (Neon_not16B VPR128:$Rd)))),
457 (INST16B VPR128:$Rd, VPR128:$Rn, VPR128:$Rm)>;
459 // Allow to match llvm.arm.* intrinsics.
460 def : Pat<(v8i8 (int_arm_neon_vbsl (v8i8 VPR64:$src),
461 (v8i8 VPR64:$Rn), (v8i8 VPR64:$Rm))),
462 (INST8B VPR64:$src, VPR64:$Rn, VPR64:$Rm)>;
463 def : Pat<(v4i16 (int_arm_neon_vbsl (v4i16 VPR64:$src),
464 (v4i16 VPR64:$Rn), (v4i16 VPR64:$Rm))),
465 (INST8B VPR64:$src, VPR64:$Rn, VPR64:$Rm)>;
466 def : Pat<(v2i32 (int_arm_neon_vbsl (v2i32 VPR64:$src),
467 (v2i32 VPR64:$Rn), (v2i32 VPR64:$Rm))),
468 (INST8B VPR64:$src, VPR64:$Rn, VPR64:$Rm)>;
469 def : Pat<(v1i64 (int_arm_neon_vbsl (v1i64 VPR64:$src),
470 (v1i64 VPR64:$Rn), (v1i64 VPR64:$Rm))),
471 (INST8B VPR64:$src, VPR64:$Rn, VPR64:$Rm)>;
472 def : Pat<(v2f32 (int_arm_neon_vbsl (v2f32 VPR64:$src),
473 (v2f32 VPR64:$Rn), (v2f32 VPR64:$Rm))),
474 (INST8B VPR64:$src, VPR64:$Rn, VPR64:$Rm)>;
475 def : Pat<(v1f64 (int_arm_neon_vbsl (v1f64 VPR64:$src),
476 (v1f64 VPR64:$Rn), (v1f64 VPR64:$Rm))),
477 (INST8B VPR64:$src, VPR64:$Rn, VPR64:$Rm)>;
478 def : Pat<(v16i8 (int_arm_neon_vbsl (v16i8 VPR128:$src),
479 (v16i8 VPR128:$Rn), (v16i8 VPR128:$Rm))),
480 (INST16B VPR128:$src, VPR128:$Rn, VPR128:$Rm)>;
481 def : Pat<(v8i16 (int_arm_neon_vbsl (v8i16 VPR128:$src),
482 (v8i16 VPR128:$Rn), (v8i16 VPR128:$Rm))),
483 (INST16B VPR128:$src, VPR128:$Rn, VPR128:$Rm)>;
484 def : Pat<(v4i32 (int_arm_neon_vbsl (v4i32 VPR128:$src),
485 (v4i32 VPR128:$Rn), (v4i32 VPR128:$Rm))),
486 (INST16B VPR128:$src, VPR128:$Rn, VPR128:$Rm)>;
487 def : Pat<(v2i64 (int_arm_neon_vbsl (v2i64 VPR128:$src),
488 (v2i64 VPR128:$Rn), (v2i64 VPR128:$Rm))),
489 (INST16B VPR128:$src, VPR128:$Rn, VPR128:$Rm)>;
490 def : Pat<(v4f32 (int_arm_neon_vbsl (v4f32 VPR128:$src),
491 (v4f32 VPR128:$Rn), (v4f32 VPR128:$Rm))),
492 (INST16B VPR128:$src, VPR128:$Rn, VPR128:$Rm)>;
493 def : Pat<(v2f64 (int_arm_neon_vbsl (v2f64 VPR128:$src),
494 (v2f64 VPR128:$Rn), (v2f64 VPR128:$Rm))),
495 (INST16B VPR128:$src, VPR128:$Rn, VPR128:$Rm)>;
498 // Additional patterns for bitwise instruction BSL
499 defm: Neon_bitwise3V_patterns<vselect, BSLvvv_8B, BSLvvv_16B>;
501 def Neon_NoBSLop : PatFrag<(ops node:$src, node:$Rn, node:$Rm),
502 (vselect node:$src, node:$Rn, node:$Rm),
503 [{ (void)N; return false; }]>;
505 // Vector Bitwise Insert if True
507 def BITvvv_8B : NeonI_3VSame_Constraint_impl<"bit", ".8b", VPR64, v8i8,
508 0b0, 0b1, 0b10, 0b00011, Neon_NoBSLop>;
509 def BITvvv_16B : NeonI_3VSame_Constraint_impl<"bit", ".16b", VPR128, v16i8,
510 0b1, 0b1, 0b10, 0b00011, Neon_NoBSLop>;
512 // Vector Bitwise Insert if False
514 def BIFvvv_8B : NeonI_3VSame_Constraint_impl<"bif", ".8b", VPR64, v8i8,
515 0b0, 0b1, 0b11, 0b00011, Neon_NoBSLop>;
516 def BIFvvv_16B : NeonI_3VSame_Constraint_impl<"bif", ".16b", VPR128, v16i8,
517 0b1, 0b1, 0b11, 0b00011, Neon_NoBSLop>;
519 // Vector Absolute Difference and Accumulate (Signed, Unsigned)
521 def Neon_uaba : PatFrag<(ops node:$Ra, node:$Rn, node:$Rm),
522 (add node:$Ra, (int_arm_neon_vabdu node:$Rn, node:$Rm))>;
523 def Neon_saba : PatFrag<(ops node:$Ra, node:$Rn, node:$Rm),
524 (add node:$Ra, (int_arm_neon_vabds node:$Rn, node:$Rm))>;
526 // Vector Absolute Difference and Accumulate (Unsigned)
527 def UABAvvv_8B : NeonI_3VSame_Constraint_impl<"uaba", ".8b", VPR64, v8i8,
528 0b0, 0b1, 0b00, 0b01111, Neon_uaba>;
529 def UABAvvv_16B : NeonI_3VSame_Constraint_impl<"uaba", ".16b", VPR128, v16i8,
530 0b1, 0b1, 0b00, 0b01111, Neon_uaba>;
531 def UABAvvv_4H : NeonI_3VSame_Constraint_impl<"uaba", ".4h", VPR64, v4i16,
532 0b0, 0b1, 0b01, 0b01111, Neon_uaba>;
533 def UABAvvv_8H : NeonI_3VSame_Constraint_impl<"uaba", ".8h", VPR128, v8i16,
534 0b1, 0b1, 0b01, 0b01111, Neon_uaba>;
535 def UABAvvv_2S : NeonI_3VSame_Constraint_impl<"uaba", ".2s", VPR64, v2i32,
536 0b0, 0b1, 0b10, 0b01111, Neon_uaba>;
537 def UABAvvv_4S : NeonI_3VSame_Constraint_impl<"uaba", ".4s", VPR128, v4i32,
538 0b1, 0b1, 0b10, 0b01111, Neon_uaba>;
540 // Vector Absolute Difference and Accumulate (Signed)
541 def SABAvvv_8B : NeonI_3VSame_Constraint_impl<"saba", ".8b", VPR64, v8i8,
542 0b0, 0b0, 0b00, 0b01111, Neon_saba>;
543 def SABAvvv_16B : NeonI_3VSame_Constraint_impl<"saba", ".16b", VPR128, v16i8,
544 0b1, 0b0, 0b00, 0b01111, Neon_saba>;
545 def SABAvvv_4H : NeonI_3VSame_Constraint_impl<"saba", ".4h", VPR64, v4i16,
546 0b0, 0b0, 0b01, 0b01111, Neon_saba>;
547 def SABAvvv_8H : NeonI_3VSame_Constraint_impl<"saba", ".8h", VPR128, v8i16,
548 0b1, 0b0, 0b01, 0b01111, Neon_saba>;
549 def SABAvvv_2S : NeonI_3VSame_Constraint_impl<"saba", ".2s", VPR64, v2i32,
550 0b0, 0b0, 0b10, 0b01111, Neon_saba>;
551 def SABAvvv_4S : NeonI_3VSame_Constraint_impl<"saba", ".4s", VPR128, v4i32,
552 0b1, 0b0, 0b10, 0b01111, Neon_saba>;
555 // Vector Absolute Difference (Signed, Unsigned)
556 defm UABDvvv : NeonI_3VSame_BHS_sizes<0b1, 0b01110, "uabd", int_arm_neon_vabdu, 0>;
557 defm SABDvvv : NeonI_3VSame_BHS_sizes<0b0, 0b01110, "sabd", int_arm_neon_vabds, 0>;
559 // Vector Absolute Difference (Floating Point)
560 defm FABDvvv: NeonI_3VSame_SD_sizes<0b1, 0b1, 0b11010, "fabd",
561 int_arm_neon_vabds, int_arm_neon_vabds,
562 int_arm_neon_vabds, v2f32, v4f32, v2f64, 0>;
564 // Vector Reciprocal Step (Floating Point)
565 defm FRECPSvvv : NeonI_3VSame_SD_sizes<0b0, 0b0, 0b11111, "frecps",
566 int_arm_neon_vrecps, int_arm_neon_vrecps,
568 v2f32, v4f32, v2f64, 0>;
570 // Vector Reciprocal Square Root Step (Floating Point)
571 defm FRSQRTSvvv : NeonI_3VSame_SD_sizes<0b0, 0b1, 0b11111, "frsqrts",
572 int_arm_neon_vrsqrts,
573 int_arm_neon_vrsqrts,
574 int_arm_neon_vrsqrts,
575 v2f32, v4f32, v2f64, 0>;
577 // Vector Comparisons
579 def Neon_cmeq : PatFrag<(ops node:$lhs, node:$rhs),
580 (Neon_cmp node:$lhs, node:$rhs, SETEQ)>;
581 def Neon_cmphs : PatFrag<(ops node:$lhs, node:$rhs),
582 (Neon_cmp node:$lhs, node:$rhs, SETUGE)>;
583 def Neon_cmge : PatFrag<(ops node:$lhs, node:$rhs),
584 (Neon_cmp node:$lhs, node:$rhs, SETGE)>;
585 def Neon_cmhi : PatFrag<(ops node:$lhs, node:$rhs),
586 (Neon_cmp node:$lhs, node:$rhs, SETUGT)>;
587 def Neon_cmgt : PatFrag<(ops node:$lhs, node:$rhs),
588 (Neon_cmp node:$lhs, node:$rhs, SETGT)>;
590 // NeonI_compare_aliases class: swaps register operands to implement
591 // comparison aliases, e.g., CMLE is alias for CMGE with operands reversed.
592 class NeonI_compare_aliases<string asmop, string asmlane,
593 Instruction inst, RegisterOperand VPRC>
594 : NeonInstAlias<asmop # "\t$Rd" # asmlane #", $Rn" # asmlane #
596 (inst VPRC:$Rd, VPRC:$Rm, VPRC:$Rn), 0b0>;
598 // Vector Comparisons (Integer)
600 // Vector Compare Mask Equal (Integer)
601 let isCommutable =1 in {
602 defm CMEQvvv : NeonI_3VSame_BHSD_sizes<0b1, 0b10001, "cmeq", Neon_cmeq, 0>;
605 // Vector Compare Mask Higher or Same (Unsigned Integer)
606 defm CMHSvvv : NeonI_3VSame_BHSD_sizes<0b1, 0b00111, "cmhs", Neon_cmphs, 0>;
608 // Vector Compare Mask Greater Than or Equal (Integer)
609 defm CMGEvvv : NeonI_3VSame_BHSD_sizes<0b0, 0b00111, "cmge", Neon_cmge, 0>;
611 // Vector Compare Mask Higher (Unsigned Integer)
612 defm CMHIvvv : NeonI_3VSame_BHSD_sizes<0b1, 0b00110, "cmhi", Neon_cmhi, 0>;
614 // Vector Compare Mask Greater Than (Integer)
615 defm CMGTvvv : NeonI_3VSame_BHSD_sizes<0b0, 0b00110, "cmgt", Neon_cmgt, 0>;
617 // Vector Compare Mask Bitwise Test (Integer)
618 defm CMTSTvvv: NeonI_3VSame_BHSD_sizes<0b0, 0b10001, "cmtst", Neon_tst, 0>;
620 // Vector Compare Mask Less or Same (Unsigned Integer)
621 // CMLS is alias for CMHS with operands reversed.
622 def CMLSvvv_8B : NeonI_compare_aliases<"cmls", ".8b", CMHSvvv_8B, VPR64>;
623 def CMLSvvv_16B : NeonI_compare_aliases<"cmls", ".16b", CMHSvvv_16B, VPR128>;
624 def CMLSvvv_4H : NeonI_compare_aliases<"cmls", ".4h", CMHSvvv_4H, VPR64>;
625 def CMLSvvv_8H : NeonI_compare_aliases<"cmls", ".8h", CMHSvvv_8H, VPR128>;
626 def CMLSvvv_2S : NeonI_compare_aliases<"cmls", ".2s", CMHSvvv_2S, VPR64>;
627 def CMLSvvv_4S : NeonI_compare_aliases<"cmls", ".4s", CMHSvvv_4S, VPR128>;
628 def CMLSvvv_2D : NeonI_compare_aliases<"cmls", ".2d", CMHSvvv_2D, VPR128>;
630 // Vector Compare Mask Less Than or Equal (Integer)
631 // CMLE is alias for CMGE with operands reversed.
632 def CMLEvvv_8B : NeonI_compare_aliases<"cmle", ".8b", CMGEvvv_8B, VPR64>;
633 def CMLEvvv_16B : NeonI_compare_aliases<"cmle", ".16b", CMGEvvv_16B, VPR128>;
634 def CMLEvvv_4H : NeonI_compare_aliases<"cmle", ".4h", CMGEvvv_4H, VPR64>;
635 def CMLEvvv_8H : NeonI_compare_aliases<"cmle", ".8h", CMGEvvv_8H, VPR128>;
636 def CMLEvvv_2S : NeonI_compare_aliases<"cmle", ".2s", CMGEvvv_2S, VPR64>;
637 def CMLEvvv_4S : NeonI_compare_aliases<"cmle", ".4s", CMGEvvv_4S, VPR128>;
638 def CMLEvvv_2D : NeonI_compare_aliases<"cmle", ".2d", CMGEvvv_2D, VPR128>;
640 // Vector Compare Mask Lower (Unsigned Integer)
641 // CMLO is alias for CMHI with operands reversed.
642 def CMLOvvv_8B : NeonI_compare_aliases<"cmlo", ".8b", CMHIvvv_8B, VPR64>;
643 def CMLOvvv_16B : NeonI_compare_aliases<"cmlo", ".16b", CMHIvvv_16B, VPR128>;
644 def CMLOvvv_4H : NeonI_compare_aliases<"cmlo", ".4h", CMHIvvv_4H, VPR64>;
645 def CMLOvvv_8H : NeonI_compare_aliases<"cmlo", ".8h", CMHIvvv_8H, VPR128>;
646 def CMLOvvv_2S : NeonI_compare_aliases<"cmlo", ".2s", CMHIvvv_2S, VPR64>;
647 def CMLOvvv_4S : NeonI_compare_aliases<"cmlo", ".4s", CMHIvvv_4S, VPR128>;
648 def CMLOvvv_2D : NeonI_compare_aliases<"cmlo", ".2d", CMHIvvv_2D, VPR128>;
650 // Vector Compare Mask Less Than (Integer)
651 // CMLT is alias for CMGT with operands reversed.
652 def CMLTvvv_8B : NeonI_compare_aliases<"cmlt", ".8b", CMGTvvv_8B, VPR64>;
653 def CMLTvvv_16B : NeonI_compare_aliases<"cmlt", ".16b", CMGTvvv_16B, VPR128>;
654 def CMLTvvv_4H : NeonI_compare_aliases<"cmlt", ".4h", CMGTvvv_4H, VPR64>;
655 def CMLTvvv_8H : NeonI_compare_aliases<"cmlt", ".8h", CMGTvvv_8H, VPR128>;
656 def CMLTvvv_2S : NeonI_compare_aliases<"cmlt", ".2s", CMGTvvv_2S, VPR64>;
657 def CMLTvvv_4S : NeonI_compare_aliases<"cmlt", ".4s", CMGTvvv_4S, VPR128>;
658 def CMLTvvv_2D : NeonI_compare_aliases<"cmlt", ".2d", CMGTvvv_2D, VPR128>;
661 def neon_uimm0_asmoperand : AsmOperandClass
664 let PredicateMethod = "isUImm<0>";
665 let RenderMethod = "addImmOperands";
668 def neon_uimm0 : Operand<i32>, ImmLeaf<i32, [{return Imm == 0;}]> {
669 let ParserMatchClass = neon_uimm0_asmoperand;
670 let PrintMethod = "printNeonUImm0Operand";
674 multiclass NeonI_cmpz_sizes<bit u, bits<5> opcode, string asmop, CondCode CC>
676 def _8B : NeonI_2VMisc<0b0, u, 0b00, opcode,
677 (outs VPR64:$Rd), (ins VPR64:$Rn, neon_uimm0:$Imm),
678 asmop # "\t$Rd.8b, $Rn.8b, $Imm",
679 [(set (v8i8 VPR64:$Rd),
680 (v8i8 (Neon_cmpz (v8i8 VPR64:$Rn), (i32 imm:$Imm), CC)))],
683 def _16B : NeonI_2VMisc<0b1, u, 0b00, opcode,
684 (outs VPR128:$Rd), (ins VPR128:$Rn, neon_uimm0:$Imm),
685 asmop # "\t$Rd.16b, $Rn.16b, $Imm",
686 [(set (v16i8 VPR128:$Rd),
687 (v16i8 (Neon_cmpz (v16i8 VPR128:$Rn), (i32 imm:$Imm), CC)))],
690 def _4H : NeonI_2VMisc<0b0, u, 0b01, opcode,
691 (outs VPR64:$Rd), (ins VPR64:$Rn, neon_uimm0:$Imm),
692 asmop # "\t$Rd.4h, $Rn.4h, $Imm",
693 [(set (v4i16 VPR64:$Rd),
694 (v4i16 (Neon_cmpz (v4i16 VPR64:$Rn), (i32 imm:$Imm), CC)))],
697 def _8H : NeonI_2VMisc<0b1, u, 0b01, opcode,
698 (outs VPR128:$Rd), (ins VPR128:$Rn, neon_uimm0:$Imm),
699 asmop # "\t$Rd.8h, $Rn.8h, $Imm",
700 [(set (v8i16 VPR128:$Rd),
701 (v8i16 (Neon_cmpz (v8i16 VPR128:$Rn), (i32 imm:$Imm), CC)))],
704 def _2S : NeonI_2VMisc<0b0, u, 0b10, opcode,
705 (outs VPR64:$Rd), (ins VPR64:$Rn, neon_uimm0:$Imm),
706 asmop # "\t$Rd.2s, $Rn.2s, $Imm",
707 [(set (v2i32 VPR64:$Rd),
708 (v2i32 (Neon_cmpz (v2i32 VPR64:$Rn), (i32 imm:$Imm), CC)))],
711 def _4S : NeonI_2VMisc<0b1, u, 0b10, opcode,
712 (outs VPR128:$Rd), (ins VPR128:$Rn, neon_uimm0:$Imm),
713 asmop # "\t$Rd.4s, $Rn.4s, $Imm",
714 [(set (v4i32 VPR128:$Rd),
715 (v4i32 (Neon_cmpz (v4i32 VPR128:$Rn), (i32 imm:$Imm), CC)))],
718 def _2D : NeonI_2VMisc<0b1, u, 0b11, opcode,
719 (outs VPR128:$Rd), (ins VPR128:$Rn, neon_uimm0:$Imm),
720 asmop # "\t$Rd.2d, $Rn.2d, $Imm",
721 [(set (v2i64 VPR128:$Rd),
722 (v2i64 (Neon_cmpz (v2i64 VPR128:$Rn), (i32 imm:$Imm), CC)))],
726 // Vector Compare Mask Equal to Zero (Integer)
727 defm CMEQvvi : NeonI_cmpz_sizes<0b0, 0b01001, "cmeq", SETEQ>;
729 // Vector Compare Mask Greater Than or Equal to Zero (Signed Integer)
730 defm CMGEvvi : NeonI_cmpz_sizes<0b1, 0b01000, "cmge", SETGE>;
732 // Vector Compare Mask Greater Than Zero (Signed Integer)
733 defm CMGTvvi : NeonI_cmpz_sizes<0b0, 0b01000, "cmgt", SETGT>;
735 // Vector Compare Mask Less Than or Equal To Zero (Signed Integer)
736 defm CMLEvvi : NeonI_cmpz_sizes<0b1, 0b01001, "cmle", SETLE>;
738 // Vector Compare Mask Less Than Zero (Signed Integer)
739 defm CMLTvvi : NeonI_cmpz_sizes<0b0, 0b01010, "cmlt", SETLT>;
741 // Vector Comparisons (Floating Point)
743 // Vector Compare Mask Equal (Floating Point)
744 let isCommutable =1 in {
745 defm FCMEQvvv : NeonI_3VSame_SD_sizes<0b0, 0b0, 0b11100, "fcmeq", Neon_cmeq,
746 Neon_cmeq, Neon_cmeq,
747 v2i32, v4i32, v2i64, 0>;
750 // Vector Compare Mask Greater Than Or Equal (Floating Point)
751 defm FCMGEvvv : NeonI_3VSame_SD_sizes<0b1, 0b0, 0b11100, "fcmge", Neon_cmge,
752 Neon_cmge, Neon_cmge,
753 v2i32, v4i32, v2i64, 0>;
755 // Vector Compare Mask Greater Than (Floating Point)
756 defm FCMGTvvv : NeonI_3VSame_SD_sizes<0b1, 0b1, 0b11100, "fcmgt", Neon_cmgt,
757 Neon_cmgt, Neon_cmgt,
758 v2i32, v4i32, v2i64, 0>;
760 // Vector Compare Mask Less Than Or Equal (Floating Point)
761 // FCMLE is alias for FCMGE with operands reversed.
762 def FCMLEvvv_2S : NeonI_compare_aliases<"fcmle", ".2s", FCMGEvvv_2S, VPR64>;
763 def FCMLEvvv_4S : NeonI_compare_aliases<"fcmle", ".4s", FCMGEvvv_4S, VPR128>;
764 def FCMLEvvv_2D : NeonI_compare_aliases<"fcmle", ".2d", FCMGEvvv_2D, VPR128>;
766 // Vector Compare Mask Less Than (Floating Point)
767 // FCMLT is alias for FCMGT with operands reversed.
768 def FCMLTvvv_2S : NeonI_compare_aliases<"fcmlt", ".2s", FCMGTvvv_2S, VPR64>;
769 def FCMLTvvv_4S : NeonI_compare_aliases<"fcmlt", ".4s", FCMGTvvv_4S, VPR128>;
770 def FCMLTvvv_2D : NeonI_compare_aliases<"fcmlt", ".2d", FCMGTvvv_2D, VPR128>;
773 multiclass NeonI_fpcmpz_sizes<bit u, bit size, bits<5> opcode,
774 string asmop, CondCode CC>
776 def _2S : NeonI_2VMisc<0b0, u, {size, 0b0}, opcode,
777 (outs VPR64:$Rd), (ins VPR64:$Rn, fpz32:$FPImm),
778 asmop # "\t$Rd.2s, $Rn.2s, $FPImm",
779 [(set (v2i32 VPR64:$Rd),
780 (v2i32 (Neon_cmpz (v2f32 VPR64:$Rn), (f32 fpz32:$FPImm), CC)))],
783 def _4S : NeonI_2VMisc<0b1, u, {size, 0b0}, opcode,
784 (outs VPR128:$Rd), (ins VPR128:$Rn, fpz32:$FPImm),
785 asmop # "\t$Rd.4s, $Rn.4s, $FPImm",
786 [(set (v4i32 VPR128:$Rd),
787 (v4i32 (Neon_cmpz (v4f32 VPR128:$Rn), (f32 fpz32:$FPImm), CC)))],
790 def _2D : NeonI_2VMisc<0b1, u, {size, 0b1}, opcode,
791 (outs VPR128:$Rd), (ins VPR128:$Rn, fpz32:$FPImm),
792 asmop # "\t$Rd.2d, $Rn.2d, $FPImm",
793 [(set (v2i64 VPR128:$Rd),
794 (v2i64 (Neon_cmpz (v2f64 VPR128:$Rn), (f32 fpz32:$FPImm), CC)))],
798 // Vector Compare Mask Equal to Zero (Floating Point)
799 defm FCMEQvvi : NeonI_fpcmpz_sizes<0b0, 0b1, 0b01101, "fcmeq", SETEQ>;
801 // Vector Compare Mask Greater Than or Equal to Zero (Floating Point)
802 defm FCMGEvvi : NeonI_fpcmpz_sizes<0b1, 0b1, 0b01100, "fcmge", SETGE>;
804 // Vector Compare Mask Greater Than Zero (Floating Point)
805 defm FCMGTvvi : NeonI_fpcmpz_sizes<0b0, 0b1, 0b01100, "fcmgt", SETGT>;
807 // Vector Compare Mask Less Than or Equal To Zero (Floating Point)
808 defm FCMLEvvi : NeonI_fpcmpz_sizes<0b1, 0b1, 0b01101, "fcmle", SETLE>;
810 // Vector Compare Mask Less Than Zero (Floating Point)
811 defm FCMLTvvi : NeonI_fpcmpz_sizes<0b0, 0b1, 0b01110, "fcmlt", SETLT>;
813 // Vector Absolute Comparisons (Floating Point)
815 // Vector Absolute Compare Mask Greater Than Or Equal (Floating Point)
816 defm FACGEvvv : NeonI_3VSame_SD_sizes<0b1, 0b0, 0b11101, "facge",
817 int_arm_neon_vacged, int_arm_neon_vacgeq,
818 int_aarch64_neon_vacgeq,
819 v2i32, v4i32, v2i64, 0>;
821 // Vector Absolute Compare Mask Greater Than (Floating Point)
822 defm FACGTvvv : NeonI_3VSame_SD_sizes<0b1, 0b1, 0b11101, "facgt",
823 int_arm_neon_vacgtd, int_arm_neon_vacgtq,
824 int_aarch64_neon_vacgtq,
825 v2i32, v4i32, v2i64, 0>;
827 // Vector Absolute Compare Mask Less Than Or Equal (Floating Point)
828 // FACLE is alias for FACGE with operands reversed.
829 def FACLEvvv_2S : NeonI_compare_aliases<"facle", ".2s", FACGEvvv_2S, VPR64>;
830 def FACLEvvv_4S : NeonI_compare_aliases<"facle", ".4s", FACGEvvv_4S, VPR128>;
831 def FACLEvvv_2D : NeonI_compare_aliases<"facle", ".2d", FACGEvvv_2D, VPR128>;
833 // Vector Absolute Compare Mask Less Than (Floating Point)
834 // FACLT is alias for FACGT with operands reversed.
835 def FACLTvvv_2S : NeonI_compare_aliases<"faclt", ".2s", FACGTvvv_2S, VPR64>;
836 def FACLTvvv_4S : NeonI_compare_aliases<"faclt", ".4s", FACGTvvv_4S, VPR128>;
837 def FACLTvvv_2D : NeonI_compare_aliases<"faclt", ".2d", FACGTvvv_2D, VPR128>;
839 // Vector halving add (Integer Signed, Unsigned)
840 defm SHADDvvv : NeonI_3VSame_BHS_sizes<0b0, 0b00000, "shadd",
841 int_arm_neon_vhadds, 1>;
842 defm UHADDvvv : NeonI_3VSame_BHS_sizes<0b1, 0b00000, "uhadd",
843 int_arm_neon_vhaddu, 1>;
845 // Vector halving sub (Integer Signed, Unsigned)
846 defm SHSUBvvv : NeonI_3VSame_BHS_sizes<0b0, 0b00100, "shsub",
847 int_arm_neon_vhsubs, 0>;
848 defm UHSUBvvv : NeonI_3VSame_BHS_sizes<0b1, 0b00100, "uhsub",
849 int_arm_neon_vhsubu, 0>;
851 // Vector rouding halving add (Integer Signed, Unsigned)
852 defm SRHADDvvv : NeonI_3VSame_BHS_sizes<0b0, 0b00010, "srhadd",
853 int_arm_neon_vrhadds, 1>;
854 defm URHADDvvv : NeonI_3VSame_BHS_sizes<0b1, 0b00010, "urhadd",
855 int_arm_neon_vrhaddu, 1>;
857 // Vector Saturating add (Integer Signed, Unsigned)
858 defm SQADDvvv : NeonI_3VSame_BHSD_sizes<0b0, 0b00001, "sqadd",
859 int_arm_neon_vqadds, 1>;
860 defm UQADDvvv : NeonI_3VSame_BHSD_sizes<0b1, 0b00001, "uqadd",
861 int_arm_neon_vqaddu, 1>;
863 // Vector Saturating sub (Integer Signed, Unsigned)
864 defm SQSUBvvv : NeonI_3VSame_BHSD_sizes<0b0, 0b00101, "sqsub",
865 int_arm_neon_vqsubs, 1>;
866 defm UQSUBvvv : NeonI_3VSame_BHSD_sizes<0b1, 0b00101, "uqsub",
867 int_arm_neon_vqsubu, 1>;
869 // Vector Shift Left (Signed and Unsigned Integer)
870 defm SSHLvvv : NeonI_3VSame_BHSD_sizes<0b0, 0b01000, "sshl",
871 int_arm_neon_vshifts, 1>;
872 defm USHLvvv : NeonI_3VSame_BHSD_sizes<0b1, 0b01000, "ushl",
873 int_arm_neon_vshiftu, 1>;
875 // Vector Saturating Shift Left (Signed and Unsigned Integer)
876 defm SQSHLvvv : NeonI_3VSame_BHSD_sizes<0b0, 0b01001, "sqshl",
877 int_arm_neon_vqshifts, 1>;
878 defm UQSHLvvv : NeonI_3VSame_BHSD_sizes<0b1, 0b01001, "uqshl",
879 int_arm_neon_vqshiftu, 1>;
881 // Vector Rouding Shift Left (Signed and Unsigned Integer)
882 defm SRSHLvvv : NeonI_3VSame_BHSD_sizes<0b0, 0b01010, "srshl",
883 int_arm_neon_vrshifts, 1>;
884 defm URSHLvvv : NeonI_3VSame_BHSD_sizes<0b1, 0b01010, "urshl",
885 int_arm_neon_vrshiftu, 1>;
887 // Vector Saturating Rouding Shift Left (Signed and Unsigned Integer)
888 defm SQRSHLvvv : NeonI_3VSame_BHSD_sizes<0b0, 0b01011, "sqrshl",
889 int_arm_neon_vqrshifts, 1>;
890 defm UQRSHLvvv : NeonI_3VSame_BHSD_sizes<0b1, 0b01011, "uqrshl",
891 int_arm_neon_vqrshiftu, 1>;
893 // Vector Maximum (Signed and Unsigned Integer)
894 defm SMAXvvv : NeonI_3VSame_BHS_sizes<0b0, 0b01100, "smax", int_arm_neon_vmaxs, 1>;
895 defm UMAXvvv : NeonI_3VSame_BHS_sizes<0b1, 0b01100, "umax", int_arm_neon_vmaxu, 1>;
897 // Vector Minimum (Signed and Unsigned Integer)
898 defm SMINvvv : NeonI_3VSame_BHS_sizes<0b0, 0b01101, "smin", int_arm_neon_vmins, 1>;
899 defm UMINvvv : NeonI_3VSame_BHS_sizes<0b1, 0b01101, "umin", int_arm_neon_vminu, 1>;
901 // Vector Maximum (Floating Point)
902 defm FMAXvvv : NeonI_3VSame_SD_sizes<0b0, 0b0, 0b11110, "fmax",
903 int_arm_neon_vmaxs, int_arm_neon_vmaxs,
904 int_arm_neon_vmaxs, v2f32, v4f32, v2f64, 1>;
906 // Vector Minimum (Floating Point)
907 defm FMINvvv : NeonI_3VSame_SD_sizes<0b0, 0b1, 0b11110, "fmin",
908 int_arm_neon_vmins, int_arm_neon_vmins,
909 int_arm_neon_vmins, v2f32, v4f32, v2f64, 1>;
911 // Vector maxNum (Floating Point) - prefer a number over a quiet NaN)
912 defm FMAXNMvvv : NeonI_3VSame_SD_sizes<0b0, 0b0, 0b11000, "fmaxnm",
913 int_aarch64_neon_vmaxnm,
914 int_aarch64_neon_vmaxnm,
915 int_aarch64_neon_vmaxnm,
916 v2f32, v4f32, v2f64, 1>;
918 // Vector minNum (Floating Point) - prefer a number over a quiet NaN)
919 defm FMINNMvvv : NeonI_3VSame_SD_sizes<0b0, 0b1, 0b11000, "fminnm",
920 int_aarch64_neon_vminnm,
921 int_aarch64_neon_vminnm,
922 int_aarch64_neon_vminnm,
923 v2f32, v4f32, v2f64, 1>;
925 // Vector Maximum Pairwise (Signed and Unsigned Integer)
926 defm SMAXPvvv : NeonI_3VSame_BHS_sizes<0b0, 0b10100, "smaxp", int_arm_neon_vpmaxs, 1>;
927 defm UMAXPvvv : NeonI_3VSame_BHS_sizes<0b1, 0b10100, "umaxp", int_arm_neon_vpmaxu, 1>;
929 // Vector Minimum Pairwise (Signed and Unsigned Integer)
930 defm SMINPvvv : NeonI_3VSame_BHS_sizes<0b0, 0b10101, "sminp", int_arm_neon_vpmins, 1>;
931 defm UMINPvvv : NeonI_3VSame_BHS_sizes<0b1, 0b10101, "uminp", int_arm_neon_vpminu, 1>;
933 // Vector Maximum Pairwise (Floating Point)
934 defm FMAXPvvv : NeonI_3VSame_SD_sizes<0b1, 0b0, 0b11110, "fmaxp",
935 int_arm_neon_vpmaxs, int_arm_neon_vpmaxs,
936 int_arm_neon_vpmaxs, v2f32, v4f32, v2f64, 1>;
938 // Vector Minimum Pairwise (Floating Point)
939 defm FMINPvvv : NeonI_3VSame_SD_sizes<0b1, 0b1, 0b11110, "fminp",
940 int_arm_neon_vpmins, int_arm_neon_vpmins,
941 int_arm_neon_vpmins, v2f32, v4f32, v2f64, 1>;
943 // Vector maxNum Pairwise (Floating Point) - prefer a number over a quiet NaN)
944 defm FMAXNMPvvv : NeonI_3VSame_SD_sizes<0b1, 0b0, 0b11000, "fmaxnmp",
945 int_aarch64_neon_vpmaxnm,
946 int_aarch64_neon_vpmaxnm,
947 int_aarch64_neon_vpmaxnm,
948 v2f32, v4f32, v2f64, 1>;
950 // Vector minNum Pairwise (Floating Point) - prefer a number over a quiet NaN)
951 defm FMINNMPvvv : NeonI_3VSame_SD_sizes<0b1, 0b1, 0b11000, "fminnmp",
952 int_aarch64_neon_vpminnm,
953 int_aarch64_neon_vpminnm,
954 int_aarch64_neon_vpminnm,
955 v2f32, v4f32, v2f64, 1>;
957 // Vector Addition Pairwise (Integer)
958 defm ADDP : NeonI_3VSame_BHSD_sizes<0b0, 0b10111, "addp", int_arm_neon_vpadd, 1>;
960 // Vector Addition Pairwise (Floating Point)
961 defm FADDP : NeonI_3VSame_SD_sizes<0b1, 0b0, 0b11010, "faddp",
965 v2f32, v4f32, v2f64, 1>;
967 // Vector Saturating Doubling Multiply High
968 defm SQDMULHvvv : NeonI_3VSame_HS_sizes<0b0, 0b10110, "sqdmulh",
969 int_arm_neon_vqdmulh, 1>;
971 // Vector Saturating Rouding Doubling Multiply High
972 defm SQRDMULHvvv : NeonI_3VSame_HS_sizes<0b1, 0b10110, "sqrdmulh",
973 int_arm_neon_vqrdmulh, 1>;
975 // Vector Multiply Extended (Floating Point)
976 defm FMULXvvv : NeonI_3VSame_SD_sizes<0b0, 0b0, 0b11011, "fmulx",
977 int_aarch64_neon_vmulx,
978 int_aarch64_neon_vmulx,
979 int_aarch64_neon_vmulx,
980 v2f32, v4f32, v2f64, 1>;
982 // Patterns to match llvm.aarch64.* intrinsic for
983 // ADDP, SMINP, UMINP, SMAXP, UMAXP having i32 as output
984 class Neon_VectorPair_v2i32_pattern<SDPatternOperator opnode, Instruction INST>
985 : Pat<(v1i32 (opnode (v2i32 VPR64:$Rn))),
987 (v2i32 (INST (v2i32 VPR64:$Rn), (v2i32 VPR64:$Rn))),
990 def : Neon_VectorPair_v2i32_pattern<int_aarch64_neon_sminv, SMINPvvv_2S>;
991 def : Neon_VectorPair_v2i32_pattern<int_aarch64_neon_uminv, UMINPvvv_2S>;
992 def : Neon_VectorPair_v2i32_pattern<int_aarch64_neon_smaxv, SMAXPvvv_2S>;
993 def : Neon_VectorPair_v2i32_pattern<int_aarch64_neon_umaxv, UMAXPvvv_2S>;
994 def : Neon_VectorPair_v2i32_pattern<int_aarch64_neon_vaddv, ADDP_2S>;
996 // Vector Immediate Instructions
998 multiclass neon_mov_imm_shift_asmoperands<string PREFIX>
1000 def _asmoperand : AsmOperandClass
1002 let Name = "NeonMovImmShift" # PREFIX;
1003 let RenderMethod = "addNeonMovImmShift" # PREFIX # "Operands";
1004 let PredicateMethod = "isNeonMovImmShift" # PREFIX;
1008 // Definition of vector immediates shift operands
1010 // The selectable use-cases extract the shift operation
1011 // information from the OpCmode fields encoded in the immediate.
1012 def neon_mod_shift_imm_XFORM : SDNodeXForm<imm, [{
1013 uint64_t OpCmode = N->getZExtValue();
1015 unsigned ShiftOnesIn;
1017 A64Imms::decodeNeonModShiftImm(OpCmode, ShiftImm, ShiftOnesIn);
1018 if (!HasShift) return SDValue();
1019 return CurDAG->getTargetConstant(ShiftImm, MVT::i32);
1022 // Vector immediates shift operands which accept LSL and MSL
1023 // shift operators with shift value in the range of 0, 8, 16, 24 (LSL),
1024 // or 0, 8 (LSLH) or 8, 16 (MSL).
1025 defm neon_mov_imm_LSL : neon_mov_imm_shift_asmoperands<"LSL">;
1026 defm neon_mov_imm_MSL : neon_mov_imm_shift_asmoperands<"MSL">;
1027 // LSLH restricts shift amount to 0, 8 out of 0, 8, 16, 24
1028 defm neon_mov_imm_LSLH : neon_mov_imm_shift_asmoperands<"LSLH">;
1030 multiclass neon_mov_imm_shift_operands<string PREFIX,
1031 string HALF, string ISHALF, code pred>
1033 def _operand : Operand<i32>, ImmLeaf<i32, pred, neon_mod_shift_imm_XFORM>
1036 "printNeonMovImmShiftOperand<A64SE::" # PREFIX # ", " # ISHALF # ">";
1038 "DecodeNeonMovImmShiftOperand<A64SE::" # PREFIX # ", " # ISHALF # ">";
1039 let ParserMatchClass =
1040 !cast<AsmOperandClass>("neon_mov_imm_" # PREFIX # HALF # "_asmoperand");
1044 defm neon_mov_imm_LSL : neon_mov_imm_shift_operands<"LSL", "", "false", [{
1046 unsigned ShiftOnesIn;
1048 A64Imms::decodeNeonModShiftImm(Imm, ShiftImm, ShiftOnesIn);
1049 return (HasShift && !ShiftOnesIn);
1052 defm neon_mov_imm_MSL : neon_mov_imm_shift_operands<"MSL", "", "false", [{
1054 unsigned ShiftOnesIn;
1056 A64Imms::decodeNeonModShiftImm(Imm, ShiftImm, ShiftOnesIn);
1057 return (HasShift && ShiftOnesIn);
1060 defm neon_mov_imm_LSLH : neon_mov_imm_shift_operands<"LSL", "H", "true", [{
1062 unsigned ShiftOnesIn;
1064 A64Imms::decodeNeonModShiftImm(Imm, ShiftImm, ShiftOnesIn);
1065 return (HasShift && !ShiftOnesIn);
1068 def neon_uimm1_asmoperand : AsmOperandClass
1071 let PredicateMethod = "isUImm<1>";
1072 let RenderMethod = "addImmOperands";
1075 def neon_uimm2_asmoperand : AsmOperandClass
1078 let PredicateMethod = "isUImm<2>";
1079 let RenderMethod = "addImmOperands";
1082 def neon_uimm8_asmoperand : AsmOperandClass
1085 let PredicateMethod = "isUImm<8>";
1086 let RenderMethod = "addImmOperands";
1089 def neon_uimm8 : Operand<i32>, ImmLeaf<i32, [{(void)Imm; return true;}]> {
1090 let ParserMatchClass = neon_uimm8_asmoperand;
1091 let PrintMethod = "printUImmHexOperand";
1094 def neon_uimm64_mask_asmoperand : AsmOperandClass
1096 let Name = "NeonUImm64Mask";
1097 let PredicateMethod = "isNeonUImm64Mask";
1098 let RenderMethod = "addNeonUImm64MaskOperands";
1101 // MCOperand for 64-bit bytemask with each byte having only the
1102 // value 0x00 and 0xff is encoded as an unsigned 8-bit value
1103 def neon_uimm64_mask : Operand<i32>, ImmLeaf<i32, [{(void)Imm; return true;}]> {
1104 let ParserMatchClass = neon_uimm64_mask_asmoperand;
1105 let PrintMethod = "printNeonUImm64MaskOperand";
1108 multiclass NeonI_mov_imm_lsl_sizes<string asmop, bit op,
1109 SDPatternOperator opnode>
1111 // shift zeros, per word
1112 def _2S : NeonI_1VModImm<0b0, op,
1114 (ins neon_uimm8:$Imm,
1115 neon_mov_imm_LSL_operand:$Simm),
1116 !strconcat(asmop, "\t$Rd.2s, $Imm$Simm"),
1117 [(set (v2i32 VPR64:$Rd),
1118 (v2i32 (opnode (timm:$Imm),
1119 (neon_mov_imm_LSL_operand:$Simm))))],
1122 let cmode = {0b0, Simm{1}, Simm{0}, 0b0};
1125 def _4S : NeonI_1VModImm<0b1, op,
1127 (ins neon_uimm8:$Imm,
1128 neon_mov_imm_LSL_operand:$Simm),
1129 !strconcat(asmop, "\t$Rd.4s, $Imm$Simm"),
1130 [(set (v4i32 VPR128:$Rd),
1131 (v4i32 (opnode (timm:$Imm),
1132 (neon_mov_imm_LSL_operand:$Simm))))],
1135 let cmode = {0b0, Simm{1}, Simm{0}, 0b0};
1138 // shift zeros, per halfword
1139 def _4H : NeonI_1VModImm<0b0, op,
1141 (ins neon_uimm8:$Imm,
1142 neon_mov_imm_LSLH_operand:$Simm),
1143 !strconcat(asmop, "\t$Rd.4h, $Imm$Simm"),
1144 [(set (v4i16 VPR64:$Rd),
1145 (v4i16 (opnode (timm:$Imm),
1146 (neon_mov_imm_LSLH_operand:$Simm))))],
1149 let cmode = {0b1, 0b0, Simm, 0b0};
1152 def _8H : NeonI_1VModImm<0b1, op,
1154 (ins neon_uimm8:$Imm,
1155 neon_mov_imm_LSLH_operand:$Simm),
1156 !strconcat(asmop, "\t$Rd.8h, $Imm$Simm"),
1157 [(set (v8i16 VPR128:$Rd),
1158 (v8i16 (opnode (timm:$Imm),
1159 (neon_mov_imm_LSLH_operand:$Simm))))],
1162 let cmode = {0b1, 0b0, Simm, 0b0};
1166 multiclass NeonI_mov_imm_with_constraint_lsl_sizes<string asmop, bit op,
1167 SDPatternOperator opnode,
1168 SDPatternOperator neonopnode>
1170 let Constraints = "$src = $Rd" in {
1171 // shift zeros, per word
1172 def _2S : NeonI_1VModImm<0b0, op,
1174 (ins VPR64:$src, neon_uimm8:$Imm,
1175 neon_mov_imm_LSL_operand:$Simm),
1176 !strconcat(asmop, "\t$Rd.2s, $Imm$Simm"),
1177 [(set (v2i32 VPR64:$Rd),
1178 (v2i32 (opnode (v2i32 VPR64:$src),
1179 (v2i32 (bitconvert (v2i32 (neonopnode timm:$Imm,
1180 neon_mov_imm_LSL_operand:$Simm)))))))],
1183 let cmode = {0b0, Simm{1}, Simm{0}, 0b1};
1186 def _4S : NeonI_1VModImm<0b1, op,
1188 (ins VPR128:$src, neon_uimm8:$Imm,
1189 neon_mov_imm_LSL_operand:$Simm),
1190 !strconcat(asmop, "\t$Rd.4s, $Imm$Simm"),
1191 [(set (v4i32 VPR128:$Rd),
1192 (v4i32 (opnode (v4i32 VPR128:$src),
1193 (v4i32 (bitconvert (v4i32 (neonopnode timm:$Imm,
1194 neon_mov_imm_LSL_operand:$Simm)))))))],
1197 let cmode = {0b0, Simm{1}, Simm{0}, 0b1};
1200 // shift zeros, per halfword
1201 def _4H : NeonI_1VModImm<0b0, op,
1203 (ins VPR64:$src, neon_uimm8:$Imm,
1204 neon_mov_imm_LSLH_operand:$Simm),
1205 !strconcat(asmop, "\t$Rd.4h, $Imm$Simm"),
1206 [(set (v4i16 VPR64:$Rd),
1207 (v4i16 (opnode (v4i16 VPR64:$src),
1208 (v4i16 (bitconvert (v4i16 (neonopnode timm:$Imm,
1209 neon_mov_imm_LSL_operand:$Simm)))))))],
1212 let cmode = {0b1, 0b0, Simm, 0b1};
1215 def _8H : NeonI_1VModImm<0b1, op,
1217 (ins VPR128:$src, neon_uimm8:$Imm,
1218 neon_mov_imm_LSLH_operand:$Simm),
1219 !strconcat(asmop, "\t$Rd.8h, $Imm$Simm"),
1220 [(set (v8i16 VPR128:$Rd),
1221 (v8i16 (opnode (v8i16 VPR128:$src),
1222 (v8i16 (bitconvert (v8i16 (neonopnode timm:$Imm,
1223 neon_mov_imm_LSL_operand:$Simm)))))))],
1226 let cmode = {0b1, 0b0, Simm, 0b1};
1231 multiclass NeonI_mov_imm_msl_sizes<string asmop, bit op,
1232 SDPatternOperator opnode>
1234 // shift ones, per word
1235 def _2S : NeonI_1VModImm<0b0, op,
1237 (ins neon_uimm8:$Imm,
1238 neon_mov_imm_MSL_operand:$Simm),
1239 !strconcat(asmop, "\t$Rd.2s, $Imm$Simm"),
1240 [(set (v2i32 VPR64:$Rd),
1241 (v2i32 (opnode (timm:$Imm),
1242 (neon_mov_imm_MSL_operand:$Simm))))],
1245 let cmode = {0b1, 0b1, 0b0, Simm};
1248 def _4S : NeonI_1VModImm<0b1, op,
1250 (ins neon_uimm8:$Imm,
1251 neon_mov_imm_MSL_operand:$Simm),
1252 !strconcat(asmop, "\t$Rd.4s, $Imm$Simm"),
1253 [(set (v4i32 VPR128:$Rd),
1254 (v4i32 (opnode (timm:$Imm),
1255 (neon_mov_imm_MSL_operand:$Simm))))],
1258 let cmode = {0b1, 0b1, 0b0, Simm};
1262 // Vector Move Immediate Shifted
1263 let isReMaterializable = 1 in {
1264 defm MOVIvi_lsl : NeonI_mov_imm_lsl_sizes<"movi", 0b0, Neon_movi>;
1267 // Vector Move Inverted Immediate Shifted
1268 let isReMaterializable = 1 in {
1269 defm MVNIvi_lsl : NeonI_mov_imm_lsl_sizes<"mvni", 0b1, Neon_mvni>;
1272 // Vector Bitwise Bit Clear (AND NOT) - immediate
1273 let isReMaterializable = 1 in {
1274 defm BICvi_lsl : NeonI_mov_imm_with_constraint_lsl_sizes<"bic", 0b1,
1278 // Vector Bitwise OR - immedidate
1280 let isReMaterializable = 1 in {
1281 defm ORRvi_lsl : NeonI_mov_imm_with_constraint_lsl_sizes<"orr", 0b0,
1285 // Additional patterns for Vector Bitwise Bit Clear (AND NOT) - immedidate
1286 // LowerBUILD_VECTOR favors lowering MOVI over MVNI.
1287 // BIC immediate instructions selection requires additional patterns to
1288 // transform Neon_movi operands into BIC immediate operands
1290 def neon_mov_imm_LSLH_transform_XFORM : SDNodeXForm<imm, [{
1291 uint64_t OpCmode = N->getZExtValue();
1293 unsigned ShiftOnesIn;
1294 (void)A64Imms::decodeNeonModShiftImm(OpCmode, ShiftImm, ShiftOnesIn);
1295 // LSLH restricts shift amount to 0, 8 which are encoded as 0 and 1
1296 // Transform encoded shift amount 0 to 1 and 1 to 0.
1297 return CurDAG->getTargetConstant(!ShiftImm, MVT::i32);
1300 def neon_mov_imm_LSLH_transform_operand
1303 unsigned ShiftOnesIn;
1305 A64Imms::decodeNeonModShiftImm(Imm, ShiftImm, ShiftOnesIn);
1306 return (HasShift && !ShiftOnesIn); }],
1307 neon_mov_imm_LSLH_transform_XFORM>;
1309 // Transform (and A, (4h Neon_movi 0xff)) -> BIC 4h (A, 0x00, LSL 8)
1310 // Transform (and A, (4h Neon_movi 0xff LSL #8)) -> BIC 4h (A, 0x00)
1311 def : Pat<(v4i16 (and VPR64:$src,
1312 (v4i16 (Neon_movi 255, neon_mov_imm_LSLH_transform_operand:$Simm)))),
1313 (BICvi_lsl_4H VPR64:$src, 0,
1314 neon_mov_imm_LSLH_transform_operand:$Simm)>;
1316 // Transform (and A, (8h Neon_movi 8h 0xff)) -> BIC 8h (A, 0x00, LSL 8)
1317 // Transform (and A, (8h Neon_movi 0xff LSL #8)) -> BIC 8h (A, 0x00)
1318 def : Pat<(v8i16 (and VPR128:$src,
1319 (v8i16 (Neon_movi 255, neon_mov_imm_LSLH_transform_operand:$Simm)))),
1320 (BICvi_lsl_8H VPR128:$src, 0,
1321 neon_mov_imm_LSLH_transform_operand:$Simm)>;
1324 multiclass Neon_bitwiseVi_patterns<SDPatternOperator opnode,
1325 SDPatternOperator neonopnode,
1327 Instruction INST8H> {
1328 def : Pat<(v8i8 (opnode VPR64:$src,
1329 (bitconvert(v4i16 (neonopnode timm:$Imm,
1330 neon_mov_imm_LSLH_operand:$Simm))))),
1331 (INST4H VPR64:$src, neon_uimm8:$Imm,
1332 neon_mov_imm_LSLH_operand:$Simm)>;
1333 def : Pat<(v1i64 (opnode VPR64:$src,
1334 (bitconvert(v4i16 (neonopnode timm:$Imm,
1335 neon_mov_imm_LSLH_operand:$Simm))))),
1336 (INST4H VPR64:$src, neon_uimm8:$Imm,
1337 neon_mov_imm_LSLH_operand:$Simm)>;
1339 def : Pat<(v16i8 (opnode VPR128:$src,
1340 (bitconvert(v8i16 (neonopnode timm:$Imm,
1341 neon_mov_imm_LSLH_operand:$Simm))))),
1342 (INST8H VPR128:$src, neon_uimm8:$Imm,
1343 neon_mov_imm_LSLH_operand:$Simm)>;
1344 def : Pat<(v4i32 (opnode VPR128:$src,
1345 (bitconvert(v8i16 (neonopnode timm:$Imm,
1346 neon_mov_imm_LSLH_operand:$Simm))))),
1347 (INST8H VPR128:$src, neon_uimm8:$Imm,
1348 neon_mov_imm_LSLH_operand:$Simm)>;
1349 def : Pat<(v2i64 (opnode VPR128:$src,
1350 (bitconvert(v8i16 (neonopnode timm:$Imm,
1351 neon_mov_imm_LSLH_operand:$Simm))))),
1352 (INST8H VPR128:$src, neon_uimm8:$Imm,
1353 neon_mov_imm_LSLH_operand:$Simm)>;
1356 // Additional patterns for Vector Vector Bitwise Bit Clear (AND NOT) - immediate
1357 defm : Neon_bitwiseVi_patterns<or, Neon_mvni, BICvi_lsl_4H, BICvi_lsl_8H>;
1359 // Additional patterns for Vector Bitwise OR - immedidate
1360 defm : Neon_bitwiseVi_patterns<or, Neon_movi, ORRvi_lsl_4H, ORRvi_lsl_8H>;
1363 // Vector Move Immediate Masked
1364 let isReMaterializable = 1 in {
1365 defm MOVIvi_msl : NeonI_mov_imm_msl_sizes<"movi", 0b0, Neon_movi>;
1368 // Vector Move Inverted Immediate Masked
1369 let isReMaterializable = 1 in {
1370 defm MVNIvi_msl : NeonI_mov_imm_msl_sizes<"mvni", 0b1, Neon_mvni>;
1373 class NeonI_mov_imm_lsl_aliases<string asmop, string asmlane,
1374 Instruction inst, RegisterOperand VPRC>
1375 : NeonInstAlias<!strconcat(asmop, "\t$Rd," # asmlane # ", $Imm"),
1376 (inst VPRC:$Rd, neon_uimm8:$Imm, 0), 0b0>;
1378 // Aliases for Vector Move Immediate Shifted
1379 def : NeonI_mov_imm_lsl_aliases<"movi", ".2s", MOVIvi_lsl_2S, VPR64>;
1380 def : NeonI_mov_imm_lsl_aliases<"movi", ".4s", MOVIvi_lsl_4S, VPR128>;
1381 def : NeonI_mov_imm_lsl_aliases<"movi", ".4h", MOVIvi_lsl_4H, VPR64>;
1382 def : NeonI_mov_imm_lsl_aliases<"movi", ".8h", MOVIvi_lsl_8H, VPR128>;
1384 // Aliases for Vector Move Inverted Immediate Shifted
1385 def : NeonI_mov_imm_lsl_aliases<"mvni", ".2s", MVNIvi_lsl_2S, VPR64>;
1386 def : NeonI_mov_imm_lsl_aliases<"mvni", ".4s", MVNIvi_lsl_4S, VPR128>;
1387 def : NeonI_mov_imm_lsl_aliases<"mvni", ".4h", MVNIvi_lsl_4H, VPR64>;
1388 def : NeonI_mov_imm_lsl_aliases<"mvni", ".8h", MVNIvi_lsl_8H, VPR128>;
1390 // Aliases for Vector Bitwise Bit Clear (AND NOT) - immediate
1391 def : NeonI_mov_imm_lsl_aliases<"bic", ".2s", BICvi_lsl_2S, VPR64>;
1392 def : NeonI_mov_imm_lsl_aliases<"bic", ".4s", BICvi_lsl_4S, VPR128>;
1393 def : NeonI_mov_imm_lsl_aliases<"bic", ".4h", BICvi_lsl_4H, VPR64>;
1394 def : NeonI_mov_imm_lsl_aliases<"bic", ".8h", BICvi_lsl_8H, VPR128>;
1396 // Aliases for Vector Bitwise OR - immedidate
1397 def : NeonI_mov_imm_lsl_aliases<"orr", ".2s", ORRvi_lsl_2S, VPR64>;
1398 def : NeonI_mov_imm_lsl_aliases<"orr", ".4s", ORRvi_lsl_4S, VPR128>;
1399 def : NeonI_mov_imm_lsl_aliases<"orr", ".4h", ORRvi_lsl_4H, VPR64>;
1400 def : NeonI_mov_imm_lsl_aliases<"orr", ".8h", ORRvi_lsl_8H, VPR128>;
1402 // Vector Move Immediate - per byte
1403 let isReMaterializable = 1 in {
1404 def MOVIvi_8B : NeonI_1VModImm<0b0, 0b0,
1405 (outs VPR64:$Rd), (ins neon_uimm8:$Imm),
1406 "movi\t$Rd.8b, $Imm",
1407 [(set (v8i8 VPR64:$Rd),
1408 (v8i8 (Neon_movi (timm:$Imm), (i32 imm))))],
1413 def MOVIvi_16B : NeonI_1VModImm<0b1, 0b0,
1414 (outs VPR128:$Rd), (ins neon_uimm8:$Imm),
1415 "movi\t$Rd.16b, $Imm",
1416 [(set (v16i8 VPR128:$Rd),
1417 (v16i8 (Neon_movi (timm:$Imm), (i32 imm))))],
1423 // Vector Move Immediate - bytemask, per double word
1424 let isReMaterializable = 1 in {
1425 def MOVIvi_2D : NeonI_1VModImm<0b1, 0b1,
1426 (outs VPR128:$Rd), (ins neon_uimm64_mask:$Imm),
1427 "movi\t $Rd.2d, $Imm",
1428 [(set (v2i64 VPR128:$Rd),
1429 (v2i64 (Neon_movi (timm:$Imm), (i32 imm))))],
1435 // Vector Move Immediate - bytemask, one doubleword
1437 let isReMaterializable = 1 in {
1438 def MOVIdi : NeonI_1VModImm<0b0, 0b1,
1439 (outs FPR64:$Rd), (ins neon_uimm64_mask:$Imm),
1441 [(set (v1i64 FPR64:$Rd),
1442 (v1i64 (Neon_movi (timm:$Imm), (i32 imm))))],
1448 // Vector Floating Point Move Immediate
1450 class NeonI_FMOV_impl<string asmlane, RegisterOperand VPRC, ValueType OpTy,
1451 Operand immOpType, bit q, bit op>
1452 : NeonI_1VModImm<q, op,
1453 (outs VPRC:$Rd), (ins immOpType:$Imm),
1454 "fmov\t$Rd" # asmlane # ", $Imm",
1455 [(set (OpTy VPRC:$Rd),
1456 (OpTy (Neon_fmovi (timm:$Imm))))],
1461 let isReMaterializable = 1 in {
1462 def FMOVvi_2S : NeonI_FMOV_impl<".2s", VPR64, v2f32, fmov32_operand, 0b0, 0b0>;
1463 def FMOVvi_4S : NeonI_FMOV_impl<".4s", VPR128, v4f32, fmov32_operand, 0b1, 0b0>;
1464 def FMOVvi_2D : NeonI_FMOV_impl<".2d", VPR128, v2f64, fmov64_operand, 0b1, 0b1>;
1467 // Vector Shift (Immediate)
1468 // Immediate in [0, 63]
1469 def imm0_63 : Operand<i32> {
1470 let ParserMatchClass = uimm6_asmoperand;
1473 // Shift Right/Left Immediate - The immh:immb field of these shifts are encoded
1477 // 8 immh:immb<6:3> = '0001xxx', <imm> is encoded in immh:immb<2:0>
1478 // 16 immh:immb<6:4> = '001xxxx', <imm> is encoded in immh:immb<3:0>
1479 // 32 immh:immb<6:5> = '01xxxxx', <imm> is encoded in immh:immb<4:0>
1480 // 64 immh:immb<6> = '1xxxxxx', <imm> is encoded in immh:immb<5:0>
1482 // The shift right immediate amount, in the range 1 to element bits, is computed
1483 // as Offset - UInt(immh:immb). The shift left immediate amount, in the range 0
1484 // to element bits - 1, is computed as UInt(immh:immb) - Offset.
1486 class shr_imm_asmoperands<string OFFSET> : AsmOperandClass {
1487 let Name = "ShrImm" # OFFSET;
1488 let RenderMethod = "addImmOperands";
1489 let DiagnosticType = "ShrImm" # OFFSET;
1492 class shr_imm<string OFFSET> : Operand<i32> {
1493 let EncoderMethod = "getShiftRightImm" # OFFSET;
1494 let DecoderMethod = "DecodeShiftRightImm" # OFFSET;
1495 let ParserMatchClass =
1496 !cast<AsmOperandClass>("shr_imm" # OFFSET # "_asmoperand");
1499 def shr_imm8_asmoperand : shr_imm_asmoperands<"8">;
1500 def shr_imm16_asmoperand : shr_imm_asmoperands<"16">;
1501 def shr_imm32_asmoperand : shr_imm_asmoperands<"32">;
1502 def shr_imm64_asmoperand : shr_imm_asmoperands<"64">;
1504 def shr_imm8 : shr_imm<"8">, ImmLeaf<i32, [{return Imm > 0 && Imm <= 8;}]>;
1505 def shr_imm16 : shr_imm<"16">, ImmLeaf<i32, [{return Imm > 0 && Imm <= 16;}]>;
1506 def shr_imm32 : shr_imm<"32">, ImmLeaf<i32, [{return Imm > 0 && Imm <= 32;}]>;
1507 def shr_imm64 : shr_imm<"64">, ImmLeaf<i32, [{return Imm > 0 && Imm <= 64;}]>;
1509 class shl_imm_asmoperands<string OFFSET> : AsmOperandClass {
1510 let Name = "ShlImm" # OFFSET;
1511 let RenderMethod = "addImmOperands";
1512 let DiagnosticType = "ShlImm" # OFFSET;
1515 class shl_imm<string OFFSET> : Operand<i32> {
1516 let EncoderMethod = "getShiftLeftImm" # OFFSET;
1517 let DecoderMethod = "DecodeShiftLeftImm" # OFFSET;
1518 let ParserMatchClass =
1519 !cast<AsmOperandClass>("shl_imm" # OFFSET # "_asmoperand");
1522 def shl_imm8_asmoperand : shl_imm_asmoperands<"8">;
1523 def shl_imm16_asmoperand : shl_imm_asmoperands<"16">;
1524 def shl_imm32_asmoperand : shl_imm_asmoperands<"32">;
1525 def shl_imm64_asmoperand : shl_imm_asmoperands<"64">;
1527 def shl_imm8 : shl_imm<"8">, ImmLeaf<i32, [{return Imm >= 0 && Imm < 8;}]>;
1528 def shl_imm16 : shl_imm<"16">, ImmLeaf<i32, [{return Imm >= 0 && Imm < 16;}]>;
1529 def shl_imm32 : shl_imm<"32">, ImmLeaf<i32, [{return Imm >= 0 && Imm < 32;}]>;
1530 def shl_imm64 : shl_imm<"64">, ImmLeaf<i32, [{return Imm >= 0 && Imm < 64;}]>;
1532 class N2VShift<bit q, bit u, bits<5> opcode, string asmop, string T,
1533 RegisterOperand VPRC, ValueType Ty, Operand ImmTy, SDNode OpNode>
1534 : NeonI_2VShiftImm<q, u, opcode,
1535 (outs VPRC:$Rd), (ins VPRC:$Rn, ImmTy:$Imm),
1536 asmop # "\t$Rd." # T # ", $Rn." # T # ", $Imm",
1537 [(set (Ty VPRC:$Rd),
1538 (Ty (OpNode (Ty VPRC:$Rn),
1539 (Ty (Neon_vdup (i32 ImmTy:$Imm))))))],
1542 multiclass NeonI_N2VShL<bit u, bits<5> opcode, string asmop> {
1543 // 64-bit vector types.
1544 def _8B : N2VShift<0b0, u, opcode, asmop, "8b", VPR64, v8i8, shl_imm8, shl> {
1545 let Inst{22-19} = 0b0001; // immh:immb = 0001xxx
1548 def _4H : N2VShift<0b0, u, opcode, asmop, "4h", VPR64, v4i16, shl_imm16, shl> {
1549 let Inst{22-20} = 0b001; // immh:immb = 001xxxx
1552 def _2S : N2VShift<0b0, u, opcode, asmop, "2s", VPR64, v2i32, shl_imm32, shl> {
1553 let Inst{22-21} = 0b01; // immh:immb = 01xxxxx
1556 // 128-bit vector types.
1557 def _16B : N2VShift<0b1, u, opcode, asmop, "16b", VPR128, v16i8, shl_imm8, shl> {
1558 let Inst{22-19} = 0b0001; // immh:immb = 0001xxx
1561 def _8H : N2VShift<0b1, u, opcode, asmop, "8h", VPR128, v8i16, shl_imm16, shl> {
1562 let Inst{22-20} = 0b001; // immh:immb = 001xxxx
1565 def _4S : N2VShift<0b1, u, opcode, asmop, "4s", VPR128, v4i32, shl_imm32, shl> {
1566 let Inst{22-21} = 0b01; // immh:immb = 01xxxxx
1569 def _2D : N2VShift<0b1, u, opcode, asmop, "2d", VPR128, v2i64, shl_imm64, shl> {
1570 let Inst{22} = 0b1; // immh:immb = 1xxxxxx
1574 multiclass NeonI_N2VShR<bit u, bits<5> opcode, string asmop, SDNode OpNode> {
1575 def _8B : N2VShift<0b0, u, opcode, asmop, "8b", VPR64, v8i8, shr_imm8,
1577 let Inst{22-19} = 0b0001;
1580 def _4H : N2VShift<0b0, u, opcode, asmop, "4h", VPR64, v4i16, shr_imm16,
1582 let Inst{22-20} = 0b001;
1585 def _2S : N2VShift<0b0, u, opcode, asmop, "2s", VPR64, v2i32, shr_imm32,
1587 let Inst{22-21} = 0b01;
1590 def _16B : N2VShift<0b1, u, opcode, asmop, "16b", VPR128, v16i8, shr_imm8,
1592 let Inst{22-19} = 0b0001;
1595 def _8H : N2VShift<0b1, u, opcode, asmop, "8h", VPR128, v8i16, shr_imm16,
1597 let Inst{22-20} = 0b001;
1600 def _4S : N2VShift<0b1, u, opcode, asmop, "4s", VPR128, v4i32, shr_imm32,
1602 let Inst{22-21} = 0b01;
1605 def _2D : N2VShift<0b1, u, opcode, asmop, "2d", VPR128, v2i64, shr_imm64,
1612 defm SHLvvi : NeonI_N2VShL<0b0, 0b01010, "shl">;
1615 defm SSHRvvi : NeonI_N2VShR<0b0, 0b00000, "sshr", sra>;
1616 defm USHRvvi : NeonI_N2VShR<0b1, 0b00000, "ushr", srl>;
1618 def Neon_High16B : PatFrag<(ops node:$in),
1619 (extract_subvector (v16i8 node:$in), (iPTR 8))>;
1620 def Neon_High8H : PatFrag<(ops node:$in),
1621 (extract_subvector (v8i16 node:$in), (iPTR 4))>;
1622 def Neon_High4S : PatFrag<(ops node:$in),
1623 (extract_subvector (v4i32 node:$in), (iPTR 2))>;
1624 def Neon_High2D : PatFrag<(ops node:$in),
1625 (extract_subvector (v2i64 node:$in), (iPTR 1))>;
1626 def Neon_High4float : PatFrag<(ops node:$in),
1627 (extract_subvector (v4f32 node:$in), (iPTR 2))>;
1628 def Neon_High2double : PatFrag<(ops node:$in),
1629 (extract_subvector (v2f64 node:$in), (iPTR 1))>;
1631 def Neon_Low16B : PatFrag<(ops node:$in),
1632 (v8i8 (extract_subvector (v16i8 node:$in),
1634 def Neon_Low8H : PatFrag<(ops node:$in),
1635 (v4i16 (extract_subvector (v8i16 node:$in),
1637 def Neon_Low4S : PatFrag<(ops node:$in),
1638 (v2i32 (extract_subvector (v4i32 node:$in),
1640 def Neon_Low2D : PatFrag<(ops node:$in),
1641 (v1i64 (extract_subvector (v2i64 node:$in),
1643 def Neon_Low4float : PatFrag<(ops node:$in),
1644 (v2f32 (extract_subvector (v4f32 node:$in),
1646 def Neon_Low2double : PatFrag<(ops node:$in),
1647 (v1f64 (extract_subvector (v2f64 node:$in),
1650 class N2VShiftLong<bit q, bit u, bits<5> opcode, string asmop, string DestT,
1651 string SrcT, ValueType DestTy, ValueType SrcTy,
1652 Operand ImmTy, SDPatternOperator ExtOp>
1653 : NeonI_2VShiftImm<q, u, opcode, (outs VPR128:$Rd),
1654 (ins VPR64:$Rn, ImmTy:$Imm),
1655 asmop # "\t$Rd." # DestT # ", $Rn." # SrcT # ", $Imm",
1656 [(set (DestTy VPR128:$Rd),
1658 (DestTy (ExtOp (SrcTy VPR64:$Rn))),
1659 (DestTy (Neon_vdup (i32 ImmTy:$Imm))))))],
1662 class N2VShiftLongHigh<bit q, bit u, bits<5> opcode, string asmop, string DestT,
1663 string SrcT, ValueType DestTy, ValueType SrcTy,
1664 int StartIndex, Operand ImmTy,
1665 SDPatternOperator ExtOp, PatFrag getTop>
1666 : NeonI_2VShiftImm<q, u, opcode, (outs VPR128:$Rd),
1667 (ins VPR128:$Rn, ImmTy:$Imm),
1668 asmop # "2\t$Rd." # DestT # ", $Rn." # SrcT # ", $Imm",
1669 [(set (DestTy VPR128:$Rd),
1672 (SrcTy (getTop VPR128:$Rn)))),
1673 (DestTy (Neon_vdup (i32 ImmTy:$Imm))))))],
1676 multiclass NeonI_N2VShLL<string prefix, bit u, bits<5> opcode, string asmop,
1678 // 64-bit vector types.
1679 def _8B : N2VShiftLong<0b0, u, opcode, asmop, "8h", "8b", v8i16, v8i8,
1681 let Inst{22-19} = 0b0001; // immh:immb = 0001xxx
1684 def _4H : N2VShiftLong<0b0, u, opcode, asmop, "4s", "4h", v4i32, v4i16,
1686 let Inst{22-20} = 0b001; // immh:immb = 001xxxx
1689 def _2S : N2VShiftLong<0b0, u, opcode, asmop, "2d", "2s", v2i64, v2i32,
1691 let Inst{22-21} = 0b01; // immh:immb = 01xxxxx
1694 // 128-bit vector types
1695 def _16B : N2VShiftLongHigh<0b1, u, opcode, asmop, "8h", "16b", v8i16, v8i8,
1696 8, shl_imm8, ExtOp, Neon_High16B> {
1697 let Inst{22-19} = 0b0001; // immh:immb = 0001xxx
1700 def _8H : N2VShiftLongHigh<0b1, u, opcode, asmop, "4s", "8h", v4i32, v4i16,
1701 4, shl_imm16, ExtOp, Neon_High8H> {
1702 let Inst{22-20} = 0b001; // immh:immb = 001xxxx
1705 def _4S : N2VShiftLongHigh<0b1, u, opcode, asmop, "2d", "4s", v2i64, v2i32,
1706 2, shl_imm32, ExtOp, Neon_High4S> {
1707 let Inst{22-21} = 0b01; // immh:immb = 01xxxxx
1710 // Use other patterns to match when the immediate is 0.
1711 def : Pat<(v8i16 (ExtOp (v8i8 VPR64:$Rn))),
1712 (!cast<Instruction>(prefix # "_8B") VPR64:$Rn, 0)>;
1714 def : Pat<(v4i32 (ExtOp (v4i16 VPR64:$Rn))),
1715 (!cast<Instruction>(prefix # "_4H") VPR64:$Rn, 0)>;
1717 def : Pat<(v2i64 (ExtOp (v2i32 VPR64:$Rn))),
1718 (!cast<Instruction>(prefix # "_2S") VPR64:$Rn, 0)>;
1720 def : Pat<(v8i16 (ExtOp (v8i8 (Neon_High16B VPR128:$Rn)))),
1721 (!cast<Instruction>(prefix # "_16B") VPR128:$Rn, 0)>;
1723 def : Pat<(v4i32 (ExtOp (v4i16 (Neon_High8H VPR128:$Rn)))),
1724 (!cast<Instruction>(prefix # "_8H") VPR128:$Rn, 0)>;
1726 def : Pat<(v2i64 (ExtOp (v2i32 (Neon_High4S VPR128:$Rn)))),
1727 (!cast<Instruction>(prefix # "_4S") VPR128:$Rn, 0)>;
1731 defm SSHLLvvi : NeonI_N2VShLL<"SSHLLvvi", 0b0, 0b10100, "sshll", sext>;
1732 defm USHLLvvi : NeonI_N2VShLL<"USHLLvvi", 0b1, 0b10100, "ushll", zext>;
1734 // Rounding/Saturating shift
1735 class N2VShift_RQ<bit q, bit u, bits<5> opcode, string asmop, string T,
1736 RegisterOperand VPRC, ValueType Ty, Operand ImmTy,
1737 SDPatternOperator OpNode>
1738 : NeonI_2VShiftImm<q, u, opcode,
1739 (outs VPRC:$Rd), (ins VPRC:$Rn, ImmTy:$Imm),
1740 asmop # "\t$Rd." # T # ", $Rn." # T # ", $Imm",
1741 [(set (Ty VPRC:$Rd), (Ty (OpNode (Ty VPRC:$Rn),
1742 (i32 ImmTy:$Imm))))],
1745 // shift right (vector by immediate)
1746 multiclass NeonI_N2VShR_RQ<bit u, bits<5> opcode, string asmop,
1747 SDPatternOperator OpNode> {
1748 def _8B : N2VShift_RQ<0b0, u, opcode, asmop, "8b", VPR64, v8i8, shr_imm8,
1750 let Inst{22-19} = 0b0001;
1753 def _4H : N2VShift_RQ<0b0, u, opcode, asmop, "4h", VPR64, v4i16, shr_imm16,
1755 let Inst{22-20} = 0b001;
1758 def _2S : N2VShift_RQ<0b0, u, opcode, asmop, "2s", VPR64, v2i32, shr_imm32,
1760 let Inst{22-21} = 0b01;
1763 def _16B : N2VShift_RQ<0b1, u, opcode, asmop, "16b", VPR128, v16i8, shr_imm8,
1765 let Inst{22-19} = 0b0001;
1768 def _8H : N2VShift_RQ<0b1, u, opcode, asmop, "8h", VPR128, v8i16, shr_imm16,
1770 let Inst{22-20} = 0b001;
1773 def _4S : N2VShift_RQ<0b1, u, opcode, asmop, "4s", VPR128, v4i32, shr_imm32,
1775 let Inst{22-21} = 0b01;
1778 def _2D : N2VShift_RQ<0b1, u, opcode, asmop, "2d", VPR128, v2i64, shr_imm64,
1784 multiclass NeonI_N2VShL_Q<bit u, bits<5> opcode, string asmop,
1785 SDPatternOperator OpNode> {
1786 // 64-bit vector types.
1787 def _8B : N2VShift_RQ<0b0, u, opcode, asmop, "8b", VPR64, v8i8, shl_imm8,
1789 let Inst{22-19} = 0b0001;
1792 def _4H : N2VShift_RQ<0b0, u, opcode, asmop, "4h", VPR64, v4i16, shl_imm16,
1794 let Inst{22-20} = 0b001;
1797 def _2S : N2VShift_RQ<0b0, u, opcode, asmop, "2s", VPR64, v2i32, shl_imm32,
1799 let Inst{22-21} = 0b01;
1802 // 128-bit vector types.
1803 def _16B : N2VShift_RQ<0b1, u, opcode, asmop, "16b", VPR128, v16i8, shl_imm8,
1805 let Inst{22-19} = 0b0001;
1808 def _8H : N2VShift_RQ<0b1, u, opcode, asmop, "8h", VPR128, v8i16, shl_imm16,
1810 let Inst{22-20} = 0b001;
1813 def _4S : N2VShift_RQ<0b1, u, opcode, asmop, "4s", VPR128, v4i32, shl_imm32,
1815 let Inst{22-21} = 0b01;
1818 def _2D : N2VShift_RQ<0b1, u, opcode, asmop, "2d", VPR128, v2i64, shl_imm64,
1824 // Rounding shift right
1825 defm SRSHRvvi : NeonI_N2VShR_RQ<0b0, 0b00100, "srshr",
1826 int_aarch64_neon_vsrshr>;
1827 defm URSHRvvi : NeonI_N2VShR_RQ<0b1, 0b00100, "urshr",
1828 int_aarch64_neon_vurshr>;
1830 // Saturating shift left unsigned
1831 defm SQSHLUvvi : NeonI_N2VShL_Q<0b1, 0b01100, "sqshlu", int_aarch64_neon_vsqshlu>;
1833 // Saturating shift left
1834 defm SQSHLvvi : NeonI_N2VShL_Q<0b0, 0b01110, "sqshl", Neon_sqrshlImm>;
1835 defm UQSHLvvi : NeonI_N2VShL_Q<0b1, 0b01110, "uqshl", Neon_uqrshlImm>;
1837 class N2VShiftAdd<bit q, bit u, bits<5> opcode, string asmop, string T,
1838 RegisterOperand VPRC, ValueType Ty, Operand ImmTy,
1840 : NeonI_2VShiftImm<q, u, opcode,
1841 (outs VPRC:$Rd), (ins VPRC:$src, VPRC:$Rn, ImmTy:$Imm),
1842 asmop # "\t$Rd." # T # ", $Rn." # T # ", $Imm",
1843 [(set (Ty VPRC:$Rd), (Ty (add (Ty VPRC:$src),
1844 (Ty (OpNode (Ty VPRC:$Rn),
1845 (Ty (Neon_vdup (i32 ImmTy:$Imm))))))))],
1847 let Constraints = "$src = $Rd";
1850 // Shift Right accumulate
1851 multiclass NeonI_N2VShRAdd<bit u, bits<5> opcode, string asmop, SDNode OpNode> {
1852 def _8B : N2VShiftAdd<0b0, u, opcode, asmop, "8b", VPR64, v8i8, shr_imm8,
1854 let Inst{22-19} = 0b0001;
1857 def _4H : N2VShiftAdd<0b0, u, opcode, asmop, "4h", VPR64, v4i16, shr_imm16,
1859 let Inst{22-20} = 0b001;
1862 def _2S : N2VShiftAdd<0b0, u, opcode, asmop, "2s", VPR64, v2i32, shr_imm32,
1864 let Inst{22-21} = 0b01;
1867 def _16B : N2VShiftAdd<0b1, u, opcode, asmop, "16b", VPR128, v16i8, shr_imm8,
1869 let Inst{22-19} = 0b0001;
1872 def _8H : N2VShiftAdd<0b1, u, opcode, asmop, "8h", VPR128, v8i16, shr_imm16,
1874 let Inst{22-20} = 0b001;
1877 def _4S : N2VShiftAdd<0b1, u, opcode, asmop, "4s", VPR128, v4i32, shr_imm32,
1879 let Inst{22-21} = 0b01;
1882 def _2D : N2VShiftAdd<0b1, u, opcode, asmop, "2d", VPR128, v2i64, shr_imm64,
1888 // Shift right and accumulate
1889 defm SSRAvvi : NeonI_N2VShRAdd<0, 0b00010, "ssra", sra>;
1890 defm USRAvvi : NeonI_N2VShRAdd<1, 0b00010, "usra", srl>;
1892 // Rounding shift accumulate
1893 class N2VShiftAdd_R<bit q, bit u, bits<5> opcode, string asmop, string T,
1894 RegisterOperand VPRC, ValueType Ty, Operand ImmTy,
1895 SDPatternOperator OpNode>
1896 : NeonI_2VShiftImm<q, u, opcode,
1897 (outs VPRC:$Rd), (ins VPRC:$src, VPRC:$Rn, ImmTy:$Imm),
1898 asmop # "\t$Rd." # T # ", $Rn." # T # ", $Imm",
1899 [(set (Ty VPRC:$Rd), (Ty (add (Ty VPRC:$src),
1900 (Ty (OpNode (Ty VPRC:$Rn), (i32 ImmTy:$Imm))))))],
1902 let Constraints = "$src = $Rd";
1905 multiclass NeonI_N2VShRAdd_R<bit u, bits<5> opcode, string asmop,
1906 SDPatternOperator OpNode> {
1907 def _8B : N2VShiftAdd_R<0b0, u, opcode, asmop, "8b", VPR64, v8i8, shr_imm8,
1909 let Inst{22-19} = 0b0001;
1912 def _4H : N2VShiftAdd_R<0b0, u, opcode, asmop, "4h", VPR64, v4i16, shr_imm16,
1914 let Inst{22-20} = 0b001;
1917 def _2S : N2VShiftAdd_R<0b0, u, opcode, asmop, "2s", VPR64, v2i32, shr_imm32,
1919 let Inst{22-21} = 0b01;
1922 def _16B : N2VShiftAdd_R<0b1, u, opcode, asmop, "16b", VPR128, v16i8, shr_imm8,
1924 let Inst{22-19} = 0b0001;
1927 def _8H : N2VShiftAdd_R<0b1, u, opcode, asmop, "8h", VPR128, v8i16, shr_imm16,
1929 let Inst{22-20} = 0b001;
1932 def _4S : N2VShiftAdd_R<0b1, u, opcode, asmop, "4s", VPR128, v4i32, shr_imm32,
1934 let Inst{22-21} = 0b01;
1937 def _2D : N2VShiftAdd_R<0b1, u, opcode, asmop, "2d", VPR128, v2i64, shr_imm64,
1943 // Rounding shift right and accumulate
1944 defm SRSRAvvi : NeonI_N2VShRAdd_R<0, 0b00110, "srsra", int_aarch64_neon_vsrshr>;
1945 defm URSRAvvi : NeonI_N2VShRAdd_R<1, 0b00110, "ursra", int_aarch64_neon_vurshr>;
1947 // Shift insert by immediate
1948 class N2VShiftIns<bit q, bit u, bits<5> opcode, string asmop, string T,
1949 RegisterOperand VPRC, ValueType Ty, Operand ImmTy,
1950 SDPatternOperator OpNode>
1951 : NeonI_2VShiftImm<q, u, opcode,
1952 (outs VPRC:$Rd), (ins VPRC:$src, VPRC:$Rn, ImmTy:$Imm),
1953 asmop # "\t$Rd." # T # ", $Rn." # T # ", $Imm",
1954 [(set (Ty VPRC:$Rd), (Ty (OpNode (Ty VPRC:$src), (Ty VPRC:$Rn),
1955 (i32 ImmTy:$Imm))))],
1957 let Constraints = "$src = $Rd";
1960 // shift left insert (vector by immediate)
1961 multiclass NeonI_N2VShLIns<bit u, bits<5> opcode, string asmop> {
1962 def _8B : N2VShiftIns<0b0, u, opcode, asmop, "8b", VPR64, v8i8, shl_imm8,
1963 int_aarch64_neon_vsli> {
1964 let Inst{22-19} = 0b0001;
1967 def _4H : N2VShiftIns<0b0, u, opcode, asmop, "4h", VPR64, v4i16, shl_imm16,
1968 int_aarch64_neon_vsli> {
1969 let Inst{22-20} = 0b001;
1972 def _2S : N2VShiftIns<0b0, u, opcode, asmop, "2s", VPR64, v2i32, shl_imm32,
1973 int_aarch64_neon_vsli> {
1974 let Inst{22-21} = 0b01;
1977 // 128-bit vector types
1978 def _16B : N2VShiftIns<0b1, u, opcode, asmop, "16b", VPR128, v16i8, shl_imm8,
1979 int_aarch64_neon_vsli> {
1980 let Inst{22-19} = 0b0001;
1983 def _8H : N2VShiftIns<0b1, u, opcode, asmop, "8h", VPR128, v8i16, shl_imm16,
1984 int_aarch64_neon_vsli> {
1985 let Inst{22-20} = 0b001;
1988 def _4S : N2VShiftIns<0b1, u, opcode, asmop, "4s", VPR128, v4i32, shl_imm32,
1989 int_aarch64_neon_vsli> {
1990 let Inst{22-21} = 0b01;
1993 def _2D : N2VShiftIns<0b1, u, opcode, asmop, "2d", VPR128, v2i64, shl_imm64,
1994 int_aarch64_neon_vsli> {
1999 // shift right insert (vector by immediate)
2000 multiclass NeonI_N2VShRIns<bit u, bits<5> opcode, string asmop> {
2001 // 64-bit vector types.
2002 def _8B : N2VShiftIns<0b0, u, opcode, asmop, "8b", VPR64, v8i8, shr_imm8,
2003 int_aarch64_neon_vsri> {
2004 let Inst{22-19} = 0b0001;
2007 def _4H : N2VShiftIns<0b0, u, opcode, asmop, "4h", VPR64, v4i16, shr_imm16,
2008 int_aarch64_neon_vsri> {
2009 let Inst{22-20} = 0b001;
2012 def _2S : N2VShiftIns<0b0, u, opcode, asmop, "2s", VPR64, v2i32, shr_imm32,
2013 int_aarch64_neon_vsri> {
2014 let Inst{22-21} = 0b01;
2017 // 128-bit vector types
2018 def _16B : N2VShiftIns<0b1, u, opcode, asmop, "16b", VPR128, v16i8, shr_imm8,
2019 int_aarch64_neon_vsri> {
2020 let Inst{22-19} = 0b0001;
2023 def _8H : N2VShiftIns<0b1, u, opcode, asmop, "8h", VPR128, v8i16, shr_imm16,
2024 int_aarch64_neon_vsri> {
2025 let Inst{22-20} = 0b001;
2028 def _4S : N2VShiftIns<0b1, u, opcode, asmop, "4s", VPR128, v4i32, shr_imm32,
2029 int_aarch64_neon_vsri> {
2030 let Inst{22-21} = 0b01;
2033 def _2D : N2VShiftIns<0b1, u, opcode, asmop, "2d", VPR128, v2i64, shr_imm64,
2034 int_aarch64_neon_vsri> {
2039 // Shift left and insert
2040 defm SLIvvi : NeonI_N2VShLIns<0b1, 0b01010, "sli">;
2042 // Shift right and insert
2043 defm SRIvvi : NeonI_N2VShRIns<0b1, 0b01000, "sri">;
2045 class N2VShR_Narrow<bit q, bit u, bits<5> opcode, string asmop, string DestT,
2046 string SrcT, Operand ImmTy>
2047 : NeonI_2VShiftImm<q, u, opcode,
2048 (outs VPR64:$Rd), (ins VPR128:$Rn, ImmTy:$Imm),
2049 asmop # "\t$Rd." # DestT # ", $Rn." # SrcT # ", $Imm",
2052 class N2VShR_Narrow_Hi<bit q, bit u, bits<5> opcode, string asmop, string DestT,
2053 string SrcT, Operand ImmTy>
2054 : NeonI_2VShiftImm<q, u, opcode, (outs VPR128:$Rd),
2055 (ins VPR128:$src, VPR128:$Rn, ImmTy:$Imm),
2056 asmop # "\t$Rd." # DestT # ", $Rn." # SrcT # ", $Imm",
2058 let Constraints = "$src = $Rd";
2061 // left long shift by immediate
2062 multiclass NeonI_N2VShR_Narrow<bit u, bits<5> opcode, string asmop> {
2063 def _8B : N2VShR_Narrow<0b0, u, opcode, asmop, "8b", "8h", shr_imm8> {
2064 let Inst{22-19} = 0b0001;
2067 def _4H : N2VShR_Narrow<0b0, u, opcode, asmop, "4h", "4s", shr_imm16> {
2068 let Inst{22-20} = 0b001;
2071 def _2S : N2VShR_Narrow<0b0, u, opcode, asmop, "2s", "2d", shr_imm32> {
2072 let Inst{22-21} = 0b01;
2075 // Shift Narrow High
2076 def _16B : N2VShR_Narrow_Hi<0b1, u, opcode, asmop # "2", "16b", "8h",
2078 let Inst{22-19} = 0b0001;
2081 def _8H : N2VShR_Narrow_Hi<0b1, u, opcode, asmop # "2", "8h", "4s",
2083 let Inst{22-20} = 0b001;
2086 def _4S : N2VShR_Narrow_Hi<0b1, u, opcode, asmop # "2", "4s", "2d",
2088 let Inst{22-21} = 0b01;
2092 // Shift right narrow
2093 defm SHRNvvi : NeonI_N2VShR_Narrow<0b0, 0b10000, "shrn">;
2095 // Shift right narrow (prefix Q is saturating, prefix R is rounding)
2096 defm QSHRUNvvi :NeonI_N2VShR_Narrow<0b1, 0b10000, "sqshrun">;
2097 defm RSHRNvvi : NeonI_N2VShR_Narrow<0b0, 0b10001, "rshrn">;
2098 defm QRSHRUNvvi : NeonI_N2VShR_Narrow<0b1, 0b10001, "sqrshrun">;
2099 defm SQSHRNvvi : NeonI_N2VShR_Narrow<0b0, 0b10010, "sqshrn">;
2100 defm UQSHRNvvi : NeonI_N2VShR_Narrow<0b1, 0b10010, "uqshrn">;
2101 defm SQRSHRNvvi : NeonI_N2VShR_Narrow<0b0, 0b10011, "sqrshrn">;
2102 defm UQRSHRNvvi : NeonI_N2VShR_Narrow<0b1, 0b10011, "uqrshrn">;
2104 def Neon_combine_2D : PatFrag<(ops node:$Rm, node:$Rn),
2105 (v2i64 (concat_vectors (v1i64 node:$Rm),
2106 (v1i64 node:$Rn)))>;
2107 def Neon_combine_8H : PatFrag<(ops node:$Rm, node:$Rn),
2108 (v8i16 (concat_vectors (v4i16 node:$Rm),
2109 (v4i16 node:$Rn)))>;
2110 def Neon_combine_4S : PatFrag<(ops node:$Rm, node:$Rn),
2111 (v4i32 (concat_vectors (v2i32 node:$Rm),
2112 (v2i32 node:$Rn)))>;
2113 def Neon_combine_4f : PatFrag<(ops node:$Rm, node:$Rn),
2114 (v4f32 (concat_vectors (v2f32 node:$Rm),
2115 (v2f32 node:$Rn)))>;
2116 def Neon_combine_2d : PatFrag<(ops node:$Rm, node:$Rn),
2117 (v2f64 (concat_vectors (v1f64 node:$Rm),
2118 (v1f64 node:$Rn)))>;
2120 def Neon_lshrImm8H : PatFrag<(ops node:$lhs, node:$rhs),
2121 (v8i16 (srl (v8i16 node:$lhs),
2122 (v8i16 (Neon_vdup (i32 node:$rhs)))))>;
2123 def Neon_lshrImm4S : PatFrag<(ops node:$lhs, node:$rhs),
2124 (v4i32 (srl (v4i32 node:$lhs),
2125 (v4i32 (Neon_vdup (i32 node:$rhs)))))>;
2126 def Neon_lshrImm2D : PatFrag<(ops node:$lhs, node:$rhs),
2127 (v2i64 (srl (v2i64 node:$lhs),
2128 (v2i64 (Neon_vdup (i32 node:$rhs)))))>;
2129 def Neon_ashrImm8H : PatFrag<(ops node:$lhs, node:$rhs),
2130 (v8i16 (sra (v8i16 node:$lhs),
2131 (v8i16 (Neon_vdup (i32 node:$rhs)))))>;
2132 def Neon_ashrImm4S : PatFrag<(ops node:$lhs, node:$rhs),
2133 (v4i32 (sra (v4i32 node:$lhs),
2134 (v4i32 (Neon_vdup (i32 node:$rhs)))))>;
2135 def Neon_ashrImm2D : PatFrag<(ops node:$lhs, node:$rhs),
2136 (v2i64 (sra (v2i64 node:$lhs),
2137 (v2i64 (Neon_vdup (i32 node:$rhs)))))>;
2139 // Normal shift right narrow is matched by IR (srl/sra, trunc, concat_vectors)
2140 multiclass Neon_shiftNarrow_patterns<string shr> {
2141 def : Pat<(v8i8 (trunc (!cast<PatFrag>("Neon_" # shr # "Imm8H") VPR128:$Rn,
2142 (i32 shr_imm8:$Imm)))),
2143 (SHRNvvi_8B VPR128:$Rn, imm:$Imm)>;
2144 def : Pat<(v4i16 (trunc (!cast<PatFrag>("Neon_" # shr # "Imm4S") VPR128:$Rn,
2145 (i32 shr_imm16:$Imm)))),
2146 (SHRNvvi_4H VPR128:$Rn, imm:$Imm)>;
2147 def : Pat<(v2i32 (trunc (!cast<PatFrag>("Neon_" # shr # "Imm2D") VPR128:$Rn,
2148 (i32 shr_imm32:$Imm)))),
2149 (SHRNvvi_2S VPR128:$Rn, imm:$Imm)>;
2151 def : Pat<(Neon_combine_2D (v1i64 VPR64:$src), (v1i64 (bitconvert
2152 (v8i8 (trunc (!cast<PatFrag>("Neon_" # shr # "Imm8H")
2153 VPR128:$Rn, (i32 shr_imm8:$Imm))))))),
2154 (SHRNvvi_16B (v2i64 (SUBREG_TO_REG (i64 0), VPR64:$src, sub_64)),
2155 VPR128:$Rn, imm:$Imm)>;
2156 def : Pat<(Neon_combine_2D (v1i64 VPR64:$src), (v1i64 (bitconvert
2157 (v4i16 (trunc (!cast<PatFrag>("Neon_" # shr # "Imm4S")
2158 VPR128:$Rn, (i32 shr_imm16:$Imm))))))),
2159 (SHRNvvi_8H (SUBREG_TO_REG (i64 0), VPR64:$src, sub_64),
2160 VPR128:$Rn, imm:$Imm)>;
2161 def : Pat<(Neon_combine_2D (v1i64 VPR64:$src), (v1i64 (bitconvert
2162 (v2i32 (trunc (!cast<PatFrag>("Neon_" # shr # "Imm2D")
2163 VPR128:$Rn, (i32 shr_imm32:$Imm))))))),
2164 (SHRNvvi_4S (SUBREG_TO_REG (i64 0), VPR64:$src, sub_64),
2165 VPR128:$Rn, imm:$Imm)>;
2168 multiclass Neon_shiftNarrow_QR_patterns<SDPatternOperator op, string prefix> {
2169 def : Pat<(v8i8 (op (v8i16 VPR128:$Rn), shr_imm8:$Imm)),
2170 (!cast<Instruction>(prefix # "_8B") VPR128:$Rn, imm:$Imm)>;
2171 def : Pat<(v4i16 (op (v4i32 VPR128:$Rn), shr_imm16:$Imm)),
2172 (!cast<Instruction>(prefix # "_4H") VPR128:$Rn, imm:$Imm)>;
2173 def : Pat<(v2i32 (op (v2i64 VPR128:$Rn), shr_imm32:$Imm)),
2174 (!cast<Instruction>(prefix # "_2S") VPR128:$Rn, imm:$Imm)>;
2176 def : Pat<(Neon_combine_2D (v1i64 VPR64:$src),
2177 (v1i64 (bitconvert (v8i8
2178 (op (v8i16 VPR128:$Rn), shr_imm8:$Imm))))),
2179 (!cast<Instruction>(prefix # "_16B")
2180 (SUBREG_TO_REG (i64 0), VPR64:$src, sub_64),
2181 VPR128:$Rn, imm:$Imm)>;
2182 def : Pat<(Neon_combine_2D (v1i64 VPR64:$src),
2183 (v1i64 (bitconvert (v4i16
2184 (op (v4i32 VPR128:$Rn), shr_imm16:$Imm))))),
2185 (!cast<Instruction>(prefix # "_8H")
2186 (SUBREG_TO_REG (i64 0), VPR64:$src, sub_64),
2187 VPR128:$Rn, imm:$Imm)>;
2188 def : Pat<(Neon_combine_2D (v1i64 VPR64:$src),
2189 (v1i64 (bitconvert (v2i32
2190 (op (v2i64 VPR128:$Rn), shr_imm32:$Imm))))),
2191 (!cast<Instruction>(prefix # "_4S")
2192 (SUBREG_TO_REG (i64 0), VPR64:$src, sub_64),
2193 VPR128:$Rn, imm:$Imm)>;
2196 defm : Neon_shiftNarrow_patterns<"lshr">;
2197 defm : Neon_shiftNarrow_patterns<"ashr">;
2199 defm : Neon_shiftNarrow_QR_patterns<int_aarch64_neon_vsqshrun, "QSHRUNvvi">;
2200 defm : Neon_shiftNarrow_QR_patterns<int_aarch64_neon_vrshrn, "RSHRNvvi">;
2201 defm : Neon_shiftNarrow_QR_patterns<int_aarch64_neon_vsqrshrun, "QRSHRUNvvi">;
2202 defm : Neon_shiftNarrow_QR_patterns<int_aarch64_neon_vsqshrn, "SQSHRNvvi">;
2203 defm : Neon_shiftNarrow_QR_patterns<int_aarch64_neon_vuqshrn, "UQSHRNvvi">;
2204 defm : Neon_shiftNarrow_QR_patterns<int_aarch64_neon_vsqrshrn, "SQRSHRNvvi">;
2205 defm : Neon_shiftNarrow_QR_patterns<int_aarch64_neon_vuqrshrn, "UQRSHRNvvi">;
2207 // Convert fix-point and float-pointing
2208 class N2VCvt_Fx<bit q, bit u, bits<5> opcode, string asmop, string T,
2209 RegisterOperand VPRC, ValueType DestTy, ValueType SrcTy,
2210 Operand ImmTy, SDPatternOperator IntOp>
2211 : NeonI_2VShiftImm<q, u, opcode,
2212 (outs VPRC:$Rd), (ins VPRC:$Rn, ImmTy:$Imm),
2213 asmop # "\t$Rd." # T # ", $Rn." # T # ", $Imm",
2214 [(set (DestTy VPRC:$Rd), (DestTy (IntOp (SrcTy VPRC:$Rn),
2215 (i32 ImmTy:$Imm))))],
2218 multiclass NeonI_N2VCvt_Fx2fp<bit u, bits<5> opcode, string asmop,
2219 SDPatternOperator IntOp> {
2220 def _2S : N2VCvt_Fx<0, u, opcode, asmop, "2s", VPR64, v2f32, v2i32,
2222 let Inst{22-21} = 0b01;
2225 def _4S : N2VCvt_Fx<1, u, opcode, asmop, "4s", VPR128, v4f32, v4i32,
2227 let Inst{22-21} = 0b01;
2230 def _2D : N2VCvt_Fx<1, u, opcode, asmop, "2d", VPR128, v2f64, v2i64,
2236 multiclass NeonI_N2VCvt_Fp2fx<bit u, bits<5> opcode, string asmop,
2237 SDPatternOperator IntOp> {
2238 def _2S : N2VCvt_Fx<0, u, opcode, asmop, "2s", VPR64, v2i32, v2f32,
2240 let Inst{22-21} = 0b01;
2243 def _4S : N2VCvt_Fx<1, u, opcode, asmop, "4s", VPR128, v4i32, v4f32,
2245 let Inst{22-21} = 0b01;
2248 def _2D : N2VCvt_Fx<1, u, opcode, asmop, "2d", VPR128, v2i64, v2f64,
2254 // Convert fixed-point to floating-point
2255 defm VCVTxs2f : NeonI_N2VCvt_Fx2fp<0, 0b11100, "scvtf",
2256 int_arm_neon_vcvtfxs2fp>;
2257 defm VCVTxu2f : NeonI_N2VCvt_Fx2fp<1, 0b11100, "ucvtf",
2258 int_arm_neon_vcvtfxu2fp>;
2260 // Convert floating-point to fixed-point
2261 defm VCVTf2xs : NeonI_N2VCvt_Fp2fx<0, 0b11111, "fcvtzs",
2262 int_arm_neon_vcvtfp2fxs>;
2263 defm VCVTf2xu : NeonI_N2VCvt_Fp2fx<1, 0b11111, "fcvtzu",
2264 int_arm_neon_vcvtfp2fxu>;
2266 multiclass Neon_sshll2_0<SDNode ext>
2268 def _v8i8 : PatFrag<(ops node:$Rn),
2269 (v8i16 (ext (v8i8 (Neon_High16B node:$Rn))))>;
2270 def _v4i16 : PatFrag<(ops node:$Rn),
2271 (v4i32 (ext (v4i16 (Neon_High8H node:$Rn))))>;
2272 def _v2i32 : PatFrag<(ops node:$Rn),
2273 (v2i64 (ext (v2i32 (Neon_High4S node:$Rn))))>;
2276 defm NI_sext_high : Neon_sshll2_0<sext>;
2277 defm NI_zext_high : Neon_sshll2_0<zext>;
2280 //===----------------------------------------------------------------------===//
2281 // Multiclasses for NeonI_Across
2282 //===----------------------------------------------------------------------===//
2286 multiclass NeonI_2VAcross_1<bit u, bits<5> opcode,
2287 string asmop, SDPatternOperator opnode>
2289 def _1h8b: NeonI_2VAcross<0b0, u, 0b00, opcode,
2290 (outs FPR16:$Rd), (ins VPR64:$Rn),
2291 asmop # "\t$Rd, $Rn.8b",
2292 [(set (v1i16 FPR16:$Rd),
2293 (v1i16 (opnode (v8i8 VPR64:$Rn))))],
2296 def _1h16b: NeonI_2VAcross<0b1, u, 0b00, opcode,
2297 (outs FPR16:$Rd), (ins VPR128:$Rn),
2298 asmop # "\t$Rd, $Rn.16b",
2299 [(set (v1i16 FPR16:$Rd),
2300 (v1i16 (opnode (v16i8 VPR128:$Rn))))],
2303 def _1s4h: NeonI_2VAcross<0b0, u, 0b01, opcode,
2304 (outs FPR32:$Rd), (ins VPR64:$Rn),
2305 asmop # "\t$Rd, $Rn.4h",
2306 [(set (v1i32 FPR32:$Rd),
2307 (v1i32 (opnode (v4i16 VPR64:$Rn))))],
2310 def _1s8h: NeonI_2VAcross<0b1, u, 0b01, opcode,
2311 (outs FPR32:$Rd), (ins VPR128:$Rn),
2312 asmop # "\t$Rd, $Rn.8h",
2313 [(set (v1i32 FPR32:$Rd),
2314 (v1i32 (opnode (v8i16 VPR128:$Rn))))],
2317 // _1d2s doesn't exist!
2319 def _1d4s: NeonI_2VAcross<0b1, u, 0b10, opcode,
2320 (outs FPR64:$Rd), (ins VPR128:$Rn),
2321 asmop # "\t$Rd, $Rn.4s",
2322 [(set (v1i64 FPR64:$Rd),
2323 (v1i64 (opnode (v4i32 VPR128:$Rn))))],
2327 defm SADDLV : NeonI_2VAcross_1<0b0, 0b00011, "saddlv", int_aarch64_neon_saddlv>;
2328 defm UADDLV : NeonI_2VAcross_1<0b1, 0b00011, "uaddlv", int_aarch64_neon_uaddlv>;
2332 multiclass NeonI_2VAcross_2<bit u, bits<5> opcode,
2333 string asmop, SDPatternOperator opnode>
2335 def _1b8b: NeonI_2VAcross<0b0, u, 0b00, opcode,
2336 (outs FPR8:$Rd), (ins VPR64:$Rn),
2337 asmop # "\t$Rd, $Rn.8b",
2338 [(set (v1i8 FPR8:$Rd),
2339 (v1i8 (opnode (v8i8 VPR64:$Rn))))],
2342 def _1b16b: NeonI_2VAcross<0b1, u, 0b00, opcode,
2343 (outs FPR8:$Rd), (ins VPR128:$Rn),
2344 asmop # "\t$Rd, $Rn.16b",
2345 [(set (v1i8 FPR8:$Rd),
2346 (v1i8 (opnode (v16i8 VPR128:$Rn))))],
2349 def _1h4h: NeonI_2VAcross<0b0, u, 0b01, opcode,
2350 (outs FPR16:$Rd), (ins VPR64:$Rn),
2351 asmop # "\t$Rd, $Rn.4h",
2352 [(set (v1i16 FPR16:$Rd),
2353 (v1i16 (opnode (v4i16 VPR64:$Rn))))],
2356 def _1h8h: NeonI_2VAcross<0b1, u, 0b01, opcode,
2357 (outs FPR16:$Rd), (ins VPR128:$Rn),
2358 asmop # "\t$Rd, $Rn.8h",
2359 [(set (v1i16 FPR16:$Rd),
2360 (v1i16 (opnode (v8i16 VPR128:$Rn))))],
2363 // _1s2s doesn't exist!
2365 def _1s4s: NeonI_2VAcross<0b1, u, 0b10, opcode,
2366 (outs FPR32:$Rd), (ins VPR128:$Rn),
2367 asmop # "\t$Rd, $Rn.4s",
2368 [(set (v1i32 FPR32:$Rd),
2369 (v1i32 (opnode (v4i32 VPR128:$Rn))))],
2373 defm SMAXV : NeonI_2VAcross_2<0b0, 0b01010, "smaxv", int_aarch64_neon_smaxv>;
2374 defm UMAXV : NeonI_2VAcross_2<0b1, 0b01010, "umaxv", int_aarch64_neon_umaxv>;
2376 defm SMINV : NeonI_2VAcross_2<0b0, 0b11010, "sminv", int_aarch64_neon_sminv>;
2377 defm UMINV : NeonI_2VAcross_2<0b1, 0b11010, "uminv", int_aarch64_neon_uminv>;
2379 defm ADDV : NeonI_2VAcross_2<0b0, 0b11011, "addv", int_aarch64_neon_vaddv>;
2383 multiclass NeonI_2VAcross_3<bit u, bits<5> opcode, bits<2> size,
2384 string asmop, SDPatternOperator opnode> {
2385 def _1s4s: NeonI_2VAcross<0b1, u, size, opcode,
2386 (outs FPR32:$Rd), (ins VPR128:$Rn),
2387 asmop # "\t$Rd, $Rn.4s",
2388 [(set (f32 FPR32:$Rd),
2389 (f32 (opnode (v4f32 VPR128:$Rn))))],
2393 defm FMAXNMV : NeonI_2VAcross_3<0b1, 0b01100, 0b00, "fmaxnmv",
2394 int_aarch64_neon_vmaxnmv>;
2395 defm FMINNMV : NeonI_2VAcross_3<0b1, 0b01100, 0b10, "fminnmv",
2396 int_aarch64_neon_vminnmv>;
2398 defm FMAXV : NeonI_2VAcross_3<0b1, 0b01111, 0b00, "fmaxv",
2399 int_aarch64_neon_vmaxv>;
2400 defm FMINV : NeonI_2VAcross_3<0b1, 0b01111, 0b10, "fminv",
2401 int_aarch64_neon_vminv>;
2403 // The followings are for instruction class (Perm)
2405 class NeonI_Permute<bit q, bits<2> size, bits<3> opcode,
2406 string asmop, RegisterOperand OpVPR, string OpS,
2407 SDPatternOperator opnode, ValueType Ty>
2408 : NeonI_Perm<q, size, opcode,
2409 (outs OpVPR:$Rd), (ins OpVPR:$Rn, OpVPR:$Rm),
2410 asmop # "\t$Rd." # OpS # ", $Rn." # OpS # ", $Rm." # OpS,
2411 [(set (Ty OpVPR:$Rd),
2412 (Ty (opnode (Ty OpVPR:$Rn), (Ty OpVPR:$Rm))))],
2415 multiclass NeonI_Perm_pat<bits<3> opcode, string asmop,
2416 SDPatternOperator opnode> {
2417 def _8b : NeonI_Permute<0b0, 0b00, opcode, asmop,
2418 VPR64, "8b", opnode, v8i8>;
2419 def _16b : NeonI_Permute<0b1, 0b00, opcode, asmop,
2420 VPR128, "16b",opnode, v16i8>;
2421 def _4h : NeonI_Permute<0b0, 0b01, opcode, asmop,
2422 VPR64, "4h", opnode, v4i16>;
2423 def _8h : NeonI_Permute<0b1, 0b01, opcode, asmop,
2424 VPR128, "8h", opnode, v8i16>;
2425 def _2s : NeonI_Permute<0b0, 0b10, opcode, asmop,
2426 VPR64, "2s", opnode, v2i32>;
2427 def _4s : NeonI_Permute<0b1, 0b10, opcode, asmop,
2428 VPR128, "4s", opnode, v4i32>;
2429 def _2d : NeonI_Permute<0b1, 0b11, opcode, asmop,
2430 VPR128, "2d", opnode, v2i64>;
2433 defm UZP1vvv : NeonI_Perm_pat<0b001, "uzp1", Neon_uzp1>;
2434 defm TRN1vvv : NeonI_Perm_pat<0b010, "trn1", Neon_trn1>;
2435 defm ZIP1vvv : NeonI_Perm_pat<0b011, "zip1", Neon_zip1>;
2436 defm UZP2vvv : NeonI_Perm_pat<0b101, "uzp2", Neon_uzp2>;
2437 defm TRN2vvv : NeonI_Perm_pat<0b110, "trn2", Neon_trn2>;
2438 defm ZIP2vvv : NeonI_Perm_pat<0b111, "zip2", Neon_zip2>;
2440 multiclass NeonI_Perm_float_pat<string INS, SDPatternOperator opnode> {
2441 def : Pat<(v2f32 (opnode (v2f32 VPR64:$Rn), (v2f32 VPR64:$Rm))),
2442 (!cast<Instruction>(INS # "_2s") VPR64:$Rn, VPR64:$Rm)>;
2444 def : Pat<(v4f32 (opnode (v4f32 VPR128:$Rn), (v4f32 VPR128:$Rm))),
2445 (!cast<Instruction>(INS # "_4s") VPR128:$Rn, VPR128:$Rm)>;
2447 def : Pat<(v2f64 (opnode (v2f64 VPR128:$Rn), (v2f64 VPR128:$Rm))),
2448 (!cast<Instruction>(INS # "_2d") VPR128:$Rn, VPR128:$Rm)>;
2451 defm : NeonI_Perm_float_pat<"UZP1vvv", Neon_uzp1>;
2452 defm : NeonI_Perm_float_pat<"UZP2vvv", Neon_uzp2>;
2453 defm : NeonI_Perm_float_pat<"ZIP1vvv", Neon_zip1>;
2454 defm : NeonI_Perm_float_pat<"ZIP2vvv", Neon_zip2>;
2455 defm : NeonI_Perm_float_pat<"TRN1vvv", Neon_trn1>;
2456 defm : NeonI_Perm_float_pat<"TRN2vvv", Neon_trn2>;
2458 // The followings are for instruction class (3V Diff)
2460 // normal long/long2 pattern
2461 class NeonI_3VDL<bit q, bit u, bits<2> size, bits<4> opcode,
2462 string asmop, string ResS, string OpS,
2463 SDPatternOperator opnode, SDPatternOperator ext,
2464 RegisterOperand OpVPR,
2465 ValueType ResTy, ValueType OpTy>
2466 : NeonI_3VDiff<q, u, size, opcode,
2467 (outs VPR128:$Rd), (ins OpVPR:$Rn, OpVPR:$Rm),
2468 asmop # "\t$Rd." # ResS # ", $Rn." # OpS # ", $Rm." # OpS,
2469 [(set (ResTy VPR128:$Rd),
2470 (ResTy (opnode (ResTy (ext (OpTy OpVPR:$Rn))),
2471 (ResTy (ext (OpTy OpVPR:$Rm))))))],
2474 multiclass NeonI_3VDL_s<bit u, bits<4> opcode,
2475 string asmop, SDPatternOperator opnode,
2476 bit Commutable = 0> {
2477 let isCommutable = Commutable in {
2478 def _8h8b : NeonI_3VDL<0b0, u, 0b00, opcode, asmop, "8h", "8b",
2479 opnode, sext, VPR64, v8i16, v8i8>;
2480 def _4s4h : NeonI_3VDL<0b0, u, 0b01, opcode, asmop, "4s", "4h",
2481 opnode, sext, VPR64, v4i32, v4i16>;
2482 def _2d2s : NeonI_3VDL<0b0, u, 0b10, opcode, asmop, "2d", "2s",
2483 opnode, sext, VPR64, v2i64, v2i32>;
2487 multiclass NeonI_3VDL2_s<bit u, bits<4> opcode, string asmop,
2488 SDPatternOperator opnode, bit Commutable = 0> {
2489 let isCommutable = Commutable in {
2490 def _8h16b : NeonI_3VDL<0b1, u, 0b00, opcode, asmop, "8h", "16b",
2491 opnode, NI_sext_high_v8i8, VPR128, v8i16, v16i8>;
2492 def _4s8h : NeonI_3VDL<0b1, u, 0b01, opcode, asmop, "4s", "8h",
2493 opnode, NI_sext_high_v4i16, VPR128, v4i32, v8i16>;
2494 def _2d4s : NeonI_3VDL<0b1, u, 0b10, opcode, asmop, "2d", "4s",
2495 opnode, NI_sext_high_v2i32, VPR128, v2i64, v4i32>;
2499 multiclass NeonI_3VDL_u<bit u, bits<4> opcode, string asmop,
2500 SDPatternOperator opnode, bit Commutable = 0> {
2501 let isCommutable = Commutable in {
2502 def _8h8b : NeonI_3VDL<0b0, u, 0b00, opcode, asmop, "8h", "8b",
2503 opnode, zext, VPR64, v8i16, v8i8>;
2504 def _4s4h : NeonI_3VDL<0b0, u, 0b01, opcode, asmop, "4s", "4h",
2505 opnode, zext, VPR64, v4i32, v4i16>;
2506 def _2d2s : NeonI_3VDL<0b0, u, 0b10, opcode, asmop, "2d", "2s",
2507 opnode, zext, VPR64, v2i64, v2i32>;
2511 multiclass NeonI_3VDL2_u<bit u, bits<4> opcode, string asmop,
2512 SDPatternOperator opnode, bit Commutable = 0> {
2513 let isCommutable = Commutable in {
2514 def _8h16b : NeonI_3VDL<0b1, u, 0b00, opcode, asmop, "8h", "16b",
2515 opnode, NI_zext_high_v8i8, VPR128, v8i16, v16i8>;
2516 def _4s8h : NeonI_3VDL<0b1, u, 0b01, opcode, asmop, "4s", "8h",
2517 opnode, NI_zext_high_v4i16, VPR128, v4i32, v8i16>;
2518 def _2d4s : NeonI_3VDL<0b1, u, 0b10, opcode, asmop, "2d", "4s",
2519 opnode, NI_zext_high_v2i32, VPR128, v2i64, v4i32>;
2523 defm SADDLvvv : NeonI_3VDL_s<0b0, 0b0000, "saddl", add, 1>;
2524 defm UADDLvvv : NeonI_3VDL_u<0b1, 0b0000, "uaddl", add, 1>;
2526 defm SADDL2vvv : NeonI_3VDL2_s<0b0, 0b0000, "saddl2", add, 1>;
2527 defm UADDL2vvv : NeonI_3VDL2_u<0b1, 0b0000, "uaddl2", add, 1>;
2529 defm SSUBLvvv : NeonI_3VDL_s<0b0, 0b0010, "ssubl", sub, 0>;
2530 defm USUBLvvv : NeonI_3VDL_u<0b1, 0b0010, "usubl", sub, 0>;
2532 defm SSUBL2vvv : NeonI_3VDL2_s<0b0, 0b0010, "ssubl2", sub, 0>;
2533 defm USUBL2vvv : NeonI_3VDL2_u<0b1, 0b0010, "usubl2", sub, 0>;
2535 // normal wide/wide2 pattern
2536 class NeonI_3VDW<bit q, bit u, bits<2> size, bits<4> opcode,
2537 string asmop, string ResS, string OpS,
2538 SDPatternOperator opnode, SDPatternOperator ext,
2539 RegisterOperand OpVPR,
2540 ValueType ResTy, ValueType OpTy>
2541 : NeonI_3VDiff<q, u, size, opcode,
2542 (outs VPR128:$Rd), (ins VPR128:$Rn, OpVPR:$Rm),
2543 asmop # "\t$Rd." # ResS # ", $Rn." # ResS # ", $Rm." # OpS,
2544 [(set (ResTy VPR128:$Rd),
2545 (ResTy (opnode (ResTy VPR128:$Rn),
2546 (ResTy (ext (OpTy OpVPR:$Rm))))))],
2549 multiclass NeonI_3VDW_s<bit u, bits<4> opcode, string asmop,
2550 SDPatternOperator opnode> {
2551 def _8h8b : NeonI_3VDW<0b0, u, 0b00, opcode, asmop, "8h", "8b",
2552 opnode, sext, VPR64, v8i16, v8i8>;
2553 def _4s4h : NeonI_3VDW<0b0, u, 0b01, opcode, asmop, "4s", "4h",
2554 opnode, sext, VPR64, v4i32, v4i16>;
2555 def _2d2s : NeonI_3VDW<0b0, u, 0b10, opcode, asmop, "2d", "2s",
2556 opnode, sext, VPR64, v2i64, v2i32>;
2559 defm SADDWvvv : NeonI_3VDW_s<0b0, 0b0001, "saddw", add>;
2560 defm SSUBWvvv : NeonI_3VDW_s<0b0, 0b0011, "ssubw", sub>;
2562 multiclass NeonI_3VDW2_s<bit u, bits<4> opcode, string asmop,
2563 SDPatternOperator opnode> {
2564 def _8h16b : NeonI_3VDW<0b1, u, 0b00, opcode, asmop, "8h", "16b",
2565 opnode, NI_sext_high_v8i8, VPR128, v8i16, v16i8>;
2566 def _4s8h : NeonI_3VDW<0b1, u, 0b01, opcode, asmop, "4s", "8h",
2567 opnode, NI_sext_high_v4i16, VPR128, v4i32, v8i16>;
2568 def _2d4s : NeonI_3VDW<0b1, u, 0b10, opcode, asmop, "2d", "4s",
2569 opnode, NI_sext_high_v2i32, VPR128, v2i64, v4i32>;
2572 defm SADDW2vvv : NeonI_3VDW2_s<0b0, 0b0001, "saddw2", add>;
2573 defm SSUBW2vvv : NeonI_3VDW2_s<0b0, 0b0011, "ssubw2", sub>;
2575 multiclass NeonI_3VDW_u<bit u, bits<4> opcode, string asmop,
2576 SDPatternOperator opnode> {
2577 def _8h8b : NeonI_3VDW<0b0, u, 0b00, opcode, asmop, "8h", "8b",
2578 opnode, zext, VPR64, v8i16, v8i8>;
2579 def _4s4h : NeonI_3VDW<0b0, u, 0b01, opcode, asmop, "4s", "4h",
2580 opnode, zext, VPR64, v4i32, v4i16>;
2581 def _2d2s : NeonI_3VDW<0b0, u, 0b10, opcode, asmop, "2d", "2s",
2582 opnode, zext, VPR64, v2i64, v2i32>;
2585 defm UADDWvvv : NeonI_3VDW_u<0b1, 0b0001, "uaddw", add>;
2586 defm USUBWvvv : NeonI_3VDW_u<0b1, 0b0011, "usubw", sub>;
2588 multiclass NeonI_3VDW2_u<bit u, bits<4> opcode, string asmop,
2589 SDPatternOperator opnode> {
2590 def _8h16b : NeonI_3VDW<0b1, u, 0b00, opcode, asmop, "8h", "16b",
2591 opnode, NI_zext_high_v8i8, VPR128, v8i16, v16i8>;
2592 def _4s8h : NeonI_3VDW<0b1, u, 0b01, opcode, asmop, "4s", "8h",
2593 opnode, NI_zext_high_v4i16, VPR128, v4i32, v8i16>;
2594 def _2d4s : NeonI_3VDW<0b1, u, 0b10, opcode, asmop, "2d", "4s",
2595 opnode, NI_zext_high_v2i32, VPR128, v2i64, v4i32>;
2598 defm UADDW2vvv : NeonI_3VDW2_u<0b1, 0b0001, "uaddw2", add>;
2599 defm USUBW2vvv : NeonI_3VDW2_u<0b1, 0b0011, "usubw2", sub>;
2601 // Get the high half part of the vector element.
2602 multiclass NeonI_get_high {
2603 def _8h : PatFrag<(ops node:$Rn),
2604 (v8i8 (trunc (v8i16 (srl (v8i16 node:$Rn),
2605 (v8i16 (Neon_vdup (i32 8)))))))>;
2606 def _4s : PatFrag<(ops node:$Rn),
2607 (v4i16 (trunc (v4i32 (srl (v4i32 node:$Rn),
2608 (v4i32 (Neon_vdup (i32 16)))))))>;
2609 def _2d : PatFrag<(ops node:$Rn),
2610 (v2i32 (trunc (v2i64 (srl (v2i64 node:$Rn),
2611 (v2i64 (Neon_vdup (i32 32)))))))>;
2614 defm NI_get_hi : NeonI_get_high;
2616 // pattern for addhn/subhn with 2 operands
2617 class NeonI_3VDN_addhn_2Op<bit q, bit u, bits<2> size, bits<4> opcode,
2618 string asmop, string ResS, string OpS,
2619 SDPatternOperator opnode, SDPatternOperator get_hi,
2620 ValueType ResTy, ValueType OpTy>
2621 : NeonI_3VDiff<q, u, size, opcode,
2622 (outs VPR64:$Rd), (ins VPR128:$Rn, VPR128:$Rm),
2623 asmop # "\t$Rd." # ResS # ", $Rn." # OpS # ", $Rm." # OpS,
2624 [(set (ResTy VPR64:$Rd),
2626 (OpTy (opnode (OpTy VPR128:$Rn),
2627 (OpTy VPR128:$Rm))))))],
2630 multiclass NeonI_3VDN_addhn_2Op<bit u, bits<4> opcode, string asmop,
2631 SDPatternOperator opnode, bit Commutable = 0> {
2632 let isCommutable = Commutable in {
2633 def _8b8h : NeonI_3VDN_addhn_2Op<0b0, u, 0b00, opcode, asmop, "8b", "8h",
2634 opnode, NI_get_hi_8h, v8i8, v8i16>;
2635 def _4h4s : NeonI_3VDN_addhn_2Op<0b0, u, 0b01, opcode, asmop, "4h", "4s",
2636 opnode, NI_get_hi_4s, v4i16, v4i32>;
2637 def _2s2d : NeonI_3VDN_addhn_2Op<0b0, u, 0b10, opcode, asmop, "2s", "2d",
2638 opnode, NI_get_hi_2d, v2i32, v2i64>;
2642 defm ADDHNvvv : NeonI_3VDN_addhn_2Op<0b0, 0b0100, "addhn", add, 1>;
2643 defm SUBHNvvv : NeonI_3VDN_addhn_2Op<0b0, 0b0110, "subhn", sub, 0>;
2645 // pattern for operation with 2 operands
2646 class NeonI_3VD_2Op<bit q, bit u, bits<2> size, bits<4> opcode,
2647 string asmop, string ResS, string OpS,
2648 SDPatternOperator opnode,
2649 RegisterOperand ResVPR, RegisterOperand OpVPR,
2650 ValueType ResTy, ValueType OpTy>
2651 : NeonI_3VDiff<q, u, size, opcode,
2652 (outs ResVPR:$Rd), (ins OpVPR:$Rn, OpVPR:$Rm),
2653 asmop # "\t$Rd." # ResS # ", $Rn." # OpS # ", $Rm." # OpS,
2654 [(set (ResTy ResVPR:$Rd),
2655 (ResTy (opnode (OpTy OpVPR:$Rn), (OpTy OpVPR:$Rm))))],
2658 // normal narrow pattern
2659 multiclass NeonI_3VDN_2Op<bit u, bits<4> opcode, string asmop,
2660 SDPatternOperator opnode, bit Commutable = 0> {
2661 let isCommutable = Commutable in {
2662 def _8b8h : NeonI_3VD_2Op<0b0, u, 0b00, opcode, asmop, "8b", "8h",
2663 opnode, VPR64, VPR128, v8i8, v8i16>;
2664 def _4h4s : NeonI_3VD_2Op<0b0, u, 0b01, opcode, asmop, "4h", "4s",
2665 opnode, VPR64, VPR128, v4i16, v4i32>;
2666 def _2s2d : NeonI_3VD_2Op<0b0, u, 0b10, opcode, asmop, "2s", "2d",
2667 opnode, VPR64, VPR128, v2i32, v2i64>;
2671 defm RADDHNvvv : NeonI_3VDN_2Op<0b1, 0b0100, "raddhn", int_arm_neon_vraddhn, 1>;
2672 defm RSUBHNvvv : NeonI_3VDN_2Op<0b1, 0b0110, "rsubhn", int_arm_neon_vrsubhn, 0>;
2674 // pattern for acle intrinsic with 3 operands
2675 class NeonI_3VDN_3Op<bit q, bit u, bits<2> size, bits<4> opcode,
2676 string asmop, string ResS, string OpS>
2677 : NeonI_3VDiff<q, u, size, opcode,
2678 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn, VPR128:$Rm),
2679 asmop # "\t$Rd." # ResS # ", $Rn." # OpS # ", $Rm." # OpS,
2681 let Constraints = "$src = $Rd";
2682 let neverHasSideEffects = 1;
2685 multiclass NeonI_3VDN_3Op_v1<bit u, bits<4> opcode, string asmop> {
2686 def _16b8h : NeonI_3VDN_3Op<0b1, u, 0b00, opcode, asmop, "16b", "8h">;
2687 def _8h4s : NeonI_3VDN_3Op<0b1, u, 0b01, opcode, asmop, "8h", "4s">;
2688 def _4s2d : NeonI_3VDN_3Op<0b1, u, 0b10, opcode, asmop, "4s", "2d">;
2691 defm ADDHN2vvv : NeonI_3VDN_3Op_v1<0b0, 0b0100, "addhn2">;
2692 defm SUBHN2vvv : NeonI_3VDN_3Op_v1<0b0, 0b0110, "subhn2">;
2694 defm RADDHN2vvv : NeonI_3VDN_3Op_v1<0b1, 0b0100, "raddhn2">;
2695 defm RSUBHN2vvv : NeonI_3VDN_3Op_v1<0b1, 0b0110, "rsubhn2">;
2697 // Patterns have to be separate because there's a SUBREG_TO_REG in the output
2699 class NarrowHighHalfPat<Instruction INST, ValueType DstTy, ValueType SrcTy,
2700 SDPatternOperator coreop>
2701 : Pat<(Neon_combine_2D (v1i64 VPR64:$src),
2702 (v1i64 (bitconvert (DstTy (coreop (SrcTy VPR128:$Rn),
2703 (SrcTy VPR128:$Rm)))))),
2704 (INST (SUBREG_TO_REG (i64 0), VPR64:$src, sub_64),
2705 VPR128:$Rn, VPR128:$Rm)>;
2708 def : NarrowHighHalfPat<ADDHN2vvv_16b8h, v8i8, v8i16,
2709 BinOpFrag<(NI_get_hi_8h (add node:$LHS, node:$RHS))>>;
2710 def : NarrowHighHalfPat<ADDHN2vvv_8h4s, v4i16, v4i32,
2711 BinOpFrag<(NI_get_hi_4s (add node:$LHS, node:$RHS))>>;
2712 def : NarrowHighHalfPat<ADDHN2vvv_4s2d, v2i32, v2i64,
2713 BinOpFrag<(NI_get_hi_2d (add node:$LHS, node:$RHS))>>;
2716 def : NarrowHighHalfPat<SUBHN2vvv_16b8h, v8i8, v8i16,
2717 BinOpFrag<(NI_get_hi_8h (sub node:$LHS, node:$RHS))>>;
2718 def : NarrowHighHalfPat<SUBHN2vvv_8h4s, v4i16, v4i32,
2719 BinOpFrag<(NI_get_hi_4s (sub node:$LHS, node:$RHS))>>;
2720 def : NarrowHighHalfPat<SUBHN2vvv_4s2d, v2i32, v2i64,
2721 BinOpFrag<(NI_get_hi_2d (sub node:$LHS, node:$RHS))>>;
2724 def : NarrowHighHalfPat<RADDHN2vvv_16b8h, v8i8, v8i16, int_arm_neon_vraddhn>;
2725 def : NarrowHighHalfPat<RADDHN2vvv_8h4s, v4i16, v4i32, int_arm_neon_vraddhn>;
2726 def : NarrowHighHalfPat<RADDHN2vvv_4s2d, v2i32, v2i64, int_arm_neon_vraddhn>;
2729 def : NarrowHighHalfPat<RSUBHN2vvv_16b8h, v8i8, v8i16, int_arm_neon_vrsubhn>;
2730 def : NarrowHighHalfPat<RSUBHN2vvv_8h4s, v4i16, v4i32, int_arm_neon_vrsubhn>;
2731 def : NarrowHighHalfPat<RSUBHN2vvv_4s2d, v2i32, v2i64, int_arm_neon_vrsubhn>;
2733 // pattern that need to extend result
2734 class NeonI_3VDL_Ext<bit q, bit u, bits<2> size, bits<4> opcode,
2735 string asmop, string ResS, string OpS,
2736 SDPatternOperator opnode,
2737 RegisterOperand OpVPR,
2738 ValueType ResTy, ValueType OpTy, ValueType OpSTy>
2739 : NeonI_3VDiff<q, u, size, opcode,
2740 (outs VPR128:$Rd), (ins OpVPR:$Rn, OpVPR:$Rm),
2741 asmop # "\t$Rd." # ResS # ", $Rn." # OpS # ", $Rm." # OpS,
2742 [(set (ResTy VPR128:$Rd),
2743 (ResTy (zext (OpSTy (opnode (OpTy OpVPR:$Rn),
2744 (OpTy OpVPR:$Rm))))))],
2747 multiclass NeonI_3VDL_zext<bit u, bits<4> opcode, string asmop,
2748 SDPatternOperator opnode, bit Commutable = 0> {
2749 let isCommutable = Commutable in {
2750 def _8h8b : NeonI_3VDL_Ext<0b0, u, 0b00, opcode, asmop, "8h", "8b",
2751 opnode, VPR64, v8i16, v8i8, v8i8>;
2752 def _4s4h : NeonI_3VDL_Ext<0b0, u, 0b01, opcode, asmop, "4s", "4h",
2753 opnode, VPR64, v4i32, v4i16, v4i16>;
2754 def _2d2s : NeonI_3VDL_Ext<0b0, u, 0b10, opcode, asmop, "2d", "2s",
2755 opnode, VPR64, v2i64, v2i32, v2i32>;
2759 defm SABDLvvv : NeonI_3VDL_zext<0b0, 0b0111, "sabdl", int_arm_neon_vabds, 1>;
2760 defm UABDLvvv : NeonI_3VDL_zext<0b1, 0b0111, "uabdl", int_arm_neon_vabdu, 1>;
2762 multiclass NeonI_Op_High<SDPatternOperator op> {
2763 def _16B : PatFrag<(ops node:$Rn, node:$Rm),
2764 (op (v8i8 (Neon_High16B node:$Rn)),
2765 (v8i8 (Neon_High16B node:$Rm)))>;
2766 def _8H : PatFrag<(ops node:$Rn, node:$Rm),
2767 (op (v4i16 (Neon_High8H node:$Rn)),
2768 (v4i16 (Neon_High8H node:$Rm)))>;
2769 def _4S : PatFrag<(ops node:$Rn, node:$Rm),
2770 (op (v2i32 (Neon_High4S node:$Rn)),
2771 (v2i32 (Neon_High4S node:$Rm)))>;
2774 defm NI_sabdl_hi : NeonI_Op_High<int_arm_neon_vabds>;
2775 defm NI_uabdl_hi : NeonI_Op_High<int_arm_neon_vabdu>;
2776 defm NI_smull_hi : NeonI_Op_High<int_arm_neon_vmulls>;
2777 defm NI_umull_hi : NeonI_Op_High<int_arm_neon_vmullu>;
2778 defm NI_qdmull_hi : NeonI_Op_High<int_arm_neon_vqdmull>;
2779 defm NI_pmull_hi : NeonI_Op_High<int_arm_neon_vmullp>;
2781 multiclass NeonI_3VDL_Abd_u<bit u, bits<4> opcode, string asmop, string opnode,
2782 bit Commutable = 0> {
2783 let isCommutable = Commutable in {
2784 def _8h8b : NeonI_3VDL_Ext<0b1, u, 0b00, opcode, asmop, "8h", "16b",
2785 !cast<PatFrag>(opnode # "_16B"),
2786 VPR128, v8i16, v16i8, v8i8>;
2787 def _4s4h : NeonI_3VDL_Ext<0b1, u, 0b01, opcode, asmop, "4s", "8h",
2788 !cast<PatFrag>(opnode # "_8H"),
2789 VPR128, v4i32, v8i16, v4i16>;
2790 def _2d2s : NeonI_3VDL_Ext<0b1, u, 0b10, opcode, asmop, "2d", "4s",
2791 !cast<PatFrag>(opnode # "_4S"),
2792 VPR128, v2i64, v4i32, v2i32>;
2796 defm SABDL2vvv : NeonI_3VDL_Abd_u<0b0, 0b0111, "sabdl2", "NI_sabdl_hi", 1>;
2797 defm UABDL2vvv : NeonI_3VDL_Abd_u<0b1, 0b0111, "uabdl2", "NI_uabdl_hi", 1>;
2799 // For pattern that need two operators being chained.
2800 class NeonI_3VDL_Aba<bit q, bit u, bits<2> size, bits<4> opcode,
2801 string asmop, string ResS, string OpS,
2802 SDPatternOperator opnode, SDPatternOperator subop,
2803 RegisterOperand OpVPR,
2804 ValueType ResTy, ValueType OpTy, ValueType OpSTy>
2805 : NeonI_3VDiff<q, u, size, opcode,
2806 (outs VPR128:$Rd), (ins VPR128:$src, OpVPR:$Rn, OpVPR:$Rm),
2807 asmop # "\t$Rd." # ResS # ", $Rn." # OpS # ", $Rm." # OpS,
2808 [(set (ResTy VPR128:$Rd),
2810 (ResTy VPR128:$src),
2811 (ResTy (zext (OpSTy (subop (OpTy OpVPR:$Rn),
2812 (OpTy OpVPR:$Rm))))))))],
2814 let Constraints = "$src = $Rd";
2817 multiclass NeonI_3VDL_Aba_v1<bit u, bits<4> opcode, string asmop,
2818 SDPatternOperator opnode, SDPatternOperator subop>{
2819 def _8h8b : NeonI_3VDL_Aba<0b0, u, 0b00, opcode, asmop, "8h", "8b",
2820 opnode, subop, VPR64, v8i16, v8i8, v8i8>;
2821 def _4s4h : NeonI_3VDL_Aba<0b0, u, 0b01, opcode, asmop, "4s", "4h",
2822 opnode, subop, VPR64, v4i32, v4i16, v4i16>;
2823 def _2d2s : NeonI_3VDL_Aba<0b0, u, 0b10, opcode, asmop, "2d", "2s",
2824 opnode, subop, VPR64, v2i64, v2i32, v2i32>;
2827 defm SABALvvv : NeonI_3VDL_Aba_v1<0b0, 0b0101, "sabal",
2828 add, int_arm_neon_vabds>;
2829 defm UABALvvv : NeonI_3VDL_Aba_v1<0b1, 0b0101, "uabal",
2830 add, int_arm_neon_vabdu>;
2832 multiclass NeonI_3VDL2_Aba_v1<bit u, bits<4> opcode, string asmop,
2833 SDPatternOperator opnode, string subop> {
2834 def _8h8b : NeonI_3VDL_Aba<0b1, u, 0b00, opcode, asmop, "8h", "16b",
2835 opnode, !cast<PatFrag>(subop # "_16B"),
2836 VPR128, v8i16, v16i8, v8i8>;
2837 def _4s4h : NeonI_3VDL_Aba<0b1, u, 0b01, opcode, asmop, "4s", "8h",
2838 opnode, !cast<PatFrag>(subop # "_8H"),
2839 VPR128, v4i32, v8i16, v4i16>;
2840 def _2d2s : NeonI_3VDL_Aba<0b1, u, 0b10, opcode, asmop, "2d", "4s",
2841 opnode, !cast<PatFrag>(subop # "_4S"),
2842 VPR128, v2i64, v4i32, v2i32>;
2845 defm SABAL2vvv : NeonI_3VDL2_Aba_v1<0b0, 0b0101, "sabal2", add,
2847 defm UABAL2vvv : NeonI_3VDL2_Aba_v1<0b1, 0b0101, "uabal2", add,
2850 // Long pattern with 2 operands
2851 multiclass NeonI_3VDL_2Op<bit u, bits<4> opcode, string asmop,
2852 SDPatternOperator opnode, bit Commutable = 0> {
2853 let isCommutable = Commutable in {
2854 def _8h8b : NeonI_3VD_2Op<0b0, u, 0b00, opcode, asmop, "8h", "8b",
2855 opnode, VPR128, VPR64, v8i16, v8i8>;
2856 def _4s4h : NeonI_3VD_2Op<0b0, u, 0b01, opcode, asmop, "4s", "4h",
2857 opnode, VPR128, VPR64, v4i32, v4i16>;
2858 def _2d2s : NeonI_3VD_2Op<0b0, u, 0b10, opcode, asmop, "2d", "2s",
2859 opnode, VPR128, VPR64, v2i64, v2i32>;
2863 defm SMULLvvv : NeonI_3VDL_2Op<0b0, 0b1100, "smull", int_arm_neon_vmulls, 1>;
2864 defm UMULLvvv : NeonI_3VDL_2Op<0b1, 0b1100, "umull", int_arm_neon_vmullu, 1>;
2866 class NeonI_3VDL2_2Op_mull<bit q, bit u, bits<2> size, bits<4> opcode,
2867 string asmop, string ResS, string OpS,
2868 SDPatternOperator opnode,
2869 ValueType ResTy, ValueType OpTy>
2870 : NeonI_3VDiff<q, u, size, opcode,
2871 (outs VPR128:$Rd), (ins VPR128:$Rn, VPR128:$Rm),
2872 asmop # "\t$Rd." # ResS # ", $Rn." # OpS # ", $Rm." # OpS,
2873 [(set (ResTy VPR128:$Rd),
2874 (ResTy (opnode (OpTy VPR128:$Rn), (OpTy VPR128:$Rm))))],
2877 multiclass NeonI_3VDL2_2Op_mull_v1<bit u, bits<4> opcode, string asmop,
2878 string opnode, bit Commutable = 0> {
2879 let isCommutable = Commutable in {
2880 def _8h16b : NeonI_3VDL2_2Op_mull<0b1, u, 0b00, opcode, asmop, "8h", "16b",
2881 !cast<PatFrag>(opnode # "_16B"),
2883 def _4s8h : NeonI_3VDL2_2Op_mull<0b1, u, 0b01, opcode, asmop, "4s", "8h",
2884 !cast<PatFrag>(opnode # "_8H"),
2886 def _2d4s : NeonI_3VDL2_2Op_mull<0b1, u, 0b10, opcode, asmop, "2d", "4s",
2887 !cast<PatFrag>(opnode # "_4S"),
2892 defm SMULL2vvv : NeonI_3VDL2_2Op_mull_v1<0b0, 0b1100, "smull2",
2894 defm UMULL2vvv : NeonI_3VDL2_2Op_mull_v1<0b1, 0b1100, "umull2",
2897 // Long pattern with 3 operands
2898 class NeonI_3VDL_3Op<bit q, bit u, bits<2> size, bits<4> opcode,
2899 string asmop, string ResS, string OpS,
2900 SDPatternOperator opnode,
2901 ValueType ResTy, ValueType OpTy>
2902 : NeonI_3VDiff<q, u, size, opcode,
2903 (outs VPR128:$Rd), (ins VPR128:$src, VPR64:$Rn, VPR64:$Rm),
2904 asmop # "\t$Rd." # ResS # ", $Rn." # OpS # ", $Rm." # OpS,
2905 [(set (ResTy VPR128:$Rd),
2907 (ResTy VPR128:$src),
2908 (OpTy VPR64:$Rn), (OpTy VPR64:$Rm))))],
2910 let Constraints = "$src = $Rd";
2913 multiclass NeonI_3VDL_3Op_v1<bit u, bits<4> opcode, string asmop,
2914 SDPatternOperator opnode> {
2915 def _8h8b : NeonI_3VDL_3Op<0b0, u, 0b00, opcode, asmop, "8h", "8b",
2916 opnode, v8i16, v8i8>;
2917 def _4s4h : NeonI_3VDL_3Op<0b0, u, 0b01, opcode, asmop, "4s", "4h",
2918 opnode, v4i32, v4i16>;
2919 def _2d2s : NeonI_3VDL_3Op<0b0, u, 0b10, opcode, asmop, "2d", "2s",
2920 opnode, v2i64, v2i32>;
2923 def Neon_smlal : PatFrag<(ops node:$Rd, node:$Rn, node:$Rm),
2925 (int_arm_neon_vmulls node:$Rn, node:$Rm))>;
2927 def Neon_umlal : PatFrag<(ops node:$Rd, node:$Rn, node:$Rm),
2929 (int_arm_neon_vmullu node:$Rn, node:$Rm))>;
2931 def Neon_smlsl : PatFrag<(ops node:$Rd, node:$Rn, node:$Rm),
2933 (int_arm_neon_vmulls node:$Rn, node:$Rm))>;
2935 def Neon_umlsl : PatFrag<(ops node:$Rd, node:$Rn, node:$Rm),
2937 (int_arm_neon_vmullu node:$Rn, node:$Rm))>;
2939 defm SMLALvvv : NeonI_3VDL_3Op_v1<0b0, 0b1000, "smlal", Neon_smlal>;
2940 defm UMLALvvv : NeonI_3VDL_3Op_v1<0b1, 0b1000, "umlal", Neon_umlal>;
2942 defm SMLSLvvv : NeonI_3VDL_3Op_v1<0b0, 0b1010, "smlsl", Neon_smlsl>;
2943 defm UMLSLvvv : NeonI_3VDL_3Op_v1<0b1, 0b1010, "umlsl", Neon_umlsl>;
2945 class NeonI_3VDL2_3Op_mlas<bit q, bit u, bits<2> size, bits<4> opcode,
2946 string asmop, string ResS, string OpS,
2947 SDPatternOperator subop, SDPatternOperator opnode,
2948 RegisterOperand OpVPR,
2949 ValueType ResTy, ValueType OpTy>
2950 : NeonI_3VDiff<q, u, size, opcode,
2951 (outs VPR128:$Rd), (ins VPR128:$src, OpVPR:$Rn, OpVPR:$Rm),
2952 asmop # "\t$Rd." # ResS # ", $Rn." # OpS # ", $Rm." # OpS,
2953 [(set (ResTy VPR128:$Rd),
2955 (ResTy VPR128:$src),
2956 (ResTy (opnode (OpTy OpVPR:$Rn), (OpTy OpVPR:$Rm))))))],
2958 let Constraints = "$src = $Rd";
2961 multiclass NeonI_3VDL2_3Op_mlas_v1<bit u, bits<4> opcode, string asmop,
2962 SDPatternOperator subop, string opnode> {
2963 def _8h16b : NeonI_3VDL2_3Op_mlas<0b1, u, 0b00, opcode, asmop, "8h", "16b",
2964 subop, !cast<PatFrag>(opnode # "_16B"),
2965 VPR128, v8i16, v16i8>;
2966 def _4s8h : NeonI_3VDL2_3Op_mlas<0b1, u, 0b01, opcode, asmop, "4s", "8h",
2967 subop, !cast<PatFrag>(opnode # "_8H"),
2968 VPR128, v4i32, v8i16>;
2969 def _2d4s : NeonI_3VDL2_3Op_mlas<0b1, u, 0b10, opcode, asmop, "2d", "4s",
2970 subop, !cast<PatFrag>(opnode # "_4S"),
2971 VPR128, v2i64, v4i32>;
2974 defm SMLAL2vvv : NeonI_3VDL2_3Op_mlas_v1<0b0, 0b1000, "smlal2",
2975 add, "NI_smull_hi">;
2976 defm UMLAL2vvv : NeonI_3VDL2_3Op_mlas_v1<0b1, 0b1000, "umlal2",
2977 add, "NI_umull_hi">;
2979 defm SMLSL2vvv : NeonI_3VDL2_3Op_mlas_v1<0b0, 0b1010, "smlsl2",
2980 sub, "NI_smull_hi">;
2981 defm UMLSL2vvv : NeonI_3VDL2_3Op_mlas_v1<0b1, 0b1010, "umlsl2",
2982 sub, "NI_umull_hi">;
2984 multiclass NeonI_3VDL_qdmlal_3Op_v2<bit u, bits<4> opcode, string asmop,
2985 SDPatternOperator opnode> {
2986 def _4s4h : NeonI_3VDL2_3Op_mlas<0b0, u, 0b01, opcode, asmop, "4s", "4h",
2987 opnode, int_arm_neon_vqdmull,
2988 VPR64, v4i32, v4i16>;
2989 def _2d2s : NeonI_3VDL2_3Op_mlas<0b0, u, 0b10, opcode, asmop, "2d", "2s",
2990 opnode, int_arm_neon_vqdmull,
2991 VPR64, v2i64, v2i32>;
2994 defm SQDMLALvvv : NeonI_3VDL_qdmlal_3Op_v2<0b0, 0b1001, "sqdmlal",
2995 int_arm_neon_vqadds>;
2996 defm SQDMLSLvvv : NeonI_3VDL_qdmlal_3Op_v2<0b0, 0b1011, "sqdmlsl",
2997 int_arm_neon_vqsubs>;
2999 multiclass NeonI_3VDL_v2<bit u, bits<4> opcode, string asmop,
3000 SDPatternOperator opnode, bit Commutable = 0> {
3001 let isCommutable = Commutable in {
3002 def _4s4h : NeonI_3VD_2Op<0b0, u, 0b01, opcode, asmop, "4s", "4h",
3003 opnode, VPR128, VPR64, v4i32, v4i16>;
3004 def _2d2s : NeonI_3VD_2Op<0b0, u, 0b10, opcode, asmop, "2d", "2s",
3005 opnode, VPR128, VPR64, v2i64, v2i32>;
3009 defm SQDMULLvvv : NeonI_3VDL_v2<0b0, 0b1101, "sqdmull",
3010 int_arm_neon_vqdmull, 1>;
3012 multiclass NeonI_3VDL2_2Op_mull_v2<bit u, bits<4> opcode, string asmop,
3013 string opnode, bit Commutable = 0> {
3014 let isCommutable = Commutable in {
3015 def _4s8h : NeonI_3VDL2_2Op_mull<0b1, u, 0b01, opcode, asmop, "4s", "8h",
3016 !cast<PatFrag>(opnode # "_8H"),
3018 def _2d4s : NeonI_3VDL2_2Op_mull<0b1, u, 0b10, opcode, asmop, "2d", "4s",
3019 !cast<PatFrag>(opnode # "_4S"),
3024 defm SQDMULL2vvv : NeonI_3VDL2_2Op_mull_v2<0b0, 0b1101, "sqdmull2",
3027 multiclass NeonI_3VDL2_3Op_qdmlal_v2<bit u, bits<4> opcode, string asmop,
3028 SDPatternOperator opnode> {
3029 def _4s8h : NeonI_3VDL2_3Op_mlas<0b1, u, 0b01, opcode, asmop, "4s", "8h",
3030 opnode, NI_qdmull_hi_8H,
3031 VPR128, v4i32, v8i16>;
3032 def _2d4s : NeonI_3VDL2_3Op_mlas<0b1, u, 0b10, opcode, asmop, "2d", "4s",
3033 opnode, NI_qdmull_hi_4S,
3034 VPR128, v2i64, v4i32>;
3037 defm SQDMLAL2vvv : NeonI_3VDL2_3Op_qdmlal_v2<0b0, 0b1001, "sqdmlal2",
3038 int_arm_neon_vqadds>;
3039 defm SQDMLSL2vvv : NeonI_3VDL2_3Op_qdmlal_v2<0b0, 0b1011, "sqdmlsl2",
3040 int_arm_neon_vqsubs>;
3042 multiclass NeonI_3VDL_v3<bit u, bits<4> opcode, string asmop,
3043 SDPatternOperator opnode_8h8b,
3044 SDPatternOperator opnode_1q1d, bit Commutable = 0> {
3045 let isCommutable = Commutable in {
3046 def _8h8b : NeonI_3VD_2Op<0b0, u, 0b00, opcode, asmop, "8h", "8b",
3047 opnode_8h8b, VPR128, VPR64, v8i16, v8i8>;
3049 def _1q1d : NeonI_3VD_2Op<0b0, u, 0b11, opcode, asmop, "1q", "1d",
3050 opnode_1q1d, VPR128, VPR64, v16i8, v1i64>;
3054 defm PMULLvvv : NeonI_3VDL_v3<0b0, 0b1110, "pmull", int_arm_neon_vmullp,
3055 int_aarch64_neon_vmull_p64, 1>;
3057 multiclass NeonI_3VDL2_2Op_mull_v3<bit u, bits<4> opcode, string asmop,
3058 string opnode, bit Commutable = 0> {
3059 let isCommutable = Commutable in {
3060 def _8h16b : NeonI_3VDL2_2Op_mull<0b1, u, 0b00, opcode, asmop, "8h", "16b",
3061 !cast<PatFrag>(opnode # "_16B"),
3065 NeonI_3VDiff<0b1, u, 0b11, opcode,
3066 (outs VPR128:$Rd), (ins VPR128:$Rn, VPR128:$Rm),
3067 asmop # "\t$Rd.1q, $Rn.2d, $Rm.2d",
3068 [(set (v16i8 VPR128:$Rd),
3069 (v16i8 (int_aarch64_neon_vmull_p64
3070 (v1i64 (scalar_to_vector
3071 (i64 (vector_extract (v2i64 VPR128:$Rn), 1)))),
3072 (v1i64 (scalar_to_vector
3073 (i64 (vector_extract (v2i64 VPR128:$Rm), 1)))))))],
3078 defm PMULL2vvv : NeonI_3VDL2_2Op_mull_v3<0b0, 0b1110, "pmull2", "NI_pmull_hi",
3081 // End of implementation for instruction class (3V Diff)
3083 // The followings are vector load/store multiple N-element structure
3084 // (class SIMD lselem).
3086 // ld1: load multiple 1-element structure to 1/2/3/4 registers.
3087 // ld2/ld3/ld4: load multiple N-element structure to N registers (N = 2, 3, 4).
3088 // The structure consists of a sequence of sets of N values.
3089 // The first element of the structure is placed in the first lane
3090 // of the first first vector, the second element in the first lane
3091 // of the second vector, and so on.
3092 // E.g. LD1_3V_2S will load 32-bit elements {A, B, C, D, E, F} sequentially into
3093 // the three 64-bit vectors list {BA, DC, FE}.
3094 // E.g. LD3_2S will load 32-bit elements {A, B, C, D, E, F} into the three
3095 // 64-bit vectors list {DA, EB, FC}.
3096 // Store instructions store multiple structure to N registers like load.
3099 class NeonI_LDVList<bit q, bits<4> opcode, bits<2> size,
3100 RegisterOperand VecList, string asmop>
3101 : NeonI_LdStMult<q, 1, opcode, size,
3102 (outs VecList:$Rt), (ins GPR64xsp:$Rn),
3103 asmop # "\t$Rt, [$Rn]",
3107 let neverHasSideEffects = 1;
3110 multiclass LDVList_BHSD<bits<4> opcode, string List, string asmop> {
3111 def _8B : NeonI_LDVList<0, opcode, 0b00,
3112 !cast<RegisterOperand>(List # "8B_operand"), asmop>;
3114 def _4H : NeonI_LDVList<0, opcode, 0b01,
3115 !cast<RegisterOperand>(List # "4H_operand"), asmop>;
3117 def _2S : NeonI_LDVList<0, opcode, 0b10,
3118 !cast<RegisterOperand>(List # "2S_operand"), asmop>;
3120 def _16B : NeonI_LDVList<1, opcode, 0b00,
3121 !cast<RegisterOperand>(List # "16B_operand"), asmop>;
3123 def _8H : NeonI_LDVList<1, opcode, 0b01,
3124 !cast<RegisterOperand>(List # "8H_operand"), asmop>;
3126 def _4S : NeonI_LDVList<1, opcode, 0b10,
3127 !cast<RegisterOperand>(List # "4S_operand"), asmop>;
3129 def _2D : NeonI_LDVList<1, opcode, 0b11,
3130 !cast<RegisterOperand>(List # "2D_operand"), asmop>;
3133 // Load multiple N-element structure to N consecutive registers (N = 1,2,3,4)
3134 defm LD1 : LDVList_BHSD<0b0111, "VOne", "ld1">;
3135 def LD1_1D : NeonI_LDVList<0, 0b0111, 0b11, VOne1D_operand, "ld1">;
3137 defm LD2 : LDVList_BHSD<0b1000, "VPair", "ld2">;
3139 defm LD3 : LDVList_BHSD<0b0100, "VTriple", "ld3">;
3141 defm LD4 : LDVList_BHSD<0b0000, "VQuad", "ld4">;
3143 // Load multiple 1-element structure to N consecutive registers (N = 2,3,4)
3144 defm LD1x2 : LDVList_BHSD<0b1010, "VPair", "ld1">;
3145 def LD1x2_1D : NeonI_LDVList<0, 0b1010, 0b11, VPair1D_operand, "ld1">;
3147 defm LD1x3 : LDVList_BHSD<0b0110, "VTriple", "ld1">;
3148 def LD1x3_1D : NeonI_LDVList<0, 0b0110, 0b11, VTriple1D_operand, "ld1">;
3150 defm LD1x4 : LDVList_BHSD<0b0010, "VQuad", "ld1">;
3151 def LD1x4_1D : NeonI_LDVList<0, 0b0010, 0b11, VQuad1D_operand, "ld1">;
3153 class NeonI_STVList<bit q, bits<4> opcode, bits<2> size,
3154 RegisterOperand VecList, string asmop>
3155 : NeonI_LdStMult<q, 0, opcode, size,
3156 (outs), (ins GPR64xsp:$Rn, VecList:$Rt),
3157 asmop # "\t$Rt, [$Rn]",
3161 let neverHasSideEffects = 1;
3164 multiclass STVList_BHSD<bits<4> opcode, string List, string asmop> {
3165 def _8B : NeonI_STVList<0, opcode, 0b00,
3166 !cast<RegisterOperand>(List # "8B_operand"), asmop>;
3168 def _4H : NeonI_STVList<0, opcode, 0b01,
3169 !cast<RegisterOperand>(List # "4H_operand"), asmop>;
3171 def _2S : NeonI_STVList<0, opcode, 0b10,
3172 !cast<RegisterOperand>(List # "2S_operand"), asmop>;
3174 def _16B : NeonI_STVList<1, opcode, 0b00,
3175 !cast<RegisterOperand>(List # "16B_operand"), asmop>;
3177 def _8H : NeonI_STVList<1, opcode, 0b01,
3178 !cast<RegisterOperand>(List # "8H_operand"), asmop>;
3180 def _4S : NeonI_STVList<1, opcode, 0b10,
3181 !cast<RegisterOperand>(List # "4S_operand"), asmop>;
3183 def _2D : NeonI_STVList<1, opcode, 0b11,
3184 !cast<RegisterOperand>(List # "2D_operand"), asmop>;
3187 // Store multiple N-element structures from N registers (N = 1,2,3,4)
3188 defm ST1 : STVList_BHSD<0b0111, "VOne", "st1">;
3189 def ST1_1D : NeonI_STVList<0, 0b0111, 0b11, VOne1D_operand, "st1">;
3191 defm ST2 : STVList_BHSD<0b1000, "VPair", "st2">;
3193 defm ST3 : STVList_BHSD<0b0100, "VTriple", "st3">;
3195 defm ST4 : STVList_BHSD<0b0000, "VQuad", "st4">;
3197 // Store multiple 1-element structures from N consecutive registers (N = 2,3,4)
3198 defm ST1x2 : STVList_BHSD<0b1010, "VPair", "st1">;
3199 def ST1x2_1D : NeonI_STVList<0, 0b1010, 0b11, VPair1D_operand, "st1">;
3201 defm ST1x3 : STVList_BHSD<0b0110, "VTriple", "st1">;
3202 def ST1x3_1D : NeonI_STVList<0, 0b0110, 0b11, VTriple1D_operand, "st1">;
3204 defm ST1x4 : STVList_BHSD<0b0010, "VQuad", "st1">;
3205 def ST1x4_1D : NeonI_STVList<0, 0b0010, 0b11, VQuad1D_operand, "st1">;
3207 def : Pat<(v2f64 (load GPR64xsp:$addr)), (LD1_2D GPR64xsp:$addr)>;
3208 def : Pat<(v2i64 (load GPR64xsp:$addr)), (LD1_2D GPR64xsp:$addr)>;
3210 def : Pat<(v4f32 (load GPR64xsp:$addr)), (LD1_4S GPR64xsp:$addr)>;
3211 def : Pat<(v4i32 (load GPR64xsp:$addr)), (LD1_4S GPR64xsp:$addr)>;
3213 def : Pat<(v8i16 (load GPR64xsp:$addr)), (LD1_8H GPR64xsp:$addr)>;
3214 def : Pat<(v16i8 (load GPR64xsp:$addr)), (LD1_16B GPR64xsp:$addr)>;
3216 def : Pat<(v1f64 (load GPR64xsp:$addr)), (LD1_1D GPR64xsp:$addr)>;
3217 def : Pat<(v1i64 (load GPR64xsp:$addr)), (LD1_1D GPR64xsp:$addr)>;
3219 def : Pat<(v2f32 (load GPR64xsp:$addr)), (LD1_2S GPR64xsp:$addr)>;
3220 def : Pat<(v2i32 (load GPR64xsp:$addr)), (LD1_2S GPR64xsp:$addr)>;
3222 def : Pat<(v4i16 (load GPR64xsp:$addr)), (LD1_4H GPR64xsp:$addr)>;
3223 def : Pat<(v8i8 (load GPR64xsp:$addr)), (LD1_8B GPR64xsp:$addr)>;
3225 def : Pat<(store (v2i64 VPR128:$value), GPR64xsp:$addr),
3226 (ST1_2D GPR64xsp:$addr, VPR128:$value)>;
3227 def : Pat<(store (v2f64 VPR128:$value), GPR64xsp:$addr),
3228 (ST1_2D GPR64xsp:$addr, VPR128:$value)>;
3230 def : Pat<(store (v4i32 VPR128:$value), GPR64xsp:$addr),
3231 (ST1_4S GPR64xsp:$addr, VPR128:$value)>;
3232 def : Pat<(store (v4f32 VPR128:$value), GPR64xsp:$addr),
3233 (ST1_4S GPR64xsp:$addr, VPR128:$value)>;
3235 def : Pat<(store (v8i16 VPR128:$value), GPR64xsp:$addr),
3236 (ST1_8H GPR64xsp:$addr, VPR128:$value)>;
3237 def : Pat<(store (v16i8 VPR128:$value), GPR64xsp:$addr),
3238 (ST1_16B GPR64xsp:$addr, VPR128:$value)>;
3240 def : Pat<(store (v1i64 VPR64:$value), GPR64xsp:$addr),
3241 (ST1_1D GPR64xsp:$addr, VPR64:$value)>;
3242 def : Pat<(store (v1f64 VPR64:$value), GPR64xsp:$addr),
3243 (ST1_1D GPR64xsp:$addr, VPR64:$value)>;
3245 def : Pat<(store (v2i32 VPR64:$value), GPR64xsp:$addr),
3246 (ST1_2S GPR64xsp:$addr, VPR64:$value)>;
3247 def : Pat<(store (v2f32 VPR64:$value), GPR64xsp:$addr),
3248 (ST1_2S GPR64xsp:$addr, VPR64:$value)>;
3250 def : Pat<(store (v4i16 VPR64:$value), GPR64xsp:$addr),
3251 (ST1_4H GPR64xsp:$addr, VPR64:$value)>;
3252 def : Pat<(store (v8i8 VPR64:$value), GPR64xsp:$addr),
3253 (ST1_8B GPR64xsp:$addr, VPR64:$value)>;
3255 // Match load/store of v1i8/v1i16/v1i32 type to FPR8/FPR16/FPR32 load/store.
3256 // FIXME: for now we have v1i8, v1i16, v1i32 legal types, if they are illegal,
3257 // these patterns are not needed any more.
3258 def : Pat<(v1i8 (load GPR64xsp:$addr)), (LSFP8_LDR $addr, 0)>;
3259 def : Pat<(v1i16 (load GPR64xsp:$addr)), (LSFP16_LDR $addr, 0)>;
3260 def : Pat<(v1i32 (load GPR64xsp:$addr)), (LSFP32_LDR $addr, 0)>;
3262 def : Pat<(store (v1i8 FPR8:$value), GPR64xsp:$addr),
3263 (LSFP8_STR $value, $addr, 0)>;
3264 def : Pat<(store (v1i16 FPR16:$value), GPR64xsp:$addr),
3265 (LSFP16_STR $value, $addr, 0)>;
3266 def : Pat<(store (v1i32 FPR32:$value), GPR64xsp:$addr),
3267 (LSFP32_STR $value, $addr, 0)>;
3270 // End of vector load/store multiple N-element structure(class SIMD lselem)
3272 // The followings are post-index vector load/store multiple N-element
3273 // structure(class SIMD lselem-post)
3274 def exact1_asmoperand : AsmOperandClass {
3275 let Name = "Exact1";
3276 let PredicateMethod = "isExactImm<1>";
3277 let RenderMethod = "addImmOperands";
3279 def uimm_exact1 : Operand<i32>, ImmLeaf<i32, [{return Imm == 1;}]> {
3280 let ParserMatchClass = exact1_asmoperand;
3283 def exact2_asmoperand : AsmOperandClass {
3284 let Name = "Exact2";
3285 let PredicateMethod = "isExactImm<2>";
3286 let RenderMethod = "addImmOperands";
3288 def uimm_exact2 : Operand<i32>, ImmLeaf<i32, [{return Imm == 2;}]> {
3289 let ParserMatchClass = exact2_asmoperand;
3292 def exact3_asmoperand : AsmOperandClass {
3293 let Name = "Exact3";
3294 let PredicateMethod = "isExactImm<3>";
3295 let RenderMethod = "addImmOperands";
3297 def uimm_exact3 : Operand<i32>, ImmLeaf<i32, [{return Imm == 3;}]> {
3298 let ParserMatchClass = exact3_asmoperand;
3301 def exact4_asmoperand : AsmOperandClass {
3302 let Name = "Exact4";
3303 let PredicateMethod = "isExactImm<4>";
3304 let RenderMethod = "addImmOperands";
3306 def uimm_exact4 : Operand<i32>, ImmLeaf<i32, [{return Imm == 4;}]> {
3307 let ParserMatchClass = exact4_asmoperand;
3310 def exact6_asmoperand : AsmOperandClass {
3311 let Name = "Exact6";
3312 let PredicateMethod = "isExactImm<6>";
3313 let RenderMethod = "addImmOperands";
3315 def uimm_exact6 : Operand<i32>, ImmLeaf<i32, [{return Imm == 6;}]> {
3316 let ParserMatchClass = exact6_asmoperand;
3319 def exact8_asmoperand : AsmOperandClass {
3320 let Name = "Exact8";
3321 let PredicateMethod = "isExactImm<8>";
3322 let RenderMethod = "addImmOperands";
3324 def uimm_exact8 : Operand<i32>, ImmLeaf<i32, [{return Imm == 8;}]> {
3325 let ParserMatchClass = exact8_asmoperand;
3328 def exact12_asmoperand : AsmOperandClass {
3329 let Name = "Exact12";
3330 let PredicateMethod = "isExactImm<12>";
3331 let RenderMethod = "addImmOperands";
3333 def uimm_exact12 : Operand<i32>, ImmLeaf<i32, [{return Imm == 12;}]> {
3334 let ParserMatchClass = exact12_asmoperand;
3337 def exact16_asmoperand : AsmOperandClass {
3338 let Name = "Exact16";
3339 let PredicateMethod = "isExactImm<16>";
3340 let RenderMethod = "addImmOperands";
3342 def uimm_exact16 : Operand<i32>, ImmLeaf<i32, [{return Imm == 16;}]> {
3343 let ParserMatchClass = exact16_asmoperand;
3346 def exact24_asmoperand : AsmOperandClass {
3347 let Name = "Exact24";
3348 let PredicateMethod = "isExactImm<24>";
3349 let RenderMethod = "addImmOperands";
3351 def uimm_exact24 : Operand<i32>, ImmLeaf<i32, [{return Imm == 24;}]> {
3352 let ParserMatchClass = exact24_asmoperand;
3355 def exact32_asmoperand : AsmOperandClass {
3356 let Name = "Exact32";
3357 let PredicateMethod = "isExactImm<32>";
3358 let RenderMethod = "addImmOperands";
3360 def uimm_exact32 : Operand<i32>, ImmLeaf<i32, [{return Imm == 32;}]> {
3361 let ParserMatchClass = exact32_asmoperand;
3364 def exact48_asmoperand : AsmOperandClass {
3365 let Name = "Exact48";
3366 let PredicateMethod = "isExactImm<48>";
3367 let RenderMethod = "addImmOperands";
3369 def uimm_exact48 : Operand<i32>, ImmLeaf<i32, [{return Imm == 48;}]> {
3370 let ParserMatchClass = exact48_asmoperand;
3373 def exact64_asmoperand : AsmOperandClass {
3374 let Name = "Exact64";
3375 let PredicateMethod = "isExactImm<64>";
3376 let RenderMethod = "addImmOperands";
3378 def uimm_exact64 : Operand<i32>, ImmLeaf<i32, [{return Imm == 64;}]> {
3379 let ParserMatchClass = exact64_asmoperand;
3382 multiclass NeonI_LDWB_VList<bit q, bits<4> opcode, bits<2> size,
3383 RegisterOperand VecList, Operand ImmTy,
3385 let Constraints = "$Rn = $wb", mayLoad = 1, neverHasSideEffects = 1,
3386 DecoderMethod = "DecodeVLDSTPostInstruction" in {
3387 def _fixed : NeonI_LdStMult_Post<q, 1, opcode, size,
3388 (outs VecList:$Rt, GPR64xsp:$wb),
3389 (ins GPR64xsp:$Rn, ImmTy:$amt),
3390 asmop # "\t$Rt, [$Rn], $amt",
3396 def _register : NeonI_LdStMult_Post<q, 1, opcode, size,
3397 (outs VecList:$Rt, GPR64xsp:$wb),
3398 (ins GPR64xsp:$Rn, GPR64noxzr:$Rm),
3399 asmop # "\t$Rt, [$Rn], $Rm",
3405 multiclass LDWB_VList_BHSD<bits<4> opcode, string List, Operand ImmTy,
3406 Operand ImmTy2, string asmop> {
3407 defm _8B : NeonI_LDWB_VList<0, opcode, 0b00,
3408 !cast<RegisterOperand>(List # "8B_operand"),
3411 defm _4H : NeonI_LDWB_VList<0, opcode, 0b01,
3412 !cast<RegisterOperand>(List # "4H_operand"),
3415 defm _2S : NeonI_LDWB_VList<0, opcode, 0b10,
3416 !cast<RegisterOperand>(List # "2S_operand"),
3419 defm _16B : NeonI_LDWB_VList<1, opcode, 0b00,
3420 !cast<RegisterOperand>(List # "16B_operand"),
3423 defm _8H : NeonI_LDWB_VList<1, opcode, 0b01,
3424 !cast<RegisterOperand>(List # "8H_operand"),
3427 defm _4S : NeonI_LDWB_VList<1, opcode, 0b10,
3428 !cast<RegisterOperand>(List # "4S_operand"),
3431 defm _2D : NeonI_LDWB_VList<1, opcode, 0b11,
3432 !cast<RegisterOperand>(List # "2D_operand"),
3436 // Post-index load multiple N-element structures from N registers (N = 1,2,3,4)
3437 defm LD1WB : LDWB_VList_BHSD<0b0111, "VOne", uimm_exact8, uimm_exact16, "ld1">;
3438 defm LD1WB_1D : NeonI_LDWB_VList<0, 0b0111, 0b11, VOne1D_operand, uimm_exact8,
3441 defm LD2WB : LDWB_VList_BHSD<0b1000, "VPair", uimm_exact16, uimm_exact32, "ld2">;
3443 defm LD3WB : LDWB_VList_BHSD<0b0100, "VTriple", uimm_exact24, uimm_exact48,
3446 defm LD4WB : LDWB_VList_BHSD<0b0000, "VQuad", uimm_exact32, uimm_exact64, "ld4">;
3448 // Post-index load multiple 1-element structures from N consecutive registers
3450 defm LD1x2WB : LDWB_VList_BHSD<0b1010, "VPair", uimm_exact16, uimm_exact32,
3452 defm LD1x2WB_1D : NeonI_LDWB_VList<0, 0b1010, 0b11, VPair1D_operand,
3453 uimm_exact16, "ld1">;
3455 defm LD1x3WB : LDWB_VList_BHSD<0b0110, "VTriple", uimm_exact24, uimm_exact48,
3457 defm LD1x3WB_1D : NeonI_LDWB_VList<0, 0b0110, 0b11, VTriple1D_operand,
3458 uimm_exact24, "ld1">;
3460 defm LD1x4WB : LDWB_VList_BHSD<0b0010, "VQuad", uimm_exact32, uimm_exact64,
3462 defm LD1x4WB_1D : NeonI_LDWB_VList<0, 0b0010, 0b11, VQuad1D_operand,
3463 uimm_exact32, "ld1">;
3465 multiclass NeonI_STWB_VList<bit q, bits<4> opcode, bits<2> size,
3466 RegisterOperand VecList, Operand ImmTy,
3468 let Constraints = "$Rn = $wb", mayStore = 1, neverHasSideEffects = 1,
3469 DecoderMethod = "DecodeVLDSTPostInstruction" in {
3470 def _fixed : NeonI_LdStMult_Post<q, 0, opcode, size,
3471 (outs GPR64xsp:$wb),
3472 (ins GPR64xsp:$Rn, ImmTy:$amt, VecList:$Rt),
3473 asmop # "\t$Rt, [$Rn], $amt",
3479 def _register : NeonI_LdStMult_Post<q, 0, opcode, size,
3480 (outs GPR64xsp:$wb),
3481 (ins GPR64xsp:$Rn, GPR64noxzr:$Rm, VecList:$Rt),
3482 asmop # "\t$Rt, [$Rn], $Rm",
3488 multiclass STWB_VList_BHSD<bits<4> opcode, string List, Operand ImmTy,
3489 Operand ImmTy2, string asmop> {
3490 defm _8B : NeonI_STWB_VList<0, opcode, 0b00,
3491 !cast<RegisterOperand>(List # "8B_operand"), ImmTy, asmop>;
3493 defm _4H : NeonI_STWB_VList<0, opcode, 0b01,
3494 !cast<RegisterOperand>(List # "4H_operand"),
3497 defm _2S : NeonI_STWB_VList<0, opcode, 0b10,
3498 !cast<RegisterOperand>(List # "2S_operand"),
3501 defm _16B : NeonI_STWB_VList<1, opcode, 0b00,
3502 !cast<RegisterOperand>(List # "16B_operand"),
3505 defm _8H : NeonI_STWB_VList<1, opcode, 0b01,
3506 !cast<RegisterOperand>(List # "8H_operand"),
3509 defm _4S : NeonI_STWB_VList<1, opcode, 0b10,
3510 !cast<RegisterOperand>(List # "4S_operand"),
3513 defm _2D : NeonI_STWB_VList<1, opcode, 0b11,
3514 !cast<RegisterOperand>(List # "2D_operand"),
3518 // Post-index load multiple N-element structures from N registers (N = 1,2,3,4)
3519 defm ST1WB : STWB_VList_BHSD<0b0111, "VOne", uimm_exact8, uimm_exact16, "st1">;
3520 defm ST1WB_1D : NeonI_STWB_VList<0, 0b0111, 0b11, VOne1D_operand, uimm_exact8,
3523 defm ST2WB : STWB_VList_BHSD<0b1000, "VPair", uimm_exact16, uimm_exact32, "st2">;
3525 defm ST3WB : STWB_VList_BHSD<0b0100, "VTriple", uimm_exact24, uimm_exact48,
3528 defm ST4WB : STWB_VList_BHSD<0b0000, "VQuad", uimm_exact32, uimm_exact64, "st4">;
3530 // Post-index load multiple 1-element structures from N consecutive registers
3532 defm ST1x2WB : STWB_VList_BHSD<0b1010, "VPair", uimm_exact16, uimm_exact32,
3534 defm ST1x2WB_1D : NeonI_STWB_VList<0, 0b1010, 0b11, VPair1D_operand,
3535 uimm_exact16, "st1">;
3537 defm ST1x3WB : STWB_VList_BHSD<0b0110, "VTriple", uimm_exact24, uimm_exact48,
3539 defm ST1x3WB_1D : NeonI_STWB_VList<0, 0b0110, 0b11, VTriple1D_operand,
3540 uimm_exact24, "st1">;
3542 defm ST1x4WB : STWB_VList_BHSD<0b0010, "VQuad", uimm_exact32, uimm_exact64,
3544 defm ST1x4WB_1D : NeonI_STWB_VList<0, 0b0010, 0b11, VQuad1D_operand,
3545 uimm_exact32, "st1">;
3547 // End of post-index vector load/store multiple N-element structure
3548 // (class SIMD lselem-post)
3550 // The followings are vector load/store single N-element structure
3551 // (class SIMD lsone).
3552 def neon_uimm0_bare : Operand<i64>,
3553 ImmLeaf<i64, [{return Imm == 0;}]> {
3554 let ParserMatchClass = neon_uimm0_asmoperand;
3555 let PrintMethod = "printUImmBareOperand";
3558 def neon_uimm1_bare : Operand<i64>,
3559 ImmLeaf<i64, [{return Imm < 2;}]> {
3560 let ParserMatchClass = neon_uimm1_asmoperand;
3561 let PrintMethod = "printUImmBareOperand";
3564 def neon_uimm2_bare : Operand<i64>,
3565 ImmLeaf<i64, [{return Imm < 4;}]> {
3566 let ParserMatchClass = neon_uimm2_asmoperand;
3567 let PrintMethod = "printUImmBareOperand";
3570 def neon_uimm3_bare : Operand<i64>,
3571 ImmLeaf<i64, [{return Imm < 8;}]> {
3572 let ParserMatchClass = uimm3_asmoperand;
3573 let PrintMethod = "printUImmBareOperand";
3576 def neon_uimm4_bare : Operand<i64>,
3577 ImmLeaf<i64, [{return Imm < 16;}]> {
3578 let ParserMatchClass = uimm4_asmoperand;
3579 let PrintMethod = "printUImmBareOperand";
3582 class NeonI_LDN_Dup<bit q, bit r, bits<3> opcode, bits<2> size,
3583 RegisterOperand VecList, string asmop>
3584 : NeonI_LdOne_Dup<q, r, opcode, size,
3585 (outs VecList:$Rt), (ins GPR64xsp:$Rn),
3586 asmop # "\t$Rt, [$Rn]",
3590 let neverHasSideEffects = 1;
3593 multiclass LDN_Dup_BHSD<bit r, bits<3> opcode, string List, string asmop> {
3594 def _8B : NeonI_LDN_Dup<0, r, opcode, 0b00,
3595 !cast<RegisterOperand>(List # "8B_operand"), asmop>;
3597 def _4H : NeonI_LDN_Dup<0, r, opcode, 0b01,
3598 !cast<RegisterOperand>(List # "4H_operand"), asmop>;
3600 def _2S : NeonI_LDN_Dup<0, r, opcode, 0b10,
3601 !cast<RegisterOperand>(List # "2S_operand"), asmop>;
3603 def _1D : NeonI_LDN_Dup<0, r, opcode, 0b11,
3604 !cast<RegisterOperand>(List # "1D_operand"), asmop>;
3606 def _16B : NeonI_LDN_Dup<1, r, opcode, 0b00,
3607 !cast<RegisterOperand>(List # "16B_operand"), asmop>;
3609 def _8H : NeonI_LDN_Dup<1, r, opcode, 0b01,
3610 !cast<RegisterOperand>(List # "8H_operand"), asmop>;
3612 def _4S : NeonI_LDN_Dup<1, r, opcode, 0b10,
3613 !cast<RegisterOperand>(List # "4S_operand"), asmop>;
3615 def _2D : NeonI_LDN_Dup<1, r, opcode, 0b11,
3616 !cast<RegisterOperand>(List # "2D_operand"), asmop>;
3619 // Load single 1-element structure to all lanes of 1 register
3620 defm LD1R : LDN_Dup_BHSD<0b0, 0b110, "VOne", "ld1r">;
3622 // Load single N-element structure to all lanes of N consecutive
3623 // registers (N = 2,3,4)
3624 defm LD2R : LDN_Dup_BHSD<0b1, 0b110, "VPair", "ld2r">;
3625 defm LD3R : LDN_Dup_BHSD<0b0, 0b111, "VTriple", "ld3r">;
3626 defm LD4R : LDN_Dup_BHSD<0b1, 0b111, "VQuad", "ld4r">;
3629 class LD1R_pattern <ValueType VTy, ValueType DTy, PatFrag LoadOp,
3631 : Pat<(VTy (Neon_vdup (DTy (LoadOp GPR64xsp:$Rn)))),
3632 (VTy (INST GPR64xsp:$Rn))>;
3634 // Match all LD1R instructions
3635 def : LD1R_pattern<v8i8, i32, extloadi8, LD1R_8B>;
3637 def : LD1R_pattern<v16i8, i32, extloadi8, LD1R_16B>;
3639 def : LD1R_pattern<v4i16, i32, extloadi16, LD1R_4H>;
3641 def : LD1R_pattern<v8i16, i32, extloadi16, LD1R_8H>;
3643 def : LD1R_pattern<v2i32, i32, load, LD1R_2S>;
3644 def : LD1R_pattern<v2f32, f32, load, LD1R_2S>;
3646 def : LD1R_pattern<v4i32, i32, load, LD1R_4S>;
3647 def : LD1R_pattern<v4f32, f32, load, LD1R_4S>;
3649 def : LD1R_pattern<v1i64, i64, load, LD1R_1D>;
3650 def : LD1R_pattern<v1f64, f64, load, LD1R_1D>;
3652 def : LD1R_pattern<v2i64, i64, load, LD1R_2D>;
3653 def : LD1R_pattern<v2f64, f64, load, LD1R_2D>;
3656 multiclass VectorList_Bare_BHSD<string PREFIX, int Count,
3657 RegisterClass RegList> {
3658 defm B : VectorList_operands<PREFIX, "B", Count, RegList>;
3659 defm H : VectorList_operands<PREFIX, "H", Count, RegList>;
3660 defm S : VectorList_operands<PREFIX, "S", Count, RegList>;
3661 defm D : VectorList_operands<PREFIX, "D", Count, RegList>;
3664 // Special vector list operand of 128-bit vectors with bare layout.
3665 // i.e. only show ".b", ".h", ".s", ".d"
3666 defm VOne : VectorList_Bare_BHSD<"VOne", 1, FPR128>;
3667 defm VPair : VectorList_Bare_BHSD<"VPair", 2, QPair>;
3668 defm VTriple : VectorList_Bare_BHSD<"VTriple", 3, QTriple>;
3669 defm VQuad : VectorList_Bare_BHSD<"VQuad", 4, QQuad>;
3671 class NeonI_LDN_Lane<bit r, bits<2> op2_1, bit op0, RegisterOperand VList,
3672 Operand ImmOp, string asmop>
3673 : NeonI_LdStOne_Lane<1, r, op2_1, op0,
3675 (ins GPR64xsp:$Rn, VList:$src, ImmOp:$lane),
3676 asmop # "\t$Rt[$lane], [$Rn]",
3680 let neverHasSideEffects = 1;
3681 let hasExtraDefRegAllocReq = 1;
3682 let Constraints = "$src = $Rt";
3685 multiclass LDN_Lane_BHSD<bit r, bit op0, string List, string asmop> {
3686 def _B : NeonI_LDN_Lane<r, 0b00, op0,
3687 !cast<RegisterOperand>(List # "B_operand"),
3688 neon_uimm4_bare, asmop> {
3689 let Inst{12-10} = lane{2-0};
3690 let Inst{30} = lane{3};
3693 def _H : NeonI_LDN_Lane<r, 0b01, op0,
3694 !cast<RegisterOperand>(List # "H_operand"),
3695 neon_uimm3_bare, asmop> {
3696 let Inst{12-10} = {lane{1}, lane{0}, 0b0};
3697 let Inst{30} = lane{2};
3700 def _S : NeonI_LDN_Lane<r, 0b10, op0,
3701 !cast<RegisterOperand>(List # "S_operand"),
3702 neon_uimm2_bare, asmop> {
3703 let Inst{12-10} = {lane{0}, 0b0, 0b0};
3704 let Inst{30} = lane{1};
3707 def _D : NeonI_LDN_Lane<r, 0b10, op0,
3708 !cast<RegisterOperand>(List # "D_operand"),
3709 neon_uimm1_bare, asmop> {
3710 let Inst{12-10} = 0b001;
3711 let Inst{30} = lane{0};
3715 // Load single 1-element structure to one lane of 1 register.
3716 defm LD1LN : LDN_Lane_BHSD<0b0, 0b0, "VOne", "ld1">;
3718 // Load single N-element structure to one lane of N consecutive registers
3720 defm LD2LN : LDN_Lane_BHSD<0b1, 0b0, "VPair", "ld2">;
3721 defm LD3LN : LDN_Lane_BHSD<0b0, 0b1, "VTriple", "ld3">;
3722 defm LD4LN : LDN_Lane_BHSD<0b1, 0b1, "VQuad", "ld4">;
3724 multiclass LD1LN_patterns<ValueType VTy, ValueType VTy2, ValueType DTy,
3725 Operand ImmOp, Operand ImmOp2, PatFrag LoadOp,
3727 def : Pat<(VTy (vector_insert (VTy VPR64:$src),
3728 (DTy (LoadOp GPR64xsp:$Rn)), (ImmOp:$lane))),
3729 (VTy (EXTRACT_SUBREG
3731 (SUBREG_TO_REG (i64 0), VPR64:$src, sub_64),
3735 def : Pat<(VTy2 (vector_insert (VTy2 VPR128:$src),
3736 (DTy (LoadOp GPR64xsp:$Rn)), (ImmOp2:$lane))),
3737 (VTy2 (INST GPR64xsp:$Rn, VPR128:$src, ImmOp2:$lane))>;
3740 // Match all LD1LN instructions
3741 defm : LD1LN_patterns<v8i8, v16i8, i32, neon_uimm3_bare, neon_uimm4_bare,
3742 extloadi8, LD1LN_B>;
3744 defm : LD1LN_patterns<v4i16, v8i16, i32, neon_uimm2_bare, neon_uimm3_bare,
3745 extloadi16, LD1LN_H>;
3747 defm : LD1LN_patterns<v2i32, v4i32, i32, neon_uimm1_bare, neon_uimm2_bare,
3749 defm : LD1LN_patterns<v2f32, v4f32, f32, neon_uimm1_bare, neon_uimm2_bare,
3752 defm : LD1LN_patterns<v1i64, v2i64, i64, neon_uimm0_bare, neon_uimm1_bare,
3754 defm : LD1LN_patterns<v1f64, v2f64, f64, neon_uimm0_bare, neon_uimm1_bare,
3757 class NeonI_STN_Lane<bit r, bits<2> op2_1, bit op0, RegisterOperand VList,
3758 Operand ImmOp, string asmop>
3759 : NeonI_LdStOne_Lane<0, r, op2_1, op0,
3760 (outs), (ins GPR64xsp:$Rn, VList:$Rt, ImmOp:$lane),
3761 asmop # "\t$Rt[$lane], [$Rn]",
3765 let neverHasSideEffects = 1;
3766 let hasExtraDefRegAllocReq = 1;
3769 multiclass STN_Lane_BHSD<bit r, bit op0, string List, string asmop> {
3770 def _B : NeonI_STN_Lane<r, 0b00, op0,
3771 !cast<RegisterOperand>(List # "B_operand"),
3772 neon_uimm4_bare, asmop> {
3773 let Inst{12-10} = lane{2-0};
3774 let Inst{30} = lane{3};
3777 def _H : NeonI_STN_Lane<r, 0b01, op0,
3778 !cast<RegisterOperand>(List # "H_operand"),
3779 neon_uimm3_bare, asmop> {
3780 let Inst{12-10} = {lane{1}, lane{0}, 0b0};
3781 let Inst{30} = lane{2};
3784 def _S : NeonI_STN_Lane<r, 0b10, op0,
3785 !cast<RegisterOperand>(List # "S_operand"),
3786 neon_uimm2_bare, asmop> {
3787 let Inst{12-10} = {lane{0}, 0b0, 0b0};
3788 let Inst{30} = lane{1};
3791 def _D : NeonI_STN_Lane<r, 0b10, op0,
3792 !cast<RegisterOperand>(List # "D_operand"),
3793 neon_uimm1_bare, asmop>{
3794 let Inst{12-10} = 0b001;
3795 let Inst{30} = lane{0};
3799 // Store single 1-element structure from one lane of 1 register.
3800 defm ST1LN : STN_Lane_BHSD<0b0, 0b0, "VOne", "st1">;
3802 // Store single N-element structure from one lane of N consecutive registers
3804 defm ST2LN : STN_Lane_BHSD<0b1, 0b0, "VPair", "st2">;
3805 defm ST3LN : STN_Lane_BHSD<0b0, 0b1, "VTriple", "st3">;
3806 defm ST4LN : STN_Lane_BHSD<0b1, 0b1, "VQuad", "st4">;
3808 multiclass ST1LN_patterns<ValueType VTy, ValueType VTy2, ValueType DTy,
3809 Operand ImmOp, Operand ImmOp2, PatFrag StoreOp,
3811 def : Pat<(StoreOp (DTy (vector_extract (VTy VPR64:$Rt), ImmOp:$lane)),
3814 (SUBREG_TO_REG (i64 0), VPR64:$Rt, sub_64),
3817 def : Pat<(StoreOp (DTy (vector_extract (VTy2 VPR128:$Rt), ImmOp2:$lane)),
3819 (INST GPR64xsp:$Rn, VPR128:$Rt, ImmOp2:$lane)>;
3822 // Match all ST1LN instructions
3823 defm : ST1LN_patterns<v8i8, v16i8, i32, neon_uimm3_bare, neon_uimm4_bare,
3824 truncstorei8, ST1LN_B>;
3826 defm : ST1LN_patterns<v4i16, v8i16, i32, neon_uimm2_bare, neon_uimm3_bare,
3827 truncstorei16, ST1LN_H>;
3829 defm : ST1LN_patterns<v2i32, v4i32, i32, neon_uimm1_bare, neon_uimm2_bare,
3831 defm : ST1LN_patterns<v2f32, v4f32, f32, neon_uimm1_bare, neon_uimm2_bare,
3834 defm : ST1LN_patterns<v1i64, v2i64, i64, neon_uimm0_bare, neon_uimm1_bare,
3836 defm : ST1LN_patterns<v1f64, v2f64, f64, neon_uimm0_bare, neon_uimm1_bare,
3839 // End of vector load/store single N-element structure (class SIMD lsone).
3842 // The following are post-index load/store single N-element instructions
3843 // (class SIMD lsone-post)
3845 multiclass NeonI_LDN_WB_Dup<bit q, bit r, bits<3> opcode, bits<2> size,
3846 RegisterOperand VecList, Operand ImmTy,
3848 let mayLoad = 1, neverHasSideEffects = 1, Constraints = "$wb = $Rn",
3849 DecoderMethod = "DecodeVLDSTLanePostInstruction" in {
3850 def _fixed : NeonI_LdOne_Dup_Post<q, r, opcode, size,
3851 (outs VecList:$Rt, GPR64xsp:$wb),
3852 (ins GPR64xsp:$Rn, ImmTy:$amt),
3853 asmop # "\t$Rt, [$Rn], $amt",
3859 def _register : NeonI_LdOne_Dup_Post<q, r, opcode, size,
3860 (outs VecList:$Rt, GPR64xsp:$wb),
3861 (ins GPR64xsp:$Rn, GPR64noxzr:$Rm),
3862 asmop # "\t$Rt, [$Rn], $Rm",
3868 multiclass LDWB_Dup_BHSD<bit r, bits<3> opcode, string List, string asmop,
3869 Operand uimm_b, Operand uimm_h,
3870 Operand uimm_s, Operand uimm_d> {
3871 defm _8B : NeonI_LDN_WB_Dup<0, r, opcode, 0b00,
3872 !cast<RegisterOperand>(List # "8B_operand"),
3875 defm _4H : NeonI_LDN_WB_Dup<0, r, opcode, 0b01,
3876 !cast<RegisterOperand>(List # "4H_operand"),
3879 defm _2S : NeonI_LDN_WB_Dup<0, r, opcode, 0b10,
3880 !cast<RegisterOperand>(List # "2S_operand"),
3883 defm _1D : NeonI_LDN_WB_Dup<0, r, opcode, 0b11,
3884 !cast<RegisterOperand>(List # "1D_operand"),
3887 defm _16B : NeonI_LDN_WB_Dup<1, r, opcode, 0b00,
3888 !cast<RegisterOperand>(List # "16B_operand"),
3891 defm _8H : NeonI_LDN_WB_Dup<1, r, opcode, 0b01,
3892 !cast<RegisterOperand>(List # "8H_operand"),
3895 defm _4S : NeonI_LDN_WB_Dup<1, r, opcode, 0b10,
3896 !cast<RegisterOperand>(List # "4S_operand"),
3899 defm _2D : NeonI_LDN_WB_Dup<1, r, opcode, 0b11,
3900 !cast<RegisterOperand>(List # "2D_operand"),
3904 // Post-index load single 1-element structure to all lanes of 1 register
3905 defm LD1R_WB : LDWB_Dup_BHSD<0b0, 0b110, "VOne", "ld1r", uimm_exact1,
3906 uimm_exact2, uimm_exact4, uimm_exact8>;
3908 // Post-index load single N-element structure to all lanes of N consecutive
3909 // registers (N = 2,3,4)
3910 defm LD2R_WB : LDWB_Dup_BHSD<0b1, 0b110, "VPair", "ld2r", uimm_exact2,
3911 uimm_exact4, uimm_exact8, uimm_exact16>;
3912 defm LD3R_WB : LDWB_Dup_BHSD<0b0, 0b111, "VTriple", "ld3r", uimm_exact3,
3913 uimm_exact6, uimm_exact12, uimm_exact24>;
3914 defm LD4R_WB : LDWB_Dup_BHSD<0b1, 0b111, "VQuad", "ld4r", uimm_exact4,
3915 uimm_exact8, uimm_exact16, uimm_exact32>;
3917 let mayLoad = 1, neverHasSideEffects = 1, hasExtraDefRegAllocReq = 1,
3918 Constraints = "$Rn = $wb, $Rt = $src",
3919 DecoderMethod = "DecodeVLDSTLanePostInstruction" in {
3920 class LDN_WBFx_Lane<bit r, bits<2> op2_1, bit op0, RegisterOperand VList,
3921 Operand ImmTy, Operand ImmOp, string asmop>
3922 : NeonI_LdStOne_Lane_Post<1, r, op2_1, op0,
3923 (outs VList:$Rt, GPR64xsp:$wb),
3924 (ins GPR64xsp:$Rn, ImmTy:$amt,
3925 VList:$src, ImmOp:$lane),
3926 asmop # "\t$Rt[$lane], [$Rn], $amt",
3932 class LDN_WBReg_Lane<bit r, bits<2> op2_1, bit op0, RegisterOperand VList,
3933 Operand ImmTy, Operand ImmOp, string asmop>
3934 : NeonI_LdStOne_Lane_Post<1, r, op2_1, op0,
3935 (outs VList:$Rt, GPR64xsp:$wb),
3936 (ins GPR64xsp:$Rn, GPR64noxzr:$Rm,
3937 VList:$src, ImmOp:$lane),
3938 asmop # "\t$Rt[$lane], [$Rn], $Rm",
3943 multiclass LD_Lane_WB_BHSD<bit r, bit op0, string List, string asmop,
3944 Operand uimm_b, Operand uimm_h,
3945 Operand uimm_s, Operand uimm_d> {
3946 def _B_fixed : LDN_WBFx_Lane<r, 0b00, op0,
3947 !cast<RegisterOperand>(List # "B_operand"),
3948 uimm_b, neon_uimm4_bare, asmop> {
3949 let Inst{12-10} = lane{2-0};
3950 let Inst{30} = lane{3};
3953 def _B_register : LDN_WBReg_Lane<r, 0b00, op0,
3954 !cast<RegisterOperand>(List # "B_operand"),
3955 uimm_b, neon_uimm4_bare, asmop> {
3956 let Inst{12-10} = lane{2-0};
3957 let Inst{30} = lane{3};
3960 def _H_fixed : LDN_WBFx_Lane<r, 0b01, op0,
3961 !cast<RegisterOperand>(List # "H_operand"),
3962 uimm_h, neon_uimm3_bare, asmop> {
3963 let Inst{12-10} = {lane{1}, lane{0}, 0b0};
3964 let Inst{30} = lane{2};
3967 def _H_register : LDN_WBReg_Lane<r, 0b01, op0,
3968 !cast<RegisterOperand>(List # "H_operand"),
3969 uimm_h, neon_uimm3_bare, asmop> {
3970 let Inst{12-10} = {lane{1}, lane{0}, 0b0};
3971 let Inst{30} = lane{2};
3974 def _S_fixed : LDN_WBFx_Lane<r, 0b10, op0,
3975 !cast<RegisterOperand>(List # "S_operand"),
3976 uimm_s, neon_uimm2_bare, asmop> {
3977 let Inst{12-10} = {lane{0}, 0b0, 0b0};
3978 let Inst{30} = lane{1};
3981 def _S_register : LDN_WBReg_Lane<r, 0b10, op0,
3982 !cast<RegisterOperand>(List # "S_operand"),
3983 uimm_s, neon_uimm2_bare, asmop> {
3984 let Inst{12-10} = {lane{0}, 0b0, 0b0};
3985 let Inst{30} = lane{1};
3988 def _D_fixed : LDN_WBFx_Lane<r, 0b10, op0,
3989 !cast<RegisterOperand>(List # "D_operand"),
3990 uimm_d, neon_uimm1_bare, asmop> {
3991 let Inst{12-10} = 0b001;
3992 let Inst{30} = lane{0};
3995 def _D_register : LDN_WBReg_Lane<r, 0b10, op0,
3996 !cast<RegisterOperand>(List # "D_operand"),
3997 uimm_d, neon_uimm1_bare, asmop> {
3998 let Inst{12-10} = 0b001;
3999 let Inst{30} = lane{0};
4003 // Post-index load single 1-element structure to one lane of 1 register.
4004 defm LD1LN_WB : LD_Lane_WB_BHSD<0b0, 0b0, "VOne", "ld1", uimm_exact1,
4005 uimm_exact2, uimm_exact4, uimm_exact8>;
4007 // Post-index load single N-element structure to one lane of N consecutive
4010 defm LD2LN_WB : LD_Lane_WB_BHSD<0b1, 0b0, "VPair", "ld2", uimm_exact2,
4011 uimm_exact4, uimm_exact8, uimm_exact16>;
4012 defm LD3LN_WB : LD_Lane_WB_BHSD<0b0, 0b1, "VTriple", "ld3", uimm_exact3,
4013 uimm_exact6, uimm_exact12, uimm_exact24>;
4014 defm LD4LN_WB : LD_Lane_WB_BHSD<0b1, 0b1, "VQuad", "ld4", uimm_exact4,
4015 uimm_exact8, uimm_exact16, uimm_exact32>;
4017 let mayStore = 1, neverHasSideEffects = 1,
4018 hasExtraDefRegAllocReq = 1, Constraints = "$Rn = $wb",
4019 DecoderMethod = "DecodeVLDSTLanePostInstruction" in {
4020 class STN_WBFx_Lane<bit r, bits<2> op2_1, bit op0, RegisterOperand VList,
4021 Operand ImmTy, Operand ImmOp, string asmop>
4022 : NeonI_LdStOne_Lane_Post<0, r, op2_1, op0,
4023 (outs GPR64xsp:$wb),
4024 (ins GPR64xsp:$Rn, ImmTy:$amt,
4025 VList:$Rt, ImmOp:$lane),
4026 asmop # "\t$Rt[$lane], [$Rn], $amt",
4032 class STN_WBReg_Lane<bit r, bits<2> op2_1, bit op0, RegisterOperand VList,
4033 Operand ImmTy, Operand ImmOp, string asmop>
4034 : NeonI_LdStOne_Lane_Post<0, r, op2_1, op0,
4035 (outs GPR64xsp:$wb),
4036 (ins GPR64xsp:$Rn, GPR64noxzr:$Rm, VList:$Rt,
4038 asmop # "\t$Rt[$lane], [$Rn], $Rm",
4043 multiclass ST_Lane_WB_BHSD<bit r, bit op0, string List, string asmop,
4044 Operand uimm_b, Operand uimm_h,
4045 Operand uimm_s, Operand uimm_d> {
4046 def _B_fixed : STN_WBFx_Lane<r, 0b00, op0,
4047 !cast<RegisterOperand>(List # "B_operand"),
4048 uimm_b, neon_uimm4_bare, asmop> {
4049 let Inst{12-10} = lane{2-0};
4050 let Inst{30} = lane{3};
4053 def _B_register : STN_WBReg_Lane<r, 0b00, op0,
4054 !cast<RegisterOperand>(List # "B_operand"),
4055 uimm_b, neon_uimm4_bare, asmop> {
4056 let Inst{12-10} = lane{2-0};
4057 let Inst{30} = lane{3};
4060 def _H_fixed : STN_WBFx_Lane<r, 0b01, op0,
4061 !cast<RegisterOperand>(List # "H_operand"),
4062 uimm_h, neon_uimm3_bare, asmop> {
4063 let Inst{12-10} = {lane{1}, lane{0}, 0b0};
4064 let Inst{30} = lane{2};
4067 def _H_register : STN_WBReg_Lane<r, 0b01, op0,
4068 !cast<RegisterOperand>(List # "H_operand"),
4069 uimm_h, neon_uimm3_bare, asmop> {
4070 let Inst{12-10} = {lane{1}, lane{0}, 0b0};
4071 let Inst{30} = lane{2};
4074 def _S_fixed : STN_WBFx_Lane<r, 0b10, op0,
4075 !cast<RegisterOperand>(List # "S_operand"),
4076 uimm_s, neon_uimm2_bare, asmop> {
4077 let Inst{12-10} = {lane{0}, 0b0, 0b0};
4078 let Inst{30} = lane{1};
4081 def _S_register : STN_WBReg_Lane<r, 0b10, op0,
4082 !cast<RegisterOperand>(List # "S_operand"),
4083 uimm_s, neon_uimm2_bare, asmop> {
4084 let Inst{12-10} = {lane{0}, 0b0, 0b0};
4085 let Inst{30} = lane{1};
4088 def _D_fixed : STN_WBFx_Lane<r, 0b10, op0,
4089 !cast<RegisterOperand>(List # "D_operand"),
4090 uimm_d, neon_uimm1_bare, asmop> {
4091 let Inst{12-10} = 0b001;
4092 let Inst{30} = lane{0};
4095 def _D_register : STN_WBReg_Lane<r, 0b10, op0,
4096 !cast<RegisterOperand>(List # "D_operand"),
4097 uimm_d, neon_uimm1_bare, asmop> {
4098 let Inst{12-10} = 0b001;
4099 let Inst{30} = lane{0};
4103 // Post-index store single 1-element structure from one lane of 1 register.
4104 defm ST1LN_WB : ST_Lane_WB_BHSD<0b0, 0b0, "VOne", "st1", uimm_exact1,
4105 uimm_exact2, uimm_exact4, uimm_exact8>;
4107 // Post-index store single N-element structure from one lane of N consecutive
4108 // registers (N = 2,3,4)
4109 defm ST2LN_WB : ST_Lane_WB_BHSD<0b1, 0b0, "VPair", "st2", uimm_exact2,
4110 uimm_exact4, uimm_exact8, uimm_exact16>;
4111 defm ST3LN_WB : ST_Lane_WB_BHSD<0b0, 0b1, "VTriple", "st3", uimm_exact3,
4112 uimm_exact6, uimm_exact12, uimm_exact24>;
4113 defm ST4LN_WB : ST_Lane_WB_BHSD<0b1, 0b1, "VQuad", "st4", uimm_exact4,
4114 uimm_exact8, uimm_exact16, uimm_exact32>;
4116 // End of post-index load/store single N-element instructions
4117 // (class SIMD lsone-post)
4119 // Neon Scalar instructions implementation
4120 // Scalar Three Same
4122 class NeonI_Scalar3Same_size<bit u, bits<2> size, bits<5> opcode, string asmop,
4124 : NeonI_Scalar3Same<u, size, opcode,
4125 (outs FPRC:$Rd), (ins FPRC:$Rn, FPRC:$Rm),
4126 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
4130 class NeonI_Scalar3Same_D_size<bit u, bits<5> opcode, string asmop>
4131 : NeonI_Scalar3Same_size<u, 0b11, opcode, asmop, FPR64>;
4133 multiclass NeonI_Scalar3Same_HS_sizes<bit u, bits<5> opcode, string asmop,
4134 bit Commutable = 0> {
4135 let isCommutable = Commutable in {
4136 def hhh : NeonI_Scalar3Same_size<u, 0b01, opcode, asmop, FPR16>;
4137 def sss : NeonI_Scalar3Same_size<u, 0b10, opcode, asmop, FPR32>;
4141 multiclass NeonI_Scalar3Same_SD_sizes<bit u, bit size_high, bits<5> opcode,
4142 string asmop, bit Commutable = 0> {
4143 let isCommutable = Commutable in {
4144 def sss : NeonI_Scalar3Same_size<u, {size_high, 0b0}, opcode, asmop, FPR32>;
4145 def ddd : NeonI_Scalar3Same_size<u, {size_high, 0b1}, opcode, asmop, FPR64>;
4149 multiclass NeonI_Scalar3Same_BHSD_sizes<bit u, bits<5> opcode,
4150 string asmop, bit Commutable = 0> {
4151 let isCommutable = Commutable in {
4152 def bbb : NeonI_Scalar3Same_size<u, 0b00, opcode, asmop, FPR8>;
4153 def hhh : NeonI_Scalar3Same_size<u, 0b01, opcode, asmop, FPR16>;
4154 def sss : NeonI_Scalar3Same_size<u, 0b10, opcode, asmop, FPR32>;
4155 def ddd : NeonI_Scalar3Same_size<u, 0b11, opcode, asmop, FPR64>;
4159 multiclass Neon_Scalar3Same_D_size_patterns<SDPatternOperator opnode,
4160 Instruction INSTD> {
4161 def : Pat<(v1i64 (opnode (v1i64 FPR64:$Rn), (v1i64 FPR64:$Rm))),
4162 (INSTD FPR64:$Rn, FPR64:$Rm)>;
4165 multiclass Neon_Scalar3Same_BHSD_size_patterns<SDPatternOperator opnode,
4170 : Neon_Scalar3Same_D_size_patterns<opnode, INSTD> {
4171 def: Pat<(v1i8 (opnode (v1i8 FPR8:$Rn), (v1i8 FPR8:$Rm))),
4172 (INSTB FPR8:$Rn, FPR8:$Rm)>;
4173 def: Pat<(v1i16 (opnode (v1i16 FPR16:$Rn), (v1i16 FPR16:$Rm))),
4174 (INSTH FPR16:$Rn, FPR16:$Rm)>;
4175 def: Pat<(v1i32 (opnode (v1i32 FPR32:$Rn), (v1i32 FPR32:$Rm))),
4176 (INSTS FPR32:$Rn, FPR32:$Rm)>;
4179 multiclass Neon_Scalar3Same_HS_size_patterns<SDPatternOperator opnode,
4181 Instruction INSTS> {
4182 def : Pat<(v1i16 (opnode (v1i16 FPR16:$Rn), (v1i16 FPR16:$Rm))),
4183 (INSTH FPR16:$Rn, FPR16:$Rm)>;
4184 def : Pat<(v1i32 (opnode (v1i32 FPR32:$Rn), (v1i32 FPR32:$Rm))),
4185 (INSTS FPR32:$Rn, FPR32:$Rm)>;
4188 multiclass Neon_Scalar3Same_SD_size_patterns<SDPatternOperator opnode,
4189 ValueType SResTy, ValueType STy,
4190 Instruction INSTS, ValueType DResTy,
4191 ValueType DTy, Instruction INSTD> {
4192 def : Pat<(SResTy (opnode (STy FPR32:$Rn), (STy FPR32:$Rm))),
4193 (INSTS FPR32:$Rn, FPR32:$Rm)>;
4194 def : Pat<(DResTy (opnode (DTy FPR64:$Rn), (DTy FPR64:$Rm))),
4195 (INSTD FPR64:$Rn, FPR64:$Rm)>;
4198 class Neon_Scalar3Same_cmp_V1_D_size_patterns<CondCode CC,
4200 : Pat<(v1i64 (Neon_cmp (v1f64 FPR64:$Rn), (v1f64 FPR64:$Rm), CC)),
4201 (INSTD FPR64:$Rn, FPR64:$Rm)>;
4203 // Scalar Three Different
4205 class NeonI_Scalar3Diff_size<bit u, bits<2> size, bits<4> opcode, string asmop,
4206 RegisterClass FPRCD, RegisterClass FPRCS>
4207 : NeonI_Scalar3Diff<u, size, opcode,
4208 (outs FPRCD:$Rd), (ins FPRCS:$Rn, FPRCS:$Rm),
4209 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
4213 multiclass NeonI_Scalar3Diff_HS_size<bit u, bits<4> opcode, string asmop> {
4214 def shh : NeonI_Scalar3Diff_size<u, 0b01, opcode, asmop, FPR32, FPR16>;
4215 def dss : NeonI_Scalar3Diff_size<u, 0b10, opcode, asmop, FPR64, FPR32>;
4218 multiclass NeonI_Scalar3Diff_ml_HS_size<bit u, bits<4> opcode, string asmop> {
4219 let Constraints = "$Src = $Rd" in {
4220 def shh : NeonI_Scalar3Diff<u, 0b01, opcode,
4221 (outs FPR32:$Rd), (ins FPR32:$Src, FPR16:$Rn, FPR16:$Rm),
4222 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
4225 def dss : NeonI_Scalar3Diff<u, 0b10, opcode,
4226 (outs FPR64:$Rd), (ins FPR64:$Src, FPR32:$Rn, FPR32:$Rm),
4227 !strconcat(asmop, "\t$Rd, $Rn, $Rm"),
4233 multiclass Neon_Scalar3Diff_HS_size_patterns<SDPatternOperator opnode,
4235 Instruction INSTS> {
4236 def : Pat<(v1i32 (opnode (v1i16 FPR16:$Rn), (v1i16 FPR16:$Rm))),
4237 (INSTH FPR16:$Rn, FPR16:$Rm)>;
4238 def : Pat<(v1i64 (opnode (v1i32 FPR32:$Rn), (v1i32 FPR32:$Rm))),
4239 (INSTS FPR32:$Rn, FPR32:$Rm)>;
4242 multiclass Neon_Scalar3Diff_ml_HS_size_patterns<SDPatternOperator opnode,
4244 Instruction INSTS> {
4245 def : Pat<(v1i32 (opnode (v1i32 FPR32:$Src), (v1i16 FPR16:$Rn), (v1i16 FPR16:$Rm))),
4246 (INSTH FPR32:$Src, FPR16:$Rn, FPR16:$Rm)>;
4247 def : Pat<(v1i64 (opnode (v1i64 FPR64:$Src), (v1i32 FPR32:$Rn), (v1i32 FPR32:$Rm))),
4248 (INSTS FPR64:$Src, FPR32:$Rn, FPR32:$Rm)>;
4251 // Scalar Two Registers Miscellaneous
4253 class NeonI_Scalar2SameMisc_size<bit u, bits<2> size, bits<5> opcode, string asmop,
4254 RegisterClass FPRCD, RegisterClass FPRCS>
4255 : NeonI_Scalar2SameMisc<u, size, opcode,
4256 (outs FPRCD:$Rd), (ins FPRCS:$Rn),
4257 !strconcat(asmop, "\t$Rd, $Rn"),
4261 multiclass NeonI_Scalar2SameMisc_SD_size<bit u, bit size_high, bits<5> opcode,
4263 def ss : NeonI_Scalar2SameMisc_size<u, {size_high, 0b0}, opcode, asmop, FPR32,
4265 def dd : NeonI_Scalar2SameMisc_size<u, {size_high, 0b1}, opcode, asmop, FPR64,
4269 multiclass NeonI_Scalar2SameMisc_D_size<bit u, bits<5> opcode, string asmop> {
4270 def dd : NeonI_Scalar2SameMisc_size<u, 0b11, opcode, asmop, FPR64, FPR64>;
4273 multiclass NeonI_Scalar2SameMisc_BHSD_size<bit u, bits<5> opcode, string asmop>
4274 : NeonI_Scalar2SameMisc_D_size<u, opcode, asmop> {
4275 def bb : NeonI_Scalar2SameMisc_size<u, 0b00, opcode, asmop, FPR8, FPR8>;
4276 def hh : NeonI_Scalar2SameMisc_size<u, 0b01, opcode, asmop, FPR16, FPR16>;
4277 def ss : NeonI_Scalar2SameMisc_size<u, 0b10, opcode, asmop, FPR32, FPR32>;
4280 class NeonI_Scalar2SameMisc_fcvtxn_D_size<bit u, bits<5> opcode, string asmop>
4281 : NeonI_Scalar2SameMisc_size<u, 0b01, opcode, asmop, FPR32, FPR64>;
4283 multiclass NeonI_Scalar2SameMisc_narrow_HSD_size<bit u, bits<5> opcode,
4285 def bh : NeonI_Scalar2SameMisc_size<u, 0b00, opcode, asmop, FPR8, FPR16>;
4286 def hs : NeonI_Scalar2SameMisc_size<u, 0b01, opcode, asmop, FPR16, FPR32>;
4287 def sd : NeonI_Scalar2SameMisc_size<u, 0b10, opcode, asmop, FPR32, FPR64>;
4290 class NeonI_Scalar2SameMisc_accum_size<bit u, bits<2> size, bits<5> opcode,
4291 string asmop, RegisterClass FPRC>
4292 : NeonI_Scalar2SameMisc<u, size, opcode,
4293 (outs FPRC:$Rd), (ins FPRC:$Src, FPRC:$Rn),
4294 !strconcat(asmop, "\t$Rd, $Rn"),
4298 multiclass NeonI_Scalar2SameMisc_accum_BHSD_size<bit u, bits<5> opcode,
4301 let Constraints = "$Src = $Rd" in {
4302 def bb : NeonI_Scalar2SameMisc_accum_size<u, 0b00, opcode, asmop, FPR8>;
4303 def hh : NeonI_Scalar2SameMisc_accum_size<u, 0b01, opcode, asmop, FPR16>;
4304 def ss : NeonI_Scalar2SameMisc_accum_size<u, 0b10, opcode, asmop, FPR32>;
4305 def dd : NeonI_Scalar2SameMisc_accum_size<u, 0b11, opcode, asmop, FPR64>;
4309 class Neon_Scalar2SameMisc_fcvtxn_D_size_patterns<SDPatternOperator opnode,
4311 : Pat<(f32 (opnode (f64 FPR64:$Rn))),
4314 multiclass Neon_Scalar2SameMisc_fcvt_SD_size_patterns<SDPatternOperator opnode,
4316 Instruction INSTD> {
4317 def : Pat<(v1i32 (opnode (f32 FPR32:$Rn))),
4319 def : Pat<(v1i64 (opnode (f64 FPR64:$Rn))),
4323 class Neon_Scalar2SameMisc_vcvt_D_size_patterns<SDPatternOperator opnode,
4325 : Pat<(v1i64 (opnode (v1f64 FPR64:$Rn))),
4328 multiclass Neon_Scalar2SameMisc_cvt_SD_size_patterns<SDPatternOperator opnode,
4330 Instruction INSTD> {
4331 def : Pat<(f32 (opnode (v1i32 FPR32:$Rn))),
4333 def : Pat<(f64 (opnode (v1i64 FPR64:$Rn))),
4337 multiclass Neon_Scalar2SameMisc_SD_size_patterns<SDPatternOperator opnode,
4339 Instruction INSTD> {
4340 def : Pat<(f32 (opnode (f32 FPR32:$Rn))),
4342 def : Pat<(f64 (opnode (f64 FPR64:$Rn))),
4346 class Neon_Scalar2SameMisc_V1_D_size_patterns<SDPatternOperator opnode,
4348 : Pat<(v1f64 (opnode (v1f64 FPR64:$Rn))),
4351 class NeonI_Scalar2SameMisc_cmpz_D_size<bit u, bits<5> opcode, string asmop>
4352 : NeonI_Scalar2SameMisc<u, 0b11, opcode,
4353 (outs FPR64:$Rd), (ins FPR64:$Rn, neon_uimm0:$Imm),
4354 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
4358 multiclass NeonI_Scalar2SameMisc_cmpz_SD_size<bit u, bits<5> opcode,
4360 def ssi : NeonI_Scalar2SameMisc<u, 0b10, opcode,
4361 (outs FPR32:$Rd), (ins FPR32:$Rn, fpz32:$FPImm),
4362 !strconcat(asmop, "\t$Rd, $Rn, $FPImm"),
4365 def ddi : NeonI_Scalar2SameMisc<u, 0b11, opcode,
4366 (outs FPR64:$Rd), (ins FPR64:$Rn, fpz32:$FPImm),
4367 !strconcat(asmop, "\t$Rd, $Rn, $FPImm"),
4372 class Neon_Scalar2SameMisc_cmpz_D_size_patterns<SDPatternOperator opnode,
4374 : Pat<(v1i64 (opnode (v1i64 FPR64:$Rn),
4375 (v1i64 (bitconvert (v8i8 Neon_AllZero))))),
4376 (INSTD FPR64:$Rn, 0)>;
4378 class Neon_Scalar2SameMisc_cmpz_D_V1_size_patterns<CondCode CC,
4380 : Pat<(v1i64 (Neon_cmpz (v1i64 FPR64:$Rn),
4381 (i32 neon_uimm0:$Imm), CC)),
4382 (INSTD FPR64:$Rn, neon_uimm0:$Imm)>;
4384 multiclass Neon_Scalar2SameMisc_cmpz_SD_size_patterns<SDPatternOperator opnode,
4386 Instruction INSTD> {
4387 def : Pat<(v1i32 (opnode (f32 FPR32:$Rn), (f32 fpz32:$FPImm))),
4388 (INSTS FPR32:$Rn, fpz32:$FPImm)>;
4389 def : Pat<(v1i64 (opnode (f64 FPR64:$Rn), (f32 fpz32:$FPImm))),
4390 (INSTD FPR64:$Rn, fpz32:$FPImm)>;
4393 multiclass Neon_Scalar2SameMisc_D_size_patterns<SDPatternOperator opnode,
4394 Instruction INSTD> {
4395 def : Pat<(v1i64 (opnode (v1i64 FPR64:$Rn))),
4399 multiclass Neon_Scalar2SameMisc_BHSD_size_patterns<SDPatternOperator opnode,
4404 : Neon_Scalar2SameMisc_D_size_patterns<opnode, INSTD> {
4405 def : Pat<(v1i8 (opnode (v1i8 FPR8:$Rn))),
4407 def : Pat<(v1i16 (opnode (v1i16 FPR16:$Rn))),
4409 def : Pat<(v1i32 (opnode (v1i32 FPR32:$Rn))),
4413 multiclass Neon_Scalar2SameMisc_narrow_HSD_size_patterns<
4414 SDPatternOperator opnode,
4417 Instruction INSTD> {
4418 def : Pat<(v1i8 (opnode (v1i16 FPR16:$Rn))),
4420 def : Pat<(v1i16 (opnode (v1i32 FPR32:$Rn))),
4422 def : Pat<(v1i32 (opnode (v1i64 FPR64:$Rn))),
4427 multiclass Neon_Scalar2SameMisc_accum_BHSD_size_patterns<
4428 SDPatternOperator opnode,
4432 Instruction INSTD> {
4433 def : Pat<(v1i8 (opnode (v1i8 FPR8:$Src), (v1i8 FPR8:$Rn))),
4434 (INSTB FPR8:$Src, FPR8:$Rn)>;
4435 def : Pat<(v1i16 (opnode (v1i16 FPR16:$Src), (v1i16 FPR16:$Rn))),
4436 (INSTH FPR16:$Src, FPR16:$Rn)>;
4437 def : Pat<(v1i32 (opnode (v1i32 FPR32:$Src), (v1i32 FPR32:$Rn))),
4438 (INSTS FPR32:$Src, FPR32:$Rn)>;
4439 def : Pat<(v1i64 (opnode (v1i64 FPR64:$Src), (v1i64 FPR64:$Rn))),
4440 (INSTD FPR64:$Src, FPR64:$Rn)>;
4443 // Scalar Shift By Immediate
4445 class NeonI_ScalarShiftImm_size<bit u, bits<5> opcode, string asmop,
4446 RegisterClass FPRC, Operand ImmTy>
4447 : NeonI_ScalarShiftImm<u, opcode,
4448 (outs FPRC:$Rd), (ins FPRC:$Rn, ImmTy:$Imm),
4449 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
4452 multiclass NeonI_ScalarShiftRightImm_D_size<bit u, bits<5> opcode,
4454 def ddi : NeonI_ScalarShiftImm_size<u, opcode, asmop, FPR64, shr_imm64> {
4456 let Inst{22} = 0b1; // immh:immb = 1xxxxxx
4457 let Inst{21-16} = Imm;
4461 multiclass NeonI_ScalarShiftRightImm_BHSD_size<bit u, bits<5> opcode,
4463 : NeonI_ScalarShiftRightImm_D_size<u, opcode, asmop> {
4464 def bbi : NeonI_ScalarShiftImm_size<u, opcode, asmop, FPR8, shr_imm8> {
4466 let Inst{22-19} = 0b0001; // immh:immb = 0001xxx
4467 let Inst{18-16} = Imm;
4469 def hhi : NeonI_ScalarShiftImm_size<u, opcode, asmop, FPR16, shr_imm16> {
4471 let Inst{22-20} = 0b001; // immh:immb = 001xxxx
4472 let Inst{19-16} = Imm;
4474 def ssi : NeonI_ScalarShiftImm_size<u, opcode, asmop, FPR32, shr_imm32> {
4476 let Inst{22-21} = 0b01; // immh:immb = 01xxxxx
4477 let Inst{20-16} = Imm;
4481 multiclass NeonI_ScalarShiftLeftImm_D_size<bit u, bits<5> opcode,
4483 def ddi : NeonI_ScalarShiftImm_size<u, opcode, asmop, FPR64, shl_imm64> {
4485 let Inst{22} = 0b1; // immh:immb = 1xxxxxx
4486 let Inst{21-16} = Imm;
4490 multiclass NeonI_ScalarShiftLeftImm_BHSD_size<bit u, bits<5> opcode,
4492 : NeonI_ScalarShiftLeftImm_D_size<u, opcode, asmop> {
4493 def bbi : NeonI_ScalarShiftImm_size<u, opcode, asmop, FPR8, shl_imm8> {
4495 let Inst{22-19} = 0b0001; // immh:immb = 0001xxx
4496 let Inst{18-16} = Imm;
4498 def hhi : NeonI_ScalarShiftImm_size<u, opcode, asmop, FPR16, shl_imm16> {
4500 let Inst{22-20} = 0b001; // immh:immb = 001xxxx
4501 let Inst{19-16} = Imm;
4503 def ssi : NeonI_ScalarShiftImm_size<u, opcode, asmop, FPR32, shl_imm32> {
4505 let Inst{22-21} = 0b01; // immh:immb = 01xxxxx
4506 let Inst{20-16} = Imm;
4510 class NeonI_ScalarShiftRightImm_accum_D_size<bit u, bits<5> opcode, string asmop>
4511 : NeonI_ScalarShiftImm<u, opcode,
4513 (ins FPR64:$Src, FPR64:$Rn, shr_imm64:$Imm),
4514 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
4517 let Inst{22} = 0b1; // immh:immb = 1xxxxxx
4518 let Inst{21-16} = Imm;
4519 let Constraints = "$Src = $Rd";
4522 class NeonI_ScalarShiftLeftImm_accum_D_size<bit u, bits<5> opcode, string asmop>
4523 : NeonI_ScalarShiftImm<u, opcode,
4525 (ins FPR64:$Src, FPR64:$Rn, shl_imm64:$Imm),
4526 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
4529 let Inst{22} = 0b1; // immh:immb = 1xxxxxx
4530 let Inst{21-16} = Imm;
4531 let Constraints = "$Src = $Rd";
4534 class NeonI_ScalarShiftImm_narrow_size<bit u, bits<5> opcode, string asmop,
4535 RegisterClass FPRCD, RegisterClass FPRCS,
4537 : NeonI_ScalarShiftImm<u, opcode,
4538 (outs FPRCD:$Rd), (ins FPRCS:$Rn, ImmTy:$Imm),
4539 !strconcat(asmop, "\t$Rd, $Rn, $Imm"),
4542 multiclass NeonI_ScalarShiftImm_narrow_HSD_size<bit u, bits<5> opcode,
4544 def bhi : NeonI_ScalarShiftImm_narrow_size<u, opcode, asmop, FPR8, FPR16,
4547 let Inst{22-19} = 0b0001; // immh:immb = 0001xxx
4548 let Inst{18-16} = Imm;
4550 def hsi : NeonI_ScalarShiftImm_narrow_size<u, opcode, asmop, FPR16, FPR32,
4553 let Inst{22-20} = 0b001; // immh:immb = 001xxxx
4554 let Inst{19-16} = Imm;
4556 def sdi : NeonI_ScalarShiftImm_narrow_size<u, opcode, asmop, FPR32, FPR64,
4559 let Inst{22-21} = 0b01; // immh:immb = 01xxxxx
4560 let Inst{20-16} = Imm;
4564 multiclass NeonI_ScalarShiftImm_cvt_SD_size<bit u, bits<5> opcode, string asmop> {
4565 def ssi : NeonI_ScalarShiftImm_size<u, opcode, asmop, FPR32, shr_imm32> {
4567 let Inst{22-21} = 0b01; // immh:immb = 01xxxxx
4568 let Inst{20-16} = Imm;
4570 def ddi : NeonI_ScalarShiftImm_size<u, opcode, asmop, FPR64, shr_imm64> {
4572 let Inst{22} = 0b1; // immh:immb = 1xxxxxx
4573 let Inst{21-16} = Imm;
4577 multiclass Neon_ScalarShiftRImm_D_size_patterns<SDPatternOperator opnode,
4578 Instruction INSTD> {
4579 def ddi : Pat<(v1i64 (opnode (v1i64 FPR64:$Rn), (i32 shr_imm64:$Imm))),
4580 (INSTD FPR64:$Rn, imm:$Imm)>;
4583 multiclass Neon_ScalarShiftLImm_D_size_patterns<SDPatternOperator opnode,
4584 Instruction INSTD> {
4585 def ddi : Pat<(v1i64 (opnode (v1i64 FPR64:$Rn), (i32 shl_imm64:$Imm))),
4586 (INSTD FPR64:$Rn, imm:$Imm)>;
4589 class Neon_ScalarShiftImm_V1_D_size_patterns<SDPatternOperator opnode,
4591 : Pat<(v1i64 (opnode (v1i64 FPR64:$Rn),
4592 (v1i64 (Neon_vdup (i32 shr_imm64:$Imm))))),
4593 (INSTD FPR64:$Rn, imm:$Imm)>;
4595 multiclass Neon_ScalarShiftLImm_BHSD_size_patterns<SDPatternOperator opnode,
4600 : Neon_ScalarShiftLImm_D_size_patterns<opnode, INSTD> {
4601 def bbi : Pat<(v1i8 (opnode (v1i8 FPR8:$Rn), (i32 shl_imm8:$Imm))),
4602 (INSTB FPR8:$Rn, imm:$Imm)>;
4603 def hhi : Pat<(v1i16 (opnode (v1i16 FPR16:$Rn), (i32 shl_imm16:$Imm))),
4604 (INSTH FPR16:$Rn, imm:$Imm)>;
4605 def ssi : Pat<(v1i32 (opnode (v1i32 FPR32:$Rn), (i32 shl_imm32:$Imm))),
4606 (INSTS FPR32:$Rn, imm:$Imm)>;
4609 class Neon_ScalarShiftLImm_accum_D_size_patterns<SDPatternOperator opnode,
4611 : Pat<(v1i64 (opnode (v1i64 FPR64:$Src), (v1i64 FPR64:$Rn),
4612 (i32 shl_imm64:$Imm))),
4613 (INSTD FPR64:$Src, FPR64:$Rn, imm:$Imm)>;
4615 class Neon_ScalarShiftRImm_accum_D_size_patterns<SDPatternOperator opnode,
4617 : Pat<(v1i64 (opnode (v1i64 FPR64:$Src), (v1i64 FPR64:$Rn),
4618 (i32 shr_imm64:$Imm))),
4619 (INSTD FPR64:$Src, FPR64:$Rn, imm:$Imm)>;
4621 multiclass Neon_ScalarShiftImm_narrow_HSD_size_patterns<
4622 SDPatternOperator opnode,
4625 Instruction INSTD> {
4626 def bhi : Pat<(v1i8 (opnode (v1i16 FPR16:$Rn), (i32 shr_imm16:$Imm))),
4627 (INSTH FPR16:$Rn, imm:$Imm)>;
4628 def hsi : Pat<(v1i16 (opnode (v1i32 FPR32:$Rn), (i32 shr_imm32:$Imm))),
4629 (INSTS FPR32:$Rn, imm:$Imm)>;
4630 def sdi : Pat<(v1i32 (opnode (v1i64 FPR64:$Rn), (i32 shr_imm64:$Imm))),
4631 (INSTD FPR64:$Rn, imm:$Imm)>;
4634 multiclass Neon_ScalarShiftImm_scvtf_SD_size_patterns<SDPatternOperator opnode,
4636 Instruction INSTD> {
4637 def ssi : Pat<(f32 (opnode (v1i32 FPR32:$Rn), (i32 shr_imm32:$Imm))),
4638 (INSTS FPR32:$Rn, imm:$Imm)>;
4639 def ddi : Pat<(f64 (opnode (v1i64 FPR64:$Rn), (i32 shr_imm64:$Imm))),
4640 (INSTD FPR64:$Rn, imm:$Imm)>;
4643 multiclass Neon_ScalarShiftImm_fcvts_SD_size_patterns<SDPatternOperator opnode,
4645 Instruction INSTD> {
4646 def ssi : Pat<(v1i32 (opnode (f32 FPR32:$Rn), (i32 shr_imm32:$Imm))),
4647 (INSTS FPR32:$Rn, imm:$Imm)>;
4648 def ddi : Pat<(v1i64 (opnode (f64 FPR64:$Rn), (i32 shr_imm64:$Imm))),
4649 (INSTD FPR64:$Rn, imm:$Imm)>;
4652 // Scalar Signed Shift Right (Immediate)
4653 defm SSHR : NeonI_ScalarShiftRightImm_D_size<0b0, 0b00000, "sshr">;
4654 defm : Neon_ScalarShiftRImm_D_size_patterns<int_aarch64_neon_vshrds_n, SSHRddi>;
4655 // Pattern to match llvm.arm.* intrinsic.
4656 def : Neon_ScalarShiftImm_V1_D_size_patterns<sra, SSHRddi>;
4658 // Scalar Unsigned Shift Right (Immediate)
4659 defm USHR : NeonI_ScalarShiftRightImm_D_size<0b1, 0b00000, "ushr">;
4660 defm : Neon_ScalarShiftRImm_D_size_patterns<int_aarch64_neon_vshrdu_n, USHRddi>;
4661 // Pattern to match llvm.arm.* intrinsic.
4662 def : Neon_ScalarShiftImm_V1_D_size_patterns<srl, USHRddi>;
4664 // Scalar Signed Rounding Shift Right (Immediate)
4665 defm SRSHR : NeonI_ScalarShiftRightImm_D_size<0b0, 0b00100, "srshr">;
4666 defm : Neon_ScalarShiftRImm_D_size_patterns<int_aarch64_neon_vsrshr, SRSHRddi>;
4668 // Scalar Unigned Rounding Shift Right (Immediate)
4669 defm URSHR : NeonI_ScalarShiftRightImm_D_size<0b1, 0b00100, "urshr">;
4670 defm : Neon_ScalarShiftRImm_D_size_patterns<int_aarch64_neon_vurshr, URSHRddi>;
4672 // Scalar Signed Shift Right and Accumulate (Immediate)
4673 def SSRA : NeonI_ScalarShiftRightImm_accum_D_size<0b0, 0b00010, "ssra">;
4674 def : Neon_ScalarShiftRImm_accum_D_size_patterns
4675 <int_aarch64_neon_vsrads_n, SSRA>;
4677 // Scalar Unsigned Shift Right and Accumulate (Immediate)
4678 def USRA : NeonI_ScalarShiftRightImm_accum_D_size<0b1, 0b00010, "usra">;
4679 def : Neon_ScalarShiftRImm_accum_D_size_patterns
4680 <int_aarch64_neon_vsradu_n, USRA>;
4682 // Scalar Signed Rounding Shift Right and Accumulate (Immediate)
4683 def SRSRA : NeonI_ScalarShiftRightImm_accum_D_size<0b0, 0b00110, "srsra">;
4684 def : Neon_ScalarShiftRImm_accum_D_size_patterns
4685 <int_aarch64_neon_vrsrads_n, SRSRA>;
4687 // Scalar Unsigned Rounding Shift Right and Accumulate (Immediate)
4688 def URSRA : NeonI_ScalarShiftRightImm_accum_D_size<0b1, 0b00110, "ursra">;
4689 def : Neon_ScalarShiftRImm_accum_D_size_patterns
4690 <int_aarch64_neon_vrsradu_n, URSRA>;
4692 // Scalar Shift Left (Immediate)
4693 defm SHL : NeonI_ScalarShiftLeftImm_D_size<0b0, 0b01010, "shl">;
4694 defm : Neon_ScalarShiftLImm_D_size_patterns<int_aarch64_neon_vshld_n, SHLddi>;
4695 // Pattern to match llvm.arm.* intrinsic.
4696 def : Neon_ScalarShiftImm_V1_D_size_patterns<shl, SHLddi>;
4698 // Signed Saturating Shift Left (Immediate)
4699 defm SQSHL : NeonI_ScalarShiftLeftImm_BHSD_size<0b0, 0b01110, "sqshl">;
4700 defm : Neon_ScalarShiftLImm_BHSD_size_patterns<int_aarch64_neon_vqshls_n,
4702 SQSHLssi, SQSHLddi>;
4703 // Pattern to match llvm.arm.* intrinsic.
4704 defm : Neon_ScalarShiftLImm_D_size_patterns<Neon_sqrshlImm, SQSHLddi>;
4706 // Unsigned Saturating Shift Left (Immediate)
4707 defm UQSHL : NeonI_ScalarShiftLeftImm_BHSD_size<0b1, 0b01110, "uqshl">;
4708 defm : Neon_ScalarShiftLImm_BHSD_size_patterns<int_aarch64_neon_vqshlu_n,
4710 UQSHLssi, UQSHLddi>;
4711 // Pattern to match llvm.arm.* intrinsic.
4712 defm : Neon_ScalarShiftLImm_D_size_patterns<Neon_uqrshlImm, UQSHLddi>;
4714 // Signed Saturating Shift Left Unsigned (Immediate)
4715 defm SQSHLU : NeonI_ScalarShiftLeftImm_BHSD_size<0b1, 0b01100, "sqshlu">;
4716 defm : Neon_ScalarShiftLImm_BHSD_size_patterns<int_aarch64_neon_vsqshlu,
4717 SQSHLUbbi, SQSHLUhhi,
4718 SQSHLUssi, SQSHLUddi>;
4720 // Shift Right And Insert (Immediate)
4721 def SRI : NeonI_ScalarShiftRightImm_accum_D_size<0b1, 0b01000, "sri">;
4722 def : Neon_ScalarShiftRImm_accum_D_size_patterns
4723 <int_aarch64_neon_vsri, SRI>;
4725 // Shift Left And Insert (Immediate)
4726 def SLI : NeonI_ScalarShiftLeftImm_accum_D_size<0b1, 0b01010, "sli">;
4727 def : Neon_ScalarShiftLImm_accum_D_size_patterns
4728 <int_aarch64_neon_vsli, SLI>;
4730 // Signed Saturating Shift Right Narrow (Immediate)
4731 defm SQSHRN : NeonI_ScalarShiftImm_narrow_HSD_size<0b0, 0b10010, "sqshrn">;
4732 defm : Neon_ScalarShiftImm_narrow_HSD_size_patterns<int_aarch64_neon_vsqshrn,
4733 SQSHRNbhi, SQSHRNhsi,
4736 // Unsigned Saturating Shift Right Narrow (Immediate)
4737 defm UQSHRN : NeonI_ScalarShiftImm_narrow_HSD_size<0b1, 0b10010, "uqshrn">;
4738 defm : Neon_ScalarShiftImm_narrow_HSD_size_patterns<int_aarch64_neon_vuqshrn,
4739 UQSHRNbhi, UQSHRNhsi,
4742 // Signed Saturating Rounded Shift Right Narrow (Immediate)
4743 defm SQRSHRN : NeonI_ScalarShiftImm_narrow_HSD_size<0b0, 0b10011, "sqrshrn">;
4744 defm : Neon_ScalarShiftImm_narrow_HSD_size_patterns<int_aarch64_neon_vsqrshrn,
4745 SQRSHRNbhi, SQRSHRNhsi,
4748 // Unsigned Saturating Rounded Shift Right Narrow (Immediate)
4749 defm UQRSHRN : NeonI_ScalarShiftImm_narrow_HSD_size<0b1, 0b10011, "uqrshrn">;
4750 defm : Neon_ScalarShiftImm_narrow_HSD_size_patterns<int_aarch64_neon_vuqrshrn,
4751 UQRSHRNbhi, UQRSHRNhsi,
4754 // Signed Saturating Shift Right Unsigned Narrow (Immediate)
4755 defm SQSHRUN : NeonI_ScalarShiftImm_narrow_HSD_size<0b1, 0b10000, "sqshrun">;
4756 defm : Neon_ScalarShiftImm_narrow_HSD_size_patterns<int_aarch64_neon_vsqshrun,
4757 SQSHRUNbhi, SQSHRUNhsi,
4760 // Signed Saturating Rounded Shift Right Unsigned Narrow (Immediate)
4761 defm SQRSHRUN : NeonI_ScalarShiftImm_narrow_HSD_size<0b1, 0b10001, "sqrshrun">;
4762 defm : Neon_ScalarShiftImm_narrow_HSD_size_patterns<int_aarch64_neon_vsqrshrun,
4763 SQRSHRUNbhi, SQRSHRUNhsi,
4766 // Scalar Signed Fixed-point Convert To Floating-Point (Immediate)
4767 defm SCVTF_N : NeonI_ScalarShiftImm_cvt_SD_size<0b0, 0b11100, "scvtf">;
4768 defm : Neon_ScalarShiftImm_scvtf_SD_size_patterns<int_aarch64_neon_vcvtfxs2fp_n,
4769 SCVTF_Nssi, SCVTF_Nddi>;
4771 // Scalar Unsigned Fixed-point Convert To Floating-Point (Immediate)
4772 defm UCVTF_N : NeonI_ScalarShiftImm_cvt_SD_size<0b1, 0b11100, "ucvtf">;
4773 defm : Neon_ScalarShiftImm_scvtf_SD_size_patterns<int_aarch64_neon_vcvtfxu2fp_n,
4774 UCVTF_Nssi, UCVTF_Nddi>;
4776 // Scalar Floating-point Convert To Signed Fixed-point (Immediate)
4777 defm FCVTZS_N : NeonI_ScalarShiftImm_cvt_SD_size<0b0, 0b11111, "fcvtzs">;
4778 defm : Neon_ScalarShiftImm_fcvts_SD_size_patterns<int_aarch64_neon_vcvtfp2fxs_n,
4779 FCVTZS_Nssi, FCVTZS_Nddi>;
4781 // Scalar Floating-point Convert To Unsigned Fixed-point (Immediate)
4782 defm FCVTZU_N : NeonI_ScalarShiftImm_cvt_SD_size<0b1, 0b11111, "fcvtzu">;
4783 defm : Neon_ScalarShiftImm_fcvts_SD_size_patterns<int_aarch64_neon_vcvtfp2fxu_n,
4784 FCVTZU_Nssi, FCVTZU_Nddi>;
4786 // Patterns For Convert Instructions Between v1f64 and v1i64
4787 class Neon_ScalarShiftImm_cvtf_v1f64_pattern<SDPatternOperator opnode,
4789 : Pat<(v1f64 (opnode (v1i64 FPR64:$Rn), (i32 shr_imm64:$Imm))),
4790 (INST FPR64:$Rn, imm:$Imm)>;
4792 class Neon_ScalarShiftImm_fcvt_v1f64_pattern<SDPatternOperator opnode,
4794 : Pat<(v1i64 (opnode (v1f64 FPR64:$Rn), (i32 shr_imm64:$Imm))),
4795 (INST FPR64:$Rn, imm:$Imm)>;
4797 def : Neon_ScalarShiftImm_cvtf_v1f64_pattern<int_arm_neon_vcvtfxs2fp,
4800 def : Neon_ScalarShiftImm_cvtf_v1f64_pattern<int_arm_neon_vcvtfxu2fp,
4803 def : Neon_ScalarShiftImm_fcvt_v1f64_pattern<int_arm_neon_vcvtfp2fxs,
4806 def : Neon_ScalarShiftImm_fcvt_v1f64_pattern<int_arm_neon_vcvtfp2fxu,
4809 // Scalar Integer Add
4810 let isCommutable = 1 in {
4811 def ADDddd : NeonI_Scalar3Same_D_size<0b0, 0b10000, "add">;
4814 // Scalar Integer Sub
4815 def SUBddd : NeonI_Scalar3Same_D_size<0b1, 0b10000, "sub">;
4817 // Pattern for Scalar Integer Add and Sub with D register only
4818 defm : Neon_Scalar3Same_D_size_patterns<add, ADDddd>;
4819 defm : Neon_Scalar3Same_D_size_patterns<sub, SUBddd>;
4821 // Patterns to match llvm.aarch64.* intrinsic for Scalar Add, Sub
4822 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vaddds, ADDddd>;
4823 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vadddu, ADDddd>;
4824 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vsubds, SUBddd>;
4825 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vsubdu, SUBddd>;
4827 // Scalar Integer Saturating Add (Signed, Unsigned)
4828 defm SQADD : NeonI_Scalar3Same_BHSD_sizes<0b0, 0b00001, "sqadd", 1>;
4829 defm UQADD : NeonI_Scalar3Same_BHSD_sizes<0b1, 0b00001, "uqadd", 1>;
4831 // Scalar Integer Saturating Sub (Signed, Unsigned)
4832 defm SQSUB : NeonI_Scalar3Same_BHSD_sizes<0b0, 0b00101, "sqsub", 0>;
4833 defm UQSUB : NeonI_Scalar3Same_BHSD_sizes<0b1, 0b00101, "uqsub", 0>;
4836 // Patterns to match llvm.aarch64.* intrinsic for
4837 // Scalar Integer Saturating Add, Sub (Signed, Unsigned)
4838 defm : Neon_Scalar3Same_BHSD_size_patterns<int_arm_neon_vqadds, SQADDbbb,
4839 SQADDhhh, SQADDsss, SQADDddd>;
4840 defm : Neon_Scalar3Same_BHSD_size_patterns<int_arm_neon_vqaddu, UQADDbbb,
4841 UQADDhhh, UQADDsss, UQADDddd>;
4842 defm : Neon_Scalar3Same_BHSD_size_patterns<int_arm_neon_vqsubs, SQSUBbbb,
4843 SQSUBhhh, SQSUBsss, SQSUBddd>;
4844 defm : Neon_Scalar3Same_BHSD_size_patterns<int_arm_neon_vqsubu, UQSUBbbb,
4845 UQSUBhhh, UQSUBsss, UQSUBddd>;
4847 // Scalar Integer Saturating Doubling Multiply Half High
4848 defm SQDMULH : NeonI_Scalar3Same_HS_sizes<0b0, 0b10110, "sqdmulh", 1>;
4850 // Scalar Integer Saturating Rounding Doubling Multiply Half High
4851 defm SQRDMULH : NeonI_Scalar3Same_HS_sizes<0b1, 0b10110, "sqrdmulh", 1>;
4853 // Patterns to match llvm.arm.* intrinsic for
4854 // Scalar Integer Saturating Doubling Multiply Half High and
4855 // Scalar Integer Saturating Rounding Doubling Multiply Half High
4856 defm : Neon_Scalar3Same_HS_size_patterns<int_arm_neon_vqdmulh, SQDMULHhhh,
4858 defm : Neon_Scalar3Same_HS_size_patterns<int_arm_neon_vqrdmulh, SQRDMULHhhh,
4861 // Scalar Floating-point Multiply Extended
4862 defm FMULX : NeonI_Scalar3Same_SD_sizes<0b0, 0b0, 0b11011, "fmulx", 1>;
4864 // Scalar Floating-point Reciprocal Step
4865 defm FRECPS : NeonI_Scalar3Same_SD_sizes<0b0, 0b0, 0b11111, "frecps", 0>;
4866 defm : Neon_Scalar3Same_SD_size_patterns<int_aarch64_neon_vrecps, f32, f32,
4867 FRECPSsss, f64, f64, FRECPSddd>;
4868 def : Pat<(v1f64 (int_arm_neon_vrecps (v1f64 FPR64:$Rn), (v1f64 FPR64:$Rm))),
4869 (FRECPSddd FPR64:$Rn, FPR64:$Rm)>;
4871 // Scalar Floating-point Reciprocal Square Root Step
4872 defm FRSQRTS : NeonI_Scalar3Same_SD_sizes<0b0, 0b1, 0b11111, "frsqrts", 0>;
4873 defm : Neon_Scalar3Same_SD_size_patterns<int_aarch64_neon_vrsqrts, f32, f32,
4874 FRSQRTSsss, f64, f64, FRSQRTSddd>;
4875 def : Pat<(v1f64 (int_arm_neon_vrsqrts (v1f64 FPR64:$Rn), (v1f64 FPR64:$Rm))),
4876 (FRSQRTSddd FPR64:$Rn, FPR64:$Rm)>;
4877 def : Pat<(v1f64 (fsqrt (v1f64 FPR64:$Rn))), (FSQRTdd FPR64:$Rn)>;
4879 // Patterns to match llvm.aarch64.* intrinsic for
4880 // Scalar Floating-point Multiply Extended,
4881 multiclass Neon_Scalar3Same_MULX_SD_size_patterns<SDPatternOperator opnode,
4883 Instruction INSTD> {
4884 def : Pat<(f32 (opnode (f32 FPR32:$Rn), (f32 FPR32:$Rm))),
4885 (INSTS FPR32:$Rn, FPR32:$Rm)>;
4886 def : Pat<(f64 (opnode (f64 FPR64:$Rn), (f64 FPR64:$Rm))),
4887 (INSTD FPR64:$Rn, FPR64:$Rm)>;
4890 defm : Neon_Scalar3Same_MULX_SD_size_patterns<int_aarch64_neon_vmulx,
4893 // Scalar Integer Shift Left (Signed, Unsigned)
4894 def SSHLddd : NeonI_Scalar3Same_D_size<0b0, 0b01000, "sshl">;
4895 def USHLddd : NeonI_Scalar3Same_D_size<0b1, 0b01000, "ushl">;
4897 // Patterns to match llvm.arm.* intrinsic for
4898 // Scalar Integer Shift Left (Signed, Unsigned)
4899 defm : Neon_Scalar3Same_D_size_patterns<int_arm_neon_vshifts, SSHLddd>;
4900 defm : Neon_Scalar3Same_D_size_patterns<int_arm_neon_vshiftu, USHLddd>;
4902 // Patterns to match llvm.aarch64.* intrinsic for
4903 // Scalar Integer Shift Left (Signed, Unsigned)
4904 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vshlds, SSHLddd>;
4905 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vshldu, USHLddd>;
4907 // Scalar Integer Saturating Shift Left (Signed, Unsigned)
4908 defm SQSHL: NeonI_Scalar3Same_BHSD_sizes<0b0, 0b01001, "sqshl", 0>;
4909 defm UQSHL: NeonI_Scalar3Same_BHSD_sizes<0b1, 0b01001, "uqshl", 0>;
4911 // Patterns to match llvm.aarch64.* intrinsic for
4912 // Scalar Integer Saturating Shift Letf (Signed, Unsigned)
4913 defm : Neon_Scalar3Same_BHSD_size_patterns<int_aarch64_neon_vqshls, SQSHLbbb,
4914 SQSHLhhh, SQSHLsss, SQSHLddd>;
4915 defm : Neon_Scalar3Same_BHSD_size_patterns<int_aarch64_neon_vqshlu, UQSHLbbb,
4916 UQSHLhhh, UQSHLsss, UQSHLddd>;
4918 // Patterns to match llvm.arm.* intrinsic for
4919 // Scalar Integer Saturating Shift Letf (Signed, Unsigned)
4920 defm : Neon_Scalar3Same_D_size_patterns<int_arm_neon_vqshifts, SQSHLddd>;
4921 defm : Neon_Scalar3Same_D_size_patterns<int_arm_neon_vqshiftu, UQSHLddd>;
4923 // Scalar Integer Rounding Shift Left (Signed, Unsigned)
4924 def SRSHLddd: NeonI_Scalar3Same_D_size<0b0, 0b01010, "srshl">;
4925 def URSHLddd: NeonI_Scalar3Same_D_size<0b1, 0b01010, "urshl">;
4927 // Patterns to match llvm.aarch64.* intrinsic for
4928 // Scalar Integer Rounding Shift Left (Signed, Unsigned)
4929 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vrshlds, SRSHLddd>;
4930 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vrshldu, URSHLddd>;
4932 // Patterns to match llvm.arm.* intrinsic for
4933 // Scalar Integer Rounding Shift Left (Signed, Unsigned)
4934 defm : Neon_Scalar3Same_D_size_patterns<int_arm_neon_vrshifts, SRSHLddd>;
4935 defm : Neon_Scalar3Same_D_size_patterns<int_arm_neon_vrshiftu, URSHLddd>;
4937 // Scalar Integer Saturating Rounding Shift Left (Signed, Unsigned)
4938 defm SQRSHL: NeonI_Scalar3Same_BHSD_sizes<0b0, 0b01011, "sqrshl", 0>;
4939 defm UQRSHL: NeonI_Scalar3Same_BHSD_sizes<0b1, 0b01011, "uqrshl", 0>;
4941 // Patterns to match llvm.aarch64.* intrinsic for
4942 // Scalar Integer Saturating Rounding Shift Left (Signed, Unsigned)
4943 defm : Neon_Scalar3Same_BHSD_size_patterns<int_aarch64_neon_vqrshls, SQRSHLbbb,
4944 SQRSHLhhh, SQRSHLsss, SQRSHLddd>;
4945 defm : Neon_Scalar3Same_BHSD_size_patterns<int_aarch64_neon_vqrshlu, UQRSHLbbb,
4946 UQRSHLhhh, UQRSHLsss, UQRSHLddd>;
4948 // Patterns to match llvm.arm.* intrinsic for
4949 // Scalar Integer Saturating Rounding Shift Left (Signed, Unsigned)
4950 defm : Neon_Scalar3Same_D_size_patterns<int_arm_neon_vqrshifts, SQRSHLddd>;
4951 defm : Neon_Scalar3Same_D_size_patterns<int_arm_neon_vqrshiftu, UQRSHLddd>;
4953 // Signed Saturating Doubling Multiply-Add Long
4954 defm SQDMLAL : NeonI_Scalar3Diff_ml_HS_size<0b0, 0b1001, "sqdmlal">;
4955 defm : Neon_Scalar3Diff_ml_HS_size_patterns<int_aarch64_neon_vqdmlal,
4956 SQDMLALshh, SQDMLALdss>;
4958 // Signed Saturating Doubling Multiply-Subtract Long
4959 defm SQDMLSL : NeonI_Scalar3Diff_ml_HS_size<0b0, 0b1011, "sqdmlsl">;
4960 defm : Neon_Scalar3Diff_ml_HS_size_patterns<int_aarch64_neon_vqdmlsl,
4961 SQDMLSLshh, SQDMLSLdss>;
4963 // Signed Saturating Doubling Multiply Long
4964 defm SQDMULL : NeonI_Scalar3Diff_HS_size<0b0, 0b1101, "sqdmull">;
4965 defm : Neon_Scalar3Diff_HS_size_patterns<int_arm_neon_vqdmull,
4966 SQDMULLshh, SQDMULLdss>;
4968 // Scalar Signed Integer Convert To Floating-point
4969 defm SCVTF : NeonI_Scalar2SameMisc_SD_size<0b0, 0b0, 0b11101, "scvtf">;
4970 defm : Neon_Scalar2SameMisc_cvt_SD_size_patterns<int_aarch64_neon_vcvtint2fps,
4973 // Scalar Unsigned Integer Convert To Floating-point
4974 defm UCVTF : NeonI_Scalar2SameMisc_SD_size<0b1, 0b0, 0b11101, "ucvtf">;
4975 defm : Neon_Scalar2SameMisc_cvt_SD_size_patterns<int_aarch64_neon_vcvtint2fpu,
4978 // Scalar Floating-point Converts
4979 def FCVTXN : NeonI_Scalar2SameMisc_fcvtxn_D_size<0b1, 0b10110, "fcvtxn">;
4980 def : Neon_Scalar2SameMisc_fcvtxn_D_size_patterns<int_aarch64_neon_fcvtxn,
4983 defm FCVTNS : NeonI_Scalar2SameMisc_SD_size<0b0, 0b0, 0b11010, "fcvtns">;
4984 defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtns,
4985 FCVTNSss, FCVTNSdd>;
4986 def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtns, FCVTNSdd>;
4988 defm FCVTNU : NeonI_Scalar2SameMisc_SD_size<0b1, 0b0, 0b11010, "fcvtnu">;
4989 defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtnu,
4990 FCVTNUss, FCVTNUdd>;
4991 def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtnu, FCVTNUdd>;
4993 defm FCVTMS : NeonI_Scalar2SameMisc_SD_size<0b0, 0b0, 0b11011, "fcvtms">;
4994 defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtms,
4995 FCVTMSss, FCVTMSdd>;
4996 def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtms, FCVTMSdd>;
4998 defm FCVTMU : NeonI_Scalar2SameMisc_SD_size<0b1, 0b0, 0b11011, "fcvtmu">;
4999 defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtmu,
5000 FCVTMUss, FCVTMUdd>;
5001 def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtmu, FCVTMUdd>;
5003 defm FCVTAS : NeonI_Scalar2SameMisc_SD_size<0b0, 0b0, 0b11100, "fcvtas">;
5004 defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtas,
5005 FCVTASss, FCVTASdd>;
5006 def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtas, FCVTASdd>;
5008 defm FCVTAU : NeonI_Scalar2SameMisc_SD_size<0b1, 0b0, 0b11100, "fcvtau">;
5009 defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtau,
5010 FCVTAUss, FCVTAUdd>;
5011 def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtau, FCVTAUdd>;
5013 defm FCVTPS : NeonI_Scalar2SameMisc_SD_size<0b0, 0b1, 0b11010, "fcvtps">;
5014 defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtps,
5015 FCVTPSss, FCVTPSdd>;
5016 def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtps, FCVTPSdd>;
5018 defm FCVTPU : NeonI_Scalar2SameMisc_SD_size<0b1, 0b1, 0b11010, "fcvtpu">;
5019 defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtpu,
5020 FCVTPUss, FCVTPUdd>;
5021 def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtpu, FCVTPUdd>;
5023 defm FCVTZS : NeonI_Scalar2SameMisc_SD_size<0b0, 0b1, 0b11011, "fcvtzs">;
5024 defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtzs,
5025 FCVTZSss, FCVTZSdd>;
5026 def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_aarch64_neon_vcvtzs,
5029 defm FCVTZU : NeonI_Scalar2SameMisc_SD_size<0b1, 0b1, 0b11011, "fcvtzu">;
5030 defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtzu,
5031 FCVTZUss, FCVTZUdd>;
5032 def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_aarch64_neon_vcvtzu,
5035 // Patterns For Convert Instructions Between v1f64 and v1i64
5036 class Neon_Scalar2SameMisc_cvtf_v1f64_pattern<SDPatternOperator opnode,
5038 : Pat<(v1f64 (opnode (v1i64 FPR64:$Rn))), (INST FPR64:$Rn)>;
5040 class Neon_Scalar2SameMisc_fcvt_v1f64_pattern<SDPatternOperator opnode,
5042 : Pat<(v1i64 (opnode (v1f64 FPR64:$Rn))), (INST FPR64:$Rn)>;
5044 def : Neon_Scalar2SameMisc_cvtf_v1f64_pattern<sint_to_fp, SCVTFdd>;
5045 def : Neon_Scalar2SameMisc_cvtf_v1f64_pattern<uint_to_fp, UCVTFdd>;
5047 def : Neon_Scalar2SameMisc_fcvt_v1f64_pattern<fp_to_sint, FCVTZSdd>;
5048 def : Neon_Scalar2SameMisc_fcvt_v1f64_pattern<fp_to_uint, FCVTZUdd>;
5050 // Scalar Floating-point Reciprocal Estimate
5051 defm FRECPE : NeonI_Scalar2SameMisc_SD_size<0b0, 0b1, 0b11101, "frecpe">;
5052 defm : Neon_Scalar2SameMisc_SD_size_patterns<int_aarch64_neon_vrecpe,
5053 FRECPEss, FRECPEdd>;
5054 def : Neon_Scalar2SameMisc_V1_D_size_patterns<int_arm_neon_vrecpe,
5057 // Scalar Floating-point Reciprocal Exponent
5058 defm FRECPX : NeonI_Scalar2SameMisc_SD_size<0b0, 0b1, 0b11111, "frecpx">;
5059 defm : Neon_Scalar2SameMisc_SD_size_patterns<int_aarch64_neon_vrecpx,
5060 FRECPXss, FRECPXdd>;
5062 // Scalar Floating-point Reciprocal Square Root Estimate
5063 defm FRSQRTE: NeonI_Scalar2SameMisc_SD_size<0b1, 0b1, 0b11101, "frsqrte">;
5064 defm : Neon_Scalar2SameMisc_SD_size_patterns<int_aarch64_neon_vrsqrte,
5065 FRSQRTEss, FRSQRTEdd>;
5066 def : Neon_Scalar2SameMisc_V1_D_size_patterns<int_arm_neon_vrsqrte,
5069 // Scalar Floating-point Round
5070 class Neon_ScalarFloatRound_pattern<SDPatternOperator opnode, Instruction INST>
5071 : Pat<(v1f64 (opnode (v1f64 FPR64:$Rn))), (INST FPR64:$Rn)>;
5073 def : Neon_ScalarFloatRound_pattern<fceil, FRINTPdd>;
5074 def : Neon_ScalarFloatRound_pattern<ffloor, FRINTMdd>;
5075 def : Neon_ScalarFloatRound_pattern<ftrunc, FRINTZdd>;
5076 def : Neon_ScalarFloatRound_pattern<frint, FRINTXdd>;
5077 def : Neon_ScalarFloatRound_pattern<fnearbyint, FRINTIdd>;
5078 def : Neon_ScalarFloatRound_pattern<frnd, FRINTAdd>;
5079 def : Neon_ScalarFloatRound_pattern<int_aarch64_neon_frintn, FRINTNdd>;
5081 // Scalar Integer Compare
5083 // Scalar Compare Bitwise Equal
5084 def CMEQddd: NeonI_Scalar3Same_D_size<0b1, 0b10001, "cmeq">;
5085 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vceq, CMEQddd>;
5087 class Neon_Scalar3Same_cmp_D_size_v1_patterns<SDPatternOperator opnode,
5090 : Pat<(v1i64 (opnode (v1i64 FPR64:$Rn), (v1i64 FPR64:$Rm), CC)),
5091 (INSTD FPR64:$Rn, FPR64:$Rm)>;
5093 def : Neon_Scalar3Same_cmp_D_size_v1_patterns<Neon_cmp, CMEQddd, SETEQ>;
5095 // Scalar Compare Signed Greather Than Or Equal
5096 def CMGEddd: NeonI_Scalar3Same_D_size<0b0, 0b00111, "cmge">;
5097 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vcge, CMGEddd>;
5098 def : Neon_Scalar3Same_cmp_D_size_v1_patterns<Neon_cmp, CMGEddd, SETGE>;
5100 // Scalar Compare Unsigned Higher Or Same
5101 def CMHSddd: NeonI_Scalar3Same_D_size<0b1, 0b00111, "cmhs">;
5102 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vchs, CMHSddd>;
5103 def : Neon_Scalar3Same_cmp_D_size_v1_patterns<Neon_cmp, CMHSddd, SETUGE>;
5105 // Scalar Compare Unsigned Higher
5106 def CMHIddd: NeonI_Scalar3Same_D_size<0b1, 0b00110, "cmhi">;
5107 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vchi, CMHIddd>;
5108 def : Neon_Scalar3Same_cmp_D_size_v1_patterns<Neon_cmp, CMHIddd, SETUGT>;
5110 // Scalar Compare Signed Greater Than
5111 def CMGTddd: NeonI_Scalar3Same_D_size<0b0, 0b00110, "cmgt">;
5112 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vcgt, CMGTddd>;
5113 def : Neon_Scalar3Same_cmp_D_size_v1_patterns<Neon_cmp, CMGTddd, SETGT>;
5115 // Scalar Compare Bitwise Test Bits
5116 def CMTSTddd: NeonI_Scalar3Same_D_size<0b0, 0b10001, "cmtst">;
5117 defm : Neon_Scalar3Same_D_size_patterns<int_aarch64_neon_vtstd, CMTSTddd>;
5118 defm : Neon_Scalar3Same_D_size_patterns<Neon_tst, CMTSTddd>;
5120 // Scalar Compare Bitwise Equal To Zero
5121 def CMEQddi: NeonI_Scalar2SameMisc_cmpz_D_size<0b0, 0b01001, "cmeq">;
5122 def : Neon_Scalar2SameMisc_cmpz_D_size_patterns<int_aarch64_neon_vceq,
5124 def : Neon_Scalar2SameMisc_cmpz_D_V1_size_patterns<SETEQ, CMEQddi>;
5126 // Scalar Compare Signed Greather Than Or Equal To Zero
5127 def CMGEddi: NeonI_Scalar2SameMisc_cmpz_D_size<0b1, 0b01000, "cmge">;
5128 def : Neon_Scalar2SameMisc_cmpz_D_size_patterns<int_aarch64_neon_vcge,
5130 def : Neon_Scalar2SameMisc_cmpz_D_V1_size_patterns<SETGE, CMGEddi>;
5132 // Scalar Compare Signed Greater Than Zero
5133 def CMGTddi: NeonI_Scalar2SameMisc_cmpz_D_size<0b0, 0b01000, "cmgt">;
5134 def : Neon_Scalar2SameMisc_cmpz_D_size_patterns<int_aarch64_neon_vcgt,
5136 def : Neon_Scalar2SameMisc_cmpz_D_V1_size_patterns<SETGT, CMGTddi>;
5138 // Scalar Compare Signed Less Than Or Equal To Zero
5139 def CMLEddi: NeonI_Scalar2SameMisc_cmpz_D_size<0b1, 0b01001, "cmle">;
5140 def : Neon_Scalar2SameMisc_cmpz_D_size_patterns<int_aarch64_neon_vclez,
5142 def : Neon_Scalar2SameMisc_cmpz_D_V1_size_patterns<SETLE, CMLEddi>;
5144 // Scalar Compare Less Than Zero
5145 def CMLTddi: NeonI_Scalar2SameMisc_cmpz_D_size<0b0, 0b01010, "cmlt">;
5146 def : Neon_Scalar2SameMisc_cmpz_D_size_patterns<int_aarch64_neon_vcltz,
5148 def : Neon_Scalar2SameMisc_cmpz_D_V1_size_patterns<SETLT, CMLTddi>;
5150 // Scalar Floating-point Compare
5152 // Scalar Floating-point Compare Mask Equal
5153 defm FCMEQ: NeonI_Scalar3Same_SD_sizes<0b0, 0b0, 0b11100, "fcmeq">;
5154 defm : Neon_Scalar3Same_SD_size_patterns<int_aarch64_neon_fceq, v1i32, f32,
5155 FCMEQsss, v1i64, f64, FCMEQddd>;
5156 def : Neon_Scalar3Same_cmp_V1_D_size_patterns<SETEQ, FCMEQddd>;
5158 // Scalar Floating-point Compare Mask Equal To Zero
5159 defm FCMEQZ: NeonI_Scalar2SameMisc_cmpz_SD_size<0b0, 0b01101, "fcmeq">;
5160 defm : Neon_Scalar2SameMisc_cmpz_SD_size_patterns<int_aarch64_neon_fceq,
5161 FCMEQZssi, FCMEQZddi>;
5162 def : Pat<(v1i64 (Neon_cmpz (v1f64 FPR64:$Rn), (f32 fpz32:$FPImm), SETEQ)),
5163 (FCMEQZddi FPR64:$Rn, fpz32:$FPImm)>;
5165 // Scalar Floating-point Compare Mask Greater Than Or Equal
5166 defm FCMGE: NeonI_Scalar3Same_SD_sizes<0b1, 0b0, 0b11100, "fcmge">;
5167 defm : Neon_Scalar3Same_SD_size_patterns<int_aarch64_neon_fcge, v1i32, f32,
5168 FCMGEsss, v1i64, f64, FCMGEddd>;
5169 def : Neon_Scalar3Same_cmp_V1_D_size_patterns<SETGE, FCMGEddd>;
5171 // Scalar Floating-point Compare Mask Greater Than Or Equal To Zero
5172 defm FCMGEZ: NeonI_Scalar2SameMisc_cmpz_SD_size<0b1, 0b01100, "fcmge">;
5173 defm : Neon_Scalar2SameMisc_cmpz_SD_size_patterns<int_aarch64_neon_fcge,
5174 FCMGEZssi, FCMGEZddi>;
5176 // Scalar Floating-point Compare Mask Greather Than
5177 defm FCMGT: NeonI_Scalar3Same_SD_sizes<0b1, 0b1, 0b11100, "fcmgt">;
5178 defm : Neon_Scalar3Same_SD_size_patterns<int_aarch64_neon_fcgt, v1i32, f32,
5179 FCMGTsss, v1i64, f64, FCMGTddd>;
5180 def : Neon_Scalar3Same_cmp_V1_D_size_patterns<SETGT, FCMGTddd>;
5182 // Scalar Floating-point Compare Mask Greather Than Zero
5183 defm FCMGTZ: NeonI_Scalar2SameMisc_cmpz_SD_size<0b0, 0b01100, "fcmgt">;
5184 defm : Neon_Scalar2SameMisc_cmpz_SD_size_patterns<int_aarch64_neon_fcgt,
5185 FCMGTZssi, FCMGTZddi>;
5187 // Scalar Floating-point Compare Mask Less Than Or Equal To Zero
5188 defm FCMLEZ: NeonI_Scalar2SameMisc_cmpz_SD_size<0b1, 0b01101, "fcmle">;
5189 defm : Neon_Scalar2SameMisc_cmpz_SD_size_patterns<int_aarch64_neon_fclez,
5190 FCMLEZssi, FCMLEZddi>;
5192 // Scalar Floating-point Compare Mask Less Than Zero
5193 defm FCMLTZ: NeonI_Scalar2SameMisc_cmpz_SD_size<0b0, 0b01110, "fcmlt">;
5194 defm : Neon_Scalar2SameMisc_cmpz_SD_size_patterns<int_aarch64_neon_fcltz,
5195 FCMLTZssi, FCMLTZddi>;
5197 // Scalar Floating-point Absolute Compare Mask Greater Than Or Equal
5198 defm FACGE: NeonI_Scalar3Same_SD_sizes<0b1, 0b0, 0b11101, "facge">;
5199 defm : Neon_Scalar3Same_SD_size_patterns<int_aarch64_neon_fcage, v1i32, f32,
5200 FACGEsss, v1i64, f64, FACGEddd>;
5201 def : Pat<(v1i64 (int_aarch64_neon_vcage (v1f64 FPR64:$Rn), (v1f64 FPR64:$Rm))),
5202 (FACGEddd FPR64:$Rn, FPR64:$Rm)>;
5204 // Scalar Floating-point Absolute Compare Mask Greater Than
5205 defm FACGT: NeonI_Scalar3Same_SD_sizes<0b1, 0b1, 0b11101, "facgt">;
5206 defm : Neon_Scalar3Same_SD_size_patterns<int_aarch64_neon_fcagt, v1i32, f32,
5207 FACGTsss, v1i64, f64, FACGTddd>;
5208 def : Pat<(v1i64 (int_aarch64_neon_vcagt (v1f64 FPR64:$Rn), (v1f64 FPR64:$Rm))),
5209 (FACGTddd FPR64:$Rn, FPR64:$Rm)>;
5211 // Scalar Floating-point Absolute Difference
5212 defm FABD: NeonI_Scalar3Same_SD_sizes<0b1, 0b1, 0b11010, "fabd">;
5213 defm : Neon_Scalar3Same_SD_size_patterns<int_aarch64_neon_vabd, f32, f32,
5214 FABDsss, f64, f64, FABDddd>;
5216 // Scalar Absolute Value
5217 defm ABS : NeonI_Scalar2SameMisc_D_size<0b0, 0b01011, "abs">;
5218 defm : Neon_Scalar2SameMisc_D_size_patterns<int_aarch64_neon_vabs, ABSdd>;
5220 // Scalar Signed Saturating Absolute Value
5221 defm SQABS : NeonI_Scalar2SameMisc_BHSD_size<0b0, 0b00111, "sqabs">;
5222 defm : Neon_Scalar2SameMisc_BHSD_size_patterns<int_arm_neon_vqabs,
5223 SQABSbb, SQABShh, SQABSss, SQABSdd>;
5226 defm NEG : NeonI_Scalar2SameMisc_D_size<0b1, 0b01011, "neg">;
5227 defm : Neon_Scalar2SameMisc_D_size_patterns<int_aarch64_neon_vneg, NEGdd>;
5229 // Scalar Signed Saturating Negate
5230 defm SQNEG : NeonI_Scalar2SameMisc_BHSD_size<0b1, 0b00111, "sqneg">;
5231 defm : Neon_Scalar2SameMisc_BHSD_size_patterns<int_arm_neon_vqneg,
5232 SQNEGbb, SQNEGhh, SQNEGss, SQNEGdd>;
5234 // Scalar Signed Saturating Accumulated of Unsigned Value
5235 defm SUQADD : NeonI_Scalar2SameMisc_accum_BHSD_size<0b0, 0b00011, "suqadd">;
5236 defm : Neon_Scalar2SameMisc_accum_BHSD_size_patterns<int_aarch64_neon_vuqadd,
5238 SUQADDss, SUQADDdd>;
5240 // Scalar Unsigned Saturating Accumulated of Signed Value
5241 defm USQADD : NeonI_Scalar2SameMisc_accum_BHSD_size<0b1, 0b00011, "usqadd">;
5242 defm : Neon_Scalar2SameMisc_accum_BHSD_size_patterns<int_aarch64_neon_vsqadd,
5244 USQADDss, USQADDdd>;
5246 def : Pat<(v1i64 (int_aarch64_neon_suqadd (v1i64 FPR64:$Src),
5247 (v1i64 FPR64:$Rn))),
5248 (SUQADDdd FPR64:$Src, FPR64:$Rn)>;
5250 def : Pat<(v1i64 (int_aarch64_neon_usqadd (v1i64 FPR64:$Src),
5251 (v1i64 FPR64:$Rn))),
5252 (USQADDdd FPR64:$Src, FPR64:$Rn)>;
5254 def : Pat<(v1i64 (int_arm_neon_vabs (v1i64 FPR64:$Rn))),
5257 def : Pat<(v1i64 (int_arm_neon_vqabs (v1i64 FPR64:$Rn))),
5258 (SQABSdd FPR64:$Rn)>;
5260 def : Pat<(v1i64 (int_arm_neon_vqneg (v1i64 FPR64:$Rn))),
5261 (SQNEGdd FPR64:$Rn)>;
5263 def : Pat<(v1i64 (sub (v1i64 (bitconvert (v8i8 Neon_AllZero))),
5264 (v1i64 FPR64:$Rn))),
5267 // Scalar Signed Saturating Extract Unsigned Narrow
5268 defm SQXTUN : NeonI_Scalar2SameMisc_narrow_HSD_size<0b1, 0b10010, "sqxtun">;
5269 defm : Neon_Scalar2SameMisc_narrow_HSD_size_patterns<int_arm_neon_vqmovnsu,
5273 // Scalar Signed Saturating Extract Narrow
5274 defm SQXTN : NeonI_Scalar2SameMisc_narrow_HSD_size<0b0, 0b10100, "sqxtn">;
5275 defm : Neon_Scalar2SameMisc_narrow_HSD_size_patterns<int_arm_neon_vqmovns,
5279 // Scalar Unsigned Saturating Extract Narrow
5280 defm UQXTN : NeonI_Scalar2SameMisc_narrow_HSD_size<0b1, 0b10100, "uqxtn">;
5281 defm : Neon_Scalar2SameMisc_narrow_HSD_size_patterns<int_arm_neon_vqmovnu,
5285 // Scalar Reduce Pairwise
5287 multiclass NeonI_ScalarPair_D_sizes<bit u, bit size, bits<5> opcode,
5288 string asmop, bit Commutable = 0> {
5289 let isCommutable = Commutable in {
5290 def _D_2D : NeonI_ScalarPair<u, {size, 0b1}, opcode,
5291 (outs FPR64:$Rd), (ins VPR128:$Rn),
5292 !strconcat(asmop, "\t$Rd, $Rn.2d"),
5298 multiclass NeonI_ScalarPair_SD_sizes<bit u, bit size, bits<5> opcode,
5299 string asmop, bit Commutable = 0>
5300 : NeonI_ScalarPair_D_sizes<u, size, opcode, asmop, Commutable> {
5301 let isCommutable = Commutable in {
5302 def _S_2S : NeonI_ScalarPair<u, {size, 0b0}, opcode,
5303 (outs FPR32:$Rd), (ins VPR64:$Rn),
5304 !strconcat(asmop, "\t$Rd, $Rn.2s"),
5310 // Scalar Reduce Addition Pairwise (Integer) with
5311 // Pattern to match llvm.arm.* intrinsic
5312 defm ADDPvv : NeonI_ScalarPair_D_sizes<0b0, 0b1, 0b11011, "addp", 0>;
5314 // Pattern to match llvm.aarch64.* intrinsic for
5315 // Scalar Reduce Addition Pairwise (Integer)
5316 def : Pat<(v1i64 (int_aarch64_neon_vpadd (v2i64 VPR128:$Rn))),
5317 (ADDPvv_D_2D VPR128:$Rn)>;
5318 def : Pat<(v1i64 (int_aarch64_neon_vaddv (v2i64 VPR128:$Rn))),
5319 (ADDPvv_D_2D VPR128:$Rn)>;
5321 // Scalar Reduce Addition Pairwise (Floating Point)
5322 defm FADDPvv : NeonI_ScalarPair_SD_sizes<0b1, 0b0, 0b01101, "faddp", 0>;
5324 // Scalar Reduce Maximum Pairwise (Floating Point)
5325 defm FMAXPvv : NeonI_ScalarPair_SD_sizes<0b1, 0b0, 0b01111, "fmaxp", 0>;
5327 // Scalar Reduce Minimum Pairwise (Floating Point)
5328 defm FMINPvv : NeonI_ScalarPair_SD_sizes<0b1, 0b1, 0b01111, "fminp", 0>;
5330 // Scalar Reduce maxNum Pairwise (Floating Point)
5331 defm FMAXNMPvv : NeonI_ScalarPair_SD_sizes<0b1, 0b0, 0b01100, "fmaxnmp", 0>;
5333 // Scalar Reduce minNum Pairwise (Floating Point)
5334 defm FMINNMPvv : NeonI_ScalarPair_SD_sizes<0b1, 0b1, 0b01100, "fminnmp", 0>;
5336 multiclass Neon_ScalarPair_SD_size_patterns<SDPatternOperator opnode,
5338 Instruction INSTD> {
5339 def : Pat<(f32 (opnode (v2f32 VPR64:$Rn))),
5341 def : Pat<(f64 (opnode (v2f64 VPR128:$Rn))),
5342 (INSTD VPR128:$Rn)>;
5345 // Patterns to match llvm.aarch64.* intrinsic for
5346 // Scalar Reduce Add, Max, Min, MaxiNum, MinNum Pairwise (Floating Point)
5347 defm : Neon_ScalarPair_SD_size_patterns<int_aarch64_neon_vpfadd,
5348 FADDPvv_S_2S, FADDPvv_D_2D>;
5350 defm : Neon_ScalarPair_SD_size_patterns<int_aarch64_neon_vpmax,
5351 FMAXPvv_S_2S, FMAXPvv_D_2D>;
5353 defm : Neon_ScalarPair_SD_size_patterns<int_aarch64_neon_vpmin,
5354 FMINPvv_S_2S, FMINPvv_D_2D>;
5356 defm : Neon_ScalarPair_SD_size_patterns<int_aarch64_neon_vpfmaxnm,
5357 FMAXNMPvv_S_2S, FMAXNMPvv_D_2D>;
5359 defm : Neon_ScalarPair_SD_size_patterns<int_aarch64_neon_vpfminnm,
5360 FMINNMPvv_S_2S, FMINNMPvv_D_2D>;
5362 def : Pat<(f32 (int_aarch64_neon_vpfadd (v4f32 VPR128:$Rn))),
5363 (FADDPvv_S_2S (v2f32
5365 (v4f32 (FADDP_4S (v4f32 VPR128:$Rn), (v4f32 VPR128:$Rn))),
5368 // Scalar by element Arithmetic
5370 class NeonI_ScalarXIndexedElemArith<string asmop, bits<4> opcode,
5371 string rmlane, bit u, bit szhi, bit szlo,
5372 RegisterClass ResFPR, RegisterClass OpFPR,
5373 RegisterOperand OpVPR, Operand OpImm>
5374 : NeonI_ScalarXIndexedElem<u, szhi, szlo, opcode,
5376 (ins OpFPR:$Rn, OpVPR:$MRm, OpImm:$Imm),
5377 asmop # "\t$Rd, $Rn, $MRm" # rmlane # "[$Imm]",
5384 class NeonI_ScalarXIndexedElemArith_Constraint_Impl<string asmop, bits<4> opcode,
5386 bit u, bit szhi, bit szlo,
5387 RegisterClass ResFPR,
5388 RegisterClass OpFPR,
5389 RegisterOperand OpVPR,
5391 : NeonI_ScalarXIndexedElem<u, szhi, szlo, opcode,
5393 (ins ResFPR:$src, OpFPR:$Rn, OpVPR:$MRm, OpImm:$Imm),
5394 asmop # "\t$Rd, $Rn, $MRm" # rmlane # "[$Imm]",
5397 let Constraints = "$src = $Rd";
5402 // Scalar Floating Point multiply (scalar, by element)
5403 def FMULssv_4S : NeonI_ScalarXIndexedElemArith<"fmul",
5404 0b1001, ".s", 0b0, 0b1, 0b0, FPR32, FPR32, VPR128, neon_uimm2_bare> {
5405 let Inst{11} = Imm{1}; // h
5406 let Inst{21} = Imm{0}; // l
5407 let Inst{20-16} = MRm;
5409 def FMULddv_2D : NeonI_ScalarXIndexedElemArith<"fmul",
5410 0b1001, ".d", 0b0, 0b1, 0b1, FPR64, FPR64, VPR128, neon_uimm1_bare> {
5411 let Inst{11} = Imm{0}; // h
5412 let Inst{21} = 0b0; // l
5413 let Inst{20-16} = MRm;
5416 // Scalar Floating Point multiply extended (scalar, by element)
5417 def FMULXssv_4S : NeonI_ScalarXIndexedElemArith<"fmulx",
5418 0b1001, ".s", 0b1, 0b1, 0b0, FPR32, FPR32, VPR128, neon_uimm2_bare> {
5419 let Inst{11} = Imm{1}; // h
5420 let Inst{21} = Imm{0}; // l
5421 let Inst{20-16} = MRm;
5423 def FMULXddv_2D : NeonI_ScalarXIndexedElemArith<"fmulx",
5424 0b1001, ".d", 0b1, 0b1, 0b1, FPR64, FPR64, VPR128, neon_uimm1_bare> {
5425 let Inst{11} = Imm{0}; // h
5426 let Inst{21} = 0b0; // l
5427 let Inst{20-16} = MRm;
5430 multiclass Neon_ScalarXIndexedElem_MUL_MULX_Patterns<
5431 SDPatternOperator opnode,
5433 ValueType ResTy, RegisterClass FPRC, ValueType OpTy, Operand OpImm,
5434 ValueType OpNTy, ValueType ExTy, Operand OpNImm> {
5436 def : Pat<(ResTy (opnode (ResTy FPRC:$Rn),
5437 (ResTy (vector_extract (OpTy VPR128:$MRm), OpImm:$Imm)))),
5438 (ResTy (INST (ResTy FPRC:$Rn), (OpTy VPR128:$MRm), OpImm:$Imm))>;
5440 def : Pat<(ResTy (opnode (ResTy FPRC:$Rn),
5441 (ResTy (vector_extract (OpNTy VPR64:$MRm), OpNImm:$Imm)))),
5442 (ResTy (INST (ResTy FPRC:$Rn),
5443 (ExTy (SUBREG_TO_REG (i64 0), VPR64:$MRm, sub_64)),
5447 def : Pat<(ResTy (opnode
5448 (ResTy (vector_extract (OpTy VPR128:$MRm), OpImm:$Imm)),
5450 (ResTy (INST (ResTy FPRC:$Rn), (OpTy VPR128:$MRm), OpImm:$Imm))>;
5452 def : Pat<(ResTy (opnode
5453 (ResTy (vector_extract (OpNTy VPR64:$MRm), OpNImm:$Imm)),
5455 (ResTy (INST (ResTy FPRC:$Rn),
5456 (ExTy (SUBREG_TO_REG (i64 0), VPR64:$MRm, sub_64)),
5460 // Patterns for Scalar Floating Point multiply (scalar, by element)
5461 defm : Neon_ScalarXIndexedElem_MUL_MULX_Patterns<fmul, FMULssv_4S,
5462 f32, FPR32, v4f32, neon_uimm2_bare, v2f32, v4f32, neon_uimm1_bare>;
5463 defm : Neon_ScalarXIndexedElem_MUL_MULX_Patterns<fmul, FMULddv_2D,
5464 f64, FPR64, v2f64, neon_uimm1_bare, v1f64, v2f64, neon_uimm0_bare>;
5466 // Patterns for Scalar Floating Point multiply extended (scalar, by element)
5467 defm : Neon_ScalarXIndexedElem_MUL_MULX_Patterns<int_aarch64_neon_vmulx,
5468 FMULXssv_4S, f32, FPR32, v4f32, neon_uimm2_bare,
5469 v2f32, v4f32, neon_uimm1_bare>;
5470 defm : Neon_ScalarXIndexedElem_MUL_MULX_Patterns<int_aarch64_neon_vmulx,
5471 FMULXddv_2D, f64, FPR64, v2f64, neon_uimm1_bare,
5472 v1f64, v2f64, neon_uimm0_bare>;
5474 // Scalar Floating Point fused multiply-add (scalar, by element)
5475 def FMLAssv_4S : NeonI_ScalarXIndexedElemArith_Constraint_Impl<"fmla",
5476 0b0001, ".s", 0b0, 0b1, 0b0, FPR32, FPR32, VPR128, neon_uimm2_bare> {
5477 let Inst{11} = Imm{1}; // h
5478 let Inst{21} = Imm{0}; // l
5479 let Inst{20-16} = MRm;
5481 def FMLAddv_2D : NeonI_ScalarXIndexedElemArith_Constraint_Impl<"fmla",
5482 0b0001, ".d", 0b0, 0b1, 0b1, FPR64, FPR64, VPR128, neon_uimm1_bare> {
5483 let Inst{11} = Imm{0}; // h
5484 let Inst{21} = 0b0; // l
5485 let Inst{20-16} = MRm;
5488 // Scalar Floating Point fused multiply-subtract (scalar, by element)
5489 def FMLSssv_4S : NeonI_ScalarXIndexedElemArith_Constraint_Impl<"fmls",
5490 0b0101, ".s", 0b0, 0b1, 0b0, FPR32, FPR32, VPR128, neon_uimm2_bare> {
5491 let Inst{11} = Imm{1}; // h
5492 let Inst{21} = Imm{0}; // l
5493 let Inst{20-16} = MRm;
5495 def FMLSddv_2D : NeonI_ScalarXIndexedElemArith_Constraint_Impl<"fmls",
5496 0b0101, ".d", 0b0, 0b1, 0b1, FPR64, FPR64, VPR128, neon_uimm1_bare> {
5497 let Inst{11} = Imm{0}; // h
5498 let Inst{21} = 0b0; // l
5499 let Inst{20-16} = MRm;
5501 // We are allowed to match the fma instruction regardless of compile options.
5502 multiclass Neon_ScalarXIndexedElem_FMA_Patterns<
5503 Instruction FMLAI, Instruction FMLSI,
5504 ValueType ResTy, RegisterClass FPRC, ValueType OpTy, Operand OpImm,
5505 ValueType OpNTy, ValueType ExTy, Operand OpNImm> {
5507 def : Pat<(ResTy (fma (ResTy FPRC:$Rn),
5508 (ResTy (vector_extract (OpTy VPR128:$MRm), OpImm:$Imm)),
5510 (ResTy (FMLAI (ResTy FPRC:$Ra),
5511 (ResTy FPRC:$Rn), (OpTy VPR128:$MRm), OpImm:$Imm))>;
5513 def : Pat<(ResTy (fma (ResTy FPRC:$Rn),
5514 (ResTy (vector_extract (OpNTy VPR64:$MRm), OpNImm:$Imm)),
5516 (ResTy (FMLAI (ResTy FPRC:$Ra),
5518 (ExTy (SUBREG_TO_REG (i64 0), VPR64:$MRm, sub_64)),
5521 // swapped fmla operands
5522 def : Pat<(ResTy (fma
5523 (ResTy (vector_extract (OpTy VPR128:$MRm), OpImm:$Imm)),
5526 (ResTy (FMLAI (ResTy FPRC:$Ra),
5527 (ResTy FPRC:$Rn), (OpTy VPR128:$MRm), OpImm:$Imm))>;
5529 def : Pat<(ResTy (fma
5530 (ResTy (vector_extract (OpNTy VPR64:$MRm), OpNImm:$Imm)),
5533 (ResTy (FMLAI (ResTy FPRC:$Ra),
5535 (ExTy (SUBREG_TO_REG (i64 0), VPR64:$MRm, sub_64)),
5539 def : Pat<(ResTy (fma (ResTy FPRC:$Rn),
5540 (fneg (ResTy (vector_extract (OpTy VPR128:$MRm), OpImm:$Imm))),
5542 (ResTy (FMLSI (ResTy FPRC:$Ra),
5543 (ResTy FPRC:$Rn), (OpTy VPR128:$MRm), OpImm:$Imm))>;
5545 def : Pat<(ResTy (fma (ResTy FPRC:$Rn),
5546 (fneg (ResTy (vector_extract (OpNTy VPR64:$MRm), OpNImm:$Imm))),
5548 (ResTy (FMLSI (ResTy FPRC:$Ra),
5550 (ExTy (SUBREG_TO_REG (i64 0), VPR64:$MRm, sub_64)),
5553 // swapped fmls operands
5554 def : Pat<(ResTy (fma
5555 (fneg (ResTy (vector_extract (OpTy VPR128:$MRm), OpImm:$Imm))),
5558 (ResTy (FMLSI (ResTy FPRC:$Ra),
5559 (ResTy FPRC:$Rn), (OpTy VPR128:$MRm), OpImm:$Imm))>;
5561 def : Pat<(ResTy (fma
5562 (fneg (ResTy (vector_extract (OpNTy VPR64:$MRm), OpNImm:$Imm))),
5565 (ResTy (FMLSI (ResTy FPRC:$Ra),
5567 (ExTy (SUBREG_TO_REG (i64 0), VPR64:$MRm, sub_64)),
5571 // Scalar Floating Point fused multiply-add and
5572 // multiply-subtract (scalar, by element)
5573 defm : Neon_ScalarXIndexedElem_FMA_Patterns<FMLAssv_4S, FMLSssv_4S,
5574 f32, FPR32, v4f32, neon_uimm2_bare, v2f32, v4f32, neon_uimm1_bare>;
5575 defm : Neon_ScalarXIndexedElem_FMA_Patterns<FMLAddv_2D, FMLSddv_2D,
5576 f64, FPR64, v2f64, neon_uimm1_bare, v1f64, v2f64, neon_uimm0_bare>;
5577 defm : Neon_ScalarXIndexedElem_FMA_Patterns<FMLAddv_2D, FMLSddv_2D,
5578 f64, FPR64, v2f64, neon_uimm1_bare, v1f64, v2f64, neon_uimm0_bare>;
5580 // Scalar Signed saturating doubling multiply long (scalar, by element)
5581 def SQDMULLshv_4H : NeonI_ScalarXIndexedElemArith<"sqdmull",
5582 0b1011, ".h", 0b0, 0b0, 0b1, FPR32, FPR16, VPR64Lo, neon_uimm2_bare> {
5583 let Inst{11} = 0b0; // h
5584 let Inst{21} = Imm{1}; // l
5585 let Inst{20} = Imm{0}; // m
5586 let Inst{19-16} = MRm{3-0};
5588 def SQDMULLshv_8H : NeonI_ScalarXIndexedElemArith<"sqdmull",
5589 0b1011, ".h", 0b0, 0b0, 0b1, FPR32, FPR16, VPR128Lo, neon_uimm3_bare> {
5590 let Inst{11} = Imm{2}; // h
5591 let Inst{21} = Imm{1}; // l
5592 let Inst{20} = Imm{0}; // m
5593 let Inst{19-16} = MRm{3-0};
5595 def SQDMULLdsv_2S : NeonI_ScalarXIndexedElemArith<"sqdmull",
5596 0b1011, ".s", 0b0, 0b1, 0b0, FPR64, FPR32, VPR64, neon_uimm1_bare> {
5597 let Inst{11} = 0b0; // h
5598 let Inst{21} = Imm{0}; // l
5599 let Inst{20-16} = MRm;
5601 def SQDMULLdsv_4S : NeonI_ScalarXIndexedElemArith<"sqdmull",
5602 0b1011, ".s", 0b0, 0b1, 0b0, FPR64, FPR32, VPR128, neon_uimm2_bare> {
5603 let Inst{11} = Imm{1}; // h
5604 let Inst{21} = Imm{0}; // l
5605 let Inst{20-16} = MRm;
5608 multiclass Neon_ScalarXIndexedElem_MUL_Patterns<
5609 SDPatternOperator opnode,
5611 ValueType ResTy, RegisterClass FPRC,
5612 ValueType OpVTy, ValueType OpTy,
5613 ValueType VecOpTy, ValueType ExTy, RegisterOperand VPRC, Operand OpImm> {
5615 def : Pat<(ResTy (opnode (OpVTy FPRC:$Rn),
5616 (OpVTy (scalar_to_vector
5617 (ExTy (vector_extract (VecOpTy VPRC:$MRm), OpImm:$Imm)))))),
5618 (ResTy (INST (OpVTy FPRC:$Rn), (VecOpTy VPRC:$MRm), OpImm:$Imm))>;
5621 def : Pat<(ResTy (opnode
5622 (OpVTy (scalar_to_vector
5623 (ExTy (vector_extract (VecOpTy VPRC:$MRm), OpImm:$Imm)))),
5625 (ResTy (INST (OpVTy FPRC:$Rn), (VecOpTy VPRC:$MRm), OpImm:$Imm))>;
5629 // Patterns for Scalar Signed saturating doubling
5630 // multiply long (scalar, by element)
5631 defm : Neon_ScalarXIndexedElem_MUL_Patterns<int_arm_neon_vqdmull,
5632 SQDMULLshv_4H, v1i32, FPR16, v1i16, i16, v4i16,
5633 i32, VPR64Lo, neon_uimm2_bare>;
5634 defm : Neon_ScalarXIndexedElem_MUL_Patterns<int_arm_neon_vqdmull,
5635 SQDMULLshv_8H, v1i32, FPR16, v1i16, i16, v8i16,
5636 i32, VPR128Lo, neon_uimm3_bare>;
5637 defm : Neon_ScalarXIndexedElem_MUL_Patterns<int_arm_neon_vqdmull,
5638 SQDMULLdsv_2S, v1i64, FPR32, v1i32, i32, v2i32,
5639 i32, VPR64Lo, neon_uimm1_bare>;
5640 defm : Neon_ScalarXIndexedElem_MUL_Patterns<int_arm_neon_vqdmull,
5641 SQDMULLdsv_4S, v1i64, FPR32, v1i32, i32, v4i32,
5642 i32, VPR128Lo, neon_uimm2_bare>;
5644 // Scalar Signed saturating doubling multiply-add long (scalar, by element)
5645 def SQDMLALshv_4H : NeonI_ScalarXIndexedElemArith_Constraint_Impl<"sqdmlal",
5646 0b0011, ".h", 0b0, 0b0, 0b1, FPR32, FPR16, VPR64Lo, neon_uimm2_bare> {
5647 let Inst{11} = 0b0; // h
5648 let Inst{21} = Imm{1}; // l
5649 let Inst{20} = Imm{0}; // m
5650 let Inst{19-16} = MRm{3-0};
5652 def SQDMLALshv_8H : NeonI_ScalarXIndexedElemArith_Constraint_Impl<"sqdmlal",
5653 0b0011, ".h", 0b0, 0b0, 0b1, FPR32, FPR16, VPR128Lo, neon_uimm3_bare> {
5654 let Inst{11} = Imm{2}; // h
5655 let Inst{21} = Imm{1}; // l
5656 let Inst{20} = Imm{0}; // m
5657 let Inst{19-16} = MRm{3-0};
5659 def SQDMLALdsv_2S : NeonI_ScalarXIndexedElemArith_Constraint_Impl<"sqdmlal",
5660 0b0011, ".s", 0b0, 0b1, 0b0, FPR64, FPR32, VPR64, neon_uimm1_bare> {
5661 let Inst{11} = 0b0; // h
5662 let Inst{21} = Imm{0}; // l
5663 let Inst{20-16} = MRm;
5665 def SQDMLALdsv_4S : NeonI_ScalarXIndexedElemArith_Constraint_Impl<"sqdmlal",
5666 0b0011, ".s", 0b0, 0b1, 0b0, FPR64, FPR32, VPR128, neon_uimm2_bare> {
5667 let Inst{11} = Imm{1}; // h
5668 let Inst{21} = Imm{0}; // l
5669 let Inst{20-16} = MRm;
5672 // Scalar Signed saturating doubling
5673 // multiply-subtract long (scalar, by element)
5674 def SQDMLSLshv_4H : NeonI_ScalarXIndexedElemArith_Constraint_Impl<"sqdmlsl",
5675 0b0111, ".h", 0b0, 0b0, 0b1, FPR32, FPR16, VPR64Lo, neon_uimm2_bare> {
5676 let Inst{11} = 0b0; // h
5677 let Inst{21} = Imm{1}; // l
5678 let Inst{20} = Imm{0}; // m
5679 let Inst{19-16} = MRm{3-0};
5681 def SQDMLSLshv_8H : NeonI_ScalarXIndexedElemArith_Constraint_Impl<"sqdmlsl",
5682 0b0111, ".h", 0b0, 0b0, 0b1, FPR32, FPR16, VPR128Lo, neon_uimm3_bare> {
5683 let Inst{11} = Imm{2}; // h
5684 let Inst{21} = Imm{1}; // l
5685 let Inst{20} = Imm{0}; // m
5686 let Inst{19-16} = MRm{3-0};
5688 def SQDMLSLdsv_2S : NeonI_ScalarXIndexedElemArith_Constraint_Impl<"sqdmlsl",
5689 0b0111, ".s", 0b0, 0b1, 0b0, FPR64, FPR32, VPR64, neon_uimm1_bare> {
5690 let Inst{11} = 0b0; // h
5691 let Inst{21} = Imm{0}; // l
5692 let Inst{20-16} = MRm;
5694 def SQDMLSLdsv_4S : NeonI_ScalarXIndexedElemArith_Constraint_Impl<"sqdmlsl",
5695 0b0111, ".s", 0b0, 0b1, 0b0, FPR64, FPR32, VPR128, neon_uimm2_bare> {
5696 let Inst{11} = Imm{1}; // h
5697 let Inst{21} = Imm{0}; // l
5698 let Inst{20-16} = MRm;
5701 multiclass Neon_ScalarXIndexedElem_MLAL_Patterns<
5702 SDPatternOperator opnode,
5703 SDPatternOperator coreopnode,
5705 ValueType ResTy, RegisterClass ResFPRC, RegisterClass FPRC,
5707 ValueType OpVTy, ValueType ExTy, RegisterOperand VPRC, Operand OpImm> {
5709 def : Pat<(ResTy (opnode
5710 (ResTy ResFPRC:$Ra),
5711 (ResTy (coreopnode (OpTy FPRC:$Rn),
5712 (OpTy (scalar_to_vector
5713 (ExTy (vector_extract (OpVTy VPRC:$MRm), OpImm:$Imm)))))))),
5714 (ResTy (INST (ResTy ResFPRC:$Ra),
5715 (OpTy FPRC:$Rn), (OpVTy VPRC:$MRm), OpImm:$Imm))>;
5718 def : Pat<(ResTy (opnode
5719 (ResTy ResFPRC:$Ra),
5721 (OpTy (scalar_to_vector
5722 (ExTy (vector_extract (OpVTy VPRC:$MRm), OpImm:$Imm)))),
5723 (OpTy FPRC:$Rn))))),
5724 (ResTy (INST (ResTy ResFPRC:$Ra),
5725 (OpTy FPRC:$Rn), (OpVTy VPRC:$MRm), OpImm:$Imm))>;
5728 // Patterns for Scalar Signed saturating
5729 // doubling multiply-add long (scalar, by element)
5730 defm : Neon_ScalarXIndexedElem_MLAL_Patterns<int_arm_neon_vqadds,
5731 int_arm_neon_vqdmull, SQDMLALshv_4H, v1i32, FPR32, FPR16, v1i16, v4i16,
5732 i32, VPR64Lo, neon_uimm2_bare>;
5733 defm : Neon_ScalarXIndexedElem_MLAL_Patterns<int_arm_neon_vqadds,
5734 int_arm_neon_vqdmull, SQDMLALshv_8H, v1i32, FPR32, FPR16, v1i16, v8i16,
5735 i32, VPR128Lo, neon_uimm3_bare>;
5736 defm : Neon_ScalarXIndexedElem_MLAL_Patterns<int_arm_neon_vqadds,
5737 int_arm_neon_vqdmull, SQDMLALdsv_2S, v1i64, FPR64, FPR32, v1i32, v2i32,
5738 i32, VPR64Lo, neon_uimm1_bare>;
5739 defm : Neon_ScalarXIndexedElem_MLAL_Patterns<int_arm_neon_vqadds,
5740 int_arm_neon_vqdmull, SQDMLALdsv_4S, v1i64, FPR64, FPR32, v1i32, v4i32,
5741 i32, VPR128Lo, neon_uimm2_bare>;
5743 // Patterns for Scalar Signed saturating
5744 // doubling multiply-sub long (scalar, by element)
5745 defm : Neon_ScalarXIndexedElem_MLAL_Patterns<int_arm_neon_vqsubs,
5746 int_arm_neon_vqdmull, SQDMLSLshv_4H, v1i32, FPR32, FPR16, v1i16, v4i16,
5747 i32, VPR64Lo, neon_uimm2_bare>;
5748 defm : Neon_ScalarXIndexedElem_MLAL_Patterns<int_arm_neon_vqsubs,
5749 int_arm_neon_vqdmull, SQDMLSLshv_8H, v1i32, FPR32, FPR16, v1i16, v8i16,
5750 i32, VPR128Lo, neon_uimm3_bare>;
5751 defm : Neon_ScalarXIndexedElem_MLAL_Patterns<int_arm_neon_vqsubs,
5752 int_arm_neon_vqdmull, SQDMLSLdsv_2S, v1i64, FPR64, FPR32, v1i32, v2i32,
5753 i32, VPR64Lo, neon_uimm1_bare>;
5754 defm : Neon_ScalarXIndexedElem_MLAL_Patterns<int_arm_neon_vqsubs,
5755 int_arm_neon_vqdmull, SQDMLSLdsv_4S, v1i64, FPR64, FPR32, v1i32, v4i32,
5756 i32, VPR128Lo, neon_uimm2_bare>;
5758 // Scalar Signed saturating doubling multiply returning
5759 // high half (scalar, by element)
5760 def SQDMULHhhv_4H : NeonI_ScalarXIndexedElemArith<"sqdmulh",
5761 0b1100, ".h", 0b0, 0b0, 0b1, FPR16, FPR16, VPR64Lo, neon_uimm2_bare> {
5762 let Inst{11} = 0b0; // h
5763 let Inst{21} = Imm{1}; // l
5764 let Inst{20} = Imm{0}; // m
5765 let Inst{19-16} = MRm{3-0};
5767 def SQDMULHhhv_8H : NeonI_ScalarXIndexedElemArith<"sqdmulh",
5768 0b1100, ".h", 0b0, 0b0, 0b1, FPR16, FPR16, VPR128Lo, neon_uimm3_bare> {
5769 let Inst{11} = Imm{2}; // h
5770 let Inst{21} = Imm{1}; // l
5771 let Inst{20} = Imm{0}; // m
5772 let Inst{19-16} = MRm{3-0};
5774 def SQDMULHssv_2S : NeonI_ScalarXIndexedElemArith<"sqdmulh",
5775 0b1100, ".s", 0b0, 0b1, 0b0, FPR32, FPR32, VPR64, neon_uimm1_bare> {
5776 let Inst{11} = 0b0; // h
5777 let Inst{21} = Imm{0}; // l
5778 let Inst{20-16} = MRm;
5780 def SQDMULHssv_4S : NeonI_ScalarXIndexedElemArith<"sqdmulh",
5781 0b1100, ".s", 0b0, 0b1, 0b0, FPR32, FPR32, VPR128, neon_uimm2_bare> {
5782 let Inst{11} = Imm{1}; // h
5783 let Inst{21} = Imm{0}; // l
5784 let Inst{20-16} = MRm;
5787 // Patterns for Scalar Signed saturating doubling multiply returning
5788 // high half (scalar, by element)
5789 defm : Neon_ScalarXIndexedElem_MUL_Patterns<int_arm_neon_vqdmulh,
5790 SQDMULHhhv_4H, v1i16, FPR16, v1i16, i16, v4i16,
5791 i32, VPR64Lo, neon_uimm2_bare>;
5792 defm : Neon_ScalarXIndexedElem_MUL_Patterns<int_arm_neon_vqdmulh,
5793 SQDMULHhhv_8H, v1i16, FPR16, v1i16, i16, v8i16,
5794 i32, VPR128Lo, neon_uimm3_bare>;
5795 defm : Neon_ScalarXIndexedElem_MUL_Patterns<int_arm_neon_vqdmulh,
5796 SQDMULHssv_2S, v1i32, FPR32, v1i32, i32, v2i32,
5797 i32, VPR64Lo, neon_uimm1_bare>;
5798 defm : Neon_ScalarXIndexedElem_MUL_Patterns<int_arm_neon_vqdmulh,
5799 SQDMULHssv_4S, v1i32, FPR32, v1i32, i32, v4i32,
5800 i32, VPR128Lo, neon_uimm2_bare>;
5802 // Scalar Signed saturating rounding doubling multiply
5803 // returning high half (scalar, by element)
5804 def SQRDMULHhhv_4H : NeonI_ScalarXIndexedElemArith<"sqrdmulh",
5805 0b1101, ".h", 0b0, 0b0, 0b1, FPR16, FPR16, VPR64Lo, neon_uimm2_bare> {
5806 let Inst{11} = 0b0; // h
5807 let Inst{21} = Imm{1}; // l
5808 let Inst{20} = Imm{0}; // m
5809 let Inst{19-16} = MRm{3-0};
5811 def SQRDMULHhhv_8H : NeonI_ScalarXIndexedElemArith<"sqrdmulh",
5812 0b1101, ".h", 0b0, 0b0, 0b1, FPR16, FPR16, VPR128Lo, neon_uimm3_bare> {
5813 let Inst{11} = Imm{2}; // h
5814 let Inst{21} = Imm{1}; // l
5815 let Inst{20} = Imm{0}; // m
5816 let Inst{19-16} = MRm{3-0};
5818 def SQRDMULHssv_2S : NeonI_ScalarXIndexedElemArith<"sqrdmulh",
5819 0b1101, ".s", 0b0, 0b1, 0b0, FPR32, FPR32, VPR64, neon_uimm1_bare> {
5820 let Inst{11} = 0b0; // h
5821 let Inst{21} = Imm{0}; // l
5822 let Inst{20-16} = MRm;
5824 def SQRDMULHssv_4S : NeonI_ScalarXIndexedElemArith<"sqrdmulh",
5825 0b1101, ".s", 0b0, 0b1, 0b0, FPR32, FPR32, VPR128, neon_uimm2_bare> {
5826 let Inst{11} = Imm{1}; // h
5827 let Inst{21} = Imm{0}; // l
5828 let Inst{20-16} = MRm;
5831 defm : Neon_ScalarXIndexedElem_MUL_Patterns<int_arm_neon_vqrdmulh,
5832 SQRDMULHhhv_4H, v1i16, FPR16, v1i16, i16, v4i16, i32,
5833 VPR64Lo, neon_uimm2_bare>;
5834 defm : Neon_ScalarXIndexedElem_MUL_Patterns<int_arm_neon_vqrdmulh,
5835 SQRDMULHhhv_8H, v1i16, FPR16, v1i16, i16, v8i16, i32,
5836 VPR128Lo, neon_uimm3_bare>;
5837 defm : Neon_ScalarXIndexedElem_MUL_Patterns<int_arm_neon_vqrdmulh,
5838 SQRDMULHssv_2S, v1i32, FPR32, v1i32, i32, v2i32, i32,
5839 VPR64Lo, neon_uimm1_bare>;
5840 defm : Neon_ScalarXIndexedElem_MUL_Patterns<int_arm_neon_vqrdmulh,
5841 SQRDMULHssv_4S, v1i32, FPR32, v1i32, i32, v4i32, i32,
5842 VPR128Lo, neon_uimm2_bare>;
5844 // Scalar general arithmetic operation
5845 class Neon_Scalar_GeneralMath2D_pattern<SDPatternOperator opnode,
5847 : Pat<(v1f64 (opnode (v1f64 FPR64:$Rn))), (INST FPR64:$Rn)>;
5849 class Neon_Scalar_GeneralMath3D_pattern<SDPatternOperator opnode,
5851 : Pat<(v1f64 (opnode (v1f64 FPR64:$Rn), (v1f64 FPR64:$Rm))),
5852 (INST FPR64:$Rn, FPR64:$Rm)>;
5854 class Neon_Scalar_GeneralMath4D_pattern<SDPatternOperator opnode,
5856 : Pat<(v1f64 (opnode (v1f64 FPR64:$Rn), (v1f64 FPR64:$Rm),
5857 (v1f64 FPR64:$Ra))),
5858 (INST FPR64:$Rn, FPR64:$Rm, FPR64:$Ra)>;
5860 def : Neon_Scalar_GeneralMath3D_pattern<fadd, FADDddd>;
5861 def : Neon_Scalar_GeneralMath3D_pattern<fmul, FMULddd>;
5862 def : Neon_Scalar_GeneralMath3D_pattern<fsub, FSUBddd>;
5863 def : Neon_Scalar_GeneralMath3D_pattern<fdiv, FDIVddd>;
5864 def : Neon_Scalar_GeneralMath3D_pattern<int_arm_neon_vabds, FABDddd>;
5865 def : Neon_Scalar_GeneralMath3D_pattern<int_arm_neon_vmaxs, FMAXddd>;
5866 def : Neon_Scalar_GeneralMath3D_pattern<int_arm_neon_vmins, FMINddd>;
5867 def : Neon_Scalar_GeneralMath3D_pattern<int_aarch64_neon_vmaxnm, FMAXNMddd>;
5868 def : Neon_Scalar_GeneralMath3D_pattern<int_aarch64_neon_vminnm, FMINNMddd>;
5870 def : Neon_Scalar_GeneralMath2D_pattern<fabs, FABSdd>;
5871 def : Neon_Scalar_GeneralMath2D_pattern<fneg, FNEGdd>;
5873 def : Neon_Scalar_GeneralMath4D_pattern<fma, FMADDdddd>;
5874 def : Neon_Scalar_GeneralMath4D_pattern<fmsub, FMSUBdddd>;
5876 // Scalar Copy - DUP element to scalar
5877 class NeonI_Scalar_DUP<string asmop, string asmlane,
5878 RegisterClass ResRC, RegisterOperand VPRC,
5880 : NeonI_ScalarCopy<(outs ResRC:$Rd), (ins VPRC:$Rn, OpImm:$Imm),
5881 asmop # "\t$Rd, $Rn." # asmlane # "[$Imm]",
5887 def DUPbv_B : NeonI_Scalar_DUP<"dup", "b", FPR8, VPR128, neon_uimm4_bare> {
5888 let Inst{20-16} = {Imm{3}, Imm{2}, Imm{1}, Imm{0}, 0b1};
5890 def DUPhv_H : NeonI_Scalar_DUP<"dup", "h", FPR16, VPR128, neon_uimm3_bare> {
5891 let Inst{20-16} = {Imm{2}, Imm{1}, Imm{0}, 0b1, 0b0};
5893 def DUPsv_S : NeonI_Scalar_DUP<"dup", "s", FPR32, VPR128, neon_uimm2_bare> {
5894 let Inst{20-16} = {Imm{1}, Imm{0}, 0b1, 0b0, 0b0};
5896 def DUPdv_D : NeonI_Scalar_DUP<"dup", "d", FPR64, VPR128, neon_uimm1_bare> {
5897 let Inst{20-16} = {Imm, 0b1, 0b0, 0b0, 0b0};
5900 multiclass NeonI_Scalar_DUP_Elt_pattern<Instruction DUPI, ValueType ResTy,
5901 ValueType OpTy, Operand OpImm,
5902 ValueType OpNTy, ValueType ExTy, Operand OpNImm> {
5903 def : Pat<(ResTy (vector_extract (OpTy VPR128:$Rn), OpImm:$Imm)),
5904 (ResTy (DUPI (OpTy VPR128:$Rn), OpImm:$Imm))>;
5906 def : Pat<(ResTy (vector_extract (OpNTy VPR64:$Rn), OpNImm:$Imm)),
5908 (ExTy (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
5912 // Patterns for vector extract of FP data using scalar DUP instructions
5913 defm : NeonI_Scalar_DUP_Elt_pattern<DUPsv_S, f32,
5914 v4f32, neon_uimm2_bare, v2f32, v4f32, neon_uimm1_bare>;
5915 defm : NeonI_Scalar_DUP_Elt_pattern<DUPdv_D, f64,
5916 v2f64, neon_uimm1_bare, v1f64, v2f64, neon_uimm0_bare>;
5918 multiclass NeonI_Scalar_DUP_Ext_Vec_pattern<Instruction DUPI,
5919 ValueType ResTy, ValueType OpTy,Operand OpLImm,
5920 ValueType NOpTy, ValueType ExTy, Operand OpNImm> {
5922 def : Pat<(ResTy (extract_subvector (OpTy VPR128:$Rn), OpLImm:$Imm)),
5923 (ResTy (DUPI VPR128:$Rn, OpLImm:$Imm))>;
5925 def : Pat<(ResTy (extract_subvector (NOpTy VPR64:$Rn), OpNImm:$Imm)),
5927 (ExTy (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
5931 // Patterns for extract subvectors of v1ix data using scalar DUP instructions.
5932 defm : NeonI_Scalar_DUP_Ext_Vec_pattern<DUPbv_B, v1i8, v16i8, neon_uimm4_bare,
5933 v8i8, v16i8, neon_uimm3_bare>;
5934 defm : NeonI_Scalar_DUP_Ext_Vec_pattern<DUPhv_H, v1i16, v8i16, neon_uimm3_bare,
5935 v4i16, v8i16, neon_uimm2_bare>;
5936 defm : NeonI_Scalar_DUP_Ext_Vec_pattern<DUPsv_S, v1i32, v4i32, neon_uimm2_bare,
5937 v2i32, v4i32, neon_uimm1_bare>;
5939 multiclass NeonI_Scalar_DUP_Copy_pattern1<Instruction DUPI, ValueType ResTy,
5940 ValueType OpTy, ValueType ElemTy,
5941 Operand OpImm, ValueType OpNTy,
5942 ValueType ExTy, Operand OpNImm> {
5944 def : Pat<(ResTy (vector_insert (ResTy undef),
5945 (ElemTy (vector_extract (OpTy VPR128:$Rn), OpImm:$Imm)),
5946 (neon_uimm0_bare:$Imm))),
5947 (ResTy (DUPI (OpTy VPR128:$Rn), OpImm:$Imm))>;
5949 def : Pat<(ResTy (vector_insert (ResTy undef),
5950 (ElemTy (vector_extract (OpNTy VPR64:$Rn), OpNImm:$Imm)),
5953 (ExTy (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
5957 multiclass NeonI_Scalar_DUP_Copy_pattern2<Instruction DUPI, ValueType ResTy,
5958 ValueType OpTy, ValueType ElemTy,
5959 Operand OpImm, ValueType OpNTy,
5960 ValueType ExTy, Operand OpNImm> {
5962 def : Pat<(ResTy (scalar_to_vector
5963 (ElemTy (vector_extract (OpTy VPR128:$Rn), OpImm:$Imm)))),
5964 (ResTy (DUPI (OpTy VPR128:$Rn), OpImm:$Imm))>;
5966 def : Pat<(ResTy (scalar_to_vector
5967 (ElemTy (vector_extract (OpNTy VPR64:$Rn), OpNImm:$Imm)))),
5969 (ExTy (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
5973 // Patterns for vector copy to v1ix and v1fx vectors using scalar DUP
5975 defm : NeonI_Scalar_DUP_Copy_pattern1<DUPdv_D,
5976 v1i64, v2i64, i64, neon_uimm1_bare,
5977 v1i64, v2i64, neon_uimm0_bare>;
5978 defm : NeonI_Scalar_DUP_Copy_pattern1<DUPsv_S,
5979 v1i32, v4i32, i32, neon_uimm2_bare,
5980 v2i32, v4i32, neon_uimm1_bare>;
5981 defm : NeonI_Scalar_DUP_Copy_pattern1<DUPhv_H,
5982 v1i16, v8i16, i32, neon_uimm3_bare,
5983 v4i16, v8i16, neon_uimm2_bare>;
5984 defm : NeonI_Scalar_DUP_Copy_pattern1<DUPbv_B,
5985 v1i8, v16i8, i32, neon_uimm4_bare,
5986 v8i8, v16i8, neon_uimm3_bare>;
5987 defm : NeonI_Scalar_DUP_Copy_pattern2<DUPdv_D,
5988 v1i64, v2i64, i64, neon_uimm1_bare,
5989 v1i64, v2i64, neon_uimm0_bare>;
5990 defm : NeonI_Scalar_DUP_Copy_pattern2<DUPsv_S,
5991 v1i32, v4i32, i32, neon_uimm2_bare,
5992 v2i32, v4i32, neon_uimm1_bare>;
5993 defm : NeonI_Scalar_DUP_Copy_pattern2<DUPhv_H,
5994 v1i16, v8i16, i32, neon_uimm3_bare,
5995 v4i16, v8i16, neon_uimm2_bare>;
5996 defm : NeonI_Scalar_DUP_Copy_pattern2<DUPbv_B,
5997 v1i8, v16i8, i32, neon_uimm4_bare,
5998 v8i8, v16i8, neon_uimm3_bare>;
6000 multiclass NeonI_Scalar_DUP_alias<string asmop, string asmlane,
6001 Instruction DUPI, Operand OpImm,
6002 RegisterClass ResRC> {
6003 def : NeonInstAlias<!strconcat(asmop, "$Rd, $Rn" # asmlane # "[$Imm]"),
6004 (DUPI ResRC:$Rd, VPR128:$Rn, OpImm:$Imm), 0b0>;
6007 // Aliases for Scalar copy - DUP element (scalar)
6008 // FIXME: This is actually the preferred syntax but TableGen can't deal with
6009 // custom printing of aliases.
6010 defm : NeonI_Scalar_DUP_alias<"mov", ".b", DUPbv_B, neon_uimm4_bare, FPR8>;
6011 defm : NeonI_Scalar_DUP_alias<"mov", ".h", DUPhv_H, neon_uimm3_bare, FPR16>;
6012 defm : NeonI_Scalar_DUP_alias<"mov", ".s", DUPsv_S, neon_uimm2_bare, FPR32>;
6013 defm : NeonI_Scalar_DUP_alias<"mov", ".d", DUPdv_D, neon_uimm1_bare, FPR64>;
6015 multiclass NeonI_SDUP<PatFrag GetLow, PatFrag GetHigh, ValueType ResTy,
6017 def : Pat<(ResTy (GetLow VPR128:$Rn)),
6018 (ResTy (EXTRACT_SUBREG (OpTy VPR128:$Rn), sub_64))>;
6019 def : Pat<(ResTy (GetHigh VPR128:$Rn)),
6020 (ResTy (DUPdv_D (OpTy VPR128:$Rn), 1))>;
6023 defm : NeonI_SDUP<Neon_Low16B, Neon_High16B, v8i8, v16i8>;
6024 defm : NeonI_SDUP<Neon_Low8H, Neon_High8H, v4i16, v8i16>;
6025 defm : NeonI_SDUP<Neon_Low4S, Neon_High4S, v2i32, v4i32>;
6026 defm : NeonI_SDUP<Neon_Low2D, Neon_High2D, v1i64, v2i64>;
6027 defm : NeonI_SDUP<Neon_Low4float, Neon_High4float, v2f32, v4f32>;
6028 defm : NeonI_SDUP<Neon_Low2double, Neon_High2double, v1f64, v2f64>;
6030 //===----------------------------------------------------------------------===//
6031 // Non-Instruction Patterns
6032 //===----------------------------------------------------------------------===//
6034 // 64-bit vector bitcasts...
6036 def : Pat<(v1i64 (bitconvert (v8i8 VPR64:$src))), (v1i64 VPR64:$src)>;
6037 def : Pat<(v2f32 (bitconvert (v8i8 VPR64:$src))), (v2f32 VPR64:$src)>;
6038 def : Pat<(v2i32 (bitconvert (v8i8 VPR64:$src))), (v2i32 VPR64:$src)>;
6039 def : Pat<(v4i16 (bitconvert (v8i8 VPR64:$src))), (v4i16 VPR64:$src)>;
6041 def : Pat<(v1i64 (bitconvert (v4i16 VPR64:$src))), (v1i64 VPR64:$src)>;
6042 def : Pat<(v2i32 (bitconvert (v4i16 VPR64:$src))), (v2i32 VPR64:$src)>;
6043 def : Pat<(v2f32 (bitconvert (v4i16 VPR64:$src))), (v2f32 VPR64:$src)>;
6044 def : Pat<(v8i8 (bitconvert (v4i16 VPR64:$src))), (v8i8 VPR64:$src)>;
6046 def : Pat<(v1i64 (bitconvert (v2i32 VPR64:$src))), (v1i64 VPR64:$src)>;
6047 def : Pat<(v2f32 (bitconvert (v2i32 VPR64:$src))), (v2f32 VPR64:$src)>;
6048 def : Pat<(v4i16 (bitconvert (v2i32 VPR64:$src))), (v4i16 VPR64:$src)>;
6049 def : Pat<(v8i8 (bitconvert (v2i32 VPR64:$src))), (v8i8 VPR64:$src)>;
6051 def : Pat<(v1i64 (bitconvert (v2f32 VPR64:$src))), (v1i64 VPR64:$src)>;
6052 def : Pat<(v2i32 (bitconvert (v2f32 VPR64:$src))), (v2i32 VPR64:$src)>;
6053 def : Pat<(v4i16 (bitconvert (v2f32 VPR64:$src))), (v4i16 VPR64:$src)>;
6054 def : Pat<(v8i8 (bitconvert (v2f32 VPR64:$src))), (v8i8 VPR64:$src)>;
6056 def : Pat<(v2f32 (bitconvert (v1i64 VPR64:$src))), (v2f32 VPR64:$src)>;
6057 def : Pat<(v2i32 (bitconvert (v1i64 VPR64:$src))), (v2i32 VPR64:$src)>;
6058 def : Pat<(v4i16 (bitconvert (v1i64 VPR64:$src))), (v4i16 VPR64:$src)>;
6059 def : Pat<(v8i8 (bitconvert (v1i64 VPR64:$src))), (v8i8 VPR64:$src)>;
6061 // ..and 128-bit vector bitcasts...
6063 def : Pat<(v2f64 (bitconvert (v16i8 VPR128:$src))), (v2f64 VPR128:$src)>;
6064 def : Pat<(v2i64 (bitconvert (v16i8 VPR128:$src))), (v2i64 VPR128:$src)>;
6065 def : Pat<(v4f32 (bitconvert (v16i8 VPR128:$src))), (v4f32 VPR128:$src)>;
6066 def : Pat<(v4i32 (bitconvert (v16i8 VPR128:$src))), (v4i32 VPR128:$src)>;
6067 def : Pat<(v8i16 (bitconvert (v16i8 VPR128:$src))), (v8i16 VPR128:$src)>;
6069 def : Pat<(v2f64 (bitconvert (v8i16 VPR128:$src))), (v2f64 VPR128:$src)>;
6070 def : Pat<(v2i64 (bitconvert (v8i16 VPR128:$src))), (v2i64 VPR128:$src)>;
6071 def : Pat<(v4i32 (bitconvert (v8i16 VPR128:$src))), (v4i32 VPR128:$src)>;
6072 def : Pat<(v4f32 (bitconvert (v8i16 VPR128:$src))), (v4f32 VPR128:$src)>;
6073 def : Pat<(v16i8 (bitconvert (v8i16 VPR128:$src))), (v16i8 VPR128:$src)>;
6075 def : Pat<(v2f64 (bitconvert (v4i32 VPR128:$src))), (v2f64 VPR128:$src)>;
6076 def : Pat<(v2i64 (bitconvert (v4i32 VPR128:$src))), (v2i64 VPR128:$src)>;
6077 def : Pat<(v4f32 (bitconvert (v4i32 VPR128:$src))), (v4f32 VPR128:$src)>;
6078 def : Pat<(v8i16 (bitconvert (v4i32 VPR128:$src))), (v8i16 VPR128:$src)>;
6079 def : Pat<(v16i8 (bitconvert (v4i32 VPR128:$src))), (v16i8 VPR128:$src)>;
6081 def : Pat<(v2f64 (bitconvert (v4f32 VPR128:$src))), (v2f64 VPR128:$src)>;
6082 def : Pat<(v2i64 (bitconvert (v4f32 VPR128:$src))), (v2i64 VPR128:$src)>;
6083 def : Pat<(v4i32 (bitconvert (v4f32 VPR128:$src))), (v4i32 VPR128:$src)>;
6084 def : Pat<(v8i16 (bitconvert (v4f32 VPR128:$src))), (v8i16 VPR128:$src)>;
6085 def : Pat<(v16i8 (bitconvert (v4f32 VPR128:$src))), (v16i8 VPR128:$src)>;
6087 def : Pat<(v2f64 (bitconvert (v2i64 VPR128:$src))), (v2f64 VPR128:$src)>;
6088 def : Pat<(v4f32 (bitconvert (v2i64 VPR128:$src))), (v4f32 VPR128:$src)>;
6089 def : Pat<(v4i32 (bitconvert (v2i64 VPR128:$src))), (v4i32 VPR128:$src)>;
6090 def : Pat<(v8i16 (bitconvert (v2i64 VPR128:$src))), (v8i16 VPR128:$src)>;
6091 def : Pat<(v16i8 (bitconvert (v2i64 VPR128:$src))), (v16i8 VPR128:$src)>;
6093 def : Pat<(v2i64 (bitconvert (v2f64 VPR128:$src))), (v2i64 VPR128:$src)>;
6094 def : Pat<(v4f32 (bitconvert (v2f64 VPR128:$src))), (v4f32 VPR128:$src)>;
6095 def : Pat<(v4i32 (bitconvert (v2f64 VPR128:$src))), (v4i32 VPR128:$src)>;
6096 def : Pat<(v8i16 (bitconvert (v2f64 VPR128:$src))), (v8i16 VPR128:$src)>;
6097 def : Pat<(v16i8 (bitconvert (v2f64 VPR128:$src))), (v16i8 VPR128:$src)>;
6099 // ...and scalar bitcasts...
6100 def : Pat<(f16 (bitconvert (v1i16 FPR16:$src))), (f16 FPR16:$src)>;
6101 def : Pat<(f32 (bitconvert (v1i32 FPR32:$src))), (f32 FPR32:$src)>;
6102 def : Pat<(f64 (bitconvert (v1i64 FPR64:$src))), (f64 FPR64:$src)>;
6103 def : Pat<(f64 (bitconvert (v1f64 FPR64:$src))), (f64 FPR64:$src)>;
6105 def : Pat<(i64 (bitconvert (v1i64 FPR64:$src))), (FMOVxd $src)>;
6106 def : Pat<(i64 (bitconvert (v1f64 FPR64:$src))), (FMOVxd $src)>;
6107 def : Pat<(i64 (bitconvert (v2i32 FPR64:$src))), (FMOVxd $src)>;
6108 def : Pat<(i64 (bitconvert (v2f32 FPR64:$src))), (FMOVxd $src)>;
6109 def : Pat<(i64 (bitconvert (v4i16 FPR64:$src))), (FMOVxd $src)>;
6110 def : Pat<(i64 (bitconvert (v8i8 FPR64:$src))), (FMOVxd $src)>;
6112 def : Pat<(i32 (bitconvert (v1i32 FPR32:$src))), (FMOVws $src)>;
6114 def : Pat<(v8i8 (bitconvert (v1i64 VPR64:$src))), (v8i8 VPR64:$src)>;
6115 def : Pat<(v4i16 (bitconvert (v1i64 VPR64:$src))), (v4i16 VPR64:$src)>;
6116 def : Pat<(v2i32 (bitconvert (v1i64 VPR64:$src))), (v2i32 VPR64:$src)>;
6118 def : Pat<(f64 (bitconvert (v8i8 VPR64:$src))), (f64 VPR64:$src)>;
6119 def : Pat<(f64 (bitconvert (v4i16 VPR64:$src))), (f64 VPR64:$src)>;
6120 def : Pat<(f64 (bitconvert (v2i32 VPR64:$src))), (f64 VPR64:$src)>;
6121 def : Pat<(f64 (bitconvert (v2f32 VPR64:$src))), (f64 VPR64:$src)>;
6122 def : Pat<(f64 (bitconvert (v1i64 VPR64:$src))), (f64 VPR64:$src)>;
6124 def : Pat<(f128 (bitconvert (v16i8 VPR128:$src))), (f128 VPR128:$src)>;
6125 def : Pat<(f128 (bitconvert (v8i16 VPR128:$src))), (f128 VPR128:$src)>;
6126 def : Pat<(f128 (bitconvert (v4i32 VPR128:$src))), (f128 VPR128:$src)>;
6127 def : Pat<(f128 (bitconvert (v2i64 VPR128:$src))), (f128 VPR128:$src)>;
6128 def : Pat<(f128 (bitconvert (v4f32 VPR128:$src))), (f128 VPR128:$src)>;
6129 def : Pat<(f128 (bitconvert (v2f64 VPR128:$src))), (f128 VPR128:$src)>;
6131 def : Pat<(v1i16 (bitconvert (f16 FPR16:$src))), (v1i16 FPR16:$src)>;
6132 def : Pat<(v1i32 (bitconvert (f32 FPR32:$src))), (v1i32 FPR32:$src)>;
6133 def : Pat<(v1i64 (bitconvert (f64 FPR64:$src))), (v1i64 FPR64:$src)>;
6134 def : Pat<(v1f64 (bitconvert (f64 FPR64:$src))), (v1f64 FPR64:$src)>;
6136 def : Pat<(v1i64 (bitconvert (i64 GPR64:$src))), (FMOVdx $src)>;
6137 def : Pat<(v1f64 (bitconvert (i64 GPR64:$src))), (FMOVdx $src)>;
6138 def : Pat<(v2i32 (bitconvert (i64 GPR64:$src))), (FMOVdx $src)>;
6139 def : Pat<(v2f32 (bitconvert (i64 GPR64:$src))), (FMOVdx $src)>;
6140 def : Pat<(v4i16 (bitconvert (i64 GPR64:$src))), (FMOVdx $src)>;
6141 def : Pat<(v8i8 (bitconvert (i64 GPR64:$src))), (FMOVdx $src)>;
6143 def : Pat<(v1i32 (bitconvert (i32 GPR32:$src))), (FMOVsw $src)>;
6145 def : Pat<(v8i8 (bitconvert (f64 FPR64:$src))), (v8i8 FPR64:$src)>;
6146 def : Pat<(v4i16 (bitconvert (f64 FPR64:$src))), (v4i16 FPR64:$src)>;
6147 def : Pat<(v2i32 (bitconvert (f64 FPR64:$src))), (v2i32 FPR64:$src)>;
6148 def : Pat<(v2f32 (bitconvert (f64 FPR64:$src))), (v2f32 FPR64:$src)>;
6149 def : Pat<(v1i64 (bitconvert (f64 FPR64:$src))), (v1i64 FPR64:$src)>;
6151 def : Pat<(v16i8 (bitconvert (f128 FPR128:$src))), (v16i8 FPR128:$src)>;
6152 def : Pat<(v8i16 (bitconvert (f128 FPR128:$src))), (v8i16 FPR128:$src)>;
6153 def : Pat<(v4i32 (bitconvert (f128 FPR128:$src))), (v4i32 FPR128:$src)>;
6154 def : Pat<(v2i64 (bitconvert (f128 FPR128:$src))), (v2i64 FPR128:$src)>;
6155 def : Pat<(v4f32 (bitconvert (f128 FPR128:$src))), (v4f32 FPR128:$src)>;
6156 def : Pat<(v2f64 (bitconvert (f128 FPR128:$src))), (v2f64 FPR128:$src)>;
6158 // Scalar Three Same
6160 def neon_uimm3 : Operand<i64>,
6161 ImmLeaf<i64, [{return Imm < 8;}]> {
6162 let ParserMatchClass = uimm3_asmoperand;
6163 let PrintMethod = "printUImmHexOperand";
6166 def neon_uimm4 : Operand<i64>,
6167 ImmLeaf<i64, [{return Imm < 16;}]> {
6168 let ParserMatchClass = uimm4_asmoperand;
6169 let PrintMethod = "printUImmHexOperand";
6173 class NeonI_Extract<bit q, bits<2> op2, string asmop,
6174 string OpS, RegisterOperand OpVPR, Operand OpImm>
6175 : NeonI_BitExtract<q, op2, (outs OpVPR:$Rd),
6176 (ins OpVPR:$Rn, OpVPR:$Rm, OpImm:$Index),
6177 asmop # "\t$Rd." # OpS # ", $Rn." # OpS #
6178 ", $Rm." # OpS # ", $Index",
6184 def EXTvvvi_8b : NeonI_Extract<0b0, 0b00, "ext", "8b",
6185 VPR64, neon_uimm3> {
6186 let Inst{14-11} = {0b0, Index{2}, Index{1}, Index{0}};
6189 def EXTvvvi_16b: NeonI_Extract<0b1, 0b00, "ext", "16b",
6190 VPR128, neon_uimm4> {
6191 let Inst{14-11} = Index;
6194 class NI_Extract<ValueType OpTy, RegisterOperand OpVPR, Instruction INST,
6196 : Pat<(OpTy (Neon_vextract (OpTy OpVPR:$Rn), (OpTy OpVPR:$Rm),
6198 (INST OpVPR:$Rn, OpVPR:$Rm, OpImm:$Imm)>;
6200 def : NI_Extract<v8i8, VPR64, EXTvvvi_8b, neon_uimm3>;
6201 def : NI_Extract<v4i16, VPR64, EXTvvvi_8b, neon_uimm3>;
6202 def : NI_Extract<v2i32, VPR64, EXTvvvi_8b, neon_uimm3>;
6203 def : NI_Extract<v1i64, VPR64, EXTvvvi_8b, neon_uimm3>;
6204 def : NI_Extract<v2f32, VPR64, EXTvvvi_8b, neon_uimm3>;
6205 def : NI_Extract<v1f64, VPR64, EXTvvvi_8b, neon_uimm3>;
6206 def : NI_Extract<v16i8, VPR128, EXTvvvi_16b, neon_uimm4>;
6207 def : NI_Extract<v8i16, VPR128, EXTvvvi_16b, neon_uimm4>;
6208 def : NI_Extract<v4i32, VPR128, EXTvvvi_16b, neon_uimm4>;
6209 def : NI_Extract<v2i64, VPR128, EXTvvvi_16b, neon_uimm4>;
6210 def : NI_Extract<v4f32, VPR128, EXTvvvi_16b, neon_uimm4>;
6211 def : NI_Extract<v2f64, VPR128, EXTvvvi_16b, neon_uimm4>;
6214 class NI_TBL<bit q, bits<2> op2, bits<2> len, bit op,
6215 string asmop, string OpS, RegisterOperand OpVPR,
6216 RegisterOperand VecList>
6217 : NeonI_TBL<q, op2, len, op,
6218 (outs OpVPR:$Rd), (ins VecList:$Rn, OpVPR:$Rm),
6219 asmop # "\t$Rd." # OpS # ", $Rn, $Rm." # OpS,
6223 // The vectors in look up table are always 16b
6224 multiclass NI_TBL_pat<bits<2> len, bit op, string asmop, string List> {
6225 def _8b : NI_TBL<0, 0b00, len, op, asmop, "8b", VPR64,
6226 !cast<RegisterOperand>(List # "16B_operand")>;
6228 def _16b : NI_TBL<1, 0b00, len, op, asmop, "16b", VPR128,
6229 !cast<RegisterOperand>(List # "16B_operand")>;
6232 defm TBL1 : NI_TBL_pat<0b00, 0b0, "tbl", "VOne">;
6233 defm TBL2 : NI_TBL_pat<0b01, 0b0, "tbl", "VPair">;
6234 defm TBL3 : NI_TBL_pat<0b10, 0b0, "tbl", "VTriple">;
6235 defm TBL4 : NI_TBL_pat<0b11, 0b0, "tbl", "VQuad">;
6237 // Table lookup extention
6238 class NI_TBX<bit q, bits<2> op2, bits<2> len, bit op,
6239 string asmop, string OpS, RegisterOperand OpVPR,
6240 RegisterOperand VecList>
6241 : NeonI_TBL<q, op2, len, op,
6242 (outs OpVPR:$Rd), (ins OpVPR:$src, VecList:$Rn, OpVPR:$Rm),
6243 asmop # "\t$Rd." # OpS # ", $Rn, $Rm." # OpS,
6246 let Constraints = "$src = $Rd";
6249 // The vectors in look up table are always 16b
6250 multiclass NI_TBX_pat<bits<2> len, bit op, string asmop, string List> {
6251 def _8b : NI_TBX<0, 0b00, len, op, asmop, "8b", VPR64,
6252 !cast<RegisterOperand>(List # "16B_operand")>;
6254 def _16b : NI_TBX<1, 0b00, len, op, asmop, "16b", VPR128,
6255 !cast<RegisterOperand>(List # "16B_operand")>;
6258 defm TBX1 : NI_TBX_pat<0b00, 0b1, "tbx", "VOne">;
6259 defm TBX2 : NI_TBX_pat<0b01, 0b1, "tbx", "VPair">;
6260 defm TBX3 : NI_TBX_pat<0b10, 0b1, "tbx", "VTriple">;
6261 defm TBX4 : NI_TBX_pat<0b11, 0b1, "tbx", "VQuad">;
6263 class NeonI_INS_main<string asmop, string Res, ValueType ResTy,
6264 RegisterClass OpGPR, ValueType OpTy, Operand OpImm>
6265 : NeonI_copy<0b1, 0b0, 0b0011,
6266 (outs VPR128:$Rd), (ins VPR128:$src, OpGPR:$Rn, OpImm:$Imm),
6267 asmop # "\t$Rd." # Res # "[$Imm], $Rn",
6268 [(set (ResTy VPR128:$Rd),
6269 (ResTy (vector_insert
6270 (ResTy VPR128:$src),
6275 let Constraints = "$src = $Rd";
6278 //Insert element (vector, from main)
6279 def INSbw : NeonI_INS_main<"ins", "b", v16i8, GPR32, i32,
6281 let Inst{20-16} = {Imm{3}, Imm{2}, Imm{1}, Imm{0}, 0b1};
6283 def INShw : NeonI_INS_main<"ins", "h", v8i16, GPR32, i32,
6285 let Inst{20-16} = {Imm{2}, Imm{1}, Imm{0}, 0b1, 0b0};
6287 def INSsw : NeonI_INS_main<"ins", "s", v4i32, GPR32, i32,
6289 let Inst{20-16} = {Imm{1}, Imm{0}, 0b1, 0b0, 0b0};
6291 def INSdx : NeonI_INS_main<"ins", "d", v2i64, GPR64, i64,
6293 let Inst{20-16} = {Imm, 0b1, 0b0, 0b0, 0b0};
6296 def : NeonInstAlias<"mov $Rd.b[$Imm], $Rn",
6297 (INSbw VPR128:$Rd, GPR32:$Rn, neon_uimm4_bare:$Imm), 0>;
6298 def : NeonInstAlias<"mov $Rd.h[$Imm], $Rn",
6299 (INShw VPR128:$Rd, GPR32:$Rn, neon_uimm3_bare:$Imm), 0>;
6300 def : NeonInstAlias<"mov $Rd.s[$Imm], $Rn",
6301 (INSsw VPR128:$Rd, GPR32:$Rn, neon_uimm2_bare:$Imm), 0>;
6302 def : NeonInstAlias<"mov $Rd.d[$Imm], $Rn",
6303 (INSdx VPR128:$Rd, GPR64:$Rn, neon_uimm1_bare:$Imm), 0>;
6305 class Neon_INS_main_pattern <ValueType ResTy,ValueType ExtResTy,
6306 RegisterClass OpGPR, ValueType OpTy,
6307 Operand OpImm, Instruction INS>
6308 : Pat<(ResTy (vector_insert
6312 (ResTy (EXTRACT_SUBREG
6313 (ExtResTy (INS (ExtResTy (SUBREG_TO_REG (i64 0), VPR64:$src, sub_64)),
6314 OpGPR:$Rn, OpImm:$Imm)), sub_64))>;
6316 def INSbw_pattern : Neon_INS_main_pattern<v8i8, v16i8, GPR32, i32,
6317 neon_uimm3_bare, INSbw>;
6318 def INShw_pattern : Neon_INS_main_pattern<v4i16, v8i16, GPR32, i32,
6319 neon_uimm2_bare, INShw>;
6320 def INSsw_pattern : Neon_INS_main_pattern<v2i32, v4i32, GPR32, i32,
6321 neon_uimm1_bare, INSsw>;
6322 def INSdx_pattern : Neon_INS_main_pattern<v1i64, v2i64, GPR64, i64,
6323 neon_uimm0_bare, INSdx>;
6325 class NeonI_INS_element<string asmop, string Res, Operand ResImm>
6326 : NeonI_insert<0b1, 0b1,
6327 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn,
6328 ResImm:$Immd, ResImm:$Immn),
6329 asmop # "\t$Rd." # Res # "[$Immd], $Rn." # Res # "[$Immn]",
6332 let Constraints = "$src = $Rd";
6337 //Insert element (vector, from element)
6338 def INSELb : NeonI_INS_element<"ins", "b", neon_uimm4_bare> {
6339 let Inst{20-16} = {Immd{3}, Immd{2}, Immd{1}, Immd{0}, 0b1};
6340 let Inst{14-11} = {Immn{3}, Immn{2}, Immn{1}, Immn{0}};
6342 def INSELh : NeonI_INS_element<"ins", "h", neon_uimm3_bare> {
6343 let Inst{20-16} = {Immd{2}, Immd{1}, Immd{0}, 0b1, 0b0};
6344 let Inst{14-11} = {Immn{2}, Immn{1}, Immn{0}, 0b0};
6345 // bit 11 is unspecified, but should be set to zero.
6347 def INSELs : NeonI_INS_element<"ins", "s", neon_uimm2_bare> {
6348 let Inst{20-16} = {Immd{1}, Immd{0}, 0b1, 0b0, 0b0};
6349 let Inst{14-11} = {Immn{1}, Immn{0}, 0b0, 0b0};
6350 // bits 11-12 are unspecified, but should be set to zero.
6352 def INSELd : NeonI_INS_element<"ins", "d", neon_uimm1_bare> {
6353 let Inst{20-16} = {Immd, 0b1, 0b0, 0b0, 0b0};
6354 let Inst{14-11} = {Immn{0}, 0b0, 0b0, 0b0};
6355 // bits 11-13 are unspecified, but should be set to zero.
6358 def : NeonInstAlias<"mov $Rd.b[$Immd], $Rn.b[$Immn]",
6359 (INSELb VPR128:$Rd, VPR128:$Rn,
6360 neon_uimm4_bare:$Immd, neon_uimm4_bare:$Immn), 0>;
6361 def : NeonInstAlias<"mov $Rd.h[$Immd], $Rn.h[$Immn]",
6362 (INSELh VPR128:$Rd, VPR128:$Rn,
6363 neon_uimm3_bare:$Immd, neon_uimm3_bare:$Immn), 0>;
6364 def : NeonInstAlias<"mov $Rd.s[$Immd], $Rn.s[$Immn]",
6365 (INSELs VPR128:$Rd, VPR128:$Rn,
6366 neon_uimm2_bare:$Immd, neon_uimm2_bare:$Immn), 0>;
6367 def : NeonInstAlias<"mov $Rd.d[$Immd], $Rn.d[$Immn]",
6368 (INSELd VPR128:$Rd, VPR128:$Rn,
6369 neon_uimm1_bare:$Immd, neon_uimm1_bare:$Immn), 0>;
6371 multiclass Neon_INS_elt_pattern<ValueType ResTy, ValueType NaTy,
6372 ValueType MidTy, Operand StImm, Operand NaImm,
6374 def : Pat<(ResTy (vector_insert
6375 (ResTy VPR128:$src),
6376 (MidTy (vector_extract
6380 (INS (ResTy VPR128:$src), (ResTy VPR128:$Rn),
6381 StImm:$Immd, StImm:$Immn)>;
6383 def : Pat <(ResTy (vector_insert
6384 (ResTy VPR128:$src),
6385 (MidTy (vector_extract
6389 (INS (ResTy VPR128:$src),
6390 (ResTy (SUBREG_TO_REG (i64 0), (NaTy VPR64:$Rn), sub_64)),
6391 StImm:$Immd, NaImm:$Immn)>;
6393 def : Pat <(NaTy (vector_insert
6395 (MidTy (vector_extract
6399 (NaTy (EXTRACT_SUBREG
6401 (ResTy (SUBREG_TO_REG (i64 0), (NaTy VPR64:$src), sub_64)),
6403 NaImm:$Immd, StImm:$Immn)),
6406 def : Pat <(NaTy (vector_insert
6408 (MidTy (vector_extract
6412 (NaTy (EXTRACT_SUBREG
6414 (ResTy (SUBREG_TO_REG (i64 0), (NaTy VPR64:$src), sub_64)),
6415 (ResTy (SUBREG_TO_REG (i64 0), (NaTy VPR64:$Rn), sub_64)),
6416 NaImm:$Immd, NaImm:$Immn)),
6420 defm : Neon_INS_elt_pattern<v4f32, v2f32, f32, neon_uimm2_bare,
6421 neon_uimm1_bare, INSELs>;
6422 defm : Neon_INS_elt_pattern<v2f64, v1f64, f64, neon_uimm1_bare,
6423 neon_uimm0_bare, INSELd>;
6424 defm : Neon_INS_elt_pattern<v16i8, v8i8, i32, neon_uimm4_bare,
6425 neon_uimm3_bare, INSELb>;
6426 defm : Neon_INS_elt_pattern<v8i16, v4i16, i32, neon_uimm3_bare,
6427 neon_uimm2_bare, INSELh>;
6428 defm : Neon_INS_elt_pattern<v4i32, v2i32, i32, neon_uimm2_bare,
6429 neon_uimm1_bare, INSELs>;
6430 defm : Neon_INS_elt_pattern<v2i64, v1i64, i64, neon_uimm1_bare,
6431 neon_uimm0_bare, INSELd>;
6433 multiclass Neon_INS_elt_float_pattern<ValueType ResTy, ValueType NaTy,
6435 RegisterClass OpFPR, Operand ResImm,
6436 SubRegIndex SubIndex, Instruction INS> {
6437 def : Pat <(ResTy (vector_insert
6438 (ResTy VPR128:$src),
6441 (INS (ResTy VPR128:$src),
6442 (ResTy (SUBREG_TO_REG (i64 0), OpFPR:$Rn, SubIndex)),
6446 def : Pat <(NaTy (vector_insert
6450 (NaTy (EXTRACT_SUBREG
6452 (ResTy (SUBREG_TO_REG (i64 0), (NaTy VPR64:$src), sub_64)),
6453 (ResTy (SUBREG_TO_REG (i64 0), (MidTy OpFPR:$Rn), SubIndex)),
6459 defm : Neon_INS_elt_float_pattern<v4f32, v2f32, f32, FPR32, neon_uimm2_bare,
6461 defm : Neon_INS_elt_float_pattern<v2f64, v1f64, f64, FPR64, neon_uimm1_bare,
6464 class NeonI_SMOV<string asmop, string Res, bit Q,
6465 ValueType OpTy, ValueType eleTy,
6466 Operand OpImm, RegisterClass ResGPR, ValueType ResTy>
6467 : NeonI_copy<Q, 0b0, 0b0101,
6468 (outs ResGPR:$Rd), (ins VPR128:$Rn, OpImm:$Imm),
6469 asmop # "\t$Rd, $Rn." # Res # "[$Imm]",
6470 [(set (ResTy ResGPR:$Rd),
6472 (ResTy (vector_extract
6473 (OpTy VPR128:$Rn), (OpImm:$Imm))),
6479 //Signed integer move (main, from element)
6480 def SMOVwb : NeonI_SMOV<"smov", "b", 0b0, v16i8, i8, neon_uimm4_bare,
6482 let Inst{20-16} = {Imm{3}, Imm{2}, Imm{1}, Imm{0}, 0b1};
6484 def SMOVwh : NeonI_SMOV<"smov", "h", 0b0, v8i16, i16, neon_uimm3_bare,
6486 let Inst{20-16} = {Imm{2}, Imm{1}, Imm{0}, 0b1, 0b0};
6488 def SMOVxb : NeonI_SMOV<"smov", "b", 0b1, v16i8, i8, neon_uimm4_bare,
6490 let Inst{20-16} = {Imm{3}, Imm{2}, Imm{1}, Imm{0}, 0b1};
6492 def SMOVxh : NeonI_SMOV<"smov", "h", 0b1, v8i16, i16, neon_uimm3_bare,
6494 let Inst{20-16} = {Imm{2}, Imm{1}, Imm{0}, 0b1, 0b0};
6496 def SMOVxs : NeonI_SMOV<"smov", "s", 0b1, v4i32, i32, neon_uimm2_bare,
6498 let Inst{20-16} = {Imm{1}, Imm{0}, 0b1, 0b0, 0b0};
6501 multiclass Neon_SMOVx_pattern <ValueType StTy, ValueType NaTy,
6502 ValueType eleTy, Operand StImm, Operand NaImm,
6503 Instruction SMOVI> {
6504 def : Pat<(i64 (sext_inreg
6506 (i32 (vector_extract
6507 (StTy VPR128:$Rn), (StImm:$Imm))))),
6509 (SMOVI VPR128:$Rn, StImm:$Imm)>;
6511 def : Pat<(i64 (sext
6512 (i32 (vector_extract
6513 (StTy VPR128:$Rn), (StImm:$Imm))))),
6514 (SMOVI VPR128:$Rn, StImm:$Imm)>;
6516 def : Pat<(i64 (sext_inreg
6517 (i64 (vector_extract
6518 (NaTy VPR64:$Rn), (NaImm:$Imm))),
6520 (SMOVI (StTy (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
6523 def : Pat<(i64 (sext_inreg
6525 (i32 (vector_extract
6526 (NaTy VPR64:$Rn), (NaImm:$Imm))))),
6528 (SMOVI (StTy (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
6531 def : Pat<(i64 (sext
6532 (i32 (vector_extract
6533 (NaTy VPR64:$Rn), (NaImm:$Imm))))),
6534 (SMOVI (StTy (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
6538 defm : Neon_SMOVx_pattern<v16i8, v8i8, i8, neon_uimm4_bare,
6539 neon_uimm3_bare, SMOVxb>;
6540 defm : Neon_SMOVx_pattern<v8i16, v4i16, i16, neon_uimm3_bare,
6541 neon_uimm2_bare, SMOVxh>;
6542 defm : Neon_SMOVx_pattern<v4i32, v2i32, i32, neon_uimm2_bare,
6543 neon_uimm1_bare, SMOVxs>;
6545 class Neon_SMOVw_pattern <ValueType StTy, ValueType NaTy,
6546 ValueType eleTy, Operand StImm, Operand NaImm,
6548 : Pat<(i32 (sext_inreg
6549 (i32 (vector_extract
6550 (NaTy VPR64:$Rn), (NaImm:$Imm))),
6552 (SMOVI (StTy (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
6555 def : Neon_SMOVw_pattern<v16i8, v8i8, i8, neon_uimm4_bare,
6556 neon_uimm3_bare, SMOVwb>;
6557 def : Neon_SMOVw_pattern<v8i16, v4i16, i16, neon_uimm3_bare,
6558 neon_uimm2_bare, SMOVwh>;
6560 class NeonI_UMOV<string asmop, string Res, bit Q,
6561 ValueType OpTy, Operand OpImm,
6562 RegisterClass ResGPR, ValueType ResTy>
6563 : NeonI_copy<Q, 0b0, 0b0111,
6564 (outs ResGPR:$Rd), (ins VPR128:$Rn, OpImm:$Imm),
6565 asmop # "\t$Rd, $Rn." # Res # "[$Imm]",
6566 [(set (ResTy ResGPR:$Rd),
6567 (ResTy (vector_extract
6568 (OpTy VPR128:$Rn), (OpImm:$Imm))))],
6573 //Unsigned integer move (main, from element)
6574 def UMOVwb : NeonI_UMOV<"umov", "b", 0b0, v16i8, neon_uimm4_bare,
6576 let Inst{20-16} = {Imm{3}, Imm{2}, Imm{1}, Imm{0}, 0b1};
6578 def UMOVwh : NeonI_UMOV<"umov", "h", 0b0, v8i16, neon_uimm3_bare,
6580 let Inst{20-16} = {Imm{2}, Imm{1}, Imm{0}, 0b1, 0b0};
6582 def UMOVws : NeonI_UMOV<"umov", "s", 0b0, v4i32, neon_uimm2_bare,
6584 let Inst{20-16} = {Imm{1}, Imm{0}, 0b1, 0b0, 0b0};
6586 def UMOVxd : NeonI_UMOV<"umov", "d", 0b1, v2i64, neon_uimm1_bare,
6588 let Inst{20-16} = {Imm, 0b1, 0b0, 0b0, 0b0};
6591 def : NeonInstAlias<"mov $Rd, $Rn.s[$Imm]",
6592 (UMOVws GPR32:$Rd, VPR128:$Rn, neon_uimm2_bare:$Imm), 0>;
6593 def : NeonInstAlias<"mov $Rd, $Rn.d[$Imm]",
6594 (UMOVxd GPR64:$Rd, VPR128:$Rn, neon_uimm1_bare:$Imm), 0>;
6596 class Neon_UMOV_pattern <ValueType StTy, ValueType NaTy, ValueType ResTy,
6597 Operand StImm, Operand NaImm,
6599 : Pat<(ResTy (vector_extract
6600 (NaTy VPR64:$Rn), NaImm:$Imm)),
6601 (SMOVI (StTy (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
6604 def : Neon_UMOV_pattern<v16i8, v8i8, i32, neon_uimm4_bare,
6605 neon_uimm3_bare, UMOVwb>;
6606 def : Neon_UMOV_pattern<v8i16, v4i16, i32, neon_uimm3_bare,
6607 neon_uimm2_bare, UMOVwh>;
6608 def : Neon_UMOV_pattern<v4i32, v2i32, i32, neon_uimm2_bare,
6609 neon_uimm1_bare, UMOVws>;
6612 (i32 (vector_extract
6613 (v16i8 VPR128:$Rn), (neon_uimm4_bare:$Imm))),
6615 (UMOVwb VPR128:$Rn, neon_uimm4_bare:$Imm)>;
6618 (i32 (vector_extract
6619 (v8i16 VPR128:$Rn), (neon_uimm3_bare:$Imm))),
6621 (UMOVwh VPR128:$Rn, neon_uimm3_bare:$Imm)>;
6623 def : Pat<(i64 (zext
6624 (i32 (vector_extract
6625 (v2i64 VPR128:$Rn), (neon_uimm1_bare:$Imm))))),
6626 (UMOVxd VPR128:$Rn, neon_uimm1_bare:$Imm)>;
6629 (i32 (vector_extract
6630 (v8i8 VPR64:$Rn), (neon_uimm3_bare:$Imm))),
6632 (UMOVwb (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64),
6633 neon_uimm3_bare:$Imm)>;
6636 (i32 (vector_extract
6637 (v4i16 VPR64:$Rn), (neon_uimm2_bare:$Imm))),
6639 (UMOVwh (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64),
6640 neon_uimm2_bare:$Imm)>;
6642 def : Pat<(i64 (zext
6643 (i32 (vector_extract
6644 (v1i64 VPR64:$Rn), (neon_uimm0_bare:$Imm))))),
6645 (UMOVxd (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64),
6646 neon_uimm0_bare:$Imm)>;
6648 // Additional copy patterns for scalar types
6649 def : Pat<(i32 (vector_extract (v1i8 FPR8:$Rn), (i64 0))),
6651 (SUBREG_TO_REG (i64 0), FPR8:$Rn, sub_8)), (i64 0))>;
6653 def : Pat<(i32 (vector_extract (v1i16 FPR16:$Rn), (i64 0))),
6655 (SUBREG_TO_REG (i64 0), FPR16:$Rn, sub_16)), (i64 0))>;
6657 def : Pat<(i32 (vector_extract (v1i32 FPR32:$Rn), (i64 0))),
6658 (FMOVws FPR32:$Rn)>;
6660 def : Pat<(i64 (vector_extract (v1i64 FPR64:$Rn), (i64 0))),
6661 (FMOVxd FPR64:$Rn)>;
6663 def : Pat<(f64 (vector_extract (v1f64 FPR64:$Rn), (i64 0))),
6666 def : Pat<(v1i8 (scalar_to_vector GPR32:$Rn)),
6667 (v1i8 (EXTRACT_SUBREG (v16i8
6668 (INSbw (v16i8 (IMPLICIT_DEF)), $Rn, (i64 0))),
6671 def : Pat<(v1i16 (scalar_to_vector GPR32:$Rn)),
6672 (v1i16 (EXTRACT_SUBREG (v8i16
6673 (INShw (v8i16 (IMPLICIT_DEF)), $Rn, (i64 0))),
6676 def : Pat<(v1i32 (scalar_to_vector GPR32:$src)),
6679 def : Pat<(v1i64 (scalar_to_vector GPR64:$src)),
6682 def : Pat<(v8i8 (scalar_to_vector GPR32:$Rn)),
6683 (v8i8 (EXTRACT_SUBREG (v16i8
6684 (INSbw (v16i8 (IMPLICIT_DEF)), $Rn, (i64 0))),
6687 def : Pat<(v4i16 (scalar_to_vector GPR32:$Rn)),
6688 (v4i16 (EXTRACT_SUBREG (v8i16
6689 (INShw (v8i16 (IMPLICIT_DEF)), $Rn, (i64 0))),
6692 def : Pat<(v2i32 (scalar_to_vector GPR32:$Rn)),
6693 (v2i32 (EXTRACT_SUBREG (v16i8
6694 (INSsw (v4i32 (IMPLICIT_DEF)), $Rn, (i64 0))),
6697 def : Pat<(v16i8 (scalar_to_vector GPR32:$Rn)),
6698 (INSbw (v16i8 (IMPLICIT_DEF)), $Rn, (i64 0))>;
6700 def : Pat<(v8i16 (scalar_to_vector GPR32:$Rn)),
6701 (INShw (v8i16 (IMPLICIT_DEF)), $Rn, (i64 0))>;
6703 def : Pat<(v4i32 (scalar_to_vector GPR32:$Rn)),
6704 (INSsw (v4i32 (IMPLICIT_DEF)), $Rn, (i64 0))>;
6706 def : Pat<(v2i64 (scalar_to_vector GPR64:$Rn)),
6707 (INSdx (v2i64 (IMPLICIT_DEF)), $Rn, (i64 0))>;
6709 def : Pat<(v2i32 (scalar_to_vector GPR32:$Rn)),
6710 (v2i32 (EXTRACT_SUBREG (v16i8
6711 (INSsw (v4i32 (IMPLICIT_DEF)), $Rn, (i64 0))),
6714 def : Pat<(v2i32 (scalar_to_vector GPR32:$Rn)),
6715 (v2i32 (EXTRACT_SUBREG (v16i8
6716 (INSsw (v4i32 (IMPLICIT_DEF)), $Rn, (i64 0))),
6719 def : Pat<(v1f64 (scalar_to_vector (f64 FPR64:$Rn))),
6722 def : Pat<(v2f64 (scalar_to_vector (f64 FPR64:$src))),
6723 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)),
6724 (f64 FPR64:$src), sub_64)>;
6726 class NeonI_DUP_Elt<bit Q, string asmop, string rdlane, string rnlane,
6727 RegisterOperand ResVPR, Operand OpImm>
6728 : NeonI_copy<Q, 0b0, 0b0000, (outs ResVPR:$Rd),
6729 (ins VPR128:$Rn, OpImm:$Imm),
6730 asmop # "\t$Rd" # rdlane # ", $Rn" # rnlane # "[$Imm]",
6736 def DUPELT16b : NeonI_DUP_Elt<0b1, "dup", ".16b", ".b", VPR128,
6738 let Inst{20-16} = {Imm{3}, Imm{2}, Imm{1}, Imm{0}, 0b1};
6741 def DUPELT8h : NeonI_DUP_Elt<0b1, "dup", ".8h", ".h", VPR128,
6743 let Inst{20-16} = {Imm{2}, Imm{1}, Imm{0}, 0b1, 0b0};
6746 def DUPELT4s : NeonI_DUP_Elt<0b1, "dup", ".4s", ".s", VPR128,
6748 let Inst{20-16} = {Imm{1}, Imm{0}, 0b1, 0b0, 0b0};
6751 def DUPELT2d : NeonI_DUP_Elt<0b1, "dup", ".2d", ".d", VPR128,
6753 let Inst{20-16} = {Imm, 0b1, 0b0, 0b0, 0b0};
6756 def DUPELT8b : NeonI_DUP_Elt<0b0, "dup", ".8b", ".b", VPR64,
6758 let Inst{20-16} = {Imm{3}, Imm{2}, Imm{1}, Imm{0}, 0b1};
6761 def DUPELT4h : NeonI_DUP_Elt<0b0, "dup", ".4h", ".h", VPR64,
6763 let Inst{20-16} = {Imm{2}, Imm{1}, Imm{0}, 0b1, 0b0};
6766 def DUPELT2s : NeonI_DUP_Elt<0b0, "dup", ".2s", ".s", VPR64,
6768 let Inst{20-16} = {Imm{1}, Imm{0}, 0b1, 0b0, 0b0};
6771 multiclass NeonI_DUP_Elt_pattern<Instruction DUPELT, ValueType ResTy,
6772 ValueType OpTy,ValueType NaTy,
6773 ValueType ExTy, Operand OpLImm,
6775 def : Pat<(ResTy (Neon_vduplane (OpTy VPR128:$Rn), OpLImm:$Imm)),
6776 (ResTy (DUPELT (OpTy VPR128:$Rn), OpLImm:$Imm))>;
6778 def : Pat<(ResTy (Neon_vduplane
6779 (NaTy VPR64:$Rn), OpNImm:$Imm)),
6781 (ExTy (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)), OpNImm:$Imm))>;
6783 defm : NeonI_DUP_Elt_pattern<DUPELT16b, v16i8, v16i8, v8i8, v16i8,
6784 neon_uimm4_bare, neon_uimm3_bare>;
6785 defm : NeonI_DUP_Elt_pattern<DUPELT8b, v8i8, v16i8, v8i8, v16i8,
6786 neon_uimm4_bare, neon_uimm3_bare>;
6787 defm : NeonI_DUP_Elt_pattern<DUPELT8h, v8i16, v8i16, v4i16, v8i16,
6788 neon_uimm3_bare, neon_uimm2_bare>;
6789 defm : NeonI_DUP_Elt_pattern<DUPELT4h, v4i16, v8i16, v4i16, v8i16,
6790 neon_uimm3_bare, neon_uimm2_bare>;
6791 defm : NeonI_DUP_Elt_pattern<DUPELT4s, v4i32, v4i32, v2i32, v4i32,
6792 neon_uimm2_bare, neon_uimm1_bare>;
6793 defm : NeonI_DUP_Elt_pattern<DUPELT2s, v2i32, v4i32, v2i32, v4i32,
6794 neon_uimm2_bare, neon_uimm1_bare>;
6795 defm : NeonI_DUP_Elt_pattern<DUPELT2d, v2i64, v2i64, v1i64, v2i64,
6796 neon_uimm1_bare, neon_uimm0_bare>;
6797 defm : NeonI_DUP_Elt_pattern<DUPELT4s, v4f32, v4f32, v2f32, v4f32,
6798 neon_uimm2_bare, neon_uimm1_bare>;
6799 defm : NeonI_DUP_Elt_pattern<DUPELT2s, v2f32, v4f32, v2f32, v4f32,
6800 neon_uimm2_bare, neon_uimm1_bare>;
6801 defm : NeonI_DUP_Elt_pattern<DUPELT2d, v2f64, v2f64, v1f64, v2f64,
6802 neon_uimm1_bare, neon_uimm0_bare>;
6804 def : Pat<(v2f32 (Neon_vdup (f32 FPR32:$Rn))),
6806 (SUBREG_TO_REG (i64 0), FPR32:$Rn, sub_32),
6808 def : Pat<(v4f32 (Neon_vdup (f32 FPR32:$Rn))),
6810 (SUBREG_TO_REG (i64 0), FPR32:$Rn, sub_32),
6812 def : Pat<(v2f64 (Neon_vdup (f64 FPR64:$Rn))),
6814 (SUBREG_TO_REG (i64 0), FPR64:$Rn, sub_64),
6817 class NeonI_DUP<bit Q, string asmop, string rdlane,
6818 RegisterOperand ResVPR, ValueType ResTy,
6819 RegisterClass OpGPR, ValueType OpTy>
6820 : NeonI_copy<Q, 0b0, 0b0001, (outs ResVPR:$Rd), (ins OpGPR:$Rn),
6821 asmop # "\t$Rd" # rdlane # ", $Rn",
6822 [(set (ResTy ResVPR:$Rd),
6823 (ResTy (Neon_vdup (OpTy OpGPR:$Rn))))],
6826 def DUP16b : NeonI_DUP<0b1, "dup", ".16b", VPR128, v16i8, GPR32, i32> {
6827 let Inst{20-16} = 0b00001;
6828 // bits 17-20 are unspecified, but should be set to zero.
6831 def DUP8h : NeonI_DUP<0b1, "dup", ".8h", VPR128, v8i16, GPR32, i32> {
6832 let Inst{20-16} = 0b00010;
6833 // bits 18-20 are unspecified, but should be set to zero.
6836 def DUP4s : NeonI_DUP<0b1, "dup", ".4s", VPR128, v4i32, GPR32, i32> {
6837 let Inst{20-16} = 0b00100;
6838 // bits 19-20 are unspecified, but should be set to zero.
6841 def DUP2d : NeonI_DUP<0b1, "dup", ".2d", VPR128, v2i64, GPR64, i64> {
6842 let Inst{20-16} = 0b01000;
6843 // bit 20 is unspecified, but should be set to zero.
6846 def DUP8b : NeonI_DUP<0b0, "dup", ".8b", VPR64, v8i8, GPR32, i32> {
6847 let Inst{20-16} = 0b00001;
6848 // bits 17-20 are unspecified, but should be set to zero.
6851 def DUP4h : NeonI_DUP<0b0, "dup", ".4h", VPR64, v4i16, GPR32, i32> {
6852 let Inst{20-16} = 0b00010;
6853 // bits 18-20 are unspecified, but should be set to zero.
6856 def DUP2s : NeonI_DUP<0b0, "dup", ".2s", VPR64, v2i32, GPR32, i32> {
6857 let Inst{20-16} = 0b00100;
6858 // bits 19-20 are unspecified, but should be set to zero.
6861 // patterns for CONCAT_VECTORS
6862 multiclass Concat_Vector_Pattern<ValueType ResTy, ValueType OpTy> {
6863 def : Pat<(ResTy (concat_vectors (OpTy VPR64:$Rn), undef)),
6864 (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)>;
6865 def : Pat<(ResTy (concat_vectors (OpTy VPR64:$Rn), (OpTy VPR64:$Rm))),
6867 (v2i64 (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
6868 (v2i64 (SUBREG_TO_REG (i64 0), VPR64:$Rm, sub_64)),
6871 def : Pat<(ResTy (concat_vectors (OpTy VPR64:$Rn), (OpTy VPR64:$Rn))),
6873 (v2i64 (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
6877 defm : Concat_Vector_Pattern<v16i8, v8i8>;
6878 defm : Concat_Vector_Pattern<v8i16, v4i16>;
6879 defm : Concat_Vector_Pattern<v4i32, v2i32>;
6880 defm : Concat_Vector_Pattern<v2i64, v1i64>;
6881 defm : Concat_Vector_Pattern<v4f32, v2f32>;
6882 defm : Concat_Vector_Pattern<v2f64, v1f64>;
6884 //patterns for EXTRACT_SUBVECTOR
6885 def : Pat<(v8i8 (extract_subvector (v16i8 VPR128:$Rn), (i64 0))),
6886 (v8i8 (EXTRACT_SUBREG VPR128:$Rn, sub_64))>;
6887 def : Pat<(v4i16 (extract_subvector (v8i16 VPR128:$Rn), (i64 0))),
6888 (v4i16 (EXTRACT_SUBREG VPR128:$Rn, sub_64))>;
6889 def : Pat<(v2i32 (extract_subvector (v4i32 VPR128:$Rn), (i64 0))),
6890 (v2i32 (EXTRACT_SUBREG VPR128:$Rn, sub_64))>;
6891 def : Pat<(v1i64 (extract_subvector (v2i64 VPR128:$Rn), (i64 0))),
6892 (v1i64 (EXTRACT_SUBREG VPR128:$Rn, sub_64))>;
6893 def : Pat<(v2f32 (extract_subvector (v4f32 VPR128:$Rn), (i64 0))),
6894 (v2f32 (EXTRACT_SUBREG VPR128:$Rn, sub_64))>;
6895 def : Pat<(v1f64 (extract_subvector (v2f64 VPR128:$Rn), (i64 0))),
6896 (v1f64 (EXTRACT_SUBREG VPR128:$Rn, sub_64))>;
6898 // The followings are for instruction class (3V Elem)
6902 class NI_2VE<bit q, bit u, bits<2> size, bits<4> opcode,
6903 string asmop, string ResS, string OpS, string EleOpS,
6904 Operand OpImm, RegisterOperand ResVPR,
6905 RegisterOperand OpVPR, RegisterOperand EleOpVPR>
6906 : NeonI_2VElem<q, u, size, opcode,
6907 (outs ResVPR:$Rd), (ins ResVPR:$src, OpVPR:$Rn,
6908 EleOpVPR:$Re, OpImm:$Index),
6909 asmop # "\t$Rd." # ResS # ", $Rn." # OpS #
6910 ", $Re." # EleOpS # "[$Index]",
6916 let Constraints = "$src = $Rd";
6919 multiclass NI_2VE_v1<bit u, bits<4> opcode, string asmop> {
6920 // vector register class for element is always 128-bit to cover the max index
6921 def _2s4s : NI_2VE<0b0, u, 0b10, opcode, asmop, "2s", "2s", "s",
6922 neon_uimm2_bare, VPR64, VPR64, VPR128> {
6923 let Inst{11} = {Index{1}};
6924 let Inst{21} = {Index{0}};
6925 let Inst{20-16} = Re;
6928 def _4s4s : NI_2VE<0b1, u, 0b10, opcode, asmop, "4s", "4s", "s",
6929 neon_uimm2_bare, VPR128, VPR128, VPR128> {
6930 let Inst{11} = {Index{1}};
6931 let Inst{21} = {Index{0}};
6932 let Inst{20-16} = Re;
6935 // Index operations on 16-bit(H) elements are restricted to using v0-v15.
6936 def _4h8h : NI_2VE<0b0, u, 0b01, opcode, asmop, "4h", "4h", "h",
6937 neon_uimm3_bare, VPR64, VPR64, VPR128Lo> {
6938 let Inst{11} = {Index{2}};
6939 let Inst{21} = {Index{1}};
6940 let Inst{20} = {Index{0}};
6941 let Inst{19-16} = Re{3-0};
6944 def _8h8h : NI_2VE<0b1, u, 0b01, opcode, asmop, "8h", "8h", "h",
6945 neon_uimm3_bare, VPR128, VPR128, VPR128Lo> {
6946 let Inst{11} = {Index{2}};
6947 let Inst{21} = {Index{1}};
6948 let Inst{20} = {Index{0}};
6949 let Inst{19-16} = Re{3-0};
6953 defm MLAvve : NI_2VE_v1<0b1, 0b0000, "mla">;
6954 defm MLSvve : NI_2VE_v1<0b1, 0b0100, "mls">;
6956 // Pattern for lane in 128-bit vector
6957 class NI_2VE_laneq<Instruction INST, Operand OpImm, SDPatternOperator op,
6958 RegisterOperand ResVPR, RegisterOperand OpVPR,
6959 RegisterOperand EleOpVPR, ValueType ResTy, ValueType OpTy,
6961 : Pat<(ResTy (op (ResTy ResVPR:$src), (OpTy OpVPR:$Rn),
6962 (OpTy (Neon_vduplane (EleOpTy EleOpVPR:$Re), (i64 OpImm:$Index))))),
6963 (INST ResVPR:$src, OpVPR:$Rn, EleOpVPR:$Re, OpImm:$Index)>;
6965 // Pattern for lane in 64-bit vector
6966 class NI_2VE_lane<Instruction INST, Operand OpImm, SDPatternOperator op,
6967 RegisterOperand ResVPR, RegisterOperand OpVPR,
6968 RegisterOperand EleOpVPR, ValueType ResTy, ValueType OpTy,
6970 : Pat<(ResTy (op (ResTy ResVPR:$src), (OpTy OpVPR:$Rn),
6971 (OpTy (Neon_vduplane (EleOpTy EleOpVPR:$Re), (i64 OpImm:$Index))))),
6972 (INST ResVPR:$src, OpVPR:$Rn,
6973 (SUBREG_TO_REG (i64 0), EleOpVPR:$Re, sub_64), OpImm:$Index)>;
6975 multiclass NI_2VE_v1_pat<string subop, SDPatternOperator op>
6977 def : NI_2VE_laneq<!cast<Instruction>(subop # "_2s4s"), neon_uimm2_bare,
6978 op, VPR64, VPR64, VPR128, v2i32, v2i32, v4i32>;
6980 def : NI_2VE_laneq<!cast<Instruction>(subop # "_4s4s"), neon_uimm2_bare,
6981 op, VPR128, VPR128, VPR128, v4i32, v4i32, v4i32>;
6983 def : NI_2VE_laneq<!cast<Instruction>(subop # "_4h8h"), neon_uimm3_bare,
6984 op, VPR64, VPR64, VPR128Lo, v4i16, v4i16, v8i16>;
6986 def : NI_2VE_laneq<!cast<Instruction>(subop # "_8h8h"), neon_uimm3_bare,
6987 op, VPR128, VPR128, VPR128Lo, v8i16, v8i16, v8i16>;
6989 // Index can only be half of the max value for lane in 64-bit vector
6991 def : NI_2VE_lane<!cast<Instruction>(subop # "_2s4s"), neon_uimm1_bare,
6992 op, VPR64, VPR64, VPR64, v2i32, v2i32, v2i32>;
6994 def : NI_2VE_lane<!cast<Instruction>(subop # "_4h8h"), neon_uimm2_bare,
6995 op, VPR64, VPR64, VPR64Lo, v4i16, v4i16, v4i16>;
6998 defm MLA_lane_v1 : NI_2VE_v1_pat<"MLAvve", Neon_mla>;
6999 defm MLS_lane_v1 : NI_2VE_v1_pat<"MLSvve", Neon_mls>;
7001 class NI_2VE_2op<bit q, bit u, bits<2> size, bits<4> opcode,
7002 string asmop, string ResS, string OpS, string EleOpS,
7003 Operand OpImm, RegisterOperand ResVPR,
7004 RegisterOperand OpVPR, RegisterOperand EleOpVPR>
7005 : NeonI_2VElem<q, u, size, opcode,
7006 (outs ResVPR:$Rd), (ins OpVPR:$Rn,
7007 EleOpVPR:$Re, OpImm:$Index),
7008 asmop # "\t$Rd." # ResS # ", $Rn." # OpS #
7009 ", $Re." # EleOpS # "[$Index]",
7016 multiclass NI_2VE_v1_2op<bit u, bits<4> opcode, string asmop> {
7017 // vector register class for element is always 128-bit to cover the max index
7018 def _2s4s : NI_2VE_2op<0b0, u, 0b10, opcode, asmop, "2s", "2s", "s",
7019 neon_uimm2_bare, VPR64, VPR64, VPR128> {
7020 let Inst{11} = {Index{1}};
7021 let Inst{21} = {Index{0}};
7022 let Inst{20-16} = Re;
7025 def _4s4s : NI_2VE_2op<0b1, u, 0b10, opcode, asmop, "4s", "4s", "s",
7026 neon_uimm2_bare, VPR128, VPR128, VPR128> {
7027 let Inst{11} = {Index{1}};
7028 let Inst{21} = {Index{0}};
7029 let Inst{20-16} = Re;
7032 // Index operations on 16-bit(H) elements are restricted to using v0-v15.
7033 def _4h8h : NI_2VE_2op<0b0, u, 0b01, opcode, asmop, "4h", "4h", "h",
7034 neon_uimm3_bare, VPR64, VPR64, VPR128Lo> {
7035 let Inst{11} = {Index{2}};
7036 let Inst{21} = {Index{1}};
7037 let Inst{20} = {Index{0}};
7038 let Inst{19-16} = Re{3-0};
7041 def _8h8h : NI_2VE_2op<0b1, u, 0b01, opcode, asmop, "8h", "8h", "h",
7042 neon_uimm3_bare, VPR128, VPR128, VPR128Lo> {
7043 let Inst{11} = {Index{2}};
7044 let Inst{21} = {Index{1}};
7045 let Inst{20} = {Index{0}};
7046 let Inst{19-16} = Re{3-0};
7050 defm MULve : NI_2VE_v1_2op<0b0, 0b1000, "mul">;
7051 defm SQDMULHve : NI_2VE_v1_2op<0b0, 0b1100, "sqdmulh">;
7052 defm SQRDMULHve : NI_2VE_v1_2op<0b0, 0b1101, "sqrdmulh">;
7054 // Pattern for lane in 128-bit vector
7055 class NI_2VE_mul_laneq<Instruction INST, Operand OpImm, SDPatternOperator op,
7056 RegisterOperand OpVPR, RegisterOperand EleOpVPR,
7057 ValueType ResTy, ValueType OpTy, ValueType EleOpTy>
7058 : Pat<(ResTy (op (OpTy OpVPR:$Rn),
7059 (OpTy (Neon_vduplane (EleOpTy EleOpVPR:$Re), (i64 OpImm:$Index))))),
7060 (INST OpVPR:$Rn, EleOpVPR:$Re, OpImm:$Index)>;
7062 // Pattern for lane in 64-bit vector
7063 class NI_2VE_mul_lane<Instruction INST, Operand OpImm, SDPatternOperator op,
7064 RegisterOperand OpVPR, RegisterOperand EleOpVPR,
7065 ValueType ResTy, ValueType OpTy, ValueType EleOpTy>
7066 : Pat<(ResTy (op (OpTy OpVPR:$Rn),
7067 (OpTy (Neon_vduplane (EleOpTy EleOpVPR:$Re), (i64 OpImm:$Index))))),
7069 (SUBREG_TO_REG (i64 0), EleOpVPR:$Re, sub_64), OpImm:$Index)>;
7071 multiclass NI_2VE_mul_v1_pat<string subop, SDPatternOperator op> {
7072 def : NI_2VE_mul_laneq<!cast<Instruction>(subop # "_2s4s"), neon_uimm2_bare,
7073 op, VPR64, VPR128, v2i32, v2i32, v4i32>;
7075 def : NI_2VE_mul_laneq<!cast<Instruction>(subop # "_4s4s"), neon_uimm2_bare,
7076 op, VPR128, VPR128, v4i32, v4i32, v4i32>;
7078 def : NI_2VE_mul_laneq<!cast<Instruction>(subop # "_4h8h"), neon_uimm3_bare,
7079 op, VPR64, VPR128Lo, v4i16, v4i16, v8i16>;
7081 def : NI_2VE_mul_laneq<!cast<Instruction>(subop # "_8h8h"), neon_uimm3_bare,
7082 op, VPR128, VPR128Lo, v8i16, v8i16, v8i16>;
7084 // Index can only be half of the max value for lane in 64-bit vector
7086 def : NI_2VE_mul_lane<!cast<Instruction>(subop # "_2s4s"), neon_uimm1_bare,
7087 op, VPR64, VPR64, v2i32, v2i32, v2i32>;
7089 def : NI_2VE_mul_lane<!cast<Instruction>(subop # "_4h8h"), neon_uimm2_bare,
7090 op, VPR64, VPR64Lo, v4i16, v4i16, v4i16>;
7093 defm MUL_lane_v1 : NI_2VE_mul_v1_pat<"MULve", mul>;
7094 defm SQDMULH_lane_v1 : NI_2VE_mul_v1_pat<"SQDMULHve", int_arm_neon_vqdmulh>;
7095 defm SQRDMULH_lane_v1 : NI_2VE_mul_v1_pat<"SQRDMULHve", int_arm_neon_vqrdmulh>;
7099 multiclass NI_2VE_v2_2op<bit u, bits<4> opcode, string asmop> {
7100 // vector register class for element is always 128-bit to cover the max index
7101 def _2s4s : NI_2VE_2op<0b0, u, 0b10, opcode, asmop, "2s", "2s", "s",
7102 neon_uimm2_bare, VPR64, VPR64, VPR128> {
7103 let Inst{11} = {Index{1}};
7104 let Inst{21} = {Index{0}};
7105 let Inst{20-16} = Re;
7108 def _4s4s : NI_2VE_2op<0b1, u, 0b10, opcode, asmop, "4s", "4s", "s",
7109 neon_uimm2_bare, VPR128, VPR128, VPR128> {
7110 let Inst{11} = {Index{1}};
7111 let Inst{21} = {Index{0}};
7112 let Inst{20-16} = Re;
7115 // _1d2d doesn't exist!
7117 def _2d2d : NI_2VE_2op<0b1, u, 0b11, opcode, asmop, "2d", "2d", "d",
7118 neon_uimm1_bare, VPR128, VPR128, VPR128> {
7119 let Inst{11} = {Index{0}};
7121 let Inst{20-16} = Re;
7125 defm FMULve : NI_2VE_v2_2op<0b0, 0b1001, "fmul">;
7126 defm FMULXve : NI_2VE_v2_2op<0b1, 0b1001, "fmulx">;
7128 class NI_2VE_mul_lane_2d<Instruction INST, Operand OpImm, SDPatternOperator op,
7129 RegisterOperand OpVPR, RegisterOperand EleOpVPR,
7130 ValueType ResTy, ValueType OpTy, ValueType EleOpTy,
7131 SDPatternOperator coreop>
7132 : Pat<(ResTy (op (OpTy OpVPR:$Rn),
7133 (OpTy (coreop (EleOpTy EleOpVPR:$Re), (EleOpTy EleOpVPR:$Re))))),
7135 (SUBREG_TO_REG (i64 0), EleOpVPR:$Re, sub_64), 0)>;
7137 multiclass NI_2VE_mul_v2_pat<string subop, SDPatternOperator op> {
7138 def : NI_2VE_mul_laneq<!cast<Instruction>(subop # "_2s4s"), neon_uimm2_bare,
7139 op, VPR64, VPR128, v2f32, v2f32, v4f32>;
7141 def : NI_2VE_mul_laneq<!cast<Instruction>(subop # "_4s4s"), neon_uimm2_bare,
7142 op, VPR128, VPR128, v4f32, v4f32, v4f32>;
7144 def : NI_2VE_mul_laneq<!cast<Instruction>(subop # "_2d2d"), neon_uimm1_bare,
7145 op, VPR128, VPR128, v2f64, v2f64, v2f64>;
7147 // Index can only be half of the max value for lane in 64-bit vector
7149 def : NI_2VE_mul_lane<!cast<Instruction>(subop # "_2s4s"), neon_uimm1_bare,
7150 op, VPR64, VPR64, v2f32, v2f32, v2f32>;
7152 def : NI_2VE_mul_lane_2d<!cast<Instruction>(subop # "_2d2d"), neon_uimm1_bare,
7153 op, VPR128, VPR64, v2f64, v2f64, v1f64,
7154 BinOpFrag<(Neon_combine_2d node:$LHS, node:$RHS)>>;
7157 defm FMUL_lane_v2 : NI_2VE_mul_v2_pat<"FMULve", fmul>;
7158 defm FMULX_lane_v2 : NI_2VE_mul_v2_pat<"FMULXve", int_aarch64_neon_vmulx>;
7160 def : Pat<(v2f32 (fmul (v2f32 (Neon_vdup (f32 FPR32:$Re))),
7161 (v2f32 VPR64:$Rn))),
7162 (FMULve_2s4s VPR64:$Rn, (SUBREG_TO_REG (i32 0), $Re, sub_32), 0)>;
7164 def : Pat<(v4f32 (fmul (v4f32 (Neon_vdup (f32 FPR32:$Re))),
7165 (v4f32 VPR128:$Rn))),
7166 (FMULve_4s4s VPR128:$Rn, (SUBREG_TO_REG (i32 0), $Re, sub_32), 0)>;
7168 def : Pat<(v2f64 (fmul (v2f64 (Neon_vdup (f64 FPR64:$Re))),
7169 (v2f64 VPR128:$Rn))),
7170 (FMULve_2d2d VPR128:$Rn, (SUBREG_TO_REG (i64 0), $Re, sub_64), 0)>;
7172 // The followings are patterns using fma
7173 // -ffp-contract=fast generates fma
7175 multiclass NI_2VE_v2<bit u, bits<4> opcode, string asmop> {
7176 // vector register class for element is always 128-bit to cover the max index
7177 def _2s4s : NI_2VE<0b0, u, 0b10, opcode, asmop, "2s", "2s", "s",
7178 neon_uimm2_bare, VPR64, VPR64, VPR128> {
7179 let Inst{11} = {Index{1}};
7180 let Inst{21} = {Index{0}};
7181 let Inst{20-16} = Re;
7184 def _4s4s : NI_2VE<0b1, u, 0b10, opcode, asmop, "4s", "4s", "s",
7185 neon_uimm2_bare, VPR128, VPR128, VPR128> {
7186 let Inst{11} = {Index{1}};
7187 let Inst{21} = {Index{0}};
7188 let Inst{20-16} = Re;
7191 // _1d2d doesn't exist!
7193 def _2d2d : NI_2VE<0b1, u, 0b11, opcode, asmop, "2d", "2d", "d",
7194 neon_uimm1_bare, VPR128, VPR128, VPR128> {
7195 let Inst{11} = {Index{0}};
7197 let Inst{20-16} = Re;
7201 defm FMLAvve : NI_2VE_v2<0b0, 0b0001, "fmla">;
7202 defm FMLSvve : NI_2VE_v2<0b0, 0b0101, "fmls">;
7204 // Pattern for lane in 128-bit vector
7205 class NI_2VEswap_laneq<Instruction INST, Operand OpImm, SDPatternOperator op,
7206 RegisterOperand ResVPR, RegisterOperand OpVPR,
7207 ValueType ResTy, ValueType OpTy,
7208 SDPatternOperator coreop>
7209 : Pat<(ResTy (op (ResTy (coreop (OpTy OpVPR:$Re), (i64 OpImm:$Index))),
7210 (ResTy ResVPR:$src), (ResTy ResVPR:$Rn))),
7211 (INST ResVPR:$src, ResVPR:$Rn, OpVPR:$Re, OpImm:$Index)>;
7213 // Pattern for lane 0
7214 class NI_2VEfma_lane0<Instruction INST, SDPatternOperator op,
7215 RegisterOperand ResVPR, ValueType ResTy>
7216 : Pat<(ResTy (op (ResTy ResVPR:$Rn),
7217 (ResTy (Neon_vdup (f32 FPR32:$Re))),
7218 (ResTy ResVPR:$src))),
7219 (INST ResVPR:$src, ResVPR:$Rn,
7220 (SUBREG_TO_REG (i32 0), $Re, sub_32), 0)>;
7222 // Pattern for lane in 64-bit vector
7223 class NI_2VEswap_lane<Instruction INST, Operand OpImm, SDPatternOperator op,
7224 RegisterOperand ResVPR, RegisterOperand OpVPR,
7225 ValueType ResTy, ValueType OpTy,
7226 SDPatternOperator coreop>
7227 : Pat<(ResTy (op (ResTy (coreop (OpTy OpVPR:$Re), (i64 OpImm:$Index))),
7228 (ResTy ResVPR:$Rn), (ResTy ResVPR:$src))),
7229 (INST ResVPR:$src, ResVPR:$Rn,
7230 (SUBREG_TO_REG (i64 0), OpVPR:$Re, sub_64), OpImm:$Index)>;
7232 // Pattern for lane in 64-bit vector
7233 class NI_2VEswap_lane_2d2d<Instruction INST, Operand OpImm,
7234 SDPatternOperator op,
7235 RegisterOperand ResVPR, RegisterOperand OpVPR,
7236 ValueType ResTy, ValueType OpTy,
7237 SDPatternOperator coreop>
7238 : Pat<(ResTy (op (ResTy (coreop (OpTy OpVPR:$Re), (OpTy OpVPR:$Re))),
7239 (ResTy ResVPR:$Rn), (ResTy ResVPR:$src))),
7240 (INST ResVPR:$src, ResVPR:$Rn,
7241 (SUBREG_TO_REG (i64 0), OpVPR:$Re, sub_64), 0)>;
7244 multiclass NI_2VE_fma_v2_pat<string subop, SDPatternOperator op> {
7245 def : NI_2VEswap_laneq<!cast<Instruction>(subop # "_2s4s"),
7246 neon_uimm2_bare, op, VPR64, VPR128, v2f32, v4f32,
7247 BinOpFrag<(Neon_vduplane node:$LHS, node:$RHS)>>;
7249 def : NI_2VEfma_lane0<!cast<Instruction>(subop # "_2s4s"),
7252 def : NI_2VEswap_laneq<!cast<Instruction>(subop # "_4s4s"),
7253 neon_uimm2_bare, op, VPR128, VPR128, v4f32, v4f32,
7254 BinOpFrag<(Neon_vduplane node:$LHS, node:$RHS)>>;
7256 def : NI_2VEfma_lane0<!cast<Instruction>(subop # "_4s4s"),
7259 def : NI_2VEswap_laneq<!cast<Instruction>(subop # "_2d2d"),
7260 neon_uimm1_bare, op, VPR128, VPR128, v2f64, v2f64,
7261 BinOpFrag<(Neon_vduplane node:$LHS, node:$RHS)>>;
7263 // Index can only be half of the max value for lane in 64-bit vector
7265 def : NI_2VEswap_lane<!cast<Instruction>(subop # "_2s4s"),
7266 neon_uimm1_bare, op, VPR64, VPR64, v2f32, v2f32,
7267 BinOpFrag<(Neon_vduplane node:$LHS, node:$RHS)>>;
7269 def : NI_2VEswap_lane_2d2d<!cast<Instruction>(subop # "_2d2d"),
7270 neon_uimm1_bare, op, VPR128, VPR64, v2f64, v1f64,
7271 BinOpFrag<(Neon_combine_2d node:$LHS, node:$RHS)>>;
7274 defm FMLA_lane_v2_s : NI_2VE_fma_v2_pat<"FMLAvve", fma>;
7276 // Pattern for lane 0
7277 class NI_2VEfms_lane0<Instruction INST, SDPatternOperator op,
7278 RegisterOperand ResVPR, ValueType ResTy>
7279 : Pat<(ResTy (op (ResTy (fneg ResVPR:$Rn)),
7280 (ResTy (Neon_vdup (f32 FPR32:$Re))),
7281 (ResTy ResVPR:$src))),
7282 (INST ResVPR:$src, ResVPR:$Rn,
7283 (SUBREG_TO_REG (i32 0), $Re, sub_32), 0)>;
7285 multiclass NI_2VE_fms_v2_pat<string subop, SDPatternOperator op>
7287 def : NI_2VEswap_laneq<!cast<Instruction>(subop # "_2s4s"),
7288 neon_uimm2_bare, op, VPR64, VPR128, v2f32, v4f32,
7289 BinOpFrag<(fneg (Neon_vduplane node:$LHS, node:$RHS))>>;
7291 def : NI_2VEswap_laneq<!cast<Instruction>(subop # "_2s4s"),
7292 neon_uimm2_bare, op, VPR64, VPR128, v2f32, v4f32,
7293 BinOpFrag<(Neon_vduplane
7294 (fneg node:$LHS), node:$RHS)>>;
7296 def : NI_2VEfms_lane0<!cast<Instruction>(subop # "_2s4s"),
7299 def : NI_2VEswap_laneq<!cast<Instruction>(subop # "_4s4s"),
7300 neon_uimm2_bare, op, VPR128, VPR128, v4f32, v4f32,
7301 BinOpFrag<(fneg (Neon_vduplane
7302 node:$LHS, node:$RHS))>>;
7304 def : NI_2VEswap_laneq<!cast<Instruction>(subop # "_4s4s"),
7305 neon_uimm2_bare, op, VPR128, VPR128, v4f32, v4f32,
7306 BinOpFrag<(Neon_vduplane
7307 (fneg node:$LHS), node:$RHS)>>;
7309 def : NI_2VEfms_lane0<!cast<Instruction>(subop # "_4s4s"),
7312 def : NI_2VEswap_laneq<!cast<Instruction>(subop # "_2d2d"),
7313 neon_uimm1_bare, op, VPR128, VPR128, v2f64, v2f64,
7314 BinOpFrag<(fneg (Neon_vduplane
7315 node:$LHS, node:$RHS))>>;
7317 def : NI_2VEswap_laneq<!cast<Instruction>(subop # "_2d2d"),
7318 neon_uimm1_bare, op, VPR128, VPR128, v2f64, v2f64,
7319 BinOpFrag<(Neon_vduplane
7320 (fneg node:$LHS), node:$RHS)>>;
7322 // Index can only be half of the max value for lane in 64-bit vector
7324 def : NI_2VEswap_lane<!cast<Instruction>(subop # "_2s4s"),
7325 neon_uimm1_bare, op, VPR64, VPR64, v2f32, v2f32,
7326 BinOpFrag<(fneg (Neon_vduplane
7327 node:$LHS, node:$RHS))>>;
7329 def : NI_2VEswap_lane<!cast<Instruction>(subop # "_2s4s"),
7330 neon_uimm1_bare, op, VPR64, VPR64, v2f32, v2f32,
7331 BinOpFrag<(Neon_vduplane
7332 (fneg node:$LHS), node:$RHS)>>;
7334 def : NI_2VEswap_lane<!cast<Instruction>(subop # "_4s4s"),
7335 neon_uimm1_bare, op, VPR128, VPR64, v4f32, v2f32,
7336 BinOpFrag<(fneg (Neon_vduplane node:$LHS, node:$RHS))>>;
7338 def : NI_2VEswap_lane<!cast<Instruction>(subop # "_4s4s"),
7339 neon_uimm1_bare, op, VPR128, VPR64, v4f32, v2f32,
7340 BinOpFrag<(Neon_vduplane (fneg node:$LHS), node:$RHS)>>;
7342 def : NI_2VEswap_lane_2d2d<!cast<Instruction>(subop # "_2d2d"),
7343 neon_uimm1_bare, op, VPR128, VPR64, v2f64, v1f64,
7344 BinOpFrag<(fneg (Neon_combine_2d
7345 node:$LHS, node:$RHS))>>;
7347 def : NI_2VEswap_lane_2d2d<!cast<Instruction>(subop # "_2d2d"),
7348 neon_uimm1_bare, op, VPR128, VPR64, v2f64, v1f64,
7349 BinOpFrag<(Neon_combine_2d
7350 (fneg node:$LHS), (fneg node:$RHS))>>;
7353 defm FMLS_lane_v2_s : NI_2VE_fms_v2_pat<"FMLSvve", fma>;
7355 // Variant 3: Long type
7356 // E.g. SMLAL : 4S/4H/H (v0-v15), 2D/2S/S
7357 // SMLAL2: 4S/8H/H (v0-v15), 2D/4S/S
7359 multiclass NI_2VE_v3<bit u, bits<4> opcode, string asmop> {
7360 // vector register class for element is always 128-bit to cover the max index
7361 def _2d2s : NI_2VE<0b0, u, 0b10, opcode, asmop, "2d", "2s", "s",
7362 neon_uimm2_bare, VPR128, VPR64, VPR128> {
7363 let Inst{11} = {Index{1}};
7364 let Inst{21} = {Index{0}};
7365 let Inst{20-16} = Re;
7368 def _2d4s : NI_2VE<0b1, u, 0b10, opcode, asmop # "2", "2d", "4s", "s",
7369 neon_uimm2_bare, VPR128, VPR128, VPR128> {
7370 let Inst{11} = {Index{1}};
7371 let Inst{21} = {Index{0}};
7372 let Inst{20-16} = Re;
7375 // Index operations on 16-bit(H) elements are restricted to using v0-v15.
7376 def _4s8h : NI_2VE<0b1, u, 0b01, opcode, asmop # "2", "4s", "8h", "h",
7377 neon_uimm3_bare, VPR128, VPR128, VPR128Lo> {
7378 let Inst{11} = {Index{2}};
7379 let Inst{21} = {Index{1}};
7380 let Inst{20} = {Index{0}};
7381 let Inst{19-16} = Re{3-0};
7384 def _4s4h : NI_2VE<0b0, u, 0b01, opcode, asmop, "4s", "4h", "h",
7385 neon_uimm3_bare, VPR128, VPR64, VPR128Lo> {
7386 let Inst{11} = {Index{2}};
7387 let Inst{21} = {Index{1}};
7388 let Inst{20} = {Index{0}};
7389 let Inst{19-16} = Re{3-0};
7393 defm SMLALvve : NI_2VE_v3<0b0, 0b0010, "smlal">;
7394 defm UMLALvve : NI_2VE_v3<0b1, 0b0010, "umlal">;
7395 defm SMLSLvve : NI_2VE_v3<0b0, 0b0110, "smlsl">;
7396 defm UMLSLvve : NI_2VE_v3<0b1, 0b0110, "umlsl">;
7397 defm SQDMLALvve : NI_2VE_v3<0b0, 0b0011, "sqdmlal">;
7398 defm SQDMLSLvve : NI_2VE_v3<0b0, 0b0111, "sqdmlsl">;
7400 multiclass NI_2VE_v3_2op<bit u, bits<4> opcode, string asmop> {
7401 // vector register class for element is always 128-bit to cover the max index
7402 def _2d2s : NI_2VE_2op<0b0, u, 0b10, opcode, asmop, "2d", "2s", "s",
7403 neon_uimm2_bare, VPR128, VPR64, VPR128> {
7404 let Inst{11} = {Index{1}};
7405 let Inst{21} = {Index{0}};
7406 let Inst{20-16} = Re;
7409 def _2d4s : NI_2VE_2op<0b1, u, 0b10, opcode, asmop # "2", "2d", "4s", "s",
7410 neon_uimm2_bare, VPR128, VPR128, VPR128> {
7411 let Inst{11} = {Index{1}};
7412 let Inst{21} = {Index{0}};
7413 let Inst{20-16} = Re;
7416 // Index operations on 16-bit(H) elements are restricted to using v0-v15.
7417 def _4s8h : NI_2VE_2op<0b1, u, 0b01, opcode, asmop # "2", "4s", "8h", "h",
7418 neon_uimm3_bare, VPR128, VPR128, VPR128Lo> {
7419 let Inst{11} = {Index{2}};
7420 let Inst{21} = {Index{1}};
7421 let Inst{20} = {Index{0}};
7422 let Inst{19-16} = Re{3-0};
7425 def _4s4h : NI_2VE_2op<0b0, u, 0b01, opcode, asmop, "4s", "4h", "h",
7426 neon_uimm3_bare, VPR128, VPR64, VPR128Lo> {
7427 let Inst{11} = {Index{2}};
7428 let Inst{21} = {Index{1}};
7429 let Inst{20} = {Index{0}};
7430 let Inst{19-16} = Re{3-0};
7434 defm SMULLve : NI_2VE_v3_2op<0b0, 0b1010, "smull">;
7435 defm UMULLve : NI_2VE_v3_2op<0b1, 0b1010, "umull">;
7436 defm SQDMULLve : NI_2VE_v3_2op<0b0, 0b1011, "sqdmull">;
7438 def : Pat<(v1f64 (scalar_to_vector (f64 FPR64:$src))),
7441 // Pattern for lane in 128-bit vector
7442 class NI_2VEL2_laneq<Instruction INST, Operand OpImm, SDPatternOperator op,
7443 RegisterOperand EleOpVPR, ValueType ResTy,
7444 ValueType OpTy, ValueType EleOpTy, ValueType HalfOpTy,
7445 SDPatternOperator hiop>
7446 : Pat<(ResTy (op (ResTy VPR128:$src),
7447 (HalfOpTy (hiop (OpTy VPR128:$Rn))),
7448 (HalfOpTy (Neon_vduplane
7449 (EleOpTy EleOpVPR:$Re), (i64 OpImm:$Index))))),
7450 (INST VPR128:$src, VPR128:$Rn, EleOpVPR:$Re, OpImm:$Index)>;
7452 // Pattern for lane in 64-bit vector
7453 class NI_2VEL2_lane<Instruction INST, Operand OpImm, SDPatternOperator op,
7454 RegisterOperand EleOpVPR, ValueType ResTy,
7455 ValueType OpTy, ValueType EleOpTy, ValueType HalfOpTy,
7456 SDPatternOperator hiop>
7457 : Pat<(ResTy (op (ResTy VPR128:$src),
7458 (HalfOpTy (hiop (OpTy VPR128:$Rn))),
7459 (HalfOpTy (Neon_vduplane
7460 (EleOpTy EleOpVPR:$Re), (i64 OpImm:$Index))))),
7461 (INST VPR128:$src, VPR128:$Rn,
7462 (SUBREG_TO_REG (i64 0), EleOpVPR:$Re, sub_64), OpImm:$Index)>;
7464 class NI_2VEL2_lane0<Instruction INST, SDPatternOperator op,
7465 ValueType ResTy, ValueType OpTy, ValueType HalfOpTy,
7466 SDPatternOperator hiop, Instruction DupInst>
7467 : Pat<(ResTy (op (ResTy VPR128:$src),
7468 (HalfOpTy (hiop (OpTy VPR128:$Rn))),
7469 (HalfOpTy (Neon_vdup (i32 GPR32:$Re))))),
7470 (INST VPR128:$src, VPR128:$Rn, (DupInst $Re), 0)>;
7472 multiclass NI_2VEL_v3_pat<string subop, SDPatternOperator op> {
7473 def : NI_2VE_laneq<!cast<Instruction>(subop # "_4s4h"), neon_uimm3_bare,
7474 op, VPR128, VPR64, VPR128Lo, v4i32, v4i16, v8i16>;
7476 def : NI_2VE_laneq<!cast<Instruction>(subop # "_2d2s"), neon_uimm2_bare,
7477 op, VPR128, VPR64, VPR128, v2i64, v2i32, v4i32>;
7479 def : NI_2VEL2_laneq<!cast<Instruction>(subop # "_4s8h"), neon_uimm3_bare,
7480 op, VPR128Lo, v4i32, v8i16, v8i16, v4i16, Neon_High8H>;
7482 def : NI_2VEL2_laneq<!cast<Instruction>(subop # "_2d4s"), neon_uimm2_bare,
7483 op, VPR128, v2i64, v4i32, v4i32, v2i32, Neon_High4S>;
7485 def : NI_2VEL2_lane0<!cast<Instruction>(subop # "_4s8h"),
7486 op, v4i32, v8i16, v4i16, Neon_High8H, DUP8h>;
7488 def : NI_2VEL2_lane0<!cast<Instruction>(subop # "_2d4s"),
7489 op, v2i64, v4i32, v2i32, Neon_High4S, DUP4s>;
7491 // Index can only be half of the max value for lane in 64-bit vector
7493 def : NI_2VE_lane<!cast<Instruction>(subop # "_4s4h"), neon_uimm2_bare,
7494 op, VPR128, VPR64, VPR64Lo, v4i32, v4i16, v4i16>;
7496 def : NI_2VE_lane<!cast<Instruction>(subop # "_2d2s"), neon_uimm1_bare,
7497 op, VPR128, VPR64, VPR64, v2i64, v2i32, v2i32>;
7499 def : NI_2VEL2_lane<!cast<Instruction>(subop # "_4s8h"), neon_uimm2_bare,
7500 op, VPR64Lo, v4i32, v8i16, v4i16, v4i16, Neon_High8H>;
7502 def : NI_2VEL2_lane<!cast<Instruction>(subop # "_2d4s"), neon_uimm1_bare,
7503 op, VPR64, v2i64, v4i32, v2i32, v2i32, Neon_High4S>;
7506 defm SMLAL_lane_v3 : NI_2VEL_v3_pat<"SMLALvve", Neon_smlal>;
7507 defm UMLAL_lane_v3 : NI_2VEL_v3_pat<"UMLALvve", Neon_umlal>;
7508 defm SMLSL_lane_v3 : NI_2VEL_v3_pat<"SMLSLvve", Neon_smlsl>;
7509 defm UMLSL_lane_v3 : NI_2VEL_v3_pat<"UMLSLvve", Neon_umlsl>;
7511 // Pattern for lane in 128-bit vector
7512 class NI_2VEL2_mul_laneq<Instruction INST, Operand OpImm, SDPatternOperator op,
7513 RegisterOperand EleOpVPR, ValueType ResTy,
7514 ValueType OpTy, ValueType EleOpTy, ValueType HalfOpTy,
7515 SDPatternOperator hiop>
7517 (HalfOpTy (hiop (OpTy VPR128:$Rn))),
7518 (HalfOpTy (Neon_vduplane
7519 (EleOpTy EleOpVPR:$Re), (i64 OpImm:$Index))))),
7520 (INST VPR128:$Rn, EleOpVPR:$Re, OpImm:$Index)>;
7522 // Pattern for lane in 64-bit vector
7523 class NI_2VEL2_mul_lane<Instruction INST, Operand OpImm, SDPatternOperator op,
7524 RegisterOperand EleOpVPR, ValueType ResTy,
7525 ValueType OpTy, ValueType EleOpTy, ValueType HalfOpTy,
7526 SDPatternOperator hiop>
7528 (HalfOpTy (hiop (OpTy VPR128:$Rn))),
7529 (HalfOpTy (Neon_vduplane
7530 (EleOpTy EleOpVPR:$Re), (i64 OpImm:$Index))))),
7532 (SUBREG_TO_REG (i64 0), EleOpVPR:$Re, sub_64), OpImm:$Index)>;
7534 // Pattern for fixed lane 0
7535 class NI_2VEL2_mul_lane0<Instruction INST, SDPatternOperator op,
7536 ValueType ResTy, ValueType OpTy, ValueType HalfOpTy,
7537 SDPatternOperator hiop, Instruction DupInst>
7539 (HalfOpTy (hiop (OpTy VPR128:$Rn))),
7540 (HalfOpTy (Neon_vdup (i32 GPR32:$Re))))),
7541 (INST VPR128:$Rn, (DupInst $Re), 0)>;
7543 multiclass NI_2VEL_mul_v3_pat<string subop, SDPatternOperator op> {
7544 def : NI_2VE_mul_laneq<!cast<Instruction>(subop # "_4s4h"), neon_uimm3_bare,
7545 op, VPR64, VPR128Lo, v4i32, v4i16, v8i16>;
7547 def : NI_2VE_mul_laneq<!cast<Instruction>(subop # "_2d2s"), neon_uimm2_bare,
7548 op, VPR64, VPR128, v2i64, v2i32, v4i32>;
7550 def : NI_2VEL2_mul_laneq<!cast<Instruction>(subop # "_4s8h"), neon_uimm3_bare,
7551 op, VPR128Lo, v4i32, v8i16, v8i16, v4i16, Neon_High8H>;
7553 def : NI_2VEL2_mul_laneq<!cast<Instruction>(subop # "_2d4s"), neon_uimm2_bare,
7554 op, VPR128, v2i64, v4i32, v4i32, v2i32, Neon_High4S>;
7556 def : NI_2VEL2_mul_lane0<!cast<Instruction>(subop # "_4s8h"),
7557 op, v4i32, v8i16, v4i16, Neon_High8H, DUP8h>;
7559 def : NI_2VEL2_mul_lane0<!cast<Instruction>(subop # "_2d4s"),
7560 op, v2i64, v4i32, v2i32, Neon_High4S, DUP4s>;
7562 // Index can only be half of the max value for lane in 64-bit vector
7564 def : NI_2VE_mul_lane<!cast<Instruction>(subop # "_4s4h"), neon_uimm2_bare,
7565 op, VPR64, VPR64Lo, v4i32, v4i16, v4i16>;
7567 def : NI_2VE_mul_lane<!cast<Instruction>(subop # "_2d2s"), neon_uimm1_bare,
7568 op, VPR64, VPR64, v2i64, v2i32, v2i32>;
7570 def : NI_2VEL2_mul_lane<!cast<Instruction>(subop # "_4s8h"), neon_uimm2_bare,
7571 op, VPR64Lo, v4i32, v8i16, v4i16, v4i16, Neon_High8H>;
7573 def : NI_2VEL2_mul_lane<!cast<Instruction>(subop # "_2d4s"), neon_uimm1_bare,
7574 op, VPR64, v2i64, v4i32, v2i32, v2i32, Neon_High4S>;
7577 defm SMULL_lane_v3 : NI_2VEL_mul_v3_pat<"SMULLve", int_arm_neon_vmulls>;
7578 defm UMULL_lane_v3 : NI_2VEL_mul_v3_pat<"UMULLve", int_arm_neon_vmullu>;
7579 defm SQDMULL_lane_v3 : NI_2VEL_mul_v3_pat<"SQDMULLve", int_arm_neon_vqdmull>;
7581 multiclass NI_qdma<SDPatternOperator op> {
7582 def _4s : PatFrag<(ops node:$Ra, node:$Rn, node:$Rm),
7584 (v4i32 (int_arm_neon_vqdmull node:$Rn, node:$Rm)))>;
7586 def _2d : PatFrag<(ops node:$Ra, node:$Rn, node:$Rm),
7588 (v2i64 (int_arm_neon_vqdmull node:$Rn, node:$Rm)))>;
7591 defm Neon_qdmlal : NI_qdma<int_arm_neon_vqadds>;
7592 defm Neon_qdmlsl : NI_qdma<int_arm_neon_vqsubs>;
7594 multiclass NI_2VEL_v3_qdma_pat<string subop, string op> {
7595 def : NI_2VE_laneq<!cast<Instruction>(subop # "_4s4h"), neon_uimm3_bare,
7596 !cast<PatFrag>(op # "_4s"), VPR128, VPR64, VPR128Lo,
7597 v4i32, v4i16, v8i16>;
7599 def : NI_2VE_laneq<!cast<Instruction>(subop # "_2d2s"), neon_uimm2_bare,
7600 !cast<PatFrag>(op # "_2d"), VPR128, VPR64, VPR128,
7601 v2i64, v2i32, v4i32>;
7603 def : NI_2VEL2_laneq<!cast<Instruction>(subop # "_4s8h"), neon_uimm3_bare,
7604 !cast<PatFrag>(op # "_4s"), VPR128Lo,
7605 v4i32, v8i16, v8i16, v4i16, Neon_High8H>;
7607 def : NI_2VEL2_laneq<!cast<Instruction>(subop # "_2d4s"), neon_uimm2_bare,
7608 !cast<PatFrag>(op # "_2d"), VPR128,
7609 v2i64, v4i32, v4i32, v2i32, Neon_High4S>;
7611 def : NI_2VEL2_lane0<!cast<Instruction>(subop # "_4s8h"),
7612 !cast<PatFrag>(op # "_4s"),
7613 v4i32, v8i16, v4i16, Neon_High8H, DUP8h>;
7615 def : NI_2VEL2_lane0<!cast<Instruction>(subop # "_2d4s"),
7616 !cast<PatFrag>(op # "_2d"),
7617 v2i64, v4i32, v2i32, Neon_High4S, DUP4s>;
7619 // Index can only be half of the max value for lane in 64-bit vector
7621 def : NI_2VE_lane<!cast<Instruction>(subop # "_4s4h"), neon_uimm2_bare,
7622 !cast<PatFrag>(op # "_4s"), VPR128, VPR64, VPR64Lo,
7623 v4i32, v4i16, v4i16>;
7625 def : NI_2VE_lane<!cast<Instruction>(subop # "_2d2s"), neon_uimm1_bare,
7626 !cast<PatFrag>(op # "_2d"), VPR128, VPR64, VPR64,
7627 v2i64, v2i32, v2i32>;
7629 def : NI_2VEL2_lane<!cast<Instruction>(subop # "_4s8h"), neon_uimm2_bare,
7630 !cast<PatFrag>(op # "_4s"), VPR64Lo,
7631 v4i32, v8i16, v4i16, v4i16, Neon_High8H>;
7633 def : NI_2VEL2_lane<!cast<Instruction>(subop # "_2d4s"), neon_uimm1_bare,
7634 !cast<PatFrag>(op # "_2d"), VPR64,
7635 v2i64, v4i32, v2i32, v2i32, Neon_High4S>;
7638 defm SQDMLAL_lane_v3 : NI_2VEL_v3_qdma_pat<"SQDMLALvve", "Neon_qdmlal">;
7639 defm SQDMLSL_lane_v3 : NI_2VEL_v3_qdma_pat<"SQDMLSLvve", "Neon_qdmlsl">;
7641 // End of implementation for instruction class (3V Elem)
7643 class NeonI_REV<string asmop, string Res, bits<2> size, bit Q, bit U,
7644 bits<5> opcode, RegisterOperand ResVPR, ValueType ResTy,
7645 SDPatternOperator Neon_Rev>
7646 : NeonI_2VMisc<Q, U, size, opcode,
7647 (outs ResVPR:$Rd), (ins ResVPR:$Rn),
7648 asmop # "\t$Rd." # Res # ", $Rn." # Res,
7649 [(set (ResTy ResVPR:$Rd),
7650 (ResTy (Neon_Rev (ResTy ResVPR:$Rn))))],
7653 def REV64_16b : NeonI_REV<"rev64", "16b", 0b00, 0b1, 0b0, 0b00000, VPR128,
7655 def REV64_8h : NeonI_REV<"rev64", "8h", 0b01, 0b1, 0b0, 0b00000, VPR128,
7657 def REV64_4s : NeonI_REV<"rev64", "4s", 0b10, 0b1, 0b0, 0b00000, VPR128,
7659 def REV64_8b : NeonI_REV<"rev64", "8b", 0b00, 0b0, 0b0, 0b00000, VPR64,
7661 def REV64_4h : NeonI_REV<"rev64", "4h", 0b01, 0b0, 0b0, 0b00000, VPR64,
7663 def REV64_2s : NeonI_REV<"rev64", "2s", 0b10, 0b0, 0b0, 0b00000, VPR64,
7666 def : Pat<(v4f32 (Neon_rev64 (v4f32 VPR128:$Rn))), (REV64_4s VPR128:$Rn)>;
7667 def : Pat<(v2f32 (Neon_rev64 (v2f32 VPR64:$Rn))), (REV64_2s VPR64:$Rn)>;
7669 def REV32_16b : NeonI_REV<"rev32", "16b", 0b00, 0b1, 0b1, 0b00000, VPR128,
7671 def REV32_8h : NeonI_REV<"rev32", "8h", 0b01, 0b1, 0b1, 0b00000, VPR128,
7673 def REV32_8b : NeonI_REV<"rev32", "8b", 0b00, 0b0, 0b1, 0b00000, VPR64,
7675 def REV32_4h : NeonI_REV<"rev32", "4h", 0b01, 0b0, 0b1, 0b00000, VPR64,
7678 def REV16_16b : NeonI_REV<"rev16", "16b", 0b00, 0b1, 0b0, 0b00001, VPR128,
7680 def REV16_8b : NeonI_REV<"rev16", "8b", 0b00, 0b0, 0b0, 0b00001, VPR64,
7683 multiclass NeonI_PairwiseAdd<string asmop, bit U, bits<5> opcode,
7684 SDPatternOperator Neon_Padd> {
7685 def 16b8h : NeonI_2VMisc<0b1, U, 0b00, opcode,
7686 (outs VPR128:$Rd), (ins VPR128:$Rn),
7687 asmop # "\t$Rd.8h, $Rn.16b",
7688 [(set (v8i16 VPR128:$Rd),
7689 (v8i16 (Neon_Padd (v16i8 VPR128:$Rn))))],
7692 def 8b4h : NeonI_2VMisc<0b0, U, 0b00, opcode,
7693 (outs VPR64:$Rd), (ins VPR64:$Rn),
7694 asmop # "\t$Rd.4h, $Rn.8b",
7695 [(set (v4i16 VPR64:$Rd),
7696 (v4i16 (Neon_Padd (v8i8 VPR64:$Rn))))],
7699 def 8h4s : NeonI_2VMisc<0b1, U, 0b01, opcode,
7700 (outs VPR128:$Rd), (ins VPR128:$Rn),
7701 asmop # "\t$Rd.4s, $Rn.8h",
7702 [(set (v4i32 VPR128:$Rd),
7703 (v4i32 (Neon_Padd (v8i16 VPR128:$Rn))))],
7706 def 4h2s : NeonI_2VMisc<0b0, U, 0b01, opcode,
7707 (outs VPR64:$Rd), (ins VPR64:$Rn),
7708 asmop # "\t$Rd.2s, $Rn.4h",
7709 [(set (v2i32 VPR64:$Rd),
7710 (v2i32 (Neon_Padd (v4i16 VPR64:$Rn))))],
7713 def 4s2d : NeonI_2VMisc<0b1, U, 0b10, opcode,
7714 (outs VPR128:$Rd), (ins VPR128:$Rn),
7715 asmop # "\t$Rd.2d, $Rn.4s",
7716 [(set (v2i64 VPR128:$Rd),
7717 (v2i64 (Neon_Padd (v4i32 VPR128:$Rn))))],
7720 def 2s1d : NeonI_2VMisc<0b0, U, 0b10, opcode,
7721 (outs VPR64:$Rd), (ins VPR64:$Rn),
7722 asmop # "\t$Rd.1d, $Rn.2s",
7723 [(set (v1i64 VPR64:$Rd),
7724 (v1i64 (Neon_Padd (v2i32 VPR64:$Rn))))],
7728 defm SADDLP : NeonI_PairwiseAdd<"saddlp", 0b0, 0b00010,
7729 int_arm_neon_vpaddls>;
7730 defm UADDLP : NeonI_PairwiseAdd<"uaddlp", 0b1, 0b00010,
7731 int_arm_neon_vpaddlu>;
7733 def : Pat<(v1i64 (int_aarch64_neon_saddlv (v2i32 VPR64:$Rn))),
7735 def : Pat<(v1i64 (int_aarch64_neon_uaddlv (v2i32 VPR64:$Rn))),
7738 multiclass NeonI_PairwiseAddAcc<string asmop, bit U, bits<5> opcode,
7739 SDPatternOperator Neon_Padd> {
7740 let Constraints = "$src = $Rd" in {
7741 def 16b8h : NeonI_2VMisc<0b1, U, 0b00, opcode,
7742 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
7743 asmop # "\t$Rd.8h, $Rn.16b",
7744 [(set (v8i16 VPR128:$Rd),
7746 (v8i16 VPR128:$src), (v16i8 VPR128:$Rn))))],
7749 def 8b4h : NeonI_2VMisc<0b0, U, 0b00, opcode,
7750 (outs VPR64:$Rd), (ins VPR64:$src, VPR64:$Rn),
7751 asmop # "\t$Rd.4h, $Rn.8b",
7752 [(set (v4i16 VPR64:$Rd),
7754 (v4i16 VPR64:$src), (v8i8 VPR64:$Rn))))],
7757 def 8h4s : NeonI_2VMisc<0b1, U, 0b01, opcode,
7758 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
7759 asmop # "\t$Rd.4s, $Rn.8h",
7760 [(set (v4i32 VPR128:$Rd),
7762 (v4i32 VPR128:$src), (v8i16 VPR128:$Rn))))],
7765 def 4h2s : NeonI_2VMisc<0b0, U, 0b01, opcode,
7766 (outs VPR64:$Rd), (ins VPR64:$src, VPR64:$Rn),
7767 asmop # "\t$Rd.2s, $Rn.4h",
7768 [(set (v2i32 VPR64:$Rd),
7770 (v2i32 VPR64:$src), (v4i16 VPR64:$Rn))))],
7773 def 4s2d : NeonI_2VMisc<0b1, U, 0b10, opcode,
7774 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
7775 asmop # "\t$Rd.2d, $Rn.4s",
7776 [(set (v2i64 VPR128:$Rd),
7778 (v2i64 VPR128:$src), (v4i32 VPR128:$Rn))))],
7781 def 2s1d : NeonI_2VMisc<0b0, U, 0b10, opcode,
7782 (outs VPR64:$Rd), (ins VPR64:$src, VPR64:$Rn),
7783 asmop # "\t$Rd.1d, $Rn.2s",
7784 [(set (v1i64 VPR64:$Rd),
7786 (v1i64 VPR64:$src), (v2i32 VPR64:$Rn))))],
7791 defm SADALP : NeonI_PairwiseAddAcc<"sadalp", 0b0, 0b00110,
7792 int_arm_neon_vpadals>;
7793 defm UADALP : NeonI_PairwiseAddAcc<"uadalp", 0b1, 0b00110,
7794 int_arm_neon_vpadalu>;
7796 multiclass NeonI_2VMisc_BHSDsize_1Arg<string asmop, bit U, bits<5> opcode> {
7797 def 16b : NeonI_2VMisc<0b1, U, 0b00, opcode,
7798 (outs VPR128:$Rd), (ins VPR128:$Rn),
7799 asmop # "\t$Rd.16b, $Rn.16b",
7802 def 8h : NeonI_2VMisc<0b1, U, 0b01, opcode,
7803 (outs VPR128:$Rd), (ins VPR128:$Rn),
7804 asmop # "\t$Rd.8h, $Rn.8h",
7807 def 4s : NeonI_2VMisc<0b1, U, 0b10, opcode,
7808 (outs VPR128:$Rd), (ins VPR128:$Rn),
7809 asmop # "\t$Rd.4s, $Rn.4s",
7812 def 2d : NeonI_2VMisc<0b1, U, 0b11, opcode,
7813 (outs VPR128:$Rd), (ins VPR128:$Rn),
7814 asmop # "\t$Rd.2d, $Rn.2d",
7817 def 8b : NeonI_2VMisc<0b0, U, 0b00, opcode,
7818 (outs VPR64:$Rd), (ins VPR64:$Rn),
7819 asmop # "\t$Rd.8b, $Rn.8b",
7822 def 4h : NeonI_2VMisc<0b0, U, 0b01, opcode,
7823 (outs VPR64:$Rd), (ins VPR64:$Rn),
7824 asmop # "\t$Rd.4h, $Rn.4h",
7827 def 2s : NeonI_2VMisc<0b0, U, 0b10, opcode,
7828 (outs VPR64:$Rd), (ins VPR64:$Rn),
7829 asmop # "\t$Rd.2s, $Rn.2s",
7833 defm SQABS : NeonI_2VMisc_BHSDsize_1Arg<"sqabs", 0b0, 0b00111>;
7834 defm SQNEG : NeonI_2VMisc_BHSDsize_1Arg<"sqneg", 0b1, 0b00111>;
7835 defm ABS : NeonI_2VMisc_BHSDsize_1Arg<"abs", 0b0, 0b01011>;
7836 defm NEG : NeonI_2VMisc_BHSDsize_1Arg<"neg", 0b1, 0b01011>;
7838 multiclass NeonI_2VMisc_BHSD_1Arg_Pattern<string Prefix,
7839 SDPatternOperator Neon_Op> {
7840 def : Pat<(v16i8 (Neon_Op (v16i8 VPR128:$Rn))),
7841 (v16i8 (!cast<Instruction>(Prefix # 16b) (v16i8 VPR128:$Rn)))>;
7843 def : Pat<(v8i16 (Neon_Op (v8i16 VPR128:$Rn))),
7844 (v8i16 (!cast<Instruction>(Prefix # 8h) (v8i16 VPR128:$Rn)))>;
7846 def : Pat<(v4i32 (Neon_Op (v4i32 VPR128:$Rn))),
7847 (v4i32 (!cast<Instruction>(Prefix # 4s) (v4i32 VPR128:$Rn)))>;
7849 def : Pat<(v2i64 (Neon_Op (v2i64 VPR128:$Rn))),
7850 (v2i64 (!cast<Instruction>(Prefix # 2d) (v2i64 VPR128:$Rn)))>;
7852 def : Pat<(v8i8 (Neon_Op (v8i8 VPR64:$Rn))),
7853 (v8i8 (!cast<Instruction>(Prefix # 8b) (v8i8 VPR64:$Rn)))>;
7855 def : Pat<(v4i16 (Neon_Op (v4i16 VPR64:$Rn))),
7856 (v4i16 (!cast<Instruction>(Prefix # 4h) (v4i16 VPR64:$Rn)))>;
7858 def : Pat<(v2i32 (Neon_Op (v2i32 VPR64:$Rn))),
7859 (v2i32 (!cast<Instruction>(Prefix # 2s) (v2i32 VPR64:$Rn)))>;
7862 defm : NeonI_2VMisc_BHSD_1Arg_Pattern<"SQABS", int_arm_neon_vqabs>;
7863 defm : NeonI_2VMisc_BHSD_1Arg_Pattern<"SQNEG", int_arm_neon_vqneg>;
7864 defm : NeonI_2VMisc_BHSD_1Arg_Pattern<"ABS", int_arm_neon_vabs>;
7866 def : Pat<(v16i8 (sub
7867 (v16i8 Neon_AllZero),
7868 (v16i8 VPR128:$Rn))),
7869 (v16i8 (NEG16b (v16i8 VPR128:$Rn)))>;
7870 def : Pat<(v8i8 (sub
7871 (v8i8 Neon_AllZero),
7873 (v8i8 (NEG8b (v8i8 VPR64:$Rn)))>;
7874 def : Pat<(v8i16 (sub
7875 (v8i16 (bitconvert (v16i8 Neon_AllZero))),
7876 (v8i16 VPR128:$Rn))),
7877 (v8i16 (NEG8h (v8i16 VPR128:$Rn)))>;
7878 def : Pat<(v4i16 (sub
7879 (v4i16 (bitconvert (v8i8 Neon_AllZero))),
7880 (v4i16 VPR64:$Rn))),
7881 (v4i16 (NEG4h (v4i16 VPR64:$Rn)))>;
7882 def : Pat<(v4i32 (sub
7883 (v4i32 (bitconvert (v16i8 Neon_AllZero))),
7884 (v4i32 VPR128:$Rn))),
7885 (v4i32 (NEG4s (v4i32 VPR128:$Rn)))>;
7886 def : Pat<(v2i32 (sub
7887 (v2i32 (bitconvert (v8i8 Neon_AllZero))),
7888 (v2i32 VPR64:$Rn))),
7889 (v2i32 (NEG2s (v2i32 VPR64:$Rn)))>;
7890 def : Pat<(v2i64 (sub
7891 (v2i64 (bitconvert (v16i8 Neon_AllZero))),
7892 (v2i64 VPR128:$Rn))),
7893 (v2i64 (NEG2d (v2i64 VPR128:$Rn)))>;
7895 multiclass NeonI_2VMisc_BHSDsize_2Args<string asmop, bit U, bits<5> opcode> {
7896 let Constraints = "$src = $Rd" in {
7897 def 16b : NeonI_2VMisc<0b1, U, 0b00, opcode,
7898 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
7899 asmop # "\t$Rd.16b, $Rn.16b",
7902 def 8h : NeonI_2VMisc<0b1, U, 0b01, opcode,
7903 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
7904 asmop # "\t$Rd.8h, $Rn.8h",
7907 def 4s : NeonI_2VMisc<0b1, U, 0b10, opcode,
7908 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
7909 asmop # "\t$Rd.4s, $Rn.4s",
7912 def 2d : NeonI_2VMisc<0b1, U, 0b11, opcode,
7913 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
7914 asmop # "\t$Rd.2d, $Rn.2d",
7917 def 8b : NeonI_2VMisc<0b0, U, 0b00, opcode,
7918 (outs VPR64:$Rd), (ins VPR64:$src, VPR64:$Rn),
7919 asmop # "\t$Rd.8b, $Rn.8b",
7922 def 4h : NeonI_2VMisc<0b0, U, 0b01, opcode,
7923 (outs VPR64:$Rd), (ins VPR64:$src, VPR64:$Rn),
7924 asmop # "\t$Rd.4h, $Rn.4h",
7927 def 2s : NeonI_2VMisc<0b0, U, 0b10, opcode,
7928 (outs VPR64:$Rd), (ins VPR64:$src, VPR64:$Rn),
7929 asmop # "\t$Rd.2s, $Rn.2s",
7934 defm SUQADD : NeonI_2VMisc_BHSDsize_2Args<"suqadd", 0b0, 0b00011>;
7935 defm USQADD : NeonI_2VMisc_BHSDsize_2Args<"usqadd", 0b1, 0b00011>;
7937 multiclass NeonI_2VMisc_BHSD_2Args_Pattern<string Prefix,
7938 SDPatternOperator Neon_Op> {
7939 def : Pat<(v16i8 (Neon_Op (v16i8 VPR128:$src), (v16i8 VPR128:$Rn))),
7940 (v16i8 (!cast<Instruction>(Prefix # 16b)
7941 (v16i8 VPR128:$src), (v16i8 VPR128:$Rn)))>;
7943 def : Pat<(v8i16 (Neon_Op (v8i16 VPR128:$src), (v8i16 VPR128:$Rn))),
7944 (v8i16 (!cast<Instruction>(Prefix # 8h)
7945 (v8i16 VPR128:$src), (v8i16 VPR128:$Rn)))>;
7947 def : Pat<(v4i32 (Neon_Op (v4i32 VPR128:$src), (v4i32 VPR128:$Rn))),
7948 (v4i32 (!cast<Instruction>(Prefix # 4s)
7949 (v4i32 VPR128:$src), (v4i32 VPR128:$Rn)))>;
7951 def : Pat<(v2i64 (Neon_Op (v2i64 VPR128:$src), (v2i64 VPR128:$Rn))),
7952 (v2i64 (!cast<Instruction>(Prefix # 2d)
7953 (v2i64 VPR128:$src), (v2i64 VPR128:$Rn)))>;
7955 def : Pat<(v8i8 (Neon_Op (v8i8 VPR64:$src), (v8i8 VPR64:$Rn))),
7956 (v8i8 (!cast<Instruction>(Prefix # 8b)
7957 (v8i8 VPR64:$src), (v8i8 VPR64:$Rn)))>;
7959 def : Pat<(v4i16 (Neon_Op (v4i16 VPR64:$src), (v4i16 VPR64:$Rn))),
7960 (v4i16 (!cast<Instruction>(Prefix # 4h)
7961 (v4i16 VPR64:$src), (v4i16 VPR64:$Rn)))>;
7963 def : Pat<(v2i32 (Neon_Op (v2i32 VPR64:$src), (v2i32 VPR64:$Rn))),
7964 (v2i32 (!cast<Instruction>(Prefix # 2s)
7965 (v2i32 VPR64:$src), (v2i32 VPR64:$Rn)))>;
7968 defm : NeonI_2VMisc_BHSD_2Args_Pattern<"SUQADD", int_aarch64_neon_suqadd>;
7969 defm : NeonI_2VMisc_BHSD_2Args_Pattern<"USQADD", int_aarch64_neon_usqadd>;
7971 multiclass NeonI_2VMisc_BHSsizes<string asmop, bit U,
7972 SDPatternOperator Neon_Op> {
7973 def 16b : NeonI_2VMisc<0b1, U, 0b00, 0b00100,
7974 (outs VPR128:$Rd), (ins VPR128:$Rn),
7975 asmop # "\t$Rd.16b, $Rn.16b",
7976 [(set (v16i8 VPR128:$Rd),
7977 (v16i8 (Neon_Op (v16i8 VPR128:$Rn))))],
7980 def 8h : NeonI_2VMisc<0b1, U, 0b01, 0b00100,
7981 (outs VPR128:$Rd), (ins VPR128:$Rn),
7982 asmop # "\t$Rd.8h, $Rn.8h",
7983 [(set (v8i16 VPR128:$Rd),
7984 (v8i16 (Neon_Op (v8i16 VPR128:$Rn))))],
7987 def 4s : NeonI_2VMisc<0b1, U, 0b10, 0b00100,
7988 (outs VPR128:$Rd), (ins VPR128:$Rn),
7989 asmop # "\t$Rd.4s, $Rn.4s",
7990 [(set (v4i32 VPR128:$Rd),
7991 (v4i32 (Neon_Op (v4i32 VPR128:$Rn))))],
7994 def 8b : NeonI_2VMisc<0b0, U, 0b00, 0b00100,
7995 (outs VPR64:$Rd), (ins VPR64:$Rn),
7996 asmop # "\t$Rd.8b, $Rn.8b",
7997 [(set (v8i8 VPR64:$Rd),
7998 (v8i8 (Neon_Op (v8i8 VPR64:$Rn))))],
8001 def 4h : NeonI_2VMisc<0b0, U, 0b01, 0b00100,
8002 (outs VPR64:$Rd), (ins VPR64:$Rn),
8003 asmop # "\t$Rd.4h, $Rn.4h",
8004 [(set (v4i16 VPR64:$Rd),
8005 (v4i16 (Neon_Op (v4i16 VPR64:$Rn))))],
8008 def 2s : NeonI_2VMisc<0b0, U, 0b10, 0b00100,
8009 (outs VPR64:$Rd), (ins VPR64:$Rn),
8010 asmop # "\t$Rd.2s, $Rn.2s",
8011 [(set (v2i32 VPR64:$Rd),
8012 (v2i32 (Neon_Op (v2i32 VPR64:$Rn))))],
8016 defm CLS : NeonI_2VMisc_BHSsizes<"cls", 0b0, int_arm_neon_vcls>;
8017 defm CLZ : NeonI_2VMisc_BHSsizes<"clz", 0b1, ctlz>;
8019 multiclass NeonI_2VMisc_Bsize<string asmop, bit U, bits<2> size,
8021 def 16b : NeonI_2VMisc<0b1, U, size, Opcode,
8022 (outs VPR128:$Rd), (ins VPR128:$Rn),
8023 asmop # "\t$Rd.16b, $Rn.16b",
8026 def 8b : NeonI_2VMisc<0b0, U, size, Opcode,
8027 (outs VPR64:$Rd), (ins VPR64:$Rn),
8028 asmop # "\t$Rd.8b, $Rn.8b",
8032 defm CNT : NeonI_2VMisc_Bsize<"cnt", 0b0, 0b00, 0b00101>;
8033 defm NOT : NeonI_2VMisc_Bsize<"not", 0b1, 0b00, 0b00101>;
8034 defm RBIT : NeonI_2VMisc_Bsize<"rbit", 0b1, 0b01, 0b00101>;
8036 def : NeonInstAlias<"mvn $Rd.16b, $Rn.16b",
8037 (NOT16b VPR128:$Rd, VPR128:$Rn), 0>;
8038 def : NeonInstAlias<"mvn $Rd.8b, $Rn.8b",
8039 (NOT8b VPR64:$Rd, VPR64:$Rn), 0>;
8041 def : Pat<(v16i8 (ctpop (v16i8 VPR128:$Rn))),
8042 (v16i8 (CNT16b (v16i8 VPR128:$Rn)))>;
8043 def : Pat<(v8i8 (ctpop (v8i8 VPR64:$Rn))),
8044 (v8i8 (CNT8b (v8i8 VPR64:$Rn)))>;
8046 def : Pat<(v16i8 (xor
8048 (v16i8 Neon_AllOne))),
8049 (v16i8 (NOT16b (v16i8 VPR128:$Rn)))>;
8050 def : Pat<(v8i8 (xor
8052 (v8i8 Neon_AllOne))),
8053 (v8i8 (NOT8b (v8i8 VPR64:$Rn)))>;
8054 def : Pat<(v8i16 (xor
8056 (v8i16 (bitconvert (v16i8 Neon_AllOne))))),
8057 (NOT16b VPR128:$Rn)>;
8058 def : Pat<(v4i16 (xor
8060 (v4i16 (bitconvert (v8i8 Neon_AllOne))))),
8062 def : Pat<(v4i32 (xor
8064 (v4i32 (bitconvert (v16i8 Neon_AllOne))))),
8065 (NOT16b VPR128:$Rn)>;
8066 def : Pat<(v2i32 (xor
8068 (v2i32 (bitconvert (v8i8 Neon_AllOne))))),
8070 def : Pat<(v2i64 (xor
8072 (v2i64 (bitconvert (v16i8 Neon_AllOne))))),
8073 (NOT16b VPR128:$Rn)>;
8075 def : Pat<(v16i8 (int_aarch64_neon_rbit (v16i8 VPR128:$Rn))),
8076 (v16i8 (RBIT16b (v16i8 VPR128:$Rn)))>;
8077 def : Pat<(v8i8 (int_aarch64_neon_rbit (v8i8 VPR64:$Rn))),
8078 (v8i8 (RBIT8b (v8i8 VPR64:$Rn)))>;
8080 multiclass NeonI_2VMisc_SDsizes<string asmop, bit U, bits<5> opcode,
8081 SDPatternOperator Neon_Op> {
8082 def 4s : NeonI_2VMisc<0b1, U, 0b10, opcode,
8083 (outs VPR128:$Rd), (ins VPR128:$Rn),
8084 asmop # "\t$Rd.4s, $Rn.4s",
8085 [(set (v4f32 VPR128:$Rd),
8086 (v4f32 (Neon_Op (v4f32 VPR128:$Rn))))],
8089 def 2d : NeonI_2VMisc<0b1, U, 0b11, opcode,
8090 (outs VPR128:$Rd), (ins VPR128:$Rn),
8091 asmop # "\t$Rd.2d, $Rn.2d",
8092 [(set (v2f64 VPR128:$Rd),
8093 (v2f64 (Neon_Op (v2f64 VPR128:$Rn))))],
8096 def 2s : NeonI_2VMisc<0b0, U, 0b10, opcode,
8097 (outs VPR64:$Rd), (ins VPR64:$Rn),
8098 asmop # "\t$Rd.2s, $Rn.2s",
8099 [(set (v2f32 VPR64:$Rd),
8100 (v2f32 (Neon_Op (v2f32 VPR64:$Rn))))],
8104 defm FABS : NeonI_2VMisc_SDsizes<"fabs", 0b0, 0b01111, fabs>;
8105 defm FNEG : NeonI_2VMisc_SDsizes<"fneg", 0b1, 0b01111, fneg>;
8107 multiclass NeonI_2VMisc_HSD_Narrow<string asmop, bit U, bits<5> opcode> {
8108 def 8h8b : NeonI_2VMisc<0b0, U, 0b00, opcode,
8109 (outs VPR64:$Rd), (ins VPR128:$Rn),
8110 asmop # "\t$Rd.8b, $Rn.8h",
8113 def 4s4h : NeonI_2VMisc<0b0, U, 0b01, opcode,
8114 (outs VPR64:$Rd), (ins VPR128:$Rn),
8115 asmop # "\t$Rd.4h, $Rn.4s",
8118 def 2d2s : NeonI_2VMisc<0b0, U, 0b10, opcode,
8119 (outs VPR64:$Rd), (ins VPR128:$Rn),
8120 asmop # "\t$Rd.2s, $Rn.2d",
8123 let Constraints = "$Rd = $src" in {
8124 def 8h16b : NeonI_2VMisc<0b1, U, 0b00, opcode,
8125 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
8126 asmop # "2\t$Rd.16b, $Rn.8h",
8129 def 4s8h : NeonI_2VMisc<0b1, U, 0b01, opcode,
8130 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
8131 asmop # "2\t$Rd.8h, $Rn.4s",
8134 def 2d4s : NeonI_2VMisc<0b1, U, 0b10, opcode,
8135 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
8136 asmop # "2\t$Rd.4s, $Rn.2d",
8141 defm XTN : NeonI_2VMisc_HSD_Narrow<"xtn", 0b0, 0b10010>;
8142 defm SQXTUN : NeonI_2VMisc_HSD_Narrow<"sqxtun", 0b1, 0b10010>;
8143 defm SQXTN : NeonI_2VMisc_HSD_Narrow<"sqxtn", 0b0, 0b10100>;
8144 defm UQXTN : NeonI_2VMisc_HSD_Narrow<"uqxtn", 0b1, 0b10100>;
8146 multiclass NeonI_2VMisc_Narrow_Patterns<string Prefix,
8147 SDPatternOperator Neon_Op> {
8148 def : Pat<(v8i8 (Neon_Op (v8i16 VPR128:$Rn))),
8149 (v8i8 (!cast<Instruction>(Prefix # 8h8b) (v8i16 VPR128:$Rn)))>;
8151 def : Pat<(v4i16 (Neon_Op (v4i32 VPR128:$Rn))),
8152 (v4i16 (!cast<Instruction>(Prefix # 4s4h) (v4i32 VPR128:$Rn)))>;
8154 def : Pat<(v2i32 (Neon_Op (v2i64 VPR128:$Rn))),
8155 (v2i32 (!cast<Instruction>(Prefix # 2d2s) (v2i64 VPR128:$Rn)))>;
8157 def : Pat<(v16i8 (concat_vectors
8159 (v8i8 (Neon_Op (v8i16 VPR128:$Rn))))),
8160 (!cast<Instruction>(Prefix # 8h16b)
8161 (SUBREG_TO_REG (i32 0), VPR64:$src, sub_64),
8164 def : Pat<(v8i16 (concat_vectors
8166 (v4i16 (Neon_Op (v4i32 VPR128:$Rn))))),
8167 (!cast<Instruction>(Prefix # 4s8h)
8168 (SUBREG_TO_REG (i32 0), VPR64:$src, sub_64),
8171 def : Pat<(v4i32 (concat_vectors
8173 (v2i32 (Neon_Op (v2i64 VPR128:$Rn))))),
8174 (!cast<Instruction>(Prefix # 2d4s)
8175 (SUBREG_TO_REG (i32 0), VPR64:$src, sub_64),
8179 defm : NeonI_2VMisc_Narrow_Patterns<"XTN", trunc>;
8180 defm : NeonI_2VMisc_Narrow_Patterns<"SQXTUN", int_arm_neon_vqmovnsu>;
8181 defm : NeonI_2VMisc_Narrow_Patterns<"SQXTN", int_arm_neon_vqmovns>;
8182 defm : NeonI_2VMisc_Narrow_Patterns<"UQXTN", int_arm_neon_vqmovnu>;
8184 multiclass NeonI_2VMisc_SHIFT<string asmop, bit U, bits<5> opcode> {
8185 let DecoderMethod = "DecodeSHLLInstruction" in {
8186 def 8b8h : NeonI_2VMisc<0b0, U, 0b00, opcode,
8188 (ins VPR64:$Rn, uimm_exact8:$Imm),
8189 asmop # "\t$Rd.8h, $Rn.8b, $Imm",
8192 def 4h4s : NeonI_2VMisc<0b0, U, 0b01, opcode,
8194 (ins VPR64:$Rn, uimm_exact16:$Imm),
8195 asmop # "\t$Rd.4s, $Rn.4h, $Imm",
8198 def 2s2d : NeonI_2VMisc<0b0, U, 0b10, opcode,
8200 (ins VPR64:$Rn, uimm_exact32:$Imm),
8201 asmop # "\t$Rd.2d, $Rn.2s, $Imm",
8204 def 16b8h : NeonI_2VMisc<0b1, U, 0b00, opcode,
8206 (ins VPR128:$Rn, uimm_exact8:$Imm),
8207 asmop # "2\t$Rd.8h, $Rn.16b, $Imm",
8210 def 8h4s : NeonI_2VMisc<0b1, U, 0b01, opcode,
8212 (ins VPR128:$Rn, uimm_exact16:$Imm),
8213 asmop # "2\t$Rd.4s, $Rn.8h, $Imm",
8216 def 4s2d : NeonI_2VMisc<0b1, U, 0b10, opcode,
8218 (ins VPR128:$Rn, uimm_exact32:$Imm),
8219 asmop # "2\t$Rd.2d, $Rn.4s, $Imm",
8224 defm SHLL : NeonI_2VMisc_SHIFT<"shll", 0b1, 0b10011>;
8226 class NeonI_SHLL_Patterns<ValueType OpTy, ValueType DesTy,
8227 SDPatternOperator ExtOp, Operand Neon_Imm,
8230 (DesTy (ExtOp (OpTy VPR64:$Rn))),
8232 (i32 Neon_Imm:$Imm))))),
8233 (!cast<Instruction>("SHLL" # suffix) VPR64:$Rn, Neon_Imm:$Imm)>;
8235 class NeonI_SHLL_High_Patterns<ValueType OpTy, ValueType DesTy,
8236 SDPatternOperator ExtOp, Operand Neon_Imm,
8237 string suffix, PatFrag GetHigh>
8240 (OpTy (GetHigh VPR128:$Rn)))),
8242 (i32 Neon_Imm:$Imm))))),
8243 (!cast<Instruction>("SHLL" # suffix) VPR128:$Rn, Neon_Imm:$Imm)>;
8245 def : NeonI_SHLL_Patterns<v8i8, v8i16, zext, uimm_exact8, "8b8h">;
8246 def : NeonI_SHLL_Patterns<v8i8, v8i16, sext, uimm_exact8, "8b8h">;
8247 def : NeonI_SHLL_Patterns<v4i16, v4i32, zext, uimm_exact16, "4h4s">;
8248 def : NeonI_SHLL_Patterns<v4i16, v4i32, sext, uimm_exact16, "4h4s">;
8249 def : NeonI_SHLL_Patterns<v2i32, v2i64, zext, uimm_exact32, "2s2d">;
8250 def : NeonI_SHLL_Patterns<v2i32, v2i64, sext, uimm_exact32, "2s2d">;
8251 def : NeonI_SHLL_High_Patterns<v8i8, v8i16, zext, uimm_exact8, "16b8h",
8253 def : NeonI_SHLL_High_Patterns<v8i8, v8i16, sext, uimm_exact8, "16b8h",
8255 def : NeonI_SHLL_High_Patterns<v4i16, v4i32, zext, uimm_exact16, "8h4s",
8257 def : NeonI_SHLL_High_Patterns<v4i16, v4i32, sext, uimm_exact16, "8h4s",
8259 def : NeonI_SHLL_High_Patterns<v2i32, v2i64, zext, uimm_exact32, "4s2d",
8261 def : NeonI_SHLL_High_Patterns<v2i32, v2i64, sext, uimm_exact32, "4s2d",
8264 multiclass NeonI_2VMisc_SD_Narrow<string asmop, bit U, bits<5> opcode> {
8265 def 4s4h : NeonI_2VMisc<0b0, U, 0b00, opcode,
8266 (outs VPR64:$Rd), (ins VPR128:$Rn),
8267 asmop # "\t$Rd.4h, $Rn.4s",
8270 def 2d2s : NeonI_2VMisc<0b0, U, 0b01, opcode,
8271 (outs VPR64:$Rd), (ins VPR128:$Rn),
8272 asmop # "\t$Rd.2s, $Rn.2d",
8275 let Constraints = "$src = $Rd" in {
8276 def 4s8h : NeonI_2VMisc<0b1, U, 0b00, opcode,
8277 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
8278 asmop # "2\t$Rd.8h, $Rn.4s",
8281 def 2d4s : NeonI_2VMisc<0b1, U, 0b01, opcode,
8282 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
8283 asmop # "2\t$Rd.4s, $Rn.2d",
8288 defm FCVTN : NeonI_2VMisc_SD_Narrow<"fcvtn", 0b0, 0b10110>;
8290 multiclass NeonI_2VMisc_Narrow_Pattern<string prefix,
8291 SDPatternOperator f32_to_f16_Op,
8292 SDPatternOperator f64_to_f32_Op> {
8294 def : Pat<(v4i16 (f32_to_f16_Op (v4f32 VPR128:$Rn))),
8295 (!cast<Instruction>(prefix # "4s4h") (v4f32 VPR128:$Rn))>;
8297 def : Pat<(v8i16 (concat_vectors
8299 (v4i16 (f32_to_f16_Op (v4f32 VPR128:$Rn))))),
8300 (!cast<Instruction>(prefix # "4s8h")
8301 (v4f32 (SUBREG_TO_REG (i32 0), VPR64:$src, sub_64)),
8302 (v4f32 VPR128:$Rn))>;
8304 def : Pat<(v2f32 (f64_to_f32_Op (v2f64 VPR128:$Rn))),
8305 (!cast<Instruction>(prefix # "2d2s") (v2f64 VPR128:$Rn))>;
8307 def : Pat<(v4f32 (concat_vectors
8309 (v2f32 (f64_to_f32_Op (v2f64 VPR128:$Rn))))),
8310 (!cast<Instruction>(prefix # "2d4s")
8311 (v4f32 (SUBREG_TO_REG (i32 0), VPR64:$src, sub_64)),
8312 (v2f64 VPR128:$Rn))>;
8315 defm : NeonI_2VMisc_Narrow_Pattern<"FCVTN", int_arm_neon_vcvtfp2hf, fround>;
8317 multiclass NeonI_2VMisc_D_Narrow<string asmop, string prefix, bit U,
8319 def 2d2s : NeonI_2VMisc<0b0, U, 0b01, opcode,
8320 (outs VPR64:$Rd), (ins VPR128:$Rn),
8321 asmop # "\t$Rd.2s, $Rn.2d",
8324 def 2d4s : NeonI_2VMisc<0b1, U, 0b01, opcode,
8325 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
8326 asmop # "2\t$Rd.4s, $Rn.2d",
8328 let Constraints = "$src = $Rd";
8331 def : Pat<(v2f32 (int_aarch64_neon_vcvtxn (v2f64 VPR128:$Rn))),
8332 (!cast<Instruction>(prefix # "2d2s") VPR128:$Rn)>;
8334 def : Pat<(v4f32 (concat_vectors
8336 (v2f32 (int_aarch64_neon_vcvtxn (v2f64 VPR128:$Rn))))),
8337 (!cast<Instruction>(prefix # "2d4s")
8338 (v4f32 (SUBREG_TO_REG (i32 0), VPR64:$src, sub_64)),
8342 defm FCVTXN : NeonI_2VMisc_D_Narrow<"fcvtxn","FCVTXN", 0b1, 0b10110>;
8344 def Neon_High4Float : PatFrag<(ops node:$in),
8345 (extract_subvector (v4f32 node:$in), (iPTR 2))>;
8347 multiclass NeonI_2VMisc_HS_Extend<string asmop, bit U, bits<5> opcode> {
8348 def 4h4s : NeonI_2VMisc<0b0, U, 0b00, opcode,
8349 (outs VPR128:$Rd), (ins VPR64:$Rn),
8350 asmop # "\t$Rd.4s, $Rn.4h",
8353 def 2s2d : NeonI_2VMisc<0b0, U, 0b01, opcode,
8354 (outs VPR128:$Rd), (ins VPR64:$Rn),
8355 asmop # "\t$Rd.2d, $Rn.2s",
8358 def 8h4s : NeonI_2VMisc<0b1, U, 0b00, opcode,
8359 (outs VPR128:$Rd), (ins VPR128:$Rn),
8360 asmop # "2\t$Rd.4s, $Rn.8h",
8363 def 4s2d : NeonI_2VMisc<0b1, U, 0b01, opcode,
8364 (outs VPR128:$Rd), (ins VPR128:$Rn),
8365 asmop # "2\t$Rd.2d, $Rn.4s",
8369 defm FCVTL : NeonI_2VMisc_HS_Extend<"fcvtl", 0b0, 0b10111>;
8371 multiclass NeonI_2VMisc_Extend_Pattern<string prefix> {
8372 def : Pat<(v4f32 (int_arm_neon_vcvthf2fp (v4i16 VPR64:$Rn))),
8373 (!cast<Instruction>(prefix # "4h4s") VPR64:$Rn)>;
8375 def : Pat<(v4f32 (int_arm_neon_vcvthf2fp
8377 (v8i16 VPR128:$Rn))))),
8378 (!cast<Instruction>(prefix # "8h4s") VPR128:$Rn)>;
8380 def : Pat<(v2f64 (fextend (v2f32 VPR64:$Rn))),
8381 (!cast<Instruction>(prefix # "2s2d") VPR64:$Rn)>;
8383 def : Pat<(v2f64 (fextend
8384 (v2f32 (Neon_High4Float
8385 (v4f32 VPR128:$Rn))))),
8386 (!cast<Instruction>(prefix # "4s2d") VPR128:$Rn)>;
8389 defm : NeonI_2VMisc_Extend_Pattern<"FCVTL">;
8391 multiclass NeonI_2VMisc_SD_Conv<string asmop, bit Size, bit U, bits<5> opcode,
8392 ValueType ResTy4s, ValueType OpTy4s,
8393 ValueType ResTy2d, ValueType OpTy2d,
8394 ValueType ResTy2s, ValueType OpTy2s,
8395 SDPatternOperator Neon_Op> {
8397 def 4s : NeonI_2VMisc<0b1, U, {Size, 0b0}, opcode,
8398 (outs VPR128:$Rd), (ins VPR128:$Rn),
8399 asmop # "\t$Rd.4s, $Rn.4s",
8400 [(set (ResTy4s VPR128:$Rd),
8401 (ResTy4s (Neon_Op (OpTy4s VPR128:$Rn))))],
8404 def 2d : NeonI_2VMisc<0b1, U, {Size, 0b1}, opcode,
8405 (outs VPR128:$Rd), (ins VPR128:$Rn),
8406 asmop # "\t$Rd.2d, $Rn.2d",
8407 [(set (ResTy2d VPR128:$Rd),
8408 (ResTy2d (Neon_Op (OpTy2d VPR128:$Rn))))],
8411 def 2s : NeonI_2VMisc<0b0, U, {Size, 0b0}, opcode,
8412 (outs VPR64:$Rd), (ins VPR64:$Rn),
8413 asmop # "\t$Rd.2s, $Rn.2s",
8414 [(set (ResTy2s VPR64:$Rd),
8415 (ResTy2s (Neon_Op (OpTy2s VPR64:$Rn))))],
8419 multiclass NeonI_2VMisc_fp_to_int<string asmop, bit Size, bit U,
8420 bits<5> opcode, SDPatternOperator Neon_Op> {
8421 defm _ : NeonI_2VMisc_SD_Conv<asmop, Size, U, opcode, v4i32, v4f32, v2i64,
8422 v2f64, v2i32, v2f32, Neon_Op>;
8425 defm FCVTNS : NeonI_2VMisc_fp_to_int<"fcvtns", 0b0, 0b0, 0b11010,
8426 int_arm_neon_vcvtns>;
8427 defm FCVTNU : NeonI_2VMisc_fp_to_int<"fcvtnu", 0b0, 0b1, 0b11010,
8428 int_arm_neon_vcvtnu>;
8429 defm FCVTPS : NeonI_2VMisc_fp_to_int<"fcvtps", 0b1, 0b0, 0b11010,
8430 int_arm_neon_vcvtps>;
8431 defm FCVTPU : NeonI_2VMisc_fp_to_int<"fcvtpu", 0b1, 0b1, 0b11010,
8432 int_arm_neon_vcvtpu>;
8433 defm FCVTMS : NeonI_2VMisc_fp_to_int<"fcvtms", 0b0, 0b0, 0b11011,
8434 int_arm_neon_vcvtms>;
8435 defm FCVTMU : NeonI_2VMisc_fp_to_int<"fcvtmu", 0b0, 0b1, 0b11011,
8436 int_arm_neon_vcvtmu>;
8437 defm FCVTZS : NeonI_2VMisc_fp_to_int<"fcvtzs", 0b1, 0b0, 0b11011, fp_to_sint>;
8438 defm FCVTZU : NeonI_2VMisc_fp_to_int<"fcvtzu", 0b1, 0b1, 0b11011, fp_to_uint>;
8439 defm FCVTAS : NeonI_2VMisc_fp_to_int<"fcvtas", 0b0, 0b0, 0b11100,
8440 int_arm_neon_vcvtas>;
8441 defm FCVTAU : NeonI_2VMisc_fp_to_int<"fcvtau", 0b0, 0b1, 0b11100,
8442 int_arm_neon_vcvtau>;
8444 multiclass NeonI_2VMisc_int_to_fp<string asmop, bit Size, bit U,
8445 bits<5> opcode, SDPatternOperator Neon_Op> {
8446 defm _ : NeonI_2VMisc_SD_Conv<asmop, Size, U, opcode, v4f32, v4i32, v2f64,
8447 v2i64, v2f32, v2i32, Neon_Op>;
8450 defm SCVTF : NeonI_2VMisc_int_to_fp<"scvtf", 0b0, 0b0, 0b11101, sint_to_fp>;
8451 defm UCVTF : NeonI_2VMisc_int_to_fp<"ucvtf", 0b0, 0b1, 0b11101, uint_to_fp>;
8453 multiclass NeonI_2VMisc_fp_to_fp<string asmop, bit Size, bit U,
8454 bits<5> opcode, SDPatternOperator Neon_Op> {
8455 defm _ : NeonI_2VMisc_SD_Conv<asmop, Size, U, opcode, v4f32, v4f32, v2f64,
8456 v2f64, v2f32, v2f32, Neon_Op>;
8459 defm FRINTN : NeonI_2VMisc_fp_to_fp<"frintn", 0b0, 0b0, 0b11000,
8460 int_aarch64_neon_frintn>;
8461 defm FRINTA : NeonI_2VMisc_fp_to_fp<"frinta", 0b0, 0b1, 0b11000, frnd>;
8462 defm FRINTP : NeonI_2VMisc_fp_to_fp<"frintp", 0b1, 0b0, 0b11000, fceil>;
8463 defm FRINTM : NeonI_2VMisc_fp_to_fp<"frintm", 0b0, 0b0, 0b11001, ffloor>;
8464 defm FRINTX : NeonI_2VMisc_fp_to_fp<"frintx", 0b0, 0b1, 0b11001, frint>;
8465 defm FRINTZ : NeonI_2VMisc_fp_to_fp<"frintz", 0b1, 0b0, 0b11001, ftrunc>;
8466 defm FRINTI : NeonI_2VMisc_fp_to_fp<"frinti", 0b1, 0b1, 0b11001, fnearbyint>;
8467 defm FRECPE : NeonI_2VMisc_fp_to_fp<"frecpe", 0b1, 0b0, 0b11101,
8468 int_arm_neon_vrecpe>;
8469 defm FRSQRTE : NeonI_2VMisc_fp_to_fp<"frsqrte", 0b1, 0b1, 0b11101,
8470 int_arm_neon_vrsqrte>;
8471 defm FSQRT : NeonI_2VMisc_fp_to_fp<"fsqrt", 0b1, 0b1, 0b11111, fsqrt>;
8473 multiclass NeonI_2VMisc_S_Conv<string asmop, bit Size, bit U,
8474 bits<5> opcode, SDPatternOperator Neon_Op> {
8475 def 4s : NeonI_2VMisc<0b1, U, {Size, 0b0}, opcode,
8476 (outs VPR128:$Rd), (ins VPR128:$Rn),
8477 asmop # "\t$Rd.4s, $Rn.4s",
8478 [(set (v4i32 VPR128:$Rd),
8479 (v4i32 (Neon_Op (v4i32 VPR128:$Rn))))],
8482 def 2s : NeonI_2VMisc<0b0, U, {Size, 0b0}, opcode,
8483 (outs VPR64:$Rd), (ins VPR64:$Rn),
8484 asmop # "\t$Rd.2s, $Rn.2s",
8485 [(set (v2i32 VPR64:$Rd),
8486 (v2i32 (Neon_Op (v2i32 VPR64:$Rn))))],
8490 defm URECPE : NeonI_2VMisc_S_Conv<"urecpe", 0b1, 0b0, 0b11100,
8491 int_arm_neon_vrecpe>;
8492 defm URSQRTE : NeonI_2VMisc_S_Conv<"ursqrte", 0b1, 0b1, 0b11100,
8493 int_arm_neon_vrsqrte>;
8496 class NeonI_Cryptoaes_2v<bits<2> size, bits<5> opcode,
8497 string asmop, SDPatternOperator opnode>
8498 : NeonI_Crypto_AES<size, opcode,
8499 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
8500 asmop # "\t$Rd.16b, $Rn.16b",
8501 [(set (v16i8 VPR128:$Rd),
8502 (v16i8 (opnode (v16i8 VPR128:$src),
8503 (v16i8 VPR128:$Rn))))],
8505 let Constraints = "$src = $Rd";
8506 let Predicates = [HasNEON, HasCrypto];
8509 def AESE : NeonI_Cryptoaes_2v<0b00, 0b00100, "aese", int_arm_neon_aese>;
8510 def AESD : NeonI_Cryptoaes_2v<0b00, 0b00101, "aesd", int_arm_neon_aesd>;
8512 class NeonI_Cryptoaes<bits<2> size, bits<5> opcode,
8513 string asmop, SDPatternOperator opnode>
8514 : NeonI_Crypto_AES<size, opcode,
8515 (outs VPR128:$Rd), (ins VPR128:$Rn),
8516 asmop # "\t$Rd.16b, $Rn.16b",
8517 [(set (v16i8 VPR128:$Rd),
8518 (v16i8 (opnode (v16i8 VPR128:$Rn))))],
8521 def AESMC : NeonI_Cryptoaes<0b00, 0b00110, "aesmc", int_arm_neon_aesmc>;
8522 def AESIMC : NeonI_Cryptoaes<0b00, 0b00111, "aesimc", int_arm_neon_aesimc>;
8524 class NeonI_Cryptosha_vv<bits<2> size, bits<5> opcode,
8525 string asmop, SDPatternOperator opnode>
8526 : NeonI_Crypto_SHA<size, opcode,
8527 (outs VPR128:$Rd), (ins VPR128:$src, VPR128:$Rn),
8528 asmop # "\t$Rd.4s, $Rn.4s",
8529 [(set (v4i32 VPR128:$Rd),
8530 (v4i32 (opnode (v4i32 VPR128:$src),
8531 (v4i32 VPR128:$Rn))))],
8533 let Constraints = "$src = $Rd";
8534 let Predicates = [HasNEON, HasCrypto];
8537 def SHA1SU1 : NeonI_Cryptosha_vv<0b00, 0b00001, "sha1su1",
8538 int_arm_neon_sha1su1>;
8539 def SHA256SU0 : NeonI_Cryptosha_vv<0b00, 0b00010, "sha256su0",
8540 int_arm_neon_sha256su0>;
8542 class NeonI_Cryptosha_ss<bits<2> size, bits<5> opcode,
8543 string asmop, SDPatternOperator opnode>
8544 : NeonI_Crypto_SHA<size, opcode,
8545 (outs FPR32:$Rd), (ins FPR32:$Rn),
8546 asmop # "\t$Rd, $Rn",
8547 [(set (v1i32 FPR32:$Rd),
8548 (v1i32 (opnode (v1i32 FPR32:$Rn))))],
8550 let Predicates = [HasNEON, HasCrypto];
8553 def SHA1H : NeonI_Cryptosha_ss<0b00, 0b00000, "sha1h", int_arm_neon_sha1h>;
8555 class NeonI_Cryptosha3_vvv<bits<2> size, bits<3> opcode, string asmop,
8556 SDPatternOperator opnode>
8557 : NeonI_Crypto_3VSHA<size, opcode,
8559 (ins VPR128:$src, VPR128:$Rn, VPR128:$Rm),
8560 asmop # "\t$Rd.4s, $Rn.4s, $Rm.4s",
8561 [(set (v4i32 VPR128:$Rd),
8562 (v4i32 (opnode (v4i32 VPR128:$src),
8564 (v4i32 VPR128:$Rm))))],
8566 let Constraints = "$src = $Rd";
8567 let Predicates = [HasNEON, HasCrypto];
8570 def SHA1SU0 : NeonI_Cryptosha3_vvv<0b00, 0b011, "sha1su0",
8571 int_arm_neon_sha1su0>;
8572 def SHA256SU1 : NeonI_Cryptosha3_vvv<0b00, 0b110, "sha256su1",
8573 int_arm_neon_sha256su1>;
8575 class NeonI_Cryptosha3_qqv<bits<2> size, bits<3> opcode, string asmop,
8576 SDPatternOperator opnode>
8577 : NeonI_Crypto_3VSHA<size, opcode,
8579 (ins FPR128:$src, FPR128:$Rn, VPR128:$Rm),
8580 asmop # "\t$Rd, $Rn, $Rm.4s",
8581 [(set (v4i32 FPR128:$Rd),
8582 (v4i32 (opnode (v4i32 FPR128:$src),
8584 (v4i32 VPR128:$Rm))))],
8586 let Constraints = "$src = $Rd";
8587 let Predicates = [HasNEON, HasCrypto];
8590 def SHA256H : NeonI_Cryptosha3_qqv<0b00, 0b100, "sha256h",
8591 int_arm_neon_sha256h>;
8592 def SHA256H2 : NeonI_Cryptosha3_qqv<0b00, 0b101, "sha256h2",
8593 int_arm_neon_sha256h2>;
8595 class NeonI_Cryptosha3_qsv<bits<2> size, bits<3> opcode, string asmop,
8596 SDPatternOperator opnode>
8597 : NeonI_Crypto_3VSHA<size, opcode,
8599 (ins FPR128:$src, FPR32:$Rn, VPR128:$Rm),
8600 asmop # "\t$Rd, $Rn, $Rm.4s",
8601 [(set (v4i32 FPR128:$Rd),
8602 (v4i32 (opnode (v4i32 FPR128:$src),
8604 (v4i32 VPR128:$Rm))))],
8606 let Constraints = "$src = $Rd";
8607 let Predicates = [HasNEON, HasCrypto];
8610 def SHA1C : NeonI_Cryptosha3_qsv<0b00, 0b000, "sha1c", int_aarch64_neon_sha1c>;
8611 def SHA1P : NeonI_Cryptosha3_qsv<0b00, 0b001, "sha1p", int_aarch64_neon_sha1p>;
8612 def SHA1M : NeonI_Cryptosha3_qsv<0b00, 0b010, "sha1m", int_aarch64_neon_sha1m>;
8615 // Patterns for handling half-precision values
8618 // Convert f16 value coming in as i16 value to f32
8619 def : Pat<(f32 (f16_to_f32 (i32 (and (i32 GPR32:$Rn), 65535)))),
8620 (FCVTsh (EXTRACT_SUBREG (FMOVsw GPR32:$Rn), sub_16))>;
8621 def : Pat<(f32 (f16_to_f32 (i32 (assertzext GPR32:$Rn)))),
8622 (FCVTsh (EXTRACT_SUBREG (FMOVsw GPR32:$Rn), sub_16))>;
8624 def : Pat<(f32 (f16_to_f32 (i32 (assertzext (i32 (
8625 f32_to_f16 (f32 FPR32:$Rn))))))),
8628 // Patterns for vector extract of half-precision FP value in i16 storage type
8629 def : Pat<(f32 (f16_to_f32 ( i32 (and (i32 (vector_extract
8630 (v4i16 VPR64:$Rn), neon_uimm2_bare:$Imm)), 65535)))),
8631 (FCVTsh (f16 (DUPhv_H
8632 (v8i16 (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
8633 neon_uimm2_bare:$Imm)))>;
8635 def : Pat<(f32 (f16_to_f32 ( i32 (and (i32 (vector_extract
8636 (v8i16 VPR128:$Rn), neon_uimm3_bare:$Imm)), 65535)))),
8637 (FCVTsh (f16 (DUPhv_H (v8i16 VPR128:$Rn), neon_uimm3_bare:$Imm)))>;
8639 // Patterns for vector insert of half-precision FP value 0 in i16 storage type
8640 def : Pat<(v8i16 (vector_insert (v8i16 VPR128:$Rn),
8641 (i32 (assertsext (i32 (fp_to_sint(f32 (f16_to_f32 (i32 0))))))),
8642 (neon_uimm3_bare:$Imm))),
8643 (v8i16 (INSELh (v8i16 VPR128:$Rn),
8644 (v8i16 (SUBREG_TO_REG (i64 0),
8645 (f16 (EXTRACT_SUBREG (f32 (FMOVsw (i32 WZR))), sub_16)),
8647 neon_uimm3_bare:$Imm, 0))>;
8649 def : Pat<(v4i16 (vector_insert (v4i16 VPR64:$Rn),
8650 (i32 (assertsext (i32 (fp_to_sint(f32 (f16_to_f32 (i32 0))))))),
8651 (neon_uimm2_bare:$Imm))),
8652 (v4i16 (EXTRACT_SUBREG
8654 (v8i16 (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
8655 (v8i16 (SUBREG_TO_REG (i64 0),
8656 (f16 (EXTRACT_SUBREG (f32 (FMOVsw (i32 WZR))), sub_16)),
8658 neon_uimm2_bare:$Imm, 0)),
8661 // Patterns for vector insert of half-precision FP value in i16 storage type
8662 def : Pat<(v8i16 (vector_insert (v8i16 VPR128:$Rn),
8663 (i32 (assertsext (i32 (fp_to_sint
8664 (f32 (f16_to_f32 (i32 (and (i32 GPR32:$src), 65535)))))))),
8665 (neon_uimm3_bare:$Imm))),
8666 (v8i16 (INSELh (v8i16 VPR128:$Rn),
8667 (v8i16 (SUBREG_TO_REG (i64 0),
8668 (f16 (EXTRACT_SUBREG (f32 (FMOVsw (i32 GPR32:$src))), sub_16)),
8670 neon_uimm3_bare:$Imm, 0))>;
8672 def : Pat<(v4i16 (vector_insert (v4i16 VPR64:$Rn),
8673 (i32 (assertsext (i32 (fp_to_sint
8674 (f32 (f16_to_f32 (i32 (and (i32 GPR32:$src), 65535)))))))),
8675 (neon_uimm2_bare:$Imm))),
8676 (v4i16 (EXTRACT_SUBREG
8678 (v8i16 (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
8679 (v8i16 (SUBREG_TO_REG (i64 0),
8680 (f16 (EXTRACT_SUBREG (f32 (FMOVsw (i32 GPR32:$src))), sub_16)),
8682 neon_uimm2_bare:$Imm, 0)),
8685 def : Pat<(v8i16 (vector_insert (v8i16 VPR128:$Rn),
8686 (i32 (vector_extract (v8i16 VPR128:$src), neon_uimm3_bare:$Imm2)),
8687 (neon_uimm3_bare:$Imm1))),
8688 (v8i16 (INSELh (v8i16 VPR128:$Rn), (v8i16 VPR128:$src),
8689 neon_uimm3_bare:$Imm1, neon_uimm3_bare:$Imm2))>;
8691 // Patterns for vector copy of half-precision FP value in i16 storage type
8692 def : Pat<(v8i16 (vector_insert (v8i16 VPR128:$Rn),
8693 (i32 (assertsext (i32 (fp_to_sint(f32 (f16_to_f32 (i32 (and (i32
8694 (vector_extract (v8i16 VPR128:$src), neon_uimm3_bare:$Imm2)),
8696 (neon_uimm3_bare:$Imm1))),
8697 (v8i16 (INSELh (v8i16 VPR128:$Rn), (v8i16 VPR128:$src),
8698 neon_uimm3_bare:$Imm1, neon_uimm3_bare:$Imm2))>;
8700 def : Pat<(v4i16 (vector_insert (v4i16 VPR64:$Rn),
8701 (i32 (assertsext (i32 (fp_to_sint(f32 (f16_to_f32 (i32 (and (i32
8702 (vector_extract (v4i16 VPR64:$src), neon_uimm3_bare:$Imm2)),
8704 (neon_uimm3_bare:$Imm1))),
8705 (v4i16 (EXTRACT_SUBREG
8707 (v8i16 (SUBREG_TO_REG (i64 0), VPR64:$Rn, sub_64)),
8708 (v8i16 (SUBREG_TO_REG (i64 0), VPR64:$src, sub_64)),
8709 neon_uimm3_bare:$Imm1, neon_uimm3_bare:$Imm2)),